fix(tui): exclude permission waits from TTFT

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

.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
@@ -100,27 +130,18 @@ jobs:
           go-version-file: go.mod
           cache: true
 
-      # govulncheck, deadcode, and golangci-lint each resolve a toolchain from
-      # their own modules, which can drop below the version go.mod requires and
-      # then fail to load our packages. Pin GOTOOLCHAIN to the go.mod toolchain
-      # so all three run under it.
-      - name: Pin toolchain from go.mod
-        run: |
-          toolchain="$(awk '/^toolchain /{print $2}' go.mod)"
-          echo "GOTOOLCHAIN=${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
@@ -130,4 +151,4 @@ jobs:
       # the repo are cleaned up incrementally (see #527).
       - name: golangci-lint (advisory)
         continue-on-error: true
-        run: go run github.com/golangci/golangci-lint/v2/cmd/[email protected] run --enable-only unused,ineffassign,staticcheck ./...
+        run: make lint-static

.github/workflows/release-artifacts.yml

@@ -97,16 +97,21 @@ jobs:
           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 the npm wrapper (@gitlawb/zero) once every platform's release
-  # assets are live — the package's postinstall downloads its binary from the
-  # GitHub Release, so publishing before the assets exist would ship a package
-  # that fails to install.
+  # 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
@@ -156,10 +161,11 @@ jobs:
             console.log(`npm ${version} supports trusted publishing.`);
           '
 
-      # Probe the exact URLs postinstall will build (internal/release/release.go
-      # asset-name scheme) as an anonymous client. This also blocks publishing
-      # while the repo is still private — private release assets 404 without
-      # auth, exactly as they would for an installing user.
+      # 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: |
@@ -173,12 +179,64 @@ jobs:
             "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::${url} is not publicly downloadable — npm postinstall would fail. Is the repo public and are all assets uploaded?" >&2
+                echo "::error::${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: Publish to npm
-        run: npm publish --access public
+      - 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

.gitignore

@@ -55,3 +55,6 @@ docs/superpowers/
 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

@@ -1 +1 @@
-{".":"0.3.0"}
+{".":"0.6.0"}

AGENTS.md

@@ -1,326 +1,69 @@
----
-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:
-
-1. Tell the agent about *your* project (drop an `AGENTS.md` in your repo).
-2. Add new specialist sub-agents.
-3. Wire Zero into the rest of your toolchain (MCP, skills, hooks, plugins).
-
-## 1. Drop a project `AGENTS.md`
-
-When Zero starts in a directory, it looks for project-level instructions and injects them into the system prompt. The lookup walks from your current working directory **up to the nearest git root** and reads the first matching file at each level — general rules at the repo root, more specific rules in sub-trees. Files are labeled with their directory in the prompt (e.g. `## Project guidelines (services/api/AGENTS.md)`).
-
-Accepted file names, in priority order at each level:
-
-| Path | Notes |
-| --- | --- |
-| `./AGENTS.md` | The classic spot — committed to your repo, shared with the team. |
-| `./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` — 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.
-
-```markdown
-# Project conventions for <your project>
-
-- Build with `make`, not `go build` directly.
-- Tests live next to the source file (`foo_test.go` next to `foo.go`).
-- Run `make lint` before opening a PR.
-- Never edit files under `third_party/` — those are vendored.
-```
-
-Tips:
-
-- Keep each file under ~8 KiB. Zero caps the **total** across all matched files at 32 KiB; everything past the cap is dropped.
-- Re-state rules in the imperative voice: "Run `make lint`", not "you should consider running the linter".
-- Don't put secrets, model IDs, or environment-specific paths in `AGENTS.md`. Use `config.json` for those.
-- 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%\Roaming\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:
-
-| Scope | Path | Shared? |
-| --- | --- | --- |
-| Built-in | compiled into Zero | yes — `worker`, `explorer`, `code-review` |
-| User | `~/.config/zero/specialists/*.md` | no — your machine only |
-| Project | `./.zero/specialists/*.md` | yes — the repo team |
-
-Project overrides user overrides built-in when names collide.
-
-A specialist is a markdown manifest with frontmatter and a system prompt:
-
-```markdown
----
-description: Reviews API changes for breaking-change risk and missing tests.
-tools: read-only,plan
----
-
-You review API changes. For every changed hunk in `internal/api/` or any file
-that ends in `_api.go`:
-
-1. Confirm the public signature is backward-compatible, or note the breaking
-   change explicitly with the migration path.
-2. Confirm a corresponding test exists in `internal/api/*_test.go` and that
-   the new behaviour is exercised.
-3. Flag any new exported symbol without a doc comment.
-
-Reply with one JSON object per finding: `{"file", "line", "severity", "message", "fix"}`.
-```
-
-CLI management (the prompt is passed inline via `--prompt`):
-
-```bash
-zero specialist list
-zero specialist show api-reviewer
-zero specialist create api-reviewer \
-    --project \
-    --description "Reviews API changes" \
-    --tools read-only,plan \
-    --prompt "$(cat api-reviewer.md)"
-zero specialist edit api-reviewer --project
-zero specialist delete api-reviewer --project
-zero specialist path                       # prints the resolved specialists directory
-```
-
-The full format spec (frontmatter fields, tool scopes, prompt conventions) is in [`docs/SPECIALISTS.md`](docs/SPECIALISTS.md).
-
-> **Roadmap.** An in-UI specialist manager (create / edit / delete / preview) is on the backlog. Today you use the `zero specialist` CLI subcommands above.
-
-## 3. Skills
-
-Skills are markdown instruction packs the agent can pull in on demand. Each skill is a directory containing a `SKILL.md`. Skills are **user-level only** in this version — there's no project-scoped skill directory yet, so anything you want shared with the team goes in `AGENTS.md` (section 1) or as a hook (section 4).
-
-Discovery root: `$ZERO_SKILLS_DIR` → `$XDG_DATA_HOME/zero/skills` → `~/.local/share/zero/skills/`. A missing directory is fine — Zero just reports "no skills".
-
-```
-~/.local/share/zero/skills/
-  run-benchmarks/
-    SKILL.md
-  write-changelog/
-    SKILL.md
-```
-
-`SKILL.md` format:
-
-```markdown
----
-description: Run the project's benchmark suite and summarize the deltas.
----
-
-# Run benchmarks
-
-1. `make bench` — captures the wall-clock and RSS before and after.
-2. `benchstat before.txt after.txt` — diffs the two.
-3. Report any regression > 5% with the function name and the previous value.
-```
-
-Only `name` and `description` are recognized in the frontmatter today. The `name` defaults to the directory name. Within a single skills root, duplicate names are resolved by lexicographic directory order. During an agent run, Zero loads the default skills directory before plugin skill roots; earlier roots win name collisions silently. Plugin-declared skills (section 6) are merged into the active agent run at plugin activation time, so bundled skills appear in the available skills list and can be loaded with the `skill` tool.
-
-The `skill` core tool lets the agent load any discovered skill by name.
-
-## 4. Hooks
-
-Hooks fire shell commands on lifecycle events. Configure them in JSON:
-
-- User: `~/.config/zero/hooks.json`
-- Project: `./.zero/hooks.json`
-
-```json
-{
-  "enabled": true,
-  "hooks": [
-    {
-      "id": "block-rm-rf",
-      "event": "beforeTool",
-      "matcher": "bash",
-      "command": "/usr/local/bin/zero-hook-block-rmrf.sh",
-      "enabled": true
-    },
-    {
-      "id": "log-session",
-      "event": "sessionStart",
-      "command": "/usr/local/bin/zero-hook-log.sh",
-      "enabled": true
-    }
-  ]
-}
-```
-
-The `args` array (when present) is passed verbatim to `exec.CommandContext`. The actual hook payload — event name, matcher, tool call id, tool name, tool input, tool output, status — is delivered to the command as **JSON on stdin**, not via `${...}` substitution. A typical handler reads stdin and decides what to do:
-
-```bash
-#!/usr/bin/env bash
-# /usr/local/bin/zero-hook-block-rmrf.sh
-set -euo pipefail
-payload="$(cat)"
-if printf '%s' "$payload" | grep -q '"input":"[^"]*rm[[:space:]]+-rf'; then
-  echo "refusing rm -rf" >&2
-  exit 1
-fi
-```
-
-Events the agent emits (in dispatch order):
-
-| Event | Fires when | Matcher allowed? |
-| --- | --- | --- |
-| `beforeTool` | A tool is about to run | yes (tool name) |
-| `afterTool` | A tool just returned | yes (tool name) |
-| `sessionStart` | A session begins | no |
-| `sessionEnd` | A session ends | no |
-| `specialistStart` | A sub-agent is spawned | yes (specialist name) |
-| `specialistStop` | A sub-agent ends | yes (specialist name) |
-
-A hook's exit code decides what happens next: `0` continues, non-zero blocks the tool call (`beforeTool`) or surfaces an error (`afterTool`). Hook execution is recorded in the audit log; the audit is reachable from the agent's view of past actions, not from a dedicated `zero doctor` check.
-
-> **Roadmap.** An in-UI hooks manager is on the backlog. Today you edit the JSON directly.
-
-## 5. MCP — Model Context Protocol
-
-Zero is both an **MCP client** (it can call external MCP servers) and an **MCP server** (other agents can call its tools).
-
-### As a client — configure MCP servers in `config.json`
-
-```json
-{
-  "mcp": {
-    "servers": {
-      "docs": {
-        "type": "stdio",
-        "command": "docs-mcp",
-        "args": ["--port", "7777"]
-      },
-      "github": {
-        "type": "http",
-        "url": "https://api.example.com/mcp",
-        "headers": { "Authorization": "Bearer YOUR_TOKEN_HERE" }
-      }
-    }
-  }
-}
-```
-
-Manage via CLI:
-
-```bash
-zero mcp add docs --type stdio -- docs-mcp --port 7777
-zero mcp add github --type http --url https://api.example.com/mcp \
-    --header "Authorization=Bearer YOUR_TOKEN_HERE"
-zero mcp list
-zero mcp check docs
-zero mcp remove github
-zero mcp oauth login github
-```
-
-Servers are merged from user and project configs (project wins on conflicts). Token-bearing values in `config.json` are sent verbatim — there is no `${env:...}` expansion — so prefer one of:
-
-- A wrapper script that sources the secret and execs the real command.
-- 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).
-
-### As a server — expose Zero's tools to another agent
-
-```bash
-zero serve --mcp
-```
-
-The server speaks MCP over stdio. Configure it from the receiving side as a `stdio` server whose command is `zero serve --mcp`.
-
-## 6. Plugins
-
-A plugin is a self-contained directory that bundles tools, hooks, and skills for one capability. Plugins live at:
-
-- User: `~/.config/zero/plugins/<id>/`
-- Project: `./.zero/plugins/<id>/`
-
-Each plugin has a `plugin.json` manifest:
-
-```json
-{
-  "id": "github-pr-review",
-  "name": "GitHub PR Review",
-  "description": "Adds review tools for GitHub PRs.",
-  "version": "1.0.0",
-  "tools": [
-    { "name": "list_prs", "command": "./tools/list_prs.sh" }
-  ],
-  "hooks": [
-    { "name": "pre-merge-check", "event": "beforeTool", "command": "./hooks/pre-merge.sh" }
-  ],
-  "skills": [
-    { "path": "./skills/review-checklist/SKILL.md" }
-  ]
-}
-```
-
-Install and manage:
-
-```bash
-zero plugins add ./github-pr-review      # copy into ~/.config/zero/plugins/ or ./.zero/plugins/
-zero plugins list
-zero plugins remove github-pr-review    # alias: rm
-```
-
-A plugin is enabled by being present in the plugins directory and disabled by removing it (or by the user setting `"enabled": false` in its `plugin.json`). Plugins are not enabled or disabled by a CLI subcommand today.
-
-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.
-
-## 7. Configuration locations
-
-Three layers, applied in order (later layers override earlier ones):
-
-| Layer | Path | 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. |
-
-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).
-
-The sandbox `additionalWriteRoots` key is **ignored in project config** by design — a checked-out repo cannot widen its own sandbox. Set it in the user config or pass `--add-dir` per-invocation.
-
-## 8. End-to-end example
-
-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/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.
+# 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.

CHANGELOG.md

@@ -6,6 +6,138 @@ 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`.
 
+## [0.6.0](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)
 
 
@@ -127,14 +259,19 @@ tagged. Until then, source builds report the version `dev`.
 ## [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.
 - Interactive `/theme` picker: bare `/theme` opens a popup that live-previews each palette as you move
   and applies on select (Esc reverts).
-- Ten built-in color themes alongside the `dark`/`light` built-ins — `dracula`, `nord`, `gruvbox`,
-  `tokyo-night`, `catppuccin`, `one-dark`, `solarized-dark`, `rose-pine`, `everforest`, and
-  `solarized-light` — selectable via `/theme <name>`, `--theme <name>`, or `ZERO_THEME`. Every palette
-  is contrast-audited to WCAG AA. The built-in light theme was reworked for legibility.
+- 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`,

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

@@ -6,7 +6,7 @@
 
 <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 1.26.5+" src="https://img.shields.io/badge/Go-1.26.5+-00ADD8?logo=go&logoColor=white">
   <img alt="25+ providers" src="https://img.shields.io/badge/providers-25+-34E2EA">
   <br>
   <strong>English</strong> | <a href="README_ZH.md">中文</a>
@@ -49,34 +49,15 @@ 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.
-
-### Bun
-
-Bun does not run dependency lifecycle scripts by default, so the `postinstall`
-that fetches the Zero binary is skipped and the first run fails with
-`No native binary found next to the npm wrapper`.
-
-The simplest fix is to trust the package after installing, which runs the
-blocked postinstall. This works for project and global installs:
-
-```bash
-# project install
-bun add @gitlawb/zero
-bun pm trust @gitlawb/zero
-
-# global install
-bun add -g @gitlawb/zero
-bun pm -g trust @gitlawb/zero
-```
-
-Alternatives: allow the postinstall up front by adding
-`"trustedDependencies": ["@gitlawb/zero"]` to your project's package.json
-before `bun add`, or run the installer manually
-(`node node_modules/@gitlawb/zero/scripts/postinstall.mjs`) on Bun versions
-that do not have `bun pm trust`.
+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
 
@@ -94,7 +75,7 @@ irm https://raw.githubusercontent.com/Gitlawb/zero/main/scripts/install.ps1 | ie
 
 ### From source
 
-Source builds require Go 1.25+.
+Source builds require Go 1.26.5+.
 
 ```bash
 git clone https://github.com/Gitlawb/zero.git
@@ -148,7 +129,17 @@ 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:
@@ -157,6 +148,39 @@ To configure Meituan LongCat (LongCat-2.0) directly, run:
 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
 `zero providers detect`.
 
@@ -174,9 +198,11 @@ 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:
 
@@ -186,6 +212,7 @@ 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 |
+| `/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 |
@@ -322,7 +349,7 @@ manifest.
 | Control | Effect |
 |---|---|
 | `NO_COLOR=<anything>` | disables color output |
-| `ZERO_THEME=<name>` | selects the startup theme (`auto`, `dark`, `light`, or a color theme like `dracula`, `nord`, `gruvbox`, `tokyo-night`, `catppuccin`, `one-dark`, `solarized-dark`, `rose-pine`, `everforest`, `solarized-light`) |
+| `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 |
@@ -339,7 +366,35 @@ 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
@@ -350,6 +405,7 @@ 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)
@@ -357,6 +413,24 @@ go run ./cmd/zero-release build --goos windows --goarch amd64 --output dist/zero
 - [Performance](docs/PERFORMANCE.md)
 - [Agent evals](docs/AGENT_EVALS.md)
 
+## Community
+
+Questions, setup help, ideas, and sharing all live in
+[GitHub Discussions](https://github.com/Gitlawb/zero/discussions):
+
+| Category | Use it for |
+|---|---|
+| [Q&A](https://github.com/Gitlawb/zero/discussions/categories/q-a) | Setup help, provider/model configuration, "how do I" questions |
+| [Ideas](https://github.com/Gitlawb/zero/discussions/categories/ideas) | Feature proposals and design discussion before any PR |
+| [Show and tell](https://github.com/Gitlawb/zero/discussions/categories/show-and-tell) | Your skills, plugins, MCP setups, themes, and workflows |
+| [Announcements](https://github.com/Gitlawb/zero/discussions/categories/announcements) | Releases and project news from the maintainers |
+
+For a good Q&A answer fast, include `zero --version`, your OS and install
+method, the provider/model in use, and `zero doctor` output. See
+[SUPPORT.md](SUPPORT.md). Bugs belong in
+[issues](https://github.com/Gitlawb/zero/issues/new/choose); security reports
+follow [SECURITY.md](SECURITY.md), never a public thread.
+
 ## Contributing
 
 Contributions are welcome. Read [CONTRIBUTING.md](CONTRIBUTING.md), run the

README_ZH.md

@@ -6,7 +6,7 @@
 
 <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 1.26.5+" src="https://img.shields.io/badge/Go-1.26.5+-00ADD8?logo=go&logoColor=white">
   <img alt="25+ providers" src="https://img.shields.io/badge/providers-25+-34E2EA">
   <br>
   <a href="README.md">English</a> | <strong>中文</strong>
@@ -38,25 +38,7 @@ npm install -g @gitlawb/zero
 zero
 ```
 
-npm 包安装一个小型包装器以及与你平台匹配的 Zero 二进制文件（从 GitHub Releases 获取）。支持 Linux、macOS 和 Windows 的 x64 和 arm64。
-
-### Bun
-
-Bun 默认不运行依赖的生命周期脚本，因此获取 Zero 二进制文件的 `postinstall` 会被跳过，首次运行会失败并显示 `No native binary found next to the npm wrapper`。
-
-最简单的解决方法是安装后信任该包，这会运行被阻止的 postinstall。项目安装和全局安装均适用：
-
-```bash
-# 项目安装
-bun add @gitlawb/zero
-bun pm trust @gitlawb/zero
-
-# 全局安装
-bun add -g @gitlawb/zero
-bun pm -g trust @gitlawb/zero
-```
-
-其他方式：在 `bun add` 之前将 `"trustedDependencies": ["@gitlawb/zero"]` 添加到项目的 package.json 以提前允许 postinstall；或在不支持 `bun pm trust` 的旧版 Bun 上手动运行安装程序（`node node_modules/@gitlawb/zero/scripts/postinstall.mjs`）。
+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
 
@@ -74,7 +56,7 @@ irm https://raw.githubusercontent.com/Gitlawb/zero/main/scripts/install.ps1 | ie
 
 ### 从源码构建
 
-源码构建需要 Go 1.25+。
+源码构建需要 Go 1.26.5+。
 
 ```bash
 git clone https://github.com/Gitlawb/zero.git
@@ -252,7 +234,7 @@ zero update           检查更新版本
 | 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`、`solarized-light`） |
+| `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 |
@@ -279,6 +261,7 @@ go run ./cmd/zero-release build --goos windows --goarch amd64 --output dist/zero
 
 - [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)

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).

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']) {
@@ -49,38 +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)) {
-  const postinstallScript = join(packageRoot, 'scripts', 'postinstall.mjs');
-  const ranByBun = process.execPath.includes('bun') || !!process.versions?.bun;
-  console.error(
-    '[zero] No native binary found next to the npm wrapper.\n' +
-      'The platform binary is fetched at install time by a postinstall script,\n' +
-      'which did not run (or was skipped) for this install.\n' +
-      '\n' +
-      'Fix it now by running the installer manually:\n' +
-      `  node "${postinstallScript}"\n` +
-      '\n' +
-      (ranByBun
-        ? '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'
-        : '') +
-      'If that fails, build from source: https://github.com/Gitlawb/zero\n' +
-      '(go run ./cmd/zero, requires Go 1.25+).',
+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

@@ -0,0 +1,341 @@
+# 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:
+
+1. Tell the agent about *your* project (drop an `AGENTS.md` in your repo).
+2. Add new specialist sub-agents.
+3. Wire Zero into the rest of your toolchain (MCP, skills, hooks, plugins).
+
+## 1. Drop a project `AGENTS.md`
+
+When Zero starts in a directory, it looks for project-level instructions and injects them into the system prompt. The lookup walks from your current working directory **up to the nearest git root** and reads the first matching file at each level — general rules at the repo root, more specific rules in sub-trees. Files are labeled with their directory in the prompt (e.g. `## Project guidelines (services/api/AGENTS.md)`).
+
+Accepted file names, in priority order at each level:
+
+| Path | Notes |
+| --- | --- |
+| `./AGENTS.md` | The classic spot — committed to your repo, shared with the team. |
+| `./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` — 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.
+
+```markdown
+# Project conventions for <your project>
+
+- Build with `make`, not `go build` directly.
+- Tests live next to the source file (`foo_test.go` next to `foo.go`).
+- Run `make lint` before opening a PR.
+- Never edit files under `third_party/` — those are vendored.
+```
+
+Tips:
+
+- Keep each file under ~8 KiB. Zero caps the **total** across all matched files at 32 KiB; everything past the cap is dropped.
+- Re-state rules in the imperative voice: "Run `make lint`", not "you should consider running the linter".
+- Don't put secrets, model IDs, or environment-specific paths in `AGENTS.md`. Use `config.json` for those.
+- 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:
+
+| Scope | Path | Shared? |
+| --- | --- | --- |
+| Built-in | compiled into Zero | yes — `worker`, `explorer`, `code-review` |
+| User | `~/.config/zero/specialists/*.md` | no — your machine only |
+| Project | `./.zero/specialists/*.md` | yes — the repo team |
+
+Project overrides user overrides built-in when names collide.
+
+A specialist is a markdown manifest with frontmatter and a system prompt:
+
+```markdown
+---
+description: Reviews API changes for breaking-change risk and missing tests.
+tools: read-only,plan
+---
+
+You review API changes. For every changed hunk in `internal/api/` or any file
+that ends in `_api.go`:
+
+1. Confirm the public signature is backward-compatible, or note the breaking
+   change explicitly with the migration path.
+2. Confirm a corresponding test exists in `internal/api/*_test.go` and that
+   the new behaviour is exercised.
+3. Flag any new exported symbol without a doc comment.
+
+Reply with one JSON object per finding: `{"file", "line", "severity", "message", "fix"}`.
+```
+
+CLI management (the prompt is passed inline via `--prompt`):
+
+```bash
+zero specialist list
+zero specialist show api-reviewer
+zero specialist create api-reviewer \
+    --project \
+    --description "Reviews API changes" \
+    --tools read-only,plan \
+    --prompt "$(cat api-reviewer.md)"
+zero specialist edit api-reviewer --project
+zero specialist delete api-reviewer --project
+zero specialist path                       # prints the resolved specialists directory
+```
+
+The full format spec (frontmatter fields, tool scopes, prompt conventions) is in [`docs/SPECIALISTS.md`](SPECIALISTS.md).
+
+> **Roadmap.** An in-UI specialist manager (create / edit / delete / preview) is on the backlog. Today you use the `zero specialist` CLI subcommands above.
+
+## 3. Skills
+
+Skills are markdown instruction packs the agent can pull in on demand. Each skill is a directory containing a `SKILL.md`. Standalone project skill directories are not supported in this version (shared project-wide skills must go in `AGENTS.md` or as a hook). However, project plugins (section 6) may bundle skills, which are merged into the active run.
+
+Discovery roots (earlier wins on name collisions):
+
+1. **Primary Zero dir** — `$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:
+
+```markdown
+---
+description: Run the project's benchmark suite and summarize the deltas.
+---
+
+# Run benchmarks
+
+1. `make bench` — captures the wall-clock and RSS before and after.
+2. `benchstat before.txt after.txt` — diffs the two.
+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. 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` tool lets the agent load any discovered skill by name (primary, agents, or plugin).
+
+## 4. Hooks
+
+Hooks fire shell commands on lifecycle events. Configure them in JSON:
+
+- User: `~/.config/zero/hooks.json`
+- Project: `./.zero/hooks.json`
+
+```json
+{
+  "enabled": true,
+  "hooks": [
+    {
+      "id": "block-rm-rf",
+      "event": "beforeTool",
+      "matcher": "bash",
+      "command": "/usr/local/bin/zero-hook-block-rmrf.sh",
+      "enabled": true
+    },
+    {
+      "id": "log-session",
+      "event": "sessionStart",
+      "command": "/usr/local/bin/zero-hook-log.sh",
+      "enabled": true
+    }
+  ]
+}
+```
+
+The `args` array (when present) is passed verbatim to `exec.CommandContext`. The actual hook payload — event name, matcher, tool call id, tool name, tool input, tool output, status — is delivered to the command as **JSON on stdin**, not via `${...}` substitution. A typical handler reads stdin and decides what to do:
+
+```bash
+#!/usr/bin/env bash
+# /usr/local/bin/zero-hook-block-rmrf.sh
+set -euo pipefail
+payload="$(cat)"
+if printf '%s' "$payload" | grep -q '"input":"[^"]*rm[[:space:]]+-rf'; then
+  echo "refusing rm -rf" >&2
+  exit 1
+fi
+```
+
+Events the agent emits (in dispatch order):
+
+| Event | Fires when | Matcher allowed? |
+| --- | --- | --- |
+| `beforeTool` | A tool is about to run | yes (tool name) |
+| `afterTool` | A tool just returned | yes (tool name) |
+| `sessionStart` | A session begins | no |
+| `sessionEnd` | A session ends | no |
+| `specialistStart` | A sub-agent is spawned | yes (specialist name) |
+| `specialistStop` | A sub-agent ends | yes (specialist name) |
+
+A hook's exit code decides what happens next: `0` continues, non-zero blocks the tool call (`beforeTool`) or surfaces an error (`afterTool`). Hook execution is recorded in the audit log; the audit is reachable from the agent's view of past actions, not from a dedicated `zero doctor` check.
+
+> **Roadmap.** An in-UI hooks manager is on the backlog. Today you edit the JSON directly.
+
+## 5. MCP — Model Context Protocol
+
+Zero is both an **MCP client** (it can call external MCP servers) and an **MCP server** (other agents can call its tools).
+
+### As a client — configure MCP servers in `config.json`
+
+```json
+{
+  "mcp": {
+    "servers": {
+      "docs": {
+        "type": "stdio",
+        "command": "docs-mcp",
+        "args": ["--port", "7777"]
+      },
+      "github": {
+        "type": "http",
+        "url": "https://api.example.com/mcp",
+        "headers": { "Authorization": "Bearer YOUR_TOKEN_HERE" }
+      }
+    }
+  }
+}
+```
+
+Manage via CLI:
+
+```bash
+zero mcp add docs --type stdio -- docs-mcp --port 7777
+zero mcp add github --type http --url https://api.example.com/mcp \
+    --header "Authorization=Bearer YOUR_TOKEN_HERE"
+zero mcp list
+zero mcp check docs
+zero mcp remove github
+zero mcp oauth login github
+```
+
+Servers are merged from user and project configs (project wins on conflicts). Token-bearing values in `config.json` are sent verbatim — there is no `${env:...}` expansion — so prefer one of:
+
+- A wrapper script that sources the secret and execs the real command.
+- 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:
+
+- User: `~/.config/zero/plugins/<id>/`
+- Project: `./.zero/plugins/<id>/`
+
+Each plugin has a `plugin.json` manifest:
+
+```json
+{
+  "id": "github-pr-review",
+  "name": "GitHub PR Review",
+  "description": "Adds review tools for GitHub PRs.",
+  "version": "1.0.0",
+  "tools": [
+    { "name": "list_prs", "command": "./tools/list_prs.sh" }
+  ],
+  "hooks": [
+    { "name": "pre-merge-check", "event": "beforeTool", "command": "./hooks/pre-merge.sh" }
+  ],
+  "skills": [
+    { "path": "./skills/review-checklist/SKILL.md" }
+  ]
+}
+```
+
+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
+```
+
+A plugin is enabled by being present in the plugins directory and disabled by removing it (or by the user setting `"enabled": false` in its `plugin.json`). Plugins are not enabled or disabled by a CLI subcommand today.
+
+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.
+
+## 7. Configuration locations
+
+Five configuration sources, in precedence order (later sources override earlier ones):
+
+| 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). |
+| 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).
+
+The sandbox `additionalWriteRoots` key is **ignored in project config** by design — a checked-out repo cannot widen its own sandbox. Set it in the user config or pass `--add-dir` per-invocation.
+
+## 8. End-to-end example
+
+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` 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](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/HOW_ZERO_WORKS.md

@@ -272,7 +272,8 @@ sequenceDiagram
         Agent->>Tool: execute(ctx, args)
         Tool-->>Agent: output, error, metadata
         Agent->>Hooks: afterTool hooks
-        Agent->>Agent: redact, truncate, classify success/failure
+        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
@@ -604,10 +605,22 @@ flowchart TD
     Prompt -- Yes --> Decision[Permission callback]
     Prompt -- No --> Run[Registry.RunWithOptions]
     Decision --> Run
-    Run --> Redact[Redact secrets + enforce output ceiling]
-    Redact --> Message[Return tool result to model]
+    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,
@@ -733,9 +746,10 @@ flowchart TD
 - **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 merges plugin skill
-  roots into the runtime skill tool and prompt-visible skill list; the base
-  `internal/skills` scanner still scans one root at a time.
+- **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

docs/INSTALL.md

@@ -4,10 +4,11 @@ Zero is distributed as:
 
 - an npm package, `@gitlawb/zero`
 - release archives on GitHub Releases
-- source builds with Go 1.25+
+- source builds with Go 1.26.5+
 
-The npm package and install scripts download a platform-specific release archive.
-They require a published GitHub Release for the requested version.
+The install scripts download a platform-specific release archive and require a
+published GitHub Release for the requested version. The npm package is
+self-contained: the platform binary installs from the npm registry.
 
 ## npm
 
@@ -16,59 +17,32 @@ 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
 
-## Bun
-
-Bun is "default-secure" and does not run lifecycle scripts of installed
-dependencies (only the installing project's own scripts), so the `postinstall`
-that fetches the Zero binary is silently skipped. The first run then fails with
-`No native binary found next to the npm wrapper`.
-
-The simplest fix is to trust the package after installing, which runs the
-blocked postinstall. This works for project and global installs:
+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
-# project install
-bun add @gitlawb/zero
-bun pm trust @gitlawb/zero
-
-# global install
-bun add -g @gitlawb/zero
-bun pm -g trust @gitlawb/zero
+node "$(npm root -g)/@gitlawb/zero/scripts/postinstall.mjs"
 ```
 
-`bun pm untrusted` (or `bun pm -g untrusted`) lists the blocked postinstalls if
-you want to inspect before trusting.
-
-Alternatively, allow the postinstall to run at install time by adding the
-package to your project's `trustedDependencies` before installing:
-
-```json
-{
-  "trustedDependencies": ["@gitlawb/zero"]
-}
-```
-
-```bash
-bun add @gitlawb/zero
-```
-
-On Bun versions that do not have `bun pm trust`, run the installer manually
-after installing:
-
-```bash
-node node_modules/@gitlawb/zero/scripts/postinstall.mjs
-```
-
-Reference: <https://bun.sh/docs/pm/lifecycle>
-
 ## Linux And macOS Script
 
 Install the latest release:
@@ -151,7 +125,7 @@ Build a local binary:
 go build -o zero ./cmd/zero
 ```
 
-Source builds require Go 1.25+.
+Source builds require Go 1.26.5+.
 
 ### Sandbox Helpers For Source Builds
 
@@ -252,8 +226,7 @@ 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

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/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.

go.mod

@@ -3,9 +3,9 @@ module github.com/Gitlawb/zero
 go 1.26.5
 
 require (
-	charm.land/bubbles/v2 v2.1.0
-	charm.land/bubbletea/v2 v2.0.7
-	charm.land/lipgloss/v2 v2.0.4
+	charm.land/bubbles/v2 v2.1.1
+	charm.land/bubbletea/v2 v2.0.8
+	charm.land/lipgloss/v2 v2.0.5
 	github.com/alecthomas/chroma/v2 v2.27.0
 	github.com/atotto/clipboard v0.1.4
 	github.com/aymanbagabas/go-udiff v0.4.1
@@ -14,18 +14,18 @@ require (
 	github.com/charmbracelet/x/term v0.2.2
 	github.com/coder/websocket v1.8.15
 	github.com/ledongthuc/pdf v0.0.0-20250511090121-5959a4027728
-	golang.org/x/sys v0.46.0
+	golang.org/x/sys v0.47.0
 	mvdan.cc/sh/v3 v3.13.1
 )
 
 require (
-	github.com/charmbracelet/ultraviolet v0.0.0-20260615092913-2399af76d5b1 // indirect
+	github.com/charmbracelet/ultraviolet v0.0.0-20260703014108-f5a850f9c2b7 // indirect
 	github.com/charmbracelet/x/exp/golden v0.0.0-20260615092313-b57e5e6d29bb // indirect
 	github.com/charmbracelet/x/termios v0.1.1 // indirect
 	github.com/charmbracelet/x/windows v0.2.2 // indirect
 	github.com/clipperhouse/displaywidth v0.11.0 // indirect
 	github.com/clipperhouse/uax29/v2 v2.7.0 // indirect
-	github.com/dlclark/regexp2/v2 v2.2.1 // indirect
+	github.com/dlclark/regexp2/v2 v2.5.0 // indirect
 	github.com/go-quicktest/qt v1.102.0 // indirect
 	github.com/lucasb-eyer/go-colorful v1.4.0 // indirect
 	github.com/mattn/go-runewidth v0.0.24 // indirect
@@ -34,5 +34,5 @@ require (
 	github.com/rogpeppe/go-internal v1.15.0 // indirect
 	github.com/xo/terminfo v0.0.0-20220910002029-abceb7e1c41e // indirect
 	golang.org/x/exp v0.0.0-20260611194520-c48552f49976 // indirect
-	golang.org/x/sync v0.21.0 // indirect
+	golang.org/x/sync v0.22.0 // indirect
 )

go.sum

@@ -1,9 +1,9 @@
-charm.land/bubbles/v2 v2.1.0 h1:YSnNh5cPYlYjPxRrzs5VEn3vwhtEn3jVGRBT3M7/I0g=
-charm.land/bubbles/v2 v2.1.0/go.mod h1:l97h4hym2hvWBVfmJDtrEHHCtkIKeTEb3TTJ4ZOB3wY=
-charm.land/bubbletea/v2 v2.0.7 h1:7qw2tTAVar7m7klOPBYfTB0mniv/RuexsYwMRNxSeL0=
-charm.land/bubbletea/v2 v2.0.7/go.mod h1:DGW2q8gvzHnOpMpZTORs0aySVHCox5C+2Svk0fci1qs=
-charm.land/lipgloss/v2 v2.0.4 h1:lcPeVtcp23SNra7lHy8iYE4UC2aIipVQ47sbGyyxR5Q=
-charm.land/lipgloss/v2 v2.0.4/go.mod h1:0653x8epbZSzdDfO/XPS1a/uYPOBeSsCssOpJOqDzik=
+charm.land/bubbles/v2 v2.1.1 h1:7r55WzBxpo/R3z98hGmY7KKPd3ET6vsf0Fb9sDHOV60=
+charm.land/bubbles/v2 v2.1.1/go.mod h1:GE6M31gaWZVXzGw73OeuTTgy4lX+OtkH0E5ymnNsHxo=
+charm.land/bubbletea/v2 v2.0.8 h1:SxTJMhCAI3lbPmy4SgX5LWZ24AdINr4I6UEqzZvYJuY=
+charm.land/bubbletea/v2 v2.0.8/go.mod h1:2SkdgoTXluXJHOUwAoRlRXF/28vklb1rFl6GcgV1/ss=
+charm.land/lipgloss/v2 v2.0.5 h1:kbNxgeeUOYv5J0YdpxFjfvf3dFvqH8Aci4zB6xqFtrY=
+charm.land/lipgloss/v2 v2.0.5/go.mod h1:9oqhxt4yxIMe6q5A4kHr44DremZk7J9UNh74GlWa5nc=
 github.com/alecthomas/assert/v2 v2.11.0 h1:2Q9r3ki8+JYXvGsDyBXwH3LcJ+WK5D0gc5E8vS6K3D0=
 github.com/alecthomas/assert/v2 v2.11.0/go.mod h1:Bze95FyfUr7x34QZrjL+XP+0qgp/zg8yS+TtBj1WA3k=
 github.com/alecthomas/chroma/v2 v2.27.0 h1:FodwmyOBgJULFYmDqibcp9pvfDLWdtPRh9v/r5BXYZs=
@@ -16,8 +16,8 @@ github.com/aymanbagabas/go-udiff v0.4.1 h1:OEIrQ8maEeDBXQDoGCbbTTXYJMYRCRO1fnodZ
 github.com/aymanbagabas/go-udiff v0.4.1/go.mod h1:0L9PGwj20lrtmEMeyw4WKJ/TMyDtvAoK9bf2u/mNo3w=
 github.com/charmbracelet/colorprofile v0.4.3 h1:QPa1IWkYI+AOB+fE+mg/5/4HRMZcaXex9t5KX76i20Q=
 github.com/charmbracelet/colorprofile v0.4.3/go.mod h1:/zT4BhpD5aGFpqQQqw7a+VtHCzu+zrQtt1zhMt9mR4Q=
-github.com/charmbracelet/ultraviolet v0.0.0-20260615092913-2399af76d5b1 h1:4+r3uOJ69ueRBt4okgEfWZeXs3BD36HcDBmOIAUlETk=
-github.com/charmbracelet/ultraviolet v0.0.0-20260615092913-2399af76d5b1/go.mod h1:f/jRa757WUmaOZrbPspXymbg/GnbF+rwe4OLsG7aXYo=
+github.com/charmbracelet/ultraviolet v0.0.0-20260703014108-f5a850f9c2b7 h1:3FmWoGNWK4STvqg0O0Aeav2T7rodWJAPeF0QpH+8gFw=
+github.com/charmbracelet/ultraviolet v0.0.0-20260703014108-f5a850f9c2b7/go.mod h1:f/jRa757WUmaOZrbPspXymbg/GnbF+rwe4OLsG7aXYo=
 github.com/charmbracelet/x/ansi v0.11.7 h1:kzv1kJvjg2S3r9KHo8hDdHFQLEqn4RBCb39dAYC84jI=
 github.com/charmbracelet/x/ansi v0.11.7/go.mod h1:9qGpnAVYz+8ACONkZBUWPtL7lulP9No6p1epAihUZwQ=
 github.com/charmbracelet/x/exp/golden v0.0.0-20260615092313-b57e5e6d29bb h1:m1Uub9uM7meZ8ofVyeRfK7M/knCi3sF2gKn9kldVW0I=
@@ -34,8 +34,8 @@ github.com/clipperhouse/uax29/v2 v2.7.0 h1:+gs4oBZ2gPfVrKPthwbMzWZDaAFPGYK72F0NJ
 github.com/clipperhouse/uax29/v2 v2.7.0/go.mod h1:EFJ2TJMRUaplDxHKj1qAEhCtQPW2tJSwu5BF98AuoVM=
 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 v2.2.1 h1:mf4KkFUj0gJuarK8P+LgiS+Lit7m9N1yAwEfPbee7R0=
-github.com/dlclark/regexp2/v2 v2.2.1/go.mod h1:avUrQvPaLz2DrFNHJF0taWAFFX2C1GMSSoeiqFjcBmU=
+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=
@@ -62,9 +62,9 @@ github.com/xo/terminfo v0.0.0-20220910002029-abceb7e1c41e h1:JVG44RsyaB9T2KIHavM
 github.com/xo/terminfo v0.0.0-20220910002029-abceb7e1c41e/go.mod h1:RbqR21r5mrJuqunuUZ/Dhy/avygyECGrLceyNeo4LiM=
 golang.org/x/exp v0.0.0-20260611194520-c48552f49976 h1:X8Hz2ImujgbmetVuW+w2YkyZChE3cBpZi2P158rTG9M=
 golang.org/x/exp v0.0.0-20260611194520-c48552f49976/go.mod h1:vnf4pv9iKZXY58sQE1L86zmNWJ4159e1RkcWiLCkeEY=
-golang.org/x/sync v0.21.0 h1:HLII4xRRTtCRkxYp4HNFF0Js/Og6q2i++KXbg0gHCwM=
-golang.org/x/sync v0.21.0/go.mod h1:9xrNwdLfx4jkKbNva9FpL6vEN7evnE43NNNJQ2LF3+0=
-golang.org/x/sys v0.46.0 h1:noSf2Fq6F8DBgS+LysIkx7rIExoNHJsxOAtPp4rthXw=
-golang.org/x/sys v0.46.0/go.mod h1:4GL1E5IUh+htKOUEOaiffhrAeqysfVGipDYzABqnCmw=
+golang.org/x/sync v0.22.0 h1:SZjpbeLmrCk4xhRSZFNZW5gFUeCeFgjekvI/+gfScek=
+golang.org/x/sync v0.22.0/go.mod h1:9xrNwdLfx4jkKbNva9FpL6vEN7evnE43NNNJQ2LF3+0=
+golang.org/x/sys v0.47.0 h1:o7XGOvZQCADBQQ4Y7VNq2dRWQR7JmOUW8Kxx4ZsNgWs=
+golang.org/x/sys v0.47.0/go.mod h1:4GL1E5IUh+htKOUEOaiffhrAeqysfVGipDYzABqnCmw=
 mvdan.cc/sh/v3 v3.13.1 h1:DP3TfgZhDkT7lerUdnp6PTGKyxxzz6T+cOlY/xEvfWk=
 mvdan.cc/sh/v3 v3.13.1/go.mod h1:lXJ8SexMvEVcHCoDvAGLZgFJ9Wsm2sulmoNEXGhYZD0=

internal/acp/agent.go

@@ -11,6 +11,9 @@ import (
 
 	"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"
@@ -23,9 +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.
@@ -60,13 +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.
@@ -83,9 +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) {
@@ -120,7 +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")
@@ -129,18 +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")
@@ -161,7 +170,17 @@ func (a *Agent) handleSessionLoad(_ context.Context, params json.RawMessage) (an
 	// a half-initialized session (registerSession sets history under the lock and
 	// reuses an already-live session rather than orphaning its in-flight turn).
 	history, historyErr := a.loadHistory(meta.SessionID)
-	sess := a.registerSession(meta.SessionID, root, history)
+	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",
@@ -169,7 +188,8 @@ func (a *Agent) handleSessionLoad(_ context.Context, params json.RawMessage) (an
 		historyErr,
 	)
 	return LoadSessionResult{
-		Modes: a.modeState(sess),
+		ConfigOptions: a.configOptions(sess),
+		Modes:         a.modeState(sess),
 	}, nil
 }
 
@@ -354,11 +374,27 @@ 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:
+		mode := agent.PermissionMode(p.Value)
+		switch mode {
+		case agent.PermissionModeAuto, agent.PermissionModeAsk:
+			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) {
@@ -370,8 +406,24 @@ 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) {
@@ -398,6 +450,76 @@ func (a *Agent) modeState(s *acpSession) *SessionModeState {
 	}
 }
 
+// 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)
+				}
+			}
+		}
+	}
+	return selected, options, restrictModels, nil
+}
+
+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."},
+		},
+	}}
+}
+
 // ---- persistence + continuity ----
 
 func (a *Agent) persistTurn(sess *acpSession, user, assistant string) error {
@@ -525,13 +647,13 @@ 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
 }
@@ -570,7 +692,18 @@ 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()
 }
@@ -581,6 +714,24 @@ 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

@@ -3,6 +3,7 @@ package acp
 import (
 	"context"
 	"encoding/json"
+	"errors"
 	"fmt"
 	"io"
 	"os"
@@ -13,6 +14,7 @@ import (
 
 	"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"
@@ -137,6 +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
@@ -156,6 +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()
@@ -181,6 +384,15 @@ 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)
+	}
 	// 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")

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/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

@@ -16,7 +16,7 @@ import (
 func TestExecuteToolCallRejectsMissingAdditionalPermissionsBeforePrompt(t *testing.T) {
 	root := t.TempDir()
 	registry := tools.NewRegistry()
-	registry.Register(tools.NewBashTool(root))
+	registry.Register(tools.NewScopedBashTool(root, nil))
 
 	promptCalls := 0
 	result, err := executeToolCall(
@@ -51,7 +51,7 @@ func TestExecuteToolCallRejectsMissingAdditionalPermissionsBeforePrompt(t *testi
 func TestExecuteToolCallRejectsAdditionalPermissionsWithoutFlagBeforePrompt(t *testing.T) {
 	root := t.TempDir()
 	registry := tools.NewRegistry()
-	registry.Register(tools.NewBashTool(root))
+	registry.Register(tools.NewScopedBashTool(root, nil))
 
 	promptCalls := 0
 	result, err := executeToolCall(
@@ -88,7 +88,7 @@ func TestExecuteToolCallRejectsAdditionalPermissionsWithoutFlagBeforePrompt(t *t
 func TestExecuteToolCallMissingAdditionalPermissionsErrorIsActionable(t *testing.T) {
 	root := t.TempDir()
 	registry := tools.NewRegistry()
-	registry.Register(tools.NewBashTool(root))
+	registry.Register(tools.NewScopedBashTool(root, nil))
 
 	result, err := executeToolCall(
 		context.Background(),
@@ -117,7 +117,7 @@ func TestExecuteToolCallMissingAdditionalPermissionsErrorIsActionable(t *testing
 func TestExecuteToolCallStillPromptsForValidAdditionalPermissions(t *testing.T) {
 	root := t.TempDir()
 	registry := tools.NewRegistry()
-	registry.Register(tools.NewBashTool(root))
+	registry.Register(tools.NewScopedBashTool(root, nil))
 
 	promptCalls := 0
 	_, err := executeToolCall(

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/command_prefix.go

@@ -501,15 +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/compaction.go

@@ -6,6 +6,7 @@ import (
 	"errors"
 	"strings"
 
+	"github.com/Gitlawb/zero/internal/trace"
 	"github.com/Gitlawb/zero/internal/zeroruntime"
 )
 
@@ -55,6 +56,10 @@ 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.
@@ -219,7 +224,7 @@ func CompactMessages(messages []zeroruntime.Message, opts CompactionOptions) (Co
 
 	// 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]...)
@@ -325,6 +330,7 @@ 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
 
 	// calibrationRatio scales the raw byte/4 token estimate toward the provider's
 	// real prompt-token count. ApproxTextTokens over-counts code-heavy content by
@@ -364,13 +370,15 @@ func (state *compactionState) calibratedTokens(raw int) int {
 	return int(float64(raw) * state.calibrationRatio)
 }
 
-func newCompactionState(options Options) *compactionState {
-	return &compactionState{
+func newCompactionState(options Options, task *taskState) *compactionState {
+	state := &compactionState{
 		enabled:      options.ContextWindow > 0,
 		threshold:    compactionThreshold(options.ContextWindow),
 		preserveLast: options.CompactionPreserveLast,
 		onUsage:      options.OnUsage,
+		task:         task,
 	}
+	return state
 }
 
 // maybeCompact runs proactive compaction at the top of a turn. It returns the
@@ -417,6 +425,7 @@ func (state *compactionState) maybeCompact(
 	compacted, err := Compact(messages, CompactionOptions{
 		PreserveLast: state.preserveLast,
 		Summarize:    summarizeClosure(ctx, provider, state.onUsage),
+		taskState:    state.task.snapshotForCompaction(messages),
 	})
 	if err != nil {
 		// Summarizer failed: keep the original history. The reactive path (or a
@@ -430,6 +439,12 @@ func (state *compactionState) maybeCompact(
 		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
 }
@@ -461,6 +476,7 @@ func (state *compactionState) recover(
 	result, compactErr := Compact(messages, CompactionOptions{
 		PreserveLast: state.preserveLast,
 		Summarize:    summarizeClosure(ctx, provider, state.onUsage),
+		taskState:    state.task.snapshotForCompaction(messages),
 	})
 	if compactErr != nil {
 		// A genuine compaction attempt was made (and failed): the budget is spent
@@ -479,7 +495,11 @@ 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 = state.calibratedTokens(estimateTokens(result) + estimateToolDefTokens(tools))
 	return result, true, nil

internal/agent/compaction_preserve.go

@@ -30,6 +30,10 @@ const (
 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 (
@@ -76,32 +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)
@@ -431,12 +417,25 @@ 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"`
+}
+
 type preservedEdit struct {
 	Path string `json:"path"`
 	Note string `json:"note,omitempty"`
@@ -464,8 +463,27 @@ 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),
+		}
+		// 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
+		}
+	}
 
 	// Plan: a fresh update_plan in middle is authoritative; otherwise carry the
 	// plan preserved by an earlier compaction.
@@ -493,12 +511,25 @@ func appendPreservedState(summary string, middle []zeroruntime.Message) string {
 		projectInstructionEntries(middle),
 	)
 
-	if block := formatPreservedState(plan, edits, tools, skills, instructions); block != "" {
+	if block := formatPreservedState(plan, task, edits, tools, skills, instructions); block != "" {
 		summary += "\n\n" + block
 	}
 	return summary
 }
 
+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
@@ -555,11 +586,11 @@ func preservedEditsToEntries(edits []preservedEdit) []skillEntry {
 
 // formatPreservedState renders state as the labelled, single-line
 // JSON block. Returns "" when there is nothing to preserve.
-func formatPreservedState(plan string, edits, tools, skills, instructions []skillEntry) string {
-	if plan == "" && len(edits) == 0 && 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})
 	}
@@ -579,15 +610,6 @@ func formatPreservedState(plan string, edits, tools, skills, instructions []skil
 	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

@@ -1,6 +1,7 @@
 package agent
 
 import (
+	"encoding/json"
 	"strings"
 	"testing"
 	"unicode/utf8"
@@ -59,6 +60,103 @@ 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"}]}`})
+	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)
+	}
+}
+
+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 +381,10 @@ 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_test.go

@@ -336,7 +336,7 @@ 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 +364,7 @@ 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,
@@ -427,7 +427,7 @@ 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))
 
 	// ContextWindow large enough that proactive compaction never triggers, so
 	// only the reactive path can save the run.
@@ -491,7 +491,7 @@ 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 +581,7 @@ 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 +628,7 @@ 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/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"})
 

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/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/guardrails.go

@@ -1,7 +1,6 @@
 package agent
 
 import (
-	"encoding/json"
 	"strconv"
 	"strings"
 
@@ -127,9 +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
@@ -183,35 +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 + `")`
 }
@@ -230,8 +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: " +
@@ -239,6 +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.
@@ -476,16 +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++
 		}

internal/agent/guardrails_test.go

@@ -138,7 +138,7 @@ 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{
@@ -169,6 +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{
@@ -229,7 +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{
@@ -260,7 +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{
@@ -288,7 +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{
@@ -318,7 +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
@@ -352,7 +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{
@@ -387,7 +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/loop.go

@@ -7,14 +7,17 @@ import (
 	"fmt"
 	"io"
 	"sort"
+	"strconv"
 	"strings"
 	"sync"
 
+	"github.com/Gitlawb/zero/internal/execution"
 	"github.com/Gitlawb/zero/internal/hooks"
 	"github.com/Gitlawb/zero/internal/redaction"
 	"github.com/Gitlawb/zero/internal/sandbox"
 	"github.com/Gitlawb/zero/internal/streamjson"
 	"github.com/Gitlawb/zero/internal/tools"
+	"github.com/Gitlawb/zero/internal/trace"
 	"github.com/Gitlawb/zero/internal/zeroruntime"
 )
 
@@ -114,6 +117,53 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 		return Result{}, errors.New("agent provider is required")
 	}
 
+	// Tracing is opt-in. When a recorder is wired, thread it into ctx so the
+	// providerio seam and reconnect helper can reach it via trace.FromContext,
+	// mark the run's start, and wrap OnUsage so token counters accumulate for
+	// free alongside the existing per-request plumbing. nil recorder leaves the
+	// loop byte-identical: FromContext returns nil, the no-op stamps cost
+	// nothing, and the OnUsage wrapper is not installed.
+	if options.Trace != nil {
+		ctx = trace.WithContext(ctx, options.Trace)
+		options.Trace.Start()
+		originalOnUsage := options.OnUsage
+		options.OnUsage = func(usage Usage) {
+			if originalOnUsage != nil {
+				originalOnUsage(usage)
+			}
+			options.Trace.Counter(trace.CounterInputTokens, int64(usage.InputTokens))
+			options.Trace.Counter(trace.CounterCachedInputTokens, int64(usage.CachedInputTokens))
+			options.Trace.Counter(trace.CounterOutputTokens, int64(usage.OutputTokens))
+		}
+	}
+
+	// Establish the run's turn session — the seam an optimized provider session
+	// (connection reuse, prewarm, native compaction) plugs into without touching
+	// this loop. Default: wrap the passed provider so the session's Stream IS
+	// provider.StreamCompletion, byte-identical to calling it directly. Opened
+	// before dispatchSessionStart so a failed open is a clean run-start error
+	// with no session hooks fired and nothing to unwind.
+	turnSessions := options.TurnSessionProvider
+	if turnSessions == nil {
+		turnSessions = zeroruntime.NewProviderTurnSessionProvider(provider, zeroruntime.ProviderCapabilities{})
+	}
+	session, sessionErr := turnSessions.OpenTurnSession(ctx)
+	if sessionErr != nil {
+		return Result{}, fmt.Errorf("agent: open turn session: %w", sessionErr)
+	}
+	// Closed via closure (not `defer session.Close()`) so after a mid-run model
+	// swap the CURRENT session is the one closed at teardown — the swap closes
+	// the old one inline. Close errors are advisory and never fail the run.
+	defer func() { _ = session.Close() }()
+	// Best-effort prewarm: the default adapter no-ops; a real session may prime
+	// a connection. Failure is ignored by contract (never fatal).
+	_ = session.Prewarm(ctx)
+	// Route ALL provider I/O — per-turn streams, mid-stream retries, the
+	// compaction summarizer, and the final-answer request — through the session
+	// by reassigning the loop's provider local. Every existing call site reads
+	// this local, so no signatures change anywhere downstream.
+	provider = sessionProvider{session: session}
+
 	maxTurns := options.MaxTurns
 	if maxTurns <= 0 {
 		maxTurns = 12
@@ -132,10 +182,46 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 	options.runPermissions = runPermissions
 	defer runPermissions.cleanup()
 
-	messages := zeroruntime.SeedMessagesWithImages(buildSystemPrompt(options), prompt, options.Images)
+	promptParts := buildSystemPromptParts(options)
+	messages := zeroruntime.SeedMessagesWithImages(promptParts.prompt, prompt, options.Images)
 
 	guards := newGuardState()
-	compactor := newCompactionState(options)
+	task := newTaskState(prompt, options.Trace)
+	compactor := newCompactionState(options, task)
+	defer func() {
+		// A final transcript comparison is observational. It records drift but
+		// never rewrites the result or changes execution after the fact.
+		task.observePlanParity(result.Messages)
+		// Tool-result observations are intentionally coalesced; this final event
+		// guarantees their aggregate counts are present even on an early return.
+		task.emit()
+	}()
+	// The execution-profile posture controller. nil Options.Profile (every
+	// existing caller) makes every observe/decide call a no-op, keeping the
+	// loop byte-identical for unprofiled runs.
+	posture := newProfileController(options.Profile)
+	// applyPosture applies at most one posture escalation when an armed trigger
+	// has fired: raise the hoisted turn ceiling, restore effort and the
+	// completion gate, stamp the counter. Shared by the end-of-turn tail and
+	// the completion-gate uncertain path (which continues the loop before the
+	// tail runs). No-op forever after the first escalation and for unprofiled
+	// runs.
+	applyPosture := func() {
+		if target, fired := posture.maybeEscalate(); fired {
+			if target.MaxTurns > maxTurns {
+				maxTurns = target.MaxTurns
+			}
+			if target.ReasoningEffort != "" {
+				options.ReasoningEffort = target.ReasoningEffort
+			} else if target.RestoreDefaultEffort {
+				options.ReasoningEffort = ""
+			}
+			if target.RestoreCompletionGate {
+				options.RequireCompletionSignal = true
+			}
+			options.Trace.Counter(trace.CounterPostureEscalations, 1)
+		}
+	}
 
 	// Background post-edit diagnostics: files changed by mutating tools are
 	// checked off the tool-call critical path and any errors are appended as a
@@ -148,13 +234,10 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 	// loaded tool's full schema; it lives only for the run (v1 within-run scope).
 	loaded := map[string]bool{}
 
-	// continueNudges counts how many times the headless completion gate
-	// (Options.RequireCompletionSignal) has re-prompted a no-tool-call turn that
-	// stopped with work still unfinished. Bounded by maxContinueNudges.
-	continueNudges := 0
-	// acceptanceRequested records that the one-time task-grounded acceptance check
-	// has already been demanded this run, so it fires at most once.
-	acceptanceRequested := false
+	// The feature-gated completion policy keeps local completion decisions and its
+	// bounded follow-up state out of the central loop. Self-correcting profiles
+	// opt into the one permitted task-grounded semantic check.
+	completionPolicy := newCompletionPolicy(options.SelfCorrect != nil)
 
 	// toolDefCache memoizes each tool's rendered JSON-schema definition across
 	// turns (a tool's advertised schema is stable for the run), so partitionTools
@@ -164,7 +247,11 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 	result = Result{Messages: copyMessages(messages)}
 	dispatchSessionStart(ctx, options)
 	defer func() {
-		dispatchSessionEnd(ctx, options, result, err)
+		// Use a context detached from ctx's cancellation: on Esc/Ctrl+C the run
+		// context is already canceled by the time this defer runs, and a canceled
+		// parent would make Hooks.Dispatch's context.WithTimeout derive an
+		// already-expired context, so the sessionEnd hook would never actually run.
+		dispatchSessionEnd(context.WithoutCancel(ctx), options, result, err)
 	}()
 	for turn := 0; turn < maxTurns; turn++ {
 		result.Turns = turn + 1
@@ -184,12 +271,34 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 		// the tool-definition tokens (they ride on every request) in its estimate.
 		// partitionTools depends only on registry/permissions/options/loaded, not on
 		// the messages, so computing it before compaction is safe.
+		toolPartitionSpan := options.Trace.Span(trace.SpanToolPartition)
 		exposed, _ := partitionToolsCached(registry, permissionMode, options, loaded, toolDefCache)
+		toolPartitionSpan.End()
+
+		// Fingerprint the exact system prompt built at run start plus this turn's
+		// provider-visible tool definitions in their emitted order. Reusing the
+		// retained prompt parts avoids rebuilding workspace context while ensuring
+		// the trace describes the same bytes carried by the request.
+		if options.Trace != nil {
+			fp := computePrefixFingerprint(buildPromptSubstringsFromParts(promptParts, exposed))
+			options.Trace.EmitPrefixHash(trace.PrefixHash{
+				SystemPromptHash:       fp.SystemPromptHash,
+				BaseInstructionsHash:   fp.BaseInstructionsHash,
+				ConfirmationPolicyHash: fp.ConfirmationPolicyHash,
+				ProjectContextHash:     fp.ProjectContextHash,
+				SkillsHash:             fp.SkillsHash,
+				ToolsHash:              fp.ToolsHash,
+				SchemaHash:             fp.SchemaHash,
+				CompletePrefixHash:     fp.CompletePrefixHash,
+			})
+		}
 
 		// PROACTIVE compaction: if the history is approaching the model's
 		// context window, summarize the oldest middle before building the
 		// request. A no-op when ContextWindow == 0 (compaction disabled).
+		compactionSpan := options.Trace.Span(trace.SpanCompaction)
 		messages = compactor.maybeCompact(ctx, provider, messages, exposed)
+		compactionSpan.End()
 		request := zeroruntime.CompletionRequest{
 			Messages:        copyMessages(messages),
 			Tools:           exposed,
@@ -255,11 +364,39 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 		// conversation-state duplication.
 		forwardedVisibleText := false
 		forwardingOpts := zeroruntime.CollectOptions{OnUsage: options.OnUsage}
-		if options.OnText != nil {
-			forwardingOpts.OnText = func(s string) { forwardedVisibleText = true; options.OnText(s) }
+		// Install text/reasoning forwarding handlers whenever EITHER a user
+		// callback OR a trace recorder is set. A headless traced run (e.g. `zero
+		// exec --trace`) sets Trace but no OnText/OnReasoning; without these
+		// handlers the stream is still collected, but FirstTokenAt would never
+		// stamp and the trace would lose its TTFT signal. The trace recorder's
+		// stamp methods are nil-safe, but we guard on `trace != nil` so a run with
+		// a user callback but no recorder still works. forwardedVisibleText stays
+		// tied to the USER callback only, preserving the stall-retry semantics
+		// (a trace-only handler forwards nothing the user would see duplicated).
+		rec := options.Trace
+		onText := options.OnText
+		if onText != nil || rec != nil {
+			forwardingOpts.OnText = func(s string) {
+				if rec != nil {
+					rec.StampFirstVisibleEvent()
+					rec.StampFirstToken()
+				}
+				if onText != nil {
+					forwardedVisibleText = true
+					onText(s)
+				}
+			}
 		}
-		if options.OnReasoning != nil {
-			forwardingOpts.OnReasoning = func(s string) { options.OnReasoning(s) }
+		onReasoning := options.OnReasoning
+		if onReasoning != nil || rec != nil {
+			forwardingOpts.OnReasoning = func(s string) {
+				if rec != nil {
+					rec.StampFirstToken()
+				}
+				if onReasoning != nil {
+					onReasoning(s)
+				}
+			}
 		}
 		if options.OnToolCallStart != nil {
 			forwardingOpts.OnToolCallStart = func(id, name string) { options.OnToolCallStart(id, name) }
@@ -297,14 +434,18 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 				if retryStreamErr != nil {
 					return collected, retryStreamErr
 				}
+				genSpan := options.Trace.Span(trace.SpanGeneration)
 				collected = zeroruntime.CollectStreamWithOptions(ctx, retryStream, zeroruntime.CollectOptions{
 					OnUsage: options.OnUsage,
 				})
+				genSpan.End()
 			}
 			return collected, nil
 		}
 
+		generationSpan := options.Trace.Span(trace.SpanGeneration)
 		collected := zeroruntime.CollectStreamWithOptions(ctx, stream, forwardingOpts)
+		generationSpan.End()
 		if collected.Error != "" {
 			updated, stop := recoverStreamError(collected)
 			collected = updated
@@ -357,7 +498,9 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 				result.Messages = copyMessages(messages)
 				return result, retryErr
 			}
+			stallGenSpan := options.Trace.Span(trace.SpanGeneration)
 			collected = zeroruntime.CollectStreamWithOptions(ctx, retryStream, forwardingOpts)
+			stallGenSpan.End()
 		}
 		if collected.Error != "" {
 			// Route a reissued stream's non-stall error through the SAME recovery as
@@ -433,67 +576,42 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 			// model's final answer ONLY when the work is actually done. Default off
 			// (RequireCompletionSignal), so interactive runs stay byte-identical.
 			if options.RequireCompletionSignal {
-				// (1) Self-report downgrade (strongest, unambiguous): the model's own
-				// final message admits it guessed / could not meet the objective. Checked
-				// FIRST so an admitted-impossible task is downgraded immediately (no wasted
-				// continue-nudges) and reports the accurate reason.
-				if reason := selfReportedIncompletion(collected.Text); reason != "" {
+				completionContext := task.completionContext(messages, guards.pendingPlanItems())
+				evaluation := completionPolicy.evaluate(collected.Text, completionContext)
+				task.observe(taskStateEvent{kind: taskStateEventCompletion, completion: evaluation})
+				switch evaluation.Decision {
+				case CompletionIncomplete:
 					result.Incomplete = true
-					result.IncompleteReason = reason
+					result.IncompleteReason = evaluation.Reason
 					result.FinalAnswer = collected.Text
 					result.Messages = copyMessages(messages)
 					return result, nil
-				}
-
-				// (2) The model stopped without a tool call while work may be unfinished:
-				//   - a continuation cue ("…Let me check the config:") is an unambiguous
-				//     mid-step stop;
-				//   - pending update_plan items are a WEAK, ambiguous signal (the model may
-				//     have finished without re-marking the last step).
-				// Nudge to continue (bounded). After the budget: a persisted continuation
-				// cue finalizes INCOMPLETE; pending-plan WITHOUT a cue does NOT (that would
-				// false-fail a completed run with stale bookkeeping) — fall through to the
-				// acceptance check / success.
-				cue := endsWithContinuationCue(collected.Text)
-				planPending := guards.pendingPlanItems()
-				if cue || planPending {
-					if continueNudges < maxContinueNudges {
-						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"
-						}
+				case CompletionUncertain:
+					// Observe the uncertainty signal and apply any escalation NOW:
+					// this branch continues the turn loop before the end-of-turn
+					// act point, so deferring would skip escalation entirely.
+					posture.observeUncertain()
+					applyPosture()
+					switch evaluation.Action {
+					case completionActionContinue:
+						options.Trace.Counter(trace.CounterCompletionNudges, 1)
 						messages = append(messages, zeroruntime.Message{
 							Role:    zeroruntime.MessageRoleUser,
-							Content: continueNudge(reason),
+							Content: continueNudge(evaluation.Reason),
+						})
+					case completionActionSemanticCheck:
+						options.Trace.Counter(trace.CounterAcceptanceChecks, 1)
+						messages = append(messages, zeroruntime.Message{
+							Role:    zeroruntime.MessageRoleUser,
+							Content: acceptanceVerificationNudge(completionContext.Objective),
 						})
-						continue
 					}
-					if cue {
-						result.Incomplete = true
-						result.IncompleteReason = "your message ended mid-step"
-						result.FinalAnswer = collected.Text
-						result.Messages = copyMessages(messages)
-						return result, nil
-					}
-					// pending-plan only, budget spent: trust the model's completion claim
-					// over stale plan bookkeeping; fall through.
-				}
-
-				// (3) Task-grounded acceptance: before accepting a "done" turn as success,
-				// require ONE acceptance check grounded in the task's stated criterion
-				// (only when self-correct is on). Rejects "well-formed == correct",
-				// "existing-tests-pass == objective met", and "result == baseline" false
-				// successes. Bounded to a single pass; a genuine post-check completion
-				// (no admission, no cue) then finalizes as success on the next turn.
-				if options.SelfCorrect != nil && !acceptanceRequested {
-					acceptanceRequested = true
-					messages = append(messages, zeroruntime.Message{
-						Role:    zeroruntime.MessageRoleUser,
-						Content: acceptanceVerificationNudge(),
-					})
 					continue
+				case CompletionComplete:
+					// Local evidence is sufficient; proceed to final diagnostics.
 				}
+			} else {
+				task.observe(taskStateEvent{kind: taskStateEventCompletion, completion: completionEvaluation{Decision: CompletionComplete}})
 			}
 			// Finalization diagnostics gate: edits from this run may still have
 			// checks in flight — the per-turn drain waits only briefly and defers
@@ -517,6 +635,11 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 		// executing so the empty-turn counter resets and plan-tracking signals
 		// stay current.
 		guards.observeTurn(collected)
+		for _, call := range collected.ToolCalls {
+			if call.Name == planToolName {
+				task.observe(taskStateEvent{kind: taskStateEventPlan, arguments: call.Arguments})
+			}
+		}
 
 		failureHint := ""
 		// turnRequestedModel records the FIRST mid-run escalation target requested
@@ -543,25 +666,32 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 			// The scan happens lazily at the run's first index — never ahead of a
 			// pending mutating call — so read-after-write ordering is preserved.
 			if index >= precomputedEnd {
-				runEnd := index
-				for runEnd < len(collected.ToolCalls) && parallelSafeToolCall(registry, collected.ToolCalls[runEnd], options) {
-					runEnd++
-				}
+				// Capability-safe consecutive run (ReadOnly+ThreadSafe, no
+				// resource-key conflicts). See extendParallelRun.
+				runEnd := extendParallelRun(registry, collected.ToolCalls, index, options)
 				if runEnd-index >= 2 {
+					batchSpan := options.Trace.Span(trace.SpanToolExecution)
 					precomputed = executeParallelReadBatch(ctx, registry, collected.ToolCalls, index, runEnd, permissionMode, options)
+					batchSpan.End()
 					precomputedStart, precomputedEnd = index, runEnd
 				}
 			}
 			if options.OnToolCall != nil {
 				options.OnToolCall(call)
 			}
+			options.Trace.StampFirstUsefulAction()
 			var toolResult ToolResult
 			var abortErr error
 			if index >= precomputedStart && index < precomputedEnd {
 				toolResult, abortErr = precomputed[index-precomputedStart].result, precomputed[index-precomputedStart].abortErr
 			} else {
+				toolSpan := options.Trace.Span(trace.SpanToolExecution)
 				toolResult, abortErr = executeToolCall(ctx, registry, call, permissionMode, options)
+				toolSpan.End()
 			}
+			options.Trace.Counter(trace.CounterToolCalls, 1)
+			recordOutputBudgetTrace(options.Trace, toolResult)
+			task.observe(taskStateEvent{kind: taskStateEventToolResult, toolResult: toolResult})
 			if options.OnToolResult != nil {
 				options.OnToolResult(toolResult)
 			}
@@ -608,6 +738,7 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 			// would misdirect the model toward JSON shape or blocked behavior.
 			retriableFailure := isRetriableToolError(toolResult)
 			outcome := guards.observeToolResult(call.Name, retriableFailure, toolResult.Output)
+			posture.observeToolOutcome(outcome, toolResult)
 			if outcome.Stop {
 				// The assistant message advertised EVERY collected tool call, but
 				// the guard halts mid-turn so the calls after this one never run.
@@ -640,7 +771,10 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 		// the assistant's tool_results stay contiguous (a user message between
 		// tool_results breaks strict provider replay). nil SelfCorrect is a no-op.
 		if options.SelfCorrect != nil && len(changedFilesThisBatch) > 0 {
-			if feedback, _ := options.SelfCorrect.AfterEdit(ctx, dedupeStrings(changedFilesThisBatch)); feedback != "" {
+			feedback, selfCorrectOutcome := options.SelfCorrect.AfterEdit(ctx, dedupeStrings(changedFilesThisBatch))
+			posture.observeSelfCorrect(selfCorrectOutcome)
+			task.observe(taskStateEvent{kind: taskStateEventVerification, verification: selfCorrectOutcome})
+			if feedback != "" {
 				messages = append(messages, zeroruntime.Message{
 					Role:    zeroruntime.MessageRoleUser,
 					Content: feedback,
@@ -653,24 +787,53 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 		// of the run. At most one switch per turn (first request wins). A switcher
 		// error is NON-FATAL — record a brief note and continue on the current
 		// model. nil switcher ⇒ requests are ignored entirely (escalation off).
-		if turnRequestedModel != "" && options.ModelSwitcher != nil {
-			newProvider, switchErr := options.ModelSwitcher(ctx, turnRequestedModel)
+		if turnRequestedModel != "" && (options.ModelSessionSwitcher != nil || options.ModelSwitcher != nil) {
+			// Resolve the escalated model's session source. The target-aware
+			// ModelSessionSwitcher is preferred: its TurnSessionProvider carries an
+			// optimized session (and capabilities) across the swap. The legacy
+			// ModelSwitcher fallback returns a bare Provider, wrapped in the
+			// default no-op session — identical to pre-seam behavior.
+			var newSessions zeroruntime.TurnSessionProvider
+			var switchErr error
+			if options.ModelSessionSwitcher != nil {
+				newSessions, switchErr = options.ModelSessionSwitcher(ctx, turnRequestedModel)
+			} else {
+				var newProvider Provider
+				newProvider, switchErr = options.ModelSwitcher(ctx, turnRequestedModel)
+				if switchErr == nil && newProvider != nil {
+					newSessions = zeroruntime.NewProviderTurnSessionProvider(newProvider, zeroruntime.ProviderCapabilities{})
+				}
+			}
 			if switchErr != nil {
 				messages = append(messages, zeroruntime.Message{
 					Role:    zeroruntime.MessageRoleUser,
 					Content: escalationFailedNoticePrefix + " (" + turnRequestedModel + "): " + switchErr.Error() + ". Continuing on " + options.Model + ".",
 				})
-			} else if newProvider != nil {
-				// Reassign the local provider so the next turn's StreamCompletion and
-				// compaction use it; update options.Model so subsequent RunOptions.Model,
-				// context-window sizing, and usage attribution follow the new model.
-				provider = newProvider
-				options.Model = turnRequestedModel
-				// KNOWN LIMITATION (deferred): the compactor's context-window budget
-				// is fixed at run start from options.ContextWindow and is NOT updated
-				// here, so a switch to a model with a different window keeps compacting
-				// against the original budget. Fixing it needs a ModelSwitcher contract
-				// change (return the new window) — out of scope for this change.
+			} else if newSessions != nil {
+				// Open the new session and close the old one, so the next turn's
+				// streams and compaction use the new model. For the default adapter
+				// the open never errors and Close is a no-op, so this stays
+				// byte-identical to a direct provider reassignment; the guarded open
+				// and prewarm matter once a session holds real state.
+				newSession, openErr := newSessions.OpenTurnSession(ctx)
+				if openErr != nil {
+					messages = append(messages, zeroruntime.Message{
+						Role:    zeroruntime.MessageRoleUser,
+						Content: escalationFailedNoticePrefix + " (" + turnRequestedModel + "): " + openErr.Error() + ". Continuing on " + options.Model + ".",
+					})
+				} else {
+					_ = session.Close()
+					session = newSession
+					_ = session.Prewarm(ctx)
+					provider = sessionProvider{session: session}
+					options.Model = turnRequestedModel
+					options.Trace.Counter(trace.CounterModelSwitches, 1)
+					// KNOWN LIMITATION (deferred): the compactor's context-window budget
+					// is fixed at run start from options.ContextWindow and is NOT updated
+					// here, so a switch to a model with a different window keeps compacting
+					// against the original budget. Fixing it needs the switcher to also
+					// report the new window — out of scope for this change.
+				}
 			}
 		}
 
@@ -704,6 +867,13 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 				Content: reminder,
 			})
 		}
+
+		// One-shot posture escalation: when an armed profile trigger fired this
+		// turn, restore the stricter knob values the profile displaced. Mutates
+		// only per-turn-read policy (the hoisted turn ceiling, reasoning effort,
+		// completion gate); never messages, model, or session. No-op forever
+		// after the first escalation and for every unprofiled run.
+		applyPosture()
 	}
 
 	if ctx.Err() != nil {
@@ -746,6 +916,28 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 	return result, nil
 }
 
+func recordOutputBudgetTrace(recorder *trace.Recorder, result ToolResult) {
+	if recorder == nil || result.Meta["output_budget_category"] == "" {
+		return
+	}
+	parseInt := func(key string) int {
+		value, _ := strconv.Atoi(result.Meta[key])
+		return value
+	}
+	spillCreated, _ := strconv.ParseBool(result.Meta["output_budget_spill_created"])
+	recorder.EmitOutputBudget(trace.OutputBudgetEvent{
+		Tool:                    result.Name,
+		Category:                result.Meta["output_budget_category"],
+		OriginalBytes:           parseInt("output_budget_original_bytes"),
+		RetainedBytes:           parseInt("output_budget_retained_bytes"),
+		EstimatedOriginalTokens: parseInt("output_budget_estimated_original_tokens"),
+		EstimatedRetainedTokens: parseInt("output_budget_estimated_retained_tokens"),
+		Truncated:               result.Truncated,
+		Reason:                  result.Meta["output_budget_reason"],
+		SpillCreated:            spillCreated,
+	})
+}
+
 func finalAnswerAfterMaxTurns(ctx context.Context, provider Provider, messages []zeroruntime.Message, toolDefs []zeroruntime.ToolDefinition, options Options) (string, []zeroruntime.Message, string) {
 	finalMessages := copyMessages(messages)
 	finalMessages = append(finalMessages, zeroruntime.Message{
@@ -764,6 +956,7 @@ func finalAnswerAfterMaxTurns(ctx context.Context, provider Provider, messages [
 	if err != nil {
 		return "", messages, ""
 	}
+	finalGenSpan := options.Trace.Span(trace.SpanGeneration)
 	collected := zeroruntime.CollectStreamWithOptions(ctx, stream, zeroruntime.CollectOptions{
 		OnText:          options.OnText,
 		OnReasoning:     options.OnReasoning,
@@ -771,6 +964,7 @@ func finalAnswerAfterMaxTurns(ctx context.Context, provider Provider, messages [
 		OnToolCallStart: options.OnToolCallStart,
 		OnToolCallDelta: options.OnToolCallDelta,
 	})
+	finalGenSpan.End()
 	if ctx.Err() != nil || collected.Error != "" || strings.TrimSpace(collected.Text) == "" {
 		return "", messages, ""
 	}
@@ -973,6 +1167,17 @@ func executeToolCall(ctx context.Context, registry *tools.Registry, call ToolCal
 			decisionCommandPrefix = grant.Prefix
 		}
 	}
+	// An explicit additional-permissions request is an elevation only when it
+	// expands the current effective policy. If the same capability was already
+	// approved for this session, reuse that grant across equivalent commands
+	// instead of prompting again merely because their command prefixes differ.
+	if toolFound && !permissionGranted && isShellCommandTool(call.Name) && options.Sandbox != nil {
+		if profile, requested, profileErr := inlineAdditionalPermissionsProfile(args, options.Cwd); profileErr == nil && requested && options.Sandbox.CoversRequestPermissions(profile) {
+			permissionGranted = true
+			decisionAction = PermissionDecisionAllowForSession
+			decisionReason = "requested sandbox capabilities already granted for this session"
+		}
+	}
 
 	var preflightDecision *sandbox.Decision
 	if toolFound && options.Sandbox != nil {
@@ -1198,21 +1403,39 @@ func executeToolCall(ctx context.Context, registry *tools.Registry, call ToolCal
 			if didRedact {
 				result.Redacted = true
 			}
+			result = registry.RebudgetAfterHook(call.Name, args, result)
 		}
 	}
+	// Stamp the sandbox risk classification for this EXECUTED call so run-policy
+	// observers can see the risk of an allowed mutation. Prefer the preflight
+	// decision's classification when the sandbox evaluated the call; otherwise
+	// run the same pure classifier the permission path uses. Denied/canceled
+	// results returned earlier keep the zero value.
+	executedRisk := sandbox.Risk{}
+	if preflightDecision != nil {
+		executedRisk = preflightDecision.Risk
+	} else if toolFound {
+		// Unknown-tool results fall through here with a nil tool; they carry a
+		// denial and keep the zero risk value.
+		executedRisk = sandbox.Classify(sandboxRequest(call.Name, tool, args, permissionGranted, permissionMode, options))
+	}
+
 	// Secret scrubbing happens at the registry boundary (the single point both
 	// the agent loop and the MCP server pass through), so result.Output is
 	// already redacted here and result.Redacted reflects whether it changed.
 	return ToolResult{
-		ToolCallID:   call.ID,
-		Name:         call.Name,
-		Status:       result.Status,
-		Output:       result.Output,
-		Meta:         result.Meta,
-		Redacted:     result.Redacted,
-		ChangedFiles: result.ChangedFiles,
-		Display:      result.Display,
-		LoadedTools:  loadedToolsFromResult(result.Meta),
+		Risk:            executedRisk,
+		ToolCallID:      call.ID,
+		Name:            call.Name,
+		Status:          result.Status,
+		Output:          result.Output,
+		Truncated:       result.Truncated,
+		Meta:            result.Meta,
+		Redacted:        result.Redacted,
+		ChangedFiles:    result.ChangedFiles,
+		ChangeSummaries: result.ChangeSummaries,
+		Display:         result.Display,
+		LoadedTools:     loadedToolsFromResult(result.Meta),
 		// A tool may signal a mid-run model escalation by carrying the target id
 		// in Meta["escalate_to_model"]. Lift it into the typed loop-level field;
 		// the Run turn loop performs the actual provider switch. Empty for every
@@ -1332,8 +1555,15 @@ func sandboxDeniedNetworkRetryCandidate(call ToolCall, args map[string]any, resu
 	if options.Sandbox == nil || !isShellCommandTool(call.Name) {
 		return false
 	}
-	if result.Meta[tools.SandboxLikelyDeniedMeta] != "true" || result.Meta[tools.SandboxDenialKindMeta] != tools.SandboxDenialKindNetwork {
-		return false
+	if result.ExecutionOutcome != nil {
+		denial := result.ExecutionOutcome.Denial
+		if denial == nil || denial.Capability.Kind != execution.CapabilityExternalNetwork {
+			return false
+		}
+	} else {
+		if result.Meta[tools.SandboxLikelyDeniedMeta] != "true" || result.Meta[tools.SandboxDenialKindMeta] != tools.SandboxDenialKindNetwork {
+			return false
+		}
 	}
 	if value, _ := args["sandbox_permissions"].(string); strings.TrimSpace(value) == string(tools.SandboxPermissionsRequireEscalated) {
 		return false
@@ -1386,6 +1616,17 @@ func sandboxRestrictedShellRetryCandidate(call ToolCall, args map[string]any, re
 }
 
 func sandboxDeniedShellResult(result tools.Result) bool {
+	if result.ExecutionOutcome != nil {
+		// Only a typed platform denial that explicitly requests unrestricted
+		// recovery may enter the legacy retry path. Narrow capability denials
+		// remain on their exact capability path.
+		denial := result.ExecutionOutcome.Denial
+		return denial != nil &&
+			denial.Source == execution.DenialSourcePlatformSandbox &&
+			denial.Capability.Kind == execution.CapabilityUnrestricted &&
+			denial.Recoverable &&
+			denial.NextAction == execution.DenialNextActionRequestApproval
+	}
 	return result.Status == tools.StatusError && result.Meta[tools.SandboxLikelyDeniedMeta] == "true"
 }
 
@@ -1514,16 +1755,18 @@ func toolResultFromPrePermissionReject(call ToolCall, result tools.Result) ToolR
 	}
 
 	return ToolResult{
-		ToolCallID:     call.ID,
-		Name:           call.Name,
-		Status:         result.Status,
-		Output:         output,
-		Meta:           meta,
-		Redacted:       result.Redacted || outputRedacted || summaryRedacted || metaRedacted,
-		ChangedFiles:   result.ChangedFiles,
-		Display:        display,
-		LoadedTools:    loadedToolsFromResult(meta),
-		RequestedModel: meta["escalate_to_model"],
+		ToolCallID:      call.ID,
+		Name:            call.Name,
+		Status:          result.Status,
+		Output:          output,
+		Truncated:       result.Truncated,
+		Meta:            meta,
+		Redacted:        result.Redacted || outputRedacted || summaryRedacted || metaRedacted,
+		ChangedFiles:    result.ChangedFiles,
+		ChangeSummaries: result.ChangeSummaries,
+		Display:         display,
+		LoadedTools:     loadedToolsFromResult(meta),
+		RequestedModel:  meta["escalate_to_model"],
 	}
 }
 
@@ -1783,14 +2026,16 @@ func askUserFallbackResult(ctx context.Context, registry *tools.Registry, call T
 			Cwd:               options.Cwd,
 		})
 		return ToolResult{
-			ToolCallID:   call.ID,
-			Name:         call.Name,
-			Status:       result.Status,
-			Output:       result.Output,
-			Meta:         result.Meta,
-			Redacted:     result.Redacted,
-			ChangedFiles: result.ChangedFiles,
-			Display:      result.Display,
+			ToolCallID:      call.ID,
+			Name:            call.Name,
+			Status:          result.Status,
+			Output:          result.Output,
+			Truncated:       result.Truncated,
+			Meta:            result.Meta,
+			Redacted:        result.Redacted,
+			ChangedFiles:    result.ChangedFiles,
+			ChangeSummaries: result.ChangeSummaries,
+			Display:         result.Display,
 		}
 	}
 	return ToolResult{
@@ -1874,7 +2119,10 @@ func requestPermission(ctx context.Context, request PermissionRequest, options O
 	if options.OnPermissionRequest == nil {
 		return PermissionDecision{Action: PermissionDecisionDeny, Reason: request.Reason}, nil
 	}
-	return options.OnPermissionRequest(ctx, request)
+	permSpan := options.Trace.Span(trace.SpanPermissionWait)
+	decision, err := options.OnPermissionRequest(ctx, request)
+	permSpan.End()
+	return decision, err
 }
 
 func normalizePermissionDecisionAction(action PermissionDecisionAction) PermissionDecisionAction {
@@ -2351,7 +2599,8 @@ func buildPermissionEvent(call ToolCall, tool tools.Tool, args map[string]any, p
 	if profile, ok, err := inlineAdditionalPermissionsProfile(args, options.Cwd); ok && err == nil {
 		scopeText = requestPermissionsScope(profile)
 	}
-	reason = userFacingPermissionReason(call.Name, args, reason)
+	decisionReason := strings.TrimSpace(reason)
+	reason = userFacingPermissionReason(call.Name, args, reason, safety.Reason)
 
 	return PermissionEvent{
 		ToolCallID:        call.ID,
@@ -2364,6 +2613,7 @@ func buildPermissionEvent(call ToolCall, tool tools.Tool, args map[string]any, p
 		Autonomy:          options.Autonomy,
 		SideEffect:        string(safety.SideEffect),
 		Reason:            reason,
+		DecisionReason:    decisionReason,
 		Scope:             scopeText,
 		Risk:              risk,
 		Block:             block,
@@ -2373,18 +2623,31 @@ func buildPermissionEvent(call ToolCall, tool tools.Tool, args map[string]any, p
 	}, true
 }
 
-func userFacingPermissionReason(toolName string, args map[string]any, reason string) string {
+func userFacingPermissionReason(toolName string, args map[string]any, reason string, fallback string) string {
 	reason = strings.TrimSpace(reason)
-	if !isShellCommandTool(toolName) || !shellCommandRequiresEscalated(args) {
+	if !isShellCommandTool(toolName) {
 		return reason
 	}
-	if justification, ok := firstStringArg(args, "justification"); ok {
-		return justification
+
+	// Policy and sandbox explanations are authoritative. The tool's static
+	// safety copy is only an internal fallback; presenting it as the reason for
+	// every ordinary shell approval makes prompts generic and obscures the cases
+	// where policy has an actionable explanation.
+	fallback = strings.TrimSpace(fallback)
+	if reason != "" && reason != fallback && reason != "tool requires approval before execution" && reason != sandbox.ReasonEscalatedSandboxRequired {
+		return reason
 	}
-	if reason == "" || reason == sandbox.ReasonEscalatedSandboxRequired {
+
+	if shellCommandRequiresEscalated(args) {
+		if justification, ok := firstStringArg(args, "justification"); ok {
+			return justification
+		}
 		return "This command needs to run outside the sandbox."
 	}
-	return reason
+
+	// Ordinary shell approvals need no invented explanation. The raw fallback is
+	// retained separately as DecisionReason for traces and policy auditing.
+	return ""
 }
 
 func grantDecisionAction(grant *sandbox.Grant) PermissionDecisionAction {
@@ -2605,19 +2868,6 @@ func permissionActionFromSandbox(action sandbox.Action) PermissionAction {
 	}
 }
 
-// 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)
-}
-
 // partitionToolsCached is partitionTools with an optional per-tool definition
 // cache. The partitioning itself (visibility, deferral, ordering) is recomputed
 // every call — it must be, because a tool's deferred state can flip mid-run (e.g.

internal/agent/loop_test.go

@@ -6,15 +6,19 @@ import (
 	"errors"
 	"fmt"
 	"os"
+	"os/exec"
 	"path/filepath"
+	"reflect"
 	"runtime"
 	"strings"
 	"testing"
 
+	"github.com/Gitlawb/zero/internal/execution"
 	"github.com/Gitlawb/zero/internal/hooks"
 	"github.com/Gitlawb/zero/internal/sandbox"
 	"github.com/Gitlawb/zero/internal/specmode"
 	"github.com/Gitlawb/zero/internal/tools"
+	"github.com/Gitlawb/zero/internal/trace"
 	"github.com/Gitlawb/zero/internal/zeroruntime"
 )
 
@@ -23,6 +27,75 @@ type mockProvider struct {
 	requests []zeroruntime.CompletionRequest
 }
 
+func TestTypedExecutionOutcomeOverridesLegacySandboxHeuristics(t *testing.T) {
+	engine := sandbox.NewEngine(sandbox.EngineOptions{WorkspaceRoot: t.TempDir(), Policy: sandbox.DefaultPolicy()})
+	call := ToolCall{Name: tools.ExecCommandToolName}
+	options := Options{Sandbox: engine}
+	legacyMeta := map[string]string{
+		tools.SandboxLikelyDeniedMeta: "true",
+		tools.SandboxDenialKindMeta:   tools.SandboxDenialKindNetwork,
+	}
+
+	applicationFailure := tools.Result{
+		Status: tools.StatusError,
+		Meta:   legacyMeta,
+		ExecutionOutcome: &execution.Outcome{
+			State: execution.StateFailed,
+			Kind:  execution.OutcomeApplicationFailure,
+			Exit:  &execution.Exit{Code: 1},
+		},
+	}
+	if sandboxDeniedNetworkRetryCandidate(call, nil, applicationFailure, options) || sandboxRestrictedShellRetryCandidate(call, nil, applicationFailure, options) {
+		t.Fatal("typed application failure must not be retried from legacy sandbox-like text")
+	}
+
+	protectedDenial := applicationFailure
+	protectedDenial.ExecutionOutcome = &execution.Outcome{
+		State: execution.StateDenied,
+		Kind:  execution.OutcomeEnforcementDenied,
+		Denial: &execution.Denial{
+			Capability: execution.Capability{Kind: execution.CapabilityProtectedMetadata, Scope: "/workspace/.zero"},
+			Source:     execution.DenialSourceConfiguredPolicy,
+			Reason:     "protected metadata",
+			NextAction: execution.DenialNextActionRequestApproval,
+		},
+	}
+	if sandboxRestrictedShellRetryCandidate(call, nil, protectedDenial, options) {
+		t.Fatal("narrow protected-metadata denial must not become an unrestricted retry")
+	}
+
+	nonRecoverableDenial := protectedDenial
+	nonRecoverableDenial.ExecutionOutcome = &execution.Outcome{
+		State: execution.StateDenied,
+		Kind:  execution.OutcomeEnforcementDenied,
+		Denial: &execution.Denial{
+			Capability:  execution.Capability{Kind: execution.CapabilityUnrestricted, Scope: "host"},
+			Source:      execution.DenialSourcePlatformSandbox,
+			Reason:      "sandbox enforcement failed closed",
+			Recoverable: false,
+			NextAction:  execution.DenialNextActionRequestApproval,
+		},
+	}
+	if sandboxRestrictedShellRetryCandidate(call, nil, nonRecoverableDenial, options) {
+		t.Fatal("non-recoverable sandbox denial must not become an unrestricted retry")
+	}
+
+	networkDenial := protectedDenial
+	networkDenial.ExecutionOutcome = &execution.Outcome{
+		State: execution.StateDenied,
+		Kind:  execution.OutcomeEnforcementDenied,
+		Denial: &execution.Denial{
+			Capability: execution.Capability{Kind: execution.CapabilityExternalNetwork},
+			Source:     execution.DenialSourcePlatformSandbox,
+			Reason:     "external network denied",
+			NextAction: execution.DenialNextActionRequestApproval,
+		},
+	}
+	if !sandboxDeniedNetworkRetryCandidate(call, nil, networkDenial, options) {
+		t.Fatal("typed external-network denial should use the narrow network approval path")
+	}
+}
+
 func (provider *mockProvider) StreamCompletion(ctx context.Context, request zeroruntime.CompletionRequest) (<-chan zeroruntime.StreamEvent, error) {
 	provider.requests = append(provider.requests, request)
 
@@ -92,6 +165,87 @@ func TestRunDispatchesSessionLifecycleHooks(t *testing.T) {
 	}
 }
 
+// cancelingProvider simulates an Esc/Ctrl+C abort: it cancels the run's own
+// context mid-stream (as the TUI's interrupt handler would) and returns no
+// events, so Run observes ctx.Err() and returns early with an already-canceled
+// context in scope for the deferred sessionEnd dispatch.
+type cancelingProvider struct {
+	cancel context.CancelFunc
+}
+
+func (provider *cancelingProvider) StreamCompletion(ctx context.Context, request zeroruntime.CompletionRequest) (<-chan zeroruntime.StreamEvent, error) {
+	provider.cancel()
+	ch := make(chan zeroruntime.StreamEvent)
+	close(ch)
+	return ch, nil
+}
+
+// TestRunDispatchesSessionEndHookAfterContextCancellation is a regression test
+// for a bug where an interrupted run (Esc/Ctrl+C) canceled ctx before the
+// deferred sessionEnd dispatch ran, so Hooks.Dispatch's context.WithTimeout(ctx, ...)
+// derived an already-canceled context and the hook command never actually launched.
+func TestRunDispatchesSessionEndHookAfterContextCancellation(t *testing.T) {
+	goBinary, err := exec.LookPath("go")
+	if err != nil {
+		// The test binary runs on the build machine, so its build-time GOROOT
+		// still identifies the toolchain that produced it.
+		goRoot := runtime.GOROOT() //nolint:staticcheck // Safe for this non-portable test binary.
+		goBinary = filepath.Join(goRoot, "bin", "go")
+		if runtime.GOOS == "windows" {
+			goBinary += ".exe"
+		}
+		if _, statErr := os.Stat(goBinary); statErr != nil {
+			t.Skipf("go binary unavailable on PATH or in GOROOT: %v", statErr)
+		}
+	}
+	audit, err := hooks.NewAuditStore(hooks.AuditStoreOptions{AuditPath: filepath.Join(t.TempDir(), "audit.jsonl")})
+	if err != nil {
+		t.Fatalf("NewAuditStore: %v", err)
+	}
+	dispatcher := hooks.NewDispatcher(hooks.DispatcherOptions{
+		Config: hooks.Config{
+			Enabled: true,
+			Hooks: []hooks.Definition{
+				{ID: "zero.session-end", Event: hooks.EventSessionEnd, Command: goBinary, Args: []string{"version"}, Enabled: true},
+			},
+		},
+		Audit: audit,
+	})
+
+	ctx, cancel := context.WithCancel(context.Background())
+	provider := &cancelingProvider{cancel: cancel}
+
+	_, err = Run(ctx, "hi", provider, Options{
+		SessionID:    "session-cancel",
+		Cwd:          t.TempDir(),
+		ProviderName: "test-provider",
+		Model:        "test-model",
+		Hooks:        dispatcher,
+	})
+	if !errors.Is(err, context.Canceled) {
+		t.Fatalf("Run err = %v, want context.Canceled", err)
+	}
+
+	events, err := audit.ReadEvents()
+	if err != nil {
+		t.Fatalf("ReadEvents: %v", err)
+	}
+	var completedStatus hooks.AuditStatus
+	found := false
+	for _, event := range events {
+		if event.Type == "hook_execution_completed" && event.Event == hooks.EventSessionEnd {
+			completedStatus = event.Status
+			found = true
+		}
+	}
+	if !found {
+		t.Fatalf("no sessionEnd hook_execution_completed audit event, events=%#v", events)
+	}
+	if completedStatus != hooks.AuditCompleted {
+		t.Fatalf("sessionEnd hook status = %q, want %q (hook must actually run despite ctx cancellation)", completedStatus, hooks.AuditCompleted)
+	}
+}
+
 type recordingWebSearchTool struct {
 	calls []map[string]any
 }
@@ -150,6 +304,18 @@ func (tool *sandboxDeniedRetryTool) Run(_ context.Context, args map[string]any)
 	return tools.Result{
 		Status: tools.StatusError,
 		Output: "touch: cannot touch '/home/user/.npm/cache': Read-only file system",
+		ExecutionOutcome: &execution.Outcome{
+			State: execution.StateDenied,
+			Kind:  execution.OutcomeEnforcementDenied,
+			Exit:  &execution.Exit{Code: 1},
+			Denial: &execution.Denial{
+				Capability:  execution.Capability{Kind: execution.CapabilityUnrestricted, Scope: "host"},
+				Source:      execution.DenialSourcePlatformSandbox,
+				Reason:      "sandbox blocked command execution",
+				Recoverable: true,
+				NextAction:  execution.DenialNextActionRequestApproval,
+			},
+		},
 		Meta: map[string]string{
 			"exit_code":                    "1",
 			tools.SandboxLikelyDeniedMeta:  "true",
@@ -725,9 +891,6 @@ func TestRunReportsTruncationFinishReason(t *testing.T) {
 	if result.FinishReason != zeroruntime.FinishReasonLength {
 		t.Fatalf("FinishReason = %q, want %q", result.FinishReason, zeroruntime.FinishReasonLength)
 	}
-	if !result.Truncated() {
-		t.Fatal("Truncated() = false, want true for a length-capped response")
-	}
 	if result.TruncationNotice() == "" {
 		t.Fatal("TruncationNotice() empty for a truncated response")
 	}
@@ -745,7 +908,7 @@ func TestRunNormalCompletionIsNotTruncated(t *testing.T) {
 	if err != nil {
 		t.Fatal(err)
 	}
-	if result.Truncated() || result.TruncationNotice() != "" {
+	if result.TruncationNotice() != "" {
 		t.Fatalf("normal completion reported as truncated: reason=%q", result.FinishReason)
 	}
 }
@@ -822,7 +985,7 @@ func TestRunEmitsUsageEvents(t *testing.T) {
 func TestRunAdvertisesRuntimeToolDefinitions(t *testing.T) {
 	root := t.TempDir()
 	registry := tools.NewRegistry()
-	registry.Register(tools.NewReadFileTool(root))
+	registry.Register(tools.NewScopedReadFileTool(root, nil))
 	provider := &mockProvider{
 		turns: [][]zeroruntime.StreamEvent{{
 			{Type: zeroruntime.StreamEventText, Content: "done"},
@@ -1029,9 +1192,9 @@ func TestRunRequestsPermissionBeforeWebSearchExecution(t *testing.T) {
 func TestRunFiltersAdvertisedTools(t *testing.T) {
 	root := t.TempDir()
 	registry := tools.NewRegistry()
-	registry.Register(tools.NewReadFileTool(root))
-	registry.Register(tools.NewGrepTool(root))
-	registry.Register(tools.NewWriteFileTool(root))
+	registry.Register(tools.NewScopedReadFileTool(root, nil))
+	registry.Register(tools.NewScopedGrepTool(root, nil))
+	registry.Register(tools.NewScopedWriteFileTool(root, nil))
 	provider := &mockProvider{
 		turns: [][]zeroruntime.StreamEvent{{
 			{Type: zeroruntime.StreamEventText, Content: "done"},
@@ -1059,7 +1222,7 @@ func TestRunFiltersAdvertisedTools(t *testing.T) {
 func TestRunRejectsFilteredToolCalls(t *testing.T) {
 	root := t.TempDir()
 	registry := tools.NewRegistry()
-	registry.Register(tools.NewReadFileTool(root))
+	registry.Register(tools.NewScopedReadFileTool(root, nil))
 	provider := &mockProvider{
 		turns: [][]zeroruntime.StreamEvent{
 			{
@@ -1096,7 +1259,7 @@ func TestRunRejectsFilteredToolCalls(t *testing.T) {
 func TestRunRejectsToolCallsOutsideEnabledList(t *testing.T) {
 	root := t.TempDir()
 	registry := tools.NewRegistry()
-	registry.Register(tools.NewReadFileTool(root))
+	registry.Register(tools.NewScopedReadFileTool(root, nil))
 	provider := &mockProvider{
 		turns: [][]zeroruntime.StreamEvent{
 			{
@@ -1134,7 +1297,7 @@ func TestRunExecutesToolCallThroughRegistry(t *testing.T) {
 	root := t.TempDir()
 	writeAgentTestFile(t, filepath.Join(root, "notes.txt"), "alpha\nbeta\n")
 	registry := tools.NewRegistry()
-	registry.Register(tools.NewReadFileTool(root))
+	registry.Register(tools.NewScopedReadFileTool(root, nil))
 	provider := &mockProvider{
 		turns: [][]zeroruntime.StreamEvent{
 			{
@@ -1175,10 +1338,50 @@ func TestRunExecutesToolCallThroughRegistry(t *testing.T) {
 	}
 }
 
+func TestRunPreservesRequestPrefixAcrossTurns(t *testing.T) {
+	root := t.TempDir()
+	writeAgentTestFile(t, filepath.Join(root, "notes.txt"), "alpha\n")
+	registry := tools.NewRegistry()
+	registry.Register(tools.NewReadFileTool(root))
+	provider := &mockProvider{turns: [][]zeroruntime.StreamEvent{
+		{
+			{Type: zeroruntime.StreamEventToolCallStart, ToolCallID: "call-1", ToolName: "read_file"},
+			{Type: zeroruntime.StreamEventToolCallDelta, ToolCallID: "call-1", ArgumentsFragment: `{"path":"notes.txt"}`},
+			{Type: zeroruntime.StreamEventToolCallEnd, ToolCallID: "call-1"},
+			{Type: zeroruntime.StreamEventDone},
+		},
+		{
+			{Type: zeroruntime.StreamEventText, Content: "done"},
+			{Type: zeroruntime.StreamEventDone},
+		},
+	}}
+
+	if _, err := Run(context.Background(), "read notes", provider, Options{
+		Cwd:       root,
+		Registry:  registry,
+		SessionID: "session-stable-prefix",
+	}); err != nil {
+		t.Fatal(err)
+	}
+	if len(provider.requests) != 2 {
+		t.Fatalf("provider requests = %d, want 2", len(provider.requests))
+	}
+	first, second := provider.requests[0], provider.requests[1]
+	if len(second.Messages) < len(first.Messages) || !reflect.DeepEqual(second.Messages[:len(first.Messages)], first.Messages) {
+		t.Fatalf("first request messages must be an exact prefix of the second:\nfirst=%#v\nsecond=%#v", first.Messages, second.Messages)
+	}
+	if !reflect.DeepEqual(first.Tools, second.Tools) {
+		t.Fatalf("tool definitions drifted between turns:\nfirst=%#v\nsecond=%#v", first.Tools, second.Tools)
+	}
+	if first.PromptCacheKey != "session-stable-prefix" || second.PromptCacheKey != first.PromptCacheKey {
+		t.Fatalf("prompt cache key must remain stable: first=%q second=%q", first.PromptCacheKey, second.PromptCacheKey)
+	}
+}
+
 func TestRunSanitizesMalformedToolCallArgumentsBeforeRetry(t *testing.T) {
 	root := t.TempDir()
 	registry := tools.NewRegistry()
-	registry.Register(tools.NewReadFileTool(root))
+	registry.Register(tools.NewScopedReadFileTool(root, nil))
 	provider := &mockProvider{
 		turns: [][]zeroruntime.StreamEvent{
 			{
@@ -1246,7 +1449,7 @@ func TestRunRecoversFirstObjectFromConcatenatedToolArgs(t *testing.T) {
 		t.Fatal(err)
 	}
 	registry := tools.NewRegistry()
-	registry.Register(tools.NewReadFileTool(root))
+	registry.Register(tools.NewScopedReadFileTool(root, nil))
 	provider := &mockProvider{
 		turns: [][]zeroruntime.StreamEvent{
 			{
@@ -1293,8 +1496,8 @@ func TestRunDefersSelfCorrectFeedbackUntilAfterToolBatch(t *testing.T) {
 		t.Fatal(err)
 	}
 	registry := tools.NewRegistry()
-	registry.Register(tools.NewWriteFileTool(root))
-	registry.Register(tools.NewReadFileTool(root))
+	registry.Register(tools.NewScopedWriteFileTool(root, nil))
+	registry.Register(tools.NewScopedReadFileTool(root, nil))
 
 	provider := &mockProvider{
 		turns: [][]zeroruntime.StreamEvent{
@@ -1369,7 +1572,7 @@ func TestRunBatchesSelfCorrectOncePerTurn(t *testing.T) {
 	// after a later call in the same turn supersedes it.
 	root := t.TempDir()
 	registry := tools.NewRegistry()
-	registry.Register(tools.NewWriteFileTool(root))
+	registry.Register(tools.NewScopedWriteFileTool(root, nil))
 
 	provider := &mockProvider{
 		turns: [][]zeroruntime.StreamEvent{
@@ -1426,7 +1629,7 @@ func TestRunBatchesSelfCorrectOncePerTurn(t *testing.T) {
 func TestRunDeniesPromptToolWithoutUnsafePermission(t *testing.T) {
 	root := t.TempDir()
 	registry := tools.NewRegistry()
-	registry.Register(tools.NewWriteFileTool(root))
+	registry.Register(tools.NewScopedWriteFileTool(root, nil))
 	provider := providerCallingWriteFileThenAnswer("write denied")
 	var permissionEvents []PermissionEvent
 
@@ -1469,7 +1672,7 @@ func TestRunDeniesPromptToolWithoutUnsafePermission(t *testing.T) {
 func TestRunRequestsPromptToolPermissionBeforeExecution(t *testing.T) {
 	root := t.TempDir()
 	registry := tools.NewRegistry()
-	registry.Register(tools.NewWriteFileTool(root))
+	registry.Register(tools.NewScopedWriteFileTool(root, nil))
 	provider := providerCallingWriteFileThenAnswer("write approved")
 	var requests []PermissionRequest
 	var permissionEvents []PermissionEvent
@@ -1525,7 +1728,7 @@ func TestRunRequestsPromptToolPermissionBeforeExecution(t *testing.T) {
 func TestRunAllowsWorkspaceWriteWithoutPromptWhenSandboxPolicyPermits(t *testing.T) {
 	root := t.TempDir()
 	registry := tools.NewRegistry()
-	registry.Register(tools.NewWriteFileTool(root))
+	registry.Register(tools.NewScopedWriteFileTool(root, nil))
 	provider := providerCallingWriteFileThenAnswer("write done")
 	var permissionEvents []PermissionEvent
 
@@ -1574,7 +1777,7 @@ func TestRunAllowsWorkspaceWriteWithoutPromptWhenSandboxPolicyPermits(t *testing
 func TestRunDeniesPromptToolWhenPermissionRequestDenied(t *testing.T) {
 	root := t.TempDir()
 	registry := tools.NewRegistry()
-	registry.Register(tools.NewWriteFileTool(root))
+	registry.Register(tools.NewScopedWriteFileTool(root, nil))
 	provider := providerCallingWriteFileThenAnswer("write denied")
 	var requests []PermissionRequest
 	var permissionEvents []PermissionEvent
@@ -1622,7 +1825,7 @@ func TestRunDeniesPromptToolWhenPermissionRequestDenied(t *testing.T) {
 func TestRunAbortsWhenPermissionRequestCanceled(t *testing.T) {
 	root := t.TempDir()
 	registry := tools.NewRegistry()
-	registry.Register(tools.NewWriteFileTool(root))
+	registry.Register(tools.NewScopedWriteFileTool(root, nil))
 	provider := providerCallingWriteFileThenAnswer("write should not continue")
 	var permissionEvents []PermissionEvent
 
@@ -1718,7 +1921,7 @@ func TestRunPersistsAlwaysAllowPermissionDecision(t *testing.T) {
 		t.Fatal(err)
 	}
 	registry := tools.NewRegistry()
-	registry.Register(tools.NewWriteFileTool(root))
+	registry.Register(tools.NewScopedWriteFileTool(root, nil))
 	provider := providerCallingWriteFileThenAnswer("write approved")
 	var permissionEvents []PermissionEvent
 	policy := sandbox.DefaultPolicy()
@@ -1790,7 +1993,7 @@ func TestRunSessionAllowSkipsMatchingPromptWithoutPersistentGrant(t *testing.T)
 		t.Fatal(err)
 	}
 	registry := tools.NewRegistry()
-	registry.Register(tools.NewWriteFileTool(root))
+	registry.Register(tools.NewScopedWriteFileTool(root, nil))
 	provider := &mockProvider{
 		turns: [][]zeroruntime.StreamEvent{
 			{
@@ -1871,7 +2074,7 @@ func TestRunSessionAllowSkipsMatchingPromptWithoutPersistentGrant(t *testing.T)
 func TestRunCommandPrefixApprovalSkipsLaterMatchingBashPrompt(t *testing.T) {
 	root := t.TempDir()
 	registry := tools.NewRegistry()
-	registry.Register(tools.NewBashTool(root))
+	registry.Register(tools.NewScopedBashTool(root, nil))
 	provider := &mockProvider{
 		turns: [][]zeroruntime.StreamEvent{
 			{
@@ -2073,7 +2276,7 @@ func TestRunPersistentCommandPrefixApprovalSkipsFutureSessionPrompt(t *testing.T
 		t.Fatal(err)
 	}
 	registry := tools.NewRegistry()
-	registry.Register(tools.NewBashTool(root))
+	registry.Register(tools.NewScopedBashTool(root, nil))
 	policy := sandbox.DefaultPolicy()
 	policy.Network = sandbox.NetworkAllow
 
@@ -2180,7 +2383,7 @@ func TestRunPersistentCommandPrefixStillPromptsForNetwork(t *testing.T) {
 		t.Fatalf("seed command prefix: %v", err)
 	}
 	registry := tools.NewRegistry()
-	registry.Register(tools.NewBashTool(root))
+	registry.Register(tools.NewScopedBashTool(root, nil))
 	provider := &mockProvider{
 		turns: [][]zeroruntime.StreamEvent{
 			{
@@ -2229,7 +2432,11 @@ func TestRunApprovedNetworkBashPromptAppliesTurnNetworkGrant(t *testing.T) {
 	root := t.TempDir()
 	command := "PATH=.:$PATH curl https://example.com"
 	if runtime.GOOS == "windows" {
-		command = "set PATH=.;%PATH% && curl https://example.com"
+		if windowsTestUsesPowerShell() {
+			command = `$env:PATH = '.;' + $env:PATH; curl.cmd https://example.com`
+		} else {
+			command = "set PATH=.;%PATH% && curl https://example.com"
+		}
 		fakeCurl := filepath.Join(root, "curl.cmd")
 		if err := os.WriteFile(fakeCurl, []byte("@echo fake curl %*\r\n"), 0o755); err != nil {
 			t.Fatal(err)
@@ -2241,7 +2448,7 @@ func TestRunApprovedNetworkBashPromptAppliesTurnNetworkGrant(t *testing.T) {
 		}
 	}
 	registry := tools.NewRegistry()
-	registry.Register(tools.NewBashTool(root))
+	registry.Register(tools.NewScopedBashTool(root, nil))
 	provider := &mockProvider{
 		turns: [][]zeroruntime.StreamEvent{
 			{
@@ -2308,7 +2515,7 @@ func TestRunApprovedNetworkBashPromptAppliesTurnNetworkGrant(t *testing.T) {
 func TestRunDoesNotOfferPrefixApprovalForUnsafeBashCommand(t *testing.T) {
 	root := t.TempDir()
 	registry := tools.NewRegistry()
-	registry.Register(tools.NewBashTool(root))
+	registry.Register(tools.NewScopedBashTool(root, nil))
 	provider := &mockProvider{
 		turns: [][]zeroruntime.StreamEvent{
 			{
@@ -2350,13 +2557,17 @@ func TestRunDoesNotOfferPrefixApprovalForUnsafeBashCommand(t *testing.T) {
 
 func TestRunPromptsForDestructiveShellInsteadOfSandboxDeny(t *testing.T) {
 	root := t.TempDir()
+	command := "echo rm -rf /"
+	if runtime.GOOS == "windows" && windowsTestUsesPowerShell() {
+		command = `Write-Output 'rm -rf /'`
+	}
 	registry := tools.NewRegistry()
-	registry.Register(tools.NewBashTool(root))
+	registry.Register(tools.NewScopedBashTool(root, nil))
 	provider := &mockProvider{
 		turns: [][]zeroruntime.StreamEvent{
 			{
 				{Type: zeroruntime.StreamEventToolCallStart, ToolCallID: "call-1", ToolName: "bash"},
-				{Type: zeroruntime.StreamEventToolCallDelta, ToolCallID: "call-1", ArgumentsFragment: `{"command":"echo rm -rf /"}`},
+				{Type: zeroruntime.StreamEventToolCallDelta, ToolCallID: "call-1", ArgumentsFragment: `{"command":` + quoteJSONString(command) + `}`},
 				{Type: zeroruntime.StreamEventToolCallEnd, ToolCallID: "call-1"},
 				{Type: zeroruntime.StreamEventDone},
 			},
@@ -2420,7 +2631,7 @@ func TestRunAlwaysAllowWithoutSandboxStillAllowsCall(t *testing.T) {
 	// prior code denied it because persistPermissionGrant errors when Sandbox==nil.
 	root := t.TempDir()
 	registry := tools.NewRegistry()
-	registry.Register(tools.NewWriteFileTool(root))
+	registry.Register(tools.NewScopedWriteFileTool(root, nil))
 	provider := providerCallingWriteFileThenAnswer("write approved")
 	var permissionEvents []PermissionEvent
 
@@ -2452,6 +2663,12 @@ func TestRunAlwaysAllowWithoutSandboxStillAllowsCall(t *testing.T) {
 	}
 }
 
+func windowsTestUsesPowerShell() bool {
+	executable, _ := tools.HostShellCommand("")
+	name := strings.ToLower(filepath.Base(executable))
+	return strings.Contains(name, "powershell") || strings.HasPrefix(name, "pwsh")
+}
+
 func containsPermissionDecision(decisions []PermissionDecisionAction, want PermissionDecisionAction) bool {
 	for _, decision := range decisions {
 		if decision == want {
@@ -2484,7 +2701,7 @@ func TestRunCancellationPreservesContextCanceledIdentity(t *testing.T) {
 func TestRunGrantsPromptToolInUnsafeMode(t *testing.T) {
 	root := t.TempDir()
 	registry := tools.NewRegistry()
-	registry.Register(tools.NewWriteFileTool(root))
+	registry.Register(tools.NewScopedWriteFileTool(root, nil))
 	provider := providerCallingWriteFileThenAnswer("write done")
 	var permissionEvents []PermissionEvent
 
@@ -2536,7 +2753,7 @@ func TestRunEmitsPermissionEventForPersistentSandboxGrant(t *testing.T) {
 		t.Fatal(err)
 	}
 	registry := tools.NewRegistry()
-	registry.Register(tools.NewWriteFileTool(root))
+	registry.Register(tools.NewScopedWriteFileTool(root, nil))
 	provider := providerCallingWriteFileThenAnswer("write done")
 	var permissionEvents []PermissionEvent
 
@@ -2715,7 +2932,7 @@ func TestRunAppliesSandboxEvenInUnsafeMode(t *testing.T) {
 	root := t.TempDir()
 	outside := filepath.Join(tempDirOutsideDefaultTemp(t), "escape.txt")
 	registry := tools.NewRegistry()
-	registry.Register(tools.NewWriteFileTool(root))
+	registry.Register(tools.NewScopedWriteFileTool(root, nil))
 	provider := providerCallingWritePathThenAnswer(outside, "sandbox handled")
 	var permissionEvents []PermissionEvent
 
@@ -2764,7 +2981,7 @@ func TestRunStopsAfterMaxTurns(t *testing.T) {
 	root := t.TempDir()
 	writeAgentTestFile(t, filepath.Join(root, "notes.txt"), "alpha")
 	registry := tools.NewRegistry()
-	registry.Register(tools.NewReadFileTool(root))
+	registry.Register(tools.NewScopedReadFileTool(root, nil))
 	provider := &mockProvider{
 		turns: [][]zeroruntime.StreamEvent{{
 			{Type: zeroruntime.StreamEventToolCallStart, ToolCallID: "call-1", ToolName: "read_file"},
@@ -2794,7 +3011,7 @@ func TestRunRequestsFinalAnswerAfterMaxTurns(t *testing.T) {
 	root := t.TempDir()
 	writeAgentTestFile(t, filepath.Join(root, "notes.txt"), "alpha")
 	registry := tools.NewRegistry()
-	registry.Register(tools.NewReadFileTool(root))
+	registry.Register(tools.NewScopedReadFileTool(root, nil))
 	provider := &mockProvider{
 		turns: [][]zeroruntime.StreamEvent{
 			{
@@ -2947,6 +3164,24 @@ func TestGitBranchForPromptResolvesRelativeWorktreeGitdir(t *testing.T) {
 	}
 }
 
+func TestGitBranchForPromptInSubdirectory(t *testing.T) {
+	root := t.TempDir()
+	gitdir := filepath.Join(root, ".git")
+	if err := os.MkdirAll(gitdir, 0o755); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(gitdir, "HEAD"), []byte("ref: refs/heads/main\n"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	subdir := filepath.Join(root, "sub", "dir")
+	if err := os.MkdirAll(subdir, 0o755); err != nil {
+		t.Fatal(err)
+	}
+	if got := gitBranchForPrompt(subdir); got != "main" {
+		t.Fatalf("gitBranchForPrompt in subdirectory = %q, want main", got)
+	}
+}
+
 func TestBuildSystemPromptInjectsWorkspaceContext(t *testing.T) {
 	cwd := t.TempDir()
 	if err := os.WriteFile(filepath.Join(cwd, "AGENTS.md"), []byte("Always run `make lint` before committing."), 0o644); err != nil {
@@ -2967,22 +3202,17 @@ func TestBuildSystemPromptInjectsHostShellContext(t *testing.T) {
 		t.Fatalf("expected operating system in environment block, got %q", prompt)
 	}
 	if runtime.GOOS == "windows" {
-		for _, want := range []string{"Windows cmd.exe syntax", "cwd argument", "MSYS binaries", "grep", "require_escalated", "use double quotes around the value"} {
-			if !strings.Contains(prompt, want) {
-				t.Fatalf("expected Windows shell guidance to mention %q, got %q", want, prompt)
-			}
+		wants := []string{"workdir/cwd", "MSYS binaries"}
+		if windowsTestUsesPowerShell() {
+			wants = append(wants, "PowerShell", "Get-ChildItem", "Select-String", "$env:NAME", "require_escalated")
+		} else {
+			wants = append(wants, "cmd.exe", "double quotes")
 		}
-		// The examples must themselves use the safe (double-quoted) form; a
-		// single-quoted example here would teach the model the exact syntax
-		// that fails under cmd.exe.
-		for _, want := range []string{`--jq ".a | b"`, `-run "A|B"`} {
+		for _, want := range wants {
 			if !strings.Contains(prompt, want) {
-				t.Fatalf("expected double-quoted example %q in Windows shell guidance, got %q", want, prompt)
+				t.Fatalf("expected selected Windows shell guidance to mention %q, got %q", want, prompt)
 			}
 		}
-		if strings.Contains(prompt, `'.a | b'`) || strings.Contains(prompt, `'A|B'`) {
-			t.Fatalf("Windows shell guidance must not show the unsafe single-quoted form, got %q", prompt)
-		}
 	} else if !strings.Contains(prompt, "/bin/sh syntax") {
 		t.Fatalf("expected POSIX shell guidance in prompt, got %q", prompt)
 	}
@@ -3035,7 +3265,7 @@ func TestBuildSystemPromptAllowsSpecModeOverride(t *testing.T) {
 func TestSpecDraftAdvertisesOnlySafeDraftTools(t *testing.T) {
 	root := t.TempDir()
 	registry := tools.NewRegistry()
-	for _, tool := range tools.CoreTools(root) {
+	for _, tool := range tools.CoreToolsScoped(root, nil) {
 		registry.Register(tool)
 	}
 	specmode.RegisterDraftTools(registry, root, nil)
@@ -3072,7 +3302,7 @@ func TestSpecDraftAdvertisesOnlySafeDraftTools(t *testing.T) {
 func TestSpecDraftDeniesHiddenToolCalls(t *testing.T) {
 	root := t.TempDir()
 	registry := tools.NewRegistry()
-	registry.Register(tools.NewWriteFileTool(root))
+	registry.Register(tools.NewScopedWriteFileTool(root, nil))
 	provider := providerCallingWriteFileThenAnswer("done")
 
 	result, err := Run(context.Background(), "draft", provider, Options{
@@ -3105,7 +3335,7 @@ func TestSpecDraftDeniesHiddenToolCalls(t *testing.T) {
 func TestSpecDraftDeniesBashToolCalls(t *testing.T) {
 	root := t.TempDir()
 	registry := tools.NewRegistry()
-	registry.Register(tools.NewBashTool(root))
+	registry.Register(tools.NewScopedBashTool(root, nil))
 	provider := &mockProvider{
 		turns: [][]zeroruntime.StreamEvent{
 			{
@@ -3246,7 +3476,7 @@ func TestRunSurfacesDroppedToolCallAlongsideValidCall(t *testing.T) {
 		t.Fatal(err)
 	}
 	registry := tools.NewRegistry()
-	registry.Register(tools.NewReadFileTool(root))
+	registry.Register(tools.NewScopedReadFileTool(root, nil))
 
 	provider := &mockProvider{
 		turns: [][]zeroruntime.StreamEvent{
@@ -3443,3 +3673,73 @@ func TestRunDoesNotFlagCleanToolOutput(t *testing.T) {
 		t.Errorf("clean output should not get a reminder, got %q", captured.Output)
 	}
 }
+
+// TestRunTracingWrapperStampsUsage verifies the per-turn tracing setup in Run:
+// a wired recorder is Started, OnUsage is wrapped so it still forwards to the
+// caller's callback AND stamps token counters, and the run completes.
+func TestRunTracingWrapperStampsUsage(t *testing.T) {
+	provider := &mockProvider{turns: [][]zeroruntime.StreamEvent{{
+		{Type: zeroruntime.StreamEventUsage, Usage: zeroruntime.Usage{InputTokens: 100, CachedInputTokens: 20, OutputTokens: 40}},
+		{Type: zeroruntime.StreamEventText, Content: "done"},
+		{Type: zeroruntime.StreamEventDone},
+	}}}
+	onUsageCalls := 0
+	rec := trace.NewRecorder("tracing-session", "run-1", "test")
+	if _, err := Run(context.Background(), "hi", provider, Options{
+		SessionID:    "tracing-session",
+		Cwd:          t.TempDir(),
+		ProviderName: "test-provider",
+		Model:        "test-model",
+		Trace:        rec,
+		OnUsage:      func(Usage) { onUsageCalls++ },
+	}); err != nil {
+		t.Fatalf("Run: %v", err)
+	}
+	tr := rec.Finish()
+	if tr.StartedAt.IsZero() {
+		t.Fatal("tracing wrapper did not Start the recorder")
+	}
+	// FirstTokenAt must stamp even though no OnText/OnReasoning user callback is
+	// set: a headless traced run (e.g. `zero exec --trace`) sets Trace but no UI
+	// callbacks, so the loop installs trace-only forwarding handlers to capture
+	// TTFT. Without them FirstTokenAt stays zero and the trace loses its signal.
+	if tr.FirstTokenAt.IsZero() {
+		t.Fatal("FirstTokenAt not stamped for a traced run with no OnText/OnReasoning callbacks")
+	}
+	if got := tr.Counter(trace.CounterInputTokens); got != 100 {
+		t.Fatalf("input token counter = %d, want 100", got)
+	}
+	if got := tr.Counter(trace.CounterCachedInputTokens); got != 20 {
+		t.Fatalf("cached input token counter = %d, want 20", got)
+	}
+	if got := tr.Counter(trace.CounterOutputTokens); got != 40 {
+		t.Fatalf("output token counter = %d, want 40", got)
+	}
+	if onUsageCalls == 0 {
+		t.Fatal("wrapped OnUsage did not forward to the caller's callback")
+	}
+}
+
+// TestRunNilTraceForwardsUsage verifies a nil recorder leaves the loop
+// byte-identical: OnUsage is not wrapped, so the caller's callback still fires
+// and nothing panics.
+func TestRunNilTraceForwardsUsage(t *testing.T) {
+	provider := &mockProvider{turns: [][]zeroruntime.StreamEvent{{
+		{Type: zeroruntime.StreamEventUsage, Usage: zeroruntime.Usage{InputTokens: 7}},
+		{Type: zeroruntime.StreamEventText, Content: "done"},
+		{Type: zeroruntime.StreamEventDone},
+	}}}
+	onUsageCalls := 0
+	if _, err := Run(context.Background(), "hi", provider, Options{
+		SessionID:    "nil-trace-session",
+		Cwd:          t.TempDir(),
+		ProviderName: "test-provider",
+		Model:        "test-model",
+		OnUsage:      func(Usage) { onUsageCalls++ },
+	}); err != nil {
+		t.Fatalf("Run: %v", err)
+	}
+	if onUsageCalls == 0 {
+		t.Fatal("OnUsage not forwarded when Trace is nil")
+	}
+}

internal/agent/output_budget_propagation_test.go

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

internal/agent/parallel_tools.go

@@ -10,12 +10,23 @@ import (
 // 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
-// auto-allowed read-only calls is executed concurrently instead; results are
+// 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
@@ -28,28 +39,117 @@ type precomputedToolResult struct {
 }
 
 // parallelSafeToolCall reports whether call may run concurrently with its
-// neighbors: the tool must exist, be side-effect-free (SideEffectRead), and be
-// auto-allowed for these args, so no interactive prompt or workspace mutation
-// is on the hot path. Loop-intercepted tools (ask_user, request_permissions)
-// and tool_search (mutates the deferred-tool set) stay sequential.
+// 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 || tool.Safety().SideEffect != tools.SideEffectRead {
+	if !found {
 		return false
 	}
-	args := map[string]any{}
-	if call.Arguments != "" {
-		if err := decodeToolArguments(call.Arguments, &args); err != nil {
-			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

internal/agent/parallel_tools_test.go

@@ -57,6 +57,20 @@ func (tool *probeTool) Parameters() tools.Schema {
 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()
@@ -92,6 +106,11 @@ 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{}) {
@@ -109,6 +128,91 @@ func TestParallelSafeToolCall(t *testing.T) {
 	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) {

internal/agent/partition_cache_stable_test.go

@@ -14,7 +14,7 @@ func TestPartitionToolsActiveAppendsLoadedToolAfterEagerBlock(t *testing.T) {
 	// "read_file".
 	root := t.TempDir()
 	registry := tools.NewRegistry()
-	registry.Register(tools.NewReadFileTool(root)) // non-deferred, "read_file"
+	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"})

internal/agent/permission_additional_perms_test.go

@@ -68,3 +68,61 @@ func TestBuildPermissionEventKeepsAllowForOrdinaryAllowedCommand(t *testing.T) {
 		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/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

@@ -14,13 +14,13 @@ import (
 // deliberately) or trimmed — the per-turn floor should not creep up silently.
 //
 // Measured baselines (2026-07): base system prompt ~3160 tokens; the tools a normal
-// (auto permission-mode) interactive turn sends ~2430 tokens. The tool figure is
-// the auto-mode advertised set, not the full core registry — higher-risk tools that
-// only advertise in other modes are excluded, matching what a real turn actually
-// pays.
+// (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
-	maxEagerToolSchemaTokens  = 2700
+	maxEagerToolSchemaTokens  = 3550
 )
 
 func TestSystemPromptTokenBudget(t *testing.T) {

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 +1,249 @@
+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")
+	}
+}
+
+// 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/reconnect.go

@@ -8,6 +8,7 @@ import (
 	"time"
 
 	"github.com/Gitlawb/zero/internal/errhint"
+	"github.com/Gitlawb/zero/internal/trace"
 	"github.com/Gitlawb/zero/internal/zeroruntime"
 )
 
@@ -74,14 +75,21 @@ func stallRetryNoticeFor(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)
 		}
@@ -90,6 +98,9 @@ func streamWithReconnect(ctx context.Context, provider Provider, request zerorun
 		}
 		stream, err = provider.StreamCompletion(ctx, request)
 		if err == nil {
+			if recorder != nil {
+				recorder.Counter(trace.CounterModelRequests, 1)
+			}
 			return stream, nil
 		}
 	}

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")
 	}
 }

internal/agent/selfcorrect.go

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

internal/agent/system_prompt.go

@@ -11,6 +11,7 @@ import (
 
 	"github.com/Gitlawb/zero/internal/config"
 	"github.com/Gitlawb/zero/internal/repomap"
+	"github.com/Gitlawb/zero/internal/tools"
 	"github.com/Gitlawb/zero/internal/workspaceseed"
 )
 
@@ -70,6 +71,22 @@ const (
 // guidelines), and the safety confirmation policy. It is built once per run so
 // every turn shares one (cacheable) system turn.
 func buildSystemPrompt(options Options) string {
+	return buildSystemPromptParts(options).prompt
+}
+
+// systemPromptParts retains both the exact joined prompt and the diagnostic
+// sections used by prompt-prefix tracing. Building these together prevents the
+// tracing path from re-reading workspace state and guarantees that the hash is
+// computed from the same prompt bytes placed in the request.
+type systemPromptParts struct {
+	prompt             string
+	baseInstructions   string
+	confirmationPolicy string
+	projectContext     string
+	skills             string
+}
+
+func buildSystemPromptParts(options Options) systemPromptParts {
 	core := strings.TrimSpace(options.SystemPrompt)
 	if core == "" {
 		core = strings.TrimSpace(coreSystemPrompt)
@@ -97,22 +114,31 @@ func buildSystemPrompt(options Options) string {
 	if user := userGuidelines(); user != "" {
 		sections = append(sections, user)
 	}
-	if ws := workspaceContext(options.Cwd); ws != "" {
-		sections = append(sections, ws)
+	project := workspaceContext(options.Cwd)
+	if project != "" {
+		sections = append(sections, project)
 	}
 	if delegation := specialistDelegationContext(options); delegation != "" {
 		sections = append(sections, delegation)
 	}
-	if skillsBlock := skillsContext(options); skillsBlock != "" {
+	skillsBlock := skillsContext(options)
+	if skillsBlock != "" {
 		sections = append(sections, skillsBlock)
 	}
 	if style := responseStyleContext(options); style != "" {
 		sections = append(sections, style)
 	}
-	if policy := strings.TrimSpace(confirmationPolicy); policy != "" {
+	policy := strings.TrimSpace(confirmationPolicy)
+	if policy != "" {
 		sections = append(sections, policy)
 	}
-	return strings.Join(sections, "\n\n")
+	return systemPromptParts{
+		prompt:             strings.Join(sections, "\n\n"),
+		baseInstructions:   core,
+		confirmationPolicy: policy,
+		projectContext:     project,
+		skills:             skillsBlock,
+	}
 }
 
 // responseStyleContext renders the operator-selected reply style (TUI /style) as
@@ -285,17 +311,13 @@ func workspaceContext(cwd string) string {
 	b.WriteString("<environment>\n")
 	b.WriteString("Working directory: " + cwd + "\n")
 	b.WriteString("Operating system: " + runtime.GOOS + "\n")
-	if runtime.GOOS == "windows" {
-		b.WriteString("Shell syntax: Windows cmd.exe syntax for exec_command/bash tools. To put | & > < etc inside an arg value, use double quotes around the value, not single quotes (single quotes do not protect metachars in cmd.exe): gh --jq \".a | b\", go test -run \"A|B\". Do not pipe to or invoke POSIX coreutils from Git for Windows (usr\\bin head/grep/tail/cat/...): they are MSYS binaries and fail under the write-restricted sandbox; use native Zero tools (grep, read_file, list_directory, glob) or cmd.exe findstr/more instead, or sandbox_permissions require_escalated only when host-level execution is truly required. Prefer the workdir/cwd argument over cd when changing directories.\n")
-	} else {
-		b.WriteString("Shell syntax: /bin/sh syntax for exec_command/bash tools; prefer the workdir/cwd argument instead of cd when changing directories.\n")
-	}
+	b.WriteString(tools.HostShellEnvironmentGuidance() + "\n")
 	if branch := gitBranchForPrompt(cwd); branch != "" {
 		b.WriteString("Git branch: " + branch + "\n")
 	}
 	b.WriteString("</environment>")
 
-	b.WriteString(projectGuidelines(cwd, findProjectGitRoot(cwd)))
+	b.WriteString(projectGuidelines(cwd, FindProjectGitRoot(cwd)))
 	if repoMap := repoMapContext(cwd); repoMap != "" {
 		b.WriteString("\n\n## Repo map\n\n" + repoMap)
 	}
@@ -542,17 +564,17 @@ func resolveDirCaseInsensitive(target, anchor string) (string, bool) {
 	return cur, true
 }
 
-// findProjectGitRoot returns the nearest ancestor of cwd that contains a
+// FindProjectGitRoot returns the nearest ancestor of cwd that contains a
 // .git entry (file or directory). Returns "" when no git root is found, so
 // the caller can fall back to cwd-only lookup.
-func findProjectGitRoot(cwd string) string {
+func FindProjectGitRoot(cwd string) string {
 	cwd = strings.TrimSpace(cwd)
 	if cwd == "" {
 		return ""
 	}
 	cur := cwd
 	for {
-		if hasGitMetadata(cur) {
+		if HasGitMetadata(cur) {
 			return cur
 		}
 		parent := filepath.Dir(cur)
@@ -563,7 +585,7 @@ func findProjectGitRoot(cwd string) string {
 	}
 }
 
-func hasGitMetadata(dir string) bool {
+func HasGitMetadata(dir string) bool {
 	gitPath := filepath.Join(dir, ".git")
 	info, err := os.Stat(gitPath)
 	if err != nil {
@@ -619,7 +641,11 @@ func repoMapContext(cwd string) string {
 // cwd, handling both a regular checkout (.git dir) and a worktree (.git file).
 // Returns "" on any problem — the prompt simply omits the branch segment.
 func gitBranchForPrompt(cwd string) string {
-	gitPath := filepath.Join(cwd, ".git")
+	gitRoot := FindProjectGitRoot(cwd)
+	if gitRoot == "" {
+		return ""
+	}
+	gitPath := filepath.Join(gitRoot, ".git")
 	info, err := os.Stat(gitPath)
 	if err != nil {
 		return ""
@@ -635,10 +661,10 @@ func gitBranchForPrompt(cwd string) string {
 			return ""
 		}
 		// In worktree mode the gitdir is often RELATIVE (e.g.
-		// "gitdir: ../.git/worktrees/<name>") — resolve it against cwd, not the
+		// "gitdir: ../.git/worktrees/<name>") — resolve it against the worktree root (gitRoot), not the
 		// process working directory, or HEAD lookup fails and we drop the branch.
 		if !filepath.IsAbs(dir) {
-			dir = filepath.Join(cwd, dir)
+			dir = filepath.Join(gitRoot, dir)
 		}
 		headPath = filepath.Join(dir, "HEAD")
 	}

internal/agent/system_prompt.md

@@ -100,6 +100,9 @@ work.
 - Use exec_command with `tty: true` for interactive terminal-style commands that
   need stdin beyond Ctrl-C. `/ps` and `/stop` are user-facing TUI commands; when
   you need to clean up a running foreground command yourself, use write_stdin.
+- write_stdin's session_id is only ever an id returned by a still-running
+  exec_command; never guess or probe ids. If you have no such session, start one
+  with exec_command, or use write_file/edit_file/apply_patch for file changes.
 - write_stdin with empty input polls an existing exec_command session, and
   `\u0003` interrupts it. Sending other stdin bytes may require approval because
   it can drive the running process beyond the original command. Non-tty sessions

internal/agent/system_prompt_test.go

@@ -424,7 +424,7 @@ func TestFindProjectGitRootIgnoresEmptyGitDirectory(t *testing.T) {
 	if err := os.MkdirAll(leaf, 0o755); err != nil {
 		t.Fatal(err)
 	}
-	if got := findProjectGitRoot(leaf); got != "" {
-		t.Fatalf("findProjectGitRoot = %q, want empty for invalid .git directory", got)
+	if got := FindProjectGitRoot(leaf); got != "" {
+		t.Fatalf("FindProjectGitRoot = %q, want empty for invalid .git directory", got)
 	}
 }

internal/agent/task_state.go

@@ -0,0 +1,359 @@
+package agent
+
+import (
+	"encoding/json"
+	"reflect"
+	"sort"
+	"strings"
+
+	"github.com/Gitlawb/zero/internal/tools"
+	"github.com/Gitlawb/zero/internal/trace"
+	"github.com/Gitlawb/zero/internal/zeroruntime"
+)
+
+type taskStatus string
+
+const (
+	taskStatusActive     taskStatus = "active"
+	taskStatusComplete   taskStatus = "complete"
+	taskStatusIncomplete taskStatus = "incomplete"
+)
+
+type taskPlanParity string
+
+const (
+	taskPlanParityUnknown  taskPlanParity = "unknown"
+	taskPlanParityMatch    taskPlanParity = "match"
+	taskPlanParityMismatch taskPlanParity = "mismatch"
+)
+
+type taskPlanItem struct {
+	Content string `json:"content"`
+	Status  string `json:"status"`
+	Notes   string `json:"notes,omitempty"`
+}
+
+type taskPlanState struct {
+	Items      []taskPlanItem `json:"items,omitempty"`
+	Pending    int            `json:"pending"`
+	InProgress int            `json:"in_progress"`
+	Completed  int            `json:"completed"`
+	Failed     int            `json:"failed"`
+}
+
+type taskToolState struct {
+	Succeeded int `json:"succeeded"`
+	Failed    int `json:"failed"`
+}
+
+type taskVerificationState struct {
+	Passed      int     `json:"passed"`
+	Failed      int     `json:"failed"`
+	LastOutcome Outcome `json:"last_outcome,omitempty"`
+}
+
+type taskStateSnapshot struct {
+	Revision           int                   `json:"revision"`
+	Objective          string                `json:"objective"`
+	Status             taskStatus            `json:"status"`
+	Plan               taskPlanState         `json:"plan"`
+	Tools              taskToolState         `json:"tools"`
+	Verification       taskVerificationState `json:"verification"`
+	ChangedFiles       []string              `json:"changed_files,omitempty"`
+	CompletionDecision CompletionDecision    `json:"completion_decision,omitempty"`
+	CompletionReason   string                `json:"completion_reason,omitempty"`
+	PlanParity         taskPlanParity        `json:"plan_parity"`
+}
+
+type taskStateEventKind int
+
+const (
+	taskStateEventPlan taskStateEventKind = iota + 1
+	taskStateEventToolResult
+	taskStateEventVerification
+	taskStateEventCompletion
+)
+
+type taskStateEvent struct {
+	kind         taskStateEventKind
+	arguments    string
+	toolResult   ToolResult
+	verification Outcome
+	completion   completionEvaluation
+}
+
+// taskState is a deterministic projection of facts the loop already observes.
+// It does not initiate work or replace the transcript; consumers must check
+// transcript parity before using its compact representation.
+type taskState struct {
+	snapshotValue taskStateSnapshot
+	changedFiles  map[string]struct{}
+	recorder      *trace.Recorder
+	planObserved  bool
+}
+
+func newTaskState(objective string, recorder *trace.Recorder) *taskState {
+	state := &taskState{
+		snapshotValue: taskStateSnapshot{
+			Objective:  strings.TrimSpace(objective),
+			Status:     taskStatusActive,
+			PlanParity: taskPlanParityUnknown,
+		},
+		changedFiles: map[string]struct{}{},
+	}
+	state.recorder = recorder
+	return state
+}
+
+func (state *taskState) observe(event taskStateEvent) bool {
+	if state == nil {
+		return false
+	}
+	changed := false
+	switch event.kind {
+	case taskStateEventPlan:
+		plan, ok := parseTaskPlan(event.arguments)
+		if !ok {
+			return false
+		}
+		state.snapshotValue.Plan = summarizeTaskPlan(plan)
+		state.planObserved = true
+		state.markActive()
+		changed = true
+	case taskStateEventToolResult:
+		state.markActive()
+		if event.toolResult.Status == tools.StatusOK {
+			state.snapshotValue.Tools.Succeeded++
+		} else {
+			state.snapshotValue.Tools.Failed++
+		}
+		for _, path := range event.toolResult.ChangedFiles {
+			path = strings.TrimSpace(path)
+			if path != "" {
+				state.changedFiles[path] = struct{}{}
+			}
+		}
+		state.snapshotValue.ChangedFiles = sortedKeys(state.changedFiles)
+		changed = true
+	case taskStateEventVerification:
+		if event.verification == OutcomeDisabled {
+			return false
+		}
+		state.markActive()
+		state.snapshotValue.Verification.LastOutcome = event.verification
+		switch event.verification {
+		case OutcomePassed:
+			state.snapshotValue.Verification.Passed++
+		case OutcomeCorrecting, OutcomeReported, OutcomeAborted:
+			state.snapshotValue.Verification.Failed++
+		}
+		changed = true
+	case taskStateEventCompletion:
+		state.snapshotValue.CompletionDecision = event.completion.Decision
+		state.snapshotValue.CompletionReason = event.completion.Reason
+		switch event.completion.Decision {
+		case CompletionComplete:
+			state.snapshotValue.Status = taskStatusComplete
+		case CompletionIncomplete:
+			state.snapshotValue.Status = taskStatusIncomplete
+		default:
+			state.snapshotValue.Status = taskStatusActive
+		}
+		changed = true
+	}
+	if changed {
+		state.snapshotValue.Revision++
+		if event.kind != taskStateEventToolResult {
+			state.emit()
+		}
+	}
+	return changed
+}
+
+func (state *taskState) markActive() {
+	state.snapshotValue.Status = taskStatusActive
+	state.snapshotValue.CompletionDecision = ""
+	state.snapshotValue.CompletionReason = ""
+}
+
+func (state *taskState) snapshot() taskStateSnapshot {
+	if state == nil {
+		return taskStateSnapshot{}
+	}
+	snapshot := state.snapshotValue
+	snapshot.Plan.Items = append([]taskPlanItem(nil), state.snapshotValue.Plan.Items...)
+	snapshot.ChangedFiles = append([]string(nil), state.snapshotValue.ChangedFiles...)
+	return snapshot
+}
+
+// observePlanParity compares only the latest plan projection with the plan tool
+// calls still present in messages. It mutates the snapshot and emits when the
+// parity value changes; objective, tool, and verification fields are not part of
+// this comparison.
+func (state *taskState) observePlanParity(messages []zeroruntime.Message) taskPlanParity {
+	if state == nil {
+		return taskPlanParityUnknown
+	}
+	transcriptPlan, found, valid := latestTaskPlan(messages)
+	tracked := state.snapshotValue.Plan.Items
+	parity := taskPlanParityMatch
+	if !valid || found != state.planObserved || (found && !reflect.DeepEqual(transcriptPlan, tracked)) {
+		parity = taskPlanParityMismatch
+	}
+	if state.snapshotValue.PlanParity != parity {
+		state.snapshotValue.PlanParity = parity
+		state.snapshotValue.Revision++
+		state.emit()
+	}
+	return parity
+}
+
+type completionContext struct {
+	Objective             string
+	PlanPending           bool
+	PlanMatchesTranscript bool
+}
+
+func (state *taskState) completionContext(messages []zeroruntime.Message, transcriptPlanPending bool) completionContext {
+	context := completionContext{PlanPending: transcriptPlanPending}
+	if state == nil {
+		return context
+	}
+	context.Objective = state.snapshotValue.Objective
+	context.PlanMatchesTranscript = state.observePlanParity(messages) == taskPlanParityMatch
+	if context.PlanMatchesTranscript {
+		context.PlanPending = state.snapshotValue.Plan.Pending+state.snapshotValue.Plan.InProgress > 0
+	}
+	return context
+}
+
+func (state *taskState) snapshotForCompaction(messages []zeroruntime.Message) *taskStateSnapshot {
+	if state == nil {
+		return nil
+	}
+	state.observePlanParity(messages)
+	snapshot := state.snapshot()
+	return &snapshot
+}
+
+func (state *taskState) emit() {
+	if state == nil || state.recorder == nil {
+		return
+	}
+	snapshot := state.snapshotValue
+	state.recorder.EmitTaskState(trace.TaskStateEvent{
+		Revision:            snapshot.Revision,
+		Status:              string(snapshot.Status),
+		PlanPending:         snapshot.Plan.Pending,
+		PlanInProgress:      snapshot.Plan.InProgress,
+		PlanCompleted:       snapshot.Plan.Completed,
+		PlanFailed:          snapshot.Plan.Failed,
+		ToolsSucceeded:      snapshot.Tools.Succeeded,
+		ToolsFailed:         snapshot.Tools.Failed,
+		VerificationPassed:  snapshot.Verification.Passed,
+		VerificationFailed:  snapshot.Verification.Failed,
+		VerificationOutcome: string(snapshot.Verification.LastOutcome),
+		ChangedFileCount:    len(snapshot.ChangedFiles),
+		CompletionDecision:  string(snapshot.CompletionDecision),
+		PlanParity:          string(snapshot.PlanParity),
+	})
+}
+
+func parseTaskPlan(arguments string) ([]taskPlanItem, bool) {
+	var object map[string]json.RawMessage
+	if json.Unmarshal([]byte(strings.TrimSpace(arguments)), &object) != nil {
+		return nil, false
+	}
+	rawPlan, ok := object["plan"]
+	if !ok || string(rawPlan) == "null" {
+		return nil, false
+	}
+	var parsed []struct {
+		ID      *string `json:"id"`
+		Content string  `json:"content"`
+		Status  string  `json:"status"`
+		Notes   string  `json:"notes"`
+	}
+	if json.Unmarshal(rawPlan, &parsed) != nil {
+		return nil, false
+	}
+	plan := make([]taskPlanItem, 0, len(parsed))
+	for _, raw := range parsed {
+		if raw.Content == "" {
+			return nil, false
+		}
+		plan = append(plan, taskPlanItem{
+			Content: raw.Content,
+			Status:  normalizeTaskPlanStatus(raw.Status),
+			Notes:   raw.Notes,
+		})
+	}
+	lastInProgress := -1
+	for index := range plan {
+		if plan[index].Status == "in_progress" {
+			if lastInProgress >= 0 {
+				plan[lastInProgress].Status = "completed"
+			}
+			lastInProgress = index
+		}
+	}
+	if len(plan) == 0 {
+		return nil, true
+	}
+	return plan, true
+}
+
+func normalizeTaskPlanStatus(status string) string {
+	switch strings.ToLower(strings.TrimSpace(status)) {
+	case "completed", "complete", "done", "finished", "resolved", "✓", "x", "[x]":
+		return "completed"
+	case "in_progress", "in-progress", "inprogress", "in progress", "active", "doing", "started", "current", "wip", "ongoing", "running":
+		return "in_progress"
+	case "failed", "fail", "error", "errored", "blocked", "cancelled", "canceled", "abandoned", "skipped":
+		return "failed"
+	default:
+		return "pending"
+	}
+}
+
+func summarizeTaskPlan(items []taskPlanItem) taskPlanState {
+	plan := taskPlanState{Items: append([]taskPlanItem(nil), items...)}
+	for _, item := range items {
+		switch item.Status {
+		case "completed":
+			plan.Completed++
+		case "in_progress":
+			plan.InProgress++
+		case "failed":
+			plan.Failed++
+		default:
+			plan.Pending++
+		}
+	}
+	return plan
+}
+
+func latestTaskPlan(messages []zeroruntime.Message) ([]taskPlanItem, bool, bool) {
+	for i := len(messages) - 1; i >= 0; i-- {
+		for j := len(messages[i].ToolCalls) - 1; j >= 0; j-- {
+			call := messages[i].ToolCalls[j]
+			if call.Name != planToolName {
+				continue
+			}
+			plan, ok := parseTaskPlan(call.Arguments)
+			return plan, true, ok
+		}
+	}
+	return nil, false, true
+}
+
+func sortedKeys(values map[string]struct{}) []string {
+	keys := make([]string, 0, len(values))
+	for value := range values {
+		keys = append(keys, value)
+	}
+	// Paths are evidence, not chronology. Stable ordering makes replay snapshots
+	// deterministic even if events originate from parallel read batches later.
+	sort.Strings(keys)
+	return keys
+}

internal/agent/task_state_test.go

@@ -0,0 +1,204 @@
+package agent
+
+import (
+	"context"
+	"reflect"
+	"testing"
+
+	"github.com/Gitlawb/zero/internal/tools"
+	"github.com/Gitlawb/zero/internal/trace"
+	"github.com/Gitlawb/zero/internal/zeroruntime"
+)
+
+func TestTaskStateReplayIsDeterministic(t *testing.T) {
+	events := []taskStateEvent{
+		{kind: taskStateEventPlan, arguments: `{"plan":[{"content":"inspect","status":"done"},{"content":"implement","status":"in progress"},{"content":"verify"}]}`},
+		{kind: taskStateEventToolResult, toolResult: ToolResult{Name: "apply_patch", Status: tools.StatusOK, ChangedFiles: []string{"a.go", "a.go", "b.go"}}},
+		{kind: taskStateEventToolResult, toolResult: ToolResult{Name: "go test", Status: tools.StatusError}},
+		{kind: taskStateEventVerification, verification: OutcomePassed},
+		{kind: taskStateEventCompletion, completion: completionEvaluation{Decision: CompletionUncertain, Reason: "pending work"}},
+	}
+
+	first := newTaskState("ship the change", nil)
+	second := newTaskState("ship the change", nil)
+	for _, event := range events {
+		first.observe(event)
+		second.observe(event)
+	}
+
+	if !reflect.DeepEqual(first.snapshot(), second.snapshot()) {
+		t.Fatalf("replaying the same events produced different snapshots:\nfirst:  %#v\nsecond: %#v", first.snapshot(), second.snapshot())
+	}
+	got := first.snapshot()
+	if got.Status != taskStatusActive || got.Plan.Pending != 1 || got.Plan.InProgress != 1 || got.Plan.Completed != 1 {
+		t.Fatalf("unexpected plan snapshot: %#v", got)
+	}
+	if got.Tools.Succeeded != 1 || got.Tools.Failed != 1 || len(got.ChangedFiles) != 2 {
+		t.Fatalf("unexpected evidence snapshot: %#v", got)
+	}
+	if got.Verification.Passed != 1 || got.Verification.Failed != 0 || got.Verification.LastOutcome != OutcomePassed {
+		t.Fatalf("unexpected verification snapshot: %#v", got.Verification)
+	}
+}
+
+func TestTaskStateCoalescesToolResultsIntoNextTraceEvent(t *testing.T) {
+	recorder := trace.NewRecorder("session", "run", "")
+	state := newTaskState("objective", recorder)
+	for range 20 {
+		state.observe(taskStateEvent{kind: taskStateEventToolResult, toolResult: ToolResult{Status: tools.StatusOK}})
+	}
+	if events := recorder.Finish().TaskStates; len(events) != 0 {
+		t.Fatalf("tool results should be coalesced, got %d trace events", len(events))
+	}
+
+	recorder = trace.NewRecorder("session", "run-2", "")
+	state = newTaskState("objective", recorder)
+	for range 20 {
+		state.observe(taskStateEvent{kind: taskStateEventToolResult, toolResult: ToolResult{Status: tools.StatusOK}})
+	}
+	state.observe(taskStateEvent{kind: taskStateEventCompletion, completion: completionEvaluation{Decision: CompletionComplete}})
+	events := recorder.Finish().TaskStates
+	if len(events) != 1 || events[0].ToolsSucceeded != 20 {
+		t.Fatalf("coalesced tool total missing from completion snapshot: %#v", events)
+	}
+}
+
+func TestRunEmitsTaskStateFromExistingLoopEvents(t *testing.T) {
+	registry := tools.NewRegistry()
+	registry.Register(tools.NewUpdatePlanTool())
+	provider := &mockProvider{turns: [][]zeroruntime.StreamEvent{
+		{
+			{Type: zeroruntime.StreamEventToolCallStart, ToolCallID: "plan-1", ToolName: planToolName},
+			{Type: zeroruntime.StreamEventToolCallDelta, ToolCallID: "plan-1", ArgumentsFragment: `{"plan":[{"content":"implement","status":"completed"}]}`},
+			{Type: zeroruntime.StreamEventToolCallEnd, ToolCallID: "plan-1"},
+			{Type: zeroruntime.StreamEventDone},
+		},
+		{
+			{Type: zeroruntime.StreamEventText, Content: "done"},
+			{Type: zeroruntime.StreamEventDone},
+		},
+	}}
+	recorder := trace.NewRecorder("session", "run", "")
+
+	result, err := Run(context.Background(), "implement the change", provider, Options{
+		Registry: registry,
+		Trace:    recorder,
+	})
+	if err != nil {
+		t.Fatalf("Run: %v", err)
+	}
+	if result.FinalAnswer != "done" {
+		t.Fatalf("final answer = %q, want done", result.FinalAnswer)
+	}
+	events := recorder.Finish().TaskStates
+	if len(events) < 3 {
+		t.Fatalf("expected plan, tool, completion, and parity snapshots, got %#v", events)
+	}
+	last := events[len(events)-1]
+	if last.Status != string(taskStatusComplete) || last.PlanCompleted != 1 || last.ToolsSucceeded != 1 || last.PlanParity != string(taskPlanParityMatch) {
+		t.Fatalf("unexpected final task snapshot: %#v", last)
+	}
+}
+
+func TestTaskStateSnapshotIsImmutable(t *testing.T) {
+	state := newTaskState("objective", nil)
+	state.observe(taskStateEvent{kind: taskStateEventPlan, arguments: `{"plan":[{"content":"one","status":"pending"}]}`})
+	state.observe(taskStateEvent{kind: taskStateEventToolResult, toolResult: ToolResult{Status: tools.StatusOK, ChangedFiles: []string{"one.go"}}})
+
+	snapshot := state.snapshot()
+	snapshot.Plan.Items[0].Content = "mutated"
+	snapshot.ChangedFiles[0] = "mutated.go"
+
+	fresh := state.snapshot()
+	if fresh.Plan.Items[0].Content != "one" || fresh.ChangedFiles[0] != "one.go" {
+		t.Fatalf("snapshot mutation leaked into task state: %#v", fresh)
+	}
+}
+
+func TestTaskStatePlanParityUsesLatestPlan(t *testing.T) {
+	state := newTaskState("objective", nil)
+	state.observe(taskStateEvent{kind: taskStateEventPlan, arguments: `{"plan":[{"content":"old","status":"completed"}]}`})
+	state.observe(taskStateEvent{kind: taskStateEventPlan, arguments: `{"plan":[{"content":"new","status":"in_progress"}]}`})
+
+	messages := []zeroruntime.Message{{Role: zeroruntime.MessageRoleAssistant, ToolCalls: []zeroruntime.ToolCall{
+		{Name: planToolName, Arguments: `{"plan":[{"content":"new","status":"in_progress"}]}`},
+	}}}
+	if parity := state.observePlanParity(messages); parity != taskPlanParityMatch {
+		t.Fatalf("parity = %q, want match", parity)
+	}
+
+	messages[0].ToolCalls[0].Arguments = `{"plan":[{"content":"different","status":"pending"}]}`
+	if parity := state.observePlanParity(messages); parity != taskPlanParityMismatch {
+		t.Fatalf("parity = %q, want mismatch", parity)
+	}
+}
+
+func TestTaskStateMatchesPlanToolNormalization(t *testing.T) {
+	state := newTaskState("objective", nil)
+	arguments := `{"plan":[{"content":"first","status":"in_progress"},{"content":"second","status":"in_progress"}]}`
+	state.observe(taskStateEvent{kind: taskStateEventPlan, arguments: arguments})
+
+	snapshot := state.snapshot()
+	if snapshot.Plan.Completed != 1 || snapshot.Plan.InProgress != 1 {
+		t.Fatalf("multiple active items were not normalized like the plan tool: %#v", snapshot.Plan)
+	}
+	messages := []zeroruntime.Message{{Role: zeroruntime.MessageRoleAssistant, ToolCalls: []zeroruntime.ToolCall{{Name: planToolName, Arguments: arguments}}}}
+	if parity := state.observePlanParity(messages); parity != taskPlanParityMatch {
+		t.Fatalf("normalized state should still match its transcript event, got %q", parity)
+	}
+
+	empty := newTaskState("objective", nil)
+	empty.observe(taskStateEvent{kind: taskStateEventPlan, arguments: `{"plan":[]}`})
+	emptyMessages := []zeroruntime.Message{{Role: zeroruntime.MessageRoleAssistant, ToolCalls: []zeroruntime.ToolCall{{Name: planToolName, Arguments: `{"plan":[]}`}}}}
+	if parity := empty.observePlanParity(emptyMessages); parity != taskPlanParityMatch {
+		t.Fatalf("explicit empty plan should match transcript, got %q", parity)
+	}
+}
+
+func TestTaskStateContextFallsBackWhenTranscriptDiffers(t *testing.T) {
+	state := newTaskState("objective", nil)
+	state.observe(taskStateEvent{kind: taskStateEventPlan, arguments: `{"plan":[{"content":"tracked","status":"completed"}]}`})
+	messages := []zeroruntime.Message{{Role: zeroruntime.MessageRoleAssistant, ToolCalls: []zeroruntime.ToolCall{
+		{Name: planToolName, Arguments: `{"plan":[{"content":"transcript","status":"pending"}]}`},
+	}}}
+
+	context := state.completionContext(messages, true)
+	if !context.PlanPending || context.PlanMatchesTranscript {
+		t.Fatalf("completion context must retain transcript truth on mismatch: %#v", context)
+	}
+	if context.Objective != "objective" {
+		t.Fatalf("objective lost from completion context: %#v", context)
+	}
+}
+
+func TestTaskStateCompactionSnapshotRetainsObjectiveOnPlanMismatch(t *testing.T) {
+	state := newTaskState("objective", nil)
+	state.observe(taskStateEvent{kind: taskStateEventPlan, arguments: `{"plan":[{"content":"verify","status":"pending"}]}`})
+	matching := []zeroruntime.Message{{Role: zeroruntime.MessageRoleAssistant, ToolCalls: []zeroruntime.ToolCall{
+		{Name: planToolName, Arguments: `{"plan":[{"content":"verify","status":"pending"}]}`},
+	}}}
+	if snapshot := state.snapshotForCompaction(matching); snapshot == nil || snapshot.Objective != "objective" {
+		t.Fatalf("matching transcript should produce compact state, got %#v", snapshot)
+	}
+
+	mismatching := append([]zeroruntime.Message(nil), matching...)
+	mismatching[0].ToolCalls = []zeroruntime.ToolCall{{Name: planToolName, Arguments: `{"plan":[{"content":"other","status":"pending"}]}`}}
+	if snapshot := state.snapshotForCompaction(mismatching); snapshot == nil || snapshot.Objective != "objective" || snapshot.PlanParity != taskPlanParityMismatch {
+		t.Fatalf("plan mismatch must retain immutable objective and mark mutable state uncorroborated, got %#v", snapshot)
+	}
+}
+
+func TestParseTaskPlanRejectsArgumentsTheToolWouldReject(t *testing.T) {
+	for _, arguments := range []string{
+		`{}`,
+		`{"plan":null}`,
+		`{"plan":[{"step":"alias is not accepted"}]}`,
+		`{"plan":[{"content":""}]}`,
+		`{"plan":[{"content":"valid"},{"content":""}]}`,
+		`{"plan":[{"id":4,"content":"valid"}]}`,
+	} {
+		if plan, ok := parseTaskPlan(arguments); ok {
+			t.Fatalf("parseTaskPlan(%s) = %#v, true; want rejected", arguments, plan)
+		}
+	}
+}

internal/agent/turn_session.go

@@ -0,0 +1,21 @@
+package agent
+
+import (
+	"context"
+
+	"github.com/Gitlawb/zero/internal/zeroruntime"
+)
+
+// sessionProvider adapts a TurnSession back to the Provider interface so the
+// loop's existing Provider-shaped call sites (streamWithReconnect, the
+// compaction summarizer, finalAnswerAfterMaxTurns) route their stream I/O
+// through the session without any signature change. For the default adapter
+// this reduces to the wrapped provider's StreamCompletion, so behavior is
+// byte-identical; an optimized session (PR8) takes effect here transparently.
+type sessionProvider struct {
+	session zeroruntime.TurnSession
+}
+
+func (s sessionProvider) StreamCompletion(ctx context.Context, request zeroruntime.CompletionRequest) (<-chan zeroruntime.StreamEvent, error) {
+	return s.session.Stream(ctx, request)
+}

internal/agent/turn_session_test.go

@@ -0,0 +1,367 @@
+package agent
+
+import (
+	"context"
+	"errors"
+	"strings"
+	"testing"
+
+	"github.com/Gitlawb/zero/internal/tools"
+	"github.com/Gitlawb/zero/internal/zeroruntime"
+)
+
+// fakeTurnSession wraps an inner provider for streaming while counting the
+// session lifecycle, so tests can observe exactly how the loop drives it.
+type fakeTurnSession struct {
+	inner    Provider
+	prewarms int
+	streams  int
+	closes   int
+	closeErr error
+}
+
+func (s *fakeTurnSession) Prewarm(context.Context) error { s.prewarms++; return nil }
+
+func (s *fakeTurnSession) Stream(ctx context.Context, request zeroruntime.CompletionRequest) (<-chan zeroruntime.StreamEvent, error) {
+	s.streams++
+	return s.inner.StreamCompletion(ctx, request)
+}
+
+func (s *fakeTurnSession) Compact(context.Context, zeroruntime.CompletionRequest) ([]zeroruntime.Message, error) {
+	return nil, zeroruntime.ErrCompactionUnsupported
+}
+
+func (s *fakeTurnSession) Close() error { s.closes++; return s.closeErr }
+
+// fakeTurnSessionProvider opens a scripted session (or fails to).
+type fakeTurnSessionProvider struct {
+	session *fakeTurnSession
+	openErr error
+	opens   int
+}
+
+func (p *fakeTurnSessionProvider) OpenTurnSession(context.Context) (zeroruntime.TurnSession, error) {
+	p.opens++
+	if p.openErr != nil {
+		return nil, p.openErr
+	}
+	return p.session, nil
+}
+
+func (p *fakeTurnSessionProvider) Capabilities() zeroruntime.ProviderCapabilities {
+	return zeroruntime.ProviderCapabilities{}
+}
+
+func singleAnswerTurns(answer string) [][]zeroruntime.StreamEvent {
+	return [][]zeroruntime.StreamEvent{{
+		{Type: zeroruntime.StreamEventText, Content: answer},
+		{Type: zeroruntime.StreamEventDone},
+	}}
+}
+
+// TestRunOpenTurnSessionFailureSurfaces verifies a failed session open is a
+// clean run-start error: no panic, no stream ever issued, nothing to close.
+func TestRunOpenTurnSessionFailureSurfaces(t *testing.T) {
+	positional := &mockProvider{turns: singleAnswerTurns("never reached")}
+	tsp := &fakeTurnSessionProvider{openErr: errors.New("handshake refused")}
+
+	_, err := Run(context.Background(), "go", positional, Options{
+		MaxTurns:            2,
+		TurnSessionProvider: tsp,
+	})
+	if err == nil {
+		t.Fatal("expected an error from a failed session open")
+	}
+	if !strings.Contains(err.Error(), "open turn session") || !strings.Contains(err.Error(), "handshake refused") {
+		t.Fatalf("expected a wrapped open-turn-session error, got %v", err)
+	}
+	if len(positional.requests) != 0 {
+		t.Fatalf("expected no provider request after a failed open, got %d", len(positional.requests))
+	}
+	if tsp.opens != 1 {
+		t.Fatalf("expected exactly one open attempt, got %d", tsp.opens)
+	}
+}
+
+// TestRunCloseFailureIsSafe verifies a Close error at teardown is swallowed:
+// the run still returns its normal result.
+func TestRunCloseFailureIsSafe(t *testing.T) {
+	inner := &mockProvider{turns: singleAnswerTurns("done")}
+	session := &fakeTurnSession{inner: inner, closeErr: errors.New("close blew up")}
+	tsp := &fakeTurnSessionProvider{session: session}
+
+	result, err := Run(context.Background(), "go", inner, Options{
+		MaxTurns:            2,
+		TurnSessionProvider: tsp,
+	})
+	if err != nil {
+		t.Fatalf("expected Close failure to be swallowed, got %v", err)
+	}
+	if result.FinalAnswer != "done" {
+		t.Fatalf("expected normal result despite Close error, got %q", result.FinalAnswer)
+	}
+	if session.closes != 1 {
+		t.Fatalf("expected exactly one Close at teardown, got %d", session.closes)
+	}
+}
+
+// TestRunStreamsThroughTurnSession verifies that when a TurnSessionProvider is
+// wired, every model request of the run flows through the session (and the
+// session is prewarmed once and closed once).
+func TestRunStreamsThroughTurnSession(t *testing.T) {
+	inner := &mockProvider{turns: singleAnswerTurns("via session")}
+	session := &fakeTurnSession{inner: inner}
+	tsp := &fakeTurnSessionProvider{session: session}
+
+	// The positional provider is a DIFFERENT mock: it must stay untouched,
+	// proving the session (not the positional provider) carried the run.
+	positional := &mockProvider{turns: singleAnswerTurns("wrong path")}
+
+	result, err := Run(context.Background(), "go", positional, Options{
+		MaxTurns:            2,
+		TurnSessionProvider: tsp,
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if result.FinalAnswer != "via session" {
+		t.Fatalf("expected the session-backed answer, got %q", result.FinalAnswer)
+	}
+	if len(positional.requests) != 0 {
+		t.Fatalf("expected the positional provider to be bypassed, got %d requests", len(positional.requests))
+	}
+	if session.streams == 0 || session.streams != len(inner.requests) {
+		t.Fatalf("expected every request to flow through the session: streams=%d inner=%d", session.streams, len(inner.requests))
+	}
+	if session.prewarms != 1 {
+		t.Fatalf("expected exactly one prewarm, got %d", session.prewarms)
+	}
+	if session.closes != 1 {
+		t.Fatalf("expected exactly one close, got %d", session.closes)
+	}
+}
+
+// TestRunDefaultTurnSessionMatchesRawProvider verifies the default (nil
+// TurnSessionProvider) path is behavior-identical to passing an explicit
+// default adapter over the same provider: same result, same request count.
+func TestRunDefaultTurnSessionMatchesRawProvider(t *testing.T) {
+	run := func(explicit bool) (Result, *mockProvider) {
+		provider := &mockProvider{turns: escalateThenAnswerTurns("identical")}
+		registry := tools.NewRegistry()
+		registry.Register(escalatingTool{target: ""})
+		options := Options{Registry: registry, MaxTurns: 4}
+		if explicit {
+			options.TurnSessionProvider = zeroruntime.NewProviderTurnSessionProvider(provider, zeroruntime.ProviderCapabilities{})
+		}
+		result, err := Run(context.Background(), "go", provider, options)
+		if err != nil {
+			t.Fatal(err)
+		}
+		return result, provider
+	}
+
+	defaultResult, defaultProvider := run(false)
+	explicitResult, explicitProvider := run(true)
+
+	if defaultResult.FinalAnswer != explicitResult.FinalAnswer {
+		t.Fatalf("final answers diverged: default=%q explicit=%q", defaultResult.FinalAnswer, explicitResult.FinalAnswer)
+	}
+	if defaultResult.Turns != explicitResult.Turns {
+		t.Fatalf("turn counts diverged: default=%d explicit=%d", defaultResult.Turns, explicitResult.Turns)
+	}
+	if len(defaultProvider.requests) != len(explicitProvider.requests) {
+		t.Fatalf("request counts diverged: default=%d explicit=%d", len(defaultProvider.requests), len(explicitProvider.requests))
+	}
+}
+
+// TestRunModelSwitchClosesSessionAndContinues verifies a mid-run model switch
+// closes the active session exactly once and streams the rest of the run on
+// the new provider.
+func TestRunModelSwitchClosesSessionAndContinues(t *testing.T) {
+	registry := tools.NewRegistry()
+	registry.Register(escalatingTool{target: "claude-opus-4.1"})
+
+	firstInner := &mockProvider{turns: escalateThenAnswerTurns("unused")}
+	firstSession := &fakeTurnSession{inner: firstInner}
+	tsp := &fakeTurnSessionProvider{session: firstSession}
+
+	secondProvider := &mockProvider{turns: singleAnswerTurns("answered after switch")}
+
+	result, err := Run(context.Background(), "go", firstInner, Options{
+		Registry:            registry,
+		Model:               "claude-sonnet-4.5",
+		MaxTurns:            4,
+		TurnSessionProvider: tsp,
+		ModelSwitcher: func(_ context.Context, _ string) (Provider, error) {
+			return secondProvider, nil
+		},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if result.FinalAnswer != "answered after switch" {
+		t.Fatalf("expected the post-switch answer, got %q", result.FinalAnswer)
+	}
+	// The original session handled exactly the escalation turn, then was closed
+	// once by the swap (the teardown defer closes the NEW session, not this one).
+	if firstSession.streams != 1 {
+		t.Fatalf("expected the original session to stream exactly the escalation turn, got %d", firstSession.streams)
+	}
+	if firstSession.closes != 1 {
+		t.Fatalf("expected the swap to close the original session exactly once, got %d", firstSession.closes)
+	}
+	if len(secondProvider.requests) != 1 {
+		t.Fatalf("expected the post-switch turn on the new provider, got %d requests", len(secondProvider.requests))
+	}
+}
+
+// TestRunModelSessionSwitcherCarriesOptimizedSession verifies the target-aware
+// switcher: the switched-to TurnSessionProvider's DISTINCT session receives the
+// post-switch streams, is prewarmed exactly once, and is closed exactly once at
+// teardown — and the legacy ModelSwitcher is never consulted when both are set.
+func TestRunModelSessionSwitcherCarriesOptimizedSession(t *testing.T) {
+	registry := tools.NewRegistry()
+	registry.Register(escalatingTool{target: "claude-opus-4.1"})
+
+	firstInner := &mockProvider{turns: escalateThenAnswerTurns("unused")}
+	firstSession := &fakeTurnSession{inner: firstInner}
+	initial := &fakeTurnSessionProvider{session: firstSession}
+
+	switchedInner := &mockProvider{turns: singleAnswerTurns("answered on the switched session")}
+	switchedSession := &fakeTurnSession{inner: switchedInner}
+	switched := &fakeTurnSessionProvider{session: switchedSession}
+
+	var switchedTo string
+	legacyCalled := false
+	result, err := Run(context.Background(), "go", firstInner, Options{
+		Registry:            registry,
+		Model:               "claude-sonnet-4.5",
+		MaxTurns:            4,
+		TurnSessionProvider: initial,
+		ModelSessionSwitcher: func(_ context.Context, modelID string) (zeroruntime.TurnSessionProvider, error) {
+			switchedTo = modelID
+			return switched, nil
+		},
+		ModelSwitcher: func(_ context.Context, _ string) (Provider, error) {
+			legacyCalled = true
+			return nil, errors.New("legacy switcher must not be consulted")
+		},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if result.FinalAnswer != "answered on the switched session" {
+		t.Fatalf("expected the switched session's answer, got %q", result.FinalAnswer)
+	}
+	if switchedTo != "claude-opus-4.1" {
+		t.Fatalf("expected switch to claude-opus-4.1, got %q", switchedTo)
+	}
+	if legacyCalled {
+		t.Fatal("legacy ModelSwitcher was consulted despite ModelSessionSwitcher being set")
+	}
+	if switched.opens != 1 {
+		t.Fatalf("expected exactly one open of the switched provider, got %d", switched.opens)
+	}
+	// The switched session carries the rest of the run: it streams the answer
+	// turn, is prewarmed exactly once at the swap, and is closed exactly once by
+	// the teardown defer.
+	if switchedSession.streams != 1 || len(switchedInner.requests) != 1 {
+		t.Fatalf("expected the switched session to stream the post-switch turn, streams=%d inner=%d", switchedSession.streams, len(switchedInner.requests))
+	}
+	if switchedSession.prewarms != 1 {
+		t.Fatalf("expected the switched session to be prewarmed exactly once, got %d", switchedSession.prewarms)
+	}
+	if switchedSession.closes != 1 {
+		t.Fatalf("expected the switched session to be closed exactly once at teardown, got %d", switchedSession.closes)
+	}
+	// The original session ends with the swap: one stream, one close.
+	if firstSession.streams != 1 || firstSession.closes != 1 {
+		t.Fatalf("expected the original session to end at the swap (streams=%d closes=%d)", firstSession.streams, firstSession.closes)
+	}
+}
+
+// TestRunModelSessionSwitcherOpenFailureIsNonFatal verifies the swap-time open
+// failure: the switcher returns a TurnSessionProvider whose OpenTurnSession
+// fails. The run stays non-fatal on the ORIGINAL session (closed only at
+// teardown), and the switch-failure note reaches the next request.
+func TestRunModelSessionSwitcherOpenFailureIsNonFatal(t *testing.T) {
+	registry := tools.NewRegistry()
+	registry.Register(escalatingTool{target: "claude-opus-4.1"})
+
+	inner := &mockProvider{turns: escalateThenAnswerTurns("recovered after failed open")}
+	session := &fakeTurnSession{inner: inner}
+	failingOpen := &fakeTurnSessionProvider{openErr: errors.New("switched handshake refused")}
+
+	result, err := Run(context.Background(), "go", inner, Options{
+		Registry:            registry,
+		Model:               "claude-sonnet-4.5",
+		MaxTurns:            4,
+		TurnSessionProvider: &fakeTurnSessionProvider{session: session},
+		ModelSessionSwitcher: func(_ context.Context, _ string) (zeroruntime.TurnSessionProvider, error) {
+			return failingOpen, nil
+		},
+	})
+	if err != nil {
+		t.Fatalf("expected a failed swap open to be non-fatal, got %v", err)
+	}
+	if result.FinalAnswer != "recovered after failed open" {
+		t.Fatalf("expected the run to recover on the original session, got %q", result.FinalAnswer)
+	}
+	if failingOpen.opens != 1 {
+		t.Fatalf("expected exactly one open attempt on the switched provider, got %d", failingOpen.opens)
+	}
+	// The original session serves both turns and is closed only at teardown —
+	// never by the failed swap.
+	if session.streams != 2 || session.closes != 1 {
+		t.Fatalf("expected the original session to serve both turns and close once (streams=%d closes=%d)", session.streams, session.closes)
+	}
+	var sawNote bool
+	for _, message := range inner.requests[1].Messages {
+		if message.Role == zeroruntime.MessageRoleUser && strings.Contains(strings.ToLower(message.Content), "could not switch") {
+			sawNote = true
+		}
+	}
+	if !sawNote {
+		t.Fatalf("expected a switch-failure note on the next turn, messages: %+v", inner.requests[1].Messages)
+	}
+}
+
+// TestRunModelSessionSwitcherErrorIsNonFatal verifies a ModelSessionSwitcher
+// error keeps the run on the current session with a transcript note — the same
+// non-fatal contract as ModelSwitcher.
+func TestRunModelSessionSwitcherErrorIsNonFatal(t *testing.T) {
+	registry := tools.NewRegistry()
+	registry.Register(escalatingTool{target: "claude-opus-4.1"})
+
+	inner := &mockProvider{turns: escalateThenAnswerTurns("recovered on original session")}
+	session := &fakeTurnSession{inner: inner}
+
+	result, err := Run(context.Background(), "go", inner, Options{
+		Registry:            registry,
+		Model:               "claude-sonnet-4.5",
+		MaxTurns:            4,
+		TurnSessionProvider: &fakeTurnSessionProvider{session: session},
+		ModelSessionSwitcher: func(_ context.Context, _ string) (zeroruntime.TurnSessionProvider, error) {
+			return nil, errors.New("session build blew up")
+		},
+	})
+	if err != nil {
+		t.Fatalf("expected session-switcher error to be non-fatal, got %v", err)
+	}
+	if result.FinalAnswer != "recovered on original session" {
+		t.Fatalf("expected the run to recover on the original session, got %q", result.FinalAnswer)
+	}
+	// Both turns stream through the original session; it closes once at teardown.
+	if session.streams != 2 || session.closes != 1 {
+		t.Fatalf("expected the original session to serve both turns and close once (streams=%d closes=%d)", session.streams, session.closes)
+	}
+	var sawNote bool
+	for _, request := range inner.requests[1].Messages {
+		if request.Role == zeroruntime.MessageRoleUser && strings.Contains(strings.ToLower(request.Content), "could not switch") {
+			sawNote = true
+		}
+	}
+	if !sawNote {
+		t.Fatalf("expected a non-fatal switch-failure note on the next turn, messages: %+v", inner.requests[1].Messages)
+	}
+}

internal/agent/types.go

@@ -3,10 +3,12 @@ package agent
 import (
 	"context"
 
+	"github.com/Gitlawb/zero/internal/execution"
 	"github.com/Gitlawb/zero/internal/hooks"
 	"github.com/Gitlawb/zero/internal/sandbox"
 	"github.com/Gitlawb/zero/internal/streamjson"
 	"github.com/Gitlawb/zero/internal/tools"
+	"github.com/Gitlawb/zero/internal/trace"
 	"github.com/Gitlawb/zero/internal/zeroruntime"
 )
 
@@ -60,17 +62,29 @@ const (
 )
 
 type ToolResult struct {
-	ToolCallID   string
-	Name         string
-	Status       tools.Status
-	Output       string
+	ToolCallID string
+	Name       string
+	Status     tools.Status
+	Output     string
+	// Truncated reports that the tool's model-visible output omitted content.
+	// The full result may be recoverable through Meta["spill_path"].
+	Truncated    bool
 	Meta         map[string]string
 	Redacted     bool
 	ChangedFiles []string
-	Display      tools.Display
+	// ChangeSummaries are non-selectable generated-tree summaries emitted by
+	// command execution; callers must not schedule per-file work from them.
+	ChangeSummaries []execution.Change
+	Display         tools.Display
 	// DenialReason categorizes why a tool call was blocked (empty when it ran).
 	// It lets a surface distinguish the cause precisely instead of parsing Output.
 	DenialReason DenialCategory
+	// Risk is the sandbox risk classification of this call, stamped for EXECUTED
+	// results so run-policy observers (the execution-profile controller) can see
+	// the risk level of an allowed mutation. It mirrors the classification the
+	// permission path already computes; denied or canceled results keep the zero
+	// value. Pure observation: nothing about permissions or sandboxing changes.
+	Risk sandbox.Risk
 	// LoadedTools carries the deferred-tool names a tool_search call asked the
 	// loop to expose next turn (lifted from Meta["load_tools"]). nil for every
 	// ordinary tool result; only tool_search populates it.
@@ -94,6 +108,54 @@ const (
 	DenialHookBlocked      DenialCategory = "hook_blocked"      // vetoed by a beforeTool hook
 )
 
+// ProfilePolicy is the loop-facing slice of a selected execution profile.
+// The surface (exec flag, TUI command) resolves a named profile into this
+// policy; the loop only ever sees the policy, never the catalog.
+type ProfilePolicy struct {
+	// Name labels trace counters and diagnostics (e.g. "fast"). The trace
+	// recorder's profile label is set by the caller at recorder construction.
+	Name string
+	// Escalate, when non-nil, arms one-shot in-run posture escalation.
+	Escalate *PostureEscalation
+}
+
+// PostureEscalation describes a one-shot escalation to stricter knob values,
+// applied mid-run when an armed trigger fires. Targets are the values the
+// selected profile DISPLACED at run start (i.e. "restore the balanced
+// posture"), so escalation can never introduce a value that was not already
+// valid for this run and model. Zero-valued targets leave that knob untouched.
+type PostureEscalation struct {
+	// MaxTurns raises the turn ceiling to this value when greater than the
+	// ceiling in effect. 0 leaves the ceiling untouched.
+	MaxTurns int
+	// ReasoningEffort replaces the run's effort when non-empty.
+	ReasoningEffort string
+	// RestoreDefaultEffort clears the run's effort override (back to the
+	// provider/model default) when true. It exists because the displaced value
+	// of a profile-filled effort is "" — which as a ReasoningEffort target
+	// means "leave untouched" — so restoring the default needs its own signal.
+	// Ignored when ReasoningEffort is non-empty.
+	RestoreDefaultEffort bool
+	// RestoreCompletionGate re-enables RequireCompletionSignal (headless
+	// completion semantics) when true.
+	RestoreCompletionGate bool
+
+	// Triggers. A zero value disables that signal entirely.
+	// OnToolFailureStreak fires when the repeated-failure guard observes a
+	// same-tool retriable-failure streak of at least this length.
+	OnToolFailureStreak int
+	// OnCompletionUncertain fires on the Nth uncertain completion evaluation
+	// (continue nudge or semantic check). Headless only: the completion gate
+	// never runs interactively.
+	OnCompletionUncertain int
+	// OnSelfCorrectFailure fires when a post-edit verification cycle reports a
+	// failing outcome (correcting, reported, or aborted).
+	OnSelfCorrectFailure bool
+	// OnRiskyMutation fires when an EXECUTED tool result carries a sandbox risk
+	// level at or above this threshold. Empty disables the signal.
+	OnRiskyMutation sandbox.RiskLevel
+}
+
 type PermissionRequest struct {
 	ToolCallID         string                     `json:"toolCallId"`
 	ToolName           string                     `json:"name"`
@@ -279,6 +341,37 @@ type Options struct {
 	// request), so every existing caller is unaffected. A returned error is
 	// non-fatal: the run continues on the current model.
 	ModelSwitcher func(ctx context.Context, modelID string) (Provider, error)
+	// TurnSessionProvider, when set, supplies the turn session the run streams
+	// through — the seam an optimized provider session (connection reuse,
+	// prewarm, native compaction) plugs into without touching the loop. nil
+	// keeps the default: the loop wraps the passed provider in a no-op session
+	// whose Stream IS provider.StreamCompletion, so behavior is byte-identical
+	// and every existing caller is unaffected.
+	TurnSessionProvider zeroruntime.TurnSessionProvider
+	// ModelSessionSwitcher, when set, is the target-aware escalation hook: the
+	// loop prefers it over ModelSwitcher, and its TurnSessionProvider keeps an
+	// optimized session (and its capabilities) across a mid-run model switch.
+	// nil falls back to ModelSwitcher, whose bare Provider is wrapped in the
+	// default no-op session — today's behavior, unchanged. Same non-fatal error
+	// contract as ModelSwitcher: a returned error records a note and the run
+	// continues on the current model and session.
+	ModelSessionSwitcher func(ctx context.Context, modelID string) (zeroruntime.TurnSessionProvider, error)
+	// Profile, when set, arms the execution-profile posture controller for this
+	// run (auto-escalation to stricter knob values on failure/uncertainty/risky
+	// mutation signals). nil — the default everywhere today — leaves the loop
+	// byte-identical: no observation, no escalation, no counters. Same opt-in
+	// convention as Trace and SelfCorrect.
+	Profile *ProfilePolicy
+	// Trace, when set, records per-turn timing for the run: the loop stamps
+	// spans (prompt build, generation, tool execution, permission wait,
+	// compaction, provider connect) and counters (model requests, tool calls,
+	// retries, tokens) into it. nil DISABLES tracing entirely — every stamp is
+	// nil-safe and the loop is byte-identical to an untraced run. The caller
+	// owns the recorder: Run stamps into it but does not Finish or emit it.
+	// A fresh Recorder is required per Run — reusing one across runs merges
+	// their spans, counters, and first-event timestamps, and Finish freezes a
+	// recorder so no further stamps take.
+	Trace *trace.Recorder
 	// SelfCorrect, when set, runs a post-edit verify-and-correct cycle after a
 	// mutating tool call: it verifies the changed files (LSP diagnostics + project
 	// tests) and feeds failures back to the model to fix, bounded by an attempt
@@ -301,8 +394,11 @@ type Options struct {
 	// continuation cue ("…Let me check the config:"). The loop then nudges the
 	// model to continue instead, bounded by maxContinueNudges (and still by
 	// MaxTurns and the run deadline); if the model keeps stalling, the run
-	// finalizes as INCOMPLETE (Result.Incomplete) rather than success. Default
-	// false leaves the loop byte-identical, so the interactive TUI is unaffected.
+	// finalizes as INCOMPLETE (Result.Incomplete) rather than success. When the
+	// run's profile also enables SelfCorrect, an otherwise-complete turn gets one
+	// task-grounded semantic check before success; profiles without SelfCorrect
+	// add no model call. Default false leaves the loop byte-identical, so the
+	// interactive TUI is unaffected.
 	RequireCompletionSignal bool
 
 	runPermissions *permissionRunState
@@ -331,13 +427,6 @@ type Result struct {
 	IncompleteReason string
 }
 
-// Truncated reports whether the final response ended abnormally (cut off at the
-// output token cap or withheld by a content filter) rather than completing
-// naturally. Callers can use it to warn the user that FinalAnswer is incomplete.
-func (result Result) Truncated() bool {
-	return result.FinishReason != ""
-}
-
 // TruncationNotice returns a user-facing warning when the final response was
 // truncated, or "" for a normal completion. Shared by the CLI and TUI so the
 // wording stays consistent.

internal/agenteval/agent_command.go

@@ -31,10 +31,6 @@ type AgentRunner interface {
 
 type AgentRunnerFunc func(context.Context, AgentRunInput) AgentRunResult
 
-func (fn AgentRunnerFunc) Run(ctx context.Context, input AgentRunInput) AgentRunResult {
-	return fn(ctx, input)
-}
-
 // defaultAgentOutputLimit caps captured stdout/stderr per stream so a chatty or
 // runaway agent cannot exhaust memory or bloat the benchmark report.
 const defaultAgentOutputLimit = 1 << 20 // 1 MiB per stream

internal/agenteval/export_test.go

@@ -0,0 +1,8 @@
+// Test seams: helpers only test code uses, kept out of the production binary.
+package agenteval
+
+import "context"
+
+func (fn AgentRunnerFunc) Run(ctx context.Context, input AgentRunInput) AgentRunResult {
+	return fn(ctx, input)
+}

internal/agenteval/score.go

@@ -125,9 +125,7 @@ func Score(suite Suite, input ScoreInput) Report {
 	if len(task.RequiredTraceEvents) > 0 {
 		report.Results = append(report.Results, scoreTraceEvents(task.RequiredTraceEvents, input))
 	}
-	for _, result := range unknownCommandResults(input.CommandResults, seenCommands, input) {
-		report.Results = append(report.Results, result)
-	}
+	report.Results = append(report.Results, unknownCommandResults(input.CommandResults, seenCommands, input)...)
 	report.finishSummary()
 	return report
 }

internal/agentinit/agentinit.go

@@ -40,14 +40,14 @@ func FormatFacts(info repoinfo.Info) string {
 	b.WriteString("Pre-computed repository facts (from a local scan — verify and fill gaps):\n")
 
 	if info.PrimaryLanguage != "" {
-		b.WriteString(fmt.Sprintf("- Primary language: %s (of %d detected)\n", info.PrimaryLanguage, info.LanguageCount))
+		fmt.Fprintf(&b, "- Primary language: %s (of %d detected)\n", info.PrimaryLanguage, info.LanguageCount)
 	}
 	if langs := topLanguages(info.Languages, 5); langs != "" {
 		b.WriteString("- Languages: " + langs + "\n")
 	}
-	b.WriteString(fmt.Sprintf("- Size: ~%d files, ~%d LOC, max dir depth %d\n", info.FileCount, info.LOCEstimate, info.MaxDepth))
+	fmt.Fprintf(&b, "- Size: ~%d files, ~%d LOC, max dir depth %d\n", info.FileCount, info.LOCEstimate, info.MaxDepth)
 	if info.WorkspaceType != "" && info.WorkspaceType != "none" {
-		b.WriteString(fmt.Sprintf("- Workspace: %s (%d packages)\n", info.WorkspaceType, info.WorkspacePackageCount))
+		fmt.Fprintf(&b, "- Workspace: %s (%d packages)\n", info.WorkspaceType, info.WorkspacePackageCount)
 	}
 	if len(info.BuildTools) > 0 {
 		b.WriteString("- Build tools: " + strings.Join(info.BuildTools, ", ") + "\n")

internal/aimlapi/client.go

@@ -0,0 +1,259 @@
+package aimlapi
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+	"net/url"
+	"strings"
+	"time"
+)
+
+// PaymentMethod selects how a partner-checkout top-up is paid for.
+type PaymentMethod string
+
+const (
+	PaymentMethodCard   PaymentMethod = "card"
+	PaymentMethodCrypto PaymentMethod = "crypto"
+)
+
+// SessionStatus is the lifecycle state of a partner-checkout session, as reported
+// by the aimlapi.com backend while polling for payment.
+type SessionStatus string
+
+const (
+	SessionStatusPendingAuth    SessionStatus = "pending_auth"
+	SessionStatusPendingPayment SessionStatus = "pending_payment"
+	SessionStatusPaid           SessionStatus = "paid"
+	SessionStatusExchanging     SessionStatus = "exchanging"
+	SessionStatusExchanged      SessionStatus = "exchanged"
+	SessionStatusCancelled      SessionStatus = "cancelled"
+	SessionStatusExpired        SessionStatus = "expired"
+	SessionStatusFailed         SessionStatus = "failed"
+)
+
+// AuthResult is the bearer token minted by a passwordless sign-in / sign-up.
+type AuthResult struct {
+	Token string `json:"token"`
+	Exp   int64  `json:"exp"`
+}
+
+// PartnerCheckoutSession is a partner-attributed top-up session; its SessionToken
+// addresses it on later pay/exchange/poll calls.
+type PartnerCheckoutSession struct {
+	ID             string        `json:"id"`
+	SessionToken   string        `json:"sessionToken"`
+	PartnerID      string        `json:"partnerId"`
+	PartnerName    *string       `json:"partnerName"`
+	UserID         *int64        `json:"userId"`
+	AmountUSDMinor *int64        `json:"amountUsdMinor"`
+	Status         SessionStatus `json:"status"`
+	IssuedKeyID    *string       `json:"issuedKeyId"`
+	ReturnURL      *string       `json:"returnUrl"`
+}
+
+// PaymentSession is the hosted payment-provider checkout; PayURL is the page the
+// user opens to pay.
+type PaymentSession struct {
+	ProviderSessionID string `json:"providerSessionId"`
+	PayURL            string `json:"payUrl"`
+}
+
+// PayResult pairs the hosted checkout with the updated partner-checkout session
+// returned when a top-up payment is initiated.
+type PayResult struct {
+	Checkout        PaymentSession         `json:"checkout"`
+	PartnerCheckout PartnerCheckoutSession `json:"partnerCheckout"`
+}
+
+// ExchangeResult holds the API key minted when a paid session is exchanged.
+type ExchangeResult struct {
+	APIKey   string `json:"apiKey"`
+	APIKeyID string `json:"apiKeyId"`
+}
+
+// TopUpByKeyRequest funds a partner-checkout session with the API key that owns
+// the account (Path A / env key), instead of an email session.
+type TopUpByKeyRequest struct {
+	SessionToken     string
+	AmountUSDMinor   int
+	PaymentSessionID string // idempotency: same id → same checkout, never a second charge
+	AutoTopUp        bool   // enroll the account in automatic top-up at checkout (card, saved off-session)
+	SuccessURL       string
+	CancelURL        string
+}
+
+// TopUpByKeyResult is the hosted checkout returned by TopUpByKey.
+type TopUpByKeyResult struct {
+	Checkout PaymentSession `json:"checkout"`
+}
+
+// APIError is a non-2xx HTTP response from the aimlapi.com API. Status carries the
+// code so callers can branch (e.g. 401 = bad key, 5xx = retryable).
+type APIError struct {
+	Message string
+	Status  int
+	Body    string
+}
+
+// Error renders the method, endpoint, status, and (when present) the response body.
+func (e APIError) Error() string {
+	if strings.TrimSpace(e.Body) != "" {
+		return fmt.Sprintf("%s: HTTP %d: %s", e.Message, e.Status, strings.TrimSpace(e.Body))
+	}
+	return fmt.Sprintf("%s: HTTP %d", e.Message, e.Status)
+}
+
+// Client talks to the aimlapi.com auth, app, and inference APIs for the onboarding
+// and partner-checkout top-up flows.
+type Client struct {
+	endpoints  Endpoints
+	httpClient *http.Client
+}
+
+func NewClient(endpoints Endpoints, httpClient *http.Client) *Client {
+	if httpClient == nil {
+		httpClient = &http.Client{Timeout: 60 * time.Second}
+	}
+	return &Client{endpoints: endpoints, httpClient: httpClient}
+}
+
+// CreateSession opens a partner-checkout session attributed to partnerID; the
+// returned SessionToken addresses it on the pay/exchange/poll calls.
+func (c *Client) CreateSession(ctx context.Context, partnerID string, partnerName string, returnURL string) (PartnerCheckoutSession, error) {
+	body := map[string]any{"partnerId": partnerID}
+	if strings.TrimSpace(partnerName) != "" {
+		body["partnerName"] = strings.TrimSpace(partnerName)
+	}
+	if strings.TrimSpace(returnURL) != "" {
+		body["returnUrl"] = strings.TrimSpace(returnURL)
+	}
+	var result PartnerCheckoutSession
+	err := c.request(ctx, http.MethodPost, strings.TrimRight(c.endpoints.AppBaseURL, "/")+"/v3/partner-checkout/sessions", "", body, &result)
+	return result, err
+}
+
+// GetSession fetches the current state of a partner-checkout session; used to poll
+// for payment completion.
+func (c *Client) GetSession(ctx context.Context, sessionToken string) (PartnerCheckoutSession, error) {
+	var result PartnerCheckoutSession
+	err := c.request(ctx, http.MethodGet, strings.TrimRight(c.endpoints.AppBaseURL, "/")+"/v3/partner-checkout/sessions/"+urlPathEscape(sessionToken), "", nil, &result)
+	return result, err
+}
+
+// Pay initiates payment for a session and returns the hosted checkout URL.
+// autoTopUp enrolls the account in automatic top-up at checkout time.
+func (c *Client) Pay(ctx context.Context, bearer string, sessionToken string, amountUSDMinor int, paymentSessionID string, method PaymentMethod, successURL string, cancelURL string, autoTopUp bool) (PayResult, error) {
+	body := map[string]any{
+		"amountUsdMinor":   amountUSDMinor,
+		"paymentSessionId": paymentSessionID,
+		"method":           method,
+	}
+	if strings.TrimSpace(successURL) != "" {
+		body["successUrl"] = strings.TrimSpace(successURL)
+	}
+	if strings.TrimSpace(cancelURL) != "" {
+		body["cancelUrl"] = strings.TrimSpace(cancelURL)
+	}
+	// Only sent when enabled: enrolls the account in automatic top-up. The backend
+	// honours this field at checkout time.
+	if autoTopUp {
+		body["autoTopUp"] = true
+	}
+	var result PayResult
+	err := c.request(ctx, http.MethodPost, strings.TrimRight(c.endpoints.AppBaseURL, "/")+"/v3/partner-checkout/sessions/"+urlPathEscape(sessionToken)+"/pay", bearer, body, &result)
+	return result, err
+}
+
+// Exchange converts a paid session into a freshly minted API key (new-account flow).
+func (c *Client) Exchange(ctx context.Context, bearer string, sessionToken string) (ExchangeResult, error) {
+	var result ExchangeResult
+	err := c.request(ctx, http.MethodPost, strings.TrimRight(c.endpoints.AppBaseURL, "/")+"/v3/partner-checkout/sessions/"+urlPathEscape(sessionToken)+"/exchange", bearer, nil, &result)
+	return result, err
+}
+
+// TopUpByKey funds a partner-checkout session using the raw API key that owns the
+// account (Path A / env key), returning the hosted checkout. It binds the session
+// to the key's account server-side, so no email session is needed and no key is
+// exchanged. req.PaymentSessionID makes a retry idempotent (same id → same
+// checkout, never a second charge).
+func (c *Client) TopUpByKey(ctx context.Context, apiKey string, req TopUpByKeyRequest) (TopUpByKeyResult, error) {
+	body := map[string]any{
+		"sessionToken":     req.SessionToken,
+		"amountUsdMinor":   req.AmountUSDMinor,
+		"paymentSessionId": req.PaymentSessionID,
+	}
+	if req.AutoTopUp {
+		body["autoTopUp"] = true
+	}
+	if strings.TrimSpace(req.SuccessURL) != "" {
+		body["successUrl"] = strings.TrimSpace(req.SuccessURL)
+	}
+	if strings.TrimSpace(req.CancelURL) != "" {
+		body["cancelUrl"] = strings.TrimSpace(req.CancelURL)
+	}
+	var result TopUpByKeyResult
+	err := c.request(ctx, http.MethodPost, c.billingV2URL("/billing/topup"), apiKey, body, &result)
+	return result, err
+}
+
+// billingV2URL builds a v2 billing API URL from the inference base. The top-up
+// endpoint lives next to the balance endpoint on the API gateway, one version up
+// (".../v1" → ".../v2/billing/..."), so it re-anchors the version while keeping any
+// host/prefix an override supplied.
+func (c *Client) billingV2URL(path string) string {
+	base := strings.TrimRight(strings.TrimSpace(c.endpoints.InferenceBaseURL), "/")
+	base = strings.TrimSuffix(base, "/v1")
+	return strings.TrimRight(base, "/") + "/v2" + path
+}
+
+func (c *Client) request(ctx context.Context, method string, endpoint string, bearer string, body any, out any) error {
+	var reader io.Reader
+	if body != nil {
+		data, err := json.Marshal(body)
+		if err != nil {
+			return err
+		}
+		reader = bytes.NewReader(data)
+	}
+	request, err := http.NewRequestWithContext(ctx, method, endpoint, reader)
+	if err != nil {
+		return err
+	}
+	request.Header.Set("Accept", "application/json")
+	if body != nil {
+		request.Header.Set("Content-Type", "application/json")
+	}
+	if strings.TrimSpace(bearer) != "" {
+		request.Header.Set("Authorization", "Bearer "+strings.TrimSpace(bearer))
+	}
+	response, err := c.httpClient.Do(request)
+	if err != nil {
+		return fmt.Errorf("network request to %s failed: %w", endpoint, err)
+	}
+	defer response.Body.Close()
+	text, readErr := io.ReadAll(io.LimitReader(response.Body, 1<<20))
+	if readErr != nil {
+		return readErr
+	}
+	if response.StatusCode < 200 || response.StatusCode >= 300 {
+		return APIError{Message: method + " " + endpoint, Status: response.StatusCode, Body: string(text)}
+	}
+	if out == nil {
+		return nil
+	}
+	if len(bytes.TrimSpace(text)) == 0 {
+		return APIError{Message: method + " " + endpoint + " returned empty body", Status: response.StatusCode}
+	}
+	if err := json.Unmarshal(text, out); err != nil {
+		return APIError{Message: method + " " + endpoint + " returned non-JSON body", Status: response.StatusCode, Body: string(text)}
+	}
+	return nil
+}
+
+func urlPathEscape(value string) string {
+	return url.PathEscape(value)
+}

internal/aimlapi/client_onboard.go

@@ -0,0 +1,106 @@
+package aimlapi
+
+import (
+	"context"
+	"fmt"
+	"net/http"
+	"strings"
+)
+
+// Additional client methods backing the interactive TUI onboarding flow. They
+// target the endpoints shipped in the aimlapi backend:
+//   - GET  {inference}/billing/balance            (Path A: validate key + balance)
+//   - PATCH{auth}/v1/auth/account                 (Path B: does the account exist?)
+//   - POST {auth}/v1/auth/sign-in/code(/verify)   (Path B existing: passwordless login)
+//   - POST {auth}/v1/auth/account/passwordless    (Path B new: email-only signup)
+//   - POST {app}/v1/keys                          (mint a key once we hold a session)
+
+// BalanceResult is the wallet balance returned by GetBalance, with the backend's
+// low-balance flag and threshold.
+type BalanceResult struct {
+	Balance             float64 `json:"balance"`
+	LowBalance          bool    `json:"lowBalance"`
+	LowBalanceThreshold float64 `json:"lowBalanceThreshold"`
+}
+
+// CheckResult reports whether an email already has an account, driving the
+// sign-in vs sign-up branch of the onboarding flow.
+type CheckResult struct {
+	// "sign-in" (account exists) or "sign-up" (no account yet).
+	Action   string  `json:"action"`
+	Provider *string `json:"provider"`
+}
+
+// CreatedKey is an API key minted by CreateKey.
+type CreatedKey struct {
+	Key string `json:"key"`
+	ID  string `json:"id"`
+}
+
+func (c *Client) authURL(path string) string {
+	return strings.TrimRight(c.endpoints.AuthBaseURL, "/") + path
+}
+
+func (c *Client) appURL(path string) string {
+	return strings.TrimRight(c.endpoints.AppBaseURL, "/") + path
+}
+
+func (c *Client) inferenceURL(path string) string {
+	return strings.TrimRight(c.endpoints.InferenceBaseURL, "/") + path
+}
+
+// GetBalance doubles as key validation: a 401 means the key is invalid, a 2xx
+// returns the wallet balance. Auth is the raw API key itself.
+func (c *Client) GetBalance(ctx context.Context, apiKey string) (BalanceResult, error) {
+	var result BalanceResult
+	err := c.request(ctx, http.MethodGet, c.inferenceURL("/billing/balance"), apiKey, nil, &result)
+	return result, err
+}
+
+// CheckAccount reports whether an email already has an aimlapi.com account
+// (Action "sign-in") or not ("sign-up").
+func (c *Client) CheckAccount(ctx context.Context, email string) (CheckResult, error) {
+	var result CheckResult
+	err := c.request(ctx, http.MethodPatch, c.authURL("/v1/auth/account"), "", map[string]any{"email": email}, &result)
+	return result, err
+}
+
+// SendSignInCode emails a 6-digit sign-in code to an existing account.
+func (c *Client) SendSignInCode(ctx context.Context, email string) error {
+	return c.request(ctx, http.MethodPost, c.authURL("/v1/auth/sign-in/code"), "", map[string]any{"email": email}, nil)
+}
+
+// VerifySignInCode exchanges an emailed code for a session bearer token.
+func (c *Client) VerifySignInCode(ctx context.Context, email string, code string) (AuthResult, error) {
+	var result AuthResult
+	err := c.request(ctx, http.MethodPost, c.authURL("/v1/auth/sign-in/code/verify"), "", map[string]any{"email": email, "code": code}, &result)
+	if err == nil && strings.TrimSpace(result.Token) == "" {
+		err = fmt.Errorf("aimlapi.com did not return an auth token")
+	}
+	return result, err
+}
+
+// CreatePasswordlessAccount registers a new email-only account and returns its
+// session bearer token.
+func (c *Client) CreatePasswordlessAccount(ctx context.Context, email string) (AuthResult, error) {
+	var result AuthResult
+	err := c.request(ctx, http.MethodPost, c.authURL("/v1/auth/account/passwordless"), "", map[string]any{"email": email}, &result)
+	if err == nil && strings.TrimSpace(result.Token) == "" {
+		err = fmt.Errorf("aimlapi.com did not return an auth token")
+	}
+	return result, err
+}
+
+// CreateKey mints a fresh API key for a session-holding user.
+func (c *Client) CreateKey(ctx context.Context, bearer string, name string) (CreatedKey, error) {
+	body := map[string]any{}
+	if strings.TrimSpace(name) != "" {
+		body["name"] = strings.TrimSpace(name)
+	}
+	var result CreatedKey
+	err := c.request(ctx, http.MethodPost, c.appURL("/v1/keys"), bearer, body, &result)
+	if err == nil && strings.TrimSpace(result.Key) == "" {
+		err = fmt.Errorf("aimlapi.com did not return an API key")
+	}
+	return result, err
+}

internal/aimlapi/client_test.go

@@ -0,0 +1,159 @@
+package aimlapi
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+)
+
+func TestCreateKeySendsAuthenticatedRequestAndDecodesKey(t *testing.T) {
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.Method != http.MethodPost || r.URL.Path != "/v1/keys" {
+			t.Fatalf("request = %s %s, want POST /v1/keys", r.Method, r.URL.Path)
+		}
+		if got := r.Header.Get("Authorization"); got != "Bearer session-token" {
+			t.Fatalf("Authorization = %q", got)
+		}
+		var body map[string]string
+		if err := json.NewDecoder(r.Body).Decode(&body); err != nil {
+			t.Fatal(err)
+		}
+		if body["name"] != "Zero CLI" {
+			t.Fatalf("name = %q", body["name"])
+		}
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = w.Write([]byte(`{"key":"sk-issued","id":"key-id"}`))
+	}))
+	defer server.Close()
+
+	client := NewClient(Endpoints{AppBaseURL: server.URL}, server.Client())
+	key, err := client.CreateKey(context.Background(), " session-token ", " Zero CLI ")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if key.Key != "sk-issued" || key.ID != "key-id" {
+		t.Fatalf("key = %#v", key)
+	}
+}
+
+func TestCreateKeyRejectsEmptyKeyResponse(t *testing.T) {
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = w.Write([]byte(`{"id":"key-id"}`))
+	}))
+	defer server.Close()
+
+	client := NewClient(Endpoints{AppBaseURL: server.URL}, server.Client())
+	if _, err := client.CreateKey(context.Background(), "session-token", "Zero"); err == nil {
+		t.Fatal("CreateKey accepted a response without an API key")
+	}
+}
+
+func TestAuthMethodsRejectEmptyTokenResponses(t *testing.T) {
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = w.Write([]byte(`{"token":""}`))
+	}))
+	defer server.Close()
+
+	client := NewClient(Endpoints{AuthBaseURL: server.URL}, server.Client())
+	tests := []struct {
+		name string
+		call func() error
+	}{
+		{
+			name: "verify sign-in code",
+			call: func() error {
+				_, err := client.VerifySignInCode(context.Background(), "[email protected]", "123456")
+				return err
+			},
+		},
+		{
+			name: "create passwordless account",
+			call: func() error {
+				_, err := client.CreatePasswordlessAccount(context.Background(), "[email protected]")
+				return err
+			},
+		},
+	}
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			if err := test.call(); err == nil {
+				t.Fatal("accepted a response without an auth token")
+			}
+		})
+	}
+}
+
+func TestClientReturnsTypedAPIError(t *testing.T) {
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		http.Error(w, "invalid key", http.StatusUnauthorized)
+	}))
+	defer server.Close()
+
+	client := NewClient(Endpoints{InferenceBaseURL: server.URL}, server.Client())
+	_, err := client.GetBalance(context.Background(), "bad-key")
+	var apiErr APIError
+	if !errors.As(err, &apiErr) {
+		t.Fatalf("error = %T %v, want APIError", err, err)
+	}
+	if apiErr.Status != http.StatusUnauthorized || apiErr.Body != "invalid key\n" {
+		t.Fatalf("APIError = %#v", apiErr)
+	}
+}
+
+func TestTypedResponseRejectsEmptySuccessBody(t *testing.T) {
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		w.WriteHeader(http.StatusNoContent)
+	}))
+	defer server.Close()
+
+	client := NewClient(Endpoints{InferenceBaseURL: server.URL}, server.Client())
+	if _, err := client.GetBalance(context.Background(), "key"); err == nil {
+		t.Fatal("GetBalance accepted an empty successful response")
+	}
+}
+
+func TestBodylessOperationAcceptsEmptySuccessBody(t *testing.T) {
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		w.WriteHeader(http.StatusNoContent)
+	}))
+	defer server.Close()
+
+	client := NewClient(Endpoints{AuthBaseURL: server.URL}, server.Client())
+	if err := client.SendSignInCode(context.Background(), "[email protected]"); err != nil {
+		t.Fatalf("SendSignInCode() error = %v", err)
+	}
+}
+
+func TestParseAmountUSD(t *testing.T) {
+	tests := []struct {
+		name    string
+		value   string
+		want    int
+		wantErr bool
+	}{
+		{name: "default", want: DefaultAmountUSDMinor},
+		{name: "minimum", value: "20", want: 2000},
+		{name: "cents", value: "25.49", want: 2549},
+		{name: "below minimum", value: "19.99", wantErr: true},
+		{name: "above maximum", value: "10001", wantErr: true},
+		{name: "not a number", value: "twenty", wantErr: true},
+		{name: "nan", value: "NaN", wantErr: true},
+		{name: "infinity", value: "Inf", wantErr: true},
+	}
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			got, err := ParseAmountUSD(test.value)
+			if (err != nil) != test.wantErr {
+				t.Fatalf("ParseAmountUSD(%q) error = %v", test.value, err)
+			}
+			if got != test.want {
+				t.Fatalf("ParseAmountUSD(%q) = %d, want %d", test.value, got, test.want)
+			}
+		})
+	}
+}

internal/aimlapi/config.go

@@ -0,0 +1,113 @@
+package aimlapi
+
+import (
+	"net/url"
+	"os"
+	"strings"
+)
+
+type Endpoints struct {
+	AuthBaseURL      string
+	AppBaseURL       string
+	InferenceBaseURL string
+	PayBaseURL       string
+	// VerificationBaseURL is the aimlapi web-app frontend base. The browser is sent
+	// back here after the partner checkout completes (see BuildPartnerReturnURL), so
+	// it must point at the same environment's front as the one the CLI is talking to
+	// to land the user there. Empty falls back to DefaultReturnURL.
+	VerificationBaseURL string
+}
+
+const (
+	DefaultPartnerID   = "part_62yQoGYDq4Yqnrj2R1iGrDNJ"
+	DefaultPartnerName = "Gitlawb"
+	PartnerHeaderName  = "X-AIMLAPI-Partner-ID"
+	// DefaultReturnURL is the https fallback the browser is sent to after the
+	// checkout / OAuth consent finishes when no frontend base is known. It must be
+	// a real web page the browser can open — NOT a custom scheme like
+	// zero://aimlapi/complete, which fails with "the scheme does not have a
+	// registered handler" because the CLI registers no OS protocol handler. The
+	// CLI learns of success by polling; this URL is purely the browser's landing.
+	DefaultReturnURL = "https://aimlapi.com/app"
+	DefaultModel     = "anthropic/claude-sonnet-5"
+
+	MinAmountUSDMinor     = 2000
+	MaxAmountUSDMinor     = 1000000
+	DefaultAmountUSDMinor = 2500
+)
+
+func ResolveEndpoints() Endpoints {
+	return Endpoints{
+		AuthBaseURL:         envOrDefault("AIMLAPI_AUTH_URL", "https://auth.aimlapi.com"),
+		AppBaseURL:          envOrDefault("AIMLAPI_APP_URL", "https://app.aimlapi.com"),
+		InferenceBaseURL:    envOrDefault("AIMLAPI_INFERENCE_URL", "https://api.aimlapi.com/v1"),
+		PayBaseURL:          envOrDefault("AIMLAPI_PAY_URL", "https://pay.aimlapi.com"),
+		VerificationBaseURL: envOrDefault("AIMLAPI_VERIFICATION_BASE_URL", "https://aimlapi.com/app"),
+	}
+}
+
+// ResolvePartnerID applies the same explicit-option, environment, and default
+// precedence everywhere partner attribution is used. Keeping checkout creation
+// and saved inference profiles on one resolver prevents them from attributing
+// the same user to different partners.
+func ResolvePartnerID(explicit string) string {
+	if value := strings.TrimSpace(explicit); value != "" {
+		return value
+	}
+	return envOrDefault("AIMLAPI_PARTNER_ID", DefaultPartnerID)
+}
+
+// WithResolvedPartnerHeader returns a copy with the effective partner ID set.
+// Header matching is case-insensitive so an existing spelling is replaced rather
+// than duplicated. Callers remain responsible for applying catalog attribution
+// only to the canonical AIMLAPI endpoint.
+func WithResolvedPartnerHeader(headers map[string]string) map[string]string {
+	resolved := make(map[string]string, len(headers)+1)
+	for key, value := range headers {
+		if strings.EqualFold(strings.TrimSpace(key), PartnerHeaderName) {
+			continue
+		}
+		resolved[key] = value
+	}
+	resolved[PartnerHeaderName] = ResolvePartnerID("")
+	return resolved
+}
+
+func BuildPartnerCheckoutReturnURLs(appBaseURL string, sessionToken string) (successURL string, cancelURL string) {
+	base := strings.TrimRight(strings.TrimSpace(appBaseURL), "/")
+	if base == "" {
+		// With no base there is nothing to anchor an absolute URL to. Return empty
+		// so the caller omits the return URLs entirely (client.Pay skips empty ones,
+		// leaving the backend's generic checkout) instead of emitting a broken
+		// relative "/checkout?...".
+		return "", ""
+	}
+	token := url.QueryEscape(sessionToken)
+	return base + "/checkout?checkout=success&partnerCheckout=1&sessionToken=" + token,
+		base + "/checkout?checkout=cancel&partnerCheckout=1&sessionToken=" + token
+}
+
+// BuildPartnerReturnURL returns the https page the browser is sent back to once
+// the partner checkout or OAuth consent completes — the aimlapi web app the user
+// opened the flow from. frontendBaseURL is normally endpoints.VerificationBaseURL
+// (the front that hosts the login/consent page), so the return follows the same
+// environment. AIMLAPI_RETURN_URL overrides it outright; an empty base falls back
+// to DefaultReturnURL. It is deliberately NOT a custom scheme (zero://…): no
+// browser can hand a custom scheme off without an OS-registered handler, which a
+// CLI does not install.
+func BuildPartnerReturnURL(frontendBaseURL string) string {
+	if override := strings.TrimSpace(os.Getenv("AIMLAPI_RETURN_URL")); override != "" {
+		return override
+	}
+	if base := strings.TrimRight(strings.TrimSpace(frontendBaseURL), "/"); base != "" {
+		return base
+	}
+	return DefaultReturnURL
+}
+
+func envOrDefault(name string, fallback string) string {
+	if value := strings.TrimSpace(os.Getenv(name)); value != "" {
+		return value
+	}
+	return fallback
+}

internal/aimlapi/config_test.go

@@ -0,0 +1,55 @@
+package aimlapi
+
+import "testing"
+
+func TestResolveEndpointsDefaults(t *testing.T) {
+	t.Setenv("AIMLAPI_AUTH_URL", "")
+	t.Setenv("AIMLAPI_APP_URL", "")
+	t.Setenv("AIMLAPI_INFERENCE_URL", "")
+	t.Setenv("AIMLAPI_PAY_URL", "")
+
+	endpoints := ResolveEndpoints()
+	if endpoints.AuthBaseURL != "https://auth.aimlapi.com" ||
+		endpoints.AppBaseURL != "https://app.aimlapi.com" ||
+		endpoints.InferenceBaseURL != "https://api.aimlapi.com/v1" ||
+		endpoints.PayBaseURL != "https://pay.aimlapi.com" {
+		t.Fatalf("unexpected endpoints: %#v", endpoints)
+	}
+}
+
+func TestResolvedPartnerHeaderUsesEnvironmentOverrideWithoutDuplicates(t *testing.T) {
+	t.Setenv("AIMLAPI_PARTNER_ID", "part_override")
+
+	headers := WithResolvedPartnerHeader(map[string]string{
+		"x-aimlapi-partner-id":       "part_catalog",
+		"X-AIMLAPI-Integration-Repo": "Gitlawb/zero",
+	})
+
+	if got := headers[PartnerHeaderName]; got != "part_override" {
+		t.Fatalf("partner header = %q, want part_override", got)
+	}
+	if _, ok := headers["x-aimlapi-partner-id"]; ok {
+		t.Fatalf("case-insensitive partner header was duplicated: %#v", headers)
+	}
+	if got := headers["X-AIMLAPI-Integration-Repo"]; got != "Gitlawb/zero" {
+		t.Fatalf("unrelated header changed: %#v", headers)
+	}
+}
+
+func TestBuildPartnerReturnURL(t *testing.T) {
+	t.Setenv("AIMLAPI_RETURN_URL", "")
+	// The frontend base (the web app the flow was opened from) is returned as-is,
+	// trailing slash trimmed — an https page the browser can actually land on.
+	if got := BuildPartnerReturnURL("https://aimlapi.com/app/"); got != "https://aimlapi.com/app" {
+		t.Fatalf("return URL = %q, want the frontend base", got)
+	}
+	// An empty base falls back to the https default, never a custom scheme.
+	if got := BuildPartnerReturnURL(""); got != DefaultReturnURL {
+		t.Fatalf("empty base return URL = %q, want %q", got, DefaultReturnURL)
+	}
+
+	t.Setenv("AIMLAPI_RETURN_URL", "https://example.test/done")
+	if got := BuildPartnerReturnURL("https://aimlapi.com/app"); got != "https://example.test/done" {
+		t.Fatalf("env override return URL = %q", got)
+	}
+}

internal/aimlapi/messages.go

@@ -0,0 +1,42 @@
+package aimlapi
+
+// UI strings shared by the first-run and provider-wizard onboarding surfaces.
+
+const (
+	// Path A — existing API key. Pick-path prompt + option labels.
+	MsgAPIKeyInputPrompt = "Enter your aimlapi.com key."
+	MsgAPIKeyInvalid     = "API key is invalid. Please make sure you enter a valid aimlapi.com key."
+	MsgPickPathPrompt    = "Do you have aimlapi.com key?"
+	MsgPickPathHaveKey   = "I already have aimlapi.com key"
+	MsgPickPathHaveHint  = "Proceed to paste the key"
+	MsgPickPathNewUser   = "I am a new user"
+	MsgPickPathNewHint   = "One click set up"
+	// Path B — email.
+	MsgEnterEmail           = "Enter your email.\nTo access aimlapi.com dashboard."
+	MsgEmailInvalid         = "Email format is incorrect."
+	MsgAccountActionInvalid = "aimlapi.com returned an unsupported account action. Please try again."
+	// %s = email
+	MsgCodeSent      = "We sent a 6-digit code to %s.\nEnter it below to continue."
+	MsgCodeIncorrect = "Code you've entered is incorrect."
+
+	// Balance check.
+	MsgLowBalance     = "Your aimlapi.com balance is running low.\nIt is recommended to top up your balance."
+	MsgEverythingRuns = "Everything is ready."
+	// Low-balance choices (spec wording).
+	MsgLowBalanceTopUp = "Sure, let's do that"
+	MsgLowBalanceSkip  = "I'll skip topping up the balance for now"
+
+	// Top-up workflow.
+	MsgTopUpPrompt    = "Add credits.\nEnter an amount (min $20)."
+	MsgAmountRequired = "Please enter a top-up amount."
+	// %s = checkout URL
+	MsgTopUpBrowserFallback = "If the browser did not open automatically please use this link to top up your account: %s"
+	MsgTopUpFailed          = "Top up failed. Please try again."
+	MsgTopUpSuccess         = "Top-up successful."
+
+	// Success — key delivered.
+	// %s = amount in dollars (e.g. "25")
+	MsgTopUpAddedFmt = "$%s has been added to your balance."
+	// %s = email
+	MsgSuccessMagicLink = "We've emailed you a magic link to %s. Use it to access your aimlapi.com account and review your usage and balance."
+)

internal/aimlapi/onboard_test.go

@@ -0,0 +1,26 @@
+package aimlapi
+
+import "testing"
+
+func TestValidEmailRejectsIncompleteDomains(t *testing.T) {
+	tests := []struct {
+		value string
+		valid bool
+	}{
+		{value: "[email protected]", valid: true},
+		{value: "[email protected]", valid: true},
+		{value: "123@123.", valid: false},
+		{value: "123@123", valid: false},
+		{value: "[email protected]", valid: false},
+		{value: "[email protected]", valid: false},
+		{value: "[email protected]", valid: false},
+		{value: "[email protected]", valid: false},
+	}
+	for _, test := range tests {
+		t.Run(test.value, func(t *testing.T) {
+			if got := ValidEmail(test.value); got != test.valid {
+				t.Fatalf("ValidEmail(%q) = %v, want %v", test.value, got, test.valid)
+			}
+		})
+	}
+}

internal/aimlapi/stream.go

@@ -0,0 +1,335 @@
+package aimlapi
+
+import (
+	"context"
+	"fmt"
+	"net/http"
+	"net/url"
+	"strings"
+)
+
+// StreamTopUpOptions drives a session-based partner-checkout top-up for the
+// interactive (TUI) onboarding. The caller already holds a user session —
+// obtained from an email-code sign-in (Path B existing) or a
+// passwordless new account (Path B sign-up) — so this skips authentication and
+// runs create-session → pay → wait-for-payment, optionally exchanging the paid
+// session into a fresh key (new accounts) versus keeping the caller's key.
+type StreamTopUpOptions struct {
+	SessionToken string // user session bearer (required)
+	AmountUSD    string // dollars; parsed via ParseAmountUSD (min $20)
+	// InferenceBaseURL pins key-bound billing and the returned provider profile to
+	// the endpoint that validated the credential. Empty uses ResolveEndpoints.
+	InferenceBaseURL string
+	Method           PaymentMethod
+	AutoTopUp        bool // enroll the account in automatic top-up at checkout time
+	Exchange         bool // exchange the paid session into a new key (new accounts)
+	Model            string
+	PartnerID        string
+	PartnerName      string
+	NoOpen           bool
+
+	// ResumeSessionToken, when set, resumes that partner-checkout session instead
+	// of opening a new one, so a retry after an ambiguous failure (a lost Pay/
+	// Exchange response) can never double-charge or strand an already-paid session.
+	ResumeSessionToken string
+	// PaymentSessionID is generated once per amount/auto-top-up intent and reused
+	// on retry so an ambiguous Pay response cannot create a second checkout.
+	PaymentSessionID string
+
+	HTTPClient  *http.Client
+	OpenBrowser func(string) error
+	OnStatus    func(Status, string)
+	// OnSession reports the partner-checkout session token the moment it exists (a
+	// fresh CreateSession, or the resumed one), so the caller can retain it and
+	// resume on a later failure. It fires with "" when the session is dead and the
+	// retained token should be dropped (a fresh run must start over).
+	OnSession func(sessionToken string)
+}
+
+// topupPhase is where StreamTopUp enters its create-checkout → poll → exchange
+// pipeline. A fresh run starts at phasePay; a resume skips ahead to whatever the
+// existing session's status makes safe.
+type topupPhase int
+
+const (
+	phasePay          topupPhase = iota // (re-)open the Stripe checkout, then poll + exchange
+	phasePoll                           // checkout already open; poll for payment, never re-pay
+	phaseExchange                       // already paid; exchange only (recovers a lost key)
+	phaseWaitExchange                   // an exchange claim is in flight; poll, never call Exchange twice
+)
+
+// StreamTopUp performs the browser-checkout top-up and reports progress via
+// OnStatus.
+func StreamTopUp(ctx context.Context, options StreamTopUpOptions) (ProvisionedKey, error) {
+	endpoints := ResolveEndpoints()
+	if baseURL := strings.TrimSpace(options.InferenceBaseURL); baseURL != "" {
+		endpoints.InferenceBaseURL = baseURL
+	}
+	client := NewClient(endpoints, options.HTTPClient)
+	partnerID := ResolvePartnerID(options.PartnerID)
+	partnerName := firstNonEmpty(options.PartnerName, DefaultPartnerName)
+	method := options.Method
+	if method != PaymentMethodCrypto {
+		method = PaymentMethodCard
+	}
+	amount, err := ParseAmountUSD(options.AmountUSD)
+	if err != nil {
+		return ProvisionedKey{}, err
+	}
+	if strings.TrimSpace(options.SessionToken) == "" {
+		return ProvisionedKey{}, fmt.Errorf("a session is required to top up")
+	}
+	if strings.TrimSpace(options.PaymentSessionID) == "" {
+		return ProvisionedKey{}, fmt.Errorf("a payment session id is required to top up")
+	}
+
+	status(options.OnStatus, StatusCreatingSession, "")
+	sessionToken, phase, err := resolveTopupSession(ctx, client, options, partnerID, partnerName, endpoints)
+	if err != nil {
+		return ProvisionedKey{}, err
+	}
+	// Retain the live token so an ambiguous failure below can resume this session
+	// rather than open a second checkout on the next attempt.
+	reportSession(options.OnSession, sessionToken)
+
+	if phase <= phasePay {
+		successURL, cancelURL := BuildPartnerCheckoutReturnURLs(endpoints.PayBaseURL, sessionToken)
+		status(options.OnStatus, StatusOpeningCheckout, "")
+		pay, err := client.Pay(ctx, options.SessionToken, sessionToken, amount, options.PaymentSessionID, method, successURL, cancelURL, options.AutoTopUp)
+		if err != nil {
+			return ProvisionedKey{}, err
+		}
+		checkoutURL, err := validatedCheckoutURL(pay.Checkout.PayURL)
+		if err != nil {
+			return ProvisionedKey{}, err
+		}
+		announceCheckout(options.OnStatus, options.OpenBrowser, options.NoOpen, checkoutURL)
+	}
+
+	paidToken := sessionToken
+	if phase <= phasePoll {
+		status(options.OnStatus, StatusWaitingPayment, "")
+		paid, err := pollUntilPaid(ctx, client, sessionToken, options.OnSession)
+		if err != nil {
+			return ProvisionedKey{}, err
+		}
+		paidToken = paid.SessionToken
+		if paid.Status == SessionStatusExchanging {
+			// Another/lost request claimed exchange while we were polling payment.
+			// Carry that fact into the provisioning branch so it follows the claim
+			// instead of issuing a second single-flight Exchange request.
+			phase = phaseWaitExchange
+		}
+	}
+
+	result := ProvisionedKey{
+		BaseURL: endpoints.InferenceBaseURL,
+		Model:   firstNonEmpty(options.Model, DefaultModel),
+	}
+	if options.Exchange {
+		status(options.OnStatus, StatusProvisioningKey, "")
+		if phase == phaseWaitExchange {
+			return ProvisionedKey{}, pollUntilExchangeSettled(ctx, client, sessionToken, options.OnSession)
+		}
+		exchange, err := client.Exchange(ctx, options.SessionToken, paidToken)
+		if err != nil {
+			return ProvisionedKey{}, err
+		}
+		if strings.TrimSpace(exchange.APIKey) == "" {
+			return ProvisionedKey{}, fmt.Errorf("aimlapi.com did not return an API key")
+		}
+		result.APIKey = exchange.APIKey
+		result.APIKeyID = exchange.APIKeyID
+	}
+	return result, nil
+}
+
+// resolveTopupSession returns the partner-checkout session token to drive and the
+// phase to enter. A fresh run opens a new session at phasePay. A resume inspects
+// the existing session and enters the earliest phase that cannot re-charge:
+//   - pending_auth    → phasePay: payment never started, so (re-)Pay is safe.
+//   - pending_payment → phasePoll: a checkout is already open; poll, never re-Pay.
+//   - paid            → phaseExchange: exchange the paid session once.
+//   - exchanging      → phaseWaitExchange: another/lost request owns the atomic
+//     exchange claim; poll it, never issue a second Exchange request.
+//   - exchanged       → terminal: the one-shot raw key is gone; point at dashboard
+//     recovery and keep the token so a re-run repeats that guidance, not a charge.
+//   - cancelled/expired/failed → terminal: drop the dead token so a re-run is fresh.
+func resolveTopupSession(ctx context.Context, client *Client, options StreamTopUpOptions, partnerID, partnerName string, endpoints Endpoints) (string, topupPhase, error) {
+	resume := strings.TrimSpace(options.ResumeSessionToken)
+	if resume == "" {
+		session, err := client.CreateSession(ctx, partnerID, partnerName, BuildPartnerReturnURL(endpoints.VerificationBaseURL))
+		if err != nil {
+			return "", phasePay, err
+		}
+		return session.SessionToken, phasePay, nil
+	}
+	session, err := client.GetSession(ctx, resume)
+	if err != nil {
+		if isTerminalSessionAPIError(err) {
+			reportSession(options.OnSession, "")
+		}
+		return "", phasePay, err
+	}
+	switch session.Status {
+	case SessionStatusPendingAuth:
+		return resume, phasePay, nil
+	case SessionStatusPendingPayment:
+		return resume, phasePoll, nil
+	case SessionStatusPaid:
+		return resume, phaseExchange, nil
+	case SessionStatusExchanging:
+		return resume, phaseWaitExchange, nil
+	case SessionStatusExchanged:
+		return "", phasePay, fmt.Errorf("session was already exchanged; rotate the key from the aimlapi.com dashboard")
+	default:
+		reportSession(options.OnSession, "") // drop the dead token → next run starts fresh
+		return "", phasePay, fmt.Errorf("payment %s; re-run the top-up to try again", session.Status)
+	}
+}
+
+func reportSession(callback func(string), sessionToken string) {
+	if callback != nil {
+		callback(sessionToken)
+	}
+}
+
+// announceCheckout opens the hosted checkout in the browser (unless suppressed)
+// and reports the URL either way, so the TUI can always show the direct link.
+func announceCheckout(onStatus func(Status, string), open func(string) error, noOpen bool, checkoutURL string) {
+	if noOpen || open == nil {
+		status(onStatus, StatusOpeningCheckout, checkoutURL)
+		return
+	}
+	if err := open(checkoutURL); err != nil {
+		status(onStatus, StatusOpeningCheckout, "Open manually: "+checkoutURL)
+		return
+	}
+	status(onStatus, StatusOpeningCheckout, checkoutURL)
+}
+
+func validatedCheckoutURL(value string) (string, error) {
+	checkoutURL := strings.TrimSpace(value)
+	parsed, err := url.Parse(checkoutURL)
+	if err != nil || !parsed.IsAbs() || !strings.EqualFold(parsed.Scheme, "https") || parsed.Hostname() == "" || parsed.User != nil {
+		return "", fmt.Errorf("payment provider did not return a valid HTTPS checkout URL")
+	}
+	return checkoutURL, nil
+}
+
+// StreamTopUpByKeyOptions drives a top-up funded with the API key that already
+// owns the account (Path A pasted key, or AIMLAPI_API_KEY). Unlike StreamTopUp it
+// needs no email session and never exchanges a key — the caller already holds one,
+// so it only funds the wallet: create-session → pay-by-key → wait-for-payment.
+type StreamTopUpByKeyOptions struct {
+	APIKey           string // raw key that owns the account (required)
+	AmountUSD        string // dollars; parsed via ParseAmountUSD (min $20)
+	InferenceBaseURL string // validated key endpoint; empty uses ResolveEndpoints
+	AutoTopUp        bool   // enroll the account in automatic top-up at checkout time
+	PartnerID        string
+	PartnerName      string
+	NoOpen           bool
+
+	// ResumeSessionToken resumes an existing partner-checkout session instead of
+	// opening a new one. PaymentSessionID is the idempotency handle carried across
+	// retries so a re-issued pay returns the same checkout, never a second charge.
+	ResumeSessionToken string
+	PaymentSessionID   string // required; generate once via NewPaymentSessionID, reuse on retry
+
+	HTTPClient  *http.Client
+	OpenBrowser func(string) error
+	OnStatus    func(Status, string)
+	OnSession   func(sessionToken string)
+}
+
+// StreamTopUpByKey funds the account behind APIKey via the hosted checkout and
+// reports progress through OnStatus. It returns an empty ProvisionedKey: the
+// pasted/env key is unchanged, only the balance grows.
+func StreamTopUpByKey(ctx context.Context, options StreamTopUpByKeyOptions) (ProvisionedKey, error) {
+	endpoints := ResolveEndpoints()
+	if baseURL := strings.TrimSpace(options.InferenceBaseURL); baseURL != "" {
+		endpoints.InferenceBaseURL = baseURL
+	}
+	client := NewClient(endpoints, options.HTTPClient)
+	partnerID := ResolvePartnerID(options.PartnerID)
+	partnerName := firstNonEmpty(options.PartnerName, DefaultPartnerName)
+	amount, err := ParseAmountUSD(options.AmountUSD)
+	if err != nil {
+		return ProvisionedKey{}, err
+	}
+	if strings.TrimSpace(options.APIKey) == "" {
+		return ProvisionedKey{}, fmt.Errorf("an API key is required to top up")
+	}
+	if strings.TrimSpace(options.PaymentSessionID) == "" {
+		return ProvisionedKey{}, fmt.Errorf("a payment session id is required to top up")
+	}
+
+	status(options.OnStatus, StatusCreatingSession, "")
+	sessionToken, phase, err := resolveByKeySession(ctx, client, options, partnerID, partnerName, endpoints)
+	if err != nil {
+		return ProvisionedKey{}, err
+	}
+	reportSession(options.OnSession, sessionToken)
+
+	if phase <= phasePay {
+		successURL, cancelURL := BuildPartnerCheckoutReturnURLs(endpoints.PayBaseURL, sessionToken)
+		status(options.OnStatus, StatusOpeningCheckout, "")
+		pay, err := client.TopUpByKey(ctx, options.APIKey, TopUpByKeyRequest{
+			SessionToken:     sessionToken,
+			AmountUSDMinor:   amount,
+			PaymentSessionID: options.PaymentSessionID,
+			AutoTopUp:        options.AutoTopUp,
+			SuccessURL:       successURL,
+			CancelURL:        cancelURL,
+		})
+		if err != nil {
+			return ProvisionedKey{}, err
+		}
+		checkoutURL, err := validatedCheckoutURL(pay.Checkout.PayURL)
+		if err != nil {
+			return ProvisionedKey{}, err
+		}
+		announceCheckout(options.OnStatus, options.OpenBrowser, options.NoOpen, checkoutURL)
+	}
+
+	if phase <= phasePoll {
+		status(options.OnStatus, StatusWaitingPayment, "")
+		if _, err := pollUntilPaid(ctx, client, sessionToken, options.OnSession); err != nil {
+			return ProvisionedKey{}, err
+		}
+	}
+	// No exchange: the account is funded and the caller keeps its existing key.
+	return ProvisionedKey{}, nil
+}
+
+// resolveByKeySession mirrors resolveTopupSession for the by-key flow. Because
+// by-key never exchanges, a paid/exchanging/exchanged session all mean "already
+// funded, done" (phaseExchange skips straight past pay + poll to the empty result).
+func resolveByKeySession(ctx context.Context, client *Client, options StreamTopUpByKeyOptions, partnerID, partnerName string, endpoints Endpoints) (string, topupPhase, error) {
+	resume := strings.TrimSpace(options.ResumeSessionToken)
+	if resume == "" {
+		session, err := client.CreateSession(ctx, partnerID, partnerName, BuildPartnerReturnURL(endpoints.VerificationBaseURL))
+		if err != nil {
+			return "", phasePay, err
+		}
+		return session.SessionToken, phasePay, nil
+	}
+	session, err := client.GetSession(ctx, resume)
+	if err != nil {
+		if isTerminalSessionAPIError(err) {
+			reportSession(options.OnSession, "")
+		}
+		return "", phasePay, err
+	}
+	switch session.Status {
+	case SessionStatusPendingAuth:
+		return resume, phasePay, nil
+	case SessionStatusPendingPayment:
+		return resume, phasePoll, nil
+	case SessionStatusPaid, SessionStatusExchanging, SessionStatusExchanged:
+		return resume, phaseExchange, nil
+	default:
+		reportSession(options.OnSession, "") // drop the dead token → next run starts fresh
+		return "", phasePay, fmt.Errorf("payment %s; re-run the top-up to try again", session.Status)
+	}
+}

internal/aimlapi/stream_test.go

@@ -0,0 +1,457 @@
+package aimlapi
+
+import (
+	"context"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+)
+
+func TestStreamTopUpByKeyFundsWithoutExchange(t *testing.T) {
+	var topupCalls, exchangeCalls int
+	var topupBody map[string]any
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		switch {
+		case r.Method == http.MethodPost && strings.HasSuffix(r.URL.Path, "/partner-checkout/sessions"):
+			_ = json.NewEncoder(w).Encode(PartnerCheckoutSession{SessionToken: "pcs_tok", Status: SessionStatusPendingAuth})
+		case r.Method == http.MethodPost && strings.HasSuffix(r.URL.Path, "/v2/billing/topup"):
+			topupCalls++
+			_ = json.NewDecoder(r.Body).Decode(&topupBody)
+			_ = json.NewEncoder(w).Encode(TopUpByKeyResult{Checkout: PaymentSession{PayURL: "https://pay/checkout"}})
+		case r.Method == http.MethodGet && strings.Contains(r.URL.Path, "/partner-checkout/sessions/"):
+			_ = json.NewEncoder(w).Encode(PartnerCheckoutSession{SessionToken: "pcs_tok", Status: SessionStatusPaid})
+		case r.Method == http.MethodPost && strings.HasSuffix(r.URL.Path, "/exchange"):
+			exchangeCalls++
+			t.Errorf("by-key must never exchange a key")
+		default:
+			t.Errorf("unexpected request: %s %s", r.Method, r.URL.Path)
+		}
+	}))
+	defer server.Close()
+	t.Setenv("AIMLAPI_APP_URL", server.URL)
+	t.Setenv("AIMLAPI_INFERENCE_URL", server.URL+"/v1")
+
+	var opened string
+	result, err := StreamTopUpByKey(context.Background(), StreamTopUpByKeyOptions{
+		APIKey:           "sk-user",
+		AmountUSD:        "20",
+		AutoTopUp:        true,
+		PaymentSessionID: "pay-1",
+		OpenBrowser:      func(u string) error { opened = u; return nil },
+	})
+	if err != nil {
+		t.Fatalf("StreamTopUpByKey = %v", err)
+	}
+	if result.APIKey != "" {
+		t.Fatalf("by-key must not mint a key, got %q", result.APIKey)
+	}
+	if topupCalls != 1 || exchangeCalls != 0 {
+		t.Fatalf("topup=%d exchange=%d, want 1/0", topupCalls, exchangeCalls)
+	}
+	if opened != "https://pay/checkout" {
+		t.Fatalf("opened checkout = %q", opened)
+	}
+	if topupBody["autoTopUp"] != true || topupBody["paymentSessionId"] != "pay-1" {
+		t.Fatalf("topup body missing autoTopUp/paymentSessionId: %#v", topupBody)
+	}
+}
+
+func TestValidatedCheckoutURLRequiresAbsoluteHTTPS(t *testing.T) {
+	tests := []struct {
+		name    string
+		value   string
+		want    string
+		wantErr bool
+	}{
+		{name: "https", value: " https://pay.example.test/checkout?id=1 ", want: "https://pay.example.test/checkout?id=1"},
+		{name: "empty", wantErr: true},
+		{name: "relative", value: "/checkout", wantErr: true},
+		{name: "http", value: "http://pay.example.test/checkout", wantErr: true},
+		{name: "file", value: "file:///tmp/checkout", wantErr: true},
+		{name: "custom protocol", value: "wallet://checkout/123", wantErr: true},
+		{name: "userinfo", value: "https://[email protected]/checkout", wantErr: true},
+	}
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			got, err := validatedCheckoutURL(test.value)
+			if (err != nil) != test.wantErr {
+				t.Fatalf("validatedCheckoutURL(%q) error = %v", test.value, err)
+			}
+			if got != test.want {
+				t.Fatalf("validatedCheckoutURL(%q) = %q, want %q", test.value, got, test.want)
+			}
+		})
+	}
+}
+
+func TestStreamTopUpByKeyPinsValidatedInferenceEndpoint(t *testing.T) {
+	validatedCalls := 0
+	validated := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		switch {
+		case r.Method == http.MethodPost && strings.HasSuffix(r.URL.Path, "/partner-checkout/sessions"):
+			_ = json.NewEncoder(w).Encode(PartnerCheckoutSession{SessionToken: "pcs_tok", Status: SessionStatusPendingAuth})
+		case r.Method == http.MethodPost && strings.HasSuffix(r.URL.Path, "/v2/billing/topup"):
+			validatedCalls++
+			_ = json.NewEncoder(w).Encode(TopUpByKeyResult{Checkout: PaymentSession{PayURL: "https://pay/checkout"}})
+		case r.Method == http.MethodGet && strings.Contains(r.URL.Path, "/partner-checkout/sessions/"):
+			_ = json.NewEncoder(w).Encode(PartnerCheckoutSession{SessionToken: "pcs_tok", Status: SessionStatusPaid})
+		default:
+			t.Errorf("unexpected validated-endpoint request: %s %s", r.Method, r.URL.Path)
+		}
+	}))
+	defer validated.Close()
+	override := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		t.Errorf("validated key leaked to process override: %s %s", r.Method, r.URL.Path)
+	}))
+	defer override.Close()
+	t.Setenv("AIMLAPI_APP_URL", validated.URL)
+	t.Setenv("AIMLAPI_INFERENCE_URL", override.URL)
+
+	_, err := StreamTopUpByKey(context.Background(), StreamTopUpByKeyOptions{
+		APIKey:           "production-key",
+		AmountUSD:        "20",
+		InferenceBaseURL: validated.URL,
+		PaymentSessionID: "payment-1",
+		OpenBrowser:      func(string) error { return nil },
+	})
+	if err != nil {
+		t.Fatalf("StreamTopUpByKey() error = %v", err)
+	}
+	if validatedCalls != 1 {
+		t.Fatalf("validated endpoint top-up calls = %d, want 1", validatedCalls)
+	}
+}
+
+func TestStreamTopUpEmailRetrySendsStablePaymentSessionID(t *testing.T) {
+	getCalls := 0
+	var paymentSessionID string
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		switch {
+		case r.Method == http.MethodGet && strings.Contains(r.URL.Path, "/partner-checkout/sessions/"):
+			status := SessionStatusPendingAuth
+			if getCalls > 0 {
+				status = SessionStatusPaid
+			}
+			getCalls++
+			_ = json.NewEncoder(w).Encode(PartnerCheckoutSession{SessionToken: "pcs_retry", Status: status})
+		case r.Method == http.MethodPost && strings.HasSuffix(r.URL.Path, "/pay"):
+			var body map[string]any
+			_ = json.NewDecoder(r.Body).Decode(&body)
+			paymentSessionID, _ = body["paymentSessionId"].(string)
+			_ = json.NewEncoder(w).Encode(PayResult{Checkout: PaymentSession{PayURL: "https://pay/checkout"}})
+		default:
+			t.Errorf("unexpected request: %s %s", r.Method, r.URL.Path)
+		}
+	}))
+	defer server.Close()
+	t.Setenv("AIMLAPI_APP_URL", server.URL)
+
+	_, err := StreamTopUp(context.Background(), StreamTopUpOptions{
+		SessionToken:       "bearer",
+		ResumeSessionToken: "pcs_retry",
+		PaymentSessionID:   "payment-stable",
+		AmountUSD:          "20",
+		NoOpen:             true,
+	})
+	if err != nil {
+		t.Fatalf("StreamTopUp() error = %v", err)
+	}
+	if paymentSessionID != "payment-stable" {
+		t.Fatalf("paymentSessionId = %q, want stable retry id", paymentSessionID)
+	}
+}
+
+// resumeServer is a partner-checkout backend fake that fails the test if the
+// re-charging endpoints (create-session / pay) are ever hit, and serves a
+// scripted GetSession status sequence so a resume can be driven deterministically.
+type resumeServer struct {
+	t          *testing.T
+	statuses   []string // GetSession status per call, last value repeats
+	getCalls   int
+	payCalls   int
+	createCall int
+	exchanged  int
+}
+
+func (rs *resumeServer) handler() http.Handler {
+	return http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		switch {
+		case r.Method == http.MethodPost && strings.HasSuffix(r.URL.Path, "/partner-checkout/sessions"):
+			rs.createCall++
+			rs.t.Errorf("resume must not create a new session")
+		case r.Method == http.MethodPost && strings.HasSuffix(r.URL.Path, "/pay"):
+			rs.payCalls++
+			rs.t.Errorf("resume must not re-pay (double charge)")
+		case r.Method == http.MethodPost && strings.HasSuffix(r.URL.Path, "/exchange"):
+			rs.exchanged++
+			_ = json.NewEncoder(w).Encode(ExchangeResult{APIKey: "sk-resumed", APIKeyID: "key_1"})
+		case r.Method == http.MethodGet:
+			status := rs.statuses[len(rs.statuses)-1]
+			if rs.getCalls < len(rs.statuses) {
+				status = rs.statuses[rs.getCalls]
+			}
+			rs.getCalls++
+			_ = json.NewEncoder(w).Encode(PartnerCheckoutSession{SessionToken: "pc_tok", Status: SessionStatus(status)})
+		default:
+			rs.t.Errorf("unexpected request: %s %s", r.Method, r.URL.Path)
+		}
+	})
+}
+
+func TestStreamTopUpResumePaidExchangesWithoutRecharge(t *testing.T) {
+	rs := &resumeServer{t: t, statuses: []string{"paid"}}
+	server := httptest.NewServer(rs.handler())
+	defer server.Close()
+	t.Setenv("AIMLAPI_APP_URL", server.URL)
+
+	var seen string
+	result, err := StreamTopUp(context.Background(), StreamTopUpOptions{
+		SessionToken:       "bearer",
+		ResumeSessionToken: "pc_tok",
+		PaymentSessionID:   "payment-1",
+		AmountUSD:          "20",
+		Exchange:           true,
+		NoOpen:             true,
+		OnSession:          func(token string) { seen = token },
+	})
+	if err != nil {
+		t.Fatalf("StreamTopUp resume = %v", err)
+	}
+	if result.APIKey != "sk-resumed" {
+		t.Fatalf("APIKey = %q, want recovered key", result.APIKey)
+	}
+	if rs.exchanged != 1 {
+		t.Fatalf("exchange calls = %d, want 1", rs.exchanged)
+	}
+	if seen != "pc_tok" {
+		t.Fatalf("OnSession token = %q, want retained pc_tok", seen)
+	}
+}
+
+func TestStreamTopUpResumePendingPaymentPollsNeverRepays(t *testing.T) {
+	// First GetSession (resume resolution) sees the still-open checkout; the next
+	// (poll) sees it settle. Pay must never be called — the fake fails the test if
+	// it is.
+	rs := &resumeServer{t: t, statuses: []string{"pending_payment", "paid"}}
+	server := httptest.NewServer(rs.handler())
+	defer server.Close()
+	t.Setenv("AIMLAPI_APP_URL", server.URL)
+
+	result, err := StreamTopUp(context.Background(), StreamTopUpOptions{
+		SessionToken:       "bearer",
+		ResumeSessionToken: "pc_tok",
+		PaymentSessionID:   "payment-1",
+		AmountUSD:          "20",
+		Exchange:           true,
+		NoOpen:             true,
+	})
+	if err != nil {
+		t.Fatalf("StreamTopUp resume = %v", err)
+	}
+	if result.APIKey != "sk-resumed" {
+		t.Fatalf("APIKey = %q, want recovered key", result.APIKey)
+	}
+	if rs.payCalls != 0 {
+		t.Fatalf("pay calls = %d, want 0", rs.payCalls)
+	}
+}
+
+func TestStreamTopUpPollingDeadSessionClearsRetainedToken(t *testing.T) {
+	rs := &resumeServer{t: t, statuses: []string{"pending_payment", "expired"}}
+	server := httptest.NewServer(rs.handler())
+	defer server.Close()
+	t.Setenv("AIMLAPI_APP_URL", server.URL)
+
+	seen := []string{}
+	_, err := StreamTopUp(context.Background(), StreamTopUpOptions{
+		SessionToken:       "bearer",
+		ResumeSessionToken: "pc_tok",
+		PaymentSessionID:   "payment-1",
+		AmountUSD:          "20",
+		Exchange:           true,
+		NoOpen:             true,
+		OnSession:          func(token string) { seen = append(seen, token) },
+	})
+	if err == nil || !strings.Contains(err.Error(), "payment expired") {
+		t.Fatalf("err = %v, want expired payment error", err)
+	}
+	if len(seen) == 0 || seen[len(seen)-1] != "" {
+		t.Fatalf("OnSession events = %#v, want terminal clear", seen)
+	}
+}
+
+func TestPollUntilPaidTerminal4xxClearsRetainedToken(t *testing.T) {
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		http.Error(w, "session rejected", http.StatusNotFound)
+	}))
+	defer server.Close()
+
+	seen := "retained"
+	_, err := pollUntilPaid(context.Background(), NewClient(Endpoints{AppBaseURL: server.URL}, server.Client()), "pc_missing", func(token string) {
+		seen = token
+	})
+	if err == nil {
+		t.Fatal("pollUntilPaid() error = nil, want terminal 4xx")
+	}
+	if seen != "" {
+		t.Fatalf("retained session = %q, want cleared", seen)
+	}
+}
+
+func TestPollUntilExchangeSettledClearsDeadSession(t *testing.T) {
+	for _, test := range []struct {
+		name       string
+		statusCode int
+		status     SessionStatus
+	}{
+		{name: "terminal API rejection", statusCode: http.StatusNotFound},
+		{name: "expired exchange claim", status: SessionStatusExpired},
+	} {
+		t.Run(test.name, func(t *testing.T) {
+			server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+				if test.statusCode != 0 {
+					http.Error(w, "session rejected", test.statusCode)
+					return
+				}
+				_ = json.NewEncoder(w).Encode(PartnerCheckoutSession{SessionToken: "pcs_dead", Status: test.status})
+			}))
+			defer server.Close()
+
+			seen := "retained"
+			err := pollUntilExchangeSettled(context.Background(), NewClient(Endpoints{AppBaseURL: server.URL}, server.Client()), "pcs_dead", func(token string) {
+				seen = token
+			})
+			if err == nil {
+				t.Fatal("pollUntilExchangeSettled() error = nil")
+			}
+			if seen != "" {
+				t.Fatalf("retained session = %q, want cleared", seen)
+			}
+		})
+	}
+}
+
+func TestResumeHelpersClearOnlyRejectedSessions(t *testing.T) {
+	tests := []struct {
+		name      string
+		status    int
+		wantClear bool
+	}{
+		{name: "terminal 4xx", status: http.StatusNotFound, wantClear: true},
+		{name: "retryable 5xx", status: http.StatusServiceUnavailable, wantClear: false},
+	}
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+				http.Error(w, "session lookup failed", test.status)
+			}))
+			defer server.Close()
+			client := NewClient(Endpoints{AppBaseURL: server.URL}, server.Client())
+
+			for _, flow := range []struct {
+				name    string
+				resolve func(func(string)) error
+			}{
+				{
+					name: "session bearer",
+					resolve: func(onSession func(string)) error {
+						_, _, err := resolveTopupSession(context.Background(), client, StreamTopUpOptions{
+							ResumeSessionToken: "pc_rejected", OnSession: onSession,
+						}, DefaultPartnerID, DefaultPartnerName, Endpoints{})
+						return err
+					},
+				},
+				{
+					name: "api key",
+					resolve: func(onSession func(string)) error {
+						_, _, err := resolveByKeySession(context.Background(), client, StreamTopUpByKeyOptions{
+							ResumeSessionToken: "pc_rejected", OnSession: onSession,
+						}, DefaultPartnerID, DefaultPartnerName, Endpoints{})
+						return err
+					},
+				},
+			} {
+				t.Run(flow.name, func(t *testing.T) {
+					seen := "retained"
+					if err := flow.resolve(func(token string) { seen = token }); err == nil {
+						t.Fatal("resolve error = nil")
+					}
+					if cleared := seen == ""; cleared != test.wantClear {
+						t.Fatalf("session cleared = %v, want %v", cleared, test.wantClear)
+					}
+				})
+			}
+		})
+	}
+}
+
+func TestStreamTopUpResumeExchangedReturnsRecoveryError(t *testing.T) {
+	rs := &resumeServer{t: t, statuses: []string{"exchanged"}}
+	server := httptest.NewServer(rs.handler())
+	defer server.Close()
+	t.Setenv("AIMLAPI_APP_URL", server.URL)
+
+	_, err := StreamTopUp(context.Background(), StreamTopUpOptions{
+		SessionToken:       "bearer",
+		ResumeSessionToken: "pc_tok",
+		PaymentSessionID:   "payment-1",
+		AmountUSD:          "20",
+		Exchange:           true,
+		NoOpen:             true,
+	})
+	if err == nil || !strings.Contains(err.Error(), "already exchanged") {
+		t.Fatalf("err = %v, want an already-exchanged recovery error", err)
+	}
+	if rs.exchanged != 0 {
+		t.Fatalf("exchange calls = %d, want 0 (no second key)", rs.exchanged)
+	}
+}
+
+func TestStreamTopUpResumeExchangingPollsWithoutSecondExchange(t *testing.T) {
+	rs := &resumeServer{t: t, statuses: []string{"exchanging", "exchanged"}}
+	server := httptest.NewServer(rs.handler())
+	defer server.Close()
+	t.Setenv("AIMLAPI_APP_URL", server.URL)
+
+	_, err := StreamTopUp(context.Background(), StreamTopUpOptions{
+		SessionToken:       "bearer",
+		ResumeSessionToken: "pc_tok",
+		PaymentSessionID:   "payment-1",
+		AmountUSD:          "20",
+		Exchange:           true,
+		NoOpen:             true,
+	})
+	if err == nil || !strings.Contains(err.Error(), "already exchanged") {
+		t.Fatalf("err = %v, want dashboard recovery guidance", err)
+	}
+	if rs.exchanged != 0 {
+		t.Fatalf("exchange calls = %d, want 0 while an exchange claim is in flight", rs.exchanged)
+	}
+	if rs.createCall != 0 || rs.payCalls != 0 {
+		t.Fatalf("resume created/paid a new checkout: create=%d pay=%d", rs.createCall, rs.payCalls)
+	}
+}
+
+func TestStreamTopUpPollObservesExchangingWithoutSecondExchange(t *testing.T) {
+	rs := &resumeServer{t: t, statuses: []string{"pending_payment", "exchanging", "exchanged"}}
+	server := httptest.NewServer(rs.handler())
+	defer server.Close()
+	t.Setenv("AIMLAPI_APP_URL", server.URL)
+
+	_, err := StreamTopUp(context.Background(), StreamTopUpOptions{
+		SessionToken:       "bearer",
+		ResumeSessionToken: "pc_tok",
+		PaymentSessionID:   "payment-1",
+		AmountUSD:          "20",
+		Exchange:           true,
+		NoOpen:             true,
+	})
+	if err == nil || !strings.Contains(err.Error(), "already exchanged") {
+		t.Fatalf("err = %v, want dashboard recovery guidance", err)
+	}
+	if rs.exchanged != 0 {
+		t.Fatalf("exchange calls = %d, want 0 after poll observed an existing claim", rs.exchanged)
+	}
+}

internal/aimlapi/topup.go

@@ -0,0 +1,186 @@
+package aimlapi
+
+import (
+	"context"
+	"crypto/rand"
+	"errors"
+	"fmt"
+	"math"
+	"net/http"
+	"strconv"
+	"strings"
+	"time"
+)
+
+// Status is a progress phase reported by StreamTopUp through its OnStatus callback.
+type Status string
+
+const (
+	StatusCreatingSession Status = "creating-session"
+	StatusOpeningCheckout Status = "opening-checkout"
+	StatusWaitingPayment  Status = "waiting-payment"
+	StatusProvisioningKey Status = "provisioning-key"
+)
+
+// ProvisionedKey is the outcome of a top-up: the (optionally minted) API key plus
+// the inference base URL and model to write into the provider profile.
+type ProvisionedKey struct {
+	APIKey   string
+	APIKeyID string
+	BaseURL  string
+	Model    string
+}
+
+const (
+	pollInterval = 3 * time.Second
+	pollTimeout  = 20 * time.Minute
+)
+
+// ParseAmountUSD parses a dollar string into USD minor units (cents), enforcing the
+// min/max top-up bounds. An empty value yields the default amount.
+func ParseAmountUSD(value string) (int, error) {
+	value = strings.TrimSpace(value)
+	if value == "" {
+		return DefaultAmountUSDMinor, nil
+	}
+	dollars, err := strconv.ParseFloat(value, 64)
+	// NaN/±Inf parse without error and slip past the min/max comparisons below
+	// (every ordered comparison with NaN is false), so reject them explicitly
+	// before the cents conversion turns them into a garbage minor-unit amount.
+	if err != nil || math.IsNaN(dollars) || math.IsInf(dollars, 0) || dollars <= 0 {
+		return 0, fmt.Errorf("invalid amount %q; pass a positive USD amount", value)
+	}
+	minor := int(dollars*100 + 0.5)
+	if minor < MinAmountUSDMinor {
+		return 0, fmt.Errorf("minimum top-up is $%d", MinAmountUSDMinor/100)
+	}
+	if minor > MaxAmountUSDMinor {
+		return 0, fmt.Errorf("maximum top-up is $%d", MaxAmountUSDMinor/100)
+	}
+	return minor, nil
+}
+
+// NewPaymentSessionID returns a random UUIDv4-shaped idempotency id for a top-up.
+// The caller generates it once per top-up intent and reuses it on retry so the
+// backend (and Stripe) coalesce retries onto the same checkout instead of a
+// second charge.
+func NewPaymentSessionID() (string, error) {
+	var b [16]byte
+	if _, err := rand.Read(b[:]); err != nil {
+		return "", err
+	}
+	b[6] = (b[6] & 0x0f) | 0x40 // version 4
+	b[8] = (b[8] & 0x3f) | 0x80 // variant 10
+	return fmt.Sprintf("%x-%x-%x-%x-%x", b[0:4], b[4:6], b[6:8], b[8:10], b[10:16]), nil
+}
+
+func pollUntilPaid(ctx context.Context, client *Client, sessionToken string, onSession func(string)) (PartnerCheckoutSession, error) {
+	deadline := time.Now().Add(pollTimeout)
+	for time.Now().Before(deadline) {
+		session, err := client.GetSession(ctx, sessionToken)
+		if err != nil {
+			// Context cancellation/deadline is terminal (also covers an in-flight
+			// request cancelled mid-poll, whose transport error wraps ctx.Err()).
+			if errors.Is(err, context.Canceled) || errors.Is(err, context.DeadlineExceeded) {
+				return PartnerCheckoutSession{}, err
+			}
+			// A 4xx API response is terminal. Everything else — a 5xx or a transient
+			// transport error (non-APIError, e.g. a dropped connection during the
+			// wait) — is retried until the poll deadline.
+			if isTerminalSessionAPIError(err) {
+				reportSession(onSession, "")
+				return PartnerCheckoutSession{}, err
+			}
+			if err := sleepContext(ctx, pollInterval); err != nil {
+				return PartnerCheckoutSession{}, err
+			}
+			continue
+		}
+		switch session.Status {
+		case SessionStatusPaid, SessionStatusExchanging:
+			return session, nil
+		case SessionStatusExchanged:
+			return PartnerCheckoutSession{}, fmt.Errorf("session was already exchanged; rotate the key from the aimlapi.com dashboard")
+		case SessionStatusCancelled, SessionStatusExpired, SessionStatusFailed:
+			// The checkout cannot recover from these states. Drop the retained token
+			// (and the TUI's paired by-key idempotency ID) now so the very next retry
+			// creates a fresh payment intent.
+			reportSession(onSession, "")
+			return PartnerCheckoutSession{}, fmt.Errorf("payment %s; re-run the top-up to try again", session.Status)
+		default:
+			if err := sleepContext(ctx, pollInterval); err != nil {
+				return PartnerCheckoutSession{}, err
+			}
+		}
+	}
+	return PartnerCheckoutSession{}, fmt.Errorf("timed out waiting for payment; re-run once the payment clears")
+}
+
+func isTerminalSessionAPIError(err error) bool {
+	var apiErr APIError
+	return errors.As(err, &apiErr) && apiErr.Status >= http.StatusBadRequest && apiErr.Status < http.StatusInternalServerError
+}
+
+// pollUntilExchangeSettled follows an exchange claim that was already in flight
+// when a retry resumed the session. Exchange is single-flight and the raw key is
+// returned only to the winning request, so issuing Exchange again is never safe.
+// Once the claim completes, direct the user to the dashboard recovery path.
+func pollUntilExchangeSettled(ctx context.Context, client *Client, sessionToken string, onSession func(string)) error {
+	deadline := time.Now().Add(pollTimeout)
+	for time.Now().Before(deadline) {
+		session, err := client.GetSession(ctx, sessionToken)
+		if err != nil {
+			if errors.Is(err, context.Canceled) || errors.Is(err, context.DeadlineExceeded) {
+				return err
+			}
+			if isTerminalSessionAPIError(err) {
+				reportSession(onSession, "")
+				return err
+			}
+			if err := sleepContext(ctx, pollInterval); err != nil {
+				return err
+			}
+			continue
+		}
+		switch session.Status {
+		case SessionStatusExchanged:
+			return fmt.Errorf("session was already exchanged; rotate the key from the aimlapi.com dashboard")
+		case SessionStatusExchanging:
+			if err := sleepContext(ctx, pollInterval); err != nil {
+				return err
+			}
+		case SessionStatusCancelled, SessionStatusExpired, SessionStatusFailed:
+			reportSession(onSession, "")
+			return fmt.Errorf("key provisioning %s; rotate the key from the aimlapi.com dashboard", session.Status)
+		default:
+			return fmt.Errorf("key provisioning returned to %s; re-run the top-up", session.Status)
+		}
+	}
+	return fmt.Errorf("timed out waiting for key provisioning; retry to check the same session")
+}
+
+func sleepContext(ctx context.Context, delay time.Duration) error {
+	timer := time.NewTimer(delay)
+	defer timer.Stop()
+	select {
+	case <-ctx.Done():
+		return ctx.Err()
+	case <-timer.C:
+		return nil
+	}
+}
+
+func status(callback func(Status, string), value Status, detail string) {
+	if callback != nil {
+		callback(value, detail)
+	}
+}
+
+func firstNonEmpty(values ...string) string {
+	for _, value := range values {
+		if value = strings.TrimSpace(value); value != "" {
+			return value
+		}
+	}
+	return ""
+}

internal/aimlapi/validation.go

@@ -0,0 +1,38 @@
+package aimlapi
+
+import (
+	"net/mail"
+	"strings"
+)
+
+// ValidEmail performs the lightweight validation used before starting the
+// passwordless onboarding request.
+func ValidEmail(value string) bool {
+	value = strings.TrimSpace(value)
+	address, err := mail.ParseAddress(value)
+	if err != nil || address.Address != value {
+		return false
+	}
+	at := strings.LastIndexByte(value, '@')
+	if at <= 0 || at == len(value)-1 {
+		return false
+	}
+	domain := value[at+1:]
+	if strings.HasPrefix(domain, ".") || strings.HasSuffix(domain, ".") || strings.Contains(domain, "..") {
+		return false
+	}
+	dot := strings.LastIndexByte(domain, '.')
+	if dot <= 0 || dot == len(domain)-1 {
+		return false
+	}
+	tld := domain[dot+1:]
+	if len(tld) < 2 {
+		return false
+	}
+	for _, char := range tld {
+		if char < 'A' || char > 'Z' && char < 'a' || char > 'z' {
+			return false
+		}
+	}
+	return true
+}

internal/background/process_posix.go

@@ -3,11 +3,10 @@
 package background
 
 import (
-	"errors"
-	"fmt"
 	"os/exec"
-	"syscall"
 	"time"
+
+	"github.com/Gitlawb/zero/internal/execution"
 )
 
 // terminationGracePeriod is how long a process has to exit after SIGTERM before
@@ -22,10 +21,7 @@ var (
 // the negative PID (the group), so any process the child forks dies with it
 // instead of being orphaned. Must be called before cmd.Start.
 func ConfigureChildProcessGroup(cmd *exec.Cmd) {
-	if cmd.SysProcAttr == nil {
-		cmd.SysProcAttr = &syscall.SysProcAttr{}
-	}
-	cmd.SysProcAttr.Setpgid = true
+	execution.ConfigureProcessGroup(cmd)
 }
 
 // terminateProcess stops a background process. It first asks politely with
@@ -41,71 +37,5 @@ func ConfigureChildProcessGroup(cmd *exec.Cmd) {
 // back to signalling only the individual PID, which also avoids reporting a false
 // success when a non-leader group-signal returns ESRCH.
 func terminateProcess(pid int) error {
-	// Guard pid <= 1: kill(-0) would target our OWN process group and kill(-1)
-	// every process we can signal. A real child PID is always > 1, so never let a
-	// bogus 0/1 expand into either.
-	if pid <= 1 {
-		return fmt.Errorf("refusing to terminate invalid pid %d", pid)
-	}
-
-	// Signal the whole group only when pid leads its own group; otherwise the
-	// individual process. See the doc comment for why a non-leader is signalled
-	// directly rather than via its (foreign) group.
-	target := pid
-	if pgid, err := syscall.Getpgid(pid); err == nil {
-		if pgid == pid {
-			target = -pid
-		}
-	} else if processGoneError(err) {
-		return nil // already gone
-	}
-
-	alive := func() bool { return syscall.Kill(target, syscall.Signal(0)) == nil }
-
-	if err := syscall.Kill(target, syscall.SIGTERM); err != nil {
-		if processGoneError(err) {
-			return nil
-		}
-		return err
-	}
-
-	// Poll liveness so we return promptly once it exits, rather than always
-	// waiting out the full grace period.
-	deadline := time.Now().Add(terminationGracePeriod)
-	for time.Now().Before(deadline) {
-		if !alive() {
-			return nil
-		}
-		time.Sleep(terminationPollInterval)
-	}
-	if !alive() {
-		return nil
-	}
-
-	// Still alive after the grace period: force-kill.
-	if err := syscall.Kill(target, syscall.SIGKILL); err != nil && !processGoneError(err) {
-		return err
-	}
-
-	// SIGKILL is asynchronous: the kernel may not have reaped it yet. Poll again
-	// so this helper only reports success once the target is actually gone —
-	// otherwise the caller gets nil while descendants are still racing.
-	deadline = time.Now().Add(terminationGracePeriod)
-	for time.Now().Before(deadline) {
-		if !alive() {
-			return nil
-		}
-		time.Sleep(terminationPollInterval)
-	}
-	if alive() {
-		return fmt.Errorf("process %d did not exit after SIGKILL", pid)
-	}
-	return nil
-}
-
-// processGoneError reports whether an error means the process group has already
-// exited (so termination is effectively done). syscall.Kill reports ESRCH when
-// no process in the target group remains.
-func processGoneError(err error) bool {
-	return errors.Is(err, syscall.ESRCH)
+	return execution.TerminateProcessTree(pid, terminationGracePeriod, terminationPollInterval)
 }

internal/background/process_posix_test.go

@@ -149,12 +149,9 @@ func TestTerminateProcessKillsForkedChildren(t *testing.T) {
 
 	// The forked child must be gone too (poll until reaped by init).
 	deadline := time.Now().Add(2 * time.Second)
-	for {
-		// Only ESRCH proves the child is gone; any other error (e.g. EPERM) would
-		// wrongly pass the test, so treat it as still-present and keep polling.
-		if errors.Is(syscall.Kill(childPID, syscall.Signal(0)), syscall.ESRCH) {
-			break // child no longer exists
-		}
+	// Only ESRCH proves the child is gone; any other error (e.g. EPERM) would
+	// wrongly pass the test, so treat it as still-present and keep polling.
+	for !errors.Is(syscall.Kill(childPID, syscall.Signal(0)), syscall.ESRCH) {
 		if time.Now().After(deadline) {
 			_ = syscall.Kill(childPID, syscall.SIGKILL)
 			t.Fatalf("forked child %d survived terminateProcess — group kill failed", childPID)

internal/background/process_windows.go

@@ -3,23 +3,16 @@
 package background
 
 import (
-	"os"
 	"os/exec"
-	"strconv"
+
+	"github.com/Gitlawb/zero/internal/execution"
 )
 
-// ConfigureChildProcessGroup is a no-op on Windows: terminateProcess uses
-// `taskkill /T` to kill the whole process tree, so no launch-time process-group
+// ConfigureChildProcessGroup is a no-op on Windows: process-tree termination is
+// delegated to execution.TerminateProcessTree, so no launch-time process-group
 // setup is required (the POSIX build sets Setpgid here instead).
-func ConfigureChildProcessGroup(cmd *exec.Cmd) {}
+func ConfigureChildProcessGroup(cmd *exec.Cmd) { execution.ConfigureProcessGroup(cmd) }
 
 func terminateProcess(pid int) error {
-	if err := exec.Command("taskkill", "/T", "/F", "/PID", strconv.Itoa(pid)).Run(); err == nil {
-		return nil
-	}
-	process, err := os.FindProcess(pid)
-	if err != nil {
-		return err
-	}
-	return process.Kill()
+	return execution.TerminateProcessTree(pid, 0, 0)
 }

internal/cli/acp.go

@@ -1,6 +1,7 @@
 package cli
 
 import (
+	"context"
 	"fmt"
 	"io"
 	"os"
@@ -9,6 +10,7 @@ import (
 	"github.com/Gitlawb/zero/internal/acp"
 	"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/tools"
 )
@@ -25,6 +27,8 @@ Usage:
 Not meant to be run interactively — point your editor's ACP / external-agent
 setting at "zero acp".`
 
+var acpSignalContext = signalContext
+
 // runACP serves ACP over stdio so an editor can drive ZERO's agent core. It
 // speaks JSON-RPC 2.0 (newline-delimited JSON) on stdin/stdout; stderr stays free
 // for human-readable diagnostics. The session lifecycle maps onto ZERO's own
@@ -40,9 +44,14 @@ func runACP(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) int
 		}
 	}
 
-	conn := acp.NewConn(deps.stdin, stdout)
+	// The process's own stdin belongs exclusively to this Conn for the life of
+	// the command, so cancellation may close it to interrupt an idle read.
+	conn := acp.NewOwnedConn(deps.stdin, stdout)
 	acp.NewAgent(conn, acp.Deps{
 		ResolveConfig: deps.resolveConfig,
+		DiscoverModels: func(ctx context.Context, profile config.ProviderProfile) ([]providermodeldiscovery.Model, error) {
+			return defaultDiscoverProviderModels(ctx, discoveryCredentialProfile(profile))
+		},
 		// deps.newProvider is wrapped in fillAppDeps to apply the stored API key,
 		// so ACP is authenticated for apiKeyStored profiles like every other
 		// surface — no ACP-specific credential handling needed.
@@ -68,9 +77,9 @@ func runACP(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) int
 		AgentInfo:            acp.Implementation{Name: "zero", Version: version},
 	})
 
-	ctx, stop := signalContext()
+	ctx, stop := acpSignalContext()
 	defer stop()
-	if err := conn.Serve(ctx); err != nil && ctx.Err() == nil {
+	if err := conn.Serve(ctx); err != nil {
 		return writeAppError(stderr, "acp: "+err.Error(), exitCrash)
 	}
 	return exitSuccess

internal/cli/acp_test.go

@@ -0,0 +1,120 @@
+package cli
+
+import (
+	"bytes"
+	"context"
+	"errors"
+	"io"
+	"strings"
+	"sync"
+	"testing"
+	"time"
+)
+
+type acpTestReader func([]byte) (int, error)
+
+func (r acpTestReader) Read(p []byte) (int, error) { return r(p) }
+
+type acpTestWriter func([]byte) (int, error)
+
+func (w acpTestWriter) Write(p []byte) (int, error) { return w(p) }
+
+func installACPTestContext(t *testing.T) context.CancelFunc {
+	t.Helper()
+	ctx, cancel := context.WithCancel(context.Background())
+	previous := acpSignalContext
+	acpSignalContext = func() (context.Context, context.CancelFunc) {
+		return ctx, cancel
+	}
+	t.Cleanup(func() {
+		cancel()
+		acpSignalContext = previous
+	})
+	return cancel
+}
+
+func TestRunACPCancellationPreservesTerminalReadError(t *testing.T) {
+	cancel := installACPTestContext(t)
+	wantErr := errors.New("transport read failed")
+	reader := acpTestReader(func(p []byte) (int, error) {
+		return copy(p, `{"jsonrpc":"1.0","id":1}`+"\n"), wantErr
+	})
+	writeStarted := make(chan struct{}, 1)
+	releaseWrite := make(chan struct{})
+	var stdout bytes.Buffer
+	writer := acpTestWriter(func(p []byte) (int, error) {
+		writeStarted <- struct{}{}
+		<-releaseWrite
+		return stdout.Write(p)
+	})
+	var stderr bytes.Buffer
+	done := make(chan int, 1)
+	go func() {
+		done <- runACP(nil, writer, &stderr, fillAppDeps(appDeps{stdin: reader}))
+	}()
+
+	select {
+	case <-writeStarted:
+	case <-time.After(time.Second):
+		t.Fatal("ACP did not reach the response write")
+	}
+	// Keep Serve in the synchronous write until cancellation is observable. The
+	// regression requires the genuine read error to win over that cancellation.
+	cancel()
+	close(releaseWrite)
+
+	select {
+	case code := <-done:
+		if code != exitCrash {
+			t.Fatalf("exit code = %d, want crash %d", code, exitCrash)
+		}
+	case <-time.After(time.Second):
+		t.Fatal("ACP did not exit after terminal read error")
+	}
+	if got := stderr.String(); !strings.Contains(got, "acp: "+wantErr.Error()) {
+		t.Fatalf("stderr = %q, want terminal read error", got)
+	}
+}
+
+func TestRunACPIdleCancellationExitsCleanly(t *testing.T) {
+	cancel := installACPTestContext(t)
+	pipeReader, pipeWriter := io.Pipe()
+	t.Cleanup(func() { _ = pipeWriter.Close() })
+	readStarted := make(chan struct{}, 1)
+	reader := &acpNotifyingReadCloser{ReadCloser: pipeReader, readStarted: readStarted}
+	var stdout, stderr bytes.Buffer
+	done := make(chan int, 1)
+	go func() {
+		done <- runACP(nil, &stdout, &stderr, fillAppDeps(appDeps{stdin: reader}))
+	}()
+
+	select {
+	case <-readStarted:
+	case <-time.After(time.Second):
+		t.Fatal("ACP did not begin reading")
+	}
+	cancel()
+
+	select {
+	case code := <-done:
+		if code != exitSuccess {
+			t.Fatalf("exit code = %d, want success %d; stderr: %s", code, exitSuccess, stderr.String())
+		}
+	case <-time.After(time.Second):
+		t.Fatal("ACP did not exit after cancellation")
+	}
+	if stderr.Len() != 0 {
+		t.Fatalf("stderr = %q, want empty", stderr.String())
+	}
+}
+
+type acpNotifyingReadCloser struct {
+	io.ReadCloser
+	readStarted chan<- struct{}
+	once        sync.Once
+}
+
+func (r *acpNotifyingReadCloser) Read(p []byte) (int, error) {
+	r.once.Do(func() { r.readStarted <- struct{}{} })
+	return r.ReadCloser.Read(p)
+}

internal/cli/agent_eval.go

@@ -515,9 +515,7 @@ func agentEvalReportFromBenchmark(suite agenteval.Suite, report agenteval.Benchm
 		if task.Agent.Truncated {
 			converted.Truncated = true
 		}
-		for _, failure := range agentEvalFailuresFromTaskReport(task) {
-			converted.Failures = append(converted.Failures, failure)
-		}
+		converted.Failures = append(converted.Failures, agentEvalFailuresFromTaskReport(task)...)
 	}
 	return converted
 }

internal/cli/app.go

@@ -17,6 +17,7 @@ import (
 
 	"github.com/Gitlawb/zero/internal/agent"
 	"github.com/Gitlawb/zero/internal/config"
+	"github.com/Gitlawb/zero/internal/execution"
 	"github.com/Gitlawb/zero/internal/hooks"
 	"github.com/Gitlawb/zero/internal/localcontrol"
 	"github.com/Gitlawb/zero/internal/mcp"
@@ -42,6 +43,7 @@ import (
 	"github.com/Gitlawb/zero/internal/worktrees"
 	"github.com/Gitlawb/zero/internal/zerogit"
 	"github.com/Gitlawb/zero/internal/zeroruntime"
+	"github.com/charmbracelet/x/term"
 )
 
 var version = "dev"
@@ -57,7 +59,11 @@ type appDeps struct {
 	// config.SetActiveProviderEnv, set in defaultAppDeps — deliberately NOT filled
 	// by fillAppDeps, so tests never mutate the process environment unless they
 	// inject it). nil ⇒ no export.
-	exportActiveProvider   func(providerName string)
+	exportActiveProvider func(providerName string)
+	// getenv reads a process environment variable (production: os.Getenv, set in
+	// defaultAppDeps — deliberately NOT filled by fillAppDeps, so tests are hermetic
+	// against ambient vars like ZERO_PROVIDER unless they inject it). nil ⇒ empty.
+	getenv                 func(string) string
 	probeProviderHealth    func(context.Context, providerhealth.Options) providerhealth.Result
 	discoverProviderModels func(context.Context, config.ProviderProfile) ([]providermodeldiscovery.Model, error)
 	detectLocalRuntimes    func(context.Context, provideronboarding.LocalDetectOptions) []provideronboarding.DetectedLocalRuntime
@@ -117,6 +123,7 @@ func defaultAppDeps() appDeps {
 		stdin:                os.Stdin,
 		userConfigPath:       config.DefaultUserConfigPath,
 		exportActiveProvider: config.SetActiveProviderEnv,
+		getenv:               os.Getenv,
 		resolveConfig: func(workspaceRoot string, overrides config.Overrides) (config.ResolvedConfig, error) {
 			options, err := config.DefaultResolveOptions(workspaceRoot)
 			if err != nil {
@@ -339,7 +346,17 @@ func runWithDeps(args []string, stdout io.Writer, stderr io.Writer, deps appDeps
 				return 0
 			}
 		}
-		if _, err := fmt.Fprintf(stdout, "zero %s\n", version); err != nil {
+		// Pipes and redirects keep the machine-readable contract exactly as it
+		// was: a single "zero <version>" line. Scripts, command substitutions,
+		// and the NPM wrapper smoke check all parse that record, so the banner
+		// is strictly a TTY affordance.
+		if !stdoutIsTerminal(stdout) {
+			if _, err := fmt.Fprintf(stdout, "zero %s\n", version); err != nil {
+				return 1
+			}
+			return 0
+		}
+		if _, err := fmt.Fprintf(stdout, "%s\n\nzero %s\n", tui.Wordmark(), version); err != nil {
 			return 1
 		}
 		return 0
@@ -357,6 +374,8 @@ func runWithDeps(args []string, stdout io.Writer, stderr io.Writer, deps appDeps
 	case "exec":
 		// Forward leading --add-dir occurrences so exec's own parser collects them.
 		return runExec(append(addDirFlagArgs(addDirs), args[1:]...), stdout, stderr, deps)
+	case "completions":
+		return runCompletions(args[1:], stdout, stderr)
 	case "daemon":
 		return runDaemon(args[1:], stdout, stderr, deps)
 	case "config":
@@ -667,6 +686,26 @@ func runInteractiveTUIWithSetup(stderr io.Writer, deps appDeps, permissionMode a
 
 	registry := newCoreRegistryScoped(workspaceRoot, scope)
 	registerLocalControlTools(registry, workspaceRoot, resolved.LocalControl)
+	executionRunner := execution.NewRunner(nil)
+	sandboxStore, err := deps.newSandboxStore()
+	if err != nil {
+		return writeAppError(stderr, "failed to initialize sandbox grants: "+err.Error(), 1)
+	}
+	if notice, err := sandboxStore.ConsumeMigrationNotice(); err != nil {
+		return writeAppError(stderr, "failed to migrate sandbox grants: "+err.Error(), 1)
+	} else if notice != "" {
+		_, _ = fmt.Fprintln(stderr, "[zero] "+notice)
+	}
+	sandboxBackend := deps.selectSandboxBackend(sandbox.BackendOptions{})
+	sandboxEngine := sandbox.NewEngine(sandbox.EngineOptions{
+		WorkspaceRoot:    workspaceRoot,
+		Policy:           applyConfiguredSandboxPolicy(sandbox.DefaultPolicy(), resolved.Sandbox),
+		Store:            sandboxStore,
+		Backend:          sandboxBackend,
+		Scope:            scope,
+		SensitiveEnvKeys: providerSensitiveEnvKeys(resolved),
+	})
+	executionRunner.SetPreparer(sandboxEngine)
 	specialistRuntime, err := registerSpecialistTools(registry, workspaceRoot, resolved.Swarm.MaxTeamSize)
 	if err != nil {
 		return writeAppError(stderr, "failed to initialize specialist tools: "+err.Error(), 1)
@@ -701,6 +740,8 @@ func runInteractiveTUIWithSetup(stderr io.Writer, deps appDeps, permissionMode a
 		mcpRuntime, err = deps.registerMCPTools(context.Background(), registry, mcpConfig, mcp.RegisterOptions{
 			PermissionStore: mcpPermissionStore,
 			Autonomy:        mcp.AutonomyLow,
+			Execution:       executionRunner,
+			WorkspaceRoot:   workspaceRoot,
 		})
 	}
 	if err != nil {
@@ -727,7 +768,7 @@ func runInteractiveTUIWithSetup(stderr io.Writer, deps appDeps, permissionMode a
 	// The interactive TUI is not worktree-reassigned, so the trust root is the
 	// launch directory itself.
 	trustRoot := workspaceRoot
-	pluginActivation := activatePlugins(workspaceRoot, registry, deps, stderr, trustRoot)
+	pluginActivation := activatePlugins(workspaceRoot, registry, deps, stderr, trustRoot, executionRunner)
 	// Ask (not Auto) is the interactive default: in Auto, ToolAdvertised exposes
 	// only PermissionAllow tools, so prompt-gated tools (write_file/edit_file/bash/
 	// apply_patch) would never be offered to the model — the TUI could neither edit
@@ -749,18 +790,6 @@ func runInteractiveTUIWithSetup(stderr io.Writer, deps appDeps, permissionMode a
 	// unchanged. The interactive surface applies no operator tool filters, so
 	// enabled/disabled are nil — matching the AgentOptions below.
 	registerToolSearchIfEligible(registry, resolved.Tools.DeferThreshold, permissionMode, nil, nil)
-	sandboxStore, err := deps.newSandboxStore()
-	if err != nil {
-		return writeAppError(stderr, "failed to initialize sandbox grants: "+err.Error(), 1)
-	}
-	sandboxBackend := deps.selectSandboxBackend(sandbox.BackendOptions{})
-	sandboxEngine := sandbox.NewEngine(sandbox.EngineOptions{
-		WorkspaceRoot: workspaceRoot,
-		Policy:        applyConfiguredSandboxPolicy(sandbox.DefaultPolicy(), resolved.Sandbox),
-		Store:         sandboxStore,
-		Backend:       sandboxBackend,
-		Scope:         scope,
-	})
 	lastKnownMCPConfig := mcpConfig
 	fileTracker := tools.NewFileTracker()
 	var scratchBaseline scratchFileBaseline
@@ -777,7 +806,7 @@ func runInteractiveTUIWithSetup(stderr io.Writer, deps appDeps, permissionMode a
 	// Build the hooks dispatcher out of the AgentOptions literal so its trust skip
 	// report can be combined with the plugin activation's, and emit at most one
 	// notice when project hooks/plugins were dropped for an untrusted workspace.
-	hookDispatcher, hookSkip := newHookDispatcherWithExtra(workspaceRoot, pluginActivation.hooks, trustRoot)
+	hookDispatcher, hookSkip := newHookDispatcherWithExtra(workspaceRoot, pluginActivation.hooks, trustRoot, executionRunner)
 	emitTrustNotice(stderr, hookSkip, pluginActivation.trustSkip, mcpSkip)
 	return deps.runTUI(context.Background(), tui.Options{
 		Cwd:                  workspaceRoot,
@@ -1115,6 +1144,17 @@ func closeSpecialistRuntime(stderr io.Writer, runtime *agentToolRuntime) {
 	}
 }
 
+// stdoutIsTerminal reports whether w is an interactive terminal. Tests pass
+// bytes.Buffer writers and pipes/redirects fail term.IsTerminal, so both take
+// the machine-readable path.
+func stdoutIsTerminal(w io.Writer) bool {
+	f, ok := w.(*os.File)
+	if !ok {
+		return false
+	}
+	return term.IsTerminal(f.Fd())
+}
+
 func writeAppError(stderr io.Writer, message string, exitCode int) int {
 	if _, err := fmt.Fprintf(stderr, "[zero] %s\n", message); err != nil {
 		return 1
@@ -1134,6 +1174,7 @@ Usage:
 
 Commands:
   exec       Run a one-shot prompt through the Go agent runtime
+  completions Generate shell completion scripts
   daemon     Manage the local background worker daemon (start/stop/status/run/attach)
   setup      Guide first-run provider setup
   config     Inspect resolved Go configuration without leaking secrets
@@ -1318,6 +1359,9 @@ Flags:
       --spec-reasoning-effort <effort>
                                     Override draft reasoning effort when --use-spec is set
       --max-turns <number>           Override the maximum agent loop turns
+      --exec-profile <name>          Apply an execution profile (balanced, fast, thorough): loop
+                                    posture only (turn budget, effort, self-correction, escalation);
+                                    composes with --mode (which picks the model) and explicit flags win
       --auto <low|medium|high>       Set exec autonomy; high enables unsafe tools
       --enabled-tools <tools>        Only expose these comma or space separated tools
       --disabled-tools <tools>       Hide these comma or space separated tools
@@ -1347,6 +1391,8 @@ Flags:
       --notify <off|bell|notify|both>
                                     Override notification mode for this run
       --no-notify                   Disable notifications for this run
+      --no-completion-gate          Accept a no-tool-call reply as final without the INCOMPLETE
+                                    downgrade (for conversational callers with an operator present)
 `)
 	return err
 }

internal/cli/app_test.go

@@ -41,6 +41,8 @@ func TestRunPrintsVersion(t *testing.T) {
 	if exitCode != 0 {
 		t.Fatalf("expected exit code 0, got %d", exitCode)
 	}
+	// Non-terminal stdout keeps the machine-readable contract: exactly one
+	// "zero <version>" record, no wordmark banner, no ANSI.
 	if got := stdout.String(); got != "zero dev\n" {
 		t.Fatalf("expected version output, got %q", got)
 	}
@@ -49,6 +51,33 @@ func TestRunPrintsVersion(t *testing.T) {
 	}
 }
 
+// TestRunVersionRedirectedFile covers the `zero --version > file` path: stdout
+// is a real *os.File but not a terminal, so the single-line contract must hold.
+func TestRunVersionRedirectedFile(t *testing.T) {
+	stdout, err := os.CreateTemp(t.TempDir(), "version-stdout")
+	if err != nil {
+		t.Fatalf("create temp stdout: %v", err)
+	}
+	defer stdout.Close()
+	var stderr bytes.Buffer
+
+	exitCode := Run([]string{"--version"}, stdout, &stderr)
+
+	if exitCode != 0 {
+		t.Fatalf("expected exit code 0, got %d", exitCode)
+	}
+	content, err := os.ReadFile(stdout.Name())
+	if err != nil {
+		t.Fatalf("read temp stdout: %v", err)
+	}
+	if got := string(content); got != "zero dev\n" {
+		t.Fatalf("expected version output, got %q", got)
+	}
+	if stderr.Len() != 0 {
+		t.Fatalf("expected empty stderr, got %q", stderr.String())
+	}
+}
+
 func TestRunPrintsHelp(t *testing.T) {
 	for _, args := range [][]string{
 		{"--help"},

internal/cli/command_center_test.go

@@ -348,6 +348,83 @@ func TestRunProvidersAddWritesCatalogProfile(t *testing.T) {
 	}
 }
 
+func TestRunProvidersAddAimlapiWritesDefaultHeaders(t *testing.T) {
+	var stdout bytes.Buffer
+	var stderr bytes.Buffer
+	configPath := filepath.Join(t.TempDir(), "zero", "config.json")
+
+	exitCode := runWithDeps([]string{"providers", "add", "aimlapi"}, &stdout, &stderr, providerSetupDeps(configPath))
+
+	if exitCode != exitSuccess {
+		t.Fatalf("expected exit code %d, got %d: %s", exitSuccess, exitCode, stderr.String())
+	}
+	cfg := readFileConfig(t, configPath)
+	if len(cfg.Providers) != 1 {
+		t.Fatalf("providers = %#v, want one provider", cfg.Providers)
+	}
+	profile := cfg.Providers[0]
+	if profile.Name != "aimlapi" ||
+		profile.CatalogID != "aimlapi" ||
+		profile.BaseURL != "https://api.aimlapi.com/v1" ||
+		profile.Model != "anthropic/claude-sonnet-5" ||
+		profile.APIKeyEnv != "AIMLAPI_API_KEY" {
+		t.Fatalf("unexpected provider profile: %#v", profile)
+	}
+	if profile.CustomHeaders["X-AIMLAPI-Partner-ID"] != "part_62yQoGYDq4Yqnrj2R1iGrDNJ" ||
+		profile.CustomHeaders["X-AIMLAPI-Integration-Repo"] != "Gitlawb/zero" {
+		t.Fatalf("missing aimlapi.com default headers: %#v", profile.CustomHeaders)
+	}
+}
+
+func TestRunProvidersAddAimlapiMixedCaseHeaderOverride(t *testing.T) {
+	var stdout bytes.Buffer
+	var stderr bytes.Buffer
+	configPath := filepath.Join(t.TempDir(), "zero", "config.json")
+
+	// A differently-cased override must replace the catalog header in place, not
+	// leave both spellings behind to race when request construction canonicalizes.
+	exitCode := runWithDeps([]string{
+		"providers", "add", "aimlapi",
+		"--header", "x-aimlapi-partner-id=part_override",
+	}, &stdout, &stderr, providerSetupDeps(configPath))
+
+	if exitCode != exitSuccess {
+		t.Fatalf("expected exit code %d, got %d: %s", exitSuccess, exitCode, stderr.String())
+	}
+	profile := readFileConfig(t, configPath).Providers[0]
+	if _, ok := profile.CustomHeaders["x-aimlapi-partner-id"]; ok {
+		t.Fatalf("lowercase override left a duplicate key: %#v", profile.CustomHeaders)
+	}
+	if profile.CustomHeaders["X-AIMLAPI-Partner-ID"] != "part_override" {
+		t.Fatalf("override did not win on the canonical key: %#v", profile.CustomHeaders)
+	}
+}
+
+func TestRunProvidersAddAimlapiOverrideDropsCatalogHeaders(t *testing.T) {
+	var stdout bytes.Buffer
+	var stderr bytes.Buffer
+	configPath := filepath.Join(t.TempDir(), "zero", "config.json")
+
+	exitCode := runWithDeps([]string{
+		"providers", "add", "aimlapi",
+		"--base-url", "https://staging.example/v1",
+		"--header", "X-Environment=staging",
+	}, &stdout, &stderr, providerSetupDeps(configPath))
+	if exitCode != exitSuccess {
+		t.Fatalf("expected exit code %d, got %d: %s", exitSuccess, exitCode, stderr.String())
+	}
+	profile := readFileConfig(t, configPath).Providers[0]
+	if _, ok := profile.CustomHeaders["X-AIMLAPI-Partner-ID"]; ok {
+		t.Fatalf("partner attribution leaked to overridden endpoint: %#v", profile.CustomHeaders)
+	}
+	if _, ok := profile.CustomHeaders["X-AIMLAPI-Integration-Repo"]; ok {
+		t.Fatalf("integration attribution leaked to overridden endpoint: %#v", profile.CustomHeaders)
+	}
+	if profile.CustomHeaders["X-Environment"] != "staging" {
+		t.Fatalf("explicit user header was dropped: %#v", profile.CustomHeaders)
+	}
+}
+
 func TestRunProvidersAddWritesCustomHeaders(t *testing.T) {
 	var stdout bytes.Buffer
 	var stderr bytes.Buffer

internal/cli/completions.go

@@ -0,0 +1,404 @@
+package cli
+
+import (
+	"fmt"
+	"io"
+	"sort"
+	"strings"
+)
+
+type completionNode struct {
+	names    []string
+	flags    []string
+	children []completionNode
+}
+
+type completionContext struct {
+	path        string
+	candidates  []string
+	transitions map[string]string
+}
+
+var completionRoot = completionNode{
+	flags: []string{"-h", "--help", "-v", "--version", "-p", "--prompt", "--add-dir", "--theme", "--skip-permissions-unsafe"},
+	children: []completionNode{
+		{names: []string{"exec"}, flags: []string{
+			"-h", "--help", "-f", "--file", "--image", "--add-dir", "--mode", "-m", "--model",
+			"--use-spec", "--spec-model", "--spec-reasoning-effort", "--max-turns", "--exec-profile",
+			"--auto", "--enabled-tools", "--disabled-tools", "--list-tools", "--profile", "-r",
+			"--reasoning-effort", "-C", "--cwd", "-w", "--worktree", "--worktree-dir", "-i",
+			"--input-format", "-o", "--output-format", "--prompt", "--resume", "--fork",
+			"--calling-session-id", "--calling-tool-use-id", "--tag", "--depth", "--session-title",
+			"--init-session-id", "--skip-permissions-unsafe", "--allow-escalation", "--self-correct",
+			"--notify", "--no-notify", "--no-completion-gate",
+		}},
+		{
+			names:    []string{"completions"},
+			flags:    []string{"-h", "--help"},
+			children: leafNodes("bash", "zsh", "fish", "powershell", "elvish"),
+		},
+		{names: []string{"daemon"}, children: leafNodes("start", "stop", "status", "run", "attach", "serve-remote", "link")},
+		{names: []string{"setup"}},
+		{names: []string{"config"}},
+		{names: []string{"models"}, children: []completionNode{{names: []string{"list", "ls"}}}},
+		{names: []string{"providers"}, children: []completionNode{
+			{names: []string{"current"}}, {names: []string{"list"}}, {names: []string{"catalog"}},
+			{names: []string{"add"}}, {names: []string{"check"}}, {names: []string{"use"}},
+			{names: []string{"remove", "rm"}}, {names: []string{"rename"}}, {names: []string{"setup"}},
+			{names: []string{"detect"}}, {names: []string{"models"}},
+		}},
+		{names: []string{"doctor"}},
+		{names: []string{"context"}},
+		{names: []string{"repo-map", "repomap"}},
+		{names: []string{"search", "find"}},
+		{names: []string{"sessions", "session"}, children: leafNodes("list", "children", "lineage", "tree", "rewind-plan", "rewind", "compact-plan")},
+		{names: []string{"spec"}, children: leafNodes("show", "approve", "reject")},
+		{names: []string{"init"}},
+		{names: []string{"specialists", "specialist"}, children: []completionNode{
+			{names: []string{"list"}}, {names: []string{"show"}}, {names: []string{"create"}},
+			{names: []string{"delete", "rm"}}, {names: []string{"edit"}}, {names: []string{"path"}},
+		}},
+		{names: []string{"plugins", "plugin"}, children: []completionNode{
+			{names: []string{"list"}}, {names: []string{"add"}}, {names: []string{"info"}}, {names: []string{"remove", "rm"}},
+		}},
+		{names: []string{"backends", "backend"}, children: leafNodes("doctor")},
+		{names: []string{"skills", "skill"}, children: []completionNode{
+			{names: []string{"list"}}, {names: []string{"add"}}, {names: []string{"info"}}, {names: []string{"remove", "rm"}},
+		}},
+		{names: []string{"tools", "tool"}, children: leafNodes("make", "list")},
+		{names: []string{"hooks"}, children: []completionNode{
+			{names: []string{"list"}}, {names: []string{"add"}}, {names: []string{"remove", "rm"}},
+			{names: []string{"enable"}}, {names: []string{"disable"}},
+		}},
+		{names: []string{"mcp"}, children: []completionNode{
+			{names: []string{"add"}}, {names: []string{"remove", "rm"}}, {names: []string{"enable"}},
+			{names: []string{"disable"}}, {names: []string{"check"}}, {names: []string{"list"}},
+			{names: []string{"permissions"}, children: leafNodes("list", "revoke", "clear")},
+			{names: []string{"tools"}, children: leafNodes("list")},
+			{names: []string{"oauth"}, children: leafNodes("login", "logout", "status")},
+		}},
+		{names: []string{"auth"}, children: leafNodes("openrouter", "chatgpt", "login", "logout", "status", "refresh")},
+		{names: []string{"sandbox"}, children: []completionNode{
+			{names: []string{"policy"}}, {names: []string{"setup"}}, {names: []string{"check"}},
+			{names: []string{"grants"}, children: leafNodes("list", "allow", "deny", "revoke", "clear")},
+		}},
+		{names: []string{"update"}},
+		{names: []string{"upgrade"}},
+		{names: []string{"worktrees", "worktree"}, children: leafNodes("prepare")},
+		{names: []string{"verify"}},
+		{names: []string{"trust"}, children: leafNodes("list", "remove")},
+		{names: []string{"eval"}, children: leafNodes("validate", "run", "bench")},
+		{names: []string{"changes", "change"}, children: []completionNode{
+			{names: []string{"inspect", "status"}}, {names: []string{"commit"}}, {names: []string{"push"}},
+			{names: []string{"pr", "pull-request"}},
+		}},
+		{names: []string{"usage"}, children: leafNodes("report")},
+		{names: []string{"cron"}, children: leafNodes("add", "list", "rm", "pause", "resume", "run")},
+		{names: []string{"repo-info", "repoinfo"}},
+		{names: []string{"serve"}},
+		{names: []string{"acp"}},
+		{names: []string{"help"}},
+		{names: []string{"version"}},
+	},
+}
+
+func leafNodes(names ...string) []completionNode {
+	nodes := make([]completionNode, 0, len(names))
+	for _, name := range names {
+		nodes = append(nodes, completionNode{names: []string{name}})
+	}
+	return nodes
+}
+
+func runCompletions(args []string, stdout io.Writer, stderr io.Writer) int {
+	if len(args) == 1 && (args[0] == "-h" || args[0] == "--help" || args[0] == "help") {
+		if err := writeCompletionsHelp(stdout); err != nil {
+			return exitCrash
+		}
+		return exitSuccess
+	}
+	if len(args) == 0 {
+		return writeExecUsageError(stderr, "shell required. Use `zero completions <bash|zsh|fish|powershell|elvish>`.")
+	}
+	if len(args) != 1 {
+		return writeExecUsageError(stderr, fmt.Sprintf("unexpected completions argument %q", args[1]))
+	}
+
+	contexts := completionContexts(completionRoot)
+	var err error
+	switch strings.ToLower(strings.TrimSpace(args[0])) {
+	case "bash":
+		err = writeBashCompletions(stdout, contexts)
+	case "zsh":
+		err = writeZshCompletions(stdout, contexts)
+	case "fish":
+		err = writeFishCompletions(stdout, contexts)
+	case "powershell":
+		err = writePowerShellCompletions(stdout, contexts)
+	case "elvish":
+		err = writeElvishCompletions(stdout, contexts)
+	default:
+		return writeExecUsageError(stderr, fmt.Sprintf("unsupported shell %q; expected bash, zsh, fish, powershell, or elvish", args[0]))
+	}
+	if err != nil {
+		return exitCrash
+	}
+	return exitSuccess
+}
+
+func writeCompletionsHelp(w io.Writer) error {
+	_, err := fmt.Fprint(w, `Generate shell completion scripts.
+
+Usage:
+  zero completions <shell>
+
+Arguments:
+  <shell>  Target shell: bash, zsh, fish, powershell, or elvish
+
+Examples:
+  source <(zero completions bash)
+  source <(zero completions zsh)
+  zero completions fish > ~/.config/fish/completions/zero.fish
+  zero completions powershell >> $PROFILE
+  eval (zero completions elvish | slurp)
+
+Flags:
+  -h, --help  Show this help
+`)
+	return err
+}
+
+func completionContexts(root completionNode) []completionContext {
+	contexts := []completionContext{}
+	var visit func(completionNode, []string)
+	visit = func(node completionNode, paths []string) {
+		candidates := append([]string{}, node.flags...)
+		transitions := map[string]string{}
+		for _, child := range node.children {
+			candidates = append(candidates, child.names...)
+			for _, path := range paths {
+				for _, name := range child.names {
+					childPath := strings.TrimSpace(path + " " + name)
+					transitions[path+"|"+name] = childPath
+				}
+			}
+		}
+		candidates = uniqueStrings(candidates)
+		for _, path := range paths {
+			contexts = append(contexts, completionContext{path: path, candidates: candidates, transitions: transitions})
+		}
+		for _, child := range node.children {
+			childPaths := make([]string, 0, len(paths)*len(child.names))
+			for _, path := range paths {
+				for _, name := range child.names {
+					childPaths = append(childPaths, strings.TrimSpace(path+" "+name))
+				}
+			}
+			visit(child, childPaths)
+		}
+	}
+	visit(root, []string{""})
+	sort.SliceStable(contexts, func(i, j int) bool { return contexts[i].path < contexts[j].path })
+	return contexts
+}
+
+func uniqueStrings(values []string) []string {
+	seen := map[string]bool{}
+	result := make([]string, 0, len(values))
+	for _, value := range values {
+		if value == "" || seen[value] {
+			continue
+		}
+		seen[value] = true
+		result = append(result, value)
+	}
+	return result
+}
+
+func allTransitions(contexts []completionContext) map[string]string {
+	transitions := map[string]string{}
+	for _, context := range contexts {
+		for key, value := range context.transitions {
+			transitions[key] = value
+		}
+	}
+	return transitions
+}
+
+func sortedTransitionKeys(transitions map[string]string) []string {
+	keys := make([]string, 0, len(transitions))
+	for key := range transitions {
+		keys = append(keys, key)
+	}
+	sort.Strings(keys)
+	return keys
+}
+
+func shellWords(values []string) string {
+	return strings.Join(values, " ")
+}
+
+func writeBashCompletions(w io.Writer, contexts []completionContext) error {
+	if _, err := fmt.Fprintln(w, "# bash completion for zero\n_zero() {\n  local cur context token i candidates\n  cur=\"${COMP_WORDS[COMP_CWORD]}\"\n  context=\"\""); err != nil {
+		return err
+	}
+	if err := writePOSIXContextLoop(w, contexts, "  for ((i = 1; i < COMP_CWORD; i++)); do", "    token=\"${COMP_WORDS[i]}\"", "  done"); err != nil {
+		return err
+	}
+	if err := writePOSIXCandidateCase(w, contexts, "  "); err != nil {
+		return err
+	}
+	_, err := fmt.Fprintln(w, "  COMPREPLY=( $(compgen -W \"$candidates\" -- \"$cur\") )\n}\ncomplete -F _zero zero")
+	return err
+}
+
+func writeZshCompletions(w io.Writer, contexts []completionContext) error {
+	if _, err := fmt.Fprintln(w, "#compdef zero\n_zero() {\n  local context token i\n  local -a candidates\n  context=\"\""); err != nil {
+		return err
+	}
+	if err := writePOSIXContextLoop(w, contexts, "  for ((i = 2; i < CURRENT; i++)); do", "    token=\"${words[i]}\"", "  done"); err != nil {
+		return err
+	}
+	if _, err := fmt.Fprintln(w, "  case \"$context\" in"); err != nil {
+		return err
+	}
+	for _, context := range contexts {
+		if _, err := fmt.Fprintf(w, "    %q) candidates=(%s) ;;\n", context.path, quotedWords(context.candidates)); err != nil {
+			return err
+		}
+	}
+	_, err := fmt.Fprintln(w, "  esac\n  compadd -- \"${candidates[@]}\"\n}\ncompdef _zero zero")
+	return err
+}
+
+func writePOSIXContextLoop(w io.Writer, contexts []completionContext, start string, token string, end string) error {
+	if _, err := fmt.Fprintln(w, start); err != nil {
+		return err
+	}
+	if _, err := fmt.Fprintln(w, token); err != nil {
+		return err
+	}
+	if _, err := fmt.Fprintln(w, "    case \"$context|$token\" in"); err != nil {
+		return err
+	}
+	transitions := allTransitions(contexts)
+	for _, key := range sortedTransitionKeys(transitions) {
+		if _, err := fmt.Fprintf(w, "      %q) context=%q ;;\n", key, transitions[key]); err != nil {
+			return err
+		}
+	}
+	if _, err := fmt.Fprintln(w, "    esac"); err != nil {
+		return err
+	}
+	_, err := fmt.Fprintln(w, end)
+	return err
+}
+
+func writePOSIXCandidateCase(w io.Writer, contexts []completionContext, indent string) error {
+	if _, err := fmt.Fprintln(w, indent+"case \"$context\" in"); err != nil {
+		return err
+	}
+	for _, context := range contexts {
+		if _, err := fmt.Fprintf(w, "%s  %q) candidates=%q ;;\n", indent, context.path, shellWords(context.candidates)); err != nil {
+			return err
+		}
+	}
+	_, err := fmt.Fprintln(w, indent+"esac")
+	return err
+}
+
+func quotedWords(values []string) string {
+	quoted := make([]string, 0, len(values))
+	for _, value := range values {
+		quoted = append(quoted, fmt.Sprintf("%q", value))
+	}
+	return strings.Join(quoted, " ")
+}
+
+func writeFishCompletions(w io.Writer, contexts []completionContext) error {
+	if _, err := fmt.Fprintln(w, "# fish completion for zero\nfunction __zero_completion_context\n    set -l context ''\n    set -l tokens (commandline -opc)"); err != nil {
+		return err
+	}
+	if _, err := fmt.Fprintln(w, "    for token in $tokens[2..-1]\n        switch \"$context|$token\""); err != nil {
+		return err
+	}
+	transitions := allTransitions(contexts)
+	for _, key := range sortedTransitionKeys(transitions) {
+		if _, err := fmt.Fprintf(w, "            case %q\n                set context %q\n", key, transitions[key]); err != nil {
+			return err
+		}
+	}
+	if _, err := fmt.Fprintln(w, "        end\n    end\n    if test -z \"$context\"\n        echo __root__\n    else\n        echo \"$context\"\n    end\nend"); err != nil {
+		return err
+	}
+	for _, context := range contexts {
+		path := context.path
+		if path == "" {
+			path = "__root__"
+		}
+		if _, err := fmt.Fprintf(w, "complete -c zero -f -n 'test (__zero_completion_context) = %q' -a %q\n", path, shellWords(context.candidates)); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+func writePowerShellCompletions(w io.Writer, contexts []completionContext) error {
+	if _, err := fmt.Fprintln(w, "# PowerShell completion for zero\nRegister-ArgumentCompleter -Native -CommandName zero -ScriptBlock {\n    param($wordToComplete, $commandAst, $cursorPosition)\n    $context = ''\n    $elements = @($commandAst.CommandElements)\n    $limit = $elements.Count\n    if ($wordToComplete -ne '' -and $limit -gt 1) { $limit-- }\n    for ($i = 1; $i -lt $limit; $i++) {\n        $token = $elements[$i].Extent.Text\n        switch (\"$context|$token\") {"); err != nil {
+		return err
+	}
+	transitions := allTransitions(contexts)
+	for _, key := range sortedTransitionKeys(transitions) {
+		if _, err := fmt.Fprintf(w, "            '%s' { $context = '%s'; break }\n", key, transitions[key]); err != nil {
+			return err
+		}
+	}
+	if _, err := fmt.Fprintln(w, "        }\n    }\n    $candidates = switch ($context) {"); err != nil {
+		return err
+	}
+	for _, context := range contexts {
+		if _, err := fmt.Fprintf(w, "        '%s' { @(%s) }\n", context.path, powershellWords(context.candidates)); err != nil {
+			return err
+		}
+	}
+	_, err := fmt.Fprintln(w, "    }\n    $candidates | Where-Object { $_ -like \"$wordToComplete*\" } | ForEach-Object {\n        [System.Management.Automation.CompletionResult]::new($_, $_, 'ParameterValue', $_)\n    }\n}")
+	return err
+}
+
+func powershellWords(values []string) string {
+	quoted := make([]string, 0, len(values))
+	for _, value := range values {
+		quoted = append(quoted, "'"+strings.ReplaceAll(value, "'", "''")+"'")
+	}
+	return strings.Join(quoted, ", ")
+}
+
+func writeElvishCompletions(w io.Writer, contexts []completionContext) error {
+	if _, err := fmt.Fprintln(w, "# Elvish completion for zero\nset edit:completion:arg-completer[zero] = {|@args|\n    var context = ''\n    for token $args[1..-1] {"); err != nil {
+		return err
+	}
+	transitions := allTransitions(contexts)
+	for index, key := range sortedTransitionKeys(transitions) {
+		keyword := "if"
+		if index > 0 {
+			keyword = "} elif"
+		}
+		if _, err := fmt.Fprintf(w, "        %s (eq $context'|'$token %q) {\n            set context = %q\n", keyword, key, transitions[key]); err != nil {
+			return err
+		}
+	}
+	if _, err := fmt.Fprintln(w, "        }\n    }"); err != nil {
+		return err
+	}
+	for index, context := range contexts {
+		keyword := "if"
+		if index > 0 {
+			keyword = "} elif"
+		}
+		if _, err := fmt.Fprintf(w, "    %s (eq $context %q) {\n        put %s\n", keyword, context.path, shellWords(context.candidates)); err != nil {
+			return err
+		}
+	}
+	_, err := fmt.Fprintln(w, "    }\n}")
+	return err
+}

internal/cli/completions_test.go

@@ -0,0 +1,177 @@
+package cli
+
+import (
+	"bytes"
+	"os/exec"
+	"strings"
+	"testing"
+)
+
+func TestCompletionsHelpAndRootHelp(t *testing.T) {
+	var stdout bytes.Buffer
+	var stderr bytes.Buffer
+	if code := runWithDeps([]string{"completions", "--help"}, &stdout, &stderr, appDeps{}); code != exitSuccess {
+		t.Fatalf("completion help exit code = %d, want %d: %s", code, exitSuccess, stderr.String())
+	}
+	for _, want := range []string{
+		"zero completions <shell>",
+		"bash, zsh, fish, powershell, or elvish",
+		"source <(zero completions bash)",
+		"~/.config/fish/completions/zero.fish",
+		"eval (zero completions elvish | slurp)",
+	} {
+		if !strings.Contains(stdout.String(), want) {
+			t.Errorf("completion help missing %q:\n%s", want, stdout.String())
+		}
+	}
+
+	stdout.Reset()
+	stderr.Reset()
+	if code := runWithDeps([]string{"--help"}, &stdout, &stderr, appDeps{}); code != exitSuccess {
+		t.Fatalf("root help exit code = %d, want %d: %s", code, exitSuccess, stderr.String())
+	}
+	if !strings.Contains(stdout.String(), "completions Generate shell completion scripts") {
+		t.Fatalf("root help does not list completions command:\n%s", stdout.String())
+	}
+}
+
+func TestCompletionsRejectsMissingUnknownAndExtraShellArguments(t *testing.T) {
+	tests := []struct {
+		name string
+		args []string
+		want string
+	}{
+		{name: "missing", args: []string{"completions"}, want: "shell required"},
+		{name: "unknown", args: []string{"completions", "nu"}, want: "unsupported shell"},
+		{name: "extra", args: []string{"completions", "bash", "extra"}, want: "unexpected completions argument"},
+	}
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			var stdout bytes.Buffer
+			var stderr bytes.Buffer
+			if code := runWithDeps(test.args, &stdout, &stderr, appDeps{}); code != exitUsage {
+				t.Fatalf("exit code = %d, want %d: %s", code, exitUsage, stderr.String())
+			}
+			if stdout.Len() != 0 {
+				t.Fatalf("stdout = %q, want empty", stdout.String())
+			}
+			if !strings.Contains(stderr.String(), test.want) {
+				t.Fatalf("stderr = %q, want substring %q", stderr.String(), test.want)
+			}
+		})
+	}
+}
+
+func TestCompletionsGeneratesEverySupportedShell(t *testing.T) {
+	tests := []struct {
+		shell       string
+		marker      string
+		syntaxShell string
+	}{
+		{shell: "bash", marker: "complete -F _zero zero", syntaxShell: "bash"},
+		{shell: "zsh", marker: "#compdef zero", syntaxShell: "zsh"},
+		{shell: "fish", marker: "complete -c zero"},
+		{shell: "powershell", marker: "Register-ArgumentCompleter -Native -CommandName zero"},
+		{shell: "elvish", marker: "edit:completion:arg-completer[zero]"},
+	}
+	for _, test := range tests {
+		t.Run(test.shell, func(t *testing.T) {
+			var stdout bytes.Buffer
+			var stderr bytes.Buffer
+			if code := runWithDeps([]string{"completions", test.shell}, &stdout, &stderr, appDeps{}); code != exitSuccess {
+				t.Fatalf("exit code = %d, want %d: %s", code, exitSuccess, stderr.String())
+			}
+			if !strings.Contains(stdout.String(), test.marker) {
+				t.Fatalf("%s completion missing marker %q:\n%s", test.shell, test.marker, stdout.String())
+			}
+			for _, want := range []string{"daemon", "mcp oauth", "sandbox grants", "--output-format"} {
+				if !strings.Contains(stdout.String(), want) {
+					t.Errorf("%s completion missing %q", test.shell, want)
+				}
+			}
+			script := stdout.String()
+			switch test.shell {
+			case "fish":
+				assertBalancedFishBlocks(t, script)
+			case "powershell", "elvish":
+				assertBalancedBraces(t, script)
+			}
+			if test.syntaxShell != "" {
+				assertNativeShellSyntax(t, test.syntaxShell, script)
+			}
+		})
+	}
+}
+
+func assertBalancedBraces(t *testing.T, script string) {
+	t.Helper()
+	if opens, closes := strings.Count(script, "{"), strings.Count(script, "}"); opens != closes {
+		t.Errorf("unbalanced braces: %d opening, %d closing", opens, closes)
+	}
+}
+
+func assertBalancedFishBlocks(t *testing.T, script string) {
+	t.Helper()
+	depth := 0
+	for _, line := range strings.Split(script, "\n") {
+		fields := strings.Fields(line)
+		if len(fields) == 0 {
+			continue
+		}
+		switch fields[0] {
+		case "function", "for", "switch", "if":
+			depth++
+		case "end":
+			depth--
+			if depth < 0 {
+				t.Fatal("fish completion closes a block before one is opened")
+			}
+		}
+	}
+	if depth != 0 {
+		t.Errorf("fish completion has %d unclosed blocks", depth)
+	}
+}
+
+func assertNativeShellSyntax(t *testing.T, shell, script string) {
+	t.Helper()
+	path, err := exec.LookPath(shell)
+	if err != nil {
+		t.Skipf("%s is not installed", shell)
+	}
+	command := exec.Command(path, "-n")
+	command.Stdin = strings.NewReader(script)
+	if output, err := command.CombinedOutput(); err != nil {
+		t.Fatalf("%s syntax check failed: %v\n%s", shell, err, output)
+	}
+}
+
+func TestCompletionTreeCoversAliasesNestingAndCommonFlags(t *testing.T) {
+	contexts := completionContexts(completionRoot)
+	byPath := make(map[string][]string, len(contexts))
+	for _, context := range contexts {
+		byPath[context.path] = context.candidates
+	}
+
+	assertCandidates(t, byPath[""], "sessions", "session", "plugins", "plugin", "worktrees", "worktree", "--add-dir", "--theme", "-p", "--prompt")
+	assertCandidates(t, byPath["exec"], "--model", "--cwd", "--worktree", "--output-format", "--resume", "--skip-permissions-unsafe")
+	assertCandidates(t, byPath["daemon"], "start", "stop", "status", "run", "attach")
+	assertCandidates(t, byPath["mcp oauth"], "login", "logout", "status")
+	assertCandidates(t, byPath["sandbox grants"], "list", "allow", "deny", "revoke", "clear")
+	assertCandidates(t, byPath["completions"], "bash", "zsh", "fish", "powershell", "elvish")
+	assertCandidates(t, byPath["plugins"], "list", "add", "info", "remove", "rm")
+	assertCandidates(t, byPath["plugin"], "list", "add", "info", "remove", "rm")
+}
+
+func assertCandidates(t *testing.T, got []string, wants ...string) {
+	t.Helper()
+	set := make(map[string]bool, len(got))
+	for _, candidate := range got {
+		set[candidate] = true
+	}
+	for _, want := range wants {
+		if !set[want] {
+			t.Errorf("candidates %v do not contain %q", got, want)
+		}
+	}
+}

internal/cli/cron_run.go

@@ -211,11 +211,12 @@ func fireJob(store *cron.Store, now func() time.Time, job cron.Job, stdout io.Wr
 			fmt.Fprintf(stderr, "warning: could not re-read job %s before persist: %v\n", job.ID, readErr)
 			return job, nil
 		}
-		if current.Status == cron.StatusPaused {
-			// Honor an external pause that landed while the job ran.
-			job.Status = cron.StatusPaused
+		current.FireCount = job.FireCount
+		current.NextRunAt = job.NextRunAt
+		if current.Status != cron.StatusPaused {
+			current.Status = job.Status
 		}
-		return job, nil
+		return current, nil
 	})
 	if errors.Is(err, cron.ErrJobNotFound) {
 		// Genuinely removed mid-run: don't recreate it (no run record, no persist).