fix(agent): branch on a dedicated-summarizer flag instead of comparing providers

Review follow-ups for #532:

- summarizeProvider now reports whether the dedicated summarizer was
  selected and summarizeClosure branches on that flag; the previous
  `chosen == provider` compared Provider interface values, which panics
  when the dynamic type is not comparable. Regression test uses a provider
  with a slice field for both the summarizer and the main provider.
- CompactionModelID's curated-default test covers the official OpenAI
  profile alongside Anthropic and Google.

Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_012JqvAf7zsP1qYFVZayKXHq

.gitattributes

@@ -0,0 +1,3 @@
+*.go text eol=lf
+*.mod text eol=lf
+*.sum text eol=lf

.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

@@ -30,6 +30,36 @@ jobs:
           go-version-file: go.mod
           cache: true
 
+      - name: Install Make on Windows
+        if: matrix.os == 'windows-latest'
+        shell: pwsh
+        run: choco install make --no-progress -y
+
+      - name: Validate quality targets with stock macOS Make
+        if: matrix.os == 'macos-latest'
+        env:
+          GIT_CONFIG_COUNT: 1
+          GIT_CONFIG_KEY_0: grep.patternType
+          GIT_CONFIG_VALUE_0: fixed
+        run: make -n vulncheck
+
+      - name: Check vulnerabilities through native Windows Make
+        if: matrix.os == 'windows-latest'
+        shell: cmd
+        run: make vulncheck
+
+      - name: Check dead code through native Windows Make
+        if: matrix.os == 'windows-latest'
+        continue-on-error: true
+        shell: cmd
+        run: make deadcode
+
+      - name: Run static lint through native Windows Make
+        if: matrix.os == 'windows-latest'
+        continue-on-error: true
+        shell: cmd
+        run: make lint-static
+
       - name: Check formatting
         if: matrix.os == 'ubuntu-latest'
         shell: bash
@@ -100,23 +130,25 @@ 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.
-      - name: Pin toolchain from go.mod
-        run: |
-          toolchain="$(awk '/^toolchain /{print $2}' go.mod)"
-          echo "GOTOOLCHAIN=${toolchain:-auto}" >> "$GITHUB_ENV"
-
       # Hard gate: fails the build when code reaches a known vulnerability. A stdlib
       # CVE is cleared by a toolchain bump (see go.mod). May also flag a newly
       # published advisory on an unrelated PR — intentional: do not ship known vulns.
       - name: govulncheck
-        run: go run golang.org/x/vuln/cmd/[email protected] ./...
+        run: make vulncheck
 
       # Advisory: reports functions unreachable from any cmd/* main so dormant
       # code is visible in CI. Non-blocking while the dormant subsystems are
       # still being wired or removed.
       - name: deadcode (advisory)
         continue-on-error: true
-        run: go run golang.org/x/tools/cmd/[email protected] -test=false ./...
+        run: make deadcode
+
+      # 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: make lint-static

.github/workflows/release-artifacts.yml

@@ -54,8 +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")"
@@ -78,13 +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: ${{ github.token }}
@@ -92,3 +96,147 @@ jobs:
           tag="${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 @@
+{".":"0.8.0"}

AGENTS.md

@@ -0,0 +1,152 @@
+# 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**: `make lint-static`.
+7. **Security**: `make vulncheck`.
+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.
+
+## 4. Common Review Blockers
+
+These classes drive multi-round reviews. Fix them before requesting review:
+
+- **Fresh base:** Rebase onto the current PR base (`main` or stacked target)
+  before review. A stale head that rolls back mainline commits is a hard
+  blocker, not a merge-time detail. Resolve conflicts by keeping upstream
+  behavior unless the PR intentionally changes it.
+- **Platform truth:** Code and tests must pass on Linux, macOS, and Windows.
+  Compare paths after canonicalization (physical/`EvalSymlinks` form); never
+  assert raw `t.TempDir()` spellings (`/var` vs `/private/var` on macOS, short
+  paths on Windows). Prefer one cross-platform function with small `GOOS`
+  checks over duplicated helpers that drift.
+- **Security edges:** Keep secrets out of argv, env dumps, and logs. Redact
+  success and error paths (including stderr). Fail closed on ownership, lease,
+  and permission checks. Do not rely on pre-open path resolution
+  (`EvalSymlinks` then open) for containment: that is a check-to-use race.
+  Bind containment at open/use time with rooted or handle-relative,
+  traversal-resistant APIs. If a no-follow API is used, apply it to every
+  traversed component and enforce the platform's reparse-point protections;
+  final-component-only no-follow is insufficient. On multi-step setup, roll back
+  only what this run created; never destroy pre-existing resources you did
+  not create; never report success when cleanup or unlock failed.
+- **Normalize before matching:** Any transform that REMOVES bytes without
+  leaving a gap is also a reassembler, so it has to run before whatever matches
+  on the result. Redacting by shape and then stripping control bytes lets a
+  credential split by a NUL, an ESC or a C1 byte pass the patterns as two
+  fragments and be rejoined on the way out; comparing a path against a form
+  produced by the same resolver the kernel just used makes a redirect agree with
+  itself. Strip, decode, or canonicalize first, then match. The unsplit value
+  passing is not evidence, so the test needs a split case.
+- **Atomic shared state:** Write a complete temporary file, then atomically
+  replace the destination so concurrent readers never see a partial write.
+  Exclusive create or a write lock alone is not enough for readers. Serialize
+  the full read-modify-write sequence for lockfiles and shared stores.
+- **Tests match the claim:** Every behavior or security-boundary change needs a
+  regression test, including the failure path. Path-sensitive logic needs at
+  least one non-Linux path case (or a hermetic fake that exercises the same
+  normalization).
+- **Honest scope:** PR description, help text, and comments must match what
+  shipped. Wire advertised entry points or shrink the claim. Do not bundle
+  unrelated fixes. Do not widen security allowlists by name alone; check
+  classification and side effects too.
+- **Host tool version floors:** Zero shells out to `git` and other host tools,
+  and the project does not declare a minimum version for any of them. A flag,
+  subcommand, or config key that is not long-established therefore raises that
+  floor silently, and the failure on an older host is an unhelpful usage error
+  rather than a clear message. Before adding one, check when it landed and say
+  so where a caller would look. Record the floor for the exact subcommand you
+  call, not for the flag in general: `--end-of-options` reached Git's shared
+  parser in 2.24, but `git rev-parse` only accepted it from 2.30, and
+  `git checkout` and `git reset` later still. Likewise `--show-current` is
+  `git branch` 2.22, and `branch.autoSetupMerge=inherit` is 2.35. Then either
+  keep a fallback for older versions or gate the call. A test that depends on a
+  newer feature must skip below its floor instead of failing, and say which
+  version it needs.
+- **Prove the test fails without the change:** A regression test earns its
+  place only if it fails on the unfixed code, for the reason it claims. Run it
+  both ways and quote the failure in the PR. The recurring miss is not a
+  missing test but a test that never reaches the behavior it names, because an
+  earlier guard rejects the input first, so it passes with and without the fix.
+  When a test targets a specific layer, call that layer directly rather than
+  relying on a public entry point that may refuse earlier.
+- **Hermetic tests:** Tests must not read or write the developer's real config,
+  cache, or state directories. Redirect every root the code under test
+  resolves, not just config. `os.UserConfigDir` and `os.UserCacheDir` read
+  `XDG_CONFIG_HOME` and `XDG_CACHE_HOME` only on Linux and the BSDs. On macOS
+  they derive from `HOME`, and on Windows from `%AppData%` and `%LocalAppData%`.
+  Setting the XDG variables alone still leaves a macOS run writing to the real
+  home, so set `HOME` as well, or take the directory as a parameter. Apply it to
+  every test that reaches the same storage, not only the one where it was first
+  needed; a test that passes on a clean machine and pollutes a real one is a
+  blocker.
+- **Error codes vary by platform and by flag combination:** A check written
+  against the code one target returns can be silently inert on another while
+  still looking correct. `openat(O_DIRECTORY|O_NOFOLLOW)` on a symlink may
+  report `ENOTDIR` or `ELOOP` depending on the platform, because the kernel can
+  reject the directory mismatch before it ever resolves the link; accept both
+  rather than assuming the one POSIX names. On Windows, `\\?\` and `\\.\`
+  prefixes are not UNC paths, and an `NtCreateFile` information class must match
+  the struct actually passed. Confirm the code each supported target returns
+  instead of reasoning from the man page.

CHANGELOG.md

@@ -6,12 +6,331 @@ 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`.
 
+## [0.8.0](https://github.com/Gitlawb/zero/compare/v0.7.0...v0.8.0) (2026-08-21)
+
+
+### Features
+
+* **security:** path-jail primitive, credential-store locking, worktree git hardening ([#891](https://github.com/Gitlawb/zero/issues/891)) ([060d38c](https://github.com/Gitlawb/zero/commit/060d38ce716feb7c5b6657b082cd35cb8dda3d75))
+* **update:** recognise Homebrew installs and add mise, go install docs ([#910](https://github.com/Gitlawb/zero/issues/910)) ([01c3498](https://github.com/Gitlawb/zero/commit/01c3498205e69d26e5f929225f9455b276444f39))
+* use Exa as default search provider ([#926](https://github.com/Gitlawb/zero/issues/926)) ([1ec7219](https://github.com/Gitlawb/zero/commit/1ec7219aee7b2840481ab0905f131644e27d2a4e))
+* **zerogit:** auto-create a conventional branch before push/pr on default branch ([#671](https://github.com/Gitlawb/zero/issues/671)) ([d065467](https://github.com/Gitlawb/zero/commit/d065467cd3ec3ed91c5dc60554c36fc4b4d6de4c))
+
+
+### Bug Fixes
+
+* **acp:** three defects a desktop ACP client hits ([#915](https://github.com/Gitlawb/zero/issues/915)) ([a78c6c3](https://github.com/Gitlawb/zero/commit/a78c6c357e1d21146d7cd2e8030a754e06f7f1f5))
+* bump Go to 1.26.6 for stdlib vulnerability fixes ([#903](https://github.com/Gitlawb/zero/issues/903)) ([dc15e82](https://github.com/Gitlawb/zero/commit/dc15e822987d8d7de467c62c375518f435b74d8c))
+* **daemon:** clean up child after startup timeout ([#774](https://github.com/Gitlawb/zero/issues/774)) ([9c29209](https://github.com/Gitlawb/zero/commit/9c292092a26c9092fd873d072b0af1b0d9d39ef0))
+* **dictation:** redact API keys from streaming transcriber errors ([#852](https://github.com/Gitlawb/zero/issues/852)) ([64a783b](https://github.com/Gitlawb/zero/commit/64a783b08caa2c7b21f188033b9ed84a14e00e84))
+* **modelregistry:** expose reasoning effort for DeepSeek V4 models ([#931](https://github.com/Gitlawb/zero/issues/931)) ([90dcfd1](https://github.com/Gitlawb/zero/commit/90dcfd127e8a6d9902ec9f4e72f6d03fef1a0fc6))
+* **providers:** discover ChatGPT capabilities ([#890](https://github.com/Gitlawb/zero/issues/890)) ([2d2450e](https://github.com/Gitlawb/zero/commit/2d2450e9a744349f0d01b1d4e9ba29c24ba5650d))
+* **sandbox:** normalize launcher names before the command-prefix denylist ([#934](https://github.com/Gitlawb/zero/issues/934)) ([6edf9a8](https://github.com/Gitlawb/zero/commit/6edf9a8b78dc030dc44598919db1c7fa9d4f809a))
+
+## [0.7.0](https://github.com/Gitlawb/zero/compare/v0.6.0...v0.7.0) (2026-08-10)
+
+
+### Features
+
+* **agent:** add PermissionModePlan for interactive read-only planning ([#853](https://github.com/Gitlawb/zero/issues/853)) ([0176ee4](https://github.com/Gitlawb/zero/commit/0176ee49ffddcff6f7edf6ae2ade4967e5b22cc3))
+* **providers:** live model lists for OpenRouter and OpenGateway ([#860](https://github.com/Gitlawb/zero/issues/860)) ([fb4f293](https://github.com/Gitlawb/zero/commit/fb4f293c20eb860ca388a10d22fd566ce282ba1f))
+* **tools:** give tool results an image channel and add view_image ([#843](https://github.com/Gitlawb/zero/issues/843)) ([cd0eb19](https://github.com/Gitlawb/zero/commit/cd0eb194e4df4b844785796c0701216740a56344))
+* **tui:** replace /retitle with local /rename ([#826](https://github.com/Gitlawb/zero/issues/826)) ([5886354](https://github.com/Gitlawb/zero/commit/588635487f22ef4d62d350140a219b7b8bb24119))
+
+
+### Bug Fixes
+
+* **agent:** end the additional_permissions retry loop ([#864](https://github.com/Gitlawb/zero/issues/864)) ([4ca4fa7](https://github.com/Gitlawb/zero/commit/4ca4fa75d31b139dbe55a3badf17da86892ab001))
+* **agent:** tell the user a prefix approval leaves the sandbox ([#885](https://github.com/Gitlawb/zero/issues/885)) ([2ec5c3a](https://github.com/Gitlawb/zero/commit/2ec5c3ae6b8fb543887bdeafdb34c4da057b8d36))
+* **cli:** report MCP servers that failed to start ([#822](https://github.com/Gitlawb/zero/issues/822)) ([#827](https://github.com/Gitlawb/zero/issues/827)) ([021281e](https://github.com/Gitlawb/zero/commit/021281ebbab3f2bbe6a77b2dab4bf121bcfe6b52))
+* **lsp:** dispatch notifications off read loop ([#759](https://github.com/Gitlawb/zero/issues/759)) ([936aa4e](https://github.com/Gitlawb/zero/commit/936aa4e77cf49738560a89395b8ba4d22263a396))
+* **mcp:** name the non-text blocks a tool result drops ([#874](https://github.com/Gitlawb/zero/issues/874)) ([027a670](https://github.com/Gitlawb/zero/commit/027a67080bea511b876c284bc911ef011eceb308))
+* **sandbox:** deny reads of Zero credential stores ([#681](https://github.com/Gitlawb/zero/issues/681)) ([55fa05f](https://github.com/Gitlawb/zero/commit/55fa05f0ab5a01b495345d43464606cd19a31b4f))
+* **sandbox:** stop the Windows write jail honouring Everyone-granted paths ([#865](https://github.com/Gitlawb/zero/issues/865)) ([91b413c](https://github.com/Gitlawb/zero/commit/91b413c514cc1607d5b4e7936780d40569239814))
+* **secrets,worktrees:** fix secret redaction leakage and prune stale worktrees ([#855](https://github.com/Gitlawb/zero/issues/855)) ([ff62f73](https://github.com/Gitlawb/zero/commit/ff62f73b7f4dc5e2472db21acba43cf6140c06d1))
+* **specialist:** make overwrites atomic ([#757](https://github.com/Gitlawb/zero/issues/757)) ([30f6d8f](https://github.com/Gitlawb/zero/commit/30f6d8f591a0b81638b4182ff56c86c8a50a5c70))
+* suppress gosec G204 warning for trusted local config commands ([#848](https://github.com/Gitlawb/zero/issues/848)) ([c52db05](https://github.com/Gitlawb/zero/commit/c52db055b0b213758a33c5c5627a411e052f8de4))
+* **tui:** a copied NUL byte no longer panics the TUI on Windows ([#876](https://github.com/Gitlawb/zero/issues/876)) ([cae0269](https://github.com/Gitlawb/zero/commit/cae0269e77e7a0a493cf799b41c653c9ff9a19de)), closes [#875](https://github.com/Gitlawb/zero/issues/875)
+* **tui:** exclude permission waits from turn time ([#820](https://github.com/Gitlawb/zero/issues/820)) ([15eb16f](https://github.com/Gitlawb/zero/commit/15eb16f736f903c94203207032f98aacbcf06318))
+* **tui:** fall back to CellMotion where AllMotion is unreliable ([#872](https://github.com/Gitlawb/zero/issues/872)) ([7fbd864](https://github.com/Gitlawb/zero/commit/7fbd864e91433441489902571ea394ebc5b7bd0d))
+* **tui:** fill alt-screen frame with theme panel background ([#850](https://github.com/Gitlawb/zero/issues/850)) ([020c28e](https://github.com/Gitlawb/zero/commit/020c28eaf95e2f38853f99fc06a13cd1568bf3c9))
+* **tui:** hide unavailable sidebar shortcut ([#821](https://github.com/Gitlawb/zero/issues/821)) ([845df6c](https://github.com/Gitlawb/zero/commit/845df6c7d28e02f475e90acfe838733eb37e6aa2))
+* use OAuth credentials in provider health checks ([#828](https://github.com/Gitlawb/zero/issues/828)) ([d37de92](https://github.com/Gitlawb/zero/commit/d37de9214b1ba86542f4f5ef3b060bb96849763b))
+
+
+### Performance Improvements
+
+* compact common test output ([#846](https://github.com/Gitlawb/zero/issues/846)) ([edf660a](https://github.com/Gitlawb/zero/commit/edf660ab85131f90a452cee51dc2c3a5cf62d632))
+* improve harness context efficiency ([#838](https://github.com/Gitlawb/zero/issues/838)) ([8e26679](https://github.com/Gitlawb/zero/commit/8e266797a8305e689806eadcf9698da2796481ab))
+
+## [0.6.0](https://github.com/Gitlawb/zero/compare/v0.5.0...v0.6.0) (2026-07-29)
+
+
+### Features
+
+* add persistent session goals ([#803](https://github.com/Gitlawb/zero/issues/803)) ([5d1869e](https://github.com/Gitlawb/zero/commit/5d1869e7bc193c97e07418b57b823905c1151d5d))
+* **providers:** add Fireworks AI provider preset ([#772](https://github.com/Gitlawb/zero/issues/772)) ([58daea7](https://github.com/Gitlawb/zero/commit/58daea7a482927e5ab35afaf97761c60bb35e113))
+* **skills:** report hash drift in skill info ([#795](https://github.com/Gitlawb/zero/issues/795)) ([9fbdc34](https://github.com/Gitlawb/zero/commit/9fbdc341551601c13434cf7b9b6f5e4ea4c483b1))
+
+
+### Bug Fixes
+
+* **acp:** interrupt idle reads on cancellation ([#782](https://github.com/Gitlawb/zero/issues/782)) ([15b7a54](https://github.com/Gitlawb/zero/commit/15b7a5475cfeefddbc59634551451a8872b996b7))
+* **cli:** include plugins info in shell completions ([#794](https://github.com/Gitlawb/zero/issues/794)) ([fa8734f](https://github.com/Gitlawb/zero/commit/fa8734f17927de8832dfe041e043cfb86f93d12a))
+* **config:** kill provider-command process tree via job object on Windows ([#690](https://github.com/Gitlawb/zero/issues/690)) ([5c58256](https://github.com/Gitlawb/zero/commit/5c58256abc29c185b449f8ce79647eb3fae61882))
+* **config:** make the provider-command timeout an upper bound ([#811](https://github.com/Gitlawb/zero/issues/811)) ([#813](https://github.com/Gitlawb/zero/issues/813)) ([1c38e8c](https://github.com/Gitlawb/zero/commit/1c38e8cba50db255f3cda8b7cf4298712b2933e1))
+* **config:** stop billing provider-command fixtures to the 5s timeout ([#810](https://github.com/Gitlawb/zero/issues/810)) ([f5fa4ff](https://github.com/Gitlawb/zero/commit/f5fa4ffc64537897aa78a56edca36d8f1dacd021))
+* **cron:** reserve job IDs atomically ([#686](https://github.com/Gitlawb/zero/issues/686)) ([d9b882e](https://github.com/Gitlawb/zero/commit/d9b882eb0a94f38de8447cb54b57860ccd8ed053))
+* **lint:** resolve static analysis baseline ([#769](https://github.com/Gitlawb/zero/issues/769)) ([583e653](https://github.com/Gitlawb/zero/commit/583e6530bd849a8efcb2b88bbe8c297e8592e62c))
+* **providers:** retry provably pre-send transport failures ([#447](https://github.com/Gitlawb/zero/issues/447)) ([#750](https://github.com/Gitlawb/zero/issues/750)) ([8acd42a](https://github.com/Gitlawb/zero/commit/8acd42a3bd743a35807d762f23a33fb61c171839))
+* **sandbox:** add platform-specific write lock around grant state file ([#752](https://github.com/Gitlawb/zero/issues/752)) ([#755](https://github.com/Gitlawb/zero/issues/755)) ([88fb7b6](https://github.com/Gitlawb/zero/commit/88fb7b6192c1ac8173089967d5e9119f92d7a101))
+* **sandbox:** clarify scoped permission prompts ([#798](https://github.com/Gitlawb/zero/issues/798)) ([9fe02ae](https://github.com/Gitlawb/zero/commit/9fe02ae52e3234cb9dae76e0c5594a287ccbb52a))
+* **sandbox:** deny reads of git's credential stores ([#815](https://github.com/Gitlawb/zero/issues/815)) ([#816](https://github.com/Gitlawb/zero/issues/816)) ([33c94ec](https://github.com/Gitlawb/zero/commit/33c94ec7a3e7987e6e36b5fd5e037ff4e6e84c26))
+* **sandbox:** preserve user config and retry denied writes ([#801](https://github.com/Gitlawb/zero/issues/801)) ([81a5e6c](https://github.com/Gitlawb/zero/commit/81a5e6cb716ba4b7ecf7ff3ff261dfa51fd01098))
+* **sandbox:** scrub the daemon token file pointer ([#677](https://github.com/Gitlawb/zero/issues/677)) ([#818](https://github.com/Gitlawb/zero/issues/818)) ([3e7692d](https://github.com/Gitlawb/zero/commit/3e7692d8f0885d346809869935c81a9cf055d89e))
+* **swarm:** close lifecycle admission on shutdown ([#776](https://github.com/Gitlawb/zero/issues/776)) ([bdbda89](https://github.com/Gitlawb/zero/commit/bdbda89f3e42d93e8058ab3e1e9ef402659530cd))
+* **tui:** fix rendering corruption over multipass + Windows Terminal ([#709](https://github.com/Gitlawb/zero/issues/709)) ([5bb1bc1](https://github.com/Gitlawb/zero/commit/5bb1bc11c3188dcea7709ff07c8b25dda3541de8))
+* **tui:** paste clipboard images with Ctrl+V ([#534](https://github.com/Gitlawb/zero/issues/534)) ([#807](https://github.com/Gitlawb/zero/issues/807)) ([ec3b822](https://github.com/Gitlawb/zero/commit/ec3b8225050eb62f0f41a9582ed2881d2f9dcc8e))
+* **tui:** reset quiet clock for new runs ([#799](https://github.com/Gitlawb/zero/issues/799)) ([a4a805a](https://github.com/Gitlawb/zero/commit/a4a805a6c82c51cb255e909bd87ca183c8805cf2))
+* **windows:** use PowerShell for command execution ([#804](https://github.com/Gitlawb/zero/issues/804)) ([518d2e5](https://github.com/Gitlawb/zero/commit/518d2e53ba3fa073090a9b41c278face85b59b64))
+
+## [0.5.0](https://github.com/Gitlawb/zero/compare/v0.4.0...v0.5.0) (2026-07-22)
+
+
+### Features
+
+* **plugins:** add zero plugins info command ([#773](https://github.com/Gitlawb/zero/issues/773)) ([2479884](https://github.com/Gitlawb/zero/commit/2479884dea49580c853a53536c3bbab22ce8a2bf))
+* **sandbox:** disable the sandbox via config ([#687](https://github.com/Gitlawb/zero/issues/687)) ([#746](https://github.com/Gitlawb/zero/issues/746)) ([a21a052](https://github.com/Gitlawb/zero/commit/a21a052a4a32ce9cb3c93c94350cd24e138532cd))
+* **sandbox:** unify command execution and enforcement ([#781](https://github.com/Gitlawb/zero/issues/781)) ([96859c9](https://github.com/Gitlawb/zero/commit/96859c9bd16f6dad4e332efc7e68178b9116118a))
+* **tui:** add /undo as an alias for /rewind ([#698](https://github.com/Gitlawb/zero/issues/698)) ([#747](https://github.com/Gitlawb/zero/issues/747)) ([8c6d302](https://github.com/Gitlawb/zero/commit/8c6d3022bd7801eaad6a0440bb21a27221887297))
+* **tui:** add isolated /btw conversations ([#748](https://github.com/Gitlawb/zero/issues/748)) ([2e267bd](https://github.com/Gitlawb/zero/commit/2e267bdbee4a77813e93a7ab51f0c575b66cee8c))
+* **tui:** permission prompt takes free-text feedback inline ([#780](https://github.com/Gitlawb/zero/issues/780)) ([3967d49](https://github.com/Gitlawb/zero/commit/3967d49d64998341af8ef6d7523fc63aeb2a5a7a))
+
+
+### Bug Fixes
+
+* **agent:** stop write_stdin session_id probing thrash ([#702](https://github.com/Gitlawb/zero/issues/702)) ([#749](https://github.com/Gitlawb/zero/issues/749)) ([fbf8598](https://github.com/Gitlawb/zero/commit/fbf85984f679058125951fe2d5e4f200e09a3e2f))
+* **cli:** warn when ZERO_PROVIDER overrides a providers-use selection ([#767](https://github.com/Gitlawb/zero/issues/767)) ([3524f79](https://github.com/Gitlawb/zero/commit/3524f795e5fdbb827166a84f03351cedfc9eba30))
+* **doctor:** detect missing native binary during runtime checks ([#450](https://github.com/Gitlawb/zero/issues/450)) ([7796022](https://github.com/Gitlawb/zero/commit/77960229b839dca856616f846839aa773f2923f7))
+* **lsp:** make the real-gopls check opt-in so a broken gopls can't fail the suite ([#684](https://github.com/Gitlawb/zero/issues/684)) ([#766](https://github.com/Gitlawb/zero/issues/766)) ([b1f4173](https://github.com/Gitlawb/zero/commit/b1f41735a7f6b7928a1875e986a4dca39d106cb9))
+* make extension installs transactional ([#762](https://github.com/Gitlawb/zero/issues/762)) ([baa4be1](https://github.com/Gitlawb/zero/commit/baa4be13ac5321da4e9f53e864dd1cd395481200))
+* **oauth:** refuse redirects on credential POSTs ([#729](https://github.com/Gitlawb/zero/issues/729)) ([#741](https://github.com/Gitlawb/zero/issues/741)) ([974fc03](https://github.com/Gitlawb/zero/commit/974fc036c2f9a194722f2e8fddbb4fdbf797effe))
+* **oauth:** validate discovered endpoints before merge/use ([#511](https://github.com/Gitlawb/zero/issues/511)) ([#739](https://github.com/Gitlawb/zero/issues/739)) ([ce4a996](https://github.com/Gitlawb/zero/commit/ce4a996ffac4482e704f0fd61b3e442398fb2401))
+* **perfbench:** absolutize the bench binary and make errored tasks first-class ([#730](https://github.com/Gitlawb/zero/issues/730)) ([dbd9443](https://github.com/Gitlawb/zero/commit/dbd94430143df6754d68551d1028ad8f15b82f1b))
+* **perfbench:** grant write tools so mutating tasks measure real edits ([#763](https://github.com/Gitlawb/zero/issues/763)) ([e1975c1](https://github.com/Gitlawb/zero/commit/e1975c1b396b3236ed648870230b4f85863b1d03))
+* **perfbench:** keep the stamped answer file out of negative oracle greps ([#737](https://github.com/Gitlawb/zero/issues/737)) ([015452c](https://github.com/Gitlawb/zero/commit/015452c1c98a39eabb021324182094e188a8bd47))
+* **providers:** stop "provider not found" for env-derived profiles ([#716](https://github.com/Gitlawb/zero/issues/716)) ([4cbd144](https://github.com/Gitlawb/zero/commit/4cbd144d11e5cb67bc5fea46f5b294562dac7a1a))
+* **sandbox:** AST second opinion for interactive-command bypasses ([#473](https://github.com/Gitlawb/zero/issues/473)) ([#745](https://github.com/Gitlawb/zero/issues/745)) ([f079b90](https://github.com/Gitlawb/zero/commit/f079b90f82dac7b7ae0864b279dc9094e31a6627))
+* **sandbox:** bind Windows elevated ACL setup to one no-follow handle ([#765](https://github.com/Gitlawb/zero/issues/765)) ([4945684](https://github.com/Gitlawb/zero/commit/4945684fa26aa5994eda59dcabedc817423c535d))
+* **sandbox:** don't auto-allow shell when re-entrancy skips wrapping ([#727](https://github.com/Gitlawb/zero/issues/727)) ([#744](https://github.com/Gitlawb/zero/issues/744)) ([6849011](https://github.com/Gitlawb/zero/commit/684901165d1b7f50a8bb4af31b1c1951e6926a79))
+* **tools:** give write_stdin's invalid-session errors the same recovery guidance ([#749](https://github.com/Gitlawb/zero/issues/749) follow-up) ([#768](https://github.com/Gitlawb/zero/issues/768)) ([da9fb50](https://github.com/Gitlawb/zero/commit/da9fb50f549f6701eb136fc99f6178d0772d9334))
+* **tools:** read_file recovers a backwards line range instead of erroring ([#779](https://github.com/Gitlawb/zero/issues/779)) ([89bdc67](https://github.com/Gitlawb/zero/commit/89bdc6719a1e1b3a3ef0e36b91a7839b3efdfba9))
+* **tui:** cache settled alt-screen transcript ([#647](https://github.com/Gitlawb/zero/issues/647)) ([d74ceb1](https://github.com/Gitlawb/zero/commit/d74ceb11271ed68a21c19248210e098f411805fb))
+* **tui:** model rows labelled by id when the description is prose; keep the sidebar under the / palette ([#775](https://github.com/Gitlawb/zero/issues/775)) ([b30c397](https://github.com/Gitlawb/zero/commit/b30c3971b40b06ba52ad649ba9dc0ad560d4be4c))
+* **tui:** stop the permission card clashing on cool themes ([#778](https://github.com/Gitlawb/zero/issues/778)) ([722bb31](https://github.com/Gitlawb/zero/commit/722bb3121682d9cd9cd4bc6c127e9014d719262a))
+
+
+### Performance Improvements
+
+* **agent:** concurrent read-only tool batches via capability gate ([#715](https://github.com/Gitlawb/zero/issues/715)) ([31d45d5](https://github.com/Gitlawb/zero/commit/31d45d5f14e915acb9946e8b8eb81632c48f126a))
+* **agent:** execution profiles with one-shot posture escalation (PR10b+PR10c) ([#740](https://github.com/Gitlawb/zero/issues/740)) ([378d538](https://github.com/Gitlawb/zero/commit/378d538e240c289e57821e9f9628f76034419d01))
+* **agent:** posture-escalation signals and controller (PR10a) ([#736](https://github.com/Gitlawb/zero/issues/736)) ([af875df](https://github.com/Gitlawb/zero/commit/af875df58775484304fc65586bd6c74552ad01a2))
+* **agent:** preserve prompt cache prefixes ([#760](https://github.com/Gitlawb/zero/issues/760)) ([739a47e](https://github.com/Gitlawb/zero/commit/739a47e3eac92c3decc8734f52a4d99c7480c3ca))
+* **openai:** optimized turn session — background prewarm and prefix telemetry (PR8) ([#723](https://github.com/Gitlawb/zero/issues/723)) ([60dc84e](https://github.com/Gitlawb/zero/commit/60dc84e7a38c5544ebc047f3cfaf4625dd1e83b5))
+* **output:** add token-aware semantic output budgeting (PR11) ([#717](https://github.com/Gitlawb/zero/issues/717)) ([e5670c4](https://github.com/Gitlawb/zero/commit/e5670c427ff39c628fb0822fd5c9317ee2174583))
+* **providers:** provider capabilities and default turn-session adapter (PR7) ([#720](https://github.com/Gitlawb/zero/issues/720)) ([30e2c3f](https://github.com/Gitlawb/zero/commit/30e2c3f7ffa1d5e487bd10b59d4e823cda191d48))
+* **turn-bench:** Phase 0 — strengthen oracles so pass rate can't be misread as correctness ([#712](https://github.com/Gitlawb/zero/issues/712)) ([727ad4d](https://github.com/Gitlawb/zero/commit/727ad4d321fab45d0cf40f8535522e3d94e55c4a))
+
+## [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`.
 

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.
@@ -55,6 +67,28 @@ An approved issue means only that the specific issue has been accepted for work.
 It does not approve unrelated future work, future pull requests, or future issues
 from the same author. Each issue is reviewed case by case.
 
+## Contribution Volume and Batching
+
+To keep triage and review sustainable while the project stabilizes, please keep
+your contribution flow to a size the core team can realistically work through:
+
+- One issue per bug. File each defect as its own issue with its own
+  reproduction. Please do not bundle several unrelated findings into a single
+  "audit" issue; separate issues are far easier to triage, approve, and close
+  individually.
+- Keep only a few items open at a time. If you already have several issues or
+  pull requests awaiting review, let those move through the queue before opening
+  more. A large simultaneous burst from one contributor slows review for
+  everyone, even when each item is valid.
+- AI-assisted contributions are welcome, held to the same bar. Automated audits
+  and generated patches are fine, but each issue and pull request must be
+  individually verified, scoped, and reproducible before you open it. Please do
+  not post bulk machine-generated findings unfiltered.
+
+These apply to all contributors and are about review bandwidth, not the value of
+any single report. Maintainers may ask you to consolidate, split, or hold
+submissions to keep the queue manageable.
+
 ## Pull Request Requirements
 
 Before opening a community pull request, make sure all of the following are true:
@@ -133,7 +167,12 @@ 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 +195,9 @@ 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.

Makefile

@@ -1,7 +1,13 @@
-# Zero build/test/lint targets. AGENTS.MD says "Build with `make`" and "Run `make
-# lint` before opening a PR" — these targets back those instructions.
+# Zero build/test/lint targets. AGENTS.md says to build and run quality checks
+# with `make` — these targets back those instructions.
 .DEFAULT_GOAL := build
-.PHONY: build build-all test test-race vet fmt fmt-check lint tidy clean help
+GO_VERSION = $(word 2,$(shell git grep -G -h "^go[[:space:]]" -- go.mod))
+GO_TOOLCHAIN = go$(GO_VERSION)
+DEADCODE_VERSION := v0.46.0
+GOLANGCI_LINT_VERSION := v2.12.2
+GOVULNCHECK_VERSION := v1.3.0
+
+.PHONY: build build-all test test-race vet fmt fmt-check lint lint-static deadcode vulncheck tidy clean baseline help
 
 # Build the main CLI binary into ./zero.
 build:
@@ -33,6 +39,22 @@ fmt-check:
 # Lint = formatting check + vet (no extra tooling required).
 lint: fmt-check vet
 
+# Versioned tools select the toolchain from their own modules when invoked with
+# package@version. Read this module's go directive directly, without invoking
+# the possibly stale Go toolchain or consulting a multi-module GOWORK. git grep
+# works with both POSIX shells and cmd.exe, including GNU Make 3.81 on macOS.
+# The target-specific export is shell-independent.
+lint-static deadcode vulncheck: export GOTOOLCHAIN = $(GO_TOOLCHAIN)
+
+lint-static:
+	go run github.com/golangci/golangci-lint/v2/cmd/golangci-lint@$(GOLANGCI_LINT_VERSION) run --enable-only unused,ineffassign,staticcheck ./...
+
+deadcode:
+	go run golang.org/x/tools/cmd/deadcode@$(DEADCODE_VERSION) -test=false ./...
+
+vulncheck:
+	go run golang.org/x/vuln/cmd/govulncheck@$(GOVULNCHECK_VERSION) ./...
+
 tidy:
 	go mod tidy
 
@@ -40,5 +62,19 @@ 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, lint-static, deadcode, vulncheck, tidy, clean, baseline"

README.md

@@ -6,8 +6,11 @@
 
 <p align="center">
   <a href="LICENSE"><img alt="license" src="https://img.shields.io/badge/license-MIT-blue"></a>
-  <img alt="Go 1.25+" src="https://img.shields.io/badge/Go-1.25+-00ADD8?logo=go&logoColor=white">
+  <img alt="Go 1.26.6+" src="https://img.shields.io/badge/Go-1.26.6+-00ADD8?logo=go&logoColor=white">
   <img alt="25+ providers" src="https://img.shields.io/badge/providers-25+-34E2EA">
+  <a href="https://discord.gg/CaQDS6wdFn"><img alt="Discord" src="https://img.shields.io/badge/Discord-join-5865F2?logo=discord&logoColor=white"></a>
+  <br>
+  <strong>English</strong> | <a href="README_ZH.md">中文</a>
 </p>
 
 Zero is an AI coding agent for your local terminal. It can inspect a repository,
@@ -47,9 +50,15 @@ npm install -g @gitlawb/zero
 zero
 ```
 
-The npm package installs a small wrapper plus the matching Zero binary for your
-platform from GitHub Releases. It supports Linux, macOS, and Windows on x64 and
-arm64.
+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.
 
 ### Install scripts
 
@@ -67,7 +76,7 @@ irm https://raw.githubusercontent.com/Gitlawb/zero/main/scripts/install.ps1 | ie
 
 ### From source
 
-Source builds require Go 1.25+.
+Source builds require Go 1.26.6+.
 
 ```bash
 git clone https://github.com/Gitlawb/zero.git
@@ -121,6 +130,56 @@ the key in the wizard:
 export OPENAI_API_KEY=sk-...
 export ANTHROPIC_API_KEY=...
 export GEMINI_API_KEY=...
+export AIMLAPI_API_KEY=...
+export LONGCAT_API_KEY=...
+export FIREWORKS_API_KEY=...
+export MINIMAX_API_KEY=...
+export MINIMAXI_API_KEY=...
+```
+
+To configure AI/ML API directly, run:
+
+```bash
+zero providers setup aimlapi --set-active
+```
+
+To configure Meituan LongCat (LongCat-2.0) directly, run:
+
+```bash
+zero providers setup longcat --set-active
+```
+
+To configure Fireworks AI directly, run:
+
+```bash
+zero providers setup fireworks --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
@@ -140,9 +199,11 @@ Useful controls:
 |---|---|
 | `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 |
+| `Ctrl+C` | cancel, exit, or return from a `/btw` conversation |
 
 Common slash commands:
 
@@ -152,6 +213,9 @@ Common slash commands:
 | `/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 |
+| `/new` | start a fresh session in place (previous session stays on disk) |
+| `/btw [question]` | ask in an isolated fork without adding the side conversation to the main session |
+| `/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 |
@@ -246,17 +310,51 @@ 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
-zero update           check for newer releases
+zero update --check   check for newer releases
+zero upgrade          download, verify, and install the latest release
 ```
 
+## Extending Zero
+
+### Project and personal instructions
+
+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 |
 |---|---|
 | `NO_COLOR=<anything>` | disables color output |
-| `ZERO_THEME=auto|dark|light` | selects the startup theme |
-| `--theme auto|dark|light` | selects the TUI theme from the CLI |
-| `/theme auto|dark|light` | switches theme inside the TUI |
+| `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 |
 
 Meaning does not rely on color alone; diffs, permissions, and statuses also use
@@ -271,7 +369,35 @@ go run ./cmd/zero-release smoke
 go run ./cmd/zero-perf-bench
 ```
 
-Cross-compile examples:
+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. The `make` targets below pin each tool to this module's Go version, so they load correctly even when your default `go` toolchain is older — running the plain `go run ...@version` form yourself can select the tool module's own (older) toolchain and fail to load this module instead.
+
+1. **Formatting**: Run `make fmt` (or `go fmt ./...`).
+2. **Vetting**: Run `make vet` (or `go vet ./...`).
+3. **Linting**: Run `make lint-static`.
+4. **Vulnerability Scan**: Run `make vulncheck`.
+
+Use the repository-managed targets rather than globally installed binaries:
+the targets apply the module's required Go toolchain as well as the reviewed,
+pinned tool versions.
+
+The installed binaries land in `$GOBIN` when it is set, otherwise in
+`$GOPATH/bin` (default `~/go/bin`). That directory must be on your `PATH` to
+run them directly. If it isn't, add it:
+
+```bash
+gobin="$(go env GOBIN)"; [ -z "$gobin" ] && gobin="$(go env GOPATH)/bin"
+case ":$PATH:" in *":$gobin:"*) ;; *) export PATH="$PATH:$gobin" ;; esac
+```
+
+### Cross-Compile Examples
 
 ```bash
 go run ./cmd/zero-release build --goos linux --goarch amd64
@@ -282,6 +408,7 @@ go run ./cmd/zero-release build --goos windows --goarch amd64 --output dist/zero
 
 - [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)
@@ -289,6 +416,26 @@ go run ./cmd/zero-release build --goos windows --goarch amd64 --output dist/zero
 - [Performance](docs/PERFORMANCE.md)
 - [Agent evals](docs/AGENT_EVALS.md)
 
+## Community
+
+Real-time chat happens on the [Discord server](https://discord.gg/CaQDS6wdFn).
+
+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. Read [CONTRIBUTING.md](CONTRIBUTING.md), run the

README_ZH.md

@@ -0,0 +1,297 @@
+<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 1.26.6+" src="https://img.shields.io/badge/Go-1.26.6+-00ADD8?logo=go&logoColor=white">
+  <img alt="25+ providers" src="https://img.shields.io/badge/providers-25+-34E2EA">
+  <a href="https://discord.gg/CaQDS6wdFn"><img alt="Discord" src="https://img.shields.io/badge/Discord-join-5865F2?logo=discord&logoColor=white"></a>
+  <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 1.26.6+。
+
+```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 |
+| `/new` | 原地开始一个新会话（之前的会话仍保留在磁盘上） |
+| `/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 --check   检查更新版本
+zero upgrade          下载、验证并安装最新版本
+```
+
+## 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)
+
+## 社区
+
+实时交流在 [Discord 服务器](https://discord.gg/CaQDS6wdFn) 进行。
+
+提问、安装帮助、想法和分享都在
+[GitHub Discussions](https://github.com/Gitlawb/zero/discussions)：
+
+| 分类 | 用途 |
+|---|---|
+| [Q&A](https://github.com/Gitlawb/zero/discussions/categories/q-a) | 安装帮助、提供商/模型配置、"如何做"类问题 |
+| [Ideas](https://github.com/Gitlawb/zero/discussions/categories/ideas) | 功能提议和 PR 之前的设计讨论 |
+| [Show and tell](https://github.com/Gitlawb/zero/discussions/categories/show-and-tell) | 你的技能、插件、MCP 配置、主题和工作流 |
+| [Announcements](https://github.com/Gitlawb/zero/discussions/categories/announcements) | 维护者发布的版本和项目动态 |
+
+## 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。

SECURITY.md

@@ -33,4 +33,4 @@ Please include, where possible:
   depend on a user explicitly disabling the sandbox (`--skip-permissions-unsafe`), are generally
   out of scope — but please report anything you are unsure about.
 - Zero sends no telemetry. Network calls go to the model/tool providers you configure (and, for
-  keyless web search, to the documented Firecrawl endpoint — see the README).
+  keyless web search, to Exa's hosted MCP endpoint at `https://mcp.exa.ai/mcp`).

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';
 
@@ -39,6 +40,11 @@ function resolveHelper(packageRoot, name) {
   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']) {
@@ -49,18 +55,229 @@ function localControlHelperManifest(packageRoot) {
   return JSON.stringify({ version: 1, helpers });
 }
 
-const packageRoot = dirname(dirname(fileURLToPath(import.meta.url)));
-const nativePath = join(packageRoot, zeroBinaryName());
-const localControlHelpers = localControlHelperManifest(packageRoot);
+// 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]
 
-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.'
+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 {

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. 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 skill add` / `zero skill remove`) always write to the primary Zero dir only; `list` / `info` search primary + `~/.agents/skills`. `skill info` reports lock source/hash when present and whether the on-disk `SKILL.md` has drifted from the recorded hash.
+
+```text
+~/.local/share/zero/skills/          # primary (install/remove 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:
@@ -269,6 +289,7 @@ Install and manage:
 ```bash
 zero plugins add ./github-pr-review      # copy into ~/.config/zero/plugins/ or ./.zero/plugins/
 zero plugins list
+zero plugins info github-pr-review       # show manifest path, enabled state, lock metadata
 zero plugins remove github-pr-review    # alias: rm
 ```
 
@@ -276,19 +297,19 @@ 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).
 
@@ -301,19 +322,20 @@ 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/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`; 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

@@ -1,13 +1,30 @@
 # Installing Zero
 
-Zero is distributed as:
+Pick whichever fits how you already install things. None of these is the
+blessed one.
 
-- an npm package, `@gitlawb/zero`
-- release archives on GitHub Releases
-- source builds with Go 1.25+
+| Method | Command | Self-update |
+| --- | --- | --- |
+| Install script (Linux, macOS) | `curl -fsSL https://raw.githubusercontent.com/Gitlawb/zero/main/scripts/install.sh \| bash` | `zero upgrade` |
+| Install script (Windows) | `irm https://raw.githubusercontent.com/Gitlawb/zero/main/scripts/install.ps1 \| iex` | `zero upgrade` |
+| npm | `npm install -g @gitlawb/zero` | `zero upgrade` (runs npm) |
+| Release archive | download from [Releases](https://github.com/Gitlawb/zero/releases) | `zero upgrade` |
+| mise | `mise use -g ubi:Gitlawb/zero` | `mise upgrade` |
+| `go install` | `go install github.com/Gitlawb/zero/cmd/zero@latest` | rerun the command |
+| Source | `go build -o zero ./cmd/zero` | rebuild |
 
-The npm package and install scripts download a platform-specific release archive.
-They require a published GitHub Release for the requested version.
+`zero upgrade` follows whatever owns the binary, with one exception it cannot
+detect. It delegates to npm for an npm install, refuses outright for a Homebrew
+keg and names `brew upgrade zero`, and replaces the binary in place for
+everything else — including a mise-managed one, which is the exception.
+
+mise is the one case it cannot detect: a mise-managed binary is an ordinary
+standalone install on disk, so `zero upgrade` will replace it and leave mise
+describing a version that is no longer there. Use `mise upgrade` instead.
+
+Release archives are the substrate for most of the above: the install scripts
+and the npm fallback both download a platform archive from a published GitHub
+Release and verify its checksum.
 
 ## npm
 
@@ -16,13 +33,31 @@ npm install -g @gitlawb/zero
 zero
 ```
 
-The package supports Linux, macOS, and Windows on x64 and arm64. It installs the
-`zero` command and downloads the matching release binary during `postinstall`.
+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.
 
 Requirements:
 
 - Node.js 18+
-- network access to npm and GitHub Releases
+- 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 Script
 
@@ -92,6 +127,45 @@ Defaults:
 - Version: latest GitHub release
 - Install path: `%LOCALAPPDATA%\zero\bin\zero.exe`
 
+## mise
+
+[mise](https://mise.jdx.dev/) installs Zero straight from the GitHub Release
+archives through its `ubi` backend, so there is nothing extra to publish and no
+registry in the middle:
+
+```bash
+mise use -g ubi:Gitlawb/zero
+```
+
+Pin a version the same way you would any other tool:
+
+```bash
+mise use -g ubi:Gitlawb/[email protected]
+```
+
+Updates come from `mise upgrade`. `zero upgrade` also works, because a
+mise-managed binary is an ordinary standalone install, but then mise's records
+describe a version that is no longer on disk. Prefer `mise upgrade`.
+
+## go install
+
+```bash
+go install github.com/Gitlawb/zero/cmd/zero@latest
+```
+
+This builds from source, so it needs Go 1.26.6+ and it does not go through the
+release archives. Two consequences worth knowing before you pick it:
+
+- On Linux you also need the sandbox helper, which is a separate binary and is
+  not installed by this command. See
+  [Sandbox Helpers For Source Builds](#sandbox-helpers-for-source-builds).
+  Without it, native sandboxing is unavailable.
+- `zero upgrade` treats the result as a standalone install and will replace the
+  binary with a release build rather than rebuilding from source. If you chose
+  `go install` deliberately, rerun it instead.
+
+macOS and Windows need no extra helper.
+
 ## From Source
 
 ```bash
@@ -106,7 +180,7 @@ Build a local binary:
 go build -o zero ./cmd/zero
 ```
 
-Source builds require Go 1.25+.
+Source builds require Go 1.26.6+.
 
 ### Sandbox Helpers For Source Builds
 
@@ -128,6 +202,61 @@ 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`:
@@ -152,8 +281,7 @@ Supported targets:
 - `linux-arm64`
 - `macos-x64`
 - `macos-arm64`
-- `windows-x64`
-- `windows-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
@@ -161,10 +289,31 @@ directory.
 
 ## Updating
 
-Check for a newer release:
+Check for a newer release, then download, verify, and install it in place:
 
 ```bash
 zero update --check
+zero upgrade
 ```
 
-Then reinstall with npm or rerun the install script for the version you want.
+See the [update guide](UPDATE.md) for update modes, flags, and platform details.
+
+`zero upgrade` behaves differently depending on how Zero was installed, because
+overwriting a binary a package manager owns leaves that manager describing a
+version that is no longer there:
+
+- **npm**: runs `npm install -g @gitlawb/zero@latest` for you.
+- **Homebrew**: refuses, and tells you to run `brew upgrade zero`. Replacing the
+  keg binary directly would be reverted by the next `brew upgrade` or
+  `brew reinstall`.
+- **everything else**: downloads the verified release archive and replaces the
+  binary in place.
+
+Homebrew is detected by the binary living inside a Cellar keg that carries
+Homebrew's own `INSTALL_RECEIPT.json`. The keg shape alone is not enough: a
+directory tree you built yourself can have exactly that shape, and refusing to
+update an install Homebrew never touched would be a real cost for a made-up
+reason. The receipt is the thing only Homebrew writes.
+
+An ordinary install under `/usr/local/bin` is left alone even on an Intel Mac
+where that is also the Homebrew prefix.

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/README.md

@@ -17,6 +17,7 @@ Zero.
 
 ## Maintainer Docs
 
+- [How Zero Works](HOW_ZERO_WORKS.md)
 - [npm wrapper smoke checklist](NPM_WRAPPER_SMOKE.md)
 - [Benchmarks](BENCHMARK.md)
 - [Performance benchmarks](PERFORMANCE.md)

docs/SPECIALISTS.md

@@ -138,3 +138,45 @@ output or stop a still-running task by id.
 If Zero is restarted while a background task is still marked `running`, the new
 manager marks that task `error` and clears its PID. This avoids sending
 `TaskStop` to a stale PID that may now belong to an unrelated process.
+
+## Recovering an Interrupted Overwrite
+
+Zero writes and flushes a complete temporary file before publishing an overwrite,
+so a write failure before publication leaves the existing manifest unchanged
+instead of truncating it. On Unix, publication uses a same-directory rename and
+preserves the existing file's permission bits.
+
+On Windows, Zero uses `ReplaceFileW` to preserve the destination DACL instead of
+silently replacing it with the temporary file's inherited DACL. `ReplaceFileW`
+is not observer-atomic: another process can briefly observe the destination path
+as absent during replacement. Zero serializes specialist loads and managed
+mutations within one process, but cannot synchronize external processes or
+editors.
+
+Windows errors 1176 (`ERROR_UNABLE_TO_MOVE_REPLACEMENT`) and 1177
+(`ERROR_UNABLE_TO_MOVE_REPLACEMENT_2`) are partial replacement failures. With
+Zero's managed backup, 1176 leaves the original names intact and needs no manual
+recovery. For 1177, Zero has moved the original aside and tries to move it back.
+That rollback almost always succeeds, and the failed write changes nothing.
+
+If the rollback itself fails — typically because another process is holding a
+lock on the file — the original is not lost, but it is left under a name Zero
+does not read:
+
+```text
+<specialist dir>/.zero-replace-<random>.backup
+```
+
+Only `*.md` files are loaded as specialists, so until that file is renamed the
+specialist will not appear in `zero specialist list` or resolve by name. Zero's
+error message includes both the backup path and destination path. Recover by
+closing whatever holds the lock and renaming the backup back:
+
+```powershell
+Move-Item .zero-replace-<random>.backup <name>.md
+```
+
+A `.zero-replace-*.backup` can also linger after a *successful* overwrite if the
+backup could not be deleted afterward. That case is reported as a warning rather
+than an error — the new manifest is already in place, and the leftover file is
+safe to delete.

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

@@ -1,16 +1,23 @@
 # Update Flow
 
 `zero update --check` checks the latest GitHub release and compares it with the
-local CLI version.
+local CLI version. `zero update --apply` (or its shorthand, `zero upgrade`)
+downloads, verifies, and installs it.
 
 ```bash
 zero update --check
 zero update --check --json
 zero update --check --repo Gitlawb/zero
 zero update --check --target windows-x64
+
+zero upgrade
+zero update --apply
 ```
 
-The command is intentionally check-only:
+`--check` and `--apply` are mutually exclusive. `zero update` requires one of
+them explicitly; `zero upgrade` is `zero update` with `--apply` implied.
+
+`--check` is check-only:
 
 - It does not replace the running binary.
 - It exits with code `0` when the check succeeds, even when an update is
@@ -19,14 +26,35 @@ The command is intentionally check-only:
 - `--json` prints the same result in a machine-readable format for scripts and
   CI.
 
+`--apply` installs the update in place:
+
+- npm installs delegate to `npm install -g @gitlawb/zero@latest`.
+- Standalone installs download the release archive, verify its checksum,
+  extract it, and atomically replace the running binary plus any installed
+  optional sandbox helpers.
+- On Windows, the running executable is renamed aside and cleaned up on the
+  next `zero update --apply` or `zero upgrade`, since it can't be overwritten
+  while running.
+- `--target` cannot be combined with `--apply`; it only applies to `--check`,
+  since applying always installs onto the current machine.
+- `--repo` and `--endpoint` are ignored when applying to an npm-managed
+  install: that path delegates to `npm install -g @gitlawb/zero@latest` and
+  takes its release from the npm registry, not from GitHub. They still apply to
+  `--check` there.
+- `--json` serializes Zero's final result. For npm-managed installs, npm may
+  also write progress output to stdout, so neither
+  `zero update --apply --json` nor `zero upgrade --json` is guaranteed to
+  produce a single parseable JSON document. Use `--check --json` for
+  machine-readable automation.
+
 Useful flags:
 
 | Flag | Purpose |
 |---|---|
-| `--repo <owner/repo>` | Check another GitHub repository. |
-| `--endpoint <url|owner/repo>` | Check a specific release API URL or repository slug. |
+| `--repo <owner/repo>` | Use another GitHub repository for `--check` and `--apply`/`upgrade`. Ignored by an npm-managed apply. |
+| `--endpoint <url\|owner/repo>` | Use a specific release API URL or repository slug for `--check` and `--apply`/`upgrade`. Ignored by an npm-managed apply. |
 | `--timeout <duration>` | Override the default release check timeout. |
-| `--target <platform-arch>` | Validate release metadata for another supported target. |
+| `--target <platform-arch>` | Validate release metadata for another supported target (`--check` only). |
 
 Supported targets are `linux-x64`, `linux-arm64`, `macos-x64`, `macos-arm64`,
 `windows-x64`, and `windows-arm64`. Without `--target`, Zero checks the current
@@ -40,5 +68,45 @@ Endpoint resolution order:
 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.
+verify its `.sha256` file. If Zero is already installed, run `zero upgrade`
+instead of reinstalling.
+
+## Windows recovery state (standalone installs)
+
+This section describes Windows only. On Linux and macOS a standalone update
+renames the staged file directly over the executable path through the
+installation directory's file descriptor, so the replacement is atomic and no
+aside copy, marker, or recovery record is ever created — a failed update leaves
+the previous binary in place and nothing to resolve.
+
+On Windows, a running executable cannot be replaced in place, so the previous
+binary is moved aside to `<binary>.zero-update-<random>.old` first. The updater
+records that exact file — bound to its filesystem identity, in per-user state
+outside the installation directory — and deletes only that recorded copy after
+the next update is verified in place. Backups it did not create are never
+removed, so a file such as `zero.exe.before-manual-patch.old` is left alone. A
+recorded copy that something else holds open (an on-access scanner, an editor)
+stays recorded and is removed by a later update instead.
+
+An update **refuses to run** while unresolved recovery state exists beside the
+binary:
+
+| State on disk | Meaning |
+|---|---|
+| `<binary>.old` (or `<binary>.<suffix>.old`) plus a `.keep` marker | A previous update could not restore the original binary. The `.old` file may be the last binary the updater verified; the installed one may be unverified. |
+| `<binary>.…old.<suffix>.recovery` | A previous update could not even write the marker, so it moved the last verified binary to that name. |
+| The binary is missing and one or more `*.old` files exist | The previous attempt was interrupted between moves. |
+
+The refusal names the paths involved and the two moves that end the state:
+either move the recovery binary back over the executable path, or — if the
+installed binary is the one you want — delete the `.keep` marker (or the
+`.recovery` copy, once you have verified the installed binary) and update again.
+
+This is deliberately fail-closed and differs from older releases, which deleted
+`<binary>.old` and proceeded. The trade-off is that anyone who can write in the
+installation directory can plant `<binary>.old` and `<binary>.old.keep` there
+and make every subsequent update refuse until an operator clears them. That is
+the safe direction: the alternative is an update that overwrites the only
+verified copy of the previous binary. If updates start refusing, inspect those
+files before removing them — an installation directory that a lower-privileged
+account can write to is itself worth fixing.

go.mod

@@ -1,32 +1,33 @@
 module github.com/Gitlawb/zero
 
-go 1.25.0
-
-toolchain go1.26.4
+go 1.26.6
 
 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/Microsoft/go-winio v0.6.2
 	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/ultraviolet v0.0.0-20260703014108-f5a850f9c2b7
 	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/image v0.45.0
+	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/x/exp/golden v0.0.0-20260615092313-b57e5e6d29bb // 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
@@ -35,5 +36,5 @@ 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 +1,11 @@
-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/Microsoft/go-winio v0.6.2 h1:F2VQgta7ecxGYO8k3ZZz3RS8fVIXVxONVUPlNERoyfY=
+github.com/Microsoft/go-winio v0.6.2/go.mod h1:yd8OoFMLzJbo9gZq8j5qaps8bJ9aShtEA8Ipt1oGCvU=
 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 +18,8 @@ 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 +34,10 @@ 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 +64,13 @@ 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/image v0.44.0 h1:+tDekMZED9+LrtB3G5xzRggpVh9CARjZqROla3R3R+I=
+golang.org/x/image v0.44.0/go.mod h1:V8K3KE9KKKE+pLpQDOeN18w9oacNSvy1tDOirTu4xtY=
+golang.org/x/image v0.45.0 h1:FMb1nTbH5H9vF55SriQHgFw5GnNL9Jg6L25BwXKzhB0=
+golang.org/x/image v0.45.0/go.mod h1:n62x/7RqlwXDvGsSU4u6IUTUf6KghUZ9Bt7cG/T9Fx4=
+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

@@ -5,11 +5,15 @@ import (
 	"encoding/base64"
 	"encoding/json"
 	"errors"
+	"log"
 	"strings"
 	"sync"
 
 	"github.com/Gitlawb/zero/internal/agent"
 	"github.com/Gitlawb/zero/internal/config"
+	"github.com/Gitlawb/zero/internal/providercatalog"
+	"github.com/Gitlawb/zero/internal/providermodelcatalog"
+	"github.com/Gitlawb/zero/internal/providermodeldiscovery"
 	"github.com/Gitlawb/zero/internal/sandbox"
 	"github.com/Gitlawb/zero/internal/sessions"
 	"github.com/Gitlawb/zero/internal/tools"
@@ -22,9 +26,10 @@ import (
 // means the editor only hosts the thread — ZERO owns BYOK and telemetry-free
 // operation.
 type Deps struct {
-	ResolveConfig func(workspaceRoot string, overrides config.Overrides) (config.ResolvedConfig, error)
-	NewProvider   func(profile config.ProviderProfile) (zeroruntime.Provider, error)
-	RunAgent      func(ctx context.Context, prompt string, provider zeroruntime.Provider, opts agent.Options) (agent.Result, error)
+	ResolveConfig  func(workspaceRoot string, overrides config.Overrides) (config.ResolvedConfig, error)
+	DiscoverModels func(context.Context, config.ProviderProfile) ([]providermodeldiscovery.Model, error)
+	NewProvider    func(profile config.ProviderProfile) (zeroruntime.Provider, error)
+	RunAgent       func(ctx context.Context, prompt string, provider zeroruntime.Provider, opts agent.Options) (agent.Result, error)
 	// BuildWorkspace builds the SCOPED tool registry and the sandbox engine for a
 	// validated workspace root, so ACP shell tools (bash/exec_command) are confined
 	// exactly like the exec surface — never run unconfined on the host.
@@ -59,13 +64,16 @@ type acpSession struct {
 	// turnMu serializes prompt turns for one session: concurrent session/prompt
 	// calls run one at a time so they can't interleave history or clobber the
 	// single cancel slot.
-	turnMu sync.Mutex
+	turnMu  sync.Mutex
+	modelMu sync.Mutex
 
-	mu      sync.Mutex
-	mode    agent.PermissionMode
-	model   string // override; "" => config default
-	cancel  context.CancelFunc
-	history []turnRecord
+	mu             sync.Mutex
+	mode           agent.PermissionMode
+	model          string
+	models         []SessionConfigOptionValue
+	restrictModels bool
+	cancel         context.CancelFunc
+	history        []turnRecord
 }
 
 // NewAgent builds the ACP server and registers its method handlers on conn.
@@ -82,9 +90,6 @@ func NewAgent(conn *Conn, deps Deps) *Agent {
 	return a
 }
 
-// Serve runs the connection read loop until the stream closes or ctx is done.
-func (a *Agent) Serve(ctx context.Context) error { return a.conn.Serve(ctx) }
-
 // ---- initialize ----
 
 func (a *Agent) handleInitialize(_ context.Context, params json.RawMessage) (any, error) {
@@ -119,7 +124,7 @@ func (a *Agent) handleInitialize(_ context.Context, params json.RawMessage) (any
 
 // ---- session lifecycle ----
 
-func (a *Agent) handleSessionNew(_ context.Context, params json.RawMessage) (any, error) {
+func (a *Agent) handleSessionNew(ctx context.Context, params json.RawMessage) (any, error) {
 	var p NewSessionParams
 	if err := json.Unmarshal(params, &p); err != nil {
 		return nil, RPCError(codeInvalidParams, "invalid session/new params")
@@ -128,18 +133,23 @@ func (a *Agent) handleSessionNew(_ context.Context, params json.RawMessage) (any
 	if err != nil {
 		return nil, RPCError(codeInvalidParams, err.Error())
 	}
-	meta, err := a.deps.Store.Create(sessions.CreateInput{Title: "ACP session", Cwd: root})
+	model, models, restrictModels, err := a.resolveModelChoices(ctx, root)
+	if err != nil {
+		return nil, RPCError(codeInternalError, "config: "+err.Error())
+	}
+	meta, err := a.deps.Store.Create(sessions.CreateInput{Title: "ACP session", Cwd: root, ModelID: model})
 	if err != nil {
 		return nil, RPCError(codeInternalError, "create session: "+err.Error())
 	}
-	sess := a.registerSession(meta.SessionID, root, nil)
+	sess := a.registerSession(meta.SessionID, root, nil, model, models, restrictModels)
 	return NewSessionResult{
-		SessionID: sess.id,
-		Modes:     a.modeState(sess),
+		SessionID:     sess.id,
+		ConfigOptions: a.configOptions(sess),
+		Modes:         a.modeState(sess),
 	}, nil
 }
 
-func (a *Agent) handleSessionLoad(_ context.Context, params json.RawMessage) (any, error) {
+func (a *Agent) handleSessionLoad(ctx context.Context, params json.RawMessage) (any, error) {
 	var p LoadSessionParams
 	if err := json.Unmarshal(params, &p); err != nil {
 		return nil, RPCError(codeInvalidParams, "invalid session/load params")
@@ -159,10 +169,27 @@ 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)
-	sess := a.registerSession(meta.SessionID, root, history)
+	history, historyErr := a.loadHistory(meta.SessionID)
+	model, models, restrictModels, err := a.resolveModelChoices(ctx, root)
+	if err != nil {
+		return nil, RPCError(codeInternalError, "config: "+err.Error())
+	}
+	if persistedModel := strings.TrimSpace(meta.ModelID); persistedModel != "" && (!restrictModels || modelChoiceExists(models, persistedModel)) {
+		model = persistedModel
+		if !modelChoiceExists(models, persistedModel) {
+			models = append(models, SessionConfigOptionValue{Value: persistedModel, Name: persistedModel})
+		}
+	}
+	sess := a.registerSession(meta.SessionID, root, history, model, models, restrictModels)
+	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),
+		ConfigOptions: a.configOptions(sess),
+		Modes:         a.modeState(sess),
 	}, nil
 }
 
@@ -253,7 +280,14 @@ 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
 }
 
@@ -262,6 +296,16 @@ func stopReasonFor(result agent.Result, err error) (string, error) {
 		if errors.Is(err, context.Canceled) {
 			return StopCancelled, nil
 		}
+		// A CLIENT THAT CANCELLED A PERMISSION PROMPT DID NOT FAIL. Only
+		// context.Canceled was matched here, so dismissing a permission dialog —
+		// a deliberate action, and for apply_patch the only refusal a client is
+		// offered — came back as JSON-RPC -32603 carrying the internal sentinel
+		// text. Editors and the desktop app both render that as a crashed turn,
+		// so declining a tool looked like ZERO falling over. It is a
+		// cancellation, and StopCancelled is what ACP has for saying so.
+		if errors.Is(err, agent.ErrPermissionApprovalCanceled) {
+			return StopCancelled, nil
+		}
 		return "", err
 	}
 	if result.FinishReason == "length" {
@@ -277,10 +321,14 @@ func stopReasonFor(result agent.Result, err error) (string, error) {
 // session/request_permission request and maps the outcome back. Failure to reach
 // the client fails closed to deny.
 func (a *Agent) requestPermission(ctx context.Context, sessionID string, req agent.PermissionRequest) (agent.PermissionDecision, error) {
+	// Built ONCE and used for both halves. Sending one list and validating the
+	// reply against another is the defect this whole path had; see
+	// decisionFromOutcome.
+	options := buildPermissionOptions(req)
 	params := RequestPermissionParams{
 		SessionID: sessionID,
 		ToolCall:  permissionToolCall(req),
-		Options:   buildPermissionOptions(req),
+		Options:   options,
 	}
 	var result RequestPermissionResult
 	if err := a.conn.Call(ctx, MethodSessionRequestPerm, params, &result); err != nil {
@@ -289,7 +337,7 @@ func (a *Agent) requestPermission(ctx context.Context, sessionID string, req age
 		}
 		return agent.PermissionDecision{Action: agent.PermissionDecisionDeny, Reason: "permission request failed: " + err.Error()}, nil
 	}
-	return decisionFromOutcome(result.Outcome, req.AvailableDecisions), nil
+	return decisionFromOutcome(result.Outcome, options), nil
 }
 
 func (a *Agent) emitPlan(registry *tools.Registry, note *notifier) {
@@ -315,9 +363,11 @@ func (a *Agent) handleSetMode(_ context.Context, params json.RawMessage) (any, e
 	if sess == nil {
 		return nil, RPCError(codeInvalidParams, "unknown session: "+p.SessionID)
 	}
+	sess.turnMu.Lock()
+	defer sess.turnMu.Unlock()
 	mode := agent.PermissionMode(p.ModeID)
 	switch mode {
-	case agent.PermissionModeAuto, agent.PermissionModeAsk:
+	case agent.PermissionModeAuto, agent.PermissionModeAsk, agent.PermissionModePlan:
 		sess.setMode(mode)
 		(&notifier{conn: a.conn, sessionID: sess.id}).currentMode(string(mode))
 		return SetSessionModeResult{}, nil
@@ -340,11 +390,31 @@ func (a *Agent) handleSetConfigOption(_ context.Context, params json.RawMessage)
 	if sess == nil {
 		return nil, RPCError(codeInvalidParams, "unknown session: "+p.SessionID)
 	}
-	if p.ConfigID != configIDModel {
+	switch p.ConfigID {
+	case configIDModel:
+		model := strings.TrimSpace(p.Value)
+		if err := a.updateModel(sess, model, sess.restrictModels); err != nil {
+			return nil, err
+		}
+	case configIDMode:
+		// Same turnMu as handleSetMode so the two advertised mode doors (session
+		// set_mode and set_config_option) serialize mode flips consistently.
+		sess.turnMu.Lock()
+		defer sess.turnMu.Unlock()
+		mode := agent.PermissionMode(p.Value)
+		switch mode {
+		case agent.PermissionModeAuto, agent.PermissionModeAsk, agent.PermissionModePlan:
+			sess.setMode(mode)
+			(&notifier{conn: a.conn, sessionID: sess.id}).currentMode(string(mode))
+		case agent.PermissionModeUnsafe:
+			return nil, RPCError(codeInvalidParams, "mode not permitted over ACP: "+p.Value)
+		default:
+			return nil, RPCError(codeInvalidParams, "unknown mode: "+p.Value)
+		}
+	default:
 		return nil, RPCError(codeInvalidParams, "unknown config option: "+p.ConfigID)
 	}
-	sess.setModel(p.Value)
-	return SetSessionConfigOptionResult{}, nil
+	return SetSessionConfigOptionResult{ConfigOptions: a.configOptions(sess)}, nil
 }
 
 func (a *Agent) handleZeroSetModel(_ context.Context, params json.RawMessage) (any, error) {
@@ -356,8 +426,24 @@ func (a *Agent) handleZeroSetModel(_ context.Context, params json.RawMessage) (a
 	if sess == nil {
 		return nil, RPCError(codeInvalidParams, "unknown session: "+p.SessionID)
 	}
-	sess.setModel(p.Model)
-	return ZeroSetModelResult{Model: p.Model}, nil
+	model := strings.TrimSpace(p.Model)
+	if err := a.updateModel(sess, model, false); err != nil {
+		return nil, err
+	}
+	return ZeroSetModelResult{Model: model}, nil
+}
+
+func (a *Agent) updateModel(sess *acpSession, model string, restrictModels bool) error {
+	sess.modelMu.Lock()
+	defer sess.modelMu.Unlock()
+	if restrictModels && !sess.hasModel(model) {
+		return RPCError(codeInvalidParams, "unknown model: "+model)
+	}
+	if _, err := a.deps.Store.UpdateModel(sess.id, model); err != nil {
+		return RPCError(codeInternalError, "save model selection: "+err.Error())
+	}
+	sess.setModel(model)
+	return nil
 }
 
 func (a *Agent) handleCancel(_ context.Context, params json.RawMessage) {
@@ -373,42 +459,121 @@ func (a *Agent) handleCancel(_ context.Context, params json.RawMessage) {
 // ---- advertising helpers ----
 
 func (a *Agent) modeState(s *acpSession) *SessionModeState {
-	// Only auto/ask are offered over ACP; Unsafe is gated to the operator (see
-	// handleSetMode) so a client can't grant itself no-prompt host access.
+	// auto/ask/plan are offered over ACP; Unsafe is gated to the operator (see
+	// handleSetMode) so a client can't grant itself no-prompt host access. Plan
+	// only narrows what a client can do (read-only, no write/shell tools), so
+	// unlike Unsafe there is no elevation risk in letting a client select it.
 	return &SessionModeState{
 		CurrentModeID: string(s.currentMode()),
 		AvailableModes: []SessionMode{
 			{ID: string(agent.PermissionModeAuto), Name: "Auto", Description: "Run safe tools automatically; ask before risky ones."},
 			{ID: string(agent.PermissionModeAsk), Name: "Ask", Description: "Ask before every tool that changes state."},
+			{ID: string(agent.PermissionModePlan), Name: "Plan", Description: "Read-only planning; write and shell tools are hidden."},
 		},
 	}
 }
 
-// ---- 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},
-			})
+// resolveModelChoices advertises authenticated live models when discovery is
+// available. The configured model is always retained as the sole fallback.
+func (a *Agent) resolveModelChoices(ctx context.Context, cwd string) (string, []SessionConfigOptionValue, bool, error) {
+	resolved, err := a.deps.ResolveConfig(cwd, config.Overrides{})
+	if err != nil {
+		return "", nil, false, err
+	}
+	selected := strings.TrimSpace(resolved.Provider.Model)
+	options := make([]SessionConfigOptionValue, 0, 8)
+	seen := make(map[string]bool)
+	add := func(id, description string) {
+		id = strings.TrimSpace(id)
+		if id == "" || seen[id] {
+			return
+		}
+		seen[id] = true
+		options = append(options, SessionConfigOptionValue{Value: id, Name: id, Description: description})
+	}
+	add(selected, "")
+	descriptor, knownProvider := providercatalog.Get(resolved.Provider.CatalogID)
+	restrictModels := knownProvider && !descriptor.Custom
+	if a.deps.DiscoverModels != nil {
+		discovered, discoverErr := a.deps.DiscoverModels(ctx, resolved.Provider)
+		if ctx.Err() != nil {
+			return "", nil, false, ctx.Err()
+		}
+		if discoverErr == nil && len(discovered) > 0 {
+			for _, model := range discovered {
+				if providermodelcatalog.ModelIDAllowedForProvider(resolved.Provider.CatalogID, model.ID) {
+					add(model.ID, model.Description)
+				}
+			}
 		}
 	}
-	sess.appendHistory(turnRecord{user: user, assistant: assistant})
+	return selected, options, restrictModels, nil
 }
 
-func (a *Agent) loadHistory(sessionID string) []turnRecord {
+func modelChoiceExists(models []SessionConfigOptionValue, model string) bool {
+	for _, option := range models {
+		if option.Value == model {
+			return true
+		}
+	}
+	return false
+}
+
+func (a *Agent) configOptions(s *acpSession) []SessionConfigOption {
+	model, models, mode := s.configState()
+	return []SessionConfigOption{{
+		ID:           configIDModel,
+		Name:         "Model",
+		Description:  "Model used for this session.",
+		Category:     configCategoryModel,
+		Type:         configOptionTypeSelect,
+		CurrentValue: model,
+		Options:      models,
+	}, {
+		ID:           configIDMode,
+		Name:         "Mode",
+		Description:  "Permission mode used for this session.",
+		Category:     configCategoryMode,
+		Type:         configOptionTypeSelect,
+		CurrentValue: string(mode),
+		Options: []SessionConfigOptionValue{
+			{Value: string(agent.PermissionModeAuto), Name: "Auto", Description: "Run safe tools automatically; ask before risky ones."},
+			{Value: string(agent.PermissionModeAsk), Name: "Ask", Description: "Ask before every tool that changes state."},
+			{Value: string(agent.PermissionModePlan), Name: "Plan", Description: "Read-only planning; write and shell tools are hidden."},
+		},
+	}}
+}
+
+// ---- persistence + continuity ----
+
+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 +609,21 @@ 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
@@ -492,13 +671,13 @@ func promptImages(blocks []ContentBlock) []zeroruntime.ImageBlock {
 // session is returned unchanged rather than orphaning its turn or resetting its
 // mode/model. history is set BEFORE publishing so no concurrent prompt can read a
 // half-initialized session.
-func (a *Agent) registerSession(id, cwd string, history []turnRecord) *acpSession {
+func (a *Agent) registerSession(id, cwd string, history []turnRecord, model string, models []SessionConfigOptionValue, restrictModels bool) *acpSession {
 	a.mu.Lock()
 	defer a.mu.Unlock()
 	if existing := a.sessions[id]; existing != nil {
 		return existing
 	}
-	sess := &acpSession{id: id, cwd: cwd, mode: agent.PermissionModeAuto, history: history}
+	sess := &acpSession{id: id, cwd: cwd, mode: agent.PermissionModeAuto, model: model, models: models, restrictModels: restrictModels, history: history}
 	a.sessions[id] = sess
 	return sess
 }
@@ -537,7 +716,18 @@ func (s *acpSession) currentMode() agent.PermissionMode {
 }
 
 func (s *acpSession) setModel(model string) {
+	model = strings.TrimSpace(model)
 	s.mu.Lock()
+	found := false
+	for _, option := range s.models {
+		if option.Value == model {
+			found = true
+			break
+		}
+	}
+	if !found && model != "" {
+		s.models = append(s.models, SessionConfigOptionValue{Value: model, Name: model})
+	}
 	s.model = model
 	s.mu.Unlock()
 }
@@ -548,6 +738,24 @@ func (s *acpSession) currentModel() string {
 	return s.model
 }
 
+func (s *acpSession) hasModel(model string) bool {
+	model = strings.TrimSpace(model)
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	for _, option := range s.models {
+		if option.Value == model {
+			return true
+		}
+	}
+	return false
+}
+
+func (s *acpSession) configState() (string, []SessionConfigOptionValue, agent.PermissionMode) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	return s.model, append([]SessionConfigOptionValue(nil), s.models...), s.mode
+}
+
 func (s *acpSession) appendHistory(rec turnRecord) {
 	s.mu.Lock()
 	s.history = append(s.history, rec)

internal/acp/agent_test.go

@@ -3,14 +3,18 @@ package acp
 import (
 	"context"
 	"encoding/json"
+	"errors"
 	"fmt"
 	"io"
+	"os"
+	"path/filepath"
 	"strings"
 	"testing"
 	"time"
 
 	"github.com/Gitlawb/zero/internal/agent"
 	"github.com/Gitlawb/zero/internal/config"
+	"github.com/Gitlawb/zero/internal/providermodeldiscovery"
 	"github.com/Gitlawb/zero/internal/sandbox"
 	"github.com/Gitlawb/zero/internal/sessions"
 	"github.com/Gitlawb/zero/internal/tools"
@@ -135,6 +139,12 @@ func TestACPEndToEndPrompt(t *testing.T) {
 	if newRes.Modes == nil || newRes.Modes.CurrentModeID != string(agent.PermissionModeAuto) {
 		t.Fatalf("expected auto mode, got %+v", newRes.Modes)
 	}
+	if len(newRes.ConfigOptions) != 2 || newRes.ConfigOptions[0].ID != configIDModel || newRes.ConfigOptions[0].CurrentValue != "fake-model" {
+		t.Fatalf("model config option = %+v, want fake-model fallback", newRes.ConfigOptions)
+	}
+	if newRes.ConfigOptions[1].ID != configIDMode || newRes.ConfigOptions[1].CurrentValue != string(agent.PermissionModeAuto) {
+		t.Fatalf("mode config option = %+v", newRes.ConfigOptions[1])
+	}
 
 	// session/prompt
 	var promptRes PromptResult
@@ -154,6 +164,201 @@ func TestACPEndToEndPrompt(t *testing.T) {
 	}
 }
 
+func TestACPModelConfigOptionsCatalogSelectionAndLoad(t *testing.T) {
+	deps := testDeps(t)
+	deps.ResolveConfig = func(_ string, o config.Overrides) (config.ResolvedConfig, error) {
+		model := "gpt-5.5"
+		if o.Provider.Model != "" {
+			model = o.Provider.Model
+		}
+		return config.ResolvedConfig{Provider: config.ProviderProfile{
+			Name: "ChatGPT", CatalogID: "chatgpt", Model: model,
+		}}, nil
+	}
+	deps.DiscoverModels = func(_ context.Context, _ config.ProviderProfile) ([]providermodeldiscovery.Model, error) {
+		return []providermodeldiscovery.Model{
+			{ID: " gpt-5.4-mini ", Description: "Fast"},
+			{ID: "gpt-5.5", Description: "duplicate configured model"},
+		}, nil
+	}
+	h := newHarness(t, deps)
+	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
+	defer cancel()
+
+	var created NewSessionResult
+	if err := h.client.Call(ctx, MethodSessionNew, NewSessionParams{Cwd: t.TempDir()}, &created); err != nil {
+		t.Fatalf("session/new: %v", err)
+	}
+	option := created.ConfigOptions[0]
+	if option.CurrentValue != "gpt-5.5" || len(option.Options) < 2 {
+		t.Fatalf("new model option = %+v", option)
+	}
+	for _, choice := range option.Options {
+		if choice.Name != choice.Value {
+			t.Fatalf("model choice name = %q, want model id %q", choice.Name, choice.Value)
+		}
+	}
+	var selected SetSessionConfigOptionResult
+	if err := h.client.Call(ctx, MethodSessionSetConfigOption, SetSessionConfigOptionParams{
+		SessionID: created.SessionID, ConfigID: configIDModel, Value: "  gpt-5.4-mini  ",
+	}, &selected); err != nil {
+		t.Fatalf("set_config_option: %v", err)
+	}
+	if got := selected.ConfigOptions[0].CurrentValue; got != "gpt-5.4-mini" {
+		t.Fatalf("selected model = %q", got)
+	}
+	if err := h.client.Call(ctx, MethodSessionSetConfigOption, SetSessionConfigOptionParams{
+		SessionID: created.SessionID, ConfigID: configIDModel, Value: "not-advertised",
+	}, &SetSessionConfigOptionResult{}); err == nil {
+		t.Fatal("unknown standard model selection was accepted")
+	}
+	h.stop()
+	h = newHarness(t, deps)
+	defer h.stop()
+	var loaded LoadSessionResult
+	if err := h.client.Call(ctx, MethodSessionLoad, LoadSessionParams{SessionID: created.SessionID}, &loaded); err != nil {
+		t.Fatalf("session/load: %v", err)
+	}
+	if len(loaded.ConfigOptions) != 2 || loaded.ConfigOptions[0].CurrentValue != "gpt-5.4-mini" {
+		t.Fatalf("load model option = %+v", loaded.ConfigOptions)
+	}
+	if loaded.ConfigOptions[1].CurrentValue != string(agent.PermissionModeAuto) {
+		t.Fatalf("load mode option = %+v", loaded.ConfigOptions[1])
+	}
+}
+
+func TestACPModelDiscoveryFailureUsesConfiguredFallbackOnly(t *testing.T) {
+	deps := testDeps(t)
+	deps.ResolveConfig = func(_ string, _ config.Overrides) (config.ResolvedConfig, error) {
+		return config.ResolvedConfig{Provider: config.ProviderProfile{
+			Name: "ChatGPT", CatalogID: "chatgpt", Model: " configured-model ",
+		}}, nil
+	}
+	deps.DiscoverModels = func(context.Context, config.ProviderProfile) ([]providermodeldiscovery.Model, error) {
+		return nil, fmt.Errorf("discovery unavailable")
+	}
+	h := newHarness(t, deps)
+	defer h.stop()
+	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
+	defer cancel()
+
+	var created NewSessionResult
+	if err := h.client.Call(ctx, MethodSessionNew, NewSessionParams{Cwd: t.TempDir()}, &created); err != nil {
+		t.Fatalf("session/new: %v", err)
+	}
+	option := created.ConfigOptions[0]
+	if option.CurrentValue != "configured-model" || len(option.Options) != 1 || option.Options[0].Value != "configured-model" {
+		t.Fatalf("fallback model option = %+v", option)
+	}
+}
+
+func TestACPCustomProviderAllowsUnadvertisedModel(t *testing.T) {
+	deps := testDeps(t)
+	deps.ResolveConfig = func(_ string, o config.Overrides) (config.ResolvedConfig, error) {
+		model := "configured-model"
+		if o.Provider.Model != "" {
+			model = o.Provider.Model
+		}
+		return config.ResolvedConfig{Provider: config.ProviderProfile{
+			Name: "Custom", CatalogID: "custom-openai-compatible", Model: model,
+		}}, nil
+	}
+	h := newHarness(t, deps)
+	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
+	defer cancel()
+
+	var created NewSessionResult
+	if err := h.client.Call(ctx, MethodSessionNew, NewSessionParams{Cwd: t.TempDir()}, &created); err != nil {
+		t.Fatalf("session/new: %v", err)
+	}
+	var selected SetSessionConfigOptionResult
+	if err := h.client.Call(ctx, MethodSessionSetConfigOption, SetSessionConfigOptionParams{
+		SessionID: created.SessionID, ConfigID: configIDModel, Value: " vendor/model ",
+	}, &selected); err != nil {
+		t.Fatalf("set custom model: %v", err)
+	}
+	if got := selected.ConfigOptions[0].CurrentValue; got != "vendor/model" {
+		t.Fatalf("custom model = %q", got)
+	}
+	h.stop()
+	h = newHarness(t, deps)
+	defer h.stop()
+	var loaded LoadSessionResult
+	if err := h.client.Call(ctx, MethodSessionLoad, LoadSessionParams{SessionID: created.SessionID}, &loaded); err != nil {
+		t.Fatalf("session/load: %v", err)
+	}
+	option := loaded.ConfigOptions[0]
+	if option.CurrentValue != "vendor/model" || !modelChoiceExists(option.Options, "vendor/model") {
+		t.Fatalf("loaded custom model option = %+v", option)
+	}
+}
+
+func TestACPModelDiscoveryFiltersProviderIncompatibleModels(t *testing.T) {
+	a := &Agent{deps: Deps{
+		ResolveConfig: func(string, config.Overrides) (config.ResolvedConfig, error) {
+			return config.ResolvedConfig{Provider: config.ProviderProfile{
+				CatalogID: "opencode-go-anthropic-compatible", Model: "minimax-m3",
+			}}, nil
+		},
+		DiscoverModels: func(context.Context, config.ProviderProfile) ([]providermodeldiscovery.Model, error) {
+			return []providermodeldiscovery.Model{{ID: "qwen3.7-plus"}, {ID: "claude-sonnet-4.5"}}, nil
+		},
+	}}
+	_, options, restricted, err := a.resolveModelChoices(context.Background(), t.TempDir())
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !restricted || !modelChoiceExists(options, "qwen3.7-plus") || modelChoiceExists(options, "claude-sonnet-4.5") {
+		t.Fatalf("filtered model options = %+v, restricted=%v", options, restricted)
+	}
+}
+
+func TestACPModelDiscoveryHonorsCancellation(t *testing.T) {
+	ctx, cancel := context.WithCancel(context.Background())
+	cancel()
+	a := &Agent{deps: Deps{
+		ResolveConfig: func(string, config.Overrides) (config.ResolvedConfig, error) {
+			return config.ResolvedConfig{Provider: config.ProviderProfile{Model: "configured-model"}}, nil
+		},
+		DiscoverModels: func(context.Context, config.ProviderProfile) ([]providermodeldiscovery.Model, error) {
+			return nil, context.Canceled
+		},
+	}}
+	if _, _, _, err := a.resolveModelChoices(ctx, t.TempDir()); !errors.Is(err, context.Canceled) {
+		t.Fatalf("resolve error = %v, want context canceled", err)
+	}
+}
+
+func TestACPConfigOptionWireSchema(t *testing.T) {
+	b, err := json.Marshal(SessionConfigOption{
+		ID: "model", Name: "Model", Description: "desc", Category: "model",
+		Type: configOptionTypeSelect, CurrentValue: "m1",
+		Options: []SessionConfigOptionValue{{Value: "m1", Name: "m1", Description: "choice"}},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	var wire map[string]any
+	if err := json.Unmarshal(b, &wire); err != nil {
+		t.Fatal(err)
+	}
+	for _, key := range []string{"type", "category", "currentValue", "options"} {
+		if _, ok := wire[key]; !ok {
+			t.Errorf("wire field %q absent: %s", key, b)
+		}
+	}
+	if _, ok := wire["value"]; ok {
+		t.Errorf("obsolete option value field present: %s", b)
+	}
+	if _, ok := wire["values"]; ok {
+		t.Errorf("obsolete values field present: %s", b)
+	}
+	choice := wire["options"].([]any)[0].(map[string]any)
+	if choice["value"] != "m1" {
+		t.Errorf("options[].value = %#v", choice["value"])
+	}
+}
+
 func TestACPUnknownSessionPromptErrors(t *testing.T) {
 	h := newHarness(t, testDeps(t))
 	defer h.stop()
@@ -179,6 +384,41 @@ func TestACPSetModeUpdatesSession(t *testing.T) {
 	if err := h.client.Call(ctx, MethodSessionSetMode, SetSessionModeParams{SessionID: newRes.SessionID, ModeID: string(agent.PermissionModeAsk)}, &SetSessionModeResult{}); err != nil {
 		t.Fatalf("set_mode ask: %v", err)
 	}
+	var configured SetSessionConfigOptionResult
+	if err := h.client.Call(ctx, MethodSessionSetConfigOption, SetSessionConfigOptionParams{
+		SessionID: newRes.SessionID, ConfigID: configIDMode, Value: string(agent.PermissionModeAuto),
+	}, &configured); err != nil {
+		t.Fatalf("set_config_option mode: %v", err)
+	}
+	if got := configured.ConfigOptions[1].CurrentValue; got != string(agent.PermissionModeAuto) {
+		t.Fatalf("configured mode = %q", got)
+	}
+	// Plan is accepted: it only narrows capability (read-only), so unlike Unsafe
+	// there is no elevation risk in letting a client select it.
+	if err := h.client.Call(ctx, MethodSessionSetMode, SetSessionModeParams{SessionID: newRes.SessionID, ModeID: string(agent.PermissionModePlan)}, &SetSessionModeResult{}); err != nil {
+		t.Fatalf("set_mode plan: %v", err)
+	}
+	// Configuration path is a separate contract from MethodSessionSetMode: the
+	// mode option is advertised via configOptions and applied by handleSetConfigOption.
+	var planConfigured SetSessionConfigOptionResult
+	if err := h.client.Call(ctx, MethodSessionSetConfigOption, SetSessionConfigOptionParams{
+		SessionID: newRes.SessionID, ConfigID: configIDMode, Value: string(agent.PermissionModePlan),
+	}, &planConfigured); err != nil {
+		t.Fatalf("set_config_option plan: %v", err)
+	}
+	if got := planConfigured.ConfigOptions[1].CurrentValue; got != string(agent.PermissionModePlan) {
+		t.Fatalf("configured mode after plan = %q, want plan", got)
+	}
+	hasPlanOption := false
+	for _, opt := range planConfigured.ConfigOptions[1].Options {
+		if opt.Value == string(agent.PermissionModePlan) {
+			hasPlanOption = true
+			break
+		}
+	}
+	if !hasPlanOption {
+		t.Fatalf("config mode options missing plan: %#v", planConfigured.ConfigOptions[1].Options)
+	}
 	// Unsafe must be rejected over ACP — a client can't self-grant no-prompt host access.
 	if err := h.client.Call(ctx, MethodSessionSetMode, SetSessionModeParams{SessionID: newRes.SessionID, ModeID: string(agent.PermissionModeUnsafe)}, &SetSessionModeResult{}); err == nil {
 		t.Fatal("expected Unsafe mode to be rejected over ACP")
@@ -189,6 +429,37 @@ func TestACPSetModeUpdatesSession(t *testing.T) {
 	}
 }
 
+// TestACPPlanModeWiresPermissionModeIntoAgentOptions confirms selecting "plan"
+// over ACP actually reaches agent.Options.PermissionMode for the next turn —
+// the same gap this test's TUI counterpart covers for /plan on.
+func TestACPPlanModeWiresPermissionModeIntoAgentOptions(t *testing.T) {
+	deps := testDeps(t)
+	var captured agent.Options
+	deps.RunAgent = func(_ context.Context, _ string, _ zeroruntime.Provider, opts agent.Options) (agent.Result, error) {
+		captured = opts
+		return agent.Result{FinalAnswer: "ok"}, nil
+	}
+
+	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)
+	}
+	if err := h.client.Call(ctx, MethodSessionSetMode, SetSessionModeParams{SessionID: newRes.SessionID, ModeID: string(agent.PermissionModePlan)}, &SetSessionModeResult{}); err != nil {
+		t.Fatalf("set_mode plan: %v", err)
+	}
+	if err := h.client.Call(ctx, MethodSessionPrompt, PromptParams{SessionID: newRes.SessionID, Prompt: []ContentBlock{TextBlock("plan it out")}}, &PromptResult{}); err != nil {
+		t.Fatalf("session/prompt: %v", err)
+	}
+	if captured.PermissionMode != agent.PermissionModePlan {
+		t.Fatalf("agent.Options.PermissionMode = %q, want plan", captured.PermissionMode)
+	}
+}
+
 // TestACPRunTurnWiresSandboxAndScopedRegistry proves the sandbox engine and the
 // scoped registry from BuildWorkspace actually reach agent.Options — i.e. ACP
 // shell tools run confined, not unconfined on the host.
@@ -242,8 +513,77 @@ 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 +591,7 @@ 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/desktopinterop_test.go

@@ -0,0 +1,217 @@
+package acp
+
+// desktopinterop_test.go covers three defects found by driving the real `zero
+// acp` binary from a desktop ACP client. Each is a case where ZERO and a
+// conforming client disagreed about the wire, and none of them failed loudly —
+// they produced a silent denial, a spurious crash, or two buttons the user
+// could not tell apart.
+//
+// A fourth, session/load restoring history without replaying it, is fixed by
+// #914 rather than here: that one also gives the replayed messages stable ids
+// and keeps session/resume replay-free, which this did not.
+
+import (
+	"errors"
+	"fmt"
+	"testing"
+
+	"github.com/Gitlawb/zero/internal/agent"
+)
+
+// ---- an offered option is an acceptable answer ----
+
+// The list the options were BUILT from and the list the answer was CHECKED
+// against were different whenever ZERO did not enumerate: the client was sent
+// Allow and Reject, and its reply was validated against an empty slice, so
+// every button it could possibly show failed closed to deny. The user clicked
+// Allow and ZERO recorded a denial, with nothing on screen to say so.
+func TestAnOfferedOptionIsAccepted(t *testing.T) {
+	// No AvailableDecisions: the fallback path, which is every permission event
+	// that is not a prompt.
+	req := agent.PermissionRequest{ToolName: "bash"}
+
+	options := buildPermissionOptions(req)
+	if len(options) == 0 {
+		t.Fatal("no options were offered")
+	}
+
+	for _, option := range options {
+		decision := decisionFromOutcome(
+			RequestPermissionOutcome{Outcome: OutcomeSelected, OptionID: option.OptionID},
+			buildPermissionOptions(req),
+		)
+		if string(decision.Action) != option.OptionID {
+			t.Fatalf("option %q was answered with %q (%s) — an offered option must be accepted",
+				option.OptionID, decision.Action, decision.Reason)
+		}
+	}
+}
+
+// The narrowing this validation exists for still holds: a client must not be
+// able to return a broader grant than it was shown.
+func TestAnOptionThatWasNotOfferedIsStillRefused(t *testing.T) {
+	req := agent.PermissionRequest{ToolName: "bash"}
+
+	decision := decisionFromOutcome(
+		RequestPermissionOutcome{Outcome: OutcomeSelected, OptionID: string(agent.PermissionDecisionAlwaysAllow)},
+		buildPermissionOptions(req),
+	)
+	if decision.Action != agent.PermissionDecisionDeny {
+		t.Fatalf("an unoffered broader grant was accepted as %q", decision.Action)
+	}
+}
+
+// The resolver is the single source of truth for both sides, which is what
+// makes the two agree by construction rather than by being kept in step.
+func TestOfferedDecisionsMatchWhatWasSent(t *testing.T) {
+	for _, req := range []agent.PermissionRequest{
+		{ToolName: "bash"},
+		{ToolName: "bash", AvailableDecisions: []agent.PermissionDecisionAction{
+			agent.PermissionDecisionAllow,
+			agent.PermissionDecisionAllowForSession,
+			agent.PermissionDecisionDeny,
+		}},
+	} {
+		offered := buildPermissionOptions(req)
+		for _, option := range buildPermissionOptions(req) {
+			if !actionOffered(option.OptionID, offered) {
+				t.Fatalf("option %q was sent but is not in the offered set %v", option.OptionID, offered)
+			}
+		}
+	}
+}
+
+// ---- two options the user can tell apart ----
+
+// request_permissions offers plain allow and strict-review allow together, and
+// both were labelled "Allow". The label is the only thing an ACP client shows,
+// so the panel presented two identical buttons where one silently enabled
+// strict auto-review of what was granted.
+func TestEveryOfferedOptionHasADistinctLabel(t *testing.T) {
+	req := agent.PermissionRequest{
+		ToolName: "request_permissions",
+		AvailableDecisions: []agent.PermissionDecisionAction{
+			agent.PermissionDecisionAllow,
+			agent.PermissionDecisionAllowStrict,
+			agent.PermissionDecisionAllowForSession,
+			agent.PermissionDecisionDeny,
+		},
+	}
+
+	seen := map[string]string{}
+	for _, option := range buildPermissionOptions(req) {
+		if previous, clash := seen[option.Name]; clash {
+			t.Fatalf("options %q and %q share the label %q — a user cannot tell them apart",
+				previous, option.OptionID, option.Name)
+		}
+		seen[option.Name] = option.OptionID
+	}
+}
+
+// The two remain distinct decisions: the label changed, the round trip did not.
+func TestStrictAllowStillRoundTrips(t *testing.T) {
+	req := agent.PermissionRequest{
+		ToolName: "request_permissions",
+		AvailableDecisions: []agent.PermissionDecisionAction{
+			agent.PermissionDecisionAllow,
+			agent.PermissionDecisionAllowStrict,
+		},
+	}
+	decision := decisionFromOutcome(
+		RequestPermissionOutcome{Outcome: OutcomeSelected, OptionID: string(agent.PermissionDecisionAllowStrict)},
+		buildPermissionOptions(req),
+	)
+	if decision.Action != agent.PermissionDecisionAllowStrict {
+		t.Fatalf("strict allow round-tripped as %q", decision.Action)
+	}
+}
+
+// ---- cancelling a permission is a cancellation, not a crash ----
+
+// Only context.Canceled was recognised, so dismissing a permission dialog came
+// back as JSON-RPC -32603 carrying the internal sentinel text. Clients render
+// that as a failed turn, so declining a tool looked like ZERO falling over —
+// and for apply_patch, dismissing is the ONLY refusal a client is offered.
+func TestCancellingAPermissionEndsTheTurnAsCancelled(t *testing.T) {
+	canceled := fmt.Errorf("%w for apply_patch", agent.ErrPermissionApprovalCanceled)
+
+	reason, err := stopReasonFor(agent.Result{}, canceled)
+	if err != nil {
+		t.Fatalf("stopReasonFor returned an error for a cancellation: %v", err)
+	}
+	if reason != StopCancelled {
+		t.Fatalf("stop reason = %q, want %q", reason, StopCancelled)
+	}
+}
+
+// A genuine failure must still be an error. Mapping too much to cancelled would
+// hide real faults behind a clean ending, which is the opposite mistake.
+func TestARealFailureIsStillAnError(t *testing.T) {
+	failure := errors.New("provider exploded")
+
+	if _, err := stopReasonFor(agent.Result{}, failure); !errors.Is(err, failure) {
+		t.Fatalf("stopReasonFor(%v) = %v, want the failure preserved", failure, err)
+	}
+}
+
+// The sentinel is matched through errors.Is, so it survives the wrapping every
+// return site applies — each one adds the tool name.
+func TestPermissionCancellationSurvivesWrapping(t *testing.T) {
+	if !errors.Is(fmt.Errorf("%w for bash", agent.ErrPermissionApprovalCanceled), agent.ErrPermissionApprovalCanceled) {
+		t.Fatal("a wrapped permission cancellation was not recognised")
+	}
+	if errors.Is(errors.New("something else"), agent.ErrPermissionApprovalCanceled) {
+		t.Fatal("an unrelated error was reported as a permission cancellation")
+	}
+}
+
+// A decision that never became an option cannot be selected.
+//
+// PermissionDecisionCancel is the case that exists today: optionKindFor drops
+// it because ACP expresses cancellation through the outcome, yet ZERO
+// enumerates it in AvailableDecisions for shell commands and apply_patch. While
+// the reply was validated against the DECISIONS rather than the sent options, a
+// client could return {"outcome":"selected","optionId":"cancel"} — an id it was
+// never shown — and abort the whole turn.
+//
+// Written over every option-less decision rather than over "cancel", so a
+// future action that optionKindFor declines to render is covered the day it is
+// added rather than the day someone remembers this test.
+func TestADecisionThatIsNotAnOptionCannotBeSelected(t *testing.T) {
+	for _, toolName := range []string{"bash", "apply_patch"} {
+		req := agent.PermissionRequest{
+			ToolName: toolName,
+			AvailableDecisions: []agent.PermissionDecisionAction{
+				agent.PermissionDecisionAllow,
+				agent.PermissionDecisionDeny,
+				agent.PermissionDecisionCancel,
+			},
+		}
+		options := buildPermissionOptions(req)
+
+		for _, action := range req.AvailableDecisions {
+			if actionOffered(string(action), options) {
+				continue // it was sent; selecting it is legitimate
+			}
+			decision := decisionFromOutcome(
+				RequestPermissionOutcome{Outcome: OutcomeSelected, OptionID: string(action)},
+				options,
+			)
+			if decision.Action != agent.PermissionDecisionDeny {
+				t.Fatalf("%s: %q was never sent as an option but was accepted as %q",
+					toolName, action, decision.Action)
+			}
+		}
+	}
+}
+
+// Cancelling is still reachable — through the outcome, which is where ACP puts
+// it. Closing the selected-id route must not close the real one.
+func TestCancellingThroughTheOutcomeStillWorks(t *testing.T) {
+	options := buildPermissionOptions(agent.PermissionRequest{ToolName: "apply_patch"})
+
+	decision := decisionFromOutcome(RequestPermissionOutcome{Outcome: OutcomeCancelled}, options)
+	if decision.Action != agent.PermissionDecisionCancel {
+		t.Fatalf("outcome=cancelled gave %q, want cancel", decision.Action)
+	}
+}

internal/acp/export_test.go

@@ -0,0 +1,11 @@
+// Test seams: helpers only test code uses, kept out of the production binary.
+package acp
+
+import "context"
+
+// Serve runs the connection read loop until the stream closes or ctx is done.
+func (a *Agent) Serve(ctx context.Context) error { return a.conn.Serve(ctx) }
+
+func ImageBlock(base64Data, mimeType string) ContentBlock {
+	return ContentBlock{Type: "image", Data: base64Data, MimeType: mimeType}
+}

internal/acp/jsonrpc.go

@@ -17,6 +17,7 @@ import (
 	"fmt"
 	"io"
 	"sync"
+	"sync/atomic"
 )
 
 // JSON-RPC 2.0 standard error codes.
@@ -80,8 +81,9 @@ type NotifyFunc func(ctx context.Context, params json.RawMessage)
 // requests/notifications — needed because ACP is bidirectional (the agent calls
 // the client for session/request_permission, fs/*, terminal/*).
 type Conn struct {
-	reader *bufio.Reader
-	w      io.Writer
+	rawReader    io.Reader // wrapped lazily in Serve, once ctx is known — see interruptibleReader
+	readerCloser io.Closer
+	w            io.Writer
 
 	writeMu sync.Mutex // serializes all writes to w
 
@@ -96,10 +98,15 @@ type Conn struct {
 	wg sync.WaitGroup // tracks in-flight inbound handlers
 }
 
-// NewConn builds a peer reading ndjson from r and writing ndjson to w.
+// NewConn builds a peer reading ndjson from r and writing ndjson to w. This
+// Conn does not own r: cancelling Serve's context cannot forcibly close it, so
+// a Read on r that's genuinely idle (never returns on its own) leaves Serve
+// blocked until r itself produces something — e.g. a test that closes r's own
+// write side to unblock it. Use NewOwnedConn when this Conn should own r's
+// lifetime.
 func NewConn(r io.Reader, w io.Writer) *Conn {
 	return &Conn{
-		reader:    bufio.NewReader(r),
+		rawReader: r,
 		w:         w,
 		handlers:  make(map[string]HandlerFunc),
 		notifiers: make(map[string]NotifyFunc),
@@ -107,6 +114,18 @@ func NewConn(r io.Reader, w io.Writer) *Conn {
 	}
 }
 
+// NewOwnedConn is like NewConn but declares that this Conn exclusively owns r:
+// if r implements io.Closer, Serve closes it on context cancellation to
+// interrupt an otherwise-idle blocking Read. Only pass a reader whose lifetime
+// this Conn should control — e.g. the process's own stdin — never a reader a
+// caller retains or shares elsewhere, since cancellation tears it down
+// permanently for every consumer, not just this Conn.
+func NewOwnedConn(r io.Reader, w io.Writer) *Conn {
+	c := NewConn(r, w)
+	c.readerCloser, _ = r.(io.Closer)
+	return c
+}
+
 // Handle registers a request handler for method.
 func (c *Conn) Handle(method string, fn HandlerFunc) { c.handlers[method] = fn }
 
@@ -128,17 +147,36 @@ func (c *Conn) Serve(ctx context.Context) error {
 		c.wg.Wait()
 	}()
 
+	interruptible := newInterruptibleReader(ctx, c.rawReader, c.readerCloser)
+	reader := bufio.NewReader(interruptible)
+
 	for {
 		// One ndjson value per line. Framing per line (rather than streaming the
 		// decoder) keeps a single malformed line from making the whole connection
 		// unrecoverable — we report -32700 and continue.
-		line, err := c.reader.ReadBytes('\n')
+		//
+		// generation brackets this SPECIFIC ReadBytes call: bufio may invoke
+		// interruptible.Read zero or more times underneath it (zero when the
+		// answer is already sitting in bufio's own buffer — including a
+		// previously-buffered, not-yet-surfaced error: bufio can return a
+		// complete line with a nil error while quietly caching an error it
+		// received ALONGSIDE that line's bytes, which then surfaces on the VERY
+		// NEXT call with no further contact with interruptible.Read at all). The
+		// generation only advances inside interruptible.Read's own cancellation
+		// branch, so "unchanged across this call" is proof this call's outcome
+		// didn't pass through that branch — whether because it was answered from
+		// the buffer or because it raced ctx.Done() and won. Only that proof
+		// makes it safe to call the outcome genuine.
+		before := interruptible.generation()
+		line, err := reader.ReadBytes('\n')
+		interrupted := interruptible.generation() != before
+
 		if len(bytes.TrimSpace(line)) > 0 {
 			c.handleLine(ctx, line)
 		}
 		if err != nil {
 			c.failAllPending(err)
-			if errors.Is(err, io.EOF) || ctx.Err() != nil {
+			if errors.Is(err, io.EOF) || interrupted {
 				return nil
 			}
 			return err
@@ -146,6 +184,100 @@ func (c *Conn) Serve(ctx context.Context) error {
 	}
 }
 
+// interruptibleReader wraps a reader so a context cancellation can unblock a
+// currently in-flight Read without ever discarding — or misattributing — the
+// real outcome of a call that wasn't actually blocked.
+//
+// Each Read runs the underlying call on its own goroutine and races it against
+// ctx.Done(). If the underlying call finishes first, its result is returned
+// exactly as received; ctx being cancelled a moment later has no bearing on a
+// call that already had its answer. If ctx.Done() fires first, closer (when
+// non-nil) is closed to nudge a genuinely blocked read, the generation counter
+// is bumped, and the call STILL waits for and returns the underlying call's own
+// result — never a synthesized one — since closing a reader's outcome is
+// whatever the reader itself reports for it, not ours to invent.
+//
+// The counter is what lets Serve tell a call that was actually raced against
+// cancellation apart from one merely answered while cancellation happened to be
+// pending — the distinction the read loop's own doc comment explains.
+type interruptibleReader struct {
+	inner  io.Reader
+	closer io.Closer
+	ctx    context.Context
+
+	closeOnce sync.Once
+	gen       atomic.Int64
+}
+
+func newInterruptibleReader(ctx context.Context, inner io.Reader, closer io.Closer) *interruptibleReader {
+	return &interruptibleReader{inner: inner, closer: closer, ctx: ctx}
+}
+
+func (r *interruptibleReader) generation() int64 { return r.gen.Load() }
+
+type interruptibleReadResult struct {
+	n   int
+	err error
+}
+
+// interruptibleReadDecision arbitrates whether an underlying Read completed or
+// cancellation claimed it first. Completion is recorded before the result is
+// published so cancellation cannot misclassify a finished read merely because
+// its goroutine has not sent the result yet.
+type interruptibleReadDecision struct {
+	once sync.Once
+}
+
+func (d *interruptibleReadDecision) complete() { d.once.Do(func() {}) }
+
+func (d *interruptibleReadDecision) interrupt() (claimed bool) {
+	d.once.Do(func() { claimed = true })
+	return claimed
+}
+
+// readInner records completion in the Read call's deferred epilogue. It also
+// converts a reader panic into a terminal error so the helper goroutine always
+// publishes a result for Read to join.
+func (r *interruptibleReader) readInner(p []byte, decision *interruptibleReadDecision) (res interruptibleReadResult) {
+	defer func() {
+		if recovered := recover(); recovered != nil {
+			res = interruptibleReadResult{err: fmt.Errorf("acp: reader panicked: %v", recovered)}
+		}
+		decision.complete()
+	}()
+	res.n, res.err = r.inner.Read(p)
+	return res
+}
+
+func (r *interruptibleReader) Read(p []byte) (int, error) {
+	resultCh := make(chan interruptibleReadResult, 1)
+	var decision interruptibleReadDecision
+	go func() {
+		resultCh <- r.readInner(p, &decision)
+	}()
+	select {
+	case res := <-resultCh:
+		return res.n, res.err
+	case <-r.ctx.Done():
+		// resultCh may not be readable yet even when inner.Read has returned: its
+		// goroutine can be descheduled between recording completion and publishing
+		// the result. Only interrupt a call that cancellation actually claims.
+		if decision.interrupt() {
+			r.gen.Add(1)
+			if r.closer != nil {
+				r.closeOnce.Do(func() { _ = r.closer.Close() })
+			}
+		}
+		// Wait for the real result rather than returning synthetically: p is
+		// still being written by the goroutine above until it does, so returning
+		// early would race that write, and the actual error (whatever closing
+		// the reader produced, or whatever the reader was already about to
+		// report) is what the caller should see.
+		res := <-resultCh
+		return res.n, res.err
+	}
+}
+
 // handleLine decodes and dispatches one ndjson frame. A parse failure replies
 // with -32700 (id null) and keeps the connection alive.
 func (c *Conn) handleLine(ctx context.Context, line []byte) {

internal/acp/jsonrpc_test.go

@@ -3,11 +3,30 @@ package acp
 import (
 	"context"
 	"encoding/json"
+	"errors"
 	"io"
+	"sync"
+	"sync/atomic"
 	"testing"
 	"time"
 )
 
+type testReadCloser struct {
+	read  func([]byte) (int, error)
+	close func() error
+}
+
+func (r testReadCloser) Read(p []byte) (int, error) { return r.read(p) }
+func (r testReadCloser) Close() error               { return r.close() }
+
+type testReader func([]byte) (int, error)
+
+func (r testReader) Read(p []byte) (int, error) { return r(p) }
+
+type testWriter func([]byte) (int, error)
+
+func (w testWriter) Write(p []byte) (int, error) { return w(p) }
+
 // connPair wires two Conns together over in-memory pipes and serves both.
 func connPair(t *testing.T) (a, b *Conn, stop func()) {
 	t.Helper()
@@ -25,6 +44,275 @@ func connPair(t *testing.T) (a, b *Conn, stop func()) {
 	}
 }
 
+func TestConnServeCancellationInterruptsIdleRead(t *testing.T) {
+	pipeReader, writer := io.Pipe()
+	t.Cleanup(func() { _ = writer.Close() })
+	// Buffered + non-blocking send, not close: interruptibleReader retries each
+	// underlying Read on its own goroutine, so if bufio ever calls this more than
+	// once, a close() here would panic on the second call.
+	readStarted := make(chan struct{}, 1)
+	closeCalls := make(chan struct{}, 2)
+	reader := testReadCloser{
+		read: func(p []byte) (int, error) {
+			select {
+			case readStarted <- struct{}{}:
+			default:
+			}
+			return pipeReader.Read(p)
+		},
+		close: func() error {
+			closeCalls <- struct{}{}
+			return pipeReader.Close()
+		},
+	}
+	conn := NewOwnedConn(reader, io.Discard)
+	ctx, cancel := context.WithCancel(context.Background())
+	done := make(chan error, 1)
+	go func() { done <- conn.Serve(ctx) }()
+
+	<-readStarted
+	cancel()
+	select {
+	case err := <-done:
+		if err != nil {
+			t.Fatalf("Serve returned %v after cancellation, want nil", err)
+		}
+	case <-time.After(time.Second):
+		t.Fatal("Serve did not return after cancelling an idle connection")
+	}
+	if got := len(closeCalls); got != 1 {
+		t.Fatalf("reader Close calls = %d, want 1", got)
+	}
+}
+
+func TestConnServePreservesTerminalReadErrorDuringCancellation(t *testing.T) {
+	for _, closable := range []bool{false, true} {
+		name := "non-closable"
+		if closable {
+			name = "closable"
+		}
+		t.Run(name, func(t *testing.T) {
+			wantErr := errors.New("read failed")
+			read := testReader(func(p []byte) (int, error) {
+				return copy(p, "not json"), wantErr
+			})
+			closeCalls := make(chan struct{}, 1)
+			var reader io.Reader = read
+			if closable {
+				reader = testReadCloser{
+					read: read,
+					close: func() error {
+						closeCalls <- struct{}{}
+						return nil
+					},
+				}
+			}
+			// Buffered + non-blocking send, not close: a later change adding a
+			// second write (e.g. an additional error frame) must not panic on a
+			// repeated close of an already-closed channel.
+			writeStarted := make(chan struct{}, 1)
+			releaseWrite := make(chan struct{})
+			writer := testWriter(func(p []byte) (int, error) {
+				select {
+				case writeStarted <- struct{}{}:
+				default:
+				}
+				<-releaseWrite
+				return len(p), nil
+			})
+			newConn := NewConn
+			if closable {
+				// Ownership matters here: this case exists to prove that even a
+				// Conn entitled to close its reader on cancellation still
+				// preserves a terminal error that was already decided.
+				newConn = NewOwnedConn
+			}
+			conn := newConn(reader, writer)
+			ctx, cancel := context.WithCancel(context.Background())
+			done := make(chan error, 1)
+			go func() { done <- conn.Serve(ctx) }()
+
+			<-writeStarted
+			cancel()
+			close(releaseWrite)
+			select {
+			case err := <-done:
+				if !errors.Is(err, wantErr) {
+					t.Fatalf("Serve returned %v, want terminal read error %v", err, wantErr)
+				}
+			case <-time.After(time.Second):
+				t.Fatal("Serve did not return after terminal read error")
+			}
+			if got := len(closeCalls); got != 0 {
+				t.Fatalf("reader Close calls = %d, want 0 after terminal read", got)
+			}
+		})
+	}
+}
+
+// TestConnServePreservesReadErrorBufferedAlongsideAPriorLine is the deterministic
+// regression test for jatmn's #782 finding: a real, non-EOF ReadBytes failure can
+// be DECIDED before cancellation ever happens, yet only SURFACE afterward, and
+// must still be reported rather than swallowed as a clean shutdown.
+//
+// bufio.Reader can obtain data AND a terminal error from ONE underlying Read call
+// (io.Reader explicitly permits returning (n>0, err) together). If a delimiter is
+// found within that data, ReadBytes returns the line with a nil error for THIS
+// call, but bufio caches the error internally and returns it — WITHOUT calling the
+// underlying reader again — on the very next ReadBytes call. This test forces that
+// exact sequence: the underlying reader hands back one complete, validly-framed
+// line together with wantErr in a SINGLE call (asserted below to be its ONLY
+// call), the loop dispatches that line through a synchronous path that blocks the
+// read loop, cancellation is asserted to have happened before the loop attempts
+// its next read, and only then is the block released. By the time the read loop
+// asks for more input, cancellation is already in effect and the underlying
+// reader is never touched again — so if Serve attributed the resulting error to
+// cancellation merely because ctx was already done, it would incorrectly return
+// nil. It must instead recognize that this read was answered entirely from
+// bufio's own cache, never raced against ctx.Done(), and report wantErr.
+func TestConnServePreservesReadErrorBufferedAlongsideAPriorLine(t *testing.T) {
+	wantErr := errors.New("read failed")
+	var readCalls atomic.Int32
+	read := testReader(func(p []byte) (int, error) {
+		readCalls.Add(1)
+		// An unsupported jsonrpc version with an id takes the SYNCHRONOUS
+		// writeError path in handleLine (not a dispatched goroutine), giving this
+		// test a reliable blocking point between processing this line and the
+		// loop's next ReadBytes call.
+		return copy(p, `{"jsonrpc":"1.0","id":1}`+"\n"), wantErr
+	})
+	writeStarted := make(chan struct{}, 1)
+	releaseWrite := make(chan struct{})
+	var releaseWriteOnce sync.Once
+	release := func() { releaseWriteOnce.Do(func() { close(releaseWrite) }) }
+	t.Cleanup(release)
+	writer := testWriter(func(p []byte) (int, error) {
+		select {
+		case writeStarted <- struct{}{}:
+		default:
+		}
+		<-releaseWrite
+		return len(p), nil
+	})
+	conn := NewConn(read, writer)
+	ctx, cancel := context.WithCancel(context.Background())
+	done := make(chan error, 1)
+	go func() { done <- conn.Serve(ctx) }()
+
+	// The synchronous write confirms the first (and, per the panic guard above,
+	// only) read has already returned and is being processed — cancelling now
+	// lands squarely in the gap between that read and the loop's next one, never
+	// racing an in-flight interruptibleReader.Read call.
+	select {
+	case <-writeStarted:
+	case <-time.After(time.Second):
+		t.Fatal("Serve did not reach the synchronous write")
+	}
+	cancel()
+	release()
+
+	select {
+	case err := <-done:
+		if !errors.Is(err, wantErr) {
+			t.Fatalf("Serve returned %v, want the buffered terminal read error %v", err, wantErr)
+		}
+	case <-time.After(time.Second):
+		t.Fatal("Serve did not return after the buffered terminal read error")
+	}
+	if got := readCalls.Load(); got != 1 {
+		t.Fatalf("underlying Read calls = %d, want exactly 1 (bufio must have served the second ReadBytes from its cache)", got)
+	}
+}
+
+// TestInterruptibleReaderPrefersDecidedResultOverCancellation is the regression
+// test for jatmn's second #782 finding: interruptibleReader.Read's select
+// between resultCh and ctx.Done() chooses pseudo-randomly when both are ready,
+// so a real, already-decided outcome (e.g. a terminal transport error) could be
+// misattributed to cancellation purely because ctx also happened to be done by
+// the time the select ran. Each iteration lets the underlying Read return (so
+// resultCh already holds the real outcome) before cancelling, maximizing the
+// odds of landing exactly in that ambiguous window; the generation counter must
+// never advance when the real, already-decided result is what gets returned.
+func TestInterruptibleReaderPrefersDecidedResultOverCancellation(t *testing.T) {
+	wantErr := errors.New("read failed")
+	for i := 0; i < 100; i++ {
+		resultReady := make(chan struct{})
+		read := testReader(func(p []byte) (int, error) {
+			defer close(resultReady)
+			return copy(p, "x"), wantErr
+		})
+		ctx, cancel := context.WithCancel(context.Background())
+		r := newInterruptibleReader(ctx, read, nil)
+
+		readDone := make(chan interruptibleReadResult, 1)
+		go func() {
+			n, err := r.Read(make([]byte, 8))
+			readDone <- interruptibleReadResult{n, err}
+		}()
+
+		<-resultReady
+		time.Sleep(time.Millisecond) // bias resultCh's send ahead of cancel
+		cancel()
+
+		res := <-readDone
+		if !errors.Is(res.err, wantErr) {
+			t.Fatalf("iteration %d: Read returned %v, want %v", i, res.err, wantErr)
+		}
+		if got := r.generation(); got != 0 {
+			t.Fatalf("iteration %d: generation = %d, want 0 (a decided result must never be attributed to cancellation)", i, got)
+		}
+		cancel()
+	}
+}
+
+// TestInterruptibleReadCompletionWinsBeforeResultPublication exercises the
+// actual inner-read wrapper at the handoff between the underlying Read and the
+// helper goroutine's result publication. Cancellation must not reclassify a
+// result returned by this wrapper as interrupted.
+func TestInterruptibleReadCompletionWinsBeforeResultPublication(t *testing.T) {
+	wantErr := errors.New("read failed")
+	read := testReader(func(p []byte) (int, error) {
+		return copy(p, "x"), wantErr
+	})
+	r := newInterruptibleReader(context.Background(), read, nil)
+	var decision interruptibleReadDecision
+	readComplete := make(chan struct{})
+	resultCh := make(chan interruptibleReadResult)
+	go func() {
+		res := r.readInner(make([]byte, 8), &decision)
+		close(readComplete)
+		resultCh <- res
+	}()
+
+	<-readComplete
+	if decision.interrupt() {
+		t.Fatal("cancellation claimed a read that completed before result publication")
+	}
+	res := <-resultCh
+	if res.n != 1 || !errors.Is(res.err, wantErr) {
+		t.Fatalf("inner read result = (%d, %v), want (1, %v)", res.n, res.err, wantErr)
+	}
+}
+
+func TestConnServeReportsReaderPanicAsTerminalError(t *testing.T) {
+	read := testReader(func([]byte) (int, error) {
+		panic("boom")
+	})
+	conn := NewConn(read, io.Discard)
+	done := make(chan error, 1)
+	go func() { done <- conn.Serve(context.Background()) }()
+
+	select {
+	case err := <-done:
+		const want = "acp: reader panicked: boom"
+		if err == nil || err.Error() != want {
+			t.Fatalf("Serve returned %v, want %q", err, want)
+		}
+	case <-time.After(time.Second):
+		t.Fatal("Serve did not return after reader panic")
+	}
+}
+
 func TestConnRequestResponse(t *testing.T) {
 	a, b, stop := connPair(t)
 	defer stop()

internal/acp/permission.go

@@ -15,17 +15,10 @@ import (
 // PermissionOptions. Only the actions ZERO actually presented (AvailableDecisions)
 // are surfaced; the optionId is the ZERO action string for a clean round-trip.
 func buildPermissionOptions(req agent.PermissionRequest) []PermissionOption {
-	actions := req.AvailableDecisions
-	if len(actions) == 0 {
-		// Sensible default if ZERO didn't enumerate: allow once / reject.
-		actions = []agent.PermissionDecisionAction{
-			agent.PermissionDecisionAllow,
-			agent.PermissionDecisionDeny,
-		}
-	}
+	actions := offeredDecisions(req)
 	options := make([]PermissionOption, 0, len(actions))
 	for _, action := range actions {
-		kind, name := optionKindFor(action)
+		kind, name := optionKindFor(action, req.PrefixApprovalEscalates)
 		if kind == "" {
 			continue // skip actions that have no clean ACP option (e.g. cancel)
 		}
@@ -38,21 +31,61 @@ func buildPermissionOptions(req agent.PermissionRequest) []PermissionOption {
 	return options
 }
 
+// offeredDecisions resolves what this request actually offers the client.
+//
+// ONE RESOLVER, USED BY BOTH SIDES, and that is the entire point of it existing.
+// The fallback below used to live inside buildPermissionOptions while
+// requestPermission validated the client's answer against the raw
+// req.AvailableDecisions. When ZERO did not enumerate — which is every
+// permission event that is not a prompt, and includes a shell call whose sandbox
+// decision came back deny — the two disagreed completely: the client was sent
+// "Allow" and "Reject", and validation was performed against an EMPTY list, so
+// every option the client could possibly click failed closed to deny with
+// "permission option was not offered".
+//
+// The user clicked Allow and ZERO recorded a denial. Nothing surfaced it,
+// because a denial is a legitimate answer — the tool was simply refused and the
+// turn carried on as though the user had rejected it.
+func offeredDecisions(req agent.PermissionRequest) []agent.PermissionDecisionAction {
+	if len(req.AvailableDecisions) > 0 {
+		return req.AvailableDecisions
+	}
+	// Sensible default if ZERO didn't enumerate: allow once / reject.
+	return []agent.PermissionDecisionAction{
+		agent.PermissionDecisionAllow,
+		agent.PermissionDecisionDeny,
+	}
+}
+
 // optionKindFor maps a ZERO decision action to an ACP PermissionOptionKind and a
 // human label. Returns an empty kind for actions that ACP expresses through the
 // outcome rather than an option (cancel).
-func optionKindFor(action agent.PermissionDecisionAction) (kind, name string) {
+//
+// escalates reports that approving a command prefix will also run the command
+// outside the sandbox. The two prefix options say so in their label, since the
+// label is all an ACP client shows and the option id carries no room for it. An
+// editor user is deciding on exactly the same terms as a terminal user, so the
+// disclosure cannot be TUI-only.
+func optionKindFor(action agent.PermissionDecisionAction, escalates bool) (kind, name string) {
 	switch action {
-	case agent.PermissionDecisionAllow, agent.PermissionDecisionAllowStrict:
+	case agent.PermissionDecisionAllow:
 		return PermAllowOnce, "Allow"
+	case agent.PermissionDecisionAllowStrict:
+		// SPLIT FROM PLAIN ALLOW because the label is the only thing an ACP
+		// client shows. Both were labelled "Allow", and request_permissions
+		// offers them together — so the panel presented two identical buttons,
+		// and whichever the user picked looked the same while one of them
+		// silently turned on strict auto-review of what was granted. The option
+		// id carries no room for the distinction, so it has to be in the name.
+		return PermAllowOnce, "Allow with strict review"
 	case agent.PermissionDecisionAllowForSession:
 		return PermAllowAlways, "Allow for this session"
 	case agent.PermissionDecisionAllowPrefix:
-		return PermAllowAlways, "Allow this command for the session"
+		return PermAllowAlways, withSandboxEscalationNote("Allow this command for the session", escalates)
 	case agent.PermissionDecisionAlwaysAllow:
 		return PermAllowAlways, "Always allow"
 	case agent.PermissionDecisionAlwaysAllowPrefix:
-		return PermAllowAlways, "Always allow this command"
+		return PermAllowAlways, withSandboxEscalationNote("Always allow this command", escalates)
 	case agent.PermissionDecisionDeny:
 		return PermRejectOnce, "Reject"
 	case agent.PermissionDecisionCancel:
@@ -64,19 +97,31 @@ func optionKindFor(action agent.PermissionDecisionAction) (kind, name string) {
 
 // decisionFromOutcome maps the client's permission outcome back to a ZERO
 // decision. A cancelled outcome cancels the run; a selected option id is the ZERO
-// action verbatim (validated against what was offered); anything unrecognized
-// fails closed to deny.
-func decisionFromOutcome(outcome RequestPermissionOutcome, offered []agent.PermissionDecisionAction) agent.PermissionDecision {
+// action verbatim (validated against what was actually sent); anything
+// unrecognized fails closed to deny.
+//
+// IT TAKES THE OPTIONS THAT WERE SENT, not the decisions the request offered,
+// and the two are not the same list. Some actions never become options —
+// PermissionDecisionCancel is dropped by optionKindFor because ACP expresses
+// cancellation through the outcome instead — yet it IS enumerated in
+// AvailableDecisions for shell commands and apply_patch, the two commonest
+// prompts. Validating against the decisions therefore accepted
+// {"outcome":"selected","optionId":"cancel"}: an identifier no client was ever
+// shown, aborting the whole turn.
+//
+// Passing the sent options makes "what we offered" and "what we accept" the
+// same slice rather than two things kept in step, so anything optionKindFor
+// declines to render is excluded structurally instead of by name.
+func decisionFromOutcome(outcome RequestPermissionOutcome, offered []PermissionOption) agent.PermissionDecision {
 	switch outcome.Outcome {
 	case OutcomeCancelled:
 		return agent.PermissionDecision{Action: agent.PermissionDecisionCancel, Reason: "client cancelled"}
 	case OutcomeSelected:
-		action := agent.PermissionDecisionAction(outcome.OptionID)
 		// Bind to what was actually offered for THIS call: a client must not be able
 		// to return a broader grant (always_allow / allow_for_session) that wasn't
 		// presented. Anything not offered fails closed to deny.
-		if actionOffered(action, offered) {
-			return agent.PermissionDecision{Action: action}
+		if actionOffered(outcome.OptionID, offered) {
+			return agent.PermissionDecision{Action: agent.PermissionDecisionAction(outcome.OptionID)}
 		}
 		return agent.PermissionDecision{Action: agent.PermissionDecisionDeny, Reason: "permission option was not offered"}
 	default:
@@ -84,9 +129,9 @@ func decisionFromOutcome(outcome RequestPermissionOutcome, offered []agent.Permi
 	}
 }
 
-func actionOffered(action agent.PermissionDecisionAction, offered []agent.PermissionDecisionAction) bool {
-	for _, a := range offered {
-		if a == action {
+func actionOffered(optionID string, offered []PermissionOption) bool {
+	for _, option := range offered {
+		if option.OptionID == optionID {
 			return true
 		}
 	}
@@ -123,3 +168,13 @@ func rawInputBytes(data []byte) json.RawMessage {
 	}
 	return json.RawMessage(data)
 }
+
+// withSandboxEscalationNote appends the escalation disclosure to a prefix
+// option's label. Kept parallel to the TUI's wording so the same consequence is
+// not described two different ways depending on which client the user runs.
+func withSandboxEscalationNote(name string, escalates bool) string {
+	if !escalates {
+		return name
+	}
+	return name + " (runs outside the sandbox)"
+}

internal/acp/permission_escalation_test.go

@@ -0,0 +1,69 @@
+package acp
+
+import (
+	"strings"
+	"testing"
+
+	"github.com/Gitlawb/zero/internal/agent"
+)
+
+func escalationRequest(escalates bool) agent.PermissionRequest {
+	return agent.PermissionRequest{
+		ToolName: "bash",
+		AvailableDecisions: []agent.PermissionDecisionAction{
+			agent.PermissionDecisionAllow,
+			agent.PermissionDecisionAllowPrefix,
+			agent.PermissionDecisionAlwaysAllowPrefix,
+			agent.PermissionDecisionDeny,
+		},
+		PrefixApprovalEscalates: escalates,
+	}
+}
+
+// An ACP client renders only Name, so if the escalation is not in the label the
+// editor user approves it blind. They face exactly the outcome a terminal user
+// faces, so the disclosure has to reach both clients or it is not a fix.
+func TestACPPrefixOptionsDiscloseLeavingTheSandbox(t *testing.T) {
+	found := 0
+	for _, option := range buildPermissionOptions(escalationRequest(true)) {
+		action := agent.PermissionDecisionAction(option.OptionID)
+		if action != agent.PermissionDecisionAllowPrefix && action != agent.PermissionDecisionAlwaysAllowPrefix {
+			if strings.Contains(option.Name, "sandbox") {
+				t.Errorf("option %q carries the disclosure but does not escalate", option.Name)
+			}
+			continue
+		}
+		found++
+		if !strings.Contains(option.Name, "outside the sandbox") {
+			t.Errorf("prefix option %q does not say the command leaves the sandbox", option.Name)
+		}
+	}
+	if found != 2 {
+		t.Fatalf("expected both prefix options, saw %d", found)
+	}
+}
+
+func TestACPPrefixOptionsStaySilentWithoutEscalation(t *testing.T) {
+	for _, option := range buildPermissionOptions(escalationRequest(false)) {
+		if strings.Contains(option.Name, "sandbox") {
+			t.Errorf("option %q warns about the sandbox when approving would not leave it", option.Name)
+		}
+	}
+}
+
+// The disclosure must live in the label only. The option id is the decision
+// action round-tripped verbatim, and decisionFromOutcome matches it against
+// what was offered, so decorating it would fail every prefix approval closed.
+func TestACPDisclosureDoesNotDisturbTheOptionIDRoundTrip(t *testing.T) {
+	request := escalationRequest(true)
+	options := buildPermissionOptions(request)
+	for _, option := range options {
+		decision := decisionFromOutcome(
+			RequestPermissionOutcome{Outcome: OutcomeSelected, OptionID: option.OptionID},
+			options,
+		)
+		if string(decision.Action) != option.OptionID {
+			t.Errorf("option id %q round-tripped to %q (%s)", option.OptionID, decision.Action, decision.Reason)
+		}
+	}
+}

internal/acp/permission_test.go

@@ -39,11 +39,14 @@ func TestBuildPermissionOptionsDefault(t *testing.T) {
 }
 
 func TestDecisionFromOutcome(t *testing.T) {
-	offered := []agent.PermissionDecisionAction{
-		agent.PermissionDecisionAllow,
-		agent.PermissionDecisionAlwaysAllow,
-		agent.PermissionDecisionDeny,
-	}
+	offered := buildPermissionOptions(agent.PermissionRequest{
+		ToolName: "bash",
+		AvailableDecisions: []agent.PermissionDecisionAction{
+			agent.PermissionDecisionAllow,
+			agent.PermissionDecisionAlwaysAllow,
+			agent.PermissionDecisionDeny,
+		},
+	})
 	if d := decisionFromOutcome(RequestPermissionOutcome{Outcome: OutcomeCancelled}, offered); d.Action != agent.PermissionDecisionCancel {
 		t.Errorf("cancelled -> %q, want cancel", d.Action)
 	}

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/acp/types.go

@@ -103,10 +103,6 @@ type ContentBlock struct {
 
 func TextBlock(text string) ContentBlock { return ContentBlock{Type: "text", Text: text} }
 
-func ImageBlock(base64Data, mimeType string) ContentBlock {
-	return ContentBlock{Type: "image", Data: base64Data, MimeType: mimeType}
-}
-
 // ---- sessions ----
 
 // McpServer mirrors the editor-provided MCP server entry. ZERO owns its own MCP
@@ -233,10 +229,6 @@ func ToolContent(block ContentBlock) ToolCallContent {
 	return ToolCallContent{Type: "content", Content: &block}
 }
 
-func ToolDiff(path, oldText, newText string) ToolCallContent {
-	return ToolCallContent{Type: "diff", Path: path, OldText: oldText, NewText: newText}
-}
-
 type ToolCallLocation struct {
 	Path string `json:"path"`
 	Line *int   `json:"line,omitempty"`
@@ -337,19 +329,22 @@ type SetSessionModeParams struct {
 
 type SetSessionModeResult struct{}
 
-// ---- session config options (model selection) ----
+// ---- session config options ----
 
 type SessionConfigOptionValue struct {
-	ID   string `json:"id"`
-	Name string `json:"name"`
+	Value       string `json:"value"`
+	Name        string `json:"name"`
+	Description string `json:"description,omitempty"`
 }
 
 type SessionConfigOption struct {
-	ID          string                     `json:"id"`
-	Name        string                     `json:"name"`
-	Description string                     `json:"description,omitempty"`
-	Value       string                     `json:"value"`
-	Values      []SessionConfigOptionValue `json:"values,omitempty"`
+	ID           string                     `json:"id"`
+	Name         string                     `json:"name"`
+	Description  string                     `json:"description,omitempty"`
+	Category     string                     `json:"category,omitempty"`
+	Type         string                     `json:"type"`
+	CurrentValue string                     `json:"currentValue"`
+	Options      []SessionConfigOptionValue `json:"options"`
 }
 
 type SetSessionConfigOptionParams struct {
@@ -373,6 +368,11 @@ type ZeroSetModelResult struct {
 	Model string `json:"model"`
 }
 
-// configIDModel is the SessionConfigOption id ZERO uses to expose model choice
-// through the standard session/set_config_option method.
-const configIDModel = "model"
+const (
+	configIDModel = "model"
+	configIDMode  = "mode"
+
+	configOptionTypeSelect = "select"
+	configCategoryModel    = "model"
+	configCategoryMode     = "mode"
+)

internal/agent/acceptance_gate_test.go

@@ -72,6 +72,26 @@ func TestSelfReportedIncompletionMatching(t *testing.T) {
 		{"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) != ""

internal/agent/additional_permissions_validation_test.go

@@ -0,0 +1,398 @@
+package agent
+
+import (
+	"context"
+	"strings"
+	"testing"
+
+	"github.com/Gitlawb/zero/internal/sandbox"
+	"github.com/Gitlawb/zero/internal/tools"
+)
+
+type normalizationRecordingShellTool struct {
+	calls []map[string]any
+}
+
+func (tool *normalizationRecordingShellTool) Name() string { return "bash" }
+func (tool *normalizationRecordingShellTool) Description() string {
+	return "record normalized shell arguments"
+}
+func (tool *normalizationRecordingShellTool) Parameters() tools.Schema {
+	return tools.Schema{Type: "object", AdditionalProperties: true}
+}
+func (tool *normalizationRecordingShellTool) Safety() tools.Safety {
+	return tools.Safety{SideEffect: tools.SideEffectShell, Permission: tools.PermissionAllow}
+}
+func (tool *normalizationRecordingShellTool) Run(_ context.Context, args map[string]any) tools.Result {
+	tool.calls = append(tool.calls, cloneArgs(args))
+	return tools.Result{Status: tools.StatusOK}
+}
+
+// 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.NewScopedBashTool(root, nil))
+
+	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.NewScopedBashTool(root, nil))
+
+	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.NewScopedBashTool(root, nil))
+
+	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)
+	}
+	assertNamesOmitRemedy(t, "missing additional_permissions", 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.NewScopedBashTool(root, nil))
+
+	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)
+	}
+}
+
+// runShellCall drives one shell tool call through the real registry (and so
+// through Registry.RunWithOptions) exactly as the agent loop would.
+func runShellCall(t *testing.T, root string, arguments string) ToolResult {
+	t.Helper()
+	registry := tools.NewRegistry()
+	registry.Register(tools.NewScopedExecCommandTool(root, nil, nil))
+	result, err := executeToolCall(
+		context.Background(),
+		registry,
+		ToolCall{ID: "c1", Name: "exec_command", Arguments: arguments},
+		PermissionModeUnsafe,
+		Options{Cwd: root},
+	)
+	if err != nil {
+		t.Fatalf("executeToolCall: %v", err)
+	}
+	return result
+}
+
+// TestAdditionalPermissionsRejectionsConverge is the regression guard for
+// #845/#847. A model that attaches a permission object to a command needing
+// none used to be told only "set sandbox_permissions", which pushed it into the
+// OTHER invalid shape (flag set, payload empty) and its "include a permission"
+// error pushed it straight back — the reported loop that burned turns on
+// `git branch --show-current`.
+//
+// The contract this pins is convergence, not tolerance: both shapes are still
+// REJECTED (nothing here grants access), but each rejection must name the one
+// follow-up call that works, and that follow-up must actually succeed.
+func TestAdditionalPermissionsRejectionsConverge(t *testing.T) {
+	root := t.TempDir()
+	const readOnly = `"cmd":"git --version"`
+
+	// Step 1 of the reported loop: a permission object with no flag.
+	first := runShellCall(t, root, `{`+readOnly+`,"additional_permissions":{"network":{"enabled":true}}}`)
+	if first.Status != tools.StatusError {
+		t.Fatalf("unflagged additional_permissions status = %q, want error (rejecting is correct)", first.Status)
+	}
+	assertNamesOmitRemedy(t, "unflagged additional_permissions", first.Output)
+
+	// Step 2 of the reported loop: the model keeps the flag but empties the
+	// payload. This must not be the dead end it was.
+	second := runShellCall(t, root, `{`+readOnly+`,"sandbox_permissions":"with_additional_permissions"}`)
+	if second.Status != tools.StatusError {
+		t.Fatalf("flag without payload status = %q, want error", second.Status)
+	}
+	assertNamesOmitRemedy(t, "flag without payload", second.Output)
+
+	// The call both errors point at must actually work.
+	converged := runShellCall(t, root, `{`+readOnly+`}`)
+	if converged.Status != tools.StatusOK {
+		t.Fatalf("the call both errors instruct the model to make failed: status=%q output=%q", converged.Status, converged.Output)
+	}
+}
+
+// assertNamesOmitRemedy pins the actionable half of the error text: the message
+// must spell out that BOTH fields can simply be dropped. "Set the flag" or
+// "include a permission" alone is what made the two errors chase each other.
+func assertNamesOmitRemedy(t *testing.T, label string, output string) {
+	t.Helper()
+	if !strings.Contains(output, "omit both sandbox_permissions and additional_permissions") {
+		t.Fatalf("%s error = %q, want it to say to omit both fields entirely", label, output)
+	}
+	if !strings.Contains(output, "retry") {
+		t.Fatalf("%s error = %q, want it to tell the model to retry", label, output)
+	}
+}
+
+// The empty-payload rejection carries the same remedy. It is pinned on the
+// validator directly because normalizeNoopAdditionalPermissions absorbs the
+// empty payload before executeToolCall reaches this branch;
+// inlineAdditionalPermissionsProfile is still the shared source of truth that
+// buildPermissionEvent and the grant path re-run, so its text must converge too.
+func TestEmptyAdditionalPermissionsErrorNamesOmitRemedy(t *testing.T) {
+	args := map[string]any{
+		"cmd":                 "git --version",
+		"sandbox_permissions": string(tools.SandboxPermissionsWithAdditionalPermissions),
+		"additional_permissions": map[string]any{
+			"network":     map[string]any{"enabled": false},
+			"file_system": map[string]any{"read": []any{}, "write": []any{}},
+		},
+	}
+	_, requested, err := inlineAdditionalPermissionsProfile(args, t.TempDir())
+	if !requested {
+		t.Fatal("requested = false, want true for an explicit with_additional_permissions call")
+	}
+	if err == nil {
+		t.Fatal("an empty additional_permissions payload was accepted; it must still be rejected")
+	}
+	assertNamesOmitRemedy(t, "empty payload", err.Error())
+}
+
+// A provider that materializes every optional property may send
+// "additional_permissions": null. That grants nothing and carries no flag, so it
+// is the field omitted — rejecting it stalled an ordinary command behind an
+// error the model could not act on.
+func TestExecuteToolCallTreatsNullAdditionalPermissionsAsOmitted(t *testing.T) {
+	result := runShellCall(t, t.TempDir(), `{"cmd":"git --version","additional_permissions":null}`)
+	if result.Status != tools.StatusOK {
+		t.Fatalf("null additional_permissions was rejected: status=%q output=%q", result.Status, result.Output)
+	}
+}
+
+// A null payload alongside the flag must still be rejected, and must land on
+// the same error as an absent payload rather than being quietly downgraded.
+func TestExecuteToolCallRejectsFlaggedNullAdditionalPermissions(t *testing.T) {
+	result := runShellCall(t, t.TempDir(),
+		`{"cmd":"git --version","sandbox_permissions":"with_additional_permissions","additional_permissions":null}`)
+	if result.Status != tools.StatusError {
+		t.Fatalf("flagged null status = %q, want error", result.Status)
+	}
+	if !strings.Contains(result.Output, "no additional_permissions object was provided") {
+		t.Fatalf("output = %q, want the missing-payload error", result.Output)
+	}
+	assertNamesOmitRemedy(t, "flagged null", result.Output)
+}
+
+func TestNormalizeProviderExpandedEmptyAdditionalPermissions(t *testing.T) {
+	for _, mode := range []string{
+		string(tools.SandboxPermissionsUseDefault),
+		string(tools.SandboxPermissionsWithAdditionalPermissions),
+	} {
+		args := map[string]any{
+			"command":             "./zero --version",
+			"sandbox_permissions": mode,
+			"additional_permissions": map[string]any{
+				"file_system": map[string]any{
+					"deny_read": []any{},
+					"entries":   []any{},
+					"read":      []any{},
+					"write":     []any{},
+				},
+				"network": map[string]any{"enabled": false},
+			},
+		}
+		normalizeNoopAdditionalPermissions(args, t.TempDir(), nil)
+		if _, exists := args["additional_permissions"]; exists {
+			t.Fatalf("mode %q retained empty additional_permissions: %#v", mode, args)
+		}
+		if got := args["sandbox_permissions"]; got != string(tools.SandboxPermissionsUseDefault) {
+			t.Fatalf("mode %q normalized to %q, want use_default", mode, got)
+		}
+	}
+}
+
+func TestExecuteToolCallNormalizesProviderExpandedEmptyAdditionalPermissions(t *testing.T) {
+	tool := &normalizationRecordingShellTool{}
+	registry := tools.NewRegistry()
+	registry.Register(tool)
+
+	result, err := executeToolCall(
+		context.Background(),
+		registry,
+		ToolCall{
+			ID:   "c1",
+			Name: "bash",
+			Arguments: `{"command":"./zero --version","sandbox_permissions":"with_additional_permissions",` +
+				`"additional_permissions":{"file_system":{"deny_read":[],"entries":[],"read":[],"write":[]},"network":{"enabled":false}}}`,
+		},
+		PermissionModeAuto,
+		Options{Cwd: t.TempDir()},
+	)
+	if err != nil {
+		t.Fatalf("executeToolCall: %v", err)
+	}
+	if result.Status != tools.StatusOK {
+		t.Fatalf("provider-expanded empty permissions were rejected: %s", result.Output)
+	}
+	if len(tool.calls) != 1 {
+		t.Fatalf("shell tool calls = %d, want 1", len(tool.calls))
+	}
+	args := tool.calls[0]
+	if _, exists := args["additional_permissions"]; exists {
+		t.Fatalf("executeToolCall retained empty additional_permissions: %#v", args)
+	}
+	if got := args["sandbox_permissions"]; got != string(tools.SandboxPermissionsUseDefault) {
+		t.Fatalf("sandbox mode = %q, want use_default", got)
+	}
+}
+
+func TestNormalizeAdditionalPermissionsPreservesRealRequest(t *testing.T) {
+	args := map[string]any{
+		"sandbox_permissions": string(tools.SandboxPermissionsWithAdditionalPermissions),
+		"additional_permissions": map[string]any{
+			"network": map[string]any{"enabled": true},
+		},
+	}
+	normalizeNoopAdditionalPermissions(args, t.TempDir(), nil)
+	if _, exists := args["additional_permissions"]; !exists {
+		t.Fatal("non-empty additional_permissions was removed")
+	}
+	if got := args["sandbox_permissions"]; got != string(tools.SandboxPermissionsWithAdditionalPermissions) {
+		t.Fatalf("sandbox mode changed to %q", got)
+	}
+}
+
+func TestNormalizeAdditionalPermissionsRemovesWorkspaceReadAlreadyCoveredBySandbox(t *testing.T) {
+	root := t.TempDir()
+	engine := sandbox.NewEngine(sandbox.EngineOptions{
+		WorkspaceRoot: root,
+		Policy:        sandbox.DefaultPolicy(),
+	})
+	args := map[string]any{
+		"command":             "./zero --version",
+		"sandbox_permissions": string(tools.SandboxPermissionsWithAdditionalPermissions),
+		"additional_permissions": map[string]any{
+			"file_system": map[string]any{
+				"entries": []any{
+					map[string]any{
+						"access": "read",
+						"path": map[string]any{
+							"type": "path",
+							"path": ".",
+						},
+					},
+				},
+				"deny_read": []any{},
+				"read":      []any{},
+				"write":     []any{},
+			},
+			"network": map[string]any{"enabled": false},
+		},
+	}
+
+	normalizeNoopAdditionalPermissions(args, root, engine)
+
+	if _, exists := args["additional_permissions"]; exists {
+		t.Fatalf("already-covered additional_permissions was retained: %#v", args)
+	}
+	if got := args["sandbox_permissions"]; got != string(tools.SandboxPermissionsUseDefault) {
+		t.Fatalf("sandbox mode changed to %q", got)
+	}
+}

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

@@ -2,6 +2,7 @@ package agent
 
 import (
 	"fmt"
+	"runtime"
 	"strings"
 
 	"github.com/Gitlawb/zero/internal/sandbox"
@@ -27,10 +28,20 @@ func proposedCommandPrefix(toolName string, args map[string]any) []string {
 		}
 		return nil
 	}
-	for _, tokens := range segments {
-		if !knownSafeCommandSegment(tokens) && sandbox.ValidCommandPrefix(tokens) {
-			return append([]string(nil), 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
@@ -38,6 +49,20 @@ func proposedCommandPrefix(toolName string, args map[string]any) []string {
 	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
@@ -168,10 +193,6 @@ func cleanPrefixRule(prefix []string) []string {
 	return cleaned
 }
 
-func safeRequestedPrefix(prefix []string, command []string) bool {
-	return safeRequestedPrefixForSegments(prefix, [][]string{command})
-}
-
 func safeRequestedPrefixForSegments(prefix []string, segments [][]string) bool {
 	if len(prefix) == 0 || !sandbox.ValidCommandPrefix(prefix) {
 		return false
@@ -196,14 +217,6 @@ func safeRequestedPrefixForSegments(prefix []string, segments [][]string) bool {
 	return matched
 }
 
-func safeShellCommandTokens(command string) ([]string, bool) {
-	segments, ok := safeShellCommandSegments(command)
-	if !ok || len(segments) != 1 {
-		return nil, false
-	}
-	return segments[0], true
-}
-
 func safeShellCommandSegments(command string) ([][]string, bool) {
 	command = strings.TrimSpace(command)
 	if command == "" {
@@ -275,7 +288,11 @@ func knownSafeCommandSegment(command []string) bool {
 	if len(command) == 0 {
 		return false
 	}
-	switch commandName(command[0]) {
+	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":
@@ -484,15 +501,3 @@ func hasStringPrefix(values []string, prefix []string) bool {
 	}
 	return true
 }
-
-func equalStringSlices(left []string, right []string) bool {
-	if len(left) != len(right) {
-		return false
-	}
-	for index := range left {
-		if left[index] != right[index] {
-			return false
-		}
-	}
-	return true
-}

internal/agent/command_prefix_test.go

@@ -1,9 +1,11 @@
 package agent
 
 import (
+	"runtime"
 	"testing"
 
 	"github.com/Gitlawb/zero/internal/sandbox"
+	"github.com/Gitlawb/zero/internal/tools"
 )
 
 func TestProposedCommandPrefixUsesSafeSimpleCommands(t *testing.T) {
@@ -35,12 +37,37 @@ 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",
@@ -72,6 +99,43 @@ func TestProposedCommandPrefixRejectsUnsafeRequestedPrefix(t *testing.T) {
 	}
 }
 
+// The launcher denylist is only worth anything if it holds on the path a user
+// actually sees. A model naming its own prefix_rule is that path: before
+// launcher names were normalized, "python3.11" was offered as a one-token
+// grant, and hasStringPrefix then matched every later python3.11 invocation.
+func TestProposedCommandPrefixRejectsLauncherSpellingsInRequestedPrefix(t *testing.T) {
+	for _, testCase := range []struct {
+		command string
+		rule    []any
+	}{
+		{"python3.11 script.py", []any{"python3.11"}},
+		{"python.exe script.py", []any{"python.exe"}},
+		{"node.exe app.js", []any{"node.exe"}},
+		{"cmd /c whoami", []any{"cmd"}},
+		{"busybox sh -c id", []any{"busybox"}},
+		{"uv run main.py", []any{"uv", "run"}},
+		{"python3.11 -c import_os", []any{"python3.11", "-c"}},
+	} {
+		got := proposedCommandPrefix("bash", map[string]any{
+			"command":     testCase.command,
+			"prefix_rule": testCase.rule,
+		})
+		if got != nil {
+			t.Errorf("prefix_rule %v for %q was offered as %#v, want rejected", testCase.rule, testCase.command, got)
+		}
+	}
+}
+
+func TestProposedCommandPrefixStillOffersOrdinaryCommands(t *testing.T) {
+	got := proposedCommandPrefix("bash", map[string]any{
+		"command":     "cargo build --release",
+		"prefix_rule": []any{"cargo", "build"},
+	})
+	if len(got) != 2 || got[0] != "cargo" || got[1] != "build" {
+		t.Fatalf("ordinary prefix rule = %#v, want [cargo build]", got)
+	}
+}
+
 func TestProposedCommandPrefixRejectsUnsafeShellForms(t *testing.T) {
 	cases := []string{
 		"cat < in > out",
@@ -120,13 +184,35 @@ 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": "ps aux | head -5"}, Options{Sandbox: engine})
+	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"})

internal/agent/compaction.go

@@ -4,8 +4,11 @@ import (
 	"context"
 	"encoding/json"
 	"errors"
+	"strconv"
 	"strings"
+	"unicode/utf8"
 
+	"github.com/Gitlawb/zero/internal/trace"
 	"github.com/Gitlawb/zero/internal/zeroruntime"
 )
 
@@ -21,19 +24,25 @@ 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).
 const summaryLabel = "[Summary of earlier conversation]"
 
-// summaryInstructions is the system prompt handed to the summarizer model.
-const summaryInstructions = "You are compacting a coding-assistant conversation to save context. " +
+// CompactionSummaryInstructions is the system prompt handed to both automatic
+// and manually requested compaction summarizers.
+const CompactionSummaryInstructions = "You are compacting a coding-assistant conversation to save context. " +
 	"Write a dense, factual summary of the conversation so far. Preserve: the user's goals and explicit constraints; " +
 	"decisions made and why; files created or modified (with paths) and key code changes; commands run and their important " +
 	"results; and anything still in progress or unresolved. Omit pleasantries. Use terse bullet points. Do not invent details. " +
@@ -50,6 +59,10 @@ type CompactionOptions struct {
 	// injected so Compact stays pure and testable; the agent loop wires it to a
 	// real provider call.
 	Summarize func(toSummarize []zeroruntime.Message) (string, error)
+	// taskState is a snapshot supplied by the running agent. Its immutable
+	// objective is always preserved; mutable fields are admitted only when its
+	// plan projection still matches the transcript.
+	taskState *taskStateSnapshot
 }
 
 // CompactionResult is the metadata-bearing result returned by CompactMessages.
@@ -65,6 +78,12 @@ type CompactionResult struct {
 	// summary message also includes summaryLabel and any preserved structured
 	// state needed by later compactions.
 	SummaryText string
+	// ProjectedChars is the text content sent to the summarizer after semantic
+	// projection, excluding provider protocol framing.
+	ProjectedChars int
+	// Truncated reports whether projection dropped content to satisfy its
+	// bounded brief or previous-summary limits.
+	Truncated bool
 	// Compacted reports whether Messages contains an injected summary.
 	Compacted bool
 }
@@ -206,15 +225,15 @@ func CompactMessages(messages []zeroruntime.Message, opts CompactionOptions) (Co
 		}, nil
 	}
 
-	summary, err := opts.Summarize(middle)
+	summaryResult, err := SummarizeCompactionMessages(middle, opts.Summarize)
 	if err != nil {
 		return CompactionResult{}, err
 	}
-	summary = strings.TrimSpace(summary)
+	summary := summaryResult.SummaryText
 
 	// Preserve structured state (active plan + loaded skills) from the elided
 	// middle verbatim, so it is not lost or paraphrased away by the prose summary.
-	content := appendPreservedState(summaryLabel+"\n"+summary, middle)
+	content := appendPreservedState(summaryLabel+"\n"+summary, middle, opts.taskState)
 
 	compacted := make([]zeroruntime.Message, 0, systemEnd+1+(len(messages)-boundary))
 	compacted = append(compacted, messages[:systemEnd]...)
@@ -228,10 +247,53 @@ func CompactMessages(messages []zeroruntime.Message, opts CompactionOptions) (Co
 		RemovedCount:   len(middle),
 		PreservedCount: len(messages) - len(middle),
 		SummaryText:    summary,
+		ProjectedChars: summaryResult.ProjectedChars,
+		Truncated:      summaryResult.Truncated,
 		Compacted:      true,
 	}, nil
 }
 
+// CompactionSummaryResult describes the shared summarization input and output.
+type CompactionSummaryResult struct {
+	SummaryText    string
+	ProjectedChars int
+	Truncated      bool
+}
+
+// SummarizeCompactionMessages applies the shared semantic projection before
+// invoking a caller-supplied summarizer. Automatic context-pressure compaction
+// and manual session compaction both use this path.
+func SummarizeCompactionMessages(messages []zeroruntime.Message, summarize func([]zeroruntime.Message) (string, error)) (CompactionSummaryResult, error) {
+	if summarize == nil {
+		return CompactionSummaryResult{}, errors.New("compaction requires a Summarize function")
+	}
+	projection := projectCompactionInput(messages)
+	summaryInput := projection.messages
+	if len(summaryInput) == 0 {
+		summaryInput = messages
+	}
+	summary, err := summarize(summaryInput)
+	if err != nil {
+		return CompactionSummaryResult{}, err
+	}
+	return CompactionSummaryResult{
+		SummaryText:    strings.TrimSpace(summary),
+		ProjectedChars: compactionMessageChars(summaryInput),
+		Truncated:      projection.truncated,
+	}, nil
+}
+
+func compactionMessageChars(messages []zeroruntime.Message) int {
+	total := 0
+	for _, message := range messages {
+		total += len(message.Content)
+		for _, call := range message.ToolCalls {
+			total += len(call.Name) + len(call.Arguments)
+		}
+	}
+	return total
+}
+
 // safeSuffixBoundary walks the preserve boundary backward (toward systemEnd) so
 // the preserved suffix begins on a user or assistant message rather than a
 // tool/tool_result message. A tool result with no preceding assistant tool call
@@ -280,6 +342,25 @@ 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.
@@ -301,15 +382,151 @@ 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)
+	task    *taskState
+	planner *contextPlanner
+	trace   *trace.Recorder
+
+	// 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
+
+	// summarizerFactory lazily builds the dedicated (cheap) summarizer provider
+	// (Options.Summarizer) for the current main model. summarizerProvider
+	// memoizes the built provider and summarizerResolved records that the
+	// factory ran (it may legitimately return no provider: the main model is
+	// already the cheap one). summarizerBroken is the sticky fallback: once a
+	// build or call fails, the run stays on the main provider until the next
+	// model switch — a misconfigured cheap model must never break (or
+	// repeatedly delay) compaction.
+	summarizerFactory  func(ctx context.Context, mainModelID string) (Provider, error)
+	summarizerProvider Provider
+	summarizerResolved bool
+	summarizerBroken   bool
+	// window is the context window the threshold was derived from. The
+	// reactive free-prune stage compares against it (the limit the provider
+	// enforces), not against the 70% proactive threshold — a pruned history
+	// between the two would still succeed on retry.
+	window int
+	// mainModel is the model the run currently streams on; the summarizer
+	// choice is made against it and remade after a mid-run switch.
+	mainModel string
 }
 
-func newCompactionState(options Options) *compactionState {
-	return &compactionState{
-		enabled:      options.ContextWindow > 0,
-		threshold:    compactionThreshold(options.ContextWindow),
-		preserveLast: options.CompactionPreserveLast,
-		onUsage:      options.OnUsage,
+// 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, task *taskState) *compactionState {
+	state := &compactionState{
+		enabled:           options.ContextWindow > 0,
+		threshold:         compactionThreshold(options.ContextWindow),
+		window:            options.ContextWindow,
+		mainModel:         options.Model,
+		preserveLast:      options.CompactionPreserveLast,
+		onUsage:           options.OnUsage,
+		task:              task,
+		planner:           newContextPlanner(contextPlannerConfig{contextWindow: options.ContextWindow, serviceTier: options.ServiceTier}),
+		trace:             options.Trace,
+		summarizerFactory: options.Summarizer,
+	}
+	return state
+}
+
+// updateWindow re-derives the compaction threshold after a mid-run model
+// switch. The low-water mark is reset: it was measured against the old
+// threshold, and keeping it could suppress a compaction the smaller window
+// now requires. <= 0 (unknown model) leaves everything unchanged.
+func (state *compactionState) updateWindow(window int) {
+	if window <= 0 {
+		return
+	}
+	state.enabled = true
+	state.window = window
+	state.threshold = compactionThreshold(window)
+	state.lowWaterMark = 0
+	if state.planner != nil {
+		state.planner.config.contextWindow = window
+	}
+}
+
+// switchModel records a mid-run model switch: the window is re-derived
+// (window <= 0 keeps the old one) and the summarizer choice is forgotten so
+// the next compaction re-resolves it against the new main model. A run that
+// started on the cheap model (no dedicated summarizer) and escalated to an
+// expensive one therefore regains cheap summaries — the exact workflow a
+// dedicated summarizer exists for.
+func (state *compactionState) switchModel(modelID string, window int) {
+	state.mainModel = modelID
+	state.updateWindow(window)
+	state.summarizerProvider = nil
+	state.summarizerResolved = false
+	state.summarizerBroken = false
+}
+
+// summarizeProvider returns the provider summarization calls should use and
+// whether it is the dedicated summarizer: that one when it applies to the
+// current main model and is healthy, else the main provider. The factory runs
+// at most once per main model; a build error marks the summarizer broken for
+// that model. Callers branch on the boolean rather than comparing Provider
+// values — interface comparison panics on a non-comparable dynamic type.
+func (state *compactionState) summarizeProvider(ctx context.Context, main Provider) (Provider, bool) {
+	if state.summarizerFactory == nil || state.summarizerBroken {
+		return main, false
+	}
+	if !state.summarizerResolved {
+		state.summarizerResolved = true
+		built, err := state.summarizerFactory(ctx, state.mainModel)
+		if err != nil {
+			state.summarizerBroken = true
+			return main, false
+		}
+		state.summarizerProvider = built
+	}
+	if state.summarizerProvider == nil {
+		return main, false
+	}
+	return state.summarizerProvider, true
+}
+
+// fitsWindow reports whether a calibrated size is safely under the limit the
+// provider enforces. The proactive threshold (70% of the window) is a trigger
+// for the paid summarizer, not the hard limit; the reactive path compares
+// against the window itself, with a small margin for estimate error, so a
+// history pruned to, say, 85% of the window takes the free retry.
+func (state *compactionState) fitsWindow(size int) bool {
+	if state.window <= 0 {
+		return size <= state.threshold
+	}
+	return size <= state.window-state.window/20
 }
 
 // maybeCompact runs proactive compaction at the top of a turn. It returns the
@@ -329,7 +546,7 @@ 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
 	}
@@ -344,9 +561,17 @@ func (state *compactionState) maybeCompact(
 	// the bodies of old, large tool results (the model has already acted on
 	// them). If that brings us back under threshold, skip the paid summarizer
 	// entirely and preserve recent turns verbatim.
+	if pruned, reclaimed := pruneSupersededReadResults(messages); reclaimed > 0 {
+		messages = pruned
+		size = state.calibratedTokens(estimateTokens(messages) + toolTokens)
+		if size <= state.threshold {
+			state.lowWaterMark = size
+			return messages
+		}
+	}
 	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
@@ -355,20 +580,27 @@ func (state *compactionState) maybeCompact(
 
 	compacted, err := Compact(messages, CompactionOptions{
 		PreserveLast: state.preserveLast,
-		Summarize:    summarizeClosure(ctx, provider, state.onUsage),
+		Summarize:    state.summarizeClosure(ctx, provider),
+		taskState:    state.task.snapshotForCompaction(messages),
 	})
 	if err != nil {
 		// Summarizer failed: keep the original history. The reactive path (or a
 		// 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
 }
@@ -390,16 +622,41 @@ func (state *compactionState) recover(
 		// context-limit error never triggers an unexpected summarization call.
 		return messages, false, nil
 	}
-	if state.reactiveAttempted {
+	if !isContextLimitError(errorMessage) {
 		return messages, false, nil
 	}
-	if !isContextLimitError(errorMessage) {
+
+	// CHEAP FIRST STAGE (mirrors maybeCompact): a context-limit failure can
+	// often be fixed for free by dropping superseded reads and pruning old
+	// tool-result bodies. It runs regardless of the one-shot reactive budget
+	// (it costs nothing and cannot loop: a second pass finds nothing left to
+	// prune). The free retry is taken only when the pruned history fits the
+	// window the provider enforces — otherwise a retry would fail the same
+	// way, so fall through to the paid summarizer in this same call.
+	toolTokens := estimateToolDefTokens(tools)
+	prunedAny := false
+	if pruned, reclaimed := pruneSupersededReadResults(messages); reclaimed > 0 {
+		messages, prunedAny = pruned, true
+	}
+	if pruned, reclaimed := pruneStaleToolOutput(messages, state.preserveLast); reclaimed > 0 {
+		messages, prunedAny = pruned, true
+	}
+	if prunedAny {
+		size := state.calibratedTokens(estimateTokens(messages) + toolTokens)
+		if state.fitsWindow(size) {
+			state.lowWaterMark = size
+			return messages, true, nil
+		}
+	}
+
+	if state.reactiveAttempted {
 		return messages, false, nil
 	}
 
 	result, compactErr := Compact(messages, CompactionOptions{
 		PreserveLast: state.preserveLast,
-		Summarize:    summarizeClosure(ctx, provider, state.onUsage),
+		Summarize:    state.summarizeClosure(ctx, provider),
+		taskState:    state.task.snapshotForCompaction(messages),
 	})
 	if compactErr != nil {
 		// A genuine compaction attempt was made (and failed): the budget is spent
@@ -418,9 +675,13 @@ 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
 }
 
@@ -428,9 +689,22 @@ func (state *compactionState) recover(
 // provider call. The summary stream intentionally does NOT forward OnText (so
 // compaction stays invisible on the user-facing surface), but it DOES forward
 // OnUsage so the summarizer's token cost is still counted by usage/budgeting.
-func summarizeClosure(ctx context.Context, provider Provider, onUsage func(Usage)) func([]zeroruntime.Message) (string, error) {
+// The call goes to the dedicated (cheap) summarizer when one applies; if that
+// call fails the summarizer is marked broken for the current model and the
+// same slice is retried once on the main provider, so the caller sees at most
+// one (main-provider-quality) error.
+func (state *compactionState) summarizeClosure(ctx context.Context, provider Provider) func([]zeroruntime.Message) (string, error) {
+	if state.planner == nil {
+		state.planner = newContextPlanner(contextPlannerConfig{})
+	}
 	return func(toSummarize []zeroruntime.Message) (string, error) {
-		return summarizeWithFallback(ctx, provider, toSummarize, onUsage)
+		chosen, dedicated := state.summarizeProvider(ctx, provider)
+		summary, err := summarizeWithPlanner(ctx, chosen, toSummarize, state.onUsage, state.planner, state.trace)
+		if err == nil || !dedicated {
+			return summary, err
+		}
+		state.summarizerBroken = true
+		return summarizeWithPlanner(ctx, provider, toSummarize, state.onUsage, state.planner, state.trace)
 	}
 }
 
@@ -442,7 +716,11 @@ func summarizeClosure(ctx context.Context, provider Provider, onUsage func(Usage
 // Non-context-limit errors (and a single message that still won't fit) surface
 // to the caller unchanged.
 func summarizeWithFallback(ctx context.Context, provider Provider, messages []zeroruntime.Message, onUsage func(Usage)) (string, error) {
-	summary, err := summarizeMessagesOnce(ctx, provider, messages, onUsage)
+	return summarizeWithPlanner(ctx, provider, messages, onUsage, newContextPlanner(contextPlannerConfig{}), nil)
+}
+
+func summarizeWithPlanner(ctx context.Context, provider Provider, messages []zeroruntime.Message, onUsage func(Usage), planner *contextPlanner, recorder *trace.Recorder) (string, error) {
+	summary, err := summarizeMessagesOnce(ctx, provider, messages, onUsage, planner, recorder)
 	if err == nil {
 		return summary, nil
 	}
@@ -451,11 +729,11 @@ func summarizeWithFallback(ctx context.Context, provider Provider, messages []ze
 	}
 
 	mid := len(messages) / 2
-	left, leftErr := summarizeWithFallback(ctx, provider, messages[:mid], onUsage)
+	left, leftErr := summarizeWithPlanner(ctx, provider, messages[:mid], onUsage, planner, recorder)
 	if leftErr != nil {
 		return "", leftErr
 	}
-	right, rightErr := summarizeWithFallback(ctx, provider, messages[mid:], onUsage)
+	right, rightErr := summarizeWithPlanner(ctx, provider, messages[mid:], onUsage, planner, recorder)
 	if rightErr != nil {
 		return "", rightErr
 	}
@@ -468,7 +746,7 @@ func summarizeWithFallback(ctx context.Context, provider Provider, messages []ze
 	combined := strings.TrimSpace(left + "\n\n" + right)
 	reduced, reduceErr := summarizeMessagesOnce(ctx, provider, []zeroruntime.Message{
 		{Role: zeroruntime.MessageRoleUser, Content: combined},
-	}, onUsage)
+	}, onUsage, planner, recorder)
 	if reduceErr != nil {
 		if isContextLimitError(reduceErr.Error()) {
 			// Even the two combined partials don't fit (extreme): fall back to the
@@ -484,15 +762,15 @@ func summarizeWithFallback(ctx context.Context, provider Provider, messages []ze
 }
 
 // summarizeMessagesOnce performs a single tool-less summarization call.
-func summarizeMessagesOnce(ctx context.Context, provider Provider, messages []zeroruntime.Message, onUsage func(Usage)) (string, error) {
-	request := zeroruntime.CompletionRequest{
-		Messages: []zeroruntime.Message{
-			{Role: zeroruntime.MessageRoleSystem, Content: summaryInstructions},
-			{Role: zeroruntime.MessageRoleUser, Content: "Summarize this conversation:\n\n" + renderTranscript(messages)},
-		},
-		// No tools: this is a plain text summarization call.
+func summarizeMessagesOnce(ctx context.Context, provider Provider, messages []zeroruntime.Message, onUsage func(Usage), planner *contextPlanner, recorder *trace.Recorder) (string, error) {
+	requestMessages := []zeroruntime.Message{
+		{Role: zeroruntime.MessageRoleSystem, Content: CompactionSummaryInstructions},
+		{Role: zeroruntime.MessageRoleUser, Content: "Summarize this conversation:\n\n" + renderTranscript(messages)},
 	}
-	stream, err := provider.StreamCompletion(ctx, request)
+	parts := systemPromptParts{prompt: CompactionSummaryInstructions, baseInstructions: CompactionSummaryInstructions}
+	plan := planner.planWithPromptParts(requestMessages, nil, "", parts)
+	recordContextPlanTrace(recorder, plan)
+	stream, err := provider.StreamCompletion(ctx, plan.Request)
 	if err != nil {
 		return "", err
 	}
@@ -509,8 +787,41 @@ func summarizeMessagesOnce(ctx context.Context, provider Provider, messages []ze
 	return summary, nil
 }
 
+// summaryToolResultClamp bounds each tool-result body in the summarizer input.
+// Tool outputs dominate the elided middle's bulk (build logs, file reads,
+// search results), yet the summary instructions only need their gist — the
+// model that ACTED on the full output already turned it into decisions the
+// surrounding messages record. 2,000 chars (~500 tokens) keeps the useful
+// head+tail of a result while cutting the summarizer input by 5-10x on
+// tool-heavy sessions; the summarizer's own cost is the price being cut.
+const summaryToolResultClamp = 2000
+
+// summaryToolArgsClamp bounds each tool-call argument blob in the transcript:
+// a write_file call carries the whole file body in its arguments.
+const summaryToolArgsClamp = 500
+
+// clampForSummary head+tail-truncates text for the summarizer transcript,
+// snapping to rune boundaries so the clamp never splits a UTF-8 sequence.
+func clampForSummary(text string, max int) string {
+	if max <= 0 || len(text) <= max {
+		return text
+	}
+	head := max * 3 / 4
+	tail := max - head
+	for head > 0 && !utf8.RuneStart(text[head]) {
+		head--
+	}
+	start := len(text) - tail
+	for start < len(text) && !utf8.RuneStart(text[start]) {
+		start++
+	}
+	return text[:head] + "\n…[" + strconv.Itoa(len(text)-head-(len(text)-start)) + " chars omitted for summarization]…\n" + text[start:]
+}
+
 // renderTranscript flattens messages into a plain-text transcript for the
 // summarizer. Secret scrubbing already happened upstream at the tool boundary.
+// Tool-result bodies and tool-call arguments are clamped (see the constants
+// above): the summary needs what was done and decided, not the raw payloads.
 func renderTranscript(messages []zeroruntime.Message) string {
 	lines := make([]string, 0, len(messages))
 	for _, message := range messages {
@@ -520,13 +831,13 @@ func renderTranscript(messages []zeroruntime.Message) string {
 			if len(message.ToolCalls) > 0 {
 				calls := make([]string, 0, len(message.ToolCalls))
 				for _, call := range message.ToolCalls {
-					calls = append(calls, call.Name+"("+call.Arguments+")")
+					calls = append(calls, call.Name+"("+clampForSummary(call.Arguments, summaryToolArgsClamp)+")")
 				}
 				line += "\n[tool calls: " + strings.Join(calls, "; ") + "]"
 			}
 			lines = append(lines, line)
 		case zeroruntime.MessageRoleTool:
-			lines = append(lines, "tool result: "+message.Content)
+			lines = append(lines, "tool result: "+clampForSummary(message.Content, summaryToolResultClamp))
 		default:
 			lines = append(lines, string(message.Role)+": "+message.Content)
 		}

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_cheap_test.go

@@ -0,0 +1,469 @@
+package agent
+
+import (
+	"context"
+	"errors"
+	"strings"
+	"testing"
+
+	"github.com/Gitlawb/zero/internal/tools"
+	"github.com/Gitlawb/zero/internal/zeroruntime"
+)
+
+// compactableHistory returns a history whose middle is large enough to trip
+// the proactive threshold for a 1000-token window.
+func compactableHistory() []zeroruntime.Message {
+	return []zeroruntime.Message{
+		{Role: zeroruntime.MessageRoleSystem, Content: "sys"},
+		{Role: zeroruntime.MessageRoleUser, Content: strings.Repeat("u", 4000)},
+		{Role: zeroruntime.MessageRoleAssistant, Content: strings.Repeat("a", 4000)},
+		{Role: zeroruntime.MessageRoleUser, Content: "u2"},
+		{Role: zeroruntime.MessageRoleAssistant, Content: "a2"},
+		{Role: zeroruntime.MessageRoleUser, Content: "u3"},
+	}
+}
+
+// Summarization must run on the dedicated (cheap) summarizer when one is
+// configured — the main provider sees no summarize traffic at all.
+func TestMaybeCompactUsesDedicatedSummarizer(t *testing.T) {
+	cheap := &mockProvider{turns: [][]zeroruntime.StreamEvent{{
+		{Type: zeroruntime.StreamEventText, Content: "CHEAP SUMMARY"},
+		{Type: zeroruntime.StreamEventDone},
+	}}}
+	main := &mockProvider{}
+	factoryCalls := 0
+	state := newCompactionState(Options{
+		ContextWindow:          1000,
+		CompactionPreserveLast: 2,
+		Summarizer: func(context.Context, string) (Provider, error) {
+			factoryCalls++
+			return cheap, nil
+		},
+	}, nil)
+
+	compacted := state.maybeCompact(context.Background(), main, compactableHistory(), nil)
+
+	if len(cheap.requests) == 0 {
+		t.Fatal("dedicated summarizer never received the summarize call")
+	}
+	if len(main.requests) != 0 {
+		t.Fatalf("main provider must see no summarize traffic, got %d requests", len(main.requests))
+	}
+	if factoryCalls != 1 {
+		t.Fatalf("summarizer factory calls = %d, want 1 (lazy, memoized)", factoryCalls)
+	}
+	if estimateTokens(compacted) >= estimateTokens(compactableHistory()) {
+		t.Fatal("compaction did not shrink the history")
+	}
+	found := false
+	for _, message := range compacted {
+		if strings.Contains(message.Content, "CHEAP SUMMARY") {
+			found = true
+		}
+	}
+	if !found {
+		t.Fatal("cheap summarizer's summary missing from the compacted history")
+	}
+}
+
+// A failing summarizer falls back to the main provider for the SAME slice and
+// stays broken for the rest of the run — no repeated failed cheap calls.
+func TestSummarizerFailureFallsBackToMainAndSticks(t *testing.T) {
+	broken := &mockProvider{turns: [][]zeroruntime.StreamEvent{{
+		{Type: zeroruntime.StreamEventError, Error: "boom: model not found"},
+	}}}
+	main := &mockProvider{turns: [][]zeroruntime.StreamEvent{
+		{{Type: zeroruntime.StreamEventText, Content: "MAIN SUMMARY"}, {Type: zeroruntime.StreamEventDone}},
+		{{Type: zeroruntime.StreamEventText, Content: "MAIN SUMMARY 2"}, {Type: zeroruntime.StreamEventDone}},
+	}}
+	state := newCompactionState(Options{
+		ContextWindow:          1000,
+		CompactionPreserveLast: 2,
+		Summarizer:             func(context.Context, string) (Provider, error) { return broken, nil },
+	}, nil)
+	summarize := state.summarizeClosure(context.Background(), main)
+	input := []zeroruntime.Message{{Role: zeroruntime.MessageRoleUser, Content: "history"}}
+
+	summary, err := summarize(input)
+	if err != nil {
+		t.Fatalf("fallback must succeed via the main provider: %v", err)
+	}
+	if summary != "MAIN SUMMARY" {
+		t.Fatalf("summary = %q, want the main provider's", summary)
+	}
+	if len(broken.requests) != 1 {
+		t.Fatalf("broken summarizer requests = %d, want exactly 1", len(broken.requests))
+	}
+	if !state.summarizerBroken {
+		t.Fatal("a failed summarizer must be marked broken for the run")
+	}
+
+	// Second summarization: straight to main, no cheap retry.
+	if _, err := summarize(input); err != nil {
+		t.Fatal(err)
+	}
+	if len(broken.requests) != 1 {
+		t.Fatalf("broken summarizer was retried (%d requests)", len(broken.requests))
+	}
+	if len(main.requests) != 2 {
+		t.Fatalf("main provider requests = %d, want 2", len(main.requests))
+	}
+}
+
+// A factory that errors (misconfigured model) is equivalent to no summarizer:
+// main provider used, factory not retried.
+func TestSummarizerFactoryErrorFallsBack(t *testing.T) {
+	main := &mockProvider{turns: [][]zeroruntime.StreamEvent{{
+		{Type: zeroruntime.StreamEventText, Content: "MAIN SUMMARY"},
+		{Type: zeroruntime.StreamEventDone},
+	}}}
+	factoryCalls := 0
+	state := newCompactionState(Options{
+		ContextWindow:          1000,
+		CompactionPreserveLast: 2,
+		Summarizer: func(context.Context, string) (Provider, error) {
+			factoryCalls++
+			return nil, errors.New("unknown model")
+		},
+	}, nil)
+	summarize := state.summarizeClosure(context.Background(), main)
+
+	summary, err := summarize([]zeroruntime.Message{{Role: zeroruntime.MessageRoleUser, Content: "history"}})
+	if err != nil || summary != "MAIN SUMMARY" {
+		t.Fatalf("summary = %q, err = %v; want main provider fallback", summary, err)
+	}
+	if factoryCalls != 1 {
+		t.Fatalf("factory calls = %d, want 1 (sticky failure)", factoryCalls)
+	}
+	if _, dedicated := state.summarizeProvider(context.Background(), main); dedicated {
+		t.Fatal("after a factory failure the run must stay on the main provider")
+	}
+	if factoryCalls != 1 {
+		t.Fatalf("factory retried after failure (%d calls)", factoryCalls)
+	}
+}
+
+// Tool payloads are clamped in the summarizer transcript: results to
+// summaryToolResultClamp, call arguments to summaryToolArgsClamp.
+func TestRenderTranscriptClampsToolPayloads(t *testing.T) {
+	bigResult := strings.Repeat("r", 50_000)
+	bigArgs := strings.Repeat("w", 20_000)
+	transcript := renderTranscript([]zeroruntime.Message{
+		{Role: zeroruntime.MessageRoleAssistant, Content: "editing", ToolCalls: []zeroruntime.ToolCall{
+			{ID: "c1", Name: "write_file", Arguments: bigArgs},
+		}},
+		{Role: zeroruntime.MessageRoleTool, ToolCallID: "c1", Content: bigResult},
+		{Role: zeroruntime.MessageRoleUser, Content: "keep going"},
+	})
+
+	if len(transcript) > summaryToolResultClamp+summaryToolArgsClamp+1024 {
+		t.Fatalf("transcript is %d bytes; clamps did not apply", len(transcript))
+	}
+	if !strings.Contains(transcript, "chars omitted for summarization") {
+		t.Fatal("clamped transcript must carry the omission marker")
+	}
+	if !strings.Contains(transcript, "keep going") {
+		t.Fatal("non-tool content must survive unclamped")
+	}
+	// Head AND tail of the clamped payloads survive.
+	if !strings.HasPrefix(transcript[strings.Index(transcript, "tool result: ")+len("tool result: "):], "r") ||
+		!strings.Contains(transcript, "rrr\n\nuser: keep going") {
+		t.Fatalf("head/tail of the tool result lost:\n%s", transcript[:200])
+	}
+}
+
+// bigToolHistory returns a history with n large tool results (~40k estimated
+// tokens each): the trailing protection window keeps the newest verbatim,
+// the older ones are prunable.
+func bigToolHistory(n int) []zeroruntime.Message {
+	bigBody := strings.Repeat("x", 160_000)
+	messages := []zeroruntime.Message{{Role: zeroruntime.MessageRoleSystem, Content: "sys"}}
+	for i := 0; i < n; i++ {
+		id := string(rune('a' + i))
+		messages = append(messages,
+			zeroruntime.Message{Role: zeroruntime.MessageRoleAssistant, ToolCalls: []zeroruntime.ToolCall{{ID: id, Name: "read_file", Arguments: "{\"path\":\"" + id + "\"}"}}},
+			zeroruntime.Message{Role: zeroruntime.MessageRoleTool, ToolCallID: id, Content: bigBody},
+		)
+	}
+	return append(messages, zeroruntime.Message{Role: zeroruntime.MessageRoleUser, Content: "go"})
+}
+
+func summaryProvider(text string) *mockProvider {
+	return &mockProvider{turns: [][]zeroruntime.StreamEvent{
+		{{Type: zeroruntime.StreamEventText, Content: text}, {Type: zeroruntime.StreamEventDone}},
+		{{Type: zeroruntime.StreamEventText, Content: text}, {Type: zeroruntime.StreamEventDone}},
+	}}
+}
+
+// A context-limit error is first answered with the free prune stage; when the
+// pruned history fits the window the retry is free (no summarizer call) and
+// the one-shot reactive budget is untouched. The paid summarizer only runs
+// when a later error finds nothing left to prune.
+func TestRecoverPrunesBeforeSummarizing(t *testing.T) {
+	provider := summaryProvider("SUMMARY")
+	// The protected trailing tool output (~40k tokens) must fit the window.
+	state := newCompactionState(Options{ContextWindow: 200_000, CompactionPreserveLast: 2}, nil)
+	messages := bigToolHistory(5)
+
+	pruned, retried, err := state.recover(context.Background(), provider, messages, nil, "maximum context length exceeded")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !retried {
+		t.Fatal("prunable history that fits must produce a free retry")
+	}
+	if len(provider.requests) != 0 {
+		t.Fatalf("free prune stage must not call the summarizer (%d requests)", len(provider.requests))
+	}
+	if state.reactiveAttempted {
+		t.Fatal("a free prune retry must not consume the reactive budget")
+	}
+	if estimateTokens(pruned) >= estimateTokens(messages) {
+		t.Fatal("prune did not shrink the history")
+	}
+
+	// Second context-limit error on the pruned history: nothing meaningfully
+	// prunable remains, so the paid summarizer fires and consumes the budget.
+	_, retried, err = state.recover(context.Background(), provider, pruned, nil, "maximum context length exceeded")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !retried {
+		t.Fatal("second recover must compact and retry")
+	}
+	if len(provider.requests) == 0 {
+		t.Fatal("second recover must reach the summarizer")
+	}
+	if !state.reactiveAttempted {
+		t.Fatal("a paid recover must consume the reactive budget")
+	}
+}
+
+// When pruning reclaims something but the pruned history still exceeds the
+// window, recover falls through to the paid summarizer in the SAME call
+// instead of spending a retry that would fail the same way.
+func TestRecoverFallsThroughToSummarizerWhenPruneIsInsufficient(t *testing.T) {
+	provider := summaryProvider("SUMMARY")
+	// A 1,000-token window: even the protected trailing output overflows it.
+	state := newCompactionState(Options{ContextWindow: 1000, CompactionPreserveLast: 2}, nil)
+	messages := bigToolHistory(5)
+
+	compacted, retried, err := state.recover(context.Background(), provider, messages, nil, "maximum context length exceeded")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !retried {
+		t.Fatal("insufficient prune must still end in a retry via summarization")
+	}
+	if len(provider.requests) == 0 {
+		t.Fatal("insufficient prune must reach the summarizer in the same recover call")
+	}
+	if !state.reactiveAttempted {
+		t.Fatal("the paid summarizer consumes the reactive budget")
+	}
+	if estimateTokens(compacted) >= estimateTokens(messages) {
+		t.Fatal("recover did not shrink the history")
+	}
+}
+
+// The free prune stage is not gated behind the one-shot budget: after a paid
+// reactive compaction, a later context-limit error still gets a free retry
+// when pruning makes the history fit.
+func TestRecoverFreePruneRunsAfterBudgetIsSpent(t *testing.T) {
+	provider := summaryProvider("SUMMARY")
+	state := newCompactionState(Options{ContextWindow: 200_000, CompactionPreserveLast: 2}, nil)
+	state.reactiveAttempted = true
+	messages := bigToolHistory(5)
+
+	pruned, retried, err := state.recover(context.Background(), provider, messages, nil, "maximum context length exceeded")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !retried || len(provider.requests) != 0 {
+		t.Fatalf("free prune must retry without the summarizer even when the budget is spent (retried=%v, requests=%d)", retried, len(provider.requests))
+	}
+	if estimateTokens(pruned) >= estimateTokens(messages) {
+		t.Fatal("prune did not shrink the history")
+	}
+	// With nothing left to prune and the budget spent, the caller keeps its
+	// original error — recover neither loops nor summarizes.
+	_, retried, err = state.recover(context.Background(), provider, pruned, nil, "maximum context length exceeded")
+	if err != nil || retried || len(provider.requests) != 0 {
+		t.Fatalf("spent budget must not summarize again (retried=%v, requests=%d, err=%v)", retried, len(provider.requests), err)
+	}
+}
+
+// The free-retry gate compares against the window the provider enforces, not
+// the 70% proactive threshold: a history pruned to ~80% of the window still
+// takes the free retry.
+func TestRecoverFreeRetryGateUsesWindowNotThreshold(t *testing.T) {
+	provider := summaryProvider("SUMMARY")
+	messages := bigToolHistory(5)
+	pruned, _ := pruneStaleToolOutput(messages, 2)
+	prunedTokens := estimateTokens(pruned)
+	// Choose a window so the pruned size sits between 70% and 95% of it.
+	window := prunedTokens * 100 / 80
+	state := newCompactionState(Options{ContextWindow: window, CompactionPreserveLast: 2}, nil)
+	if prunedTokens <= state.threshold {
+		t.Fatalf("test setup: pruned size %d must exceed the proactive threshold %d", prunedTokens, state.threshold)
+	}
+
+	_, retried, err := state.recover(context.Background(), provider, messages, nil, "maximum context length exceeded")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !retried || len(provider.requests) != 0 {
+		t.Fatalf("a history that fits the window must take the free retry (retried=%v, requests=%d)", retried, len(provider.requests))
+	}
+}
+
+// After a mid-run model switch the summarizer is re-resolved against the new
+// main model: a run that started on the cheap model (no dedicated summarizer)
+// and escalated gets a cheap summarizer, and a previously broken one is
+// retried for the new model.
+func TestSwitchModelReresolvesSummarizer(t *testing.T) {
+	cheap := summaryProvider("CHEAP SUMMARY")
+	main := summaryProvider("MAIN SUMMARY")
+	var factoryModels []string
+	state := newCompactionState(Options{
+		ContextWindow:          1000,
+		CompactionPreserveLast: 2,
+		Model:                  "cheap-model",
+		Summarizer: func(_ context.Context, mainModelID string) (Provider, error) {
+			factoryModels = append(factoryModels, mainModelID)
+			if mainModelID == "cheap-model" {
+				return nil, nil // main already is the cheap model
+			}
+			return cheap, nil
+		},
+	}, nil)
+
+	if _, dedicated := state.summarizeProvider(context.Background(), main); dedicated {
+		t.Fatal("on the cheap model the main provider must summarize")
+	}
+	if _, dedicated := state.summarizeProvider(context.Background(), main); dedicated || len(factoryModels) != 1 {
+		t.Fatalf("factory must run once per main model, calls=%v", factoryModels)
+	}
+
+	state.switchModel("big-model", 400_000)
+	if state.threshold != compactionThreshold(400_000) || state.window != 400_000 {
+		t.Fatalf("switchModel must re-derive the window, got threshold=%d window=%d", state.threshold, state.window)
+	}
+	if got, dedicated := state.summarizeProvider(context.Background(), main); !dedicated || got != Provider(cheap) {
+		t.Fatal("after escalating to an expensive model the dedicated cheap summarizer must be used")
+	}
+	if len(factoryModels) != 2 || factoryModels[1] != "big-model" {
+		t.Fatalf("factory must be re-invoked with the new main model, calls=%v", factoryModels)
+	}
+
+	// A broken summarizer is retried after the next switch.
+	state.summarizerBroken = true
+	if _, dedicated := state.summarizeProvider(context.Background(), main); dedicated {
+		t.Fatal("a broken summarizer must fall back to main")
+	}
+	state.switchModel("other-model", 0)
+	if got, dedicated := state.summarizeProvider(context.Background(), main); !dedicated || got != Provider(cheap) {
+		t.Fatal("a model switch must clear the broken flag and rebuild the summarizer")
+	}
+}
+
+func TestCompactionUpdateWindow(t *testing.T) {
+	state := newCompactionState(Options{ContextWindow: 100_000, CompactionPreserveLast: 2}, nil)
+	state.lowWaterMark = 50_000
+	before := state.threshold
+
+	state.updateWindow(0) // unknown model: unchanged
+	if state.threshold != before || state.lowWaterMark != 50_000 {
+		t.Fatal("updateWindow(0) must be a no-op")
+	}
+
+	state.updateWindow(10_000)
+	if state.threshold >= before {
+		t.Fatalf("threshold = %d, want re-derived below %d", state.threshold, before)
+	}
+	if state.lowWaterMark != 0 {
+		t.Fatal("low-water mark must reset on a window change")
+	}
+	if !state.enabled {
+		t.Fatal("a positive window must keep compaction enabled")
+	}
+}
+
+// End-to-end: after escalate_model switches providers, the loop re-derives the
+// context window via ContextWindowFor, visible in the OnContext report.
+func TestRunEscalationRederivesContextWindow(t *testing.T) {
+	registry := tools.NewRegistry()
+	registry.Register(escalatingTool{target: "big-model"})
+	switched := &mockProvider{turns: [][]zeroruntime.StreamEvent{{
+		{Type: zeroruntime.StreamEventText, Content: "done"},
+		{Type: zeroruntime.StreamEventDone},
+	}}}
+	provider := &mockProvider{turns: [][]zeroruntime.StreamEvent{{
+		{Type: zeroruntime.StreamEventToolCallStart, ToolCallID: "call-1", ToolName: "escalate"},
+		{Type: zeroruntime.StreamEventToolCallDelta, ToolCallID: "call-1", ArgumentsFragment: `{}`},
+		{Type: zeroruntime.StreamEventToolCallEnd, ToolCallID: "call-1"},
+		{Type: zeroruntime.StreamEventDone},
+	}}}
+
+	var windows []int
+	result, err := Run(context.Background(), "go", provider, Options{
+		Registry:      registry,
+		ContextWindow: 100_000,
+		Model:         "small-model",
+		ModelSwitcher: func(context.Context, string) (Provider, error) { return switched, nil },
+		ContextWindowFor: func(modelID string) int {
+			if modelID == "big-model" {
+				return 400_000
+			}
+			return 0
+		},
+		OnContext: func(breakdown ContextBreakdown) { windows = append(windows, breakdown.ContextWindow) },
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if result.FinalAnswer != "done" {
+		t.Fatalf("final answer = %q", result.FinalAnswer)
+	}
+	if len(windows) < 2 {
+		t.Fatalf("expected at least 2 OnContext reports, got %d", len(windows))
+	}
+	if windows[0] != 100_000 {
+		t.Fatalf("turn 1 window = %d, want the original 100000", windows[0])
+	}
+	if windows[len(windows)-1] != 400_000 {
+		t.Fatalf("post-escalation window = %d, want 400000", windows[len(windows)-1])
+	}
+}
+
+// nonComparableProvider has a slice field, so comparing two Provider interface
+// values holding it would panic; the summarizer path must never compare them.
+type nonComparableProvider struct {
+	inner *mockProvider
+	tags  []string
+}
+
+func (p nonComparableProvider) StreamCompletion(ctx context.Context, request zeroruntime.CompletionRequest) (<-chan zeroruntime.StreamEvent, error) {
+	return p.inner.StreamCompletion(ctx, request)
+}
+
+func TestSummarizeClosureNeverComparesProviderValues(t *testing.T) {
+	broken := &mockProvider{turns: [][]zeroruntime.StreamEvent{{
+		{Type: zeroruntime.StreamEventError, Error: "boom: model not found"},
+	}}}
+	main := nonComparableProvider{inner: summaryProvider("MAIN SUMMARY"), tags: []string{"x"}}
+	state := newCompactionState(Options{
+		ContextWindow:          1000,
+		CompactionPreserveLast: 2,
+		Summarizer:             func(context.Context, string) (Provider, error) { return nonComparableProvider{inner: broken}, nil },
+	}, nil)
+	summary, err := state.summarizeClosure(context.Background(), main)([]zeroruntime.Message{{Role: zeroruntime.MessageRoleUser, Content: "history"}})
+	if err != nil || summary != "MAIN SUMMARY" {
+		t.Fatalf("fallback with non-comparable providers must not panic and must use main, got %q err=%v", summary, err)
+	}
+	// And with no dedicated summarizer at all the closure still works.
+	state = newCompactionState(Options{ContextWindow: 1000, CompactionPreserveLast: 2}, nil)
+	if _, err := state.summarizeClosure(context.Background(), main)([]zeroruntime.Message{{Role: zeroruntime.MessageRoleUser, Content: "history"}}); err != nil {
+		t.Fatal(err)
+	}
+}

internal/agent/compaction_metadata_test.go

@@ -43,8 +43,14 @@ func TestCompactMessagesReturnsMetadataForManualCompaction(t *testing.T) {
 	if result.SummaryText != "manual summary" {
 		t.Fatalf("SummaryText = %q, want trimmed summary", result.SummaryText)
 	}
-	if len(captured) != 3 || captured[0].Content != "first question" || captured[2].Content != "second question" {
-		t.Fatalf("summarized middle = %#v, want the three non-preserved non-system messages", captured)
+	if result.ProjectedChars != compactionMessageChars(captured) || result.ProjectedChars == 0 {
+		t.Fatalf("ProjectedChars = %d, want %d", result.ProjectedChars, compactionMessageChars(captured))
+	}
+	if result.Truncated {
+		t.Fatal("Truncated = true for a projection within its limits")
+	}
+	if len(captured) != 1 || !strings.Contains(captured[0].Content, "first question") || !strings.Contains(captured[0].Content, "second question") {
+		t.Fatalf("summarized projection = %#v, want intent from the non-preserved middle", captured)
 	}
 	if len(result.Messages) != 4 {
 		t.Fatalf("compacted message count = %d, want 4", len(result.Messages))
@@ -96,11 +102,46 @@ func TestCompactMessagesNoopReturnsUncompactedMetadata(t *testing.T) {
 	if result.SummaryText != "" {
 		t.Fatalf("SummaryText = %q, want empty", result.SummaryText)
 	}
+	if result.ProjectedChars != 0 || result.Truncated {
+		t.Fatalf("no-op projection metadata = (%d, %t), want zero values", result.ProjectedChars, result.Truncated)
+	}
 	if !reflect.DeepEqual(result.Messages, messages) {
 		t.Fatalf("Messages changed on no-op: %#v", result.Messages)
 	}
 }
 
+func TestCompactMessagesReturnsTruncatedProjectionMetadata(t *testing.T) {
+	messages := []zeroruntime.Message{{Role: zeroruntime.MessageRoleSystem, Content: "system"}}
+	for index := range 20 {
+		messages = append(messages, zeroruntime.Message{
+			Role:    zeroruntime.MessageRoleUser,
+			Content: strings.Repeat("contextword ", 256) + string(rune('a'+index)),
+		})
+	}
+	messages = append(messages,
+		zeroruntime.Message{Role: zeroruntime.MessageRoleAssistant, Content: "recent answer"},
+		zeroruntime.Message{Role: zeroruntime.MessageRoleUser, Content: "latest question"},
+	)
+
+	var captured []zeroruntime.Message
+	result, err := CompactMessages(messages, CompactionOptions{
+		PreserveLast: 2,
+		Summarize: func(toSummarize []zeroruntime.Message) (string, error) {
+			captured = append([]zeroruntime.Message(nil), toSummarize...)
+			return "bounded summary", nil
+		},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !result.Compacted || !result.Truncated {
+		t.Fatalf("compaction metadata = compacted %t, truncated %t; want both true", result.Compacted, result.Truncated)
+	}
+	if result.ProjectedChars != compactionMessageChars(captured) || result.ProjectedChars == 0 {
+		t.Fatalf("ProjectedChars = %d, want %d", result.ProjectedChars, compactionMessageChars(captured))
+	}
+}
+
 func TestCompactMessagesPropagatesSummarizeError(t *testing.T) {
 	messages := []zeroruntime.Message{
 		{Role: zeroruntime.MessageRoleSystem, Content: "system"},

internal/agent/compaction_preserve.go

@@ -20,6 +20,20 @@ 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
+	// maxTaskObjectiveBytes keeps the original objective recognizable after
+	// compaction without allowing an unusually large prompt to recreate the
+	// context pressure the compaction just removed.
+	maxTaskObjectiveBytes = 512
 )
 
 const (
@@ -66,32 +80,14 @@ func extractLatestPlan(messages []zeroruntime.Message) string {
 // ({"plan":[{content,status,...}]}) as terse status-tagged bullet lines. Returns
 // "" on malformed arguments or an empty plan.
 func formatPlanArguments(arguments string) string {
-	var parsed struct {
-		Plan []struct {
-			Content string `json:"content"`
-			Step    string `json:"step"`
-			Status  string `json:"status"`
-			Notes   string `json:"notes"`
-		} `json:"plan"`
-	}
-	if err := json.Unmarshal([]byte(strings.TrimSpace(arguments)), &parsed); err != nil {
+	plan, ok := parseTaskPlan(arguments)
+	if !ok {
 		return ""
 	}
-	lines := make([]string, 0, len(parsed.Plan))
-	for _, item := range parsed.Plan {
-		content := strings.TrimSpace(item.Content)
-		if content == "" {
-			content = strings.TrimSpace(item.Step)
-		}
-		if content == "" {
-			continue
-		}
-		status := strings.TrimSpace(item.Status)
-		if status == "" {
-			status = "pending"
-		}
-		line := "- [" + status + "] " + content
-		if notes := strings.TrimSpace(item.Notes); notes != "" {
+	lines := make([]string, 0, len(plan))
+	for _, item := range plan {
+		line := "- [" + item.Status + "] " + item.Content
+		if notes := item.Notes; notes != "" {
 			line += "\n  Notes: " + notes
 		}
 		lines = append(lines, line)
@@ -100,12 +96,136 @@ 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 files mutated by editing 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{}
+	noteByPath := 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)
+		}
+	}
+	for _, message := range messages {
+		if message.Role != zeroruntime.MessageRoleTool || len(message.ChangedFiles) == 0 {
+			continue
+		}
+		for _, path := range message.ChangedFiles {
+			path = strings.TrimSpace(path)
+			if path == "" {
+				continue
+			}
+			sequence = append(sequence, path)
+			noteByPath[path] = editNote(message.Content)
+		}
+	}
+	order := lastSeenOrder(sequence)
+	if len(order) == 0 {
+		return nil
+	}
+
+	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 +430,35 @@ func capBody(body string) string {
 // preservedState is the JSON shape of the carried-across-compaction block.
 type preservedState struct {
 	Plan                string                 `json:"plan,omitempty"`
+	Task                *preservedTaskState    `json:"task,omitempty"`
+	RecentEdits         []preservedEdit        `json:"recent_edits,omitempty"`
 	Tools               []preservedTool        `json:"tools,omitempty"`
 	Skills              []preservedSkill       `json:"skills,omitempty"`
 	ProjectInstructions []preservedInstruction `json:"project_instructions,omitempty"`
 }
 
+type preservedTaskState struct {
+	Objective           string              `json:"objective"`
+	Status              taskStatus          `json:"status,omitempty"`
+	Pending             int                 `json:"pending,omitempty"`
+	InProgress          int                 `json:"in_progress,omitempty"`
+	Completed           int                 `json:"completed,omitempty"`
+	Failed              int                 `json:"failed,omitempty"`
+	VerificationPassed  int                 `json:"verification_passed,omitempty"`
+	VerificationFailed  int                 `json:"verification_failed,omitempty"`
+	VerificationOutcome Outcome             `json:"verification_outcome,omitempty"`
+	Constraints         []string            `json:"constraints,omitempty"`
+	ChangedFiles        []string            `json:"changed_files,omitempty"`
+	UnresolvedFailures  []taskFailureState  `json:"unresolved_failures,omitempty"`
+	Approvals           []taskApprovalState `json:"approvals,omitempty"`
+	Artifacts           []taskArtifactState `json:"artifacts,omitempty"`
+}
+
+type preservedEdit struct {
+	Path string `json:"path"`
+	Note string `json:"note,omitempty"`
+}
+
 type preservedTool struct {
 	Name string `json:"name"`
 	Body string `json:"body"`
@@ -337,8 +481,39 @@ type preservedInstruction struct {
 // may live only inside the injected summary message, which on a later compaction
 // lands in middle with no real tool calls left to extract. Fresh tool calls and
 // instruction blocks override the carried-forward copy by name/source.
-func appendPreservedState(summary string, middle []zeroruntime.Message) string {
+func appendPreservedState(summary string, middle []zeroruntime.Message, taskSnapshot *taskStateSnapshot) string {
 	priorState := parsePreservedStateBlock(latestSummaryContent(middle))
+	task := priorState.Task
+	if taskSnapshot != nil {
+		task = &preservedTaskState{
+			Objective:          capTaskObjective(taskSnapshot.Objective),
+			Constraints:        mergeBoundedComparable(priorTaskConstraints(priorState.Task), taskSnapshot.Constraints, maxTaskConstraints),
+			ChangedFiles:       mergeBoundedComparable(priorTaskChangedFiles(priorState.Task), taskSnapshot.ChangedFiles, maxTaskEvidenceEntries),
+			UnresolvedFailures: mergeTaskFailures(priorTaskFailures(priorState.Task), taskSnapshot.UnresolvedFailures, taskSnapshot.ResolvedFailureKeys),
+			Approvals:          mergeBoundedComparable(priorTaskApprovals(priorState.Task), taskSnapshot.Approvals, maxTaskEvidenceEntries),
+			Artifacts:          mergeBoundedComparable(priorTaskArtifacts(priorState.Task), taskSnapshot.Artifacts, maxTaskEvidenceEntries),
+		}
+		// Plan parity corroborates only the mutable task projection. The objective
+		// comes directly from the run prompt and is immutable, so it must survive
+		// even after compaction removes the plan tool call needed for comparison.
+		if taskSnapshot.PlanParity == taskPlanParityMatch {
+			task.Status = taskSnapshot.Status
+			task.Pending = taskSnapshot.Plan.Pending
+			task.InProgress = taskSnapshot.Plan.InProgress
+			task.Completed = taskSnapshot.Plan.Completed
+			task.Failed = taskSnapshot.Plan.Failed
+			task.VerificationPassed = taskSnapshot.Verification.Passed
+			task.VerificationFailed = taskSnapshot.Verification.Failed
+			task.VerificationOutcome = taskSnapshot.Verification.LastOutcome
+		}
+	}
+	freshConstraints := explicitConstraintsFromMessages(middle)
+	if len(freshConstraints) > 0 {
+		if task == nil {
+			task = &preservedTaskState{}
+		}
+		task.Constraints = mergeBoundedComparable(task.Constraints, freshConstraints, maxTaskConstraints)
+	}
 
 	// Plan: a fresh update_plan in middle is authoritative; otherwise carry the
 	// plan preserved by an earlier compaction.
@@ -347,6 +522,12 @@ 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 +541,178 @@ func appendPreservedState(summary string, middle []zeroruntime.Message) string {
 		projectInstructionEntries(middle),
 	)
 
-	if block := formatPreservedState(plan, tools, skills, instructions); block != "" {
+	if block := formatPreservedState(plan, task, edits, tools, skills, instructions); block != "" {
 		summary += "\n\n" + block
 	}
 	return summary
 }
 
+func explicitConstraintsFromMessages(messages []zeroruntime.Message) []string {
+	var constraints []string
+	for _, message := range messages {
+		if message.Role != zeroruntime.MessageRoleUser || strings.Contains(message.Content, preservedStateLabel) {
+			continue
+		}
+		if _, body := projectInstructionBlock(message.Content); body != "" {
+			continue
+		}
+		constraints = mergeBoundedComparable(constraints, extractExplicitConstraints(message.Content), maxTaskConstraints)
+	}
+	return constraints
+}
+
+func priorTaskConstraints(task *preservedTaskState) []string {
+	if task == nil {
+		return nil
+	}
+	return task.Constraints
+}
+
+func priorTaskChangedFiles(task *preservedTaskState) []string {
+	if task == nil {
+		return nil
+	}
+	return task.ChangedFiles
+}
+
+func priorTaskFailures(task *preservedTaskState) []taskFailureState {
+	if task == nil {
+		return nil
+	}
+	return task.UnresolvedFailures
+}
+
+func priorTaskApprovals(task *preservedTaskState) []taskApprovalState {
+	if task == nil {
+		return nil
+	}
+	return task.Approvals
+}
+
+func priorTaskArtifacts(task *preservedTaskState) []taskArtifactState {
+	if task == nil {
+		return nil
+	}
+	return task.Artifacts
+}
+
+func mergeTaskFailures(older, newer []taskFailureState, resolved []string) []taskFailureState {
+	resolvedSet := make(map[string]struct{}, len(resolved))
+	for _, key := range resolved {
+		resolvedSet[key] = struct{}{}
+	}
+	merged := make([]taskFailureState, 0, len(older)+len(newer))
+	for _, failure := range append(append([]taskFailureState(nil), older...), newer...) {
+		if _, ok := resolvedSet[failure.Key]; ok {
+			continue
+		}
+		withoutSameKey := make([]taskFailureState, 0, len(merged))
+		for _, existing := range merged {
+			if existing.Key != failure.Key {
+				withoutSameKey = append(withoutSameKey, existing)
+			}
+		}
+		merged = append(withoutSameKey, failure)
+	}
+	if len(merged) > maxTaskEvidenceEntries {
+		merged = merged[len(merged)-maxTaskEvidenceEntries:]
+	}
+	return merged
+}
+
+func mergeBoundedComparable[T comparable](older, newer []T, limit int) []T {
+	merged := make([]T, 0, len(older)+len(newer))
+	seen := make(map[T]struct{}, len(older)+len(newer))
+	for _, value := range append(append([]T(nil), older...), newer...) {
+		if _, ok := seen[value]; ok {
+			continue
+		}
+		seen[value] = struct{}{}
+		merged = append(merged, value)
+	}
+	if len(merged) > limit {
+		merged = merged[len(merged)-limit:]
+	}
+	return merged
+}
+
+func capTaskObjective(objective string) string {
+	objective = strings.TrimSpace(objective)
+	if len(objective) <= maxTaskObjectiveBytes {
+		return objective
+	}
+	const suffix = "…"
+	limit := maxTaskObjectiveBytes - len(suffix)
+	for limit > 0 && !utf8.RuneStart(objective[limit]) {
+		limit--
+	}
+	return strings.TrimSpace(objective[:limit]) + suffix
+}
+
+// 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, task *preservedTaskState, edits, tools, skills, instructions []skillEntry) string {
+	if plan == "" && task == nil && len(edits) == 0 && len(tools) == 0 && len(skills) == 0 && len(instructions) == 0 {
 		return ""
 	}
-	state := preservedState{Plan: plan}
+	state := preservedState{Plan: plan, Task: task}
+	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})
 	}
@@ -389,15 +729,6 @@ func formatPreservedState(plan string, tools, skills, instructions []skillEntry)
 	return preservedStateLabel + "\n" + string(encoded)
 }
 
-// parsePreservedState recovers the plan + skills from a prior summary's preserved
-// block. JSON escaping makes this lossless even when a skill body contains
-// markdown headings, code fences, or quotes. Returns ("", nil) when absent or
-// malformed.
-func parsePreservedState(summaryContent string) (string, []skillEntry) {
-	state := parsePreservedStateBlock(summaryContent)
-	return state.Plan, preservedSkillsToEntries(state.Skills)
-}
-
 func parsePreservedStateBlock(summaryContent string) preservedState {
 	idx := strings.LastIndex(summaryContent, preservedStateLabel)
 	if idx < 0 {

internal/agent/compaction_preserve_test.go

@@ -1,10 +1,12 @@
 package agent
 
 import (
+	"encoding/json"
 	"strings"
 	"testing"
 	"unicode/utf8"
 
+	"github.com/Gitlawb/zero/internal/tools"
 	"github.com/Gitlawb/zero/internal/zeroruntime"
 )
 
@@ -59,6 +61,169 @@ func TestCompactPreservesActivePlan(t *testing.T) {
 	}
 }
 
+func TestCompactPreservesBoundedTaskContext(t *testing.T) {
+	objective := strings.Repeat("世", maxTaskObjectiveBytes)
+	task := newTaskState(objective, nil)
+	task.observe(taskStateEvent{kind: taskStateEventPlan, arguments: `{"plan":[{"content":"write code","status":"in_progress"},{"content":"add tests","status":"pending"}]}`})
+	task.observe(taskStateEvent{kind: taskStateEventToolResult, arguments: `{"patch":"*** Update File: internal/db.go"}`, toolResult: ToolResult{
+		Name: "apply_patch", Status: tools.StatusOK, Output: "Done!", ChangedFiles: []string{"internal/db.go"},
+	}})
+	messages := stateConversation()
+	compacted, err := Compact(messages, CompactionOptions{
+		PreserveLast: 2,
+		Summarize:    func([]zeroruntime.Message) (string, error) { return "SUMMARY", nil },
+		taskState:    task.snapshotForCompaction(messages),
+	})
+	if err != nil {
+		t.Fatalf("Compact: %v", err)
+	}
+	state := parsePreservedStateBlock(compacted[1].Content)
+	if state.Task == nil || state.Task.Status != taskStatusActive || state.Task.InProgress != 1 || state.Task.Pending != 1 {
+		t.Fatalf("unexpected compact task state: %#v", state.Task)
+	}
+	if len(state.Task.Objective) > maxTaskObjectiveBytes || !utf8.ValidString(state.Task.Objective) {
+		t.Fatalf("objective was not safely bounded: %d bytes %q", len(state.Task.Objective), state.Task.Objective)
+	}
+	if len(state.Task.ChangedFiles) != 1 || state.Task.ChangedFiles[0] != "internal/db.go" {
+		t.Fatalf("changed files were not preserved: %#v", state.Task.ChangedFiles)
+	}
+}
+
+func TestCompactPreservesRuntimeEvidenceAcrossRepeatedCompaction(t *testing.T) {
+	task := newTaskState("Please keep the change focused.", nil)
+	task.observe(taskStateEvent{kind: taskStateEventToolResult, arguments: `{"cmd":"go test ./..."}`, toolResult: ToolResult{
+		Name: "exec_command", Status: tools.StatusError, Output: "Error: tests failed",
+		Meta: map[string]string{"spill_path": ".zero/artifacts/tests.txt"},
+	}})
+	task.observe(taskStateEvent{kind: taskStateEventPermission, permission: PermissionEvent{
+		ToolName: "exec_command", DecisionAction: PermissionDecisionAllowForSession, Scope: "/tmp",
+	}})
+	messages := stateConversation()
+
+	first, err := Compact(messages, CompactionOptions{
+		PreserveLast: 2,
+		Summarize:    func([]zeroruntime.Message) (string, error) { return "FIRST", nil },
+		taskState:    task.snapshotForCompaction(messages),
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	task.observe(taskStateEvent{kind: taskStateEventToolResult, arguments: `{"cmd":"go test ./..."}`, toolResult: ToolResult{
+		Name: "exec_command", Status: tools.StatusError, Output: "Error: tests still fail with newer evidence",
+		Meta: map[string]string{"spill_path": ".zero/artifacts/tests.txt"},
+	}})
+	secondInput := append(append([]zeroruntime.Message{}, first...),
+		zeroruntime.Message{Role: zeroruntime.MessageRoleUser, Content: "Never add background memory models."},
+		zeroruntime.Message{Role: zeroruntime.MessageRoleAssistant, Content: "understood"},
+		zeroruntime.Message{Role: zeroruntime.MessageRoleUser, Content: "continue"},
+		zeroruntime.Message{Role: zeroruntime.MessageRoleAssistant, Content: "working"},
+	)
+	second, err := Compact(secondInput, CompactionOptions{
+		PreserveLast: 2,
+		Summarize:    func([]zeroruntime.Message) (string, error) { return "SECOND", nil },
+		taskState:    task.snapshotForCompaction(secondInput),
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	state := parsePreservedStateBlock(second[1].Content)
+	if state.Task == nil || len(state.Task.UnresolvedFailures) != 1 || len(state.Task.Approvals) != 1 || len(state.Task.Artifacts) != 1 {
+		t.Fatalf("runtime evidence did not survive repeated compaction: %#v", state.Task)
+	}
+	if got := state.Task.UnresolvedFailures[0].Summary; got != "Error: tests still fail with newer evidence" {
+		t.Fatalf("newer failure evidence did not replace the older summary: %#v", state.Task.UnresolvedFailures)
+	}
+	for _, want := range []string{"Please keep the change focused.", "Never add background memory models."} {
+		if !containsString(state.Task.Constraints, want) {
+			t.Fatalf("constraint %q missing after repeated compaction: %#v", want, state.Task.Constraints)
+		}
+	}
+}
+
+func containsString(values []string, want string) bool {
+	for _, value := range values {
+		if value == want {
+			return true
+		}
+	}
+	return false
+}
+
+func TestCompactPreservesObjectiveAfterPlanParityMismatch(t *testing.T) {
+	prior := preservedState{Task: &preservedTaskState{Objective: "stale objective", Status: taskStatusActive, Pending: 1}}
+	encoded, err := json.Marshal(prior)
+	if err != nil {
+		t.Fatal(err)
+	}
+	messages := []zeroruntime.Message{
+		{Role: zeroruntime.MessageRoleSystem, Content: "system"},
+		{Role: zeroruntime.MessageRoleUser, Content: summaryLabel + "\nold\n\n" + preservedStateLabel + "\n" + string(encoded)},
+		{Role: zeroruntime.MessageRoleAssistant, Content: "continuing"},
+		{Role: zeroruntime.MessageRoleUser, Content: "more"},
+		{Role: zeroruntime.MessageRoleAssistant, Content: "working"},
+		{Role: zeroruntime.MessageRoleUser, Content: "again"},
+		{Role: zeroruntime.MessageRoleAssistant, Content: "done"},
+	}
+	compacted, err := Compact(messages, CompactionOptions{
+		PreserveLast: 2,
+		Summarize:    func([]zeroruntime.Message) (string, error) { return "SUMMARY", nil },
+		taskState: &taskStateSnapshot{
+			Objective:  "current objective",
+			PlanParity: taskPlanParityMismatch,
+		},
+	})
+	if err != nil {
+		t.Fatalf("Compact: %v", err)
+	}
+	state := parsePreservedStateBlock(compacted[1].Content)
+	if state.Task == nil || state.Task.Objective != "current objective" {
+		t.Fatalf("immutable objective was lost on plan mismatch: %#v", state.Task)
+	}
+	if state.Task.Status != "" || state.Task.Pending != 0 {
+		t.Fatalf("uncorroborated mutable task fields survived plan mismatch: %#v", state.Task)
+	}
+}
+
+func TestTaskObjectiveSurvivesRepeatedCompactionWithoutPlanRefresh(t *testing.T) {
+	task := newTaskState("keep this objective", nil)
+	task.observe(taskStateEvent{kind: taskStateEventPlan, arguments: `{"plan":[{"content":"write code","status":"in_progress"},{"content":"add tests","status":"pending"}]}`})
+	messages := stateConversation()
+
+	first, err := Compact(messages, CompactionOptions{
+		PreserveLast: 2,
+		Summarize:    func([]zeroruntime.Message) (string, error) { return "FIRST", nil },
+		taskState:    task.snapshotForCompaction(messages),
+	})
+	if err != nil {
+		t.Fatalf("first Compact: %v", err)
+	}
+	secondInput := append(append([]zeroruntime.Message{}, first...),
+		zeroruntime.Message{Role: zeroruntime.MessageRoleUser, Content: "more"},
+		zeroruntime.Message{Role: zeroruntime.MessageRoleAssistant, Content: "working"},
+		zeroruntime.Message{Role: zeroruntime.MessageRoleUser, Content: "again"},
+		zeroruntime.Message{Role: zeroruntime.MessageRoleAssistant, Content: "done"},
+	)
+	snapshot := task.snapshotForCompaction(secondInput)
+	if snapshot.PlanParity != taskPlanParityMismatch {
+		t.Fatalf("plan call should be absent after first compaction, parity=%q", snapshot.PlanParity)
+	}
+	second, err := Compact(secondInput, CompactionOptions{
+		PreserveLast: 2,
+		Summarize:    func([]zeroruntime.Message) (string, error) { return "SECOND", nil },
+		taskState:    snapshot,
+	})
+	if err != nil {
+		t.Fatalf("second Compact: %v", err)
+	}
+	state := parsePreservedStateBlock(second[1].Content)
+	if state.Task == nil || state.Task.Objective != "keep this objective" {
+		t.Fatalf("objective lost after second compaction: %#v", state.Task)
+	}
+	if state.Task.Status != "" || state.Task.InProgress != 0 {
+		t.Fatalf("uncorroborated mutable fields survived second compaction: %#v", state.Task)
+	}
+}
+
 func TestCompactPreservesLoadedSkills(t *testing.T) {
 	summary := compactStateConversation(t, stateConversation())
 	if !strings.Contains(summary, preservedStateLabel) {
@@ -283,12 +448,10 @@ func TestExtractLatestPlanReturnsMostRecent(t *testing.T) {
 	}
 }
 
-func TestFormatPlanArgumentsAcceptsStepAlias(t *testing.T) {
+func TestFormatPlanArgumentsRejectsUnsupportedStepAlias(t *testing.T) {
 	got := formatPlanArguments(`{"plan":[{"step":"write failing test","status":"in_progress"},{"content":"keep existing shape","status":"pending"}]}`)
-	for _, want := range []string{"- [in_progress] write failing test", "- [pending] keep existing shape"} {
-		if !strings.Contains(got, want) {
-			t.Fatalf("expected %q in formatted plan, got %q", want, got)
-		}
+	if got != "" {
+		t.Fatalf("unsupported alias should reject the whole plan like update_plan, got %q", got)
 	}
 }
 

internal/agent/compaction_projection.go

@@ -0,0 +1,315 @@
+package agent
+
+import (
+	"encoding/json"
+	"fmt"
+	"regexp"
+	"strings"
+
+	"github.com/Gitlawb/zero/internal/zeroruntime"
+)
+
+const (
+	compactionUserWordBudget       = 256
+	compactionAssistantWordBudget  = 200
+	compactionToolCallsPerTurn     = 8
+	compactionToolArgumentBytes    = 120
+	compactionErrorBytes           = 150
+	compactionResultBytes          = 1200
+	compactionPreviousSummaryBytes = 16 * 1024
+	compactionBriefMaxBytes        = 24 * 1024
+)
+
+var compactionWordPattern = regexp.MustCompile(`[\p{L}\p{N}]+`)
+
+type compactionProjection struct {
+	messages  []zeroruntime.Message
+	truncated bool
+}
+
+// projectCompactionInput removes reconstructible tool output before the model
+// summary call. It retains a bounded transcript of user intent, assistant
+// decisions, tool actions, and concise errors. Exact durable state is appended
+// separately from the original messages after the summary returns.
+func projectCompactionInput(messages []zeroruntime.Message) compactionProjection {
+	sections := make([]string, 0, len(messages))
+	previousSummary := ""
+	truncated := false
+	for index, message := range messages {
+		switch message.Role {
+		case zeroruntime.MessageRoleUser:
+			content := message.Content
+			if marker := strings.Index(content, preservedStateLabel); marker >= 0 {
+				content = content[:marker]
+			}
+			trimmed := strings.TrimSpace(content)
+			if strings.HasPrefix(trimmed, summaryLabel) {
+				fullSummary := strings.TrimSpace(strings.TrimPrefix(trimmed, summaryLabel))
+				previousSummary = clipHeadTailBytes(fullSummary, compactionPreviousSummaryBytes)
+				truncated = truncated || len(fullSummary) > compactionPreviousSummaryBytes
+				continue
+			}
+			if content = clipInformativeWords(content, compactionUserWordBudget); content != "" {
+				sections = append(sections, fmt.Sprintf("[user #%d]\n%s", index, content))
+			}
+		case zeroruntime.MessageRoleAssistant:
+			lines := make([]string, 0, 1+len(message.ToolCalls))
+			if content := clipInformativeWords(message.Content, compactionAssistantWordBudget); content != "" {
+				lines = append(lines, content)
+			}
+			calls := message.ToolCalls
+			omitted := 0
+			if len(calls) > compactionToolCallsPerTurn {
+				omitted = len(calls) - compactionToolCallsPerTurn
+				calls = calls[omitted:]
+			}
+			if omitted > 0 {
+				lines = append(lines, fmt.Sprintf("* (%d earlier tool calls omitted)", omitted))
+			}
+			for _, call := range calls {
+				lines = append(lines, compactionToolCallLine(call))
+			}
+			if len(lines) > 0 {
+				sections = append(sections, fmt.Sprintf("[assistant #%d]\n%s", index, strings.Join(lines, "\n")))
+			}
+		case zeroruntime.MessageRoleTool:
+			// New tool-result messages carry the execution status directly, as the
+			// source ToolResult does. Keep the text check for older session history
+			// created before Message exposed IsError.
+			name := toolNameForResult(messages, index)
+			mustPreserve := name == "ask_user" || len(message.ChangedFiles) > 0
+			if !mustPreserve && !message.IsError && !isLikelyToolError(message.Content) {
+				continue
+			}
+			if name == "" {
+				name = "tool"
+			}
+			switch {
+			case message.IsError || isLikelyToolError(message.Content):
+				sections = append(sections, fmt.Sprintf("[tool_error #%d] %s\n%s", index, name, clipBytes(firstLine(message.Content), compactionErrorBytes)))
+			case name == "ask_user":
+				sections = append(sections, fmt.Sprintf("[user_answer #%d] %s\n%s", index, name, clipBytes(message.Content, compactionResultBytes)))
+			default:
+				sections = append(sections, fmt.Sprintf("[tool_result #%d] %s changed %s\n%s", index, name,
+					clipBytes(strings.Join(message.ChangedFiles, ", "), compactionToolArgumentBytes), clipBytes(firstLine(message.Content), compactionErrorBytes)))
+			}
+		}
+	}
+	if len(sections) == 0 && previousSummary == "" {
+		return compactionProjection{}
+	}
+	brief, briefTruncated := capCompactionBrief(strings.Join(sections, "\n\n"))
+	truncated = truncated || briefTruncated
+	if previousSummary != "" {
+		brief = "[previous summary]\n" + previousSummary + "\n\n" + brief
+	}
+	return compactionProjection{
+		messages:  []zeroruntime.Message{{Role: zeroruntime.MessageRoleUser, Content: strings.TrimSpace(brief)}},
+		truncated: truncated,
+	}
+}
+
+func capCompactionBrief(brief string) (string, bool) {
+	if len(brief) <= compactionBriefMaxBytes {
+		return brief, false
+	}
+	marker := "\n\n...[middle transcript omitted to fit compaction budget]...\n\n"
+	available := compactionBriefMaxBytes - len(marker)
+	// Keep both chronological edges: the oldest material has not necessarily
+	// been summarized before, while the newest material is most actionable.
+	headBytes := available * 2 / 5
+	tailBytes := available - headBytes
+	head := clipPrefixAtBoundary(brief, headBytes)
+	tail := clipSuffixAtBoundary(brief, tailBytes)
+	return strings.TrimSpace(head) + marker + strings.TrimSpace(tail), true
+}
+
+func compactionToolCallLine(call zeroruntime.ToolCall) string {
+	detail := ""
+	var arguments map[string]any
+	if json.Unmarshal([]byte(strings.TrimSpace(call.Arguments)), &arguments) == nil {
+		for _, key := range []string{"file_path", "path", "url", "query", "pattern", "prompt", "description", "question"} {
+			if value, ok := arguments[key].(string); ok && strings.TrimSpace(value) != "" {
+				detail = value
+				break
+			}
+		}
+		if detail == "" && call.Name == "ask_user" {
+			detail = askUserQuestionsDetail(arguments)
+		}
+		if detail == "" && call.Name == "apply_patch" {
+			if patch, ok := arguments["patch"].(string); ok {
+				detail = strings.Join(patchPaths(patch), ", ")
+			}
+		}
+	}
+	if detail == "" {
+		detail = commandFromArguments(call.Arguments)
+	}
+	if detail == "" {
+		return "* " + call.Name
+	}
+	return fmt.Sprintf("* %s %q", call.Name, clipBytes(detail, compactionToolArgumentBytes))
+}
+
+func isLikelyToolError(content string) bool {
+	prefix := strings.TrimSpace(content)
+	if len(prefix) > compactionErrorBytes {
+		prefix = prefix[:compactionErrorBytes]
+	}
+	lower := strings.ToLower(prefix)
+	for _, prefix := range []string{"error:", "tool error:", "failed:", "permission denied", "command failed"} {
+		if strings.HasPrefix(lower, prefix) {
+			return true
+		}
+	}
+	return false
+}
+
+func clipInformativeWords(text string, limit int) string {
+	flat := strings.Join(strings.Fields(text), " ")
+	if flat == "" || limit <= 0 {
+		return ""
+	}
+	count := 0
+	for _, location := range compactionWordPattern.FindAllStringIndex(flat, -1) {
+		word := strings.ToLower(flat[location[0]:location[1]])
+		if compactionStopWords[word] {
+			continue
+		}
+		count++
+		if count > limit {
+			end := location[0]
+			return strings.TrimSpace(flat[:end]) + "...(truncated)"
+		}
+	}
+	return flat
+}
+
+func clipBytes(text string, limit int) string {
+	text = strings.TrimSpace(text)
+	if limit <= 0 {
+		return ""
+	}
+	if len(text) <= limit {
+		return text
+	}
+	if limit <= 3 {
+		end := limit
+		for end > 0 && end < len(text) && text[end]&0xc0 == 0x80 {
+			end--
+		}
+		return text[:end]
+	}
+	end := limit - 3
+	for end > 0 && end < len(text) && text[end]&0xc0 == 0x80 {
+		end--
+	}
+	return strings.TrimSpace(text[:end]) + "..."
+}
+
+func clipHeadTailBytes(text string, limit int) string {
+	text = strings.TrimSpace(text)
+	if len(text) <= limit {
+		return text
+	}
+	if limit <= 0 {
+		return ""
+	}
+	marker := "\n...[middle omitted]...\n"
+	if limit <= len(marker) {
+		return clipBytes(text, limit)
+	}
+	available := limit - len(marker)
+	head := clipPrefixAtBoundary(text, available/2)
+	tail := clipSuffixAtBoundary(text, available-available/2)
+	return strings.TrimSpace(head) + marker + strings.TrimSpace(tail)
+}
+
+func clipPrefixAtBoundary(text string, limit int) string {
+	if limit <= 0 {
+		return ""
+	}
+	if len(text) <= limit {
+		return text
+	}
+	end := limit
+	for end > 0 && text[end]&0xc0 == 0x80 {
+		end--
+	}
+	if boundary := strings.LastIndex(text[:end], "\n\n["); boundary > 0 {
+		end = boundary
+	}
+	return text[:end]
+}
+
+func clipSuffixAtBoundary(text string, limit int) string {
+	if limit <= 0 {
+		return ""
+	}
+	if len(text) <= limit {
+		return text
+	}
+	start := len(text) - limit
+	for start < len(text) && text[start]&0xc0 == 0x80 {
+		start++
+	}
+	if boundary := strings.Index(text[start:], "\n\n["); boundary >= 0 {
+		start += boundary + 2
+	}
+	return text[start:]
+}
+
+func askUserQuestionsDetail(arguments map[string]any) string {
+	raw, ok := arguments["questions"].([]any)
+	if !ok {
+		return ""
+	}
+	questions := make([]string, 0, len(raw))
+	for _, item := range raw {
+		object, ok := item.(map[string]any)
+		if !ok {
+			continue
+		}
+		if question, ok := object["question"].(string); ok && strings.TrimSpace(question) != "" {
+			questions = append(questions, strings.TrimSpace(question))
+		}
+	}
+	return strings.Join(questions, " | ")
+}
+
+func patchPaths(patch string) []string {
+	var paths []string
+	seen := map[string]bool{}
+	for _, line := range strings.Split(patch, "\n") {
+		for _, prefix := range []string{"*** Add File: ", "*** Update File: ", "*** Delete File: ", "*** Move to: "} {
+			if !strings.HasPrefix(line, prefix) {
+				continue
+			}
+			path := strings.TrimSpace(strings.TrimPrefix(line, prefix))
+			if path != "" && !seen[path] {
+				seen[path] = true
+				paths = append(paths, path)
+			}
+		}
+	}
+	return paths
+}
+
+var compactionStopWords = map[string]bool{
+	"a": true, "an": true, "the": true, "is": true, "are": true, "was": true, "were": true,
+	"be": true, "been": true, "being": true, "have": true, "has": true, "had": true,
+	"do": true, "does": true, "did": true, "will": true, "would": true, "could": true,
+	"should": true, "may": true, "might": true, "shall": true, "can": true, "need": true, "must": true,
+	"to": true, "of": true, "in": true, "for": true, "on": true, "with": true, "at": true,
+	"by": true, "from": true, "as": true, "into": true, "through": true, "during": true,
+	"before": true, "after": true, "above": true, "below": true, "between": true, "under": true, "over": true,
+	"and": true, "but": true, "or": true, "nor": true, "not": true, "so": true, "yet": true,
+	"both": true, "either": true, "neither": true, "each": true, "every": true, "all": true,
+	"any": true, "few": true, "more": true, "most": true, "other": true, "some": true, "such": true, "no": true,
+	"that": true, "this": true, "these": true, "those": true, "it": true, "its": true,
+	"i": true, "me": true, "my": true, "we": true, "our": true, "you": true, "your": true,
+	"he": true, "him": true, "his": true, "she": true, "her": true, "they": true, "them": true, "their": true,
+	"who": true, "which": true, "what": true, "if": true, "then": true, "than": true,
+	"when": true, "where": true, "how": true, "just": true, "also": true,
+}

internal/agent/compaction_projection_test.go

@@ -0,0 +1,130 @@
+package agent
+
+import (
+	"strings"
+	"testing"
+	"unicode/utf8"
+
+	"github.com/Gitlawb/zero/internal/zeroruntime"
+)
+
+func TestClipBytesHandlesTinyLimitsOnRuneBoundaries(t *testing.T) {
+	for _, limit := range []int{0, 1, 2, 3} {
+		if clipped := clipBytes("世界", limit); len(clipped) > limit || !utf8.ValidString(clipped) {
+			t.Fatalf("clipBytes limit %d returned invalid result %q", limit, clipped)
+		}
+	}
+}
+
+func TestBoundaryClippersHandleNonPositiveLimits(t *testing.T) {
+	for _, clip := range []func(string, int) string{clipPrefixAtBoundary, clipSuffixAtBoundary} {
+		for _, limit := range []int{-1, 0} {
+			if got := clip("世界", limit); got != "" {
+				t.Fatalf("clipper limit %d = %q, want empty", limit, got)
+			}
+		}
+	}
+}
+
+func TestProjectCompactionInputDropsRecoverableToolBodies(t *testing.T) {
+	largeBody := strings.Repeat("recoverable source line\n", 1000)
+	messages := []zeroruntime.Message{
+		{Role: zeroruntime.MessageRoleUser, Content: "Inspect the parser and preserve its current behavior."},
+		{Role: zeroruntime.MessageRoleAssistant, Content: "I will inspect before editing.", ToolCalls: []zeroruntime.ToolCall{{ID: "read", Name: "read_file", Arguments: `{"path":"internal/parser.go"}`}}},
+		{Role: zeroruntime.MessageRoleTool, ToolCallID: "read", Content: largeBody},
+		{Role: zeroruntime.MessageRoleAssistant, ToolCalls: []zeroruntime.ToolCall{{ID: "test", Name: "exec_command", Arguments: `{"cmd":"go test ./internal/parser"}`}}},
+		{Role: zeroruntime.MessageRoleTool, ToolCallID: "test", Content: "Error: parser regression\nlong diagnostic details"},
+	}
+
+	projection := projectCompactionInput(messages)
+	projected := projection.messages
+	if len(projected) != 1 {
+		t.Fatalf("projection = %#v, want one compact message", projected)
+	}
+	content := projected[0].Content
+	for _, want := range []string{"Inspect the parser", `read_file "internal/parser.go"`, `exec_command "go test ./internal/parser"`, "Error: parser regression"} {
+		if !strings.Contains(content, want) {
+			t.Fatalf("projection missing %q: %q", want, content)
+		}
+	}
+	if strings.Contains(content, "recoverable source line") || strings.Contains(content, "long diagnostic details") {
+		t.Fatalf("projection retained reconstructible or overlong tool output: %q", content)
+	}
+	if estimateTokens(projected)*20 >= estimateTokens(messages) {
+		t.Fatalf("projection did not materially reduce tokens: before=%d after=%d", estimateTokens(messages), estimateTokens(projected))
+	}
+}
+
+func TestProjectCompactionInputRetainsStructurallyMarkedToolErrors(t *testing.T) {
+	messages := []zeroruntime.Message{
+		{Role: zeroruntime.MessageRoleAssistant, ToolCalls: []zeroruntime.ToolCall{{ID: "git", Name: "exec_command", Arguments: `{"cmd":"git status"}`}}},
+		{Role: zeroruntime.MessageRoleTool, ToolCallID: "git", Content: "fatal: not a git repository", IsError: true},
+	}
+
+	projected := projectCompactionInput(messages).messages
+	if len(projected) != 1 || !strings.Contains(projected[0].Content, "fatal: not a git repository") {
+		t.Fatalf("structured tool error was dropped from projection: %#v", projected)
+	}
+}
+
+func TestProjectCompactionInputRetainsAskUserExchangeAndSemanticArguments(t *testing.T) {
+	messages := []zeroruntime.Message{
+		{Role: zeroruntime.MessageRoleAssistant, ToolCalls: []zeroruntime.ToolCall{{ID: "ask", Name: "ask_user", Arguments: `{"questions":[{"question":"Which database should remain?"}]}`}}},
+		{Role: zeroruntime.MessageRoleTool, ToolCallID: "ask", Content: "Which database should remain?: Postgres only; do not touch MySQL."},
+		{Role: zeroruntime.MessageRoleAssistant, ToolCalls: []zeroruntime.ToolCall{
+			{ID: "patch", Name: "apply_patch", Arguments: `{"patch":"*** Begin Patch\n*** Update File: internal/db.go\n*** End Patch"}`},
+			{ID: "fetch", Name: "web_fetch", Arguments: `{"url":"https://example.com/spec"}`},
+			{ID: "task", Name: "task", Arguments: `{"prompt":"Inspect the Windows implementation"}`},
+		}},
+		{Role: zeroruntime.MessageRoleTool, ToolCallID: "patch", Content: "Done!", ChangedFiles: []string{"internal/db.go"}},
+	}
+
+	content := projectCompactionInput(messages).messages[0].Content
+	for _, want := range []string{"Which database should remain?", "Postgres only; do not touch MySQL", "internal/db.go", "https://example.com/spec", "Inspect the Windows implementation"} {
+		if !strings.Contains(content, want) {
+			t.Fatalf("projection missing %q: %q", want, content)
+		}
+	}
+}
+
+func TestProjectCompactionInputCapsToolCallsPerTurn(t *testing.T) {
+	calls := make([]zeroruntime.ToolCall, 12)
+	for index := range calls {
+		calls[index] = zeroruntime.ToolCall{Name: "read_file", Arguments: `{"path":"file.go"}`}
+	}
+	projected := projectCompactionInput([]zeroruntime.Message{{Role: zeroruntime.MessageRoleAssistant, ToolCalls: calls}}).messages
+	content := projected[0].Content
+	if strings.Count(content, "* read_file") != compactionToolCallsPerTurn || !strings.Contains(content, "4 earlier tool calls omitted") {
+		t.Fatalf("tool-call tail was not bounded: %q", content)
+	}
+}
+
+func TestProjectCompactionInputCarriesPreviousSummaryOutsideBriefCap(t *testing.T) {
+	previous := "keep the earlier architecture decision\n" + strings.Repeat("prior detail ", 2000) + "\nkeep the newest prior decision"
+	messages := []zeroruntime.Message{{Role: zeroruntime.MessageRoleUser, Content: summaryLabel + "\n" + previous + "\n\n" + preservedStateLabel + "\n{}"}}
+	for index := range 500 {
+		messages = append(messages, zeroruntime.Message{Role: zeroruntime.MessageRoleAssistant, Content: strings.Repeat("later transcript detail ", 20)})
+		messages = append(messages, zeroruntime.Message{Role: zeroruntime.MessageRoleUser, Content: "later request " + strings.Repeat("context ", 20) + string(rune('a'+index%26))})
+	}
+
+	projection := projectCompactionInput(messages)
+	content := projection.messages[0].Content
+	if !strings.Contains(content, "[previous summary]\nkeep the earlier architecture decision") || !strings.Contains(content, "keep the newest prior decision") {
+		t.Fatalf("previous summary was lost when the brief tail was capped: %q", content)
+	}
+	if !strings.Contains(content, "middle transcript omitted to fit compaction budget") {
+		t.Fatalf("oversized brief was not head/tail capped: %q", content)
+	}
+	if !projection.truncated {
+		t.Fatal("projection did not structurally report truncation")
+	}
+}
+
+func TestProjectCompactionInputDoesNotInferTruncationFromUserText(t *testing.T) {
+	projection := projectCompactionInput([]zeroruntime.Message{{
+		Role: zeroruntime.MessageRoleUser, Content: "Keep this literal marker: [middle omitted]",
+	}})
+	if projection.truncated {
+		t.Fatal("literal user text produced a false truncation signal")
+	}
+}

internal/agent/compaction_recent_edits_test.go

@@ -0,0 +1,128 @@
+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"},
+		{Role: zeroruntime.MessageRoleAssistant, ToolCalls: []zeroruntime.ToolCall{
+			{ID: "e3", Name: "apply_patch", Arguments: `{"patch":"*** Update File: internal/baz.go"}`},
+		}},
+		{Role: zeroruntime.MessageRoleTool, ToolCallID: "e3", Content: "Done!", ChangedFiles: []string{"internal/baz.go"}},
+	}
+
+	edits := recentEdits(messages)
+	if len(edits) != 3 {
+		t.Fatalf("expected 3 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])
+	}
+	if edits[2].name != "internal/baz.go" || edits[2].body != "Done!" {
+		t.Fatalf("third edit = %#v, want apply_patch changed file with its note", edits[2])
+	}
+}
+
+// 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/compaction_test.go

@@ -7,6 +7,7 @@ import (
 	"testing"
 
 	"github.com/Gitlawb/zero/internal/tools"
+	"github.com/Gitlawb/zero/internal/trace"
 	"github.com/Gitlawb/zero/internal/zeroruntime"
 )
 
@@ -58,12 +59,15 @@ func TestCompactKeepsSystemAndPreservedSuffix(t *testing.T) {
 	if last.Content != "most recent question" {
 		t.Fatalf("expected most recent message preserved, got %q", last.Content)
 	}
-	// The summarized middle excludes system and preserved suffix.
-	if len(captured) != 3 {
-		t.Fatalf("expected 3 summarized messages, got %d: %#v", len(captured), captured)
+	// The summarizer receives a compact semantic projection of the middle rather
+	// than reconstructible raw tool output.
+	if len(captured) != 1 || captured[0].Role != zeroruntime.MessageRoleUser {
+		t.Fatalf("expected one projected summary input, got %#v", captured)
 	}
-	if captured[0].Content != "first question" {
-		t.Fatalf("expected oldest non-system message first, got %#v", captured[0])
+	for _, want := range []string{"[user #0]", "first question", "[assistant #1]", "first answer", "second question"} {
+		if !strings.Contains(captured[0].Content, want) {
+			t.Fatalf("projected summary input missing %q: %q", want, captured[0].Content)
+		}
 	}
 	// Compaction must shrink the conversation.
 	if estimateTokens(result) >= estimateTokens(messages) {
@@ -336,7 +340,7 @@ func TestRunProactiveCompactionTriggers(t *testing.T) {
 	}
 
 	registry := tools.NewRegistry()
-	registry.Register(tools.NewReadFileTool(t.TempDir()))
+	registry.Register(tools.NewScopedReadFileTool(t.TempDir(), nil))
 
 	result, err := Run(context.Background(), strings.Repeat("y", 8000), provider, Options{
 		Registry:               registry,
@@ -364,7 +368,7 @@ func TestRunNoCompactionWhenContextWindowZero(t *testing.T) {
 		},
 	}
 	registry := tools.NewRegistry()
-	registry.Register(tools.NewReadFileTool(t.TempDir()))
+	registry.Register(tools.NewScopedReadFileTool(t.TempDir(), nil))
 
 	_, err := Run(context.Background(), strings.Repeat("y", 8000), provider, Options{
 		Registry:       registry,
@@ -379,6 +383,43 @@ func TestRunNoCompactionWhenContextWindowZero(t *testing.T) {
 	}
 }
 
+func TestMaybeCompactStopsAfterSupersededReadsReachThreshold(t *testing.T) {
+	body := strings.Repeat("same source line\n", 1000)
+	messages := []zeroruntime.Message{
+		{Role: zeroruntime.MessageRoleSystem, Content: "system"},
+		toolCallWithArgs("old", "read_file", `{"path":"a.go"}`),
+		{Role: zeroruntime.MessageRoleTool, ToolCallID: "old", Content: body},
+		toolCallWithArgs("new", "read_file", `{"path":"a.go"}`),
+		{Role: zeroruntime.MessageRoleTool, ToolCallID: "new", Content: body},
+	}
+	pruned, reclaimed := pruneSupersededReadResults(messages)
+	if reclaimed <= 0 {
+		t.Fatal("test setup did not reclaim superseded read output")
+	}
+	prunedSize := estimateTokens(pruned)
+	if estimateTokens(messages) <= prunedSize {
+		t.Fatal("test setup does not cross the compaction threshold")
+	}
+
+	provider := &summarizeRecordingProvider{}
+	state := &compactionState{
+		enabled:      true,
+		threshold:    prunedSize,
+		preserveLast: 2,
+	}
+	got := state.maybeCompact(context.Background(), provider, messages, nil)
+
+	if provider.summarizeCalls != 0 {
+		t.Fatalf("cheap pruning should avoid the summarizer, got %d calls", provider.summarizeCalls)
+	}
+	if state.lowWaterMark != prunedSize {
+		t.Fatalf("lowWaterMark = %d, want %d", state.lowWaterMark, prunedSize)
+	}
+	if !strings.Contains(got[2].Content, "superseded identical") || got[4].Content != body {
+		t.Fatalf("unexpected pruned messages: %#v", got)
+	}
+}
+
 // reactiveProvider builds up history with a tool call on turn 1, then errors
 // with a context-limit message on turn 2 (once the history is large enough that
 // compaction can shrink it), then succeeds on the same-turn retry. Summary
@@ -388,6 +429,7 @@ type reactiveProvider struct {
 	summarizeCalls int
 	turnRequests   int
 	failedOnce     bool
+	connectError   bool
 	bigText        string
 	finalText      string
 }
@@ -409,6 +451,9 @@ func (provider *reactiveProvider) StreamCompletion(ctx context.Context, request
 		return streamEvents(toolTurnWithText(provider.bigText, "1", "read_file", `{"path":"x"}`)), nil
 	case provider.turnRequests == 2 && !provider.failedOnce:
 		provider.failedOnce = true
+		if provider.connectError {
+			return nil, errors.New("context_length_exceeded")
+		}
 		return streamEvents([]zeroruntime.StreamEvent{
 			{Type: zeroruntime.StreamEventError, Error: "This model's maximum context length is 1000 tokens. Please reduce the length of the messages."},
 		}), nil
@@ -417,6 +462,39 @@ func (provider *reactiveProvider) StreamCompletion(ctx context.Context, request
 	}
 }
 
+func TestRunConnectErrorCompactionRecordsReplacementPlan(t *testing.T) {
+	provider := &reactiveProvider{
+		connectError: true,
+		bigText:      strings.Repeat("b", 6000),
+		finalText:    "recovered",
+	}
+	registry := tools.NewRegistry()
+	registry.Register(tools.NewScopedReadFileTool(t.TempDir(), nil))
+	recorder := trace.NewRecorder("connect-compaction", "run-1", "test")
+	var contextPlans []ContextBreakdown
+
+	result, err := Run(context.Background(), strings.Repeat("z", 6000), provider, Options{
+		Registry:               registry,
+		PermissionMode:         PermissionModeUnsafe,
+		ContextWindow:          10_000_000,
+		CompactionPreserveLast: 2,
+		Trace:                  recorder,
+		OnContext: func(breakdown ContextBreakdown) {
+			contextPlans = append(contextPlans, breakdown)
+		},
+	})
+	if err != nil || result.FinalAnswer != "recovered" {
+		t.Fatalf("connect-error compaction failed: result=%#v err=%v", result, err)
+	}
+	if len(contextPlans) != 3 || contextPlans[2].MessageTokens >= contextPlans[1].MessageTokens {
+		t.Fatalf("replacement request context not reported: %#v", contextPlans)
+	}
+	gotTrace := recorder.Finish()
+	if len(gotTrace.PrefixHashes) != 4 || gotTrace.PrefixHashes[3].InvalidationReason != "unchanged" {
+		t.Fatalf("replacement request trace not recorded: %#v", gotTrace.PrefixHashes)
+	}
+}
+
 func TestRunReactiveCompactionRecovers(t *testing.T) {
 	provider := &reactiveProvider{
 		bigText:   strings.Repeat("b", 6000),
@@ -427,7 +505,9 @@ func TestRunReactiveCompactionRecovers(t *testing.T) {
 	// the compaction closure. read_file also returns a sizeable result that
 	// bloats the history.
 	registry := tools.NewRegistry()
-	registry.Register(tools.NewReadFileTool(t.TempDir()))
+	registry.Register(tools.NewScopedReadFileTool(t.TempDir(), nil))
+	recorder := trace.NewRecorder("reactive-compaction", "run-1", "test")
+	var contextPlans []ContextBreakdown
 
 	// ContextWindow large enough that proactive compaction never triggers, so
 	// only the reactive path can save the run.
@@ -436,6 +516,10 @@ func TestRunReactiveCompactionRecovers(t *testing.T) {
 		PermissionMode:         PermissionModeUnsafe,
 		ContextWindow:          10_000_000,
 		CompactionPreserveLast: 2,
+		Trace:                  recorder,
+		OnContext: func(breakdown ContextBreakdown) {
+			contextPlans = append(contextPlans, breakdown)
+		},
 	})
 	if err != nil {
 		t.Fatalf("expected reactive compaction to recover the run, got error: %v", err)
@@ -446,6 +530,25 @@ func TestRunReactiveCompactionRecovers(t *testing.T) {
 	if provider.summarizeCalls == 0 {
 		t.Fatal("expected reactive compaction to invoke the summarizer")
 	}
+	if len(contextPlans) != 3 {
+		t.Fatalf("main request context callbacks = %d, want initial turns plus compacted retry", len(contextPlans))
+	}
+	if contextPlans[2].MessageTokens >= contextPlans[1].MessageTokens {
+		t.Fatalf("retry context was not updated after compaction: before=%d after=%d", contextPlans[1].MessageTokens, contextPlans[2].MessageTokens)
+	}
+	gotTrace := recorder.Finish()
+	if len(gotTrace.PrefixHashes) != 4 {
+		t.Fatalf("prefix evidence count = %d, want two turns, compaction, and retry: %#v", len(gotTrace.PrefixHashes), gotTrace.PrefixHashes)
+	}
+	if gotTrace.PrefixHashes[0].InvalidationReason != "initial" ||
+		gotTrace.PrefixHashes[1].InvalidationReason != "unchanged" ||
+		gotTrace.PrefixHashes[2].InvalidationReason != "initial" ||
+		gotTrace.PrefixHashes[3].InvalidationReason != "unchanged" {
+		t.Fatalf("unexpected request-specific invalidation evidence: %#v", gotTrace.PrefixHashes)
+	}
+	if gotTrace.PrefixHashes[2].CompletePrefixHash == gotTrace.PrefixHashes[1].CompletePrefixHash {
+		t.Fatalf("compaction fingerprint reused the main request prefix: %#v", gotTrace.PrefixHashes)
+	}
 }
 
 // midStreamReactiveProvider forwards some text BEFORE surfacing a context-limit
@@ -491,7 +594,7 @@ func TestRunReactiveRetryDoesNotDoubleEmitText(t *testing.T) {
 		finalText:   "recovered",
 	}
 	registry := tools.NewRegistry()
-	registry.Register(tools.NewReadFileTool(t.TempDir()))
+	registry.Register(tools.NewScopedReadFileTool(t.TempDir(), nil))
 
 	var deltas []string
 	result, err := Run(context.Background(), strings.Repeat("z", 6000), provider, Options{
@@ -581,7 +684,7 @@ func TestCompactNeverProducesConsecutiveUserMessages(t *testing.T) {
 }
 
 func TestRecoverNoopDoesNotConsumeReactiveBudget(t *testing.T) {
-	st := newCompactionState(Options{ContextWindow: 1000, CompactionPreserveLast: 2})
+	st := newCompactionState(Options{ContextWindow: 1000, CompactionPreserveLast: 2}, nil)
 	provider := &mockProvider{turns: [][]zeroruntime.StreamEvent{{
 		{Type: zeroruntime.StreamEventText, Content: "SUMMARY"}, {Type: zeroruntime.StreamEventDone},
 	}}}
@@ -628,7 +731,7 @@ func TestRecoverNoopDoesNotConsumeReactiveBudget(t *testing.T) {
 }
 
 func TestRecoverDisabledIsNoop(t *testing.T) {
-	st := newCompactionState(Options{ContextWindow: 0})
+	st := newCompactionState(Options{ContextWindow: 0}, nil)
 	msgs := []zeroruntime.Message{{Role: zeroruntime.MessageRoleUser, Content: "x"}}
 	called := false
 	// recover must not invoke the provider/summarize when disabled, even on a

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, context completionContext) 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 || context.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 +1,75 @@
+package agent
+
+import (
+	"strings"
+	"testing"
+)
+
+func TestCompletionPolicyLocalEvidenceDecidesWithoutSemanticCheck(t *testing.T) {
+	policy := newCompletionPolicy(false)
+
+	complete := policy.evaluate("Done. All required checks pass.", completionContext{})
+	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.", completionContext{})
+	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:", completionContext{PlanPending: 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:", completionContext{PlanPending: 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.", completionContext{PlanPending: 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.", completionContext{PlanPending: 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.", completionContext{})
+	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.", completionContext{})
+	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")
+	}
+}
+
+func TestAcceptanceVerificationNudgeIncludesBoundedObjective(t *testing.T) {
+	objective := "make the requested behavior observable"
+	nudge := acceptanceVerificationNudge(objective)
+	if !strings.Contains(nudge, "Task objective: "+objective) {
+		t.Fatalf("semantic check was not grounded in the objective: %q", nudge)
+	}
+}

internal/agent/context_measurement.go

@@ -23,6 +23,24 @@ type ContextBreakdown struct {
 	TotalTokens   int     // SystemTokens + ToolTokens + MessageTokens
 	ContextWindow int     // model context window; 0 when unknown
 	UsedFraction  float64 // TotalTokens / ContextWindow; 0 when window unknown
+	// Blocks explain why each model-visible category is present without exposing
+	// its content. The planner currently preserves all three categories.
+	Blocks []ContextBlock
+	// CompletePrefixHash identifies the provider-visible system+tool prefix.
+	CompletePrefixHash string
+	// PrefixInvalidationReason is "initial", "unchanged", or a stable comma-
+	// separated list of changed prefix categories.
+	PrefixInvalidationReason string
+}
+
+// ContextBlock is content-free evidence for one model-visible context category.
+type ContextBlock struct {
+	Kind        string
+	Tokens      int
+	CacheClass  string
+	Authority   string
+	Recoverable bool
+	Reason      string
 }
 
 // MeasureContext estimates the per-category token footprint of a request: the
@@ -41,6 +59,24 @@ func MeasureContext(messages []zeroruntime.Message, tools []zeroruntime.ToolDefi
 		ContextWindow: contextWindow,
 	}
 	breakdown.TotalTokens = breakdown.SystemTokens + breakdown.ToolTokens + breakdown.MessageTokens
+	if breakdown.SystemTokens > 0 {
+		breakdown.Blocks = append(breakdown.Blocks, ContextBlock{
+			Kind: "system", Tokens: breakdown.SystemTokens, CacheClass: "stable_prefix",
+			Authority: "required", Reason: "instructions, workspace policy, and safety policy",
+		})
+	}
+	if breakdown.ToolTokens > 0 {
+		breakdown.Blocks = append(breakdown.Blocks, ContextBlock{
+			Kind: "tools", Tokens: breakdown.ToolTokens, CacheClass: "stable_prefix",
+			Authority: "capability", Reason: "currently permitted and selected tool definitions",
+		})
+	}
+	if breakdown.MessageTokens > 0 {
+		breakdown.Blocks = append(breakdown.Blocks, ContextBlock{
+			Kind: "conversation", Tokens: breakdown.MessageTokens, CacheClass: "dynamic_tail",
+			Authority: "transcript", Reason: "valid provider replay history and current task context",
+		})
+	}
 	if contextWindow > 0 {
 		breakdown.UsedFraction = float64(breakdown.TotalTokens) / float64(contextWindow)
 	}

internal/agent/context_measurement_test.go

@@ -32,6 +32,9 @@ func TestMeasureContextSplitsByCategory(t *testing.T) {
 	if breakdown.UsedFraction != wantFraction {
 		t.Fatalf("UsedFraction = %v, want %v", breakdown.UsedFraction, wantFraction)
 	}
+	if len(breakdown.Blocks) != 3 {
+		t.Fatalf("Blocks = %#v, want system/tools/conversation", breakdown.Blocks)
+	}
 }
 
 func TestMeasureContextUnknownWindowHasZeroFraction(t *testing.T) {

internal/agent/context_planner.go

@@ -0,0 +1,185 @@
+package agent
+
+import (
+	"strings"
+
+	"github.com/Gitlawb/zero/internal/zeroruntime"
+)
+
+type contextPlannerConfig struct {
+	contextWindow  int
+	promptCacheKey string
+	serviceTier    string
+	promptParts    systemPromptParts
+}
+
+// contextPlanner is the single seam for constructing model-visible requests.
+// Planning is deterministic and does not retrieve content, execute tools, or
+// change permissions. The initial implementation deliberately preserves every
+// message and tool definition while making composition and cache drift
+// inspectable; later selection policies must continue to cross this seam.
+type contextPlanner struct {
+	config          contextPlannerConfig
+	previousPrefix  prefixFingerprint
+	hasPrevious     bool
+	toolSnapshotKey string
+	toolSnapshot    []zeroruntime.ToolDefinition
+}
+
+type contextPlan struct {
+	Request           zeroruntime.CompletionRequest
+	Breakdown         ContextBreakdown
+	PrefixFingerprint prefixFingerprint
+}
+
+func newContextPlanner(config contextPlannerConfig) *contextPlanner {
+	return &contextPlanner{config: config}
+}
+
+// Plan returns a provider request snapshot plus content-free accounting.
+// It intentionally performs no relevance filtering: preserving current model
+// capability is the baseline contract for future planner policies.
+func (planner *contextPlanner) Plan(messages []zeroruntime.Message, toolDefs []zeroruntime.ToolDefinition, reasoningEffort string) contextPlan {
+	return planner.plan(messages, toolDefs, reasoningEffort, planner.config.promptParts)
+}
+
+// planWithPromptParts plans a request whose stable system sections differ from
+// the planner's configured main-run prompt, such as a compaction summary call.
+func (planner *contextPlanner) planWithPromptParts(messages []zeroruntime.Message, toolDefs []zeroruntime.ToolDefinition, reasoningEffort string, parts systemPromptParts) contextPlan {
+	return planner.plan(messages, toolDefs, reasoningEffort, parts)
+}
+
+func (planner *contextPlanner) plan(messages []zeroruntime.Message, toolDefs []zeroruntime.ToolDefinition, reasoningEffort string, parts systemPromptParts) contextPlan {
+	requestMessages := copyMessages(messages)
+	requestSystemPrompt := leadingSystemContent(requestMessages)
+	if parts.prompt != requestSystemPrompt {
+		// Component boundaries from a different request would make the detailed
+		// invalidation reason misleading. Retain the exact request prefix and
+		// classify it conservatively as an unstructured system prompt instead.
+		parts = systemPromptParts{prompt: requestSystemPrompt}
+	}
+	fingerprint := computePrefixFingerprint(buildPromptSubstringsFromParts(parts, toolDefs))
+	request := zeroruntime.CompletionRequest{
+		Messages:        requestMessages,
+		Tools:           planner.snapshotTools(toolDefs, fingerprint),
+		ReasoningEffort: reasoningEffort,
+		ServiceTier:     planner.config.serviceTier,
+		PromptCacheKey:  planner.config.promptCacheKey,
+	}
+	breakdown := MeasureContext(request.Messages, request.Tools, planner.config.contextWindow)
+	breakdown.CompletePrefixHash = fingerprint.CompletePrefixHash
+	var previous *prefixFingerprint
+	if planner.hasPrevious {
+		previous = &planner.previousPrefix
+	}
+	breakdown.PrefixInvalidationReason = explainPrefixChange(previous, fingerprint)
+	planner.previousPrefix = fingerprint
+	planner.hasPrevious = true
+	return contextPlan{
+		Request:           request,
+		Breakdown:         breakdown,
+		PrefixFingerprint: fingerprint,
+	}
+}
+
+// snapshotTools freezes provider-visible tool definitions when their semantic
+// fingerprint changes. Stable turns reuse the frozen schemas instead of
+// recursively cloning every map and slice for every request.
+func (planner *contextPlanner) snapshotTools(toolDefs []zeroruntime.ToolDefinition, fingerprint prefixFingerprint) []zeroruntime.ToolDefinition {
+	key := fingerprint.ToolsHash + "\x00" + fingerprint.SchemaHash
+	if planner.toolSnapshotKey != key {
+		planner.toolSnapshot = copyToolDefinitions(toolDefs)
+		planner.toolSnapshotKey = key
+	}
+	return planner.toolSnapshot
+}
+
+func copyToolDefinitions(toolDefs []zeroruntime.ToolDefinition) []zeroruntime.ToolDefinition {
+	if toolDefs == nil {
+		return nil
+	}
+	copied := make([]zeroruntime.ToolDefinition, len(toolDefs))
+	for index, definition := range toolDefs {
+		copied[index] = definition
+		copied[index].Parameters = copySchemaMap(definition.Parameters)
+	}
+	return copied
+}
+
+func copySchemaMap(schema map[string]any) map[string]any {
+	if schema == nil {
+		return nil
+	}
+	copied := make(map[string]any, len(schema))
+	for key, value := range schema {
+		copied[key] = copySchemaValue(value)
+	}
+	return copied
+}
+
+func copySchemaValue(value any) any {
+	switch typed := value.(type) {
+	case map[string]any:
+		return copySchemaMap(typed)
+	case []any:
+		copied := make([]any, len(typed))
+		for index, item := range typed {
+			copied[index] = copySchemaValue(item)
+		}
+		return copied
+	case []string:
+		return append([]string(nil), typed...)
+	default:
+		return value
+	}
+}
+
+func leadingSystemContent(messages []zeroruntime.Message) string {
+	contents := make([]string, 0, 1)
+	for _, message := range messages {
+		if message.Role != zeroruntime.MessageRoleSystem {
+			break
+		}
+		contents = append(contents, message.Content)
+	}
+	return strings.Join(contents, "\n\n")
+}
+
+func explainPrefixChange(previous *prefixFingerprint, current prefixFingerprint) string {
+	if previous == nil {
+		return "initial"
+	}
+	if previous.CompletePrefixHash == current.CompletePrefixHash {
+		return "unchanged"
+	}
+	reasons := make([]string, 0, 3)
+	if previous.SystemPromptHash != current.SystemPromptHash {
+		parts := make([]string, 0, 4)
+		if previous.BaseInstructionsHash != current.BaseInstructionsHash {
+			parts = append(parts, "base_instructions")
+		}
+		if previous.ConfirmationPolicyHash != current.ConfirmationPolicyHash {
+			parts = append(parts, "confirmation_policy")
+		}
+		if previous.ProjectContextHash != current.ProjectContextHash {
+			parts = append(parts, "project_context")
+		}
+		if previous.SkillsHash != current.SkillsHash {
+			parts = append(parts, "skills")
+		}
+		if len(parts) == 0 {
+			parts = append(parts, "system_prompt")
+		}
+		reasons = append(reasons, parts...)
+	}
+	if previous.ToolsHash != current.ToolsHash {
+		reasons = append(reasons, "tools")
+	}
+	if previous.SchemaHash != current.SchemaHash {
+		reasons = append(reasons, "schema")
+	}
+	if len(reasons) == 0 {
+		return "prefix_changed"
+	}
+	return strings.Join(reasons, ",")
+}

internal/agent/context_planner_test.go

@@ -0,0 +1,187 @@
+package agent
+
+import (
+	"reflect"
+	"testing"
+
+	"github.com/Gitlawb/zero/internal/zeroruntime"
+)
+
+func TestContextPlannerPreservesProviderRequest(t *testing.T) {
+	messages := []zeroruntime.Message{
+		{Role: zeroruntime.MessageRoleSystem, Content: "system"},
+		{Role: zeroruntime.MessageRoleUser, Content: "inspect this", Images: []zeroruntime.ImageBlock{{MediaType: "image/png", Data: []byte{1, 2, 3}}}},
+		{Role: zeroruntime.MessageRoleAssistant, Content: "working", ToolCalls: []zeroruntime.ToolCall{{ID: "call-1", Name: "read_file", Arguments: `{"path":"main.go"}`}}, Reasoning: []zeroruntime.ReasoningBlock{{Provider: "test", Type: "thinking", Signature: "sig"}}},
+	}
+	toolDefs := []zeroruntime.ToolDefinition{{
+		Name: "read_file", Description: "Read a file",
+		Parameters: map[string]any{
+			"type":       "object",
+			"required":   []string{"path"},
+			"properties": map[string]any{"path": map[string]any{"type": "string"}},
+			"anyOf":      []any{map[string]any{"type": "object"}},
+		},
+	}}
+	planner := newContextPlanner(contextPlannerConfig{
+		contextWindow:  128_000,
+		promptCacheKey: "session-1",
+		serviceTier:    "priority",
+		promptParts:    systemPromptParts{prompt: "system", baseInstructions: "system"},
+	})
+
+	plan := planner.Plan(messages, toolDefs, "medium")
+	want := zeroruntime.CompletionRequest{
+		Messages:        copyMessages(messages),
+		Tools:           toolDefs,
+		ReasoningEffort: "medium",
+		ServiceTier:     "priority",
+		PromptCacheKey:  "session-1",
+	}
+	if !reflect.DeepEqual(plan.Request, want) {
+		t.Fatalf("planned request changed provider input:\n got: %#v\nwant: %#v", plan.Request, want)
+	}
+	stable := planner.Plan(messages, toolDefs, "medium")
+	if &stable.Request.Tools[0] != &plan.Request.Tools[0] {
+		t.Fatal("stable tool fingerprint rebuilt the immutable schema snapshot")
+	}
+
+	// A plan owns its request snapshot; later caller mutation cannot change its
+	// messages or its recursively nested tool schema.
+	messages[1].Content = "changed"
+	messages[1].Images[0].Data[0] = 9
+	toolDefs[0].Name = "changed_tool"
+	toolDefs[0].Parameters["required"].([]string)[0] = "other"
+	toolDefs[0].Parameters["properties"].(map[string]any)["path"].(map[string]any)["type"] = "integer"
+	toolDefs[0].Parameters["anyOf"].([]any)[0].(map[string]any)["type"] = "array"
+	if plan.Request.Messages[1].Content != "inspect this" || plan.Request.Messages[1].Images[0].Data[0] != 1 {
+		t.Fatalf("planned messages alias caller state: %#v", plan.Request.Messages[1])
+	}
+	if plan.Request.Tools[0].Name != "read_file" ||
+		plan.Request.Tools[0].Parameters["required"].([]string)[0] != "path" ||
+		plan.Request.Tools[0].Parameters["properties"].(map[string]any)["path"].(map[string]any)["type"] != "string" ||
+		plan.Request.Tools[0].Parameters["anyOf"].([]any)[0].(map[string]any)["type"] != "object" {
+		t.Fatalf("planned tools alias caller state: %#v", plan.Request.Tools[0])
+	}
+
+	next := planner.Plan(messages, toolDefs, "medium")
+	if next.PrefixFingerprint == plan.PrefixFingerprint || next.Request.Tools[0].Name != "changed_tool" {
+		t.Fatalf("changed tool schema did not refresh snapshot: %#v", next)
+	}
+	if plan.Request.Tools[0].Name != "read_file" {
+		t.Fatalf("refresh mutated earlier request: %#v", plan.Request.Tools[0])
+	}
+}
+
+func TestContextPlannerReportsInspectableBlocks(t *testing.T) {
+	planner := newContextPlanner(contextPlannerConfig{contextWindow: 100_000})
+	plan := planner.Plan([]zeroruntime.Message{
+		{Role: zeroruntime.MessageRoleSystem, Content: filler(400)},
+		{Role: zeroruntime.MessageRoleUser, Content: filler(200)},
+	}, []zeroruntime.ToolDefinition{{Name: "read_file", Description: "Read a file"}}, "")
+
+	if plan.Breakdown.PrefixInvalidationReason != "initial" || plan.Breakdown.CompletePrefixHash == "" {
+		t.Fatalf("prefix evidence = %#v", plan.Breakdown)
+	}
+	if len(plan.Breakdown.Blocks) != 3 {
+		t.Fatalf("blocks = %#v, want system/tools/conversation", plan.Breakdown.Blocks)
+	}
+	wantKinds := []string{"system", "tools", "conversation"}
+	total := 0
+	for index, block := range plan.Breakdown.Blocks {
+		if block.Kind != wantKinds[index] || block.Tokens <= 0 || block.Reason == "" || block.Authority == "" || block.CacheClass == "" {
+			t.Fatalf("block %d = %#v", index, block)
+		}
+		total += block.Tokens
+	}
+	if total != plan.Breakdown.TotalTokens {
+		t.Fatalf("block tokens = %d, want total %d", total, plan.Breakdown.TotalTokens)
+	}
+}
+
+func TestContextPlannerFingerprintsActualRequestSystemPrompt(t *testing.T) {
+	planner := newContextPlanner(contextPlannerConfig{promptParts: systemPromptParts{
+		prompt: "configured system", baseInstructions: "configured system",
+	}})
+	messages := []zeroruntime.Message{
+		{Role: zeroruntime.MessageRoleSystem, Content: "request-specific system"},
+		{Role: zeroruntime.MessageRoleUser, Content: "task"},
+	}
+
+	plan := planner.Plan(messages, nil, "")
+	want := computePrefixFingerprint(buildPromptSubstringsFromParts(
+		systemPromptParts{prompt: "request-specific system"}, nil,
+	))
+	if plan.PrefixFingerprint != want {
+		t.Fatalf("fingerprint describes configured prompt instead of request:\n got: %#v\nwant: %#v", plan.PrefixFingerprint, want)
+	}
+}
+
+func TestContextPlannerExplainsPrefixInvalidation(t *testing.T) {
+	planner := newContextPlanner(contextPlannerConfig{promptParts: systemPromptParts{
+		prompt: "system", baseInstructions: "base", projectContext: "project",
+	}})
+	messages := []zeroruntime.Message{{Role: zeroruntime.MessageRoleSystem, Content: "system"}, {Role: zeroruntime.MessageRoleUser, Content: "task"}}
+	tools := []zeroruntime.ToolDefinition{{Name: "read_file", Description: "Read", Parameters: map[string]any{"type": "object"}}}
+
+	if got := planner.Plan(messages, tools, "").Breakdown.PrefixInvalidationReason; got != "initial" {
+		t.Fatalf("first reason = %q, want initial", got)
+	}
+	if got := planner.Plan(messages, tools, "").Breakdown.PrefixInvalidationReason; got != "unchanged" {
+		t.Fatalf("stable reason = %q, want unchanged", got)
+	}
+	tools[0].Description = "Read an exact file"
+	if got := planner.Plan(messages, tools, "").Breakdown.PrefixInvalidationReason; got != "tools" {
+		t.Fatalf("tool reason = %q, want tools", got)
+	}
+	tools[0].Parameters["required"] = []any{"path"}
+	if got := planner.Plan(messages, tools, "").Breakdown.PrefixInvalidationReason; got != "schema" {
+		t.Fatalf("schema reason = %q, want schema", got)
+	}
+	planner.config.promptParts.prompt = "system with changed project"
+	planner.config.promptParts.projectContext = "changed project"
+	messages[0].Content = "system with changed project"
+	if got := planner.Plan(messages, tools, "").Breakdown.PrefixInvalidationReason; got != "project_context" {
+		t.Fatalf("project reason = %q, want project_context", got)
+	}
+}
+
+func TestExplainPrefixChangeNamesEveryChangedComponent(t *testing.T) {
+	baseline := prefixFingerprint{
+		SystemPromptHash:       "system",
+		BaseInstructionsHash:   "base",
+		ConfirmationPolicyHash: "policy",
+		ProjectContextHash:     "project",
+		SkillsHash:             "skills",
+		ToolsHash:              "tools",
+		SchemaHash:             "schema",
+		CompletePrefixHash:     "complete",
+	}
+	tests := []struct {
+		name   string
+		mutate func(*prefixFingerprint)
+		want   string
+	}{
+		{name: "base instructions", mutate: func(value *prefixFingerprint) {
+			value.SystemPromptHash, value.BaseInstructionsHash = "changed", "changed"
+		}, want: "base_instructions"},
+		{name: "confirmation policy", mutate: func(value *prefixFingerprint) {
+			value.SystemPromptHash, value.ConfirmationPolicyHash = "changed", "changed"
+		}, want: "confirmation_policy"},
+		{name: "project context", mutate: func(value *prefixFingerprint) {
+			value.SystemPromptHash, value.ProjectContextHash = "changed", "changed"
+		}, want: "project_context"},
+		{name: "skills", mutate: func(value *prefixFingerprint) { value.SystemPromptHash, value.SkillsHash = "changed", "changed" }, want: "skills"},
+		{name: "unclassified system prompt", mutate: func(value *prefixFingerprint) { value.SystemPromptHash = "changed" }, want: "system_prompt"},
+		{name: "tools and schema", mutate: func(value *prefixFingerprint) { value.ToolsHash, value.SchemaHash = "changed", "changed" }, want: "tools,schema"},
+	}
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			current := baseline
+			test.mutate(&current)
+			current.CompletePrefixHash = "changed"
+			if got := explainPrefixChange(&baseline, current); got != test.want {
+				t.Fatalf("reason = %q, want %q", got, test.want)
+			}
+		})
+	}
+}

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

@@ -150,7 +150,7 @@ func assertNoDeferredDiscoveryMessage(t *testing.T, request zeroruntime.Completi
 func TestPartitionToolsInactiveIsByteIdenticalAndDropsToolSearch(t *testing.T) {
 	root := t.TempDir()
 	registry := tools.NewRegistry()
-	registry.Register(tools.NewReadFileTool(root))
+	registry.Register(tools.NewScopedReadFileTool(root, nil))
 	registry.Register(fakeDeferredTool{name: "mcp__srv__a", desc: "tool a"})
 	registry.Register(fakeToolSearchTool{})
 
@@ -305,8 +305,8 @@ func TestPartitionToolsActiveExcludesDisabledDeferredFromDiscoveryAndExposed(t *
 func TestPartitionToolsActiveHidesUnloadedExposesLoaded(t *testing.T) {
 	root := t.TempDir()
 	registry := tools.NewRegistry()
-	registry.Register(tools.NewReadFileTool(root)) // non-deferred builtin
-	registry.Register(fakeToolSearchTool{})        // non-deferred, must stay exposed
+	registry.Register(tools.NewScopedReadFileTool(root, nil)) // non-deferred builtin
+	registry.Register(fakeToolSearchTool{})                   // non-deferred, must stay exposed
 	registry.Register(fakeDeferredTool{name: "mcp__srv__alpha", desc: "alpha tool"})
 	registry.Register(fakeDeferredTool{name: "mcp__srv__beta", desc: "beta tool"})
 
@@ -341,15 +341,15 @@ func TestPartitionToolsActiveHidesUnloadedExposesLoaded(t *testing.T) {
 	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__alpha") {
+		t.Fatalf("stable tool_search catalog must retain 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)
 	}
 }
 
-func TestPartitionToolsActiveNothingHiddenEmptyReminder(t *testing.T) {
+func TestPartitionToolsActiveNothingHiddenKeepsStableCatalog(t *testing.T) {
 	registry := tools.NewRegistry()
 	registry.Register(fakeDeferredTool{name: "mcp__srv__alpha", desc: "alpha"})
 	registry.Register(fakeDeferredTool{name: "mcp__srv__beta", desc: "beta"})
@@ -368,9 +368,8 @@ func TestPartitionToolsActiveNothingHiddenEmptyReminder(t *testing.T) {
 	if !exposedNames["tool_search"] {
 		t.Fatalf("expected tool_search exposed on active path, got %#v", exposed)
 	}
-	// No hidden tools means no dynamic discovery text is needed.
-	if reminder != "" {
-		t.Fatalf("expected empty discovery text when nothing is hidden, got %q", reminder)
+	if !strings.Contains(reminder, "mcp__srv__alpha") || !strings.Contains(reminder, "mcp__srv__beta") {
+		t.Fatalf("stable catalog must retain all deferred names after loading, got %q", reminder)
 	}
 }
 
@@ -798,3 +797,98 @@ func TestDisabledToolSearchFallsBackToEager(t *testing.T) {
 		t.Fatalf("deferred tool must be callable under eager fallback, got status=%s output=%q", result.Status, result.Output)
 	}
 }
+
+// fakeDeferredMutatorTool is a deferred-eligible tool with mutating Safety
+// (SideEffectWrite), standing in for a real write/mutator MCP tool that would
+// be hidden behind tool_search once deferral activates.
+type fakeDeferredMutatorTool struct{ name string }
+
+func (t fakeDeferredMutatorTool) Name() string        { return t.name }
+func (t fakeDeferredMutatorTool) Description() string { return "mutates the workspace, deferred" }
+func (t fakeDeferredMutatorTool) Parameters() tools.Schema {
+	return tools.Schema{Type: "object", AdditionalProperties: false}
+}
+func (t fakeDeferredMutatorTool) Safety() tools.Safety {
+	return tools.Safety{SideEffect: tools.SideEffectWrite, Permission: tools.PermissionPrompt, Reason: "mutates"}
+}
+func (t fakeDeferredMutatorTool) Run(_ context.Context, _ map[string]any) tools.Result {
+	return tools.Result{Status: tools.StatusOK, Output: "mutated"}
+}
+func (t fakeDeferredMutatorTool) Deferred() bool { return true }
+
+// TestPlanModeToolSearchNeverLeaksDeferredMutatorSchema guards the concern
+// raised in PR #642 review (jatmn, P2): that tool_search filters deferred
+// candidates only by EnabledTools/DisabledTools, not by plan-mode visibility,
+// so a plan-mode model could call `tool_search select:<deferred write tool>`
+// and receive that tool's full schema even though a direct call to it is
+// correctly denied the following turn.
+//
+// tool_search's own Safety is SideEffectNone, and ToolAdvertisedForPermissionMode (the
+// same gate executeToolCall uses to deny a direct call) requires
+// SideEffect==Read to advertise a tool in plan mode. That means tool_search
+// itself is never advertised, never activates deferral (loaderUsable in
+// partitionToolsCached requires ToolAdvertised(loader, permissionMode)), and
+// is denied at dispatch like any other hidden tool if a stale/adversarial
+// call reaches it anyway — so a deferred mutator's schema can never reach the
+// model through tool_search while in plan mode. This test pins that
+// end-to-end: tool_search is absent from the advertised tool list, and a
+// forced call to it is denied before rendering any tool schema.
+func TestPlanModeToolSearchNeverLeaksDeferredMutatorSchema(t *testing.T) {
+	root := t.TempDir()
+	registry := tools.NewRegistry()
+	registry.Register(tools.NewReadFileTool(root))
+	registry.Register(fakeDeferredMutatorTool{name: "mcp__srv__mutate"})
+	registry.Register(fakeDeferredMutatorTool{name: "mcp__srv__mutate2"})
+	registry.Register(tools.NewToolSearchTool(registry))
+
+	provider := &mockProvider{turns: [][]zeroruntime.StreamEvent{
+		{ // turn 1: force a call to tool_search even though it should not be advertised.
+			{Type: zeroruntime.StreamEventToolCallStart, ToolCallID: "c1", ToolName: tools.ToolSearchToolName},
+			{Type: zeroruntime.StreamEventToolCallDelta, ToolCallID: "c1", ArgumentsFragment: `{"query":"select:mcp__srv__mutate"}`},
+			{Type: zeroruntime.StreamEventToolCallEnd, ToolCallID: "c1"},
+			{Type: zeroruntime.StreamEventDone},
+		},
+		{ // turn 2: final answer.
+			{Type: zeroruntime.StreamEventText, Content: "done"},
+			{Type: zeroruntime.StreamEventDone},
+		},
+	}}
+
+	result, err := Run(context.Background(), "plan", provider, Options{
+		Registry:       registry,
+		PermissionMode: PermissionModePlan,
+		DeferThreshold: 2, // 2 deferred mutators registered => eligible for deferral.
+		MaxTurns:       2,
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	// tool_search (and the deferred mutators) must not be advertised in plan
+	// mode's turn 1 tool list at all.
+	for _, def := range provider.requests[0].Tools {
+		if def.Name == tools.ToolSearchToolName {
+			t.Fatalf("plan mode must not advertise tool_search, got %#v", provider.requests[0].Tools)
+		}
+		if def.Name == "mcp__srv__mutate" || def.Name == "mcp__srv__mutate2" {
+			t.Fatalf("plan mode must not advertise a deferred mutator, got %#v", provider.requests[0].Tools)
+		}
+	}
+
+	// The forced call must be denied outright, never a loaded-schema result:
+	// the tool result must not mention the mutator's name or carry a
+	// load_tools signal.
+	var toolMessage string
+	for _, message := range result.Messages {
+		if message.Role == zeroruntime.MessageRoleTool {
+			toolMessage = message.Content
+			break
+		}
+	}
+	if !strings.Contains(toolMessage, "not available in plan mode") {
+		t.Fatalf("expected tool_search call denied in plan mode, got %q", toolMessage)
+	}
+	if strings.Contains(toolMessage, "mcp__srv__mutate") {
+		t.Fatalf("denial must not leak the deferred mutator's name/schema, got %q", toolMessage)
+	}
+}

internal/agent/denial_test.go

@@ -11,7 +11,7 @@ import (
 func TestExecuteToolCallCategorizesFilteredDenial(t *testing.T) {
 	root := t.TempDir()
 	registry := tools.NewRegistry()
-	registry.Register(tools.NewReadFileTool(root))
+	registry.Register(tools.NewScopedReadFileTool(root, nil))
 
 	result, _ := executeToolCall(context.Background(), registry, ToolCall{ID: "c1", Name: "read_file", Arguments: `{"path":"x"}`}, PermissionModeAuto, Options{
 		DisabledTools: []string{"read_file"},

internal/agent/escalate_loop_test.go

@@ -144,9 +144,10 @@ func TestRunSwitchesProviderOnEscalationRequest(t *testing.T) {
 	var switchedTo string
 	switchCount := 0
 	result, err := Run(context.Background(), "go", firstProvider, Options{
-		Registry: registry,
-		Model:    "claude-sonnet-4.5",
-		MaxTurns: 4,
+		Registry:    registry,
+		Model:       "claude-sonnet-4.5",
+		ServiceTier: "priority",
+		MaxTurns:    4,
 		ModelSwitcher: func(_ context.Context, modelID string) (Provider, error) {
 			switchCount++
 			switchedTo = modelID
@@ -167,9 +168,15 @@ func TestRunSwitchesProviderOnEscalationRequest(t *testing.T) {
 	if len(firstProvider.requests) != 1 {
 		t.Fatalf("expected first provider to handle exactly the escalation turn, got %d requests", len(firstProvider.requests))
 	}
+	if got := firstProvider.requests[0].ServiceTier; got != "priority" {
+		t.Fatalf("initial request service tier = %q, want priority", got)
+	}
 	if len(secondProvider.requests) != 1 {
 		t.Fatalf("expected second provider to handle the post-switch turn, got %d requests", len(secondProvider.requests))
 	}
+	if got := secondProvider.requests[0].ServiceTier; got != "" {
+		t.Fatalf("post-switch request retained service tier %q", got)
+	}
 	if result.FinalAnswer != "answered on the upgraded model" {
 		t.Fatalf("expected final answer from the swapped provider, got %q", result.FinalAnswer)
 	}
@@ -211,9 +218,10 @@ func TestRunEscalationSwitcherErrorIsNonFatal(t *testing.T) {
 	provider := &mockProvider{turns: escalateThenAnswerTurns("recovered")}
 
 	result, err := Run(context.Background(), "go", provider, Options{
-		Registry: registry,
-		Model:    "claude-sonnet-4.5",
-		MaxTurns: 4,
+		Registry:    registry,
+		Model:       "claude-sonnet-4.5",
+		ServiceTier: "priority",
+		MaxTurns:    4,
 		ModelSwitcher: func(_ context.Context, _ string) (Provider, error) {
 			return nil, errors.New("provider build blew up")
 		},
@@ -225,6 +233,9 @@ func TestRunEscalationSwitcherErrorIsNonFatal(t *testing.T) {
 	if len(provider.requests) != 2 {
 		t.Fatalf("expected the run to continue on the original provider, got %d requests", len(provider.requests))
 	}
+	if got := provider.requests[1].ServiceTier; got != "priority" {
+		t.Fatalf("failed-switch request service tier = %q, want priority", got)
+	}
 	if result.FinalAnswer != "recovered" {
 		t.Fatalf("expected final answer after non-fatal switch error, got %q", result.FinalAnswer)
 	}

internal/agent/export_test.go

@@ -0,0 +1,41 @@
+// Test seams: helpers only test code uses, kept out of the production binary.
+package agent
+
+import (
+	"github.com/Gitlawb/zero/internal/tools"
+	"github.com/Gitlawb/zero/internal/zeroruntime"
+)
+
+func equalStringSlices(left []string, right []string) bool {
+	if len(left) != len(right) {
+		return false
+	}
+	for index := range left {
+		if left[index] != right[index] {
+			return false
+		}
+	}
+	return true
+}
+
+// parsePreservedState recovers the plan + skills from a prior summary's preserved
+// block. JSON escaping makes this lossless even when a skill body contains
+// markdown headings, code fences, or quotes. Returns ("", nil) when absent or
+// malformed.
+func parsePreservedState(summaryContent string) (string, []skillEntry) {
+	state := parsePreservedStateBlock(summaryContent)
+	return state.Plan, preservedSkillsToEntries(state.Skills)
+}
+
+// partitionTools builds the per-turn advertised tool list and optional
+// tool_search discovery text. INACTIVE (DeferThreshold <= 0 or the eligible count is
+// below it): every visible tool is exposed with its full schema EXCEPT tool_search
+// (dropped so it is never advertised when it cannot help), and the discovery text is
+// empty — byte-identical to the pre-deferral output. ACTIVE: a deferred-eligible
+// tool is exposed only when loaded[name]; otherwise it is hidden and searchable
+// through tool_search. Non-deferred tools (including tool_search) are always
+// exposed. The exposed slice is alpha-sorted by name, matching the legacy order
+// so the inactive path is stable.
+func partitionTools(registry *tools.Registry, permissionMode PermissionMode, options Options, loaded map[string]bool) ([]zeroruntime.ToolDefinition, string) {
+	return partitionToolsCached(registry, permissionMode, options, loaded, nil)
+}

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

@@ -1,7 +1,6 @@
 package agent
 
 import (
-	"encoding/json"
 	"strconv"
 	"strings"
 
@@ -22,6 +21,12 @@ 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
@@ -43,8 +48,12 @@ 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
@@ -117,9 +126,8 @@ func endsWithContinuationCue(text string) bool {
 // 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":
+	switch normalizeTaskPlanStatus(status) {
+	case "completed", "failed":
 		return false
 	default:
 		return true
@@ -173,35 +181,112 @@ var successNegationTails = []string{
 	"reproduce", "spot any", "locate any",
 }
 
-// selfReportedIncompletion returns a short reason when the model's final text
-// admits it guessed or could not meet the objective, else "". Case-insensitive.
-func selfReportedIncompletion(text string) string {
-	lower := strings.ToLower(text)
-	for _, phrase := range selfReportPhrases {
-		if strings.Contains(lower, phrase) {
-			return selfReportReason(phrase)
+// 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)
 		}
 	}
-	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(lower[start:], stem)
-			if rel < 0 {
-				break
+	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)
 			}
-			abs := start + rel
-			tail := strings.TrimSpace(lower[abs+len(stem):])
-			if !hasAnyPrefix(tail, successNegationTails) {
-				return selfReportReason(strings.TrimSpace(stem) + " …")
+		}
+		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)
 			}
-			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 + `")`
 }
@@ -220,8 +305,8 @@ func hasAnyPrefix(s string, prefixes []string) bool {
 // 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 + " — " +
+func acceptanceVerificationNudge(objective string) string {
+	nudge := "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: " +
@@ -229,6 +314,10 @@ func acceptanceVerificationNudge() string {
 		"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."
+	if objective = capTaskObjective(objective); objective != "" {
+		nudge += "\n\nTask objective: " + objective
+	}
+	return nudge
 }
 
 // toolFailureHintMarker is a stable substring for tests.
@@ -354,8 +443,13 @@ 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.
@@ -414,8 +508,9 @@ 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++
@@ -427,11 +522,16 @@ 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
@@ -455,16 +555,12 @@ func (state *guardState) pendingPlanItems() bool {
 // 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 {
+	plan, ok := parseTaskPlan(arguments)
+	if !ok {
 		return
 	}
 	pending := 0
-	for _, item := range parsed.Plan {
+	for _, item := range plan {
 		if planStatusRemaining(item.Status) {
 			pending++
 		}
@@ -485,10 +581,12 @@ 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{
@@ -101,11 +109,36 @@ 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")
 	registry := tools.NewRegistry()
-	registry.Register(tools.NewReadFileTool(root))
+	registry.Register(tools.NewScopedReadFileTool(root, nil))
 
 	provider := &mockProvider{
 		turns: [][]zeroruntime.StreamEvent{
@@ -136,6 +169,37 @@ func TestRunResetsEmptyTurnCounterOnToolCall(t *testing.T) {
 	}
 }
 
+// The whole point of #702's fix: the unknown-session error's normalized
+// signature must not vary with the session id, so a model probing ids 1, 2,
+// 3, … produces one repeated signature the failure guard can count. If the id
+// leaked into the first 80 normalized chars, each probe would reset the streak
+// and the halt would never fire.
+func TestUnknownExecSessionErrorSignatureIsIDInvariant(t *testing.T) {
+	a := errorSignature(tools.UnknownExecSessionError(1))
+	b := errorSignature(tools.UnknownExecSessionError(999999))
+	if a != b {
+		t.Fatalf("unknown-session signature varies with id:\n  %q\n  %q", a, b)
+	}
+}
+
+// End to end: with an id-invariant signature, probing a different unknown id
+// each turn now trips the repeated-failure halt at toolFailureStopAt, where
+// before the fix it never would.
+func TestUnknownExecSessionProbingTripsFailureHalt(t *testing.T) {
+	var state guardState
+	var stoppedAt int
+	for i := 1; i <= toolFailureStopAt; i++ {
+		out := state.observeToolResult(tools.WriteStdinToolName, true, tools.UnknownExecSessionError(i))
+		if out.Stop {
+			stoppedAt = i
+			break
+		}
+	}
+	if stoppedAt != toolFailureStopAt {
+		t.Fatalf("probing distinct unknown ids stopped at %d, want %d", stoppedAt, toolFailureStopAt)
+	}
+}
+
 func TestGuardStateResetsToolOnlyStreakOnEmptyNonToolTurn(t *testing.T) {
 	var state guardState
 	toolOnly := zeroruntime.CollectedStream{
@@ -196,7 +260,7 @@ func TestRunInjectsPlanNotCalledReminderForMultiStepTask(t *testing.T) {
 	root := t.TempDir()
 	writeAgentTestFile(t, root+"/notes.txt", "alpha")
 	registry := tools.NewRegistry()
-	registry.Register(tools.NewReadFileTool(root))
+	registry.Register(tools.NewScopedReadFileTool(root, nil))
 
 	provider := &mockProvider{
 		turns: [][]zeroruntime.StreamEvent{
@@ -227,7 +291,7 @@ func TestRunDoesNotInjectPlanReminderForTrivialTask(t *testing.T) {
 	root := t.TempDir()
 	writeAgentTestFile(t, root+"/notes.txt", "alpha")
 	registry := tools.NewRegistry()
-	registry.Register(tools.NewReadFileTool(root))
+	registry.Register(tools.NewScopedReadFileTool(root, nil))
 
 	provider := &mockProvider{
 		turns: [][]zeroruntime.StreamEvent{
@@ -255,7 +319,7 @@ func TestRunDoesNotInjectNotCalledReminderWhenPlanUsed(t *testing.T) {
 	root := t.TempDir()
 	writeAgentTestFile(t, root+"/notes.txt", "alpha")
 	registry := tools.NewRegistry()
-	registry.Register(tools.NewReadFileTool(root))
+	registry.Register(tools.NewScopedReadFileTool(root, nil))
 	registry.Register(tools.NewUpdatePlanTool())
 
 	provider := &mockProvider{
@@ -285,7 +349,7 @@ func TestRunInjectsStalePlanReminderAfterManyToolCalls(t *testing.T) {
 	root := t.TempDir()
 	writeAgentTestFile(t, root+"/notes.txt", "alpha")
 	registry := tools.NewRegistry()
-	registry.Register(tools.NewReadFileTool(root))
+	registry.Register(tools.NewScopedReadFileTool(root, nil))
 	registry.Register(tools.NewUpdatePlanTool())
 
 	// Turn 1 calls update_plan (so the not-called reminder never triggers), then
@@ -319,7 +383,7 @@ func TestRunStalePlanReminderIsOneShotPerInterval(t *testing.T) {
 	root := t.TempDir()
 	writeAgentTestFile(t, root+"/notes.txt", "alpha")
 	registry := tools.NewRegistry()
-	registry.Register(tools.NewReadFileTool(root))
+	registry.Register(tools.NewScopedReadFileTool(root, nil))
 	registry.Register(tools.NewUpdatePlanTool())
 
 	turns := [][]zeroruntime.StreamEvent{
@@ -354,7 +418,7 @@ func TestRunInjectsToolOnlyProgressReminder(t *testing.T) {
 	root := t.TempDir()
 	writeAgentTestFile(t, root+"/notes.txt", "alpha")
 	registry := tools.NewRegistry()
-	registry.Register(tools.NewReadFileTool(root))
+	registry.Register(tools.NewScopedReadFileTool(root, nil))
 
 	turns := make([][]zeroruntime.StreamEvent, 0, toolOnlyProgressReminderAt+1)
 	for i := 0; i < toolOnlyProgressReminderAt; i++ {

internal/agent/output_budget_propagation_test.go

@@ -0,0 +1,255 @@
+package agent
+
+import (
+	"context"
+	"os"
+	"path/filepath"
+	"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-repeat" && index+1 < len(os.Args) {
+			repeats, err := strconv.Atoi(os.Args[index+1])
+			if err != nil {
+				os.Exit(2)
+			}
+			if _, err := os.Stdout.WriteString(strings.Repeat("hook feedback ", repeats)); err != nil {
+				os.Exit(2)
+			}
+			os.Exit(0)
+		}
+		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 {
+	return outputBudgetHookDispatcherFor("propagation_output", 200)
+}
+
+func outputBudgetHookDispatcherFor(toolName string, repeats int) *hooks.Dispatcher {
+	feedback := strings.Repeat("hook feedback ", repeats)
+	hookArgs := []string{
+		"-test.run=TestOutputBudgetHookHelperProcess",
+		"--",
+		"--zero-output-budget-hook",
+		feedback,
+	}
+	if repeats > 1_000 {
+		hookArgs = []string{
+			"-test.run=TestOutputBudgetHookHelperProcess",
+			"--",
+			"--zero-output-budget-hook-repeat",
+			strconv.Itoa(repeats),
+		}
+	}
+	return hooks.NewDispatcher(hooks.DispatcherOptions{Config: hooks.Config{
+		Enabled: true,
+		Hooks: []hooks.Definition{{
+			ID:      "large-feedback",
+			Event:   hooks.EventAfterTool,
+			Matcher: toolName,
+			Command: os.Args[0],
+			Args:    hookArgs,
+			Enabled: true,
+		}},
+	}})
+}
+
+func TestExecuteToolCallCommitsReadObservationAfterFinalBoundary(t *testing.T) {
+	root := t.TempDir()
+	targetPath := filepath.Join(root, "target.txt")
+	if err := os.WriteFile(targetPath, []byte("known content\n"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	linkPath := filepath.Join(root, "small.txt")
+	if err := os.Symlink(targetPath, linkPath); err != nil {
+		t.Skipf("symlink unavailable: %v", err)
+	}
+	resolvedPath, err := filepath.EvalSymlinks(linkPath)
+	if err != nil {
+		t.Fatal(err)
+	}
+	registry := tools.NewRegistry()
+	registry.Register(tools.NewReadFileTool(root))
+	tracker := tools.NewFileTracker()
+
+	result, abortErr := executeToolCall(context.Background(), registry, ToolCall{
+		ID: "read-small", Name: "read_file", Arguments: `{"path":"small.txt"}`,
+	}, PermissionModeAuto, Options{Cwd: root, FileTracker: tracker})
+	if abortErr != nil || result.Status != tools.StatusOK {
+		t.Fatalf("executeToolCall: result=%#v err=%v", result, abortErr)
+	}
+	if !tracker.SeenWhole(resolvedPath) {
+		t.Fatal("unmodified final read output was not committed as observed")
+	}
+}
+
+func TestExecuteToolCallDoesNotCommitReadObservationBeforeHookBudget(t *testing.T) {
+	t.Setenv("ZERO_TOOL_OUTPUT_CEILING_TOKENS", "80")
+	root := t.TempDir()
+	targetPath := filepath.Join(root, "target.txt")
+	if err := os.WriteFile(targetPath, []byte("known content\n"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	linkPath := filepath.Join(root, "small.txt")
+	if err := os.Symlink(targetPath, linkPath); err != nil {
+		t.Skipf("symlink unavailable: %v", err)
+	}
+	resolvedPath, err := filepath.EvalSymlinks(linkPath)
+	if err != nil {
+		t.Fatal(err)
+	}
+	registry := tools.NewRegistry()
+	registry.Register(tools.NewReadFileTool(root))
+	tracker := tools.NewFileTracker()
+
+	result, abortErr := executeToolCall(context.Background(), registry, ToolCall{
+		ID: "read-hook", Name: "read_file", Arguments: `{"path":"small.txt"}`,
+	}, PermissionModeAuto, Options{
+		Cwd: root, FileTracker: tracker, Hooks: outputBudgetHookDispatcherFor("read_file", 20_000),
+	})
+	if abortErr != nil {
+		t.Fatalf("executeToolCall: %v", abortErr)
+	}
+	if !result.Truncated {
+		t.Fatalf("precondition: hook feedback was not truncated: %#v", result)
+	}
+	if tracker.SeenWhole(resolvedPath) {
+		t.Fatal("read observation committed before hook output was budgeted")
+	}
+}
+
+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.NewScopedReadFileTool(root, nil)) // 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 +1,128 @@
+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")
+	}
+}
+
+func TestBuildPermissionEventReusesCoveredAdditionalPermissions(t *testing.T) {
+	call := ToolCall{ID: "call_3", Name: "bash", Arguments: `{}`}
+	args := map[string]any{
+		"command":             "npm install",
+		"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(promptShellTool{}, args, true, decision) {
+		t.Fatal("an already-covered capability must not trigger another prompt")
+	}
+	event, ok := buildPermissionEvent(call, promptShellTool{}, args, true, PermissionModeAsk, Options{}, decision)
+	if !ok {
+		t.Fatal("covered capability should still produce an auditable allow event")
+	}
+	if event.Action != PermissionActionAllow {
+		t.Fatalf("Action = %q, want allow for an already-covered capability", event.Action)
+	}
+}
+
+func TestBuildPermissionEventOmitsGenericShellReasonFromPrompt(t *testing.T) {
+	call := ToolCall{ID: "call_4", Name: "bash", Arguments: `{}`}
+	args := map[string]any{"command": "make test"}
+	decision := &sandbox.Decision{Action: sandbox.ActionPrompt, Reason: "runs a shell command"}
+
+	event, ok := buildPermissionEvent(call, promptShellTool{}, args, false, PermissionModeAsk, Options{}, decision)
+	if !ok {
+		t.Fatal("generic shell approval must produce an event")
+	}
+	if event.Reason != "" {
+		t.Fatalf("user-facing reason = %q, want omitted generic tool text", event.Reason)
+	}
+	if event.DecisionReason != "runs a shell command" {
+		t.Fatalf("decision reason = %q, want raw audit reason retained", event.DecisionReason)
+	}
+}
+
+func TestBuildPermissionEventPrefersPolicyReasonOverModelJustification(t *testing.T) {
+	call := ToolCall{ID: "call_5", Name: "bash", Arguments: `{}`}
+	args := map[string]any{
+		"command":             "curl https://example.test",
+		"sandbox_permissions": string(tools.SandboxPermissionsRequireEscalated),
+		"justification":       "Run without sandboxing.",
+	}
+	decision := &sandbox.Decision{Action: sandbox.ActionPrompt, Reason: sandbox.ReasonNetworkBlocked}
+
+	event, ok := buildPermissionEvent(call, promptShellTool{}, args, false, PermissionModeAsk, Options{}, decision)
+	if !ok {
+		t.Fatal("policy approval must produce an event")
+	}
+	if event.Reason != sandbox.ReasonNetworkBlocked {
+		t.Fatalf("user-facing reason = %q, want policy reason %q", event.Reason, sandbox.ReasonNetworkBlocked)
+	}
+}

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/prefix_escalation_disclosure_test.go

@@ -0,0 +1,118 @@
+package agent
+
+import (
+	"path/filepath"
+	"testing"
+
+	"github.com/Gitlawb/zero/internal/sandbox"
+	"github.com/Gitlawb/zero/internal/tools"
+)
+
+// escalatingEngine allows unsandboxed escalation: no denied reads, so
+// UnsandboxedExecutionAllowed is true and a prefix approval would rewrite the
+// call to require_escalated.
+func escalatingEngine(t *testing.T) *sandbox.Engine {
+	t.Helper()
+	policy := sandbox.DefaultPolicy()
+	policy.DenyRead = nil
+	return sandbox.NewEngine(sandbox.EngineOptions{WorkspaceRoot: t.TempDir(), Policy: policy})
+}
+
+// nonEscalatingEngine has a denied read, which is exactly the condition under
+// which escalation is refused, so no disclosure should be made.
+func nonEscalatingEngine(t *testing.T) *sandbox.Engine {
+	t.Helper()
+	root := t.TempDir()
+	policy := sandbox.DefaultPolicy()
+	policy.DenyRead = []string{filepath.Join(root, "secret.txt")}
+	return sandbox.NewEngine(sandbox.EngineOptions{WorkspaceRoot: root, Policy: policy})
+}
+
+// THE POINT OF THE FILE. The prompt's disclosure and the args rewrite must
+// agree on every input, because the defect being fixed is precisely that the
+// two disagreed: the rewrite escalated out of the sandbox while the prompt said
+// only "allow command prefix for session".
+//
+// Asserting agreement rather than asserting each side's expected value is what
+// makes this hold up. If someone later widens or narrows the rewrite's guards
+// and forgets the flag, the two diverge here and this fails, which a pair of
+// independent expected-value tests would not catch.
+func TestPrefixEscalationDisclosureMatchesTheArgsRewrite(t *testing.T) {
+	escalating := Options{Sandbox: escalatingEngine(t)}
+	denied := Options{Sandbox: nonEscalatingEngine(t)}
+
+	cases := []struct {
+		name     string
+		toolName string
+		args     map[string]any
+		options  Options
+	}{
+		{"plain bash command", "bash", map[string]any{"command": "go test ./..."}, escalating},
+		{"plain exec_command", "exec_command", map[string]any{"cmd": "go build ./..."}, escalating},
+		{"not a shell tool", "read_file", map[string]any{"path": "go.mod"}, escalating},
+		{"no sandbox configured", "bash", map[string]any{"command": "go test ./..."}, Options{}},
+		{"escalation refused by denied reads", "bash", map[string]any{"command": "go test ./..."}, denied},
+		{
+			"call already asks for escalation",
+			"bash",
+			map[string]any{"command": "go test ./...", "sandbox_permissions": string(tools.SandboxPermissionsRequireEscalated)},
+			escalating,
+		},
+		{
+			"call already asks for additional permissions",
+			"bash",
+			map[string]any{"command": "go test ./...", "sandbox_permissions": string(tools.SandboxPermissionsWithAdditionalPermissions)},
+			escalating,
+		},
+		// An unrecognized value is NOT an opt-out: the rewrite overwrites it, so
+		// the disclosure has to fire. Included because the obvious reading of
+		// the guards says otherwise, and a case asserting the wrong thing here
+		// would have passed silently.
+		{
+			"unrecognized sandbox_permissions value",
+			"bash",
+			map[string]any{"command": "go test ./...", "sandbox_permissions": "all"},
+			escalating,
+		},
+		{"empty args", "bash", map[string]any{}, escalating},
+		{"nil args", "bash", nil, escalating},
+	}
+
+	for _, testCase := range cases {
+		t.Run(testCase.name, func(t *testing.T) {
+			disclosed := shellPrefixApprovalEscalates(testCase.toolName, testCase.args, testCase.options)
+
+			// What the user's approval would actually be turned into.
+			rewritten := shellExecutionArgsForApproval(testCase.toolName, testCase.args, PermissionDecisionAllowPrefix, testCase.options)
+			escalates := shellCommandRequiresEscalated(rewritten) && !shellCommandRequiresEscalated(testCase.args)
+
+			if disclosed != escalates {
+				t.Fatalf("prompt discloses escalation=%v but approving actually escalates=%v; the prompt and the rewrite disagree, which is the consent gap this guards", disclosed, escalates)
+			}
+		})
+	}
+}
+
+// The two prefix decisions are the only ones that escalate, so a disclosure
+// attached to a request offering neither would be a false alarm. Checked
+// against the same predicate the rewrite consults.
+func TestOnlyPrefixDecisionsEscalate(t *testing.T) {
+	escalating := map[PermissionDecisionAction]bool{
+		PermissionDecisionAllowPrefix:       true,
+		PermissionDecisionAlwaysAllowPrefix: true,
+	}
+	for _, action := range []PermissionDecisionAction{
+		PermissionDecisionAllow,
+		PermissionDecisionAllowStrict,
+		PermissionDecisionAllowForSession,
+		PermissionDecisionAllowPrefix,
+		PermissionDecisionAlwaysAllowPrefix,
+		PermissionDecisionAlwaysAllow,
+		PermissionDecisionDeny,
+		PermissionDecisionCancel,
+	} {
+		if got := shellPrefixApprovalBypassesSandbox(action); got != escalating[action] {
+			t.Errorf("decision %q bypasses sandbox=%v, want %v", action, got, escalating[action])
+		}
+	}
+}

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 +1,87 @@
+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 ~3230 tokens. The tool figure is
+// the auto-mode advertised set, including exec_command: auto mode must expose both
+// process creation and write_stdin process interaction or the latter is unusable
+// without a session id. Higher-risk tools that do not opt into auto remain excluded.
+const (
+	maxBaseSystemPromptTokens = 3500
+	// Raised deliberately from 3550 on 2026-08-08. view_image (#843) took the set
+	// to 3588, and #867 trimmed it back to 3578: still 28 tokens over.
+	//
+	// The overrun is a rounding error rather than a regression. view_image costs
+	// 82 tokens, the second cheapest tool in the set, against exec_command's 909
+	// and request_permissions' 397; it is simply the one that crossed the line.
+	// Raised rather than paid down because the alternative was trimming a
+	// description to claw back 28 tokens, degrading a tool's usability to satisfy
+	// a line drawn against a 2026-07 measurement that predates two tools.
+	//
+	// The ratchet did its job: it caught the creep and forced a decision instead
+	// of a drift. That is the point of it, so keep raising it deliberately rather
+	// than reflexively.
+	maxEagerToolSchemaTokens = 3650
+)
+
+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 {
+		// NOT "defer a tool": this test pins DeferThreshold at 0, so marking a
+		// tool deferred leaves it exposed here and changes nothing. Deferral is
+		// still worth doing for real sessions, it just cannot move this number.
+		// The levers that do are a smaller schema, one fewer core tool, or a
+		// deliberate raise.
+		t.Fatalf("eager tool schemas are %d tokens, over the %d ceiling — trim a schema, drop a core tool, or raise the ceiling deliberately (deferring will NOT help: this test disables deferral)", got, maxEagerToolSchemaTokens)
+	}
+}
+
+func TestAgentAdvertisesBothEditTools(t *testing.T) {
+	registry := tools.NewRegistry()
+	for _, tool := range tools.CoreToolsScoped(t.TempDir(), nil) {
+		registry.Register(tool)
+	}
+	exposed, _ := partitionTools(registry, PermissionModeAsk, Options{}, map[string]bool{})
+	names := make(map[string]bool, len(exposed))
+	for _, definition := range exposed {
+		names[definition.Name] = true
+	}
+	if !names["apply_patch"] {
+		t.Fatal("agent must retain apply_patch for multi-hunk changes")
+	}
+	if !names["edit_file"] {
+		t.Fatal("agent must receive edit_file for targeted exact replacements")
+	}
+}

internal/agent/prompt_fingerprint.go

@@ -0,0 +1,244 @@
+package agent
+
+import (
+	"crypto/sha256"
+	"encoding/hex"
+	"encoding/json"
+	"fmt"
+	"sort"
+	"strings"
+
+	"github.com/Gitlawb/zero/internal/zeroruntime"
+)
+
+// promptSubstrings are the cacheable components of the prompt a run sends to a
+// model. systemPrompt is the exact joined system-message content. The remaining
+// fields retain useful diagnostic boundaries within that prompt and tool list.
+type promptSubstrings struct {
+	systemPrompt       string
+	baseInstructions   string
+	confirmationPolicy string
+	projectContext     string
+	skills             string
+	tools              string
+	schema             string
+}
+
+// buildPromptSubstrings builds the prompt and retains both its exact joined
+// form and narrower diagnostic components. Callers that already built the
+// prompt should use buildPromptSubstringsFromParts to avoid repeating workspace
+// reads.
+func buildPromptSubstrings(options Options, exposed []zeroruntime.ToolDefinition) promptSubstrings {
+	return buildPromptSubstringsFromParts(buildSystemPromptParts(options), exposed)
+}
+
+func buildPromptSubstringsFromParts(parts systemPromptParts, exposed []zeroruntime.ToolDefinition) promptSubstrings {
+	toolsSubstr, schemaSubstr := toolSubstrings(exposed)
+
+	return promptSubstrings{
+		systemPrompt:       parts.prompt,
+		baseInstructions:   parts.baseInstructions,
+		confirmationPolicy: parts.confirmationPolicy,
+		projectContext:     parts.projectContext,
+		skills:             parts.skills,
+		tools:              toolsSubstr,
+		schema:             schemaSubstr,
+	}
+}
+
+// toolSubstrings preserves the emitted tool order. Names and descriptions are
+// fingerprinted separately from schemas so either kind of drift is identifiable.
+// Order is part of request identity: silently sorting here would hide a provider-
+// visible reordering and could report a false cache hit.
+func toolSubstrings(exposed []zeroruntime.ToolDefinition) (toolsSubstr, schemaSubstr string) {
+	if len(exposed) == 0 {
+		return "", ""
+	}
+	var toolsSB, schemaSB strings.Builder
+	for _, def := range exposed {
+		writeLengthPrefixed(&toolsSB, def.Name)
+		writeLengthPrefixed(&toolsSB, def.Description)
+		// 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.
+		writeLengthPrefixed(&schemaSB, def.Name)
+		writeLengthPrefixed(&schemaSB, canonicalSchemaString(def.Parameters))
+	}
+	return toolsSB.String(), schemaSB.String()
+}
+
+func writeLengthPrefixed(sb *strings.Builder, value string) {
+	fmt.Fprintf(sb, "%d:", len(value))
+	sb.WriteString(value)
+}
+
+// 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 builds and fingerprints a prompt for callers that do
+// not already have one. The agent loop uses computePrefixFingerprint with the
+// prompt parts it built once at run start, ensuring tracing observes the exact
+// request content without repeating workspace I/O.
+func ComputePrefixFingerprint(options Options, exposed []zeroruntime.ToolDefinition) prefixFingerprint {
+	return computePrefixFingerprint(buildPromptSubstrings(options, exposed))
+}
+
+func computePrefixFingerprint(subs promptSubstrings) prefixFingerprint {
+	systemPrompt := sha256hex(subs.systemPrompt)
+	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{
+		systemPrompt, toolsH, schema,
+	}, "|"))
+	return prefixFingerprint{
+		SystemPromptHash:       systemPrompt,
+		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 {
+	SystemPromptHash       string
+	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 +1,264 @@
+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")
+	}
+}
+
+// Tool order and descriptions are provider-visible and therefore part of the
+// cacheable prefix identity.
+func TestComputePrefixFingerprintToolsHashMatchesEmittedOrderAndDescriptions(t *testing.T) {
+	opts := Options{Cwd: t.TempDir(), SystemPrompt: "core"}
+	a := []zeroruntime.ToolDefinition{
+		{Name: "read_file", Description: "read", Parameters: map[string]any{"schema": "schema-read"}},
+		{Name: "grep", Description: "search", Parameters: map[string]any{"schema": "schema-grep"}},
+	}
+	b := []zeroruntime.ToolDefinition{
+		{Name: "grep", Description: "search", Parameters: map[string]any{"schema": "schema-grep"}},
+		{Name: "read_file", Description: "read", Parameters: map[string]any{"schema": "schema-read"}},
+	}
+	fpA := ComputePrefixFingerprint(opts, a)
+	fpB := ComputePrefixFingerprint(opts, b)
+	if fpA.ToolsHash == fpB.ToolsHash {
+		t.Fatalf("ToolsHash must change when emitted order changes: a=%s b=%s", fpA.ToolsHash, fpB.ToolsHash)
+	}
+	c := append([]zeroruntime.ToolDefinition(nil), a...)
+	c[0].Description = "read a file"
+	fpC := ComputePrefixFingerprint(opts, c)
+	if fpA.ToolsHash == fpC.ToolsHash {
+		t.Fatalf("ToolsHash must change when a description 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)
+	}
+}
+
+// CompletePrefixHash represents the exact joined system prompt and the ordered
+// tool definitions. Narrower system-section hashes remain diagnostic only.
+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.SystemPromptHash,
+		fp.ToolsHash,
+		fp.SchemaHash,
+	}, "|"))
+	if fp.CompletePrefixHash != expected {
+		t.Fatalf("CompletePrefixHash must equal sha256hex of the exact system, tools, and schema hashes:\n  got:      %s\n  expected: %s", fp.CompletePrefixHash, expected)
+	}
+}
+
+func TestComputePrefixFingerprintCoversEverySystemPromptSection(t *testing.T) {
+	base := Options{Cwd: t.TempDir(), SystemPrompt: "core", Model: "gpt-5"}
+	changed := base
+	changed.ResponseStyle = "concise"
+
+	fpBase := ComputePrefixFingerprint(base, nil)
+	fpChanged := ComputePrefixFingerprint(changed, nil)
+	if fpBase.SystemPromptHash == fpChanged.SystemPromptHash {
+		t.Fatal("SystemPromptHash must change when any rendered system section changes")
+	}
+	if fpBase.CompletePrefixHash == fpChanged.CompletePrefixHash {
+		t.Fatal("CompletePrefixHash must include the exact full system prompt")
+	}
+}
+
+func TestComputePrefixFingerprintCoversTransientSystemPrompt(t *testing.T) {
+	base := Options{Cwd: t.TempDir(), SystemPrompt: "core"}
+	changed := base
+	changed.TransientSystemPrompt = "turn-only guidance"
+
+	fpBase := ComputePrefixFingerprint(base, nil)
+	fpChanged := ComputePrefixFingerprint(changed, nil)
+	if fpBase.SystemPromptHash == fpChanged.SystemPromptHash {
+		t.Fatal("SystemPromptHash must change for transient system guidance")
+	}
+	if fpBase.CompletePrefixHash == fpChanged.CompletePrefixHash {
+		t.Fatal("CompletePrefixHash must cover transient system guidance")
+	}
+}
+
+// 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.systemPrompt != buildSystemPrompt(opts) {
+		t.Fatal("systemPrompt substring must equal the exact joined system prompt")
+	}
+	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.SystemPromptHash != sha256hex(subs.systemPrompt) {
+		t.Fatalf("SystemPromptHash in fingerprint must equal sha256hex of the joined prompt")
+	}
+	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/prune.go

@@ -121,3 +121,67 @@ func toolNameForResult(messages []zeroruntime.Message, index int) string {
 func isPrunedPlaceholder(content string) bool {
 	return strings.HasPrefix(strings.TrimSpace(content), "[pruned ")
 }
+
+var supersedableReadTools = map[string]bool{
+	"read_file":          true,
+	"read_minified_file": true,
+	"grep":               true,
+	"glob":               true,
+	"list_directory":     true,
+}
+
+// pruneSupersededReadResults removes only older byte-identical results of the
+// same read call. It runs when context pressure already requires a rewrite, so
+// it does not trade a warm provider-cache prefix for a small apparent saving.
+func pruneSupersededReadResults(messages []zeroruntime.Message) ([]zeroruntime.Message, int) {
+	type seenResultKey struct {
+		call    string
+		content string
+	}
+	seen := map[seenResultKey]struct{}{}
+	out := messages
+	reclaimed := 0
+	copied := false
+
+	for index := len(messages) - 1; index >= 0; index-- {
+		message := messages[index]
+		if message.Role != zeroruntime.MessageRoleTool || isPrunedPlaceholder(message.Content) {
+			continue
+		}
+		call, ok := toolCallForResult(messages, index)
+		if !ok || !supersedableReadTools[call.Name] {
+			continue
+		}
+		key := seenResultKey{
+			call:    call.Name + "\x00" + strings.TrimSpace(call.Arguments),
+			content: message.Content,
+		}
+		if _, exists := seen[key]; exists {
+			placeholder := fmt.Sprintf("[superseded identical %s result — the newer result remains in context]", call.Name)
+			if !copied {
+				out = append([]zeroruntime.Message(nil), messages...)
+				copied = true
+			}
+			out[index].Content = placeholder
+			reclaimed += max(0, ApproxTextTokens(message.Content)-ApproxTextTokens(placeholder))
+			continue
+		}
+		seen[key] = struct{}{}
+	}
+	return out, reclaimed
+}
+
+func toolCallForResult(messages []zeroruntime.Message, resultIndex int) (zeroruntime.ToolCall, bool) {
+	id := messages[resultIndex].ToolCallID
+	if id == "" {
+		return zeroruntime.ToolCall{}, false
+	}
+	for index := resultIndex - 1; index >= 0; index-- {
+		for _, call := range messages[index].ToolCalls {
+			if call.ID == id {
+				return call, true
+			}
+		}
+	}
+	return zeroruntime.ToolCall{}, false
+}

internal/agent/prune_test.go

@@ -25,6 +25,13 @@ func toolCallMsg(id, name string) zeroruntime.Message {
 	}
 }
 
+func toolCallWithArgs(id, name, args string) zeroruntime.Message {
+	return zeroruntime.Message{
+		Role:      zeroruntime.MessageRoleAssistant,
+		ToolCalls: []zeroruntime.ToolCall{{ID: id, Name: name, Arguments: args}},
+	}
+}
+
 func TestPruneSkipsSmallSessions(t *testing.T) {
 	// Total reclaimable is under the gate → no pruning.
 	msgs := []zeroruntime.Message{
@@ -120,3 +127,69 @@ func TestPruneDoesNotMutateInput(t *testing.T) {
 		t.Fatal("pruneStaleToolOutput must not mutate the caller's slice")
 	}
 }
+
+func TestPruneSupersededReadResultsOnlyDropsIdenticalOlderReads(t *testing.T) {
+	body := strings.Repeat("same source line\n", 1000)
+	messages := []zeroruntime.Message{
+		toolCallWithArgs("old", "read_file", `{"path":"a.go"}`),
+		{Role: zeroruntime.MessageRoleTool, ToolCallID: "old", Content: body},
+		toolCallWithArgs("different", "read_file", `{"path":"b.go"}`),
+		{Role: zeroruntime.MessageRoleTool, ToolCallID: "different", Content: body},
+		toolCallWithArgs("new", "read_file", `{"path":"a.go"}`),
+		{Role: zeroruntime.MessageRoleTool, ToolCallID: "new", Content: body},
+	}
+	out, reclaimed := pruneSupersededReadResults(messages)
+	if reclaimed <= 0 || !strings.Contains(out[1].Content, "superseded identical") {
+		t.Fatalf("older duplicate was not pruned: reclaimed=%d output=%q", reclaimed, out[1].Content)
+	}
+	if out[3].Content != body || out[5].Content != body {
+		t.Fatal("different call and newest result must remain exact")
+	}
+	if messages[1].Content != body {
+		t.Fatal("input slice was mutated")
+	}
+}
+
+func TestPruneSupersededReadResultsKeepsChangedAndMutatingResults(t *testing.T) {
+	messages := []zeroruntime.Message{
+		toolCallWithArgs("r1", "read_file", `{"path":"a.go"}`),
+		{Role: zeroruntime.MessageRoleTool, ToolCallID: "r1", Content: "version one"},
+		toolCallWithArgs("r2", "read_file", `{"path":"a.go"}`),
+		{Role: zeroruntime.MessageRoleTool, ToolCallID: "r2", Content: "version two"},
+		toolCallWithArgs("e1", "edit_file", `{"path":"a.go"}`),
+		{Role: zeroruntime.MessageRoleTool, ToolCallID: "e1", Content: "edited"},
+		toolCallWithArgs("e2", "edit_file", `{"path":"a.go"}`),
+		{Role: zeroruntime.MessageRoleTool, ToolCallID: "e2", Content: "edited"},
+	}
+	out, reclaimed := pruneSupersededReadResults(messages)
+	if reclaimed != 0 {
+		t.Fatalf("unsafe results were pruned: %d", reclaimed)
+	}
+	for index := range messages {
+		if out[index].Content != messages[index].Content {
+			t.Fatalf("message %d changed unexpectedly", index)
+		}
+	}
+}
+
+func TestPruneSupersededReadResultsFindsMatchingBodyAcrossChangedRead(t *testing.T) {
+	bodyA := strings.Repeat("version a\n", 100)
+	bodyB := strings.Repeat("version b\n", 100)
+	messages := []zeroruntime.Message{
+		toolCallWithArgs("old-a", "read_file", `{"path":"a.go"}`),
+		{Role: zeroruntime.MessageRoleTool, ToolCallID: "old-a", Content: bodyA},
+		toolCallWithArgs("middle-b", "read_file", `{"path":"a.go"}`),
+		{Role: zeroruntime.MessageRoleTool, ToolCallID: "middle-b", Content: bodyB},
+		toolCallWithArgs("new-a", "read_file", `{"path":"a.go"}`),
+		{Role: zeroruntime.MessageRoleTool, ToolCallID: "new-a", Content: bodyA},
+	}
+
+	out, reclaimed := pruneSupersededReadResults(messages)
+
+	if reclaimed <= 0 || !strings.Contains(out[1].Content, "superseded identical") {
+		t.Fatalf("older matching body was not pruned across a changed result: reclaimed=%d output=%q", reclaimed, out[1].Content)
+	}
+	if out[3].Content != bodyB || out[5].Content != bodyA {
+		t.Fatal("changed middle result and newest matching result must remain exact")
+	}
+}

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

@@ -180,7 +180,7 @@ func TestInlineAdditionalPermissionsRequiresPromptUnderAutoAllow(t *testing.T) {
 		},
 	}
 	decision := &sandbox.Decision{Action: sandbox.ActionAllow, AutoAllowed: true}
-	if !shouldRequestPermission(tools.NewBashTool(t.TempDir()), args, false, decision) {
+	if !shouldRequestPermission(tools.NewScopedBashTool(t.TempDir(), nil), args, false, decision) {
 		t.Fatal("inline additional permissions must request approval even when a sandboxed shell would otherwise auto-allow")
 	}
 }