test(tui): cover the /undo numeric-sequence form (#698 review)

Add a parseCommand("/undo 123") assertion so both the latest and numeric
<sequence> forms of the /rewind contract are exercised through the alias.

.github/ISSUE_TEMPLATE/config.yml

@@ -1,5 +1,11 @@
 blank_issues_enabled: false
 contact_links:
+  - name: Questions and setup help
+    url: https://github.com/Gitlawb/zero/discussions/categories/q-a
+    about: Ask in Discussions Q&A. Include zero --version, your OS, provider/model, and zero doctor output for setup problems.
+  - name: Feature ideas
+    url: https://github.com/Gitlawb/zero/discussions/categories/ideas
+    about: Propose and discuss ideas in Discussions before opening a PR (Zero is in controlled development, see CONTRIBUTING.md).
   - name: Security vulnerability
     url: https://github.com/Gitlawb/zero/security/advisories/new
     about: Report security issues privately via GitHub Security Advisories (see SECURITY.md). Do not file a public issue.

.github/ISSUE_TEMPLATE/feature_request.yml

@@ -1,12 +1,22 @@
-name: Feature request
-description: Propose a new capability or change
+name: Feature request (previously discussed)
+description: File only after your idea gained traction in Discussions Ideas or a maintainer asked for a tracking issue
 labels: ["enhancement"]
 body:
   - type: markdown
     attributes:
       value: |
-        Please do not begin implementation until the issue has been reviewed and marked with the
+        New feature ideas start in [Discussions Ideas](https://github.com/Gitlawb/zero/discussions/categories/ideas),
+        not here. This form is for the step after that: an idea that gained traction in Discussions,
+        or one a maintainer asked to have tracked as an issue. Please link the discussion below.
+        Do not begin implementation until the issue has been reviewed and marked with the
         `issue-approved` label (see CONTRIBUTING.md).
+  - type: input
+    id: discussion
+    attributes:
+      label: Ideas discussion
+      description: Link to the Discussions Ideas thread (or note which maintainer asked for this issue).
+    validations:
+      required: true
   - type: textarea
     id: problem
     attributes:

.github/workflows/ci.yml

@@ -100,9 +100,10 @@ jobs:
           go-version-file: go.mod
           cache: true
 
-      # govulncheck and deadcode resolve a toolchain from their own modules, which
-      # can drop below the version go.mod requires and then fail to load our
-      # packages. Pin GOTOOLCHAIN to the go.mod toolchain so both run under it.
+      # govulncheck, deadcode, and golangci-lint each resolve a toolchain from
+      # their own modules, which can drop below the version go.mod requires and
+      # then fail to load our packages. Pin GOTOOLCHAIN to the go.mod toolchain
+      # so all three run under it.
       - name: Pin toolchain from go.mod
         run: |
           toolchain="$(awk '/^toolchain /{print $2}' go.mod)"
@@ -120,3 +121,13 @@ jobs:
       - name: deadcode (advisory)
         continue-on-error: true
         run: go run golang.org/x/tools/cmd/[email protected] -test=false ./...
+
+      # Advisory: catches what deadcode's whole-program reachability analysis
+      # doesn't, unused private functions/assignments reachable within a
+      # package but never actually called, plus staticcheck-style correctness
+      # and readability issues. Scoped to a few linters rather than the full
+      # default battery, and non-blocking, while the existing findings across
+      # the repo are cleaned up incrementally (see #527).
+      - name: golangci-lint (advisory)
+        continue-on-error: true
+        run: go run github.com/golangci/golangci-lint/v2/cmd/[email protected] run --enable-only unused,ineffassign,staticcheck ./...

.github/workflows/release-artifacts.yml

@@ -20,10 +20,15 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
+        # `zero-release package` builds + smoke-tests natively (host==target), so each
+        # target platform needs its own native runner. These five cover the platforms
+        # install.sh / install.ps1 / the npm wrapper resolve.
         os:
-          - ubuntu-latest
-          - macos-latest
-          - windows-latest
+          - ubuntu-latest      # linux-x64
+          - ubuntu-24.04-arm   # linux-arm64
+          - macos-latest       # macos-arm64 (Apple Silicon)
+          - macos-15-intel     # macos-x64 (Intel)
+          - windows-latest     # windows-x64
 
     steps:
       - name: Checkout
@@ -49,8 +54,10 @@ jobs:
       # is published under the pushed git tag. If they disagree, install.sh/install.ps1
       # and the npm postinstall all build a download URL that 404s. Fail fast so a
       # forgotten version bump can never publish mis-versioned, un-installable assets.
+      # Matches on the ref (not event_name) because release-please dispatches this
+      # workflow at the tag ref — its GITHUB_TOKEN-created tag cannot fire `push`.
       - name: Check tag matches package.json version
-        if: ${{ github.event_name == 'push' && startsWith(github.ref, 'refs/tags/v') }}
+        if: ${{ startsWith(github.ref, 'refs/tags/v') }}
         shell: bash
         run: |
           pkg_version="$(node -p "require('./package.json').version")"
@@ -73,13 +80,15 @@ jobs:
           path: dist/release/*
           if-no-files-found: error
 
-      # On a `v*` tag push, publish the packaged archive + .sha256 as GitHub Release
-      # assets so the documented `scripts/install.sh` / `install.ps1` path (which reads
-      # /releases/latest) works. Uses the preinstalled `gh` (no new action dependency);
-      # each matrix OS creates the release if absent (best-effort) then uploads its own
-      # assets with --clobber. NOTE: maintainers must push the tag — this never tags.
+      # On a `v*` tag ref (pushed manually, or dispatched at the tag by the
+      # release-please workflow), publish the packaged archive + .sha256 as GitHub
+      # Release assets so the documented `scripts/install.sh` / `install.ps1` path
+      # (which reads /releases/latest) works. Uses the preinstalled `gh` (no new
+      # action dependency); each matrix OS creates the release if absent
+      # (best-effort — release-please normally creates it first) then uploads its
+      # own assets with --clobber. Tags come from merging the release-please PR.
       - name: Publish to GitHub Release
-        if: ${{ github.event_name == 'push' && startsWith(github.ref, 'refs/tags/v') }}
+        if: ${{ startsWith(github.ref, 'refs/tags/v') }}
         shell: bash
         env:
           GH_TOKEN: ${{ github.token }}
@@ -87,3 +96,147 @@ jobs:
           tag="${GITHUB_REF_NAME}"
           gh release create "$tag" --repo "$GITHUB_REPOSITORY" --title "$tag" --generate-notes 2>/dev/null || true
           gh release upload "$tag" dist/release/* --repo "$GITHUB_REPOSITORY" --clobber
+
+  # Publishes @gitlawb/zero to npm: first the five platform payloads (versions
+  # of the SAME package at suffixed versions like X.Y.Z-linux-x64, carrying the
+  # native binary + vendored helpers), then the wrapper at X.Y.Z whose
+  # optionalDependencies alias those platform versions. Platform versions MUST
+  # publish first — the wrapper pins them exactly, so the reverse order would
+  # open a window where installs resolve aliases that 404. See
+  # docs/NPM_PACKAGING.md for the full contract.
+  #
+  # Auth is npm OIDC trusted publishing: no NPM_TOKEN secret. The package's
+  # npmjs.com settings must list this repo + workflow file as a trusted
+  # publisher, and provenance is then generated automatically. Note npm only
+  # allows configuring a trusted publisher on an EXISTING package, so the very
+  # first @gitlawb/zero publish is a one-time manual `npm publish` bootstrap.
+  # Because the platform payloads are versions of the same package, they reuse
+  # that single trusted-publisher configuration.
+  publish-npm:
+    name: Publish npm package
+    needs: package
+    if: ${{ startsWith(github.ref, 'refs/tags/v') }}
+    runs-on: ubuntu-latest
+    # Manual approval gate: this environment carries a required-reviewer rule
+    # (Settings -> Environments -> npm-publish), so the job pauses until a
+    # maintainer approves it in the Actions UI before anything reaches npm.
+    environment:
+      name: npm-publish
+      url: https://www.npmjs.com/package/@gitlawb/zero
+    permissions:
+      contents: read
+      id-token: write # OIDC token exchange for npm trusted publishing + provenance
+    steps:
+      - name: Checkout
+        uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5 # v4
+
+      # No registry-url here on purpose: setup-node's registry-url writes an
+      # .npmrc line referencing $NODE_AUTH_TOKEN, and npm hard-fails on the
+      # unset env var. Trusted publishing needs no token at all.
+      - name: Setup Node
+        uses: actions/setup-node@49933ea5288caeca8642d1e84afbd3f7d6820020 # v4
+        with:
+          node-version: 24
+
+      # OIDC trusted publishing needs npm >= 11.5.1. Use the npm bundled with
+      # the pinned Node toolchain (Node 24 ships >= 11.5.1) rather than pulling
+      # npm@latest from the registry inside the publish path — an unpinned
+      # install there would be avoidable supply-chain exposure. Fail fast if a
+      # future runner/toolchain change ever regresses the bundled npm.
+      - name: Check npm supports trusted publishing
+        run: |
+          npm --version | node -e '
+            const version = require("fs").readFileSync(0, "utf8").trim();
+            const [major, minor, patch] = version.split(".").map(Number);
+            const ok =
+              major > 11 ||
+              (major === 11 && (minor > 5 || (minor === 5 && patch >= 1)));
+            if (!ok) {
+              console.error(
+                `::error::bundled npm ${version} is too old for OIDC trusted ` +
+                  `publishing (needs >= 11.5.1) — bump node-version in this job.`,
+              );
+              process.exit(1);
+            }
+            console.log(`npm ${version} supports trusted publishing.`);
+          '
+
+      # Probe the exact URLs the wrapper's first-run downloader builds
+      # (internal/release/release.go asset-name scheme) as an anonymous client
+      # — install.sh/install.ps1 and `zero update` read them too. This also
+      # blocks publishing while the repo is still private: private release
+      # assets 404 without auth, exactly as they would for an installing user.
+      - name: Verify release assets are downloadable
+        shell: bash
+        run: |
+          version="$(node -p "require('./package.json').version")"
+          base="https://github.com/${GITHUB_REPOSITORY}/releases/download/v${version}"
+          for asset in \
+            "zero-v${version}-linux-x64.tar.gz" \
+            "zero-v${version}-linux-arm64.tar.gz" \
+            "zero-v${version}-macos-x64.tar.gz" \
+            "zero-v${version}-macos-arm64.tar.gz" \
+            "zero-v${version}-windows-x64.zip"; do
+            for url in "${base}/${asset}" "${base}/${asset}.sha256"; do
+              if ! curl -fsSLI -o /dev/null "$url"; then
+                echo "::error::${url} is not publicly downloadable — the wrapper's fallback downloader and install.sh would fail. Is the repo public and are all assets uploaded?" >&2
+                exit 1
+              fi
+            done
+          done
+          echo "all release assets for v${version} are publicly downloadable."
+
+      - name: Download release artifacts
+        uses: actions/download-artifact@d3f86a106a0bac45b974a628896c90dbdf5c8093 # v4.3.0
+        with:
+          pattern: zero-*
+          path: dist/artifacts
+          merge-multiple: true
+
+      - name: Assemble npm packages
+        run: node scripts/npm/build-platform-packages.mjs --artifacts-dir dist/artifacts --out-dir dist/npm
+
+      # Platform versions are semver PRERELEASES of the wrapper version
+      # (X.Y.Z-linux-x64), so they must never publish under the default
+      # `latest` dist-tag — that would make `npm install @gitlawb/zero`
+      # deliver a platform payload as the CLI. Publish them under a dedicated
+      # tag and assert `latest` survived before the wrapper goes out.
+      - name: Publish platform packages to npm
+        shell: bash
+        run: |
+          version="$(node -p "require('./package.json').version")"
+          for dir in dist/npm/platforms/*/; do
+            echo "publishing ${dir}"
+            (cd "$dir" && npm publish --access public --tag platform)
+          done
+          # Tolerate a failed lookup (registry lag, or no latest tag yet on a
+          # bootstrap publish): under set -e a bare failure here would abort
+          # AFTER the platform payloads went out, leaving no wrapper and a
+          # re-run that fails on republishing existing versions. An empty
+          # value falls through the guard below.
+          latest="$(npm view @gitlawb/zero dist-tags.latest || true)"
+          case "$latest" in
+            "${version}-"*)
+              echo "::error::dist-tag latest points at platform version ${latest} — a platform publish clobbered it. Fix the dist-tag before the wrapper publish." >&2
+              exit 1
+              ;;
+          esac
+          echo "dist-tag latest is ${latest} (unchanged by platform publishes)."
+
+      - name: Publish wrapper to npm
+        shell: bash
+        run: |
+          version="$(node -p "require('./package.json').version")"
+          (cd dist/npm/wrapper && npm publish --access public)
+          # Confirm the wrapper owns `latest` (retry: registry metadata can lag
+          # a publish by a few seconds).
+          for attempt in 1 2 3 4 5; do
+            latest="$(npm view @gitlawb/zero dist-tags.latest || true)"
+            if [ "$latest" = "$version" ]; then
+              echo "dist-tag latest is ${version}."
+              exit 0
+            fi
+            sleep 5
+          done
+          echo "::error::dist-tag latest is ${latest}, expected ${version} after the wrapper publish." >&2
+          exit 1

.github/workflows/release-please.yml

@@ -0,0 +1,35 @@
+name: Release Please
+
+on:
+  push:
+    branches:
+      - main
+
+permissions:
+  contents: write # push the release-PR branch, create the tag + GitHub Release
+  pull-requests: write # open/update the release PR
+  issues: write # create/apply release-tracking labels on the release PR
+  actions: write # dispatch release-artifacts.yml once a release is cut
+
+jobs:
+  release-please:
+    name: Release PR / tag
+    runs-on: ubuntu-latest
+    steps:
+      - name: Run release-please
+        id: release
+        uses: googleapis/release-please-action@a02a34c4d625f9be7cb89156071d8567266a2445 # v4.2.0
+
+      # Tags created with the default GITHUB_TOKEN never fire `on: push: tags`
+      # workflows (GitHub suppresses workflows triggered by workflow-created
+      # events), so release-artifacts.yml would not run for a release-please
+      # tag. workflow_dispatch invoked through the API is exempt from that
+      # suppression — dispatch the artifacts build at the new tag explicitly.
+      - name: Build and publish release artifacts
+        if: ${{ steps.release.outputs.release_created }}
+        env:
+          GH_TOKEN: ${{ github.token }}
+        run: |
+          gh workflow run release-artifacts.yml \
+            --repo "$GITHUB_REPOSITORY" \
+            --ref "${{ steps.release.outputs.tag_name }}"

.github/workflows/zero-action-smoke.yml

@@ -4,12 +4,14 @@ on:
   pull_request:
     paths:
       - action.yml
+      - docs/GITHUB_ACTION.md
       - .github/workflows/zero-action-smoke.yml
   push:
     branches:
       - main
     paths:
       - action.yml
+      - docs/GITHUB_ACTION.md
       - .github/workflows/zero-action-smoke.yml
 
 jobs:
@@ -29,7 +31,8 @@ jobs:
           import sys, yaml
 
           with open("action.yml", "r", encoding="utf-8") as handle:
-              doc = yaml.safe_load(handle)
+              action_raw = handle.read()
+          doc = yaml.safe_load(action_raw)
 
           assert isinstance(doc, dict), "action.yml must be a mapping"
           assert doc.get("name"), "action must declare a name"
@@ -65,6 +68,19 @@ jobs:
           assert "--skip-permissions-unsafe" not in serialized, \
               "action must never pass --skip-permissions-unsafe"
 
+          run_zero = next((step for step in steps if step.get("name") == "Run ZERO"), None)
+          assert run_zero is not None, "action.yml must contain a Run ZERO step"
+          run_zero_script = run_zero.get("run") or ""
+          assert 'export "${ZERO_INPUT_API_KEY_ENV}=${ZERO_INPUT_API_KEY}"' in run_zero_script, \
+              "Run ZERO must export the provider key for the current step"
+          assert "later steps inherit" not in action_raw.lower(), \
+              "provider keys must not be documented as inherited by later steps"
+          for line in run_zero_script.splitlines():
+              if "GITHUB_ENV" in line and (
+                  "ZERO_INPUT_API_KEY" in line or "ZERO_INPUT_API_KEY_ENV" in line
+              ):
+                  raise AssertionError("provider key must not be persisted to GITHUB_ENV")
+
           print(f"action.yml OK: {len(inputs)} inputs, {len(outputs)} outputs, {len(steps)} steps")
           PY
 

.gitignore

@@ -27,6 +27,10 @@ report.[0-9]_.[0-9]_.[0-9]_.[0-9]_.json
 .cache
 *.tsbuildinfo
 
+# Go build/module caches (GOCACHE/GOMODCACHE pointed here for local runs)
+.gocache/
+.gomodcache/
+
 # IntelliJ based IDEs
 .idea
 
@@ -37,8 +41,20 @@ report.[0-9]_.[0-9]_.[0-9]_.[0-9]_.json
 
 /zero
 /zero.exe
+/zero-windows-command-runner.exe
+/zero-windows-sandbox-setup.exe
+/zero-skills
+/.zero-binary-version
 /snake-game.html
 
 # Superpowers skill artifacts (brainstorm mockups, specs, plans) — local only
 .superpowers/
 docs/superpowers/
+
+# Termux/Android build artifacts
+zero-termux-arm64
+zero-linux-sandbox
+zero-seccomp
+
+# Generated benchmark reports are evidence for one configuration, never tree state
+internal/perfbench/reports/*.json

.release-please-manifest.json

@@ -0,0 +1 @@
+{".":"0.4.0"}

AGENTS.md

@@ -0,0 +1,71 @@
+# Repository Guidelines for Zero
+
+These instructions apply to all work in this repository. For the user-facing
+guide to extending Zero with specialists, hooks, plugins, MCP, and skills, see
+[docs/EXTENDING.md](docs/EXTENDING.md).
+
+## 1. Contribution and Pull Request Rules
+
+- Before opening any pull request, **all contributors**—including maintainers,
+  community contributors, and coding agents—must read and follow
+  [CONTRIBUTING.md](CONTRIBUTING.md).
+- Community pull requests require an existing parent issue with the
+  `issue-approved` label. Team members may open pull requests through the
+  internal development process described in `CONTRIBUTING.md`.
+- Keep each change focused on the approved or assigned scope. Do not include
+  unrelated fixes, refactors, formatting churn, generated output, or existing
+  local changes in the same commit or pull request.
+- Discuss new implementation languages, runtimes, major dependency changes,
+  and broad architectural rewrites with maintainers before implementation.
+- Pull request descriptions must explain what changed and why, link the parent
+  issue when required, and list the tests or verification performed. Include
+  screenshots or a short recording for user-visible UI changes when practical.
+
+## 2. Repository and Implementation Conventions
+
+- Use the Go version declared in `go.mod`. Do not hardcode a different local
+  toolchain version in scripts or documentation.
+- Use the repository build and release commands (`make` and
+  `go run ./cmd/zero-release ...`) instead of inventing parallel build flows.
+- Keep tests beside their source files (`foo_test.go` next to `foo.go`). Add a
+  regression test for behavior changes and run affected concurrent code under
+  the race detector.
+- Never edit files under `third_party/`; they are vendored.
+- Prefer one cross-platform function with small conditional checks over
+  duplicated platform-specific helpers when the behavior can remain unified.
+- Do not commit generated benchmark reports from
+  `internal/perfbench/reports/*.json`; reports are configuration-specific
+  evidence, not repository state.
+- Preserve the user's working tree. Do not overwrite, delete, stage, or commit
+  unrelated tracked or untracked files.
+
+## 3. Required Validation
+
+Run validation from the repository root before committing, opening a pull
+request, or completing an implementation task:
+
+1. **Formatting check**: `make fmt-check`. If it fails, format with
+   `go fmt ./...` (or `make fmt`) and run the check again.
+2. **Vet**: `go vet ./...` (or `make vet`).
+3. **Tests**: `go test ./...`. Use `make test` for the full race-enabled suite,
+   or run focused tests with `-race`, when concurrency is affected.
+4. **Build**: `go run ./cmd/zero-release build`.
+5. **Smoke test**: `go run ./cmd/zero-release smoke`.
+6. **Advisory lint**:
+   `go run github.com/golangci/golangci-lint/v2/cmd/[email protected] run --enable-only unused,ineffassign,staticcheck ./...`.
+7. **Security**:
+   `go run golang.org/x/vuln/cmd/[email protected] ./...`.
+8. **Diff hygiene**: `git diff HEAD --check` (covers staged and unstaged
+   tracked changes).
+
+`make lint` currently runs the formatting check and `go vet`; it does **not**
+run golangci-lint. The pinned golangci-lint job is advisory in CI while the
+existing repository-wide backlog is cleaned up. Fix findings introduced by or
+related to the current change. Unrelated pre-existing advisory findings do not
+justify expanding a focused pull request; report them separately.
+
+Formatting, vet, tests, build, smoke, diff hygiene, and govulncheck are hard
+requirements. If a related check fails, fix it. If a required check cannot run
+because of the environment or fails for an unrelated external reason, report
+the exact failure and obtain maintainer direction rather than silently ignoring
+it.

CHANGELOG.md

@@ -6,12 +6,196 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 aims to follow [Semantic Versioning](https://semver.org/spec/v2.0.0.html) once the first release is
 tagged. Until then, source builds report the version `dev`.
 
+## [0.4.0](https://github.com/Gitlawb/zero/compare/v0.3.0...v0.4.0) (2026-07-17)
+
+
+### Features
+
+* **aimlapi:** AI/ML API provider with guided onboarding (top-up + key issuance) ([#655](https://github.com/Gitlawb/zero/issues/655)) ([6b9c2f0](https://github.com/Gitlawb/zero/commit/6b9c2f0c083e02bcd3aad68fb7af0501a9e7cd61))
+* **cli:** show ZERO wordmark on --version ([#673](https://github.com/Gitlawb/zero/issues/673)) ([9acb411](https://github.com/Gitlawb/zero/commit/9acb4113cc3337b3f361a16278e9cf11ca105e34))
+* **cli:** wire MCP serve WorkspaceRoot and --add-dir scope ([#694](https://github.com/Gitlawb/zero/issues/694)) ([75a78e7](https://github.com/Gitlawb/zero/commit/75a78e715bf23154f69c8c79cb58eb4b535b2a2a))
+* **npm:** ship the native binary as platform optionalDependencies ([#626](https://github.com/Gitlawb/zero/issues/626)) ([5e1405d](https://github.com/Gitlawb/zero/commit/5e1405d0b7abff5b3ccb3cfdb66d64d6d3322922))
+* **perf:** emit prompt-prefix hash fingerprint per turn ([#704](https://github.com/Gitlawb/zero/issues/704)) ([1c5c6e7](https://github.com/Gitlawb/zero/commit/1c5c6e78a8a0e228bdf53d7be90934fbce9d98c3))
+* **providers:** add AI/ML API preset (rebased onto main) ([#621](https://github.com/Gitlawb/zero/issues/621)) ([d66a9dd](https://github.com/Gitlawb/zero/commit/d66a9dda69c32aa59f4bea903cfefe00d4b7adef))
+* **providers:** refresh MiniMax model coverage ([#665](https://github.com/Gitlawb/zero/issues/665)) ([fa3052a](https://github.com/Gitlawb/zero/commit/fa3052a1422a4ad30a3a6295829564f42dee31a8))
+* **skills:** discover shared ~/.agents/skills with multi-root skill loading ([#696](https://github.com/Gitlawb/zero/issues/696)) ([7d57999](https://github.com/Gitlawb/zero/commit/7d579996b43741a2e57e3e38fcdd8484fcfc34e9))
+* **tui:** Ctrl+X leader chords and emacs menu navigation ([#699](https://github.com/Gitlawb/zero/issues/699)) ([7f669f4](https://github.com/Gitlawb/zero/commit/7f669f455021be51319ee3b8298cd48a17f745c7))
+* **tui:** press up to edit queued messages ([#656](https://github.com/Gitlawb/zero/issues/656)) ([4c986d3](https://github.com/Gitlawb/zero/commit/4c986d327b5ed26338d295c23bea5839681db8d0))
+
+
+### Bug Fixes
+
+* **acp:** make truncateHint rune-safe ([#614](https://github.com/Gitlawb/zero/issues/614)) ([ddc4927](https://github.com/Gitlawb/zero/commit/ddc4927aac5544bf4dd2c46615aeda5a81c96576))
+* **agent,tui:** resolve git branch detection when starting Zero in subdirectories ([#613](https://github.com/Gitlawb/zero/issues/613)) ([0184581](https://github.com/Gitlawb/zero/commit/0184581ec234ea414e0a47bc33e8b0f4ddfb497b))
+* **agent:** raise default and deep-mode turn budgets ([#650](https://github.com/Gitlawb/zero/issues/650)) ([635c93a](https://github.com/Gitlawb/zero/commit/635c93af51ebc20f3e0917917e55a79edfe27c35))
+* **cli:** prevent consuming positional arguments as flag values ([#619](https://github.com/Gitlawb/zero/issues/619)) ([5b4f48d](https://github.com/Gitlawb/zero/commit/5b4f48d2dcb66402c13bde0c3cfe9c9371da19fb))
+* **config:** enforce MCP trust boundary so project config cannot override user disable ([#609](https://github.com/Gitlawb/zero/issues/609)) ([4d8c31c](https://github.com/Gitlawb/zero/commit/4d8c31cf16a3080344e3be7039408fcae71c075d)), closes [#512](https://github.com/Gitlawb/zero/issues/512)
+* **config:** surface unknown/typo'd config fields instead of silently dropping them ([#645](https://github.com/Gitlawb/zero/issues/645)) ([893b7b4](https://github.com/Gitlawb/zero/commit/893b7b424cc203a2fcf92327a4e25c84286a90e0))
+* **cron:** prevent cron job Mutate from clobbering concurrent updates ([#630](https://github.com/Gitlawb/zero/issues/630)) ([e4bd703](https://github.com/Gitlawb/zero/commit/e4bd703cfb28dab2dfa3c2ddba46237e1bb2e164))
+* **daemon:** handle os.ErrPermission as collision during O_EXCL lock creation ([#616](https://github.com/Gitlawb/zero/issues/616)) ([8ea5384](https://github.com/Gitlawb/zero/commit/8ea53841a5d54c37a153779ae86bea010659433c))
+* **exec:** stop false INCOMPLETE downgrades on conversational final messages ([#608](https://github.com/Gitlawb/zero/issues/608)) ([b6117af](https://github.com/Gitlawb/zero/commit/b6117af86d6bc87a4ee66910e99d76cb16b03fed))
+* harden MCP credential boundaries ([#597](https://github.com/Gitlawb/zero/issues/597)) ([fdddb05](https://github.com/Gitlawb/zero/commit/fdddb05ba84b1600ae6c3a20028bf83afe474c44))
+* **hooks:** fail closed on launch failures for beforeTool hooks ([#629](https://github.com/Gitlawb/zero/issues/629)) ([dc06fe7](https://github.com/Gitlawb/zero/commit/dc06fe72caf45f72d2cba1e8a835c0f5b405c1e8))
+* **hooks:** run sessionEnd hooks after Esc/Ctrl+C interrupts ([#606](https://github.com/Gitlawb/zero/issues/606)) ([824ecdb](https://github.com/Gitlawb/zero/commit/824ecdbcf9c467c35ef4e2666770fdadcb5bf402))
+* **keyring:** pass generic password via stdin on macOS ([#574](https://github.com/Gitlawb/zero/issues/574)) ([91ea6de](https://github.com/Gitlawb/zero/commit/91ea6ded7503538834a84d090f78670a363c62d3))
+* **lock:** prevent POSIX lock file overwrite and leak on Windows/Unix ([#628](https://github.com/Gitlawb/zero/issues/628)) ([da41c3a](https://github.com/Gitlawb/zero/commit/da41c3a75b782d6e0836fe13346321e40a90fbb4))
+* **openai:** omit prompt_cache_key for openai-compatible providers ([#636](https://github.com/Gitlawb/zero/issues/636)) ([1af5882](https://github.com/Gitlawb/zero/commit/1af58828eb3c22567599c000736c913a290959d2)), closes [#624](https://github.com/Gitlawb/zero/issues/624)
+* **plugins:** resolve relative executable paths against plugin root ([#627](https://github.com/Gitlawb/zero/issues/627)) ([2efe6d5](https://github.com/Gitlawb/zero/commit/2efe6d539e29374b3ef39c2290bdea81f33a228b))
+* **sandbox:** remove windowsWriteRestricted flag to fix DenyRead bypass ([#612](https://github.com/Gitlawb/zero/issues/612)) ([3d96ac7](https://github.com/Gitlawb/zero/commit/3d96ac7e55c760a97f28c0e6ceaf1ec3b4ab717a))
+* **sandbox:** scrub dynamic credential env vars ([#682](https://github.com/Gitlawb/zero/issues/682)) ([9043bae](https://github.com/Gitlawb/zero/commit/9043baedcff7776c7373645b57563cac06b31847))
+* **sandbox:** scrub sensitive credentials from sandbox environment ([#660](https://github.com/Gitlawb/zero/issues/660)) ([6fc1220](https://github.com/Gitlawb/zero/commit/6fc1220f6ac66fb3ae67b637cbbed7068d2213c0))
+* **sandbox:** unblock git fetch/commit/add under the write-restricted sandbox ([#654](https://github.com/Gitlawb/zero/issues/654)) ([5c4815a](https://github.com/Gitlawb/zero/commit/5c4815a66ed07d9cf90b825adfd936d3ac07639d))
+* **sandbox:** use WRITE_RESTRICTED token when no DenyRead paths are configured ([#658](https://github.com/Gitlawb/zero/issues/658)) ([a5d2e32](https://github.com/Gitlawb/zero/commit/a5d2e327c8681671aa8a9e5378801215b747edcf))
+* **securefile,credstore:** call Sync on temp file before close and rename ([#631](https://github.com/Gitlawb/zero/issues/631)) ([212734a](https://github.com/Gitlawb/zero/commit/212734adf3b5f982e22385161205aa1afe4634fe))
+* **securefile:** reclaim stale lock files to prevent permanent DOS ([#615](https://github.com/Gitlawb/zero/issues/615)) ([8536cc8](https://github.com/Gitlawb/zero/commit/8536cc87f7a885f9e436d6ef28f5f325201623dc))
+* **swarm:** wait for job.Runs directly in scheduler skip test ([#667](https://github.com/Gitlawb/zero/issues/667)) ([1bb6b57](https://github.com/Gitlawb/zero/commit/1bb6b5745af90321d1e657a12a1976cded5dd1bd))
+* **tools:** classify silent wrapped Windows command failures as sandbox denials ([#659](https://github.com/Gitlawb/zero/issues/659)) ([8bd9742](https://github.com/Gitlawb/zero/commit/8bd9742fa95c41b93ea4e718628aed0ff3ae9dd0))
+* **tools:** preserve SysProcAttr during PTY fallback ([#618](https://github.com/Gitlawb/zero/issues/618)) ([f78b36c](https://github.com/Gitlawb/zero/commit/f78b36c770daa4577a9f99265b18a354454e36eb))
+* **tui:** resolve pending askUser callbacks to prevent runner hangs ([#620](https://github.com/Gitlawb/zero/issues/620)) ([aa73a76](https://github.com/Gitlawb/zero/commit/aa73a76f1bd1b6fe97bac2fbff2d61b7474139f2))
+* **tui:** stop the composer cursor blinking while typing or unfocused ([#672](https://github.com/Gitlawb/zero/issues/672)) ([2b42cd5](https://github.com/Gitlawb/zero/commit/2b42cd567b96d6f7e0818594e53ff31cce1e42e9))
+* **windows:** resolve absolute path for taskkill to prevent hijacking ([#617](https://github.com/Gitlawb/zero/issues/617)) ([2db00ee](https://github.com/Gitlawb/zero/commit/2db00ee3d57db97e0fbe23cb8628e8bcb47f6f09))
+
+
+### Performance Improvements
+
+* **tools:** add explicit effect metadata for safe concurrency ([#705](https://github.com/Gitlawb/zero/issues/705)) ([8ef8576](https://github.com/Gitlawb/zero/commit/8ef8576df7d0775a5b815bc0776a794cadb75c34))
+
+## [0.3.0](https://github.com/Gitlawb/zero/compare/v0.2.0...v0.3.0) (2026-07-09)
+
+
+### Features
+
+* gate project-scoped hooks, plugins, and MCP servers behind workspace trust ([#529](https://github.com/Gitlawb/zero/issues/529)) ([a880ce8](https://github.com/Gitlawb/zero/commit/a880ce80a6ec72da511fb9bdf6dd69291c72a64b))
+* **modelregistry:** infer reasoning efforts for Hunyuan and vendor-prefixed model ids ([#599](https://github.com/Gitlawb/zero/issues/599)) ([92a92ce](https://github.com/Gitlawb/zero/commit/92a92ceb29dc63f5216ab140ef9a6dd9afe17df8))
+* **providers:** OAuth login profiles and list-first /provider manager ([#560](https://github.com/Gitlawb/zero/issues/560)) ([1655056](https://github.com/Gitlawb/zero/commit/16550569c1be615cfaf244dab25909ec37f6dee6))
+* **tui:** remember recent provider+model selections in /model picker ([#568](https://github.com/Gitlawb/zero/issues/568)) ([d0c4e62](https://github.com/Gitlawb/zero/commit/d0c4e62cd429a0614c15296934c740a08bc0e07b))
+* **tui:** show CLI version on the startup home screen ([#538](https://github.com/Gitlawb/zero/issues/538)) ([fd69233](https://github.com/Gitlawb/zero/commit/fd69233e334f1823a06b5794085a9255b3abdfa8))
+* voice dictation (speech-to-text) ([#557](https://github.com/Gitlawb/zero/issues/557)) ([87158a1](https://github.com/Gitlawb/zero/commit/87158a1c90b4f91fc5f2bb8178ebaf46d7654680))
+
+
+### Bug Fixes
+
+* address bugs found in a multi-agent codebase audit ([#481](https://github.com/Gitlawb/zero/issues/481)) ([008bc9b](https://github.com/Gitlawb/zero/commit/008bc9b3f3ba13c7d4822b9559b020f381ff555b))
+* **agent:** keep tools exposed for max-turn finalization ([#533](https://github.com/Gitlawb/zero/issues/533)) ([3f0503b](https://github.com/Gitlawb/zero/commit/3f0503bc2312ae29d5ade784d8824dc9a3524958))
+* **auth:** persist OpenRouter API key after login ([#595](https://github.com/Gitlawb/zero/issues/595)) ([2a062aa](https://github.com/Gitlawb/zero/commit/2a062aa4014c6cd5e20a57dad4a685f88966f109))
+* bump Go to 1.26.5 for crypto/tls fix (GO-2026-5856) ([#607](https://github.com/Gitlawb/zero/issues/607)) ([a7cfb99](https://github.com/Gitlawb/zero/commit/a7cfb99fed7b88ebc09a2f251cb82864d3c2cade))
+* gitignore Windows sandbox helpers and npm version marker ([#578](https://github.com/Gitlawb/zero/issues/578)) ([25653f6](https://github.com/Gitlawb/zero/commit/25653f686c016f95b81ceb7ff5d5452d37c4d4f3))
+* **mcp:** silence startup warning for unconfigured default servers ([#563](https://github.com/Gitlawb/zero/issues/563)) ([302f58b](https://github.com/Gitlawb/zero/commit/302f58bb5f2a03ec7230354ed4747e4e55c16c50))
+* **mcp:** skip RFC 8414 discovery when OAuth endpoints are preconfigured ([#586](https://github.com/Gitlawb/zero/issues/586)) ([8a52d98](https://github.com/Gitlawb/zero/commit/8a52d98cad7cd0086dee9aede4ce477e432bd385))
+* **modelregistry:** reject oversized models.dev cache responses ([#602](https://github.com/Gitlawb/zero/issues/602)) ([66a6396](https://github.com/Gitlawb/zero/commit/66a63964149fb2f07e646e5f1987627c5cd9ac28))
+* **provider:** stop dropping custom no-auth providers on restart ([#558](https://github.com/Gitlawb/zero/issues/558)) ([ba99fa8](https://github.com/Gitlawb/zero/commit/ba99fa8d487fb28f2700e0ff10b2a25c75303cf7))
+* **sandbox:** Windows-appropriate suggestions when blocking interactive commands ([#414](https://github.com/Gitlawb/zero/issues/414)) ([ba4c007](https://github.com/Gitlawb/zero/commit/ba4c00755dc1c31b3dcca18e50b20f62c6bf5d1f))
+* **tools:** block MSYS and WSL shells under the Windows sandbox ([#587](https://github.com/Gitlawb/zero/issues/587)) ([0666818](https://github.com/Gitlawb/zero/commit/066681855b80e3baf5d07d7397610b25f724e353))
+* **tools:** platform-specific pager suggestions, quote/caret-safe cd detection ([#543](https://github.com/Gitlawb/zero/issues/543)) ([8b248f4](https://github.com/Gitlawb/zero/commit/8b248f4e1198dc86ab332a697fba4cf520823cbd))
+* **tools:** Windows cmd.exe quoting guidance and clipboard escaping fix ([#468](https://github.com/Gitlawb/zero/issues/468)) ([f10ed0c](https://github.com/Gitlawb/zero/commit/f10ed0c893ce6de08923f143d681ba96f0fcfe3a))
+* **tui:** bypass toggleSidebar and toggleMouse global shortcuts when composer is non-empty ([#576](https://github.com/Gitlawb/zero/issues/576)) ([c7346fb](https://github.com/Gitlawb/zero/commit/c7346fbcf01fe7e70ed2fcfccbf9965e5985727b))
+* **tui:** paste protection for Termux char-by-char paste ([#573](https://github.com/Gitlawb/zero/issues/573)) ([8e9149f](https://github.com/Gitlawb/zero/commit/8e9149f1a296bebebf95cae9dd2a7c5156c9dbb6))
+* **tui:** update picker_test for switchProviderModel 4-value return ([#589](https://github.com/Gitlawb/zero/issues/589)) ([8f15650](https://github.com/Gitlawb/zero/commit/8f156506dc92449bd24caa14e36f28477ba00fff))
+* **update:** clearer error on unsupported release platform (android/termux) ([#603](https://github.com/Gitlawb/zero/issues/603)) ([1fc9b2d](https://github.com/Gitlawb/zero/commit/1fc9b2d25c79e089f34cb7b5b6a7f7c7b8233123))
+* **update:** support safe symlink extraction during updates ([#575](https://github.com/Gitlawb/zero/issues/575)) ([ce9cb91](https://github.com/Gitlawb/zero/commit/ce9cb912958a3d9ae0b052bfebb849c28e0e719b))
+* warn about untracked scratch files left behind after a run ([#571](https://github.com/Gitlawb/zero/issues/571)) ([062328b](https://github.com/Gitlawb/zero/commit/062328b632a2d27353a6d47c522bbd22d7282539))
+
+
+### Performance Improvements
+
+* **grep:** stop content scan after head limit ([#601](https://github.com/Gitlawb/zero/issues/601)) ([8a05e64](https://github.com/Gitlawb/zero/commit/8a05e6486f7a63281b869064db03dfc5531e6a04))
+
+## [0.2.0](https://github.com/Gitlawb/zero/compare/v0.1.0...v0.2.0) (2026-07-06)
+
+
+### Features
+
+* add --auto flag for LLM-generated commit messages ([#423](https://github.com/Gitlawb/zero/issues/423)) ([b0abde7](https://github.com/Gitlawb/zero/commit/b0abde7d0697e808480cd59d69a6f4d0c6320475))
+* add zero changes push and pr subcommands, and extra repo-info metrics ([#391](https://github.com/Gitlawb/zero/issues/391)) ([2312abe](https://github.com/Gitlawb/zero/commit/2312abe5ddd95f4c6ef373cfb61cc03092f48cdd))
+* agent quality, caching, retry, and tooling upgrades ([#506](https://github.com/Gitlawb/zero/issues/506)) ([3c81fea](https://github.com/Gitlawb/zero/commit/3c81fea22873ee3df7fc97b10cb4f77792706c4b))
+* **agent:** curb over-engineering the solution in the editing discipline ([#517](https://github.com/Gitlawb/zero/issues/517)) ([f4c998a](https://github.com/Gitlawb/zero/commit/f4c998ac30c4f07ff313a2d706791e857293be49)), closes [#516](https://github.com/Gitlawb/zero/issues/516)
+* **agent:** inject per-user config.UserConfigDir()/zero/ZERO.md guidelines into system prompt ([#475](https://github.com/Gitlawb/zero/issues/475)) ([7b10aab](https://github.com/Gitlawb/zero/commit/7b10aab74bf14a01166b2cea22deab79bba9850b))
+* **openai:** forward prompt_cache_key for server-side prefix cache routing ([#515](https://github.com/Gitlawb/zero/issues/515)) ([87e7e69](https://github.com/Gitlawb/zero/commit/87e7e69afd18b5539579856f3a61c6a95bc445ae))
+* **providers:** add `zero providers models` to discover a provider's models ([#386](https://github.com/Gitlawb/zero/issues/386)) ([0bc8074](https://github.com/Gitlawb/zero/commit/0bc8074c97b0310e4a9d70c3f967003ee5e8a59f))
+* **providers:** add KiloCode and OpenCode provider support ([#388](https://github.com/Gitlawb/zero/issues/388)) ([b1ccb6d](https://github.com/Gitlawb/zero/commit/b1ccb6d9c1875377f5e5ea81a1304edd1e41ab4f))
+* **providers:** add Meituan LongCat catalog preset ([#424](https://github.com/Gitlawb/zero/issues/424)) ([b4275e3](https://github.com/Gitlawb/zero/commit/b4275e350472b2490212bf814709819d354c1216))
+* **providers:** split minimax zai into intl cn ([#398](https://github.com/Gitlawb/zero/issues/398)) ([aaad4d2](https://github.com/Gitlawb/zero/commit/aaad4d271270f41af837b6f3b60ae80beba0c645))
+* require manual approval before npm publish + drop release-as pin ([#369](https://github.com/Gitlawb/zero/issues/369)) ([bd89a1f](https://github.com/Gitlawb/zero/commit/bd89a1f451643c1b65ec803070abc7b116631ebe))
+* **sandbox:** unelevated Windows fallback tier instead of prompts-only degrade ([#427](https://github.com/Gitlawb/zero/issues/427)) ([b9ddd6f](https://github.com/Gitlawb/zero/commit/b9ddd6f42138312a1fee8d8bb67c46c8eb1dea2f))
+* support shift enter for composer newlines ([#462](https://github.com/Gitlawb/zero/issues/462)) ([daf65e0](https://github.com/Gitlawb/zero/commit/daf65e0af9a040314d4ab337b0ad59c55416b7bc))
+* **tui:** /loop — repeat a prompt or command on an interval or self-paced ([#502](https://github.com/Gitlawb/zero/issues/502)) ([387fe67](https://github.com/Gitlawb/zero/commit/387fe67ee7cd81317c9c969f5906a4437080fea3))
+* **tui:** add search/filter to provider picker in setup wizard ([#400](https://github.com/Gitlawb/zero/issues/400)) ([2fcea71](https://github.com/Gitlawb/zero/commit/2fcea71778d23e050c93409c471aef45b68c1621))
+* **update:** add zero upgrade command to apply self-updates ([#461](https://github.com/Gitlawb/zero/issues/461)) ([5f36349](https://github.com/Gitlawb/zero/commit/5f36349c1884e81fa9bc66bb5fe813b627e897b7))
+
+
+### Bug Fixes
+
+* **action:** keep provider key scoped to zero step ([#448](https://github.com/Gitlawb/zero/issues/448)) ([407a927](https://github.com/Gitlawb/zero/commit/407a92739ff508cba32d2c12b3f36f0efcdd54c3))
+* add android platform support for Termux npm install ([#455](https://github.com/Gitlawb/zero/issues/455)) ([9bd93c6](https://github.com/Gitlawb/zero/commit/9bd93c62f8d57fb74057284aa66a1b6e1429dcdd)), closes [#449](https://github.com/Gitlawb/zero/issues/449)
+* **agent:** reject a malformed additional_permissions payload before prompting ([#453](https://github.com/Gitlawb/zero/issues/453)) ([e4f760e](https://github.com/Gitlawb/zero/commit/e4f760ee8bd57299cd2fcb37e8e23130037c2607))
+* allow non-TLS connections to private-network provider endpoints ([#444](https://github.com/Gitlawb/zero/issues/444)) ([1d86384](https://github.com/Gitlawb/zero/commit/1d8638466ca31517eb9db2b9353d3dce1cbeeabc))
+* **auth:** route zero auth login chatgpt to the dedicated ChatGPT flow ([#443](https://github.com/Gitlawb/zero/issues/443)) ([305a62c](https://github.com/Gitlawb/zero/commit/305a62c954ca6cec00bc58d5398f933415156aff))
+* **config:** fall back to a usable saved provider instead of forcing full re-onboarding ([#410](https://github.com/Gitlawb/zero/issues/410)) ([c60ad87](https://github.com/Gitlawb/zero/commit/c60ad8729f79bb841114d352ee2d2fe29d5d0e41))
+* **config:** let a gateway ANTHROPIC_BASE_URL resolve as anthropic-compatible ([#497](https://github.com/Gitlawb/zero/issues/497)) ([30dd7c3](https://github.com/Gitlawb/zero/commit/30dd7c3112ad22d42fa12b5addd4e38f4beda42a)), closes [#479](https://github.com/Gitlawb/zero/issues/479)
+* **config:** unbrick first-run setup — default google/anthropic models, enter setup on fixable config errors ([#385](https://github.com/Gitlawb/zero/issues/385)) ([72eed06](https://github.com/Gitlawb/zero/commit/72eed06b4f94c43d75d31fe54a58d2f566de059e))
+* **config:** use ~/.config on macOS and enter setup when no provider ([#371](https://github.com/Gitlawb/zero/issues/371)) ([#372](https://github.com/Gitlawb/zero/issues/372)) ([027a8f2](https://github.com/Gitlawb/zero/commit/027a8f2768b17b89f5c8270887f156e2ccda69ea))
+* **docs:** rename AGENTS.MD > AGENTS.md ([#438](https://github.com/Gitlawb/zero/issues/438)) ([4266baf](https://github.com/Gitlawb/zero/commit/4266baf222df583ed2078b776687f12d496475b5))
+* **gemini:** strip unsupported JSON Schema fields from tool declarations ([#374](https://github.com/Gitlawb/zero/issues/374)) ([39e7100](https://github.com/Gitlawb/zero/commit/39e7100674150144a1152e3110c64c7cf0321d64)), closes [#373](https://github.com/Gitlawb/zero/issues/373)
+* **install:** persist install dir to user PATH on Windows ([#407](https://github.com/Gitlawb/zero/issues/407)) ([bdb1b0e](https://github.com/Gitlawb/zero/commit/bdb1b0ecd15859b1712a6037d296dace7f9c3c3f))
+* **mcp:** block cross-origin credential redirects ([#396](https://github.com/Gitlawb/zero/issues/396)) ([f915f70](https://github.com/Gitlawb/zero/commit/f915f70e5a3096e2419fa8d961a0f84a626fa4a9))
+* **oauth:** treat Windows ERROR_ACCESS_DENIED as lock contention in createSecretFile ([#445](https://github.com/Gitlawb/zero/issues/445)) ([d05e914](https://github.com/Gitlawb/zero/commit/d05e9148a7f79f67d1d3c31fca2775f21fbd331e))
+* **openai:** handle Ollama reasoning stream deltas ([#486](https://github.com/Gitlawb/zero/issues/486)) ([f6c0606](https://github.com/Gitlawb/zero/commit/f6c060631e18e082dda24cc4dc0903c31c2120d6))
+* preserve conversation context in exec prompts ([#460](https://github.com/Gitlawb/zero/issues/460)) ([949ee43](https://github.com/Gitlawb/zero/commit/949ee43f71e5cb7fab4695c5cb7b442fe4ecfbf7))
+* **provider-wizard:** allow multiple custom OpenAI-compatible providers ([#403](https://github.com/Gitlawb/zero/issues/403)) ([3fbbd28](https://github.com/Gitlawb/zero/commit/3fbbd28e4c586822cc4312c86232d94befe56e87))
+* **sandbox:** fix nested pipe creation under the Windows restricted token ([#456](https://github.com/Gitlawb/zero/issues/456)) ([563a6db](https://github.com/Gitlawb/zero/commit/563a6dbe91e65d5daeefd7626e8a77e30a6d8fb2))
+* **sandbox:** gate /tmp test assertions on GOOS, not path existence ([#426](https://github.com/Gitlawb/zero/issues/426)) ([f653dca](https://github.com/Gitlawb/zero/commit/f653dcac363fb69ad7be5b35e6e0fa6d2bce476d))
+* **sandbox:** self-heal a corrupt unelevated setup marker ([#437](https://github.com/Gitlawb/zero/issues/437)) ([8d0c5fe](https://github.com/Gitlawb/zero/commit/8d0c5feccb8bdbfb015df0508aa6e3bcbd1fd0e8))
+* **specialist:** cap max specialist nesting depth ([#491](https://github.com/Gitlawb/zero/issues/491)) ([177442c](https://github.com/Gitlawb/zero/commit/177442cfe4015bd8df04cc9894f98b468ee796d4))
+* Termux/Android support — PRoot scroll, SIGSYS sandbox, build docs ([#509](https://github.com/Gitlawb/zero/issues/509)) ([0f69d99](https://github.com/Gitlawb/zero/commit/0f69d995e9b586b774f66c066b21abab5e03024a))
+* **tools:** Block MSYS coreutils under Windows sandbox ([#476](https://github.com/Gitlawb/zero/issues/476)) ([81aad58](https://github.com/Gitlawb/zero/commit/81aad58d97839e51c068e2f08907618991fdc3fb))
+* **tools:** CRLF line ending mismatch in edit_file tool on Windows ([#378](https://github.com/Gitlawb/zero/issues/378)) ([33dc7ae](https://github.com/Gitlawb/zero/commit/33dc7ae2cc82c5389675531e1416856dae7151ce))
+* **tools:** fix cmd.exe /S/C corrupting commands with embedded quotes ([#465](https://github.com/Gitlawb/zero/issues/465)) ([190241b](https://github.com/Gitlawb/zero/commit/190241bd593f43211b766e0b13c8e89802d4bb37))
+* **tools:** flag piped POSIX utilities before running on Windows ([#412](https://github.com/Gitlawb/zero/issues/412)) ([5658a36](https://github.com/Gitlawb/zero/commit/5658a366274fc59a9d5336b06a21019c9c25cbf1))
+* **tools:** make grep and glob respect run cancellation ([#464](https://github.com/Gitlawb/zero/issues/464)) ([ba6c026](https://github.com/Gitlawb/zero/commit/ba6c0264697b7d7ed479f6e782fba9700a481e3d))
+* **tools:** require permission before web_search requests ([#382](https://github.com/Gitlawb/zero/issues/382)) ([960db96](https://github.com/Gitlawb/zero/commit/960db9660e4e31dc588fe8f7d6f116ff5e225566))
+* **tui:** compose help overlay through the viewport overlay pipeline ([#421](https://github.com/Gitlawb/zero/issues/421)) ([5b2b4de](https://github.com/Gitlawb/zero/commit/5b2b4dea1aaf9e0f68baa25e97e83296fb17b1a2))
+* **tui:** keep the profile name on /model switch so the stored key resolves ([#441](https://github.com/Gitlawb/zero/issues/441)) ([9134148](https://github.com/Gitlawb/zero/commit/9134148f4df3e4e556fba6c2f8babfdf6fcfeee1)), closes [#440](https://github.com/Gitlawb/zero/issues/440)
+* **tui:** resolve every permission request so the agent can't deadlock ([#397](https://github.com/Gitlawb/zero/issues/397)) ([952788f](https://github.com/Gitlawb/zero/commit/952788f72d32957659fe004521fcc8372b9ba9b4))
+* **tui:** show an M suffix for million-scale token counts ([#457](https://github.com/Gitlawb/zero/issues/457)) ([0562e3b](https://github.com/Gitlawb/zero/commit/0562e3bef7df2328610a48a1e81632a8da4aec64))
+* **tui:** title /model rows by model name, not the catalog description ([#395](https://github.com/Gitlawb/zero/issues/395)) ([cdf9d83](https://github.com/Gitlawb/zero/commit/cdf9d839ae57a729f292f36f7c5b0c67b41b288d))
+
+
+### Performance Improvements
+
+* cache TUI model registry ([#496](https://github.com/Gitlawb/zero/issues/496)) ([e7d88b4](https://github.com/Gitlawb/zero/commit/e7d88b4b518049733da25a8447c00144bd1da518))
+* universal tool-output ceiling with spill + async post-edit diagnostics ([#518](https://github.com/Gitlawb/zero/issues/518)) ([95ccd5b](https://github.com/Gitlawb/zero/commit/95ccd5bc327f6fb464ff0239f7229de789f578dc))
+
+## 0.1.0 (2026-07-02)
+
+
+### Features
+
+* publish zero to npm via release-please ([#367](https://github.com/Gitlawb/zero/issues/367)) ([8eccc26](https://github.com/Gitlawb/zero/commit/8eccc2669887bc38d35bc16a315c888e4d9ec43a))
+* **tui:** FILES sidebar panel with click-to-select and file drill-in ([#365](https://github.com/Gitlawb/zero/issues/365)) ([142c548](https://github.com/Gitlawb/zero/commit/142c548c89a8652ce300e64ddf1228ee36df7606))
+
+
+### Bug Fixes
+
+* **auth:** propagate credentials to every provider-build surface and pin children to the live provider ([#366](https://github.com/Gitlawb/zero/issues/366)) ([6e0a665](https://github.com/Gitlawb/zero/commit/6e0a665118fe0e09c4b07d482dd18f86045acd2b))
+
 ## [Unreleased]
 
 ### Added
+- Shared multi-agent skills discovery: when present, `~/.agents/skills` is searched after the primary
+  Zero skills dir (and before plugin skill roots). `zero skills list` / `info` and the runtime `skill`
+  tool share one multi-root discovery path; install/remove/lock still target only the Zero skills directory.
 - `SECURITY.md` with a private vulnerability-reporting path, `CODE_OF_CONDUCT.md`, this changelog, and
   GitHub issue/PR templates.
-- `--theme {auto|dark|light}` flag for the TUI (previously only the `ZERO_THEME` env var existed).
+- Interactive `/theme` picker: bare `/theme` opens a popup that live-previews each palette as you move
+  and applies on select (Esc reverts).
+- Twelve built-in color themes alongside the `dark`/`light` built-ins — `dracula`, `nord`, `gruvbox`,
+  `tokyo-night`, `catppuccin`, `one-dark`, `solarized-dark`, `rose-pine`, `everforest`,
+  `solarized-light`, `dune`, and `neon` — selectable via `/theme <name>`, `--theme <name>`, or
+  `ZERO_THEME`. Every palette is contrast-audited to WCAG AA, and the new presets are additionally
+  audited after xterm-256 downsampling; see [docs/THEMES.md](docs/THEMES.md). The built-in light
+  theme was reworked for legibility.
+- `--theme <name>` flag for the TUI, accepting `auto` or any registered theme (previously only the
+  `ZERO_THEME` env var existed).
 - "Accessibility / Appearance" section in the README documenting `NO_COLOR`, `ZERO_THEME`, `/theme`,
   and `ZERO_NO_FADE`.
 
@@ -33,5 +217,4 @@ tagged. Until then, source builds report the version `dev`.
 - The inert `/input-style` slash command (it had no backend).
 
 ### Fixed
-- README/`go.mod` Go-version mismatch and other stale docs claims surfaced by the product/UX audit
-  (`docs/audit/2026-06-21-product-ux-audit.md`).
+- README/`go.mod` Go-version mismatch and other stale public-release docs claims.

CONTRIBUTING.md

@@ -8,6 +8,13 @@ stabilizes. Community participation is still welcome, especially through bug
 reports, feature requests, reproduction details, logs, and discussion on existing
 issues.
 
+Questions and setup help belong in
+[GitHub Discussions](https://github.com/Gitlawb/zero/discussions) (see
+[SUPPORT.md](SUPPORT.md)), not in the issue tracker. Feature ideas start in the
+[Ideas](https://github.com/Gitlawb/zero/discussions/categories/ideas) category;
+an idea that gains traction there is the path to an approved issue and,
+eventually, an accepted PR.
+
 Please read this policy before opening a pull request. Pull requests that do not
 follow it may be closed without review.
 
@@ -33,7 +40,11 @@ unless the issue has been explicitly approved for outside contribution.
 There are three primary ways to contribute:
 
 1. Open a bug report with clear reproduction steps.
-2. Open a feature request describing the problem, use case, and expected behavior.
+2. Propose a feature in
+   [Discussions Ideas](https://github.com/Gitlawb/zero/discussions/categories/ideas),
+   describing the problem, use case, and expected behavior. A feature-request
+   issue is filed only after the idea gains traction there or a maintainer asks
+   for one.
 3. Add useful context to an existing issue, such as logs, screenshots, examples,
    affected versions, or reproduction cases.
 
@@ -46,7 +57,8 @@ All community pull requests require an approved parent issue.
 
 Before starting implementation work:
 
-1. Open an issue using the appropriate issue template.
+1. For a bug, open an issue using the bug-report template. For a feature, start
+   in Discussions Ideas; the feature-request issue comes after the discussion.
 2. Describe the bug, feature request, or functional change clearly.
 3. Wait for the core team to review the issue.
 4. Do not open a pull request unless the issue has the `issue-approved` label.
@@ -133,7 +145,12 @@ Clear bug reports are the most useful way to help the project during this phase.
 
 ## Submitting a Feature Request
 
-When filing a feature request, please include:
+Feature requests start as a thread in
+[Discussions Ideas](https://github.com/Gitlawb/zero/discussions/categories/ideas),
+not as an issue. The feature-request issue form exists for the step AFTER that:
+when an idea has gained traction in Discussions or a maintainer has asked for an
+issue to track it. Whether in the discussion or the eventual issue, please
+include:
 
 - The problem you are trying to solve.
 - The behavior or functionality you are requesting.
@@ -156,5 +173,9 @@ contribution phase.
 
 ## Getting Help
 
-If you are unsure whether a change is appropriate, open an issue first and ask.
-Please keep discussion focused, specific, and respectful of maintainer time.
+General questions, setup help, and "is this a bug or my config" uncertainty all
+go to
+[Discussions Q&A](https://github.com/Gitlawb/zero/discussions/categories/q-a).
+If you are unsure whether a proposed code change is appropriate, raise it in
+Discussions Ideas or on the relevant issue first and ask. Please keep discussion
+focused, specific, and respectful of maintainer time.

LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Gitlawb
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

Makefile

@@ -1,7 +1,7 @@
-# Zero build/test/lint targets. AGENTS.MD says "Build with `make`" and "Run `make
+# Zero build/test/lint targets. AGENTS.md says "Build with `make`" and "Run `make
 # lint` before opening a PR" — these targets back those instructions.
 .DEFAULT_GOAL := build
-.PHONY: build build-all test test-race vet fmt fmt-check lint tidy clean help
+.PHONY: build build-all test test-race vet fmt fmt-check lint tidy clean baseline help
 
 # Build the main CLI binary into ./zero.
 build:
@@ -40,5 +40,19 @@ clean:
 	rm -f zero
 	go clean ./...
 
+# Run the per-turn benchmark harness over the checked-in baseline manifest and
+# write the JSON result to internal/perfbench/reports/baseline.json. Requires a
+# built `zero` binary and a model; set ZERO_BENCH_MODEL (required) and
+# ZERO_BENCH_BINARY (defaults to ./zero) to configure the run. The report is
+# machine-specific and regenerated, not hand-edited.
+baseline: build
+	@if [ -z "$(ZERO_BENCH_MODEL)" ]; then echo "Set ZERO_BENCH_MODEL (and optionally ZERO_BENCH_BINARY) before running 'make baseline'"; exit 2; fi
+	@ZERO_BIN="$${ZERO_BENCH_BINARY:-./zero}"; \
+	go run ./cmd/zero-perf-bench turn \
+		--suite internal/perfbench/manifests/baseline.json \
+		--model $(ZERO_BENCH_MODEL) \
+		--binary "$$ZERO_BIN" \
+		--output internal/perfbench/reports/baseline.json
+
 help:
-	@echo "Targets: build (default), build-all, test, test-quick, vet, fmt, fmt-check, lint, tidy, clean"
+	@echo "Targets: build (default), build-all, test, test-quick, vet, fmt, fmt-check, lint, tidy, clean, baseline"

README.md

@@ -1,363 +1,432 @@
-<div align="center">
+<p align="center">
+  <img src="docs/assets/zero-logo.png" alt="Zero" width="385">
+</p>
 
-```
-███████╗ ███████╗ ██████╗   ██████╗
-╚══███╔╝ ██╔════╝ ██╔══██╗ ██╔═══██╗
-  ███╔╝  █████╗   ██████╔╝ ██║   ██║
- ███╔╝   ██╔══╝   ██╔══██╗ ██║   ██║
-███████╗ ███████╗ ██║  ██║ ╚██████╔╝
-╚══════╝ ╚══════╝ ╚═╝  ╚═╝  ╚═════╝
-```
+<p align="center"><strong>A terminal coding agent you own.</strong></p>
 
-### The terminal coding agent you fully own.
+<p align="center">
+  <a href="LICENSE"><img alt="license" src="https://img.shields.io/badge/license-MIT-blue"></a>
+  <img alt="Go 1.26.5+" src="https://img.shields.io/badge/Go-1.26.5+-00ADD8?logo=go&logoColor=white">
+  <img alt="25+ providers" src="https://img.shields.io/badge/providers-25+-34E2EA">
+  <br>
+  <strong>English</strong> | <a href="README_ZH.md">中文</a>
+</p>
 
-**Any model. Any provider. Your rules.**
-
-![go](https://img.shields.io/badge/Go-1.25-00ADD8?logo=go&logoColor=white)
-![providers](https://img.shields.io/badge/providers-25%2B-34E2EA)
-![tests](https://img.shields.io/badge/test%20files-200%2B-43D17A)
-![status](https://img.shields.io/badge/status-active%20development-E8B84B)
-
-</div>
-
----
-
-Zero is an AI coding agent that lives in your terminal. It runs a full agentic loop — reading, searching, editing, and executing in your repo — against **whatever model you choose**: frontier APIs, fast cloud inference, or a local model on your own machine. One interface, no vendor lock-in, no telemetry phoning home.
+Zero is an AI coding agent for your local terminal. It can inspect a repository,
+edit files, run commands, use browser/terminal helpers, and keep durable local
+sessions while you choose the model and the permission level.
 
 ```bash
-zero                                          # interactive TUI
-zero exec "fix the failing test in ./pkg"     # headless one-shot
-zero exec -o stream-json < turns.jsonl        # programmatic, for scripts & CI
+zero
+zero exec "fix the failing test in ./pkg"
+zero exec --output-format stream-json < turns.jsonl
 ```
 
-> Zero treats the **model as a swappable, per-task choice** and **never mutates your system without a permission decision**.
-
 ## Why Zero
 
-- 🔌 **25+ providers, one interface** — OpenAI, Anthropic, Gemini, Ollama (local & cloud), LM Studio, OpenRouter, Groq, DeepSeek, Mistral, xAI, Qwen, Kimi, GitHub Models, and any OpenAI- or Anthropic-compatible endpoint. Switch mid-session with `/model`.
-- 🖥️ **A TUI that feels premium** — truecolor Bubble Tea interface with a first-run setup wizard, searchable live model picker, scrollback, themes, image input, and slash commands for everything.
-- 🤖 **Headless & scriptable** — `zero exec` with `text` / `json` / `stream-json` I/O, session resume & fork, isolated `--worktree` runs, and meaningful exit codes. Built for CI.
-- 🧠 **Subagents** — delegate to built-in `worker`, `explorer`, and `code-review` specialists (or generate your own) that run as real background tasks, even out-of-process.
-- 📋 **Spec mode** — have the agent draft a spec first, review and approve it, *then* let it build. No more runaway sessions.
-- 📈 **Mid-run model escalation** — start cheap, and let the agent request a stronger model only when it hits a wall (`--allow-escalation`).
-- 🗺️ **Repo intelligence** — deterministic repo maps, workspace indexing, and context-budget reports keep the agent grounded in *your* codebase, not hallucinations.
-- ⏰ **Scheduled agents** — `zero cron` runs file-backed, dependency-free agent jobs on a schedule.
-- 🛡️ **Safe by default** — permission-gated mutations, autonomy ceilings, sandbox policy (writes stay inside the workspace unless you grant extra directories with `--add-dir` / `/add-dir`), and secret redaction everywhere. Unsafe mode is an explicit, loudly-labeled opt-in.
-- 💾 **Durable sessions** — append-only local event store with full-text search, resume, fork, and rewind. Your history never leaves your disk.
-- 🧩 **Extensible** — skills, plugins, hooks, and MCP (Zero is both an MCP client *and* an MCP server).
+- **Use the model you want.** Bring OpenAI, Anthropic, Gemini, Groq, OpenRouter,
+  DeepSeek, Mistral, xAI, Qwen, Kimi, GitHub Models, Ollama, LM Studio, or any
+  OpenAI-/Anthropic-compatible endpoint.
+- **Stay in control.** File writes, shell commands, network access, and
+  out-of-workspace writes go through Zero's permission and sandbox policy.
+- **Works in the terminal.** The TUI has model/provider pickers, image input,
+  slash commands, live plan/tool rendering, scrollback, themes, and resume/fork
+  support.
+- **Works without the TUI.** `zero exec` is scriptable, supports text/JSON/
+  stream-JSON I/O, isolated worktrees, spec-first runs, and meaningful exit
+  codes for CI.
+- **Keeps context local.** Sessions are stored on disk, searchable, resumable,
+  and never uploaded as telemetry by Zero.
+- **Extensible when you need it.** Use MCP servers, skills, plugins, hooks, and
+  specialist subagents from the same CLI.
 
-## Quick start
+## Install
 
-```bash
-# run from source (requires Go 1.25+)
-go run ./cmd/zero
-```
-
-Or, once the first release is published, install the prebuilt binary via npm — the wrapper
-downloads the right binary for your platform:
+### npm
 
 ```bash
 npm install -g @gitlawb/zero
 zero
 ```
 
-> **Pre-built binaries are coming soon.** The npm package and the `scripts/install.sh`
-> (Linux/macOS) / `scripts/install.ps1` (Windows) installers all download GitHub Release
-> assets, which require the first release to be published. Until then, build from source with
-> `go run ./cmd/zero` (or `go build -o zero ./cmd/zero`).
+The npm package is a small wrapper whose platform build (Linux and macOS on
+x64/arm64, Windows on x64 — including the browser/terminal control helpers)
+installs as an optional dependency straight from the npm registry — no install
+scripts, no downloads outside npm. Bun, pnpm, and yarn work the same way with
+no trust or approval steps. Installs that skip optional dependencies
+(`--omit=optional`) still work: the wrapper fetches the binary from the
+matching GitHub Release whenever it is missing. Windows on ARM runs the x64
+build under emulation. See [docs/NPM_PACKAGING.md](docs/NPM_PACKAGING.md) for
+how the package is put together.
 
-First launch opens a **guided setup wizard** — pick a provider, paste a key, choose a model, done. Or do it non-interactively:
+### Install scripts
+
+Linux/macOS:
 
 ```bash
-export OPENAI_API_KEY=sk-...      # or ANTHROPIC_API_KEY, GEMINI_API_KEY, GROQ_API_KEY, ...
-zero setup                        # guided first-run provider setup
-zero doctor                       # verify config, keys, and connectivity
+curl -fsSL https://raw.githubusercontent.com/Gitlawb/zero/main/scripts/install.sh | bash
 ```
 
-Local models need no key at all:
+Windows PowerShell:
+
+```powershell
+irm https://raw.githubusercontent.com/Gitlawb/zero/main/scripts/install.ps1 | iex
+```
+
+### From source
+
+Source builds require Go 1.26.5+.
 
 ```bash
-# Ollama or LM Studio running locally? Zero finds them.
+git clone https://github.com/Gitlawb/zero.git
+cd zero
+go run ./cmd/zero
+```
+
+Release installers and the npm wrapper require published GitHub Release assets.
+If you are testing before the first public release, build from source:
+
+```bash
+go build -o zero ./cmd/zero
+```
+
+On Linux, build the sandbox helper too if you want native sandboxing:
+
+```bash
+go build -o zero-linux-sandbox ./cmd/zero-linux-sandbox
+go build -o zero-seccomp ./cmd/zero-seccomp   # optional compatibility wrapper
+```
+
+Put `zero` and `zero-linux-sandbox` in the same directory on `PATH`
+(`~/.local/bin` is a good default). macOS does not need an extra helper binary.
+Windows source builds can use the main `zero.exe` as their sandbox helper; release
+archives still ship standalone Windows helper executables.
+
+More install details: [docs/INSTALL.md](docs/INSTALL.md).
+
+## First Run
+
+Start the TUI:
+
+```bash
+zero
+```
+
+The setup wizard helps you pick a provider and model. You can also configure
+providers from the command line:
+
+```bash
+zero setup
 zero providers list
+zero models list
+zero doctor
 ```
 
-## The TUI
-
-Type to chat, **Enter** to send. `/` opens command suggestions, **Shift+Tab** cycles permission modes, **Ctrl+C** exits.
-
-| | |
-|---|---|
-| `/model` `/provider` | switch model or provider mid-session (searchable picker) |
-| `/spec` `/plan` | spec-mode drafting and live plan view |
-| `/image` | attach images for vision models |
-| `/resume` `/rewind` | time-travel across sessions |
-| `/compact` `/context` | manage the context window |
-| `/permissions` `/tools` | inspect what the agent can touch |
-| `/add-dir` | grant an extra write directory for the session, or list current write roots |
-| `/theme` `/style` | make it yours |
-| `/doctor` `/config` | health and config without leaving the chat |
-
-Turn-completion notifications (terminal bell / OSC-9) ping you when the agent finishes or needs input — go make coffee.
-
-## Accessibility & appearance
-
-Zero ships sensible defaults and a few environment/command controls. Meaning never depends on color
-alone — diffs carry `+`/`−` signs, permission outcomes use `✓`/`✗` glyphs plus a text `PERMISSION`
-badge, and permission modes show text labels — so the UI stays readable when color is stripped.
-
-| Control | What it does |
-|---|---|
-| `NO_COLOR` (any non-empty value) | Disables color, per [no-color.org](https://no-color.org). Zero honors **any** non-empty value (`NO_COLOR=1`, `yes`, `true`, …); bold/underline still render. |
-| `ZERO_THEME=auto\|dark\|light` | Selects the palette at startup. `auto` (default) detects the terminal background. |
-| `--theme auto\|dark\|light` | Same as `ZERO_THEME`, as a launch flag (takes precedence over the env var). |
-| `/theme [auto\|dark\|light]` | Switches the palette live inside the TUI. |
-| `ZERO_NO_FADE=1` | Reduce-motion opt-out: disables the streaming-text fade animation. Fade also auto-disables over SSH/tmux and on low-color/no-TTY terminals. |
-
-## Headless `exec`
+For API providers, set the matching environment variable before setup or enter
+the key in the wizard:
 
 ```bash
-# one-shot
-zero exec "explain internal/agent/loop.go and suggest one improvement"
+export OPENAI_API_KEY=sk-...
+export ANTHROPIC_API_KEY=...
+export GEMINI_API_KEY=...
+export AIMLAPI_API_KEY=...
+export LONGCAT_API_KEY=...
+export MINIMAX_API_KEY=...
+export MINIMAXI_API_KEY=...
+```
 
-# pick a model and mode preset per task
-zero exec --model claude-sonnet-4.5 --mode deep "refactor the session store"
+To configure AI/ML API directly, run:
 
-# spec-first: draft → review → approve → build
+```bash
+zero providers setup aimlapi --set-active
+```
+
+To configure Meituan LongCat (LongCat-2.0) directly, run:
+
+```bash
+zero providers setup longcat --set-active
+```
+
+MiniMax presets use the Anthropic-compatible endpoints for the global and China
+regions:
+
+```bash
+zero providers add minimax --set-active
+zero providers add minimaxi-cn --set-active
+```
+
+To use the OpenAI-compatible endpoints instead, add a custom compatible profile
+for the required region:
+
+```bash
+zero providers add custom-openai-compatible \
+  --name minimax-openai \
+  --model MiniMax-M3 \
+  --base-url https://api.minimax.io/v1 \
+  --api-key-env MINIMAX_API_KEY \
+  --set-active
+
+zero providers add custom-openai-compatible \
+  --name minimax-cn-openai \
+  --model MiniMax-M3 \
+  --base-url https://api.minimaxi.com/v1 \
+  --api-key-env MINIMAXI_API_KEY \
+  --set-active
+```
+
+For local models, run Ollama or LM Studio and then use `zero setup` or
+`zero providers detect`.
+
+## Daily Use
+
+### Interactive TUI
+
+```bash
+zero
+```
+
+Useful controls:
+
+| Control | Action |
+|---|---|
+| `Enter` | send the prompt |
+| `/` | open slash-command suggestions |
+| `Ctrl+X` then letter | common slash commands (e.g. `m` → `/model`; `Ctrl+X` `?` for full list) |
+| `Ctrl+P` / `Ctrl+N` | previous / next item in menus (arrows still work) |
+| `Shift+Tab` | cycle permission mode |
+| `Ctrl+B` | show/hide the sidebar |
+| `Ctrl+C` | cancel or exit |
+
+Common slash commands:
+
+| Command | Purpose |
+|---|---|
+| `/model`, `/provider` | switch the active model/provider |
+| `/spec`, `/plan` | draft and review a plan before building |
+| `/image` | attach an image for vision-capable models |
+| `/resume`, `/rewind` | continue or roll back local sessions |
+| `/loop` | repeat a prompt or custom `/command` on an interval (`/loop 5m /babysit-prs`) or self-paced |
+| `/compact`, `/context` | manage context usage |
+| `/permissions`, `/tools` | inspect available tools and policy |
+| `/add-dir` | allow an extra write directory for this session |
+| `/theme`, `/doctor`, `/config` | adjust appearance and inspect setup |
+
+### Headless `exec`
+
+```bash
+zero exec "explain internal/agent/loop.go"
+zero exec --model claude-sonnet-4.5 "refactor the config loader"
 zero exec --use-spec "add rate limiting to the API client"
+zero exec --worktree "try the migration in an isolated worktree"
+zero exec --resume
+zero exec --fork <session-id> "try the other approach"
+```
 
-# run in an isolated git worktree, escalate model only if needed
-zero exec -w --allow-escalation "migrate the config loader to v2"
+Programmatic use:
 
-# multi-turn programmatic I/O over stdio
+```bash
 zero exec --input-format stream-json --output-format stream-json < turns.jsonl
-
-# resume or fork any previous session
-zero exec --resume            # latest
-zero exec --fork <session-id> "now try the other approach"
 ```
 
-Key flags: `-m/--model` · `--mode <smart|deep|fast|large|precise>` · `--image` · `--use-spec` · `--auto <low|medium|high>` · `--enabled-tools/--disabled-tools` · `-w/--worktree` · `--add-dir <path>` (repeatable) · `--resume/--fork` · `--allow-escalation` · `--notify` · `-o <text|json|stream-json>`.
+The stream-JSON contract is documented in
+[docs/STREAM_JSON_PROTOCOL.md](docs/STREAM_JSON_PROTOCOL.md).
 
-stdout carries **only** program output; logs go to stderr. Full contract in [`docs/STREAM_JSON_PROTOCOL.md`](docs/STREAM_JSON_PROTOCOL.md).
+## Safety Model
 
-## Commands
+Zero is designed to make side effects visible.
 
+- Workspace reads are allowed by default.
+- File writes are limited to the workspace unless you grant another directory.
+- Shell commands, network access, destructive commands, and elevated actions are
+  permission-gated.
+- `--add-dir <path>` and `/add-dir <path>` grant additional write roots without
+  giving the agent the whole filesystem.
+- Unsafe/autonomous modes are explicit opt-ins.
+- Secrets are redacted from tool output and logs where Zero controls the surface.
+
+Example:
+
+```bash
+zero --add-dir ../docs-site
+zero exec --add-dir ../shared "update both repos"
 ```
+
+Sandbox behavior can be inspected with:
+
+```bash
+zero sandbox policy
+zero sandbox grants list
+```
+
+## Web And Local Control
+
+Zero includes local file/search/edit/shell tools, `web_fetch` for public URLs,
+and MCP support for additional tools.
+
+For local dev servers, use shell commands such as `curl` through `exec_command`
+so the normal sandbox and permission policy applies. Long-running commands stay
+attached to a background terminal session and can be listed or stopped from the
+TUI.
+
+The npm package also includes browser and terminal helper packages used by local
+browser/terminal tools. Source builds can use the same helpers when they are on
+`PATH` or configured in Zero's local-control settings.
+
+## Common Commands
+
+```text
 zero                  interactive TUI
-zero exec             one-shot / scripted agent runs
-zero setup            guided first-run provider setup
-zero models           model registry (capabilities, context, cost)
-zero providers        provider profiles + 25-provider catalog
-zero doctor           config, key, and connectivity health checks
-zero context          workspace context-budget report
-zero repo-map         deterministic repository map for agent context
-zero repo-info        local (network-free) repository characterizer
-zero search | find    full-text search over local session history
-zero sessions         session lineage inspection
-zero spec             review & approve saved spec-mode drafts
-zero specialist       manage subagent profiles
-zero skills           markdown instruction skills
-zero plugins          plugin manifests
-zero hooks            lifecycle hook configuration
-zero mcp              MCP client settings
-zero serve --mcp      expose Zero's tools over MCP stdio
-zero sandbox          sandbox policy & persistent grants
-zero worktrees        isolated git worktrees for agent runs
-zero verify           detect & run local verification checks
-zero changes          inspect & commit local git changes
+zero exec             one-shot or scripted agent run
+zero setup            first-run provider setup
+zero auth             OAuth/login helpers for supported providers
+zero models           model registry and capabilities
+zero providers        provider profiles and detection
+zero doctor           setup, key, and connectivity checks
+zero context          context-budget report
+zero repo-map         deterministic repository map
+zero repo-info        local repository summary
+zero search | find    search local session history
+zero sessions         inspect, resume, fork, and rewind sessions
+zero spec             manage spec-mode drafts
+zero specialist       manage specialist subagents
+zero skills           manage markdown instruction skills
+zero plugins          manage plugins
+zero hooks            manage lifecycle hooks
+zero mcp              manage MCP servers and tools
+zero serve --mcp      expose Zero tools over MCP stdio
+zero sandbox          inspect sandbox policy and grants
+zero worktrees        prepare isolated git worktrees
+zero verify           detect and run local verification checks
+zero changes          inspect and commit local git changes
 zero usage            token usage and estimated cost
-zero cron             scheduled agent jobs (file-backed, dep-free)
+zero cron             scheduled agent jobs
 zero update           check for newer releases
 ```
 
-## Providers
+## Extending Zero
 
-Bring your own key — or no key at all for local runtimes.
+### Project and personal instructions
 
-| Tier | Providers |
+Zero appends project-specific guidance to the system prompt from the first
+`AGENTS.md`, `ZERO.md`, or `.zero/AGENTS.md` file found in each directory from
+the git root down to your current working directory (checked in that order
+per directory). Files are injected general-to-specific, capped at 8 KiB per
+file and 32 KiB total.
+
+A personal `ZERO.md` under `config.UserConfigDir()/zero/ZERO.md`
+(`$XDG_CONFIG_HOME/zero/ZERO.md` or `~/.config/zero/ZERO.md` on Linux/macOS,
+`%AppData%\Roaming\zero\ZERO.md` on Windows) applies across every workspace, ahead of any project guidelines.
+
+### Plugins
+
+Plugins are discovered from `~/.config/zero/plugins/<name>/plugin.json` (user
+scope — `$XDG_CONFIG_HOME` or `~/.config` on every OS, independent of the
+`config.UserConfigDir()` path used above) and `<cwd>/.zero/plugins/<name>/plugin.json`
+(project scope — resolved from the current working directory, not the repo
+root), and managed with `zero plugins`. A manifest can declare:
+
+- `tools` — custom tools (`command`, `args`, `inputSchema`, and a
+  `permission` of `prompt` or `deny`; `allow` is honored only when manifest tool
+  auto-approval is enabled)
+- `hooks` — commands run on `beforeTool`, `afterTool`, `sessionStart`, or
+  `sessionEnd`
+- `prompts` and `skills` — additional prompt/skill files
+
+MCP servers (`zero mcp`) and standalone markdown skills (`zero skills`) use
+the same extension points and can also be wired up outside of a plugin
+manifest.
+
+## Appearance And Accessibility
+
+| Control | Effect |
 |---|---|
-| **Frontier APIs** | OpenAI · Anthropic · Google Gemini |
-| **Fast cloud inference** | Groq · OpenRouter · Together AI · DeepSeek · Mistral · xAI · NVIDIA NIM |
-| **Local — no key, no cloud** | Ollama · LM Studio |
-| **More clouds** | Ollama Cloud · DashScope (Qwen) · Moonshot (Kimi) · MiniMax · Z.ai · Venice · GitHub Models · and more |
-| **Enterprise (catalog)** | Amazon Bedrock · Vertex AI *(adapters in progress)* |
-| **Anything else** | any OpenAI-compatible or Anthropic-compatible endpoint |
+| `NO_COLOR=<anything>` | disables color output |
+| `ZERO_THEME=<name>` | selects the startup theme (`auto`, `dark`, `light`, or a color theme like `dracula`, `nord`, `gruvbox`, `tokyo-night`, `catppuccin`, `one-dark`, `solarized-dark`, `rose-pine`, `everforest`, `neon`, `solarized-light`, `dune`) |
+| `--theme <name>` | selects the TUI theme from the CLI (same names) |
+| `/theme` | opens the theme picker inside the TUI (live preview; `/theme <name>` switches directly) |
+| `ZERO_NO_FADE=1` | disables streaming fade animation |
 
-The model registry tracks each model's capabilities, context window, and cost — and the live model picker discovers what your provider actually serves.
-
-**Slow or stalling streams.** Every provider aborts a stream that goes completely silent (no data and no keep-alive) for `ZERO_STREAM_IDLE_TIMEOUT` (default **5 minutes**) so a hung connection can't block the agent forever. SSE keep-alives reset the timer, so a heartbeating-but-slow backend is never killed. If a slow cloud or reasoning model still trips it, raise the limit — `ZERO_STREAM_IDLE_TIMEOUT=10m` (accepts a Go duration or bare seconds; `0`/`off` disables the watchdog).
-
-## Tools
-
-| Tool | Purpose | Side effect |
-|---|---|---|
-| `read_file` · `list_directory` · `grep` · `glob` | explore & search | read |
-| `web_fetch` | fetch docs & references | network |
-| `update_plan` · `ask_user` | plan & clarify | none |
-| `write_file` · `edit_file` · `apply_patch` | create & modify | write (gated) |
-| `exec_command` · `write_stdin` | run shell commands, keep long-running processes alive, poll/send input | shell (gated on start/input) |
-| `bash` | run one-shot legacy shell commands | shell (gated) |
-| `Task` · `TaskOutput` · `TaskStop` | delegate to specialist subagents | per-tool gating |
-| `GenerateSpecialist` | create new subagent manifests | write (gated) |
-| `skill` | load markdown instruction skills | read |
-| `tool_search` | lazily load deferred tools (large MCP sets stay cheap) | none |
-| `escalate_model` | request a stronger model mid-run | gated by `--allow-escalation` |
-
-Every mutating tool routes through the permission policy **before** any side effect.
-
-### Web search & scraping (free, no API key)
-
-`web_fetch` is built in and always available: it pulls a single public remote URL
-into clean markdown locally, with no third party. For localhost/private dev-server
-URLs, use shell commands such as `curl` through `exec_command` so sandbox network
-permission applies. For local dev servers, start the server in the foreground with
-`exec_command` and keep the returned `session_id`; use `write_stdin` to poll or
-interrupt it. Use `/ps` to list running background terminals and `/stop` or
-`/stop <session_id>` to close them. Interactive terminal-style commands can set
-`tty: true` on `exec_command` so stdin works like a terminal session.
-
-For **search**, JS-rendered scraping,
-whole-site crawls, PDF→markdown, and structured extraction, Zero ships the
-[Firecrawl](https://firecrawl.dev) **keyless** MCP server **enabled by default**, so
-web search and scraping work out of the box with **no setup and no API key** (1,000
-free credits/month). It is a normal MCP server, so you stay in control:
-
-```bash
-zero mcp tools list           # see the firecrawl_* tools it exposes
-zero mcp disable firecrawl    # turn it off — web_fetch stays as the local floor
-```
-
-- **Privacy / disclosure:** keyless requests route through `firecrawl.dev`. If you
-  prefer nothing leaves your machine, `zero mcp disable firecrawl` and rely on
-  `web_fetch`, or self-host below.
-- **Self-host (unlimited, private):** Firecrawl is open source (AGPL-3.0). Run your
-  own instance and override the default URL — Zero only *calls* it over the network,
-  so the AGPL never reaches into Zero's own code:
-  ```jsonc
-  // config.json
-  { "mcp": { "servers": { "firecrawl": { "type": "http", "url": "http://localhost:3002/v2/mcp" } } } }
-  ```
-- **Bring your own key (higher limits):** add an auth header over the default:
-  ```bash
-  zero mcp add firecrawl --type http --url https://mcp.firecrawl.dev/v2/mcp \
-    --header "Authorization: Bearer fc-your-key"
-  ```
-
-An MCP server that can't be reached at startup (e.g. Firecrawl on an offline
-machine) is **skipped with a warning, not fatal** — Zero still launches and the
-rest of its tools, including `web_fetch`, keep working.
-
-### Extra write directories (`--add-dir`)
-
-Zero confines writes to the workspace by default. To let the agent write somewhere
-else, pass the repeatable `--add-dir` flag — it works for both the interactive TUI
-and `zero exec`:
-
-```bash
-zero --add-dir ~/Desktop/scratch                       # launch the TUI with an extra write root
-zero exec --add-dir ../sibling-repo "update both repos"
-```
-
-In the TUI, `/add-dir <path>` grants a directory mid-session (session-only), and a
-bare `/add-dir` lists the current write roots. To persist extra roots across
-sessions, set `sandbox.additionalWriteRoots` in the **global** user config
-(`~/.config/zero/config.json`); the key is deliberately ignored in project config
-so a checked-out repo can't widen its own sandbox. Flag and config sources merge
-as a union.
-
-Granted roots must already exist (the filesystem root is rejected), symlinks are
-resolved when the grant is made, and the same per-root symlink-traversal checks
-that protect the workspace apply to each extra root. Relative paths in tool calls
-still resolve against the workspace only, and network and destructive-shell policy
-are unchanged. A write denied outside all roots returns an error that suggests
-`/add-dir`.
-
-Two extra hardening/diagnostic flags are available in the `sandbox` config block,
-both **off by default** and safe to leave unset:
-
-- `sandbox.blockUnixSockets` (Linux) — asks the Linux sandbox helper to install a
-  best-effort seccomp filter that denies `AF_UNIX` socket creation in the
-  sandboxed command. Linux packages also continue to ship `zero-seccomp` as a
-  compatibility wrapper for existing direct uses; new sandbox plans do not depend
-  on that external wrapper.
-- `sandbox.monitorDenials` (macOS) — tails the unified log for this run's seatbelt
-  denials and appends them to a command's stderr as a `<sandbox_blocks>` block
-  so blocked operations are visible. A no-op on OS versions that do not deliver
-  seatbelt denials to the queryable log.
-
-## Architecture
-
-```
-   TUI (Bubble Tea)      headless exec       MCP server      cron runner
-        └──────────────────────┬──────────────────────┘
-                  surface-agnostic agent core
-            (loop · typed event stream · tool registry)
-   ┌──────────┬──────────┬───────────┬───────────┬────────────┬──────────┐
- providers   tools     sessions    specialist   repo intel   permissions
- + catalog   registry  + search    + background + workspace  + sandbox
- + registry            + rewind      tasks        index      + redaction
-```
-
-- **Surface-agnostic core** — the agent loop streams text + tool calls and emits one typed event stream consumed identically by the TUI, `exec`, the MCP server, and cron.
-- **Edges are interfaces** — `Provider`, `Tool`, `SessionStore`, and the permission policy are swappable.
-- **Model is data** — capabilities, cost, and routing live in the registry, never hard-coded.
-- **Pure Go** — one static binary per platform; the npm wrapper just delegates to it.
-
-## Project layout
-
-```
-cmd/
-  zero/                 production CLI entrypoint
-  zero-release/         release builder + smoke tests
-  zero-perf-bench/      performance benchmarks
-  zero-pr-review/       deterministic PR review helper
-internal/
-  agent/ zeroruntime/   agent loop & runtime orchestration
-  cli/                  command surface (exec, doctor, cron, ...)
-  tui/                  Bubble Tea terminal interface
-  providers/ providercatalog/ providermodelcatalog/
-  modelregistry/        capabilities, context windows, cost
-  tools/                read/write/edit/bash/grep/glob/patch/...
-  specialist/ background/  subagents + out-of-process tasks
-  sessions/ search/     append-only store, full-text search
-  repomap/ repoinfo/ workspaceindex/ contextreport/
-  specmode/ cron/ skills/ plugins/ hooks/ mcp/
-  sandbox/ redaction/ secrets/   safety surfaces
-  doctor/ providerhealth/ verify/ selfverify/
-docs/                   PRD, protocols, install/update/perf
-scripts/                installers
-```
+Meaning does not rely on color alone; diffs, permissions, and statuses also use
+text or glyph markers.
 
 ## Development
 
 ```bash
-go test ./...                     # full test suite (200+ test files)
-go run ./cmd/zero-release build   # compile the release binary
-go run ./cmd/zero-release smoke   # smoke-test it
-go run ./cmd/zero-perf-bench      # perf benchmarks (docs/PERFORMANCE.md)
+go test ./...
+go run ./cmd/zero-release build
+go run ./cmd/zero-release smoke
+go run ./cmd/zero-perf-bench
+```
 
-# cross-compile
+Experimental: `ZERO_OPENAI_TURN_SESSION=1` enables the optimized OpenAI turn
+session (background connection prewarm + request-prefix telemetry) for headless
+`zero exec` runs against official OpenAI profiles. Off by default; `0`/`false`
+disable. A/B-benchmark it by running the same `zero-perf-bench` suite with the
+variable unset and set.
+
+### Code Quality and Security Checks
+
+Before committing any changes, ensure all Go code quality and security checks pass. Pinned `go run` commands matching CI constraints can be used directly without prior installation:
+
+1. **Formatting**: Run `go fmt ./...` (or `make fmt`).
+2. **Vetting**: Run `go vet ./...` (or `make vet`).
+3. **Linting**: Run `go run github.com/golangci/golangci-lint/v2/cmd/[email protected] run --enable-only unused,ineffassign,staticcheck ./...`.
+4. **Vulnerability Scan**: Run `go run golang.org/x/vuln/cmd/[email protected] ./...`.
+
+If you prefer to install these tools globally on your path, you can run:
+
+```bash
+# Install golangci-lint
+go install github.com/golangci/golangci-lint/v2/cmd/[email protected]
+
+# Install govulncheck
+go install golang.org/x/vuln/cmd/[email protected]
+```
+
+### Cross-Compile Examples
+
+```bash
 go run ./cmd/zero-release build --goos linux --goarch amd64
 go run ./cmd/zero-release build --goos windows --goarch amd64 --output dist/zero.exe
 ```
 
 ## Documentation
 
-- [Product Requirements (PRD)](docs/PRD.md) — vision, full feature spec, roadmap
-- [Stream-JSON protocol](docs/STREAM_JSON_PROTOCOL.md) — headless I/O contract
-- [Specialists](docs/SPECIALISTS.md) — subagent manifests, Task tools, background state
-- [Install](docs/INSTALL.md) · [Update](docs/UPDATE.md) · [Performance](docs/PERFORMANCE.md)
+- [Install](docs/INSTALL.md)
+- [Update flow](docs/UPDATE.md)
+- [Themes](docs/THEMES.md)
+- [Stream-JSON protocol](docs/STREAM_JSON_PROTOCOL.md)
+- [Specialists](docs/SPECIALISTS.md)
+- [GitHub Action](docs/GITHUB_ACTION.md)
+- [Benchmarks](docs/BENCHMARK.md)
+- [Performance](docs/PERFORMANCE.md)
+- [Agent evals](docs/AGENT_EVALS.md)
+
+## Community
+
+Questions, setup help, ideas, and sharing all live in
+[GitHub Discussions](https://github.com/Gitlawb/zero/discussions):
+
+| Category | Use it for |
+|---|---|
+| [Q&A](https://github.com/Gitlawb/zero/discussions/categories/q-a) | Setup help, provider/model configuration, "how do I" questions |
+| [Ideas](https://github.com/Gitlawb/zero/discussions/categories/ideas) | Feature proposals and design discussion before any PR |
+| [Show and tell](https://github.com/Gitlawb/zero/discussions/categories/show-and-tell) | Your skills, plugins, MCP setups, themes, and workflows |
+| [Announcements](https://github.com/Gitlawb/zero/discussions/categories/announcements) | Releases and project news from the maintainers |
+
+For a good Q&A answer fast, include `zero --version`, your OS and install
+method, the provider/model in use, and `zero doctor` output. See
+[SUPPORT.md](SUPPORT.md). Bugs belong in
+[issues](https://github.com/Gitlawb/zero/issues/new/choose); security reports
+follow [SECURITY.md](SECURITY.md), never a public thread.
 
 ## Contributing
 
-Contributions are welcome — see [CONTRIBUTING.md](CONTRIBUTING.md). Run `go test ./...` and the relevant build or smoke command before opening a PR.
+Contributions are welcome. Read [CONTRIBUTING.md](CONTRIBUTING.md), run the
+relevant tests, and open a focused pull request.
+
+Security reports should follow [SECURITY.md](SECURITY.md).
 
 ## License
 
-License is being finalized; a `LICENSE` file will be added before the public release.
-
----
-
-<div align="center">
-<sub><b>Zero</b> — one terminal · every model</sub>
-</div>
+Zero is released under the [MIT License](LICENSE).

README_ZH.md

@@ -0,0 +1,280 @@
+<p align="center">
+  <img src="docs/assets/zero-logo.png" alt="Zero" width="385">
+</p>
+
+<p align="center"><strong>一个独属于你的终端编码智能体。</strong></p>
+
+<p align="center">
+  <a href="LICENSE"><img alt="license" src="https://img.shields.io/badge/license-MIT-blue"></a>
+  <img alt="Go 1.26.5+" src="https://img.shields.io/badge/Go-1.26.5+-00ADD8?logo=go&logoColor=white">
+  <img alt="25+ providers" src="https://img.shields.io/badge/providers-25+-34E2EA">
+  <br>
+  <a href="README.md">English</a> | <strong>中文</strong>
+</p>
+
+Zero 是一个用于本地终端的 AI 编码智能体。它可以检查仓库、编辑文件、运行命令、使用浏览器/终端辅助工具，并在你选择模型和权限级别的同时保持持久的本地会话。
+
+```bash
+zero
+zero exec "修复 ./pkg 中失败的测试"
+zero exec --output-format stream-json < turns.jsonl
+```
+
+## 为什么选择 Zero
+
+- **使用你想要的模型。** 支持 OpenAI、Anthropic、Gemini、Groq、OpenRouter、DeepSeek、Mistral、xAI、Qwen、Kimi、GitHub Models、Ollama、LM Studio，或任何 OpenAI/Anthropic 兼容端点。
+- **保持控制权。** 文件写入、Shell 命令、网络访问和工作区外写入都经过 Zero 的权限和沙箱策略。
+- **在终端中工作。** TUI 具有模型/提供商选择器、图片输入、斜杠命令、实时计划/工具渲染、回滚滚动、主题以及恢复/分叉支持。
+- **无 TUI 也能工作。** `zero exec` 可脚本化，支持文本/JSON/stream-JSON I/O、隔离的工作树、规范优先运行，以及用于 CI 的有意义的退出码。
+- **保持上下文本地化。** 会话存储在磁盘上，可搜索、可恢复，且 Zero 不会作为遥测数据上传。
+- **可扩展。** 使用 MCP 服务器、技能、插件、钩子和来自同一 CLI 的专业子智能体。
+
+## 安装
+
+### npm
+
+```bash
+npm install -g @gitlawb/zero
+zero
+```
+
+npm 包是一个小型包装器，其平台构建（Linux 和 macOS 的 x64/arm64、Windows 的 x64，包含浏览器/终端控制辅助工具）作为可选依赖直接从 npm registry 安装——没有安装脚本，也不会从 npm 之外下载任何内容。Bun、pnpm 和 yarn 的行为完全一致，无需任何信任或批准步骤。跳过可选依赖的安装（`--omit=optional`）也能工作：只要二进制文件缺失，包装器就会从对应的 GitHub Release 下载。Windows on ARM 通过模拟运行 x64 构建。详见 [docs/NPM_PACKAGING.md](docs/NPM_PACKAGING.md)。
+
+### 安装脚本
+
+Linux/macOS：
+
+```bash
+curl -fsSL https://raw.githubusercontent.com/Gitlawb/zero/main/scripts/install.sh | bash
+```
+
+Windows PowerShell：
+
+```powershell
+irm https://raw.githubusercontent.com/Gitlawb/zero/main/scripts/install.ps1 | iex
+```
+
+### 从源码构建
+
+源码构建需要 Go 1.26.5+。
+
+```bash
+git clone https://github.com/Gitlawb/zero.git
+cd zero
+go run ./cmd/zero
+```
+
+发布安装器和 npm 包需要已发布的 GitHub Release 资源。如果你在首次公开发布之前进行测试，请从源码构建：
+
+```bash
+go build -o zero ./cmd/zero
+```
+
+在 Linux 上，如果你需要原生沙箱，还需要构建沙箱辅助程序：
+
+```bash
+go build -o zero-linux-sandbox ./cmd/zero-linux-sandbox
+go build -o zero-seccomp ./cmd/zero-seccomp   # 可选的兼容性包装器
+```
+
+将 `zero` 和 `zero-linux-sandbox` 放在 `PATH` 上的同一目录中（`~/.local/bin` 是一个好的默认选择）。macOS 不需要额外的辅助二进制文件。Windows 源码构建可以使用主 `zero.exe` 作为沙箱辅助程序；发布包仍然附带独立的 Windows 辅助可执行文件。
+
+更多安装细节：[docs/INSTALL.md](docs/INSTALL.md)。
+
+## 首次运行
+
+启动 TUI：
+
+```bash
+zero
+```
+
+设置向导帮助你选择提供商和模型。你也可以从命令行配置提供商：
+
+```bash
+zero setup
+zero providers list
+zero models list
+zero doctor
+```
+
+对于 API 提供商，在设置之前设置匹配的环境变量或在向导中输入密钥：
+
+```bash
+export OPENAI_API_KEY=sk-...
+export ANTHROPIC_API_KEY=...
+export GEMINI_API_KEY=...
+export LONGCAT_API_KEY=...
+```
+
+要直接配置美团 LongCat（LongCat-2.0），运行：
+
+```bash
+zero providers setup longcat --set-active
+```
+
+对于本地模型，运行 Ollama 或 LM Studio，然后使用 `zero setup` 或 `zero providers detect`。
+
+## 日常使用
+
+### 交互式 TUI
+
+```bash
+zero
+```
+
+常用控制：
+
+| 控制 | 操作 |
+|---|---|
+| `Enter` | 发送提示 |
+| `/` | 打开斜杠命令建议 |
+| `Shift+Tab` | 切换权限模式 |
+| `Ctrl+B` | 显示/隐藏侧边栏 |
+| `Ctrl+C` | 取消或退出 |
+
+常用斜杠命令：
+
+| 命令 | 用途 |
+|---|---|
+| `/model`、`/provider` | 切换活动模型/提供商 |
+| `/spec`、`/plan` | 在构建之前起草和审查计划 |
+| `/image` | 为视觉模型附加图片 |
+| `/resume`、`/rewind` | 继续或回滚本地会话 |
+| `/compact`、`/context` | 管理上下文使用 |
+| `/permissions`、`/tools` | 检查可用工具和策略 |
+| `/add-dir` | 为此会话授予额外的写入目录 |
+| `/theme`、`/doctor`、`/config` | 调整外观和检查设置 |
+
+### 无头 `exec` 模式
+
+```bash
+zero exec "解释 internal/agent/loop.go"
+zero exec --model claude-sonnet-4.5 "重构配置加载器"
+zero exec --use-spec "为 API 客户端添加速率限制"
+zero exec --worktree "在隔离的工作树中尝试迁移"
+zero exec --resume
+zero exec --fork <session-id> "尝试另一种方法"
+```
+
+编程使用：
+
+```bash
+zero exec --input-format stream-json --output-format stream-json < turns.jsonl
+```
+
+Stream-JSON 协议文档在 [docs/STREAM_JSON_PROTOCOL.md](docs/STREAM_JSON_PROTOCOL.md)。
+
+## 安全模型
+
+Zero 旨在使副作用可见。
+
+- 工作区读取默认允许。
+- 文件写入限制在工作区内，除非你授予其他目录。
+- Shell 命令、网络访问、破坏性命令和提权操作需要权限授权。
+- `--add-dir <path>` 和 `/add-dir <path>` 授予额外的写入根目录，而不会给智能体整个文件系统。
+- 不安全/自主模式是显式选择加入。
+- 在 Zero 控制的界面上，密钥会从工具输出和日志中被脱敏。
+
+示例：
+
+```bash
+zero --add-dir ../docs-site
+zero exec --add-dir ../shared "更新两个仓库"
+```
+
+沙箱行为可以通过以下方式检查：
+
+```bash
+zero sandbox policy
+zero sandbox grants list
+```
+
+## Web 和本地控制
+
+Zero 包含本地文件/搜索/编辑/Shell 工具、用于公共 URL 的 `web_fetch`，以及用于额外工具的 MCP 支持。
+
+对于本地开发服务器，使用 Shell 命令（如通过 `exec_command` 的 `curl`），这样正常的沙箱和权限策略就会生效。长时间运行的命令保持附加到后台终端会话，可以从 TUI 中列出或停止。
+
+npm 包还包含本地浏览器/终端工具使用的浏览器和终端辅助包。源码构建可以在它们位于 `PATH` 上或在 Zero 的本地控制设置中配置时使用相同的辅助工具。
+
+## 常用命令
+
+```text
+zero                  交互式 TUI
+zero exec             一次性或脚本化智能体运行
+zero setup            首次运行提供商设置
+zero auth             支持提供商的 OAuth/登录辅助
+zero models           模型注册表和能力
+zero providers        提供商配置和检测
+zero doctor           设置、密钥和连接检查
+zero context          上下文预算报告
+zero repo-map         确定性仓库映射
+zero repo-info        本地仓库摘要
+zero search | find    搜索本地会话历史
+zero sessions         检查、恢复、分叉和回滚会话
+zero spec             管理规范模式草稿
+zero specialist       管理专业子智能体
+zero skills           管理 Markdown 指令技能
+zero plugins          管理插件
+zero hooks            管理生命周期钩子
+zero mcp              管理 MCP 服务器和工具
+zero serve --mcp      通过 MCP stdio 暴露 Zero 工具
+zero sandbox          检查沙箱策略和授权
+zero worktrees        准备隔离的 Git 工作树
+zero verify           检测和运行本地验证检查
+zero changes          检查和提交本地 Git 变更
+zero usage            Token 使用量和估算成本
+zero cron             定时智能体任务
+zero update           检查更新版本
+```
+
+## 外观和无障碍
+
+| 控制 | 效果 |
+|---|---|
+| `NO_COLOR=<任意值>` | 禁用颜色输出 |
+| `ZERO_THEME=<名称>` | 选择启动主题（`auto`、`dark`、`light`，或颜色主题如 `dracula`、`nord`、`gruvbox`、`tokyo-night`、`catppuccin`、`one-dark`、`solarized-dark`、`rose-pine`、`everforest`、`neon`、`solarized-light`、`dune`） |
+| `--theme <名称>` | 从 CLI 选择 TUI 主题（相同名称） |
+| `/theme` | 在 TUI 中打开主题选择器（实时预览；`/theme <名称>` 直接切换） |
+| `ZERO_NO_FADE=1` | 禁用流式淡入动画 |
+
+含义不仅仅依赖于颜色；差异、权限和状态也使用文本或标记符号。
+
+## 开发
+
+```bash
+go test ./...
+go run ./cmd/zero-release build
+go run ./cmd/zero-release smoke
+go run ./cmd/zero-perf-bench
+```
+
+交叉编译示例：
+
+```bash
+go run ./cmd/zero-release build --goos linux --goarch amd64
+go run ./cmd/zero-release build --goos windows --goarch amd64 --output dist/zero.exe
+```
+
+## 文档
+
+- [安装](docs/INSTALL.md)
+- [更新流程](docs/UPDATE.md)
+- [主题](docs/THEMES.md)
+- [Stream-JSON 协议](docs/STREAM_JSON_PROTOCOL.md)
+- [专家](docs/SPECIALISTS.md)
+- [GitHub Action](docs/GITHUB_ACTION.md)
+- [基准测试](docs/BENCHMARK.md)
+- [性能](docs/PERFORMANCE.md)
+- [智能体评估](docs/AGENT_EVALS.md)
+
+## 贡献
+
+欢迎贡献。阅读 [CONTRIBUTING.md](CONTRIBUTING.md)，运行相关测试，然后提交一个聚焦的拉取请求。
+
+安全报告应遵循 [SECURITY.md](SECURITY.md)。
+
+## 许可证
+
+Zero 基于 [MIT 许可证](LICENSE) 发布。

SUPPORT.md

@@ -0,0 +1,36 @@
+# Getting help with Zero
+
+## Questions, setup help, configuration
+
+Ask in [GitHub Discussions](https://github.com/Gitlawb/zero/discussions). Use
+[Q&A](https://github.com/Gitlawb/zero/discussions/categories/q-a) so answers can
+be marked and found by the next person.
+
+The fastest way to a useful answer is to include:
+
+- `zero --version` output
+- your OS (Windows, macOS, Linux) and how you installed Zero (npm, install
+  script, source)
+- the provider and model you are using (`zero providers list`)
+- `zero doctor` output for setup problems
+- the exact command you ran and the exact error text
+
+## Bugs
+
+If Zero does something wrong (crashes, wrong output, broken behavior), file a
+[bug report](https://github.com/Gitlawb/zero/issues/new/choose). If you are not
+sure whether it is a bug or a setup problem, start in Discussions and we will
+promote it to an issue if it is one.
+
+## Feature ideas
+
+Start a thread in
+[Ideas](https://github.com/Gitlawb/zero/discussions/categories/ideas). Zero is
+in controlled development (see CONTRIBUTING.md), so discussing an idea first is
+also the path to getting a PR accepted.
+
+## Security issues
+
+Never file a public issue or discussion. Report privately via
+[GitHub Security Advisories](https://github.com/Gitlawb/zero/security/advisories/new)
+(see SECURITY.md).

action.yml

@@ -27,14 +27,14 @@ inputs:
   api-key:
     description: >-
       The provider API key. Pass it from a repository or organization secret;
-      it is exported into the step environment and never written to the log.
+      it is exported only for the ZERO step and never written to the log.
     required: false
     default: ""
   api-key-env:
     description: >-
       Name of the environment variable the chosen provider reads its key from
       (for example OPENAI_API_KEY or ANTHROPIC_API_KEY). When set together with
-      api-key, that variable is exported for the run. Leave the action
+      api-key, that variable is exported only for the ZERO step. Leave the action
       provider-agnostic by passing the env name your provider expects rather
       than hardcoding one here.
     required: false
@@ -180,6 +180,7 @@ runs:
         ZERO_INPUT_ADD_DIR: ${{ inputs.add-dir }}
         ZERO_INPUT_WORKTREE: ${{ inputs.worktree }}
         ZERO_INPUT_OUTPUT_FORMAT: ${{ inputs.output-format }}
+        ACTION_PATH: ${{ github.action_path }}
         RUNNER_TEMP: ${{ runner.temp }}
       run: |
         set -euo pipefail
@@ -192,10 +193,9 @@ runs:
             echo "::error::api-key-env must be a valid environment variable name." >&2
             exit 2
           fi
-          # Export in-process so the zero invocation in this same run block sees
-          # it, and append to GITHUB_ENV so any later steps inherit it too.
+          # Export in-process for this Run ZERO step so zero can read it; do
+          # not persist provider credentials to later job steps.
           export "${ZERO_INPUT_API_KEY_ENV}=${ZERO_INPUT_API_KEY}"
-          printf '%s=%s\n' "${ZERO_INPUT_API_KEY_ENV}" "${ZERO_INPUT_API_KEY}" >> "${GITHUB_ENV}"
         fi
         if [ -n "${ZERO_INPUT_PROVIDER}" ]; then
           export ZERO_PROVIDER="${ZERO_INPUT_PROVIDER}"
@@ -248,10 +248,9 @@ runs:
         echo "exit-code=${code}" >> "${GITHUB_OUTPUT}"
         echo "output-file=${output_file}" >> "${GITHUB_OUTPUT}"
 
-        # Best-effort one-line summary: last non-empty line of the captured output.
-        summary="$(grep -v '^[[:space:]]*$' "${output_file}" | tail -n 1 || true)"
-        # Keep the summary single-line and bounded for the step output.
-        summary="$(printf '%s' "${summary}" | cut -c1-280)"
+        # Best-effort one-line summary. Structured output ends with lifecycle
+        # metadata, so parse final/error events instead of tailing stdout.
+        summary="$(node "${ACTION_PATH}/scripts/action-summary.mjs" "${ZERO_INPUT_OUTPUT_FORMAT}" "${output_file}" || true)"
         {
           echo "summary<<ZERO_EOF"
           printf '%s\n' "${summary}"

bin/zero.js

@@ -2,6 +2,7 @@
 
 import { spawnSync } from 'node:child_process';
 import { existsSync } from 'node:fs';
+import { createRequire } from 'node:module';
 import { dirname, join } from 'node:path';
 import { fileURLToPath } from 'node:url';
 
@@ -9,18 +10,283 @@ function zeroBinaryName(platform = process.platform) {
   return platform === 'win32' ? 'zero.exe' : 'zero';
 }
 
-const packageRoot = dirname(dirname(fileURLToPath(import.meta.url)));
-const nativePath = join(packageRoot, zeroBinaryName());
+function helperShimNames(name, platform = process.platform) {
+  if (platform === 'win32') {
+    return [`${name}.cmd`, `${name}.exe`, name];
+  }
+  return [name];
+}
 
-if (!existsSync(nativePath)) {
-  console.error(
-    '[zero] No native binary found next to the npm wrapper. Reinstall the zero package or run `go run ./cmd/zero-release build` from the repository.'
+function commandForShim(path, platform = process.platform) {
+  if (platform === 'win32' && path.toLowerCase().endsWith('.cmd')) {
+    return {
+      command: process.env.ComSpec || 'cmd.exe',
+      prefixArgs: ['/d', '/s', '/c', `"${path.replace(/"/g, '""')}"`],
+    };
+  }
+  return { command: path, prefixArgs: [] };
+}
+
+function resolveHelper(packageRoot, name) {
+  const binDir = join(packageRoot, 'node_modules', '.bin');
+  for (const shimName of helperShimNames(name)) {
+    const candidate = join(binDir, shimName);
+    if (!existsSync(candidate)) continue;
+    return {
+      ...commandForShim(candidate),
+      pathPrepend: [binDir],
+    };
+  }
+  return null;
+}
+
+// Legacy path: helpers installed as npm dependencies of the wrapper. Current
+// installs vendor the helpers inside the platform package instead (a helpers/
+// directory next to the binary, found by the Go side on its own), so this
+// manifest is usually empty — it only matters for a downloaded-binary install
+// whose wrapper still carries helper dependencies.
+function localControlHelperManifest(packageRoot) {
+  const helpers = {};
+  for (const name of ['agent-browser', 'tuistory']) {
+    const helper = resolveHelper(packageRoot, name);
+    if (helper) helpers[name] = helper;
+  }
+  if (Object.keys(helpers).length === 0) return '';
+  return JSON.stringify({ version: 1, helpers });
+}
+
+// Mirrors internal/cli/observability.go parseDoctorArgs + writeDoctorHelp so the
+// wrapper's missing-binary doctor fallback matches the Go doctor CLI surface.
+const DOCTOR_HELP = `Usage:
+  zero doctor [flags]
+
+Runs Go backend health checks for config and provider setup.
+
+Flags:
+      --json            Print JSON report
+      --connectivity    Include provider endpoint connectivity probe when available
+  -h, --help            Show this help
+`;
+
+const EXIT_USAGE = 2;
+
+function installedByBun() {
+  if (process.env.ZERO_WRAPPER_SIMULATE_BUN === '1') {
+    return true;
+  }
+  return process.execPath.includes('bun') || !!process.versions?.bun;
+}
+
+function bunRecoveryParagraph() {
+  return (
+    'You installed with Bun, which does not run dependency lifecycle scripts\n' +
+    'by default. Trust the package to run the blocked postinstall:\n' +
+    '  bun pm trust @gitlawb/zero       (project install)\n' +
+    '  bun pm -g trust @gitlawb/zero    (global install)\n' +
+    'On Bun versions without `bun pm trust`, add\n' +
+    '  "trustedDependencies": ["@gitlawb/zero"]\n' +
+    'to your project package.json and reinstall.\n' +
+    '\n'
   );
+}
+
+function buildFromSourceParagraph() {
+  return (
+    'If this platform has no prebuilt binary, build from source:\n' +
+    'https://github.com/Gitlawb/zero (go run ./cmd/zero, requires Go 1.26+).'
+  );
+}
+
+function missingBinaryContextParagraph() {
+  return (
+    'Normally npm installs it as an optional dependency of @gitlawb/zero\n' +
+    `(@gitlawb/zero-${process.platform}-${process.arch}), and the wrapper can\n` +
+    'also download it from the GitHub Release when it is missing.'
+  );
+}
+
+function missingNativeRecoveryParagraphs(postinstallScript) {
+  const ranByBun = installedByBun();
+  return (
+    missingBinaryContextParagraph() +
+    '\n' +
+    '\n' +
+    'Things to try:\n' +
+    '  - reinstall without omitting optional dependencies:\n' +
+    '      npm install -g @gitlawb/zero\n' +
+    '  - run the downloader manually (needs write access to the package dir):\n' +
+    `      node "${postinstallScript}"\n` +
+    '\n' +
+    (ranByBun ? bunRecoveryParagraph() : '') +
+    buildFromSourceParagraph()
+  );
+}
+
+function formatGenericMissingNativeBinaryMessage(postinstallScript) {
+  return (
+    '[zero] No native binary is available for this install.\n' +
+    missingNativeRecoveryParagraphs(postinstallScript)
+  );
+}
+
+function parseDoctorArgs(args) {
+  let json = false;
+  for (const arg of args) {
+    switch (arg) {
+      case '-h':
+      case '--help':
+      case 'help':
+        return { kind: 'help' };
+      case '--json':
+        json = true;
+        break;
+      case '--connectivity':
+        break;
+      default:
+        return { kind: 'error', message: `unknown doctor flag ${JSON.stringify(arg)}` };
+    }
+  }
+  return { kind: 'run', json };
+}
+
+function missingNativeDoctorJSONReport(postinstallScript) {
+  return {
+    generatedAt: new Date().toISOString().replace(/\.\d{3}Z$/, 'Z'),
+    ok: false,
+    checks: [
+      {
+        id: 'runtime.go',
+        label: 'Go runtime',
+        status: 'fail',
+        message: 'Native zero binary is missing next to the npm wrapper.',
+        details: {
+          remedy: `node "${postinstallScript}"`,
+        },
+      },
+    ],
+  };
+}
+
+function missingNativeDoctorTextReport(postinstallScript) {
+  return (
+    'Zero doctor report (' +
+    new Date().toISOString() +
+    ')\n' +
+    'Overall: fail\n' +
+    '[fail] runtime.go - Native zero binary is missing next to the npm wrapper.\n' +
+    '  remedy: node "' +
+    postinstallScript +
+    '"\n' +
+    '\n' +
+    missingNativeRecoveryParagraphs(postinstallScript)
+  );
+}
+
+// The platform payload is a version of @gitlawb/zero itself, installed under
+// an npm: alias (see docs/NPM_PACKAGING.md). The alias name is derived from
+// process.platform/process.arch, so an unsupported platform simply fails to
+// resolve and we fall through to the downloader.
+function platformPackageBinary() {
+  const aliasedName = `@gitlawb/zero-${process.platform}-${process.arch}`;
+  let manifestPath;
+  try {
+    manifestPath = createRequire(import.meta.url).resolve(`${aliasedName}/package.json`);
+  } catch {
+    return null;
+  }
+  const candidate = join(dirname(manifestPath), zeroBinaryName());
+  return existsSync(candidate) ? candidate : null;
+}
+
+// Mirrors the release matrix (internal/release/release.go): the platforms a
+// GitHub Release asset exists for. android maps to the linux asset; there is
+// no windows-arm64 asset (Windows on ARM runs the x64 build under emulation).
+function releaseAssetAvailable() {
+  const platform = { linux: 'linux', android: 'linux', darwin: 'macos', win32: 'windows' }[
+    process.platform
+  ];
+  const arch = process.arch === 'x64' || process.arch === 'arm64' ? process.arch : null;
+  if (!platform || !arch) return false;
+  return !(platform === 'windows' && arch === 'arm64');
+}
+
+// Fallback for installs without a platform package (--omit=optional, package
+// managers that skip optionalDependencies): fetch the binary next to the
+// wrapper with the same checksum-verified downloader that used to run as
+// postinstall. Deliberately NOT cached on failure — a transient network error
+// self-heals on the next run — so this retries on every invocation until a
+// binary is in place. Platforms with no release asset skip the attempt.
+function downloadMissingBinary(packageRoot) {
+  const downloadScript = join(packageRoot, 'scripts', 'postinstall.mjs');
+  if (!existsSync(downloadScript) || !releaseAssetAvailable()) return;
+  console.error(
+    '[zero] no platform package installed — fetching the native binary from the GitHub Release (retried on each run until it succeeds).',
+  );
+  spawnSync(process.execPath, [downloadScript], {
+    stdio: ['ignore', 'inherit', 'inherit'],
+  });
+}
+
+const packageRoot = dirname(dirname(fileURLToPath(import.meta.url)));
+
+function resolveNativeBinary() {
+  const fromPlatformPackage = platformPackageBinary();
+  if (fromPlatformPackage) return fromPlatformPackage;
+  const downloaded = join(packageRoot, zeroBinaryName());
+  if (existsSync(downloaded)) return downloaded;
+  downloadMissingBinary(packageRoot);
+  return existsSync(downloaded) ? downloaded : null;
+}
+
+const nativePath = resolveNativeBinary();
+
+if (!nativePath) {
+  const postinstallScript = join(packageRoot, 'scripts', 'postinstall.mjs');
+
+  // `zero doctor` is the diagnostic command: when the native binary is missing
+  // it's the one invocation that must NOT bail with the generic wrapper error,
+  // because that's exactly the blind alley issue #405 reports. Instead, surface
+  // a doctor-shaped FAIL line for the runtime so the user's diagnostic finds the
+  // real cause. We branch on a literal 'doctor' subcommand only (matching `zero
+  // doctor` and `zero doctor --connectivity`), preserving the existing bail for
+  // every other invocation (exec, providers list, TUI, --version, etc.).
+  const argv = process.argv.slice(2);
+  const isDoctor = argv.length > 0 && argv[0] === 'doctor';
+  if (isDoctor) {
+    const parsed = parseDoctorArgs(argv.slice(1));
+    if (parsed.kind === 'help') {
+      process.stdout.write(DOCTOR_HELP);
+      process.exit(0);
+    }
+    if (parsed.kind === 'error') {
+      process.stderr.write(`[zero] ${parsed.message}\n`);
+      process.exit(EXIT_USAGE);
+    }
+
+    if (parsed.json) {
+      process.stdout.write(JSON.stringify(missingNativeDoctorJSONReport(postinstallScript), null, 2) + '\n');
+      process.exit(1);
+    }
+
+    process.stdout.write(missingNativeDoctorTextReport(postinstallScript) + '\n');
+    process.exit(1);
+  }
+
+  console.error(formatGenericMissingNativeBinaryMessage(postinstallScript));
   process.exit(1);
 }
 
+const env = { ...process.env };
+const localControlHelpers = localControlHelperManifest(packageRoot);
+if (localControlHelpers) {
+  env.ZERO_LOCAL_CONTROL_HELPERS = localControlHelpers;
+} else {
+  delete env.ZERO_LOCAL_CONTROL_HELPERS;
+}
+
 const child = spawnSync(nativePath, process.argv.slice(2), {
   stdio: 'inherit',
+  env,
 });
 
 if (child.error) {

cmd/zero-perf-bench/main.go

@@ -39,6 +39,9 @@ func run(args []string, getenv func(string) string, stdout io.Writer, stderr io.
 	if len(args) > 0 && args[0] == "tasks" {
 		return runTasksCommand(args[1:], getenv, stdout, stderr)
 	}
+	if len(args) > 0 && args[0] == "turn" {
+		return runTurnCommand(args[1:], getenv, stdout, stderr)
+	}
 	options, err := parseArgs(args, getenv)
 	if err != nil {
 		_, _ = fmt.Fprintln(stderr, err.Error())
@@ -206,6 +209,7 @@ func helpText() string {
 	return strings.Join([]string{
 		"Usage: zero-perf-bench [options]",
 		"       zero-perf-bench tasks [options]   (Terminal-Bench-style task harness; see `tasks --help`)",
+		"       zero-perf-bench turn  [options]   (per-turn tracing benchmark; see `turn --help`)",
 		"",
 		"Options:",
 		"  --iterations <n>             Measured samples to collect (default: 5)",

cmd/zero-perf-bench/main_test.go

@@ -4,6 +4,8 @@ import (
 	"bytes"
 	"strings"
 	"testing"
+
+	"github.com/Gitlawb/zero/internal/perfbench"
 )
 
 func TestParseArgsUsesCliAndEnvOverrides(t *testing.T) {
@@ -92,3 +94,18 @@ func TestHelpText(t *testing.T) {
 func emptyEnv(string) string {
 	return ""
 }
+
+func TestTurnExitCodeFailsWhenEveryTaskErrored(t *testing.T) {
+	var stderr bytes.Buffer
+	allErrored := perfbench.TurnBenchResult{TasksAttempted: 3, TasksErrored: 3}
+	if code := turnExitCode(allErrored, &stderr); code != 1 {
+		t.Fatalf("exit=%d, want 1 when every task errored", code)
+	}
+	if !strings.Contains(stderr.String(), "no accepted benchmark sample") {
+		t.Fatalf("stderr missing explanation: %q", stderr.String())
+	}
+	partial := perfbench.TurnBenchResult{TasksAttempted: 3, TasksErrored: 2}
+	if code := turnExitCode(partial, &stderr); code != 0 {
+		t.Fatalf("exit=%d, want 0 for partial errors (summary surfaces them)", code)
+	}
+}

cmd/zero-perf-bench/turn.go

@@ -0,0 +1,279 @@
+package main
+
+import (
+	"bytes"
+	"context"
+	"fmt"
+	"io"
+	"os"
+	"path/filepath"
+	"strings"
+
+	"github.com/Gitlawb/zero/internal/execprofile"
+	"github.com/Gitlawb/zero/internal/perfbench"
+)
+
+// turnOptions configures the `zero-perf-bench turn` subcommand: the per-turn
+// benchmark harness that runs ZERO headlessly with --trace, parses each turn's
+// NDJSON trace, and records per-span latency plus the top controllable latency
+// sources — the Phase 0 baseline's "do not proceed until" criterion.
+type turnOptions struct {
+	SuitePath   string
+	Model       string
+	Mode        string
+	ExecProfile string
+	SelfCorrect bool
+	Binary      string
+	Iterations  int
+	Version     string
+	Commit      string
+	Output      string
+	JSON        bool
+	DryRun      bool
+	Help        bool
+}
+
+func runTurnCommand(args []string, getenv func(string) string, stdout io.Writer, stderr io.Writer) int {
+	options, err := parseTurnArgs(args, getenv)
+	if err != nil {
+		_, _ = fmt.Fprintln(stderr, err.Error())
+		return 2
+	}
+	if options.Help {
+		_, _ = fmt.Fprint(stdout, turnHelpText())
+		return 0
+	}
+
+	set, err := perfbench.LoadTaskSet(options.SuitePath)
+	if err != nil {
+		_, _ = fmt.Fprintln(stderr, "[zero] Turn benchmark failed: "+err.Error())
+		return 1
+	}
+
+	// The dry-run path records a zero-iteration run without a binary, so the
+	// manifest loads and the report path is exercised in CI without a model.
+	if options.DryRun {
+		_, _ = fmt.Fprintln(stdout, "[zero] turn benchmark: dry run (no agent invoked)")
+		return 0
+	}
+
+	binary, err := perfbench.ResolveBinary(options.Binary)
+	if err != nil {
+		_, _ = fmt.Fprintln(stderr, "[zero] Turn benchmark failed: "+err.Error())
+		return 2
+	}
+
+	result, err := perfbench.RunTurnBench(context.Background(), set, perfbench.TurnBenchConfig{
+		Model:       options.Model,
+		Mode:        options.Mode,
+		ExecProfile: options.ExecProfile,
+		SelfCorrect: options.SelfCorrect,
+		Version:     options.Version,
+		Commit:      options.Commit,
+		Iterations:  options.Iterations,
+		Runner:      perfbench.NewTurnExecRunner(binary),
+	})
+	if err != nil {
+		_, _ = fmt.Fprintln(stderr, "[zero] Turn benchmark failed: "+err.Error())
+		return 1
+	}
+
+	if options.Output != "" {
+		if err := writeTurnReport(options.Output, result); err != nil {
+			_, _ = fmt.Fprintln(stderr, "[zero] Turn benchmark failed: "+err.Error())
+			return 1
+		}
+	}
+	if options.JSON {
+		if err := perfbench.WriteTurnBenchJSON(stdout, result); err != nil {
+			_, _ = fmt.Fprintln(stderr, "[zero] Turn benchmark failed: "+err.Error())
+			return 1
+		}
+		return turnExitCode(result, stderr)
+	}
+	_, _ = fmt.Fprintln(stdout, perfbench.FormatTurnBenchSummary(result))
+	return turnExitCode(result, stderr)
+}
+
+// turnExitCode fails the command when every attempted task errored (no
+// iteration produced an accepted benchmark sample): such a report measures
+// nothing, and exiting 0 would let a broken configuration (missing binary, bad
+// path, failing harness step) pass as a clean baseline. Partial errors keep
+// exit 0 — the summary surfaces them loudly and the surviving samples are
+// still valid measurements.
+func turnExitCode(result perfbench.TurnBenchResult, stderr io.Writer) int {
+	if result.TasksAttempted > 0 && result.TasksErrored == result.TasksAttempted {
+		_, _ = fmt.Fprintln(stderr, "[zero] Turn benchmark failed: every task errored with no accepted benchmark sample (see warnings); the report contains no valid measurements")
+		return 1
+	}
+	return 0
+}
+
+func parseTurnArgs(args []string, getenv func(string) string) (turnOptions, error) {
+	options := turnOptions{
+		Iterations: 1,
+		Version:    strings.TrimSpace(getenv("ZERO_BENCH_VERSION")),
+		Commit:     strings.TrimSpace(getenv("ZERO_BENCH_COMMIT")),
+	}
+	for index := 0; index < len(args); index++ {
+		arg := args[index]
+		flag, inlineValue := splitFlagValue(arg)
+		switch flag {
+		case "--suite":
+			value, next, err := readOptionValue(args, inlineValue, index, flag)
+			if err != nil {
+				return options, err
+			}
+			options.SuitePath = value
+			index = next
+		case "--model":
+			value, next, err := readOptionValue(args, inlineValue, index, flag)
+			if err != nil {
+				return options, err
+			}
+			options.Model = value
+			index = next
+		case "--mode":
+			value, next, err := readOptionValue(args, inlineValue, index, flag)
+			if err != nil {
+				return options, err
+			}
+			options.Mode = value
+			index = next
+		case "--exec-profile":
+			value, next, err := readOptionValue(args, inlineValue, index, flag)
+			if err != nil {
+				return options, err
+			}
+			options.ExecProfile = value
+			index = next
+		case "--binary":
+			value, next, err := readOptionValue(args, inlineValue, index, flag)
+			if err != nil {
+				return options, err
+			}
+			options.Binary = value
+			index = next
+		case "--iterations":
+			value, next, err := readOptionValue(args, inlineValue, index, flag)
+			if err != nil {
+				return options, err
+			}
+			parsed, err := parsePositiveInteger(flag, value)
+			if err != nil {
+				return options, err
+			}
+			options.Iterations = parsed
+			index = next
+		case "--version":
+			value, next, err := readOptionValue(args, inlineValue, index, flag)
+			if err != nil {
+				return options, err
+			}
+			options.Version = value
+			index = next
+		case "--commit":
+			value, next, err := readOptionValue(args, inlineValue, index, flag)
+			if err != nil {
+				return options, err
+			}
+			options.Commit = value
+			index = next
+		case "--output":
+			value, next, err := readOptionValue(args, inlineValue, index, flag)
+			if err != nil {
+				return options, err
+			}
+			options.Output = value
+			index = next
+		case "--self-correct":
+			if strings.Contains(arg, "=") {
+				return options, fmt.Errorf("%s does not accept a value", flag)
+			}
+			options.SelfCorrect = true
+		case "--json":
+			if strings.Contains(arg, "=") {
+				return options, fmt.Errorf("%s does not accept a value", flag)
+			}
+			options.JSON = true
+		case "--dry-run":
+			if strings.Contains(arg, "=") {
+				return options, fmt.Errorf("%s does not accept a value", flag)
+			}
+			options.DryRun = true
+		case "-h", "--help":
+			if strings.Contains(arg, "=") {
+				return options, fmt.Errorf("%s does not accept a value", flag)
+			}
+			options.Help = true
+		default:
+			return options, fmt.Errorf("unknown option: %s", arg)
+		}
+	}
+	if options.Help {
+		return options, nil
+	}
+	if strings.TrimSpace(options.SuitePath) == "" {
+		return options, fmt.Errorf("--suite is required")
+	}
+	if strings.TrimSpace(options.Model) == "" && !options.DryRun {
+		return options, fmt.Errorf("--model is required (or pass --dry-run)")
+	}
+	// Validate the profile here, before anything spawns: an unknown name would
+	// otherwise make every child exit with a usage error in milliseconds, and
+	// those near-zero walls would be recorded as valid latency samples in an
+	// exit-0 report — exactly the misread-as-improvement trap the harness
+	// closed for spawn failures. Normalizing to the catalog name also keeps
+	// the stamped execProfile canonical (Lookup is case-insensitive), so two
+	// captures of the same posture always compare equal.
+	if raw := strings.TrimSpace(options.ExecProfile); raw != "" {
+		profile, ok := execprofile.Lookup(raw)
+		if !ok {
+			return options, fmt.Errorf("unknown execution profile %q for --exec-profile. Valid profiles: %s", raw, strings.Join(execprofile.Names(), ", "))
+		}
+		options.ExecProfile = profile.Name
+	}
+	return options, nil
+}
+
+func writeTurnReport(path string, result perfbench.TurnBenchResult) error {
+	if dir := filepath.Dir(path); dir != "" {
+		if err := os.MkdirAll(dir, 0o755); err != nil {
+			return err
+		}
+	}
+	var buffer bytes.Buffer
+	if err := perfbench.WriteTurnBenchJSON(&buffer, result); err != nil {
+		return err
+	}
+	return os.WriteFile(path, buffer.Bytes(), 0o644)
+}
+
+func turnHelpText() string {
+	return strings.Join([]string{
+		"Usage: zero-perf-bench turn [options]",
+		"",
+		"Runs ZERO headlessly with --trace against a per-turn benchmark task set and",
+		"records per-span latency plus the top controllable latency sources (the",
+		"Phase 0 baseline's \"do not proceed until\" criterion). Each task is a fresh",
+		"`zero exec` process, so iterations are cold-start samples; a warm path needs",
+		"an in-process runner (future work).",
+		"",
+		"Options:",
+		"  --suite <path>      Task set JSON file (required)",
+		"  --model <model>     Model to run (required unless --dry-run)",
+		"  --mode <name>       Exec mode preset to apply",
+		"  --exec-profile <name>",
+		"                      Execution profile for every task (fast|balanced|thorough);",
+		"                      forwarded to zero exec and stamped into the result",
+		"  --self-correct      Enable the post-edit verify-and-correct loop",
+		"  --binary <path>     Path to the `zero` binary (default: zero on PATH / repo root)",
+		"  --iterations <n>    Times to run each task (default: 1)",
+		"  --version <v>       Record the ZERO version (default: $ZERO_BENCH_VERSION)",
+		"  --commit <sha>      Record the ZERO commit (default: $ZERO_BENCH_COMMIT)",
+		"  --output <path>     Write the JSON result to path",
+		"  --json              Print only the JSON result",
+		"  --dry-run           Load the manifest and exit without invoking the agent",
+		"  -h, --help          Show this help",
+	}, "\n") + "\n"
+}

cmd/zero-perf-bench/turn_test.go

@@ -0,0 +1,65 @@
+package main
+
+import (
+	"strings"
+	"testing"
+)
+
+func TestParseTurnArgsExecProfile(t *testing.T) {
+	getenv := func(string) string { return "" }
+	options, err := parseTurnArgs([]string{
+		"--suite", "suite.json",
+		"--model", "m",
+		"--exec-profile", "fast",
+	}, getenv)
+	if err != nil {
+		t.Fatalf("parseTurnArgs: %v", err)
+	}
+	if options.ExecProfile != "fast" {
+		t.Fatalf("ExecProfile = %q, want fast", options.ExecProfile)
+	}
+
+	// Inline form.
+	options, err = parseTurnArgs([]string{
+		"--suite=suite.json",
+		"--model=m",
+		"--exec-profile=thorough",
+	}, getenv)
+	if err != nil {
+		t.Fatalf("parseTurnArgs inline: %v", err)
+	}
+	if options.ExecProfile != "thorough" {
+		t.Fatalf("ExecProfile = %q, want thorough", options.ExecProfile)
+	}
+}
+
+// The bench validates the profile BEFORE spawning anything: an unknown name
+// would otherwise make every child exit with a usage error in milliseconds
+// and record those near-zero walls as valid latency samples in an exit-0
+// report. Lookup normalization also keeps the stamped name canonical so two
+// captures of the same posture compare equal.
+func TestParseTurnArgsExecProfileValidatesAndNormalizes(t *testing.T) {
+	getenv := func(string) string { return "" }
+
+	options, err := parseTurnArgs([]string{
+		"--suite", "suite.json", "--model", "m", "--exec-profile", "FAST",
+	}, getenv)
+	if err != nil {
+		t.Fatalf("parseTurnArgs: %v", err)
+	}
+	if options.ExecProfile != "fast" {
+		t.Fatalf("ExecProfile = %q, want the canonical fast", options.ExecProfile)
+	}
+
+	_, err = parseTurnArgs([]string{
+		"--suite", "suite.json", "--model", "m", "--exec-profile", "blanced",
+	}, getenv)
+	if err == nil {
+		t.Fatal("an unknown profile must fail parse, before anything spawns")
+	}
+	for _, name := range []string{"balanced", "fast", "thorough"} {
+		if !strings.Contains(err.Error(), name) {
+			t.Fatalf("error must list %q, got %q", name, err.Error())
+		}
+	}
+}

docs/EXTENDING.md

@@ -1,14 +1,8 @@
----
-description: Extending Zero — how to write AGENTS.md files, custom specialists, skills, hooks, MCP servers, and plugins for the open-source CLI coding agent.
-globs: "*.go, *.js, *.md, *.json, *.toml, *.yaml, *.yml"
-alwaysApply: false
----
-
 # Extending Zero
 
-Zero is an open-source terminal coding agent in the same family as Claude Code, Codex, and Droid. Out of the box it does the obvious things — read, edit, run, search — but the design point of the project is that **every surface is configurable**. This document is the user-facing guide for that configuration.
+Zero is an open-source terminal coding agent. Out of the box it does the obvious things — read, edit, run, search — but the design point of the project is that **every surface is configurable**. This document is the user-facing guide for that configuration.
 
-If you only want to *use* Zero, the [README](README.md) is enough. This page is for the other three jobs:
+If you only want to *use* Zero, the [README](../README.md) is enough. This page is for the other three jobs:
 
 1. Tell the agent about *your* project (drop an `AGENTS.md` in your repo).
 2. Add new specialist sub-agents.
@@ -26,7 +20,7 @@ Accepted file names, in priority order at each level:
 | `./ZERO.md` | Brand-specific alias. Same format, lower priority. |
 | `./.zero/AGENTS.md` | Project-local, hidden, gitignored. Personal notes that stay out of git. |
 
-Matching is **case-insensitive** on the basename, so `AGENTS.md`, `Agents.md`, and `agents.md` resolve to the same file on Windows and macOS. The git-tracked filename in this repo is `AGENTS.MD` (uppercase `MD`) — keep that on case-sensitive filesystems (Linux, the WSL filesystem, or a CI runner) to match what the loader looks for.
+Matching is **case-insensitive** on the basename, so `AGENTS.md`, `Agents.md`, and `agents.md` resolve to the same file on Windows and macOS. The git-tracked filename in this repo is `AGENTS.md` — keep that on case-sensitive filesystems (Linux, the WSL filesystem, or a CI runner) to match what the loader looks for.
 
 Both files use the same format. YAML frontmatter is optional; the markdown body is loaded as instructions for the agent. Zero reads the file once at session start, so changes take effect on the next `zero` launch — not mid-session.
 
@@ -47,6 +41,12 @@ Tips:
 - In a monorepo, drop a narrower `AGENTS.md` in each sub-tree (e.g. `services/api/AGENTS.md`). Zero picks those up automatically when you launch from inside the sub-tree.
 - A YAML frontmatter block (`---\n...\n---`) at the top is preserved verbatim in the injected prompt but is not parsed for `globs:` or `alwaysApply:` scoping today — keep the body self-contained.
 
+### Personal guidelines, across every project
+
+For preferences that follow *you*, not a specific repo (tone, tooling habits, workflow), drop a `ZERO.md` in your user config directory: `~/.config/zero/ZERO.md` on Linux/macOS, `%AppData%\zero\ZERO.md` on Windows — the same directory as `config.json` and your personal specialists. Same format and 8 KiB cap as the project files above, and the same case-insensitive basename match.
+
+This file is injected as its own `## User guidelines` section, before the project's `AGENTS.md`/`ZERO.md`, and is labeled as personal preference in the prompt: project guidelines are the later, more specific instruction and take precedence over it when the two conflict.
+
 ## 2. Custom specialists
 
 Specialists are Zero's sub-agents. Three scopes, in priority order:
@@ -94,22 +94,32 @@ zero specialist delete api-reviewer --project
 zero specialist path                       # prints the resolved specialists directory
 ```
 
-The full format spec (frontmatter fields, tool scopes, prompt conventions) is in [`docs/SPECIALISTS.md`](docs/SPECIALISTS.md).
+The full format spec (frontmatter fields, tool scopes, prompt conventions) is in [`docs/SPECIALISTS.md`](SPECIALISTS.md).
 
 > **Roadmap.** An in-UI specialist manager (create / edit / delete / preview) is on the backlog. Today you use the `zero specialist` CLI subcommands above.
 
 ## 3. Skills
 
-Skills are markdown instruction packs the agent can pull in on demand. Each skill is a directory containing a `SKILL.md`. Skills are **user-level only** in this version — there's no project-scoped skill directory yet, so anything you want shared with the team goes in `AGENTS.md` (section 1) or as a hook (section 4).
+Skills are markdown instruction packs the agent can pull in on demand. Each skill is a directory containing a `SKILL.md`. Standalone project skill directories are not supported in this version (shared project-wide skills must go in `AGENTS.md` or as a hook). However, project plugins (section 6) may bundle skills, which are merged into the active run.
 
-Discovery root: `$ZERO_SKILLS_DIR` → `$XDG_DATA_HOME/zero/skills` → `~/.local/share/zero/skills/`. A missing directory is fine — Zero just reports "no skills".
+Discovery roots (earlier wins on name collisions):
 
-```
-~/.local/share/zero/skills/
+1. **Primary Zero dir** — `$ZERO_SKILLS_DIR` if set, else `$XDG_DATA_HOME/zero/skills`, else `~/.local/share/zero/skills/`
+2. **Shared multi-agent dir** — `~/.agents/skills/` when present (read-only discovery; never an install target)
+3. **Plugin skill roots** — skills bundled by active plugins (section 6)
+
+A missing directory is fine — Zero just omits it. Management commands (`zero skills add` / `remove` / `lock`) always write to the primary Zero dir only; `list` / `info` search primary + `~/.agents/skills`.
+
+```text
+~/.local/share/zero/skills/          # primary (install/remove/lock target)
   run-benchmarks/
     SKILL.md
   write-changelog/
     SKILL.md
+
+~/.agents/skills/                    # optional shared multi-agent root
+  shared-review/
+    SKILL.md
 ```
 
 `SKILL.md` format:
@@ -126,9 +136,9 @@ description: Run the project's benchmark suite and summarize the deltas.
 3. Report any regression > 5% with the function name and the previous value.
 ```
 
-Only `name` and `description` are recognized in the frontmatter today. The `name` defaults to the directory name. If two skills declare the same name, the one in the lexicographically-first directory wins — duplicates are dropped silently. Plugin-declared skills (section 6) are not yet merged into the loader, so a `skills:` entry inside a plugin's `plugin.json` is not visible to the `skill` tool today.
+Only `name` and `description` are recognized in the frontmatter today. The `name` defaults to the directory name. Within a single skills root, duplicate names are resolved by lexicographic directory order. Across roots, Zero loads the primary directory first, then `~/.agents/skills`, then plugin skill roots; earlier roots win name collisions. Plugin-declared skills (section 6) are merged into the active agent run at plugin activation time, so bundled skills appear in the available skills list and can be loaded with the `skill` tool.
 
-The `skill` core tool lets the agent load any discovered skill by name.
+The `skill` tool lets the agent load any discovered skill by name (primary, agents, or plugin).
 
 ## 4. Hooks
 
@@ -229,14 +239,24 @@ Servers are merged from user and project configs (project wins on conflicts). To
 - A `--header` value produced by command substitution (`"Authorization=Bearer $(print-token)"`) in a private shell config that you keep out of git.
 - A secret manager that injects the env var your MCP server reads on its own (the `command` and `args` then run inside that environment).
 
+Project MCP config remains supported for shared team servers, but a project layer
+that changes an existing server's target (`url`, `command`, or `args`) does not
+inherit user credentials. Clear or replace `headers`, `env`, or `oauth` explicitly,
+or use a new server name. OAuth tokens are bound to the resolved server identity,
+so changing a server target can require `zero mcp oauth login <server>` again.
+OAuth server names ending in `.<32 hex chars>` are reserved for token storage.
+
 ### As a server — expose Zero's tools to another agent
 
 ```bash
 zero serve --mcp
+zero serve --mcp -C /path/to/workspace
+zero serve --mcp --add-dir /path/to/extra
 ```
 
 The server speaks MCP over stdio. Configure it from the receiving side as a `stdio` server whose command is `zero serve --mcp`.
 
+`-C/--cwd` sets the workspace root exposed as MCP resources and used by core tools. `--add-dir` (repeatable) widens that resource/tool scope beyond the workspace without granting the sandbox temp roots used by interactive runs.
 ## 6. Plugins
 
 A plugin is a self-contained directory that bundles tools, hooks, and skills for one capability. Plugins live at:
@@ -276,19 +296,19 @@ A plugin is enabled by being present in the plugins directory and disabled by re
 
 Plugin commands run with the plugin directory as their working directory. Use relative paths; the loader resolves them at activation time.
 
-> **Roadmap.** An in-UI plugins manager (browse, install, enable / disable) is on the backlog. Today you use the `zero plugins` CLI subcommands above. Skills declared inside a plugin's `plugin.json` are not yet merged into the `skill` tool's discovery (see section 3).
+> **Roadmap.** An in-UI plugins manager (browse, install, enable / disable) is on the backlog. Today you use the `zero plugins` CLI subcommands above.
 
 ## 7. Configuration locations
 
-Three layers, applied in order (later layers override earlier ones):
+Five configuration sources, in precedence order (later sources override earlier ones):
 
-| Layer | Path | Notes |
+| Source | Path / Key | Notes |
 | --- | --- | --- |
 | Built-in defaults | compiled in | Lowest priority. |
 | User config | `~/.config/zero/config.json` | Your machine. Never committed. |
 | Project config | `./.zero/config.json` | The repo. Committed (or not, your call). |
-| CLI flags | `--model`, `--mode`, ... | Highest priority, per-invocation. |
 | Environment | `ZERO_*` | Provider commands, secrets, skills dir override. |
+| CLI flags | `--model`, `--mode`, ... | Highest priority, per-invocation. |
 
 The user config holds things that should follow the user across projects (default provider, default model, theme). The project config holds things the team agreed on (provider catalog, sandbox policies, model restrictions).
 
@@ -301,20 +321,20 @@ A team that wants every contributor's Zero to behave the same way commits:
 - `AGENTS.md` — project conventions, build commands, do-not-edit lists.
 - `.zero/config.json` — provider catalog, default model, allowed tools.
 - `.zero/specialists/api-reviewer.md` — the team's PR-review specialist.
-- `.zero/hooks.json` — block `rm -rf` and `git push --force` on `beforeTool`.
+- `.zero/hooks.json` — block `rm -rf` on `beforeTool`.
 - `.zero/plugins/internal-tooling/` — a plugin that adds the team's internal CLI tools to the agent's toolset.
 
 Each contributor adds only:
 
 - `~/.config/zero/config.json` — their personal API keys, theme, default mode.
+- `~/.config/zero/ZERO.md` — personal preferences that follow them across every project (see section 1).
 - `~/.local/share/zero/skills/` — personal skills they keep across projects.
 
 That's it. Run `zero` from the repo root and the agent has the team's full instruction set, every contributor's personal setup, and nothing else.
 
 ## 9. Reference
 
-- [README](README.md) — install, quickstart, command reference.
-- [docs/SPECIALISTS.md](docs/SPECIALISTS.md) — full specialist manifest spec.
-- [docs/PRD.md](docs/PRD.md) — product requirements, full feature surface.
-- [docs/STREAM_JSON_PROTOCOL.md](docs/STREAM_JSON_PROTOCOL.md) — `zero exec` I/O contract.
-- [docs/INSTALL.md](docs/INSTALL.md) — install from source or release.
+- [README](../README.md) — install, quickstart, command reference.
+- [docs/SPECIALISTS.md](SPECIALISTS.md) — full specialist manifest spec.
+- [docs/STREAM_JSON_PROTOCOL.md](STREAM_JSON_PROTOCOL.md) — `zero exec` I/O contract.
+- [docs/INSTALL.md](INSTALL.md) — install from source or release.

docs/GITHUB_ACTION.md

@@ -41,8 +41,9 @@ jobs:
 > provider-neutral: `api-key-env` is the environment variable name the chosen
 > provider reads its key from (for example `ANTHROPIC_API_KEY`, `GEMINI_API_KEY`,
 > `OPENROUTER_API_KEY`), and `api-key` is the secret value. The action exports
-> that variable for the run and never prints it. If your repository already
-> commits a `.zero/config.json` with an active provider, you can omit `provider`.
+> that variable only for the ZERO step and never prints it. If your repository
+> already commits a `.zero/config.json` with an active provider, you can omit
+> `provider`.
 
 ## Inputs
 
@@ -51,8 +52,8 @@ jobs:
 | `prompt` | one of `prompt`/`prompt-file` | `""` | The instruction for ZERO to execute. |
 | `prompt-file` | one of `prompt`/`prompt-file` | `""` | Path (relative to `working-directory`) to a file whose contents are the prompt. |
 | `provider` | no | `""` | Provider id to activate (e.g. `openai`, `anthropic`, `gemini`, `ollama`, or any compatible endpoint). |
-| `api-key` | no | `""` | The provider API key. Pass from a secret; never logged. |
-| `api-key-env` | no | `""` | Env var name the provider reads its key from (e.g. `OPENAI_API_KEY`). Exported with `api-key`. |
+| `api-key` | no | `""` | The provider API key. Pass from a secret; exported only for the ZERO step and never logged. |
+| `api-key-env` | no | `""` | Env var name the provider reads its key from (e.g. `OPENAI_API_KEY`). Exported only for the ZERO step when used with `api-key`. |
 | `model` | no | `""` | Model id. Defaults to the resolved provider's default. |
 | `mode` | no | `""` | Run mode (`zero exec --mode`), e.g. `smart`, `deep`, `fast`. |
 | `auto` | no | `low` | Autonomy ceiling (`zero exec --auto`): `low`, `medium`, or `high`. Conservative by default. |
@@ -162,9 +163,10 @@ jobs:
 ## Security notes
 
 - **Secrets are passed as Action secrets and never logged.** The action exports
-  `api-key` (under `api-key-env`) and `slack-webhook-url` into the step
-  environment only. They are never echoed, and ZERO redacts secret-shaped
-  strings from its own output.
+  `api-key` (under `api-key-env`) only for the ZERO step. `slack-webhook-url`
+  is passed only to the Slack post step. They are never echoed or persisted to
+  later workflow steps, and ZERO redacts secret-shaped strings from its own
+  output.
 - **The sandbox is always active.** This action never passes
   `--skip-permissions-unsafe`, so writes stay inside the checked-out repository
   (plus any roots you grant with `add-dir`). Unsafe mode is never enabled

docs/HOW_ZERO_WORKS.md

@@ -0,0 +1,801 @@
+# How Zero Works
+
+This document explains Zero's architecture at a high level: the layers, module
+boundaries, data flow, and runtime concepts behind the interactive TUI and the
+headless `zero exec` mode.
+
+It is written as a progressive map: each high-level section introduces the
+concepts you need for orientation, then points to the later section that explains
+the mechanics in more detail.
+
+## Big Picture
+
+Zero is a Go CLI application with two main user-facing surfaces:
+
+- **Interactive TUI** (`zero`): a Bubble Tea application that renders chat,
+  tool cards, permission prompts, model/provider pickers, and session controls.
+- **Headless exec** (`zero exec ...`): a scriptable command runner that uses the
+  same agent loop and tools, but writes text, JSON, or stream-JSON events.
+
+Both surfaces converge on the same core agent runtime:
+
+```mermaid
+flowchart TD
+    User[User / automation] --> CLI[cmd/zero -> internal/cli]
+
+    CLI --> TUI[Interactive TUI<br/>internal/tui]
+    CLI --> Exec[Headless exec<br/>internal/cli/exec.go]
+    CLI --> ACP[ACP / editor bridge<br/>internal/acp]
+
+    TUI --> Agent[Agent loop<br/>internal/agent]
+    Exec --> Agent
+    ACP --> Agent
+
+    Agent --> Provider[Provider runtime<br/>internal/zeroruntime.Provider]
+    Provider --> APIs[OpenAI / Anthropic / Gemini / compatible APIs]
+
+    Agent --> Registry[Tool registry<br/>internal/tools.Registry]
+    Registry --> CoreTools[Core tools<br/>read/edit/shell/search/plan]
+    Registry --> MCP[MCP tools<br/>internal/mcp]
+    Registry --> Specialists[Specialist tools<br/>internal/specialist]
+    Registry --> Plugins[Plugin tools<br/>internal/plugins]
+    Registry --> LocalControl[Browser / terminal / desktop helpers]
+
+    Agent --> Sandbox[Permission + sandbox engine<br/>internal/sandbox]
+    Agent --> Hooks[Hooks dispatcher<br/>internal/hooks]
+    Agent --> Sessions[Session event store<br/>internal/sessions]
+```
+
+The important split is:
+
+- `internal/tui` is the **presentation layer**. For the mechanics, see
+  [Presentation Layer: Bubble Tea TUI](#presentation-layer-bubble-tea-tui) and
+  [How the TUI and Agent Loop Connect](#how-the-tui-and-agent-loop-connect).
+- `internal/agent` is the **reasoning/tool loop**. For the per-turn mechanics,
+  see [Agent Loop](#agent-loop), [Tool Execution Lifecycle](#tool-execution-lifecycle),
+  and [Context Budget and Compaction](#context-budget-and-compaction).
+- `internal/tools`, `internal/sandbox`, and provider packages are the **runtime
+  capability layer**. For the details, see [Runtime Provider Layer](#runtime-provider-layer),
+  [Tools and Capabilities](#tools-and-capabilities), and
+  [Permission and Sandbox Model](#permission-and-sandbox-model).
+- `internal/sessions` is the **durable event log** used for resume, fork,
+  rewind, sub-agents, and history. See [Session Storage and History](#session-storage-and-history).
+
+## Architecture Spine
+
+The easiest way to understand Zero is not as a set of independent packages, but
+as one spine with several mechanisms attached to it:
+
+```mermaid
+flowchart LR
+    Surface[Surface<br/>TUI / exec / ACP] --> Run[Run request<br/>prompt + runtime options]
+    Run --> Agent[Agent loop<br/>turns, tools, recovery]
+    Agent --> Capabilities[Capabilities<br/>providers + tools + sandbox]
+    Capabilities --> Events[Runtime events<br/>text, tool calls, permissions, usage]
+    Events --> Surface
+    Events --> Sessions[Session log<br/>resume, fork, rewind, history]
+
+    Config[Config / guidelines / skills] --> Run
+    Extensions[MCP / specialists / plugins / hooks] --> Capabilities
+```
+
+The **TUI + agent loop** is the main story because it explains the normal user
+experience: type a prompt, watch the model stream, approve tool use, see file
+changes, and receive a final answer. Other mechanisms are still important, but
+most of them attach to the same spine:
+
+| Mechanism | Where it attaches | What it changes |
+| --- | --- | --- |
+| Config, provider profiles, model registry | Before a run starts | Which provider/model is used and what defaults apply. |
+| Project/user guidelines, skills, specialists context | Prompt construction | What instructions and available capabilities the model sees. |
+| Tool registry, MCP, plugins, local-control helpers | Capability layer | Which actions the model may request. |
+| Permission policy, sandbox, command-prefix grants | Tool execution | Whether a requested action can run, needs approval, or is denied. |
+| Hooks | Tool lifecycle | Optional before/after observers or blockers around tool calls. |
+| Sessions, checkpoints, compaction metadata | Event persistence/context | What can be resumed, forked, rewound, or summarized for later turns. |
+| Stream-JSON | Surface layer | How the same runtime events are emitted for automation instead of rendered in the TUI. |
+
+So this document focuses first on the TUI/agent-loop bridge, then explains the
+supporting mechanisms in terms of where they connect to that bridge.
+
+Use this as the reading path:
+
+| If you see this concept in a high-level flow | Read this detail section |
+| --- | --- |
+| Surface submits a run | [Startup Flow](#startup-flow), then [How the TUI and Agent Loop Connect](#how-the-tui-and-agent-loop-connect) |
+| Agent builds context and advances turns | [Agent Loop](#agent-loop) |
+| “Maybe compact context” | [Context Budget and Compaction](#context-budget-and-compaction) |
+| Model requests a tool | [Tool Execution Lifecycle](#tool-execution-lifecycle), then [Tools and Capabilities](#tools-and-capabilities) |
+| Tool needs approval or sandbox access | [Permission and Sandbox Model](#permission-and-sandbox-model) |
+| Events appear in the UI or stream-JSON | [How the TUI and Agent Loop Connect](#how-the-tui-and-agent-loop-connect), [Stream-JSON Protocol](#stream-json-protocol) |
+| Resume, fork, rewind, or sub-agent history | [Session Storage and History](#session-storage-and-history) |
+
+## Main Packages
+
+The sections above describe the architecture in concepts. If you want to connect
+those concepts to implementation details, use this table as the code map before
+reading the deeper flows below.
+
+| Package | Role |
+| --- | --- |
+| `cmd/zero` | Tiny binary entrypoint; calls `internal/cli.Run`. |
+| `internal/cli` | Command parsing, config resolution, provider creation, registry setup, TUI/exec/ACP launch paths. |
+| `internal/tui` | Bubble Tea model/update/view code, transcript rendering, composer, prompts, slash commands, setup wizard. |
+| `internal/agent` | Core loop that sends messages to the model, streams output, executes tools, handles permissions, compaction, retries, and final answers. |
+| `internal/zeroruntime` | Provider-neutral message, tool-call, stream, usage, and image types. Providers implement `Provider.StreamCompletion`. |
+| `internal/providers` | Factory that turns a resolved provider profile into an OpenAI, Anthropic, Gemini, Codex, or compatible provider adapter. |
+| `internal/tools` | Tool interface, registry, core read/write/shell/network tools, local-control wrappers, tool safety metadata, redaction, output ceilings. |
+| `internal/sandbox` | Permission policy, path/network checks, grant store, command-prefix approvals, and platform sandbox backends. |
+| `internal/sessions` | Local append-only session storage: `metadata.json` + `events.jsonl`. |
+| `internal/config` | User/project config resolution, active provider, preferences, sandbox/tool settings. |
+| `internal/mcp` | MCP server config, client runtime, permission store, and MCP tool registration. |
+| `internal/specialist` / `internal/swarm` | Sub-agent manifests and team/member orchestration exposed as tools. |
+| `internal/localcontrol` / `internal/browser` | Runtime helpers used by config-gated local-control tool wrappers. |
+| `internal/plugins` / `internal/skills` / `internal/hooks` | Extension surfaces loaded into the agent context, tool registry, or tool lifecycle. |
+| `internal/streamjson` | Machine-readable protocol for `zero exec --input-format/--output-format stream-json`. |
+
+## Startup Flow
+
+### Interactive TUI (`zero`)
+
+```mermaid
+sequenceDiagram
+    participant Main as cmd/zero
+    participant CLI as internal/cli
+    participant Config as internal/config
+    participant TUI as internal/tui
+    participant Agent as internal/agent
+
+    Main->>CLI: Run(args)
+    CLI->>Config: resolve config for workspace
+    CLI->>CLI: build provider
+    CLI->>CLI: build tool registry
+    CLI->>CLI: register core, config-gated local-control, specialists, MCP, plugins
+    CLI->>CLI: create sandbox engine + session store
+    CLI->>TUI: run Bubble Tea with Options
+    TUI->>TUI: Init / Update / View event loop
+    TUI->>Agent: agent.Run on submitted prompt
+```
+
+The TUI owns terminal interaction. It tracks composer state, transcript rows,
+scrolling, permission dialogs, model/provider pickers, setup wizard state, and
+visual updates. When the user submits a prompt, the TUI builds `agent.Options`
+with callbacks like `OnText`, `OnToolCall`, `OnToolResult`, `OnPermission`,
+`OnAskUser`, and `OnUsage`. Those callbacks convert runtime events into visible
+transcript rows and session events.
+
+### Headless exec (`zero exec`)
+
+```mermaid
+flowchart TD
+    A[Parse exec flags] --> B[Resolve config + provider]
+    B --> C[Build tool registry]
+    C --> D[Resolve prompt / stream-json input / images]
+    D --> E[Prepare or resume session]
+    E --> F[Create stream-json/text/json writer]
+    F --> G[agent.Run]
+    G --> H[Write final output and exit code]
+```
+
+`zero exec` uses the same `agent.Run` loop as the TUI, but without interactive
+terminal state. It can persist sessions, resume/fork existing sessions, run in
+stream-JSON mode for automation, and always requires a stronger completion
+signal so a no-tool assistant response is not treated as success while work is
+still pending.
+
+## Agent Loop
+
+The core loop lives in `internal/agent.Run`. Conceptually, every turn follows
+this cycle:
+
+```mermaid
+flowchart TD
+    Start[Build system prompt + user prompt] --> Expose[Partition visible tools]
+    Expose --> Compact[Maybe compact context]
+    Compact --> Request[Send CompletionRequest to provider]
+    Request --> Stream[Collect streamed text, reasoning, usage, tool calls]
+
+    Stream --> HasTools{Tool calls?}
+    HasTools -- No --> FinalCheck{Final answer acceptable?}
+    FinalCheck -- Yes --> Done[Return Result]
+    FinalCheck -- No --> Nudge[Append user nudge and continue]
+
+    HasTools -- Yes --> Execute[Execute tool calls]
+    Execute --> Permission[Permission / sandbox / hooks]
+    Permission --> ToolResult[Append tool result messages]
+    ToolResult --> Diagnostics[Post-edit diagnostics / self-correct]
+    Diagnostics --> Expose
+```
+
+Key responsibilities of the loop:
+
+1. **Build prompt state** — combines the system prompt, project/user guidelines,
+   loaded skills/specialist instructions, the current user prompt, and images. The
+   resulting message list is the model-facing context that later tool results are
+   appended to. Resumed session history is replayed by the caller before a new run,
+   not passed as a separate `agent.Run` history field.
+2. **Expose tools** — starts from the registry built by the CLI, then filters
+   tool definitions by permission mode, enabled or disabled filters, and
+   deferred-loading rules. Tool schemas also count toward context budget, so they
+   matter for compaction decisions.
+3. **Call provider** — sends a provider-neutral `zeroruntime.CompletionRequest`.
+4. **Stream events** — forwards text, reasoning, tool-call deltas, and usage to
+   the caller via callbacks.
+5. **Execute tools** — decodes tool arguments, checks filters, applies
+   permissions and sandbox policy, dispatches hooks, runs the registered tool,
+   redacts output, and appends a tool message. See
+   [Tool Execution Lifecycle](#tool-execution-lifecycle).
+6. **Recover and guard** — handles malformed tool calls, repeated failures,
+   empty turns, context pressure, stream reconnect/retry, and max-turn fallback.
+   Context pressure is handled by [Context Budget and Compaction](#context-budget-and-compaction).
+7. **Finalize** — returns a final answer or a structured incomplete/stop reason.
+
+The model never directly mutates the workspace. It asks for tool calls; Zero
+validates and executes those calls through the registry, permission policy, and
+sandbox engine.
+
+## Tool Execution Lifecycle
+
+A tool call is part of the model conversation, not a side channel that bypasses
+it. The model proposes an action, Zero decides whether that action is allowed to
+run, the tool returns a structured result, and that result becomes context for the
+next provider request.
+
+```mermaid
+sequenceDiagram
+    participant Model as LLM / provider
+    participant Agent as agent.Run
+    participant Policy as filters + permissions + sandbox
+    participant Hooks as hooks
+    participant Tool as tool implementation
+    participant Surface as TUI / exec callbacks
+    participant Transcript as model conversation + session events
+
+    Agent->>Model: request with messages + available tool schemas
+    Model-->>Agent: assistant text and/or tool_call(name, args)
+    Agent-->>Surface: stream text / tool-call preview
+    Agent->>Agent: collect and decode tool arguments
+    Agent->>Policy: check tool filters, permission mode, sandbox policy
+
+    alt denied before execution
+        Policy-->>Agent: denied result
+        Agent-->>Surface: tool failed/denied event
+        Agent->>Transcript: append tool result explaining denial
+    else needs user approval
+        Policy-->>Surface: permission request
+        Surface-->>Policy: allow / deny / remember decision
+        Policy-->>Agent: decision
+    end
+
+    alt approved to execute
+        Agent->>Hooks: beforeTool hooks
+        Hooks-->>Agent: allow / block / annotate
+        Agent->>Tool: execute(ctx, args)
+        Tool-->>Agent: output, error, metadata
+        Agent->>Hooks: afterTool hooks
+        Agent->>Agent: redact secrets, apply semantic output budget, enforce byte ceiling
+        Agent->>Agent: classify success/failure
+        Agent-->>Surface: OnToolResult callback
+        Agent->>Transcript: append tool result message
+    end
+
+    Agent->>Model: next request includes prior assistant tool_call + tool result
+    Model-->>Agent: continue with more tools or final answer
+```
+
+Conceptually, one requested tool goes through these phases:
+
+1. **Proposal** — the provider streams assistant content and may emit one or more
+   tool calls. Zero can show the partial tool arguments in the TUI while they are
+   still streaming, but nothing has executed yet.
+2. **Decode and gate** — the agent validates JSON arguments, checks whether the
+   tool is enabled, applies permission mode, asks the sandbox engine about path,
+   network, shell, or command-prefix constraints, and may invoke permission
+   callbacks.
+3. **Execute** — the selected registry entry runs locally, through MCP, as a
+   specialist/sub-agent, or through a plugin/local-control helper. Hooks can run
+   before or after this step.
+4. **Normalize result** — Zero turns the outcome into an agent `ToolResult`: a
+   success, failure, denial, cancellation, timeout, or validation error with
+   output/metadata that may be redacted or truncated.
+5. **Append to context** — the result is appended as a tool-result message paired
+   with the assistant tool call. The next model request includes that result, so
+   the model can inspect it and decide what to do next.
+6. **Render and persist** — the result is surfaced to the TUI, stream-JSON
+   writer, or ACP client. TUI and exec record rich tool/session events for
+   resume/history; ACP currently persists conversational messages and streams
+   tool/permission updates to the client.
+
+Tool success and failure are both informative context. A successful `read_file`
+result gives the model file contents; a failed `edit_file` result gives the model
+an error such as “old string not found”; a denied shell command tells the model
+that the action is not currently permitted. The model is expected to use that
+feedback to continue, choose another tool, ask the user, or explain the blocker.
+
+The judgement that “information is sufficient” is made by the model, not by the
+TUI. Zero controls the loop mechanics: it keeps appending messages, enforces
+safety, retries or nudges around malformed/empty/incomplete turns, and stops on
+hard limits. But the semantic decision to call another tool versus produce a
+final answer comes from the next provider response. In other words:
+
+```mermaid
+flowchart TD
+    Result[Tool result appended to conversation] --> Next[Next provider request]
+    Next --> Decide{Model response}
+    Decide -- More information/action needed --> MoreTools[Request another tool]
+    MoreTools --> Result
+    Decide -- User input needed --> AskUser[Call ask_user or explain question]
+    AskUser --> Result
+    Decide -- Enough information --> Final[Final assistant answer]
+    Decide -- Invalid / empty / unsafe --> Guard[Agent guardrail, retry, denial, or stop reason]
+    Guard --> Next
+```
+
+This is why tool output is usually more than UI decoration. It is part of the
+working memory for the next model turn, subject to context-window pressure and
+compaction. The TUI may summarize or prettify it for the user, but the agent loop
+also stores the machine-facing result that the model sees.
+
+## Context Budget and Compaction
+
+The `Maybe compact context` step in the agent-loop diagram is Zero's context-window
+pressure valve. It is not a separate memory engine; it rewrites the current
+model-facing conversation so the next provider request fits within the configured
+context window.
+
+Compaction is enabled only when the run knows a positive `ContextWindow`. When it
+is enabled, the loop does this near the top of each turn:
+
+```mermaid
+flowchart TD
+    Tools[Partition visible tools] --> Estimate[Estimate messages + tool-definition tokens]
+    Estimate --> Under{Under threshold?}
+    Under -- Yes --> Keep[Keep context as-is]
+    Under -- No --> Grown{Grown past low-water mark?}
+    Grown -- No --> Keep
+    Grown -- Yes --> Prune[Prune stale large tool outputs]
+    Prune --> Recheck{Back under threshold?}
+    Recheck -- Yes --> KeepPruned[Use pruned context]
+    Recheck -- No --> Summarize[Summarize oldest middle]
+    Summarize --> Replace[system messages + summary + recent suffix]
+```
+
+The trigger is intentionally approximate:
+
+- Zero estimates request size using message text, images, tool-call arguments, and
+  the tool definitions advertised on that turn.
+- The proactive threshold is roughly **two-thirds of the context window**.
+- Provider usage from completed turns is used to calibrate the rough estimate over
+  time, so the budget gauge and compaction decision track the provider more
+  closely after a few turns.
+- A low-water mark prevents the loop from re-summarizing the same already-compacted
+  history on every turn.
+
+When compaction actually runs, it tries to preserve the parts most likely to be
+needed by the model:
+
+- leading system messages stay verbatim;
+- the most recent turns stay verbatim (`CompactionPreserveLast`, defaulting to the
+  agent's normal preserve count);
+- the preserved suffix is widened to a provider-valid turn boundary so tool-result
+  messages are not replayed without their corresponding assistant tool call;
+- the older middle is summarized into one user-role message labeled
+  `[Summary of earlier conversation]`;
+- structured state that should not be paraphrased away, such as active plan/loaded
+  skill state, is preserved alongside the prose summary.
+
+There are two compaction paths:
+
+1. **Proactive compaction** runs before the provider call when the estimated
+   request is over threshold.
+2. **Reactive compaction** runs after a provider returns a context-limit style
+   error. Zero compacts once and retries that same turn; if the provider still
+   rejects the request, the error is surfaced rather than looping forever.
+
+Compaction failures are conservative. If the summarizer fails during proactive
+compaction, Zero keeps the original messages and tries to continue. It does not
+drop conversation history just to make room.
+
+## Presentation Layer: Bubble Tea TUI
+
+`internal/tui` is a Bubble Tea application. The usual Bubble Tea shape applies:
+
+```mermaid
+flowchart LR
+    Init[model.Init] --> Update[model.Update]
+    Update --> View[model.View]
+    View --> Terminal[Terminal frame]
+    Terminal --> Msg[tea.Msg: key, mouse, window, agent event]
+    Msg --> Update
+```
+
+The TUI is intentionally not the agent brain. It is responsible for:
+
+- collecting user input from the composer;
+- rendering assistant text, reasoning, tool cards, permissions, plans, usage,
+  files, and sidebars;
+- handling slash commands and modal surfaces;
+- converting agent callbacks into Bubble Tea messages;
+- persisting visible session events;
+- allowing live provider/model/permission changes.
+
+When a prompt is submitted, the TUI starts a background command that invokes
+`agent.Run`. Streaming callbacks send messages back into the Bubble Tea update
+loop, so rendering remains responsive while the agent is working.
+
+## How the TUI and Agent Loop Connect
+
+A common source of confusion is that Zero has two event loops running at once:
+
+1. **Bubble Tea loop** — owns terminal input, in-memory UI state, and frames.
+2. **Agent loop** — owns model turns, tool execution, permissions, retries, and
+   stop conditions.
+
+They are connected by `tea.Cmd` and `agent.Options` callbacks. The TUI never
+runs provider/tool logic directly inside `Update`; instead, submit handling gates
+and dispatches a prompt, the TUI starts a background command, and that command
+calls `agent.Run` with callbacks that send runtime events back to the Bubble Tea
+program. Function names in this section are illustrative code anchors, not API
+boundaries.
+
+```mermaid
+sequenceDiagram
+    participant User
+    participant TUI as Bubble Tea model<br/>internal/tui
+    participant Cmd as tea.Cmd goroutine<br/>runAgentWithOptions
+    participant Agent as agent.Run<br/>internal/agent
+    participant Provider as Provider.StreamCompletion
+    participant Tools as Tool registry + sandbox
+    participant Store as Session store
+
+    User->>TUI: type prompt / press Enter
+    TUI->>TUI: validate/gate submit and append visible user row
+    TUI->>Cmd: launch tea.Cmd with runID + cancellable context
+    Cmd->>Agent: agent.Run(prompt, provider, options)
+    Agent->>Provider: stream completion request
+    Provider-->>Agent: text / reasoning / tool-call deltas
+    Agent-->>Cmd: OnText / OnReasoning / OnToolCallStart callbacks
+    Cmd-->>TUI: runtimeMessageSink(tea.Msg)
+    TUI->>TUI: Update appends/updates transcript rows
+    TUI-->>User: View renders new frame
+
+    Agent->>Tools: execute collected tool call
+    Tools->>Tools: permission, sandbox, hooks, redaction
+    Tools-->>Agent: tool result
+    Agent-->>Cmd: OnToolCall / OnToolResult / OnPermission / OnUsage
+    Cmd->>Store: batch session events for resume/history
+    Cmd-->>TUI: final agentResponseMsg
+    TUI->>TUI: mark run complete, flush rows/events, unlock composer
+```
+
+In the TUI path, the same event can have three representations:
+
+| Runtime thing | Agent representation | TUI representation | Session representation |
+| --- | --- | --- | --- |
+| Assistant text | streamed provider text forwarded through `OnText` | incremental assistant row text | assistant message event |
+| Reasoning preview | `OnReasoning` callback | collapsible/preview reasoning row | usually not the same as final assistant text |
+| Tool call | `agent.ToolCall` from collected provider stream | tool card, specialist card, or plan panel update | `EventToolCall` |
+| Tool result | `agent.ToolResult` | result card/status, changed files, plan/specialist updates | `EventToolResult` |
+| Permission prompt/audit | `PermissionRequest` / `PermissionEvent` | modal prompt or permission row | permission request/decision events |
+| Usage/context | provider usage + measured context | context gauge / usage metadata | usage event |
+
+### One interactive turn, end to end
+
+For a normal prompt in the interactive TUI:
+
+1. Submit handling validates composer state and command gates, then dispatches
+   the prompt to the launch path.
+2. The launch path appends immediate UI rows, creates a cancellable run context,
+   and starts `runAgent` as a `tea.Cmd`; run bookkeeping assigns the `runID` used
+   to correlate later events.
+3. `runAgentWithOptions` copies the current UI-controlled runtime choices into
+   `agent.Options`: provider/model labels, permission mode, cwd, session ID,
+   images, registry, response style, context window, and self-correction hooks.
+4. The command wraps agent callbacks. For example:
+   - `OnText` sends incremental assistant text into the TUI message sink.
+   - `OnToolCallStart` / `OnToolCallDelta` show live tool-argument streaming.
+   - `OnToolCall` creates visible tool rows and appends session events.
+   - `OnToolResult` creates result rows, updates plan/specialist side panels, and
+     records changed files.
+   - `OnPermissionRequest` sends a prompt to the TUI and blocks the agent loop
+     until the user answers or the run context is cancelled.
+   - `OnAskUser` does the same for model-initiated clarification questions.
+5. `agent.Run` builds provider-neutral messages and tool definitions, streams a
+   provider response, executes any requested tools through the registry/sandbox,
+   appends tool results back into the model conversation, and repeats until it has
+   a final answer or a stop condition.
+6. When `agent.Run` returns, the command emits `agentResponseMsg`. The Bubble Tea
+   loop marks the run complete, flushes queued rows/session events, updates usage
+   and file state, and re-enables the composer.
+
+The important architectural point: **the TUI is stateful and reactive, but the
+agent loop is the source of runtime truth for model/tool progress**. Presentation
+state is derived from callback events and persisted session events; workspace
+changes happen only through tools executed by the agent runtime.
+
+### Feature flows use the same bridge
+
+For one interactive turn, the supporting systems do not run beside the TUI/agent
+loop; they run inside or around it:
+
+```mermaid
+flowchart TD
+    Submit[TUI submit] --> Options[Build agent.Options]
+    Options --> Context[Assemble context<br/>config, guidelines, history, skills]
+    Context --> ProviderCall[Provider stream]
+    ProviderCall --> StreamEvents[Text/reasoning/tool-call events]
+    StreamEvents --> Render[TUI render callbacks]
+
+    ProviderCall --> ToolCall{Tool call?}
+    ToolCall -- No --> Final[Final response]
+    ToolCall -- Yes --> ToolGate[Tool gate<br/>filters, permissions, sandbox]
+    ToolGate --> HooksBefore[beforeTool hooks]
+    HooksBefore --> Execute[Tool / MCP / specialist / plugin]
+    Execute --> HooksAfter[afterTool hooks]
+    HooksAfter --> ToolResult[Tool result message]
+    ToolResult --> Persist[Session events]
+    ToolResult --> Context
+
+    Final --> Persist
+    Persist --> Resume[Resume / fork / rewind / parent metadata]
+```
+
+Most interactive features are variations of that same bridge rather than separate
+architectures:
+
+- **Slash commands** either mutate TUI state directly (`/model`, `/permission`,
+  `/help`) or start short `tea.Cmd` jobs that later return a message to `Update`
+  (`/doctor`, `/search`, `/compact`, session commands).
+- **Permissions** sit between a model-requested tool call and actual execution:
+  the agent blocks on `OnPermissionRequest`, the TUI surfaces a modal and sends
+  the decision back through a channel, and the session log records the
+  request/decision events.
+- **Plans, specialists/sub-agents, and ask-user prompts** all follow the same
+  runtime pattern: the model requests a capability, the agent records the
+  call/result in its conversation, and the TUI mirrors the visible state as plan
+  panels, specialist cards, or question forms. Specialist session metadata can
+  carry parent/root and sub-agent/spec fields, with richer linkage depending on
+  the specialist or swarm path.
+- **Streaming tool writes** start as provider tool-call deltas before execution;
+  the TUI shows the arguments live, then replaces/augments them with the actual
+  `ToolResult` after the tool runs.
+- **Compaction** is an agent-loop context-management step; the TUI may show that
+  it happened, but the purpose is to keep the next provider request within budget.
+- **Stream-JSON** is another surface over the same callbacks: `zero exec`
+  serializes the events as JSON protocol events instead of Bubble Tea views.
+
+So a feature-specific deep dive should usually explain: what user event starts it,
+which `tea.Msg`/callback carries it, which agent/tool event is authoritative, and
+which session events make it resumable.
+
+## Runtime Provider Layer
+
+Providers are hidden behind this interface in `internal/zeroruntime`:
+
+```go
+type Provider interface {
+    StreamCompletion(ctx context.Context, request CompletionRequest) (<-chan StreamEvent, error)
+}
+```
+
+The agent loop only knows about provider-neutral messages, tool definitions,
+stream events, reasoning blocks, images, and token usage. Provider adapters in
+`internal/providers/*` translate those neutral types to provider-specific APIs:
+
+- OpenAI and OpenAI-compatible APIs;
+- Anthropic and Anthropic-compatible APIs;
+- Gemini;
+- Codex / ChatGPT OAuth-backed runtime.
+
+`internal/providers.New` resolves the configured model through the model
+registry/catalog, normalizes API model IDs and base URLs, attaches API keys or
+OAuth resolvers, and returns the concrete adapter.
+
+## Tools and Capabilities
+
+Tools implement a common interface and are registered by name in
+`internal/tools.Registry`.
+
+```mermaid
+flowchart TD
+    AgentToolCall[Tool call from model] --> Decode[Decode JSON args]
+    Decode --> Filter[Enabled/disabled + permission-mode filter]
+    Filter --> Precheck[Pre-permission rejecters]
+    Precheck --> SandboxEval[Sandbox Evaluate]
+    SandboxEval --> Prompt{Needs user approval?}
+    Prompt -- Yes --> Decision[Permission callback]
+    Prompt -- No --> Run[Registry.RunWithOptions]
+    Decision --> Run
+    Run --> Redact[Redact secrets]
+    Redact --> Budget[Token-aware semantic output budget]
+    Budget --> Ceiling[Existing hard byte ceiling]
+    Ceiling --> Message[Return one tool result to model]
+```
+
+Oversized results use deterministic, provider-neutral estimated-token budgets.
+Policies retain useful structure for files, search matches, tests, process logs,
+diffs, and worker conclusions; tools without a declared category use a UTF-8-safe
+head/tail fallback. The estimate is intentionally conservative for non-ASCII
+text and is not exact provider tokenization. Existing byte ceilings remain the
+authoritative safety limit. When the existing spill mechanism can persist the
+complete redacted text received by the budgeting layer, the result includes its
+safe spill reference. This does not imply capture of subprocess bytes already
+discarded by a tool's established internal buffer.
+
+Core tool groups include:
+
+- **Read-only tools**: file reads, directory listing, glob, grep, LSP navigation,
+  skills, `ask_user`, permission requests.
+- **Write tools**: `write_file`, `edit_file`, `apply_patch`, `update_plan`.
+- **Shell tools**: `exec_command`, `write_stdin`, legacy `bash`.
+- **Network tools**: `web_fetch` and optional web search backend.
+- **Local-control tools**: config-gated wrappers for browser, terminal, desktop,
+  and artifact helpers backed by packages such as `internal/localcontrol` and
+  `internal/browser`.
+- **Extension tools**: MCP tools, specialists/sub-agents, swarm tools, plugins.
+
+Tool definitions include safety metadata: side-effect class, permission level,
+reason, and whether prompt-gated tools are advertised in auto-like modes. The
+agent uses this metadata plus per-turn filters and deferred-loading state to
+choose which registered tools the model can see and whether a call can run
+immediately.
+
+## Permission and Sandbox Model
+
+Zero has two related but separate checks:
+
+1. **Tool permission metadata** says whether a tool is allowed, denied, or needs
+   approval.
+2. **Sandbox policy** evaluates the actual request: path scope, network access,
+   destructive shell patterns, explicit escalation, persistent/session grants,
+   and platform isolation availability.
+
+```mermaid
+flowchart TD
+    Call[Tool call] --> Safety[Tool Safety metadata]
+    Call --> Scope[Workspace/additional roots scope]
+    Safety --> Engine[Sandbox Engine Evaluate]
+    Scope --> Engine
+    Engine --> Allow[Allow]
+    Engine --> Prompt[Prompt user]
+    Engine --> Deny[Deny]
+    Prompt --> Grant[Turn/session/persistent grant]
+    Grant --> Execute[Execute tool]
+    Allow --> Execute
+    Deny --> Result[Return denied tool result]
+```
+
+The sandbox engine can:
+
+- auto-allow workspace writes for safe file-edit tools;
+- block reads/writes outside the allowed scope;
+- prompt for recoverable outside-workspace access;
+- deny or prompt for network depending on policy;
+- detect destructive shell commands;
+- run shell commands through a platform backend when available;
+- remember session or persistent grants;
+- support narrow command-prefix approvals.
+
+The user-facing TUI and `zero exec` mode both route permission requests through
+callbacks, but the TUI can show interactive modals while headless mode generally
+returns non-interactive fallback behavior unless a protocol/client handles it.
+
+## Session Storage and History
+
+Sessions live locally under Zero's data directory. A session directory contains:
+
+- `metadata.json` — session ID, title, cwd, provider/model, parent/root IDs,
+  spec/sub-agent fields, timestamps, and event count.
+- `events.jsonl` — append-only event log with messages, tool calls/results,
+  permissions, usage, checkpoints, forks, compaction, and specs. TUI and exec use
+  the richer event stream; ACP currently records conversational message history
+  while streaming tool and permission updates to the client.
+
+```mermaid
+flowchart LR
+    TUI[TUI callbacks] --> Store[internal/sessions.Store]
+    Exec[exec recorder] --> Store
+    Store --> Metadata[metadata.json]
+    Store --> Events[events.jsonl]
+    Events --> Resume[resume]
+    Events --> Fork[fork]
+    Events --> Rewind[rewind/checkpoints]
+    Events --> Children[parent/root metadata]
+```
+
+This is why sessions can be resumed, forked, rewound, and inspected without a
+remote service. Zero stores the local event history; providers only receive the
+messages needed for the current completion request.
+
+## Stream-JSON Protocol
+
+`zero exec --output-format stream-json` emits newline-delimited JSON events for
+integrations and CI. The protocol is implemented in `internal/streamjson` and is
+specified in [STREAM_JSON_PROTOCOL.md](STREAM_JSON_PROTOCOL.md).
+
+Common output event types include:
+
+- `run_start`, `text`, `reasoning`, `tool_call`, `tool_result`;
+- `permission`, `permission_request`, `permission_decision`;
+- `checkpoint`, `restore`, `usage`, `final`, `warning`, `error`, `run_end`.
+
+Input stream-JSON can provide `prompt` or `message` events and image payloads.
+The CLI parses those events, resolves a single prompt, validates/decodes images,
+and passes the result into the same agent loop used everywhere else.
+
+## Extensibility Flow
+
+Zero's extension surfaces are loaded before the agent loop starts:
+
+```mermaid
+flowchart TD
+    Config[Config resolution] --> Registry[Tool registry]
+    Registry --> Core[Core tools]
+    Registry --> MCP[MCP server tools]
+    Registry --> Specialist[Specialist / Task tools]
+    Registry --> Plugin[Plugin tools]
+
+    Config --> Prompt[System prompt context]
+    Prompt --> Guidelines[AGENTS.md / ZERO.md]
+    Prompt --> Skills[Skills list]
+    Prompt --> SpecialistContext[Specialist descriptions]
+
+    ToolLifecycle[Tool lifecycle] --> Hooks[beforeTool / afterTool hooks]
+```
+
+- **Project/user guidelines** become prompt context.
+- **Skills** are instruction packs the model can load on demand.
+- **Specialists** are sub-agents callable through the `Task` tool.
+- **MCP servers** contribute external tools.
+- **Plugins** can add tools, hooks, and skill roots. Bootstrap always registers a
+  multi-root skill tool: primary Zero skills dir, optional `~/.agents/skills`,
+  then plugin skill roots (earlier wins). `internal/skills` owns that merge via
+  `LoadFromRoots` / `DiscoveryRoots`; single-root `Load` remains for install/write.
+- **Hooks** can observe or block tool lifecycle events.
+
+## End-to-End Data Flow
+
+A typical interactive edit request looks like this:
+
+```mermaid
+sequenceDiagram
+    participant U as User
+    participant T as TUI
+    participant A as Agent loop
+    participant P as Provider
+    participant R as Tool registry
+    participant S as Sandbox
+    participant FS as Workspace/session files
+
+    U->>T: Type prompt and press Enter
+    T->>A: agent.Run(prompt, provider, options)
+    A->>P: CompletionRequest(messages, tool definitions)
+    P-->>A: Stream text + tool call edit_file
+    A->>S: Evaluate edit_file args
+    S-->>A: allow/prompt/deny decision
+    A->>T: permission callback if prompt needed
+    T-->>A: user decision
+    A->>R: Run edit_file
+    R->>FS: Write workspace file
+    R-->>A: Tool result + changed files
+    A->>P: Next CompletionRequest with tool result
+    P-->>A: Final answer
+    A-->>T: Result + callbacks
+    T->>FS: Append session events
+    T-->>U: Render final transcript
+```
+
+The same request in `zero exec` follows the same provider/tool/sandbox path, but
+streaming events are written to stdout/stderr instead of rendered in Bubble Tea.
+
+## Mental Model
+
+When you work on Zero, start by locating the change on the four-layer spine from
+[Architecture Spine](#architecture-spine) — Surface, Agent loop, Capabilities,
+Persistence — then jump to the package that owns that boundary:
+
+- UI behavior? Start in `internal/tui`.
+- Command-line behavior? Start in `internal/cli`.
+- Model/tool-turn behavior? Start in `internal/agent`.
+- A new capability? Register a tool in `internal/tools` or an extension surface.
+- Safety policy? Start in `internal/sandbox` and agent permission wiring.
+- Resume/history behavior? Start in `internal/sessions` and session recorders.

docs/INSTALL.md

@@ -1,32 +1,55 @@
-# Zero Install Scripts
+# Installing Zero
 
-> **Status: pre-release.** No GitHub release has been published yet, so the install scripts below
-> (which download release assets from `Gitlawb/zero`'s Releases) will fail with a 404 until the first
-> release exists. Until then, build from source: `go run ./cmd/zero` or `go build -o zero ./cmd/zero`
-> (requires Go 1.25+). The instructions below describe how the scripts work once releases are published.
+Zero is distributed as:
 
-Zero release archives are published as:
+- an npm package, `@gitlawb/zero`
+- release archives on GitHub Releases
+- source builds with Go 1.26.5+
 
-- `zero-v<version>-linux-<arch>.tar.gz`
-- `zero-v<version>-macos-<arch>.tar.gz`
-- `zero-v<version>-windows-<arch>.zip`
+The install scripts download a platform-specific release archive and require a
+published GitHub Release for the requested version. The npm package is
+self-contained: the platform binary installs from the npm registry.
 
-Each archive must have a matching `.sha256` file. The install scripts download both files and verify the checksum before copying the binary.
-
-Maintainers can verify the release directory before upload:
+## npm
 
 ```bash
-go run ./cmd/zero-release package
-go run ./cmd/zero-release verify
+npm install -g @gitlawb/zero
+zero
 ```
 
-`verify:release` requires every archive in `dist/release` to have a same-directory `.sha256` file whose contents are:
+The package supports Linux and macOS on x64 and arm64, and Windows on x64
+(Windows on ARM runs the x64 build under emulation). It installs a small
+`zero` wrapper plus, as an optional dependency, a platform payload with the
+native binary and the bundled browser/terminal control helpers. There are no
+install scripts and nothing is downloaded from outside the npm registry, so
+the install is silent and works identically under npm, Bun, pnpm, and yarn —
+no trust or approval steps. See [NPM_PACKAGING.md](NPM_PACKAGING.md) for the
+package architecture.
 
-```text
-<sha256>  <archive-name>
+Requirements:
+
+- Node.js 18+
+- network access to npm
+
+If the install skipped optional dependencies (`npm install --omit=optional`,
+or a package manager configured to do so), the wrapper falls back to
+downloading the binary from the matching GitHub Release, with checksum
+verification. The fetch is retried on each run until a binary is in place (a
+transient network failure heals itself; an install directory the current user
+cannot write to keeps failing — rerun the install or the command below with
+sufficient permissions). To trigger the fetch manually:
+
+```bash
+node "$(npm root -g)/@gitlawb/zero/scripts/postinstall.mjs"
 ```
 
-## Linux And macOS
+## Linux And macOS Script
+
+Install the latest release:
+
+```bash
+curl -fsSL https://raw.githubusercontent.com/Gitlawb/zero/main/scripts/install.sh | bash
+```
 
 From a checkout:
 
@@ -38,18 +61,14 @@ Install a specific version:
 
 ```bash
 ZERO_VERSION=0.1.0 scripts/install.sh
+scripts/install.sh --version 0.1.0
 ```
 
-Install to a custom directory:
+Install somewhere else:
 
 ```bash
 ZERO_INSTALL_DIR="$HOME/bin" scripts/install.sh
-```
-
-Install from a fork or internal repository:
-
-```bash
-scripts/install.sh --repo owner/repo
+scripts/install.sh --install-dir "$HOME/bin"
 ```
 
 Defaults:
@@ -58,9 +77,15 @@ Defaults:
 - Version: latest GitHub release
 - Install path: `~/.local/bin/zero`
 
-Requirements: `curl` or `wget`, `tar`, and `shasum` or `sha256sum`.
+Requirements: Bash, `curl` or `wget`, `tar`, and `shasum` or `sha256sum`.
 
-## Windows
+## Windows PowerShell Script
+
+Install the latest release:
+
+```powershell
+irm https://raw.githubusercontent.com/Gitlawb/zero/main/scripts/install.ps1 | iex
+```
 
 From a checkout:
 
@@ -74,30 +99,145 @@ Install a specific version:
 powershell -ExecutionPolicy Bypass -File scripts/install.ps1 -Version 0.1.0
 ```
 
-Install to a custom directory:
+Install somewhere else:
 
 ```powershell
 powershell -ExecutionPolicy Bypass -File scripts/install.ps1 -InstallDir "$env:USERPROFILE\bin"
 ```
 
-Install from a fork or internal repository:
-
-```powershell
-powershell -ExecutionPolicy Bypass -File scripts/install.ps1 -Repository owner/repo
-```
-
 Defaults:
 
 - Repository: `Gitlawb/zero`
 - Version: latest GitHub release
 - Install path: `%LOCALAPPDATA%\zero\bin\zero.exe`
 
+## From Source
+
+```bash
+git clone https://github.com/Gitlawb/zero.git
+cd zero
+go run ./cmd/zero
+```
+
+Build a local binary:
+
+```bash
+go build -o zero ./cmd/zero
+```
+
+Source builds require Go 1.26.5+.
+
+### Sandbox Helpers For Source Builds
+
+Release archives include the platform sandbox helpers. If you build directly
+from source, build the helpers you need:
+
+Linux:
+
+```bash
+go build -o zero ./cmd/zero
+go build -o zero-linux-sandbox ./cmd/zero-linux-sandbox
+go build -o zero-seccomp ./cmd/zero-seccomp
+```
+
+Put `zero` and `zero-linux-sandbox` in the same directory on `PATH`, for example
+`~/.local/bin`. `zero-seccomp` is kept as a compatibility wrapper; the sandbox
+helper applies the Unix-socket filter itself when that sandbox option is enabled.
+Linux native sandboxing also requires Bubblewrap to be installed.
+
+macOS uses the system sandbox and does not need an extra helper binary.
+
+### Termux (Android)
+
+Zero can run natively on Android via [Termux](https://termux.dev/). Build with
+`GOOS=android` to avoid the `faccessat2` syscall that is blocked by Samsung's
+seccomp filter on Android:
+
+```bash
+# Install Go in Termux
+pkg install golang
+
+# Build Zero for Android
+git clone https://github.com/Gitlawb/zero.git
+cd zero
+CGO_ENABLED=0 GOOS=android GOARCH=arm64 go build -ldflags="-s -w" -o zero ./cmd/zero
+
+# Move into PATH
+mv zero ~/.local/bin/
+```
+
+> **Why `GOOS=android`?** Go 1.26+ detects `runtime.GOOS == "android"` and skips
+> the `faccessat2` syscall inside `os/exec.findExecutable`, falling back to
+> permission-bit checks. Without this flag, Android's seccomp sends SIGSYS and
+> kills the process whenever Zero looks up a binary on `PATH` (git, sh, etc.).
+
+**DNS.** Android does not expose `/etc/resolv.conf`. Go's pure-Go DNS resolver
+needs one. Use `proot` to bind-mount Termux's resolver config:
+
+```bash
+pkg install proot
+proot -b "$PREFIX/etc/resolv.conf:/etc/resolv.conf" zero
+```
+
+Create a wrapper at `~/.local/bin/zero` to avoid typing proot every time:
+
+```bash
+#!/data/data/com.termux/files/usr/bin/bash
+exec proot -b "$PREFIX/etc/resolv.conf:/etc/resolv.conf" ~/.local/bin/zero.bin "$@"
+```
+
+**Scroll.** On native Termux (not under PRoot), mouse scrolling works out of the
+box. The TUI uses Bubble Tea's `AllMotion` mouse mode by default. If you run Zero
+inside PRoot (e.g. through proot-distro), the scroll fix activates `CellMotion`
+to avoid PRoot's ptrace interference with the 1003 escape sequence.
+
+**Providers.** Zero works with any OpenAI-compatible provider on Termux. For
+example, to use OpenCode Zen's free tier:
+
+```bash
+zero providers add opencode \
+  --name opencode \
+  --model deepseek-v4-flash-free \
+  --base-url https://opencode.ai/zen/v1 \
+  --set-active
+```
+
+Windows source builds can use the main `zero.exe` as the command runner and setup
+helper through Zero's built-in self-dispatch path. If you want a release-style
+layout anyway, build the standalone helper executables next to `zero.exe`:
+
+```powershell
+go build -o zero.exe ./cmd/zero
+go build -o zero-windows-command-runner.exe ./cmd/zero-windows-command-runner
+go build -o zero-windows-sandbox-setup.exe ./cmd/zero-windows-sandbox-setup
+```
+
+## Release Archive Format
+
+Release archives are named:
+
+- `zero-v<version>-linux-<arch>.tar.gz`
+- `zero-v<version>-macos-<arch>.tar.gz`
+- `zero-v<version>-windows-<arch>.zip`
+
+Supported targets:
+
+- `linux-x64`
+- `linux-arm64`
+- `macos-x64`
+- `macos-arm64`
+- `windows-x64` (Windows on ARM runs this build under emulation)
+
+Each archive must have a matching `.sha256` file. The install scripts download
+both files, verify the checksum, and then copy the binary into the install
+directory.
+
 ## Updating
 
-Check whether a newer release exists:
+Check for a newer release:
 
 ```bash
 zero update --check
 ```
 
-For M2, updates are check-only. Re-run the installer to replace the local binary after reviewing the release.
+Then reinstall with npm or rerun the install script for the version you want.

docs/M1_HEADLESS_EXEC_PRD.md

@@ -1,87 +0,0 @@
-# M1 Headless Exec PRD
-
-Status: Draft
-Owner: Vasanth
-Milestone: M1
-Scope: `zero exec` CLI surface
-
-## Goal
-
-Ship the first production-shaped headless command for Zero:
-
-```sh
-zero exec "inspect this repo and summarize the next fix"
-```
-
-This command should be useful for local scripts, CI smoke runs, and later VS Code integration without pulling in the full future protocol, session, MCP, sandbox, or plugin roadmap.
-
-## Non-Goals
-
-- No daemon mode.
-- No JSON-RPC input protocol.
-- No MCP, plugin, marketplace, or subagent support.
-- No session resume/fork/search.
-- No new provider adapters in this slice.
-- No architecture rewrite of the agent loop.
-
-## User Stories
-
-1. As a CLI user, I can run one prompt headlessly and receive the assistant answer on stdout.
-2. As a script author, I can choose `text` or JSONL output.
-3. As a reviewer, I can see tool activity without corrupting text stdout.
-4. As a model tester, I can override the configured model for one run.
-5. As a cautious user, prompt-gated tools stay disabled unless I pass an explicit unsafe flag.
-
-## CLI Contract
-
-```sh
-zero exec [prompt...] [options]
-```
-
-Options for M1:
-
-- `-f, --file <path>`: read the prompt from a file.
-- `-m, --model <model>`: override the configured model for this run.
-- `-C, --cwd <path>`: run from another working directory.
-- `-o, --output-format <text|json>`: choose stdout format.
-- `--skip-permissions-unsafe`: grant prompt-gated tools for this run.
-
-The existing `zero -p "prompt"` shortcut remains as a compatibility path and should call the same runner.
-
-## Output Contract
-
-Text mode:
-
-- stdout contains assistant text only.
-- stderr contains tool rows, warnings, and errors.
-
-JSON mode:
-
-- stdout is newline-delimited JSON events.
-- Required event types for this slice: `run_start`, `text`, `tool_call`, `tool_result`, `final`, `warning`, `error`, `done`.
-
-## Exit Codes
-
-- `0`: success
-- `1`: unexpected crash
-- `2`: usage error
-- `3`: provider/config/runtime error
-- `4`: reserved for future tool failure policy
-- `5`: reserved for future permission-denied policy
-
-## Permission Behavior
-
-Default headless mode advertises only tools marked `permission: "allow"`.
-
-`--skip-permissions-unsafe` advertises tools that are not marked `deny` and grants prompt-gated tools through `ToolRegistry.run`. It must not bypass schema validation or direct tool execution safeguards.
-
-## Acceptance Criteria
-
-- `go test ./...` passes.
-- `go run ./cmd/zero-release build` passes.
-- `go run ./cmd/zero-release smoke` passes.
-- `zero exec --help` documents the M1 flags.
-- Usage errors do not initialize providers.
-- Provider failures return exit code `3`.
-- Normal mode does not advertise `bash`, `write_file`, `edit_file`, or `apply_patch`.
-- Unsafe mode still uses the registry execution path.

docs/NPM_PACKAGING.md

@@ -0,0 +1,162 @@
+# npm packaging
+
+How `@gitlawb/zero` is put together on npm, why it is shaped this way, and the
+rules the release pipeline must follow. Read this before touching
+`package.json`, `bin/zero.js`, `scripts/postinstall.mjs`,
+`scripts/npm/build-platform-packages.mjs`, or the npm-publish steps of
+`release-artifacts.yml`.
+
+## Goals
+
+A `npm install -g @gitlawb/zero` must be **silent and self-contained**:
+
+- No `EBADENGINE` warnings — ours or from any package in the dependency tree.
+- No install scripts anywhere in the tree, so npm's `allow-scripts` gating,
+  Bun's blocked-by-default lifecycle scripts, and pnpm's strict mode all
+  install zero without prompts, trust ceremonies, or broken binaries.
+- No network fetches outside the npm registry at install time. GitHub being
+  down or rate-limited must not break `npm install`.
+- Browser control (`agent-browser`) and terminal control (`tuistory`) work out
+  of the box.
+
+## Architecture
+
+The model is the one used by Codex, esbuild, and Biome: a tiny wrapper package
+plus per-platform payloads carrying the native binaries.
+
+```
+@gitlawb/zero                    <- wrapper: bin/zero.js + optionalDependencies
+├─ @gitlawb/zero-darwin-arm64 -> npm:@gitlawb/zero@{version}-darwin-arm64
+├─ @gitlawb/zero-darwin-x64   -> npm:@gitlawb/zero@{version}-darwin-x64
+├─ @gitlawb/zero-linux-arm64  -> npm:@gitlawb/zero@{version}-linux-arm64
+├─ @gitlawb/zero-linux-x64    -> npm:@gitlawb/zero@{version}-linux-x64
+└─ @gitlawb/zero-win32-x64    -> npm:@gitlawb/zero@{version}-win32-x64
+```
+
+- The platform "packages" are **versions of the same `@gitlawb/zero` package**,
+  published at suffixed versions (`0.4.0-linux-x64`) and referenced through
+  `npm:` aliases in `optionalDependencies`. One package name means one npm
+  trusted-publisher configuration — no new publish credentials per platform.
+  The alias suffixes use Node's `process.platform`/`process.arch` names so
+  `bin/zero.js` can derive its platform package directly.
+- Each platform version sets `os` and `cpu`, so npm installs exactly one of
+  them and skips the rest.
+- A platform payload is assembled from the platform's release archive by
+  `scripts/npm/build-platform-packages.mjs` and contains the `zero` binary,
+  the platform's sandbox helpers, and the vendored `helpers/` tree (see
+  below). It has no `bin`, no scripts, and no dependencies — it is inert on
+  its own; the wrapper execs the binary out of it.
+- The published wrapper has **no scripts and no dependencies** — the
+  assembly script strips both from the repo `package.json` and injects the
+  exact-version `optionalDependencies`. (The repo `package.json` keeps
+  `agent-browser`/`tuistory` in `dependencies` purely as the version pins for
+  the vendored helpers tree; they are never installed by consumers.)
+- There is **no windows-arm64 build** (matches the release matrix); Windows on
+  ARM runs the x64 build under emulation via the first-run fallback.
+
+### Binary resolution in `bin/zero.js`
+
+1. Resolve `@gitlawb/zero-<platform>-<arch>` and exec the `zero` binary from
+   it. This wins over any previously downloaded copy — the platform version is
+   pinned to the wrapper release.
+2. Fall back to a binary previously downloaded next to the wrapper.
+3. If neither exists (`--omit=optional`, package managers that skip optional
+   dependencies), run the **fallback downloader**: `scripts/postinstall.mjs`,
+   the exact logic that used to run as a postinstall hook (HTTPS-only, SHA-256
+   verified against the release's own checksum file, no zip-slip), invoked by
+   the wrapper itself. Failures are deliberately not cached — the fetch
+   retries on every run until a binary is in place, so a transient network
+   error self-heals. Platforms with no release asset (anything outside the
+   matrix above, including windows-arm64) skip the attempt entirely rather
+   than probing the network each run.
+4. If the download is impossible too, print build-from-source guidance.
+
+There is deliberately **no `scripts.postinstall`** in any published
+`package.json`. The downloader exists only as a first-run fallback.
+
+### Vendored helpers (`helpers/`)
+
+`agent-browser` (Apache-2.0, vercel-labs/agent-browser) and `tuistory` are
+**vendored binaries/packages inside the platform payload**, not npm
+dependencies of the wrapper:
+
+- As a dependency, agent-browser's `engines: { node: ">=24", pnpm: ">=11" }`
+  and postinstall script produce `EBADENGINE` and `allow-scripts` warnings for
+  every installer. As a vendored tree it produces none, because npm never
+  resolves it.
+- `zero-release package` already stages the helpers tree into every release
+  archive (`stageLocalControlHelpers` runs `npm ci` from the repo's
+  `package.json` pins + lockfile on the native builder). The assembly script
+  reuses that staged tree, so npm installs and `install.sh` installs get
+  identical helpers.
+- The Go binary resolves helpers from `<binary dir>/helpers/node_modules/.bin`
+  on its own (`internal/localcontrol/browser.go`, `adjacentHelper`) — no
+  wrapper involvement, no configuration.
+- **Symlink materialization:** `npm pack` silently drops symlinks, and npm's
+  own `.bin` shims are symlinks on POSIX. The assembly script rewrites every
+  `.bin` symlink into a relocatable `#!/bin/sh` exec shim and dereferences any
+  other symlink, so the vendored tree survives publishing. (Verified
+  empirically; nested `node_modules` under `helpers/` IS packed — only the
+  package-root `node_modules` is always ignored.)
+- **Binary pruning:** agent-browser ships one ~11 MB native binary per
+  platform (7 total). The assembly script keeps only the payload's own
+  platform binary (plus the musl variant on linux, which its launcher detects
+  at runtime) — about 65 MB saved per platform package.
+
+## Publishing rules (release pipeline)
+
+These are the invariants `release-artifacts.yml` must hold. Breaking the first
+one is user-visible immediately.
+
+1. **Platform versions must never become `latest`.** `0.4.0-linux-x64` is a
+   semver *prerelease* of `0.4.0`; publishing it without an explicit
+   non-`latest` dist-tag would clobber `latest` and users would install a
+   platform payload as the CLI. The workflow publishes platform versions with
+   `--tag platform` and asserts `latest` survived before the wrapper publish,
+   then asserts `latest` equals the wrapper version after it. Both lookups
+   tolerate a failed `npm view` (registry lag, or no `latest` tag yet on a
+   bootstrap publish) — a missing value must not abort the job between the
+   platform and wrapper publishes, which would strand a half-published
+   release.
+2. **Platform versions publish before the wrapper.** The wrapper's
+   `optionalDependencies` pin exact suffixed versions; publishing the wrapper
+   first would create a window where installs resolve aliases that 404.
+3. **Exact-version pinning.** The wrapper at `X.Y.Z` references platform
+   versions `X.Y.Z-<platform>-<arch>` exactly — never ranges — so a wrapper
+   and its binaries can never skew.
+4. All publishes go through the existing npm OIDC trusted-publishing flow and
+   the `npm-publish` environment gate; no tokens. Because the platform
+   payloads are versions of the same package, they reuse the single
+   trusted-publisher configuration.
+
+`zero update` keeps working unchanged: it detects an npm install by finding a
+`package.json` named `@gitlawb/zero` next to the running binary — true inside
+a platform payload too — and updates via `npm install -g @gitlawb/zero@latest`.
+
+## Runbook: bumping the vendored helpers
+
+The vendored helper versions are pinned by the repo `package.json`
+`dependencies` + `package-lock.json`; the release build vendors exactly what
+the lockfile resolves. No `^range` surprises ship: a release carries what its
+lockfile pinned at build time.
+
+- Routine bump: `npm install agent-browser@<version>` (or `tuistory@…`) in the
+  repo to update pin + lockfile, verify the browser/terminal tools end-to-end,
+  ship with the next release.
+- Security response: if a vendored helper publishes a security fix, the bump
+  is release-worthy on its own — cut a patch release of zero carrying only the
+  pin change.
+- Attribution: agent-browser is Apache-2.0; its LICENSE ships inside the
+  vendored package directory (`helpers/node_modules/agent-browser/`), which
+  satisfies redistribution attribution.
+
+## History / rationale
+
+Until v0.3.x the npm package was a wrapper with a `postinstall` downloader and
+`agent-browser`/`tuistory` as regular dependencies. That produced three
+warnings on every install (`EBADENGINE` from agent-browser's `node >=24`
+engines pin, and `allow-scripts` warnings for both postinstall scripts), broke
+silently under Bun's default script blocking, and coupled `npm install` to
+GitHub Releases availability. The platform-package model removes the warnings
+structurally — there is nothing left in the tree for a package manager to warn
+about — rather than asking users to approve or suppress them.

docs/NPM_WRAPPER_SMOKE.md

@@ -1,8 +1,8 @@
 # npm Wrapper Smoke Checklist
 
 Run this checklist when a PR changes npm distribution files such as
-`package.json`, `bin/zero.js`, Go build or release tooling, or the npm `bin`
-wrapper.
+`package.json`, `bin/zero.js`, `scripts/npm/build-platform-packages.mjs`,
+Go build or release tooling, or the npm `bin` wrapper.
 
 ## Required Checks
 
@@ -21,17 +21,52 @@ go run ./cmd/zero-release build
 go run ./cmd/zero-release smoke
 ```
 
+## End-To-End Assembly Smoke
+
+Build a real archive for the host platform and assemble the npm payloads from
+it (this is exactly what the publish job does):
+
+```bash
+go run ./cmd/zero-release package
+node scripts/npm/build-platform-packages.mjs \
+  --artifacts-dir dist/release --out-dir dist/npm --only <platform>-<arch>
+```
+
+Then simulate an installed layout and run the wrapper through each resolution
+path:
+
+```bash
+# platform-package path
+mkdir -p /tmp/zero-sim/node_modules/@gitlawb
+cp -R dist/npm/wrapper /tmp/zero-sim/node_modules/@gitlawb/zero
+cp -R dist/npm/platforms/zero-<platform>-<arch> \
+  "/tmp/zero-sim/node_modules/@gitlawb/zero-<platform>-<arch>"
+node /tmp/zero-sim/node_modules/@gitlawb/zero/bin/zero.js --version
+
+# first-run download fallback (delete the platform package first)
+rm -rf "/tmp/zero-sim/node_modules/@gitlawb/zero-<platform>-<arch>"
+node /tmp/zero-sim/node_modules/@gitlawb/zero/bin/zero.js --version
+```
+
 ## Checklist
 
 - `package.json` has the expected package name (`@gitlawb/zero`), version,
-  `bin.zero` entry, and exactly one `scripts` entry — the `postinstall` hook.
-- `scripts/postinstall.mjs` resolves the correct release asset name/URL per
-  platform (`ZERO_INSTALL_DRY_RUN=1` prints the plan), verifies the downloaded
-  archive's SHA-256, and extracts only the known binary basenames into place.
-  `ZERO_SKIP_DOWNLOAD=1` opts out cleanly (exit 0) and an unsupported
-  platform/arch is a non-fatal skip.
-- The wrapper binary resolves through the package `bin` entry and
-  `node_modules/.bin` in a package-install smoke test.
+  `bin.zero` entry, and NO `scripts` entries — the published package must be
+  free of lifecycle scripts (see NPM_PACKAGING.md).
+- `scripts/npm/build-platform-packages.mjs` emits platform payloads whose
+  `package.json` is `@gitlawb/zero@<version>-<platform>-<arch>` with matching
+  `os`/`cpu`, contains the executable binary and the vendored `helpers/` tree
+  with regular-file (non-symlink) `.bin` shims, and a wrapper payload with no
+  scripts, no dependencies, and the full five-alias `optionalDependencies`
+  matrix.
+- `scripts/postinstall.mjs` (the first-run fallback) resolves the correct
+  release asset name/URL per platform (`ZERO_INSTALL_DRY_RUN=1` prints the
+  plan), verifies the downloaded archive's SHA-256, and extracts only the
+  known binary basenames into place. `ZERO_SKIP_DOWNLOAD=1` opts out cleanly
+  (exit 0) and an unsupported platform/arch is a non-fatal skip.
+- The wrapper prefers the platform package binary over a previously downloaded
+  one, and `agent-browser`/`tuistory` run from the platform package's
+  `helpers/node_modules/.bin` shims.
 - The built binary exits 0 for `zero --version` or `zero --help`.
 - `zero --version` reports `zero <package.json version>`.
 - Release packaging still emits the expected archive and checksum names when

docs/PERFORMANCE.md

@@ -1,6 +1,6 @@
 # Zero Performance Benchmarks
 
-The M2 performance harness tracks three release-facing signals:
+The performance harness tracks three release-facing signals:
 
 - Cold start: process startup time for `zero --version`.
 - Binary first output: time from spawning the built `zero --version` command to

docs/PRD.md

@@ -1,602 +0,0 @@
-# Zero — Product Requirements Document (v1.0)
-
-**Status:** DRAFT v1.0
-**Target:** Compete with Claude Code + Cursor + OpenCode
-**Timeline:** 6-12 months
-**Stack:** Go native core + npm distribution wrapper
-
----
-
-## 1. Vision
-
-**Zero is a terminal-first, editor-aware, multi-provider AI coding platform.**
-
-It combines the best of:
-- **Claude Code** — terminal-native agentic coding
-- **Cursor** — IDE integration + smart editing
-- **OpenCode** — multi-provider + headless
-
-**One Go-native agent core. Three surfaces. Zero lock-in.**
-
----
-
-## 2. Core Surfaces
-
-### 2.1 Terminal (TUI) — Primary
-- Interactive coding agent in terminal
-- Streaming, append-style (like Claude Code)
-- Bubble Tea + Lip Gloss-based
-- Full keyboard control
-- Mouse support
-
-### 2.2 Editor (VS Code Extension)
-- Same agent core, different UI
-- Inline suggestions + chat panel
-- File editing with diff view
-- Diagnostics integration
-- LSP-aware
-
-### 2.3 Headless (CLI/CI)
-- `zero exec "prompt"` for scripts
-- `zero serve` for daemon mode
-- JSON-RPC over stdio/WebSocket
-- MCP server support
-
----
-
-## 3. Architecture
-
-```
-┌─────────────────────────────────────────────────┐
-│              ZERO PLATFORM                        │
-├─────────────────────────────────────────────────┤
-│                                                  │
-│  ┌──────────┐  ┌──────────┐  ┌──────────┐     │
-│  │   TUI    │  │   VSCode │  │ Headless │     │
-│  │ Surface  │  │ Surface  │  │ Surface  │     │
-│  └────┬─────┘  └────┬─────┘  └────┬─────┘     │
-│       │             │             │            │
-│       └─────────────┼─────────────┘            │
-│                     │                          │
-│              ┌──────▼──────┐                   │
-│              │  Agent Core │                   │
-│              │  (ReAct)    │                   │
-│              └──────┬──────┘                   │
-│                     │                          │
-│  ┌──────┬──────┬────┴────┬──────┬──────┐      │
-│  │Tools │Perms │Sessions │Provi │Memory│      │
-│  └──────┴──────┴─────────┴──────┴──────┘      │
-│                                                  │
-└─────────────────────────────────────────────────┘
-```
-
-**Layering Rules:**
-- Agent Core never imports a surface
-- Surfaces are thin shells (I/O + rendering)
-- All state persists to local files
-- No cloud dependency
-
----
-
-## 4. Functional Requirements (24 Features)
-
-### F1 — Agent Core (P0)
-- ReAct loop: think → tool → result → think
-- Stream text + tool calls + thinking
-- Parallel tool execution where safe
-- Bound turns (max 100)
-- Clean interrupt (Esc/Ctrl-C)
-- Event stream: `text | tool_call | tool_result | thinking | usage | plan_update | error | turn_end`
-
-### F2 — Tools (P0)
-
-**Core (8):**
-- `read_file` — read with line ranges, images
-- `write_file` — create/overwrite
-- `edit_file` — exact-string edit (uniqueness guard)
-- `bash` — run shell command
-- `grep` — ripgrep content search
-- `glob` — file pattern matching
-- `apply_patch` — multi-hunk patches
-- `update_plan` — todo list
-
-**Advanced (4):**
-- `web_fetch` / `web_search` — web access
-- `task` — spawn subagent
-- `structured_output` — JSON-Schema validated
-- `notebook_edit` — Jupyter notebooks
-
-**Tool Requirements:**
-- Go typed parameter schemas -> JSON Schema
-- Side-effect class: `read | write | shell | network | out_of_workspace`
-- Structured results: `{ status, output, truncated?, meta? }`
-- Atomic execution
-- Permission gating
-
-### F3 — Providers & Model Registry (P0)
-- OpenAI-compatible (works with OpenAI, OpenRouter, Ollama, etc.)
-- Anthropic (Claude)
-- Google Gemini
-- Azure OpenAI
-- AWS Bedrock
-- xAI (Grok)
-
-**Model Registry:**
-
-```go
-type ModelEntry struct {
-  ID              string
-  Name            string
-  Provider        string
-  APIProviders    []string
-  ContextLimits   ContextLimits
-  ReasoningEffort ReasoningEffort
-  Capabilities    ModelCapabilities
-  Cost            ModelCost
-  Tier            string
-  MatchPatterns   []string
-}
-```
-
-- Per-task model selection (`-m`)
-- Provider rotation/failover
-- Cost tracking
-- "Spec model" for planning
-- Fallback model on failure
-
-### F4 — Terminal UI (P0)
-- **Layout:** header (cwd · git · model) → tool rows → diff → response → input → footer
-- **Streaming:** append-only transcript model with live tail updates
-- **Markdown:** full CommonMark + syntax highlighting
-- **Tool rows:** collapsible with one-line summaries
-- **Input:** `/` commands, `@` files, `!` bash, `Esc` interrupt
-- **Robustness:** resize, wide-char, ANSI, non-TTY fallback
-- **Splash:** ZERO wordmark
-- **Themes:** light/dark + custom
-
-### F5 — Sessions (P0)
-- Persist to `~/.local/share/zero/sessions/<id>/`
-- `meta.json` + `events.jsonl` + `messages.jsonl` + `checkpoints/`
-- `--resume [id]` (default: latest)
-- `--fork <id>` (copy with lineage)
-- Search: `zero search <query>` over messages/tools/results
-- Atomic writes (write-temp-then-rename)
-- Schema versioning
-- Export to Markdown/JSON
-
-### F6 — Configuration (P0)
-- Layered: defaults → user → project → env → CLI
-- `~/.config/zero/config.json` (user)
-- `./.zero/config.json` (project)
-- Runtime overlay (`--settings <path>`)
-- Feature flags for staged rollout
-- Validation on load
-
-### F7 — MCP (Model Context Protocol) (P0)
-- Connect to stdio/HTTP/SSE servers
-- Surface MCP tools into registry
-- `zero mcp add/remove/list`
-- Per-server and per-tool permissions
-- Persistent grants with revoke/clear
-- MCP server mode (zero as MCP server)
-
-### F8 — Headless / Programmatic (P0)
-- `zero exec [prompt]` — one-shot
-- `zero serve` — daemon mode
-- `--input-format stream-json` — line-delimited events
-- `--input-format stream-jsonrpc` — JSON-RPC protocol
-- `-o text|json` output formats
-- Exit codes: `0=success, 1=crash, 2=usage, 3=provider, 4=tool, 5=permission, 6=budget`
-
-### F9 — Permissions & Autonomy (P0)
-- **Side-effect classes:** read, write, shell, network, out_of_workspace
-- **Autonomy levels:** low (ask), medium (smart), high (auto)
-- `--skip-permissions-unsafe` (loud warning)
-- TUI: inline prompts (allow/deny/always)
-- Headless: default-deny unless granted
-- Persistent grants with list/revoke/clear
-- Audit log in `events.jsonl`
-
-**Permission Matrix:**
-
-| Tool Class | low | medium | high |
-|------------|-----|--------|------|
-| Read | allow | allow | allow |
-| Workspace write | prompt | allow | allow |
-| Shell | prompt | prompt | allow |
-| Network | prompt | allow | allow |
-| Out-of-workspace | deny | prompt | prompt |
-
-### F10 — Observability (P0)
-- Structured logging (stdout/stderr split)
-- Secret redaction (API keys, tokens, passwords)
-- No vendor telemetry (local only)
-- `zero doctor` — health check command
-- Crash reports (opt-in, local)
-- Performance metrics (TTFT, tokens/sec)
-
-### F11 — Subagents (P1)
-- `task` tool spawns scoped subagent
-- Built-in agents: code-review, security-review, refactor, test-gen
-- Recursion depth limit
-- Parent session linkage
-- Shared/isolated budgets
-
-### F12 — Plugins (P1)
-- User/project-scoped plugins
-- Commands, tools, prompts, hooks
-- Plugin manifest: `plugin.json`
-- `zero plugin install <path|url>`
-- No marketplace in v1 (manual install)
-
-### F13 — Distribution (P0)
-- **Single binary** via Go cross-compilation
-- npm package as a thin installer/launcher for the platform binary
-- Cross-platform: Linux, macOS, Windows
-- `zero update [--check]` self-update
-- Checksum verification
-- Rollback on failure
-
-### F14 — VS Code Extension (P1)
-- Same agent core via stream protocol
-- Inline chat panel
-- Diff view in editor
-- Diagnostics integration
-- LSP awareness
-- File watching
-
-### F15 — Memory & Context (P1)
-- `AGENTS.md` hierarchical memory
-- `@import` for modular rules
-- Auto-generated from `init` command
-- Project conventions learning
-- Skills system (reusable instruction packs)
-- Context budgeting (token tracking)
-- Auto-compaction when full
-
-### F16 — Hooks System (P1)
-- Lifecycle hooks: `PreToolUse`, `PostToolUse`, `SessionStart`, `SessionEnd`
-- User-defined shell commands
-- JSON output for conditional flow
-- `.zero/hooks/` directory
-
-### F17 — Slash Commands (P0)
-
-**Built-in (21):**
-- Session: `/clear`, `/compact`, `/context`, `/add-dir`, `/rewind`, `/resume`, `/exit`
-- Model: `/model`, `/effort`, `/style`, `/permissions`
-- Project: `/init`, `/memory`, `/skills`
-- Extensibility: `/mcp`, `/hooks`, `/agents`
-- Meta: `/help`, `/config`, `/doctor`, `/feedback`
-
-**Custom:** `./.zero/commands/*.md`
-
-### F18 — Git Integration (P1)
-- Worktree isolation per task
-- Auto-commit with messages
-- Branch management
-- PR creation
-- Diff review
-- Conflict resolution
-
-### F19 — Testing & Verification (P1)
-- `run_tests` tool
-- Self-verification loop
-- Test runner integration
-- Coverage reporting
-- Diagnostics (typecheck, lint)
-- LSP integration
-
-### F20 — Web Access (P1)
-- `web_fetch` — fetch URL content
-- `web_search` — search engine
-- Provider/permission gated
-- Rate limiting
-- Caching
-
-### F21 — Sandbox (P2)
-- Linux: bubblewrap
-- macOS: sandbox-exec
-- Windows: AppContainer
-- Filesystem policies (workspace write confinement + user-granted extra roots via `--add-dir`)
-- Network policies
-- Unix socket controls
-
-### F22 — Multi-modal (P2)
-- Image input (screenshots, diagrams)
-- PDF parsing
-- Voice input (optional)
-- File attachments
-- Paste support
-
-### F23 — Telemetry (P2)
-- Local-only metrics
-- Token usage tracking
-- Cost reports
-- Performance profiling
-- Crash reports (opt-in)
-- NO vendor telemetry
-
-### F24 — Advanced Features (P2)
-- Code review automation
-- Auto-refactor
-- Dependency updates
-- Security scanning
-- Documentation generation
-
----
-
-## 5. Data Models
-
-### 5.1 Headless Event Stream
-
-```json
-{ "type": "run_start", "runId": "r1", "sessionId": "s1", "cwd": "/repo", "model": "..." }
-{ "type": "text", "delta": "..." }
-{ "type": "tool_call", "id": "t1", "name": "bash", "args": {...}, "sideEffect": "shell" }
-{ "type": "permission", "id": "t1", "decision": "needs_approval", "reason": "shell" }
-{ "type": "tool_result", "id": "t1", "status": "ok", "output": "...", "truncated": false }
-{ "type": "usage", "promptTokens": 1234, "completionTokens": 567, "costUsd": 0.012 }
-{ "type": "plan_update", "plan": [{ "id": "1", "content": "...", "status": "in_progress" }] }
-{ "type": "turn_end", "stopReason": "end_turn" }
-{ "type": "error", "code": "rate_limit", "message": "...", "recoverable": true }
-{ "type": "run_end", "status": "success", "exitCode": 0 }
-```
-
-### 5.2 Session On Disk
-
-```
-~/.local/share/zero/sessions/<id>/
-  meta.json        # { schemaVersion, id, parentId?, title, cwd, model, createdAt, updatedAt, tokens, cost }
-  events.jsonl     # append-only AgentEvent stream
-  messages.jsonl   # provider-facing messages
-  checkpoints/     # snapshots for restore/fork/compact
-```
-
-### 5.3 Tool Contract
-
-```go
-type Tool interface {
-  Name() string
-  Description() string
-  Parameters() JSONSchema
-  SideEffect() SideEffect
-  Execute(ctx ToolContext, args map[string]any) ToolResult
-}
-
-type ToolResult struct {
-  Status    string
-  Output    string
-  Truncated bool
-  Meta      map[string]any
-}
-```
-
-### 5.4 Config Schema
-
-```json
-{
-  "model": "claude-sonnet-4-6",
-  "providers": {
-    "anthropic": { "apiKey": "env:AN..._KEY" },
-    "openai": { "apiKey": "env:OP..._KEY" }
-  },
-  "permissions": {
-    "autonomy": "medium",
-    "grants": [{ "tool": "bash", "pattern": "git status", "decision": "allow" }]
-  },
-  "ui": { "theme": "dark", "syntaxHighlighting": true }
-}
-```
-
----
-
-## 6. Non-Functional Requirements
-
-### Performance
-- TTFT < 500ms after provider's first byte
-- No full-transcript repaint per token
-- Cold start < 300ms (warm cache)
-- Tool execution < 100ms overhead
-- Session resume < 1s
-
-### Reliability
-- Crash-free TUI under resize, wide-char, non-TTY
-- Clean interrupt (no half-written state)
-- Atomic session writes
-- Recovery from malformed JSONL
-- Provider failover
-
-### Security
-- No mutation without permission decision (hard invariant)
-- API keys never logged
-- Secret redaction everywhere
-- Headless stdout free of logs
-- Workspace boundary by default
-- Sandboxing for untrusted code
-
-### Portability
-- Linux, macOS, Windows
-- Any OpenAI-compatible endpoint
-- x86_64, ARM64
-- Single binary distribution
-
-### Testability
-- Mock provider for unit tests
-- Fake MCP server/client
-- Fake terminal/PTY
-- Sandbox simulator
-- Deterministic event replay
-
----
-
-## 7. Milestones
-
-### M0 — Foundation (Weeks 1-2)
-- Go module setup
-- Agent core skeleton
-- 8 core tools (read, write, edit, bash, grep, glob, apply_patch, update_plan)
-- OpenAI-compatible provider
-- Basic Bubble Tea TUI
-- Permission gate (read vs mutate)
-- Session persistence
-
-**Exit:** `zero` runs, can read/edit files, basic chat works.
-
-### M1 — Multi-Provider (Weeks 3-4)
-- Anthropic provider
-- Gemini provider
-- Model registry
-- Per-task model selection
-- Provider rotation
-- Cost tracking
-- Headless `exec` mode
-
-**Exit:** Switch between Claude/GPT/Gemini, see cost, run headless.
-
-### M2 — Polish (Weeks 5-8)
-- Advanced TUI features
-- Full markdown + syntax highlighting
-- Tool result truncation/pagination
-- Auto-compaction
-- Context budgeting
-- 20 slash commands
-- Doctor command
-- Search command
-
-**Exit:** Production-ready TUI, smooth streaming, all commands work.
-
-### M3 — Distribution (Weeks 9-10)
-- Single binary build
-- Self-update
-- Cross-platform CI
-- VS Code extension MVP
-- Documentation site
-
-**Exit:** `zero` ships as one binary, works on Win/Mac/Linux.
-
-### M4 — Extensibility (Weeks 11-14)
-- MCP client + permissions
-- Hooks system
-- Skills system
-- Memory (AGENTS.md)
-- Plugin loading
-- Subagents (built-in: review, security)
-
-**Exit:** Plugins work, MCP servers connect, memory persists.
-
-### M5 — Advanced (Weeks 15-20)
-- Web fetch/search
-- Git worktrees
-- Self-verification
-- Diagnostics (LSP, typecheck)
-- Testing integration
-- Code review automation
-
-**Exit:** Zero can review code, run tests, manage git.
-
-### M6 — Sandbox & Security (Weeks 21-24)
-- Linux sandbox (bubblewrap)
-- macOS sandbox
-- Windows AppContainer
-- Network policies
-- Filesystem policies
-
-**Exit:** Untrusted code runs safely in sandbox.
-
-### Beyond v1.0
-- Multi-modal (images, PDFs)
-- Voice input
-- Cloud sync (optional)
-- Team features
-- Marketplace (if demand)
-
----
-
-## 8. Risks & Mitigations
-
-| # | Risk | Mitigation |
-|---|------|------------|
-| R1 | Scope creep | Stick to milestones, defer features |
-| R2 | Provider API changes | Abstract via interface, test with mocks |
-| R3 | TUI complexity | Use Bubble Tea/Lip Gloss, follow mature terminal agent UX patterns |
-| R4 | Performance | Benchmark early, optimize streaming |
-| R5 | Security vulnerabilities | Permission gate from day 1, security review |
-| R6 | Cross-platform bugs | CI on Win/Mac/Linux from M0 |
-| R7 | Breaking changes | Schema versioning, migration tools |
-
----
-
-## 9. Success Metrics
-
-### v1.0 (6 months)
-- 5+ providers supported
-- 12+ tools
-- 100% test coverage on core
-- < 300ms cold start
-- Single binary < 50MB
-- Works on Win/Mac/Linux
-- 1000+ GitHub stars
-- 50+ contributors
-
-### v2.0 (12 months)
-- VS Code extension published
-- MCP server mode
-- Plugin ecosystem (20+ plugins)
-- 10K+ active users
-- Compete with Claude Code on benchmarks
-
----
-
-## 10. Out of Scope (v1.0)
-
-- Cloud sync / hosted backend
-- Relay / remote SSH execution
-- Multi-agent mission mode
-- Plugin marketplace
-- Auto-generated wikis
-- Git authorship notes
-- Vendor telemetry
-- Voice input
-- Mobile support
-
----
-
-## 11. Open Decisions
-
-1. **TUI library:** Bubble Tea + Lip Gloss — **Decide: Go-native TUI** (single-binary friendly, mature terminal ecosystem)
-2. **Storage:** JSONL files vs SQLite — **Decide: JSONL** (inspectable, simple)
-3. **Package structure:** Single vs monorepo — **Decide: Single** until M4
-4. **Provider abstraction:** Direct API vs wrapper — **Decide: Direct** (portability)
-5. **Patch strategy:** apply_patch vs edit_file — **Decide: Both** (apply_patch for bulk, edit_file for small)
-
----
-
-## 12. Team & Timeline
-
-**Team (3 people):**
-- You (lead + TUI + tools)
-- Backend dev (providers + core)
-- Designer (VS Code extension + docs)
-
-**Timeline:**
-- M0-M1: 4 weeks (foundation)
-- M2-M3: 6 weeks (polish + distribution)
-- M4-M5: 8 weeks (extensibility)
-- M6: 4 weeks (sandbox)
-- **Total v1.0: 6 months**
-
----
-
-## 13. Why This Will Work
-
-1. **Proven patterns** — Based on clean-room analysis of existing terminal coding-agent UX patterns
-2. **Multi-provider** — Unique advantage vs Claude Code/Cursor
-3. **Open source** — Community contributions
-4. **Go** — Native binary, fast startup, strong concurrency, cross-platform CLI ergonomics
-5. **npm distribution wrapper** — Keeps JavaScript ecosystem installation while the runtime stays native
-6. **Clear milestones** — Ship early, iterate
-
----
-
-**This is the real PRD. 13 sections, 24 features, 6 months to v1.0.**

docs/README.md

@@ -0,0 +1,24 @@
+# Zero Documentation
+
+This directory contains public documentation for using, extending, and releasing
+Zero.
+
+## User Docs
+
+- [Install](INSTALL.md)
+- [Update flow](UPDATE.md)
+- [OAuth logins and subscription-backed providers](oauth-subscriptions.md)
+
+## Automation And Integrations
+
+- [Stream-JSON protocol](STREAM_JSON_PROTOCOL.md)
+- [GitHub Action](GITHUB_ACTION.md)
+- [Specialists](SPECIALISTS.md)
+
+## Maintainer Docs
+
+- [How Zero Works](HOW_ZERO_WORKS.md)
+- [npm wrapper smoke checklist](NPM_WRAPPER_SMOKE.md)
+- [Benchmarks](BENCHMARK.md)
+- [Performance benchmarks](PERFORMANCE.md)
+- [Offline agent evals](AGENT_EVALS.md)

docs/STREAM_JSON_PROTOCOL.md

@@ -41,6 +41,7 @@ events when present.
 
 ```json
 { "schemaVersion": 2, "type": "run_start", "runId": "run_20260603_abc123", "sessionId": "zero_20260603100000_abc123", "cwd": "/repo", "provider": "openai", "model": "gpt-4.1", "apiModel": "gpt-4.1" }
+{ "schemaVersion": 2, "type": "reasoning", "runId": "run_20260603_abc123", "delta": "Thinking..." }
 { "schemaVersion": 2, "type": "text", "runId": "run_20260603_abc123", "delta": "..." }
 { "schemaVersion": 2, "type": "tool_call", "runId": "run_20260603_abc123", "id": "call_1", "name": "read_file", "args": { "path": "README.md" }, "sideEffect": "read" }
 { "schemaVersion": 2, "type": "permission_request", "runId": "run_20260603_abc123", "id": "call_2", "name": "write_file", "action": "prompt", "permission": "prompt", "permissionMode": "ask", "sideEffect": "write", "reason": "Creates or overwrites files." }
@@ -51,6 +52,10 @@ events when present.
 { "schemaVersion": 2, "type": "run_end", "runId": "run_20260603_abc123", "status": "success", "exitCode": 0 }
 ```
 
+`reasoning` events carry live model reasoning/status deltas for providers that
+stream them separately from answer text. They are liveness/progress events only:
+they are not folded into `text` or the final answer.
+
 Permission events may include structured sandbox metadata:
 
 ```json

docs/THEMES.md

@@ -0,0 +1,39 @@
+# TUI Themes
+
+Zero's TUI ships a set of built-in color themes. Pick one with `--theme <name>`,
+the `ZERO_THEME` environment variable, or the `/theme <name>` command while
+running. `auto` (the default) follows the terminal's detected background.
+
+Run `/theme` with no argument to open a picker: move through the list to
+live-preview each theme, press Enter to apply the highlighted one, or Esc to
+cancel and restore the previously active theme. Run `/theme list` to print
+the active theme and the registered names without opening the picker.
+
+## Dune (`dune`)
+
+A warm sand-and-cream palette: sand/cream surface, charcoal ink, and a soft
+amber accent.
+
+## Neon (`neon`)
+
+A neon-on-black palette: pitch-black surface, bright green ink, and a cyan accent.
+
+## Adding a theme
+
+Every theme is a `palette` (a table of color hex tokens) plus one entry in
+`themeRegistry`, both in `internal/tui/theme_palettes.go`. `buildTheme` in
+`internal/tui/theme.go` turns a `palette` into the resolved `lipgloss.Style`
+set every renderer reads from the active `zeroTheme`. Adding a theme means
+adding a new `palette{...}` literal, a `themeRegistry` entry, and test
+coverage for the new palette (see below).
+
+Registry-wide tests in `internal/tui/theme_select_test.go` assert the basic
+WCAG AA text tokens, the gray-ramp order, the diff word-span pairs, and the
+selected-row band for every entry. The rendered-surface invariants beyond
+those (permission surfaces, selected-row secondary text, diff gutters, and
+the xterm-256 downsampling checks) are asserted per palette, not against the
+whole registry: `TestExtendedThemeContrastInvariants` and
+`TestExtendedThemeANSI256Contrast` enumerate the palettes they cover. A new
+theme must be added to those tests (or given equivalent palette-specific
+assertions), or CI can stay green while its permission, selected-row, and
+diff surfaces ship unreadable.

docs/UPDATE.md

@@ -1,38 +1,44 @@
-# Zero Update Flow
+# Update Flow
 
-`zero update --check` checks the latest GitHub release for `Gitlawb/zero` and compares it with the local CLI version.
+`zero update --check` checks the latest GitHub release and compares it with the
+local CLI version.
 
 ```bash
 zero update --check
 zero update --check --json
 zero update --check --repo Gitlawb/zero
-zero update --check --endpoint https://api.github.com/repos/Gitlawb/zero/releases/latest
-zero update --check --timeout 3s
 zero update --check --target windows-x64
 ```
 
-For M2 this command is intentionally check-only:
+The command is intentionally check-only:
 
 - It does not replace the running binary.
-- It exits with code `0` when the check succeeds, even when an update is available.
+- It exits with code `0` when the check succeeds, even when an update is
+  available.
 - It exits with code `1` when the release check cannot be completed.
-- `--json` prints the same result in a machine-readable format for scripts and CI.
-- `--repo <owner/repo>` checks a different GitHub repository when no endpoint is provided.
-- `--endpoint <url|owner/repo>` checks a specific release API URL or repository slug.
-- `--timeout <duration>` overrides the default release check timeout.
-- `--target <platform-arch>` validates release metadata for another supported target from the current machine.
-- Release checks time out after 5 seconds by default.
-- It validates that the latest release includes the expected archive and matching `.sha256` asset for the selected platform target.
+- `--json` prints the same result in a machine-readable format for scripts and
+  CI.
 
-Supported targets are `linux-x64`, `linux-arm64`, `macos-x64`, `macos-arm64`, `windows-x64`, and `windows-arm64`. Without `--target`, Zero checks the current platform.
+Useful flags:
 
-The release endpoint resolves in this order:
+| Flag | Purpose |
+|---|---|
+| `--repo <owner/repo>` | Check another GitHub repository. |
+| `--endpoint <url|owner/repo>` | Check a specific release API URL or repository slug. |
+| `--timeout <duration>` | Override the default release check timeout. |
+| `--target <platform-arch>` | Validate release metadata for another supported target. |
 
-- `--endpoint` from the CLI, or `Options.Endpoint` when calling `update.Check` from code.
-- `ZERO_UPDATE_RELEASE_URL` from the environment.
-- `--repo` from the CLI, or `Options.Repository` when calling `update.Check` from code.
-- `https://api.github.com/repos/Gitlawb/zero/releases/latest`.
+Supported targets are `linux-x64`, `linux-arm64`, `macos-x64`, `macos-arm64`,
+`windows-x64`, and `windows-arm64`. Without `--target`, Zero checks the current
+platform.
 
-`--endpoint`, `Options.Endpoint`, and `ZERO_UPDATE_RELEASE_URL` may be a full URL or an `owner/repo` slug.
+Endpoint resolution order:
 
-Installer scripts download the matching release asset for the local platform and verify its `.sha256` file. If Zero is already installed, use this command before re-running the installer.
+1. `--endpoint`
+2. `ZERO_UPDATE_RELEASE_URL`
+3. `--repo`
+4. `https://api.github.com/repos/Gitlawb/zero/releases/latest`
+
+Installer scripts download the matching release asset for the local platform and
+verify its `.sha256` file. If Zero is already installed, run `zero update --check`
+before reinstalling.

docs/WORK_SPLIT_PRD.md

@@ -1,286 +0,0 @@
-# Zero WorkSplit PRD v3: Go-Native Product Ownership Plan
-
-Status: Draft v3.0
-Date: 2026-06-06
-Timeline: Active sprint plan toward v0.1, then v1.0 hardening
-
-Team:
-
-- Vasanth
-- Gnanam
-- Anandan
-
-## 1. Product Goal
-
-Zero is a terminal-first AI coding agent built in Go.
-
-The goal is not just to remove the old TypeScript runtime. The goal is to make Zero feel like a mature daily-driver CLI agent:
-
-- fast startup
-- strong command UX
-- safe autonomy
-- clear permission handling
-- useful local tools
-- reliable provider/runtime behavior
-- installable, updateable, and testable on real developer machines
-
-This document replaces the older migration-era work split. The Go runtime is now the main product baseline, so future work should focus on product depth rather than repeating the old M0 migration plan.
-
-## 2. Current Baseline On Main
-
-The current main branch is Go-native for the runtime and CLI.
-
-Implemented baseline:
-
-- `cmd/zero` is the primary CLI entrypoint.
-- `cmd/zero-pr-review` exists for PR automation workflows.
-- Core runtime packages live under `internal/`.
-- Current CLI surface includes `exec`, `config`, `models`, `providers`, `doctor`, `search`, `sessions`, `plugins`, `hooks`, `mcp`, `sandbox`, `update`, `worktrees`, `verify`, `changes`, `serve`, `help`, and `version`.
-- Core areas exist for agent flow, provider/model registry, tools, TUI, sessions, search, sandbox, self-verification, stream-json, usage, redaction, hooks, MCP, plugins, update, worktrees, and git/change handling.
-- TypeScript is no longer part of the repository baseline. The npm package keeps a small Node wrapper that delegates to the Go binary.
-
-Still needed before Zero feels production-ready:
-
-- Permission and sandbox UX must become clearer and harder to misuse.
-- TUI command flows need polish beyond basic rendering.
-- Platform behavior, especially Windows sandbox/update/install paths, needs stronger product ownership.
-- Review automation should produce useful, repeatable output for maintainers.
-- Release trust needs checksums, install smoke tests, and update verification.
-
-## 3. Ownership Rules
-
-- Every feature has one DRI.
-- UI, runtime, and platform work should be split by contract, not by vague file ownership.
-- Every PR must expose user-runnable behavior or a concrete contract consumed by another owner.
-- Docs, CI, and website work can support a product PR, but they should not be the whole lane for any owner.
-- Anandan must own platform product work, not only docs/site/CI.
-- Shared contracts should be agreed before large UI or integration work lands.
-- Keep PRs large enough to move the product forward, but still reviewable.
-
-## 4. Team Roles
-
-| Owner | Primary Role | Owns | Should Not Own Alone |
-| --- | --- | --- | --- |
-| Vasanth | Product UX and TUI owner | Bubble Tea TUI, startup screen, command palette, slash command UX, tool rendering, permission prompts, sandbox block rendering, verification/test/git output UX, themes, daily CLI feel | Provider internals, release pipeline, platform adapters |
-| Gnanam | Runtime core owner | Provider/model registry, provider factory, agent/runtime protocols, stream-json, sessions/search backends, config/doctor/usage/redaction, MCP/hooks/plugins backend, sandbox policy/grants, permission event contract | TUI composition, release packaging, platform-specific install trust |
-| Anandan | Platform product owner | Install/update/release trust, binary/package smoke tests, platform sandbox adapters, Windows behavior, PR/review automation CLI, CI as product verification, performance and release checks | TUI rendering, provider semantics, model registry ownership |
-
-## 5. Anandan Platform Product Rule
-
-Anandan should not be assigned only docs/site/CI forever. His lane must produce platform behavior that users and maintainers can run.
-
-Accepted Anandan work:
-
-- `zero sandbox policy` improvements and platform-specific sandbox behavior.
-- Windows sandbox adapter behavior and clear fallback reporting.
-- `zero update` verification, release metadata validation, and install smoke tests.
-- Package and binary checksums.
-- PR automation CLI workflows that reviewers can run locally or in CI.
-- CI jobs that prove real Zero commands work after installation.
-
-Not accepted as standalone Anandan work:
-
-- Docs-only updates.
-- Website-only updates.
-- Future VS Code notes with no current CLI behavior.
-- CI matrix changes that do not prove a user-facing command or release flow.
-- Release notes or checklists without executable validation.
-
-Docs and CI are still useful, but they must attach to real platform product behavior.
-
-## 6. Shared Contracts
-
-| Contract | DRI | Consumers | Purpose |
-| --- | --- | --- | --- |
-| Provider stream events | Gnanam | Vasanth, Anandan | Lets TUI and automation render model output, tool calls, and errors consistently. |
-| Permission and sandbox events | Gnanam | Vasanth, Anandan | Lets TUI ask for decisions and platform code enforce them without bypasses. |
-| Session event schema | Gnanam | Vasanth, Anandan | Lets TUI, resume, search, and automation share one session model. |
-| Command metadata registry | Vasanth | Gnanam, Anandan | Lets slash commands, help, shell completions, and docs use one command source. |
-| Platform/release contract | Anandan | Vasanth, Gnanam | Defines install, update, checksum, smoke, and platform capability results. |
-| Stream-json protocol | Gnanam | Anandan | Lets CI and PR automation consume Zero output without scraping text UI. |
-
-## 7. Active Workstreams
-
-### A. Safe Autonomy
-
-Top priority. Zero must be able to run tools confidently while showing users what is happening and asking when risk increases.
-
-Planned PRs:
-
-- Vasanth: `feat/tui-sandbox-permissions`
-  - permission prompt UX
-  - allow, deny, and always-allow flows
-  - sandbox block rendering
-  - `/permissions` or equivalent permission state UI
-- Gnanam: `feat/runtime-permission-events`
-  - structured permission request events
-  - sandbox grant contract
-  - stream-json support
-  - tests for denial, approval, and block cases
-- Anandan: `feat/windows-sandbox-platform`
-  - platform sandbox capability checks
-  - stronger Windows behavior or explicit safe fallback
-  - `zero sandbox policy --json` output useful enough for CI and support
-  - platform tests where possible
-
-Done when:
-
-- High-autonomy tool attempts produce a clear prompt, decision, and result in TUI and headless modes.
-- Denied commands do not run.
-- Sandbox blocks are visible and actionable.
-- Platform fallback is explicit, tested, and not silently unsafe.
-
-### B. Command UX And Daily CLI
-
-The CLI should feel like a product, not a collection of backend commands.
-
-Planned PRs:
-
-- Vasanth: `feat/tui-command-polish`
-  - `/model`, `/provider`, `/context`, `/doctor`, `/verify`, `/resume`, and `/compact` become clear interactive flows
-  - command results render consistently
-  - startup screen stays minimal until the user asks for panels
-- Gnanam: `feat/runtime-command-contracts`
-  - stable command data APIs
-  - provider/model/session/config result structs
-  - error types suitable for TUI rendering
-- Anandan: `feat/install-update-user-flow`
-  - `zero update` and install flows are user-runnable
-  - package smoke command verifies installed binary behavior
-  - CI proves the packaged command starts and reports version/help
-
-Done when:
-
-- A new user can run Zero, inspect config/provider/model state, and understand the next action without reading source.
-
-### C. Verification And Review Automation
-
-Zero should help maintainers verify changes and review PRs with repeatable local commands.
-
-Planned PRs:
-
-- Vasanth: `feat/tui-verification-results`
-  - render verify/test/git/review results in TUI
-  - show pass/fail/blocked states clearly
-- Gnanam: `feat/runtime-self-verify-contract`
-  - structured verification results
-  - stable test/change event model
-  - redacted logs for model and automation use
-- Anandan: `feat/pr-review-cli-workflow`
-  - make `cmd/zero-pr-review` or `zero review` useful for changed files, validations, and summary output
-  - support local and CI invocation
-  - avoid brittle text scraping where stream-json can be used
-
-Done when:
-
-- A maintainer can run one local command before review and get useful changed-file, validation, and risk output.
-
-### D. Release Trust And Distribution
-
-This is Anandan's main product lane after platform sandbox work.
-
-Planned PRs:
-
-- Anandan: `feat/release-trust`
-  - release metadata validation
-  - checksums
-  - install smoke tests
-  - update verification
-  - clear error output when release artifacts are invalid
-- Vasanth:
-  - consume release/update results in CLI UX where needed
-- Gnanam:
-  - ensure config, errors, and logs are structured and redacted
-
-Done when:
-
-- Users can install, run, verify, and update Zero with confidence.
-- Maintainers can prove release artifacts work before publishing.
-
-### E. Extensibility UX
-
-This comes after safe autonomy and daily CLI polish.
-
-Planned PRs:
-
-- Vasanth:
-  - `/mcp`, `/hooks`, `/plugins`, and related UI flows
-- Gnanam:
-  - harden backend contracts and lifecycle behavior
-- Anandan:
-  - automation and docs only after executable behavior exists
-
-Done when:
-
-- Extensibility feels discoverable and safe from the terminal.
-
-## 8. Immediate Next PRs
-
-These are the next reviewable slices. Do not use the old migration-era next steps.
-
-1. Vasanth: `feat/tui-sandbox-permissions`
-2. Gnanam: `feat/runtime-permission-events`
-3. Anandan: `feat/windows-sandbox-platform`
-4. Anandan: `feat/update-release-verification`
-5. Vasanth: `feat/tui-command-polish`
-6. Gnanam: `feat/session-rewind-compaction-contract`
-
-Recommended first parallel set:
-
-- Vasanth starts the TUI permission prompt and block display.
-- Gnanam defines the permission event/grant contract.
-- Anandan owns Windows/platform sandbox reporting and tests.
-
-## 9. Command Ownership Matrix
-
-| Command Area | Vasanth | Gnanam | Anandan |
-| --- | --- | --- | --- |
-| `exec` | TUI and interactive UX | Runtime execution contract | Package smoke and platform validation |
-| `sandbox` | Prompt and block UX | Policy, grants, and events | Platform adapters and Windows behavior |
-| `update` | Display and user messaging | Config and redacted errors | Primary owner of update verification |
-| `verify` | Result rendering | Verification backend contract | CI and release integration |
-| `changes` | TUI rendering | Change detection backend | PR automation integration |
-| `mcp`, `hooks`, `plugins`, `serve` | Discoverability and UI | Backend lifecycle and contracts | Automation/docs after behavior exists |
-| `sessions`, `search` | Resume/search UX | Storage, query, and event model | Stream-json and automation consumers |
-
-## 10. Definition Of Done
-
-For every code PR:
-
-- Add or update tests for the changed behavior.
-- Run `go test ./...` when Go files changed.
-- Run the relevant build command when CLI entrypoints changed.
-- Run package smoke checks when install/update/npm wrapper files changed.
-- Include manual CLI output or screenshots only when they prove a user-facing flow.
-
-For docs-only PRs:
-
-- State clearly that no runtime behavior changed.
-- Do not use docs-only PRs as a substitute for platform product work.
-
-For Anandan's PRs specifically:
-
-- Must include one of:
-  - user-runnable command behavior
-  - install/update/release artifact verification
-  - platform adapter behavior
-  - CI that executes real installed Zero commands
-- Docs/CI can support the PR, but cannot be the entire value.
-
-## 11. Review Rotation
-
-- Vasanth PRs:
-  - Gnanam reviews runtime contracts.
-  - Anandan reviews platform/release impact when touched.
-- Gnanam PRs:
-  - Vasanth reviews user-facing contract impact.
-  - Anandan reviews platform/protocol impact when touched.
-- Anandan PRs:
-  - Vasanth reviews user-facing CLI behavior.
-  - Gnanam reviews runtime contract and safety impact.
-
-## 12. Retired Historical Plan
-
-The old M0-M6 TypeScript-to-Go migration schedule is now historical.
-
-Do not assign new work from the old migration next-steps section. The Go migration foundation has already landed on main. Future work should move Zero from "ported" to "mature terminal coding agent."

docs/audit/2026-06-10-deep-audit-status.md

@@ -1,208 +0,0 @@
-# Deep-audit status ledger (re-verified against HEAD on branch fix/audit-batch, 2026-06-13)
-
-Total verified: 175 of 175 extracted findings
-
-| Status | Count |
-|---|---|
-| fixed | 97 |
-| open | 66 |
-| partial | 12 |
-| obsolete | 0 |
-
-<!--
-Counts are the post-re-verification view: of the 87 rows the body still marks
-open/partial (69 open + 18 partial), the re-verification found 9 already fixed on
-`main` (3 open + 6 partial; listed in 2026-06-13-reverification.md), leaving 66
-open + 12 partial = 78 genuinely pending, and fixed 88 + 9 = 97. The per-finding
-rows below retain their original verdicts pending a row-level reclassification pass.
--->
-
-
-> **Authoritative re-verification (2026-06-13, branch `fix/audit-batch` off current `main`).**
-> All 87 rows the 06-11 snapshot marked `open`/`partial` were re-checked against the ACTUAL
-> current code by independent agents (the 06-11 ledger was verified on `fix/audit-critical-high`,
-> which has since diverged). Every `already_fixed` verdict was adversarially re-checked by a
-> second agent trying to prove the defect was still present — none were refuted.
-> **Result: 78 genuinely pending** (66 still_open + 12 partial); **9 rows were already fixed on
-> `main`** and are reclassified to `fixed` here. Full per-finding table:
-> [`2026-06-13-reverification.md`](2026-06-13-reverification.md).
-> Pending by severity: **high 2, medium 42, low 34**. Counts above reflect this re-verification
-> plus this branch's own fixes (scanner.go:81, compaction.go:58, worktrees.go:247).
-
-| Sev | Status | Location | Evidence |
-|---|---|---|---|
-| critical | fixed | internal/search/search.go:346 | search.go:363-428 findMatch maps lowered offsets back via lowerWithOffsets + snapRuneStart clamps context; /tmp probe of audit repro (100x'Ⱥ'+' hello') returns match {201 206}, valid UTF-8, no panic |
-| critical | fixed | internal/mcp/protocol.go:69 | 82b9c8d: const maxMessageBytes=64MiB (protocol.go:18) bounds both framing paths — Content-Length values above the cap are rejected (:90-92) and the newline-delimited readLine errors past the cap (:78-80); stdio transport switched to newline-delimited JSON. Covered by protocol_test.go; go test green |
-| high | fixed | internal/cli/usage.go:121 | usage.go:123-131: diff defaults to zerogit.DiffStat{}; resolveWorkspaceRoot/inspectChanges errors are ignored (degrade to net-LOC 0) instead of returning |
-| high | fixed | internal/tui/transcript.go:137 | transcript.go:121-142 keys are run-scoped ("%d:%d:%s" with row.runID) plus effectiveToolRowID ordinal suffix; model.go:1501/1553/1572/1584 set runID on rows |
-| high | fixed | internal/tui/model.go:853 | model.go:1034-1043 commandExit now calls cancelRun(), sets exiting, and defers tea.Quit while flushRunIDs is non-empty; deferred quit fires after flush at model.go:645-647 |
-| high | fixed | internal/tui/model.go:599 | internal/tui/flush.go:11-30 settled-row flush frontier prints history to native scrollback via tea.Println; model.go transcriptView() renders only live tail from m.flushed; covered by internal/tui/flush_test.go |
-| high | fixed | internal/tui/model.go:512 | model.go:614,633-637 flush goes to flushSessionID recorded at cancel time (model.go:1334) via appendSessionEventsTo (tui/session.go:76), not the active session |
-| high | fixed | internal/tui/session_controls.go:176 | session_controls.go:182-184 blocks /rewind while len(m.flushRunIDs) > 0 ('cannot rewind while a cancelled run is still flushing'); covered by tui/session_test.go |
-| high | fixed | internal/tui/model.go:584 | flush.go:71-120 settles rows into native scrollback via tea.Println; View()/transcriptView renders only rows from m.flushed (model.go:751-784) |
-| high | fixed | internal/tui/transcript.go:137 | transcript.go:121-141 bakes runID into every key; within-run repeats get ordinal suffix via effectiveToolRowID (transcript.go:148, used at model.go:1493-1496 with callSeq) |
-| high | fixed | internal/tui/rendering.go:525 | live cap 16 kept for tidy tail only; flushed scrollback cards use flushCardBodyMaxLines=400 (rendering.go:154-158) and Ctrl+O detailed view is uncapped (renderRowDetailed bodyCap:0, rendering.go:165; capCardLines treats <=0 as no cap, rendering.go:718-720) |
-| high | fixed | internal/tui/transcript.go:135 | transcript.go:121-141 transcriptRowKey bakes runID into every key ("%d:%d:%s"); model.go:1501/1553 set row.runID and effectiveToolRowID suffixes within-run repeats |
-| high | fixed | internal/tui/model.go:512 | model.go:614-637 — flushRunIDs maps cancelled run -> session id captured at cancel time (model.go:1334); flush goes via appendSessionEventsTo(flushSessionID,…) (session.go:76) when active session differs |
-| high | fixed | internal/tui/run.go:26 | flush.go:71-120 settled-row frontier prints history to native scrollback via tea.Println; model.go:769 View renders only the unflushed live tail |
-| high | fixed | internal/tui/transcript.go:139 | transcript.go:125 key is now "%d:%d:%s" (kind:runID:id) for all keyed kinds; transcript.go:148 effectiveToolRowID adds per-run ordinal suffix for within-run repeats |
-| high | fixed | internal/tui/model.go:311 | model.go:355-360 Ctrl+C sets exiting and returns nil (no Quit) while flushRunIDs non-empty; deferred tea.Quit fires at model.go:645-647 once the cancelled run's session events flush |
-| high | fixed | internal/tui/model.go:853 | model.go:1034-1043 commandExit now calls cancelRun(), sets m.exiting, and defers tea.Quit while len(m.flushRunIDs) > 0 — same protection as Ctrl+C |
-| high | fixed | internal/cli/extensions.go:186 | internal/cli/extensions.go:191 passes Servers through redactMCPServerConfigs (:205-224) which replaces every env/header value with [REDACTED]; asserted by extensions_test.go:108 (go test ok) |
-| high | fixed | internal/mcp/protocol.go:42 | 82b9c8d: read() now accepts newline-delimited JSON frames (isJSONStart→decodeMessage) and only falls back to Content-Length headers for back-compat; write() emits one JSON message per line. Covered by protocol_test.go; go test green |
-| high | fixed | internal/streamjson/streamjson.go:310 | 3142de6: secretPatterns now anchored — sk keys require a word boundary plus a 16-char body and bearer requires a 16-char token — so 'task-list' and prose like 'bearer of bad news' are no longer redacted while real keys/tokens still are. Covered by streamjson_test.go |
-| high | fixed | internal/cli/cron.go:112 | 7a47059: cronAdd now consumes an explicit positional expr (cron.go:113-118) BEFORE applying recipe defaults, so a schedule given alongside --recipe is honored; extra positionals error. Covered by cron_test.go |
-| high | fixed | internal/providerhealth/providerhealth.go:285 | 600057d: default probe now uses newConnectivityClient — CheckRedirect re-validates every hop (capped at 5) and DialContext validates+pins the resolved IP, dialing the IP literal to close the DNS-rebinding TOCTOU. Covered by connectivity_client_test.go |
-| high | fixed | internal/sessions/store.go:547 | 12a3f62 + 8c75074: ReadEvents now drops a torn final record (last non-empty line) only when the file lacks a trailing newline (tornTailPossible); a fully-flushed corrupt line, or any malformed earlier line, still fails loudly. Covered by store_test.go |
-| high | partial | internal/cli/cron_run.go:173 | 7a47059: fireJob now re-reads the job before persisting and preserves an external pause/removal mid-run (cron_run.go:170-181), closing the pause-clobber. But internal/cron/store.go still has no flock/CAS, so the concurrent-scheduler race (two schedulers both seeing NextRunAt due) remains — single-scheduler is the supported model |
-| high | fixed | internal/tools/bash.go:104 | 0dbb5e0: hardenProcessLifetime (bash_proc_unix.go) sets Setpgid, a Cancel that SIGKILLs the whole process group (negative pid), and a WaitDelay backstop so a backgrounded child cannot outlive the timeout or hang Run(); Windows sets WaitDelay only. Covered by bash_tool_test.go |
-| high | fixed | internal/sandbox/runner.go:262 | 2eeaea7: sandbox-exec profile write rule now also allows the standard device nodes (/dev/null, …) and temp trees (/tmp, /private/var/folders, …) for parity with bubblewrap's --dev/--tmpfs; the workspace stays the only writable project location. Covered by runner_test.go |
-| high | fixed | internal/tools/apply_patch.go:161 | b6182bf: patch headers are now parsed with hunk line-counting (git-apply semantics), so a hunk-body line like '-- /etc/old' is no longer mistaken for a file header and a valid patch is accepted. Covered by write_tools_test.go |
-| high | fixed | internal/tools/grep.go:224 | b6182bf: glob now matches against the path relative to the SEARCH directory (ripgrep semantics) via filepath.Rel(target,path); a single explicit file matches by base name. probe path=subdir glob=*.go now matches subdir/a.go. Covered by file_tools_test.go |
-| high | fixed | internal/providers/openai/types.go:3 | 46d495c (+7044d36/6664f71): chatCompletionRequest gained reasoning_effort; zeroruntime.CompletionRequest.ReasoningEffort is threaded through the agent loop and mapped per provider — OpenAI reasoning_effort, Anthropic extended thinking, Gemini thinking budget; CLI gates the value against the resolved model. Covered by openai/anthropic/gemini provider_test.go |
-| high | fixed | internal/specialist/exec.go:166 | 39a9c55: BuildArgs always prefixes the session title with the specialist name and falls back to the bare name when description is empty, so AgentName is non-empty and a description-less run resumes. Covered by specialist/exec_test.go |
-| high | fixed | internal/agent/loop.go:233 | 95a6a8e: Run now sets result.FinishReason from collected.FinishReason (and from the after-max-turns final answer); Result.Truncated()/TruncationNotice() drive an exec-writer warning and a TUI system row when a response was cut off. Covered by loop_test.go |
-| high | partial | internal/hooks/hooks.go:372 | 17a91d9: a Dispatcher that runs configured hooks and audits each via NewAuditStore now exists and is tested (dispatch.go + dispatch_test.go), closing the 'zero non-test callers' gap for Select/audit. But it has ZERO production callers — not wired into the beforeTool/afterTool tool lifecycle — and runHooks still supports only `list` (extensions.go:94); no dispatch/edit CLI surface |
-| low | fixed | internal/cli/exec_writer.go:412 | exec_writer.go:415 and :404 now call cutRuneBoundary (:420-431) which backs up to the last UTF-8 rune start before slicing; added in audit-fix commit 6fb0231 |
-| low | fixed | internal/mcp/server.go:129 | server.go:129-132 adds `case "ping": return server.writeResult(message.ID, map[string]any{})` (introduced in commit 6fb0231, present at HEAD) |
-| low | fixed | internal/search/search.go:309 | search.go:318-322: resolveSessions now returns fmt.Errorf("zero session not found: %s", sessionID) when store.Get returns nil instead of an empty success |
-| low | fixed | internal/doctor/doctor.go:134 | doctor.go:129-141: providerConfigCheck now reports apiKey presence ("set"/"not set") based on profile.APIKey != "", with comment explaining the change; doctor tests pass |
-| low | fixed | internal/tui/transcript.go:303 | truncateTUIOutput cuts via rune-boundary cutRunes (transcript.go:302,307-318); titles use cutRunes (session.go:112, spec_mode.go:305) and cutRuneBoundary (exec_sessions.go:90, exec_writer.go:420) |
-| low | fixed | go.mod:10 | go.mod:7-12 direct requires are now only bubbles/bubbletea/lipgloss/x-sys; glamour is gone entirely and x/ansi (line 18) and termenv (line 31) are demoted to // indirect |
-| low | fixed | internal/tui/model.go:525 | agentResponseMsg completion path now invokes m.runCancel() before nil-ing it (model.go:651-658, with explanatory comment about the per-prompt leak) |
-| low | fixed | internal/imageinput/imageinput.go:58 | imageinput.go:56-66 os.Open failure now wraps real cause ("cannot open image %s: %w") and read failure "cannot read image %s: %w"; only a stat miss reports "not found" (:38) |
-| low | fixed | internal/tui/session.go:93 | All three paths rune-safe: internal/tui/session.go:112 cutRunes(..., tuiSessionTitleLimit), internal/tui/spec_mode.go:305 cutRunes, internal/cli/exec_sessions.go:90 cutRuneBoundary(title, 80) |
-| low | fixed | internal/cli/exec_writer.go:412 | internal/cli/exec_writer.go:404 (truncateForStatus) and :415 (truncateForStreamJSONOutput) both use cutRuneBoundary defined at exec_writer.go:420-431 (backs up to utf8.RuneStart) |
-| low | fixed | internal/sessions/replay.go:237 | replay.go:245-256 cutPromptRuneBoundary (utf8.RuneStart backoff) now used at replay.go:174-177 and 238; titles rune-safe via cutRuneBoundary (cli/exec_sessions.go:90) and cutRunes (tui/session.go:112) |
-| low | fixed | internal/tui/rendering.go:68 | footerText/commandFooterText/formatCommandFooterText/defaultCommandFooterText (present at b144cf9~1 rendering.go) are absent at HEAD; dead actionAppendToolCall/ToolResult enum values also removed from transcript.go |
-| low | fixed | internal/tui/transcript.go:303 | cutRunes lands cuts on rune boundaries (transcript.go:305-318), used by truncateTUIOutput for tool-result text (transcript.go:302) and session titles (session.go:112, spec_mode.go:305) |
-| low | fixed | internal/tui/model.go:853 | commandExit cancels the run, sets exiting, and defers quit while flushRunIDs is non-empty (model.go:1034-1043); deferred tea.Quit fires when the last flush drains (model.go:645-647), and new prompts are gated while exiting (model.go:1012-1014) |
-| low | fixed | internal/tui/session.go:93 | cutRunes (transcript.go:307-318) cuts on rune boundaries; used by tuiSessionTitle (session.go:112) and truncateTUIOutput (transcript.go:302) for tool-result text |
-| low | fixed | internal/tui/rendering.go:482 | rendering.go:526 — permission card action row now renders "[esc] cancel run", matching the KeyEsc fall-through to m.cancelRun() (model.go:391-393) |
-| low | fixed | internal/cli/update.go:37 | internal/update/update.go:136-142: normalizeVersionTag failure now falls back to currentVersion="0.0.0" so the check proceeds to the network call (landed in #158) |
-| low | fixed | internal/notify/notify.go:134 | notify.go:137 now drops `(r >= 0x80 && r <= 0x9f)` with comment that C1 includes single-byte CSI/OSC/ST introducers (landed in #158) |
-| low | fixed | internal/tui/model.go:512 | model.go:616-624 the flushRunIDs (cancelled-run) branch of agentResponseMsg now records each usage event via recordUsageEvent → m.usageTracker.Record (session_controls.go:294) |
-| low | fixed | internal/tui/rendering.go:490 | rendering.go:560-564 the ask card now echoes the in-progress answer: "❯ " + fill(zeroTheme.ink).Render(input) + accent cursor, appended to the card lines |
-| low | fixed | internal/tui/view.go:227 | view.go:230-235 now matches `path == home` or strings.HasPrefix(path, home+os.PathSeparator) — a path-boundary check with a comment citing the /Users/alice2 sibling case |
-| low | fixed | internal/sandbox/risk.go:53 | defect not present: risk.go:11-50 patterns are package-level regexp.MustCompile compiled once at init (identical at audit commit b144cf9); per call only MatchString runs |
-| low | fixed | internal/modelregistry/catalog.go:36 | catalog.go:36 claude-haiku-3.5 extra capabilities now []ModelCapability{ModelCapabilityPromptCache} — ModelCapabilityVision removed (base withBaseCapabilities adds no vision), so SupportsVision is false |
-| low | fixed | internal/providers/providerio/retry.go:107 | retry.go:108-110 ShouldRetryStatus now returns true for `code == 529` alongside 429/503 (doc comment lines 99-107 updated to cover Anthropic overloaded), consistent with ClassifiedError's 529 rate-limit classification at providerio.go:224 |
-| low | fixed | internal/tui/spec_mode.go:301 | spec_mode.go:305 and session.go:112 now use cutRunes (transcript.go:307-318 backs off to utf8.RuneStart boundary); no byte-index title slicing remains |
-| low | fixed | internal/agent/system_prompt.go:89 | system_prompt.go:102-107 now walks cut back with `for cut > 0 && !utf8.RuneStart(content[cut])` before slicing, so truncation lands on a rune boundary (matches audit doc's rune-safe truncation fix claim) |
-| low | fixed | internal/config/writer.go:53 | internal/config/writer.go:55-74 — writeConfigFile now does os.CreateTemp + Chmod(0600) + write + os.Rename (atomic temp+rename, with explanatory comment at :53); go test ./internal/config -run Upsert passes |
-| low | fixed | internal/cli/exec_sessions.go:88 | exec_sessions.go:87-91: title = cutRuneBoundary(title, 80); helper (exec_writer.go:420-431) verified rune-safe by /tmp probe |
-| low | fixed | internal/cli/exec_writer.go:412 | exec_writer.go:404 (truncateForStatus/stderr) and :415 (truncateForStreamJSONOutput) both use cutRuneBoundary, which backs up to a utf8.RuneStart boundary |
-| low | open | internal/streamjson/streamjson.go:30 | repo-wide grep finds EventRestore only at its declaration (streamjson.go:30) — no emitter and not even test references; streamjson.go untouched by 6fb0231 or later merges |
-| low | open | internal/cli/cron_run.go:150 | cron_run.go:145-151: rec.Error still set only from errBuf on code!=0; outBuf (where stream-json error events land) is captured but never read |
-| low | open | internal/cli/exec_sessions.go:55 | exec_sessions.go:55 still constructs sessions.NewStore(sessions.StoreOptions{}) directly, and exec.go:375-389 PrepareExecOptions still omits the Store field; deps.newSessionStore is used by spec/usage/observability but not the exec run path |
-| low | fixed | internal/secrets/scanner.go:35 | [reverified 2026-06-13] already fixed on main: scanner.go:37 pattern is now `sk-[A-Za-z0-9_-]{20,}` (allows -/_ so sk-proj-/sk-svcacct- match), with comment at :35-36. Covered by scanner_test.go TestScanDetectsModernPrefixedOpenAIKeys |
-| low | open | internal/redaction/redaction.go:222 | redaction.go:221-235 pointer/map ptr added to context.seen and never deleted after subtree recursion, so a second (sibling) visit returns CircularReference |
-| low | open | internal/redaction/redaction.go:174 | redaction.go:174 `interface{ StackTrace() fmt.Stringer }` unchanged; grep finds no implementer in the module and go.mod has no pkg/errors or similar dep |
-| low | open | internal/zerogit/zerogit.go:417 | zerogit.go:418-420 fallback EnvRunner adapter still discards env (`_ []string`), so stagedSnapshotDiff's GIT_INDEX_FILE isolation is lost with a caller-supplied Runner |
-| low | open | internal/tui/commands_test.go:46 | internal/tui/commands_test.go:46 still defines commandTestStringSliceContains; repo-wide grep finds no callers, only the definition |
-| low | open | internal/tui/session_test.go:215 | internal/tui/session_test.go:215 'runtimeMessages := []tea.Msg{}' appended in RuntimeMessageSink at line 226 but never read in the test body (only runtimeMessageCh is consumed) |
-| low | open | internal/sessions/store.go:555 | store.go:554-556 timestamp() still time.RFC3339 (second precision); List ties broken lexically by SessionID (store.go:331-336), not actual recency, so Latest() can still pick the wrong session on same-second ties |
-| low | open | internal/tui/session_controls.go:165 | session_controls.go:165 still only increments compactRequests; sole reads are its own status display (lines 266, 280) with "Backend: state: pending integration" — no compaction is ever triggered by it |
-| low | open | internal/tui/model.go:837 | model.go:1012-1014 silently drops prompt submits while m.exiting (return m, nil — no feedback row); no exiting/flushRunIDs indicator anywhere in statusLine (view.go:113-139), composerLine, or any render path |
-| low | open | internal/tui/command_center.go:17 | tui/command_center.go:17-21 passes only Now/Runtime/Provider (no UserConfig/ProjectConfig) vs cli/observability.go:64-71; empty paths warn at doctor.go:120 and :273 |
-| low | open | internal/background/process_posix.go:38 | process_posix.go:39-47 still polls signal-0 on os.FindProcess(pid) while the Wait goroutine (specialist/exec.go:653) reaps the child; line 50 Kill() can hit a recycled pid |
-| low | open | internal/tools/registry.go:108 | repo-wide grep: OnSandboxDecision only at registry.go:19,108,116 — no caller or test ever sets it |
-| low | open | internal/sandbox/grants.go:146 | grants.go unchanged since #77: Lookup trims key (grants.go:146) but readState validates raw file keys via trimming ValidateToolName (grants.go:225-227,273) and keeps the raw map key, so a whitespace-padded key validates yet never matches |
-| low | open | internal/tools/read_file.go:89 | probe: max_lines=5 -> Truncated=true with no marker beyond '(lines 1-5 of 20)' header (read_file.go:102-111); rune-width sub-claim not reproducible (ASCII strconv.Itoa padding) |
-| low | open | internal/providers/factory.go:118 | factory.go:115 still `if !explicitProvider \|\| isImplicitOpenAI(...)`; isImplicitOpenAI (factory.go:159-164) requires TrimSpace(ProviderKind)=="" && TrimSpace(Provider)=="", which contradicts explicitProvider==true (factory.go:147-157), so it can never return true at its sole call site |
-| low | open | internal/providercatalog/catalog.go:46 | catalog.go:46 Public still declared; no constructor or descriptor ever sets it (always false); the only read is the vacuous guard in catalog_test.go:105, which already existed before the audit (added in 4accec5/#141) — no production read or write |
-| low | open | internal/verify/verify.go:215 | verify.go:215 RunLoop still defined, sole caller verify_test.go:207; production loop is selfverify.Run (selfverify.go:53-74) used by cli/workflows.go |
-| low | open | internal/specialist/output_tool.go:244 | output_tool.go:244 formatTaskOutput has zero callers (prod or tests); only formatTaskOutputSummary is used (output_tool.go:154) |
-| low | open | internal/specialist/exec.go:346 | exec.go:347-350: on SetPID error calls cleanupBackgroundPromptFile (runtime.go:73 deletes file/dir) and returns err with no recordSpecialistStop/UpdateStatus while child keeps running |
-| low | open | internal/agent/loop.go:188 | helpers.go:85 sets collected.Error = ctx.Err().Error(); loop.go:189-191 returns errors.New(collected.Error) before the ctx.Err() identity return at loop.go:193-195, so errors.Is(err, context.Canceled) still fails for that path |
-| low | open | internal/agent/types.go:183 | types.go:183 OnContext field and loop.go:102-103 emit site exist, but repo-wide grep finds no "OnContext:" assignment and no MeasureContext/ContextBreakdown use in cli/tui/cmd/contextreport (only tests) |
-| low | open | internal/agent/loop.go:1072 | loop.go:1120-1122 and loop.go:1129-1131 both still assign schema["items"] = propertyToRuntimeMap(*property.Items) |
-| low | open | internal/agent/loop.go:473 | loop.go:530-531 set requestEvent.GrantMatched/Grant but requestEvent is never read after the switch (post-run event rebuilt at loop.go:558); fallback at loop.go:510-511 unreachable since buildPermissionEvent always returns ok=true under the shouldRequestPermission gate (Prompt-permission tools, loop.go:751-758 vs 877-894) |
-| low | open | internal/agent/loop.go:183 | loop.go:184-186 retry CollectStreamWithOptions still passes only OnUsage; comment at loop.go:178-183 documents the OnText omission as deliberate anti-duplication, but streaming surfaces still receive only the aborted partial text |
-| low | open | internal/zeroruntime/helpers.go:104 | helpers.go:104 still does `collected.Text += event.Content` per StreamEventText (no strings.Builder); unchanged since audit commit b144cf9 |
-| low | open | internal/zerocommands/sandbox_snapshots.go:104 | grep for SandboxPlan/Decision/Policy/Backend/Risk/Violation/Hook/Plugin/MCPServer snapshots finds no consumers outside internal/zerocommands + tests; `zero sandbox policy --json` serializes raw sandbox.Policy/Backend/BackendPlan instead (internal/cli/sandbox.go:90-105) |
-| low | open | internal/config/contracts.go:12 | internal/config/contracts.go still defines ContractGap/DefaultContractGaps/FindContractGapsByMilestone; only reference repo-wide is internal/config/contracts_test.go |
-| low | open | internal/skills/skills.go:91 | skills.Duplicates (skills.go:91) and skills.Get (skills.go:217) have zero callers outside internal/skills; cli/skills.go uses List and tools/skill.go uses Load, so collisions are still dropped without any user-facing warning (dedup is now documented/deterministic but unsurfaced) |
-| low | open | internal/plugins/plugins.go:893 | internal/plugins/plugins.go:893 allows only beforeTool/afterTool/sessionStart/sessionEnd, while internal/hooks/hooks.go:812 also accepts specialistStart/specialistStop — sets still disagree |
-| low | open | internal/config/resolver.go:136 | internal/config/resolver.go:143, :173 and :502 all gate on `MaxTurns > 0` with no error path, so a configured maxTurns of 0 or negative is still silently ignored (falls back to default 30) |
-| low | open | internal/cli/exec.go:353 | exec.go:375 PrepareExec called without Store (defaults to NewStore in sessions/exec_session.go:58), exec_sessions.go:55 direct NewStore, exec.go:396 time.Now() instead of deps.now |
-| low | open | internal/zerocommands/contracts.go:166 | contracts.go:166 APIKeySet checks only profile.APIKey, ignoring AuthHeaderValue (valid sole credential per config/types.go:49-51); rendered "not set" at cli/command_center.go:328 |
-| low | partial | internal/cli/cron.go:275 | promptExcerpt fixed via cutRuneBoundary(p,47) at cron.go:275, but cronTruncate at cron_run.go:179-184 still returns s[:max]+"…" (byte slice, can split a rune in persisted run errors) |
-| low | partial | internal/tui/options.go:28 | app.go:413 production caller still sets none of the three, but newModel now self-defaults a usage.Tracker when nil (model.go:239-242) so usage works; effort/style are settable at runtime via /effort and /style |
-| low | partial | internal/tui/rendering.go:62 | Footer system, actionAppendToolCall/Result, and dead theme tokens (selRow/statusOk/statusErr/line2) are removed, but listCommandNames/formatCommandHelpLines remain test-only (commands.go:297,306), resumeText's args!="" branch is still unreachable (command_center.go:44-54; sole caller session.go:122 passes ""), OnContext is still wired by no production caller (agent/loop.go:102), and stale footer comments persist (commands.go:254, autocomplete.go:64) |
-| low | partial | internal/zerogit/zerogit.go:206 | truncateSubject now rune-based (zerogit.go:464-472, no invalid UTF-8), but ValidateMessage still byte-based `len(firstLine) > 72` (zerogit.go:207) so valid non-ASCII subjects are rejected |
-| low | partial | internal/cli/cron.go:275 | promptExcerpt fixed (cron.go:275 uses cutRuneBoundary) but error helper cronTruncate still byte-slices: internal/cli/cron_run.go:183 'return s[:max] + "…"', used for run errors at cron_run.go:150 |
-| low | partial | internal/tui/command_center.go:44 | Footer helpers now render (rendering.go:690-700, bash/grep footers 892-905/969-971), but resumeText's args!="" branch (command_center.go:43-54) is still unreachable (sole caller session.go:122 passes "") and stale footer comments persist (autocomplete.go:64, commands.go:254) |
-| low | partial | internal/cli/cron.go:275 | promptExcerpt fixed via cutRuneBoundary (cron.go:275), but cronTruncate still byte-slices `s[:max]` at cron_run.go:182, splitting runes in run-record error text |
-| low | partial | internal/tui/rendering.go:62 | footer builders (footerText/commandFooterText/etc.) and the four theme tokens (selRow/statusOk/statusErr/line2/panel2) are fully removed (zero grep hits), but formatCommandHelpLines/listCommandNames (commands.go:297-308) remain production code referenced only by tests, and stale "footer advertises" comments persist (commands.go:254, autocomplete.go:64) |
-| medium | fixed | internal/update/update.go:135 | update.go:136-142: normalizeVersionTag error now falls back to currentVersion="0.0.0" (commented for dev builds) instead of returning the error |
-| medium | fixed | internal/tui/session.go:190 | session.go:219-227 rehydrates role "ask_user"/"ask_user_answers" rows via askUserRequestFromPayload (session.go:349) and askUserAnswersText (session.go:380) |
-| medium | fixed | internal/tui/model.go:1067 | cancelRun appends rowSystem "Run cancelled." plus any partial streamed answer (model.go:1336-1343); asserted by flush_test.go:105 |
-| medium | fixed | internal/search/search.go:345 | search.go:363-407 findMatch maps lowered offsets back via lowerWithOffsets offset table; snapRuneStart (:417) guards context slicing; tests pass |
-| medium | fixed | internal/search/search.go:108 | search.go:111-118 LoadIndex error now `result.SkippedSessions++; continue` instead of returning; SkippedSessions reported in Result (:61) |
-| medium | fixed | .github/workflows/ci.yml:31 | .github/workflows/ci.yml:31-44 adds 'Check formatting' (gofmt -l, fails on output) and 'Vet' (go vet ./...) steps; gofmt -l . is clean at HEAD |
-| medium | fixed | internal/tui/transcript.go:303 | internal/tui/transcript.go:302 truncateTUIOutput now calls cutRunes (rune-boundary cut, transcript.go:307-318); non-ASCII test TestCutRunesNeverSplitsUTF8 at flush_test.go:158 |
-| medium | fixed | internal/cli/usage.go:46 | usage.go:47-51 collectUsageData now does set.skipped++ and continue on ReadEvents failure, so a corrupt session is skipped instead of aborting |
-| medium | fixed | internal/tui/model.go:1081 | cancelRun appends partial answer then rowSystem "Run cancelled." marker (model.go:1336-1343), shared by Esc (model.go:392) and Ctrl+C (model.go:355) paths |
-| medium | fixed | internal/tui/model.go:227 | WindowSizeMsg sets m.input.Width = maxInt(20, chatWidth(msg.Width)-14) so long input scrolls with cursor visible (model.go:556-558) |
-| medium | fixed | internal/tui/view.go:394 | view.go:412-431 now requires a hunk header match or BOTH '--- ' and '+++ ' file headers; a lone '---' line no longer triggers the diff renderer |
-| medium | fixed | internal/tui/rendering.go:261 | rendering.go:286-330 preserves explicit newlines and each line's leading indent (tabs→4 spaces), continuations keep the indent; only intra-line space runs collapse |
-| medium | fixed | internal/tui/model.go:456 | WindowSizeMsg resizes composer and prints title bar at real width (model.go:553-566); flush refuses to freeze history before a real width arrives (flush.go:72-77); width/tier recomputed every render (view.go:17-19) and history is native scrollback, so only the ephemeral pre-size frame uses the 96-col default |
-| medium | fixed | internal/tui/session_controls.go:176 | session_controls.go:182-184 — handleRewindCommand returns "cannot rewind while a cancelled run is still flushing" when len(m.flushRunIDs) > 0 |
-| medium | fixed | internal/tui/session.go:190 | model.go:1461-1476 persists ask_user question payload AND ask_user_answers; session.go:218-227 rehydrates both via askUserRequestFromPayload/askUserAnswersText (session.go:349, 380) |
-| medium | fixed | internal/tui/model.go:853 | model.go:1034-1043 — commandExit calls cancelRun(), sets exiting, defers tea.Quit until flushRunIDs drains (model.go:645); spec_mode.go:25 gates /spec on m.pending \|\| m.exiting |
-| medium | fixed | internal/tui/spec_mode.go:22 | spec_mode.go:25 handleSpecCommand gates on `if m.pending \|\| m.exiting` and refuses to start, with comment citing the post-Ctrl+C flush window |
-| medium | fixed | internal/tui/model.go:1081 | model.go:1340-1343 cancelRun appends the partial streamed answer plus rowSystem "Run cancelled." to m.transcript (live), in addition to the session-log marker |
-| medium | fixed | internal/tui/model.go:610 | model.go:769 transcriptView loops only `for index := m.flushed` (live tail; settled rows flushed once to scrollback per flush.go), and rendering.go:184-192 memoizes stable row renders in the LRU staticRenderCache (render_cache.go:19) |
-| medium | fixed | internal/tui/session.go:93 | session.go:112 tuiSessionTitle uses cutRunes; transcript.go:302/307-318 truncateTUIOutput cuts on the last rune boundary at or before the byte limit |
-| medium | fixed | internal/tui/model.go:227 | model.go:558 WindowSizeMsg handler sets `m.input.Width = maxInt(20, chatWidth(msg.Width)-14)` so long input scrolls horizontally with the cursor visible |
-| medium | fixed | internal/providers/providerio/providerio.go:184 | providerio.go:142-154 onComment now forwards a keepAlive marker to the consumer, which calls resetIdle() at providerio.go:203 before the keepAlive continue (line 204-205); providerio tests pass |
-| medium | fixed | internal/modelregistry/catalog.go:27 | catalog.go:27 gpt-4.1 entry now reads ContextLimits{ContextWindow: 1_047_576, MaxOutputTokens: 32_768}, matching mini/nano |
-| medium | fixed | internal/tui/run.go:41 | run.go:42-46 now prints `zero: tui error: <err>` to os.Stderr before returning 1 |
-| medium | open | internal/mcp/client.go:302 | client.go:295-317 readLoop dispatches any id-bearing message to pending[id] without checking message.Method; grep confirms no 'ping' handling anywhere in client.go |
-| medium | open | internal/mcp/client.go:246 | client.go:246 `client.writer.write(...)` is a blocking pipe write executed while client.mu is held (:228-:255) with no ctx awareness; file unchanged since the audit (only the response wait, which predates the audit, is ctx-aware) |
-| medium | open | internal/mcp/client.go:98 | client.go:98-99 `var stderr bytes.Buffer; cmd.Stderr = &stderr` — plain unbounded buffer attached for the process lifetime, only read on initialize failure |
-| medium | open | internal/mcp/network_client.go:621 | network_client.go:603-632 decodeSSERPCMessage stops at the first non-empty 'message' event (sets found, returns false); caller at :210 errors on id mismatch instead of scanning for the matching response — unchanged since audit |
-| medium | open | internal/mcp/network_client.go:637 | network_client.go:637 `scanner.Buffer(make([]byte, 0, 64*1024), 1024*1024)`; ErrTooLong propagates via readStream->failPending (:578-:592) setting a permanent streamErr with no reconnect, so request() (:234) fails forever |
-| medium | open | internal/cli/cron_run.go:115 | cron_run.go:102-126 reconcileOverdue still reschedules any active job with NextRunAt before the current minute without firing; no RunNow/RunOnce sentinel exists in internal/cron, and the file is unchanged since pre-audit commit 0f8ea50 |
-| medium | open | internal/sessions/lineage.go:140 | lineage.go:140 tree() still calls ListChildren per node, which calls store.List() (lineage.go:74) re-reading every metadata.json from disk (store.go:312-327, no cache); file unchanged since pre-audit commits |
-| medium | open | internal/cli/exec_sessions.go:114 | exec_sessions.go:117-123 still early-returns once recorder.err is set and stores AppendEvent errors there; recorder.err is never read anywhere (exec.go/exec_spec.go check only writer.err) |
-| medium | open | internal/tui/session_controls.go:117 | m.responseStyle still only displayed (command_views.go:155); /theme //input-style render "does not have a backend setting yet" (rendering.go:47); /compact shows "state: pending integration" (session_controls.go:272); all still in commandDefinitions help/autocomplete (commands.go:189-230) |
-| medium | fixed | internal/tui/session_controls.go:42 | 46d495c/7044d36/6664f71: zeroruntime.CompletionRequest now carries ReasoningEffort, threaded from agent.Options through every request the loop builds and mapped per provider (OpenAI reasoning_effort, Anthropic extended thinking with thinking-block replay, Gemini thinkingBudget). Covered by provider tests |
-| medium | fixed | internal/secrets/scanner.go:81 | [reverified+fixed 2026-06-13, fix/audit-batch] Redact now replaces longest matches first (copies findings, sorts by len(Match) desc) so a containing PRIVATE KEY block is redacted before any nested shorter match corrupts its exact string. Covered by scanner_test.go TestRedactNestedSecretStillRemovesWholeBlock (fails pre-fix: leaks BEGIN/END header) |
-| medium | fixed | internal/redaction/redaction.go:362 | [reverified 2026-06-13] already fixed on main: urlWithCredsPattern is a package-level var (redaction.go:367) and redactURLPasswords (:369) uses it via ReplaceAllStringFunc — no per-call regexp.MustCompile remains |
-| medium | open | internal/zerogit/zerogit.go:212 | zerogit.go:223 `path := strings.TrimSpace(line[3:])` — no C-quote unescaping and no `old -> new` rename split; no -z porcelain used |
-| medium | fixed | internal/worktrees/worktrees.go:247 | [reverified+fixed 2026-06-13, fix/audit-batch] defaultRunGit now captures stdout/stderr into separate bytes.Buffers and returns CommandResult{Stdout, Stderr}; git warnings no longer pollute parsed rev-parse output and error messages get real stderr. Covered by worktrees_test.go TestDefaultRunGitSeparatesStdoutAndStderr |
-| medium | open | internal/sessions/store.go:632 | store.go:632-635 writeMetadata still os.WriteFile temp + os.Rename with no file/dir Sync; zero Sync() calls anywhere in internal/sessions; appendEventLocked (store.go:512-517) also unsynced |
-| medium | open | internal/tui/session_controls.go:235 | session_controls.go:244 returns lastCheckpoint-1 while the tool_call is persisted immediately before its checkpoint (model.go:1505 then 1527), so truncate (rewind.go:205 keeps seq<=target) retains the dangling tool_call |
-| medium | open | internal/sessions/store.go:387 | store.go:387-391 Fork re-appends every parent event (no provider_usage skip) via AppendEvent which stamps a fresh CreatedAt (store.go:495-501); BuildReport has no event-id dedupe |
-| medium | open | internal/usage/report.go:17 | report.go:17-21 usageEventPayload has only token fields and report.go:102 prices from session Metadata.ModelID, while exec.go:510-512 still persists payload["model"] under --allow-escalation — never read |
-| medium | open | internal/cli/exec_sessions.go:115 | exec_sessions.go:117-121 append still early-returns once recorder.err is set, and grep shows sessionRecorder.err is never read in exec.go (no surfacing at run end) |
-| medium | open | internal/tui/session_controls.go:117 | m.responseStyle (session_controls.go:117) is only echoed in /style and /config text (command_views.go:155); grep shows it never reaches agent.Options/system prompt (runAgentWithOptions model.go:1366-1397 sets no style) |
-| medium | open | internal/cron/schedule.go:171 | Probe at HEAD: Next("30 1 * * *", 2026-11-01 01:30 EDT) = 01:30 EST same day (repeated hour); schedule.go:171-201 guards only spring-forward gaps |
-| medium | open | internal/cli/cron.go:112 | cron.go:112-114 sets expr=r.Expr while expr is still ""; cron.go:123 only uses positional[0] if expr=="" — user's expr silently ignored, no warning |
-| medium | open | internal/background/process_posix.go:50 | process_posix.go:29/50 signal only pid (no kill(-pid)); launchBackgroundProcess (internal/specialist/exec.go:642) sets no Setpgid; Windows still tree-kills via taskkill /T (process_windows.go:12) |
-| medium | open | internal/background/manager.go:462 | manager.go:462-471 normalizeLoadedTask rewrites Status==running to error/PID=0/ExitCode=-1; loadTasks persists it to disk at manager.go:419 |
-| medium | open | internal/tui/session_controls.go:117 | session_controls.go:117 still only stores m.responseStyle; sole readers are display strings (command_views.go:155, session_controls.go:120/132) — runAgentWithOptions (model.go:1366+) never threads it into agent.Options or the system prompt, so /style still changes nothing about responses |
-| medium | open | internal/tools/registry.go:224 | registry.go:190-236 Core* sets still omit escalate_model; only exec --allow-escalation registers it (cli/exec.go:157, predates audit), TUI path (cli/app.go:449) never does |
-| medium | open | internal/sandbox/safe_command.go:138 | file unchanged since #109; splitShellSegments (safe_command.go:431) is quote-unaware — probe: `git commit -m "use top \| less"` flagged as 'less', `echo "a; vim b"` flagged as 'vim' |
-| medium | open | internal/providers/gemini/types.go:88 | gemini/types.go:88-93 usageMetadata still has only Prompt/Candidates/Thoughts/Total token counts (no cachedContentTokenCount); grep shows CachedInputTokens is set only in openai/provider.go:253 and anthropic/provider.go:274, never in gemini |
-| medium | open | internal/providers/openai/tool_state.go:95 | tool_state.go:95-100 closeOpen still emits StreamEventToolCallDropped for calls with empty id (and applyDelta:58 never streams them); zeroruntime/helpers.go:141-207 empty-ID/synthetic-key collector support remains unreachable from the adapter |
-| medium | open | internal/specialist/exec.go:150 | exec.go:150 and :195 still hardcode "--auto","high"; cli/exec_tools.go:84-85 maps high -> agent.PermissionModeUnsafe |
-| medium | open | internal/specialist/accounting.go:72 | accounting.go:31/80 check (ReadEvents) then :47/91 append under separate store locks; concurrent callers remain at exec.go:336 (onExit) and output_tool.go:150 (TaskOutput); no serializing mutex in package |
-| medium | open | internal/specialist/streamer.go:43 | streamer.go:43 still caps scanner at 1024*1024; ErrTooLong fails ParseStream -> runChildProcess err (exec.go:624-627) aborts whole run; tool_call Args still emitted untruncated (cli/exec_writer.go:99) |
-| medium | open | internal/agent/loop.go:468 | loop.go:525-528 AlwaysAllow still returns deniedPermissionResult on persistPermissionGrant error; loop.go:778-779 still errors unconditionally when options.Sandbox == nil |
-| medium | partial | internal/agent/compaction.go:58 | [reverified+fixed 2026-06-13, fix/audit-batch] estimateTokens now adds len(message.Images)*imageTokenEstimate (flat per-image cost, ~1k; raw byte size is a poor proxy for vision tiles) — Images are no longer uncounted. Covered by compaction_test.go TestEstimateTokensCountsImages. REMAINING: maybeCompact still has no tool-definition term in its estimate |
-| medium | open | internal/agent/compaction.go:376 | compaction.go:376 summarizeMessagesOnce uses zeroruntime.CollectStream with no OnUsage; comment at compaction.go:314-315 documents OnText/OnUsage are intentionally not forwarded |
-| medium | open | internal/config/resolver.go:95 | ResolvedConfig.MCP set at resolver.go:102 has no readers (grep: zero non-test uses); MCP registration uses ResolveMCP (resolver.go:109-124) which ignores options.ProviderCommand, so ZERO_PROVIDER_COMMAND MCP servers are still dropped |
-| medium | open | internal/hooks/hooks.go:503 | internal/hooks/hooks.go:503 — append still calls store.ReadEvents() (full file read + JSON parse) on every append just to compute the next Sequence |
-| medium | open | internal/cli/app.go:149 | app.go:177-190: comment says trailing non-flag args "still are" ignored; rest scanned only for --add-dir then dropped before runInteractiveTUI |
-| medium | open | internal/cli/exec.go:500 | exec.go:524-532: interrupted branch emits errorEvent/runEnd only for execOutputStreamJSON; json path prints "Interrupted." to stderr, no terminal JSON event |
-| medium | open | internal/cli/exec.go:182 | exec.go:186-193: only stream-json is special-cased; -o json falls to writeExecToolList plain text (exec_tools.go:100-110 "Tools visible to model:") |
-| medium | partial | internal/tui/transcript.go:113 | Full-slice copy removed — in-place append, unkeyed rows skip the scan (transcript.go:92-101) — but hasTranscriptRow still linearly scans all rows per keyed tool/permission/ask append (transcript.go:103-114), so tool-heavy rehydration stays O(n²) |
-| medium | partial | internal/tui/transcript.go:113 | O(n) full-slice copy removed (in-place append, transcript.go:96-100), but hasTranscriptRow (transcript.go:103-114) still linearly scans all rows per keyed (tool/permission) row and m.transcript is never pruned, so tool-heavy rehydration remains O(n^2) |
-| medium | partial | internal/tui/model.go:569 | transcript half fixed: interim text appended as non-final assistant row on tool call/error/cancel (model.go:710-715, 681-683, 1340-1342); but session log still records only FinalAnswer (model.go:1638-1644) — interim segments never persisted, so /resume loses them |
-| medium | partial | internal/tui/transcript.go:117 | re-render gone (View renders only unflushed tail, model.go:769; render_cache.go memoizes) and copy-on-append removed (transcript.go:96-100); but hasTranscriptRow still scans the whole transcript per keyed append (transcript.go:103-114) — O(n²) cumulative for tool rows |
-| medium | partial | internal/tui/transcript.go:113 | Copy fixed: transcript.go:96-100 appends in place; but hasTranscriptRow (transcript.go:103-114) still linear-scans all rows per keyed (tool/permission/ask) row, so tool-heavy /resume rehydration remains O(N²) |
-| medium | partial | internal/tui/transcript.go:117 | transcript.go:100 full-slice copy removed (in-place append; non-keyed rows now O(1) via empty-key fast path), but hasTranscriptRow transcript.go:108-113 still linear-scans all rows per tool/permission/ask append — tool-heavy appends/rehydration remain quadratic |
-| medium | partial | internal/tui/rendering.go:261 | rendering.go:286-303 now preserves explicit newlines and each line's leading indentation (code blocks survive), but rendering.go:305 still runs strings.Fields(body) per line, collapsing intra-line space runs so mid-line column alignment is still lost |

docs/audit/2026-06-13-reverification.md

@@ -1,117 +0,0 @@
-# Audit re-verification against current `main` (2026-06-13)
-
-Every pending (open/partial) row from `2026-06-10-deep-audit-status.md` was re-checked
-against the **actual current code** by an independent agent (the ledger was verified on a
-since-diverged branch). Every `already_fixed` verdict was then adversarially re-checked by a
-second agent trying to prove the defect was still present; **none were refuted** (all 9 hold).
-
-| Verdict | Count |
-|---|---|
-| still_open | 66 |
-| partial | 12 |
-| already_fixed (ledger over-counted) | 9 |
-| **genuinely pending (still_open + partial)** | **78** |
-
-Pending by severity: high 2, medium 42, low 34.
-
-## Already fixed on `main` (reclassify to fixed) (9)
-
-| Sev | Location (current) | Rationale |
-|---|---|---|
-| medium | `internal/cli/cron.go:106-128` | The original defect (recipe set `expr=r.Expr` before the positional was consumed, silently ignoring a user-supplied expr) is gone: cronAdd now consumes the positional expr first at cron.go:113-114 (`if len(positional)==1 && expr=="" { expr=positional[0] }`), and the recipe only fills expr if still empty (cron.go:122-123); extra positionals error. Confirmed by TestCronAddExplicitExprOverridesRecipe (cron_test.go:95-118). |
-| low | `internal/cli/cron_run.go:198-205` | The original gap was that cronTruncate byte-sliced `s[:max]+"…"` and could split a multi-byte rune; current cronTruncate (cron_run.go:204) now returns `cutRuneBoundary(s, max) + "…"`, and cutRuneBoundary (exec_writer.go:420-431) backs up to a utf8.RuneStart boundary, so the rune-split defect is gone. promptExcerpt (cron.go:278) was already rune-safe. |
-| low | `internal/cli/cron_run.go:198-205` | Duplicate of the same underlying issue: the cited byte-slice `return s[:max] + "…"` for run errors (used at cron_run.go:151) no longer exists; cronTruncate now calls cutRuneBoundary(s, max) which lands on a UTF-8 rune boundary, so persisted run-error text can no longer be split mid-rune. |
-| low | `internal/cli/cron_run.go:198-205` | Duplicate of the same underlying issue: cronTruncate no longer byte-slices `s[:max]`; it uses the rune-safe cutRuneBoundary helper (cron_run.go:204), with a comment explaining the cut lands on a UTF-8 rune boundary, closing the run-record invalid-UTF-8 gap. |
-| low | `internal/redaction/redaction.go:217-231 (pointer), :232-252 (map)` | The pointer case now does context.seen[ptr]=struct{}{} (line 225) then delete(context.seen, ptr) after redactReflect returns (line 230); the map case mirrors this (line 240 add, line 251 delete). Comments at lines 226-228 explicitly state the pointer is dropped after recursion so a shared non-cyclic reference reached via a sibling branch is not mistaken for a cycle, exactly the originally-described defect, which is gone. |
-| low | `internal/sandbox/grants.go:392-406` | readState now rebuilds a new `normalized` map keyed by `key := strings.TrimSpace(name)` (grants.go:394, stored at :403) instead of retaining the raw padded file key, and Lookup indexes with `strings.TrimSpace(toolName)` (grants.go:172), so a whitespace-padded file key both validates AND matches. The ledger's premise ('unchanged since #77') is stale — this normalization was added in commit 3f3de1a (#170). |
-| low | `internal/tui/session_controls.go:196-241` | handleCompactCommand no longer just increments compactRequests: when sessionCompactor or hasSessionBackedCompactor() is available it sets compactInFlight and returns tea.Batch(m.runCompact(), ...), and runCompact()->compactActiveSession() actually calls sessionStore.PlanCompaction/RecordCompaction (replay.go:145/181). The 'pending integration' status string is gone (only a negative-assertion test references it), and sessionStore is always non-nil (model.go:272-274), so real compaction is triggered. |
-| low | `internal/zerocommands/contracts.go:169` | ProviderSnapshotFromProfile now sets APIKeySet via `strings.TrimSpace(profile.APIKey) != "" \|\| strings.TrimSpace(profile.AuthHeaderValue) != ""` (line 169) with a comment (lines 166-168) explaining auth-header-only profiles must not render as 'not set'; config/types.go:51 confirms AuthHeaderValue is a valid sole credential, and command_center.go:317/334 render apiKeyState(provider.APIKeySet) off this fixed boolean. |
-| low | `internal/zerogit/zerogit.go:209` | The remaining gap in the partial finding (ValidateMessage still byte-based `len(firstLine) > 72`) is gone: line 209 now uses `utf8.RuneCountInString(firstLine) > 72` with an explanatory comment, and truncateSubject (line 466) is already rune-based, so valid non-ASCII subjects are no longer rejected. |
-
-## Genuinely partial (a real gap remains) (12)
-
-| Sev | Location (current) | Rationale |
-|---|---|---|
-| high | `internal/cli/cron_run.go:176-188 (re-read fix) + internal/cron/store.go:170-178 (Update, no lock)` | fireJob does re-read via store.Get and preserves an external pause (current.Status == StatusPaused at cron_run.go:186-187) and handles ErrJobNotFound/transient errors, closing the pause-clobber; but cron/store.go Update still just calls writeJob (temp+rename) with no flock/CAS anywhere in the package, so the concurrent-scheduler read-modify-write race remains exactly as the ledger noted. |
-| high | `internal/hooks/dispatch.go:112 (Dispatch); wired at internal/agent/loop.go:584,620 and constructed at internal/cli/hook_dispatch.go:51` | The primary gap is closed: the Dispatcher now has production callers and IS wired into the tool lifecycle - constructed in app.go/exec.go/exec_spec.go, assigned to agent Options.Hooks, and invoked at loop.go:584 (beforeTool, with veto via blockedByHookResult) and loop.go:620 (afterTool feedback). But the original finding's CLI-surface gap remains: runHooks in internal/cli/extensions.go:88-135 still supports only `list`/help with no dispatch/edit subcommand. |
-| medium | `internal/agent/compaction.go:280 (maybeCompact) / 83-96 (estimateTokens)` | The image-counting fix is present (estimateTokens line 92 adds len(message.Images)*imageTokenEstimate), but the remaining gap stands: maybeCompact's trigger uses estimateTokens(messages), which has no tool-definition term, and is called at loop.go:85 with no tools (the exposed tool defs are built afterward at loop.go:87). The tool-definition estimator estimateToolTokens exists only in the separate MeasureContext reporting path (context_measurement.go:53), not in the compaction trigger. |
-| medium | `internal/tui/rendering.go:292-336 (gap at :311)` | wrapPlainText now preserves explicit newlines (split on \n, :297) and each line's leading indentation (TrimLeft + reconstructed indent, :302-305), so code blocks survive. But the word loop at :311 still iterates strings.Fields(body), which collapses intra-line space runs, so mid-line column alignment within a line is still lost exactly as the finding describes. |
-| medium | `internal/tui/transcript.go:103-114` | The full-slice copy is genuinely gone (in-place append at transcript.go:100 with a comment about the old O(n²)), but hasTranscriptRow (transcript.go:108-112) still linear-scans every existing row computing transcriptRowKey per keyed tool/permission/ask append, and m.transcript is never pruned, so rehydration via appendTranscriptRow (session.go:145-146) remains O(n²). |
-| medium | `internal/tui/transcript.go:103-114` | Duplicate of the same finding: copy-on-append removed (transcript.go:96-100) but hasTranscriptRow (transcript.go:108-112) still linearly scans all rows per keyed row and the transcript is never bounded, so tool-heavy rehydration stays O(n²). |
-| medium | `internal/tui/transcript.go:103-114` | Duplicate of the transcript.go:113 finding: append is in-place (transcript.go:100), but hasTranscriptRow (transcript.go:108-112) still linear-scans all rows per keyed (tool/permission/ask) row, so tool-heavy /resume rehydration remains O(n²). |
-| medium | `internal/tui/transcript.go:103-114` | View re-render and copy-on-append are fixed (in-place append at transcript.go:100, empty-key fast path at transcript.go:105-107), but hasTranscriptRow (transcript.go:108-112) still scans the whole transcript per keyed append — O(n²) cumulative for tool rows during rehydration. |
-| medium | `internal/tui/transcript.go:103-114` | Duplicate of the transcript.go:117 finding: the full-slice copy is removed and non-keyed rows are O(1) via the empty-key fast path (transcript.go:105-107), but hasTranscriptRow (transcript.go:108-112) still linear-scans all rows per tool/permission/ask append, so those appends/rehydration remain quadratic. |
-| low | `internal/agent/loop.go:553-580, 611-616` | The main "requestEvent never read after the switch" claim is now fixed (requestEvent is passed to emitDeniedPermission/deniedPermissionResult on the denial paths at 571-572,577-578), but requestEvent.GrantMatched/Grant set at 574-575 in the AlwaysAllow success case remain dead writes (the post-run event at 611-616 is rebuilt fresh) and the fallbackPermissionEvent at 555 is still unreachable because shouldRequestPermission only fires for PermissionPrompt tools, for which buildPermissionEvent never returns ok=false (false only at 1071/1076 for non-Prompt/Allow). |
-| low | `internal/tui/commands.go:297,306; internal/tui/command_center.go:43-55; internal/agent/loop.go:105-106; internal/tui/commands.go:254` | The main fix landed: footer builders, theme tokens (selRow/statusOk/statusErr/line2/panel2), and actionAppendToolCall/Result are all gone (zero grep hits). But the cited gaps remain: listCommandNames (commands.go:297) and formatCommandHelpLines (commands.go:306) are referenced only by test files; resumeText's args!="" branch (command_center.go:45-55) is still unreachable since its sole non-test caller session.go:122 passes ""; OnContext (agent/loop.go:105-106) is emitted but has no production assignment; and the stale footer comment persists at commands.go:254. |
-| low | `internal/tui/commands.go:297,306; internal/tui/command_center.go:43-55; internal/agent/loop.go:105-106; internal/tui/commands.go:254` | Duplicate of the other rendering.go:62 row; same underlying code. Footer builders and the theme tokens are fully removed, but formatCommandHelpLines/listCommandNames (commands.go:297-308) remain production-defined yet referenced only by tests, and a stale 'footer advertises' comment persists at commands.go:254 (the autocomplete.go:64 stale comment cited in the older row is now gone, but commands.go:254 still has one). |
-
-## Genuinely still open (66)
-
-| Sev | Location (current) | Rationale |
-|---|---|---|
-| medium | `internal/agent/compaction.go:422 (summarizeMessagesOnce)` | summarizeMessagesOnce still calls plain zeroruntime.CollectStream (compaction.go:422), which is CollectStreamWithOptions with empty CollectOptions{} (helpers.go:78), so no OnUsage callback fires; the collected Usage is discarded and the deliberate non-forwarding is now documented at summarizeClosure (compaction.go:359-361). Compaction-call token usage is still not surfaced/tracked. |
-| medium | `internal/agent/loop.go:567-573, 928-931` | The AlwaysAllow case still calls deniedPermissionResult/emitDeniedPermission when persistPermissionGrant errors (loop.go:569-573), and persistPermissionGrant (now at 928-931) still errors unconditionally when options.Sandbox == nil, so choosing AlwaysAllow without a configured sandbox denies the already-approved call; both sub-claims persist (only relocated from the stale 778-779 reference). |
-| medium | `internal/background/manager.go:462-471 (and persistTaskLocked call at :419)` | normalizeLoadedTask still unconditionally rewrites task.Status==StatusRunning to StatusError, sets PID=0 and ExitCode=-1, and marks changed=true; loadTasks then persists this to disk via persistTaskLocked at line 419, with no process-liveness check or reconciliation. The defect is present essentially as the auditor described. |
-| medium | `internal/background/process_posix.go:29,50` | SIGTERM (line 29) and Kill() (line 50) still target only the single pid with no kill(-pid)/negative-pid group signal, and launchBackgroundProcess (specialist/exec.go:650) sets no SysProcAttr{Setpgid:true} (grep finds zero process-group setup), while Windows still tree-kills via taskkill /T /F (process_windows.go:12). All three claims hold. |
-| medium | `internal/cli/app.go:200-220` | In the `--skip-permissions-unsafe` case, the code re-splits args[1:] via splitLeadingAddDirFlags and only iterates `rest` to reject a misplaced --add-dir (lines 215-219); it then calls runInteractiveTUI with merged addDirs only, never forwarding `rest`, so any trailing non-flag args are still silently dropped. The in-code comment at lines 208-209 explicitly confirms this ('were ignored on this path... and still are'), matching the original finding verbatim. |
-| medium | `internal/cli/cron_run.go:103-128` | reconcileOverdue still reschedules any strictly-overdue active job (NextRunAt before nowMin) to sched.Next without firing it (lines 116-126); grep confirms no RunNow/RunOnce/FireNow sentinel exists in internal/cron, so the only way to fire backlog is the --catch-up flag that skips reconcileOverdue entirely. |
-| medium | `internal/cli/exec.go:191-199` | The --list-tools branch special-cases only execOutputStreamJSON (exec.go:192-194); -o json (execOutputJSON) falls through to writeExecToolList, which calls formatExecToolList and prints the plain-text "Tools visible to model:" listing (exec_tools.go:100-110) rather than any JSON output. |
-| medium | `internal/cli/exec.go:531-545` | The interrupted branch still emits writer.errorEvent/runEnd only when outputFormat == execOutputStreamJSON (exec.go:535-540); for -o json it just does fmt.Fprintln(stderr, "Interrupted.") at exec.go:542 and returns exitInterrupted, producing no terminal JSON event on stdout (unlike the provider-error path which writes type:error/type:done for json). |
-| medium | `internal/cli/exec_sessions.go:116-124` | execSessionRecorder.append still early-returns when recorder.err != nil (line 117) and latches AppendEvent failures into recorder.err (line 120); a package-wide grep shows recorder.err/sessionRecorder.err is read only inside append itself, never by exec.go or exec_spec.go (which check only writer.err), so persistence errors are silently swallowed at run end. |
-| medium | `internal/cli/exec_sessions.go:116-124` | Duplicate of the same recorder.err defect: append (lines 116-124) early-returns once recorder.err is set and stores AppendEvent errors there, but sessionRecorder.err is never read in exec.go or exec_spec.go, so a failed session append is never surfaced to the user at run end. |
-| medium | `internal/config/resolver.go:102, 110-124` | ResolvedConfig.MCP is still set at resolver.go:102 but has zero readers: every `.MCP` access in internal/cli is the config.MCPConfig type or the separate resolveMCPConfig/ResolveMCP path, and no resolveConfig caller reads resolved.MCP. MCP registration runs through ResolveMCP (resolver.go:110-124), which only merges config-file paths and Overrides and never calls LoadProviderCommand/touches options.ProviderCommand (asserted intentional by TestResolveMCPDoesNotRunProviderCommand), so ZERO_PROVIDER_COMMAND MCP servers are still dropped. |
-| medium | `internal/cron/schedule.go:171-201` | Empirically confirmed: on the 2026-11-01 US fall-back day Next("30 1 * * *") returns 01:30 EDT and then 01:30 EST again, firing the daily job twice during the repeated wall-clock hour. The Next loop and its forward-progress guard (schedule.go:182-198) only handle spring-forward non-existent instants and contain no de-duplication for the ambiguous repeated fall-back hour. |
-| medium | `internal/hooks/hooks.go:503` | AuditStore.append still calls store.ReadEvents() at line 503, which os.ReadFile's the entire JSONL audit file and json.Unmarshal's every line on every single append, solely to scan for the highest existing Sequence (lines 507-513); no cached/in-memory sequence counter was introduced, so the O(n) full re-read per append is unchanged. |
-| medium | `internal/mcp/client.go:246` | client.writer.write(...) at line 246 still executes while client.mu is held (locked at :228, unlocked at :255), and writer.write (protocol.go:148-163) does a blocking Write+Flush on the stdin pipe with no ctx awareness; only the response wait at :257 is ctx-aware, unchanged from the audit. |
-| medium | `internal/mcp/client.go:295-318` | readLoop still dispatches any id-bearing message to pending[id] (lines 302-316) with no check of message.Method; grep finds no 'ping' or Method handling anywhere in client.go, so server-initiated requests are still ignored/misrouted exactly as described. |
-| medium | `internal/mcp/client.go:98-99` | Still 'var stderr bytes.Buffer; cmd.Stderr = &stderr' — a plain unbounded buffer (no LimitReader/cap), never reset after startup, attached for the whole process lifetime and only read on initialize failure at line 114. |
-| medium | `internal/mcp/network_client.go:612-642 (decodeSSERPCMessage), caller at :219` | decodeSSERPCMessage still decodes the first non-empty 'message' SSE event, sets found=true and returns false (stops scanning) without checking the message ID; the synchronous caller networkClient.request at line 219 does `if !rpcIDMatches(message.ID, id)` and errors on mismatch instead of scanning for the matching response. Defect unchanged (caller line is now :219, was the stale :210). |
-| medium | `internal/mcp/network_client.go:646 (scanSSEEvents), failPending at :587-593, request gate at :243` | scanSSEEvents still calls scanner.Buffer(make([]byte, 0, 64*1024), 1024*1024) with the 1MiB cap; a bufio.Scanner ErrTooLong flows through readStream into failPending (:593) which sets a permanent client.streamErr, and request() returns that streamErr forever (:243-246) with no reconnect path (openStream is only called once at connect, :77). |
-| medium | `internal/providers/gemini/types.go:103-108` | The usageMetadata struct still declares only PromptTokenCount, CandidatesTokenCount, ThoughtsTokenCount, and TotalTokenCount with no cachedContentTokenCount field; provider.go:225-228 reads only those into input/output/reasoning tokens and emitDone (provider.go:282-286) builds TokenUsage without CachedInputTokens, which is set only in openai/provider.go:253 and anthropic/provider.go:330. |
-| medium | `internal/providers/openai/tool_state.go:95-101 (and applyDelta:58)` | applyDelta still returns early when call.id=="" (line 58), and closeOpen still only emits StreamEventToolCallDropped for empty-id/empty-name calls (lines 95-99) instead of streaming them with a synthetic key; every Start/Delta/End the adapter emits carries call.id, so the zeroruntime collector's empty-ID/synthetic-key support (helpers.go start/delta/end at ~173-216) is never reached from the OpenAI adapter. The provider_test still asserts such calls are dropped and never started, confirming the behavior is unchanged, and tool_state.go's last commit (4d2ce6b) predates the 2026-06-10 audit. |
-| medium | `internal/sandbox/safe_command.go:431` | splitShellSegments still uses a strings.NewReplacer that blindly turns \|, ;, &&, \|\|, $(, ), backtick into segment boundaries with zero quote awareness; a live probe confirms `git commit -m "use top \| less"` is flagged as 'less' (interactive=true) and `echo "a; vim b"` is flagged as 'vim', exactly as the original finding described. The file is unchanged since commit #109 and contains no quote-state parsing. |
-| medium | `internal/sessions/lineage.go:140` | tree() (lineage.go:128-154) still calls store.ListChildren(sessionID) at lineage.go:140 once per node, and ListChildren calls store.List() at lineage.go:74; List() (store.go:312-338) re-reads the whole sessions dir via os.ReadDir+Get->readMetadata (store.go:674-684, a fresh os.ReadFile+json.Unmarshal with no cache), so an N-node tree re-reads every metadata.json N times. The defect is present exactly as described. |
-| medium | `internal/sessions/store.go:428-432 (Fork copy loop), 526-567 (appendEventLocked); internal/usage/report.go:74-119 (BuildReport)` | Fork's loop at store.go:428-431 re-appends every parent event (no provider_usage/EventUsage skip) via AppendEvent, which stamps a fresh CreatedAt/ID in appendEventLocked (537-543), and BuildReport (report.go:74-119) aggregates all EventUsage events with no event-id dedupe, so forked usage events are double-counted. |
-| medium | `internal/sessions/store.go:686-701 (writeMetadata), 549-559 (appendEventLocked write path)` | writeMetadata still does os.WriteFile to a temp file + os.Rename (store.go:693-696) with no file or directory fsync, and a repo-wide grep finds zero Sync() calls anywhere in internal/sessions/; appendEventLocked's append write (549-557) is likewise unsynced. |
-| medium | `internal/specialist/accounting.go:80-91 (appendSpecialistUsageRollup) and :31-47 (recordSpecialistStop)` | The dedup is still a non-atomic check-then-append: specialistEventExists calls store.ReadEvents (which acquires NO lock, store.go:569) and only the later AppendEvent takes the per-session lock for its own write, so concurrent callers (exec.go:343 onExit and output_tool.go:150 TaskOutput) can both pass the existence check and each append a duplicate event. No serializing mutex spans the check and append in the specialist package, and the store's lockSession/appendEventLocked machinery is not used by the accounting path. |
-| medium | `internal/specialist/exec.go:150 (and :202)` | Both BuildArgs (:150) and BuildResumeArgs (:202) still hardcode args append "--auto","high", and internal/cli/exec_tools.go:84-85 maps "high" -> agent.PermissionModeUnsafe, so specialist children always run in unsafe permission mode. |
-| medium | `internal/specialist/streamer.go:43` | ParseStream still calls scanner.Buffer(make([]byte,0,64*1024), 1024*1024), so a line over 1MB yields bufio.ErrTooLong; exec.go:631-634 returns that error from runChildProcess aborting the whole run, and the producer at cli/exec_writer.go:99 (parseToolCallArgs) still emits tool_call Args with no length cap. All three elements of the original finding are unchanged. |
-| medium | `internal/tools/registry.go:198-245` | The Core* tool sets (CoreReadOnlyTools/CoreWriteTools/CoreShellTools at :224/CoreNetworkTools/CoreTools) still omit escalate_model; it is registered only under exec --allow-escalation (cli/exec.go:156-157), while the TUI registry built by newCoreRegistryScoped (cli/app.go:436,531-537) never registers it — app.go only mentions escalate_model in help text at :733. |
-| medium | `internal/tui/model.go:2022-2028 (session log) vs model.go:854/802/1721 (transcript)` | Interim narration is appended to the live transcript on tool call (model.go:852-857), error (801-803), and cancel (1720-1722), but the session log only persists result.FinalAnswer via EventMessage (model.go:2022-2028); onText (1781-1786) and sendAgentText (2064-2069) record no session event, so the gap the ledger flagged as 'partial' is unchanged — interim segments are still never persisted and /resume loses them. |
-| medium | `internal/tui/session_controls.go:148-177; internal/tui/model.go:1746-1778` | m.responseStyle is still only stored by handleStyleCommand (session_controls.go:156) and echoed in display strings (session_controls.go:159/171, command_views.go:275); runAgentWithOptions threads ReasoningEffort/SystemPrompt/PermissionMode into options but never references responseStyle, and agent.Options has no style field (no non-display ResponseStyle reader exists), so /style still changes nothing about responses. |
-| medium | `internal/tui/session_controls.go:148-177; internal/tui/model.go:1746-1778` | Duplicate of the responseStyle finding: m.responseStyle is set at session_controls.go:156 but runAgentWithOptions (model.go:1746-1778) never assigns it into agent.Options or the system prompt; the only readers are display strings (command_views.go:275, session_controls.go:159/171), so the style preference is inert. |
-| medium | `internal/tui/session_controls.go:148-177; internal/tui/model.go:1746-1778` | Duplicate of the responseStyle finding verified once: handleStyleCommand only stores m.responseStyle and grep shows no non-display consumer; runAgentWithOptions sets ReasoningEffort/SystemPrompt/etc. on agent.Options but never the style, so /style remains cosmetic. |
-| medium | `internal/tui/session_controls.go:318 (resolveRewindTarget); internal/sessions/rewind.go:205` | resolveRewindTarget still returns lastCheckpoint-1, while OnToolCall appends EventToolCall then the EventSessionCheckpoint right after it (model.go:1886-1911) so the tool_call gets the immediately-preceding sequence; TruncateEvents keeps ev.Sequence<=target (rewind.go:205), so the tool_call (seq==target) is retained while its checkpoint (seq>target) is dropped, leaving the dangling tool_call described in the finding. |
-| medium | `internal/usage/report.go:17-21 (struct) and :102 (pricing); writer at internal/cli/exec.go:521-524` | usageEventPayload (report.go:17-21) still has only PromptTokens/CompletionTokens/TotalTokens with no model field, and BuildReport at report.go:102 still prices via modelBySession[event.SessionID] from session Metadata.ModelID; meanwhile exec.go:521-522 still persists payload["model"]=currentModel under --allow-escalation, and a repo-wide grep confirms no consumer reads that per-event model key (the only usageEventPayload reader silently drops it on unmarshal). |
-| medium | `internal/zerogit/zerogit.go:225` | parseStatus still does `path := strings.TrimSpace(line[3:])` with no C-quote unescaping and no `old -> new` rename split, and the caller at line 119 runs `status --short --untracked-files=all` without `-z`/porcelain, so quoted or renamed paths are still mis-parsed exactly as described. |
-| low | `internal/agent/loop.go:1305-1316` | propertyToRuntimeMap still assigns schema["items"] = propertyToRuntimeMap(*property.Items) twice for the same property.Items != nil guard (lines 1305-1307 and 1314-1316); the second assignment is a redundant duplicate writing an identical value. |
-| low | `internal/agent/loop.go:189-191` | The reactive-retry CollectStreamWithOptions call still passes only OnUsage and omits OnText (comment at 183-188 documents this as deliberate anti-duplication), so streaming surfaces still receive only the aborted partial text on the retried turn exactly as the finding describes. |
-| low | `internal/agent/loop.go:194-201` | When ctx is canceled, internal/zeroruntime/helpers.go:89 sets collected.Error = ctx.Err().Error(); loop.go:194-197 returns errors.New(collected.Error) before the ctx.Err() identity return at 198-201, so the returned error no longer wraps the sentinel and errors.Is(err, context.Canceled) still fails on that path. |
-| low | `internal/agent/types.go:196 (field); internal/agent/loop.go:105-106 (emit site); internal/agent/context_measurement.go:31 (MeasureContext/ContextBreakdown)` | The OnContext callback field (types.go:196) and its emit site (loop.go:105-106 calling MeasureContext) still exist, but a repo-wide grep finds ZERO 'OnContext:' assignments anywhere (test or prod) and no use of MeasureContext/ContextBreakdown outside internal/agent tests; cli/tui/cmd and the contextreport package never reference them (contextreport only calls agent.BuildSystemPromptPreview), so the callback remains permanently a no-op exactly as the original finding described. |
-| low | `internal/background/process_posix.go:38-50` | terminateProcess still does os.FindProcess(pid) (line 24) and polls liveness via signal-0 in processAlive (lines 39-47, 59-61), while launchBackgroundProcess (specialist/exec.go:659-673) still reaps the child in a concurrent command.Wait() goroutine, so the PID can be reaped and recycled before process.Kill() at line 50 fires. Defect unchanged. |
-| low | `internal/cli/cron_run.go:146-152` | rec.Error is still set only from errBuf (cronTruncate(strings.TrimSpace(errBuf.String()),500) when code!=0); outBuf is declared and passed to exec at line 146-147 but never read anywhere in fireJob, so stream-json error events on stdout are still discarded. |
-| low | `internal/cli/exec.go:383-398 (PrepareExec call), exec.go:404 (time.Now), internal/cli/exec_sessions.go:55 (direct NewStore)` | All three sub-claims hold: the sessions.PrepareExec call at exec.go:383-398 omits the Store field so PrepareExec falls back to NewStore(StoreOptions{}) at exec_session.go:58-61; preflightExecSession still does sessions.NewStore(StoreOptions{}) directly at exec_sessions.go:55; and runExec uses streamjson.CreateRunID(time.Now()) at exec.go:404 instead of the injectable deps.now (which is available and used at exec.go:142/186). |
-| low | `internal/cli/exec_sessions.go:55` | preflightExecSession still calls sessions.NewStore(sessions.StoreOptions{}) directly at line 55, and exec.go's PrepareExec call (exec.go:383-397) omits the Store field so sessions/exec_session.go:58-60 defaults to NewStore again; deps.newSessionStore is used by spec/usage/observability/sessions paths but never by the exec run path in exec.go. |
-| low | `internal/config/contracts.go:5,12,65` | ContractGap (line 5), DefaultContractGaps (line 12), and FindContractGapsByMilestone (line 65) are all still defined, and a repo-wide grep for these symbols (plus ContractOwnerRuntime) finds references ONLY in internal/config/contracts_test.go. No production caller exists, so the dead-code defect persists exactly as described. |
-| low | `internal/config/resolver.go:144,177,506` | All three MaxTurns sites (mergeConfig :144, mergeProjectConfig :177, applyOverrides :506) still gate purely on `if src/overrides.MaxTurns > 0` with no error path, and Resolve() has no MaxTurns validation, so a configured 0 or negative maxTurns is silently dropped and falls back to defaultMaxTurns=30 (resolver.go:16,22). A negative-value error path exists for tools.deferThreshold (:57-59) but none was added for MaxTurns. |
-| low | `internal/plugins/plugins.go:894` | parseHookEvent's switch (plugins.go:894) still accepts only HookBeforeTool/HookAfterTool/HookSessionStart/HookSessionEnd and the HookEvent consts (plugins.go:39-42) omit specialist events entirely, whereas hooks.go parseEvent (hooks.go:812) accepts those four plus EventSpecialistStart/EventSpecialistStop, so the two validation sets still disagree exactly as described. |
-| low | `internal/providercatalog/catalog.go:46` | The `Public bool` field is still declared on Descriptor (catalog.go:46) but is never assigned anywhere in the codebase: there are zero `Public:` struct-literal assignments and zero `.Public =` writes. The only read remains the vacuous guard at catalog_test.go:106 (`if descriptor.Public && !descriptor.RequiresAuth`), which is dead code since Public is always false. |
-| low | `internal/providers/factory.go:115` | At the sole call site (line 115) the `\|\|` short-circuits so isImplicitOpenAI is only evaluated when explicitProvider==true, which (per explicitProviderKind at 147-157) requires ProviderKind!="" or Provider!=""; but isImplicitOpenAI (159-164) requires both ProviderKind=="" and Provider=="", so it can never return true there — the term remains dead code exactly as described. |
-| low | `internal/redaction/redaction.go:174` | Line 174 still declares `var stackTracer interface{ StackTrace() fmt.Stringer }`; a repo-wide grep finds StackTrace only at redaction.go:174 and :176 (no implementer in the module), and go.mod has no pkg/errors-style dependency, so this errors.As branch is dead code that can never match exactly as the finding described. |
-| low | `internal/sessions/store.go:615-617 (timestamp), 331-336 (List), 340-346 (Latest)` | timestamp() at store.go:616 still formats with time.RFC3339 (second precision), and List() at 332-334 breaks UpdatedAt ties lexically by SessionID, so Latest() (line 345 returns sessions[0]) can still pick the wrong session on same-second ties exactly as described. |
-| low | `internal/skills/skills.go:91 (Duplicates); consumers: internal/cli/skills.go:62, internal/plugins/activate.go:376` | skills.Duplicates and skills.Get still have zero callers outside internal/skills (Get has zero callers anywhere, even tests), and the duplicate-collision data is never surfaced: cli/skills.go runSkillsList calls skills.List with no warning, and the skill tool discards the dups at activate.go:376 with `merged, _ := MergedSkillsLoaded(...)`. Although plugins.mergeSkills now computes a parallel DuplicateName slice that MergedSkills returns, every call site drops it, so shadowed/dropped skills are still resolved silently with no user-facing warning, exactly the defect described. |
-| low | `internal/specialist/exec.go:353-358` | On manager.SetPID failure the code only calls executor.cleanupBackgroundPromptFile and returns the error; it does not call recordSpecialistStop or manager.UpdateStatus(StatusError) while the already-launched child keeps running, unlike the launch-error path at :347-352 which does record a stop. |
-| low | `internal/specialist/output_tool.go:244` | formatTaskOutput is still dead code: a codebase-wide grep finds only its own definition at :244 and no callers in prod or tests; the only used formatter is formatTaskOutputSummary, called from readOutput at :154. |
-| low | `internal/streamjson/streamjson.go:30` | Repo-wide grep for EventRestore (and the literal "restore" EventType) returns only the declaration at streamjson.go:30 — no emitter and no test reference, whereas the sibling EventCheckpoint is emitted at cli/exec_writer.go:119 and referenced in streamjson_test.go:435. The constant remains dead exactly as the original finding described. |
-| low | `internal/tools/read_file.go:98-121` | The truncation path is unchanged from main (empty git diff): when max_lines truncates, it sets Result.Truncated=true (line 120) but the only in-output signal is the '(lines 1-5 of 20)' range header (line 114) — no distinct truncation marker is appended to Output (line 119). The rune-width sub-claim remains non-reproducible because line-number padding uses ASCII digits via strconv.Itoa (lines 105-109), exactly as the original finding concluded. |
-| low | `internal/tools/registry.go:19` | Repo-wide grep finds OnSandboxDecision only at registry.go:19 (declaration), :113 (nil check), and :121 (invocation) — zero struct-literal (OnSandboxDecision:) or dot-assignment (.OnSandboxDecision =) sites anywhere in non-test or test code, so the field is still never set by any caller. |
-| low | `internal/tui/command_center.go:45-55` | resumeText's args!="" branch (command_center.go:45-55) is still dead: the only production caller, handleResumeCommand at session.go:121-122, passes "" (the non-empty case is handled separately via resolveResumeSession at session.go:125), so the branch remains unreachable; one stale 'footer advertises' comment also persists at commands.go:254 (the autocomplete.go one is gone). |
-| low | `internal/tui/command_center.go:17-24` | doctorText() still constructs doctor.Run(doctor.Options{Now, Runtime:"go", Provider}) with no UserConfig/ProjectConfig (command_center.go:18-22), unlike the CLI path at observability.go:64-71; so configFilesCheck (doctor.go:111-123) gets empty paths and emits the StatusWarn 'No explicit Zero config files were inspected.' exactly as the finding describes. |
-| low | `internal/tui/commands_test.go:46` | The function commandTestStringSliceContains is still defined at line 46, and a repo-wide grep across all *.go files returns only the definition itself with zero call sites, matching the original dead-helper finding exactly. |
-| low | `internal/tui/model.go:1363-1365` | handleSubmit still drops a prompt while m.exiting with a bare `return m, nil` and no feedback row, and no render path surfaces an exiting/flushRunIDs indicator: statusLine (view.go:137-156) shows only provider+usage, and composerLine's hint switch (model.go:1130-1134) only shows 'esc stop' when m.pending. |
-| low | `internal/tui/options.go:34,39-40 (UsageTracker/ReasoningEffort/ResponseStyle); internal/cli/app.go:484-516; internal/tui/model.go:277-279` | The exact residual gap the ledger evidence describes is still present verbatim: the production caller's tui.Options{...} block in app.go (484-516) sets none of UsageTracker, ReasoningEffort, or ResponseStyle. The mitigations cited also still hold — newModel self-defaults a usage.Tracker when options.UsageTracker is nil (model.go:277-279) and effort/style remain runtime-settable via /effort (session_controls.go:81) and /style (session_controls.go:156) — so the code matches the partial state described, with the same unset-fields gap unchanged. |
-| low | `internal/tui/session_test.go:215` | The local slice `runtimeMessages := []tea.Msg{}` (line 215) is still only appended to in the RuntimeMessageSink closure (line 226) and never read or asserted anywhere in the test body; the test exclusively consumes the channel runtimeMessageCh via receiveRuntimeMessage, leaving the slice write-only exactly as the finding described. |
-| low | `internal/verify/verify.go:215` | RunLoop is still defined at verify.go:215 and a repo-wide grep finds only two references: the definition and its sole caller verify_test.go:207. The production retry loop is selfverify.Run (selfverify.go:53), which re-implements the attempt loop calling verify.Run directly (selfverify.go:75) and is the one wired into cli/workflows.go, so RunLoop plus LoopOptions/LoopReport remain dead production code. |
-| low | `internal/zerocommands/sandbox_snapshots.go:82-104 (SandboxPlanSnapshot/SandboxDecisionSnapshot); raw serialization at internal/cli/sandbox.go:92-94,146-148` | Grep confirms SandboxPlanSnapshot/Decision/Policy/Backend/Risk/Violation snapshots are referenced only inside internal/zerocommands and its tests (no production consumer; the only new consumers in tui/command_views.go are SandboxGrantSnapshots and ProviderSnapshot, which are different types). `zero sandbox policy --json` in internal/cli/sandbox.go still builds its payload from raw zeroSandbox.Policy/Backend/BackendPlan (lines 92-94 and 146-148) instead of the snapshot constructors, exactly as the original evidence described. |
-| low | `internal/zerogit/zerogit.go:419-423` | The fallback EnvRunner adapter at lines 420-422 is still `func(ctx, dir, _ []string, args ...string)` which discards the env slice and calls `runGit(ctx, dir, args...)`, so a caller-supplied Runner with no RunGitEnv still loses stagedSnapshotDiff's GIT_INDEX_FILE isolation. |
-| low | `internal/agent/loop.go:194-201 (origin internal/zeroruntime/helpers.go:89)` | helpers.go:89 still sets collected.Error = ctx.Err().Error() (stringifying the cancellation, losing identity), and loop.go:194-196 returns errors.New(collected.Error) BEFORE the identity-preserving `return result, ctx.Err()` at loop.go:198-200, so on the stream-cancellation path the returned error is a plain string error and errors.Is(err, context.Canceled) still fails. |

docs/audit/2026-06-20-deep-audit.md

@@ -1,649 +0,0 @@
-# Zero — Deep Adversarial Code Audit (2026-06-20)
-
-- **Target:** `github.com/Gitlawb/zero` @ `origin/main` commit `bfbdbb1` (a terminal coding agent: Go, Bubble Tea v2 TUI, surface-agnostic core feeding TUI / headless `exec` / MCP server / cron).
-- **Toolchain:** Go 1.25.0 (toolchain go1.26.4). 72 packages.
-- **Method:** Fresh checkout of `origin/main` (the audit-batch worktree lags main). Recon → build/vet/`-race` gates → 13-subsystem deep read (each reconciling its slice of the prior audit docs) → adversarial re-verification of every finding by independent skeptics (2 for high, 1 otherwise) that re-read guards/locks/callers to try to *disprove* it; only survivors are reported. Findings are grounded in current code with `path:line` + quoted evidence; several were empirically reproduced with throwaway tests.
-- **Scope note:** Audit only — no source was modified.
-
-## 1. Executive summary
-
-The tree is in markedly better shape than the 2026-06-10 audit: of ~150 prior findings, **64 are fully fixed and 14 partially fixed** in current `main` (atomic config write, durable session metadata, cross-process cron/hooks/oauth locks, MCP framing + alloc cap, SSRF DNS-rebind pinning, sandbox quote-aware interactive detection, FinishReason consumption, transcript O(n²) + per-frame re-render, UTF-8 rune-safe truncation across the board, secret-pattern anchoring, and more). Build and vet are clean; the race detector found **zero data races** across 67/68 packages.
-
-The residue clusters into four recurring root causes (§5). The single most important new item is **one security finding (M10):** the provider health-probe forwards `x-api-key`/`x-goog-api-key`/custom headers across an HTTP redirect to a public host — a credential-exfil vector the recently-added DNS-rebind/internal-redirect hardening does not cover. The lone high-severity item (**H1**) is a trust/correctness gap: four of six documented, user-configurable hook events are never dispatched. The remaining mediums are durability (events.jsonl not fsync'd; PTY last-chunk drop), cron edge cases (DST double-fire, claim double-grant, `--run-now` cancelled), connection-lifecycle gaps (remote-bridge idle deadline, Shutdown can't close bridge conns, MCP SSE never reconnects), a sandbox over-block false-positive, a stream-json prose-mangling regex, and a TUI streaming O(n²).
-
-### Severity rollup
-
-| Severity | Count |
-|---|---|
-| Critical | 0 |
-| High | 1 |
-| Medium | 15 |
-| Low | 17 |
-| Info | 2 |
-| **Total (surviving)** | **35** |
-
-Prior-audit reconciliation: **64 fixed · 14 partial · 12 still-open** (§4). Two drafted findings were dropped in adversarial review (§4.4).
-
-## 2. Build / vet / test results (verbatim)
-
-Run in the fresh `origin/main` checkout, Go-native only (per AGENTS.md), with test isolation `HOME`/`XDG_CONFIG_HOME` pointed at temp dirs and `CI=1`.
-
-```
-$ go build ./...
-(no output) — exit 0  ✅ clean
-
-$ go vet ./...
-(no output) — exit 0  ✅ clean
-
-$ go test ./... -race -count=1
-ok   <67 packages>            ✅
---- FAIL: TestIndependentExecCommandConstructorsShareDefaultManager (1.01s)
-    exec_command_test.go:43: expected shared manager to find completed session, got ... session_id:"1000" ...
---- FAIL: TestExecCommandReturnsSessionAndWriteStdinPollsCompletion (1.02s)
-    exec_command_test.go:82: expected exit_code 0, got ... session_id:"1000" ...
---- FAIL: TestExecCommandReturnsExitCodeWhenCommandCompletesDuringInitialYield (1.01s)
-    exec_command_test.go:99: completed command must not return session_id, got ... session_id:"1000" ...
-FAIL github.com/Gitlawb/zero/internal/tools   13.664s
-exit 1
-```
-
-**Race detector: 0 data races reported in any package.** The 3 `internal/tools` failures are **`-race`-only timing flakes, not product defects**: they pass cleanly 3× without `-race` (`go test ./internal/tools -run <them> -count=3` → `ok`), and fail under `-race` only because the test's fixed 10ms yield window is too short when the binary runs ~5–10× slower under race instrumentation — the command is still running when the window elapses, so the tool *correctly* returns a still-running `session_id` instead of an `exit_code`. This is a test-robustness gap (the suite is not `-race`-clean) recorded under Confidence notes (§7); it is the same exec-session timing family that recently churned in PRs #270/#273. (A genuine, separate `-race`-detectable data race in this package — `lastUsedAt` — is reported as **L15**, found by inspection because no test exercises it.)
-
-## 3. Findings by severity
-
-> Each finding was re-verified against current code by independent skeptics; the evidence below is what withstood that challenge. Locations are `path:line` relative to the repo root.
-
-### H1 · Four of six hook events (sessionStart/sessionEnd/specialistStart/specialistStop) are validated and user-configurable but NEVER dispatched
-- **Severity / category:** high / dead-inert  · **Subsystem:** hooks
-- **Location:** `internal/hooks/dispatch.go:112; internal/agent/loop.go:914,937`
-- **Evidence:**
-```
-KnownEvents() returns all six (hooks.go:888) and IsValidEvent gates `zero hooks add` (hooks_manage.go:297-298) advertising all six in help (hooks_manage.go:315 "beforeTool, afterTool, sessionStart, sessionEnd, specialistStart, specialistStop"). But the only two production call sites of Dispatcher.Dispatch are loop.go:915 (`Event: hooks.EventBeforeTool`) and loop.go:938 (`Event: hooks.EventAfterTool`). grep over internal/+cmd/ for `.Dispatch(` finds exactly those two; no dispatch ever passes EventSessionStart/EventSessionEnd/EventSpecialistStart/EventSpecialistStop (the specialist lifecycle in internal/specialist/accounting.go:38,60 emits sessions.EventSpecialistStart/Stop to the SESSION store, never to hooks.Dispatcher).
-```
-- **Impact:** Across a whole session a user (or plugin) who configures a sessionStart/sessionEnd/specialistStart/specialistStop hook — e.g. a sessionStart hook that loads secrets/env, or a sessionEnd hook that flushes/cleans up, or a specialist audit hook — gets a green `zero hooks add` and a hook that `zero hooks list` shows as enabled, yet the command is silently never executed for the entire run. This is a correctness/trust gap: lifecycle automation users believe is wired (cleanup, audit, env priming) just does not happen, with no error surfaced. This is the same class as the prior 2026-06-10 high finding 'loaded but never executed' — the beforeTool/afterTool half was fixed; the other four events remai
-- **Fix:** Either (a) actually dispatch these events from their lifecycle points: call options.Hooks.Dispatch with EventSessionStart at session/agent-run start, EventSessionEnd at run teardown, and EventSpecialistStart/Stop alongside the sessions events in internal/specialist/accounting.go; or (b) if only beforeTool/afterTool are intended to be runnable, remove the other four from KnownEvents()/IsValidEvent and the CLI help so users cannot configure dead hooks. Option (a) is preferable since the events are already a documented surface.
-
-### M1 · streamjson api_key secret pattern still mangles ordinary prose/output in every stream-json event
-- **Severity / category:** medium / robustness  · **Subsystem:** cli-usage-streamjson
-- **Location:** `internal/streamjson/streamjson.go:314`
-- **Evidence:**
-```
-regexp.MustCompile(`(?i)(api[_-]?key["'=:\s]+)[^"',\s)]+`)  — the ["'=:\s]+ class lets a plain space follow apikey, so the next word is captured. Verified: redactString("The function apikey: foo is documented here") -> "...[REDACTED] is documented..."; redactString("The user's apiKey value spans...") -> "...apiKey [REDACTED] spans...". The sibling sk-/bearer patterns were anchored, but this one was not.
-```
-- **Impact:** FormatEvent runs redactValue->redactString over EVERY string field of EVERY emitted stream-json event (text-event Delta = model output, tool_result Output, final Text, warning/error Message). Any time the model's answer, a tool's output, a diff, a log line, or documentation contains the literal token "api_key:"/"apiKey "/"api-key=" followed by a word, a chunk of that non-secret content is silently replaced with [REDACTED] in the machine-readable protocol a downstream consumer parses. Over a session that discusses configs/auth (common for a coding agent) this corrupts arbitrary output with no warning. This is the unfinished half of prior finding M48.
-- **Fix:** Require an actual credential shape after the key marker rather than 'any next word': drop bare whitespace as a value delimiter (use `(api[_-]?key)\s*[=:]\s*["']?` for the prefix) and require a credential-length body, e.g. `[A-Za-z0-9._-]{12,}` (mirror the sk-/bearer anchoring already applied). Add a prose regression case ("the api_key: foo setting" must survive) to TestRedactStringDoesNotOverMatchProse.
-
-### M2 · ResolvedConfig.MCP is computed but never read; ZERO_PROVIDER_COMMAND MCP servers are silently dropped
-- **Severity / category:** medium / correctness  · **Subsystem:** config-plugins-skills
-- **Location:** `internal/config/resolver.go:106`
-- **Evidence:**
-```
-Resolve() merges provider-command output (LoadProviderCommand at resolver.go:54 + mergeConfig at :58, which calls mergeMCPConfig) and sets `MCP: cfg.MCP` (resolver.go:106). But `grep -rn 'resolved.MCP' internal/cli` returns nothing — no caller reads ResolvedConfig.MCP. Runtime MCP registration goes through the separate ResolveMCP (resolver.go:115-133), which only seeds DefaultMCPServers + merges UserConfigPath/ProjectConfigPath/Overrides.MCP and never calls LoadProviderCommand or touches options.ProviderCommand.
-```
-- **Impact:** A provider command (ZERO_PROVIDER_COMMAND) that emits an mcp.servers block has those servers folded into Resolve()'s result but the runtime registers from ResolveMCP, so provider-command-supplied MCP servers are never registered for the whole session — the user sees no MCP tools from them and no error. ResolvedConfig.MCP is also dead weight.
-- **Fix:** Have ResolveMCP also merge LoadProviderCommand(options.ProviderCommand).MCP (between the file merges and the overrides merge), or make the runtime read resolved.MCP from the single Resolve() path. Drop ResolvedConfig.MCP if it stays unread.
-
-### M3 · Project-checked-in slash commands auto-load and expand-to-prompt with no provenance marker or confirmation; Command.Project is dead
-- **Severity / category:** medium / robustness  · **Subsystem:** config-plugins-skills
-- **Location:** `internal/tui/user_commands.go:29`
-- **Evidence:**
-```
-model.go:473 `usercommands.Load(usercommands.DefaultPaths(cwd, userConfigDir))` auto-loads `<cwd>/.zero/commands/*.md` at TUI startup. handleUserCommand (user_commands.go:29-33) does `prompt := usercommands.Expand(cmd.Template, args)` then `m.launchPrompt(prompt)` — no confirmation, no display of whether the command is project- or user-sourced. The provenance field exists (usercommands.go:120 `Project: project`) but `grep -rn '\.Project' internal/tui internal/usercommands` shows it is only ever WRITTEN, never read. Project commands shadow user commands of the same name (Load merges project after user, byName overwrite).
-```
-- **Impact:** Opening Zero inside a cloned/untrusted repo silently loads that repo's `.zero/commands/*.md`. A repo can name a command after a common one (`/test`, `/fix`, `/pr`) so the user invoking it expands attacker-authored instructions and submits them to the model verbatim, driving the tool-enabled agent — with no indication the command body came from the repo rather than the user's own config. The Project flag built to distinguish this is never surfaced.
-- **Fix:** Surface provenance: in the command list/help and on first invocation of a project-sourced command, show it is repo-sourced (and ideally one-time confirm), using the already-present Command.Project flag; or scope auto-load of project commands behind a trust prompt like other workspace-trust gates.
-
-### M4 · DST fall-back double-fire: the collapse guard is dead on the real scheduler path (sub-minute `fired`)
-- **Severity / category:** medium / correctness  · **Subsystem:** cron
-- **Location:** `internal/cron/schedule.go:205`
-- **Evidence:**
-```
-Guard: `if sameWallClockMinute(t, after) && after.Truncate(time.Minute).Equal(after) { t = t.Add(time.Minute) } else { return t }`. But fireJob feeds raw wall-clock time: `fired := now()` (cron_run.go:136) then `sched.Next(fired)` (cron_run.go:190), and reconcileOverdue uses `sched.Next(now())` (cron_run.go:122) — none truncate to the minute. With real time.Now() the second clause `after.Truncate(time.Minute).Equal(after)` is false, so the collapse never engages. Reproduced: Next(2026-11-01 01:30:15 EDT) for `30 1 * * *` = 2026-11-01 01:30:00 EST (06:30 UTC) — same day, one absolute hour later.
-```
-- **Impact:** On the annual US/EU DST fall-back day, a daily job scheduled in the repeated wall-clock hour (e.g. `30 1 * * *` in a DST zone, NextRunAt stored in that location) runs twice that day: after firing at 01:30 EDT the post-exec advance computes NextRunAt=01:30 EST one hour later, and the next tick fires it again (claimFire's `current.NextRunAt.After(fired)` does not block it because the second 01:30 EST instant is not after the stored 01:30 EST). Duplicate scheduled agent run = duplicate model-token spend. The unit tests pass because they pass a minute-aligned `after`; the production caller never does.
-- **Fix:** Truncate the fire instant to the minute before computing the next slot, so the guard's precondition holds for the real caller: in fireJob set `fired := now().Truncate(time.Minute)` (or pass a truncated value into sched.Next at cron_run.go:122/190 and claimFire). Alternatively make Next collapse the fall-back repeat whenever the candidate shares wall-clock minute with `after` AND is a later absolute time, independent of after's sub-minute precision, adjusting the strictly-after contract accordingly.
-
-### M5 · claimFire grants the slot to BOTH concurrent schedulers when the schedule cannot advance (invalid/exhausted spec)
-- **Severity / category:** medium / concurrency  · **Subsystem:** cron
-- **Location:** `internal/cli/cron_run.go:256`
-- **Evidence:**
-```
-claimFire only advances NextRunAt when the expr parses AND Next is non-zero: `if sched, perr := cron.Parse(current.Expr); perr == nil { if nxt := sched.Next(fired); !nxt.IsZero() { current.NextRunAt = nxt } }` but unconditionally leaves `claimed := true` for an active, due job. For an unadvanceable schedule NextRunAt is left unchanged, so a second caller still sees Status==active and `!NextRunAt.After(fired)` and also returns claimed=true. Reproduced by calling claimFire twice with no fire in between: `claim1=true claim2=true`.
-```
-- **Impact:** Two concurrent schedulers (the documented `cron run --once` under system cron overlapping a forever-mode `cron run`, cron_run.go:61-63) both claim and both run `zero exec` for the same fire of a job whose Expr is impossible/exhausted (e.g. `0 0 30 2 *` stored directly per TestCronRunPausesUnadvanceableJob, a hand-edited metadata.json, or a finite spec that has run out near the 9-year search cap). Both spawn an agent before either pauses the job in its post-exec Mutate. This is the exact double-fire the claim mechanism exists to prevent; the prior concurrent test masks it only because the fake exec returns instantly so the first fire pauses the job before the second claims — a real minutes-lo
-- **Fix:** Make claimFire fail the claim for an unadvanceable schedule: if Parse errors or Next(fired) is zero, set `claimed = false` (let a single deterministic caller pick it up) or advance NextRunAt to a sentinel/pause it inside the same locked Mutate so the loser observes the change. The winner's post-exec block already pauses it; the claim just needs to not hand the same un-advanced slot to a second runner.
-
-### M6 · `cron add --run-now` job is silently cancelled by forever-mode startup reconcile instead of firing
-- **Severity / category:** medium / correctness  · **Subsystem:** cron
-- **Location:** `internal/cli/cron_run.go:118`
-- **Evidence:**
-```
-cronAdd --run-now sets `next = now()` (cron.go:159) and prints `next run <now>`. reconcileOverdue (run at forever-startup when !catchUp, cron_run.go:71) treats any NextRunAt before the current minute as strictly overdue and reschedules without firing: `nowMin := now().Truncate(time.Minute) ... if !j.NextRunAt.Before(nowMin) { continue } ... j.NextRunAt = nxt; store.Update(j)`. The Job struct carries no run-now marker, so the immediate request is indistinguishable from an ordinary backlog. Reproduced: --run-now job added at 08:00, `cron run` started at 08:05 → NextRunAt moved to 09:00, FireCount=0, no output.
-```
-- **Impact:** A user who runs `zero cron add ... --run-now` and then starts the foreground scheduler more than a minute later never gets the promised immediate run; the job silently waits until its next regular slot. Contradicts both the --run-now contract and the message cronAdd printed. Still open from the prior audit (docs/audit/2026-06-10-deep-audit.md:2208).
-- **Fix:** Persist a one-shot intent (e.g. a RunNow bool on Job cleared on first fire, or set NextRunAt to a future-safe sentinel that reconcile fires rather than skips). In reconcileOverdue, exempt jobs explicitly marked run-now (fire them in the first fireDue pass) instead of treating their now() NextRunAt as stale backlog.
-
-### M7 · Remote bridge clears the connection deadline before the daemon handshake, so an authenticated-but-idle peer pins a connection slot forever
-- **Severity / category:** medium / resource-perf  · **Subsystem:** daemon
-- **Location:** `internal/daemon/remote/bridge.go:223`
-- **Evidence:**
-```
-// Clear the handshake deadline: a session may stream...
-_ = conn.SetDeadline(time.Time{})
-...
-b.server.ServeConn(conn) // performs the daemon handshake + one command
---- server.go:225-231 (reached via ServeConn) ---
-func (s *Server) handleConn(conn net.Conn) {
-	defer conn.Close()
-	hello, err := ReadControl(conn) // no deadline
-```
-- **Impact:** The bridge bounds only the auth handshake (line 181) and then fully clears the conn deadline. After a peer passes auth it is handed to ServeConn -> handleConn, which does ReadControl(conn) for the daemon hello and again for the command with NO read deadline and no per-command timeout. A peer that completes the auth handshake (valid token) but then never sends the daemon hello (or sends hello but never the command) blocks its handler goroutine indefinitely while holding one of the b.sem slots (default MaxConnections=32). 32 such stalled connections exhaust the bridge: every other remote client is refused 'at capacity' (bridge.go:164-168) until process restart. This is an authenticated-client
-- **Fix:** Do not fully clear the deadline before ServeConn. Either keep a (larger) per-command deadline for the daemon hello+command exchange in the session path, or set a fresh deadline covering the daemon handshake and clear it only once streaming begins. Simplest: in bridge.go before ServeConn, set conn.SetDeadline(now+handshakeTimeout) for the session path and let streamToClient clear it, OR add a read deadline inside handleConn around the two ReadControl handshake calls.
-
-### M8 · Server.Shutdown cannot close remote bridge connections (ServeConn bypasses trackConn), so a stalled remote handshake-read survives Shutdown
-- **Severity / category:** medium / robustness  · **Subsystem:** daemon
-- **Location:** `internal/daemon/server.go:222`
-- **Evidence:**
-```
-func (s *Server) ServeConn(conn net.Conn) { s.handleConn(conn) }
---- vs the local accept loop (server.go:136-141) ---
-s.trackConn(conn)
-s.wg.Add(1)
-go func() { defer s.wg.Done(); defer s.untrackConn(conn); s.handleConn(conn) }()
---- Shutdown only closes tracked conns (server.go:181-183) ---
-for c := range s.conns { _ = c.Close() }
-```
-- **Impact:** The D3 fix (Shutdown closes every open connection so a handler blocked on a read returns) only covers connections accepted by the LOCAL Serve loop, which calls trackConn. The remote bridge enters via ServeConn, which calls handleConn directly and never registers the conn in s.conns. The streaming phase is still interruptible (streamToClient selects on <-s.done), but a remote connection blocked in the pre-stream ReadControl handshake (combined with the no-deadline gap above) is NOT closed by Shutdown and is NOT covered by s.wg, so it leaks until process exit and keeps occupying a bridge slot through the entire graceful-drain window. The asymmetry means the carefully-built D3/D4/D5 shutdown gu
-- **Fix:** Have ServeConn register the conn for shutdown closure too: call s.trackConn(conn)/defer s.untrackConn(conn) inside ServeConn (or expose a tracked variant the bridge uses), so Shutdown's conns-close loop unblocks a stalled remote handshake read as it does for local connections. Pair with a handshake read deadline (previous finding) for defense in depth.
-
-### M9 · MCP SSE client never reconnects: any stream close permanently kills the session
-- **Severity / category:** medium / robustness  · **Subsystem:** mcp
-- **Location:** `internal/mcp/network_client.go:517`
-- **Evidence:**
-```
-readStream end: `client.failPending(fmt.Errorf("MCP SSE stream closed for server %s", client.server.Name))`; failPending sets `client.streamErr = err` (line 593); request() gate: `if client.streamErr != nil { ... return err }` (lines 243-246). openStream is invoked exactly once, only from connectRemoteSSE at connect (line 77) — grep finds no other caller and no reconnect loop. registry.go connects each server once at startup and reuses the client for the whole session with no reconnect-on-error.
-```
-- **Impact:** For an SSE-transport MCP server, the persistent GET stream closing for ANY reason during a session — clean server-side EOF (the normal `MCP SSE stream closed` path), an idle/proxy timeout, a network blip, or an over-8MiB SSE line tripping bufio.ErrTooLong — sets a permanent streamErr. Every subsequent tools/call and tools/list for that server then returns that error for the rest of the process lifetime; the only recovery is restarting Zero (or `zero mcp enable` per tui hint). A long agent session that idles past a server keep-alive window silently loses all of that server's tools.
-- **Fix:** On a non-cancelled stream termination, attempt a bounded re-open: clear streamErr/endpointURL, call openStream again with backoff (e.g. mirror agent/reconnect.go's maxStreamReconnects + exponential backoff), and only set a permanent streamErr after retries are exhausted or ctx is done. Re-send `notifications/initialized` after a successful re-open so the server re-establishes session state.
-
-### M10 · providerhealth probe re-sends x-api-key / x-goog-api-key / CustomHeaders to a cross-host redirect target (credential exfil to a public attacker host)
-- **Severity / category:** medium / security  · **Subsystem:** oauth-providers
-- **Location:** `internal/providerhealth/providerhealth.go:430`
-- **Evidence:**
-```
-CheckRedirect: func(req *http.Request, via []*http.Request) error { if len(via) >= maxConnectivityRedirects {...}; return validateEndpoint(req.Context(), req.URL.String(), resolver) }  — re-validates the redirect HOST against the private/special-use blocklist but never strips auth headers. applyAuth (line 529-561) sets DefaultAuthHeader "x-api-key" (Anthropic), "x-goog-api-key" (Google), and arbitrary profile.CustomHeaders (lines 539/548/558).
-```
-- **Impact:** Go's stdlib only auto-strips Authorization/Cookie/WWW-Authenticate when a redirect crosses hosts; it does NOT strip x-api-key, x-goog-api-key, or arbitrary custom headers. The DNS-rebinding/internal-redirect half of the prior high finding is now fixed (safeDialContext pins validated IPs; CheckRedirect blocks internal targets), but a configured baseURL that returns a 3xx to an attacker-controlled PUBLIC host (a compromised/MITM'd/typo'd proxy endpoint, or a hostile OpenAI-compatible gateway the user pointed at) passes the blocklist and receives the provider API key verbatim. Production callers (observability.go:58, provider_setup.go:99, setup.go:116) pass no HTTPClient, so newConnectivityClie
-- **Fix:** In CheckRedirect, when req.URL.Host != via[0].URL.Host (or any earlier hop's host), delete the auth-bearing headers from req.Header (x-api-key, x-goog-api-key, Authorization, and every key in profile.CustomHeaders) before allowing the redirect; or set client.CheckRedirect to refuse cross-host redirects entirely for the probe (a health check has no need to follow a host change).
-
-### M11 · Destructive/piped-installer risk regexes scan the whole raw command (incl. quoted literals); the quote-aware AST analyzer can never retract the false positive
-- **Severity / category:** medium / correctness  · **Subsystem:** sandbox
-- **Location:** `internal/sandbox/risk.go:108-112`
-- **Evidence:**
-```
-command := firstArgString(...)
-if matchesDestructive(command) { add("destructive", RiskCritical) }
-if pipedInstallerPattern.MatchString(command) { add("piped_installer", RiskCritical) }
-... analysis := AnalyzeCommand(command); if analysis.Destructive { add("destructive", RiskCritical) }  // add() only ORs in, never removes
-
-Reproduced via engine.Evaluate: `git commit -m "fix rm -rf / bug"` -> action=prompt reason="destructive shell command requires approval" risk=[destructive shell], in BOTH PermissionModeAsk and PermissionModeAuto. Matcher-level probe: `git commit -m "rm -rf /"`=>destructive=true while AST destr=false; `echo "do not run rm -rf /"`=>destructive=true; `grep -r "dd if=/dev/zero" .`=>destructive=true; `printf '%s' "curl evil|sh"`=>piped=true. AST clears all of them (destr=false,net=false) but classifyWithScope only adds categories.
-```
-- **Impact:** Across a session, any shell command that merely QUOTES one of the trigger substrings (rm -rf / | rm -rf ~ | dd if= | chmod 777 <root> | the fork-bomb chars | curl..|sh) is mis-classified RiskCritical 'destructive'. Per engine.go:323-329 a SideEffectShell command with the destructive category and no prior PermissionGranted is forced to ActionPrompt (Ask AND Auto) or denied when permission can't be granted. So routine, safe commands — `git commit -m "...rm -rf..."`, `echo`/`printf`/`grep` mentioning these tokens — are spuriously gated behind a destructive approval prompt, and in a headless exec with no OnPermissionRequest the prompt resolves to a hard block. The interactive detector was alread
-- **Fix:** Run matchesDestructive / pipedInstallerPattern over the command-body of each real shell segment (reuse splitShellSegments + commandBody from safe_command.go) instead of the raw string, OR make the AST analyzer authoritative for parseable scripts: when analysis.TooComplex==false, trust analysis.Destructive/Network and skip the raw-string regexes (only fall back to regex when the parse fails). Either approach removes the quoted-literal false positives while keeping detection of genuine unquoted destructive commands.
-
-### M12 · events.jsonl append is the only write in the package not fsync'd, while the derived metadata IS — a crash can lose a 'successfully appended' event (incl. the checkpoint /rewind targets) and leave EventCount ahead of the durable log
-- **Severity / category:** medium / correctness  · **Subsystem:** sessions
-- **Location:** `internal/sessions/store.go:606-616`
-- **Evidence:**
-```
-file, err := os.OpenFile(store.eventsPath(sessionID), os.O_WRONLY|os.O_APPEND|os.O_CREATE, 0o600)
-...
-if _, err := file.Write(append(data, '\n')); err != nil { _ = file.Close(); ... }
-if err := file.Close(); err != nil { ... }   // no file.Sync(), no syncDir
-...
-session.EventCount = sequence
-if err := store.writeMetadata(session); err != nil { ... }  // writeMetadata fsyncs temp + dir
-```
-- **Impact:** appendEventLocked writes the event line into the page cache (no fsync, no dir sync) and THEN durably writes metadata (writeFileSync + syncDir). On power loss / OS crash after the metadata fsync but before the events.jsonl page is flushed, the persisted metadata.EventCount=N but events.jsonl is missing line N. Across a session this silently drops the last appended event even though AppendEvent returned success: a just-recorded EventSessionCheckpoint vanishes (its blob, written via writeFileAtomicSync, IS durable but is now unreferenced, so /rewind cannot target it and the file content is orphaned), and the user's last turn (message/tool_call/usage) is lost on resume. It also inverts the seque
-- **Fix:** Make the append durable like every other write: after file.Write, call file.Sync() before Close (and ideally syncDir(sessionPath) on first create of events.jsonl). Cheapest: replace the open/append with an fsync'd append helper, e.g. f.Write(...); if err := f.Sync(); err != nil {...}; f.Close(). This restores the ordering guarantee that the log is at least as durable as the metadata derived from it.
-
-### M13 · Mailbox stale-lock break compares the lock file to itself, not to the snapshot observed as stale — can delete a healthy fresh lock (split-brain writers)
-- **Severity / category:** medium / concurrency  · **Subsystem:** specialist-swarm
-- **Location:** `internal/swarm/mailbox.go:371`
-- **Evidence:**
-```
-if info, statErr := os.Stat(lockPath); statErr == nil && time.Since(info.ModTime()) > lockStaleAfter {
-	stale, _ := os.ReadFile(lockPath)
-	if data, rerr := os.ReadFile(lockPath); rerr == nil && string(data) == string(stale) {
-		os.Remove(lockPath)
-	}
-	continue
-}
-```
-- **Impact:** The comment claims the break is 'conditional on the file's content being unchanged since we observed it,' but `stale` and `data` are two consecutive reads of the SAME file microseconds apart, so the equality check is essentially always true and provides no protection. If a holder crashes and leaves a stale lock, two (or more) waiters can each observe it stale and each call os.Remove(); meanwhile a third writer may legitimately re-create the lock via O_EXCL between the two removes. The second waiter's Remove then deletes that fresh, validly-held lock, after which both that holder and the waiter (on its next O_EXCL create) hold the lock simultaneously — two concurrent writers to the same per-a
-- **Fix:** Capture the lock content at Stat time and only remove if it is byte-identical AND the ModTime is unchanged at remove time; better, make the break ownership-aware like release(): read the token, then remove via an atomic rename-then-unlink or re-stat-and-verify-mtime guard. Simplest robust fix: snapshot (content, modTime) when first seen stale, sleep a short grace, re-stat, and only Remove if both still match — so a freshly rotated lock (new mtime/content) is never deleted.
-
-### M14 · PTY exec session can drop its final output chunk on exit (copy goroutine unsynchronized with markDone), then the session is removed
-- **Severity / category:** medium / concurrency  · **Subsystem:** tools
-- **Location:** `internal/tools/exec_command.go:515-526 (with exec_pty_linux.go:34-37)`
-- **Evidence:**
-```
-startSession's Wait goroutine: `err := command.Wait(); ... if session.cleanup != nil { session.cleanup() }; plan.Cleanup(); cancel(); session.markDone(...)`. For a TTY session the only thing copying PTY output into the buffer is the fire-and-forget goroutine in exec_pty_linux.go:34 `go func() { _, _ = io.Copy(output, master) }()`. command.Stdout/Stderr are the *os.File slave (exec starts NO copy goroutine and WaitDelay does not cover our own goroutine), so command.Wait() returns as soon as the process exits — before io.Copy has necessarily drained the master FD. markDone then closes `done`. collect (line 541-545) does a final drainString after doneClosed, but if the copy goroutine has not yet written the tail, that drain misses it; run()/RunWithOptions then call `tool.manager.remove(session.id)` (lines 465, 676) so the session is gone and no later poll can recover the lost tail.
-```
-- **Impact:** Across a session, the last lines a TTY (tty:true) command prints right before it exits can be silently dropped from the tool result the model sees — e.g. a test runner's final PASS/FAIL summary or a build's last error line — and because the exited session is immediately removed, the model cannot poll for it. Pipe-mode (tty:false) is unaffected because exec's own Wait drains its os.Pipe. Linux-only (other platforms fall back to pipe via exec_pty_fallback.go).
-- **Fix:** Make the PTY copy goroutine observable: have startPTYProcess return a `copied <-chan struct{}` (closed when io.Copy returns), and in the Wait goroutine block on it (with a bounded timeout matching bashWaitDelay) AFTER cleanup() closes master and BEFORE markDone, so all buffered PTY output is in the buffer before `done` is closed and the session is eligible for removal.
-
-### M15 · Streaming live block re-parses the entire growing answer markdown on every spinner/fade tick (cumulative O(N^2) per turn)
-- **Severity / category:** medium / resource-perf  · **Subsystem:** tui-render
-- **Location:** `internal/tui/model.go:2079`
-- **Evidence:**
-```
-interimBlock: `lines := renderAssistantMarkdownText(text, assistantMeasure(width), width)` where `text := strings.TrimRight(m.streamingText, "\n")`. streamingText only grows (model.go:1185 `m.streamingText += msg.delta`) and is cleared solely at turn end (1411/1522/3208). interimBlock is called from the per-frame body item closure (transcript_selection.go:245 `lines: viewLines(m.interimBlock(width))`). The spinner self-reschedules while pending (model.go:1212-1238, gated on m.pending) and the fade ticks every 150ms (streaming_fade.go:79), so the full markdown parse+wrap runs ~12x/sec.
-```
-- **Impact:** Settled transcript rows are memoized via defaultRenderCache (rendering.go:190-197), but the streaming block is NOT cached. For a long single answer (e.g. a model emitting a multi-hundred/thousand-line file or table as prose), each ~80-150ms tick re-runs renderAssistantMarkdownText over the whole accumulated buffer — fence scanning, table parsing, per-line inline wrap with lipgloss.Width. Cost per frame is O(current length); over the turn the work is O(N^2/delta). On a slow/large stream the TUI visibly chews CPU and the frame rate drops the longer the answer gets, exactly when responsiveness matters most.
-- **Fix:** Memoize the rendered streaming lines keyed by (streamingText length or a content hash, width); only re-render when streamingText or width actually changed since the last frame. The fade is applied per-line AFTER the markdown render (styleStreamingLine, model.go:2087), so the expensive markdown step can be cached while the cheap per-line color pass still runs each tick. Alternatively bound the markdown render to the visible tail (last height lines) since the viewport clips anyway.
-
-### L1 · Reactive-compaction retry and max-turns final-answer call bypass streamWithReconnect (no transient-disconnect retry)
-- **Severity / category:** low / resource-perf  · **Subsystem:** agent-loop
-- **Location:** `internal/agent/loop.go:205,253,518`
-- **Evidence:**
-```
-stream, err = provider.StreamCompletion(ctx, request) // reactive connect-failure retry
-retryStream, retryStreamErr := provider.StreamCompletion(ctx, retryRequest) // reactive mid-stream retry
-stream, err := provider.StreamCompletion(ctx, ...) // finalAnswerAfterMaxTurns
-```
-- **Impact:** The main turn connect goes through streamWithReconnect (2 retries with backoff on EOF/reset/502/503/timeout), but the post-compaction retry calls and the max-turns final-answer call call provider.StreamCompletion directly. A single transient upstream hiccup on exactly those calls (e.g. the final summary after a long autonomous/cron run, or the retried turn right after a context-limit compaction) fails the whole run with no reconnect, re-burning the entire session's tokens — the precise failure mode streamWithReconnect was added to prevent.
-- **Fix:** Route these three calls through streamWithReconnect(ctx, provider, request, reconnectNoticeFor(options)) as well; they are pre-content connects so no OnText duplication risk, matching the existing reconnect contract.
-
-### L2 · OnContext / MeasureContext / estimateToolTokens context-utilization pipeline has no production consumer
-- **Severity / category:** low / dead-inert  · **Subsystem:** agent-loop
-- **Location:** `internal/agent/loop.go:164-166`
-- **Evidence:**
-```
-if options.OnContext != nil {
-	options.OnContext(MeasureContext(messages, request.Tools, options.ContextWindow))
-}
-```
-- **Impact:** Repo-wide grep shows zero `OnContext:` assignments outside agent tests (cli/tui/exec never set it), so MeasureContext and its private estimateToolTokens (a near-duplicate of compaction.go's estimateToolDefTokens) are computed for nobody in production and ship as permanently inert plumbing. Carries a maintenance hazard (two divergent tool-token estimators) and a tiny per-turn branch. Matches the prior low finding — still open.
-- **Fix:** Either wire OnContext from the TUI/exec to drive a context-utilization bar, or delete OnContext + MeasureContext + ContextBreakdown + estimateToolTokens and keep the single estimateToolDefTokens used by compaction.
-
-### L3 · zero --skip-permissions-unsafe still silently drops trailing positional args (e.g. a prompt)
-- **Severity / category:** low / robustness  · **Subsystem:** cli-usage-streamjson
-- **Location:** `internal/cli/app.go:234`
-- **Evidence:**
-```
-moreDirs, rest, err := splitLeadingAddDirFlags(args[1:]) ... for _, arg := range rest { if arg == "--add-dir"... { return error } } ... return runInteractiveTUI(stderr, deps, agent.PermissionModeUnsafe, append(append([]string{}, addDirs...), moreDirs...)). `rest` is iterated only to reject a misplaced --add-dir; every other trailing arg is discarded and the interactive TUI (which takes no positional prompt) is launched.
-```
-- **Impact:** `zero --skip-permissions-unsafe "fix the bug"` silently discards the prompt and drops the user into the interactive TUI with no error or warning, so a scripted/one-shot unsafe invocation appears to hang/ignore input. Documented in-code as intentional ('were ignored on this path... and still are'), but it is still a silent UX drop and unchanged since the prior audit (prior M81 / app.go:149).
-- **Fix:** If `rest` contains any non-flag tokens, fail loudly ("--skip-permissions-unsafe launches the interactive TUI and takes no prompt; use `zero --skip-permissions-unsafe exec -p ...` or pipe input") instead of discarding them, matching the loud-rejection style already used for misplaced --add-dir two lines above.
-
-### L4 · usercommands frontmatter model:/agent:/mode: are parsed and documented but never applied
-- **Severity / category:** low / dead-inert  · **Subsystem:** config-plugins-skills
-- **Location:** `internal/usercommands/usercommands.go:125`
-- **Evidence:**
-```
-parseCommand fills cmd.Model (usercommands.go:125) and cmd.Agent (usercommands.go:126-129 from `agent:`/`mode:`), and the package doc (lines 8-9) advertises `model:` as a real override. But the sole invocation path, handleUserCommand (tui/user_commands.go:29-33), calls launchPrompt(prompt) — and `launchPrompt(prompt string)` (model.go:3020) takes only the prompt; cmd.Model/cmd.Agent are never read by any caller (`grep` for the fields outside the struct/parse shows only the parse assignment).
-```
-- **Impact:** A user who writes `model: <x>` or `agent: <x>` in a `.zero/commands/foo.md` frontmatter, per the documented feature, gets it silently ignored — the command always runs on the current model/agent. The fields are inert surface that mislead users into thinking per-command model routing works.
-- **Fix:** Either honor cmd.Model/cmd.Agent in handleUserCommand (route the launched prompt through the model/agent switch before launchPrompt) or remove the fields and the package-doc `model:` line.
-
-### L5 · Runtime skill tool discards duplicate-name collisions (no user-facing warning)
-- **Severity / category:** low / dead-inert  · **Subsystem:** config-plugins-skills
-- **Location:** `internal/plugins/activate.go:376`
-- **Evidence:**
-```
-skillTool.Run does `merged, _ := MergedSkillsLoaded(tool.defaultDir, tool.pluginRoots)` (activate.go:376), discarding the DuplicateName slice that mergeSkills computes (activate.go:311). The CLI `zero skills list` path now DOES surface dups (cli/skills.go:69-81), but the agent-facing skill tool silently drops a shadowed same-named skill.
-```
-- **Impact:** When a plugin skill and a user skill (or two plugin roots) share a frontmatter name, the loser is silently dropped at agent runtime and the model/user is never told which skill body it actually got — a shadowed skill disappears with no diagnostic during a session.
-- **Fix:** When the skill tool resolves a name that had a collision, append a one-line note to the tool Result (or emit a startup warning) listing the shadowed Loser path, reusing the already-returned DuplicateName slice.
-
-### L6 · pause/resume do an unlocked Get then a locked Update (non-atomic read-modify-write)
-- **Severity / category:** low / concurrency  · **Subsystem:** cron
-- **Location:** `internal/cli/cron.go:215`
-- **Evidence:**
-```
-cronSetStatus: `job, err := store.Get(args[0])` (Get takes no lock, store.go:150) ... `job.Status = status; store.Update(job)` (Update takes the per-job lock, store.go:170-183). cronResume (cron.go:234-253) is the same Get-then-Update shape, overwriting the whole struct including FireCount/NextRunAt from the pre-read snapshot.
-```
-- **Impact:** If a fireJob post-exec Mutate lands between resume's unlocked Get and resume's locked Update, the fire's FireCount++ and advanced NextRunAt are clobbered by resume's stale snapshot (resume recomputes NextRunAt from now() so scheduling stays sane, but FireCount silently regresses). The pause race is benign because fireJob re-reads Status under its lock and honors an external pause; resume has no such re-read. Low frequency (user-driven), cosmetic FireCount loss, no double-fire.
-- **Fix:** Route pause/resume through store.Mutate so the read and the field update happen under the same per-job lock: read current, set Status (and for resume recompute NextRunAt from current.Expr), preserving current.FireCount, and write — instead of Get+Update on a snapshot.
-
-### L7 · Single-instance lock refuses startup on PID reuse (bare-PID lock file, no identity/start-time)
-- **Severity / category:** low / robustness  · **Subsystem:** daemon
-- **Location:** `internal/daemon/lock.go:58`
-- **Evidence:**
-```
-pid, perr := readPidFile(path)
-if perr == nil && pid > 0 && isAlive(pid) {
-	return nil, fmt.Errorf("%w (pid %d)", ErrAlreadyRunning, pid)
-}
-```
-- **Impact:** The lock file stores only a bare PID. If a daemon dies uncleanly and the OS later recycles its exact PID for an unrelated live process, acquireLock sees isAlive(pid)==true and refuses to start with ErrAlreadyRunning, even though no zero daemon is running. The reclaim path (reclaimStaleLock, D6) is sound for the genuinely-dead case and the rename-aside race is well handled, but it never triggers here because the recycled PID reads as alive. Recovery requires manual lock removal. Low probability (exact PID reuse) and self-inflicted only after an unclean crash, but it is a real availability edge the bare-PID format cannot disambiguate.
-- **Fix:** Record more than the PID in the lock file (e.g. PID + process start time, or a daemon-specific marker / boot id) and treat a live PID as the holder only when the recorded identity matches. On mismatch, treat as stale and reclaim via the existing rename-aside path.
-
-### L8 · Plugin hook event set disagrees with hooks package: plugins reject specialist events and cannot register them
-- **Severity / category:** low / robustness  · **Subsystem:** hooks
-- **Location:** `internal/plugins/plugins.go:38-42,893-899; internal/plugins/activate.go:412-423`
-- **Evidence:**
-```
-hooks.go parseEvent (hooks.go:901-911) accepts six events. plugins.go HookEvent consts (lines 38-42) declare only HookBeforeTool/HookAfterTool/HookSessionStart/HookSessionEnd, and parseHookEvent (plugins.go:893-898) switches on exactly those four, rejecting specialist* with "Expected beforeTool, afterTool, sessionStart, or sessionEnd." mapHookEvent (activate.go:412-423) likewise maps only those four and returns ok=false otherwise.
-```
-- **Impact:** A plugin manifest declaring a specialistStart/specialistStop hook is rejected as a schema error even though hooks.json (user/project) accepts the same value. The two config surfaces for the identical Dispatcher have divergent valid-event sets, so plugin authors hit an inconsistent contract. Low because (per the finding above) those events do not actually fire from any surface today, so the practical effect is a confusing validation error rather than lost execution. This is the still-open prior reverification low (2026-06-13 line 97).
-- **Fix:** Once the dead events are resolved (finding above), make the two sets agree: drive plugin validation/mapping off hooks.KnownEvents()/IsValidEvent (or add HookSpecialistStart/Stop consts + cases) so a single source of truth defines which events are valid, eliminating the drift.
-
-### L9 · Stdio request write blocks under client.mu with no ctx awareness
-- **Severity / category:** low / robustness  · **Subsystem:** mcp
-- **Location:** `internal/mcp/client.go:282`
-- **Evidence:**
-```
-request() locks client.mu at line 264 and, still holding it, calls `client.writer.write(rpcMessage{...})` at line 282 before unlocking at 291. writer.write (protocol.go:148-163) does a blocking bufio Write+Flush to the child's stdin pipe with no context. Only the response wait (select at 293) is ctx-aware.
-```
-- **Impact:** If a misbehaving child stops draining its stdin, the OS pipe buffer (~64 KiB) fills and write() blocks while holding client.mu, serializing and stalling every other in-flight/queued request on that server with no ctx-cancel escape. Mitigated because Close() uses a separate closeMu (not mu) and closes stdin, which unblocks the stuck write — so it is recoverable rather than a hard deadlock; impact is limited to a pathological non-reading child.
-- **Fix:** Keep id allocation + pending registration under mu but perform the write outside mu (or write under a dedicated writeMu), and bound it: run the write in a goroutine and select on ctx.Done() so a stalled pipe surfaces ctx.Err() to the caller instead of blocking peers.
-
-### L10 · Child stderr capture keeps the head, dropping the likely-relevant tail on crash
-- **Severity / category:** low / robustness  · **Subsystem:** mcp
-- **Location:** `internal/mcp/client.go:106`
-- **Evidence:**
-```
-boundedBuffer.Write: `if remaining := b.cap - b.buf.Len(); remaining > 0 { ... b.buf.Write(p[:remaining]) ... }` — once cap (64 KiB) is reached, all further bytes are discarded; only the earliest bytes are retained. connectStdio reads stderr only on initialize failure (lines 150-154).
-```
-- **Impact:** Diagnostics quality: a server that prints a verbose startup banner (>64 KiB) and then dies with an error message at the end shows only the banner in the wrapped initialize error; the actual fatal line is dropped. Not a correctness/security issue, but it can make MCP startup failures hard to diagnose.
-- **Fix:** Prefer a ring buffer that retains the LAST cap bytes (or head+tail split) so the terminal error is preserved; the bound is unchanged.
-
-### L11 · ChatGPT account-id silently omitted (no warning) when the id_token claim exists but is the wrong shape
-- **Severity / category:** low / robustness  · **Subsystem:** oauth-providers
-- **Location:** `internal/provideroauth/chatgpt.go:227`
-- **Evidence:**
-```
-if ns, ok := claims[chatgptAuthClaimNamespace].(map[string]any); ok { if value, ok := ns[chatgptAccountClaim].(string); ok { return strings.TrimSpace(value), nil } }  ... return "", nil  — a present-but-non-string nested claim (or a present namespace with a missing key) falls through to `return "", nil`, the same as 'no id_token'. The caller (line 154-162) only warns on claimErr!=nil and only sets Account when account!="".
-```
-- **Impact:** If OpenAI ever changes the claim's JSON shape (e.g. nests it one level deeper, or makes it a number), extraction returns ("", nil) with no error: token.Account stays empty, the chatgpt-account-id header is omitted on every Codex call (codex.go:172 sets it only when ok&&account!=""), and the user gets Cloudflare 401s on every request with zero diagnostic — neither the login warning nor a parse error fires. The current 401 is then indistinguishable from an expired token.
-- **Fix:** Distinguish 'claim absent' from 'claim present but unusable': when claims[ns] exists but yields no non-empty string account id, return a non-nil error (e.g. 'account-id claim present but not a string') so the existing warning path at chatgpt.go:159 fires and the user is told to re-auth.
-
-### L12 · Latest()/List() order by second-granularity UpdatedAt with a SessionID lexical tiebreak; same-second updates with user-supplied IDs can resume the wrong session
-- **Severity / category:** low / correctness  · **Subsystem:** sessions
-- **Location:** `internal/sessions/store.go:332-337,708-710`
-- **Evidence:**
-```
-func (store *Store) timestamp() string { return store.now().UTC().Format(time.RFC3339) }  // second precision
-...
-if sessions[left].UpdatedAt == sessions[right].UpdatedAt { return sessions[left].SessionID < sessions[right].SessionID }
-return sessions[left].UpdatedAt > sessions[right].UpdatedAt
-```
-- **Impact:** UpdatedAt is RFC3339 (1s resolution). When two sessions are touched in the same wall-clock second (rapid back-to-back exec runs, or cron firing several jobs), the tiebreak is lexical SessionID, not real recency. exec_sessions.go:76 store.Latest() (used to resume the most recent session) and session.go:170 LatestResumable() can therefore land on the wrong session. For store-generated IDs the ID embeds UnixNano so lexical order roughly tracks recency, but for caller-supplied SessionIDs (Create/Fork/Child accept an explicit SessionID) the order is arbitrary, so `--continue`-style resume can attach to a stale session. Low: narrow timing window and mostly mitigated for auto-generated IDs.
-- **Fix:** Persist a higher-resolution UpdatedAt (RFC3339Nano) so same-second ties are broken by true time, and keep SessionID only as a final deterministic tiebreak; or carry a monotonically increasing per-store update counter into the sort key.
-
-### L13 · Scheduler leaks a per-job context when a bounded job completes via MaxRuns
-- **Severity / category:** low / resource-perf  · **Subsystem:** specialist-swarm
-- **Location:** `internal/swarm/scheduler.go:280`
-- **Evidence:**
-```
-runs := job.incRuns()
-if max := job.schedule.MaxRuns; max > 0 && runs >= max {
-	return
-}
-```
-- **Impact:** Each scheduled job gets its own ctx, cancel := context.WithCancel(s.ctx) (Add, line 182). On the MaxRuns-completion path run() returns without ever calling job.cancel(); only Cancel() and Close() invoke it. The derived context (and the goroutine the runtime spawns to propagate parent cancellation) is retained until the whole scheduler/Swarm context is cancelled at session end. A long session that adds many bounded daily/interval jobs accumulates one leaked context per completed job. Bounded by session lifetime and small, but it is a real, avoidable leak the codebase elsewhere is careful about.
-- **Fix:** Add `defer job.cancel()` at the top of run() (alongside the existing `defer s.wg.Done()` / `defer s.forget(job.id)`), so the job's context is released on every exit path including MaxRuns completion.
-
-### L14 · terminateProcess returns success without killing surviving group members when the leader PID was already reaped
-- **Severity / category:** low / robustness  · **Subsystem:** specialist-swarm
-- **Location:** `internal/background/process_posix.go:55`
-- **Evidence:**
-```
-target := pid
-if pgid, err := syscall.Getpgid(pid); err == nil {
-	if pgid == pid {
-		target = -pid
-	}
-} else if processGoneError(err) {
-	return nil // already gone
-}
-```
-- **Impact:** launchBackgroundProcess (specialist/exec.go:797) runs a command.Wait() goroutine that reaps the background specialist leader as soon as it exits. If the leader exits on its own but had forked a detached grandchild into the same process group, a later TaskStop calls terminateProcess(pid): Getpgid(reaped pid) returns ESRCH, the code treats it as 'already gone' and returns nil — never signalling the surviving group. TaskStop reports success while grandchildren (e.g. a dev server the sub-agent started and the parent zero exec exited without reaping) keep running. The full-tree kill the M6 fix added only works while the leader is still alive to identify the group.
-- **Fix:** Persist the child's pgid at launch (it equals the pid under Setpgid) and, when the leader is gone, fall back to syscall.Kill(-pgid, ...) to sweep any survivors instead of returning nil immediately on a reaped leader.
-
-### L15 · Data race on execSession.lastUsedAt: prune reads it under manager.mu while touch writes it under session.mu
-- **Severity / category:** low / concurrency  · **Subsystem:** tools
-- **Location:** `internal/tools/exec_command.go:123 vs :278`
-- **Evidence:**
-```
-sessionToPruneLocked (called from store under manager.mu) sorts with `sessions[i].lastUsedAt.Before(sessions[j].lastUsedAt)` (line 123), reading lastUsedAt with only manager.mu held. touch() writes it under a DIFFERENT lock: `session.mu.Lock(); session.lastUsedAt = time.Now(); session.mu.Unlock()` (lines 277-279). snapshot() correctly reads it under session.mu (line 291), but the prune path does not. The two paths use disjoint mutexes, so there is no happens-before between the write and the read.
-```
-- **Impact:** When a new exec_command at the session cap triggers eviction (store -> sessionToPruneLocked) concurrently with a write_stdin poll calling touch() on another live session (or with removeCompletedLater/stopAll churn), this is a genuine data race on a time.Time (a multi-word struct), which `go -race` would flag and which can yield a torn/garbage timestamp that mis-selects the prune victim. Unexercised by current tests (no test drives concurrent touch+prune), so it is latent rather than observed.
-- **Fix:** Read lastUsedAt through the session lock in the prune comparator (snapshot a `[]struct{id int; last time.Time}` by calling a small locked getter per session before sorting), or store lastUsedAt as an atomic.Int64 of UnixNano accessed the same way from both touch and prune.
-
-### L16 · truncateRunes/middleTruncate truncate by rune COUNT but are used with display-WIDTH budgets — wide/CJK content overshoots its budget
-- **Severity / category:** low / correctness  · **Subsystem:** tui-render
-- **Location:** `internal/tui/view.go:918`
-- **Evidence:**
-```
-truncateRunes: `runes := []rune(text); if len(runes) <= limit { return text }; ... return string(runes[:limit-1]) + "…"` — purely rune-count based. middleTruncate (startup.go:138-153) likewise slices `runes[:front]`/`runes[len-back:]`. Both are called with display-width budgets: diffCardBody `truncateRunes(strings.TrimPrefix(line, "+"), textBudget)` (rendering.go:1511), then diffBodyLine pads `pad := textBudget - lipgloss.Width(text)` (rendering.go:1536); toolCardHead `truncateRunes(arg, maxInt(12, width/3))` (1336) and `middleTruncate(shown, maxInt(16, width/2))` (1328); grep/read/bash card budgets (1594/1628/1711).
-```
-- **Impact:** For a diff/grep/tool line of double-width (CJK/emoji) characters, truncateRunes keeps up to `limit` runes each ~2 cells wide, so the kept text can be ~2x its display budget. In diffBodyLine the band pad then computes negative and is skipped, so the add/del tint band no longer fills to the card edge and rows of CJK code render with uneven/short solid bands. Overflow past the card width is contained because toolCard re-clamps every body line with the width-aware fitStyledLine (rendering.go:1387) and the head with fitStyledLine (1380); the residual impact is cosmetic misalignment of the diff/tool band for wide-char content, not layout breakage.
-- **Fix:** Add a display-width-aware truncate (loop accumulating lipgloss.Width per rune until budget, reserving 1 cell for the ellipsis) and use it where the argument is a width budget; or route these unstyled budgets through the existing width-aware fitStyledLine/truncateStyledLine primitives instead of rune-count truncateRunes.
-
-### L17 · Cancelled run's usage events/rows are applied to whatever session is active when the goroutine returns, even after /resume switched sessions
-- **Severity / category:** low / correctness  · **Subsystem:** tui-render
-- **Location:** `internal/tui/model.go:1319`
-- **Evidence:**
-```
-In the agentResponseMsg `msg.runID != m.activeRunID` (flushing) branch, the session-EVENT flush is correctly routed to the recorded session (`if flushSessionID == m.activeSession.SessionID { appendSessionEvents } else { appendSessionEventsTo(flushSessionID, ...) }`, model.go:1335-1338). But the usage events just above are applied unconditionally to the current model: `for _, event := range msg.usageEvents { m, usageRows = m.recordUsageEvent(...); for _, row := range usageRows { m.transcript = appendTranscriptRow(m.transcript, row) } }` (1319-1325) — no flushSessionID gating.
-```
-- **Impact:** If the user cancels an in-flight run and then /resume's a DIFFERENT session before the cancelled run's goroutine returns its accumulated response, the cancelled run's usage tracker delta and any usage transcript rows land in the now-active different session's in-memory transcript/usage readout. Narrow timing window, in-memory display only (the durable session-event flush is already correctly routed), so impact is a transient mis-attributed usage line, not persisted corruption.
-- **Fix:** Only record usage events and append usage rows when `flushSessionID == m.activeSession.SessionID`; when the run was recording into another session, persist its usage to that session's tracker (or drop the display rows) rather than mutating the active model.
-
-### I1 · Local control socket handleConn has no read deadline (owner-only mitigates)
-- **Severity / category:** info / robustness  · **Subsystem:** daemon
-- **Location:** `internal/daemon/server.go:228`
-- **Evidence:**
-```
-hello, err := ReadControl(conn)
-if err != nil { return }
-...
-cmd, err := ReadControl(conn) // both reads unbounded
-```
-- **Impact:** The locally-accepted control connection performs the hello + command reads with no deadline. Unlike the remote bridge this is gated by an owner-only Unix socket (secureSocketParent/hardenSocketFile), so the only peer able to stall it is the socket owner — the same principal that controls the daemon — making it a non-threat in practice. Noted because the same unbounded-read code is shared with the remote path (where it IS exploitable, see the medium findings); a defensive deadline here would make handleConn safe regardless of how the conn was obtained.
-- **Fix:** Optionally add a modest read deadline around the two handshake ReadControl calls in handleConn so the protection is intrinsic to the function rather than relying on the caller (local socket perms / remote bridge handshake bound).
-
-### I2 · Audit append failures are silently swallowed by the dispatcher
-- **Severity / category:** info / robustness  · **Subsystem:** hooks
-- **Location:** `internal/hooks/dispatch.go:223-249`
-- **Evidence:**
-```
-recordStarted (line 227 `_, _ = dispatcher.audit.AppendStarted(...)`) and recordCompleted (line 240 `_, _ = dispatcher.audit.AppendCompleted(...)`) discard the error. AppendStarted/AppendCompleted can fail on lock-acquire timeout (lock.go:86 "timed out acquiring audit lock" after 10s) or any I/O/marshal error.
-```
-- **Impact:** Over a session, if the cross-process audit lock is contended/stranded (a crashed holder's lock not yet 60s stale, or a permissions issue under XDG_DATA_HOME), every hook still runs and the tool proceeds normally, but the audit JSONL silently gains no record — there is no log line, metric, or diagnostic that the security/compliance audit trail dropped events. For a feature whose stated purpose is an auditable record of policy-hook executions, silent gaps undermine the trail. Informational because the dispatcher fails OPEN by design and the audit store is optional.
-- **Fix:** Surface a one-time/rate-limited warning (or a diagnostic on DispatchOutcome) when an audit append errors, so an operator can tell the audit trail is incomplete rather than assuming the absence of a record means the hook did not run.
-
-## 4. Reconciliation with prior audit docs
-
-Reconciled against `docs/audit/2026-06-10-deep-audit.md`, `2026-06-10-deep-audit-status.md`, and `2026-06-13-reverification.md`. Each prior item for an audited area was re-checked against current code. Summary: **64 fixed, 14 partial, 12 still-open** (deduplicated across overlapping subsystem reports). The still-open and partial items are the ones that matter for remediation; the 12 still-open are all re-derived above as current findings (or noted as out-of-scope dead code).
-
-#### Still open (12)
-
-| Subsystem | Prior finding | Location | Evidence |
-|---|---|---|---|
-| agent-loop | Dead stores to requestEvent after always-allow, and unreachable fallbackPermissionEvent | `internal/agent/loop.go:473` | In the AlwaysAllow case requestEvent.GrantMatched/Grant are still set at loop.go:776-778 but the model-visible event is rebuilt fresh from the sandbox decision at 830-840, and fallbackPermissionEvent (loop.go:1694) is st |
-| agent-loop | Reactive mid-stream retry never forwards retried assistant text to OnText | `internal/agent/loop.go:183` | The reactive retry still collects with only OnUsage (loop.go:264-266), deliberately omitting OnText to avoid duplicating the already-forwarded partial; documented at 258-263. Behavior unchanged (intentional trade-off). |
-| agent-loop | OnContext / MeasureContext context-utilization pipeline has no consumer on any surface | `internal/agent/types.go:183` | Grep finds no `OnContext:` assignment outside agent tests; MeasureContext is only called at loop.go:165 behind the never-set OnContext guard. Reported again above as a current low finding. |
-| cli-usage-streamjson | zero --skip-permissions-unsafe silently discards all trailing arguments (M81) | `internal/cli/app.go:149` | app.go:234-243 — `rest` is only scanned for misplaced --add-dir; all other trailing args are dropped and the TUI is launched. Confirmed still-open at 2026-06-13 reverification and unchanged in current code. |
-| config-plugins-skills | ResolvedConfig.MCP is computed but never read; provider-command MCP servers are silently dropped | `internal/config/resolver.go:95` | resolver.go:106 still sets MCP: cfg.MCP with zero readers in internal/cli (grep 'resolved.MCP' empty); ResolveMCP (resolver.go:115-133) still seeds defaults+file+overrides only and never calls LoadProviderCommand, so ZER |
-| config-plugins-skills | maxTurns<=0 is silently ignored | `internal/config/resolver.go:144,177,506` | All three merge sites still gate on `> 0` (resolver.go:152,191,539) and Resolve() still has no MaxTurns validation, unlike deferThreshold (:66-68) and maxTeamSize (:70-72); a 0/negative value is silently replaced by defa |
-| config-plugins-skills | plugin vs standalone hook-event validators disagree | `internal/plugins/plugins.go:894` | plugins.parseHookEvent (plugins.go:894) + mapHookEvent (activate.go:412) still accept only the four non-specialist events; hooks.parseEvent/IsValidEvent (hooks.go:888,906) still accept six. The sets still disagree. |
-| config-plugins-skills | config/contracts.go (ContractGap API) is unused dead code | `internal/config/contracts.go:12` | Per the 2026-06-13 reverification (line 95), ContractGap/DefaultContractGaps/FindContractGapsByMilestone are referenced only by contracts_test.go; no production change in this area since (outside my named focus but noted |
-| cron | cron run start-up reconcile silently cancels --run-now jobs | `internal/cli/cron_run.go:115` | reconcileOverdue (cron_run.go:105-130) still reschedules any NextRunAt before nowMin without firing, with no run-now exemption, and Job has no run-now marker. Reproduced: --run-now job (NextRunAt=now at add) pushed to ne |
-| hooks | Plugin/hooks event-set disagreement (plugin parseHookEvent accepts only 4 events; hooks accepts 6) | `internal/plugins/plugins.go:894 (2026-06-13 reverification l` | plugins.go:38-42 still declares only 4 HookEvent consts and parseHookEvent (plugins.go:893-898) still switches on those 4, while hooks.go parseEvent (hooks.go:901-911) accepts all 6. Re-derived above as a low finding. |
-| sandbox | OnSandboxDecision callback option is set up to fire but never wired by any caller (dead code) | `internal/tools/registry.go:108` | This field lives in internal/tools/registry.go, outside the sandbox package's audited scope; the 2026-06-13 reverification (line 107) confirms repo-wide grep still finds OnSandboxDecision only at the declaration, nil-che |
-| sessions | Second-granularity RFC3339 timestamps make Latest()/list ordering pick the wrong session on same-second ties | `internal/sessions/store.go:555` | store.go:709 still formats with time.RFC3339 (1s) and List tiebreak (332-336) is lexical SessionID; re-reported as the low finding above. |
-
-#### Partially fixed (14)
-
-| Subsystem | Prior finding | Location | Evidence |
-|---|---|---|---|
-| cli-usage-streamjson | streamjson secret patterns mangle ordinary text in every stream-json output event (M48) | `internal/streamjson/streamjson.go:310` | sk- and bearer patterns are now word-boundary + length anchored (streamjson.go:313,317) and TestRedactStringDoesNotOverMatchProse covers task-list/bearer; but the api[_-]?key pattern at streamjson.go:314 still over-match |
-| config-plugins-skills | Hooks subsystem is loaded and listed but never executed (no event dispatch, no edit command) | `internal/hooks/hooks.go:372` | Now partially wired: beforeTool/afterTool DO dispatch from the agent loop (internal/agent/loop.go:914,937) and a CLI edit surface exists (internal/cli/hooks_manage.go `zero hooks add/...`). But the other four events (ses |
-| config-plugins-skills | skills.Get/Duplicates unused; duplicate-name collisions never warned | `internal/skills/skills.go:91 + consumers` | skills.Duplicates is now consumed by the CLI list path (internal/cli/skills.go:69-81 emits per-collision warnings to stderr), so collisions ARE surfaced for `zero skills list`. But the agent-facing skill tool still drops |
-| cron | Cron store has no inter-process locking: paused/edited jobs clobbered; concurrent schedulers double-fire | `internal/cli/cron_run.go:173` | A real cross-process lock now exists (internal/cron/lock.go: O_EXCL sibling <id>.lock with atomic stale reclaim) and is taken by Update/Mutate/Remove/AppendRun (store.go:174,196,219,266). fireJob now claims-before-fire u |
-| cron | Cron Next double-fires wall-clock schedules on DST fall-back (repeated hour) | `internal/cron/schedule.go:171` | A fall-back collapse guard was added (schedule.go:191-209, sameWallClockMinute) and is covered by TestNextDSTFallBackDoesNotRepeatHour. But the guard only engages when `after` is exactly minute-aligned (schedule.go:205 ` |
-| hooks | Hooks subsystem is loaded and listed but never executed (no event dispatch) | `internal/hooks/hooks.go:372 (2026-06-10 deep audit)` | The core gap is closed for the two tool events: Dispatcher.Dispatch is wired into the agent loop at internal/agent/loop.go:914 (beforeTool, with veto routed through blockedByHookResult at loop.go:955) and loop.go:937 (af |
-| mcp | Client readLoop misroutes server-to-client requests as responses (id collision) and never answers ping | `internal/mcp/client.go:302` | client.go now has a single readLoop (331-355) that dispatches strictly by `client.pending[id]` and DROPS unmatched id-bearing messages (responses==nil) instead of corrupting a waiter, and skips id==nil notifications (338 |
-| mcp | Stdio request write blocks indefinitely while holding client.mu and ignores ctx cancellation | `internal/mcp/client.go:246` | The unbounded RESPONSE wait was fixed: mu is released at client.go:291 before the ctx-aware select on the response channel (293-311), so a hung server no longer holds the lock during the wait. But the WRITE itself (line  |
-| mcp | 1 MiB SSE scanner token cap permanently kills the SSE session on large messages | `internal/mcp/network_client.go:637` | maxSSEEventBytes raised to 8MiB (network_client.go:658) and scanner.Buffer uses it (662), plus an aggregate multi-line data cap (707-713), so the common large-message case no longer trips ErrTooLong. But a single SSE lin |
-| oauth-providers | Health probe SSRF blocklist and credentials defeated by redirects and DNS rebinding | `internal/providerhealth/providerhealth.go:285` | DNS rebinding TOCTOU is fixed: safeDialContext (providerhealth.go:442-474) re-resolves at dial time, validates every resolved addr, and dials the validated IP literal via dialValidatedAddrs. Internal-redirect escape is f |
-| oauth-providers | OpenAI adapter drops tool calls lacking an id; agent loops forever on retry | `internal/providers/openai/tool_state.go:95` | tool_state.go:235-240 closeOpen now emits StreamEventToolCallDropped for an id-less/name-less call that streamed an id/name/arguments, and loop.go:300-306 converts a dropped-call-with-no-other-calls turn into a single re |
-| sessions | No fsync before rename/append: metadata.json can be empty after a crash, silently hiding the session | `internal/sessions/store.go:632` | writeMetadata (store.go:779-804) now writes a temp file via writeFileSync (fsyncs data), os.Rename, then syncDir(parent) — metadata is fully crash-safe. BUT the sibling events.jsonl append (store.go:606-616) still has no |
-| specialist-swarm | Post-SIGTERM grace polling can SIGKILL a recycled PID | `internal/background/process_posix.go:38-50` | terminateProcess now resolves pgid up front and signals the group (-pid) only when pid leads its own group, returning nil early if Getpgid reports the leader already gone (process_posix.go:55-61); group-leader targeting  |
-| tools | escalate_model tool is implemented and registry-aware but never included in any Core* tool set | `internal/tools/registry.go:224` | Still absent from CoreReadOnly/Write/Shell/Network/CoreTools sets (registry.go:204-266) and from the TUI registry (cli/app.go:660 newCoreRegistryScoped only iterates CoreToolsScoped). It IS now registered, but only in he |
-
-#### Fixed (64)
-
-| Subsystem | Prior finding | Location | Evidence |
-|---|---|---|---|
-| agent-loop | Truncated/filtered responses (FinishReason) are produced by every provider but never consumed by the agent loop | `internal/agent/loop.go:233` | FinishReason is now consumed at loop.go:284 (result.FinishReason = collected.FinishReason) and loop.go:502/539, exposed via Result.TruncationNotice() (types.go:278-292), and surfaced to users in both internal/tui/model.g |
-| agent-loop | "Always allow" permission decision is converted into a denial when grant persistence fails (always, when Options.Sandbox | `internal/agent/loop.go:468` | PermissionDecisionAlwaysAllow case (loop.go:763-779) now sets permissionGranted=true unconditionally and only attempts persistPermissionGrant when options.Sandbox != nil, ignoring a persist error (the call is honored reg |
-| agent-loop | estimateTokens ignores image attachments and the compaction trigger ignores tool definitions, so the context budget unde | `internal/agent/compaction.go:58` | estimateTokens adds len(message.Images)*imageTokenEstimate (compaction.go:114); maybeCompact now takes the exposed tools and counts estimateToolDefTokens(tools) into both the threshold and shrink checks (compaction.go:33 |
-| agent-loop | Compaction summarizer provider calls bypass OnUsage — their token spend is invisible to usage accounting | `internal/agent/compaction.go:376` | summarizeMessagesOnce now calls CollectStreamWithOptions(ctx, stream, CollectOptions{OnUsage: onUsage}) (compaction.go:501); onUsage is threaded from options.OnUsage via newCompactionState -> summarizeClosure. |
-| agent-loop | Mid-stream context cancellation is returned as a flattened errors.New string; the ctx.Err() identity check is unreachabl | `internal/agent/loop.go:188` | loop.go:272-275 now checks `if ctx.Err() != nil { return result, ctx.Err() }` BEFORE the `if collected.Error != ""` return at 276-279, so a canceled stream returns the wrapped context.Canceled and errors.Is succeeds. |
-| agent-loop | Duplicate items-schema assignment in propertyToRuntimeMap | `internal/agent/loop.go:1072` | propertyToRuntimeMap (loop.go:1978-1980) assigns schema["items"] exactly once under `if property.Items != nil`. |
-| agent-loop | Project-guidelines truncation can split a multibyte UTF-8 rune in the system prompt | `internal/agent/system_prompt.go:89` | projectGuidelines walks back to a rune boundary before cutting: `for cut > 0 && !utf8.RuneStart(content[cut]) { cut-- }` (system_prompt.go:255-258); compaction_preserve.go capBody does the same (304-307). |
-| cli-usage-streamjson | Ctrl+C with -o json ends the JSON event stream with no error/done terminal event (M82) | `internal/cli/exec.go:500` | exec.go:567-595 emits a terminal error + done event for both execOutputStreamJSON (errorEvent+runEnd) and execOutputJSON (writeJSONLine type=error then type=done, exit_code=130) on context cancellation. |
-| cli-usage-streamjson | zero exec --list-tools -o json ignores the requested JSON format and prints plain text (M83) | `internal/cli/exec.go:182` | exec.go:206-219 branches on outputFormat: stream-json -> writeExecStreamJSONFinal, json -> writeExecToolListJSON (exec_tools.go:103), else plain text. |
-| cli-usage-streamjson | Per-event 'model' field persisted under --allow-escalation is never read by the usage report, mispricing escalated runs  | `internal/usage/report.go:17` | report.go:138-141 prefers payload.Model (set on escalation runs) before falling back to modelBySession; exec.go:555-556 writes payload["model"]=currentModel only under options.allowEscalation, and currentModel is reassig |
-| cli-usage-streamjson | One corrupt session aborts the entire zero usage report (M101) | `internal/cli/usage.go:46` | usage.go:47-51 — a ReadEvents error increments set.skipped and continues instead of returning, so a single unreadable session no longer aborts the report. (Minor: set.skipped is never surfaced to the user — info-level on |
-| cli-usage-streamjson | Fork duplicates provider_usage events, double-counting usage in zero usage report (M100) | `internal/sessions/store.go:387` | store.go:431-436 — Fork explicitly skips EventUsage when copying parent events into the fork ('Do NOT copy usage accounting into the fork'), so report aggregation over parent+fork no longer double-counts. |
-| cli-usage-streamjson | zero usage hard-fails outside a git repository (M72) | `internal/cli/usage.go:121` | usage.go:122-131 — net-LOC is best-effort: resolveWorkspaceRoot/inspectChanges errors are swallowed and DiffStat degrades to zero, so the token report still renders outside a git repo. |
-| cli-usage-streamjson | zero mcp list --json leaks MCP server env/header secret values (M74) | `internal/cli/extensions.go:186` | extensions.go:220-223 redacts via redactMCPServerConfigs (env/header values -> [REDACTED], URL creds via redactMCPURL at 236-256) and then RedactValue before writePrettyJSON; the plain-text path also runs RedactString (2 |
-| cli-usage-streamjson | redactURLPasswords recompiles its regexp on every RedactString call (M93) | `internal/redaction/redaction.go:362` | urlWithCredsPattern is now a package-level var (redaction.go:424) compiled once; redactURLPasswords (426) reuses it. |
-| cli-usage-streamjson | redaction opaque/custom-scheme Authorization not redacted (M12) | `internal/redaction/redaction.go:92` | headerPattern makes the scheme optional and captures the whole value to EOL (redaction.go:92, 186-194); opaque_auth_test.go verifies opaque tokens, custom schemes, and Proxy-Authorization are redacted while a known schem |
-| cli-usage-streamjson | secrets.Redact leaves a private key un-redacted when an inner pattern matches inside it (M90) | `internal/secrets/scanner.go:81` | scanner.go:91-99 replaces longest matches first (order sorted by len desc) so the whole PEM PRIVATE KEY block is redacted before any shorter nested match corrupts it. |
-| cli-usage-streamjson | Secret scanner misses modern OpenAI key formats / private key block header (M158) | `internal/secrets/scanner.go:35` | scanner.go:41 openai_key body is `sk-[A-Za-z0-9_-]{20,}` (covers sk-proj-/sk-svcacct- and -/_), and scanner.go:44 matches the full BEGIN..END PRIVATE KEY block. |
-| cli-usage-streamjson | Stream-json / status / tool-output truncation slices strings mid-rune (M153/M155/M165) | `internal/cli/exec_writer.go:412` | exec_writer.go:401-431 — truncateForStatus and truncateForStreamJSONOutput both call cutRuneBoundary, which walks back to the last UTF-8 rune boundary at or before n. |
-| config-plugins-skills | config writer performs a non-atomic in-place write of the user config | `internal/config/writer.go:53` | writeConfigFile (writer.go:149-184) now uses os.CreateTemp in the target dir, tmp.Chmod(0o600), Write, Close, then os.Rename(tmpPath, path) with a deferred os.Remove cleanup (lines 161-182) — the documented write-to-temp |
-| config-plugins-skills | parseThinkTags setting silently dropped on every profile merge (L17) | `internal/config/resolver.go (mergeProfile)` | mergeProfile now preserves it: resolver.go:420-422 `if next.ParseThinkTags != nil { base.ParseThinkTags = next.ParseThinkTags }`, and it is read end-to-end by providers/factory.go:97-98 (parseThinkTagsForProfile) into op |
-| config-plugins-skills | Plugin-declared skill paths are not merged into skill lookup (inert) | `internal/skills/skills.go (Load single-root)` | Plugin skills are now merged at runtime: plugins.NewSkillTool (activate.go:340) overlays default dir + plugin SkillRoots via mergeSkills, and it is registered when SkillRoots are present (cli/plugin_activate.go:53-54), i |
-| cron | cron add <expr> --recipe R silently discards the user's cron expression | `internal/cli/cron.go:112` | cronAdd now consumes the positional expression BEFORE applying recipe defaults: `if len(positional) == 1 && expr == "" { expr = positional[0] }` (cron.go:113-115) precedes the recipe block whose guard is `if expr == "" { |
-| cron | cron run records lose the failure reason: exec errors go to the discarded stdout stream | `internal/cli/cron_run.go:150` | fireJob now recovers the failure detail from the stream-json error event on stdout: `if streamErr := extractStreamJSONError(outBuf.String()); streamErr != "" { detail = streamErr }` then `rec.Error = cronTruncate(detail, |
-| cron | promptExcerpt/cronTruncate byte-slice strings mid-rune (invalid UTF-8) | `internal/cli/cron.go:275` | promptExcerpt now cuts on a rune boundary: `cutRuneBoundary(p, 47) + "…"` (cron.go:278), and cronTruncate does `cutRuneBoundary(s, max) + "…"` (cron_run.go:272) instead of the prior raw byte slice. |
-| hooks | AuditStore.append is O(n^2): re-reads and re-parses the entire audit JSONL on every append | `internal/hooks/hooks.go:503 (2026-06-10) / reverified still-` | append() no longer calls ReadEvents(); it calls lastSequence() (hooks.go:515) which os.Open + Stat + ReadAt only the trailing 8KB window (hooks.go:550-602: `const tailWindow = 8 * 1024`, `file.ReadAt(buf, start)`), parse |
-| hooks | Cross-process sequence collision on the shared audit JSONL (read-then-append not atomic across processes) | `internal/hooks/hooks.go append (implied by 2026-06-13 line 6` | append() now wraps the lastSequence()+write in a cross-process O_EXCL file lock via store.lockAudit() (hooks.go:509-513 -> lock.go:33-90), with stale-lock reclaim (lock.go:97-108) mirroring cron/oauth. TestAuditStoreSequ |
-| mcp | Unbounded Content-Length allocation lets a peer crash the whole process | `internal/mcp/protocol.go:69` | protocol.go:17 defines maxMessageBytes=64MiB; readHeaderFramed caps Content-Length (lines 121-123: `if parsed > maxMessageBytes { return ... exceeds ... limit }`) before `body := make([]byte, contentLength)` (138); readL |
-| mcp | MCP stdio transport uses LSP Content-Length framing instead of MCP newline-delimited JSON | `internal/mcp/protocol.go:42` | read() (protocol.go:54-76) now reads newline-delimited JSON by default (isJSONStart -> decodeMessage) and only falls back to LSP Content-Length framing when the first line is not JSON; write() (148-163) emits one JSON ob |
-| mcp | zero mcp list --json leaks MCP server env/header secret values | `internal/cli/extensions.go:186` | extensions.go:222 wraps cfg.Servers in redactMCPServerConfigs (236-255), which replaces every Env and Header value with `[REDACTED]` and runs redactMCPURL on the URL (252) before writePrettyJSON; redactMCPURL (mcp_tools. |
-| mcp | Child stderr captured into an unbounded bytes.Buffer for the entire server lifetime | `internal/mcp/client.go:98` | stderr is now `&boundedBuffer{cap: maxStderrCapture}` (client.go:134, cap=64KiB at line 91); boundedBuffer.Write (103-115) discards bytes past cap. (Tail-vs-head retention is a separate low finding, not unboundedness.) |
-| mcp | Streamable-HTTP SSE response decoding takes the first event instead of the matching response | `internal/mcp/network_client.go:621` | decodeSSERPCMessage (network_client.go:612-652) now skips events whose decoded message has a non-empty Method (server-initiated notifications/requests) — `if candidate.Method != "" { return true }` (636-638) — and return |
-| oauth-providers | Gemini adapter never reads cachedContentTokenCount; CachedInputTokens always 0 | `internal/providers/gemini/types.go:88` | gemini/types.go:107 now declares `CachedContentTokenCount int \`json:"cachedContentTokenCount"\``; provider.go:225 `state.cachedTokens = payload.UsageMetadata.CachedContentTokenCount` and provider.go:283 emits `CachedInp |
-| oauth-providers | HTTP 529 classified as rate-limit but excluded from the 429/503 retry policy | `internal/providers/providerio/retry.go:107` | retry.go:108-110 `func ShouldRetryStatus(code int) bool { return code == http.StatusTooManyRequests // code == http.StatusServiceUnavailable // code == 529 }` — 529 now retried; doc comment updated to include Anthropic ' |
-| sandbox | macOS sandbox-exec profile blocks /tmp and /dev/null, breaking most wrapped shell commands | `internal/sandbox/runner.go:262` | runner.go:412-436 sandboxWritableDevices (/dev/null,/dev/zero,/dev/stdin,...) and sandboxWritableSubpaths (/tmp,/private/tmp,/var/tmp,/var/folders,/dev/fd) are now emitted by seatbeltWriteRule (runner.go:559-581, AllowTe |
-| sandbox | apply_patch misparses unified-diff body lines starting with '--- '/'+++ ' as file paths | `internal/tools/apply_patch.go:161` | risk.go:291-327 patchHeaderPaths now tracks hunk state (inHunk, oldRemaining/newRemaining from parsePatchHunkCounts) and decrements per body line, so '--- '/'+++ ' header parsing at risk.go:319-324 is only reached OUTSID |
-| sandbox | Interactive-command detector flags pager/REPL names inside quoted arguments (splitShellSegments not quote-aware) | `internal/sandbox/safe_command.go:138` | safe_command.go:474-583 splitShellSegments is now a full quote/substitution-aware scanner (inSingle/inDouble state, substStack for $()/backtick, backslash escapes); operators inside quotes are literal. Detection also anc |
-| sandbox | Grant store prefix/substring safety: Lookup trims while stored writer key may not be normalized (trim asymmetry => load  | `internal/sandbox/grants.go:146` | grants.go:556-568 readState rebuilds a `normalized` map keyed by `key := strings.TrimSpace(name)` and re-keys every bucket; normalizeStoredGrant(key, grant) compares against the trimmed key (grants.go:614). Lookup/Revoke |
-| sandbox | Destructive/network/installer classification recompiles N regexes per shell call | `internal/sandbox/risk.go:53` | risk.go:12-55 destructiveCommandPattern, pipedInstallerPattern, unparseableNetworkPattern, and destructiveExtraPatterns are all package-level regexp.MustCompile vars compiled once at init, not per call. |
-| sessions | Torn append (crash mid-write) permanently bricks a session: ReadEvents hard-fails on any malformed JSONL line | `internal/sessions/store.go:547` | store.go:677-704 now computes tornTailPossible (trailing byte != newline) and lastNonEmpty, and breaks (drops the partial record) only when the malformed line is the last non-empty line AND the file lacks a trailing newl |
-| sessions | Fork duplicates provider_usage events and rewrites their timestamps, double-counting and re-dating usage in `zero usage  | `internal/sessions/store.go:387` | Fork (store.go:430-442) explicitly `if event.Type == EventUsage { continue }` so usage events are never copied into the fork; usage/report.go BuildReport iterates every session's events (cli/usage.go collectUsageData), s |
-| sessions | `zero sessions tree` is O(nodes x sessions) disk reads (full store re-list per tree node) | `internal/sessions/lineage.go:140` | Tree (lineage.go:127-167) now does a single store.List() up front, indexes byID and childrenByParent in memory, and treeFrom (169-189) recurses over the in-memory maps with no per-node disk scan. |
-| sessions | Naive byte-slice truncation can split multi-byte UTF-8 runes in prompts and titles (replay) | `internal/sessions/replay.go:237` | replay.go:382-393 cutPromptRuneBoundary backs up while !utf8.RuneStart(s[n]) before slicing; it is used at the prompt-truncation sites (310-314) and payloadPreview (376), so replay/preview truncation is rune-safe. |
-| sessions | execSessionRecorder.append latches AppendEvent errors into recorder.err but no caller reads it (silent session-persisten | `internal/cli/exec_sessions.go:116-124` | warnIfRecordingFailed (exec_sessions.go:128-132) now surfaces recorder.err to stderr and is deferred by both runExec (exec.go:464) and the spec-draft path (exec_spec.go:90). |
-| specialist-swarm | Specialist sessions launched without a description are unresumable (AgentName never recorded / recorded as garbage) | `internal/specialist/exec.go:166` | exec.go:250-255 now always emits --session-title: when Description is empty it passes the bare manifest name (`args = append(args, "--session-title", name)`), so specialistAgentName resolves to the real specialist name a |
-| specialist-swarm | Specialist children always run with --auto high (PermissionModeUnsafe): one Task approval silently grants unprompted she | `internal/specialist/exec.go:150` | specialistAutonomy() (exec.go:125-132) maps only an explicit 'unsafe' parent mode to 'high' and everything else (including empty) to 'low'; BuildArgs/BuildResumeArgs (exec.go:230,282) call it with input.PermissionMode. T |
-| specialist-swarm | Race: duplicate specialist stop/usage accounting events — check-then-append dedupe is not atomic | `internal/specialist/accounting.go:72` | recordSpecialistStop and appendSpecialistUsageRollup now route through appendSpecialistEventOnce → Store.AppendEventUnlessExists (accounting.go:60,105,160-172), which performs the existence check and the append atomicall |
-| specialist-swarm | Foreground specialist run fails entirely when a child stream-json line exceeds 1 MiB | `internal/specialist/streamer.go:43` | ParseStream (streamer.go:41-73) now uses bufio.NewReader + ReadString('\n') with no per-line cap, and the live foreground path runChildProcess (exec.go:735-759) likewise uses bufio.NewReader.ReadString. The 1 MiB Scanner |
-| specialist-swarm | SetPID failure path deletes the prompt file under a running child and records no terminal accounting/status | `internal/specialist/exec.go:346` | runBackground SetPID-failure branch (exec.go:451-466) now terminates the orphaned child (background.TerminateProcess(pid), joining any kill error), calls manager.UpdateStatus(StatusError) and recordSpecialistStop, in add |
-| specialist-swarm | POSIX background-task kill terminates only the leader PID; the specialist's subprocess tree leaks on SIGKILL escalation | `internal/background/process_posix.go:50` | launchBackgroundProcess now calls background.ConfigureChildProcessGroup(command) before Start (exec.go:791), which sets Setpgid (process_posix.go:24-29), and terminateProcess signals the negative PID (the whole group) wh |
-| specialist-swarm | formatTaskOutput is dead code | `internal/specialist/output_tool.go:244` | grep across the repo finds no `formatTaskOutput(` definition or call; the only formatter in use is formatTaskOutputSummary, called from readOutput (output_tool.go:154). The dead function was removed. |
-| tools | bash tool timeout does not kill the command when a child inherits the output pipes | `internal/tools/bash.go:104` | bash.go:119 now calls hardenProcessLifetime(command); bash_proc_unix.go:28-44 sets Setpgid + WaitDelay=bashWaitDelay(2s) + Cancel that SIGKILLs the whole negative-pid process group, so a backgrounded child can no longer  |
-| tools | grep glob filter is applied to root-relative paths, silently dropping matches under a subdirectory search | `internal/tools/grep.go:224` | grepFiles now matches the glob against the path relative to the SEARCH directory: grep.go:307-311 computes `globPath = filepath.ToSlash(rel)` via `filepath.Rel(target, path)` and matches globMatcher against that, so `glo |
-| tools | OnSandboxDecision callback option is set up to fire but is never wired by any caller (dead code) | `internal/tools/registry.go:108` | registry.go no longer declares or invokes OnSandboxDecision anywhere; RunOptions (registry.go:15-43) has no such field and RunWithOptions (registry.go:93-171) propagates decisions via res.SandboxDecision instead. The dea |
-| tools | read_file returns Truncated but no truncation marker in output; rune-width line numbering can misalign | `internal/tools/read_file.go:89` | read_file.go:118-123 now appends an explicit marker on truncation: `[truncated: N more line(s) in the requested range not shown; set start_line=... to continue]`. Line-number padding uses ASCII strconv.Itoa width (read_f |
-| tools | lsp_navigate workspace confinement (recently fixed) — verify | `internal/tools/lsp_navigate.go` | lsp_navigate.go:85 resolves the model path via resolveScopedReadPath BEFORE reading or handing it to the LSP manager (line 90 readWorkspaceFile(absPath), line 97 req.Path=absPath), so a '..'/absolute path cannot open a f |
-| tui-render | wrapPlainText collapses internal whitespace, destroying code indentation/alignment in assistant answers | `internal/tui/rendering.go:261 (also 2026-06-13 reverify rend` | rendering.go:399-409 now detects preformatted/columnar bodies (an internal run of >=2 spaces) and emits them verbatim via splitPreservingWidth (459-479) instead of strings.Fields, preserving every space; leading indent i |
-| tui-render | appendTranscriptRow O(n) copy + O(n) dedup scan per row → O(n^2) append and /resume rehydration | `internal/tui/transcript.go:113/117 (multiple rows)` | appendTranscriptRow now appends in place (transcript.go:99-108, comment notes the old O(n^2) is gone). The prior reverification gap — hasTranscriptRow still linear-scanning per keyed row during rehydration — is closed fo |
-| tui-render | Full transcript re-rendered from scratch every frame, including per-line regex parsing | `internal/tui/model.go:610/124` | renderRowMode/renderRowDetailed now memoize per-row output via the LRU defaultRenderCache (rendering.go:172-197) keyed on width/flush/bodyCap/cwd/row fields/rc hints (render_cache.go:126-174); only rows tied to the activ |
-| tui-render | looksLikeDiff false-positives on any output containing a line starting with '---' | `internal/tui/view.go:394` | view.go:897-916 now requires a real hunk header (hunkHeaderPattern) OR BOTH '+++ ' and '--- ' file headers before returning true; a lone '---' line no longer hijacks bash/generic output into the diff renderer. |
-| tui-render | truncateTUIOutput byte-slices UTF-8 mid-rune, emitting invalid UTF-8 | `internal/tui/transcript.go:303` | truncateTUIOutput (transcript.go:436-445) now calls cutRunes (449-460), which walks back to the nearest UTF-8 rune boundary via utf8.RuneStart before slicing, so the transcript/session log can no longer receive a split m |
-| tui-render | Cancelled-run event flush appends to whichever session is active at flush time (cross-session contamination) | `internal/tui/model.go:512` | flushRunIDs now maps runID -> recording sessionID (model.go:161, set at cancel 3177-3180), and agentResponseMsg routes the durable flush to that session: `if flushSessionID == m.activeSession.SessionID { appendSessionEve |
-| tui-render | Transcript dedupe key omits runID — repeated provider tool-call IDs silently drop later tool rows | `internal/tui/transcript.go:135/137/139` | transcriptRowKey bakes runID into every key, e.g. tool rows return fmt.Sprintf("%d:%d:%s", row.kind, row.runID, row.id) (transcript.go:157-181), with a comment explaining Gemini's repeating synthetic gemini_tool_N ids; w |
-| tui-render | Resize unhandled beyond storing width/height; composer not width-bounded (blind typing) | `internal/tui/model.go:456/227` | tea.WindowSizeMsg (model.go:1249-1267) stores width/height, resets the streaming fade line-age mapping that a re-wrap would invalidate (1256-1257), sizes the composer for horizontal scroll `m.input.SetWidth(maxInt(20, ch |
-### 4.4 Findings dropped in adversarial review
-
-Two drafted findings did not survive re-verification and are explicitly withdrawn:
-
-| Subsystem | Claimed | Why dropped |
-|---|---|---|
-| agent-loop | Compactor context-window budget frozen at run start, not updated on mid-run model escalation | Mechanics accurate, but the user-visible failure mode is unreachable today — the escalation path that would change the window does not feed back into the frozen budget in a way that produces wrong behavior. Not a defect. |
-| config-plugins-skills | `maxTurns <= 0` silently dropped with no validation error | Disproven: a 0/negative `maxTurns` is replaced by the documented default (30) by design, not silently mis-handled; there is no incorrect behavior. (The "no explicit validation error" observation is retained only as the low-priority still-open consistency note, not as a defect.) |
-
-## 5. Root-cause synthesis
-
-The 35 surviving findings cluster into four underlying causes:
-
-**A. Pattern/substring matching on un-structured text where a structure-aware path exists but isn't authoritative.** The sandbox destructive/piped-installer classifier scans the *raw* command string (M11) even though a quote-aware AST analyzer runs alongside — but the analyzer can only *add* risk categories, never retract a regex false positive. The stream-json `api_key` redactor (M1) matches "any next word" after the marker. Both over-match real content (a `git commit -m "…rm -rf…"`, prose containing `apiKey foo`). Fix posture: make the structured parser authoritative when the parse succeeds; anchor patterns to credential/command shape.
-
-**B. "Success" returned before the durable/drain step completes.** `AppendEvent` returns success after a page-cache write with no `fsync`, while the *derived* metadata is fsync'd — so a crash can leave `EventCount` ahead of the durable log and lose the last event, including the checkpoint `/rewind` targets (M12). A PTY exec session's `command.Wait()` returns and the session is removed before the fire-and-forget copy goroutine has drained the master FD, dropping the command's final output (M14). The hook dispatcher discards audit-append errors (I2). Fix posture: tie the success signal to completion of the durable write / output drain.
-
-**C. Surfaces shipped ahead of (or drifting from) their wiring.** Four hook events are validated, documented, and `zero hooks add`-able but never dispatched (H1); `ResolvedConfig.MCP` is computed but unread so provider-command MCP servers are dropped (M2); usercommands `model:`/`agent:` frontmatter is parsed/documented but ignored (L4); project slash-commands auto-load with a built-but-unused provenance flag (M3); `OnContext`/`MeasureContext` has no consumer (L2); plugin vs hooks event-set validators disagree (L8); the runtime skill tool drops dup-name collisions silently (L5). Fix posture: a single source of truth for valid events; either wire the surface or remove it from config/CLI/docs.
-
-**D. Lifecycle-edge handling asymmetric with the happy path.** Idle/close/exhausted/redirect edges are under-handled: the remote bridge clears the conn deadline before the daemon handshake so an authenticated-idle peer pins a slot forever (M7) and `Shutdown` can't even close bridge conns because `ServeConn` bypasses `trackConn` (M8); the MCP SSE client never reconnects after any stream close (M9); cron double-fires on DST fall-back (M4), double-grants a claim when the schedule can't advance (M5), and silently cancels `--run-now` on reconcile (M6); the swarm scheduler leaks a per-job context on `MaxRuns` completion (L13); the mailbox stale-lock break compares the file to itself instead of the observed-stale snapshot (M13); and the health probe forwards credentials across a cross-host redirect (M10 — the security item, an edge of the same "redirect handling" family the DNS-rebind fix addressed for the internal case but not the cross-host-public-credential case).
-
-## 6. Prioritized remediation order
-
-1. **M10 — strip auth headers on cross-host redirect in the health probe** (only security finding; credential exfil to a public host).
-2. **H1 — wire or remove the four un-dispatched hook events** (trust/correctness; silent no-op of a documented automation surface).
-3. **M12 — `fsync` the events.jsonl append** (data-loss of the last event, including `/rewind` checkpoints).
-4. **M11 — make the sandbox AST authoritative / segment the destructive regexes** (false-positive hard-blocks safe commands in headless exec).
-5. **M14 — synchronize the PTY copy goroutine with session exit** (model loses a command's final output).
-6. **M4 / M5 / M6 — cron: DST double-fire, claim double-grant on unadvanceable spec, `--run-now` reconcile** (duplicate/again missed scheduled agent runs = token spend / broken contract).
-7. **M7 / M8 / M9 — connection lifecycle: bridge handshake deadline, `ServeConn` tracking for Shutdown, MCP SSE bounded reconnect** (slot exhaustion / lost MCP tools mid-session).
-8. **M1 — anchor the stream-json `api_key` redactor** (corrupts machine-readable output).
-9. **M2 / M3 / M13 / M15 — provider-command MCP drop, project-command provenance, mailbox stale-break, TUI streaming O(n²)**.
-10. **Lows / infos** as hygiene — notably **L15** (genuine `lastUsedAt` data race; add a `-race` test), **L1** (route the three remaining stream connects through `streamWithReconnect`), and the dead/inert cleanups (L2/L4/L5/L8).
-
-## 7. Confidence notes (what could not be fully verified)
-
-- **`-race` suite is not green:** 3 `internal/tools` exec-session tests fail under `-race` due to fixed-timing yield-window assumptions (characterized in §2 as flakes, confirmed by passing 3× without `-race`). I could not make those tests deterministic without modifying source (audit only). No data race underlies them; the one real race in that package (**L15**) was found by inspection because no test exercises concurrent `touch()`+prune.
-- **Platform-specific backends only skimmed:** Windows sandbox/process files (`*_windows.go`), `landlock_linux.go`, `seccomp_*.go`, `log_monitor_darwin.go`, and the OS keyring backends (`secret_access_*.go`) were read shallowly — outside the symlink/parsing/risk/locking focus. A dedicated Windows/Linux-LSM pass is advisable.
-- **Not deep-read (out of stated focus):** `internal/tools/web_fetch.go` & `web_search.go` (network tools), Anthropic provider internals, the plugin/skill **install/checksum/lockfile** paths, and TUI overlays (`mouse.go`, picker/wizard/onboarding, `command_center.go` specifics). Build/vet cover them; behavior was not audited.
-- **Daemon prior-audit reconciliation is necessarily empty:** the `D1–D11` labels referenced in current daemon code comments come from a review not present in the three on-disk `docs/audit/*.md`, so "fixed since" for the daemon is against code comments, not those docs. The daemon findings here (M7/M8/L7/I1) are re-derived against current code.
-- **Reproductions:** M4, M5, M6, M11 were empirically reproduced with throwaway tests against the current tree (then removed). M10, M12, M14, M9, H1 are verified by code reading + caller/grep tracing, not a live exploit. Severity calibration for M10 is *medium* (requires a malicious/compromised/typo'd `baseURL` that 3xx-redirects to an attacker host) rather than high; if a deployment commonly points providers at untrusted gateways, treat it as high.
-
----
-
-## 8. Remediation status — `fix/audit-2026-06-20`
-
-Implemented in staged commits on branch `fix/audit-2026-06-20` (one finding/tight-cluster per commit; each gated build/vet/gofmt + `-race` on the affected package; full `-race` suite green on the final tree). Each finding was re-confirmed against current code before fixing.
-
-### Fixed (15 findings + 1 test-robustness)
-
-| ID | Sev | Status | Note |
-|---|---|---|---|
-| M10 | medium | ✅ Fixed | strip auth headers (x-api-key/x-goog-api-key/custom) on cross-host redirect in the health probe |
-| M14 | medium | ✅ Fixed | join the PTY copy goroutine before markDone/remove so a command's final output isn't dropped on exit (also fixes the flaky `TestExecCommandTTYSessionAcceptsInputOnLinux`) |
-| M12 | medium | ✅ Fixed | fsync events.jsonl append so the log is as durable as its metadata |
-| M4 | medium | ✅ Fixed | feed `sched.Next` a minute-aligned instant so the DST fall-back collapse guard engages |
-| M5 | medium | ✅ Fixed | pause an unadvanceable job inside the claim so two schedulers can't both fire it |
-| M7 | medium | ✅ Fixed | bound the daemon control handshake with a read deadline (idle remote peer can't pin a slot) |
-| M8 | medium | ✅ Fixed | `ServeConn` tracks the conn so `Shutdown` can close a stalled remote handshake |
-| M13 | medium | ✅ Fixed | atomic rename-with-verify mailbox stale-lock break (no self-compare; no split-brain) |
-| M1 | medium | ✅ Fixed | anchor the streamjson `api_key` redaction so prose isn't mangled |
-| I1 | info | ✅ Fixed | local control socket handshake now deadline-bounded (same fix as M7) |
-| L15 | low | ✅ Fixed | snapshot exec-session `lastUsedAt` under its lock — fixes a real `-race` data race |
-| L13 | low | ✅ Fixed | `defer job.cancel()` releases the scheduler's per-job context on MaxRuns completion |
-| L1 | low | ✅ Fixed | route the post-compaction / max-turns connects through `streamWithReconnect` |
-| L3 | low | ✅ Fixed | `--skip-permissions-unsafe` rejects trailing positional args instead of dropping them |
-| L11 | low | ✅ Fixed | error on a present-but-wrong-shape chatgpt account-id claim so the login warning fires |
-| (report §2) | — | ✅ Fixed | de-flaked the 3 exec-session timing tests so `go test ./... -race` is green |
-
-### Deferred (larger/riskier than a clean stage — rationale)
-
-| ID | Sev | Why deferred |
-|---|---|---|
-| H1 | high | Wiring `sessionStart/End` + `specialistStart/Stop` correctly is a cross-surface design change: the dispatcher lives only in `agent.Options` (fires per-tool), so session boundaries differ across exec/TUI/daemon and the specialist events need the parent dispatcher threaded through the child boundary. The audit's alternative (delete the events) would disable a documented feature (against the guardrails). Needs a focused design pass, not a surgical edit. |
-| M9 | medium | MCP SSE reconnect is risky concurrency (re-open + re-`initialize`d session state) and not deterministically unit-testable; warrants its own change with a fault-injection harness. |
-| M11 | medium | Making the sandbox AST authoritative is security-sensitive — a wrong call under-blocks a genuinely destructive command. Needs an extensive "still blocks real destructive commands" matrix before flipping the precedence; the guardrails forbid weakening confinement to land a fix. |
-| M2 | medium | Merging provider-command MCP into `ResolveMCP` is a real wiring decision (which resolve path the runtime should read); deferred to avoid guessing the intended single source of truth. |
-| M3, M6 | medium | Project-command provenance marker (UX) and cron `--run-now` one-shot marker (schema field) are design choices better made deliberately. |
-| M15 | medium | Streaming-block memoization is perf-only (no correctness risk) and fiddly in the TUI render hot path. |
-| L2, L4, L5, L6, L7, L8, L9, L10, L12, L14, L16, L17, I2 | low/info | Lower-value hygiene / cross-cutting (dead-code removal decisions, cross-platform lock identity, RFC3339Nano format change, TUI width math, minor concurrency hardening). Batchable in a follow-up. |
-
-### Skipped (not a real defect)
-- The two findings the audit itself dropped in §4.4 (compactor frozen-budget; `maxTurns<=0`) were already withdrawn and not implemented.

docs/audit/2026-06-21-product-ux-audit.md

@@ -1,444 +0,0 @@
-# Zero — Product / UX / TUI Audit (2026-06-21)
-
-- **Target:** `github.com/Gitlawb/zero` @ `origin/main` commit `f7ee189` — a terminal coding agent (Go, Bubble Tea v2 TUI; one event stream feeding TUI / headless `exec` / MCP server / cron).
-- **Lens:** product, UX, TUI/visual correctness, and open-source-release readiness, judged from a **stranger who just cloned the repo**. The code/security audit is separate (`docs/audit/2026-06-20-deep-audit.md`) and is **not** repeated here.
-- **Scope note:** audit only — no source changed.
-
-## 1. Executive summary
-
-Zero is a **functionally mature** agent with a thoughtfully-built TUI (memoized rendering, width tiers, light/dark themes, a real reduce-motion opt-out, color-independent diff/permission encoding, and a genuinely good guided setup wizard). The gap is the **release/packaging and last-mile UX layer**, and it is not yet shippable as open source.
-
-Two **release blockers**: there is **no `LICENSE` file** (the repo is legally all-rights-reserved despite being marketed as open source), and the **headline binary-install path is dead** — `scripts/install.sh` / `install.ps1` resolve `Gitlawb/zero`'s GitHub Releases, which 404 (no release has ever been published), so the only working install is `go run` — which then trips a **Go-version mismatch** (README says 1.24, `go.mod` requires 1.25). A new user on a machine without Go 1.25 cannot get to first run.
-
-Beyond that, the recurring themes are **docs/claims drifting from the code** (a `/usage` slash command and a `--theme` flag that don't exist; a keyless local-model on-ramp the health probe blocks), **error messages that don't guide recovery** (no-key errors dump a raw OpenAI URL instead of pointing at `zero setup`; `doctor` reports "Overall: pass" with no credential and leaks Go `map[...]` syntax), and **low-contrast theme tokens carrying real content** (faint/faintest fail WCAG while holding line numbers, diff headers, and help text).
-
-### Severity rollup (35 findings, post-dedup)
-
-| Severity | Count |
-|---|---|
-| Critical | 2 |
-| High | 11 |
-| Medium | 11 |
-| Low | 5 |
-| Info | 6 |
-
-(Counts merge cross-area duplicates — LICENSE, install-404, Go-version, and the no-provider-setup-pointer cluster each appeared from multiple area finders.)
-
-## 2. Environment & method
-
-- **Build:** `go build ./cmd/zero` → clean on go1.26.4; binary run as `/tmp/zero` with **isolated config** (`HOME`/`XDG_CONFIG_HOME` → temp dirs) so findings reflect a fresh machine, not local state.
-- **Ran (real output captured):** `zero --help`, `doctor`, `exec` (no provider / bad key / unreachable URL / empty prompt), `setup <provider> --verify` (no key / wrong key, openai + ollama), `providers check --connectivity`, `config`, `version`; bad-flag / bad-mode / bad-effort / malformed-config error paths; the existing TUI visual tests (`command_card_contrast`, `theme_select`/luminance, `width_tiers`, `wrap_whitespace`, `theme_reprobe`) — all pass; throwaway Go tests measuring display-cell widths of the truncation helpers with CJK input; WCAG contrast ratios computed from the theme hex tokens; a `creack/pty` capture of a monochrome onboarding frame to diff SGR codes across `NO_COLOR` values.
-- **Terminal profiles:** color/`NO_COLOR`/profile behavior was checked via the `colorprofile` dependency + one pty capture; **truecolor/256/16-color and a physically light terminal were not driven frame-by-frame** (see Confidence notes).
-- **Could not capture:** live interactive TUI frames — Bubble Tea v2's alternate-screen buffer yields only escape setup/teardown to `script`/`expect`. Rendered first-run wizard, resize redraw, streaming fade, and narrow-tier painting are assessed **from code + the passing visual tests** and marked unverified where relevant.
-- **Adversarial pass:** every finding faced 1–2 skeptics that re-ran/re-read to disprove it; 2 were dropped (see §A.dropped). Several of my own first-pass guesses were corrected (exit codes are correct; meaning never depends on color alone; a reduce-motion opt-out does exist).
-
-## 3. "First 5 minutes" — new-user walkthrough & verdict
-
-1. **Clone, read README.** Strong pitch ("terminal coding agent you fully own", "no telemetry"). Two install options offered as peers: a binary one-liner and `go run ./cmd/zero`.
-2. **Try the binary one-liner** (`scripts/install.sh`) — the natural choice without Go. → **Dead-ends** with a curl 404 / "could not read tag_name" (no GitHub Release exists). [C1]
-3. **Fall back to `go run ./cmd/zero`** per the README ("requires Go 1.24+"). On Go 1.24 → **build error** `go.mod requires go >= 1.25.0`. [H2] (Works on 1.25+.)
-4. **First launch with no key.** Headless `exec` prints a clear, exit-coded error — but it names only `OPENAI_MODEL`/`OPENAI_API_KEY` and a raw `platform.openai.com` URL, **never `zero setup` or `zero auth`** (the commands built for this moment). [H/M cluster] The guided `zero setup` wizard itself is excellent [I1] — if the user discovers it.
-5. **Verify setup** (`zero setup openai --verify`) with no key yet → "**The provider rejected the API key**" — but no key was sent. [M1] Or try the README's headlined **keyless local-model path** (Ollama at localhost) and run `doctor`/`--verify` → "**localhost hosts are blocked**" + "make sure the server is running" (self-contradictory). [H1]
-6. **Run `zero doctor`** to sanity-check → "**Overall: pass**" even with no provider credential [H], and an unreadable `missing: map[gopls:...]` Go-map blob [H].
-
-**Verdict: a stranger cannot reliably reach a working first session in 5 minutes.** Not because the agent is weak — it's the install/license/version/onboarding-error plumbing around it. Once a user is past setup with a real key, the core loop and TUI are solid.
-
-## 4. Findings by area
-
-### A. First-run & onboarding
-
-#### C1 · Both README binary-install paths point at a repo that 404s — `scripts/install.sh` and `install.ps1` dead-end for every new user
-
-**Status (2026-06-21):** Fixed (in-repo) — `622e4b4`. README/INSTALL demote the binary path to "coming soon" and lead with `go run`; release-artifacts.yml now publishes a GitHub Release on a `v*` tag. Publishing the release + making the repo public are MANUAL maintainer steps (repo owner/name already matched go.mod).
-- **Severity:** critical · also: G-oss-readiness
-- **Where:** `README.md:55-57, docs/INSTALL.md:52, scripts/install.sh:4 (ZERO_REPO default Gitlawb/zero)`
-- **Evidence:** curl -s -o /dev/null -w '%{http_code}' https://api.github.com/repos/Gitlawb/zero -> 404. /releases, /releases/latest, /tags all return {"message":"Not Found","status":"404"}. install.sh:106 resolves api_url="${ZERO_GITHUB_API%/}/repos/${ZERO_REPO}/releases/latest" then install.sh:111 `[ -n "$tag" ] || fail "could not read tag_name from $api_url"`. With curl --fail (install.sh:80) the API call itself errors out first. README presents this as the 'or install a release binary' path alongside go run.
-- **Impact:** A stranger who picks the README's headline binary-install path (the natural choice on a machine without Go) hits an immediate, opaque failure ('could not read tag_name' or a curl --fail HTTP error). There is no published release, so the documented one-liner cannot succeed for anyone. Only the `go run ./cmd/zero` path works.
-- **Fix:** Publish at least one GitHub release with the documented archive+.sha256 assets under the correct public repo, and fix the ZERO_REPO default / README links to that repo's real owner/name. Until releases exist, demote the binary-install path in README/INSTALL to 'coming soon' and lead with `go run ./cmd/zero`.
-
-#### C2 · No LICENSE file; README says license 'being finalized' — OSS-release blocker (confirming established fact, still true @ f7ee189)
-
-**Status (2026-06-21):** Skipped — license choice deferred by the maintainer ("leave the license portion now"). LICENSE / README §License / package.json license remain a MANUAL step.
-- **Severity:** critical · also: G-oss-readiness
-- **Where:** `README.md:330-332; repo root (no LICENSE* file)`
-- **Evidence:** `ls LICENSE*` -> no matches. README.md:332: 'License is being finalized; a `LICENSE` file will be added before the public release.'
-- **Impact:** A new user cloning the repo has no legal grant to use, modify, or redistribute. For an 'open-source' project this blocks adoption/redistribution and is a release blocker.
-- **Fix:** Add a LICENSE file (e.g. Apache-2.0/MIT) and update README §License to name it.
-
-#### H1 · Local-model verification/health commands fail with a misleading 'localhost hosts are blocked' error — breaks the README's headlined keyless path
-
-**Status (2026-06-21):** Fixed — `24c8de3`. Loopback allowed only for a user-configured local base_url; redirects + other private ranges stay blocked.
-- **Severity:** high
-- **Where:** `internal/providerhealth/providerhealth.go:374-377 (also :81,:94 loopback prefixes); surfaced by `zero setup <local> --verify`, `zero providers check <local> --connectivity`, `zero doctor``
-- **Evidence:** `zero setup ollama --verify` -> `[zero] setup verification failed: Could not reach the provider endpoint. Check the base URL ... and that the server is running. (provider connectivity URL is unsafe: localhost hosts are blocked)`. `zero providers check ollama --connectivity` -> status: fail / `provider.connectivity: provider connectivity URL is unsafe: localhost hosts are blocked` / exit 3. validateEndpoint() hard-rejects normalized=='localhost' and 127.0.0.0/8 + ::1. Guard is confined to providerhealth probe path (grep shows blockedAddrReason/endpointSafetyError only in providerhealth + unrelated web_fetch.go) so actual `zero exec` chat again
-- **Impact:** README §Why Zero and Quick start headline 'Local models need no key at all' as the easiest on-ramp. A new user who runs Ollama/LM Studio at localhost:11434 and then runs setup --verify / doctor / providers check (all suggested) sees FAIL with a self-contradictory message ('make sure the server is running' + 'localhost is blocked'), and the 'next' hint tells them to re-run the exact command that will keep failing. The keyless p
-- **Fix:** Allow loopback/localhost in the provider connectivity probe for explicitly user-configured local providers (the SSRF guard is meant for fetched/remote URLs, not the user's own configured base_url). At minimum, special-case local provider kinds to skip the loopback block, and drop the contradictory 'server is running' wrapper when the real cause is the loopback policy.
-
-#### H2 · go.mod requires `go 1.25.0` (+ toolchain go1.26.4) but README badge and Quick start say Go 1.24 — a Go 1.24 user cannot build from source
-
-**Status (2026-06-21):** Fixed — `622e4b4`. README badge + quick start now say Go 1.25+.
-- **Severity:** high · also: G-oss-readiness
-- **Where:** `go.mod:3 (`go 1.25.0`), go.mod:4 (`toolchain go1.26.4`); README.md:16 badge 'Go-1.24', README.md:52 'requires Go 1.24+'`
-- **Evidence:** go.mod: `go 1.25.0` / `toolchain go1.26.4`. README.md:16 `![go](https://img.shields.io/badge/Go-1.24-...)`, README.md:52 `# run from source (requires Go 1.24+)`. A Go 1.24.x toolchain refuses a module declaring `go 1.25.0` unless auto-toolchain download is enabled (GOTOOLCHAIN=auto + network); offline/pinned 1.24 users get a 'go.mod requires go >= 1.25.0' build error.
-- **Impact:** The README explicitly invites Go 1.24 users to `go run ./cmd/zero` (the ONLY working install path given the missing releases). They hit a build failure on the documented version, with no hint that 1.25 is actually required. Builds confirmed working only on go1.26.4 in this env.
-- **Fix:** Make the docs match go.mod: change the badge and Quick-start note to 'requires Go 1.25+'. (Or, if 1.24 support is intended, lower the go.mod directive to 1.24 and verify it compiles.)
-
-#### M1 · `zero setup <provider> --verify` with NO key set reports 'The provider rejected the API key' — but no key was ever sent
-
-**Status (2026-06-21):** Fixed — `42b7123`. `--verify` distinguishes "no API key found" from "rejected".
-- **Severity:** medium
-- **Where:** `internal/cli/setup.go:59-77 (verify branch); verified via runtime`
-- **Evidence:** With OPENAI_API_KEY unset: `zero setup openai --verify` -> '... setup verification failed: The provider rejected the API key. Double-check the key value (and that it is for this provider), then re-run setup. (Provider openai requires API credentials.)' — identical message to the WRONG-key case (OPENAI_API_KEY=sk-wrongkey123).
-- **Impact:** Confusing for a first-timer who ran setup without exporting a key yet (a very common ordering): the tool says the key was 'rejected'/'double-check the key value' when the real problem is there is no key at all. Sends them debugging a key they never set.
-- **Fix:** Distinguish the missing-credentials case from the rejected-credentials case: when no key/env is present, say 'No API key found — export OPENAI_API_KEY (or pass --api-key-env) then re-run', and reserve 'rejected the key' for actual 401s.
-
-#### I1 · The guided setup wizard copy itself is well-written and self-explanatory (source-level; live frames unverified)
-
-**Status (2026-06-21):** Info — no action (positive finding).
-- **Severity:** info
-- **Where:** `internal/tui/onboarding.go:1191-1650`
-- **Evidence:** Strings include 'Welcome to Zero' / 'A terminal agent for changing real code.' (1202-1204), a Safety panel 'Zero asks before running shell commands or changing files. Unsafe mode stays off unless you explicitly enable it. Default: ask before risky work.' (1191-1196), clear step headers ('Choose a provider', 'Choose a model', 'Paste your <name> API key' with 'Saved keys stay in your user config' / 'Blank uses <ENV> from your shell'), model-ID examples (1311), and a consistent footer ('↑/↓ choose  Enter continue  q quit'). Non-interactive `zero setup <p>` output is also good: prints config path + concrete 'next:' steps + a runnable 'try this:'
-- **Impact:** Positive: the wizard and non-interactive setup are a strong on-ramp; the onboarding problems above are in the surrounding install/license/version/verify plumbing, not the wizard copy.
-- **Fix:** None — keep. (Recommend capturing real wizard frames via a pty harness in CI since alt-screen content isn't visible to script/expect, to guard the rendering that could not be verified here.)
-
-
-### B. Interaction flows & UX
-
-#### H3 · /input-style is a fully inert command — registered, in /help, autocompletes, but does nothing
-
-**Status (2026-06-21):** Fixed — `1b2e453`. Command + kind + handler + dead helper removed.
-- **Severity:** high
-- **Where:** `internal/tui/commands.go:288-296 (registry) + internal/tui/model.go:2964-2969 (handler) + internal/tui/rendering.go:43-53 (output)`
-- **Evidence:** Handler: `case commandInputStyle: ... text: shellOnlyCommandText(command.name)`. shellOnlyCommandText renders a warning card whose only body line is "This control is available in the TUI but does not have a backend setting yet." Captured from a test I ran: `command-card input-style / status: warning / State / This control is available in the TUI but does not have a backend setting yet. / hint: use /help to inspect active commands`. grep confirms NO other input-style logic exists: `grep -rn 'input-style|inputStyle|InputStyle' internal --include='*.go' | grep -v _test` returns only the registry entry + this handler. Yet it is listed in /help (S
-- **Impact:** A new user discovers /input-style via `/` autocomplete or /help (where its description, "Show input style state", implies it surfaces real configuration), runs it, and gets a warning card admitting it's a no-op stub. This is the one remaining surfaced-but-inert slash command (the sibling stubs /style, /compact, /theme, /effort flagged in the earlier code audit are now actually wired on current main). It erodes trust in the com
-- **Fix:** Remove the /input-style entry from commandDefinitions (commands.go) and drop the commandInputStyle kind + its case in model.go, so /help and autocomplete stop advertising it. Re-add it only when a backend setting exists.
-
-#### H4 · README's TUI slash-command table advertises /usage, which is not a slash command (resolves to 'unknown command')
-
-**Status (2026-06-21):** Fixed — `e68d6cb`. `/usage` removed from the README slash table.
-- **Severity:** high
-- **Where:** `README.md:89 (`| \`/doctor\` \`/usage\` \`/config\` | health, cost, and config without leaving the chat |`) under the "## The TUI" slash-command table`
-- **Evidence:** The registry has no /usage (full name+alias dump from commands.go contains /doctor, /config, but no /usage). I ran a test: `parseCommand("/usage").kind == commandUnknown` (PASS), and `formatGroupedCommandHelp()` does NOT contain /usage (`HELP has ... /usage=false`). `zero --help` shows `usage  Summarize token usage and estimated cost` exists ONLY as a CLI subcommand. So a user who types `/usage` in-chat (per the README) hits the commandUnknown path (model.go:2976) -> "unknown command: /usage" (unless they happen to have authored .zero/commands/usage.md, which a new user has not). The table's caption literally promises "...cost...without leavi
-- **Impact:** A brand-new user reads the only slash-command reference in the README, types /usage to check token cost mid-session, and gets an error. The advertised in-chat cost readout does not exist as a slash command; the real feature is the separate `zero usage` CLI subcommand, which contradicts the table's 'without leaving the chat' promise.
-- **Fix:** Either add a /usage slash command (kind+handler showing the same token/cost summary as the `zero usage` subcommand) or remove `/usage` from the README:89 table and document `zero usage` as a CLI command instead.
-
-#### L1 · Ctrl+E in the composer is hijacked by mouse-release toggle, breaking the emacs end-of-line binding (while Ctrl+A=home still works)
-
-**Status (2026-06-21):** Deferred — low (Ctrl+E composer binding).
-- **Severity:** low
-- **Where:** `internal/tui/model.go:818-825 (top-level Ctrl+E intercept) vs internal/tui/composer.go:223-225 (composer Ctrl+E = end-of-line)`
-- **Evidence:** The top-level key switch handles `case keyCtrl(msg, 'e'):` and unconditionally toggles m.mouseReleased then `return`s, before the composer dispatch (model.go:1140 `applyComposerKey`) is ever reached. composer.go:223 declares `case keyIs(msg, tea.KeyEnd) || keyCtrl(msg, 'e'): state.cursor = composerLineEnd(state)` — that Ctrl+E arm is therefore unreachable from the composer. By contrast Ctrl+A/K/U/W are NOT intercepted at top level (`grep -n "keyCtrl(msg, 'a'|'e'|'k'|'u'|'w')" model.go` returns ONLY line 818 = Ctrl+E), so they reach the composer normally. The `?`-help overlay (keybinding_help.go:53) documents Ctrl+E only as 'release the mouse
-- **Impact:** A user with emacs/readline muscle memory: Ctrl+A correctly jumps to line start, but Ctrl+E (expected: jump to line end) instead toggles mouse-capture mode and prints a 'Mouse released' system notice — a surprising, asymmetric behavior in the input editor.
-- **Fix:** Either remove the dead Ctrl+E arm from composer.go:223 (leaving only tea.KeyEnd), or gate the top-level Ctrl+E mouse-toggle so it only fires on an empty/inactive composer (like the `?` overlay does at model.go:982) and let Ctrl+E reach the composer when editing text; then update keybinding_help.go accordingly.
-
-#### I2 · Two permission decisions share hotkey 'y' but are mutually exclusive by tool type (no actual collision)
-
-**Status (2026-06-21):** Info — no action (no real collision).
-- **Severity:** info
-- **Where:** `internal/tui/permission_prompt.go:45-48 + internal/agent/loop.go:1737-1742,1798-1805`
-- **Evidence:** permissionOptions maps both AlwaysAllowPrefix and AlwaysAllow to hotkey "y". The dedup in permissionOptions is keyed by decision action, not hotkey, so if both appeared in one prompt both 'y' rows would render. However availablePermissionDecisions only appends AlwaysAllowPrefix for shell tools with a command prefix (loop.go:1738), and AlwaysAllow only when permissionSupportsPersistentDecision(toolName) is true — which returns false for `bash/exec_command/write_stdin/apply_patch` (loop.go:1800). So a shell prompt gets AlwaysAllowPrefix=y and a non-shell prompt gets AlwaysAllow=y; they never co-occur.
-- **Impact:** None today — verified mutually exclusive by tool type. Noting it because the safety relies on a non-local invariant (the loop.go gating); a future tool that is both a shell tool AND persistent-decision-eligible would surface two 'y' rows where the first match wins, silently shadowing the second.
-- **Fix:** Optional defensive measure: de-duplicate by hotkey in permissionOptions (or assign distinct hotkeys) so the invariant is enforced at the render layer rather than depending on the agent's decision-set gating.
-
-
-### C. TUI rendering & visual correctness
-
-#### M5 · middleTruncate (diff card file paths) overflows the budget on wide-rune paths
-
-**Status (2026-06-21):** Deferred — explicitly out of scope for this PR (wide-rune truncation budget; outer width clip already prevents frame breakage per I4).
-- **Severity:** medium
-- **Where:** `internal/tui/startup.go:138 middleTruncate (used for the diff card path head at internal/tui/rendering.go:1471 with budget innerWidth/2)`
-- **Evidence:** middleTruncate is also rune-count based (`if len(runes) <= limit`, then string(runes[:front])+"…"+string(runes[len-back:])). Ran the real fn: `middleTruncate(strings.Repeat("路",40), 20)` returns 39 cells for a 20-cell budget (test log: `middleTruncate(20) -> cells=39 (want <=20)`).
-- **Impact:** A file path containing CJK characters in the diff card header renders ~2x its budget; the outer fitStyledLine re-clips it so the card edge survives, but the head/right-aligned counts (+N/-N) collide or the path is cut harder than intended — the path becomes harder to read for non-Latin filenames.
-- **Fix:** Make middleTruncate measure with lipgloss.Width and cut front/back halves with splitAtWidth on the reversed/forward string, or fall back to the width-aware truncateRunes once that is fixed.
-
-#### I4 · Wide-rune handling is correct in prose/markdown/table/user-prompt wrap and at the outer card fit — overflow does NOT break the card frame
-
-**Status (2026-06-21):** Info — no action (verified good).
-- **Severity:** info
-- **Where:** `internal/tui/rendering.go:381 wrapPlainText / :442 splitAtWidth / :463 splitPreservingWidth; internal/tui/startup.go:213 fitStyledLine / :223 truncateStyledLine`
-- **Evidence:** These all measure via lipgloss.Width / per-glyph glyphWidth (e.g. truncateStyledLine line 260 accumulates glyphWidth and also tracks OSC-8 link state so truncation can't leak a hyperlink). toolCard re-fits every body line to innerWidth (rendering.go:1387) so even the buggy truncateRunes output cannot push the card past the terminal width — verified: a 100-col CJK diff card produced all lines at exactly 100 cells. Assistant prose is also capped at 96 cols (assistantMeasureCap, rendering.go:360) for readability on wide terminals. WindowSizeMsg (model.go:1249) resizes composer and resets fade state cleanly.
-- **Impact:** Positive: the layout is robust against frame-breaking; the wide-rune bug manifests as silent content loss inside the card rather than a corrupted UI. Documents why the high-severity finding is content-loss, not layout-break.
-- **Fix:** No change needed; noted so the truncateRunes fix is scoped to content fidelity, and to record that interactive frames themselves remain UNVERIFIED (no pty).
-
-
-### C/E. Color, theme & accessibility
-
-#### H5 · faintest fails WCAG even the 3.0 UI floor in BOTH themes; it carries functional content (line numbers, diff @@/+++/--- headers, tool args, separators)
-
-**Status (2026-06-21):** Fixed — `cc3b5e3`. faintest raised to WCAG AA (>=4.5) in both themes.
-- **Severity:** high
-- **Where:** `internal/tui/theme.go:135 (dark faintest #3a3a40), :168 (light faintest #9b9ba3); consumed as zeroTheme.faintest / diffMeta / addLineNum / delLineNum / toolArg`
-- **Evidence:** Computed WCAG ratios: light faintest #9b9ba3 on panel #ececed = 2.34; dark faintest #3a3a40 on panel #0e0e10 = 1.71, and on the unpainted canvas #070708 = 1.78. AA-normal needs 4.5, the relaxed UI/large floor is 3.0 — these are below even 3.0. faintest renders diffMeta (theme.go:223: '@@ hunks, +++/--- headers') and the diff gutter line numbers (theme.go:226-227), which are functional, not decorative.
-- **Impact:** On a brand-new user's terminal, diff hunk headers, line numbers, and tool argument hints are near-invisible to anyone with even mild low vision, and washed out for everyone on a bright screen. The existing test passes because TestLightPaletteContrastAndHierarchy only asserts a luminance DELTA (panelL-inkL>=0.5) and monotonic ordering — it never checks a WCAG ratio, so these sub-floor values ship green.
-- **Fix:** Darken light faintest toward ~#6e6e76 (>=3.0 on #ececed) and lighten dark faintest toward ~#5a5a62 (>=3.0 on #0e0e10/#070708); then add a test asserting actual WCAG ratio >= 3.0 for every token used to carry text, not just a luminance delta.
-
-#### H6 · faint is sub-AA in both themes yet carries instructional/navigation text (help footer, MCP wizard hints, composer placeholder, working timer)
-
-**Status (2026-06-21):** Fixed — `cc3b5e3`. faint raised to WCAG AA in both themes.
-- **Severity:** high
-- **Where:** `internal/tui/theme.go:134 (dark faint #5b5b63), :167 (light faint #78787f); used at keybinding_help.go:89, mcp_add_wizard_view.go:26-160, mcp_manager.go:362/443, model.go:2111/2151/2231`
-- **Evidence:** WCAG: dark faint #5b5b63 on panel = 2.87 (below the 3.0 UI floor); light faint #78787f on panel = 3.71 (below AA-normal 4.5). It renders real instructions: keybinding_help.go:70 'keybindingHelpFooter = "? or Esc to close · /help for slash commands"' shown via zeroTheme.faint (the user's only on-screen cue for how to exit the help overlay), mcp_manager.go:362 the nav legend 'type search up/down navigate Enter action Esc close', and the composer placeholder (model.go:2231).
-- **Impact:** The dismiss hint, navigation legends, MCP setup guidance, and input placeholder are hard to read for low-vision users and dim for everyone — a new user trying to learn the keys is reading the help in the lowest-contrast color in the UI.
-- **Fix:** Raise faint to clear 4.5 (light) / >=3.0 (dark) on its panel, e.g. dark #6f6f78, light #5f5f66. Reserve sub-AA grays strictly for non-text decoration (rules/borders), which already have separate tokens (line/line2).
-
-#### M2 · Light-theme accent (4.34) is below AA for normal text, despite a test that claims to verify light accent contrast
-
-**Status (2026-06-21):** Fixed — `cc3b5e3`. Light accent #4d7a08->#477006 (AA); theme test now asserts a true WCAG ratio.
-- **Severity:** medium
-- **Where:** `internal/tui/theme.go:169 (light accent #4d7a08); test internal/tui/theme_select_test.go:83`
-- **Evidence:** WCAG: light accent #4d7a08 on panel #ececed = 4.34 (AA-normal needs 4.5). accent renders the user gutter '❯' (theme.go:38/210), bash prompt (theme.go:48/218), spinner, and focus. The test at theme_select_test.go:83 only asserts 'panelL - relLum(accent) < 0.25' (a luminance delta, not a contrast ratio), so 4.34 passes. Dark accent #caff3f is fine (16.4).
-- **Impact:** On a light terminal the brand prompt glyph and bash gutter sit just under the readability bar for small text. Borderline rather than broken, but it is the most-repeated colored glyph on screen.
-- **Fix:** Darken light accent slightly (e.g. #436a07 → ~4.7) to clear AA-normal, and change the test to assert a real WCAG ratio (>=4.5 for text-bearing accent) instead of a luminance delta.
-
-#### M3 · NO_COLOR honored only for strconv.ParseBool-style values; NO_COLOR=yes / NO_COLOR=anything leaves the UI in full color, violating the no-color.org spec
-
-**Status (2026-06-21):** Fixed — `7fc296b`. NO_COLOR with any non-empty value forces the Ascii profile.
-- **Severity:** medium
-- **Where:** `dependency github.com/charmbracelet/[email protected] env.go:115-117 (envNoColor uses strconv.ParseBool); zero never sets WithColorProfile (internal/tui/run.go:25-35), so it inherits this`
-- **Evidence:** Drove /tmp/zero under a pty across NO_COLOR variants and diffed emitted SGR: 'NO_COLOR=1' -> colorSGR=false (306 bytes, monochrome), 'NO_COLOR=true' -> false, but 'NO_COLOR=yes' -> colorSGR=true (676 bytes, full color) and 'NO_COLOR=foo' -> colorSGR=true. Source: colorprofile env.go:116 'noColor, _ := strconv.ParseBool(env.get("NO_COLOR"))'. The no-color.org spec says NO_COLOR present with ANY non-empty value must disable color. (NO_COLOR=1 correctly produced a readable bold-only frame — color degrades cleanly when the value is recognized.)
-- **Impact:** A user who follows the common convention 'export NO_COLOR=yes' (or any non-bool truthy value) still gets a full-color TUI, defeating their accessibility/preference setting. Silent and surprising.
-- **Fix:** Don't rely on the dependency's ParseBool: detect NO_COLOR yourself (present and non-empty => force colorprofile.Ascii via tea.WithColorProfile) in run.go, matching no-color.org. At minimum document that only NO_COLOR=1/true work today.
-
-#### M4 · Documented-in-code --theme flag does not exist; Options.Theme is read but never populated, so the only non-interactive theme control is the undocumented ZERO_THEME env var
-
-**Status (2026-06-21):** Fixed — `3b73074`. Real `--theme {auto|dark|light}` flag wired to Options.Theme; stale comment corrected.
-- **Severity:** medium
-- **Where:** `internal/tui/theme.go:12 and theme_select.go:18 (both claim a '--theme flag'); internal/tui/options.go:52-54 (Theme field); internal/tui/model.go:559 (reads it)`
-- **Evidence:** Binary: 'zero --theme light -p hi' -> 'unknown command "--theme"'. 'zero --help' global Flags list has no theme flag. grep for assignments to Options.Theme / '.Theme =' across cmd/ and internal/ returns ZERO hits — the field is only ever READ (model.go:559 resolveThemeMode(options.Theme, ...)). Yet theme_select.go:18 comments 'the --theme flag, threaded via Options.Theme' and theme.go:12 says startup selection happens via '--theme'.
-- **Impact:** A user reading the code (or expecting the documented flag) cannot pick a theme at launch via --theme. They must discover ZERO_THEME (documented nowhere) or use /theme after the TUI is already up. The stale comments actively mislead.
-- **Fix:** Either wire a real --theme {auto|dark|light} flag that sets Options.Theme, or delete the --theme references from theme.go:12 and theme_select.go:18 and document ZERO_THEME + /theme as the supported controls.
-
-#### L2 · Accessibility/customization controls (NO_COLOR, ZERO_THEME, ZERO_NO_FADE, motion opt-out) are completely undocumented
-
-**Status (2026-06-21):** Fixed — `3cd0a50`. README "Accessibility & appearance" section documents NO_COLOR/ZERO_THEME/--theme//theme/ZERO_NO_FADE.
-- **Severity:** low
-- **Where:** `README.md / docs/ (grep for NO_COLOR|ZERO_THEME|ZERO_NO_FADE|--theme|reduce.motion returns nothing); controls exist at streaming_fade.go:28-44 and theme_select.go:19`
-- **Evidence:** grep -rni 'NO_COLOR|ZERO_THEME|ZERO_NO_FADE|--theme|reduce.motion' README.md docs/ => no matches. The controls themselves are real and tested: streaming_fade.go:28-44 honors ZERO_NO_FADE and auto-disables fade over SSH/tmux/16-color/no-TTY; /theme and ZERO_THEME switch palettes.
-- **Impact:** A user who finds the streaming-text fade or color distracting (motion sensitivity) has no way to know ZERO_NO_FADE exists; a NO_COLOR/light-terminal user can't discover ZERO_THEME. The a11y story is decent in code but invisible to the new user it would help.
-- **Fix:** Add a short 'Accessibility / Appearance' section to the README listing NO_COLOR (note the value caveat), ZERO_THEME=auto|dark|light, /theme, and ZERO_NO_FADE (reduce-motion opt-out).
-
-#### I3 · Meaning does not depend on color alone, and there IS a reduce-motion opt-out — both verified
-
-**Status (2026-06-21):** Info — no action (verified good).
-- **Severity:** info
-- **Where:** `internal/tui/rendering.go:1512/1516 (diff signs), :1200-1203 (permission ✓/✗), view.go:277-285 (mode labels), :996 (PERMISSION text badge); streaming_fade.go:28-44 (ZERO_NO_FADE)`
-- **Evidence:** Diffs emit explicit sign columns: rendering.go:1512 diffBodyLine(...,"+",...) for adds and :1516 "−" for dels, so add/del survive color-stripping. Permission outcome uses glyphs ✓ (rendering.go:1200) / ✗ (:1203), the card is gated by a text 'PERMISSION' badge (:996), and permission MODE shows a text label 'auto-approve'/'ask'/'unsafe' (view.go:277-285) not color-only. Streaming fade has a reduce-motion path: ZERO_NO_FADE plus auto-disable on SSH/tmux/ANSI/no-TTY (streaming_fade.go:29-43); the spinner has no separate opt-out but is a single glyph, not a fade. NO_COLOR=1 pty frame rendered fully readable in bold-only.
-- **Impact:** Positive: the core surfaces (diffs, permissions, modes) remain interpretable for colorblind users and under NO_COLOR. Recorded so it's not re-flagged; the only motion gap is the spinner having no independent disable.
-- **Fix:** Optional: extend the reduce-motion opt-out to also still the spinner glyph when ZERO_NO_FADE (or a NO_COLOR/reduce-motion signal) is set, for full motion-sensitivity coverage.
-
-
-### D. Consistency & polish
-
-#### H7 · Bad/missing API key produces a mangled, self-contradicting error paragraph (Redact corrupts the provider's own help text)
-
-**Status (2026-06-21):** Fixed — `8b9b479`. 401/403 lead with an actionable curated message; Redact only scrubs token-shaped Bearer words.
-- **Severity:** high
-- **Where:** `internal/providers/providerio/providerio.go:234-249 (Redact); surfaced via exec/-p with no/invalid key`
-- **Evidence:** Live: `/tmp/zero -p "hi" --model not-a-real-model` -> `[zero] auth error: You didn't provide an API key. You need to provide your API key in an Authorization header using authorization [REDACTED] (i.e. Authorization: authorization [REDACTED] or as the password field (with blank username) if you're accessing the API from your browser and are prompted for a username and password. You can obtain an API key from https://platform.openai.com/account/api-keys.` The raw 401 prose contains the literal word 'Bearer' twice as INSTRUCTION text; Redact() (line 242-247: `if EqualFold(TrimRight(words[index],":"),"Bearer")` -> rewrite this+next word) treats
-- **Impact:** A brand-new user whose only mistake is a missing/typo'd key gets an incomprehensible, grammatically-broken paragraph that (a) reads like a bug, (b) gives browser-username/password advice irrelevant to a CLI, (c) names platform.openai.com even when they aren't trying to use OpenAI, and (d) never tells them to run `zero auth` / `zero setup`. This is the single most likely error a first-run user hits.
-- **Fix:** Don't pass raw upstream auth-error bodies through verbatim. For 401/403 emit a curated message: `auth error: your <provider> API key is missing or invalid. Run \`zero auth\` or set <ENV>=...`. Keep the raw body only behind a --verbose/debug flag. Also tighten Redact so the 'Bearer' heuristic only fires on token-shaped following words (e.g. matches sk-/long base64), not arbitrary prose.
-
-#### H8 · doctor leaks raw Go map[...] syntax for ANY nested-map detail — confirmed on lsp.servers AND config.validation (shared formatDetails bug)
-
-**Status (2026-06-21):** Fixed — `bb4b348`. formatDetailValue renders nested maps as "k: v"; no more raw map[...].
-- **Severity:** high
-- **Where:** `internal/doctor/doctor.go:277 (formatDetails uses %v on the value)`
-- **Evidence:** lsp.servers: `missing: map[gopls:install with \`go install golang.org/x/tools/gopls@latest\` ... pyright-langserver:install with \`npm install -g pyright\` ...] | present: map[rust-analyzer:on PATH]`. config.validation (with a malformed config.json): `/tmp/zbad/.zero/config.json: map[col:4 error:invalid config JSON: invalid character 't' looking for beginning of object key string line:1]` — the actual error sentence is buried inside a Go map dump. Root cause line 277: `fmt.Sprintf("%s: %v", ..., redaction.RedactValue(value, ...))` where value is a map[string]any.
-- **Impact:** Builds on the established lsp leak: this is not isolated to one check — every doctor detail whose value is a nested map renders as unreadable Go syntax. The MOST useful diagnostics (which LSP to install, what's wrong with your config + the line/col) are exactly the ones mangled into `map[...]`. New users running `zero doctor` to fix a problem get gibberish.
-- **Fix:** In formatDetails, special-case map[string]any values: render them as indented `key: value` sub-lines (or `k=v, k=v`) instead of `%v`. For config.validation specifically, hoist line/col/error into a one-line human string (`line 1, col 4: invalid character 't' ...`).
-
-#### M6 · Inconsistent validation of enum-like flags: --mode rejects bad values with a great message, --reasoning-effort silently ignores garbage
-
-**Status (2026-06-21):** Deferred — medium not in this PR's scope (crit/high + OSS files + contrast). Tracked for a follow-up (validate --reasoning-effort like --mode).
-- **Severity:** medium
-- **Where:** `internal/cli/exec.go:902 (mode) vs exec.go:965-998 (reasoning effort)`
-- **Evidence:** `zero -p hi --mode bogus` -> `[zero] unknown mode "bogus". Valid modes: smart, deep, fast, large, precise.` (exit 2 — exemplary microcopy). But `zero -p hi --reasoning-effort wat` -> `gpt-4.1 does not support reasoning effort; ignoring --reasoning-effort wat` then proceeds; `wat` is never validated as a value (the message only fires because gpt-4.1 lacks the capability — on a model that DOES support effort, an invalid value like `wat` would pass through unchecked).
-- **Impact:** Inconsistent error contract across sibling flags. A typo in --reasoning-effort is silently swallowed instead of corrected with the valid set (low/medium/high), so users think their setting applied when it didn't.
-- **Fix:** Validate --reasoning-effort against the allowed enum the same way --mode does, with a `Valid efforts: low, medium, high.` message; only fall back to the capability-skip notice when the value itself is valid.
-
-#### L3 · Brand capitalization inconsistent across one CLI surface (ZERO / Zero / zero) and `zero version` prints no real version
-
-**Status (2026-06-21):** Deferred — low (brand capitalization + real version string).
-- **Severity:** low
-- **Where:** ``zero --help` header vs body; `zero version``
-- **Evidence:** `zero --help` first line: `ZERO terminal coding agent`; the 10 command descriptions all say `Zero` (e.g. 'List Zero model registry entries'); `zero version` -> `zero dev`. So one help screen shows three casings, and the release binary reports its version as the literal word `dev`.
-- **Impact:** Cosmetic inconsistency a design-conscious user notices immediately; `zero dev` on a fresh clone gives no way to tell which build they have when filing issues.
-- **Fix:** Pick one brand casing (recommend 'Zero') and use it everywhere including the --help header. Stamp a real version (git describe / tag) into the build via ldflags so `zero version` is meaningful.
-
-#### I5 · Empty states and existing polish tests are genuinely clean
-
-**Status (2026-06-21):** Info — no action (verified good).
-- **Severity:** info
-- **Where:** `providers/sessions/skills/plugins/usage commands; internal/tui/*_test.go`
-- **Evidence:** Empty-state copy is clear and path-aware: `No Zero skills found in /tmp/.../skills.`, `No local Zero plugins loaded.`, `No Zero sessions found.`, and `usage` renders a tidy zero table with `n/a (net LOC <= 0)` guards. `providers` even flags `api key: not set`. Visual tests TestLightPaletteContrastAndHierarchy / TestWidthTierSegments / TestViewNeverExceedsTerminalWidth / TestWrapPlainTextPreservesAlignedWhitespace / TestThemeAutoReProbesBackground / TestHelpRoutesThroughStyledCard / TestStyleCommandCardContentRowTwoTonesCommands all PASS.
-- **Impact:** Positive: card alignment, width tiers, light/dark contrast, whitespace-preserving wrap, and theme re-probe are regression-guarded and look correct. The polish problems are concentrated in error/warning microcopy, not layout or empty states.
-- **Fix:** No change needed; keep these tests. Minor terminology nit to consider later: `config` says `active provider: none` while `providers` lists openai as the default keyless profile — reconcile the two so 'active' vs 'default' is unambiguous.
-
-
-### F. Product completeness & robustness
-
-#### H9 · `zero doctor` reports "Overall: pass" with NO provider credential — the one tool meant to verify keys gives a false all-clear
-
-**Status (2026-06-21):** Fixed — `bb4b348`. provider.config fails (not pass) for a remote provider with no credential; keyless local stays pass.
-- **Severity:** high
-- **Where:** `internal/doctor/doctor.go:136-156 (providerConfigCheck) + :83-84 (report.OK only flips on StatusFail) + :188 (connectivity is warn/skipped). Reproduced via `zero doctor`.`
-- **Evidence:** With unset OPENAI_API_KEY and empty config: `Overall: pass` ... `[pass] provider.config - Provider config loaded for openai.\n  ... credentialConfigured: not set ... model: gpt-4.1` ... `[warn] provider.connectivity - Connectivity probe skipped. Run zero doctor --connectivity`. In code, providerConfigCheck returns StatusPass whenever the profile is non-empty (the built-in openai/gpt-4.1 default is non-empty), regardless of `credential := "not set"`; report.OK is false ONLY if some check is StatusFail (line 83-84).
-- **Impact:** README says `zero doctor` will "verify config, keys, and connectivity." A brand-new user who runs it before their first prompt sees a green "Overall: pass", concludes setup is done, then `zero exec` immediately fails with a raw upstream auth error. The diagnostic actively misleads instead of catching the single most common new-user misconfiguration.
-- **Fix:** In providerConfigCheck, when credential=="not set" AND the provider kind requires a key (i.e. not a local/keyless runtime like ollama/lmstudio), return StatusWarn (or StatusFail) with help text pointing to `zero setup` / `zero auth`. That makes Overall reflect "not actually usable yet" and gives the user the next action.
-
-#### H10 · No-provider / missing-key error never names zero's own onboarding (`zero setup`/`zero auth`) — it dumps a raw OpenAI HTTP message telling the user to visit platform.openai.com
-
-**Status (2026-06-21):** Fixed — `42b7123`. exec + TUI no-provider errors point at `zero setup` / `zero auth`.
-- **Severity:** high · also: A-onboarding, D-polish
-- **Where:** `internal/providers/providerio/providerio.go:218-232 (ClassifiedError) + internal/config/resolver.go:740 (silent default to openai/gpt-4.1). Reproduced via `zero exec "hello"` with no config.`
-- **Evidence:** `zero exec "hello"` (no key) → stderr: `[zero] auth error: You didn't provide an API key. You need to provide your API key in an Authorization header ... You can obtain an API key from https://platform.openai.com/account/api-keys.` (exit 3). The README's literal first example is `zero exec "fix the failing test in ./pkg"`. grep for `zero setup|zero auth|no provider configured` across internal/cli, internal/agent, internal/providers error paths returns nothing in this path — the only zero-added text is the `auth error: ` prefix.
-- **Impact:** An exec-first / CI user who copy-pastes the README's headline command before running the wizard gets a wall of OpenAI-branded text steering them to OpenAI's website, with zero mention that zero has a `zero setup` wizard or `zero auth`, and no hint that it silently defaulted to OpenAI. Onboarding dead-ends for anyone not using the interactive TUI.
-- **Fix:** When the resolved profile has no credential (credential not set) and the request fails auth, prepend a zero-owned line before the upstream text, e.g.: `No provider credential found. Run \`zero setup\` (interactive) or set OPENAI_API_KEY / ANTHROPIC_API_KEY / GEMINI_API_KEY. Local models need no key — see \`zero providers list\`.` Detect the no-credential case at the agent/exec boundary rather than relying on the upstream 401 b
-
-#### M7 · Provider errors (auth / rate limit) are classified but give the user no recovery action
-
-**Status (2026-06-21):** Largely addressed by H7 (`8b9b479`) — auth errors now name a recovery action (`zero auth`). Remaining per-class recovery hints DEFERRED.
-- **Severity:** medium
-- **Where:** `internal/providers/providerio/providerio.go:218-231 (ClassifiedError); rendered raw in TUI via internal/tui/model.go:1403-1409 (rowError text = msg.err.Error()) and to exec stderr.`
-- **Evidence:** ClassifiedError only adds a category prefix: `auth error: `, `rate limit error: `, `provider error: `. The TUI error row is `transcriptRow{kind: rowError, text: msg.err.Error(), final: true}` — the verbatim classified string, no appended next-step. Confirmed live: rate-limit/auth bodies are the raw upstream message with only the prefix.
-- **Impact:** On a 429 the user sees `rate limit error: <upstream text>` with no guidance (wait? `/model` to a cheaper or different provider? it already retried?). On auth failure, no pointer to `/provider`/`zero auth`. The failure is legible as a *category* but not *recoverable* without the user already knowing zero's surface. (Reconnect for transient disconnects in reconnect.go DOES surface a `[connection lost — reconnecting N/2…]` notice
-- **Fix:** Append a short, category-specific recovery hint when rendering rowError / exec error: auth → `Run \`zero auth\` or check your API key, or /provider to switch.`; rate limit → `Zero already retried with backoff; wait or use /model to switch provider.` Keep it one line; the classifier already knows the category.
-
-#### M8 · TUI has no slash command to discover or manage subagents/cron/skills/plugins — features the README markets prominently
-
-**Status (2026-06-21):** Deferred — larger feature (discovery slash commands for subagents/cron/skills/plugins); out of this PR's scope.
-- **Severity:** medium
-- **Where:** `internal/tui/commands.go (full registry: /spec and /mcp exist; no /specialist, /cron, /skills, /plugins, /subagent). CLI subcommands exist (zero specialist|cron|skills|plugins per `zero --help`).`
-- **Evidence:** grep over commands.go for `spec|specialist|cron|mcp|skill|plugin|subagent` matches only `/mcp` (:174) and `/spec` (:197). README §Why Zero headlines "Subagents", "Spec mode", "Scheduled agents", and "Extensible — skills, plugins, hooks"; the TUI command table only lists /spec /plan. The ? overlay (keybinding_help.go:58-61) references specialist *cards* but only as a reactive drill-in, not a way to list/create them.
-- **Impact:** A user living in the interactive TUI (the default `zero` with no args) can type `/` and will never surface cron, skills, plugins, or specialist management. These read as CLI-only features even though README presents them as first-class. Discoverability of half the headline feature set is gated on reading the README's Commands section and dropping out of the TUI.
-- **Fix:** Add thin TUI slash commands that shell into the existing CLI surfaces (or at least informational `/specialist`, `/cron`, `/skills`, `/plugins` that print status + the CLI command to manage them), and list them in `/help`. Even read-only listing closes the discoverability gap.
-
-#### I6 · Strong robustness wins worth preserving: network-down framing, mid-stream reconnect, output truncation, and width-safe rendering
-
-**Status (2026-06-21):** Info — no action (verified good).
-- **Severity:** info
-- **Where:** `internal/providers (upstream-unreachable message), internal/agent/reconnect.go, internal/tools/exec_command.go:802-812, internal/tui/rendering.go:1161-1233; tests in internal/tui.`
-- **Evidence:** Unreachable base URL → `[zero] upstream unreachable: the model server could not connect to 127.0.0.1:9 (connection refused). The request never reached the model — this is a network failure ... Verify the host is reachable (DNS/proxy/VPN/firewall) ...` (exit 3). reconnect.go retries connect-time disconnects (eof/reset/timeout/502/503) twice with backoff and surfaces `[connection lost — reconnecting N/2…]`. exec output truncates head+tail with `\n[zero] output truncated\n`. Tool cards cap at 16 live / 400 flushed lines and collapse noisy output. Visual tests PASS: TestViewNeverExceedsTerminalWidth, TestTinyTierSingleSegmentAndRailLessCards, Tes
-- **Impact:** These are genuinely good for a new user: network failures are legible and actionable, transient hiccups self-heal, and huge tool output won't flood the transcript or overflow narrow terminals. Listed so they are not regressed by fixes to the items above.
-- **Fix:** None — keep. The findings above should reuse this same quality bar (the unreachable-host message is the model for what auth/rate-limit/no-key errors should look like).
-
-
-### G. Open-source readiness
-
-#### H11 · No SECURITY.md and no private vulnerability-reporting path
-
-**Status (2026-06-21):** Fixed — `634f23f`. SECURITY.md with a private GitHub-advisory reporting path.
-- **Severity:** high
-- **Where:** `/tmp/zero-ux/ (no SECURITY.md at root or .github); CONTRIBUTING.md (no disclosure section)`
-- **Evidence:** find for security.md => none. grep -niE 'vulnerab|disclos|CVE|report.*privately' across CONTRIBUTING.md/README.md/INSTALL.md => only an unrelated 'security issue' mention in CONTRIBUTING:92 and a firecrawl privacy note. For a coding agent that executes shell, edits files, and handles API keys/secrets, there is no documented way to report a vuln.
-- **Impact:** A security researcher who finds a sandbox-escape or secret-leak in an agent that runs arbitrary commands has no private channel and will either drop a public issue (0-day exposure) or go silent. GitHub also won't surface a 'Security policy' for the repo.
-- **Fix:** Add SECURITY.md (root or .github/) with a private reporting address or GitHub Security Advisories link and a supported-versions/response-time note.
-
-#### M9 · No issue templates and no PR template (.github/) despite CONTRIBUTING mandating their use
-
-**Status (2026-06-21):** Fixed — `634f23f`. .github issue (bug/feature/config) + PR templates added.
-- **Severity:** medium
-- **Where:** `/tmp/zero-ux/.github/ (only workflows/); CONTRIBUTING.md:49 ('Open an issue using the appropriate issue template')`
-- **Evidence:** find .github -type f => only the 4 workflow YAMLs. No .github/ISSUE_TEMPLATE/, no PULL_REQUEST_TEMPLATE. Yet CONTRIBUTING.md:49 says 'Open an issue using the appropriate issue template' and the whole policy hinges on an `issue-approved` label workflow.
-- **Impact:** A newcomer told to 'use the appropriate issue template' finds none — bug reports arrive unstructured (the exact low-signal reports CONTRIBUTING says will be closed), and the documented contribution process is self-contradictory on first contact.
-- **Fix:** Add .github/ISSUE_TEMPLATE/bug_report.yml + feature_request.yml (matching the fields CONTRIBUTING already enumerates at lines 122-145) and a PULL_REQUEST_TEMPLATE.md that prompts for the `Fixes #123` issue link.
-
-#### M10 · No CODE_OF_CONDUCT.md
-
-**Status (2026-06-21):** Fixed — `634f23f`. CODE_OF_CONDUCT.md (Contributor Covenant 2.1) added.
-- **Severity:** medium
-- **Where:** `/tmp/zero-ux/ (none at root or .github)`
-- **Evidence:** find for code_of_conduct* => none.
-- **Impact:** Standard OSS-readiness/community-health item missing; GitHub's community profile flags it and some users/orgs treat its absence as a maturity signal. Low functional impact but expected for a public release.
-- **Fix:** Add CODE_OF_CONDUCT.md (e.g. Contributor Covenant) with a contact for enforcement.
-
-#### M11 · No CHANGELOG and source builds report version 'dev' — `zero update` has nothing to compare against
-
-**Status (2026-06-21):** Partially fixed — `634f23f` adds CHANGELOG.md; build-time version stamping (still `dev`) is DEFERRED.
-- **Severity:** medium
-- **Where:** `/tmp/zero-ux/ (no CHANGELOG); internal/update/update.go:138-141; `/tmp/zero --version` => 'zero dev'`
-- **Evidence:** `/tmp/zero --version` => `zero dev`. update.go:138-141: 'Source/dev builds carry a non-semver version ("dev"); ... currentVersion = "0.0.0"'. No CHANGELOG file anywhere. README.md:147 advertises `zero update` and INSTALL.md:91-98 documents `zero update --check`.
-- **Impact:** A user running a source build sees version 'dev' and `zero update` treats it as 0.0.0 (will always claim an update is available once any release exists). With no CHANGELOG, a user can't see what changed between versions before updating. Weakens the documented update flow.
-- **Fix:** Add a CHANGELOG.md (Keep-a-Changelog style) and document the release/versioning process; consider stamping the real version via -ldflags in the release builder so installed binaries report a semver.
-
-#### L4 · README has no screenshot/GIF for a TUI-first product
-
-**Status (2026-06-21):** Deferred — MANUAL maintainer step (add a screenshot/GIF; not auto-generated).
-- **Severity:** low
-- **Where:** `/tmp/zero-ux/README.md:1-21 (only ASCII logo + shields.io badges)`
-- **Evidence:** grep -niE '!\[|\.png|\.gif|demo|screenshot|asciinema' README.md => only the 4 shields.io badge images; no product screenshot or terminal recording. README.md:38 promises 'A TUI that feels premium'.
-- **Impact:** A newcomer evaluating a 'premium TUI' coding agent on the README/repo page can't see what it looks like before investing in a build — significant conversion/first-impression cost for a terminal-UI product.
-- **Fix:** Add a screenshot or asciinema/GIF of the TUI (setup wizard + a chat turn) near the top of the README.
-
-#### L5 · npm package name is the un-scoped generic 'zero' — almost certainly unpublishable as-is
-
-**Status (2026-06-21):** Deferred — low (npm package name).
-- **Severity:** low
-- **Where:** `/tmp/zero-ux/package.json:2`
-- **Evidence:** package.json: `"name": "zero"`. README.md:279 calls the npm wrapper a real install path ('the npm wrapper just delegates to it'). bin/zero.js exists as the wrapper.
-- **Impact:** `npm publish` of a bare common word like `zero` will collide with an existing package / be rejected, so the advertised npm distribution can't ship under this name. Minor since npm isn't the primary documented path, but it's a release blocker for that channel.
-- **Fix:** Scope the package (e.g. @gitlawb/zero) or pick an available name, and document the actual npm install command in the README (currently none is shown).
-
-### Dropped in adversarial verification
-
-| Claim | Why dropped |
-|---|---|
-| Wide-rune (CJK/emoji) diff/code/grep content truncated by rune-count, "half the line silently lost" | The mechanism (rune-count budget in `truncateRunes`) is real, but the **outer `fitStyledLine(line, innerWidth)` clip** (`rendering.go:1387`) is display-width-aware and absorbs the over-selection: an end-to-end width sweep showed **zero** content loss vs a width-correct budget and the ellipsis is preserved. No user-visible harm → invalid as a UX defect (kept only as a minor `middleTruncate` budget note, M). |
-| Default `firecrawl` MCP server 401s on "every first run", breaking the keyless web-search promise | **Split verdict** — one verifier reproduced an **intermittent ~20%** keyless 401 (Firecrawl-side flakiness, surfaced as a mangled Bearer-redacted warning); another got HTTP 200 + 25 tools registered and the warning not emitted by `zero setup`. Not reproducible enough to assert → recorded as Info below, not a finding. |
-
-> **Info (intermittent, unverified):** the keyless `firecrawl` default endpoint returned a 401 on a minority of identical requests in one run; if it flakes, the warning is rendered mangled (the Bearer-redaction eats the message tail). Worth a real-world soak test of the "no API key" web-search claim before launch.
-
-## 5. Open-source readiness checklist
-
-| Item | Status | Note |
-|---|---|---|
-| **LICENSE** | ❌ FAIL | Missing; README says "being finalized." **Release blocker.** No SPDX headers, no `package.json` license field either. (C2) |
-| Working install path | ❌ FAIL | Binary install (`install.sh`/`.ps1`) 404s — no GitHub Release; CI never creates one. `go run` works but needs Go 1.25. (C1) |
-| README — newcomer quality | ⚠️ PARTIAL | Good what-it-is/quickstart/config prose + telemetry/privacy statement; but stale Go-version, a `/usage` slash + `--theme` flag that don't exist, and **no screenshot/GIF** for a TUI-first product. |
-| CONTRIBUTING.md | ⚠️ PARTIAL | Present and detailed — but mandates issue/PR templates that don't exist. |
-| CODE_OF_CONDUCT.md | ❌ FAIL | Missing. |
-| SECURITY.md + vuln path | ❌ FAIL | Missing; no private disclosure channel — a baseline expectation for a tool that runs shell/edits files. (H) |
-| Issue templates / PR template | ❌ FAIL | `.github/` has only workflows. |
-| CHANGELOG + versioning | ❌ FAIL | No CHANGELOG; binaries report `dev` → `zero update` always claims an update. |
-| CI coverage | ✅ PASS | Strong: 3-OS matrix, gofmt/vet/govulncheck/deadcode. |
-| Telemetry / privacy statement | ✅ PASS | "No telemetry" stated; Firecrawl keyless-routing disclosed. |
-| Dependency-license compatibility | ✅ PASS | No copyleft deps linked into the binary; Firecrawl AGPL is network-only and README's reasoning is correct. |
-| Reproducible build | ⚠️ PARTIAL | `go build` reproducible on Go 1.25+; binary distribution non-functional. |
-
-## 6. Root-cause synthesis
-
-1. **The release/packaging layer was never finished.** The product is mature but the OSS scaffolding isn't: no LICENSE, no published release (install scripts dead-end), no SECURITY/CoC/CHANGELOG/templates, version pinned at `dev`. Code shipped ahead of distribution. → C1, C2, and most of area G.
-2. **Docs/marketing copy drifted from the implementation.** README claims Go 1.24 (needs 1.25), a `/usage` slash command and a `--theme` flag that don't exist, and a keyless local-model on-ramp the SSRF probe blocks; accessibility env vars are undocumented. The pitch outran the code. → H1, H2, H4, M (theme flag), L (a11y env docs).
-3. **Raw core errors reach users without a "what to do next" UX layer.** The surface-agnostic core's classified/raw errors surface verbatim: no-key dumps an OpenAI URL instead of `zero setup`; "rejected the key" when none was sent; `doctor` says "pass" with no credential and prints Go `map[...]`; provider errors carry no recovery action. → the no-provider cluster, M1, doctor findings, provider-error finding.
-4. **Theme aesthetics undercut the readability bar.** `faint`/`faintest` fail WCAG yet carry functional content (line numbers, diff `@@`/`+++`/`---`, help footer, placeholders); light `accent` is sub-AA. → the contrast findings.
-
-## 7. Prioritized fix order for the OSS launch
-
-**MUST ship (launch blockers):**
-1. **Add a `LICENSE`** and name it in README. (C2)
-2. **Make an install path actually work** — publish a GitHub Release with the documented assets under the real repo (and fix `ZERO_REPO`/links), or demote binary-install to "coming soon" and lead with `go run`. (C1)
-3. **Fix the Go-version claim** (one-line: README → "Go 1.25+"). (H2)
-4. **Add `SECURITY.md`** with a private vuln-reporting path. (H)
-5. **Stop misleading on first contact:** remove the inert `/input-style` (H3) and the README's nonexistent `/usage` slash + `--theme` flag (H4, M); or wire them.
-6. **Unblock the keyless local-model on-ramp** — allow loopback in the provider connectivity probe for user-configured local providers. (H1)
-7. **Onboarding errors guide recovery:** no-provider/bad-key → point to `zero setup`/`zero auth`, distinguish *missing* key from *rejected* key, and don't dump a raw OpenAI URL. (no-provider cluster, M1)
-8. **`doctor` must not say "Overall: pass" with no credential**, and must stop leaking Go `map[...]`. (H, H)
-
-**SHOULD (strongly recommended before a public launch):**
-- Lift `faint`/`faintest` (and light `accent`) to meet a contrast bar — they carry real content. (H, H, M)
-- Add `CODE_OF_CONDUCT.md`, `CHANGELOG.md`, issue/PR templates; honor `NO_COLOR=<any value>` per spec; document `NO_COLOR`/`ZERO_THEME`/`ZERO_NO_FADE`. (G + M/L cluster)
-- Add a README screenshot/GIF (TUI-first product). (L)
-- Give the TUI discoverability of cron/specialists/skills/plugins (slash commands or a `/help` pointer). (M)
-- Provider errors: append a recovery action; pre-soak-test the keyless Firecrawl path.
-
-**NICE-TO-HAVE:** scope the npm package name; fix the `Ctrl+E` composer binding shadowed by mouse-toggle; normalize brand capitalization + stamp a real `version`; `middleTruncate` wide-rune budget. (L cluster)
-
-## 8. Confidence notes (what could not be fully verified)
-
-- **Live interactive TUI frames are UNVERIFIED.** Bubble Tea v2's alt-screen defeats `script`/`expect` (only escape setup/teardown captured). The rendered first-run wizard, resize redraw, streaming fade, and narrow-tier painting are assessed from render code + the **passing** visual tests, not captured frames. One color check did succeed via `creack/pty` + an OSC-11 reply (monochrome onboarding frame). **Recommend a pty golden-frame harness in CI** to lock the rendering this audit could only read.
-- **Contrast** ratios are computed deterministically from the theme hex tokens (solid), but "readable on a *physically* light terminal" was exercised with a dark OSC-11 reply, so the light-palette verdict rests on the hex values + the luminance tests, not a real light terminal.
-- **The Go 1.24 build failure** is reasoned from the `go.mod` directive (only go1.26.4 present in this env), not reproduced on a real 1.24 toolchain.
-- **Truecolor / 256 / 16-color downsampling** is library-handled (`colorprofile` writer) and verified by code, not by per-profile pty captures.
-- **GitHub Release existence** for `Gitlawb/zero` was probed via the public API (404 for repo/releases/tags) but the environment has no authenticated GitHub access; the repo may be private/unpublished rather than absent — either way the documented install path cannot succeed for a public stranger today.
-- **Firecrawl keyless 401** could not be settled (intermittent; see §dropped) — needs a real soak test.

docs/design/ZERO_LIME_TUI_REBUILD_PROMPT.md

@@ -1,281 +0,0 @@
-# ZERO · Lime TUI rebuild — staged prompt
-
-You are working in a fresh checkout of `github.com/Gitlawb/zero` (Go 1.24,
-Bubble Tea v1.3.10, Lip Gloss, Bubbles v1.0.0, Glamour — all in go.mod; add NO
-new dependencies).
-
-Mission: **remove both existing TUI designs entirely** — the default cyan
-shell (`internal/tui/theme.go` palette + the centered splash in `startup.go`)
-AND the whole `zeroline` skin system — and replace them with ONE design:
-**Lime**, the near-black, lime-accent chat surface prototyped in
-`docs/design/zero_tui_lime.html` (commit the prototype there first; it is the
-visual source of truth). The prototype is a web mock — this spec is the
-authoritative terminal translation of it. Behavior must not regress: every
-slash command, prompt flow, overlay, and safety property that works today
-must work after.
-
-This is a fresh clone: file names below were verified against a recent
-snapshot, but ALWAYS recon before editing. If a path or symbol has moved,
-adapt — do not blindly apply.
-
-**Gate (run after EVERY stage, commit + tag before continuing):**
-```
-go vet ./...
-go test ./...
-go run ./cmd/zero-release build
-```
-Stage 4 additionally runs `go run ./cmd/zero-release smoke`. Never tag a red
-tree. AGENTS.MD applies: Go-native commands only.
-
----
-
-## Design spec — Lime
-
-### Palette — `internal/tui/theme.go` becomes the ONLY place colors exist
-| Token      | Hex       | Use |
-|------------|-----------|-----|
-| bg         | `#070708` | canvas (terminal default bg; do not paint full-bleed) |
-| panel      | `#0e0e10` | card backgrounds (tool/diff/read/grep/sessions rows) |
-| panel2     | `#121215` | card header rows, picker rows |
-| panel3     | `#17171b` | selected/hovered row bg |
-| line       | `#242429` | default borders, rules, status separators |
-| line2      | `#2e2e34` | emphasized borders |
-| ink        | `#ececee` | primary text |
-| muted      | `#8b8b93` | secondary text, assistant interim prose |
-| faint      | `#5b5b63` | hints, metadata |
-| faintest   | `#3a3a40` | line numbers, separators, tool args |
-| accent     | `#caff3f` | brand lime: prompts, badge, spinner, focus, final rail |
-| green      | `#5dd1a4` | success, diff add sign, ✓ |
-| red        | `#ff7a7a` | errors, diff del sign, ✗, deny |
-| amber      | `#ffc25c` | permission surfaces, warnings, auto badge |
-| blue       | `#7db4ff` | grep file locations, local-model dot |
-| addBg      | `#15201d` | diff added-line bg (solid stand-in for green @9% over panel) |
-| delBg      | `#241819` | diff deleted-line bg (solid stand-in for red @9%) |
-| permBg     | `#1c1915` | permission card bg (amber @6% over panel) |
-| selBg      | `#1d2114` | selected picker/suggestion row (accent @8%) |
-| addInk     | `#bdeed7` | added-line text |
-| delInk     | `#f2c4c4` | deleted-line text |
-| onAccent   | `#000000` | text on accent or amber fills |
-
-Terminals cannot alpha-blend or glow — the four solid tints above ARE the
-translation of the prototype's rgba/glow effects; do not attempt gradients,
-shadows, or blur. Truecolor hex throughout; lipgloss downsamples on limited
-terminals (keep the existing explanatory comment style in theme.go).
-
-Keep the `tuiTheme` single-value pattern but rebuild the field set around
-these tokens (style per role: badge, userPrompt, sayText, finalRail, toolName,
-toolTarget, toolArg, autoTag, statusOk/Err, diffAdd/Del/Meta, permBadge,
-permRisk, grepLoc, bashPrompt, stGroup, etc.). No hex outside theme.go.
-
-### Layout — single-column chat surface (matches the existing architecture)
-```
-zero / ~/dev/zero                    anthropic/claude-sonnet-4-5 · 200K
-────────────────────────────────────────────────────────────────────────
-❯ add a --version flag to the cli                              (user)
-I'll add a --version flag…▌                          (interim, muted)
-╭ grep internal/cli  flag\.|RegisterFlag                ⠋ ──────────╮
-│ internal/cli/root.go:41  fs := flag.NewFlagSet("zero"…           │
-╰ 3 matches ───────────────────────────────────────────────────────╯
-╭ edit internal/cli/root.go                    auto  ✓ ────────────╮
-│ internal/cli/root.go            new file          +6 −1          │
-│   41 │   fs.BoolVar(&o.Version, "version", false, …   (addBg)    │
-╰──────────────────────────────────────────────────────────────────╯
-│ Done — the CLI now prints its version.       (final: accent rail)
-● done · 2 tools · 8.4s                                  (done line)
-
-❯ describe a task for zero…                                (composer)
-────────────────────────────────────────────────────────────────────────
-● anthropic │ claude-sonnet-4-5 │        │ 12.4K tok │ interactive │ ⏵⏵ auto
-```
-Render order stays: title line → rule → transcript → composer → rule →
-status line. NO alt-screen change, NO panes, NO mouse — keep `run.go`
-exactly as-is including its no-mouse-capture comment.
-
-### Title bar (replaces `headerBar`)
-Left: badge — `lipgloss` style bg=accent fg=onAccent bold, content ` 0 ` —
-then `zero` (ink, bold, lowercase), ` / ` (faintest), cwd via existing
-`shortenPath` (muted), git branch (faint) when present. Right:
-`provider/model` (ink) + ` · ctx` (faint) where ctx comes from
-`modelregistry.ContextLimits.ContextWindow` formatted like `200K`. Reuse the
-existing width-candidate fallback pattern (`startupHeaderLine`) so segments
-drop gracefully: full → drop ctx → drop cwd → badge+model only. Below it,
-one rule line in `line`.
-
-### Empty state (DELETE the splash entirely)
-Remove `startupView`, `startupHeader`, `commandChips`, the figlet
-`zeroLogoLines`, gap math, and the `showSplash` field/branch. The app boots
-into the chat surface; when the transcript is empty, the stream area shows,
-vertically centered:
-- a 5–6 line block-art `0` in accent (derive the glyph from the `O` columns
-  of the old figlet constant before deleting it),
-- tagline (muted): `a std-lib-first coding agent · bring your own key · no lock-in`,
-- hint (faint): `running zero against <provider/model>` with the model in ink,
-- three suggestion chips, one per line: `❯` (accent bold) + suggestion (ink)
-  inside a `line`-bordered rounded box; pressing `1`–`3` while the composer
-  is empty inserts that suggestion. Define suggestions as a string slice in
-  the tui package (e.g. add a --version flag / explain internal/agent/loop.go /
-  fix the failing test in internal/tools).
-
-### Stream blocks (restyle `transcript.go` / `rendering.go` rows)
-- **user** (`rowUser`): `❯ ` accent bold + ink text. Blank line above turns.
-- **assistant interim** (`rowAssistant` while streaming): muted text, wrapped
-  to `min(width-4, 74)` columns, trailing `▌` cursor in accent while the
-  turn is pending (drives off existing pending state — no new ticker).
-- **final answer** (last assistant row of a turn): a `│` accent rail gutter
-  on every wrapped line + ink text, same 74-col measure. If distinguishing
-  interim vs final is not already possible from row data, mark the row at
-  append time — do not re-parse text.
-- **done line** (end of a turn): `●` green (red on error) + faint
-  `done · N tools · Xs` with faintest `·` separators, derived from data the
-  model already has (tool rows this turn + existing usage/timing if present;
-  omit segments that don't exist rather than inventing counters).
-- **system / error notes** (`rowSystem`/`rowError` equivalents): one-line
-  bordered note — system: faint on panel with `line` border; denial/error:
-  red text, red-tinted border. Keep content unchanged.
-
-### Tool cards (replaces `titledCard`/tool rows for `rowToolCall`+`rowToolResult`)
-Rounded border card on panel bg. Head row on panel2 with a bottom rule:
-status glyph + tool name (ink bold) + target (muted, middle-truncated) +
-one-line arg via the existing `argHint` (faintest, right side) + optional
-`auto` tag (amber text in amber-bordered mini box) when the call was
-auto-approved by mode or a stored grant. Status glyph: running =
-`bubbles/spinner` MiniDot in accent · ok = `✓` green · error = `✗` red.
-Border tint follows state: line → accent-ish `#5a6b2e` while running →
-red-mix on error (define both as theme tokens). Bodies by result shape,
-reusing the existing detection where present:
-- **diff** (`looksLikeDiff`): head row = path (ink) + `NEW FILE` tag (green
-  on addBg) when applicable + right-aligned `+N` green / `−N` red counts;
-  body lines = 4-col right-aligned line number (faintest) + sign col
-  (green/red) + text — added rows on addBg in addInk, deleted on delBg in
-  delInk, context muted. Keep the 16-line cap + `… N more lines` footer.
-- **read**: same gutter, muted text, head shows path + `Lstart–Lend` (faint).
-- **bash**: cmd row `❯` accent bold + command (ink) above a rule; output
-  muted (stderr-ish in delInk), indented 2; footer `exit 0` green /
-  `exit N` red when an exit code is in the result.
-- **grep**: rows `file:line` (blue) + text (muted, truncated); footer faint
-  `N matches`.
-- Anything else: current generic output rendering, restyled to muted on panel.
-
-### Permission card (restyle the existing `pendingPermission` prompt — keys unchanged)
-Card with amber-mixed border on permBg. Top row: `PERMISSION` badge (onAccent
-on amber, bold) + `risk: <low|medium|high>` (amber) from the request's
-existing risk data. Body: `<tool>` (amber bold) + target (ink) + reason
-(muted). Action row of bordered key-chips: `[a] allow once` (onAccent on
-accent), `[y] always` (ink, accent border), `[d] deny` (ink, red on hover is
-web-only — just red-bordered), `[esc]` hint faint. After resolution collapse
-to one faint line: `allowed once · <tool>` green / `always · <tool>` green /
-`denied · <tool>` red. Ask-user and spec-review prompts get the same card
-language with `line` borders. Do NOT alter `handlePermissionKey`,
-`nextPermissionMode`, or grant-store semantics.
-
-### Composer + status line
-Composer: `❯` accent bold + the existing `bubbles/textinput`, borderless
-(remove the bordered input block), placeholder `describe a task for zero…`,
-switching to `running… ctrl+c to interrupt` while pending; right-aligned
-faint hint `run ↵` (idle, only when input is non-empty) / `esc stop`
-(pending). Rules above composer and above status line in `line`.
-Status line groups separated by ` │ ` (line color): `●` accent + provider ·
-model id · flexible gap · `<X.XK> tok` (existing usage segment data,
-keep cost when priced) · `interactive` · permission mode (reuse
-`modeLabel()`: `⏵⏵ auto-approve` green / `ask` amber / `unsafe` red) ·
-optional green `always: <n> grants` when the sandbox grant store exposes a
-cheap count (omit otherwise — do not add plumbing).
-
-### Overlays (restyle in place, identical key handling)
-- Autocomplete: rows on panel, selected row on selBg with accent `❯` marker,
-  name ink / desc faint.
-- Picker (`/model`, `/provider`, etc.): bordered panel; rows = provider dot
-  (accent remote, blue local) + id (ink) + right meta `ctx · KEY_ENV`
-  (faint/faintest) when the registry/provider catalog already exposes them;
-  selected row selBg. Title row + `↑/↓ · ⏎ · esc` hints faint.
-- Sessions: `/resume` (alias `/sessions`) keeps its existing flow; restyle
-  its list as stacked cards — id (accent) + age (faint) top row, title (ink),
-  meta line (faint with faintest dots). No new persistence behavior.
-
-### Explicitly OUT of scope
-The prototype's `text/json` toggle visualizes the headless
-`--output-format stream-json` CLI mode (see docs/STREAM_JSON_PROTOCOL.md) —
-not a TUI feature. The TweaksPanel in the HTML is a design harness — ignore
-it. No mouse support, no model dropdown (keyboard picker is the translation).
-
----
-
-## Adaptive requirements (treat as acceptance criteria)
-- Width tiers, re-evaluated on every `WindowSizeMsg`:
-  **≥100** full spec · **80–99** drop tool-arg column, header ctx, status
-  `interactive` group · **60–79** drop diff/read line-number gutters, badge
-  renders as `0` without padding, status keeps provider+tokens+mode only ·
-  **<58** (current `minStartupWidth`) single-segment header, cards lose
-  side borders (keep top/bottom rules), composer + mode line only.
-- Text measure: say/final wrap at `min(width-4, 74)`; paths middle-truncate
-  (`internal/…/root.go`); never emit a styled line wider than the terminal
-  (extend the existing `fitStyledLine`/`truncateRunes` helpers — reuse, don't
-  fork).
-- Height: when the terminal is shorter than the frame, transcript trims from
-  the top (existing inline behavior) — verify no overlap of composer/status.
-- Color: nothing may rely on truecolor to be legible — check every pairing
-  at 256 colors mentally; the four tint tokens must remain darker than ink.
-- Tests for the tier logic: table-driven across widths {58, 70, 80, 100, 120}
-  asserting which segments render, using the existing ANSI-stripping helpers.
-
----
-
-## Stages
-
-### Stage 0 — recon + delete the `zeroline` design system  *(tag: `tui-lime-s0`)*
-Recon first: `grep -rni zeroline --include='*.go' .` and read
-`internal/tui/options.go`, `internal/cli/app.go`. Then remove completely:
-`internal/zeroline/` (whole package + tests), `internal/tui/zeroline_view.go`
-(+ test), `internal/cli/zeroline.go`, the `case "zeroline":` dispatch and the
-zeroline help-text line in `internal/cli/app.go`, the skin parameter on
-`runInteractiveTUIWithSkin` (rename to `runInteractiveTUI`), the `Skin`,
-`ThemeVariant`, `ThemeDark` fields in `tui.Options`, the model's `skin`
-field and every `m.skin == "zeroline"` branch (View dispatch, `/theme`
-special-case, update paths, `picker.go`, `image_attach.go`). Update or
-delete affected tests; the zeroline grep must return zero hits. Gate, tag.
-
-### Stage 1 — Lime palette on the existing layout  *(tag: `tui-lime-s1`)*
-Rewrite `theme.go` to the token table. Restyle every current render site
-(header, transcript rows, titledCard/diffCard, overlays, status line,
-prompts) to the new tokens WITHOUT changing structure yet — this isolates
-the string-assertion churn in `model_test.go` / `session_test.go` /
-`command_polish_test.go` from the layout work. Prefer asserting plain
-content via the ANSI-stripping helpers over styled bytes. Gate, tag.
-
-### Stage 2 — chat-surface components  *(tag: `tui-lime-s2`)*
-Title bar + rule, empty state replacing the splash (splash code deleted),
-stream blocks (user/interim/final/done/notes), tool cards with all four
-result bodies, MiniDot spinner, composer + status line redesign. Add
-rendering tests per block type with synthetic rows. Gate, tag.
-
-### Stage 3 — interactive surfaces  *(tag: `tui-lime-s3`)*
-Permission card + resolved collapse + auto tag, ask-user and spec-review
-cards, autocomplete/picker restyle with model-row metadata, `/resume`
-sessions card list, suggestion-chip `1–3` insertion. Key handling and
-semantics byte-identical to before — assert that in tests where they
-already exist. Gate, tag.
-
-### Stage 4 — adaptive polish + docs  *(tag: `tui-lime-s4`)*
-Implement + test the width tiers, sweep remaining render sites (spec mode,
-command center, doctor output if styled) for stray old-palette styles,
-update `internal/cli/app.go` help text and README's TUI section, manual run
-`go run ./cmd/zero` at 58/80/100/120 cols (no wrapped rules, no orphan ANSI,
-prompts render correctly). Full gate + `go run ./cmd/zero-release smoke`.
-Gate, tag.
-
----
-
-## Hard constraints
-- No new module dependencies; charmbracelet stack only. No mouse capture.
-- Zero behavior change to permissions, grants, ask-user, spec review,
-  sessions, slash commands, autocomplete, image attach, compaction, rewind.
-- All colors in `theme.go`; renderers consume named styles only.
-- Reuse existing helpers (`fitStyledLine`, `truncateRunes`, `argHint`,
-  `shortenPath`, header candidates) instead of writing parallel ones.
-- Comment style matches the repo: explain WHY at non-obvious decisions.
-
-## Definition of done
-One design in the tree (no zeroline references, no splash, no cyan tokens),
-the Lime chat surface matching this spec at ≥100 cols and degrading per the
-width tiers below that, every pre-existing flow keyboard-reachable and
-visually consistent, `go test ./...` + release build + smoke fully green.

docs/oauth-subscriptions.md

@@ -170,7 +170,7 @@ preset (no API key — the proxy authenticates):
       "name": "chatgpt",
       "catalogID": "chatgpt-proxy",     // OpenAI-compatible, local, no key
       "baseURL": "http://localhost:10531/v1", // override for your proxy's port
-      "model": "gpt-5"                   // whatever model your proxy serves
+      "model": "gpt-5.5"                 // whatever model your proxy serves
     }
   ]
 }

docs/superpowers/plans/2026-06-12-agent-eval-benchmark-harness.md

@@ -1,304 +0,0 @@
-# Agent Eval Benchmark Harness Implementation Plan
-
-> **For agentic workers:** REQUIRED SUB-SKILL: Use superpowers:subagent-driven-development (recommended) or superpowers:executing-plans to implement this plan task-by-task. Steps use checkbox (`- [ ]`) syntax for tracking.
-
-**Goal:** Build a local, offline-testable benchmark harness that copies eval fixtures, runs an agent command per task, scores the resulting workspace, and reports aggregate results.
-
-**Architecture:** Keep `internal/agenteval.Runner` as the scoring primitive. Add a `Harness` layer that materializes fixture workspaces into a work directory, initializes a clean Git baseline, invokes an injectable `AgentRunner`, then calls `Runner.Run`. The CLI exposes this as `zero eval bench`, while `zero eval run` remains the lower-level scorer for an already-mutated workspace.
-
-**Tech Stack:** Go 1.24 standard library, existing `internal/agenteval`, existing `internal/cli`, Git CLI for workspace baselines.
-
----
-
-## File Structure
-
-- Create `internal/agenteval/materialize.go`: fixture path resolution, directory copy, Git init/baseline helpers.
-- Create `internal/agenteval/materialize_test.go`: fixture copy and Git baseline tests.
-- Create `internal/agenteval/agent_command.go`: command-template based agent runner with `{prompt}`, `{workspace}`, `{task_id}` placeholders.
-- Create `internal/agenteval/agent_command_test.go`: placeholder substitution, cwd, exit/error behavior.
-- Create `internal/agenteval/benchmark.go`: benchmark orchestration and aggregate report.
-- Create `internal/agenteval/benchmark_test.go`: end-to-end harness with fake agent runner and sample fixtures.
-- Modify `internal/cli/agent_eval.go`: parse `zero eval bench` and wire `Harness`.
-- Modify `internal/cli/agent_eval_test.go`: bench parsing, JSON/text output, report writing.
-- Modify `docs/AGENT_EVALS.md`: document `bench` mode and real local command examples.
-
-## Task 1: Fixture Materializer
-
-**Files:**
-- Create: `internal/agenteval/materialize.go`
-- Create: `internal/agenteval/materialize_test.go`
-
-- [ ] **Step 1: Write failing tests**
-
-```go
-func TestMaterializeTaskCopiesFixtureAndInitializesGit(t *testing.T) {
-	suitePath := filepath.Join("testdata", "sample_suite.json")
-	suite, err := LoadSuite(suitePath)
-	if err != nil {
-		t.Fatal(err)
-	}
-	task := suite.Tasks[0]
-	workRoot := t.TempDir()
-
-	workspace, err := Materializer{}.MaterializeTask(context.Background(), suitePath, task, MaterializeInput{WorkRoot: workRoot})
-	if err != nil {
-		t.Fatalf("MaterializeTask: %v", err)
-	}
-	if _, err := os.Stat(filepath.Join(workspace.Path, "go.mod")); err != nil {
-		t.Fatalf("fixture was not copied: %v", err)
-	}
-	if output, err := exec.Command("git", "-C", workspace.Path, "status", "--porcelain").CombinedOutput(); err != nil || strings.TrimSpace(string(output)) != "" {
-		t.Fatalf("workspace baseline is dirty: err=%v output=%s", err, output)
-	}
-}
-```
-
-- [ ] **Step 2: Run test to verify it fails**
-
-Run: `go test ./internal/agenteval -run TestMaterializeTask -count=1`
-Expected: FAIL because `Materializer` is undefined.
-
-- [ ] **Step 3: Implement minimal materializer**
-
-Create a `Materializer` with:
-- `MaterializeTask(ctx, suitePath string, task Task, input MaterializeInput) (Workspace, error)`
-- `MaterializeInput{WorkRoot string}`
-- `Workspace{Path string, TaskID string, FixturePath string}`
-- Resolve relative `task.WorkspaceFixture` from `filepath.Dir(suitePath)`.
-- Copy directories recursively using stdlib only.
-- Skip `.git` while copying.
-- Run `git init`, `git add .`, and `git commit -m "baseline"` with local user config/env.
-- Return clear errors for missing fixture, absolute fixture escapes, and empty work root.
-
-- [ ] **Step 4: Verify**
-
-Run:
-```bash
-go test ./internal/agenteval -run 'TestMaterializeTask|TestMaterializer' -count=1
-```
-
-## Task 2: Agent Command Runner
-
-**Files:**
-- Create: `internal/agenteval/agent_command.go`
-- Create: `internal/agenteval/agent_command_test.go`
-
-- [ ] **Step 1: Write failing tests**
-
-```go
-func TestCommandAgentRunnerExpandsPlaceholders(t *testing.T) {
-	dir := t.TempDir()
-	script := filepath.Join(dir, "agent.bat")
-	if err := os.WriteFile(script, []byte("@echo %1|%2|%3> out.txt\r\n"), 0o755); err != nil {
-		t.Fatal(err)
-	}
-	runner := CommandAgentRunner{Command: []string{script, "{task_id}", "{workspace}", "{prompt}"}}
-	result := runner.Run(context.Background(), AgentRunInput{TaskID: "task-a", WorkspacePath: dir, Prompt: "fix bug"})
-	if result.ExitCode != 0 || result.Error != "" {
-		t.Fatalf("Run = %#v", result)
-	}
-	data, err := os.ReadFile(filepath.Join(dir, "out.txt"))
-	if err != nil {
-		t.Fatal(err)
-	}
-	if !strings.Contains(string(data), "task-a") || !strings.Contains(string(data), dir) || !strings.Contains(string(data), "fix bug") {
-		t.Fatalf("placeholders not expanded: %q", data)
-	}
-}
-```
-
-- [ ] **Step 2: Run test to verify it fails**
-
-Run: `go test ./internal/agenteval -run TestCommandAgentRunner -count=1`
-Expected: FAIL because `CommandAgentRunner` is undefined.
-
-- [ ] **Step 3: Implement runner**
-
-Add:
-- `type AgentRunInput struct { TaskID, Prompt, WorkspacePath string }`
-- `type AgentRunResult struct { ExitCode int; Stdout, Stderr, Error string }`
-- `type AgentRunner interface { Run(context.Context, AgentRunInput) AgentRunResult }`
-- `type CommandAgentRunner struct { Command []string }`
-- `Run` executes without shell interpolation, with `cmd.Dir = WorkspacePath`.
-- Replace `{prompt}`, `{workspace}`, `{task_id}` in every arg.
-- Empty command returns `ExitCode:-1` and `Error:"agent command is required"`.
-
-- [ ] **Step 4: Verify**
-
-Run: `go test ./internal/agenteval -run TestCommandAgentRunner -count=1`
-
-## Task 3: Benchmark Harness
-
-**Files:**
-- Create: `internal/agenteval/benchmark.go`
-- Create: `internal/agenteval/benchmark_test.go`
-
-- [ ] **Step 1: Write failing tests**
-
-```go
-func TestHarnessRunsTaskFromFixtureAndScoresResult(t *testing.T) {
-	suitePath := filepath.Join("testdata", "sample_suite.json")
-	suite, err := LoadSuite(suitePath)
-	if err != nil {
-		t.Fatal(err)
-	}
-	harness := Harness{
-		Materializer: Materializer{},
-		Agent: agentRunnerFunc(func(ctx context.Context, input AgentRunInput) AgentRunResult {
-			target := filepath.Join(input.WorkspacePath, "docs", "STREAM_JSON_PROTOCOL.md")
-			err := os.WriteFile(target, []byte("updated"), 0o644)
-			if err != nil {
-				return AgentRunResult{ExitCode: -1, Error: err.Error()}
-			}
-			return AgentRunResult{ExitCode: 0}
-		}),
-		Runner: Runner{},
-	}
-	report := harness.Run(context.Background(), suitePath, suite, BenchmarkInput{
-		TaskID:   "document-stream-json-verify-events",
-		WorkRoot: t.TempDir(),
-	})
-	if report.Summary.TotalTasks != 1 || report.Summary.PassedTasks != 1 || !report.OK {
-		t.Fatalf("report = %#v", report)
-	}
-}
-```
-
-- [ ] **Step 2: Run test to verify it fails**
-
-Run: `go test ./internal/agenteval -run TestHarness -count=1`
-Expected: FAIL because `Harness` is undefined.
-
-- [ ] **Step 3: Implement harness**
-
-Add:
-- `BenchmarkInput{TaskID, WorkRoot string; KeepWorkspaces bool}`
-- `BenchmarkReport{Contract, SuiteID, OK, Summary, Tasks}`
-- `BenchmarkTaskReport{TaskID, WorkspacePath, FixturePath, Agent AgentRunResult, Report Report}`
-- `BenchmarkSummary{TotalTasks, PassedTasks, FailedTasks, BlockedTasks, ErrorTasks int}`
-- `Harness{Materializer Materializer; Agent AgentRunner; Runner Runner}`
-- If no `Agent`, produce blocked task reports with message `agent command is required`.
-- Select one task when `TaskID` set, otherwise all tasks.
-- For each task: materialize, run agent, if agent fails mark blocked; otherwise score with `Runner.Run`.
-- Remove each materialized task workspace after scoring unless `BenchmarkInput.KeepWorkspaces` is true.
-
-- [ ] **Step 4: Verify**
-
-Run: `go test ./internal/agenteval -run 'TestHarness|TestBenchmark' -count=1`
-
-## Task 4: CLI `zero eval bench`
-
-**Files:**
-- Modify: `internal/cli/agent_eval.go`
-- Modify: `internal/cli/agent_eval_test.go`
-
-- [ ] **Step 1: Write failing tests**
-
-```go
-func TestRunEvalBenchJSONModePassesHarnessOptions(t *testing.T) {
-	var stdout, stderr bytes.Buffer
-	exitCode := runWithDeps([]string{
-		"eval", "bench",
-		"--suite", "evals/context.json",
-		"--task", "edit-reader",
-		"--work-root", "D:\\tmp\\zero-evals",
-		"--agent-command", "zero", "exec", "{prompt}",
-		"--json",
-	}, &stdout, &stderr, appDeps{
-		runAgentEval: func(ctx context.Context, options agentEvalOptions) (agentEvalReport, error) {
-			if options.Mode != "bench" || options.TaskID != "edit-reader" || options.WorkRoot == "" || len(options.AgentCommand) != 3 {
-				t.Fatalf("unexpected bench options: %#v", options)
-			}
-			return agentEvalReport{Suite: "quality-context", Status: "pass", OK: true, Total: 1, Passed: 1}, nil
-		},
-	})
-	if exitCode != exitSuccess {
-		t.Fatalf("exit=%d stderr=%s", exitCode, stderr.String())
-	}
-}
-```
-
-- [ ] **Step 2: Run test to verify it fails**
-
-Run: `go test ./internal/cli -run TestRunEvalBench -count=1`
-Expected: FAIL because bench mode is not parsed.
-
-- [ ] **Step 3: Implement CLI**
-
-Add parser support:
-- Modes: `validate`, `run`, `bench`
-- Bench flags: `--work-root <path>`, `--agent-command <argv...>`, `--keep-workspaces`
-- `--task`, `--report-dir`, `--json` work in bench mode.
-- `--workspace` remains run-only.
-- Default work root for bench: temp dir with prefix `zero-eval-`.
-- Default `AgentCommand` empty -> harness report blocked rather than usage error.
-- Convert benchmark aggregate to existing CLI report shape.
-
-- [ ] **Step 4: Verify**
-
-Run:
-```bash
-go test ./internal/cli -run 'TestRunEvalBench|TestRunEvalRun' -count=1
-```
-
-## Task 5: Docs and Manual Examples
-
-**Files:**
-- Modify: `docs/AGENT_EVALS.md`
-
-- [ ] **Step 1: Document modes**
-
-Add a `bench` subsection:
-- `validate`: schema only
-- `run`: score an already-mutated worktree
-- `bench`: copy fixture, run agent command, score result
-
-- [ ] **Step 2: Include examples**
-
-Document:
-```bash
-go run ./cmd/zero eval bench \
-  --suite internal/agenteval/testdata/sample_suite.json \
-  --task document-stream-json-verify-events \
-  --work-root /tmp/zero-evals \
-  --agent-command zero exec --cwd {workspace} {prompt}
-```
-
-Also document a deterministic fake-agent example for local testing.
-
-- [ ] **Step 3: Verify**
-
-Run:
-```bash
-go run ./cmd/zero eval --suite internal/agenteval/testdata/sample_suite.json
-go test ./...
-```
-
-## Task 6: Integration Validation
-
-**Files:**
-- Modify as needed only in files above.
-
-- [ ] **Step 1: Run focused validation**
-
-```bash
-go test -count=1 ./internal/agenteval ./internal/cli
-```
-
-- [ ] **Step 2: Run fixture validation**
-
-```bash
-go test ./...
-```
-
-from `internal/agenteval/testdata/fixtures/zero-mini`.
-
-- [ ] **Step 3: Run full repo validation**
-
-```bash
-gofmt -l internal/agenteval internal/cli
-git diff --check
-go test ./...
-go run ./cmd/zero-release build
-go run ./cmd/zero-release smoke
-```

docs/superpowers/plans/2026-06-12-agent-eval-d-level-upgrades.md

@@ -1,181 +0,0 @@
-# Agent Eval D-Level Upgrades Implementation Plan
-
-> **For agentic workers:** REQUIRED SUB-SKILL: Use superpowers:subagent-driven-development (recommended) or superpowers:executing-plans to implement this plan task-by-task. Steps use checkbox (`- [ ]`) syntax for tracking.
-
-**Goal:** Upgrade the local agent eval harness from a smoke benchmark into a quality signal that can compare models, score traces, catch context mistakes, and preserve regression artifacts.
-
-**Architecture:** Keep the existing `internal/agenteval` harness as the core. Add small schema fields to `Task`, score them through `Score` and `Harness.Run`, then expose model matrix and full benchmark report data through `zero eval bench`. Avoid provider calls or UI work.
-
-**Tech Stack:** Go 1.24, standard library JSON/process/git helpers, existing `go test ./...` validation.
-
----
-
-### Task 1: Suite Schema And Rubric Checks
-
-**Files:**
-- Modify: `internal/agenteval/suite.go`
-- Modify: `internal/agenteval/score.go`
-- Test: `internal/agenteval/agenteval_test.go`
-
-- [ ] **Step 1: Write failing schema/rubric tests**
-
-Add tests that load and normalize `tags`, `difficulty`, and `forbiddenChangedFiles`, reject malformed forbidden paths, and fail scoring when a forbidden file is touched.
-
-- [ ] **Step 2: Run tests to verify failure**
-
-Run: `go test ./internal/agenteval -run "LoadSuite|Validate|Forbidden" -count=1`
-
-- [ ] **Step 3: Implement minimal schema and scoring**
-
-Add these fields:
-
-```go
-Tags                  []string `json:"tags,omitempty"`
-Difficulty            string   `json:"difficulty,omitempty"`
-ForbiddenChangedFiles []string `json:"forbiddenChangedFiles,omitempty"`
-```
-
-Normalize file lists, validate duplicate/malformed entries, and add a `forbidden_changed_files` result.
-
-- [ ] **Step 4: Verify green**
-
-Run: `go test ./internal/agenteval -count=1`
-
-### Task 2: Agent Trace Scoring
-
-**Files:**
-- Create: `internal/agenteval/trace.go`
-- Test: `internal/agenteval/trace_test.go`
-- Modify: `internal/agenteval/benchmark.go`
-
-- [ ] **Step 1: Write failing trace tests**
-
-Cover JSONL events such as:
-
-```json
-{"type":"tool","name":"read_file"}
-{"event":"verify","name":"go-test"}
-```
-
-Require deterministic event keys like `tool:read_file` and `verify:go-test`.
-
-- [ ] **Step 2: Run tests to verify failure**
-
-Run: `go test ./internal/agenteval -run Trace -count=1`
-
-- [ ] **Step 3: Implement parser and benchmark scoring**
-
-Parse agent stdout line by line, ignore non-JSON noise, and append a `trace` result when a task declares `requiredTraceEvents`.
-
-- [ ] **Step 4: Verify green**
-
-Run: `go test ./internal/agenteval -count=1`
-
-### Task 3: Context Quality Checks
-
-**Files:**
-- Create: `internal/agenteval/context_checks.go`
-- Test: `internal/agenteval/context_checks_test.go`
-- Modify: `internal/agenteval/benchmark.go`
-
-- [ ] **Step 1: Write failing context tests**
-
-Cover a task that declares:
-
-```json
-"contextChecks": {
-  "requiredFiles": ["docs/STREAM_JSON_PROTOCOL.md"],
-  "forbiddenFiles": ["node_modules/cache.txt"]
-}
-```
-
-- [ ] **Step 2: Run tests to verify failure**
-
-Run: `go test ./internal/agenteval -run Context -count=1`
-
-- [ ] **Step 3: Implement fixture/workspace context checks**
-
-Check required files exist under the materialized workspace and forbidden files do not. Append a `context` result before summary finalization.
-
-- [ ] **Step 4: Verify green**
-
-Run: `go test ./internal/agenteval -count=1`
-
-### Task 4: Model Matrix Benchmarking
-
-**Files:**
-- Modify: `internal/agenteval/agent_command.go`
-- Modify: `internal/agenteval/benchmark.go`
-- Test: `internal/agenteval/agent_command_test.go`
-- Test: `internal/agenteval/benchmark_test.go`
-
-- [ ] **Step 1: Write failing model-matrix tests**
-
-Assert `{model}` expands in agent command argv and `BenchmarkInput{Models: []string{"a","b"}}` runs every task for each model.
-
-- [ ] **Step 2: Run tests to verify failure**
-
-Run: `go test ./internal/agenteval -run "Model|Harness" -count=1`
-
-- [ ] **Step 3: Implement model propagation**
-
-Add `Model` to `AgentRunInput` and `BenchmarkTaskReport`; add `Models []string` to `BenchmarkInput`; loop task/model pairs deterministically.
-
-- [ ] **Step 4: Verify green**
-
-Run: `go test ./internal/agenteval -count=1`
-
-### Task 5: CLI Flags And Regression Artifacts
-
-**Files:**
-- Modify: `internal/cli/agent_eval.go`
-- Test: `internal/cli/agent_eval_test.go`
-
-- [ ] **Step 1: Write failing CLI tests**
-
-Cover repeated `--model`, comma-separated `--models`, help text, and report-dir JSON containing nested benchmark detail.
-
-- [ ] **Step 2: Run tests to verify failure**
-
-Run: `go test ./internal/cli -run Eval -count=1`
-
-- [ ] **Step 3: Implement CLI wiring**
-
-Pass models into `agenteval.BenchmarkInput`, include benchmark detail in `agentEvalReport`, and prefix failure IDs with model when present.
-
-- [ ] **Step 4: Verify green**
-
-Run: `go test ./internal/cli -count=1`
-
-### Task 6: Suite Expansion And Docs
-
-**Files:**
-- Modify: `internal/agenteval/testdata/sample_suite.json`
-- Modify: `docs/AGENT_EVALS.md`
-
-- [ ] **Step 1: Write failing suite expectation**
-
-Update the sample suite test to require richer coverage count and metadata fields.
-
-- [ ] **Step 2: Run tests to verify failure**
-
-Run: `go test ./internal/agenteval -run SampleSuite -count=1`
-
-- [ ] **Step 3: Expand the fixture suite and docs**
-
-Add more tasks using the existing `zero-mini` fixture, plus examples for `--model`, `{model}`, `requiredTraceEvents`, `contextChecks`, and report artifacts.
-
-- [ ] **Step 4: Verify green**
-
-Run: `go test ./internal/agenteval ./internal/cli -count=1`
-
-### Final Verification
-
-- [ ] Run `gofmt -l internal/agenteval internal/cli`
-- [ ] Run `git diff --check`
-- [ ] Run `go vet ./...`
-- [ ] Run `go test ./...`
-- [ ] Run `go run ./cmd/zero eval --suite internal/agenteval/testdata/sample_suite.json`
-- [ ] Run `go run ./cmd/zero eval bench --suite internal/agenteval/testdata/sample_suite.json --task document-stream-json-verify-events --model test-model --json --agent-command powershell -NoProfile -Command "& { param(`$ws) Set-Content -LiteralPath (Join-Path `$ws 'docs/STREAM_JSON_PROTOCOL.md') -Value updated; Write-Output '{\"type\":\"tool\",\"name\":\"read_file\"}' }" "{workspace}"`
-- [ ] Run `go run ./cmd/zero-release build`
-- [ ] Commit and push to `feat/agent-eval-harness`

docs/superpowers/specs/2026-06-10-additional-write-roots-design.md

@@ -1,175 +0,0 @@
-# Additional write roots (`--add-dir` / `/add-dir`)
-
-**Date:** 2026-06-10
-**Status:** Approved
-
-## Problem
-
-Zero's sandbox confines all writes to the workspace root, with no escape hatch.
-`EnforceWorkspace` is hardcoded `true` in `DefaultPolicy()`
-(`internal/sandbox/types.go`) and `SandboxConfig` exposes only `maxAutonomy`
-and `network`. When a user asks the agent to save output outside the workspace
-(e.g. "save this to ~/Desktop/dev/untitled folder"), every layer denies it:
-
-1. **Policy engine** — `Engine.Evaluate` → `validateWorkspacePaths`
-   (`internal/sandbox/paths.go`) denies tool requests whose paths fall outside
-   the workspace root.
-2. **macOS seatbelt** — `sandboxExecProfile` (`internal/sandbox/runner.go`)
-   emits `(allow file-write* (subpath "<workspaceRoot>"))`, so shell commands
-   fail with `Operation not permitted`.
-3. **Linux bubblewrap** — `bubblewrapCommandPlan` binds only the workspace
-   read-write; system paths are read-only binds.
-4. **File tools** — `resolveWorkspacePath` / `recheckWorkspaceWriteTarget`
-   (`internal/tools/workspace.go`) confine `write_file`, `edit`, `read_file`,
-   `list_directory`, `glob`, `grep`, and the bash tool's `cwd`.
-
-The agent's only recourse is workarounds (zip in workspace, ask the user to
-move it), and the denial surfaces as a raw OS error with no guidance.
-
-## Goal
-
-Let users grant zero write access to specific directories outside the
-workspace — at launch (flag/config) and mid-session (TUI command) — without
-weakening the sandbox elsewhere. Decided UX: **launch flag + runtime command**
-(prompt-on-denial may layer on later).
-
-## Non-goals
-
-- No interactive prompt-on-denial flow in this iteration.
-- No persistence of `/add-dir` grants across sessions (config is the durable
-  surface).
-- No widening of network or destructive-shell policy.
-- No `enforceWorkspace: false` config knob.
-
-## Design
-
-### 1. Core type: `sandbox.Scope`
-
-New concurrency-safe type in `internal/sandbox` (e.g. `scope.go`):
-
-```go
-type Scope struct {
-    mu            sync.RWMutex
-    workspaceRoot string   // normalized at construction
-    extraRoots    []string // normalized, deduplicated
-}
-
-func NewScope(workspaceRoot string, extras []string) (*Scope, error)
-func (s *Scope) Roots() []string            // workspace first, then extras
-func (s *Scope) Add(path string) (string, error) // returns normalized root
-func (s *Scope) validate(path string) *pathViolation // package-internal
-```
-
-`Add` normalization and guardrails:
-
-- Expand a leading `~` to the user home directory.
-- Make absolute, `filepath.Clean`, then `filepath.EvalSymlinks`.
-- The path must exist and be a directory (no auto-create).
-- Reject the filesystem root `/` (and Windows volume roots) — granting it
-  would disable confinement entirely.
-- Adding the workspace root or a path inside any existing root is a
-  no-op (idempotent success, not an error).
-
-`Validate` generalizes today's `validateWorkspacePath`: a path is allowed if
-it validates under **any** root, including the per-segment symlink-traversal
-check against that root. The existing single-root function remains as the
-per-root helper.
-
-One `*Scope` is constructed at startup and shared by the engine, the command
-runner, and the tool registry, so a mid-session `Add` is immediately visible
-to all layers.
-
-### 2. Policy engine
-
-- `EngineOptions` gains `Scope *Scope`; `NewEngine` builds a workspace-only
-  scope when nil (back-compat for tests/embedders).
-- `Evaluate`'s workspace check (engine.go) calls `scope.Validate` per
-  requested path instead of `validateWorkspacePaths(root, request)`.
-- Risk classification (`risk.go`) treats paths inside any scope root as
-  in-workspace, so writes to granted roots are not classified
-  `out_of_workspace`/critical.
-- Denial reason becomes actionable:
-  `"<path> is outside the workspace (use /add-dir or --add-dir to allow writes there)"`.
-
-### 3. OS runners
-
-- `sandboxExecProfile` emits one `(subpath "...")` per scope root inside the
-  single `(allow file-write* ...)` rule.
-- `bubblewrapCommandPlan` adds `--bind <root> <root>` (read-write, real path)
-  for each extra root. The workspace keeps its `/workspace` remount; extra
-  roots keep their host paths so absolute paths in commands work unchanged.
-- `resolveCommandDir` accepts a command `cwd` inside any scope root.
-- Profiles are built per command plan from `scope.Roots()`, so commands run
-  after `/add-dir` get the widened profile with no restart.
-
-### 4. File tools
-
-- New narrow interface in `internal/tools`:
-  `type PathScope interface { Roots() []string }` (satisfied by
-  `*sandbox.Scope`) to avoid a hard dependency direction problem.
-- `resolveWorkspacePath` and `recheckWorkspaceWriteTarget` gain scope-aware
-  variants: absolute paths are resolved against the first root that contains
-  them; **relative paths still resolve against the workspace root only**.
-- Tool constructors (`CoreReadOnlyTools`, `CoreWriteTools`, `CoreShellTools`,
-  `CoreTools` in `registry.go`) accept the scope (workspace-only default kept
-  via a convenience wrapper so existing call sites/tests keep compiling).
-- Result: `write_file`, `edit`, `read_file`, `list_directory`, `glob`, `grep`,
-  and bash `cwd` all work consistently inside extra roots. Write access
-  implies read access within a granted root.
-
-### 5. Config, CLI, TUI
-
-- **Config:** `SandboxConfig` gains
-  `AdditionalWriteRoots []string \`json:"additionalWriteRoots,omitempty"\``.
-  The key is honored from the **global user config**
-  (`os.UserConfigDir()/zero/config.json`) so a grant can apply to every
-  project, and from CLI flags/overrides. It is **deliberately excluded from
-  the project config allowlist** (`mergeProjectConfig`): a repo-controlled
-  `.zero/config.json` must not be able to widen write access outside the
-  workspace. Sources are merged as a union (append + dedupe), not replace.
-- **CLI:** repeatable `--add-dir <path>` flag on `zero` (TUI) and `zero exec`.
-  Effective set = union(flag, config), normalized through `Scope` at startup;
-  an invalid root fails fast with a clear error.
-- **TUI:** new `/add-dir` command (`internal/tui/commands.go`):
-  - `/add-dir` (bare) — lists current write roots.
-  - `/add-dir <path>` — calls `scope.Add`; on success appends a system line:
-    `write access added: <path> (this session only)`; on failure shows the
-    normalization error.
-- **Observability:** `zero sandbox` status output and sandbox snapshots
-  (`internal/zerocommands/sandbox_snapshots.go`) list extra roots.
-
-### 6. Security invariants
-
-- Extra roots only widen *file-write* (and file-tool read) access under the
-  named directories; network and destructive-shell policy are untouched.
-- Symlink-traversal protection applies per root exactly as it does for the
-  workspace today.
-- `/` and volume roots are rejected; roots are symlink-resolved before being
-  trusted so a symlinked grant cannot silently point elsewhere later.
-- Runtime grants are session-only; durable grants require editing config.
-
-### 7. Testing
-
-- **Scope unit tests:** normalization (`~`, relative, symlink), rejection of
-  `/`, missing dirs, files; idempotent adds; multi-root `Validate` including
-  symlink-traversal escapes from one root into another denied area.
-- **Runner tests:** seatbelt profile string contains one `subpath` per root;
-  bubblewrap args contain the extra `--bind`s; `resolveCommandDir` accepts
-  extra-root cwd (extend `runner_test.go`).
-- **Engine tests:** out-of-workspace write allowed when inside an extra root;
-  still denied (with the new actionable message) when outside all roots.
-- **Tool tests:** `write_file`/`edit`/`read_file` in an extra root succeed;
-  relative paths never resolve against extra roots (extend
-  `file_tools_test.go`).
-- **TUI tests:** `parseCommand` recognizes `/add-dir`; handler list/add/error
-  paths.
-- **Config/CLI tests:** flag-config union, fail-fast on invalid root.
-- Platform-gated integration tests follow the existing sandbox test patterns.
-
-## Implementation order
-
-1. `sandbox.Scope` + tests.
-2. Engine + risk + runner integration (flag-less, scope seeded empty).
-3. File-tool scope plumbing.
-4. Config key + CLI flags.
-5. TUI `/add-dir` command + status output.

docs/superpowers/specs/2026-06-11-permission-scope-grants-design.md

@@ -1,107 +0,0 @@
-# Permission-UX Phase 1 — scope-enforced grants
-
-Status: approved (2026-06-11). Branch: `feat/permission-ux` (continues Phase 0, commit `742ff48`).
-
-## Problem
-
-Phase 0 derives a human-readable *scope* for a tool call (the file, directory, or
-cwd it touches) and shows it on the permission card and the persisted decision
-row. But the grant that an "always allow" writes is still **tool-wide**:
-
-- `persistPermissionGrant` (agent/loop.go) stores only `ToolName`.
-- the grant file is keyed by tool name: `state.Grants[toolName]`.
-- enforcement (`engine.Decide` → `store.Lookup(toolName, autonomy)`, engine.go:100)
-  matches on tool name alone.
-
-So clicking "always allow" on a write to `src/main.go` silently authorizes *every*
-future write — the blind-yes the Phase-0 card was written to warn against.
-
-## Goal
-
-Persist and enforce the scope so an "always" grant covers exactly what the card
-showed: that file, or that directory subtree — nothing else.
-
-## Decisions
-
-1. **Match semantics:** a `file` scope re-matches only its exact path; a `dir`
-   scope matches that directory and any descendant; an empty scope (a call with no
-   path-like arg, e.g. `bash` with no `cwd`) stays tool-wide. A tool-wide *request*
-   is **not** covered by a narrower grant — it re-prompts (fail-safe).
-2. **Path anchoring:** scopes are normalized to an absolute, cleaned path at grant
-   time, anchored to the workspace root. Matching compares absolute paths, so
-   `./src/main.go` == `src/main.go` == `/proj/src/main.go`, and a grant made in
-   `/proj-a` never matches `/proj-b`. The card still displays the short relative form.
-3. **Storage:** `grants` becomes `map[toolName][]Grant`, one record per
-   (scope, decision). Schema bumps `1 → 2`; a v1 file is migrated by reading each
-   entry as a single tool-wide grant, so existing grants keep working unchanged.
-4. **Precedence:** among grants that cover a request, a covering **deny wins**
-   (regardless of autonomy); otherwise the most-specific covering **allow**
-   (file > dir > tool-wide; ties broken by longer path, then newest) whose
-   `MaxAutonomy` satisfies the request.
-
-## Design
-
-### Scope derivation — single source of truth (`internal/sandbox/grant_scope.go`)
-
-`DeriveScope(toolName, args) (raw string, kind ScopeKind)` is the one place that
-knows which args are path-like, reusing Phase 0's key priority (most specific
-first): `path`,`file` → `file`; `directory`,`dir`,`cwd` → `dir`. Empty / `.` →
-tool-wide. `resolveScopeAbs(raw, workspaceRoot)` cleans and absolutizes (relative
-anchored to the workspace root; falls back to `filepath.Abs` when the root is
-empty). `agent.permissionScope` (Phase-0 display) is refactored to call
-`DeriveScope`, so the displayed scope and the stored/matched scope can never diverge.
-
-`ScopeKind`: `""` (tool-wide) | `"file"` | `"dir"`.
-
-### Data model (`internal/sandbox/grants.go`)
-
-`Grant` gains `Scope` (absolute path, `omitempty`) and `ScopeKind` (`omitempty`).
-`grantFile.Grants` becomes `map[string][]Grant`; `grantSchemaVersion` = 2.
-`GrantInput` gains `Scope` (raw) + `ScopeKind`.
-
-### Grant creation
-
-The agent derives `(raw, kind)` via `DeriveScope` and passes them in `GrantInput`.
-`engine.Grant` resolves `Scope` to absolute using its configured `workspaceRoot`
-(the engine already holds one — engine.go:33) before delegating to `store.Grant`,
-which appends to the tool's slice (replacing a same-scope entry). CLI grants
-(`zero sandbox grants allow|deny <tool>`) leave the scope empty → tool-wide, as today.
-
-### Matching (`engine.Decide` → `store.Lookup`)
-
-`Decide` derives the request's absolute scope from `request.Args` +
-`request.WorkspaceRoot` (already populated at engine.go:63) and passes it to a
-scope-aware `Lookup(toolName, reqScopeAbs, autonomy)`. `grantCovers` implements the
-match semantics above; `Lookup` applies the deny-wins / most-specific-allow
-precedence and the per-grant autonomy check. The surrounding allow/deny/ceiling
-flow (engine.go:99-126) is unchanged.
-
-### Migration & a fixed bug
-
-`readState` peeks at `schemaVersion`: v1 (`map[string]Grant`) entries become single
-tool-wide grants; v2 reads `map[string][]Grant`; anything else still errors with
-"unsupported schemaVersion". Keys are canonicalized (trimmed) on read so a
-whitespace-padded tool name matches at lookup — closing audit finding #90
-(grants.go:146). `writeState` always emits v2.
-
-### CLI surface
-
-`List()` flattens to `[]Grant` (sorted by tool, then scope). `FormatGrantList`
-shows the scope (`*` for tool-wide). `Revoke(toolName)` removes all of a tool's
-grants and returns the count. Per-scope revoke is out of scope.
-
-## Testing
-
-- `grant_scope_test` / `grants_test`: `DeriveScope` kinds; `resolveScopeAbs` (relative
-  anchored, `./` equivalence, absolute passthrough, cross-project non-match);
-  `grantCovers` matrix (tool-wide/file/dir × equal/descendant/sibling/parent/
-  empty-request); deny-precedence; per-grant + policy autonomy; v1→v2 migration
-  round-trip; whitespace-key canonicalization; `FormatGrantList` scope rendering.
-- `engine_test`: covered scoped request auto-allows; uncovered sibling re-prompts;
-  dir deny blocks a subtree request; legacy tool-wide grant still allows.
-- `agent`: `persistPermissionGrant` forwards scope+kind; `TestPermissionScope`
-  stays green (display unchanged).
-
-## Out of scope
-
-Card UI changes (Phase 0), per-scope revoke, glob/pattern scopes.

go.mod

@@ -1,31 +1,31 @@
 module github.com/Gitlawb/zero
 
-go 1.25.0
-
-toolchain go1.26.4
+go 1.26.5
 
 require (
-	charm.land/bubbles/v2 v2.1.0
-	charm.land/bubbletea/v2 v2.0.7
-	charm.land/lipgloss/v2 v2.0.4
+	charm.land/bubbles/v2 v2.1.1
+	charm.land/bubbletea/v2 v2.0.8
+	charm.land/lipgloss/v2 v2.0.5
 	github.com/alecthomas/chroma/v2 v2.27.0
 	github.com/atotto/clipboard v0.1.4
+	github.com/aymanbagabas/go-udiff v0.4.1
 	github.com/charmbracelet/colorprofile v0.4.3
 	github.com/charmbracelet/x/ansi v0.11.7
+	github.com/charmbracelet/x/term v0.2.2
+	github.com/coder/websocket v1.8.15
 	github.com/ledongthuc/pdf v0.0.0-20250511090121-5959a4027728
-	golang.org/x/sys v0.46.0
+	golang.org/x/sys v0.47.0
 	mvdan.cc/sh/v3 v3.13.1
 )
 
 require (
-	github.com/charmbracelet/ultraviolet v0.0.0-20260615092913-2399af76d5b1 // indirect
+	github.com/charmbracelet/ultraviolet v0.0.0-20260703014108-f5a850f9c2b7 // indirect
 	github.com/charmbracelet/x/exp/golden v0.0.0-20260615092313-b57e5e6d29bb // indirect
-	github.com/charmbracelet/x/term v0.2.2 // indirect
 	github.com/charmbracelet/x/termios v0.1.1 // indirect
 	github.com/charmbracelet/x/windows v0.2.2 // indirect
 	github.com/clipperhouse/displaywidth v0.11.0 // indirect
 	github.com/clipperhouse/uax29/v2 v2.7.0 // indirect
-	github.com/dlclark/regexp2/v2 v2.2.1 // indirect
+	github.com/dlclark/regexp2/v2 v2.5.0 // indirect
 	github.com/go-quicktest/qt v1.102.0 // indirect
 	github.com/lucasb-eyer/go-colorful v1.4.0 // indirect
 	github.com/mattn/go-runewidth v0.0.24 // indirect
@@ -34,5 +34,5 @@ require (
 	github.com/rogpeppe/go-internal v1.15.0 // indirect
 	github.com/xo/terminfo v0.0.0-20220910002029-abceb7e1c41e // indirect
 	golang.org/x/exp v0.0.0-20260611194520-c48552f49976 // indirect
-	golang.org/x/sync v0.21.0 // indirect
+	golang.org/x/sync v0.22.0 // indirect
 )

go.sum

@@ -1,9 +1,9 @@
-charm.land/bubbles/v2 v2.1.0 h1:YSnNh5cPYlYjPxRrzs5VEn3vwhtEn3jVGRBT3M7/I0g=
-charm.land/bubbles/v2 v2.1.0/go.mod h1:l97h4hym2hvWBVfmJDtrEHHCtkIKeTEb3TTJ4ZOB3wY=
-charm.land/bubbletea/v2 v2.0.7 h1:7qw2tTAVar7m7klOPBYfTB0mniv/RuexsYwMRNxSeL0=
-charm.land/bubbletea/v2 v2.0.7/go.mod h1:DGW2q8gvzHnOpMpZTORs0aySVHCox5C+2Svk0fci1qs=
-charm.land/lipgloss/v2 v2.0.4 h1:lcPeVtcp23SNra7lHy8iYE4UC2aIipVQ47sbGyyxR5Q=
-charm.land/lipgloss/v2 v2.0.4/go.mod h1:0653x8epbZSzdDfO/XPS1a/uYPOBeSsCssOpJOqDzik=
+charm.land/bubbles/v2 v2.1.1 h1:7r55WzBxpo/R3z98hGmY7KKPd3ET6vsf0Fb9sDHOV60=
+charm.land/bubbles/v2 v2.1.1/go.mod h1:GE6M31gaWZVXzGw73OeuTTgy4lX+OtkH0E5ymnNsHxo=
+charm.land/bubbletea/v2 v2.0.8 h1:SxTJMhCAI3lbPmy4SgX5LWZ24AdINr4I6UEqzZvYJuY=
+charm.land/bubbletea/v2 v2.0.8/go.mod h1:2SkdgoTXluXJHOUwAoRlRXF/28vklb1rFl6GcgV1/ss=
+charm.land/lipgloss/v2 v2.0.5 h1:kbNxgeeUOYv5J0YdpxFjfvf3dFvqH8Aci4zB6xqFtrY=
+charm.land/lipgloss/v2 v2.0.5/go.mod h1:9oqhxt4yxIMe6q5A4kHr44DremZk7J9UNh74GlWa5nc=
 github.com/alecthomas/assert/v2 v2.11.0 h1:2Q9r3ki8+JYXvGsDyBXwH3LcJ+WK5D0gc5E8vS6K3D0=
 github.com/alecthomas/assert/v2 v2.11.0/go.mod h1:Bze95FyfUr7x34QZrjL+XP+0qgp/zg8yS+TtBj1WA3k=
 github.com/alecthomas/chroma/v2 v2.27.0 h1:FodwmyOBgJULFYmDqibcp9pvfDLWdtPRh9v/r5BXYZs=
@@ -16,8 +16,8 @@ github.com/aymanbagabas/go-udiff v0.4.1 h1:OEIrQ8maEeDBXQDoGCbbTTXYJMYRCRO1fnodZ
 github.com/aymanbagabas/go-udiff v0.4.1/go.mod h1:0L9PGwj20lrtmEMeyw4WKJ/TMyDtvAoK9bf2u/mNo3w=
 github.com/charmbracelet/colorprofile v0.4.3 h1:QPa1IWkYI+AOB+fE+mg/5/4HRMZcaXex9t5KX76i20Q=
 github.com/charmbracelet/colorprofile v0.4.3/go.mod h1:/zT4BhpD5aGFpqQQqw7a+VtHCzu+zrQtt1zhMt9mR4Q=
-github.com/charmbracelet/ultraviolet v0.0.0-20260615092913-2399af76d5b1 h1:4+r3uOJ69ueRBt4okgEfWZeXs3BD36HcDBmOIAUlETk=
-github.com/charmbracelet/ultraviolet v0.0.0-20260615092913-2399af76d5b1/go.mod h1:f/jRa757WUmaOZrbPspXymbg/GnbF+rwe4OLsG7aXYo=
+github.com/charmbracelet/ultraviolet v0.0.0-20260703014108-f5a850f9c2b7 h1:3FmWoGNWK4STvqg0O0Aeav2T7rodWJAPeF0QpH+8gFw=
+github.com/charmbracelet/ultraviolet v0.0.0-20260703014108-f5a850f9c2b7/go.mod h1:f/jRa757WUmaOZrbPspXymbg/GnbF+rwe4OLsG7aXYo=
 github.com/charmbracelet/x/ansi v0.11.7 h1:kzv1kJvjg2S3r9KHo8hDdHFQLEqn4RBCb39dAYC84jI=
 github.com/charmbracelet/x/ansi v0.11.7/go.mod h1:9qGpnAVYz+8ACONkZBUWPtL7lulP9No6p1epAihUZwQ=
 github.com/charmbracelet/x/exp/golden v0.0.0-20260615092313-b57e5e6d29bb h1:m1Uub9uM7meZ8ofVyeRfK7M/knCi3sF2gKn9kldVW0I=
@@ -32,8 +32,10 @@ github.com/clipperhouse/displaywidth v0.11.0 h1:lBc6kY44VFw+TDx4I8opi/EtL9m20WSE
 github.com/clipperhouse/displaywidth v0.11.0/go.mod h1:bkrFNkf81G8HyVqmKGxsPufD3JhNl3dSqnGhOoSD/o0=
 github.com/clipperhouse/uax29/v2 v2.7.0 h1:+gs4oBZ2gPfVrKPthwbMzWZDaAFPGYK72F0NJv2v7Vk=
 github.com/clipperhouse/uax29/v2 v2.7.0/go.mod h1:EFJ2TJMRUaplDxHKj1qAEhCtQPW2tJSwu5BF98AuoVM=
-github.com/dlclark/regexp2/v2 v2.2.1 h1:mf4KkFUj0gJuarK8P+LgiS+Lit7m9N1yAwEfPbee7R0=
-github.com/dlclark/regexp2/v2 v2.2.1/go.mod h1:avUrQvPaLz2DrFNHJF0taWAFFX2C1GMSSoeiqFjcBmU=
+github.com/coder/websocket v1.8.15 h1:6B2JPeOGlpff2Uz6vOEH1Vzpi0iUz20A+lPVhPHtNUA=
+github.com/coder/websocket v1.8.15/go.mod h1:NX3SzP+inril6yawo5CQXx8+fk145lPDC6pumgx0mVg=
+github.com/dlclark/regexp2/v2 v2.5.0 h1:liNiWIPCvCh5HBcYfsjd+P16AG79fwd6T1Toy2gOtEA=
+github.com/dlclark/regexp2/v2 v2.5.0/go.mod h1:avUrQvPaLz2DrFNHJF0taWAFFX2C1GMSSoeiqFjcBmU=
 github.com/go-quicktest/qt v1.102.0 h1:HSQxCeh5YZH3EL3W39ixjtyaEhcWSXQHtHnMBzSs474=
 github.com/go-quicktest/qt v1.102.0/go.mod h1:p4lGIVX+8Wa6ZPNDvqcxq36XpUDLh42FLetFU7odllI=
 github.com/google/go-cmp v0.7.0 h1:wk8382ETsv4JYUZwIsn6YpYiWiBsYLSJiTsyBybVuN8=
@@ -60,9 +62,9 @@ github.com/xo/terminfo v0.0.0-20220910002029-abceb7e1c41e h1:JVG44RsyaB9T2KIHavM
 github.com/xo/terminfo v0.0.0-20220910002029-abceb7e1c41e/go.mod h1:RbqR21r5mrJuqunuUZ/Dhy/avygyECGrLceyNeo4LiM=
 golang.org/x/exp v0.0.0-20260611194520-c48552f49976 h1:X8Hz2ImujgbmetVuW+w2YkyZChE3cBpZi2P158rTG9M=
 golang.org/x/exp v0.0.0-20260611194520-c48552f49976/go.mod h1:vnf4pv9iKZXY58sQE1L86zmNWJ4159e1RkcWiLCkeEY=
-golang.org/x/sync v0.21.0 h1:HLII4xRRTtCRkxYp4HNFF0Js/Og6q2i++KXbg0gHCwM=
-golang.org/x/sync v0.21.0/go.mod h1:9xrNwdLfx4jkKbNva9FpL6vEN7evnE43NNNJQ2LF3+0=
-golang.org/x/sys v0.46.0 h1:noSf2Fq6F8DBgS+LysIkx7rIExoNHJsxOAtPp4rthXw=
-golang.org/x/sys v0.46.0/go.mod h1:4GL1E5IUh+htKOUEOaiffhrAeqysfVGipDYzABqnCmw=
+golang.org/x/sync v0.22.0 h1:SZjpbeLmrCk4xhRSZFNZW5gFUeCeFgjekvI/+gfScek=
+golang.org/x/sync v0.22.0/go.mod h1:9xrNwdLfx4jkKbNva9FpL6vEN7evnE43NNNJQ2LF3+0=
+golang.org/x/sys v0.47.0 h1:o7XGOvZQCADBQQ4Y7VNq2dRWQR7JmOUW8Kxx4ZsNgWs=
+golang.org/x/sys v0.47.0/go.mod h1:4GL1E5IUh+htKOUEOaiffhrAeqysfVGipDYzABqnCmw=
 mvdan.cc/sh/v3 v3.13.1 h1:DP3TfgZhDkT7lerUdnp6PTGKyxxzz6T+cOlY/xEvfWk=
 mvdan.cc/sh/v3 v3.13.1/go.mod h1:lXJ8SexMvEVcHCoDvAGLZgFJ9Wsm2sulmoNEXGhYZD0=

internal/acp/agent.go

@@ -5,6 +5,7 @@ import (
 	"encoding/base64"
 	"encoding/json"
 	"errors"
+	"log"
 	"strings"
 	"sync"
 
@@ -159,8 +160,14 @@ func (a *Agent) handleSessionLoad(_ context.Context, params json.RawMessage) (an
 	// Load history BEFORE publishing the session so no concurrent prompt observes
 	// a half-initialized session (registerSession sets history under the lock and
 	// reuses an already-live session rather than orphaning its in-flight turn).
-	history := a.loadHistory(meta.SessionID)
+	history, historyErr := a.loadHistory(meta.SessionID)
 	sess := a.registerSession(meta.SessionID, root, history)
+	a.warnPersistence(
+		&notifier{conn: a.conn, sessionID: sess.id},
+		"load session history",
+		"Could not load session history. The session is open, but earlier turns may be missing until storage recovers.",
+		historyErr,
+	)
 	return LoadSessionResult{
 		Modes: a.modeState(sess),
 	}, nil
@@ -253,7 +260,14 @@ func (a *Agent) runTurn(ctx context.Context, sess *acpSession, userText string,
 	if stopErr != nil {
 		return "", RPCError(codeInternalError, stopErr.Error())
 	}
-	a.persistTurn(sess, userText, result.FinalAnswer)
+	if err := a.persistTurn(sess, userText, result.FinalAnswer); err != nil {
+		a.warnPersistence(
+			note,
+			"save session history",
+			"Could not save session history. This turn is available in memory, but future resume may miss it until storage recovers.",
+			err,
+		)
+	}
 	return reason, nil
 }
 
@@ -386,29 +400,34 @@ func (a *Agent) modeState(s *acpSession) *SessionModeState {
 
 // ---- persistence + continuity ----
 
-func (a *Agent) persistTurn(sess *acpSession, user, assistant string) {
-	if a.deps.Store != nil {
-		_, _ = a.deps.Store.AppendEvent(sess.id, sessions.AppendEventInput{
-			Type:    sessions.EventMessage,
-			Payload: map[string]any{"role": "user", "content": user},
-		})
-		if assistant != "" {
-			_, _ = a.deps.Store.AppendEvent(sess.id, sessions.AppendEventInput{
-				Type:    sessions.EventMessage,
-				Payload: map[string]any{"role": "assistant", "content": assistant},
-			})
-		}
-	}
-	sess.appendHistory(turnRecord{user: user, assistant: assistant})
-}
-
-func (a *Agent) loadHistory(sessionID string) []turnRecord {
+func (a *Agent) persistTurn(sess *acpSession, user, assistant string) error {
+	defer sess.appendHistory(turnRecord{user: user, assistant: assistant})
 	if a.deps.Store == nil {
 		return nil
 	}
+	events := []sessions.AppendEventInput{
+		{
+			Type:    sessions.EventMessage,
+			Payload: map[string]any{"role": "user", "content": user},
+		},
+	}
+	if assistant != "" {
+		events = append(events, sessions.AppendEventInput{
+			Type:    sessions.EventMessage,
+			Payload: map[string]any{"role": "assistant", "content": assistant},
+		})
+	}
+	_, err := a.deps.Store.AppendEvents(sess.id, events)
+	return err
+}
+
+func (a *Agent) loadHistory(sessionID string) ([]turnRecord, error) {
+	if a.deps.Store == nil {
+		return nil, nil
+	}
 	events, err := a.deps.Store.ReadEvents(sessionID)
 	if err != nil {
-		return nil
+		return nil, err
 	}
 	var records []turnRecord
 	var pendingUser string
@@ -444,7 +463,21 @@ func (a *Agent) loadHistory(sessionID string) []turnRecord {
 	if havePending {
 		records = append(records, turnRecord{user: pendingUser})
 	}
-	return records
+	return records, nil
+}
+
+func (a *Agent) warnPersistence(note *notifier, action string, message string, err error) {
+	if err == nil {
+		return
+	}
+	sessionID := ""
+	if note != nil {
+		sessionID = note.sessionID
+	}
+	log.Printf("zero acp: failed to %s for session %s: %v", action, sessionID, err)
+	if note != nil {
+		note.text("\n\n[zero warning] " + message + "\n")
+	}
 }
 
 // buildPrompt prepends prior conversation as context, since agent.Run drives a

internal/acp/agent_test.go

@@ -5,6 +5,8 @@ import (
 	"encoding/json"
 	"fmt"
 	"io"
+	"os"
+	"path/filepath"
 	"strings"
 	"testing"
 	"time"
@@ -242,8 +244,77 @@ func TestACPRejectsInvalidCwd(t *testing.T) {
 	}
 }
 
+func TestACPPromptWarnsWhenTurnPersistenceFails(t *testing.T) {
+	deps := testDeps(t)
+	h := newHarness(t, deps)
+	defer h.stop()
+	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
+	defer cancel()
+
+	var newRes NewSessionResult
+	if err := h.client.Call(ctx, MethodSessionNew, NewSessionParams{Cwd: t.TempDir(), McpServers: []McpServer{}}, &newRes); err != nil {
+		t.Fatalf("session/new: %v", err)
+	}
+	metadataPath := filepath.Join(deps.Store.RootDir, newRes.SessionID, sessions.MetadataFile)
+	if err := os.Remove(metadataPath); err != nil {
+		t.Fatalf("remove metadata: %v", err)
+	}
+
+	var promptRes PromptResult
+	if err := h.client.Call(ctx, MethodSessionPrompt, PromptParams{
+		SessionID: newRes.SessionID,
+		Prompt:    []ContentBlock{TextBlock("hi")},
+	}, &promptRes); err != nil {
+		t.Fatalf("session/prompt: %v", err)
+	}
+	if promptRes.StopReason != StopEndTurn {
+		t.Fatalf("stopReason = %q, want %q", promptRes.StopReason, StopEndTurn)
+	}
+	got := drainTextUntil(t, h.updates, func(text string) bool {
+		return strings.Contains(text, "Hello from ZERO") &&
+			strings.Contains(text, "Could not save session history")
+	})
+	if !strings.Contains(got, "Could not save session history") {
+		t.Fatalf("streamed text = %q, want persistence warning", got)
+	}
+}
+
+func TestACPLoadWarnsWhenHistoryReadFails(t *testing.T) {
+	deps := testDeps(t)
+	cwd := t.TempDir()
+	meta, err := deps.Store.Create(sessions.CreateInput{Title: "ACP session", Cwd: cwd})
+	if err != nil {
+		t.Fatalf("create session: %v", err)
+	}
+	eventsPath := filepath.Join(deps.Store.RootDir, meta.SessionID, sessions.EventsFile)
+	if err := os.WriteFile(eventsPath, []byte("{bad json}\n"), 0o600); err != nil {
+		t.Fatalf("write corrupt events: %v", err)
+	}
+	h := newHarness(t, deps)
+	defer h.stop()
+	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
+	defer cancel()
+
+	if err := h.client.Call(ctx, MethodSessionLoad, LoadSessionParams{SessionID: meta.SessionID, Cwd: cwd, McpServers: []McpServer{}}, &LoadSessionResult{}); err != nil {
+		t.Fatalf("session/load: %v", err)
+	}
+	got := drainTextUntil(t, h.updates, func(text string) bool {
+		return strings.Contains(text, "Could not load session history")
+	})
+	if !strings.Contains(got, "Could not load session history") {
+		t.Fatalf("streamed text = %q, want load warning", got)
+	}
+}
+
 // drainText collects streamed chunks for a short window and concatenates them.
 func drainText(t *testing.T, ch <-chan string) string {
+	t.Helper()
+	return drainTextUntil(t, ch, func(text string) bool {
+		return strings.Contains(text, "Hello from ZERO")
+	})
+}
+
+func drainTextUntil(t *testing.T, ch <-chan string, done func(string) bool) string {
 	t.Helper()
 	var b strings.Builder
 	deadline := time.After(2 * time.Second)
@@ -251,7 +322,7 @@ func drainText(t *testing.T, ch <-chan string) string {
 		select {
 		case s := <-ch:
 			b.WriteString(s)
-			if strings.Contains(b.String(), "Hello from ZERO") {
+			if done(b.String()) {
 				return b.String()
 			}
 		case <-deadline:

internal/acp/translate.go

@@ -3,6 +3,7 @@ package acp
 import (
 	"encoding/json"
 	"strings"
+	"unicode/utf8"
 
 	"github.com/Gitlawb/zero/internal/agent"
 	"github.com/Gitlawb/zero/internal/tools"
@@ -70,8 +71,8 @@ func primaryArgHint(rawArgs string) string {
 func truncateHint(s string) string {
 	s = strings.TrimSpace(s)
 	const max = 60
-	if len(s) > max {
-		return s[:max] + "…"
+	if utf8.RuneCountInString(s) > max {
+		return string([]rune(s)[:max]) + "…"
 	}
 	return s
 }

internal/acp/translate_test.go

@@ -1,7 +1,9 @@
 package acp
 
 import (
+	"strings"
 	"testing"
+	"unicode/utf8"
 
 	"github.com/Gitlawb/zero/internal/agent"
 	"github.com/Gitlawb/zero/internal/tools"
@@ -126,3 +128,35 @@ func TestPromptText(t *testing.T) {
 		t.Fatalf("promptText = %q", got)
 	}
 }
+
+func TestToolTitleTruncateHintRuneSafe(t *testing.T) {
+	// A 61-character string containing multi-byte UTF-8 runes (emojis / CJK characters).
+	// We want to verify that it is truncated without cutting any runes or producing invalid UTF-8.
+	longPath := "📁/项目/非常长的路径名称/测试/🚀/emoji-and-cjk-characters-which-are-very-long-and-exceed-sixty-characters"
+
+	// Create JSON args for read_file
+	rawArgs := `{"path":"` + longPath + `"}`
+	got := toolTitle("read_file", rawArgs)
+
+	expectedPrefix := "read_file "
+	if !strings.HasPrefix(got, expectedPrefix) {
+		t.Fatalf("expected title to start with %q, got %q", expectedPrefix, got)
+	}
+
+	hint := strings.TrimPrefix(got, expectedPrefix)
+	// Hint should end with the ellipsis character
+	if !strings.HasSuffix(hint, "…") {
+		t.Fatalf("expected truncated hint to end with ellipsis, got %q", hint)
+	}
+
+	// Check that we don't have invalid UTF-8 runes
+	if !utf8.ValidString(hint) {
+		t.Fatalf("truncated hint is not a valid UTF-8 string: %q", hint)
+	}
+
+	// The rune count of the hint (excluding ellipsis) should be exactly 60
+	runes := []rune(strings.TrimSuffix(hint, "…"))
+	if len(runes) != 60 {
+		t.Fatalf("expected exactly 60 runes before ellipsis, got %d (hint: %q)", len(runes), hint)
+	}
+}

internal/agent/acceptance_gate_test.go

@@ -0,0 +1,248 @@
+package agent
+
+import (
+	"context"
+	"strings"
+	"testing"
+
+	"github.com/Gitlawb/zero/internal/tools"
+	"github.com/Gitlawb/zero/internal/zeroruntime"
+)
+
+func enabledSelfCorrector(t *testing.T) *SelfCorrector {
+	t.Helper()
+	// checker/verifier nil → no LSP / project-test pass runs; we only need a
+	// non-nil SelfCorrector to enable the task-grounded acceptance gate.
+	return NewSelfCorrector(t.TempDir(), nil, nil, SelfCorrectConfig{Enabled: true})
+}
+
+// BUG #2(a): a final message that admits the model guessed / could not meet the
+// objective must be reported as INCOMPLETE, never success — and immediately
+// (an admitted guess is not worth re-prompting).
+func TestAcceptanceSelfReportAdmissionIsIncomplete(t *testing.T) {
+	provider := &mockProvider{turns: [][]zeroruntime.StreamEvent{
+		textTurn("I couldn't analyze the board image, so I wrote the common opening move e2e4 as my best guess."),
+	}}
+
+	result, err := Run(context.Background(), "write the best chess move", provider, Options{
+		Registry:                tools.NewRegistry(),
+		MaxTurns:                10,
+		RequireCompletionSignal: true,
+		SelfCorrect:             enabledSelfCorrector(t),
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !result.Incomplete {
+		t.Fatalf("admitted guess must be Incomplete, got success; final=%q", result.FinalAnswer)
+	}
+	if !strings.Contains(result.IncompleteReason, "objective was not met") {
+		t.Fatalf("IncompleteReason = %q, want it to cite the admission", result.IncompleteReason)
+	}
+	if len(provider.requests) != 1 {
+		t.Fatalf("an admission should downgrade immediately (1 turn), got %d", len(provider.requests))
+	}
+}
+
+// BUG #2(a): the inability detector must generalize over the verb the model uses
+// (so re-phrasing can't defeat it) while NOT misreading success-y negations
+// ("could not find any issues") as admissions.
+func TestSelfReportedIncompletionMatching(t *testing.T) {
+	cases := []struct {
+		text  string
+		admit bool
+	}{
+		// real chess re-run wording that slipped a fixed phrase list:
+		{"…which I cannot do without proper image analysis capabilities.", true},
+		{"I cannot visually analyze the chess board image.", true},
+		{"I could not recover the original values, so this is my best guess.", true},
+		{"I was unable to complete the optimization.", true},
+		{"The result may not be correct.", true},
+		// multi-occurrence: an early guarded use ("could not find any") must NOT mask a
+		// later genuine admission with the same stem (review #5):
+		{"I could not find any examples in the codebase, so I could not implement the feature.", true},
+		// genuine successes that must NOT be flagged:
+		{"I implemented the function and all tests pass.", false},
+		{"I verified the fix; I could not find any remaining issues.", false},
+		{"I cannot reproduce the bug, so the fix holds.", false},
+		{"Done — the output matches the required format and values.", false},
+		// behavior descriptions that must NOT be flagged — these phrases describe
+		// implemented behavior, not an admission (review #1, removed from the list):
+		{"The parser will fall back to UTF-8 when no encoding is set.", false},
+		{"I replaced the placeholder value with the computed key.", false},
+		{"The retry uses exponential backoff as a fallback.", false},
+		{"Without proper error handling the function would crash, so I added it.", false},
+		// real false positive (conversational recap): retelling a PAST exchange is
+		// not an admission about the current objective — the sentence carries its
+		// own narrative context ("You asked …", "at the time"):
+		{"You asked if I could work autonomously on zero if you gave me a GitHub account — I was honest that my sandbox had no repo, no build system, and no network/integration wired up yet, so I couldn't actually do it at the time.", false},
+		{"When we last spoke, I couldn't reach the GitHub API from the sandbox.", false},
+		{"Last session I was unable to build the project; that is what we should retry now.", false},
+		// quoted admissions (the model citing its own earlier message or a log
+		// line) must NOT be flagged:
+		{`My earlier reply said "I couldn't do it" but the tooling is now installed.`, false},
+		{"The run log contains `I cannot access the network`, which is expected inside the sandbox.", false},
+		// current-objective admissions must STILL be flagged, including past tense
+		// and multi-sentence finals whose other sentences are narrative:
+		{"I couldn't complete the refactor.", true},
+		{"I hit an error I could not resolve, so the migration is unfinished.", true},
+		{"You asked me to port the parser. I couldn't finish the second half.", true},
+		// unbalanced delimiters must not swallow the tail (only BALANCED quoted
+		// spans are stripped), so an admission after a stray quote still fires:
+		{"The log ends with a stray \" and I couldn't finish the migration.", true},
+		{"There is an unterminated `code span and I cannot proceed with the fix.", true},
+		{"The doc opens with “a curly quote and I was unable to complete the task.", true},
+	}
+	for _, c := range cases {
+		got := selfReportedIncompletion(c.text) != ""
+		if got != c.admit {
+			t.Errorf("selfReportedIncompletion(%q) admitted=%v, want %v", c.text, got, c.admit)
+		}
+	}
+}
+
+// BUG #2(a): a "cannot analyze / cannot determine" admission (present tense) must
+// be caught even when the model also stamps a bogus "PASS" based on a shape check.
+// Mirrors the chess-best-move final ("I cannot analyze the position… PASS: I wrote
+// the move and verified by reading it back").
+func TestAcceptanceCannotPhrasingIsIncomplete(t *testing.T) {
+	provider := &mockProvider{turns: [][]zeroruntime.StreamEvent{
+		textTurn("I cannot analyze the specific position from the image, so I cannot determine the best move. PASS: I wrote e2e4 to the file and verified it by reading it back."),
+	}}
+
+	result, err := Run(context.Background(), "write the best move", provider, Options{
+		Registry:                tools.NewRegistry(),
+		MaxTurns:                10,
+		RequireCompletionSignal: true,
+		SelfCorrect:             enabledSelfCorrector(t),
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !result.Incomplete {
+		t.Fatalf("a 'cannot analyze/determine' admission must be Incomplete despite the bogus PASS; final=%q", result.FinalAnswer)
+	}
+}
+
+// A pending/in_progress update_plan item is an AMBIGUOUS signal (the model may have
+// finished without re-marking the step), so it must NOT by itself force INCOMPLETE.
+// The gate nudges, but if the model keeps claiming completion with no continuation
+// cue and no admission, the run finalizes as SUCCESS — trusting the completion claim
+// over stale plan bookkeeping. (Only a continuation cue or self-report admission
+// forces INCOMPLETE; see TestCompletionGate* and TestAcceptance*Admission*.)
+func TestPendingPlanAloneDoesNotForceIncomplete(t *testing.T) {
+	registry := tools.NewRegistry()
+	registry.Register(tools.NewUpdatePlanTool())
+	done := "All set." // confident; no cue, no admission — only the plan is stale
+	provider := &mockProvider{turns: [][]zeroruntime.StreamEvent{
+		planTurn("completed", "in_progress"),
+		textTurn(done), textTurn(done), textTurn(done), textTurn(done), textTurn(done),
+	}}
+
+	result, err := Run(context.Background(), "do the multi-step task", provider, Options{
+		Registry:                registry,
+		MaxTurns:                10,
+		RequireCompletionSignal: true,
+		SelfCorrect:             enabledSelfCorrector(t),
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if result.Incomplete {
+		t.Fatalf("pending-plan alone must NOT force Incomplete (stale bookkeeping != unfinished); reason=%q", result.IncompleteReason)
+	}
+	if result.FinalAnswer != done {
+		t.Fatalf("final answer = %q, want the completion claim %q", result.FinalAnswer, done)
+	}
+	if !someRequestContains(provider.requests, continueNudgeMarker) {
+		t.Fatalf("expected at least one continue-nudge before accepting completion")
+	}
+}
+
+// BUG #2(b): when self-correct is on, the run must pass a one-time task-grounded
+// acceptance check before success. A model that, when challenged, confirms its
+// work meets the criterion (no admission) then finalizes as success — no regression.
+func TestAcceptanceGroundedCompletionSucceeds(t *testing.T) {
+	provider := &mockProvider{turns: [][]zeroruntime.StreamEvent{
+		textTurn("I implemented the function."),
+		textTurn("PASS. I ran the verification and it passes; the output meets the task requirement."),
+	}}
+
+	result, err := Run(context.Background(), "implement the function", provider, Options{
+		Registry:                tools.NewRegistry(),
+		MaxTurns:                10,
+		RequireCompletionSignal: true,
+		SelfCorrect:             enabledSelfCorrector(t),
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if result.Incomplete {
+		t.Fatalf("a verified completion must succeed, got Incomplete (%q)", result.IncompleteReason)
+	}
+	if result.FinalAnswer != "PASS. I ran the verification and it passes; the output meets the task requirement." {
+		t.Fatalf("final answer = %q", result.FinalAnswer)
+	}
+	// Exactly one acceptance pass: the first "done" turn was challenged, the second confirmed.
+	if len(provider.requests) != 2 {
+		t.Fatalf("expected 2 turns (1 acceptance challenge + 1 confirmation), got %d", len(provider.requests))
+	}
+	if !someRequestContains(provider.requests, acceptanceNudgeMarker) {
+		t.Fatalf("expected the task-grounded acceptance check to be demanded")
+	}
+}
+
+// With the gate OFF (interactive/TUI default), even an admission is accepted as the
+// final answer exactly as before — proving no behavior change for non-headless callers.
+func TestAcceptanceGateOffLeavesAdmissionAsSuccess(t *testing.T) {
+	admission := "I couldn't determine the value, so this is a guess."
+	provider := &mockProvider{turns: [][]zeroruntime.StreamEvent{
+		textTurn(admission),
+	}}
+
+	result, err := Run(context.Background(), "do a thing", provider, Options{
+		Registry: tools.NewRegistry(),
+		MaxTurns: 10,
+		// RequireCompletionSignal false, SelfCorrect nil — the legacy path.
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if result.Incomplete {
+		t.Fatalf("legacy path must never set Incomplete; final=%q", result.FinalAnswer)
+	}
+	if result.FinalAnswer != admission {
+		t.Fatalf("final answer = %q, want the text verbatim (legacy behavior)", result.FinalAnswer)
+	}
+	if len(provider.requests) != 1 {
+		t.Fatalf("legacy path must not re-prompt; got %d turns", len(provider.requests))
+	}
+}
+
+// The acceptance check is scoped to runs with self-correct on: a headless run
+// WITHOUT a SelfCorrector accepts a confident completion immediately (the (b) gate
+// is inert), so enabling RequireCompletionSignal alone adds no acceptance turn.
+func TestAcceptanceGateRequiresSelfCorrect(t *testing.T) {
+	provider := &mockProvider{turns: [][]zeroruntime.StreamEvent{
+		textTurn("Done. Implemented and it works."),
+	}}
+
+	result, err := Run(context.Background(), "implement it", provider, Options{
+		Registry:                tools.NewRegistry(),
+		MaxTurns:                10,
+		RequireCompletionSignal: true,
+		// SelfCorrect deliberately nil → acceptance gate (b) must not fire.
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if result.Incomplete {
+		t.Fatalf("confident completion without self-correct must succeed; reason=%q", result.IncompleteReason)
+	}
+	if len(provider.requests) != 1 {
+		t.Fatalf("acceptance gate must not fire without self-correct; got %d turns", len(provider.requests))
+	}
+	if someRequestContains(provider.requests, acceptanceNudgeMarker) {
+		t.Fatalf("acceptance nudge must not be injected when self-correct is off")
+	}
+}

internal/agent/additional_permissions_validation_test.go

@@ -0,0 +1,146 @@
+package agent
+
+import (
+	"context"
+	"strings"
+	"testing"
+
+	"github.com/Gitlawb/zero/internal/tools"
+)
+
+// A shell call that sets sandbox_permissions: with_additional_permissions but
+// omits (or malforms) additional_permissions can never succeed: the same
+// validation runs again at grant time regardless of the user's decision. It
+// must be rejected as a plain tool error BEFORE any permission prompt, not
+// surfaced as a confusing "approved but denied" prompt outcome.
+func TestExecuteToolCallRejectsMissingAdditionalPermissionsBeforePrompt(t *testing.T) {
+	root := t.TempDir()
+	registry := tools.NewRegistry()
+	registry.Register(tools.NewBashTool(root))
+
+	promptCalls := 0
+	result, err := executeToolCall(
+		context.Background(),
+		registry,
+		ToolCall{
+			ID:        "c1",
+			Name:      "bash",
+			Arguments: `{"command":"echo hi","sandbox_permissions":"with_additional_permissions"}`,
+		},
+		PermissionModeAuto,
+		Options{
+			OnPermissionRequest: func(ctx context.Context, request PermissionRequest) (PermissionDecision, error) {
+				promptCalls++
+				return PermissionDecision{Action: PermissionDecisionAllow}, nil
+			},
+		},
+	)
+	if err != nil {
+		t.Fatalf("executeToolCall: %v", err)
+	}
+	if result.Status != tools.StatusError {
+		t.Fatalf("status = %q, want error", result.Status)
+	}
+	if promptCalls != 0 {
+		t.Fatalf("OnPermissionRequest called %d times, want 0 (must reject before prompting)", promptCalls)
+	}
+}
+
+// The mismatched-flag case (additional_permissions present without the
+// sandbox_permissions flag) must be rejected the same way.
+func TestExecuteToolCallRejectsAdditionalPermissionsWithoutFlagBeforePrompt(t *testing.T) {
+	root := t.TempDir()
+	registry := tools.NewRegistry()
+	registry.Register(tools.NewBashTool(root))
+
+	promptCalls := 0
+	result, err := executeToolCall(
+		context.Background(),
+		registry,
+		ToolCall{
+			ID:        "c1",
+			Name:      "bash",
+			Arguments: `{"command":"echo hi","additional_permissions":{"network":{"enabled":true}}}`,
+		},
+		PermissionModeAuto,
+		Options{
+			OnPermissionRequest: func(ctx context.Context, request PermissionRequest) (PermissionDecision, error) {
+				promptCalls++
+				return PermissionDecision{Action: PermissionDecisionAllow}, nil
+			},
+		},
+	)
+	if err != nil {
+		t.Fatalf("executeToolCall: %v", err)
+	}
+	if result.Status != tools.StatusError {
+		t.Fatalf("status = %q, want error", result.Status)
+	}
+	if promptCalls != 0 {
+		t.Fatalf("OnPermissionRequest called %d times, want 0 (must reject before prompting)", promptCalls)
+	}
+}
+
+// The rejection message must show a concrete, correctly-shaped example: a
+// model that gets this wrong once (attaching sandbox_permissions out of habit
+// from an earlier call, without a valid payload) needs enough in the error to
+// self-correct on retry instead of repeating the identical mistake.
+func TestExecuteToolCallMissingAdditionalPermissionsErrorIsActionable(t *testing.T) {
+	root := t.TempDir()
+	registry := tools.NewRegistry()
+	registry.Register(tools.NewBashTool(root))
+
+	result, err := executeToolCall(
+		context.Background(),
+		registry,
+		ToolCall{
+			ID:        "c1",
+			Name:      "bash",
+			Arguments: `{"command":"ping example.com","sandbox_permissions":"with_additional_permissions"}`,
+		},
+		PermissionModeAuto,
+		Options{},
+	)
+	if err != nil {
+		t.Fatalf("executeToolCall: %v", err)
+	}
+	if !strings.Contains(result.Output, `{"network": {"enabled": true}}`) {
+		t.Fatalf("output = %q, want a concrete additional_permissions example", result.Output)
+	}
+	if !strings.Contains(result.Output, "omit sandbox_permissions") {
+		t.Fatalf("output = %q, want guidance that the flag can simply be omitted", result.Output)
+	}
+}
+
+// A well-formed additional_permissions payload must still reach the normal
+// permission prompt: this fix only rejects malformed payloads.
+func TestExecuteToolCallStillPromptsForValidAdditionalPermissions(t *testing.T) {
+	root := t.TempDir()
+	registry := tools.NewRegistry()
+	registry.Register(tools.NewBashTool(root))
+
+	promptCalls := 0
+	_, err := executeToolCall(
+		context.Background(),
+		registry,
+		ToolCall{
+			ID:   "c1",
+			Name: "bash",
+			Arguments: `{"command":"echo hi","sandbox_permissions":"with_additional_permissions",` +
+				`"additional_permissions":{"network":{"enabled":true}}}`,
+		},
+		PermissionModeAuto,
+		Options{
+			OnPermissionRequest: func(ctx context.Context, request PermissionRequest) (PermissionDecision, error) {
+				promptCalls++
+				return PermissionDecision{Action: PermissionDecisionDeny}, nil
+			},
+		},
+	)
+	if err != nil {
+		t.Fatalf("executeToolCall: %v", err)
+	}
+	if promptCalls != 1 {
+		t.Fatalf("OnPermissionRequest called %d times, want exactly 1 for a valid elevation request", promptCalls)
+	}
+}

internal/agent/ask_user_test.go

@@ -141,9 +141,11 @@ func TestRunAskUserCancellationAbortsRun(t *testing.T) {
 	registry := registryWithAskUser()
 	args := `{"questions":[{"question":"Which framework?"}]}`
 	provider := providerCallingAskUserThenAnswer(args, "done")
+	var toolResults []ToolResult
 
 	result, err := Run(context.Background(), "clarify", provider, Options{
-		Registry: registry,
+		Registry:     registry,
+		OnToolResult: func(result ToolResult) { toolResults = append(toolResults, result) },
 		OnAskUser: func(_ context.Context, _ AskUserRequest) (AskUserResponse, error) {
 			return AskUserResponse{}, context.Canceled
 		},
@@ -158,6 +160,9 @@ func TestRunAskUserCancellationAbortsRun(t *testing.T) {
 	if len(provider.requests) != 1 {
 		t.Fatalf("expected the run to stop after the canceled ask_user (1 turn), got %d", len(provider.requests))
 	}
+	if len(toolResults) != 1 || toolResults[0].ToolCallID == "" {
+		t.Fatalf("cancellation must emit exactly one result for the call, got %#v", toolResults)
+	}
 	// The recorded tool result must reflect cancellation, not a synthetic answer.
 	var toolMsg string
 	for _, m := range result.Messages {

internal/agent/async_diagnostics.go

@@ -0,0 +1,185 @@
+package agent
+
+import (
+	"context"
+	"path/filepath"
+	"sort"
+	"strings"
+	"sync"
+	"time"
+)
+
+// Post-edit LSP diagnostics used to run synchronously inside edit_file /
+// write_file: every edit blocked its tool result on a fresh-publish wait with
+// a quiet-period debounce (≥300ms floor that reset on each publish, 10s cap) —
+// the largest recurring latency tax in an edit-heavy session. Collection is
+// now asynchronous: a file is enqueued the moment a mutating tool finishes, a
+// single worker checks it while the rest of the tool batch (and any
+// self-correct verification) runs, and the loop drains completed results just
+// before building the NEXT provider request, appending errors as a user-role
+// nudge. The model always issues another request after an edit turn — it has
+// to read its tool results — so it still sees introduced errors at the same
+// decision point as before, without any tool call stalling on the language
+// server.
+//
+// A file re-edited before its check runs is checked once in its final state
+// (the checker re-reads the file); a file re-edited after its check completed
+// is re-enqueued and the newer result wins. When the worker is still busy at
+// drain time, the drain gives up after a short wait and the results are
+// delivered on the following turn instead — they stay accurate because any
+// further edit re-enqueues the file.
+
+// asyncDiagnosticsDrainTimeout bounds how long a turn waits on the in-flight
+// check before deferring delivery to the next turn. Most checks finish well
+// under it (debounce floor + analysis); a cold language server on a large
+// repo does not get to stall the loop the way the old inline 10s cap could.
+// A var so tests can shorten the wait.
+var asyncDiagnosticsDrainTimeout = 3 * time.Second
+
+// asyncDiagnosticsFinalDrainTimeout is the wait used when the run is about to
+// FINALIZE (natural completion or the max-turns summary): there is no later
+// turn to defer to, so the gate waits out the old inline-era budget rather
+// than silently dropping an error the last edit introduced. A var so tests
+// can shorten the wait.
+var asyncDiagnosticsFinalDrainTimeout = 10 * time.Second
+
+// asyncDiagnosticsNudge prefixes the drained diagnostics blocks; phrased like
+// the old inline block so the model reacts the same way.
+const asyncDiagnosticsNudge = "Diagnostics after your recent edits (fix any errors you introduced):\n"
+
+type asyncDiagnostics struct {
+	check         func(context.Context, string) string
+	workspaceRoot string
+
+	mu      sync.Mutex
+	queue   []string
+	queued  map[string]bool
+	results map[string]string
+	// working is non-nil while the worker goroutine runs and is closed when it
+	// exits, so drain can wait for quiescence without polling.
+	working chan struct{}
+}
+
+// newAsyncDiagnostics returns nil when check is nil, and every method on a nil
+// receiver is a no-op — callers wire it unconditionally.
+func newAsyncDiagnostics(check func(context.Context, string) string, workspaceRoot string) *asyncDiagnostics {
+	if check == nil {
+		return nil
+	}
+	return &asyncDiagnostics{
+		check:         check,
+		workspaceRoot: workspaceRoot,
+		queued:        map[string]bool{},
+		results:       map[string]string{},
+	}
+}
+
+// enqueue schedules changed files for a background check. Paths are
+// workspace-relative from Result.ChangedFiles (absolute when under a granted
+// extra write root), mirroring SelfCorrector's resolution.
+func (diagnostics *asyncDiagnostics) enqueue(ctx context.Context, files []string) {
+	if diagnostics == nil || len(files) == 0 {
+		return
+	}
+	diagnostics.mu.Lock()
+	defer diagnostics.mu.Unlock()
+	for _, file := range files {
+		path := diagnostics.absPath(file)
+		if diagnostics.queued[path] {
+			continue
+		}
+		diagnostics.queued[path] = true
+		// A completed result for this file describes a pre-edit state now.
+		delete(diagnostics.results, path)
+		diagnostics.queue = append(diagnostics.queue, path)
+	}
+	if diagnostics.working == nil && len(diagnostics.queue) > 0 {
+		done := make(chan struct{})
+		diagnostics.working = done
+		go diagnostics.work(ctx, done)
+	}
+}
+
+func (diagnostics *asyncDiagnostics) work(ctx context.Context, done chan struct{}) {
+	defer close(done)
+	for {
+		diagnostics.mu.Lock()
+		if len(diagnostics.queue) == 0 || ctx.Err() != nil {
+			diagnostics.working = nil
+			diagnostics.mu.Unlock()
+			return
+		}
+		path := diagnostics.queue[0]
+		diagnostics.queue = diagnostics.queue[1:]
+		// Un-mark before checking so an edit that lands DURING the check
+		// re-enqueues the file and it is checked again in its final state.
+		delete(diagnostics.queued, path)
+		diagnostics.mu.Unlock()
+
+		block := diagnostics.check(ctx, path)
+
+		diagnostics.mu.Lock()
+		if block != "" {
+			diagnostics.results[path] = block
+		} else {
+			delete(diagnostics.results, path)
+		}
+		diagnostics.mu.Unlock()
+	}
+}
+
+// drain waits briefly for the worker to go quiet, then formats and consumes
+// all completed error blocks as one nudge. Returns "" when there is nothing
+// to report or the worker is still busy (results then surface next turn).
+func (diagnostics *asyncDiagnostics) drain(ctx context.Context) string {
+	return diagnostics.drainWithin(ctx, asyncDiagnosticsDrainTimeout)
+}
+
+// drainFinal is drain with the finalization budget: called before the run's
+// last model request, where an undelivered result would otherwise be lost.
+func (diagnostics *asyncDiagnostics) drainFinal(ctx context.Context) string {
+	return diagnostics.drainWithin(ctx, asyncDiagnosticsFinalDrainTimeout)
+}
+
+func (diagnostics *asyncDiagnostics) drainWithin(ctx context.Context, timeout time.Duration) string {
+	if diagnostics == nil {
+		return ""
+	}
+	diagnostics.mu.Lock()
+	busy := diagnostics.working
+	diagnostics.mu.Unlock()
+	if busy != nil {
+		timer := time.NewTimer(timeout)
+		defer timer.Stop()
+		select {
+		case <-busy:
+		case <-timer.C:
+			return ""
+		case <-ctx.Done():
+			return ""
+		}
+	}
+	diagnostics.mu.Lock()
+	defer diagnostics.mu.Unlock()
+	if len(diagnostics.results) == 0 {
+		return ""
+	}
+	paths := make([]string, 0, len(diagnostics.results))
+	for path := range diagnostics.results {
+		paths = append(paths, path)
+	}
+	sort.Strings(paths)
+	blocks := make([]string, 0, len(paths))
+	for _, path := range paths {
+		blocks = append(blocks, diagnostics.results[path])
+	}
+	diagnostics.results = map[string]string{}
+	return asyncDiagnosticsNudge + strings.Join(blocks, "\n")
+}
+
+func (diagnostics *asyncDiagnostics) absPath(path string) string {
+	if filepath.IsAbs(path) {
+		return path
+	}
+	return filepath.Join(diagnostics.workspaceRoot, path)
+}

internal/agent/async_diagnostics_test.go

@@ -0,0 +1,338 @@
+package agent
+
+import (
+	"context"
+	"path/filepath"
+	"strings"
+	"sync"
+	"testing"
+	"time"
+
+	"github.com/Gitlawb/zero/internal/tools"
+	"github.com/Gitlawb/zero/internal/zeroruntime"
+)
+
+func TestAsyncDiagnosticsNilCollectorNoOps(t *testing.T) {
+	var diagnostics *asyncDiagnostics
+	if diagnostics != newAsyncDiagnostics(nil, "/ws") {
+		t.Fatal("nil check must yield a nil collector")
+	}
+	diagnostics.enqueue(context.Background(), []string{"a.go"})
+	if nudge := diagnostics.drain(context.Background()); nudge != "" {
+		t.Fatalf("nil collector drain = %q, want empty", nudge)
+	}
+}
+
+func TestAsyncDiagnosticsCollectsAndDrainsOnce(t *testing.T) {
+	// A real absolute root: a literal "/ws" is not filepath.IsAbs on Windows
+	// (no drive letter), which is exactly what the assertions below check.
+	root := t.TempDir()
+	var mu sync.Mutex
+	var checked []string
+	check := func(_ context.Context, absPath string) string {
+		mu.Lock()
+		checked = append(checked, absPath)
+		mu.Unlock()
+		if filepath.Base(absPath) == "clean.go" {
+			return ""
+		}
+		return "ERR " + filepath.Base(absPath)
+	}
+	diagnostics := newAsyncDiagnostics(check, root)
+	diagnostics.enqueue(context.Background(), []string{"broken.go", "clean.go"})
+
+	nudge := diagnostics.drain(context.Background())
+	if !strings.HasPrefix(nudge, asyncDiagnosticsNudge) {
+		t.Fatalf("nudge missing prefix: %q", nudge)
+	}
+	if !strings.Contains(nudge, "ERR broken.go") || strings.Contains(nudge, "clean.go") {
+		t.Fatalf("nudge = %q, want broken.go errors only", nudge)
+	}
+	mu.Lock()
+	for _, path := range checked {
+		if !filepath.IsAbs(path) || !strings.HasPrefix(path, root+string(filepath.Separator)) {
+			t.Fatalf("check received %q, want path under workspace root %q", path, root)
+		}
+	}
+	mu.Unlock()
+	if second := diagnostics.drain(context.Background()); second != "" {
+		t.Fatalf("second drain = %q, want empty (results are consumed)", second)
+	}
+}
+
+// A check still running at drain time defers delivery to the next drain
+// instead of blocking the turn.
+func TestAsyncDiagnosticsSlowCheckDefersToNextDrain(t *testing.T) {
+	previous := asyncDiagnosticsDrainTimeout
+	asyncDiagnosticsDrainTimeout = 20 * time.Millisecond
+	defer func() { asyncDiagnosticsDrainTimeout = previous }()
+
+	release := make(chan struct{})
+	check := func(_ context.Context, absPath string) string {
+		<-release
+		return "ERR " + filepath.Base(absPath)
+	}
+	diagnostics := newAsyncDiagnostics(check, "/ws")
+	diagnostics.enqueue(context.Background(), []string{"slow.go"})
+
+	start := time.Now()
+	if nudge := diagnostics.drain(context.Background()); nudge != "" {
+		t.Fatalf("drain during in-flight check = %q, want empty", nudge)
+	}
+	if elapsed := time.Since(start); elapsed > time.Second {
+		t.Fatalf("drain blocked %v, want bounded by the (shortened) timeout", elapsed)
+	}
+
+	close(release)
+	deadline := time.Now().Add(2 * time.Second)
+	for {
+		if nudge := diagnostics.drain(context.Background()); nudge != "" {
+			if !strings.Contains(nudge, "ERR slow.go") {
+				t.Fatalf("deferred nudge = %q", nudge)
+			}
+			return
+		}
+		if time.Now().After(deadline) {
+			t.Fatal("deferred diagnostics never delivered")
+		}
+		time.Sleep(5 * time.Millisecond)
+	}
+}
+
+// Re-enqueueing a file after its check completed replaces the stale result.
+func TestAsyncDiagnosticsReEditReplacesResult(t *testing.T) {
+	var mu sync.Mutex
+	response := "ERR old"
+	check := func(context.Context, string) string {
+		mu.Lock()
+		defer mu.Unlock()
+		return response
+	}
+	diagnostics := newAsyncDiagnostics(check, "/ws")
+
+	diagnostics.enqueue(context.Background(), []string{"a.go"})
+	waitForIdle(t, diagnostics)
+	mu.Lock()
+	response = "ERR new"
+	mu.Unlock()
+	diagnostics.enqueue(context.Background(), []string{"a.go"})
+
+	nudge := diagnostics.drain(context.Background())
+	if !strings.Contains(nudge, "ERR new") || strings.Contains(nudge, "ERR old") {
+		t.Fatalf("nudge = %q, want only the re-check result", nudge)
+	}
+}
+
+func waitForIdle(t *testing.T, diagnostics *asyncDiagnostics) {
+	t.Helper()
+	deadline := time.Now().Add(2 * time.Second)
+	for {
+		diagnostics.mu.Lock()
+		busy := diagnostics.working
+		diagnostics.mu.Unlock()
+		if busy == nil {
+			return
+		}
+		if time.Now().After(deadline) {
+			t.Fatal("worker never went idle")
+		}
+		time.Sleep(2 * time.Millisecond)
+	}
+}
+
+// changedFilesTool is a fake mutating tool reporting a changed file without
+// touching the filesystem.
+type changedFilesTool struct{}
+
+func (changedFilesTool) Name() string        { return "fake_edit" }
+func (changedFilesTool) Description() string { return "test mutating tool" }
+func (changedFilesTool) Parameters() tools.Schema {
+	return tools.Schema{Type: "object", Properties: map[string]tools.PropertySchema{}}
+}
+func (changedFilesTool) Safety() tools.Safety {
+	return tools.Safety{SideEffect: tools.SideEffectWrite, Permission: tools.PermissionAllow, Reason: "test"}
+}
+func (changedFilesTool) Run(context.Context, map[string]any) tools.Result {
+	return tools.Result{Status: tools.StatusOK, Output: "edited", ChangedFiles: []string{"main.go"}}
+}
+
+// End-to-end through Run: an edit turn does NOT carry diagnostics in the tool
+// result (the old inline path), but the next request contains the background
+// diagnostics nudge, before the model finalizes.
+func TestRunDeliversAsyncDiagnosticsNudgeNextTurn(t *testing.T) {
+	registry := tools.NewRegistry()
+	registry.Register(changedFilesTool{})
+	provider := &mockProvider{
+		turns: [][]zeroruntime.StreamEvent{
+			{
+				{Type: zeroruntime.StreamEventToolCallStart, ToolCallID: "call-1", ToolName: "fake_edit"},
+				{Type: zeroruntime.StreamEventToolCallDelta, ToolCallID: "call-1", ArgumentsFragment: `{}`},
+				{Type: zeroruntime.StreamEventToolCallEnd, ToolCallID: "call-1"},
+				{Type: zeroruntime.StreamEventDone},
+			},
+			{
+				{Type: zeroruntime.StreamEventText, Content: "done"},
+				{Type: zeroruntime.StreamEventDone},
+			},
+		},
+	}
+	root := t.TempDir()
+	var checkedPath string
+	var mu sync.Mutex
+
+	result, err := Run(context.Background(), "fix main.go", provider, Options{
+		Registry: registry,
+		Cwd:      root,
+		FileDiagnostics: func(_ context.Context, absPath string) string {
+			mu.Lock()
+			checkedPath = absPath
+			mu.Unlock()
+			return "main.go:1:1 error: boom"
+		},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if result.FinalAnswer != "done" {
+		t.Fatalf("final answer = %q", result.FinalAnswer)
+	}
+	if len(provider.requests) != 2 {
+		t.Fatalf("expected 2 provider turns, got %d", len(provider.requests))
+	}
+
+	// The tool result itself must NOT block on / embed diagnostics.
+	for _, message := range provider.requests[1].Messages {
+		if message.Role == zeroruntime.MessageRoleTool && strings.Contains(message.Content, "boom") {
+			t.Fatalf("diagnostics leaked into the tool result: %q", message.Content)
+		}
+	}
+	// The nudge must be present in the second request as a user message.
+	found := false
+	for _, message := range provider.requests[1].Messages {
+		if message.Role == zeroruntime.MessageRoleUser && strings.HasPrefix(message.Content, asyncDiagnosticsNudge) {
+			if !strings.Contains(message.Content, "boom") {
+				t.Fatalf("nudge missing diagnostics: %q", message.Content)
+			}
+			found = true
+		}
+	}
+	if !found {
+		t.Fatal("no async diagnostics nudge in the follow-up request")
+	}
+	mu.Lock()
+	defer mu.Unlock()
+	if want := filepath.Join(root, "main.go"); checkedPath != want {
+		t.Fatalf("checked path = %q, want %q", checkedPath, want)
+	}
+}
+
+// When the run hits the maxTurns ceiling right after an edit turn, the
+// final-answer request must still carry the diagnostics nudge — otherwise an
+// error introduced by the last edit would go unreported in the summary.
+func TestRunDrainsAsyncDiagnosticsBeforeMaxTurnsFinalAnswer(t *testing.T) {
+	registry := tools.NewRegistry()
+	registry.Register(changedFilesTool{})
+	provider := &mockProvider{
+		turns: [][]zeroruntime.StreamEvent{
+			{
+				{Type: zeroruntime.StreamEventToolCallStart, ToolCallID: "call-1", ToolName: "fake_edit"},
+				{Type: zeroruntime.StreamEventToolCallDelta, ToolCallID: "call-1", ArgumentsFragment: `{}`},
+				{Type: zeroruntime.StreamEventToolCallEnd, ToolCallID: "call-1"},
+				{Type: zeroruntime.StreamEventDone},
+			},
+			{
+				{Type: zeroruntime.StreamEventText, Content: "summary"},
+				{Type: zeroruntime.StreamEventDone},
+			},
+		},
+	}
+
+	result, err := Run(context.Background(), "fix main.go", provider, Options{
+		Registry: registry,
+		Cwd:      t.TempDir(),
+		MaxTurns: 1,
+		FileDiagnostics: func(context.Context, string) string {
+			return "main.go:1:1 error: boom"
+		},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if result.FinalAnswer != "summary" {
+		t.Fatalf("final answer = %q", result.FinalAnswer)
+	}
+	if len(provider.requests) != 2 {
+		t.Fatalf("expected the max-turns final-answer request, got %d requests", len(provider.requests))
+	}
+	found := false
+	for _, message := range provider.requests[1].Messages {
+		if message.Role == zeroruntime.MessageRoleUser && strings.HasPrefix(message.Content, asyncDiagnosticsNudge) {
+			found = true
+		}
+	}
+	if !found {
+		t.Fatal("max-turns final-answer request missing the diagnostics nudge")
+	}
+}
+
+// When the model finalizes while a slow check is still in flight (the
+// per-turn drain missed it), the finalization gate must wait it out and give
+// the model one more turn with the nudge instead of dropping the errors.
+func TestRunFinalizationGateDeliversLateDiagnostics(t *testing.T) {
+	previous := asyncDiagnosticsDrainTimeout
+	asyncDiagnosticsDrainTimeout = time.Millisecond
+	defer func() { asyncDiagnosticsDrainTimeout = previous }()
+
+	registry := tools.NewRegistry()
+	registry.Register(changedFilesTool{})
+	provider := &mockProvider{
+		turns: [][]zeroruntime.StreamEvent{
+			{
+				{Type: zeroruntime.StreamEventToolCallStart, ToolCallID: "call-1", ToolName: "fake_edit"},
+				{Type: zeroruntime.StreamEventToolCallDelta, ToolCallID: "call-1", ArgumentsFragment: `{}`},
+				{Type: zeroruntime.StreamEventToolCallEnd, ToolCallID: "call-1"},
+				{Type: zeroruntime.StreamEventDone},
+			},
+			{
+				{Type: zeroruntime.StreamEventText, Content: "done"},
+				{Type: zeroruntime.StreamEventDone},
+			},
+			{
+				{Type: zeroruntime.StreamEventText, Content: "done after fix"},
+				{Type: zeroruntime.StreamEventDone},
+			},
+		},
+	}
+
+	result, err := Run(context.Background(), "fix main.go", provider, Options{
+		Registry: registry,
+		Cwd:      t.TempDir(),
+		FileDiagnostics: func(context.Context, string) string {
+			// Slower than the (shortened) per-turn drain, well under the
+			// finalization budget.
+			time.Sleep(100 * time.Millisecond)
+			return "main.go:1:1 error: boom"
+		},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if result.FinalAnswer != "done after fix" {
+		t.Fatalf("final answer = %q, want the post-nudge answer", result.FinalAnswer)
+	}
+	if len(provider.requests) != 3 {
+		t.Fatalf("expected a third request carrying the late nudge, got %d requests", len(provider.requests))
+	}
+	found := false
+	for _, message := range provider.requests[2].Messages {
+		if message.Role == zeroruntime.MessageRoleUser && strings.HasPrefix(message.Content, asyncDiagnosticsNudge) {
+			if !strings.Contains(message.Content, "boom") {
+				t.Fatalf("nudge missing diagnostics: %q", message.Content)
+			}
+			found = true
+		}
+	}
+	if !found {
+		t.Fatal("finalization gate did not deliver the late diagnostics nudge")
+	}
+}

internal/agent/command_prefix.go

@@ -1,9 +1,12 @@
 package agent
 
 import (
+	"fmt"
+	"runtime"
 	"strings"
 
 	"github.com/Gitlawb/zero/internal/sandbox"
+	"github.com/Gitlawb/zero/internal/tools"
 	"mvdan.cc/sh/v3/syntax"
 )
 
@@ -15,43 +18,130 @@ func proposedCommandPrefix(toolName string, args map[string]any) []string {
 	if !ok {
 		return nil
 	}
-	tokens, ok := safeShellCommandTokens(command)
+	segments, ok := safeShellCommandSegments(command)
 	if !ok {
 		return nil
 	}
 	if requested, ok := requestedPrefixRule(args); ok {
-		if safeRequestedPrefix(requested, tokens) {
+		if safeRequestedPrefixForSegments(requested, segments) {
 			return requested
 		}
 		return nil
 	}
-	if !sandbox.ValidCommandPrefix(tokens) {
+	// Only propose approving a prefix of one segment when every other segment
+	// in the command is independently known-safe. Once a prefix is approved,
+	// shellExecutionArgsForApproval escalates the whole command (every
+	// segment, not just the approved one) to bypass the sandbox, so offering a
+	// prefix that leaves an MSYS-prone (or otherwise unsafe) segment uncovered
+	// would let that segment run unsandboxed without ever being reviewed.
+	for index, tokens := range segments {
+		if knownSafeCommandSegment(tokens) {
+			continue
+		}
+		if !sandbox.ValidCommandPrefix(tokens) || !otherSegmentsKnownSafe(segments, index) {
+			return nil
+		}
+		return append([]string(nil), tokens...)
+	}
+	if len(segments) == 0 || !sandbox.ValidCommandPrefix(segments[0]) {
 		return nil
 	}
-	return append([]string(nil), tokens...)
+	return append([]string(nil), segments[0]...)
+}
+
+// otherSegmentsKnownSafe reports whether every segment other than the one at
+// skip is known-safe on its own.
+func otherSegmentsKnownSafe(segments [][]string, skip int) bool {
+	for index, tokens := range segments {
+		if index == skip {
+			continue
+		}
+		if !knownSafeCommandSegment(tokens) {
+			return false
+		}
+	}
+	return true
 }
 
 func matchCommandPrefix(toolName string, args map[string]any, options Options) (sandbox.CommandPrefixGrant, bool, bool) {
 	if !isShellCommandTool(toolName) || options.Sandbox == nil {
 		return sandbox.CommandPrefixGrant{}, false, false
 	}
+	if shellCommandAdditionalPermissionsRequested(args) {
+		return sandbox.CommandPrefixGrant{}, false, false
+	}
 	command, ok := firstStringArg(args, "command", "cmd", "script", "shell")
 	if !ok {
 		return sandbox.CommandPrefixGrant{}, false, false
 	}
-	tokens, ok := safeShellCommandTokens(command)
+	segments, ok := safeShellCommandSegments(command)
 	if !ok {
 		return sandbox.CommandPrefixGrant{}, false, false
 	}
-	if grant, ok := options.Sandbox.LookupCommandPrefix(toolName, tokens); ok {
-		return grant, true, false
+	var matched sandbox.CommandPrefixGrant
+	matchedAny := false
+	matchedSession := false
+	for _, tokens := range segments {
+		if grant, ok := options.Sandbox.LookupCommandPrefix(toolName, tokens); ok {
+			if !matchedAny {
+				matched = grant
+			}
+			matchedAny = true
+			continue
+		}
+		if grant, ok := options.Sandbox.LookupCommandPrefixForSession(toolName, tokens); ok {
+			if !matchedAny {
+				matched = grant
+			}
+			matchedAny = true
+			matchedSession = true
+			continue
+		}
+		if knownSafeCommandSegment(tokens) {
+			continue
+		}
+		return sandbox.CommandPrefixGrant{}, false, false
 	}
-	if grant, ok := options.Sandbox.LookupCommandPrefixForSession(toolName, tokens); ok {
-		return grant, true, true
+	if matchedAny {
+		return matched, true, matchedSession
 	}
 	return sandbox.CommandPrefixGrant{}, false, false
 }
 
+func shellExecutionArgsForApproval(toolName string, args map[string]any, action PermissionDecisionAction, options Options) map[string]any {
+	if !isShellCommandTool(toolName) || !shellPrefixApprovalBypassesSandbox(action) {
+		return args
+	}
+	if options.Sandbox == nil || !options.Sandbox.UnsandboxedExecutionAllowed() {
+		return args
+	}
+	if shellCommandAdditionalPermissionsRequested(args) || shellCommandRequiresEscalated(args) {
+		return args
+	}
+	planned := cloneArgs(args)
+	if planned == nil {
+		planned = map[string]any{}
+	}
+	planned["sandbox_permissions"] = string(tools.SandboxPermissionsRequireEscalated)
+	return planned
+}
+
+func shellPrefixApprovalBypassesSandbox(action PermissionDecisionAction) bool {
+	return action == PermissionDecisionAllowPrefix || action == PermissionDecisionAlwaysAllowPrefix
+}
+
+func shellCommandRequiresEscalated(args map[string]any) bool {
+	raw, ok := args["sandbox_permissions"]
+	if !ok || raw == nil {
+		return false
+	}
+	value, ok := raw.(string)
+	if !ok {
+		value = fmt.Sprint(raw)
+	}
+	return strings.TrimSpace(value) == string(tools.SandboxPermissionsRequireEscalated)
+}
+
 func isShellCommandTool(toolName string) bool {
 	return toolName == "bash" || toolName == "exec_command"
 }
@@ -103,33 +193,81 @@ func cleanPrefixRule(prefix []string) []string {
 	return cleaned
 }
 
-func safeRequestedPrefix(prefix []string, command []string) bool {
-	if len(prefix) == 0 || !sandbox.ValidCommandPrefix(prefix) || len(prefix) > len(command) {
+func safeRequestedPrefixForSegments(prefix []string, segments [][]string) bool {
+	if len(prefix) == 0 || !sandbox.ValidCommandPrefix(prefix) {
 		return false
 	}
-	for index := range prefix {
-		if prefix[index] != command[index] {
+	matched := false
+	for _, command := range segments {
+		if len(prefix) > len(command) {
+			if knownSafeCommandSegment(command) {
+				continue
+			}
 			return false
 		}
+		if hasStringPrefix(command, prefix) {
+			matched = true
+			continue
+		}
+		if knownSafeCommandSegment(command) {
+			continue
+		}
+		return false
 	}
-	return true
+	return matched
 }
 
-func safeShellCommandTokens(command string) ([]string, bool) {
+func safeShellCommandSegments(command string) ([][]string, bool) {
 	command = strings.TrimSpace(command)
 	if command == "" {
 		return nil, false
 	}
 	file, err := syntax.NewParser().Parse(strings.NewReader(command), "")
-	if err != nil || len(file.Stmts) != 1 {
+	if err != nil || len(file.Stmts) == 0 {
 		return nil, false
 	}
-	stmt := file.Stmts[0]
-	if stmt.Negated || stmt.Background || stmt.Coprocess || stmt.Disown || stmt.Semicolon.IsValid() || len(stmt.Redirs) > 0 {
+	segments := make([][]string, 0, len(file.Stmts))
+	for _, stmt := range file.Stmts {
+		if !collectSafeShellStatement(stmt, &segments) {
+			return nil, false
+		}
+	}
+	if len(segments) == 0 {
 		return nil, false
 	}
-	call, ok := stmt.Cmd.(*syntax.CallExpr)
-	if !ok || len(call.Assigns) > 0 || len(call.Args) == 0 {
+	return segments, true
+}
+
+func collectSafeShellStatement(stmt *syntax.Stmt, segments *[][]string) bool {
+	if stmt == nil || stmt.Negated || stmt.Background || stmt.Coprocess || stmt.Disown || len(stmt.Redirs) > 0 {
+		return false
+	}
+	return collectSafeShellCommand(stmt.Cmd, segments)
+}
+
+func collectSafeShellCommand(cmd syntax.Command, segments *[][]string) bool {
+	switch node := cmd.(type) {
+	case *syntax.CallExpr:
+		tokens, ok := literalCallTokens(node)
+		if !ok {
+			return false
+		}
+		*segments = append(*segments, tokens)
+		return true
+	case *syntax.BinaryCmd:
+		switch node.Op {
+		case syntax.AndStmt, syntax.OrStmt, syntax.Pipe:
+		default:
+			return false
+		}
+		return collectSafeShellStatement(node.X, segments) && collectSafeShellStatement(node.Y, segments)
+	default:
+		return false
+	}
+}
+
+func literalCallTokens(call *syntax.CallExpr) ([]string, bool) {
+	if call == nil || len(call.Assigns) > 0 || len(call.Args) == 0 {
 		return nil, false
 	}
 	tokens := make([]string, 0, len(call.Args))
@@ -138,17 +276,232 @@ func safeShellCommandTokens(command string) ([]string, bool) {
 			return nil, false
 		}
 		lit, ok := word.Parts[0].(*syntax.Lit)
-		if !ok {
+		if !ok || strings.ContainsAny(lit.Value, "*?[]{}") {
 			return nil, false
 		}
 		tokens = append(tokens, lit.Value)
 	}
-	if !sandbox.ValidCommandPrefix(tokens) {
-		return nil, false
-	}
 	return tokens, true
 }
 
+func knownSafeCommandSegment(command []string) bool {
+	if len(command) == 0 {
+		return false
+	}
+	name := commandName(command[0])
+	if runtime.GOOS == "windows" && tools.MsysProneCommandName(name) {
+		return false
+	}
+	switch name {
+	case "cat", "cd", "cut", "echo", "expr", "false", "grep", "head", "id",
+		"ls", "nl", "paste", "pwd", "rev", "seq", "stat", "tail", "tr",
+		"true", "uname", "uniq", "wc", "which", "whoami":
+		return true
+	case "base64":
+		return safeBase64Command(command[1:])
+	case "find":
+		return safeFindCommand(command[1:])
+	case "rg":
+		return safeRipgrepCommand(command[1:])
+	case "sed":
+		return safeSedCommand(command[1:])
+	case "git":
+		return safeGitCommand(command)
+	default:
+		return false
+	}
+}
+
+func safeBase64Command(args []string) bool {
+	for _, arg := range args {
+		if arg == "-o" || arg == "--output" || strings.HasPrefix(arg, "--output=") || (strings.HasPrefix(arg, "-o") && arg != "-o") {
+			return false
+		}
+	}
+	return true
+}
+
+func safeFindCommand(args []string) bool {
+	unsafe := map[string]bool{
+		"-exec": true, "-execdir": true, "-ok": true, "-okdir": true,
+		"-delete": true,
+		"-fls":    true, "-fprint": true, "-fprint0": true, "-fprintf": true,
+	}
+	for _, arg := range args {
+		if unsafe[arg] {
+			return false
+		}
+	}
+	return true
+}
+
+func safeRipgrepCommand(args []string) bool {
+	for _, arg := range args {
+		if arg == "--search-zip" || arg == "-z" || arg == "--pre" || arg == "--hostname-bin" ||
+			strings.HasPrefix(arg, "--pre=") || strings.HasPrefix(arg, "--hostname-bin=") {
+			return false
+		}
+	}
+	return true
+}
+
+func safeSedCommand(args []string) bool {
+	if len(args) < 2 || len(args) > 3 || args[0] != "-n" {
+		return false
+	}
+	return validSedPrintArg(args[1])
+}
+
+func validSedPrintArg(arg string) bool {
+	if !strings.HasSuffix(arg, "p") {
+		return false
+	}
+	body := strings.TrimSuffix(arg, "p")
+	if body == "" {
+		return false
+	}
+	for _, part := range strings.Split(body, ",") {
+		if part == "" {
+			return false
+		}
+		for _, r := range part {
+			if r < '0' || r > '9' {
+				return false
+			}
+		}
+	}
+	return true
+}
+
+func safeGitCommand(command []string) bool {
+	subIndex, subcommand, ok := gitSubcommand(command)
+	if !ok {
+		return false
+	}
+	if gitHasUnsafeGlobalOption(command[1:subIndex]) {
+		return false
+	}
+	args := command[subIndex+1:]
+	switch subcommand {
+	case "status", "log", "diff", "show":
+		return gitArgsReadOnly(args)
+	case "branch":
+		return gitArgsReadOnly(args) && gitBranchReadOnly(args)
+	default:
+		return false
+	}
+}
+
+func gitSubcommand(command []string) (int, string, bool) {
+	for index := 1; index < len(command); index++ {
+		arg := command[index]
+		if gitOptionConsumesValue(arg) {
+			index++
+			continue
+		}
+		if gitOptionHasInlineValue(arg) || arg == "--" || strings.HasPrefix(arg, "-") {
+			continue
+		}
+		switch arg {
+		case "status", "log", "diff", "show", "branch":
+			return index, arg, true
+		default:
+			return 0, "", false
+		}
+	}
+	return 0, "", false
+}
+
+func gitOptionConsumesValue(arg string) bool {
+	switch arg {
+	case "-C", "-c", "--config-env", "--exec-path", "--git-dir", "--namespace", "--super-prefix", "--work-tree":
+		return true
+	default:
+		return false
+	}
+}
+
+func gitOptionHasInlineValue(arg string) bool {
+	return strings.HasPrefix(arg, "--config-env=") ||
+		strings.HasPrefix(arg, "--exec-path=") ||
+		strings.HasPrefix(arg, "--git-dir=") ||
+		strings.HasPrefix(arg, "--namespace=") ||
+		strings.HasPrefix(arg, "--super-prefix=") ||
+		strings.HasPrefix(arg, "--work-tree=") ||
+		((strings.HasPrefix(arg, "-C") || strings.HasPrefix(arg, "-c")) && len(arg) > 2)
+}
+
+func gitHasUnsafeGlobalOption(args []string) bool {
+	for index := 0; index < len(args); index++ {
+		arg := args[index]
+		switch {
+		case strings.HasPrefix(arg, "--upload-pack"):
+			return true
+		case arg == "-C" || strings.HasPrefix(arg, "-C"):
+			return true
+		case gitOptionConsumesValue(arg):
+			index++
+		}
+	}
+	return false
+}
+
+func gitArgsReadOnly(args []string) bool {
+	for _, arg := range args {
+		switch {
+		case strings.HasPrefix(arg, "--output="), strings.HasPrefix(arg, "--exec="):
+			return false
+		case arg == "--output", arg == "--exec":
+			return false
+		}
+	}
+	return true
+}
+
+func gitBranchReadOnly(args []string) bool {
+	if len(args) == 0 {
+		return true
+	}
+	for _, arg := range args {
+		if strings.HasPrefix(arg, "--format=") {
+			continue
+		}
+		switch arg {
+		case "--list", "-l", "--show-current", "-a", "--all", "-r", "--remotes", "-v", "-vv", "--verbose":
+			continue
+		default:
+			return false
+		}
+	}
+	return true
+}
+
+func commandName(raw string) string {
+	name := strings.TrimSpace(raw)
+	if index := strings.LastIndexAny(name, `/\`); index >= 0 {
+		name = name[index+1:]
+	}
+	name = strings.ToLower(name)
+	for _, suffix := range []string{".exe", ".cmd", ".bat", ".com"} {
+		if strings.HasSuffix(name, suffix) {
+			return strings.TrimSuffix(name, suffix)
+		}
+	}
+	return name
+}
+
+func hasStringPrefix(values []string, prefix []string) bool {
+	if len(prefix) == 0 || len(prefix) > len(values) {
+		return false
+	}
+	for index := range prefix {
+		if values[index] != prefix[index] {
+			return false
+		}
+	}
+	return true
+}
+
 func equalStringSlices(left []string, right []string) bool {
 	if len(left) != len(right) {
 		return false

internal/agent/command_prefix_test.go

@@ -1,6 +1,12 @@
 package agent
 
-import "testing"
+import (
+	"runtime"
+	"testing"
+
+	"github.com/Gitlawb/zero/internal/sandbox"
+	"github.com/Gitlawb/zero/internal/tools"
+)
 
 func TestProposedCommandPrefixUsesSafeSimpleCommands(t *testing.T) {
 	got := proposedCommandPrefix("bash", map[string]any{"command": "go test ./..."})
@@ -29,6 +35,60 @@ func TestProposedCommandPrefixHonorsValidatedRequestedPrefix(t *testing.T) {
 	}
 }
 
+func TestProposedCommandPrefixSupportsSegmentedCommands(t *testing.T) {
+	got := proposedCommandPrefix("bash", map[string]any{"command": "ps aux | head -5"})
+	if runtime.GOOS == "windows" {
+		// head is MSYS-prone on Windows (#458), so proposedCommandPrefix must
+		// not offer "ps aux" as a reusable prefix here: approving it would
+		// escalate the whole command, including the uncovered head segment,
+		// to bypass the sandbox unreviewed. See
+		// TestProposedCommandPrefixRejectsPrefixLeavingUnsafeTailUncovered for
+		// the platform-independent regression coverage of this behavior.
+		if got != nil {
+			t.Fatalf("expected no prefix on Windows because head is MSYS-prone, got %#v", got)
+		}
+		return
+	}
+	want := []string{"ps", "aux"}
+	if !equalStringSlices(got, want) {
+		t.Fatalf("prefix = %#v, want %#v", got, want)
+	}
+}
+
+// TestProposedCommandPrefixRejectsPrefixLeavingUnsafeTailUncovered guards
+// against proposedCommandPrefix offering to approve one segment of a
+// multi-segment command (e.g. "ps aux") while a different segment (e.g. "npm
+// install") is not known-safe. shellExecutionArgsForApproval escalates the
+// entire command once any prefix is approved, so an uncovered unsafe segment
+// would bypass the sandbox unreviewed. Uses npm, which is never known-safe on
+// any platform, so the assertion does not depend on runtime.GOOS.
+func TestProposedCommandPrefixRejectsPrefixLeavingUnsafeTailUncovered(t *testing.T) {
+	if got := proposedCommandPrefix("bash", map[string]any{"command": "ps aux && npm install"}); got != nil {
+		t.Fatalf("expected no prefix because npm segment is not known-safe, got %#v", got)
+	}
+}
+
+func TestProposedCommandPrefixHonorsRequestedPrefixAcrossSegments(t *testing.T) {
+	got := proposedCommandPrefix("bash", map[string]any{
+		"command":     "git status --short && git status --branch",
+		"prefix_rule": []any{"git", "status"},
+	})
+	want := []string{"git", "status"}
+	if !equalStringSlices(got, want) {
+		t.Fatalf("prefix = %#v, want %#v", got, want)
+	}
+}
+
+func TestProposedCommandPrefixRejectsRequestedPrefixThatDoesNotCoverSegments(t *testing.T) {
+	got := proposedCommandPrefix("bash", map[string]any{
+		"command":     "ps aux && npm install",
+		"prefix_rule": []any{"ps", "aux"},
+	})
+	if got != nil {
+		t.Fatalf("partial requested prefix should be rejected, got %#v", got)
+	}
+}
+
 func TestProposedCommandPrefixRejectsUnsafeRequestedPrefix(t *testing.T) {
 	got := proposedCommandPrefix("bash", map[string]any{
 		"command":     "git status --short",
@@ -41,7 +101,6 @@ func TestProposedCommandPrefixRejectsUnsafeRequestedPrefix(t *testing.T) {
 
 func TestProposedCommandPrefixRejectsUnsafeShellForms(t *testing.T) {
 	cases := []string{
-		"echo hi && echo bye",
 		"cat < in > out",
 		"FOO=bar go test",
 		"echo $(whoami)",
@@ -85,6 +144,47 @@ func TestProposedCommandPrefixRejectsUnsafeLaunchers(t *testing.T) {
 	}
 }
 
+func TestMatchCommandPrefixCoversSegmentedCommandWithSafeTail(t *testing.T) {
+	engine := sandbox.NewEngine(sandbox.EngineOptions{WorkspaceRoot: t.TempDir()})
+	engine.GrantCommandPrefixForSession("bash", []string{"ps", "aux"})
+	// head is MSYS-prone on Windows (#458) and must not count as a known-safe tail.
+	command := "ps aux | head -5"
+	if runtime.GOOS == "windows" {
+		command = "ps aux | echo ok"
+	}
+
+	grant, ok, session := matchCommandPrefix("bash", map[string]any{"command": command}, Options{Sandbox: engine})
+	if !ok || !session || !equalStringSlices(grant.Prefix, []string{"ps", "aux"}) {
+		t.Fatalf("match = %#v ok=%v session=%v, want session ps aux prefix", grant, ok, session)
+	}
+}
+
+func TestKnownSafeCommandSegmentRejectsMsysProneOnWindows(t *testing.T) {
+	if runtime.GOOS != "windows" {
+		t.Skip("windows-only known-safe MSYS guard")
+	}
+	for _, command := range [][]string{{"head", "-5"}, {"cat", "file.txt"}, {"grep", "pat"}} {
+		if knownSafeCommandSegment(command) {
+			t.Fatalf("expected %q to be unsafe on Windows, got known-safe", command)
+		}
+	}
+	if !knownSafeCommandSegment([]string{"echo", "ok"}) {
+		t.Fatal("expected echo to remain known-safe on Windows")
+	}
+	if !tools.MsysProneCommandName("head") {
+		t.Fatal("expected head to be MSYS-prone")
+	}
+}
+
+func TestMatchCommandPrefixRejectsUncoveredSegment(t *testing.T) {
+	engine := sandbox.NewEngine(sandbox.EngineOptions{WorkspaceRoot: t.TempDir()})
+	engine.GrantCommandPrefixForSession("bash", []string{"ps", "aux"})
+
+	if grant, ok, session := matchCommandPrefix("bash", map[string]any{"command": "ps aux && npm install"}, Options{Sandbox: engine}); ok {
+		t.Fatalf("match = %#v session=%v, want no match because npm segment is uncovered", grant, session)
+	}
+}
+
 func TestProposedCommandPrefixRejectsRequestedUnsafeLauncherPrefix(t *testing.T) {
 	got := proposedCommandPrefix("bash", map[string]any{
 		"command":     "find . -type f",

internal/agent/compaction.go

@@ -6,6 +6,7 @@ import (
 	"errors"
 	"strings"
 
+	"github.com/Gitlawb/zero/internal/trace"
 	"github.com/Gitlawb/zero/internal/zeroruntime"
 )
 
@@ -21,12 +22,17 @@ import (
 // the agent loop owns the provider wiring.
 
 // defaultCompactionPreserveLast is how many trailing messages are kept verbatim
-// when the caller does not specify a preserve count.
-const defaultCompactionPreserveLast = 8
+// when the caller does not specify a preserve count. Kept in sync with the replay
+// reconstruction default in internal/sessions/replay.go.
+const defaultCompactionPreserveLast = 6
 
 // compactionTriggerRatio is the fraction of the context window at which
-// proactive compaction kicks in (top of each turn).
-const compactionTriggerRatio = 0.8
+// proactive compaction kicks in (top of each turn). NOTE: this is speculative
+// tuning — validate against the per-turn latency/compaction metrics (roadmap 5.2)
+// before trusting it. Measured fixed overhead is ~7.4k tokens (not the ~16k the
+// roadmap assumed), so context fills slower; if data shows 0.7 triggers redundant
+// summarization, raise it back toward 0.8.
+const compactionTriggerRatio = 0.7
 
 // summaryLabel prefixes the injected summary so it is unmistakable in the
 // transcript (and so tests can assert on it).
@@ -280,6 +286,25 @@ func isContextLimitError(message string) bool {
 	return false
 }
 
+// isStreamTimeoutError reports whether a streamed error is a provider stream
+// idle/stall timeout (providerio surfaces these as "idle timeout after …" /
+// "no output for …" / "stream stalled"). Such a turn produced nothing, so when
+// no content was forwarded yet it can be safely re-issued on a fresh connection
+// (see the stall-retry in the agent loop). Substring + case-insensitive to
+// tolerate the prefixed "provider stream error: …" wrapping.
+func isStreamTimeoutError(message string) bool {
+	lowered := strings.ToLower(strings.TrimSpace(message))
+	if lowered == "" {
+		return false
+	}
+	for _, needle := range []string{"idle timeout after", "no output for", "stream stalled"} {
+		if strings.Contains(lowered, needle) {
+			return true
+		}
+	}
+	return false
+}
+
 // compactionState carries the per-run state the agent loop needs to compact a
 // conversation safely. It is created once per Run and is a no-op whenever
 // options.ContextWindow <= 0.
@@ -301,6 +326,43 @@ type compactionState struct {
 	// OnText is deliberately NOT forwarded (compaction stays invisible to the user),
 	// but its token COST must still be counted so usage reports and budgets include it.
 	onUsage func(Usage)
+
+	// calibrationRatio scales the raw byte/4 token estimate toward the provider's
+	// real prompt-token count. ApproxTextTokens over-counts code-heavy content by
+	// ~15-20%, which would trip compaction early (at ~60% of true capacity). It
+	// starts at 1.0 and converges via an EMA as each turn reports actual usage, so
+	// later turns compact nearer to real capacity. Zero is treated as 1.0.
+	calibrationRatio float64
+}
+
+// calibrate folds one turn's (rawEstimate, actualPromptTokens) sample into the
+// running calibration ratio. A single sample is clamped to a sane band so an
+// outlier (a huge cache-read turn, a provider-overhead spike) can't skew the
+// estimate enough to disable or thrash compaction.
+func (state *compactionState) calibrate(rawEstimate int, actualPromptTokens int) {
+	if !state.enabled || rawEstimate <= 0 || actualPromptTokens <= 0 {
+		return
+	}
+	sample := float64(actualPromptTokens) / float64(rawEstimate)
+	if sample < 0.5 {
+		sample = 0.5
+	} else if sample > 2.0 {
+		sample = 2.0
+	}
+	if state.calibrationRatio <= 0 {
+		state.calibrationRatio = 1.0
+	}
+	const alpha = 0.3 // weight on the newest sample; smooths jitter across turns
+	state.calibrationRatio = state.calibrationRatio*(1-alpha) + sample*alpha
+}
+
+// calibratedTokens applies the learned ratio to a raw estimate. Before any sample
+// arrives (ratio unset) it returns the raw estimate unchanged.
+func (state *compactionState) calibratedTokens(raw int) int {
+	if state.calibrationRatio <= 0 {
+		return raw
+	}
+	return int(float64(raw) * state.calibrationRatio)
 }
 
 func newCompactionState(options Options) *compactionState {
@@ -329,7 +391,7 @@ func (state *compactionState) maybeCompact(
 	// the messages; both the threshold check and the shrink check below use the
 	// same term so they stay consistent.
 	toolTokens := estimateToolDefTokens(tools)
-	size := estimateTokens(messages) + toolTokens
+	size := state.calibratedTokens(estimateTokens(messages) + toolTokens)
 	if size <= state.threshold {
 		return messages
 	}
@@ -346,7 +408,7 @@ func (state *compactionState) maybeCompact(
 	// entirely and preserve recent turns verbatim.
 	if pruned, reclaimed := pruneStaleToolOutput(messages, state.preserveLast); reclaimed > 0 {
 		messages = pruned
-		size = estimateTokens(messages) + toolTokens
+		size = state.calibratedTokens(estimateTokens(messages) + toolTokens)
 		if size <= state.threshold {
 			state.lowWaterMark = size
 			return messages
@@ -362,13 +424,19 @@ func (state *compactionState) maybeCompact(
 		// later turn) can try again; we never drop messages on failure here.
 		return messages
 	}
-	newSize := estimateTokens(compacted) + toolTokens
+	newSize := state.calibratedTokens(estimateTokens(compacted) + toolTokens)
 	if newSize >= size {
 		// Compaction did not actually shrink anything (e.g. nothing to
 		// summarize). Leave the history untouched and don't churn next turn.
 		state.lowWaterMark = size
 		return messages
 	}
+	// Only count a compaction when it actually shrank the history, so the
+	// compaction counter reflects real context reductions rather than paid
+	// no-ops that left the token budget untouched.
+	if r := trace.FromContext(ctx); r != nil {
+		r.Counter(trace.CounterCompactionCount, 1)
+	}
 	state.lowWaterMark = newSize
 	return compacted
 }
@@ -418,9 +486,13 @@ func (state *compactionState) recover(
 	// one-shot budget now so a provider that keeps returning context-limit errors
 	// after a successful compaction can't loop forever. Store the low-water mark in
 	// the SAME combined (messages + tool-defs) domain maybeCompact uses, so the
-	// proactive shrink-guard compares like with like.
+	// proactive shrink-guard compares like with like. Count it now that the shrink
+	// is confirmed, so the counter mirrors maybeCompact's real-reduction policy.
+	if r := trace.FromContext(ctx); r != nil {
+		r.Counter(trace.CounterCompactionCount, 1)
+	}
 	state.reactiveAttempted = true
-	state.lowWaterMark = estimateTokens(result) + estimateToolDefTokens(tools)
+	state.lowWaterMark = state.calibratedTokens(estimateTokens(result) + estimateToolDefTokens(tools))
 	return result, true, nil
 }
 

internal/agent/compaction_calibrate_test.go

@@ -0,0 +1,50 @@
+package agent
+
+import "testing"
+
+func TestCalibrationConvergesTowardActual(t *testing.T) {
+	state := &compactionState{enabled: true}
+
+	// Before any sample, the estimate is unchanged.
+	if got := state.calibratedTokens(1000); got != 1000 {
+		t.Fatalf("uncalibrated tokens = %d, want 1000", got)
+	}
+
+	// The estimator over-counts: our raw estimate is 1000 but the provider reports
+	// 850 (a 0.85 ratio). Feeding that sample repeatedly should pull the ratio
+	// toward 0.85, so a 1000-token raw estimate calibrates downward.
+	for range 20 {
+		state.calibrate(1000, 850)
+	}
+	got := state.calibratedTokens(1000)
+	if got >= 1000 || got < 820 || got > 880 {
+		t.Fatalf("calibrated tokens = %d, want ~850 after convergence", got)
+	}
+}
+
+func TestCalibrateIgnoresDegenerateSamples(t *testing.T) {
+	state := &compactionState{enabled: true}
+	state.calibrate(0, 500) // zero estimate
+	state.calibrate(500, 0) // zero actual
+	if state.calibrationRatio != 0 {
+		t.Fatalf("degenerate samples must not move the ratio, got %v", state.calibrationRatio)
+	}
+	// Disabled state never calibrates.
+	disabled := &compactionState{enabled: false}
+	disabled.calibrate(1000, 850)
+	if disabled.calibrationRatio != 0 {
+		t.Fatalf("disabled compaction must not calibrate, got %v", disabled.calibrationRatio)
+	}
+}
+
+func TestCalibrateClampsOutliers(t *testing.T) {
+	state := &compactionState{enabled: true}
+	// A wild outlier (10x) is clamped to 2.0, so one bad sample can't blow up the
+	// ratio and disable compaction.
+	for range 50 {
+		state.calibrate(1000, 10000)
+	}
+	if state.calibrationRatio > 2.0 {
+		t.Fatalf("ratio should be clamped at 2.0, got %v", state.calibrationRatio)
+	}
+}

internal/agent/compaction_preserve.go

@@ -20,6 +20,16 @@ const (
 	toolNameUpdatePlan = "update_plan"
 	toolNameToolSearch = "tool_search"
 	toolNameSkill      = "skill"
+	toolNameWriteFile  = "write_file"
+	toolNameEditFile   = "edit_file"
+)
+
+// maxRecentEdits caps how many edited files are carried across a compaction, and
+// maxEditNoteBytes caps each edit's one-line note, so a long editing run can't
+// bloat the preserved-state block.
+const (
+	maxRecentEdits   = 20
+	maxEditNoteBytes = 160
 )
 
 const (
@@ -100,12 +110,123 @@ func formatPlanArguments(arguments string) string {
 }
 
 // skillEntry is a named preserved body. It began as loaded-skill state and is
-// reused for loaded tools and project instruction blocks.
+// reused for loaded tools, project instruction blocks, and recent file edits.
 type skillEntry struct {
 	name string
 	body string
 }
 
+// recentEdits returns the files mutated by write_file/edit_file calls in messages
+// — latest note per path, in last-seen order — as skillEntry{name: path, body:
+// note}. After compaction elides the editing turns, this tells the model WHAT it
+// changed in each file (from the tool's result) so it need not re-read to
+// rediscover its own footprint. Capped at maxRecentEdits paths.
+//
+// Ordering is by LAST edit, not first: re-editing a file moves it to the newest
+// position so the tail cap below keeps the file the model most recently touched
+// rather than an earlier, now-stale entry.
+func recentEdits(messages []zeroruntime.Message) []skillEntry {
+	pathByID := map[string]string{}
+	sequence := make([]string, 0)
+	for _, message := range messages {
+		for _, call := range message.ToolCalls {
+			if call.Name != toolNameWriteFile && call.Name != toolNameEditFile {
+				continue
+			}
+			path := editPathFromArguments(call.Arguments)
+			if path == "" {
+				continue
+			}
+			if call.ID != "" {
+				pathByID[call.ID] = path
+			}
+			sequence = append(sequence, path)
+		}
+	}
+	order := lastSeenOrder(sequence)
+	if len(order) == 0 {
+		return nil
+	}
+
+	noteByPath := map[string]string{}
+	for _, message := range messages {
+		if message.Role != zeroruntime.MessageRoleTool || message.ToolCallID == "" {
+			continue
+		}
+		if path, ok := pathByID[message.ToolCallID]; ok {
+			noteByPath[path] = editNote(message.Content)
+		}
+	}
+
+	// Keep the most recent maxRecentEdits paths (the tail of last-seen order).
+	if len(order) > maxRecentEdits {
+		order = order[len(order)-maxRecentEdits:]
+	}
+	entries := make([]skillEntry, 0, len(order))
+	for _, path := range order {
+		entries = append(entries, skillEntry{name: path, body: noteByPath[path]})
+	}
+	return entries
+}
+
+// lastSeenOrder dedupes paths keeping each at its LAST occurrence, preserving
+// chronological order otherwise. A re-edited path therefore lands at the newest
+// position rather than staying pinned to where it first appeared.
+func lastSeenOrder(paths []string) []string {
+	lastIdx := make(map[string]int, len(paths))
+	for i, p := range paths {
+		lastIdx[p] = i
+	}
+	order := make([]string, 0, len(lastIdx))
+	for i, p := range paths {
+		if lastIdx[p] == i {
+			order = append(order, p)
+		}
+	}
+	return order
+}
+
+// editPathFromArguments pulls the target file path from a write_file/edit_file
+// call's JSON arguments (path/file/file_path/filename aliases). Returns "" on
+// malformed arguments or no path.
+func editPathFromArguments(arguments string) string {
+	var parsed struct {
+		Path     string `json:"path"`
+		File     string `json:"file"`
+		FilePath string `json:"file_path"`
+		Filename string `json:"filename"`
+	}
+	if err := json.Unmarshal([]byte(strings.TrimSpace(arguments)), &parsed); err != nil {
+		return ""
+	}
+	for _, candidate := range []string{parsed.Path, parsed.File, parsed.FilePath, parsed.Filename} {
+		if s := strings.TrimSpace(candidate); s != "" {
+			return s
+		}
+	}
+	return ""
+}
+
+// editNote reduces an edit tool's result to a single short line (its first
+// non-empty line, byte-capped on a rune boundary) for the preserved summary.
+func editNote(content string) string {
+	for _, line := range strings.Split(content, "\n") {
+		s := strings.TrimSpace(line)
+		if s == "" {
+			continue
+		}
+		if len(s) > maxEditNoteBytes {
+			limit := maxEditNoteBytes
+			for limit > 0 && !utf8.RuneStart(s[limit]) {
+				limit--
+			}
+			s = s[:limit] + "…"
+		}
+		return s
+	}
+	return ""
+}
+
 // loadedSkills returns the skills loaded via the skill tool in messages — the
 // latest body per name, in first-seen order — matching each skill tool call to
 // its tool result by id.
@@ -310,11 +431,17 @@ func capBody(body string) string {
 // preservedState is the JSON shape of the carried-across-compaction block.
 type preservedState struct {
 	Plan                string                 `json:"plan,omitempty"`
+	RecentEdits         []preservedEdit        `json:"recent_edits,omitempty"`
 	Tools               []preservedTool        `json:"tools,omitempty"`
 	Skills              []preservedSkill       `json:"skills,omitempty"`
 	ProjectInstructions []preservedInstruction `json:"project_instructions,omitempty"`
 }
 
+type preservedEdit struct {
+	Path string `json:"path"`
+	Note string `json:"note,omitempty"`
+}
+
 type preservedTool struct {
 	Name string `json:"name"`
 	Body string `json:"body"`
@@ -347,6 +474,12 @@ func appendPreservedState(summary string, middle []zeroruntime.Message) string {
 		plan = priorState.Plan
 	}
 
+	// Recent edits: merge edits preserved earlier with fresh write_file/edit_file
+	// results in middle (newer note per path wins AND moves to the newest
+	// position), so the tail cap keeps the files the model most recently touched
+	// across repeated compactions rather than dropping a just-re-edited file.
+	edits := capRecentEdits(mergeRecentEdits(preservedEditsToEntries(priorState.RecentEdits), recentEdits(middle)))
+
 	// Tools: preserve deferred tool_search schemas from the transcript. Fresh
 	// loads override older carried copies by name.
 	tools := mergeSkillEntries(preservedToolsToEntries(priorState.Tools), loadedToolSchemas(middle))
@@ -360,19 +493,76 @@ func appendPreservedState(summary string, middle []zeroruntime.Message) string {
 		projectInstructionEntries(middle),
 	)
 
-	if block := formatPreservedState(plan, tools, skills, instructions); block != "" {
+	if block := formatPreservedState(plan, edits, tools, skills, instructions); block != "" {
 		summary += "\n\n" + block
 	}
 	return summary
 }
 
+// mergeRecentEdits overlays fresh edits onto edits preserved by an earlier
+// compaction. Unlike mergeSkillEntries (which keeps refreshed entries in their
+// original slot), a path touched again by a fresh edit MOVES to the newest
+// position and takes the fresh note, so capRecentEdits — which keeps only the
+// most-recent tail — never drops a file the model just re-edited. Paths present
+// only in the older set keep their relative order, ahead of the fresh ones.
+func mergeRecentEdits(older, newer []skillEntry) []skillEntry {
+	if len(newer) == 0 {
+		return older
+	}
+	freshBody := make(map[string]string, len(newer))
+	freshOrder := make([]string, 0, len(newer))
+	for _, e := range newer {
+		if _, ok := freshBody[e.name]; !ok {
+			freshOrder = append(freshOrder, e.name)
+		}
+		freshBody[e.name] = e.body
+	}
+	merged := make([]skillEntry, 0, len(older)+len(freshOrder))
+	// Older entries that were NOT re-edited keep their position, oldest first.
+	for _, e := range older {
+		if _, refreshed := freshBody[e.name]; refreshed {
+			continue // re-appended at its newest position below
+		}
+		merged = append(merged, e)
+	}
+	// Fresh edits follow, in last-seen order, each carrying the fresh note — so
+	// the most recently touched files sit at the tail the cap preserves.
+	for _, name := range freshOrder {
+		merged = append(merged, skillEntry{name: name, body: freshBody[name]})
+	}
+	return merged
+}
+
+// capRecentEdits bounds the preserved edit list to the most recent maxRecentEdits
+// entries after a merge, so repeated compactions can't grow it without limit.
+func capRecentEdits(entries []skillEntry) []skillEntry {
+	if len(entries) > maxRecentEdits {
+		return entries[len(entries)-maxRecentEdits:]
+	}
+	return entries
+}
+
+func preservedEditsToEntries(edits []preservedEdit) []skillEntry {
+	entries := make([]skillEntry, 0, len(edits))
+	for _, e := range edits {
+		if e.Path == "" {
+			continue
+		}
+		entries = append(entries, skillEntry{name: e.Path, body: e.Note})
+	}
+	return entries
+}
+
 // formatPreservedState renders state as the labelled, single-line
 // JSON block. Returns "" when there is nothing to preserve.
-func formatPreservedState(plan string, tools, skills, instructions []skillEntry) string {
-	if plan == "" && len(tools) == 0 && len(skills) == 0 && len(instructions) == 0 {
+func formatPreservedState(plan string, edits, tools, skills, instructions []skillEntry) string {
+	if plan == "" && len(edits) == 0 && len(tools) == 0 && len(skills) == 0 && len(instructions) == 0 {
 		return ""
 	}
 	state := preservedState{Plan: plan}
+	for _, e := range edits {
+		state.RecentEdits = append(state.RecentEdits, preservedEdit{Path: e.name, Note: e.body})
+	}
 	for _, t := range tools {
 		state.Tools = append(state.Tools, preservedTool{Name: t.name, Body: t.body})
 	}

internal/agent/compaction_recent_edits_test.go

@@ -0,0 +1,121 @@
+package agent
+
+import (
+	"fmt"
+	"strings"
+	"testing"
+
+	"github.com/Gitlawb/zero/internal/zeroruntime"
+)
+
+// recentEdits extracts each mutated file's path and a one-line note from the
+// matching tool result, latest note per path in last-seen order.
+func TestRecentEditsExtractsPathsAndNotes(t *testing.T) {
+	messages := []zeroruntime.Message{
+		{Role: zeroruntime.MessageRoleAssistant, ToolCalls: []zeroruntime.ToolCall{
+			{ID: "e1", Name: "write_file", Arguments: `{"path":"internal/foo.go","content":"package foo"}`},
+		}},
+		{Role: zeroruntime.MessageRoleTool, ToolCallID: "e1", Content: "Wrote internal/foo.go (12 lines)"},
+		{Role: zeroruntime.MessageRoleAssistant, ToolCalls: []zeroruntime.ToolCall{
+			{ID: "e2", Name: "edit_file", Arguments: `{"path":"internal/bar.go","old_string":"a","new_string":"b"}`},
+		}},
+		{Role: zeroruntime.MessageRoleTool, ToolCallID: "e2", Content: "Applied edit to internal/bar.go"},
+	}
+
+	edits := recentEdits(messages)
+	if len(edits) != 2 {
+		t.Fatalf("expected 2 edited files, got %d: %#v", len(edits), edits)
+	}
+	if edits[0].name != "internal/foo.go" || !strings.Contains(edits[0].body, "12 lines") {
+		t.Fatalf("first edit = %#v, want foo.go with its note", edits[0])
+	}
+	if edits[1].name != "internal/bar.go" || !strings.Contains(edits[1].body, "Applied edit") {
+		t.Fatalf("second edit = %#v, want bar.go with its note", edits[1])
+	}
+}
+
+// After compaction elides the editing turns, the preserved-state block still
+// names the edited files and what changed, so the model needn't re-read them.
+func TestCompactionPreservesRecentEdits(t *testing.T) {
+	messages := []zeroruntime.Message{
+		{Role: zeroruntime.MessageRoleSystem, Content: "system"},
+		{Role: zeroruntime.MessageRoleUser, Content: "add a flag"},
+		{Role: zeroruntime.MessageRoleAssistant, Content: "editing", ToolCalls: []zeroruntime.ToolCall{
+			{ID: "e1", Name: "write_file", Arguments: `{"path":"cmd/main.go","content":"..."}`},
+		}},
+		{Role: zeroruntime.MessageRoleTool, ToolCallID: "e1", Content: "Wrote cmd/main.go (adds --version flag)"},
+		{Role: zeroruntime.MessageRoleAssistant, Content: "done"},
+		{Role: zeroruntime.MessageRoleUser, Content: "continue"},
+		{Role: zeroruntime.MessageRoleAssistant, Content: "continuing"},
+	}
+	summary := compactStateConversation(t, messages)
+
+	state := parsePreservedStateBlock(summary)
+	if len(state.RecentEdits) != 1 {
+		t.Fatalf("expected 1 preserved edit, got %#v", state.RecentEdits)
+	}
+	if state.RecentEdits[0].Path != "cmd/main.go" || !strings.Contains(state.RecentEdits[0].Note, "--version") {
+		t.Fatalf("preserved edit = %#v, want cmd/main.go + its note", state.RecentEdits[0])
+	}
+}
+
+// A fresh edit note for a path overrides the one carried from an earlier
+// compaction (newer wins) and moves the path to the newest position rather than
+// duplicating it or leaving it pinned to its old slot.
+func TestRecentEditsMergeNewerWins(t *testing.T) {
+	prior := preservedState{RecentEdits: []preservedEdit{{Path: "a.go", Note: "old note"}, {Path: "z.go", Note: "z"}}}
+	older := preservedEditsToEntries(prior.RecentEdits)
+	newer := []skillEntry{{name: "a.go", body: "new note"}, {name: "b.go", body: "added"}}
+
+	merged := mergeRecentEdits(older, newer)
+	// z.go (untouched) keeps its lead; a.go is re-edited so it takes the new note
+	// and moves behind z.go, ahead of the brand-new b.go: [z.go, a.go, b.go].
+	if len(merged) != 3 {
+		t.Fatalf("expected z.go + refreshed a.go + b.go, got %#v", merged)
+	}
+	if merged[0].name != "z.go" {
+		t.Fatalf("untouched z.go should stay first, got %#v", merged)
+	}
+	if merged[1].name != "a.go" || merged[1].body != "new note" {
+		t.Fatalf("re-edited a.go should move to the newest edits with the new note, got %#v", merged[1])
+	}
+	if merged[2].name != "b.go" {
+		t.Fatalf("brand-new b.go should be last, got %#v", merged[2])
+	}
+}
+
+// Regression: once more than maxRecentEdits distinct files are tracked, re-editing
+// an early file must move it into the capped tail rather than leaving it to be
+// dropped — otherwise the next compaction omits exactly the file the model most
+// recently touched.
+func TestRecentEditsCapKeepsReeditedFile(t *testing.T) {
+	// Earlier compaction preserved exactly maxRecentEdits files, f0 the oldest.
+	older := make([]skillEntry, 0, maxRecentEdits)
+	for i := 0; i < maxRecentEdits; i++ {
+		older = append(older, skillEntry{name: fmt.Sprintf("f%d.go", i), body: "old"})
+	}
+	// Fresh window re-edits the OLDEST file (f0) and adds a brand-new file, so the
+	// merged list exceeds the cap and something must be dropped.
+	newer := []skillEntry{{name: "f0.go", body: "re-edited"}, {name: "fresh.go", body: "new"}}
+
+	merged := capRecentEdits(mergeRecentEdits(older, newer))
+	if len(merged) != maxRecentEdits {
+		t.Fatalf("cap should keep %d entries, got %d", maxRecentEdits, len(merged))
+	}
+
+	var f0 *skillEntry
+	for i := range merged {
+		if merged[i].name == "f0.go" {
+			f0 = &merged[i]
+		}
+		if merged[i].name == "f1.go" {
+			t.Fatalf("f1.go (now the least-recently touched) should have been dropped, got %#v", merged)
+		}
+	}
+	if f0 == nil {
+		t.Fatalf("re-edited f0.go must survive the cap, got %#v", merged)
+	}
+	if f0.body != "re-edited" {
+		t.Fatalf("surviving f0.go should carry the fresh note, got %q", f0.body)
+	}
+}

internal/agent/completion_gate_test.go

@@ -0,0 +1,205 @@
+package agent
+
+import (
+	"context"
+	"strings"
+	"testing"
+
+	"github.com/Gitlawb/zero/internal/tools"
+	"github.com/Gitlawb/zero/internal/zeroruntime"
+)
+
+// someRequestContains reports whether any message sent to the provider across all
+// turns contains substr — used to assert a continue nudge was actually injected.
+func someRequestContains(requests []zeroruntime.CompletionRequest, substr string) bool {
+	for _, req := range requests {
+		for _, msg := range req.Messages {
+			if strings.Contains(msg.Content, substr) {
+				return true
+			}
+		}
+	}
+	return false
+}
+
+// planTurn is a turn that calls update_plan with the given item statuses (reusing
+// the package's shared toolTurn helper).
+func planTurn(statuses ...string) []zeroruntime.StreamEvent {
+	items := make([]string, len(statuses))
+	for i, s := range statuses {
+		items[i] = `{"content":"step ` + s + `","status":"` + s + `"}`
+	}
+	return toolTurn("plan", "update_plan", `{"plan":[`+strings.Join(items, ",")+`]}`)
+}
+
+// BUG #1 regression: a no-tool-call turn that ends mid-step while plan items are
+// still pending must NOT be accepted as success. The loop must re-prompt (bounded)
+// and, if the model keeps stalling, finalize as INCOMPLETE — never success.
+func TestCompletionGatePendingPlanContinuesThenIncomplete(t *testing.T) {
+	// Mirrors the git-multibranch failure: plan with pending steps, then the model
+	// keeps emitting "…Let me check the SSH configuration:" without acting.
+	cue := "Now I need to configure the SSH server. Let me check the current SSH configuration:"
+	registry := tools.NewRegistry()
+	registry.Register(tools.NewUpdatePlanTool())
+
+	provider := &mockProvider{turns: [][]zeroruntime.StreamEvent{
+		planTurn("completed", "pending", "pending"),
+		textTurn(cue), textTurn(cue), textTurn(cue), textTurn(cue), textTurn(cue),
+	}}
+
+	result, err := Run(context.Background(), "set up a git server", provider, Options{
+		Registry:                registry,
+		MaxTurns:                10,
+		RequireCompletionSignal: true,
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !result.Incomplete {
+		t.Fatalf("expected Incomplete=true (model stalled with pending plan), got false; final=%q turns=%d", result.FinalAnswer, result.Turns)
+	}
+	// 1 plan turn + maxContinueNudges(3) nudged turns + 1 final stalling turn = 5.
+	// Critically it did NOT stop at the first text turn (request 2) as success.
+	if len(provider.requests) != 1+maxContinueNudges+1 {
+		t.Fatalf("expected %d provider turns (1 plan + %d nudges + 1 final), got %d",
+			1+maxContinueNudges+1, maxContinueNudges, len(provider.requests))
+	}
+	if !someRequestContains(provider.requests, continueNudgeMarker) {
+		t.Fatalf("expected a continue nudge (%q) to be injected into the conversation", continueNudgeMarker)
+	}
+}
+
+// A genuinely-complete single-turn answer (no plan, no continuation cue) must
+// still finalize as success — the gate must not break short/read-only tasks.
+func TestCompletionGateAcceptsGenuineCompletion(t *testing.T) {
+	provider := &mockProvider{turns: [][]zeroruntime.StreamEvent{
+		textTurn("The file contains 42 lines."),
+	}}
+
+	result, err := Run(context.Background(), "count the lines", provider, Options{
+		Registry:                tools.NewRegistry(),
+		MaxTurns:                10,
+		RequireCompletionSignal: true,
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if result.Incomplete {
+		t.Fatalf("genuine completion wrongly marked Incomplete; final=%q", result.FinalAnswer)
+	}
+	if result.FinalAnswer != "The file contains 42 lines." {
+		t.Fatalf("final answer = %q, want the completed answer", result.FinalAnswer)
+	}
+	if len(provider.requests) != 1 {
+		t.Fatalf("expected exactly 1 turn (no spurious re-prompt), got %d", len(provider.requests))
+	}
+}
+
+// A continuation-cue turn triggers a re-prompt, but once the model actually
+// finishes (clean answer, no cue, no pending plan) the run exits as success — the
+// nudge gives the model a path to a legitimate completion.
+func TestCompletionGateContinuesOnCueThenSucceeds(t *testing.T) {
+	provider := &mockProvider{turns: [][]zeroruntime.StreamEvent{
+		textTurn("Let me read the file:"),
+		textTurn("Done. The file has 42 lines."),
+	}}
+
+	result, err := Run(context.Background(), "count the lines", provider, Options{
+		Registry:                tools.NewRegistry(),
+		MaxTurns:                10,
+		RequireCompletionSignal: true,
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if result.Incomplete {
+		t.Fatalf("run wrongly marked Incomplete after the model completed; final=%q", result.FinalAnswer)
+	}
+	if result.FinalAnswer != "Done. The file has 42 lines." {
+		t.Fatalf("final answer = %q, want the completed answer", result.FinalAnswer)
+	}
+	if len(provider.requests) != 2 {
+		t.Fatalf("expected 2 turns (1 cue re-prompted + 1 completion), got %d", len(provider.requests))
+	}
+	if !someRequestContains(provider.requests, continueNudgeMarker) {
+		t.Fatalf("expected a continue nudge after the continuation-cue turn")
+	}
+}
+
+// With the gate OFF (the interactive/TUI default), a continuation-cue turn is
+// accepted as the final answer exactly as before — guaranteeing no behavior
+// change for non-headless callers.
+func TestCompletionGateOffPreservesLegacyBehavior(t *testing.T) {
+	cue := "Let me check the config:"
+	provider := &mockProvider{turns: [][]zeroruntime.StreamEvent{
+		textTurn(cue),
+	}}
+
+	result, err := Run(context.Background(), "do a thing", provider, Options{
+		Registry: tools.NewRegistry(),
+		MaxTurns: 10,
+		// RequireCompletionSignal deliberately left false.
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if result.Incomplete {
+		t.Fatalf("legacy path must never set Incomplete; final=%q", result.FinalAnswer)
+	}
+	if result.FinalAnswer != cue {
+		t.Fatalf("final answer = %q, want %q (legacy: text-only turn is the answer)", result.FinalAnswer, cue)
+	}
+	if len(provider.requests) != 1 {
+		t.Fatalf("legacy path must not re-prompt; got %d turns", len(provider.requests))
+	}
+}
+
+// review #6: the continuation-cue detector must catch a mid-line action announcement
+// that stops on a colon (the git-multibranch case), without flagging genuine closers
+// — a recommendation, a plain summary colon, or a sign-off.
+func TestContinuationCueMatching(t *testing.T) {
+	cases := []struct {
+		text string
+		cue  bool
+	}{
+		{"Now I need to configure the SSH server. Let me check the current SSH configuration:", true},
+		{"Let me read the file:", true},
+		{"Now I'll run the tests:", true},
+		{"Next, I suggest reviewing the changes.", false}, // recommendation, no colon
+		{"Here is the summary:", false},                   // summary colon, no action lead-in
+		{"Let me know if you need anything:", false},      // sign-off
+		{"The function is implemented and all tests pass.", false},
+	}
+	for _, c := range cases {
+		if got := endsWithContinuationCue(c.text); got != c.cue {
+			t.Errorf("endsWithContinuationCue(%q) = %v, want %v", c.text, got, c.cue)
+		}
+	}
+}
+
+// review #4: a run that loops to the MaxTurns ceiling (always calling a tool, so it
+// never reaches the no-tool-call gate) was reported as success. Under the headless
+// gate, a max-turns cutoff is INCOMPLETE — the agent was stopped mid-run, not done.
+func TestMaxTurnsCutoffIsIncompleteUnderGate(t *testing.T) {
+	registry := tools.NewRegistry()
+	registry.Register(tools.NewUpdatePlanTool())
+	toolEvery := toolTurn("c", "update_plan", `{"plan":[{"content":"step","status":"in_progress"}]}`)
+	provider := &mockProvider{turns: [][]zeroruntime.StreamEvent{
+		toolEvery, toolEvery, toolEvery, toolEvery,
+	}}
+
+	result, err := Run(context.Background(), "keep working", provider, Options{
+		Registry:                registry,
+		MaxTurns:                2,
+		RequireCompletionSignal: true,
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !result.Incomplete {
+		t.Fatalf("a max-turns cutoff under the gate must be Incomplete; final=%q", result.FinalAnswer)
+	}
+	if !strings.Contains(result.IncompleteReason, "max-turns") {
+		t.Fatalf("IncompleteReason = %q, want it to cite max-turns", result.IncompleteReason)
+	}
+}

internal/agent/completion_policy.go

@@ -0,0 +1,86 @@
+package agent
+
+// CompletionDecision is the deterministic outcome of evaluating a text-only
+// assistant turn. Uncertain means the loop needs one bounded follow-up action
+// before it can finalize the run.
+type CompletionDecision string
+
+const (
+	CompletionUncertain  CompletionDecision = "uncertain"
+	CompletionComplete   CompletionDecision = "complete"
+	CompletionIncomplete CompletionDecision = "incomplete"
+)
+
+type completionAction string
+
+const (
+	completionActionNone          completionAction = ""
+	completionActionContinue      completionAction = "continue"
+	completionActionSemanticCheck completionAction = "semantic_check"
+)
+
+type completionEvaluation struct {
+	Decision CompletionDecision
+	Action   completionAction
+	Reason   string
+}
+
+// completionPolicy owns the small amount of state needed to keep uncertain
+// decisions bounded. requireSemanticCheck is set only for run profiles that
+// already opted into self-correction; ordinary completion remains entirely
+// local and adds no model call.
+type completionPolicy struct {
+	continueNudges         int
+	requireSemanticCheck   bool
+	semanticCheckRequested bool
+}
+
+func newCompletionPolicy(requireSemanticCheck bool) *completionPolicy {
+	return &completionPolicy{requireSemanticCheck: requireSemanticCheck}
+}
+
+func (policy *completionPolicy) evaluate(text string, planPending bool) completionEvaluation {
+	// A direct admission is strong local evidence and takes precedence over all
+	// weaker signals, avoiding both continuation nudges and semantic checks.
+	if reason := selfReportedIncompletion(text); reason != "" {
+		return completionEvaluation{Decision: CompletionIncomplete, Reason: reason}
+	}
+
+	// A continuation cue is strong unfinished evidence; a pending plan is only
+	// weak evidence because bookkeeping can be stale. Both get a bounded chance
+	// to continue, but only a persisted cue is ultimately incomplete.
+	cue := endsWithContinuationCue(text)
+	if cue || planPending {
+		if policy.continueNudges < maxContinueNudges {
+			policy.continueNudges++
+			reason := "your message ended mid-step"
+			if !cue {
+				reason = "pending plan items remain — finish them, or mark them complete with update_plan if you are done"
+			}
+			return completionEvaluation{
+				Decision: CompletionUncertain,
+				Action:   completionActionContinue,
+				Reason:   reason,
+			}
+		}
+		if cue {
+			return completionEvaluation{
+				Decision: CompletionIncomplete,
+				Reason:   "your message ended mid-step",
+			}
+		}
+	}
+
+	// Profiles with self-correction enabled require one task-grounded semantic
+	// check. It is requested at most once per run; the next locally complete turn
+	// is accepted without another model call.
+	if policy.requireSemanticCheck && !policy.semanticCheckRequested {
+		policy.semanticCheckRequested = true
+		return completionEvaluation{
+			Decision: CompletionUncertain,
+			Action:   completionActionSemanticCheck,
+		}
+	}
+
+	return completionEvaluation{Decision: CompletionComplete, Action: completionActionNone}
+}

internal/agent/completion_policy_test.go

@@ -0,0 +1,64 @@
+package agent
+
+import "testing"
+
+func TestCompletionPolicyLocalEvidenceDecidesWithoutSemanticCheck(t *testing.T) {
+	policy := newCompletionPolicy(false)
+
+	complete := policy.evaluate("Done. All required checks pass.", false)
+	if complete.Decision != CompletionComplete {
+		t.Fatalf("confident completion decision = %q, want %q", complete.Decision, CompletionComplete)
+	}
+
+	incomplete := policy.evaluate("I couldn't verify the result, so this is my best guess.", false)
+	if incomplete.Decision != CompletionIncomplete {
+		t.Fatalf("admitted failure decision = %q, want %q", incomplete.Decision, CompletionIncomplete)
+	}
+}
+
+func TestCompletionPolicyPreservesBoundedPlanStallProtection(t *testing.T) {
+	policy := newCompletionPolicy(false)
+	for attempt := 0; attempt < maxContinueNudges; attempt++ {
+		got := policy.evaluate("Let me inspect the remaining configuration:", true)
+		if got.Decision != CompletionUncertain || got.Action != completionActionContinue {
+			t.Fatalf("attempt %d = (%q, %q), want uncertain continue", attempt+1, got.Decision, got.Action)
+		}
+	}
+
+	got := policy.evaluate("Let me inspect the remaining configuration:", true)
+	if got.Decision != CompletionIncomplete {
+		t.Fatalf("decision after nudge budget = %q, want %q", got.Decision, CompletionIncomplete)
+	}
+}
+
+func TestCompletionPolicyTreatsPendingPlanAsWeakEvidence(t *testing.T) {
+	policy := newCompletionPolicy(false)
+	for attempt := 0; attempt < maxContinueNudges; attempt++ {
+		got := policy.evaluate("All set.", true)
+		if got.Decision != CompletionUncertain || got.Action != completionActionContinue {
+			t.Fatalf("attempt %d = (%q, %q), want uncertain continue", attempt+1, got.Decision, got.Action)
+		}
+	}
+
+	got := policy.evaluate("All set.", true)
+	if got.Decision != CompletionComplete {
+		t.Fatalf("stale plan decision after nudge budget = %q, want %q", got.Decision, CompletionComplete)
+	}
+}
+
+func TestCompletionPolicyAllowsExactlyOneRequiredSemanticCheck(t *testing.T) {
+	policy := newCompletionPolicy(true)
+
+	first := policy.evaluate("Implemented and tested.", false)
+	if first.Decision != CompletionUncertain || first.Action != completionActionSemanticCheck {
+		t.Fatalf("first decision = (%q, %q), want uncertain semantic check", first.Decision, first.Action)
+	}
+
+	second := policy.evaluate("PASS. The result meets the task criterion.", false)
+	if second.Decision != CompletionComplete {
+		t.Fatalf("post-check decision = %q, want %q", second.Decision, CompletionComplete)
+	}
+	if second.Action == completionActionSemanticCheck {
+		t.Fatal("semantic check was requested more than once")
+	}
+}

internal/agent/confirmation_policy.md

@@ -81,3 +81,11 @@ Otherwise, confirm before executing.
 6. Avoid redundant confirmations if you already confirmed and there is no material new risk.
 7. Write/shell operations inside the workspace do not need the same scrutiny as operations outside it.
 8. Interactive programs (editors, pagers, REPLs, `ssh`/`psql`/`mysql` without a command, `top`/`htop`, `git rebase -i`) will be blocked because they hang the agent; always use the non-interactive alternative.
+
+## Sandbox Permission Requests
+
+- Prefer sandboxed additional permissions over full escalation: use `sandbox_permissions: "with_additional_permissions"` with the required file-system or network permission when that can satisfy the command.
+- Use `sandbox_permissions: "require_escalated"` only when the command must run outside the sandbox, such as inspecting or controlling host/global state that the sandbox intentionally hides.
+- If a sandboxed shell command returns only sandbox-local state when the user asked for host/global state, do not answer from that misleading output. Rerun the needed command with `sandbox_permissions: "require_escalated"`, a short user-facing `justification`, and a narrow `prefix_rule` when appropriate.
+- For `require_escalated`, include a short user-facing `justification`. When a reusable approval is appropriate, include a narrow `prefix_rule`; do not suggest broad shell, interpreter, or launcher prefixes.
+- Shell commands are evaluated as independent command segments across `|`, `&&`, `||`, and `;`. A reusable prefix must apply safely to the command segment it is meant to authorize.

internal/agent/decode_args_test.go

@@ -0,0 +1,99 @@
+package agent
+
+import "testing"
+
+// decodeToolArguments must (a) decode a normal single-object payload unchanged,
+// (b) tolerate a weak model packing multiple concatenated top-level JSON objects
+// into one arguments string by decoding the FIRST object and ignoring the rest
+// (the minimax-m3 "invalid character '{' after top-level value" failure), and
+// (c) still reject a genuinely malformed/truncated first object.
+func TestDecodeToolArguments(t *testing.T) {
+	t.Run("single object decodes normally", func(t *testing.T) {
+		var a map[string]any
+		if err := decodeToolArguments(`{"path":"x"}`, &a); err != nil {
+			t.Fatalf("err = %v, want nil", err)
+		}
+		if a["path"] != "x" {
+			t.Fatalf("decoded %v, want path=x", a)
+		}
+	})
+
+	t.Run("concatenated objects decode the first, ignore trailing", func(t *testing.T) {
+		var b map[string]any
+		if err := decodeToolArguments(`{"path":"first"}{"path":"second"}`, &b); err != nil {
+			t.Fatalf("multi-object should decode the first, got err = %v", err)
+		}
+		if b["path"] != "first" {
+			t.Fatalf("decoded %v, want the first object (path=first)", b)
+		}
+	})
+
+	t.Run("distinct tools' args concatenated (real minimax case): first object used", func(t *testing.T) {
+		var c map[string]any
+		if err := decodeToolArguments(`{"plan":[1,2]}{"cmd":"go version"}`, &c); err != nil {
+			t.Fatalf("err = %v, want nil", err)
+		}
+		if _, ok := c["plan"]; !ok || c["cmd"] != nil {
+			t.Fatalf("decoded %v, want only the first object's keys", c)
+		}
+	})
+
+	t.Run("truncated first object still errors", func(t *testing.T) {
+		var d map[string]any
+		if err := decodeToolArguments(`{"path":`, &d); err == nil {
+			t.Fatal("truncated JSON must still error")
+		}
+	})
+
+	t.Run("non-JSON trailing garbage still errors (not masked)", func(t *testing.T) {
+		var g map[string]any
+		if err := decodeToolArguments(`{"x":1}xyz`, &g); err == nil {
+			t.Fatal("a valid first object followed by non-JSON garbage must still error")
+		}
+	})
+
+	t.Run("recoverableToolArguments: only whole-JSON trailing is tolerated", func(t *testing.T) {
+		cases := []struct {
+			in    string
+			ok    bool
+			first string
+		}{
+			{`{"a":1}`, true, `{"a":1}`},
+			{`{"a":1}{"b":2}`, true, `{"a":1}`},
+			{`{"a":1}  {"b":2}  `, true, `{"a":1}`},
+			{`{"a":1}xyz`, false, ""},
+			{`{"a":`, false, ""},
+			{`   `, false, ""},
+		}
+		for _, c := range cases {
+			first, ok := recoverableToolArguments(c.in)
+			if ok != c.ok || (ok && first != c.first) {
+				t.Fatalf("recoverableToolArguments(%q) = (%q,%v), want (%q,%v)", c.in, first, ok, c.first, c.ok)
+			}
+		}
+	})
+
+	t.Run("historySafeToolCalls recovers the first object, not {}", func(t *testing.T) {
+		out := historySafeToolCalls([]ToolCall{
+			{ID: "c1", Name: "read_file", Arguments: `{"path":"README.md"}{"path":"AGENTS.md"}`}, // concatenated -> first
+			{ID: "c2", Name: "read_file", Arguments: `{"path":"x"}xyz`},                          // garbage -> {}
+			{ID: "c3", Name: "read_file", Arguments: `{"path":"ok"}`},                            // valid -> unchanged
+		})
+		if out[0].Arguments != `{"path":"README.md"}` {
+			t.Fatalf("concatenated args should record the first object, got %q", out[0].Arguments)
+		}
+		if out[1].Arguments != "{}" {
+			t.Fatalf("garbage-trailing args should record {}, got %q", out[1].Arguments)
+		}
+		if out[2].Arguments != `{"path":"ok"}` {
+			t.Fatalf("valid args should be unchanged, got %q", out[2].Arguments)
+		}
+	})
+
+	t.Run("whitespace-only errors (caller guards empty)", func(t *testing.T) {
+		var e map[string]any
+		if err := decodeToolArguments("   ", &e); err == nil {
+			t.Fatal("whitespace-only must error")
+		}
+	})
+}

internal/agent/deferred_loop_test.go

@@ -129,6 +129,24 @@ func (fakeToolSearchTool) Run(_ context.Context, _ map[string]any) tools.Result
 	return tools.Result{Status: tools.StatusOK, Output: "ok"}
 }
 
+func toolDefinitionByName(definitions []zeroruntime.ToolDefinition, name string) *zeroruntime.ToolDefinition {
+	for i := range definitions {
+		if definitions[i].Name == name {
+			return &definitions[i]
+		}
+	}
+	return nil
+}
+
+func assertNoDeferredDiscoveryMessage(t *testing.T, request zeroruntime.CompletionRequest) {
+	t.Helper()
+	for _, message := range request.Messages {
+		if message.Role == zeroruntime.MessageRoleUser && strings.Contains(message.Content, "Deferred tools:") {
+			t.Fatalf("deferred discovery must not be appended as a user message: %q", message.Content)
+		}
+	}
+}
+
 func TestPartitionToolsInactiveIsByteIdenticalAndDropsToolSearch(t *testing.T) {
 	root := t.TempDir()
 	registry := tools.NewRegistry()
@@ -219,8 +237,8 @@ func TestPartitionToolsBelowThresholdInactive(t *testing.T) {
 // FIX 6(a): a deferred tool removed by DisabledTools must NOT count toward the
 // eligible total. With N deferred registered and threshold == N, disabling one
 // drops the surviving eligible count to N-1 < threshold, so the partition takes
-// the INACTIVE path (empty reminder, all VISIBLE tools exposed with full schemas,
-// the disabled one filtered out entirely).
+// the INACTIVE path (empty discovery text, all VISIBLE tools exposed with full
+// schemas, the disabled one filtered out entirely).
 func TestPartitionToolsDisabledDeferredDropsBelowThresholdInactive(t *testing.T) {
 	registry := tools.NewRegistry()
 	registry.Register(fakeDeferredTool{name: "mcp__srv__alpha", desc: "alpha"})
@@ -246,9 +264,9 @@ func TestPartitionToolsDisabledDeferredDropsBelowThresholdInactive(t *testing.T)
 }
 
 // FIX 6(b): with deferral ACTIVE, a DisabledTools-hidden deferred tool must never
-// appear in the reminder NOR be exposed — it is filtered out before the partition
+// influence discovery NOR be exposed — it is filtered out before the partition
 // even considers it.
-func TestPartitionToolsActiveExcludesDisabledDeferredFromReminderAndExposed(t *testing.T) {
+func TestPartitionToolsActiveExcludesDisabledDeferredFromDiscoveryAndExposed(t *testing.T) {
 	registry := tools.NewRegistry()
 	registry.Register(fakeDeferredTool{name: "mcp__srv__alpha", desc: "alpha"})
 	registry.Register(fakeDeferredTool{name: "mcp__srv__beta", desc: "beta"})
@@ -256,25 +274,31 @@ func TestPartitionToolsActiveExcludesDisabledDeferredFromReminderAndExposed(t *t
 	registry.Register(fakeToolSearchTool{}) // usable loader => deferral can activate
 
 	// 3 deferred, disable beta => 2 surviving eligible, threshold 2 => active.
-	exposed, reminder := partitionTools(registry, PermissionModeAuto, Options{
+	exposed, discovery := partitionTools(registry, PermissionModeAuto, Options{
 		DeferThreshold: 2,
 		DisabledTools:  []string{"mcp__srv__beta"},
 	}, map[string]bool{})
 
-	if reminder == "" {
-		t.Fatalf("expected active path with a non-empty reminder for the unloaded tools")
+	if discovery == "" {
+		t.Fatalf("expected active path with discovery text for the unloaded tools")
 	}
-	if strings.Contains(reminder, "mcp__srv__beta") {
-		t.Fatalf("disabled deferred tool must not appear in the reminder, got %q", reminder)
+	if strings.Contains(discovery, "mcp__srv__beta") {
+		t.Fatalf("disabled deferred tool must not appear in discovery text, got %q", discovery)
 	}
 	for _, def := range exposed {
 		if def.Name == "mcp__srv__beta" {
 			t.Fatalf("disabled deferred tool must not be exposed, got %#v", exposed)
 		}
 	}
-	// The two surviving deferred tools are hidden (unloaded) and named in the reminder.
-	if !strings.Contains(reminder, "mcp__srv__alpha") || !strings.Contains(reminder, "mcp__srv__gamma") {
-		t.Fatalf("reminder must list the surviving unloaded deferred tools, got %q", reminder)
+	search := toolDefinitionByName(exposed, tools.ToolSearchToolName)
+	if search == nil {
+		t.Fatalf("tool_search must be exposed on active path, got %#v", exposed)
+	}
+	if search.Description != discovery {
+		t.Fatalf("tool_search description must carry discovery text\n got: %q\nwant: %q", search.Description, discovery)
+	}
+	if !strings.Contains(discovery, "mcp") {
+		t.Fatalf("discovery text must list the surviving deferred source, got %q", discovery)
 	}
 }
 
@@ -289,7 +313,7 @@ func TestPartitionToolsActiveHidesUnloadedExposesLoaded(t *testing.T) {
 	loaded := map[string]bool{"mcp__srv__alpha": true}
 
 	// 2 eligible deferred tools, threshold 2 => active.
-	exposed, reminder := partitionTools(registry, PermissionModeAuto, Options{DeferThreshold: 2}, loaded)
+	exposed, discovery := partitionTools(registry, PermissionModeAuto, Options{DeferThreshold: 2}, loaded)
 
 	exposedNames := map[string]bool{}
 	for _, def := range exposed {
@@ -307,14 +331,21 @@ func TestPartitionToolsActiveHidesUnloadedExposesLoaded(t *testing.T) {
 	if exposedNames["mcp__srv__beta"] {
 		t.Fatalf("unloaded deferred tool must be hidden from exposed, got %#v", exposed)
 	}
-	if reminder == "" {
-		t.Fatalf("expected a non-empty reminder for the hidden tool")
+	if discovery == "" {
+		t.Fatalf("expected non-empty discovery text for the hidden tool")
 	}
-	if !strings.Contains(reminder, "mcp__srv__beta") {
-		t.Fatalf("expected reminder to list the hidden tool, got %q", reminder)
+	search := toolDefinitionByName(exposed, tools.ToolSearchToolName)
+	if search == nil {
+		t.Fatalf("expected tool_search exposed on active path, got %#v", exposed)
 	}
-	if strings.Contains(reminder, "mcp__srv__alpha") {
-		t.Fatalf("loaded tool must not appear in the reminder, got %q", reminder)
+	if search.Description != discovery || !strings.Contains(search.Description, "mcp") {
+		t.Fatalf("tool_search description must carry discovery source, got %q", search.Description)
+	}
+	if strings.Contains(search.Description, "mcp__srv__alpha") {
+		t.Fatalf("tool_search description must not list already-loaded tool names, got %q", search.Description)
+	}
+	if !strings.Contains(search.Description, "mcp__srv__beta") {
+		t.Fatalf("tool_search description must list exact hidden tool names, got %q", search.Description)
 	}
 }
 
@@ -337,9 +368,9 @@ func TestPartitionToolsActiveNothingHiddenEmptyReminder(t *testing.T) {
 	if !exposedNames["tool_search"] {
 		t.Fatalf("expected tool_search exposed on active path, got %#v", exposed)
 	}
-	// BuildDeferredReminder returns "" for no hidden lines.
+	// No hidden tools means no dynamic discovery text is needed.
 	if reminder != "" {
-		t.Fatalf("expected empty reminder when nothing is hidden, got %q", reminder)
+		t.Fatalf("expected empty discovery text when nothing is hidden, got %q", reminder)
 	}
 }
 
@@ -376,7 +407,8 @@ func TestRunLoadsDeferredToolThenAdvertisesNextTurn(t *testing.T) {
 		t.Fatalf("expected two turns, got %d", len(provider.requests))
 	}
 
-	// Turn 1: alpha is hidden (not advertised) and the reminder is a trailing user msg.
+	// Turn 1: alpha is hidden (not advertised), and discovery lives on the
+	// tool_search definition rather than a trailing user message.
 	turn1Tools := map[string]bool{}
 	for _, def := range provider.requests[0].Tools {
 		turn1Tools[def.Name] = true
@@ -384,9 +416,13 @@ func TestRunLoadsDeferredToolThenAdvertisesNextTurn(t *testing.T) {
 	if turn1Tools["mcp__srv__alpha"] {
 		t.Fatalf("turn 1 must not advertise the not-yet-loaded deferred tool")
 	}
-	last1 := provider.requests[0].Messages[len(provider.requests[0].Messages)-1]
-	if last1.Role != zeroruntime.MessageRoleUser || !strings.Contains(last1.Content, "tool_search") {
-		t.Fatalf("turn 1 request must end with the deferred-tools reminder, got role=%s content=%q", last1.Role, last1.Content)
+	assertNoDeferredDiscoveryMessage(t, provider.requests[0])
+	search1 := toolDefinitionByName(provider.requests[0].Tools, tools.ToolSearchToolName)
+	if search1 == nil {
+		t.Fatalf("turn 1 must expose tool_search, got %#v", provider.requests[0].Tools)
+	}
+	if !strings.Contains(search1.Description, "Tool discovery") || !strings.Contains(search1.Description, "mcp") {
+		t.Fatalf("tool_search description must carry deferred discovery, got %q", search1.Description)
 	}
 
 	// Turn 2: alpha is now loaded and advertised with a full schema.
@@ -401,20 +437,20 @@ func TestRunLoadsDeferredToolThenAdvertisesNextTurn(t *testing.T) {
 		t.Fatalf("beta was never loaded; it must stay hidden in turn 2")
 	}
 
-	// The reminder must NOT persist into the returned message history.
+	// The discovery text must NOT persist into the returned message history.
 	for _, m := range result.Messages {
-		if m.Role == zeroruntime.MessageRoleUser && strings.Contains(m.Content, "Call tool_search") {
-			t.Fatalf("the deferred-tools reminder must not be persisted in result.Messages")
+		if m.Role == zeroruntime.MessageRoleUser && strings.Contains(m.Content, "Tool discovery") {
+			t.Fatalf("deferred discovery must not be persisted in result.Messages")
 		}
 	}
 }
 
-// TestRunReactiveRetryKeepsLoadedDeferredToolAndReminder drives a mid-run
+// TestRunReactiveRetryKeepsLoadedDeferredToolAndDiscovery drives a mid-run
 // context-limit error that triggers reactive compaction+retry while deferral is
 // ACTIVE and a deferred tool is already loaded. The retried turn must NOT be
-// degraded to the empty-loaded/no-reminder state: it must still advertise the
-// loaded tool's FULL schema and carry the <system-reminder> for the hidden tool.
-func TestRunReactiveRetryKeepsLoadedDeferredToolAndReminder(t *testing.T) {
+// degraded to the empty-loaded/no-discovery state: it must still advertise the
+// loaded tool's FULL schema and keep tool_search discovery for hidden tools.
+func TestRunReactiveRetryKeepsLoadedDeferredToolAndDiscovery(t *testing.T) {
 	registry := tools.NewRegistry()
 	// load_signal asks the loop to load mcp__srv__alpha next turn.
 	registry.Register(loadSignalTool{value: "mcp__srv__alpha"})
@@ -492,24 +528,20 @@ func TestRunReactiveRetryKeepsLoadedDeferredToolAndReminder(t *testing.T) {
 		t.Fatalf("retry must advertise the loaded tool's FULL schema, got %#v", alpha.Parameters)
 	}
 
-	// The retry must carry the deferred-tools reminder as its trailing user
-	// message (the hidden tool is mcp__srv__beta) — not the no-reminder state.
-	last := retry.Messages[len(retry.Messages)-1]
-	if last.Role != zeroruntime.MessageRoleUser || !strings.Contains(last.Content, "tool_search") {
-		t.Fatalf("retry request must end with the deferred-tools reminder, got role=%s content=%q", last.Role, last.Content)
+	assertNoDeferredDiscoveryMessage(t, retry)
+	search := toolDefinitionByName(retry.Tools, tools.ToolSearchToolName)
+	if search == nil {
+		t.Fatalf("retry must expose tool_search for still-hidden tools, got %#v", retry.Tools)
 	}
-	if !strings.Contains(last.Content, "mcp__srv__beta") {
-		t.Fatalf("retry reminder must list the still-hidden tool mcp__srv__beta, got %q", last.Content)
-	}
-	if strings.Contains(last.Content, "mcp__srv__alpha") {
-		t.Fatalf("loaded tool must not appear in the retry reminder, got %q", last.Content)
+	if !strings.Contains(search.Description, "Tool discovery") || !strings.Contains(search.Description, "mcp") {
+		t.Fatalf("retry tool_search description must carry deferred discovery, got %q", search.Description)
 	}
 
-	// The reminder must NOT persist into the returned message history, even on
+	// The discovery text must NOT persist into the returned message history, even on
 	// the reactive-retry path.
 	for _, m := range result.Messages {
-		if m.Role == zeroruntime.MessageRoleUser && strings.Contains(m.Content, "Call tool_search") {
-			t.Fatalf("the deferred-tools reminder must not be persisted in result.Messages")
+		if m.Role == zeroruntime.MessageRoleUser && strings.Contains(m.Content, "Tool discovery") {
+			t.Fatalf("deferred discovery must not be persisted in result.Messages")
 		}
 	}
 }
@@ -544,15 +576,15 @@ func (provider *connectErrorProvider) StreamCompletion(_ context.Context, reques
 	return ch, nil
 }
 
-// TestRunConnectTimeReactiveRetryKeepsLoadedDeferredToolAndReminder mirrors
-// TestRunReactiveRetryKeepsLoadedDeferredToolAndReminder but drives the
+// TestRunConnectTimeReactiveRetryKeepsLoadedDeferredToolAndDiscovery mirrors
+// TestRunReactiveRetryKeepsLoadedDeferredToolAndDiscovery but drives the
 // CONNECT-TIME error path: StreamCompletion itself returns a non-nil
 // context-limit error (rather than a mid-stream StreamEventError). With deferral
 // ACTIVE and a deferred tool already loaded, the rebuilt retry request must still
-// advertise the loaded tool's FULL schema AND carry the deferred-tools reminder —
-// i.e. the first reactive block reuses exposed/reminder, not the empty-loaded
+// advertise the loaded tool's FULL schema AND carry tool_search discovery —
+// i.e. the first reactive block reuses exposed, not the empty-loaded
 // partition.
-func TestRunConnectTimeReactiveRetryKeepsLoadedDeferredToolAndReminder(t *testing.T) {
+func TestRunConnectTimeReactiveRetryKeepsLoadedDeferredToolAndDiscovery(t *testing.T) {
 	registry := tools.NewRegistry()
 	registry.Register(loadSignalTool{value: "mcp__srv__alpha"})
 	registry.Register(fakeDeferredTool{name: "mcp__srv__alpha", desc: "alpha tool"})
@@ -625,22 +657,19 @@ func TestRunConnectTimeReactiveRetryKeepsLoadedDeferredToolAndReminder(t *testin
 		t.Fatalf("retry must advertise the loaded tool's FULL schema, got %#v", alpha.Parameters)
 	}
 
-	// The retry must carry the deferred-tools reminder as its trailing user message.
-	last := retry.Messages[len(retry.Messages)-1]
-	if last.Role != zeroruntime.MessageRoleUser || !strings.Contains(last.Content, "tool_search") {
-		t.Fatalf("retry request must end with the deferred-tools reminder, got role=%s content=%q", last.Role, last.Content)
+	assertNoDeferredDiscoveryMessage(t, retry)
+	search := toolDefinitionByName(retry.Tools, tools.ToolSearchToolName)
+	if search == nil {
+		t.Fatalf("retry must expose tool_search for still-hidden tools, got %#v", retry.Tools)
 	}
-	if !strings.Contains(last.Content, "mcp__srv__beta") {
-		t.Fatalf("retry reminder must list the still-hidden tool mcp__srv__beta, got %q", last.Content)
-	}
-	if strings.Contains(last.Content, "mcp__srv__alpha") {
-		t.Fatalf("loaded tool must not appear in the retry reminder, got %q", last.Content)
+	if !strings.Contains(search.Description, "Tool discovery") || !strings.Contains(search.Description, "mcp") {
+		t.Fatalf("retry tool_search description must carry deferred discovery, got %q", search.Description)
 	}
 
-	// The reminder must NOT persist into the returned message history.
+	// The discovery text must NOT persist into the returned message history.
 	for _, m := range result.Messages {
-		if m.Role == zeroruntime.MessageRoleUser && strings.Contains(m.Content, "Call tool_search") {
-			t.Fatalf("the deferred-tools reminder must not be persisted in result.Messages")
+		if m.Role == zeroruntime.MessageRoleUser && strings.Contains(m.Content, "Tool discovery") {
+			t.Fatalf("deferred discovery must not be persisted in result.Messages")
 		}
 	}
 }
@@ -649,8 +678,7 @@ func TestRunConnectTimeReactiveRetryKeepsLoadedDeferredToolAndReminder(t *testin
 // regression: N deferred tools are registered alongside tool_search, the operator
 // allowlists ONLY the N deferred names (NOT tool_search), and DeferThreshold == N.
 // Deferral must ACTIVATE, the partition must STILL expose tool_search (the gateway
-// to the allowlisted tools), and a tool_search call must be DISPATCH-ALLOWED — so
-// the reminder never points the model at an unreachable tool.
+// to the allowlisted tools), and a tool_search call must be DISPATCH-ALLOWED.
 func TestAllowlistedDeferredToolsKeepToolSearchReachable(t *testing.T) {
 	registry := tools.NewRegistry()
 	deferredNames := []string{"mcp__srv__alpha", "mcp__srv__beta"}
@@ -667,14 +695,14 @@ func TestAllowlistedDeferredToolsKeepToolSearchReachable(t *testing.T) {
 		EnabledTools: append([]string{}, deferredNames...),
 	}
 
-	exposed, reminder := partitionTools(registry, PermissionModeAuto, options, map[string]bool{})
+	exposed, discovery := partitionTools(registry, PermissionModeAuto, options, map[string]bool{})
 
-	// Deferral active => non-empty reminder advertising the hidden deferred tools.
-	if reminder == "" {
-		t.Fatalf("expected deferral active (non-empty reminder) at threshold N, got empty")
+	// Deferral active => non-empty tool_search discovery for hidden deferred tools.
+	if discovery == "" {
+		t.Fatalf("expected deferral active (non-empty discovery) at threshold N, got empty")
 	}
-	if !strings.Contains(reminder, "tool_search") {
-		t.Fatalf("reminder must instruct the model to call tool_search, got %q", reminder)
+	if !strings.Contains(discovery, "tool_search") {
+		t.Fatalf("discovery must instruct the model to call tool_search, got %q", discovery)
 	}
 
 	// FIX 2(a): tool_search is exposed even though the allowlist omits it.
@@ -685,6 +713,10 @@ func TestAllowlistedDeferredToolsKeepToolSearchReachable(t *testing.T) {
 	if !exposedNames["tool_search"] {
 		t.Fatalf("tool_search must be exposed on the active path despite the allowlist omitting it, got %#v", exposed)
 	}
+	search := toolDefinitionByName(exposed, tools.ToolSearchToolName)
+	if search == nil || search.Description != discovery {
+		t.Fatalf("tool_search description must carry discovery text, got %#v discovery=%q", search, discovery)
+	}
 	// The allowlisted-but-unloaded deferred tools are hidden (not exposed).
 	for _, name := range deferredNames {
 		if exposedNames[name] {

internal/agent/file_diagnostics.go

@@ -0,0 +1,62 @@
+package agent
+
+import (
+	"context"
+	"os"
+	"path/filepath"
+	"strings"
+	"time"
+
+	"github.com/Gitlawb/zero/internal/lsp"
+)
+
+// fileDiagnosticsTimeout bounds one inline post-edit diagnostics check so a
+// slow or wedged language server can never hang a tool call; on timeout the
+// edit simply reports without a diagnostics block.
+const fileDiagnosticsTimeout = 10 * time.Second
+
+// NewFileDiagnostics adapts an *lsp.Manager to the per-edit inline diagnostics
+// callback (tools.RunOptions.Diagnostics): it reads the just-written file,
+// checks it against the file's language server, and formats error-severity
+// diagnostics for the model. Warnings and hints are excluded — nagging about
+// style on every edit is noise, while a type error the edit just introduced is
+// exactly what the model should see before its next step. Diagnostics are
+// rendered with workspace-relative paths: the absolute path would put the
+// local username/home directory into the model prompt and session transcript
+// on every edit. Returns nil when manager is nil, disabling inline diagnostics
+// entirely.
+func NewFileDiagnostics(manager *lsp.Manager, workspaceRoot string) func(context.Context, string) string {
+	if manager == nil {
+		return nil
+	}
+	return func(ctx context.Context, absPath string) string {
+		text, err := os.ReadFile(absPath)
+		if err != nil {
+			return ""
+		}
+		checkCtx, cancel := context.WithTimeout(ctx, fileDiagnosticsTimeout)
+		defer cancel()
+		diagnostics, err := manager.Check(checkCtx, absPath, string(text))
+		if err != nil {
+			return ""
+		}
+		errors := lsp.FilterBySeverity(diagnostics, lsp.SeverityError)
+		if len(errors) == 0 {
+			return ""
+		}
+		return lsp.FormatDiagnostics(diagnosticsDisplayPath(workspaceRoot, absPath), errors)
+	}
+}
+
+// diagnosticsDisplayPath renders absPath relative to the workspace root for
+// model-facing output, falling back to the file's base name when the path is
+// outside the workspace (a bare name still identifies the file without
+// exposing the directory layout).
+func diagnosticsDisplayPath(workspaceRoot, absPath string) string {
+	if workspaceRoot != "" {
+		if rel, err := filepath.Rel(workspaceRoot, absPath); err == nil && !strings.HasPrefix(rel, "..") {
+			return rel
+		}
+	}
+	return filepath.Base(absPath)
+}

internal/agent/file_diagnostics_test.go

@@ -0,0 +1,24 @@
+package agent
+
+import (
+	"path/filepath"
+	"testing"
+)
+
+// Diagnostics are model-facing: absolute paths would leak the local username
+// and directory layout into the prompt and session transcript on every edit.
+func TestDiagnosticsDisplayPath(t *testing.T) {
+	root := filepath.Join("/Users", "someone", "project")
+	cases := []struct {
+		root, abs, want string
+	}{
+		{root, filepath.Join(root, "internal", "a.go"), filepath.Join("internal", "a.go")},
+		{root, filepath.Join("/etc", "other.go"), "other.go"}, // outside root -> base name only
+		{"", filepath.Join("/home", "user", "x.go"), "x.go"},  // no root -> base name only
+	}
+	for _, c := range cases {
+		if got := diagnosticsDisplayPath(c.root, c.abs); got != c.want {
+			t.Errorf("diagnosticsDisplayPath(%q, %q) = %q, want %q", c.root, c.abs, got, c.want)
+		}
+	}
+}

internal/agent/guardrails.go

@@ -1,6 +1,7 @@
 package agent
 
 import (
+	"encoding/json"
 	"strconv"
 	"strings"
 
@@ -21,6 +22,12 @@ const (
 	// calls have executed since the last update_plan call.
 	staleToolCallThreshold = 10
 
+	// stalePlanTurnThreshold injects the same one-shot reminder once this many
+	// turns have passed since the last update_plan while plan items are still
+	// pending — the turn-based complement to staleToolCallThreshold, catching a
+	// plan that drifts stale across many turns that each make few tool calls.
+	stalePlanTurnThreshold = 8
+
 	// toolOnlyProgressReminderAt injects a one-shot progress nudge after this
 	// many consecutive turns contain tool calls but no visible assistant text.
 	// It does not stop the run; it tells the model to synthesize what it already
@@ -42,10 +49,275 @@ const (
 	toolFailureHintAt = 2
 	// toolFailureStopAt halts the run after a tool fails this many times in a row
 	// with the same error, so NO model (weak or strong) burns turns looping on a
-	// bad call.
-	toolFailureStopAt = 4
+	// bad call. Set to 6 (not 4): a corrective hint fires at toolFailureHintAt (2),
+	// and a model iterating on a genuinely tricky edit can legitimately fail a few
+	// times after the hint while converging — stopping at 4 cut those runs short.
+	// The streak still resets the moment the tool succeeds or hits a different
+	// error, so this only affects true same-error loops.
+	toolFailureStopAt = 6
+
+	// maxContinueNudges bounds how many times the headless completion gate
+	// (Options.RequireCompletionSignal) re-prompts a model that stopped without a
+	// tool call while work clearly remained. Once spent, the run finalizes as
+	// INCOMPLETE rather than nudging forever (and it is still bounded by maxTurns
+	// and the run deadline).
+	maxContinueNudges = 3
 )
 
+// continueNudgeMarker is a stable substring for tests.
+const continueNudgeMarker = "the task is not finished"
+
+// continueNudge tells a model that stopped without a tool call — while work
+// clearly remained — to keep going, or to mark the plan complete and summarize if
+// it is genuinely done. The second path gives it a clean route to a legitimate
+// completion (a finished plan + no continuation cue then exits as success).
+func continueNudge(reason string) string {
+	return "You stopped without calling a tool, but " + continueNudgeMarker + " (" + reason + "). " +
+		"Do not stop here: take the next concrete action with a tool now. " +
+		"If you are genuinely finished, first mark the plan complete with update_plan, then give your final summary."
+}
+
+// endsWithContinuationCue reports whether an assistant message ends mid-thought —
+// the model announcing its OWN next action and then stopping on a colon
+// ("…Let me check the SSH configuration:") rather than concluding. It requires
+// BOTH a trailing colon AND an action lead-in on the last line, so genuine closers
+// are NOT flagged: a forward-looking recommendation ("Next, I suggest reviewing
+// the changes."), a sign-off ("Let me know if you need anything"), or a summary
+// that merely ends in a colon ("Here is the summary:"). A bare trailing colon alone
+// is too common in legitimate final answers to use as a signal.
+func endsWithContinuationCue(text string) bool {
+	trimmed := strings.TrimSpace(text)
+	if trimmed == "" {
+		return false
+	}
+	lines := strings.Split(trimmed, "\n")
+	last := ""
+	for i := len(lines) - 1; i >= 0; i-- {
+		if s := strings.TrimSpace(lines[i]); s != "" {
+			last = strings.ToLower(s)
+			break
+		}
+	}
+	if !strings.HasSuffix(last, ":") {
+		return false
+	}
+	// Inspect the final clause (after the last sentence break) so a mid-line action
+	// announcement is caught ("…configure the server. Let me check the config:")
+	// while a plain summary colon ("Here is the summary:") or a recommendation is not.
+	clause := last
+	if idx := strings.LastIndex(last, ". "); idx >= 0 {
+		clause = strings.TrimSpace(last[idx+2:])
+	}
+	if strings.HasPrefix(clause, "let me know") { // sign-off, not a mid-step action
+		return false
+	}
+	for _, cue := range []string{
+		"let me ", "let's ", "now i'll ", "now i will ", "now let me ", "let me now ",
+		"i'll now ", "i will now ", "next i'll ", "next, i'll ", "first, i'll ", "first let me ",
+	} {
+		if strings.HasPrefix(clause, cue) {
+			return true
+		}
+	}
+	return false
+}
+
+// planStatusRemaining reports whether a raw update_plan status string represents
+// unfinished work. Anything not clearly completed/failed (incl. empty/unknown,
+// which the update_plan tool coerces to "pending") counts as remaining, matching
+// that tool's own status normalization.
+func planStatusRemaining(status string) bool {
+	switch strings.ToLower(strings.TrimSpace(status)) {
+	case "completed", "complete", "done", "finished", "resolved", "✓", "x", "[x]",
+		"failed", "fail", "error", "errored", "blocked", "cancelled", "canceled", "abandoned", "skipped":
+		return false
+	default:
+		return true
+	}
+}
+
+// acceptanceNudgeMarker is a stable substring for tests; it is embedded verbatim
+// in acceptanceVerificationNudge.
+const acceptanceNudgeMarker = "verify your work against the task's stated success criterion"
+
+// selfReportPhrases are high-signal admissions of guessing/fabrication or stated
+// uncertainty about the produced result. They are matched by plain substring with
+// NO context guard, so every entry must be FIRST-PERSON or unambiguously about the
+// model's own output — NOT a phrase that also describes implemented behavior.
+// (Earlier versions included "fall back to" / "placeholder value" / "as a
+// fallback" / bare "best guess" / "without proper", which match legitimate final
+// answers like "the parser will fall back to UTF-8" or "I replaced the placeholder
+// value", wrongly downgrading completed runs — so those are removed. Admissions of
+// INABILITY are handled separately by inabilityStems, which carry a guard.)
+var selfReportPhrases = []string{
+	// first-person admissions of guessing / fabrication / assumption
+	"i guessed", "my best guess", "i'm guessing", "this is a guess", "just a guess",
+	"i made it up", "i fabricated", "i assumed a", "i had to assume",
+	// first-person lack of capability (inability stems also cover cannot/could-not)
+	"i do not have the ability", "i don't have the ability", "i lack the ability",
+	"no way for me to", "not possible for me to",
+	// stated uncertainty about the correctness of the produced result
+	"may not be correct", "might not be correct", "may be incorrect", "might be incorrect",
+	"this may not work", "this might not work", "not fully functional", "not fully working",
+}
+
+// inabilityStems are first-person "I cannot / can't / could not / am unable to /
+// do not have" stems. Matching the STEM (not a fixed verb) generalizes over
+// whatever action the model claims it could not perform — "analyze", "determine",
+// "do", "complete", "verify", "see", etc. — so the detector is not defeated by
+// re-phrasing (the chess case slipped a fixed "cannot analyze" list by writing
+// "…which I cannot do without proper image analysis capabilities").
+var inabilityStems = []string{
+	"i cannot ", "i can't ", "i can not ", "i could not ", "i couldn't ",
+	"i am unable to", "i'm unable to", "i was unable to", "i wasn't able to",
+	"i was not able to", "i do not have", "i don't have", "unable to ",
+	"without being able to",
+}
+
+// successNegationTails are negated phrasings that indicate SUCCESS, not an
+// admission ("I could not find any remaining issues", "I cannot reproduce the
+// bug"). When an inability stem is immediately followed by one of these, it is not
+// treated as an admission, so a clean result is not misreported as INCOMPLETE.
+var successNegationTails = []string{
+	"find any", "found any", "find a ", "see any", "detect any", "identify any",
+	"reproduce", "spot any", "locate any",
+}
+
+// narrativeMarkers flag a sentence as RETELLING a past exchange rather than
+// reporting the outcome of the current objective, so an inability admission in
+// that sentence is about THEN, not NOW. Grounded in a real false positive: a
+// conversational recap ("You asked if I could work autonomously … I was honest
+// that my sandbox had no repo … so I couldn't actually do it at the time") was
+// downgraded to INCOMPLETE on the quoted "i couldn't " stem even though the
+// current turn's objective (summarize where we left off) was fully met. Every
+// marker is second-person or explicitly past-referential, so a first-person
+// admission about the current task ("I couldn't complete the refactor") is
+// never masked.
+var narrativeMarkers = []string{
+	"you asked", "you said", "you wanted", "you mentioned",
+	"we discussed", "we talked", "we covered",
+	"when we last", "last time", "last session", "previous session",
+	"previous conversation", "earlier session", "earlier conversation",
+	"at the time", "back then",
+}
+
+// stripQuoted removes spans enclosed in double quotes (straight or curly) or
+// backticks, so an admission the model merely QUOTES — its own earlier message,
+// a log line, an error string — cannot fire the detector. Only BALANCED spans
+// are removed: an opening delimiter that never closes is kept as literal text,
+// so a stray quote cannot swallow the rest of the message (and with it a
+// genuine admission the detector must see). Single quotes are left alone: they
+// are overwhelmingly apostrophes ("couldn't"), and treating them as quote
+// delimiters would swallow the text between two contractions.
+func stripQuoted(s string) string {
+	var b strings.Builder
+	var span strings.Builder // pending text since the open delimiter, kept if it never closes
+	open := rune(0)
+	for _, r := range s {
+		switch {
+		case open != 0:
+			if (open == '"' && r == '"') || (open == '“' && r == '”') || (open == '`' && r == '`') {
+				open = 0
+				span.Reset()
+				continue
+			}
+			span.WriteRune(r)
+		case r == '"' || r == '“' || r == '`':
+			open = r
+			span.WriteRune(r)
+		default:
+			b.WriteRune(r)
+		}
+	}
+	b.WriteString(span.String()) // dangling delimiter: restore the span verbatim
+	return b.String()
+}
+
+// admissionSentences splits lowered text into sentence-ish fragments so the
+// detector judges each claim in its own context. Newlines split too (markdown
+// bullets are separate claims); the exact boundaries only need to keep an
+// admission next to its own narrative/negation context, not be grammatical.
+func admissionSentences(lower string) []string {
+	return strings.FieldsFunc(lower, func(r rune) bool {
+		return r == '.' || r == '!' || r == '?' || r == '\n'
+	})
+}
+
+// selfReportedIncompletion returns a short reason when the model's final text
+// admits it guessed or could not meet the objective, else "". Case-insensitive.
+// Matching is per sentence, after dropping quoted spans, and a sentence that
+// retells a past exchange (narrativeMarkers) is skipped entirely — an admission
+// must be the model's own report about the CURRENT objective, not general
+// language that merely resembles one.
+func selfReportedIncompletion(text string) string {
+	for _, sentence := range admissionSentences(strings.ToLower(stripQuoted(text))) {
+		if containsAny(sentence, narrativeMarkers) {
+			continue
+		}
+		for _, phrase := range selfReportPhrases {
+			if strings.Contains(sentence, phrase) {
+				return selfReportReason(phrase)
+			}
+		}
+		for _, stem := range inabilityStems {
+			// Scan EVERY occurrence of the stem, not just the first: an earlier
+			// success-negation use ("I could not find any examples, so I could not
+			// implement it") must not mask a later genuine admission with the same stem.
+			for start := 0; ; {
+				rel := strings.Index(sentence[start:], stem)
+				if rel < 0 {
+					break
+				}
+				abs := start + rel
+				tail := strings.TrimSpace(sentence[abs+len(stem):])
+				if !hasAnyPrefix(tail, successNegationTails) {
+					return selfReportReason(strings.TrimSpace(stem) + " …")
+				}
+				start = abs + len(stem)
+			}
+		}
+	}
+	return ""
+}
+
+func containsAny(s string, subs []string) bool {
+	for _, sub := range subs {
+		if strings.Contains(s, sub) {
+			return true
+		}
+	}
+	return false
+}
+
+func selfReportReason(marker string) string {
+	return `the final message admits the objective was not met ("` + marker + `")`
+}
+
+func hasAnyPrefix(s string, prefixes []string) bool {
+	for _, p := range prefixes {
+		if strings.HasPrefix(s, p) {
+			return true
+		}
+	}
+	return false
+}
+
+// acceptanceVerificationNudge forces a TASK-GROUNDED acceptance check before a run
+// may finalize as success. It explicitly rejects the three false-success patterns
+// the bug-hunt surfaced: well-formed output treated as correct; pre-existing tests
+// passing treated as the objective being met; and a result that merely matches a
+// baseline the task asked to beat or improve. General — no task-specific content.
+func acceptanceVerificationNudge() string {
+	return "Before this task can be marked complete, " + acceptanceNudgeMarker + " — " +
+		"NOT the shape or format of your output, NOT that pre-existing tests pass, and NOT that your " +
+		"result merely matches a baseline you were asked to beat or improve. " +
+		"Re-read the original task, then run a concrete check that exercises the actual requirement: " +
+		"execute the program or tests that demonstrate the required behavior, or directly probe the specific " +
+		"thing the task asked you to produce, recover, fix, or optimize. " +
+		"If that check passes, reply PASS and cite the evidence. " +
+		"If it does not pass — or you cannot run such a check — say so plainly and keep working; do not claim success."
+}
+
 // toolFailureHintMarker is a stable substring for tests.
 const toolFailureHintMarker = "kept failing with the same error"
 
@@ -169,14 +441,23 @@ type guardState struct {
 	emptyTurns               int
 	totalToolCalls           int
 	toolCallsSincePlanUpdate int
-	planEverCalled           bool
-	notCalledReminderSent    bool
+	// turnsSincePlanUpdate counts turns (not individual tool calls) since the last
+	// update_plan, so a plan that goes stale across many low-tool-call turns is
+	// still caught — the tool-call counter alone can take many turns to trip when
+	// the model makes only one call per turn.
+	turnsSincePlanUpdate  int
+	planEverCalled        bool
+	notCalledReminderSent bool
 	// staleReminderSent records whether the stale reminder has already fired for
 	// the current stale interval. It is cleared when a plan update opens a new
 	// interval, making the reminder one-shot per interval rather than per turn.
 	staleReminderSent    bool
 	toolOnlyTurns        int
 	toolOnlyReminderSent bool
+	// planItemsPending is the number of remaining (pending/in_progress) items in
+	// the most recent update_plan call, so the headless completion gate can tell
+	// whether work is unfinished when the model stops without a tool call.
+	planItemsPending int
 	// toolFailures tracks consecutive same-error failures per tool, keyed by tool
 	// name, so the loop can hint then halt instead of looping forever.
 	toolFailures map[string]*toolFailureRecord
@@ -225,8 +506,9 @@ func (state *guardState) observeToolResult(name string, failed bool, output stri
 func (state *guardState) observeTurn(collected zeroruntime.CollectedStream) (stop bool) {
 	hasToolCalls := len(collected.ToolCalls) > 0
 	hasVisibleText := strings.TrimSpace(collected.Text) != ""
+	hasReasoning := collected.HasReasoning || len(collected.ReasoningBlocks) > 0
 
-	if hasToolCalls || hasVisibleText {
+	if hasToolCalls || hasVisibleText || hasReasoning {
 		state.emptyTurns = 0
 	} else {
 		state.emptyTurns++
@@ -238,13 +520,21 @@ func (state *guardState) observeTurn(collected zeroruntime.CollectedStream) (sto
 		state.toolOnlyReminderSent = false
 	}
 
+	// One turn has passed; the plan-update below resets this to 0 when the model
+	// refreshes the plan this turn.
+	state.turnsSincePlanUpdate++
+
 	for _, call := range collected.ToolCalls {
 		state.totalToolCalls++
 		if call.Name == planToolName {
 			state.planEverCalled = true
 			state.toolCallsSincePlanUpdate = 0
+			state.turnsSincePlanUpdate = 0
 			// A fresh plan update opens a new stale interval.
 			state.staleReminderSent = false
+			// Record how many items remain so the completion gate knows whether
+			// work is unfinished if the model later stops without a tool call.
+			state.observePlanUpdate(call.Arguments)
 		} else {
 			state.toolCallsSincePlanUpdate++
 		}
@@ -253,6 +543,33 @@ func (state *guardState) observeTurn(collected zeroruntime.CollectedStream) (sto
 	return state.emptyTurns >= maxEmptyTurns
 }
 
+// pendingPlanItems reports whether the most recent update_plan call still has
+// unfinished (pending/in_progress) items. False when no plan was ever recorded.
+func (state *guardState) pendingPlanItems() bool {
+	return state.planItemsPending > 0
+}
+
+// observePlanUpdate parses an update_plan call's raw arguments and records how
+// many items are still remaining. Malformed arguments leave the prior count
+// unchanged (best-effort — the plan panel itself tolerates the same).
+func (state *guardState) observePlanUpdate(arguments string) {
+	var parsed struct {
+		Plan []struct {
+			Status string `json:"status"`
+		} `json:"plan"`
+	}
+	if json.Unmarshal([]byte(arguments), &parsed) != nil {
+		return
+	}
+	pending := 0
+	for _, item := range parsed.Plan {
+		if planStatusRemaining(item.Status) {
+			pending++
+		}
+	}
+	state.planItemsPending = pending
+}
+
 func (state *guardState) progressReminder() string {
 	if state.toolOnlyReminderSent || state.toolOnlyTurns < toolOnlyProgressReminderAt {
 		return ""
@@ -266,10 +583,12 @@ func (state *guardState) progressReminder() string {
 // number of turns completed so far).
 func (state *guardState) planReminder(turn int) string {
 	// STALE reminder takes priority: a long run without a plan update is the
-	// stronger signal. One-shot per stale interval.
-	if state.planEverCalled &&
-		!state.staleReminderSent &&
-		state.toolCallsSincePlanUpdate >= staleToolCallThreshold {
+	// stronger signal. Fires on either the tool-call streak OR a turn streak with
+	// pending items (so a plan drifting stale across many low-call turns is caught,
+	// while a fully-completed plan is left alone). One-shot per stale interval.
+	if state.planEverCalled && !state.staleReminderSent &&
+		(state.toolCallsSincePlanUpdate >= staleToolCallThreshold ||
+			(state.turnsSincePlanUpdate >= stalePlanTurnThreshold && state.planItemsPending > 0)) {
 		state.staleReminderSent = true
 		return planStaleReminder(state.toolCallsSincePlanUpdate)
 	}

internal/agent/guardrails_test.go

@@ -22,6 +22,14 @@ func textTurn(content string) []zeroruntime.StreamEvent {
 	}
 }
 
+// reasoningTurn produces live reasoning without visible assistant text.
+func reasoningTurn(content string) []zeroruntime.StreamEvent {
+	return []zeroruntime.StreamEvent{
+		{Type: zeroruntime.StreamEventReasoning, Content: content},
+		{Type: zeroruntime.StreamEventDone},
+	}
+}
+
 // toolTurn produces a turn that calls a named tool with the given args JSON.
 func toolTurn(callID string, toolName string, args string) []zeroruntime.StreamEvent {
 	return []zeroruntime.StreamEvent{
@@ -101,6 +109,31 @@ func TestRunResetsEmptyTurnCounterOnVisibleOutput(t *testing.T) {
 	}
 }
 
+func TestRunResetsEmptyTurnCounterOnReasoning(t *testing.T) {
+	provider := &mockProvider{
+		turns: [][]zeroruntime.StreamEvent{
+			reasoningTurn("thinking 1"),
+			reasoningTurn("thinking 2"),
+			reasoningTurn("thinking 3"),
+			textTurn("done"),
+		},
+	}
+
+	result, err := Run(context.Background(), "go", provider, Options{
+		Registry: tools.NewRegistry(),
+		MaxTurns: 12,
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if result.FinalAnswer != "done" {
+		t.Fatalf("expected reasoning-only turns to keep the run live until final answer, got %q", result.FinalAnswer)
+	}
+	if len(provider.requests) != 4 {
+		t.Fatalf("expected 4 turns, got %d", len(provider.requests))
+	}
+}
+
 func TestRunResetsEmptyTurnCounterOnToolCall(t *testing.T) {
 	root := t.TempDir()
 	writeAgentTestFile(t, root+"/notes.txt", "alpha")

internal/agent/member_auto_test.go

@@ -0,0 +1,51 @@
+package agent
+
+import (
+	"context"
+	"testing"
+
+	"github.com/Gitlawb/zero/internal/tools"
+)
+
+type safetyTool struct {
+	name   string
+	safety tools.Safety
+}
+
+func (t safetyTool) Name() string        { return t.name }
+func (t safetyTool) Description() string { return "test tool" }
+func (t safetyTool) Parameters() tools.Schema {
+	return tools.Schema{Type: "object", AdditionalProperties: false}
+}
+func (t safetyTool) Safety() tools.Safety { return t.safety }
+func (t safetyTool) Run(context.Context, map[string]any) tools.Result {
+	return tools.Result{Status: tools.StatusOK}
+}
+
+// member-auto advertises the in-workspace mutators a headless member needs to
+// build (write/edit + shell) that plain Auto hides, but NOT network or denied
+// tools — the sandbox still gates the advertised ones at call time.
+func TestToolAdvertisedMemberAuto(t *testing.T) {
+	write := safetyTool{name: "write_file", safety: tools.Safety{SideEffect: tools.SideEffectWrite, Permission: tools.PermissionPrompt}}
+	shell := safetyTool{name: "bash", safety: tools.Safety{SideEffect: tools.SideEffectShell, Permission: tools.PermissionPrompt}}
+	read := safetyTool{name: "read_file", safety: tools.Safety{SideEffect: tools.SideEffectRead, Permission: tools.PermissionAllow}}
+	network := safetyTool{name: "net_tool", safety: tools.Safety{SideEffect: tools.SideEffectNetwork, Permission: tools.PermissionPrompt}}
+	denied := safetyTool{name: "blocked", safety: tools.Safety{SideEffect: tools.SideEffectRead, Permission: tools.PermissionDeny}}
+
+	// Plain Auto hides prompt-requiring mutators (the read-only member problem).
+	if ToolAdvertised(write, PermissionModeAuto) || ToolAdvertised(shell, PermissionModeAuto) {
+		t.Fatal("Auto must NOT advertise write/shell prompt tools")
+	}
+
+	for _, tool := range []tools.Tool{write, shell, read} {
+		if !ToolAdvertised(tool, PermissionModeMemberAuto) {
+			t.Fatalf("member-auto must advertise %q", tool.Name())
+		}
+	}
+	if ToolAdvertised(network, PermissionModeMemberAuto) {
+		t.Fatal("member-auto must NOT advertise a network prompt tool")
+	}
+	if ToolAdvertised(denied, PermissionModeMemberAuto) {
+		t.Fatal("member-auto must NOT advertise a denied tool")
+	}
+}

internal/agent/output_budget_propagation_test.go

@@ -0,0 +1,167 @@
+package agent
+
+import (
+	"context"
+	"os"
+	"strconv"
+	"strings"
+	"testing"
+
+	"github.com/Gitlawb/zero/internal/hooks"
+	"github.com/Gitlawb/zero/internal/tools"
+	"github.com/Gitlawb/zero/internal/trace"
+	"github.com/Gitlawb/zero/internal/zeroruntime"
+)
+
+type propagationOutputTool struct {
+	output string
+}
+
+func TestOutputBudgetHookHelperProcess(t *testing.T) {
+	for index, arg := range os.Args {
+		if arg != "--zero-output-budget-hook" || index+1 >= len(os.Args) {
+			continue
+		}
+		if _, err := os.Stdout.WriteString(os.Args[index+1]); err != nil {
+			os.Exit(2)
+		}
+		os.Exit(0)
+	}
+}
+
+func largeOutputBudgetHookDispatcher() *hooks.Dispatcher {
+	feedback := strings.Repeat("hook feedback ", 200)
+	return hooks.NewDispatcher(hooks.DispatcherOptions{Config: hooks.Config{
+		Enabled: true,
+		Hooks: []hooks.Definition{{
+			ID:      "large-feedback",
+			Event:   hooks.EventAfterTool,
+			Matcher: "propagation_output",
+			Command: os.Args[0],
+			Args: []string{
+				"-test.run=TestOutputBudgetHookHelperProcess",
+				"--",
+				"--zero-output-budget-hook",
+				feedback,
+			},
+			Enabled: true,
+		}},
+	}})
+}
+
+func (tool propagationOutputTool) Name() string        { return "propagation_output" }
+func (tool propagationOutputTool) Description() string { return "returns output for propagation tests" }
+func (tool propagationOutputTool) Parameters() tools.Schema {
+	return tools.Schema{Type: "object", AdditionalProperties: false}
+}
+func (tool propagationOutputTool) Safety() tools.Safety {
+	return tools.Safety{SideEffect: tools.SideEffectRead, Permission: tools.PermissionAllow, Reason: "test read"}
+}
+func (tool propagationOutputTool) Run(context.Context, map[string]any) tools.Result {
+	return tools.Result{Status: tools.StatusOK, Output: tool.output}
+}
+
+func TestExecuteToolCallPropagatesOutputTruncation(t *testing.T) {
+	t.Setenv("TMPDIR", t.TempDir())
+	t.Setenv("ZERO_TOOL_OUTPUT_CEILING_TOKENS", "80")
+	registry := tools.NewRegistry()
+	registry.Register(propagationOutputTool{output: strings.Repeat("large output\n", 1000)})
+
+	result, abortErr := executeToolCall(context.Background(), registry, ToolCall{
+		ID:        "call-budget",
+		Name:      "propagation_output",
+		Arguments: `{}`,
+	}, PermissionModeAuto, Options{Cwd: t.TempDir()})
+	if abortErr != nil {
+		t.Fatalf("executeToolCall abort error: %v", abortErr)
+	}
+	if !result.Truncated {
+		t.Fatalf("agent ToolResult lost tools.Result.Truncated: %#v", result)
+	}
+	if result.Meta["spill_path"] == "" {
+		t.Fatalf("agent ToolResult lost spill metadata: %#v", result.Meta)
+	}
+}
+
+func TestRecordOutputBudgetTraceUsesOnlyCompactMetadata(t *testing.T) {
+	recorder := trace.NewRecorder("session", "run", "")
+	recorder.Start()
+	recordOutputBudgetTrace(recorder, ToolResult{
+		Name:      "grep",
+		Truncated: true,
+		Output:    "SECRET OUTPUT MUST NOT ENTER TRACE",
+		Meta: map[string]string{
+			"output_budget_category":                  "search",
+			"output_budget_original_bytes":            "1000",
+			"output_budget_retained_bytes":            "100",
+			"output_budget_estimated_original_tokens": "250",
+			"output_budget_estimated_retained_tokens": "25",
+			"output_budget_reason":                    "semantic_search_budget",
+			"output_budget_spill_created":             "true",
+			"spill_path":                              "/secret/path/not-for-trace",
+		},
+	})
+	events := recorder.Finish().OutputBudgets
+	if len(events) != 1 {
+		t.Fatalf("events = %#v", events)
+	}
+	event := events[0]
+	if event.Tool != "grep" || event.Category != "search" || event.OriginalBytes != 1000 || event.RetainedBytes != 100 || !event.SpillCreated {
+		t.Fatalf("unexpected trace event: %#v", event)
+	}
+}
+
+func TestExecuteToolCallRebudgetsOversizedAfterToolFeedback(t *testing.T) {
+	t.Setenv("ZERO_TOOL_OUTPUT_CEILING_TOKENS", "80")
+	registry := tools.NewRegistry()
+	registry.Register(propagationOutputTool{output: "tool output"})
+	dispatcher := largeOutputBudgetHookDispatcher()
+
+	result, abortErr := executeToolCall(context.Background(), registry, ToolCall{
+		ID:        "call-hook-budget",
+		Name:      "propagation_output",
+		Arguments: `{}`,
+	}, PermissionModeAuto, Options{Hooks: dispatcher})
+	if abortErr != nil {
+		t.Fatalf("executeToolCall abort error: %v", abortErr)
+	}
+	if !result.Truncated || len(result.Output) > 80*4 {
+		t.Fatalf("afterTool feedback bypassed output budget: truncated=%t bytes=%d meta=%#v", result.Truncated, len(result.Output), result.Meta)
+	}
+	if result.Meta["output_budget_category"] == "" || result.Meta["output_budget_retained_bytes"] != strconv.Itoa(len(result.Output)) {
+		t.Fatalf("post-hook budget metadata does not describe final output: %#v", result.Meta)
+	}
+}
+
+func TestRunTraceReflectsPostHookBudget(t *testing.T) {
+	t.Setenv("ZERO_TOOL_OUTPUT_CEILING_TOKENS", "80")
+	registry := tools.NewRegistry()
+	registry.Register(propagationOutputTool{output: "tool output"})
+	dispatcher := largeOutputBudgetHookDispatcher()
+	provider := &mockProvider{turns: [][]zeroruntime.StreamEvent{
+		{
+			{Type: zeroruntime.StreamEventToolCallStart, ToolCallID: "call-hook-trace", ToolName: "propagation_output"},
+			{Type: zeroruntime.StreamEventToolCallDelta, ToolCallID: "call-hook-trace", ArgumentsFragment: `{}`},
+			{Type: zeroruntime.StreamEventToolCallEnd, ToolCallID: "call-hook-trace"},
+			{Type: zeroruntime.StreamEventDone},
+		},
+		{{Type: zeroruntime.StreamEventText, Content: "done"}, {Type: zeroruntime.StreamEventDone}},
+	}}
+	recorder := trace.NewRecorder("session", "run", "")
+	var toolResults []ToolResult
+	if _, err := Run(context.Background(), "budget hook", provider, Options{
+		Registry:     registry,
+		Hooks:        dispatcher,
+		Trace:        recorder,
+		OnToolResult: func(result ToolResult) { toolResults = append(toolResults, result) },
+	}); err != nil {
+		t.Fatalf("Run: %v", err)
+	}
+	if len(toolResults) != 1 || !toolResults[0].Truncated {
+		t.Fatalf("tool result = %#v, want one truncated post-hook result", toolResults)
+	}
+	events := recorder.Finish().OutputBudgets
+	if len(events) != 1 || !events[0].Truncated || events[0].RetainedBytes != len(toolResults[0].Output) {
+		t.Fatalf("trace does not describe final post-hook output: events=%#v result=%#v", events, toolResults[0])
+	}
+}

internal/agent/parallel_tools.go

@@ -0,0 +1,197 @@
+package agent
+
+import (
+	"context"
+	"sync"
+
+	"github.com/Gitlawb/zero/internal/tools"
+)
+
+// Parallel read-ahead for tool batches. When a turn requests several
+// independent lookups (read_file + grep + glob is the common shape), executing
+// them one after another serializes pure I/O waits. A consecutive run of
+// capability-safe read-only calls is executed concurrently instead; results are
+// then consumed in the original call order, so guard counters, message
+// ordering, abort semantics, and the surface's call/result event pairing are
+// byte-identical to sequential execution. Runs never span a mutating call: a
+// read that follows a write must observe the write, so eligibility is decided
+// per consecutive run, not per batch.
+//
+// Eligibility uses the PR5 tool-effect contract (tools.CapabilitiesOf):
+//
+//	Effect == ReadOnly
+//	AND ThreadSafe == true
+//	AND auto-allowed (no interactive permission prompt on the hot path)
+//	AND no resource-key conflict with an earlier call in the same concurrent window
+//
+// Unknown, mutators, interactive tools, and non-thread-safe reads stay
+// sequential. Empty resource keys do not conflict (ThreadSafe is the safety
+// gate); shared non-empty keys force a batch boundary.
+
+// maxParallelReadTools bounds concurrent read-only tool executions in a turn.
+const maxParallelReadTools = 8
+
+// precomputedToolResult is one parallel read-ahead execution, keyed back to
+// its batch index by the caller.
+type precomputedToolResult struct {
+	result   ToolResult
+	abortErr error
+}
+
+// parallelSafeToolCall reports whether call may run concurrently with its
+// neighbors under the PR5 capability contract. Loop-intercepted tools
+// (ask_user, request_permissions) and tool_search (mutates the deferred-tool
+// set) always stay sequential.
+func parallelSafeToolCall(registry *tools.Registry, call ToolCall, options Options) bool {
+	switch call.Name {
+	case "ask_user", tools.RequestPermissionsToolName, tools.ToolSearchToolName:
+		return false
+	}
+	tool, found := registry.Get(call.Name)
+	if !found {
+		return false
+	}
+	caps := tools.CapabilitiesOf(tool)
+	// Fail-closed: only audited concurrent-safe pure reads.
+	if caps.Effect != tools.EffectReadOnly || !caps.ThreadSafe {
+		return false
+	}
+	args, ok := decodeCallArgs(call)
+	if !ok {
+		return false
+	}
+	return effectivePermission(tool, args) == tools.PermissionAllow
+}
+
+// decodeCallArgs decodes tool call JSON arguments. Returns false on malformed
+// input so the call stays sequential (never panics into the parallel path).
+func decodeCallArgs(call ToolCall) (map[string]any, bool) {
+	args := map[string]any{}
+	if call.Arguments == "" {
+		return args, true
+	}
+	if err := decodeToolArguments(call.Arguments, &args); err != nil {
+		return nil, false
+	}
+	return args, true
+}
+
+// resourceKeysForCall returns the conflict keys for a call, or nil when the
+// tool has no ResourceKeys function / no keys for these args.
+func resourceKeysForCall(registry *tools.Registry, call ToolCall) []string {
+	tool, found := registry.Get(call.Name)
+	if !found {
+		return nil
+	}
+	caps := tools.CapabilitiesOf(tool)
+	if caps.ResourceKeys == nil {
+		return nil
+	}
+	args, ok := decodeCallArgs(call)
+	if !ok {
+		return nil
+	}
+	return caps.ResourceKeys(args)
+}
+
+// resourceKeysConflict reports whether two key sets share any non-empty key.
+// Empty key sets never conflict: ThreadSafe is the eligibility gate; keys only
+// refine conflict detection among tools that declare specific resources.
+func resourceKeysConflict(a, b []string) bool {
+	if len(a) == 0 || len(b) == 0 {
+		return false
+	}
+	seen := make(map[string]struct{}, len(a))
+	for _, key := range a {
+		if key == "" {
+			continue
+		}
+		seen[key] = struct{}{}
+	}
+	for _, key := range b {
+		if key == "" {
+			continue
+		}
+		if _, ok := seen[key]; ok {
+			return true
+		}
+	}
+	return false
+}
+
+// extendParallelRun returns the exclusive end index of a consecutive parallel-
+// safe run starting at start. The run stops before the first call that is not
+// parallel-safe or that conflicts on resource keys with any earlier call in
+// the same window (so two read_file calls on the same path stay sequential).
+func extendParallelRun(registry *tools.Registry, calls []ToolCall, start int, options Options) int {
+	if start >= len(calls) || !parallelSafeToolCall(registry, calls[start], options) {
+		return start
+	}
+	end := start + 1
+	keysWindow := [][]string{resourceKeysForCall(registry, calls[start])}
+	for end < len(calls) {
+		if !parallelSafeToolCall(registry, calls[end], options) {
+			break
+		}
+		nextKeys := resourceKeysForCall(registry, calls[end])
+		conflict := false
+		for _, prev := range keysWindow {
+			if resourceKeysConflict(prev, nextKeys) {
+				conflict = true
+				break
+			}
+		}
+		if conflict {
+			break
+		}
+		keysWindow = append(keysWindow, nextKeys)
+		end++
+	}
+	return end
+}
+
+// executeParallelReadBatch runs calls[start:end] concurrently (bounded by
+// maxParallelReadTools) and returns results indexed relative to start. All
+// execution-side callbacks that can fire inside executeToolCall are serialized
+// behind one mutex: a permission prompt (a sandbox preflight can demand one
+// even for an auto-allowed read) must never appear twice at once on an
+// interactive front-end, and OnPermission event handlers append to shared
+// session-recording state without their own locking — two batched reads under
+// a granted extra root would otherwise race (the pre-batch serial loop never
+// had two callbacks in flight at once).
+func executeParallelReadBatch(ctx context.Context, registry *tools.Registry, calls []ToolCall, start, end int, permissionMode PermissionMode, options Options) []precomputedToolResult {
+	batchOptions := options
+	var callbackMutex sync.Mutex
+	if options.OnPermissionRequest != nil {
+		inner := options.OnPermissionRequest
+		batchOptions.OnPermissionRequest = func(ctx context.Context, request PermissionRequest) (PermissionDecision, error) {
+			callbackMutex.Lock()
+			defer callbackMutex.Unlock()
+			return inner(ctx, request)
+		}
+	}
+	if options.OnPermission != nil {
+		inner := options.OnPermission
+		batchOptions.OnPermission = func(event PermissionEvent) {
+			callbackMutex.Lock()
+			defer callbackMutex.Unlock()
+			inner(event)
+		}
+	}
+
+	results := make([]precomputedToolResult, end-start)
+	semaphore := make(chan struct{}, maxParallelReadTools)
+	var waitGroup sync.WaitGroup
+	for index := start; index < end; index++ {
+		waitGroup.Add(1)
+		go func(index int) {
+			defer waitGroup.Done()
+			semaphore <- struct{}{}
+			defer func() { <-semaphore }()
+			result, abortErr := executeToolCall(ctx, registry, calls[index], permissionMode, batchOptions)
+			results[index-start] = precomputedToolResult{result: result, abortErr: abortErr}
+		}(index)
+	}
+	waitGroup.Wait()
+	return results
+}

internal/agent/parallel_tools_test.go

@@ -0,0 +1,308 @@
+package agent
+
+import (
+	"context"
+	"fmt"
+	"sync"
+	"testing"
+	"time"
+
+	"github.com/Gitlawb/zero/internal/tools"
+	"github.com/Gitlawb/zero/internal/zeroruntime"
+)
+
+// probeTool records execution overlap and ordering so tests can assert what
+// actually ran concurrently.
+type probeTool struct {
+	name       string
+	sideEffect tools.SideEffect
+	delay      time.Duration
+	// shared, when set, receives this probe's start/end entries too, so a test
+	// can observe ordering ACROSS probes (e.g. reads vs a write barrier).
+	shared *probeLog
+
+	mu        sync.Mutex
+	active    int
+	maxActive int
+	log       []string
+}
+
+// probeLog is a mutex-guarded event log shared between probes.
+type probeLog struct {
+	mu      sync.Mutex
+	entries []string
+}
+
+func (log *probeLog) append(entry string) {
+	log.mu.Lock()
+	defer log.mu.Unlock()
+	log.entries = append(log.entries, entry)
+}
+
+func (log *probeLog) snapshot() []string {
+	log.mu.Lock()
+	defer log.mu.Unlock()
+	return append([]string(nil), log.entries...)
+}
+
+func (tool *probeTool) Name() string        { return tool.name }
+func (tool *probeTool) Description() string { return "test probe tool" }
+func (tool *probeTool) Parameters() tools.Schema {
+	return tools.Schema{
+		Type:                 "object",
+		Properties:           map[string]tools.PropertySchema{"id": {Type: "string"}},
+		AdditionalProperties: false,
+	}
+}
+func (tool *probeTool) Safety() tools.Safety {
+	return tools.Safety{SideEffect: tool.sideEffect, Permission: tools.PermissionAllow, Reason: "test"}
+}
+
+// Capabilities maps the probe's SideEffect into the PR5 capability contract so
+// parallelSafeToolCall (CapabilitiesOf) can classify reads vs mutators. Without
+// this, probes default to EffectUnknown and never enter the concurrent path.
+func (tool *probeTool) Capabilities() tools.ToolCapabilities {
+	switch tool.sideEffect {
+	case tools.SideEffectRead:
+		return tools.ToolCapabilities{Effect: tools.EffectReadOnly, ThreadSafe: true}
+	case tools.SideEffectWrite:
+		return tools.ToolCapabilities{Effect: tools.EffectWorkspaceWrite, ThreadSafe: false}
+	default:
+		return tools.UnknownCapabilities()
+	}
+}
+func (tool *probeTool) Run(_ context.Context, args map[string]any) tools.Result {
+	id, _ := args["id"].(string)
+	tool.mu.Lock()
+	tool.active++
+	if tool.active > tool.maxActive {
+		tool.maxActive = tool.active
+	}
+	tool.log = append(tool.log, "start:"+id)
+	tool.mu.Unlock()
+	if tool.shared != nil {
+		tool.shared.append("start:" + id)
+	}
+	time.Sleep(tool.delay)
+	if tool.shared != nil {
+		tool.shared.append("end:" + id)
+	}
+	tool.mu.Lock()
+	tool.active--
+	tool.log = append(tool.log, "end:"+id)
+	tool.mu.Unlock()
+	return tools.Result{Status: tools.StatusOK, Output: "probe " + id}
+}
+
+func probeCallEvents(callID, toolName, id string) []zeroruntime.StreamEvent {
+	return []zeroruntime.StreamEvent{
+		{Type: zeroruntime.StreamEventToolCallStart, ToolCallID: callID, ToolName: toolName},
+		{Type: zeroruntime.StreamEventToolCallDelta, ToolCallID: callID, ArgumentsFragment: fmt.Sprintf(`{"id":%q}`, id)},
+		{Type: zeroruntime.StreamEventToolCallEnd, ToolCallID: callID},
+	}
+}
+
+func TestParallelSafeToolCall(t *testing.T) {
+	registry := tools.NewRegistry()
+	registry.Register(&probeTool{name: "probe_read", sideEffect: tools.SideEffectRead})
+	registry.Register(&probeTool{name: "probe_write", sideEffect: tools.SideEffectWrite})
+	// Read-only but not ThreadSafe: must stay sequential under the PR5 gate.
+	registry.Register(&keyedProbeTool{
+		probeTool:  probeTool{name: "probe_read_serial", sideEffect: tools.SideEffectRead},
+		threadSafe: false,
+	})
+
+	call := func(name, args string) ToolCall { return ToolCall{ID: "c", Name: name, Arguments: args} }
+	if !parallelSafeToolCall(registry, call("probe_read", `{"id":"a"}`), Options{}) {
+		t.Fatal("auto-allowed read tool must be parallel-safe")
+	}
+	if parallelSafeToolCall(registry, call("probe_write", `{"id":"a"}`), Options{}) {
+		t.Fatal("mutating tool must not be parallel-safe")
+	}
+	if parallelSafeToolCall(registry, call("unknown_tool", `{}`), Options{}) {
+		t.Fatal("unknown tool must not be parallel-safe")
+	}
+	if parallelSafeToolCall(registry, call("probe_read", `{"id":`), Options{}) {
+		t.Fatal("undecodable arguments must not be parallel-safe")
+	}
+	if parallelSafeToolCall(registry, call("ask_user", `{}`), Options{}) {
+		t.Fatal("loop-intercepted tools must stay sequential")
+	}
+	if parallelSafeToolCall(registry, call("probe_read_serial", `{"id":"a"}`), Options{}) {
+		t.Fatal("ReadOnly without ThreadSafe must not be parallel-safe")
+	}
+}
+
+func TestResourceKeysConflict(t *testing.T) {
+	if resourceKeysConflict(nil, []string{"file:a"}) {
+		t.Fatal("empty vs non-empty must not conflict")
+	}
+	if resourceKeysConflict([]string{"file:a"}, nil) {
+		t.Fatal("non-empty vs empty must not conflict")
+	}
+	if resourceKeysConflict([]string{"file:a"}, []string{"file:b"}) {
+		t.Fatal("distinct keys must not conflict")
+	}
+	if !resourceKeysConflict([]string{"file:a", "file:b"}, []string{"file:b"}) {
+		t.Fatal("shared key must conflict")
+	}
+	if resourceKeysConflict([]string{"", "file:a"}, []string{""}) {
+		t.Fatal("empty-string keys must be ignored")
+	}
+}
+
+func TestExtendParallelRunResourceKeyBoundary(t *testing.T) {
+	// Two keyed reads on the same path must not share a concurrent window;
+	// distinct paths may batch together.
+	registry := tools.NewRegistry()
+	registry.Register(&keyedProbeTool{
+		probeTool:  probeTool{name: "keyed_read", sideEffect: tools.SideEffectRead},
+		threadSafe: true,
+		keys: func(args map[string]any) []string {
+			id, _ := args["id"].(string)
+			if id == "" {
+				return nil
+			}
+			return []string{"file:" + id}
+		},
+	})
+	calls := []ToolCall{
+		{ID: "1", Name: "keyed_read", Arguments: `{"id":"a"}`},
+		{ID: "2", Name: "keyed_read", Arguments: `{"id":"b"}`},
+		{ID: "3", Name: "keyed_read", Arguments: `{"id":"a"}`}, // conflicts with call 0
+	}
+	// start=0: a and b share no key → run covers [0,2)
+	if end := extendParallelRun(registry, calls, 0, Options{}); end != 2 {
+		t.Fatalf("extend from 0 = %d, want 2 (stop before second file:a)", end)
+	}
+	// start=2: single remaining call
+	if end := extendParallelRun(registry, calls, 2, Options{}); end != 3 {
+		t.Fatalf("extend from 2 = %d, want 3", end)
+	}
+	// Empty keys never conflict: three keyless safe reads form one window.
+	registry.Register(&probeTool{name: "keyless_read", sideEffect: tools.SideEffectRead})
+	keyless := []ToolCall{
+		{ID: "1", Name: "keyless_read", Arguments: `{"id":"x"}`},
+		{ID: "2", Name: "keyless_read", Arguments: `{"id":"y"}`},
+		{ID: "3", Name: "keyless_read", Arguments: `{"id":"z"}`},
+	}
+	if end := extendParallelRun(registry, keyless, 0, Options{}); end != 3 {
+		t.Fatalf("keyless extend = %d, want 3", end)
+	}
+}
+
+// keyedProbeTool is a probe with explicit ResourceKeys / ThreadSafe for
+// planner unit tests.
+type keyedProbeTool struct {
+	probeTool
+	threadSafe bool
+	keys       func(args map[string]any) []string
+}
+
+func (tool *keyedProbeTool) Capabilities() tools.ToolCapabilities {
+	caps := tools.ToolCapabilities{
+		Effect:     tools.EffectReadOnly,
+		ThreadSafe: tool.threadSafe,
+	}
+	if tool.keys != nil {
+		caps.ResourceKeys = tool.keys
+	}
+	// Honor mutator side effects from the embedded probe when used as write.
+	if tool.sideEffect == tools.SideEffectWrite {
+		caps.Effect = tools.EffectWorkspaceWrite
+		caps.ThreadSafe = false
+	}
+	return caps
+}
+
+func TestRunExecutesConsecutiveReadsConcurrently(t *testing.T) {
+	probe := &probeTool{name: "probe_read", sideEffect: tools.SideEffectRead, delay: 60 * time.Millisecond}
+	registry := tools.NewRegistry()
+	registry.Register(probe)
+
+	turnOne := append(probeCallEvents("call-1", "probe_read", "a"), probeCallEvents("call-2", "probe_read", "b")...)
+	turnOne = append(turnOne, probeCallEvents("call-3", "probe_read", "c")...)
+	turnOne = append(turnOne, zeroruntime.StreamEvent{Type: zeroruntime.StreamEventDone})
+	provider := &mockProvider{
+		turns: [][]zeroruntime.StreamEvent{
+			turnOne,
+			{{Type: zeroruntime.StreamEventText, Content: "done"}, {Type: zeroruntime.StreamEventDone}},
+		},
+	}
+
+	var results []ToolResult
+	_, err := Run(context.Background(), "probe", provider, Options{
+		Registry:     registry,
+		OnToolResult: func(result ToolResult) { results = append(results, result) },
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if probe.maxActive < 2 {
+		t.Fatalf("consecutive read-only calls must overlap, max concurrency was %d", probe.maxActive)
+	}
+	// Results must still be recorded in original call order.
+	if len(results) != 3 || results[0].ToolCallID != "call-1" || results[1].ToolCallID != "call-2" || results[2].ToolCallID != "call-3" {
+		t.Fatalf("tool results out of order: %#v", results)
+	}
+	messages := provider.requests[1].Messages
+	var toolOrder []string
+	for _, message := range messages {
+		if message.Role == zeroruntime.MessageRoleTool {
+			toolOrder = append(toolOrder, message.ToolCallID)
+		}
+	}
+	if len(toolOrder) != 3 || toolOrder[0] != "call-1" || toolOrder[1] != "call-2" || toolOrder[2] != "call-3" {
+		t.Fatalf("recorded tool messages out of order: %v", toolOrder)
+	}
+}
+
+func TestRunParallelReadsNeverSpanMutatingCall(t *testing.T) {
+	shared := &probeLog{}
+	read := &probeTool{name: "probe_read", sideEffect: tools.SideEffectRead, delay: 30 * time.Millisecond, shared: shared}
+	write := &probeTool{name: "probe_write", sideEffect: tools.SideEffectWrite, shared: shared}
+	registry := tools.NewRegistry()
+	registry.Register(read)
+	registry.Register(write)
+
+	turnOne := append(probeCallEvents("call-1", "probe_read", "r1"), probeCallEvents("call-2", "probe_read", "r2")...)
+	turnOne = append(turnOne, probeCallEvents("call-3", "probe_write", "w")...)
+	turnOne = append(turnOne, probeCallEvents("call-4", "probe_read", "r3")...)
+	turnOne = append(turnOne, probeCallEvents("call-5", "probe_read", "r4")...)
+	turnOne = append(turnOne, zeroruntime.StreamEvent{Type: zeroruntime.StreamEventDone})
+	provider := &mockProvider{
+		turns: [][]zeroruntime.StreamEvent{
+			turnOne,
+			{{Type: zeroruntime.StreamEventText, Content: "done"}, {Type: zeroruntime.StreamEventDone}},
+		},
+	}
+
+	_, err := Run(context.Background(), "probe", provider, Options{Registry: registry})
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	// Cross-probe ordering on the SHARED log: the write must start only after
+	// both first-batch reads finished, and both second-batch reads must start
+	// only after the write finished — batches never cross a mutating call.
+	log := shared.snapshot()
+	index := func(entry string) int {
+		for i, e := range log {
+			if e == entry {
+				return i
+			}
+		}
+		t.Fatalf("entry %q missing from shared log: %v", entry, log)
+		return -1
+	}
+	writeStart, writeEnd := index("start:w"), index("end:w")
+	if firstBatchMaxEnd := max(index("end:r1"), index("end:r2")); writeStart < firstBatchMaxEnd {
+		t.Fatalf("write started before the first read batch finished: %v", log)
+	}
+	if secondBatchMinStart := min(index("start:r3"), index("start:r4")); secondBatchMinStart < writeEnd {
+		t.Fatalf("second read batch started before the write finished: %v", log)
+	}
+	if read.maxActive < 2 {
+		t.Fatalf("reads within a batch must overlap, max concurrency was %d", read.maxActive)
+	}
+}

internal/agent/partition_cache_stable_test.go

@@ -0,0 +1,44 @@
+package agent
+
+import (
+	"testing"
+
+	"github.com/Gitlawb/zero/internal/tools"
+)
+
+func TestPartitionToolsActiveAppendsLoadedToolAfterEagerBlock(t *testing.T) {
+	// Cache-stability property: loading a deferred tool must APPEND it after the
+	// always-eager tools, not sort it into the middle of them — otherwise the tool
+	// definitions (the provider's cacheable prefix) shift on every load. The old
+	// behaviour alpha-sorted the whole array, so "mcp__srv__alpha" jumped ahead of
+	// "read_file".
+	root := t.TempDir()
+	registry := tools.NewRegistry()
+	registry.Register(tools.NewReadFileTool(root)) // non-deferred, "read_file"
+	registry.Register(fakeToolSearchTool{})
+	registry.Register(fakeDeferredTool{name: "mcp__srv__alpha", desc: "alpha"})
+	registry.Register(fakeDeferredTool{name: "mcp__srv__beta", desc: "beta"})
+
+	exposed, _ := partitionTools(registry, PermissionModeAuto, Options{DeferThreshold: 2}, map[string]bool{"mcp__srv__alpha": true})
+
+	pos := func(name string) int {
+		for i := range exposed {
+			if exposed[i].Name == name {
+				return i
+			}
+		}
+		return -1
+	}
+	readPos, searchPos, alphaPos := pos("read_file"), pos("tool_search"), pos("mcp__srv__alpha")
+	if readPos < 0 || searchPos < 0 || alphaPos < 0 {
+		t.Fatalf("expected read_file, tool_search and loaded alpha all exposed, got %#v", exposed)
+	}
+	// The eager tools must precede the loaded deferred tool (appended, not inserted).
+	if readPos >= alphaPos || searchPos >= alphaPos {
+		t.Fatalf("loaded deferred tool must be appended after the eager block; got read=%d search=%d alpha=%d", readPos, searchPos, alphaPos)
+	}
+	// beta was never loaded — it stays hidden.
+	if pos("mcp__srv__beta") != -1 {
+		t.Fatalf("unloaded deferred tool must stay hidden, got %#v", exposed)
+	}
+}

internal/agent/partition_cache_test.go

@@ -0,0 +1,63 @@
+package agent
+
+import (
+	"context"
+	"reflect"
+	"testing"
+
+	"github.com/Gitlawb/zero/internal/tools"
+	"github.com/Gitlawb/zero/internal/zeroruntime"
+)
+
+// countingSchemaTool records how many times its Parameters() schema is read, so a
+// test can prove the definition cache avoids re-rendering across turns.
+type countingSchemaTool struct {
+	name  string
+	calls *int
+}
+
+func (t countingSchemaTool) Name() string        { return t.name }
+func (t countingSchemaTool) Description() string { return "counts schema reads" }
+func (t countingSchemaTool) Parameters() tools.Schema {
+	*t.calls++
+	return tools.Schema{Type: "object", AdditionalProperties: false, Properties: map[string]tools.PropertySchema{
+		"x": {Type: "string"},
+	}}
+}
+func (t countingSchemaTool) Safety() tools.Safety {
+	return tools.Safety{SideEffect: tools.SideEffectRead, Permission: tools.PermissionAllow}
+}
+func (t countingSchemaTool) Run(_ context.Context, _ map[string]any) tools.Result {
+	return tools.Result{Status: tools.StatusOK}
+}
+
+// The cache renders a tool's schema once and reuses it across calls, and its
+// output is identical to the uncached path.
+func TestPartitionToolsCacheRendersOnceAndMatchesUncached(t *testing.T) {
+	calls := 0
+	registry := tools.NewRegistry()
+	registry.Register(countingSchemaTool{name: "alpha", calls: &calls})
+	registry.Register(countingSchemaTool{name: "beta", calls: &calls})
+
+	uncached, _ := partitionTools(registry, PermissionModeAuto, Options{}, map[string]bool{})
+
+	base := calls
+	cache := map[string]zeroruntime.ToolDefinition{}
+	first, _ := partitionToolsCached(registry, PermissionModeAuto, Options{}, map[string]bool{}, cache)
+	rendersAfterFirst := calls - base
+	second, _ := partitionToolsCached(registry, PermissionModeAuto, Options{}, map[string]bool{}, cache)
+	rendersAfterSecond := calls - base
+
+	if rendersAfterFirst != 2 {
+		t.Fatalf("first cached call should render 2 tools once each, rendered %d times", rendersAfterFirst)
+	}
+	if rendersAfterSecond != 2 {
+		t.Fatalf("second cached call must reuse the cache (no new schema reads), total renders = %d", rendersAfterSecond)
+	}
+	if !reflect.DeepEqual(first, second) {
+		t.Fatal("cached calls must return identical definitions")
+	}
+	if !reflect.DeepEqual(first, uncached) {
+		t.Fatal("cached output must match the uncached partitionTools output")
+	}
+}

internal/agent/permission_additional_perms_test.go

@@ -0,0 +1,70 @@
+package agent
+
+import (
+	"context"
+	"testing"
+
+	"github.com/Gitlawb/zero/internal/sandbox"
+	"github.com/Gitlawb/zero/internal/tools"
+)
+
+// promptShellTool is a minimal shell-style tool whose Safety requires a prompt,
+// standing in for bash so buildPermissionEvent can be exercised directly.
+type promptShellTool struct{}
+
+func (promptShellTool) Name() string        { return "bash" }
+func (promptShellTool) Description() string { return "run a shell command" }
+func (promptShellTool) Parameters() tools.Schema {
+	return tools.Schema{Type: "object", AdditionalProperties: false}
+}
+func (promptShellTool) Safety() tools.Safety {
+	return tools.Safety{SideEffect: tools.SideEffectShell, Permission: tools.PermissionPrompt, Reason: "runs a shell command"}
+}
+func (promptShellTool) Run(context.Context, map[string]any) tools.Result {
+	return tools.Result{Status: tools.StatusOK}
+}
+
+// A command whose sandbox decision is Allow but that explicitly requests
+// additional sandbox permissions is an ELEVATION: it must be surfaced as a
+// prompt, never Action=allow. Previously it carried allow while the loop still
+// blocked on OnPermissionRequest, and the TUI (which renders only prompts)
+// dropped it → the run deadlocked forever. This is the exact gpt-5.5
+// folder-creation shape captured in the reported hang.
+func TestBuildPermissionEventPromptsForAdditionalPermissions(t *testing.T) {
+	call := ToolCall{ID: "call_1", Name: "bash", Arguments: `{}`}
+	args := map[string]any{
+		"command":             "mkdir foo && cd foo && go mod init foo",
+		"sandbox_permissions": string(tools.SandboxPermissionsWithAdditionalPermissions),
+	}
+	// Base command allowed by the sandbox…
+	decision := &sandbox.Decision{Action: sandbox.ActionAllow}
+
+	event, ok := buildPermissionEvent(call, promptShellTool{}, args, false, PermissionModeAsk, Options{}, decision)
+	if !ok {
+		t.Fatal("buildPermissionEvent returned ok=false; the elevation request must produce an event")
+	}
+	if event.Action != PermissionActionPrompt {
+		t.Fatalf("Action = %q, want prompt — an additional-permissions elevation must ask the user, not auto-allow", event.Action)
+	}
+	// Sanity: shouldRequestPermission agrees the loop blocks on this.
+	if !shouldRequestPermission(promptShellTool{}, args, false, decision) {
+		t.Fatal("shouldRequestPermission must be true for an additional-permissions request")
+	}
+}
+
+// Without the additional-permissions flag, a sandbox-allowed command is NOT a
+// prompt (shouldRequestPermission is false; it just runs) — the override must
+// not over-reach and prompt for ordinary allowed commands.
+func TestBuildPermissionEventKeepsAllowForOrdinaryAllowedCommand(t *testing.T) {
+	call := ToolCall{ID: "call_2", Name: "bash", Arguments: `{}`}
+	args := map[string]any{"command": "ls"}
+	decision := &sandbox.Decision{Action: sandbox.ActionAllow}
+
+	event, _ := buildPermissionEvent(call, promptShellTool{}, args, false, PermissionModeAsk, Options{}, decision)
+	if event.Action == PermissionActionPrompt {
+		t.Fatalf("Action = prompt for an ordinary allowed command; want allow (no spurious prompt)")
+	}
+	if shouldRequestPermission(promptShellTool{}, args, false, decision) {
+		t.Fatal("shouldRequestPermission must be false for an ordinary sandbox-allowed command")
+	}
+}

internal/agent/plan_staleness_test.go

@@ -0,0 +1,64 @@
+package agent
+
+import (
+	"strings"
+	"testing"
+
+	"github.com/Gitlawb/zero/internal/zeroruntime"
+)
+
+func planUpdateTurn(status string) zeroruntime.CollectedStream {
+	return zeroruntime.CollectedStream{ToolCalls: []zeroruntime.ToolCall{
+		{Name: "update_plan", Arguments: `{"plan":[{"content":"step one","status":"` + status + `"}]}`},
+	}}
+}
+
+// Turn-based staleness: a plan that goes many turns without an update — while
+// items are still pending — draws the stale reminder even though few tool calls
+// have run (the tool-call streak alone would not have tripped).
+func TestPlanReminderFiresOnTurnStaleness(t *testing.T) {
+	state := newGuardState()
+	state.observeTurn(planUpdateTurn("in_progress"))
+
+	// Text-only turns advance the turn counter without adding tool calls.
+	for range stalePlanTurnThreshold {
+		state.observeTurn(zeroruntime.CollectedStream{Text: "still working…"})
+	}
+
+	if state.toolCallsSincePlanUpdate >= staleToolCallThreshold {
+		t.Fatalf("precondition: the tool-call trigger must not fire (%d calls)", state.toolCallsSincePlanUpdate)
+	}
+	if got := state.planReminder(stalePlanTurnThreshold + 2); !strings.Contains(got, planStaleReminderMarker) {
+		t.Fatalf("expected a turn-based stale reminder, got %q", got)
+	}
+}
+
+// A fully-completed plan is NOT stale: no pending items means the turn-based
+// trigger stays quiet even after many turns, so a finished plan isn't nagged.
+func TestPlanReminderSkipsTurnStalenessWhenComplete(t *testing.T) {
+	state := newGuardState()
+	state.observeTurn(planUpdateTurn("completed"))
+
+	for range stalePlanTurnThreshold + 3 {
+		state.observeTurn(zeroruntime.CollectedStream{Text: "wrapping up…"})
+	}
+
+	if got := state.planReminder(stalePlanTurnThreshold + 5); got != "" {
+		t.Fatalf("a completed plan must not draw a stale reminder, got %q", got)
+	}
+}
+
+// A plan update resets the turn counter, so the stale reminder does not fire
+// right after a fresh update.
+func TestPlanUpdateResetsTurnStaleness(t *testing.T) {
+	state := newGuardState()
+	state.observeTurn(planUpdateTurn("in_progress"))
+	for range stalePlanTurnThreshold - 1 {
+		state.observeTurn(zeroruntime.CollectedStream{Text: "working…"})
+	}
+	// Refresh the plan — resets the turn streak.
+	state.observeTurn(planUpdateTurn("in_progress"))
+	if got := state.planReminder(stalePlanTurnThreshold + 1); got != "" {
+		t.Fatalf("a just-refreshed plan must not be stale, got %q", got)
+	}
+}

internal/agent/profile_controller.go

@@ -0,0 +1,118 @@
+package agent
+
+import (
+	"github.com/Gitlawb/zero/internal/sandbox"
+)
+
+// profileController observes per-turn run signals and decides at most one
+// posture escalation per run for the armed execution profile. A nil policy (or
+// a policy without Escalate) makes every method a no-op, so runs without a
+// profile are byte-identical to before — the same opt-in convention as Trace
+// and SelfCorrect.
+//
+// The controller only ever CHANGES loop policy knobs (turn ceiling, reasoning
+// effort, completion gate) through the values the loop applies at its act
+// point; it never injects messages, never touches the model or session, and
+// never interacts with the model-initiated escalate_model path.
+type profileController struct {
+	policy *PostureEscalation
+
+	escalated     bool
+	uncertainSeen int
+
+	failureTripped bool
+	riskTripped    bool
+	scTripped      bool
+	uncertainTrip  bool
+}
+
+// newProfileController is nil-safe: a nil ProfilePolicy (the default for every
+// existing caller) or a nil Escalate yields a controller that observes and
+// decides nothing.
+func newProfileController(policy *ProfilePolicy) *profileController {
+	if policy == nil {
+		return &profileController{}
+	}
+	return &profileController{policy: policy.Escalate}
+}
+
+// observeToolOutcome watches two signals from an executed tool call: the
+// repeated-failure guard's streak (reusing the guard's own counting, not a
+// second failure model) and the sandbox risk classification of an executed
+// result.
+func (c *profileController) observeToolOutcome(outcome toolFailureOutcome, result ToolResult) {
+	if c.policy == nil || c.escalated {
+		return
+	}
+	if c.policy.OnToolFailureStreak > 0 && outcome.Count >= c.policy.OnToolFailureStreak {
+		c.failureTripped = true
+	}
+	if threshold := riskRank(c.policy.OnRiskyMutation); threshold > 0 &&
+		result.DenialReason == DenialNone &&
+		riskRank(result.Risk.Level) >= threshold {
+		// The call EXECUTED (was not denied): partial failures count too, since
+		// the mutation ran. An unrecognized threshold ranks 0 and disables the
+		// signal instead of matching everything.
+		c.riskTripped = true
+	}
+}
+
+// observeUncertain counts uncertain completion evaluations (continue nudges and
+// semantic checks). Headless-only by construction: the completion gate never
+// runs interactively.
+func (c *profileController) observeUncertain() {
+	if c.policy == nil || c.escalated || c.policy.OnCompletionUncertain <= 0 {
+		return
+	}
+	c.uncertainSeen++
+	if c.uncertainSeen >= c.policy.OnCompletionUncertain {
+		c.uncertainTrip = true
+	}
+}
+
+// observeSelfCorrect watches the post-edit verification outcome. Any failing
+// outcome (correcting, reported, aborted) is a signal; passed and disabled are
+// not.
+func (c *profileController) observeSelfCorrect(outcome Outcome) {
+	if c.policy == nil || c.escalated || !c.policy.OnSelfCorrectFailure {
+		return
+	}
+	switch outcome {
+	case OutcomeCorrecting, OutcomeReported, OutcomeAborted:
+		c.scTripped = true
+	}
+}
+
+// maybeEscalate reports the escalation target exactly once, the first time any
+// armed trigger has fired. Subsequent calls return false for the rest of the
+// run: escalation is one-shot and never de-escalates, so no cooldown state is
+// needed.
+func (c *profileController) maybeEscalate() (PostureEscalation, bool) {
+	if c.policy == nil || c.escalated {
+		return PostureEscalation{}, false
+	}
+	if !(c.failureTripped || c.riskTripped || c.scTripped || c.uncertainTrip) {
+		return PostureEscalation{}, false
+	}
+	c.escalated = true
+	return *c.policy, true
+}
+
+// riskRank orders sandbox risk levels for threshold comparison. Unknown levels
+// rank 0: an unrecognized RESULT level can never meet a valid threshold, and an
+// unrecognized THRESHOLD is rejected before comparison (rank 0 disables the
+// signal) so a typo in a profile can never make every result match.
+func riskRank(level sandbox.RiskLevel) int {
+	switch level {
+	case sandbox.RiskLow:
+		return 1
+	case sandbox.RiskMedium:
+		return 2
+	case sandbox.RiskHigh:
+		return 3
+	case sandbox.RiskCritical:
+		return 4
+	default:
+		return 0
+	}
+}

internal/agent/profile_controller_test.go

@@ -0,0 +1,345 @@
+package agent
+
+import (
+	"context"
+	"strings"
+	"testing"
+
+	"github.com/Gitlawb/zero/internal/sandbox"
+	"github.com/Gitlawb/zero/internal/tools"
+	"github.com/Gitlawb/zero/internal/trace"
+	"github.com/Gitlawb/zero/internal/zeroruntime"
+)
+
+func TestProfileControllerNilPolicyIsNoOp(t *testing.T) {
+	for _, controller := range []*profileController{
+		newProfileController(nil),
+		newProfileController(&ProfilePolicy{Name: "balanced"}), // Escalate nil
+	} {
+		controller.observeToolOutcome(toolFailureOutcome{Count: 99}, ToolResult{Status: tools.StatusOK, Risk: sandbox.Risk{Level: sandbox.RiskCritical}})
+		controller.observeUncertain()
+		controller.observeSelfCorrect(OutcomeAborted)
+		if _, fired := controller.maybeEscalate(); fired {
+			t.Fatal("nil policy must never escalate")
+		}
+	}
+}
+
+func TestProfileControllerEscalatesOnFailureStreak(t *testing.T) {
+	controller := newProfileController(&ProfilePolicy{Escalate: &PostureEscalation{OnToolFailureStreak: 2, MaxTurns: 80}})
+	controller.observeToolOutcome(toolFailureOutcome{Count: 1}, ToolResult{Status: tools.StatusError})
+	if _, fired := controller.maybeEscalate(); fired {
+		t.Fatal("streak below threshold must not escalate")
+	}
+	controller.observeToolOutcome(toolFailureOutcome{Count: 2}, ToolResult{Status: tools.StatusError})
+	target, fired := controller.maybeEscalate()
+	if !fired || target.MaxTurns != 80 {
+		t.Fatalf("expected escalation with the policy target, got fired=%t target=%+v", fired, target)
+	}
+}
+
+func TestProfileControllerEscalatesOnRiskyMutation(t *testing.T) {
+	controller := newProfileController(&ProfilePolicy{Escalate: &PostureEscalation{OnRiskyMutation: sandbox.RiskHigh}})
+	// Medium risk on an executed result: below threshold.
+	controller.observeToolOutcome(toolFailureOutcome{}, ToolResult{Status: tools.StatusOK, Risk: sandbox.Risk{Level: sandbox.RiskMedium}})
+	if _, fired := controller.maybeEscalate(); fired {
+		t.Fatal("medium risk must not trip a high threshold")
+	}
+	// Critical risk on a DENIED result: never counts (the mutation never ran).
+	controller.observeToolOutcome(toolFailureOutcome{}, ToolResult{Status: tools.StatusError, DenialReason: DenialSandboxBlock, Risk: sandbox.Risk{Level: sandbox.RiskCritical}})
+	if _, fired := controller.maybeEscalate(); fired {
+		t.Fatal("a denied result must not trip the risky-mutation signal")
+	}
+	// Critical risk on an executed PARTIAL FAILURE (error status, no denial):
+	// the mutation ran, so it counts.
+	controller.observeToolOutcome(toolFailureOutcome{}, ToolResult{Status: tools.StatusError, Risk: sandbox.Risk{Level: sandbox.RiskCritical}})
+	if _, fired := controller.maybeEscalate(); !fired {
+		t.Fatal("an executed partial failure at critical risk must trip a high threshold")
+	}
+}
+
+func TestProfileControllerIgnoresInvalidRiskThreshold(t *testing.T) {
+	controller := newProfileController(&ProfilePolicy{Escalate: &PostureEscalation{OnRiskyMutation: sandbox.RiskLevel("hgih")}})
+	controller.observeToolOutcome(toolFailureOutcome{}, ToolResult{Status: tools.StatusOK, Risk: sandbox.Risk{Level: sandbox.RiskCritical}})
+	if _, fired := controller.maybeEscalate(); fired {
+		t.Fatal("an unrecognized threshold must disable the signal, not match every result")
+	}
+}
+
+func TestProfileControllerEscalatesOnSelfCorrectFailure(t *testing.T) {
+	controller := newProfileController(&ProfilePolicy{Escalate: &PostureEscalation{OnSelfCorrectFailure: true}})
+	controller.observeSelfCorrect(OutcomePassed)
+	controller.observeSelfCorrect(OutcomeDisabled)
+	if _, fired := controller.maybeEscalate(); fired {
+		t.Fatal("passing/disabled outcomes must not escalate")
+	}
+	controller.observeSelfCorrect(OutcomeReported)
+	if _, fired := controller.maybeEscalate(); !fired {
+		t.Fatal("a failing verification outcome must escalate")
+	}
+}
+
+func TestProfileControllerEscalatesOnUncertainCompletion(t *testing.T) {
+	controller := newProfileController(&ProfilePolicy{Escalate: &PostureEscalation{OnCompletionUncertain: 2}})
+	controller.observeUncertain()
+	if _, fired := controller.maybeEscalate(); fired {
+		t.Fatal("first uncertain evaluation must not trip a threshold of 2")
+	}
+	controller.observeUncertain()
+	if _, fired := controller.maybeEscalate(); !fired {
+		t.Fatal("second uncertain evaluation must escalate")
+	}
+}
+
+func TestProfileControllerEscalatesAtMostOnce(t *testing.T) {
+	controller := newProfileController(&ProfilePolicy{Escalate: &PostureEscalation{OnToolFailureStreak: 1}})
+	controller.observeToolOutcome(toolFailureOutcome{Count: 1}, ToolResult{Status: tools.StatusError})
+	if _, fired := controller.maybeEscalate(); !fired {
+		t.Fatal("expected first escalation")
+	}
+	controller.observeToolOutcome(toolFailureOutcome{Count: 5}, ToolResult{Status: tools.StatusOK, Risk: sandbox.Risk{Level: sandbox.RiskCritical}})
+	controller.observeUncertain()
+	if _, fired := controller.maybeEscalate(); fired {
+		t.Fatal("escalation must be one-shot per run")
+	}
+}
+
+// TestPostureEscalationOnUncertainCompletion proves the uncertain-path act
+// point: the completion gate's continue branch escapes the end-of-turn tail,
+// so escalation must apply before the continue. A cue turn under the gate
+// counts as uncertain, escalates, and the very next request carries the target
+// effort.
+func TestPostureEscalationOnUncertainCompletion(t *testing.T) {
+	provider := &mockProvider{turns: [][]zeroruntime.StreamEvent{
+		{
+			{Type: zeroruntime.StreamEventText, Content: "Let me read the file:"},
+			{Type: zeroruntime.StreamEventDone},
+		},
+		{
+			{Type: zeroruntime.StreamEventText, Content: "Done. All set."},
+			{Type: zeroruntime.StreamEventDone},
+		},
+	}}
+
+	recorder := trace.NewRecorder("posture-session", "run-2", "fast")
+	result, err := Run(context.Background(), "go", provider, Options{
+		Registry:                tools.NewRegistry(),
+		MaxTurns:                10,
+		RequireCompletionSignal: true,
+		Trace:                   recorder,
+		Profile: &ProfilePolicy{
+			Name: "fast",
+			Escalate: &PostureEscalation{
+				ReasoningEffort:       "high",
+				OnCompletionUncertain: 1,
+			},
+		},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if result.FinalAnswer != "Done. All set." {
+		t.Fatalf("expected the completed answer, got %q", result.FinalAnswer)
+	}
+	if len(provider.requests) != 2 {
+		t.Fatalf("expected 2 requests, got %d", len(provider.requests))
+	}
+	if provider.requests[1].ReasoningEffort != "high" {
+		t.Fatalf("expected the post-escalation request to carry the target effort, got %q", provider.requests[1].ReasoningEffort)
+	}
+	tr := recorder.Finish()
+	if got := tr.Counter(trace.CounterPostureEscalations); got != 1 {
+		t.Fatalf("posture_escalations = %d, want 1", got)
+	}
+}
+
+// A profile that filled the run's effort displaced the provider default ("");
+// RestoreDefaultEffort lets escalation restore that default, which a plain ""
+// ReasoningEffort target cannot express (it means "leave untouched").
+func TestPostureEscalationRestoresDefaultEffort(t *testing.T) {
+	provider := &mockProvider{turns: [][]zeroruntime.StreamEvent{
+		{
+			{Type: zeroruntime.StreamEventText, Content: "Let me read the file:"},
+			{Type: zeroruntime.StreamEventDone},
+		},
+		{
+			{Type: zeroruntime.StreamEventText, Content: "Done. All set."},
+			{Type: zeroruntime.StreamEventDone},
+		},
+	}}
+
+	result, err := Run(context.Background(), "go", provider, Options{
+		Registry:                tools.NewRegistry(),
+		MaxTurns:                10,
+		ReasoningEffort:         "low",
+		RequireCompletionSignal: true,
+		Profile: &ProfilePolicy{
+			Name: "fast",
+			Escalate: &PostureEscalation{
+				RestoreDefaultEffort:  true,
+				OnCompletionUncertain: 1,
+			},
+		},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if result.FinalAnswer != "Done. All set." {
+		t.Fatalf("expected the completed answer, got %q", result.FinalAnswer)
+	}
+	if len(provider.requests) != 2 {
+		t.Fatalf("expected 2 requests, got %d", len(provider.requests))
+	}
+	if provider.requests[0].ReasoningEffort != "low" {
+		t.Fatalf("pre-escalation request effort = %q, want the profile's low", provider.requests[0].ReasoningEffort)
+	}
+	if provider.requests[1].ReasoningEffort != "" {
+		t.Fatalf("post-escalation request effort = %q, want the restored provider default (empty)", provider.requests[1].ReasoningEffort)
+	}
+}
+
+// failingProfileTool always errors with the same signature so the repeated-
+// failure guard builds a streak.
+type failingProfileTool struct{}
+
+func (failingProfileTool) Name() string        { return "flaky_probe" }
+func (failingProfileTool) Description() string { return "always fails identically (test)" }
+func (failingProfileTool) Parameters() tools.Schema {
+	return tools.Schema{Type: "object", AdditionalProperties: false}
+}
+func (failingProfileTool) Safety() tools.Safety {
+	return tools.Safety{SideEffect: tools.SideEffectRead, Permission: tools.PermissionAllow}
+}
+func (failingProfileTool) Run(context.Context, map[string]any) tools.Result {
+	return tools.Result{Status: tools.StatusError, Output: "probe exploded: same signature"}
+}
+
+// TestPostureEscalationRaisesTurnCeilingMidRun proves the end-to-end act path:
+// a Fast-style run with a 2-turn ceiling hits a failure streak, escalates, and
+// finishes on a turn the original ceiling would have cut off, with the effort
+// target applied to the post-escalation request and the counter emitted.
+func TestPostureEscalationRaisesTurnCeilingMidRun(t *testing.T) {
+	registry := tools.NewRegistry()
+	registry.Register(failingProfileTool{})
+
+	toolTurn := []zeroruntime.StreamEvent{
+		{Type: zeroruntime.StreamEventToolCallStart, ToolCallID: "c1", ToolName: "flaky_probe"},
+		{Type: zeroruntime.StreamEventToolCallEnd, ToolCallID: "c1"},
+		{Type: zeroruntime.StreamEventDone},
+	}
+	provider := &mockProvider{turns: [][]zeroruntime.StreamEvent{
+		toolTurn,
+		toolTurn,
+		{
+			{Type: zeroruntime.StreamEventText, Content: "recovered after escalation"},
+			{Type: zeroruntime.StreamEventDone},
+		},
+	}}
+
+	recorder := trace.NewRecorder("posture-session", "run-1", "fast")
+	result, err := Run(context.Background(), "go", provider, Options{
+		Registry: registry,
+		MaxTurns: 2,
+		Trace:    recorder,
+		Profile: &ProfilePolicy{
+			Name: "fast",
+			Escalate: &PostureEscalation{
+				MaxTurns:            4,
+				ReasoningEffort:     "high",
+				OnToolFailureStreak: 2,
+			},
+		},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if result.FinalAnswer != "recovered after escalation" {
+		t.Fatalf("expected the run to continue past the original 2-turn ceiling, got %q (turns=%d)", result.FinalAnswer, result.Turns)
+	}
+	if len(provider.requests) != 3 {
+		t.Fatalf("expected 3 requests (ceiling raised from 2 to 4), got %d", len(provider.requests))
+	}
+	if provider.requests[2].ReasoningEffort != "high" {
+		t.Fatalf("expected the post-escalation request to carry the target effort, got %q", provider.requests[2].ReasoningEffort)
+	}
+	tr := recorder.Finish()
+	if got := tr.Counter(trace.CounterPostureEscalations); got != 1 {
+		t.Fatalf("posture_escalations = %d, want 1", got)
+	}
+}
+
+// TestPostureEscalationAbsentWithoutProfile pins the no-regression invariant at
+// the loop level: the identical failing script without a profile ends at the
+// original ceiling with the max-turns answer.
+func TestPostureEscalationAbsentWithoutProfile(t *testing.T) {
+	registry := tools.NewRegistry()
+	registry.Register(failingProfileTool{})
+
+	toolTurn := []zeroruntime.StreamEvent{
+		{Type: zeroruntime.StreamEventToolCallStart, ToolCallID: "c1", ToolName: "flaky_probe"},
+		{Type: zeroruntime.StreamEventToolCallEnd, ToolCallID: "c1"},
+		{Type: zeroruntime.StreamEventDone},
+	}
+	provider := &mockProvider{turns: [][]zeroruntime.StreamEvent{toolTurn, toolTurn}}
+
+	result, err := Run(context.Background(), "go", provider, Options{
+		Registry: registry,
+		MaxTurns: 2,
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(provider.requests) > 3 {
+		t.Fatalf("unprofiled run must not extend the ceiling, got %d requests", len(provider.requests))
+	}
+	if !strings.Contains(result.FinalAnswer, "maximum number of turns") {
+		t.Fatalf("unprofiled run must end with the max-turns answer, got %q", result.FinalAnswer)
+	}
+}
+
+// TestExecuteToolCallClassifiesRiskWithoutSandbox verifies the executed-risk
+// stamp on the unsandboxed path: a read-only tool classifies low.
+func TestExecuteToolCallClassifiesRiskWithoutSandbox(t *testing.T) {
+	registry := tools.NewRegistry()
+	registry.Register(failingProfileTool{})
+
+	result, abortErr := executeToolCall(
+		context.Background(),
+		registry,
+		ToolCall{ID: "c1", Name: "flaky_probe"},
+		PermissionModeAuto,
+		Options{},
+	)
+	if abortErr != nil {
+		t.Fatal(abortErr)
+	}
+	if result.Risk.Level == "" {
+		t.Fatal("executed result must carry a risk classification")
+	}
+	if riskRank(result.Risk.Level) > riskRank(sandbox.RiskMedium) {
+		t.Fatalf("read-only tool classified %q, want at most medium", result.Risk.Level)
+	}
+}
+
+// TestUnknownToolResultCarriesZeroRisk verifies a call that never executed a
+// tool (unknown name, nil tool) keeps the zero risk value instead of panicking
+// or inventing a classification.
+func TestUnknownToolResultCarriesZeroRisk(t *testing.T) {
+	registry := tools.NewRegistry()
+	result, abortErr := executeToolCall(
+		context.Background(),
+		registry,
+		ToolCall{ID: "c1", Name: "no_such_tool"},
+		PermissionModeAuto,
+		Options{},
+	)
+	if abortErr != nil {
+		t.Fatal(abortErr)
+	}
+	if result.Status == tools.StatusOK {
+		t.Fatal("expected an error result for an unknown tool")
+	}
+	if result.Risk.Level != "" {
+		t.Fatalf("not-executed result must keep the zero risk value, got %q", result.Risk.Level)
+	}
+}

internal/agent/prompt_budget_test.go

@@ -0,0 +1,51 @@
+package agent
+
+import (
+	"testing"
+
+	"github.com/Gitlawb/zero/internal/tools"
+)
+
+// These ceilings are a deliberate ratchet on Zero's fixed per-turn overhead — the
+// system prompt and the eager tool schemas that ride on EVERY request. They are
+// set ~10% above the current measured cost, using ApproxTextTokens (the same
+// estimate the compaction loop and /context use, ~non-whitespace-bytes/4). A change
+// that pushes either past its ceiling must be justified (and the ceiling raised
+// deliberately) or trimmed — the per-turn floor should not creep up silently.
+//
+// Measured baselines (2026-07): base system prompt ~3160 tokens; the tools a normal
+// (auto permission-mode) interactive turn sends ~2430 tokens. The tool figure is
+// the auto-mode advertised set, not the full core registry — higher-risk tools that
+// only advertise in other modes are excluded, matching what a real turn actually
+// pays.
+const (
+	maxBaseSystemPromptTokens = 3500
+	maxEagerToolSchemaTokens  = 2700
+)
+
+func TestSystemPromptTokenBudget(t *testing.T) {
+	// Minimal render: a model is set (so the session block renders) but no Cwd, so
+	// the workspace map, project guidelines, and repo map are excluded — this is the
+	// fixed base every session pays regardless of the workspace it runs in.
+	prompt := buildSystemPrompt(Options{Model: "claude-opus-4-8"})
+	got := ApproxTextTokens(prompt)
+	t.Logf("base system prompt: %d tokens (%d bytes)", got, len(prompt))
+	if got > maxBaseSystemPromptTokens {
+		t.Fatalf("base system prompt is %d tokens, over the %d ceiling — trim it or raise the ceiling deliberately", got, maxBaseSystemPromptTokens)
+	}
+}
+
+func TestEagerToolSchemaTokenBudget(t *testing.T) {
+	registry := tools.NewRegistry()
+	for _, tool := range tools.CoreToolsScoped(t.TempDir(), nil) {
+		registry.Register(tool)
+	}
+	// Options{} keeps DeferThreshold at 0, so deferral is inactive and every core
+	// tool is exposed eagerly — exactly what a plugin-free session sends each turn.
+	exposed, _ := partitionTools(registry, PermissionModeAuto, Options{}, map[string]bool{})
+	got := estimateToolDefTokens(exposed)
+	t.Logf("eager core tool schemas: %d tokens across %d tools", got, len(exposed))
+	if got > maxEagerToolSchemaTokens {
+		t.Fatalf("eager tool schemas are %d tokens, over the %d ceiling — defer a tool or raise the ceiling deliberately", got, maxEagerToolSchemaTokens)
+	}
+}

internal/agent/prompt_fingerprint.go

@@ -0,0 +1,300 @@
+package agent
+
+import (
+	"crypto/sha256"
+	"encoding/hex"
+	"encoding/json"
+	"fmt"
+	"sort"
+	"strings"
+
+	"github.com/Gitlawb/zero/internal/zeroruntime"
+)
+
+// promptSubstrings are the seven cacheable sub-components of the prompt a run
+// sends to a model. Each is the literal content the sub-component would emit
+// into the system prompt or tool list, hashed so a downstream observer can
+// detect drift turn-over-turn. The substrings are produced by
+// buildPromptSubstrings and consumed by ComputePrefixFingerprint.
+type promptSubstrings struct {
+	baseInstructions   string
+	confirmationPolicy string
+	projectContext     string
+	skills             string
+	tools              string
+	schema             string
+}
+
+// buildPromptSubstrings assembles the seven cacheable sub-components of the
+// prompt without joining them. The system prompt sections are produced by the
+// same private builders that buildSystemPrompt uses, so the substrings are
+// byte-identical to what would have appeared in the joined prompt. The tools
+// and schema substrings are derived from the partitioned tool list (caller-
+// provided) so this helper does not need to re-run the partition.
+//
+// This is the seam ComputePrefixFingerprint reads from. It exists as a
+// separate function (rather than returning both prompt and substrings from
+// buildSystemPrompt) so existing callers of buildSystemPrompt are unaffected
+// and the substrings helper is independently testable.
+func buildPromptSubstrings(options Options, exposed []zeroruntime.ToolDefinition) promptSubstrings {
+	core := strings.TrimSpace(options.SystemPrompt)
+	if core == "" {
+		core = strings.TrimSpace(coreSystemPrompt)
+	}
+	if core == "" {
+		core = fallbackSystemPrompt
+	}
+
+	// The confirmation policy is appended unconditionally in buildSystemPrompt;
+	// the substring here is the same constant.
+	policy := strings.TrimSpace(confirmationPolicy)
+
+	// projectContext is the joined output of workspaceContext, which itself
+	// walks the AGENTS.md / ZERO.md / .zero/AGENTS.md chain. Hashing the
+	// joined string captures every guideline file's contribution and the
+	// ordering between them.
+	project := workspaceContext(options.Cwd)
+
+	// skills is the joined output of skillsContext, with each skill on its own
+	// line. The substring is the literal text the model would see.
+	skills := skillsContext(options)
+
+	// tools and schema are derived from the partitioned tool list. The list
+	// passed in is already in stable order (partitionToolsCached is the
+	// canonical stable partitioner) so concatenating in slice order is
+	// sufficient — we sort names again defensively in case a future caller
+	// hands us a list from a different partitioner.
+	toolsSubstr, schemaSubstr := toolSubstrings(exposed)
+
+	return promptSubstrings{
+		baseInstructions:   core,
+		confirmationPolicy: policy,
+		projectContext:     project,
+		skills:             skills,
+		tools:              toolsSubstr,
+		schema:             schemaSubstr,
+	}
+}
+
+// toolSubstrings extracts a stable, name-sorted digest of the partitioned tool
+// list. The tools substring is "name1\nname2\n..." (sorted) so a reordering of
+// the partition produces a different hash. The schema substring is the
+// concatenation of each tool's JSON schema in name order, so a schema change
+// is detected independently of the tool ordering.
+func toolSubstrings(exposed []zeroruntime.ToolDefinition) (toolsSubstr, schemaSubstr string) {
+	if len(exposed) == 0 {
+		return "", ""
+	}
+	names := make([]string, 0, len(exposed))
+	byName := make(map[string]zeroruntime.ToolDefinition, len(exposed))
+	for _, def := range exposed {
+		names = append(names, def.Name)
+		byName[def.Name] = def
+	}
+	sort.Strings(names)
+	var toolsSB, schemaSB strings.Builder
+	for i, name := range names {
+		if i > 0 {
+			toolsSB.WriteByte('\n')
+		}
+		toolsSB.WriteString(name)
+		def := byName[name]
+		// Parameters is rendered to a canonical JSON string per tool. The
+		// schema-render cache in internal/agent (used by partitionToolsCached)
+		// guarantees the same render is byte-identical across turns for the
+		// same tool, so this substring is a stable hash input.
+		schemaSB.WriteString(name)
+		schemaSB.WriteByte('\n')
+		schemaSB.WriteString(canonicalSchemaString(def.Parameters))
+		schemaSB.WriteByte('\n')
+	}
+	return toolsSB.String(), schemaSB.String()
+}
+
+// canonicalSchemaString renders a tool's parameter schema to a stable string.
+// tools.ToolDef.Parameters is a map[string]any in the common case; Go's
+// fmt.Sprintf("%v", m) iterates a map in random order, which would produce
+// a different hash for the same Parameters value across calls and defeat
+// the fingerprint. encoding/json marshals maps with keys sorted
+// alphabetically, so json.Marshal is the primary stable render. The
+// fallback for the rare non-JSON-compatible value (functions, channels,
+// NaN/Inf floats, cyclic references) is a stable key-sorted stringifier
+// that walks the value with sorted keys for maps and a leading
+// "__non_json:" prefix so a future schema change is visible in the trace
+// (a SchemaHash collision from the same value) rather than a silent hash
+// drift.
+//
+// Note: SchemaHash is a stability signal, not a wire-identity signal. The
+// bytes json.Marshal produces for a Go map[string]any may differ from
+// the bytes the provider actually sends on the wire (provider encoders
+// use their own JSON conventions; an Anthropic schema may serialize with
+// different whitespace, key order, or number formatting). A consumer
+// must NOT assume SchemaHash matches the provider's wire schema — only
+// that two turns with the same Go-side Parameters produce the same
+// SchemaHash. This is sufficient for the trace's purpose (drift
+// detection) but not for a content-based equality check.
+func canonicalSchemaString(params map[string]any) string {
+	if len(params) == 0 {
+		return ""
+	}
+	data, err := json.Marshal(params)
+	if err != nil {
+		return "__non_json:" + stableStringify(params)
+	}
+	return string(data)
+}
+
+// stableStringify renders v to a deterministic string. Maps have their keys
+// sorted alphabetically; slices are walked in order; primitives use their
+// natural Go format. The function is recursive but bounded by the size of
+// the input, so a cyclic reference is not reachable here (the json.Marshal
+// fallback is hit when json.Marshal itself returns an error, which for the
+// common case is a non-JSON-compatible value, not a cycle — cycles are
+// rare in tool schemas and acceptable to mis-render in the fallback path
+// since the trace's contract is "the hash is stable for the same input,"
+// not "the fallback is lossless"). Used only when json.Marshal fails.
+func stableStringify(v any) string {
+	var sb strings.Builder
+	writeStable(&sb, v)
+	return sb.String()
+}
+
+func writeStable(sb *strings.Builder, v any) {
+	switch x := v.(type) {
+	case nil:
+		sb.WriteString("null")
+	case bool:
+		if x {
+			sb.WriteString("true")
+		} else {
+			sb.WriteString("false")
+		}
+	case string:
+		sb.WriteByte('"')
+		sb.WriteString(x)
+		sb.WriteByte('"')
+	case float64:
+		// Use %g for compact, deterministic float rendering. NaN and Inf are
+		// not JSON-encodable (which is why we are in the fallback path) and
+		// render as "NaN" / "+Inf" / "-Inf" — distinct, stable strings.
+		fmt.Fprintf(sb, "%g", x)
+	case float32:
+		fmt.Fprintf(sb, "%g", float64(x))
+	case int:
+		fmt.Fprintf(sb, "%d", x)
+	case int64:
+		fmt.Fprintf(sb, "%d", x)
+	case int32:
+		fmt.Fprintf(sb, "%d", x)
+	case uint:
+		fmt.Fprintf(sb, "%d", x)
+	case uint64:
+		fmt.Fprintf(sb, "%d", x)
+	case uint32:
+		fmt.Fprintf(sb, "%d", x)
+	case map[string]any:
+		sb.WriteByte('{')
+		keys := make([]string, 0, len(x))
+		for k := range x {
+			keys = append(keys, k)
+		}
+		sort.Strings(keys)
+		for i, k := range keys {
+			if i > 0 {
+				sb.WriteByte(',')
+			}
+			sb.WriteByte('"')
+			sb.WriteString(k)
+			sb.WriteString(`":`)
+			writeStable(sb, x[k])
+		}
+		sb.WriteByte('}')
+	case []any:
+		sb.WriteByte('[')
+		for i, item := range x {
+			if i > 0 {
+				sb.WriteByte(',')
+			}
+			writeStable(sb, item)
+		}
+		sb.WriteByte(']')
+	default:
+		// Last-resort: include the type so distinct values produce distinct
+		// strings even if their default format collides. Without the type
+		// tag, fmt.Sprintf("%v", x) for two different types could produce
+		// the same bytes (rare, but the type prefix makes it impossible).
+		fmt.Fprintf(sb, "<%T:%v>", x, x)
+	}
+}
+
+// ComputePrefixFingerprint returns a trace.PrefixHash (a 7-field fingerprint
+// of the prompt prefix) for one turn of a run. The seven sub-hashes are
+// independent SHA-256s of the corresponding sub-component; the complete-prefix
+// hash is a SHA-256 of the canonical concatenation of the other six, so any
+// sub-component drift is observable both individually and in aggregate.
+//
+// I/O side effects: buildPromptSubstrings calls workspaceContext, which runs
+// git (gitBranchForPrompt, FindProjectGitRoot) and reads the AGENTS.md /
+// ZERO.md / .zero/AGENTS.md chain plus the repo map. The same workspace-
+// context work is performed separately by buildSystemPrompt a few lines
+// later in the request-build path, so on every traced turn the file and
+// git reads happen twice. The duplication is intentional for this PR (a
+// follow-up will pass substrings through) and is a known perf cost only on
+// the opt-in trace path.
+//
+// A stable CompletePrefixHash across turns means the four captured
+// sub-components (baseInstructions, confirmationPolicy, projectContext,
+// skills) are byte-identical. It does NOT rule out drift in the seven
+// uncaptured sections of buildSystemPrompt (modelPromptAddendum,
+// sessionRuntimeContext, approvedCommandPrefixContext, workspaceSeedContext,
+// userGuidelines, specialistDelegationContext, responseStyleContext). For
+// default Options the four captured substrings are the full prompt; for
+// non-default Options the seven uncaptured sections may contribute, and a
+// consumer correlating CompletePrefix stability with cached_input_tokens
+// must cross-check the model_switches counter (modelPromptAddendum changes
+// on a model switch) to disambiguate "no drift" from "drift in an
+// uncaptured section."
+
+func ComputePrefixFingerprint(options Options, exposed []zeroruntime.ToolDefinition) prefixFingerprint {
+	subs := buildPromptSubstrings(options, exposed)
+	base := sha256hex(subs.baseInstructions)
+	policy := sha256hex(subs.confirmationPolicy)
+	project := sha256hex(subs.projectContext)
+	skills := sha256hex(subs.skills)
+	toolsH := sha256hex(subs.tools)
+	schema := sha256hex(subs.schema)
+	complete := sha256hex(strings.Join([]string{
+		base, policy, project, skills, toolsH, schema,
+	}, "|"))
+	return prefixFingerprint{
+		BaseInstructionsHash:   base,
+		ConfirmationPolicyHash: policy,
+		ProjectContextHash:     project,
+		SkillsHash:             skills,
+		ToolsHash:              toolsH,
+		SchemaHash:             schema,
+		CompletePrefixHash:     complete,
+	}
+}
+
+// prefixFingerprint is the agent-side shape of a prompt-prefix fingerprint. It
+// is converted to a trace.PrefixHash at the loop boundary (see EmitPrefixHash
+// in loop.go) so the trace package does not need to import this one. The field
+// names match the trace.PrefixHash JSON tags 1:1.
+type prefixFingerprint struct {
+	BaseInstructionsHash   string
+	ConfirmationPolicyHash string
+	ProjectContextHash     string
+	SkillsHash             string
+	ToolsHash              string
+	SchemaHash             string
+	CompletePrefixHash     string
+}
+
+// sha256hex returns the hex-encoded SHA-256 of s. Empty input produces the
+// hash of the empty string, which is a constant; callers that want to
+// distinguish "absent" from "empty" should check s before calling.
+func sha256hex(s string) string {
+	sum := sha256.Sum256([]byte(s))
+	return hex.EncodeToString(sum[:])
+}

internal/agent/prompt_fingerprint_test.go

@@ -0,0 +1,243 @@
+package agent
+
+import (
+	"strings"
+	"testing"
+
+	"github.com/Gitlawb/zero/internal/zeroruntime"
+)
+
+// TestComputePrefixFingerprintStableAcrossCalls asserts the headline property
+// the trace is meant to expose: two ComputePrefixFingerprint calls over the
+// same Options and the same exposed tool list produce byte-identical hashes.
+// This is the regression catch for "did anyone introduce non-determinism into
+// the prompt prefix."
+func TestComputePrefixFingerprintStableAcrossCalls(t *testing.T) {
+	opts := Options{Cwd: t.TempDir(), SystemPrompt: "core"}
+	exposed := []zeroruntime.ToolDefinition{
+		{Name: "read_file", Parameters: map[string]any{"type": "object"}},
+		{Name: "grep", Parameters: map[string]any{"type": "object"}},
+	}
+	first := ComputePrefixFingerprint(opts, exposed)
+	second := ComputePrefixFingerprint(opts, exposed)
+	if first != second {
+		t.Fatalf("fingerprint must be stable across calls with the same input:\n  first=%+v\n second=%+v", first, second)
+	}
+	if first.CompletePrefixHash == "" {
+		t.Fatalf("CompletePrefixHash must be non-empty for a non-trivial prompt")
+	}
+}
+
+// TestComputePrefixFingerprintToolsHashIsNameSetSensitive asserts the
+// name-set sensitivity of the ToolsHash. The test name used to read as
+// "reorder changes hash" but the assertion is the opposite: the hash is
+// name-set sensitive, not name-order sensitive (toolSubstrings sorts
+// names before hashing). The partitioner is expected to produce a stable
+// order, so a name-order change is not a hash signal — a name-set change
+// is. The test asserts both: permutation of the same tools produces the
+// same hash, and adding a tool produces a different hash.
+func TestComputePrefixFingerprintToolsHashIsNameSetSensitive(t *testing.T) {
+	opts := Options{Cwd: t.TempDir(), SystemPrompt: "core"}
+	a := []zeroruntime.ToolDefinition{
+		{Name: "read_file", Parameters: map[string]any{"schema": "schema-read"}},
+		{Name: "grep", Parameters: map[string]any{"schema": "schema-grep"}},
+	}
+	b := []zeroruntime.ToolDefinition{
+		{Name: "grep", Parameters: map[string]any{"schema": "schema-grep"}},
+		{Name: "read_file", Parameters: map[string]any{"schema": "schema-read"}},
+	}
+	// toolSubstrings sorts names internally, so the ToolsHash is identical
+	// for a permutation of the same tools. We assert that here to document
+	// the contract: name-order independence, name-set sensitivity.
+	fpA := ComputePrefixFingerprint(opts, a)
+	fpB := ComputePrefixFingerprint(opts, b)
+	if fpA.ToolsHash != fpB.ToolsHash {
+		t.Fatalf("ToolsHash must be name-set sensitive, not name-order sensitive: a=%s b=%s", fpA.ToolsHash, fpB.ToolsHash)
+	}
+	// A different tool set must produce a different ToolsHash.
+	c := append([]zeroruntime.ToolDefinition{}, a...)
+	c = append(c, zeroruntime.ToolDefinition{Name: "bash", Parameters: map[string]any{"schema": "schema-bash"}})
+	fpC := ComputePrefixFingerprint(opts, c)
+	if fpA.ToolsHash == fpC.ToolsHash {
+		t.Fatalf("ToolsHash must change when the tool set changes: a=%s c=%s", fpA.ToolsHash, fpC.ToolsHash)
+	}
+}
+
+// TestComputePrefixFingerprintSchemaChangeChangesSchemaHash asserts that a
+// change to a tool's Parameters schema is visible in the SchemaHash. A schema
+// edit that should be cache-stable (e.g. field reorder, doc tweak) will move
+// the hash and surface in the trace; a schema edit that legitimately
+// invalidates the cache also moves the hash, which is the right behavior.
+func TestComputePrefixFingerprintSchemaChangeChangesSchemaHash(t *testing.T) {
+	opts := Options{Cwd: t.TempDir(), SystemPrompt: "core"}
+	a := []zeroruntime.ToolDefinition{
+		{Name: "read_file", Parameters: map[string]any{"version": "v1"}},
+	}
+	b := []zeroruntime.ToolDefinition{
+		{Name: "read_file", Parameters: map[string]any{"version": "v2"}},
+	}
+	fpA := ComputePrefixFingerprint(opts, a)
+	fpB := ComputePrefixFingerprint(opts, b)
+	if fpA.SchemaHash == fpB.SchemaHash {
+		t.Fatalf("SchemaHash must change when tool Parameters change: a=%s b=%s", fpA.SchemaHash, fpB.SchemaHash)
+	}
+}
+
+// TestComputePrefixFingerprintSystemPromptChangeChangesBaseHash asserts the
+// most common drift path: a system-prompt edit (e.g. an agent.md update, a
+// model addendum change) must move the base-instructions hash, which moves
+// the complete-prefix hash, which is what the trace is for.
+func TestComputePrefixFingerprintSystemPromptChangeChangesBaseHash(t *testing.T) {
+	optsA := Options{Cwd: t.TempDir(), SystemPrompt: "core v1"}
+	optsB := Options{Cwd: t.TempDir(), SystemPrompt: "core v2"}
+	exposed := []zeroruntime.ToolDefinition{{Name: "noop", Parameters: map[string]any{}}}
+	fpA := ComputePrefixFingerprint(optsA, exposed)
+	fpB := ComputePrefixFingerprint(optsB, exposed)
+	if fpA.BaseInstructionsHash == fpB.BaseInstructionsHash {
+		t.Fatalf("BaseInstructionsHash must change when SystemPrompt changes: a=%s b=%s", fpA.BaseInstructionsHash, fpB.BaseInstructionsHash)
+	}
+	if fpA.CompletePrefixHash == fpB.CompletePrefixHash {
+		t.Fatalf("CompletePrefixHash must change when any sub-hash changes: a=%s b=%s", fpA.CompletePrefixHash, fpB.CompletePrefixHash)
+	}
+}
+
+// TestComputePrefixFingerprintCompleteIsAggregateOfSubHashes asserts the
+// canonical-join property: CompletePrefixHash must depend on every sub-hash.
+// This is the regression catch for "did anyone reorder or drop a sub-hash
+// from the canonical join without updating CompletePrefixHash."
+func TestComputePrefixFingerprintCompleteIsAggregateOfSubHashes(t *testing.T) {
+	opts := Options{Cwd: t.TempDir(), SystemPrompt: "core"}
+	exposed := []zeroruntime.ToolDefinition{{Name: "read_file", Parameters: map[string]any{"k": "v"}}}
+	fp := ComputePrefixFingerprint(opts, exposed)
+	// Manually compute what the canonical join should be.
+	expected := sha256hex(strings.Join([]string{
+		fp.BaseInstructionsHash,
+		fp.ConfirmationPolicyHash,
+		fp.ProjectContextHash,
+		fp.SkillsHash,
+		fp.ToolsHash,
+		fp.SchemaHash,
+	}, "|"))
+	if fp.CompletePrefixHash != expected {
+		t.Fatalf("CompletePrefixHash must equal sha256hex of the canonical join of the other six:\n  got:      %s\n  expected: %s", fp.CompletePrefixHash, expected)
+	}
+}
+
+// TestBuildPromptSubstringsDefaultOptions asserts the invariants of the
+// substrings helper for default Options: which substrings are non-empty,
+// which are empty, and that the substring-to-hash round-trip is lossless.
+func TestBuildPromptSubstringsDefaultOptions(t *testing.T) {
+	opts := Options{
+		Cwd:          t.TempDir(),
+		SystemPrompt: "test core",
+	}
+	subs := buildPromptSubstrings(opts, nil)
+	// Invariants the trace depends on (default Options):
+	//   1. baseInstructions substring equals the core system prompt bytes.
+	//   2. confirmationPolicy substring is non-empty (the embedded policy
+	//      is always present, post TrimSpace).
+	//   3. skills, tools, and schema substrings are empty (no skills or
+	//      tools configured for default Options).
+	if subs.baseInstructions != "test core" {
+		t.Fatalf("baseInstructions substring must equal the core system prompt: got %q want %q", subs.baseInstructions, "test core")
+	}
+	if subs.confirmationPolicy == "" {
+		t.Fatalf("confirmationPolicy substring must be non-empty (the embedded policy is always present)")
+	}
+	if subs.skills != "" {
+		t.Fatalf("skills substring must be empty for default Options: got %q", subs.skills)
+	}
+	// Round-trip: the corresponding fingerprint hashes must equal
+	// sha256hex of the substrings, with no truncation or reformatting
+	// between the two layers.
+	fp := ComputePrefixFingerprint(opts, nil)
+	if fp.BaseInstructionsHash != sha256hex(subs.baseInstructions) {
+		t.Fatalf("BaseInstructionsHash in fingerprint must equal sha256hex of the substring")
+	}
+	if fp.ConfirmationPolicyHash != sha256hex(subs.confirmationPolicy) {
+		t.Fatalf("ConfirmationPolicyHash in fingerprint must equal sha256hex of the substring")
+	}
+}
+
+// TestCanonicalSchemaStringStableForMapParameters asserts the headline
+// property the trace needs: the same map[string]any produces the same
+// canonical string across calls, even though Go's fmt.Sprintf("%v", m)
+// iterates a map in random order. json.Marshal sorts map keys
+// alphabetically, which is the fix.
+func TestCanonicalSchemaStringStableForMapParameters(t *testing.T) {
+	params := map[string]any{
+		"description": "read a file",
+		"properties": map[string]any{
+			"path": map[string]any{"type": "string"},
+		},
+		"required": []any{"path"},
+		"type":     "object",
+	}
+	first := canonicalSchemaString(params)
+	second := canonicalSchemaString(params)
+	if first != second {
+		t.Fatalf("canonicalSchemaString must be stable across calls for the same Parameters:\n  first=%s\n second=%s", first, second)
+	}
+	if first == "" {
+		t.Fatal("canonicalSchemaString must produce a non-empty string for a non-empty map")
+	}
+	// Two maps with the same key/value pairs (different declaration
+	// order in source) must produce identical canonical strings. This
+	// is the property that defeats Go's map iteration randomization.
+	reordered := map[string]any{
+		"type":     "object",
+		"required": []any{"path"},
+		"properties": map[string]any{
+			"path": map[string]any{"type": "string"},
+		},
+		"description": "read a file",
+	}
+	if canonicalSchemaString(params) != canonicalSchemaString(reordered) {
+		t.Fatal("canonicalSchemaString must produce identical output for maps with the same key/value pairs in different declaration orders")
+	}
+}
+
+// TestCanonicalSchemaStringFallbackStableForNonJSONValue exercises the
+// non-JSON fallback in canonicalSchemaString: when Parameters contains a
+// value json.Marshal cannot encode (here, a channel), the function must
+// fall back to a key-sorted stringification that is stable across calls
+// and across map-iteration order. The previous fallback used
+// fmt.Sprintf("%v", params), which iterates the map in random order and
+// produced a different hash for the same value across calls — defeating
+// the fingerprint. The current fallback is stable by construction.
+func TestCanonicalSchemaStringFallbackStableForNonJSONValue(t *testing.T) {
+	ch := make(chan int) // channels are not JSON-encodable
+	params := map[string]any{
+		"description": "a tool with a non-JSON value",
+		"channel":     ch,
+		"z_first":     1,
+		"a_second":    2,
+	}
+	first := canonicalSchemaString(params)
+	second := canonicalSchemaString(params)
+	if first == "" {
+		t.Fatal("canonicalSchemaString must produce a non-empty string for the fallback path")
+	}
+	if first != second {
+		t.Fatalf("canonicalSchemaString fallback must be stable across calls:\n  first=%s\n second=%s", first, second)
+	}
+	// A re-ordered map (same key/value pairs, different declaration order)
+	// must produce the same fallback string. This is the property the
+	// key-sorted stringifier provides that fmt.Sprintf("%v", m) does not.
+	reordered := map[string]any{
+		"a_second":    2,
+		"z_first":     1,
+		"channel":     ch,
+		"description": "a tool with a non-JSON value",
+	}
+	if canonicalSchemaString(params) != canonicalSchemaString(reordered) {
+		t.Fatalf("canonicalSchemaString fallback must produce identical output for maps with the same key/value pairs in different declaration orders:\n  first=%s\n second=%s", first, canonicalSchemaString(reordered))
+	}
+	// The fallback must include the "__non_json:" prefix so a consumer
+	// can tell the bytes are not a JSON-marshaled schema (a hash collision
+	// between a json.Marshal result and a stableStringify result is
+	// possible in theory; the prefix makes the source observable).
+	if !strings.HasPrefix(first, "__non_json:") {
+		t.Fatalf("canonicalSchemaString fallback must start with the __non_json: prefix, got: %s", first)
+	}
+}

internal/agent/reconnect.go

@@ -3,9 +3,12 @@ package agent
 import (
 	"context"
 	"fmt"
+	"math/rand"
 	"strings"
 	"time"
 
+	"github.com/Gitlawb/zero/internal/errhint"
+	"github.com/Gitlawb/zero/internal/trace"
 	"github.com/Gitlawb/zero/internal/zeroruntime"
 )
 
@@ -17,10 +20,20 @@ import (
 // times. We retry ONLY the connect (not a partially-consumed stream), so no
 // already-forwarded OnText is ever duplicated.
 const (
-	maxStreamReconnects = 2
-	streamReconnectBase = 500 * time.Millisecond
+	// maxStreamReconnects is how many times the connect is re-issued after a
+	// transient disconnect. 4 (not 2): with jittered exponential backoff this rides
+	// out a multi-second network blip (~0.5s + 1s + 2s + 4s worst case) instead of
+	// dying on a 2s hiccup and re-burning every token on a restart.
+	maxStreamReconnects = 4
+	// streamReconnectMax caps a single backoff so the tail attempts don't wait
+	// minutes on a long outage.
+	streamReconnectMax = 8 * time.Second
 )
 
+// streamReconnectBase is the first-retry delay (doubled each subsequent attempt).
+// A var, not a const, so tests can shrink it to keep the exhaustion path fast.
+var streamReconnectBase = 500 * time.Millisecond
+
 // reconnectNotifier is called before each retry with the 1-based attempt number
 // and the max, so the caller can surface a "Reconnecting N/max" notice. Nil is
 // fine.
@@ -40,6 +53,21 @@ func reconnectNoticeFor(options Options) reconnectNotifier {
 	}
 }
 
+// stallRetryNoticeFor builds a notifier for the loop's content-stall retry (a
+// stream that connected and produced only transient output, then went silent).
+// Distinct wording from reconnectNoticeFor's "connection lost" because the
+// connection was fine — the model stalled. Surfaced through OnReasoning, the
+// non-content channel, so it never folds into the answer text. Nil when there
+// is no reasoning sink (the retry still happens silently).
+func stallRetryNoticeFor(options Options) reconnectNotifier {
+	if options.OnReasoning == nil {
+		return nil
+	}
+	return func(attempt, max int) {
+		options.OnReasoning(fmt.Sprintf("\n[no output — model stalled; retrying %d/%d…]\n", attempt, max))
+	}
+}
+
 // streamWithReconnect issues request via provider.StreamCompletion and, on a
 // transient disconnect error, retries the connect up to maxStreamReconnects
 // times with exponential backoff. It returns the live stream on success, or the
@@ -47,22 +75,32 @@ func reconnectNoticeFor(options Options) reconnectNotifier {
 // already past its deadline is returned immediately (no retry) — those have
 // their own handling (compaction for context-limit, image-rejection, etc.).
 func streamWithReconnect(ctx context.Context, provider Provider, request zeroruntime.CompletionRequest, notify reconnectNotifier) (<-chan zeroruntime.StreamEvent, error) {
+	recorder := trace.FromContext(ctx)
 	stream, err := provider.StreamCompletion(ctx, request)
 	if err == nil {
+		if recorder != nil {
+			recorder.Counter(trace.CounterModelRequests, 1)
+		}
 		return stream, nil
 	}
 	for attempt := 1; attempt <= maxStreamReconnects; attempt++ {
 		if !shouldReconnect(ctx, err) {
 			return nil, err
 		}
+		if recorder != nil {
+			recorder.Counter(trace.CounterReconnectCount, 1)
+		}
 		if notify != nil {
 			notify(attempt, maxStreamReconnects)
 		}
-		if waitErr := sleepWithContext(ctx, backoffFor(attempt)); waitErr != nil {
+		if waitErr := sleepWithContext(ctx, jitteredBackoff(attempt)); waitErr != nil {
 			return nil, err // ctx cancelled while waiting; surface the original error
 		}
 		stream, err = provider.StreamCompletion(ctx, request)
 		if err == nil {
+			if recorder != nil {
+				recorder.Counter(trace.CounterModelRequests, 1)
+			}
 			return stream, nil
 		}
 	}
@@ -81,6 +119,19 @@ func shouldReconnect(ctx context.Context, err error) bool {
 	if isContextLimitError(msg) || isImageRejectionError(err) {
 		return false
 	}
+	// HTTP 5xx statuses are non-reconnectable and must be excluded BEFORE the
+	// transport substring match below — otherwise "504 Gateway Timeout" would slip
+	// through on the generic "timeout" needle. 503 already exhausted
+	// providerio.SendWithRetry (retrying is a redundant double-retry); 500/502/504
+	// are non-idempotent by providerio's rule — the completion POST may already
+	// have reached the model before the upstream/gateway gave up, so replaying the
+	// connect risks duplicate billable work. Digit-boundary matched so an
+	// incidental "504" inside a latency/id number is not mistaken for a status.
+	if errhint.HasStatusCode(msg, "500", "502", "503", "504") {
+		return false
+	}
+	// Transport-level disconnects only. A genuine transport failure (EOF, reset,
+	// refused, timeout) means no response was received, which is safe to reconnect.
 	for _, needle := range []string{
 		"eof",
 		"connection reset",
@@ -91,8 +142,6 @@ func shouldReconnect(ctx context.Context, err error) bool {
 		"timed out",
 		"temporarily unavailable",
 		"i/o timeout",
-		"503",
-		"502",
 		"server closed",
 		"unexpected end",
 	} {
@@ -103,14 +152,31 @@ func shouldReconnect(ctx context.Context, err error) bool {
 	return false
 }
 
+// backoffFor is the deterministic exponential base delay for a 1-based attempt,
+// capped at streamReconnectMax. Jitter is layered on separately (jitteredBackoff).
 func backoffFor(attempt int) time.Duration {
 	d := streamReconnectBase
 	for i := 1; i < attempt; i++ {
+		if d >= streamReconnectMax {
+			return streamReconnectMax
+		}
 		d *= 2
 	}
+	if d > streamReconnectMax {
+		d = streamReconnectMax
+	}
 	return d
 }
 
+// jitteredBackoff adds up to 50% random jitter on top of backoffFor so concurrent
+// runs (swarm members, a cron fleet) that all trip on the same outage don't
+// reconnect in lockstep and hammer a recovering endpoint. Never shorter than the
+// deterministic base, so backoff still grows attempt over attempt.
+func jitteredBackoff(attempt int) time.Duration {
+	base := backoffFor(attempt)
+	return base + time.Duration(rand.Int63n(int64(base/2)+1))
+}
+
 func sleepWithContext(ctx context.Context, d time.Duration) error {
 	timer := time.NewTimer(d)
 	defer timer.Stop()

internal/agent/reconnect_test.go

@@ -44,6 +44,10 @@ func TestStreamWithReconnectRecoversFromTransientDisconnect(t *testing.T) {
 }
 
 func TestStreamWithReconnectGivesUpAfterMax(t *testing.T) {
+	// Shrink the backoff so exhausting all retries stays fast (real base would
+	// sleep ~7.5s across 4 attempts).
+	defer func(orig time.Duration) { streamReconnectBase = orig }(streamReconnectBase)
+	streamReconnectBase = time.Millisecond
 	// Always fails with a disconnect error → exhausts retries and returns it.
 	p := &flakyProvider{failBefore: 99, failErr: errors.New("connection reset by peer")}
 	_, err := streamWithReconnect(context.Background(), p, zeroruntime.CompletionRequest{}, nil)
@@ -87,7 +91,7 @@ func TestShouldReconnectClassification(t *testing.T) {
 	ctx := context.Background()
 	disconnects := []string{
 		"unexpected EOF", "connection reset by peer", "broken pipe",
-		"i/o timeout", "503 Service Unavailable", "server closed the connection",
+		"i/o timeout", "server closed the connection", "connection refused",
 	}
 	for _, m := range disconnects {
 		if !shouldReconnect(ctx, errors.New(m)) {
@@ -97,6 +101,13 @@ func TestShouldReconnectClassification(t *testing.T) {
 	notDisconnects := []string{
 		"context length exceeded", "invalid api key", "model not found",
 		"400 bad request: unsupported parameter",
+		// HTTP 5xx statuses are handled by providerio.SendWithRetry (503) or are
+		// non-idempotent (500/502/504); the reconnect path must not double-retry
+		// them. "504 Gateway Timeout" must NOT slip through on the generic "timeout"
+		// transport signal — a gateway timeout is non-idempotent, since the upstream
+		// may have already processed the completion POST before giving up.
+		"503 Service Unavailable", "502 Bad Gateway",
+		"504 Gateway Timeout", "500 Internal Server Error",
 	}
 	for _, m := range notDisconnects {
 		if shouldReconnect(ctx, errors.New(m)) {
@@ -112,6 +123,24 @@ func TestBackoffGrows(t *testing.T) {
 	if backoffFor(2) != 2*streamReconnectBase {
 		t.Fatalf("attempt 2 backoff = %v, want %v", backoffFor(2), 2*streamReconnectBase)
 	}
+	// The exponential base is capped so late attempts don't wait minutes.
+	if got := backoffFor(20); got != streamReconnectMax {
+		t.Fatalf("attempt 20 backoff = %v, want cap %v", got, streamReconnectMax)
+	}
+}
+
+func TestJitteredBackoffStaysInBounds(t *testing.T) {
+	// Jitter never drops below the deterministic base and never exceeds base*1.5,
+	// so backoff still grows attempt over attempt while decorrelating retries.
+	for attempt := 1; attempt <= 5; attempt++ {
+		base := backoffFor(attempt)
+		for i := 0; i < 200; i++ {
+			got := jitteredBackoff(attempt)
+			if got < base || got > base+base/2 {
+				t.Fatalf("attempt %d jittered backoff %v out of [%v, %v]", attempt, got, base, base+base/2)
+			}
+		}
+	}
 }
 
 func TestReconnectNoticeRoutesThroughReasoning(t *testing.T) {

internal/agent/request_permissions_test.go

@@ -15,7 +15,7 @@ import (
 
 func TestRequestPermissionsTurnGrantAllowsLaterToolAndCleansUp(t *testing.T) {
 	workspace := t.TempDir()
-	outside := t.TempDir()
+	outside := tempDirOutsideDefaultTemp(t)
 	target := filepath.Join(outside, "vegetables.txt")
 	scope, err := sandbox.NewScope(workspace, nil)
 	if err != nil {
@@ -120,6 +120,71 @@ func TestRequestPermissionsTurnGrantAllowsLaterToolAndCleansUp(t *testing.T) {
 	}
 }
 
+func tempDirOutsideDefaultTemp(t *testing.T) string {
+	t.Helper()
+	dir, err := os.MkdirTemp(".", ".zero-sandbox-outside-")
+	if err != nil {
+		t.Fatalf("MkdirTemp outside default temp: %v", err)
+	}
+	t.Cleanup(func() { _ = os.RemoveAll(dir) })
+	abs, err := filepath.Abs(dir)
+	if err != nil {
+		t.Fatalf("Abs(%q): %v", dir, err)
+	}
+	if resolved, err := filepath.EvalSymlinks(abs); err == nil {
+		return resolved
+	}
+	return filepath.Clean(abs)
+}
+
+func TestInlineAdditionalPermissionsGrantAllowsNetworkCommand(t *testing.T) {
+	workspace := t.TempDir()
+	engine := sandbox.NewEngine(sandbox.EngineOptions{
+		WorkspaceRoot: workspace,
+		Policy:        sandbox.DefaultPolicy(),
+	})
+	args := map[string]any{
+		"sandbox_permissions": string(tools.SandboxPermissionsWithAdditionalPermissions),
+		"additional_permissions": map[string]any{
+			"network": map[string]any{"enabled": true},
+		},
+	}
+	cleanup, err := grantInlineAdditionalPermissions(args, sandbox.PermissionGrantScopeTurn, Options{
+		Sandbox: engine,
+		Cwd:     workspace,
+	})
+	if err != nil {
+		t.Fatalf("grantInlineAdditionalPermissions: %v", err)
+	}
+	defer cleanup()
+
+	decision := engine.Evaluate(context.Background(), sandbox.Request{
+		WorkspaceRoot:     workspace,
+		ToolName:          "exec_command",
+		SideEffect:        sandbox.SideEffectShell,
+		Permission:        sandbox.PermissionPrompt,
+		PermissionGranted: true,
+		PermissionMode:    sandbox.PermissionModeAsk,
+		Args:              map[string]any{"cmd": "python3 -m http.server 8080 --bind 127.0.0.1"},
+	})
+	if decision.Action != sandbox.ActionAllow {
+		t.Fatalf("inline network grant should allow shell network command, got %#v", decision)
+	}
+}
+
+func TestInlineAdditionalPermissionsRequiresPromptUnderAutoAllow(t *testing.T) {
+	args := map[string]any{
+		"sandbox_permissions": string(tools.SandboxPermissionsWithAdditionalPermissions),
+		"additional_permissions": map[string]any{
+			"network": map[string]any{"enabled": true},
+		},
+	}
+	decision := &sandbox.Decision{Action: sandbox.ActionAllow, AutoAllowed: true}
+	if !shouldRequestPermission(tools.NewBashTool(t.TempDir()), args, false, decision) {
+		t.Fatal("inline additional permissions must request approval even when a sandboxed shell would otherwise auto-allow")
+	}
+}
+
 func TestRequestPermissionsDenyReturnsEmptyGrantAndContinues(t *testing.T) {
 	workspace := t.TempDir()
 	registry := tools.NewRegistry()

internal/agent/selfcorrect.go

@@ -9,6 +9,7 @@ import (
 
 	"github.com/Gitlawb/zero/internal/lsp"
 	"github.com/Gitlawb/zero/internal/redaction"
+	"github.com/Gitlawb/zero/internal/trace"
 	"github.com/Gitlawb/zero/internal/verify"
 )
 
@@ -133,7 +134,9 @@ func (sc *SelfCorrector) inspect(ctx context.Context, changedFiles []string) Cor
 	}
 
 	if sc.cfg.IncludeTests && sc.verifier != nil {
+		verifySpan := trace.FromContext(ctx).Span(trace.SpanVerification)
 		vr, err := sc.verifier.Verify(ctx)
+		verifySpan.End()
 		if err != nil {
 			// "No plan / no tests" is not an error (DetectPlan returns an empty
 			// plan), so a non-nil error means verification could not run at all —

internal/agent/skills_context_test.go

@@ -0,0 +1,88 @@
+package agent
+
+import (
+	"strconv"
+	"strings"
+	"testing"
+)
+
+func TestSkillsContextCapsLongList(t *testing.T) {
+	// 60 skills with realistic trigger-rich descriptions (~180 chars each, under
+	// the 200-rune truncation) total well past the 4096-byte list budget, so the
+	// overflow summary must kick in while the block stays bounded.
+	longDesc := strings.Repeat("use when the request touches deployments, release notes, or verification; ", 2) + "number "
+	skills := make([]SkillInfo, 0, 60)
+	for i := 0; i < 60; i++ {
+		n := strconv.Itoa(i)
+		skills = append(skills, SkillInfo{Name: "skill-" + n, Description: longDesc + n})
+	}
+	got := skillsContext(Options{Skills: skills})
+	if len(got) > skillsContextListBudget+1000 {
+		t.Fatalf("skills block should stay bounded near the %d-byte budget, got %d bytes:\n%s", skillsContextListBudget, len(got), got)
+	}
+	if !strings.Contains(got, "more (call skill") {
+		t.Fatalf("expected an overflow summary line, got:\n%s", got)
+	}
+	// The first skill is always listed regardless of budget.
+	if !strings.Contains(got, "- skill-0:") {
+		t.Fatalf("expected the first skill to always be listed, got:\n%s", got)
+	}
+}
+
+// A realistic mid-size skill set (20 skills, trigger-rich descriptions) must be
+// listed in FULL — no overflow summary. This pins the fix for the old 640-byte
+// budget, under which skills past ~#6 were invisible to the model and therefore
+// never triggered.
+func TestSkillsContextListsRealisticSetInFull(t *testing.T) {
+	desc := "Use when the user asks about deployments, release notes, or pre-merge verification runs."
+	skills := make([]SkillInfo, 0, 20)
+	for i := 0; i < 20; i++ {
+		skills = append(skills, SkillInfo{Name: "skill-" + strconv.Itoa(i), Description: desc})
+	}
+	got := skillsContext(Options{Skills: skills})
+	if strings.Contains(got, "more (call skill") {
+		t.Fatalf("20 described skills must all be listed without overflow, got:\n%s", got)
+	}
+	for i := 0; i < 20; i++ {
+		if !strings.Contains(got, "- skill-"+strconv.Itoa(i)+":") {
+			t.Fatalf("skill-%d missing from the list:\n%s", i, got)
+		}
+	}
+}
+
+func TestSkillsContext(t *testing.T) {
+	if got := skillsContext(Options{}); got != "" {
+		t.Fatalf("no skills should yield an empty section, got %q", got)
+	}
+	got := skillsContext(Options{Skills: []SkillInfo{
+		{Name: "commit-writer", Description: "Write a conventional-commit message."},
+		{Name: "  ", Description: "nameless, should be skipped"},
+		{Name: "reviewer"},
+	}})
+	if !strings.Contains(got, "<available_skills>") || !strings.Contains(got, "</available_skills>") {
+		t.Fatalf("missing available_skills block: %q", got)
+	}
+	if !strings.Contains(got, "- commit-writer: Write a conventional-commit message.") {
+		t.Fatalf("missing commit-writer line: %q", got)
+	}
+	if !strings.Contains(got, "- reviewer\n") {
+		t.Fatalf("reviewer (no description) line missing: %q", got)
+	}
+	if strings.Contains(got, "nameless") {
+		t.Fatalf("nameless entry should be skipped: %q", got)
+	}
+}
+
+func TestSystemPromptIncludesSkillsOnlyWhenInstalled(t *testing.T) {
+	with := buildSystemPrompt(Options{Skills: []SkillInfo{
+		{Name: "commit-writer", Description: "Write a commit message."},
+	}})
+	if !strings.Contains(with, "<available_skills>") || !strings.Contains(with, "skill tool") {
+		t.Fatalf("expected available_skills guidance in system prompt: %q", with)
+	}
+	// Default (no skills) must reproduce the prior prompt: no skills block.
+	without := buildSystemPrompt(Options{})
+	if strings.Contains(without, "<available_skills>") {
+		t.Fatalf("available_skills block must not appear without skills")
+	}
+}

internal/agent/stall_retry_test.go

@@ -0,0 +1,168 @@
+package agent
+
+import (
+	"context"
+	"sync/atomic"
+	"testing"
+
+	"github.com/Gitlawb/zero/internal/tools"
+	"github.com/Gitlawb/zero/internal/zeroruntime"
+)
+
+// stallProvider connects successfully (HTTP 200) but the stream emits a stall/idle
+// timeout error for the first stallBefore calls, then succeeds with "done". The
+// pre-stall emission simulates different partial-output shapes:
+//   - partialText: forwarded PROSE (StreamEventText) — must NOT be retried
+//     (re-issuing would duplicate visible answer text).
+//   - reasoningText: transient reasoning preview — IS retried.
+//   - partialToolCall: a tool call started (+ arg delta) but never ended, the
+//     "froze mid-write_file" case — IS retried (the incomplete call is never
+//     executed or committed).
+type stallProvider struct {
+	calls           int32
+	stallBefore     int32
+	partialText     string
+	reasoningText   string
+	partialToolCall string
+}
+
+func (p *stallProvider) StreamCompletion(_ context.Context, _ zeroruntime.CompletionRequest) (<-chan zeroruntime.StreamEvent, error) {
+	n := atomic.AddInt32(&p.calls, 1)
+	ch := make(chan zeroruntime.StreamEvent, 5)
+	if n <= p.stallBefore {
+		if p.partialText != "" {
+			ch <- zeroruntime.StreamEvent{Type: zeroruntime.StreamEventText, Content: p.partialText}
+		}
+		if p.reasoningText != "" {
+			ch <- zeroruntime.StreamEvent{Type: zeroruntime.StreamEventReasoning, Content: p.reasoningText}
+		}
+		if p.partialToolCall != "" {
+			// A tool call that starts and streams a partial argument fragment but
+			// never gets StreamEventToolCallEnd, then the stream errors — exactly
+			// the gpt-5.x "began the big write_file then froze" shape.
+			ch <- zeroruntime.StreamEvent{Type: zeroruntime.StreamEventToolCallStart, ToolCallID: "tc_1", ToolName: p.partialToolCall}
+			ch <- zeroruntime.StreamEvent{Type: zeroruntime.StreamEventToolCallDelta, ToolCallID: "tc_1", ArgumentsFragment: `{"path":"x.html","content":"<!doctype`}
+		}
+		ch <- zeroruntime.StreamEvent{
+			Type:  zeroruntime.StreamEventError,
+			Error: "provider stream error: no output for 6m (the model produced nothing)",
+		}
+		close(ch)
+		return ch, nil
+	}
+	ch <- zeroruntime.StreamEvent{Type: zeroruntime.StreamEventText, Content: "done"}
+	ch <- zeroruntime.StreamEvent{Type: zeroruntime.StreamEventDone}
+	close(ch)
+	return ch, nil
+}
+
+// A no-output stall is re-issued on a fresh connection and recovers — this is the
+// macOS stale-pooled-connection hang turned into an automatic recovery.
+func TestRunRetriesStreamStallWithNoOutput(t *testing.T) {
+	p := &stallProvider{stallBefore: 1}
+	result, err := Run(context.Background(), "go", p, Options{Registry: tools.NewRegistry()})
+	if err != nil {
+		t.Fatalf("a no-output stall should retry to success, got %v", err)
+	}
+	if result.FinalAnswer != "done" {
+		t.Fatalf("final answer = %q, want %q", result.FinalAnswer, "done")
+	}
+	if got := atomic.LoadInt32(&p.calls); got != 2 {
+		t.Fatalf("want 2 calls (1 stall + 1 retry), got %d", got)
+	}
+}
+
+// A stall AFTER partial output must NOT be retried — re-issuing would duplicate
+// the already-streamed text. It surfaces the error instead.
+func TestRunDoesNotRetryStallAfterPartialOutput(t *testing.T) {
+	p := &stallProvider{stallBefore: 1, partialText: "partial"}
+	_, err := Run(context.Background(), "go", p, Options{Registry: tools.NewRegistry()})
+	if err == nil {
+		t.Fatal("a stall after partial output must NOT be retried; want an error")
+	}
+	if got := atomic.LoadInt32(&p.calls); got != 1 {
+		t.Fatalf("partial-then-stall must not retry, got %d calls", got)
+	}
+}
+
+// A stall after only transient REASONING was forwarded (no prose text, no
+// completed tool call) IS retried: reasoning is a "thinking" preview, not
+// committed answer text, so re-streaming it on a clearly-signalled retry is
+// acceptable and recovers the common "reasoned then froze" stall. Only
+// forwarded final prose (OnText) blocks the retry (collected.Text is empty
+// here, so this exercises the transient-vs-prose gate specifically).
+func TestRunRetriesStallAfterOnlyReasoning(t *testing.T) {
+	p := &stallProvider{stallBefore: 1, reasoningText: "thinking hard"}
+	result, err := Run(context.Background(), "go", p, Options{
+		Registry:    tools.NewRegistry(),
+		OnReasoning: func(string) {},
+	})
+	if err != nil {
+		t.Fatalf("a stall after only reasoning should retry to success, got %v", err)
+	}
+	if result.FinalAnswer != "done" {
+		t.Fatalf("final answer = %q, want %q", result.FinalAnswer, "done")
+	}
+	if got := atomic.LoadInt32(&p.calls); got != 2 {
+		t.Fatalf("want 2 calls (1 stall + 1 retry), got %d", got)
+	}
+}
+
+// A stall mid-way through an INCOMPLETE tool call (started + partial args, never
+// ended) IS retried — the exact gpt-5.x / ollama "began the big write_file then
+// froze" case. The incomplete call is never executed (a turn returns on the
+// stall error before dispatch) and never committed to history, so re-issuing is
+// safe and recovers to success. This is the primary scenario this change fixes.
+func TestRunRetriesStallAfterIncompleteToolCall(t *testing.T) {
+	starts := 0
+	p := &stallProvider{stallBefore: 1, reasoningText: "planning", partialToolCall: "write_file"}
+	result, err := Run(context.Background(), "go", p, Options{
+		Registry:        tools.NewRegistry(),
+		OnReasoning:     func(string) {},
+		OnToolCallStart: func(string, string) { starts++ },
+		OnToolCallDelta: func(string, string) {},
+	})
+	if err != nil {
+		t.Fatalf("a stall mid-incomplete-tool-call should retry to success, got %v", err)
+	}
+	if result.FinalAnswer != "done" {
+		t.Fatalf("final answer = %q, want %q", result.FinalAnswer, "done")
+	}
+	if got := atomic.LoadInt32(&p.calls); got != 2 {
+		t.Fatalf("want 2 calls (1 stall + 1 retry), got %d", got)
+	}
+	if starts != 1 {
+		t.Fatalf("the incomplete tool call should have been forwarded once before the stall, got %d starts", starts)
+	}
+}
+
+// A persistent stall surfaces an error after exhausting the capped retries.
+func TestRunGivesUpAfterMaxStallRetries(t *testing.T) {
+	p := &stallProvider{stallBefore: 99}
+	_, err := Run(context.Background(), "go", p, Options{Registry: tools.NewRegistry()})
+	if err == nil {
+		t.Fatal("a persistent stall must surface an error after exhausting retries")
+	}
+	if got := atomic.LoadInt32(&p.calls); got != int32(1+maxStreamStallRetries) {
+		t.Fatalf("want %d calls (1 + %d retries), got %d", 1+maxStreamStallRetries, maxStreamStallRetries, got)
+	}
+}
+
+func TestIsStreamTimeoutError(t *testing.T) {
+	timeouts := []string{
+		"provider stream error: no output for 10m (the model produced nothing)",
+		"provider stream error: idle timeout after 5m0s (upstream stopped sending data)",
+		"stream stalled (upstream kept the connection alive but produced no output)",
+	}
+	for _, m := range timeouts {
+		if !isStreamTimeoutError(m) {
+			t.Fatalf("want timeout-classified: %q", m)
+		}
+	}
+	notTimeouts := []string{"", "context length exceeded", "rate limit error: slow down", "model not found"}
+	for _, m := range notTimeouts {
+		if isStreamTimeoutError(m) {
+			t.Fatalf("must NOT classify as timeout: %q", m)
+		}
+	}
+}

internal/agent/system_prompt.go

@@ -5,9 +5,11 @@ import (
 	"os"
 	"path/filepath"
 	"runtime"
+	"strconv"
 	"strings"
 	"unicode/utf8"
 
+	"github.com/Gitlawb/zero/internal/config"
 	"github.com/Gitlawb/zero/internal/repomap"
 	"github.com/Gitlawb/zero/internal/workspaceseed"
 )
@@ -37,6 +39,14 @@ const fallbackSystemPrompt = "You are Zero, a terminal coding agent. Help with t
 // general-to-specific order.
 var projectContextFiles = []string{"AGENTS.md", "ZERO.md", ".zero/AGENTS.md"}
 
+// userContextFile is the per-user instruction file, resolved under
+// config.UserConfigDir()/zero/ alongside the rest of Zero's per-user config
+// (config.json, commands, specialists) so users can keep personal guidance out
+// of individual repositories.
+const userContextFile = "ZERO.md"
+
+var userConfigDirForPrompt = config.UserConfigDir
+
 // maxProjectContextBytes caps how much of a single project doc is injected so
 // a large guidelines file can't blow the context budget.
 const maxProjectContextBytes = 8 << 10 // 8 KiB per file
@@ -80,12 +90,22 @@ func buildSystemPrompt(options Options) string {
 	if seed := workspaceSeedContext(options.Cwd); seed != "" {
 		sections = append(sections, seed)
 	}
+	// User guidelines are injected before workspace/project guidelines so the
+	// project's AGENTS.md/ZERO.md is the later, more specific instruction
+	// block. See userGuidelines for the explicit precedence note carried in
+	// the section text itself.
+	if user := userGuidelines(); user != "" {
+		sections = append(sections, user)
+	}
 	if ws := workspaceContext(options.Cwd); ws != "" {
 		sections = append(sections, ws)
 	}
 	if delegation := specialistDelegationContext(options); delegation != "" {
 		sections = append(sections, delegation)
 	}
+	if skillsBlock := skillsContext(options); skillsBlock != "" {
+		sections = append(sections, skillsBlock)
+	}
 	if style := responseStyleContext(options); style != "" {
 		sections = append(sections, style)
 	}
@@ -164,6 +184,75 @@ func specialistDelegationContext(options Options) string {
 	return b.String()
 }
 
+// skillsContext lists the reusable skills the model can pull in on demand via the
+// skill tool, so it invokes the right one on the first try instead of guessing a
+// name, failing, and reading the error. It mirrors specialistDelegationContext:
+// names + one-line descriptions only (never skill bodies), and it renders nothing
+// when no skills are installed, so a skill-less run reproduces the previous prompt
+// byte-for-byte.
+// skillsContextListBudget bounds the bytes spent listing individual skills so a
+// workspace with an extreme number of them can't bloat every turn's prompt.
+// Skills past the budget are summarized as a count rather than dropped — the
+// model can still load any of them by name (an unknown name returns the full
+// list from the skill tool).
+//
+// The budget is deliberately generous — roughly 18 skills with maximally long
+// (200-rune, ASCII) descriptions, or 40+ with typical terser ones; non-ASCII
+// descriptions consume it faster since the budget counts bytes. This list is
+// the model's ONLY discovery surface, so a skill squeezed out of it effectively
+// never triggers. The previous 640-byte cap silently broke invocation for
+// anyone with more than ~6 skills — the model cannot match a request against a
+// skill it cannot see.
+const skillsContextListBudget = 4096
+
+func skillsContext(options Options) string {
+	if len(options.Skills) == 0 {
+		return ""
+	}
+	var b strings.Builder
+	b.WriteString("<available_skills>\n")
+	b.WriteString("Reusable, on-demand instruction sets you can load with the skill tool. Before acting on a request, scan this list; when the request matches a skill's name or description, call skill with its exact name FIRST and follow the loaded guidance — do not guess names, do not skip a matching skill, and do not substitute your own approach for its instructions.\n")
+	listed, spent, omitted := 0, 0, 0
+	for _, info := range options.Skills {
+		name := strings.TrimSpace(info.Name)
+		if name == "" {
+			continue
+		}
+		line := "- " + name
+		if desc := strings.TrimSpace(info.Description); desc != "" {
+			line += ": " + truncateForSkillLine(desc)
+		}
+		line += "\n"
+		// Always list at least one skill; past the budget, summarize the remainder
+		// as a count instead of silently dropping it.
+		if listed > 0 && spent+len(line) > skillsContextListBudget {
+			omitted++
+			continue
+		}
+		b.WriteString(line)
+		spent += len(line)
+		listed++
+	}
+	if omitted > 0 {
+		b.WriteString("- …and " + strconv.Itoa(omitted) + " more (call skill with a name; an unknown name lists them all)\n")
+	}
+	b.WriteString("</available_skills>")
+	return b.String()
+}
+
+// truncateForSkillLine keeps a skill's one-line description short so a single
+// verbose description can't dominate the skills-list budget. The cap is roomy on
+// purpose: the description carries the skill's trigger conditions ("use when…"),
+// and truncating those is what stops the model from ever invoking the skill.
+func truncateForSkillLine(desc string) string {
+	const maxDescRunes = 200
+	runes := []rune(desc)
+	if len(runes) <= maxDescRunes {
+		return desc
+	}
+	return strings.TrimSpace(string(runes[:maxDescRunes])) + "…"
+}
+
 func sessionRuntimeContext(options Options) string {
 	provider := strings.TrimSpace(options.ProviderName)
 	model := strings.TrimSpace(options.Model)
@@ -197,7 +286,7 @@ func workspaceContext(cwd string) string {
 	b.WriteString("Working directory: " + cwd + "\n")
 	b.WriteString("Operating system: " + runtime.GOOS + "\n")
 	if runtime.GOOS == "windows" {
-		b.WriteString("Shell syntax: Windows cmd.exe syntax for exec_command/bash tools; prefer the workdir/cwd argument instead of cd when changing directories.\n")
+		b.WriteString("Shell syntax: Windows cmd.exe syntax for exec_command/bash tools. To put | & > < etc inside an arg value, use double quotes around the value, not single quotes (single quotes do not protect metachars in cmd.exe): gh --jq \".a | b\", go test -run \"A|B\". Do not pipe to or invoke POSIX coreutils from Git for Windows (usr\\bin head/grep/tail/cat/...): they are MSYS binaries and fail under the write-restricted sandbox; use native Zero tools (grep, read_file, list_directory, glob) or cmd.exe findstr/more instead, or sandbox_permissions require_escalated only when host-level execution is truly required. Prefer the workdir/cwd argument over cd when changing directories.\n")
 	} else {
 		b.WriteString("Shell syntax: /bin/sh syntax for exec_command/bash tools; prefer the workdir/cwd argument instead of cd when changing directories.\n")
 	}
@@ -206,7 +295,7 @@ func workspaceContext(cwd string) string {
 	}
 	b.WriteString("</environment>")
 
-	b.WriteString(projectGuidelines(cwd, findProjectGitRoot(cwd)))
+	b.WriteString(projectGuidelines(cwd, FindProjectGitRoot(cwd)))
 	if repoMap := repoMapContext(cwd); repoMap != "" {
 		b.WriteString("\n\n## Repo map\n\n" + repoMap)
 	}
@@ -250,13 +339,7 @@ func projectGuidelines(cwd, gitRoot string) string {
 		if remaining := maxProjectContextTotalBytes - totalUsed; remaining < limit {
 			limit = remaining
 		}
-		if len(content) > limit {
-			cut := limit
-			for cut > 0 && !utf8.RuneStart(content[cut]) {
-				cut--
-			}
-			content = content[:cut] + "\n… (truncated)"
-		}
+		content = truncateGuidelineContent(content, limit)
 		label := projectGuidelineLabel(match, gitRoot)
 		b.WriteString("\n\n## Project guidelines (" + label + ")\n\n" + content)
 		totalUsed += len(content)
@@ -264,6 +347,92 @@ func projectGuidelines(cwd, gitRoot string) string {
 	return b.String()
 }
 
+// userGuidelines returns the per-user ZERO.md instructions block, if present.
+// The file lives in config.UserConfigDir()/zero/ next to Zero's other
+// per-user config; the basename match is case-insensitive so a file saved as
+// zero.md still resolves on case-sensitive filesystems, mirroring the project
+// guideline loader. The section carries an explicit precedence note because
+// this is a global, personal preferences file: it is injected earlier in the
+// prompt than the project's AGENTS.md/ZERO.md (see buildSystemPrompt), and
+// the note keeps that precedence unambiguous even if a model weighs later
+// context more heavily than section order alone implies.
+func userGuidelines() string {
+	configDir, err := userConfigDirForPrompt()
+	if err != nil {
+		return ""
+	}
+	configDir = strings.TrimSpace(configDir)
+	if configDir == "" {
+		return ""
+	}
+	match := findCaseInsensitiveFile(filepath.Join(configDir, "zero"), userContextFile)
+	if match == "" {
+		return ""
+	}
+	data, err := os.ReadFile(match)
+	if err != nil {
+		return ""
+	}
+	content := strings.TrimSpace(string(data))
+	if content == "" {
+		return ""
+	}
+	content = truncateGuidelineContent(content, maxProjectContextBytes)
+	return "## User guidelines (" + filepath.Base(match) + ")\n\n" +
+		"These are the operator's personal preferences, not project policy. " +
+		"Where they conflict with a repository's project guidelines below (AGENTS.md/ZERO.md), the project guidelines take precedence.\n\n" +
+		content
+}
+
+// truncationMarker is appended to guideline content that was cut short.
+const truncationMarker = "\n… (truncated)"
+
+// truncateGuidelineContent caps content at limit bytes without splitting a
+// UTF-8 rune, appending a truncation marker when anything was cut. Space for
+// the marker is reserved before choosing the cut point. When limit is too
+// small to fit the marker itself, the marker is dropped instead, so the
+// returned string never exceeds limit for any non-negative limit.
+func truncateGuidelineContent(content string, limit int) string {
+	if len(content) <= limit {
+		return content
+	}
+	if limit <= 0 {
+		return ""
+	}
+	cut := limit - len(truncationMarker)
+	if cut < 0 {
+		cut = 0
+	}
+	for cut > 0 && !utf8.RuneStart(content[cut]) {
+		cut--
+	}
+	if truncated := content[:cut] + truncationMarker; len(truncated) <= limit {
+		return truncated
+	}
+	// limit is smaller than the marker itself: fall back to a hard,
+	// rune-safe cut at limit bytes with no marker rather than exceeding it.
+	end := limit
+	for end > 0 && !utf8.RuneStart(content[end]) {
+		end--
+	}
+	return content[:end]
+}
+
+// findCaseInsensitiveFile returns the on-disk path of the regular file in dir
+// whose name matches basename case-insensitively, or "" when absent.
+func findCaseInsensitiveFile(dir, basename string) string {
+	entries, err := os.ReadDir(dir)
+	if err != nil {
+		return ""
+	}
+	for _, e := range entries {
+		if !e.IsDir() && strings.EqualFold(e.Name(), basename) {
+			return filepath.Join(dir, e.Name())
+		}
+	}
+	return ""
+}
+
 // projectGuidelineDirs returns the directory chain from gitRoot to cwd
 // (inclusive), in root-to-leaf order. If gitRoot is empty or unreachable
 // from cwd, the chain collapses to [cwd].
@@ -310,29 +479,22 @@ func projectGuidelineLabel(match, gitRoot string) string {
 
 // findProjectContextFile returns the first matching project guideline file
 // in dir. Lookup is case-insensitive on the basename and on the actual
-// filename returned, so a git-tracked file like AGENTS.MD (uppercase MD)
-// still resolves to its true filename on case-sensitive filesystems — the
-// returned path is what should appear in the project guidelines label.
+// filename returned, so a git-tracked file like AGENTS.md still resolves to
+// its true filename on case-sensitive filesystems — the returned path is what
+// should appear in the project guidelines label.
 // Returns "" when nothing matches.
 func findProjectContextFile(dir string) string {
 	for _, name := range projectContextFiles {
-		baseLower := strings.ToLower(filepath.Base(name))
 		// Walk to the file's parent through dir with case-insensitive segment
-		// matching, then find a regular file with the same lowercased
-		// basename. This works on both case-sensitive and case-insensitive
-		// filesystems and always returns the on-disk filename.
+		// matching, then find a regular file with the same basename. This works
+		// on both case-sensitive and case-insensitive filesystems and always
+		// returns the on-disk filename.
 		parent, ok := resolveDirCaseInsensitive(filepath.Dir(filepath.Join(dir, name)), dir)
 		if !ok {
 			continue
 		}
-		entries, err := os.ReadDir(parent)
-		if err != nil {
-			continue
-		}
-		for _, e := range entries {
-			if !e.IsDir() && strings.ToLower(e.Name()) == baseLower {
-				return filepath.Join(parent, e.Name())
-			}
+		if match := findCaseInsensitiveFile(parent, filepath.Base(name)); match != "" {
+			return match
 		}
 	}
 	return ""
@@ -380,17 +542,17 @@ func resolveDirCaseInsensitive(target, anchor string) (string, bool) {
 	return cur, true
 }
 
-// findProjectGitRoot returns the nearest ancestor of cwd that contains a
+// FindProjectGitRoot returns the nearest ancestor of cwd that contains a
 // .git entry (file or directory). Returns "" when no git root is found, so
 // the caller can fall back to cwd-only lookup.
-func findProjectGitRoot(cwd string) string {
+func FindProjectGitRoot(cwd string) string {
 	cwd = strings.TrimSpace(cwd)
 	if cwd == "" {
 		return ""
 	}
 	cur := cwd
 	for {
-		if hasGitMetadata(cur) {
+		if HasGitMetadata(cur) {
 			return cur
 		}
 		parent := filepath.Dir(cur)
@@ -401,7 +563,7 @@ func findProjectGitRoot(cwd string) string {
 	}
 }
 
-func hasGitMetadata(dir string) bool {
+func HasGitMetadata(dir string) bool {
 	gitPath := filepath.Join(dir, ".git")
 	info, err := os.Stat(gitPath)
 	if err != nil {
@@ -457,7 +619,11 @@ func repoMapContext(cwd string) string {
 // cwd, handling both a regular checkout (.git dir) and a worktree (.git file).
 // Returns "" on any problem — the prompt simply omits the branch segment.
 func gitBranchForPrompt(cwd string) string {
-	gitPath := filepath.Join(cwd, ".git")
+	gitRoot := FindProjectGitRoot(cwd)
+	if gitRoot == "" {
+		return ""
+	}
+	gitPath := filepath.Join(gitRoot, ".git")
 	info, err := os.Stat(gitPath)
 	if err != nil {
 		return ""
@@ -473,10 +639,10 @@ func gitBranchForPrompt(cwd string) string {
 			return ""
 		}
 		// In worktree mode the gitdir is often RELATIVE (e.g.
-		// "gitdir: ../.git/worktrees/<name>") — resolve it against cwd, not the
+		// "gitdir: ../.git/worktrees/<name>") — resolve it against the worktree root (gitRoot), not the
 		// process working directory, or HEAD lookup fails and we drop the branch.
 		if !filepath.IsAbs(dir) {
-			dir = filepath.Join(cwd, dir)
+			dir = filepath.Join(gitRoot, dir)
 		}
 		headPath = filepath.Join(dir, "HEAD")
 	}

internal/agent/system_prompt.md

@@ -16,7 +16,12 @@ work.
   implementation, verification, and a clear summary.
 - Only stop to ask the user (via the ask_user tool) when you are genuinely
   blocked on a decision that is theirs to make and that you cannot resolve from
-  the code, the request, or sensible defaults.
+  the code, the request, or sensible defaults. When the answer is likely one of a
+  small set, include 2-4 suggested `options` and mark the best as `recommended`
+  (it must be one of the options) so the user can pick quickly; give each option a
+  short `optionDescriptions` line when a one-word label needs context, and a short
+  `header` (2-3 words) per question as its tab label when you ask several. Omit
+  options for genuinely open-ended questions.
 
 ## Workflow: plan then act
 
@@ -26,15 +31,25 @@ work.
    tests, or config before you modify behavior. Never edit a file you have not
    read.
 2. **Plan.** For multi-step work, call update_plan with an ordered checklist and
-   keep it current as you go. Skip the plan for trivial one-step tasks.
+   keep it live. The plan bar is the user's progress signal — call update_plan
+   after EACH concrete unit of work (every file written, every command run), not
+   just at coarse milestones: mark the finished item completed and the next one
+   in_progress before you start it. A plan stuck at 0/N while files are landing is
+   a bug, not economy — these calls are cheap, expected, and the update_plan cards
+   are hidden from the transcript, so frequent updates cost the user nothing and
+   never clutter the conversation. Keep at most one item in_progress, and never
+   batch the updates to the end of the turn. Skip the plan for trivial one-step
+   tasks.
 3. **Implement.** Make focused changes that match the surrounding code's style,
    naming, and conventions. Prefer the smallest change that fully solves the
    problem. Avoid broad refactors, unrelated rewrites, dependency churn, and
    formatting-only edits unless the user asked for them.
 4. **Verify.** Verify after edits; see the testing gate below. This is
    mandatory.
-5. **Summarize.** Give a concise final response: what changed, why, which files
-   matter, and what validation ran. State plainly what was not verified.
+5. **Summarize.** Close with a summary scaled to the work: a trivial fix earns one
+   line; a substantial change earns a few short bullets covering what changed, where
+   it lives, and what you did and did not verify. Lead with the outcome, and never
+   pad — do not restate the task or narrate steps the user already watched you take.
 
 ## Editing discipline
 
@@ -45,7 +60,13 @@ work.
 - Make one tool call per file. Do not batch multi-file writes into a single
   shell or script invocation.
 - For edits to existing files, prefer edit_file or apply_patch with minimal,
-  targeted diffs. Match the existing indentation, imports, and idioms.
+  targeted diffs. Match the existing indentation, imports, and idioms. Match the
+  file's comment density: do not add explanatory comments unless the user asks or
+  the code is already comment-dense.
+- Solve the problem as posed, not a more general version of it. Add no
+  speculative abstraction, configurability, or handling for cases that cannot
+  occur, and nothing the user did not ask for. A small diff can still be
+  over-built; if a 200-line solution could be 50, rewrite it.
 - Preserve behavior you were not asked to change. Do not delete or rewrite code
   you did not author unless the task requires it; if you must, say so.
 
@@ -63,8 +84,13 @@ work.
 
 ## Tool use
 
-- Use tools to act, not to narrate. Do not announce each call; just do the work
-  and explain the outcome.
+- Lead a multi-step task with a one- or two-sentence plain-language preamble on
+  your approach, so the user can follow what you're about to do. Then keep a brief
+  running account: drop a short, single-line note before each SIGNIFICANT step
+  (e.g. "Now the stylesheet.", "Let me wire up the cart and filters.") and explain
+  the outcome. Use tools to act, not to narrate — don't announce every individual
+  call or read; narrate only the steps a person would care about, and skip
+  narration entirely for trivial one-step tasks.
 - Run independent, read-only lookups together when you can, rather than one at a
   time, to move faster.
 - exec_command is for commands that have no native tool (build, test, git,