test(sandbox): give the cross-platform anchor tests native paths

Moving these beside the other BuildWindowsACLPlan tests made them run on
Linux and macOS, where their Windows path literals stop meaning what they
say: pathWithinRoot is filepath.Rel underneath, a backslash is an ordinary
character off Windows, and C:\workspace\.git\hooks is then one component
that is not under C:\workspace, so every anchor came back empty. The
neighbouring tests get away with such literals because they only compare
strings they built the same way and never ask whether one contains another.

Paths are built with filepath.Join from temp roots now, so the containment
question is asked in the separator the running platform actually uses.

.coderabbit.yaml

@@ -29,6 +29,9 @@ reviews:
     - path: "tests/**/*.ts"
       instructions: |
         Review tests for Bun test idioms, deterministic behavior, and meaningful coverage of safety decisions, provider/runtime flows, session events, and TUI command behavior.
+    - path: "**/*.go"
+      instructions: |
+        Do not request Plan 9 or other unsupported GOOS/GOARCH shims. The five native prebuilt release targets are linux-x64, linux-arm64, macos-arm64, macos-x64, and windows-x64. Android/Termux is also a supported runtime through the documented android-arm64 source build and Linux artifact fallback. Other GOOS/GOARCH combinations are not a review criterion.
   tools:
     github-checks:
       timeout_ms: 300000

.github/workflows/ci.yml

@@ -6,10 +6,109 @@ on:
     branches:
       - main
 
+concurrency:
+  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref }}
+  cancel-in-progress: ${{ github.event_name == 'pull_request' }}
+
+permissions:
+  contents: read
+
 jobs:
+  quality:
+    name: Code Quality & Lint
+    runs-on: ubuntu-latest
+    timeout-minutes: 10
+    permissions:
+      contents: read
+
+    steps:
+      - name: Checkout
+        uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5 # v4.2.2
+        with:
+          persist-credentials: false
+
+      - name: Setup Go
+        uses: actions/setup-go@40f1582b2485089dde7abd97c1529aa768e1baff # v5.3.0
+        with:
+          go-version-file: go.mod
+          cache: true
+
+      - name: Check formatting
+        run: make fmt-check
+
+      - name: Vet
+        run: go vet ./...
+
+      - name: deadcode (advisory)
+        continue-on-error: true
+        run: make deadcode
+
+      - name: golangci-lint (advisory)
+        continue-on-error: true
+        run: make lint-static
+
+  test:
+    name: Unit Tests & Coverage
+    runs-on: ubuntu-latest
+    timeout-minutes: 15
+    permissions:
+      contents: read
+
+    steps:
+      - name: Checkout
+        uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5 # v4.2.2
+        with:
+          persist-credentials: false
+
+      - name: Setup Go
+        uses: actions/setup-go@40f1582b2485089dde7abd97c1529aa768e1baff # v5.3.0
+        with:
+          go-version-file: go.mod
+          cache: true
+
+      - name: Test with coverage
+        run: |
+          go test -count=1 -coverprofile=coverage.out -covermode=atomic ./...
+
+      - name: Publish coverage summary
+        if: always()
+        shell: bash
+        run: |
+          if [ -f coverage.out ]; then
+            go tool cover -func=coverage.out | awk '
+            BEGIN {
+              total = "unknown"
+              rows = ""
+            }
+            $1 == "total:" {
+              total = $3
+              next
+            }
+            {
+              rows = rows sprintf("| `%s` | `%s` | **%s** |\n", $1, $2, $3)
+            }
+            END {
+              print "### 📊 Test Coverage Summary: **" total "**"
+              print "<details><summary>Coverage per function</summary>\n"
+              print "| File | Function | Coverage |"
+              print "| :--- | :--- | :--- |"
+              printf "%s", rows
+              print "</details>"
+            }' >> "$GITHUB_STEP_SUMMARY"
+          fi
+
+      - name: Upload coverage artifact
+        if: always()
+        uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02 # v4.6.2
+        with:
+          name: zero-coverage-report
+          path: coverage.out
+          if-no-files-found: warn
+
   smoke:
     name: Smoke (${{ matrix.os }})
     runs-on: ${{ matrix.os }}
+    timeout-minutes: 15
     strategy:
       fail-fast: false
       matrix:
@@ -18,64 +117,22 @@ jobs:
           - macos-latest
           - windows-latest
 
+    permissions:
+      contents: read
+
     steps:
       - name: Checkout
-        uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5 # v4
+        uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5 # v4.2.2
         with:
           persist-credentials: false
 
       - name: Setup Go
-        uses: actions/setup-go@40f1582b2485089dde7abd97c1529aa768e1baff
+        uses: actions/setup-go@40f1582b2485089dde7abd97c1529aa768e1baff # v5.3.0
         with:
           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
-        run: |
-          unformatted="$(gofmt -l .)"
-          if [ -n "$unformatted" ]; then
-            echo "gofmt needed on:" >&2
-            echo "$unformatted" >&2
-            exit 1
-          fi
-
-      - name: Vet
-        if: matrix.os == 'ubuntu-latest'
-        run: go vet ./...
-
-      - name: Test
+      - name: Test (Fast Non-Race)
         run: go test ./...
 
       - name: Build binary
@@ -84,9 +141,15 @@ jobs:
       - name: Smoke binary
         run: go run ./cmd/zero-release smoke
 
-  performance:
-    name: Performance Smoke
+  race:
+    name: Race Detector
     runs-on: ubuntu-latest
+    permissions:
+      contents: read
+    # A race the detector newly reports can present as a deadlock rather than a
+    # failure, and the default job timeout is six hours. This job is expected to
+    # be the longest in the workflow, so the ceiling is generous but finite.
+    timeout-minutes: 30
 
     steps:
       - name: Checkout
@@ -100,55 +163,70 @@ jobs:
           go-version-file: go.mod
           cache: true
 
+      # Invoke the Makefile target rather than repeating its flags. `make test`
+      # is already `go test ./... -race -count=1`, so running it here is what
+      # stops the project's declared test command and what CI actually runs
+      # from drifting apart again. The race detector needs cgo, which the
+      # hosted ubuntu runner supplies without an extra setup step.
+      - name: Test with the race detector
+        run: make test
+
+  performance:
+    name: Performance Smoke
+    runs-on: ubuntu-latest
+    timeout-minutes: 10
+    permissions:
+      contents: read
+
+    steps:
+      - name: Checkout
+        uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5 # v4.2.2
+        with:
+          persist-credentials: false
+
+      - name: Setup Go
+        uses: actions/setup-go@40f1582b2485089dde7abd97c1529aa768e1baff # v5.3.0
+        with:
+          go-version-file: go.mod
+          cache: true
+
       - name: Build binary
         run: go run ./cmd/zero-release build
 
       - name: Performance smoke
-        run: go run ./cmd/zero-perf-bench --output dist/perf/perf-bench.json --ci
+        run: |
+          mkdir -p dist/perf
+          go run ./cmd/zero-perf-bench --output dist/perf/perf-bench.json --ci
 
       - name: Upload performance report
         if: always()
-        uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02 # v4
+        uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02 # v4.6.2
         with:
           name: zero-performance-smoke
           path: dist/perf/perf-bench.json
           if-no-files-found: warn
 
   security:
-    name: Security & code health
+    name: Security & Code Health
     runs-on: ubuntu-latest
+    timeout-minutes: 10
+    permissions:
+      contents: read
 
     steps:
       - name: Checkout
-        uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5 # v4
+        uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5 # v4.2.2
         with:
           persist-credentials: false
 
       - name: Setup Go
-        uses: actions/setup-go@40f1582b2485089dde7abd97c1529aa768e1baff
+        uses: actions/setup-go@40f1582b2485089dde7abd97c1529aa768e1baff # v5.3.0
         with:
           go-version-file: go.mod
           cache: true
 
-      # 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
+      - name: govulncheck (linux)
         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: make deadcode
-
-      # Advisory: catches what deadcode's whole-program reachability analysis
-      # doesn't, unused private functions/assignments reachable within a
-      # package but never actually called, plus staticcheck-style correctness
-      # and readability issues. Scoped to a few linters rather than the full
-      # default battery, and non-blocking, while the existing findings across
-      # the repo are cleaned up incrementally (see #527).
-      - name: golangci-lint (advisory)
-        continue-on-error: true
-        run: make lint-static
+      - name: govulncheck (windows)
+        run: make vulncheck-windows

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

@@ -14,6 +14,9 @@ on:
       - docs/GITHUB_ACTION.md
       - .github/workflows/zero-action-smoke.yml
 
+permissions:
+  contents: read
+
 jobs:
   validate:
     name: Validate action.yml

.release-please-manifest.json

@@ -1 +1 @@
-{".":"0.7.0"}
+{".":"0.8.0"}

AGENTS.md

@@ -81,6 +81,13 @@ These classes drive multi-round reviews. Fix them before requesting review:
   assert raw `t.TempDir()` spellings (`/var` vs `/private/var` on macOS, short
   paths on Windows). Prefer one cross-platform function with small `GOOS`
   checks over duplicated helpers that drift.
+- **Supported targets only:** CONTRIBUTING.md names the five native prebuilt
+  release artifacts (`linux-x64`, `linux-arm64`, `macos-arm64`, `macos-x64`,
+  `windows-x64`) and the additional supported Android/Termux source-build and
+  artifact-fallback path. Other `GOOS`/`GOARCH` combinations, including
+  `plan9`, are unsupported and are not a review criterion. Do not block a PR
+  for failing to compile on an unsupported `GOOS`/`GOARCH`. Tests may use
+  `plan9` as a fixture for "unsupported".
 - **Security edges:** Keep secrets out of argv, env dumps, and logs. Redact
   success and error paths (including stderr). Fail closed on ownership, lease,
   and permission checks. Do not rely on pre-open path resolution
@@ -91,6 +98,14 @@ These classes drive multi-round reviews. Fix them before requesting review:
   final-component-only no-follow is insufficient. On multi-step setup, roll back
   only what this run created; never destroy pre-existing resources you did
   not create; never report success when cleanup or unlock failed.
+- **Normalize before matching:** Any transform that REMOVES bytes without
+  leaving a gap is also a reassembler, so it has to run before whatever matches
+  on the result. Redacting by shape and then stripping control bytes lets a
+  credential split by a NUL, an ESC or a C1 byte pass the patterns as two
+  fragments and be rejoined on the way out; comparing a path against a form
+  produced by the same resolver the kernel just used makes a redirect agree with
+  itself. Strip, decode, or canonicalize first, then match. The unsplit value
+  passing is not evidence, so the test needs a split case.
 - **Atomic shared state:** Write a complete temporary file, then atomically
   replace the destination so concurrent readers never see a partial write.
   Exclusive create or a write lock alone is not enough for readers. Serialize
@@ -103,3 +118,42 @@ These classes drive multi-round reviews. Fix them before requesting review:
   shipped. Wire advertised entry points or shrink the claim. Do not bundle
   unrelated fixes. Do not widen security allowlists by name alone; check
   classification and side effects too.
+- **Host tool version floors:** Zero shells out to `git` and other host tools,
+  and the project does not declare a minimum version for any of them. A flag,
+  subcommand, or config key that is not long-established therefore raises that
+  floor silently, and the failure on an older host is an unhelpful usage error
+  rather than a clear message. Before adding one, check when it landed and say
+  so where a caller would look. Record the floor for the exact subcommand you
+  call, not for the flag in general: `--end-of-options` reached Git's shared
+  parser in 2.24, but `git rev-parse` only accepted it from 2.30, and
+  `git checkout` and `git reset` later still. Likewise `--show-current` is
+  `git branch` 2.22, and `branch.autoSetupMerge=inherit` is 2.35. Then either
+  keep a fallback for older versions or gate the call. A test that depends on a
+  newer feature must skip below its floor instead of failing, and say which
+  version it needs.
+- **Prove the test fails without the change:** A regression test earns its
+  place only if it fails on the unfixed code, for the reason it claims. Run it
+  both ways and quote the failure in the PR. The recurring miss is not a
+  missing test but a test that never reaches the behavior it names, because an
+  earlier guard rejects the input first, so it passes with and without the fix.
+  When a test targets a specific layer, call that layer directly rather than
+  relying on a public entry point that may refuse earlier.
+- **Hermetic tests:** Tests must not read or write the developer's real config,
+  cache, or state directories. Redirect every root the code under test
+  resolves, not just config. `os.UserConfigDir` and `os.UserCacheDir` read
+  `XDG_CONFIG_HOME` and `XDG_CACHE_HOME` only on Linux and the BSDs. On macOS
+  they derive from `HOME`, and on Windows from `%AppData%` and `%LocalAppData%`.
+  Setting the XDG variables alone still leaves a macOS run writing to the real
+  home, so set `HOME` as well, or take the directory as a parameter. Apply it to
+  every test that reaches the same storage, not only the one where it was first
+  needed; a test that passes on a clean machine and pollutes a real one is a
+  blocker.
+- **Error codes vary by platform and by flag combination:** A check written
+  against the code one target returns can be silently inert on another while
+  still looking correct. `openat(O_DIRECTORY|O_NOFOLLOW)` on a symlink may
+  report `ENOTDIR` or `ELOOP` depending on the platform, because the kernel can
+  reject the directory mismatch before it ever resolves the link; accept both
+  rather than assuming the one POSIX names. On Windows, `\\?\` and `\\.\`
+  prefixes are not UNC paths, and an `NtCreateFile` information class must match
+  the struct actually passed. Confirm the code each supported target returns
+  instead of reasoning from the man page.

CHANGELOG.md

@@ -6,6 +6,27 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 aims to follow [Semantic Versioning](https://semver.org/spec/v2.0.0.html) once the first release is
 tagged. Until then, source builds report the version `dev`.
 
+## [0.8.0](https://github.com/Gitlawb/zero/compare/v0.7.0...v0.8.0) (2026-08-21)
+
+
+### Features
+
+* **security:** path-jail primitive, credential-store locking, worktree git hardening ([#891](https://github.com/Gitlawb/zero/issues/891)) ([060d38c](https://github.com/Gitlawb/zero/commit/060d38ce716feb7c5b6657b082cd35cb8dda3d75))
+* **update:** recognise Homebrew installs and add mise, go install docs ([#910](https://github.com/Gitlawb/zero/issues/910)) ([01c3498](https://github.com/Gitlawb/zero/commit/01c3498205e69d26e5f929225f9455b276444f39))
+* use Exa as default search provider ([#926](https://github.com/Gitlawb/zero/issues/926)) ([1ec7219](https://github.com/Gitlawb/zero/commit/1ec7219aee7b2840481ab0905f131644e27d2a4e))
+* **zerogit:** auto-create a conventional branch before push/pr on default branch ([#671](https://github.com/Gitlawb/zero/issues/671)) ([d065467](https://github.com/Gitlawb/zero/commit/d065467cd3ec3ed91c5dc60554c36fc4b4d6de4c))
+
+
+### Bug Fixes
+
+* **acp:** three defects a desktop ACP client hits ([#915](https://github.com/Gitlawb/zero/issues/915)) ([a78c6c3](https://github.com/Gitlawb/zero/commit/a78c6c357e1d21146d7cd2e8030a754e06f7f1f5))
+* bump Go to 1.26.6 for stdlib vulnerability fixes ([#903](https://github.com/Gitlawb/zero/issues/903)) ([dc15e82](https://github.com/Gitlawb/zero/commit/dc15e822987d8d7de467c62c375518f435b74d8c))
+* **daemon:** clean up child after startup timeout ([#774](https://github.com/Gitlawb/zero/issues/774)) ([9c29209](https://github.com/Gitlawb/zero/commit/9c292092a26c9092fd873d072b0af1b0d9d39ef0))
+* **dictation:** redact API keys from streaming transcriber errors ([#852](https://github.com/Gitlawb/zero/issues/852)) ([64a783b](https://github.com/Gitlawb/zero/commit/64a783b08caa2c7b21f188033b9ed84a14e00e84))
+* **modelregistry:** expose reasoning effort for DeepSeek V4 models ([#931](https://github.com/Gitlawb/zero/issues/931)) ([90dcfd1](https://github.com/Gitlawb/zero/commit/90dcfd127e8a6d9902ec9f4e72f6d03fef1a0fc6))
+* **providers:** discover ChatGPT capabilities ([#890](https://github.com/Gitlawb/zero/issues/890)) ([2d2450e](https://github.com/Gitlawb/zero/commit/2d2450e9a744349f0d01b1d4e9ba29c24ba5650d))
+* **sandbox:** normalize launcher names before the command-prefix denylist ([#934](https://github.com/Gitlawb/zero/issues/934)) ([6edf9a8](https://github.com/Gitlawb/zero/commit/6edf9a8b78dc030dc44598919db1c7fa9d4f809a))
+
 ## [0.7.0](https://github.com/Gitlawb/zero/compare/v0.6.0...v0.7.0) (2026-08-10)
 
 

CONTRIBUTING.md

@@ -126,6 +126,31 @@ not enough. Dependency changes need a clear project benefit, such as fixing a
 specific bug, addressing a security issue, improving compatibility, or supporting
 an approved feature.
 
+## Supported platforms
+
+Zero publishes five native prebuilt artifacts from
+`.github/workflows/release-artifacts.yml`:
+
+- `linux-x64` (`linux/amd64`)
+- `linux-arm64` (`linux/arm64`)
+- `macos-arm64` (`darwin/arm64`)
+- `macos-x64` (`darwin/amd64`)
+- `windows-x64` (`windows/amd64`)
+
+The shell installer and npm wrapper resolve the applicable artifacts from that
+list. The PowerShell installer supports the published `windows-x64` artifact;
+there is no native Windows ARM64 artifact. Windows on ARM users can run the x64
+build under emulation or build from source.
+
+Android/Termux is also supported even though it has no distinct release
+artifact. The npm wrapper maps Android to the Linux artifact, and
+`docs/INSTALL.md` documents the native `android/arm64` source build needed on
+devices whose seccomp policy blocks the Linux build's syscall path.
+
+Other `GOOS`/`GOARCH` values, including Plan 9, are unsupported. They are not a
+review criterion. Do not block a pull request for failing to compile on an
+unsupported target. Tests may use `plan9` as a fixture for an unsupported OS.
+
 ## What We Prioritize
 
 The core team will prioritize:

Makefile

@@ -7,7 +7,7 @@ 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
+.PHONY: build build-all test test-race vet fmt fmt-check lint lint-static deadcode vulncheck vulncheck-windows tidy clean baseline help
 
 # Build the main CLI binary into ./zero.
 build:
@@ -44,7 +44,8 @@ lint: fmt-check vet
 # 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 deadcode vulncheck vulncheck-windows: export GOTOOLCHAIN = $(GO_TOOLCHAIN)
+lint-static deadcode vulncheck vulncheck-windows: export GOWORK = off
 
 lint-static:
 	go run github.com/golangci/golangci-lint/v2/cmd/golangci-lint@$(GOLANGCI_LINT_VERSION) run --enable-only unused,ineffassign,staticcheck ./...
@@ -55,6 +56,11 @@ deadcode:
 vulncheck:
 	go run golang.org/x/vuln/cmd/govulncheck@$(GOVULNCHECK_VERSION) ./...
 
+vulncheck-windows:
+	mkdir -p "$(CURDIR)/.cache/gobin"
+	GOBIN="$(CURDIR)/.cache/gobin" go install golang.org/x/vuln/cmd/govulncheck@$(GOVULNCHECK_VERSION)
+	GOOS=windows "$(CURDIR)/.cache/gobin/govulncheck" ./...
+
 tidy:
 	go mod tidy
 
@@ -77,4 +83,4 @@ baseline: build
 		--output internal/perfbench/reports/baseline.json
 
 help:
-	@echo "Targets: build (default), build-all, test, test-quick, vet, fmt, fmt-check, lint, lint-static, deadcode, vulncheck, tidy, clean, baseline"
+	@echo "Targets: build (default), build-all, test, test-quick, vet, fmt, fmt-check, lint, lint-static, deadcode, vulncheck, vulncheck-windows, 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 1.26.5+" src="https://img.shields.io/badge/Go-1.26.5+-00ADD8?logo=go&logoColor=white">
+  <img alt="Go 1.26.6+" src="https://img.shields.io/badge/Go-1.26.6+-00ADD8?logo=go&logoColor=white">
   <img alt="25+ providers" src="https://img.shields.io/badge/providers-25+-34E2EA">
   <a href="https://discord.gg/CaQDS6wdFn"><img alt="Discord" src="https://img.shields.io/badge/Discord-join-5865F2?logo=discord&logoColor=white"></a>
   <br>
@@ -26,7 +26,7 @@ zero exec --output-format stream-json < turns.jsonl
 ## Why Zero
 
 - **Use the model you want.** Bring OpenAI, Anthropic, Gemini, Groq, OpenRouter,
-  DeepSeek, Mistral, xAI, Qwen, Kimi, GitHub Models, Ollama, LM Studio, or any
+  DeepSeek, Mistral, xAI, Qwen, Kimi, GitHub Models, Ollama, LM Studio, Atomic Chat, or any
   OpenAI-/Anthropic-compatible endpoint.
 - **Stay in control.** File writes, shell commands, network access, and
   out-of-workspace writes go through Zero's permission and sandbox policy.
@@ -76,7 +76,7 @@ irm https://raw.githubusercontent.com/Gitlawb/zero/main/scripts/install.ps1 | ie
 
 ### From source
 
-Source builds require Go 1.26.5+.
+Source builds require Go 1.26.6+.
 
 ```bash
 git clone https://github.com/Gitlawb/zero.git
@@ -182,8 +182,12 @@ zero providers add custom-openai-compatible \
   --set-active
 ```
 
-For local models, run Ollama or LM Studio and then use `zero setup` or
-`zero providers detect`.
+For local models, run Ollama, LM Studio, or the [Atomic Chat](https://atomic.chat)
+desktop app, then use `zero setup` or `zero providers detect`. For Atomic Chat,
+load a model and enable its local OpenAI-compatible API (default
+`http://127.0.0.1:1337/v1`). Choose `atomic-chat-local`; detection includes the
+loaded model ID in the add command. If no usable ID is discovered, load a model
+and retry. Model IDs requiring shell-specific quoting use interactive setup.
 
 ## Daily Use
 
@@ -273,6 +277,19 @@ zero sandbox grants list
 Zero includes local file/search/edit/shell tools, `web_fetch` for public URLs,
 and MCP support for additional tools.
 
+`web_fetch` refuses loopback, private and other special-use addresses: it checks
+the URL before asking permission, resolves the host, and dials the address it
+validated so a name cannot resolve to something else in between.
+
+If `HTTP_PROXY`/`HTTPS_PROXY` is set, `web_fetch` and the provider connectivity
+probe use it, and that last step changes: the proxy is dialed and the target
+hostname is sent to it, so the proxy decides which address the request actually
+reaches. The URL is still checked and resolved locally first, but a proxy that
+answers differently can reach a private service. A forward proxy already sees
+and can rewrite every request through it, so this is the trust you accept by
+configuring one. Leave the variables unset for the checks to be enforced end to
+end.
+
 For local dev servers, use shell commands such as `curl` through `exec_command`
 so the normal sandbox and permission policy applies. Long-running commands stay
 attached to a background terminal session and can be listed or stopped from the
@@ -371,9 +388,13 @@ go run ./cmd/zero-perf-bench
 
 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.
+`zero exec` runs against official OpenAI profiles. Off by default; `0`, `false`,
+or `off` disable it. A/B-benchmark it by running the same `zero-perf-bench` suite
+with the variable unset and set.
+
+Native ChatGPT Responses sessions are enabled by default. Set
+`ZERO_CHATGPT_TURN_SESSION=0`, `false`, or `off` to restore stateless HTTP/SSE
+transport.
 
 ### Code Quality and Security Checks
 

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 1.26.5+" src="https://img.shields.io/badge/Go-1.26.5+-00ADD8?logo=go&logoColor=white">
+  <img alt="Go 1.26.6+" src="https://img.shields.io/badge/Go-1.26.6+-00ADD8?logo=go&logoColor=white">
   <img alt="25+ providers" src="https://img.shields.io/badge/providers-25+-34E2EA">
   <a href="https://discord.gg/CaQDS6wdFn"><img alt="Discord" src="https://img.shields.io/badge/Discord-join-5865F2?logo=discord&logoColor=white"></a>
   <br>
@@ -23,7 +23,7 @@ zero exec --output-format stream-json < turns.jsonl
 
 ## 为什么选择 Zero
 
-- **使用你想要的模型。** 支持 OpenAI、Anthropic、Gemini、Groq、OpenRouter、DeepSeek、Mistral、xAI、Qwen、Kimi、GitHub Models、Ollama、LM Studio，或任何 OpenAI/Anthropic 兼容端点。
+- **使用你想要的模型。** 支持 OpenAI、Anthropic、Gemini、Groq、OpenRouter、DeepSeek、Mistral、xAI、Qwen、Kimi、GitHub Models、Ollama、LM Studio、Atomic Chat，或任何 OpenAI/Anthropic 兼容端点。
 - **保持控制权。** 文件写入、Shell 命令、网络访问和工作区外写入都经过 Zero 的权限和沙箱策略。
 - **在终端中工作。** TUI 具有模型/提供商选择器、图片输入、斜杠命令、实时计划/工具渲染、回滚滚动、主题以及恢复/分叉支持。
 - **无 TUI 也能工作。** `zero exec` 可脚本化，支持文本/JSON/stream-JSON I/O、隔离的工作树、规范优先运行，以及用于 CI 的有意义的退出码。
@@ -57,7 +57,7 @@ irm https://raw.githubusercontent.com/Gitlawb/zero/main/scripts/install.ps1 | ie
 
 ### 从源码构建
 
-源码构建需要 Go 1.26.5+。
+源码构建需要 Go 1.26.6+。
 
 ```bash
 git clone https://github.com/Gitlawb/zero.git
@@ -114,7 +114,11 @@ export LONGCAT_API_KEY=...
 zero providers setup longcat --set-active
 ```
 
-对于本地模型，运行 Ollama 或 LM Studio，然后使用 `zero setup` 或 `zero providers detect`。
+对于本地模型，运行 Ollama、LM Studio 或 [Atomic Chat](https://atomic.chat) 桌面应用，
+然后使用 `zero setup` 或 `zero providers detect`。使用 Atomic Chat 时，请先加载模型并启用
+本地 OpenAI 兼容 API（默认地址为 `http://127.0.0.1:1337/v1`），再选择 `atomic-chat-local`。
+检测生成的添加命令会包含已加载的模型 ID；如果未发现可用的 ID，请加载模型后重试。
+对于需要特定 Shell 转义的模型 ID，请使用交互式设置。
 
 ## 日常使用
 

SECURITY.md

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

docs/HOW_ZERO_WORKS.md

@@ -627,7 +627,12 @@ Core tool groups include:
   skills, `ask_user`, permission requests.
 - **Write tools**: `write_file`, `edit_file`, `apply_patch`, `update_plan`.
 - **Shell tools**: `exec_command`, `write_stdin`, legacy `bash`.
-- **Network tools**: `web_fetch` and optional web search backend.
+- **Network tools**: `web_fetch` and optional web search backend. `web_fetch`
+  validates the URL before the permission prompt, refuses loopback/private/
+  special-use addresses, and dials the address it resolved. With
+  `HTTP_PROXY`/`HTTPS_PROXY` set, the proxy is dialed instead and resolves the
+  target itself, so destination enforcement ends at the proxy; see the note in
+  the README.
 - **Local-control tools**: config-gated wrappers for browser, terminal, desktop,
   and artifact helpers backed by packages such as `internal/localcontrol` and
   `internal/browser`.

docs/INSTALL.md

@@ -1,14 +1,30 @@
 # Installing Zero
 
-Zero is distributed as:
+Pick whichever fits how you already install things. None of these is the
+blessed one.
 
-- an npm package, `@gitlawb/zero`
-- release archives on GitHub Releases
-- source builds with Go 1.26.5+
+| Method | Command | Self-update |
+| --- | --- | --- |
+| Install script (Linux, macOS) | `curl -fsSL https://raw.githubusercontent.com/Gitlawb/zero/main/scripts/install.sh \| bash` | `zero upgrade` |
+| Install script (Windows) | `irm https://raw.githubusercontent.com/Gitlawb/zero/main/scripts/install.ps1 \| iex` | `zero upgrade` |
+| npm | `npm install -g @gitlawb/zero` | `zero upgrade` (runs npm) |
+| Release archive | download from [Releases](https://github.com/Gitlawb/zero/releases) | `zero upgrade` |
+| mise | `mise use -g ubi:Gitlawb/zero` | `mise upgrade` |
+| `go install` | `go install github.com/Gitlawb/zero/cmd/zero@latest` | rerun the command |
+| Source | `go build -o zero ./cmd/zero` | rebuild |
 
-The 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.
+`zero upgrade` follows whatever owns the binary, with one exception it cannot
+detect. It delegates to npm for an npm install, refuses outright for a Homebrew
+keg and names `brew upgrade zero`, and replaces the binary in place for
+everything else — including a mise-managed one, which is the exception.
+
+mise is the one case it cannot detect: a mise-managed binary is an ordinary
+standalone install on disk, so `zero upgrade` will replace it and leave mise
+describing a version that is no longer there. Use `mise upgrade` instead.
+
+Release archives are the substrate for most of the above: the install scripts
+and the npm fallback both download a platform archive from a published GitHub
+Release and verify its checksum.
 
 ## npm
 
@@ -111,6 +127,45 @@ Defaults:
 - Version: latest GitHub release
 - Install path: `%LOCALAPPDATA%\zero\bin\zero.exe`
 
+## mise
+
+[mise](https://mise.jdx.dev/) installs Zero straight from the GitHub Release
+archives through its `ubi` backend, so there is nothing extra to publish and no
+registry in the middle:
+
+```bash
+mise use -g ubi:Gitlawb/zero
+```
+
+Pin a version the same way you would any other tool:
+
+```bash
+mise use -g ubi:Gitlawb/[email protected]
+```
+
+Updates come from `mise upgrade`. `zero upgrade` also works, because a
+mise-managed binary is an ordinary standalone install, but then mise's records
+describe a version that is no longer on disk. Prefer `mise upgrade`.
+
+## go install
+
+```bash
+go install github.com/Gitlawb/zero/cmd/zero@latest
+```
+
+This builds from source, so it needs Go 1.26.6+ and it does not go through the
+release archives. Two consequences worth knowing before you pick it:
+
+- On Linux you also need the sandbox helper, which is a separate binary and is
+  not installed by this command. See
+  [Sandbox Helpers For Source Builds](#sandbox-helpers-for-source-builds).
+  Without it, native sandboxing is unavailable.
+- `zero upgrade` treats the result as a standalone install and will replace the
+  binary with a release build rather than rebuilding from source. If you chose
+  `go install` deliberately, rerun it instead.
+
+macOS and Windows need no extra helper.
+
 ## From Source
 
 ```bash
@@ -125,7 +180,7 @@ Build a local binary:
 go build -o zero ./cmd/zero
 ```
 
-Source builds require Go 1.26.5+.
+Source builds require Go 1.26.6+.
 
 ### Sandbox Helpers For Source Builds
 
@@ -242,3 +297,23 @@ zero upgrade
 ```
 
 See the [update guide](UPDATE.md) for update modes, flags, and platform details.
+
+`zero upgrade` behaves differently depending on how Zero was installed, because
+overwriting a binary a package manager owns leaves that manager describing a
+version that is no longer there:
+
+- **npm**: runs `npm install -g @gitlawb/zero@latest` for you.
+- **Homebrew**: refuses, and tells you to run `brew upgrade zero`. Replacing the
+  keg binary directly would be reverted by the next `brew upgrade` or
+  `brew reinstall`.
+- **everything else**: downloads the verified release archive and replaces the
+  binary in place.
+
+Homebrew is detected by the binary living inside a Cellar keg that carries
+Homebrew's own `INSTALL_RECEIPT.json`. The keg shape alone is not enough: a
+directory tree you built yourself can have exactly that shape, and refusing to
+update an install Homebrew never touched would be a real cost for a made-up
+reason. The receipt is the thing only Homebrew writes.
+
+An ordinary install under `/usr/local/bin` is left alone even on an Intel Mac
+where that is also the Homebrew prefix.

docs/UPDATE.md

@@ -70,3 +70,53 @@ Endpoint resolution order:
 Installer scripts download the matching release asset for the local platform and
 verify its `.sha256` file. If Zero is already installed, run `zero upgrade`
 instead of reinstalling.
+
+## Windows recovery state (standalone installs)
+
+This section describes Windows only. On Linux and macOS a standalone update
+renames the staged file directly over the executable path through the
+installation directory's file descriptor, so the replacement is atomic and no
+aside copy, marker, or recovery record is ever created — a failed update leaves
+the previous binary in place and nothing to resolve.
+
+On Windows, a running executable cannot be replaced in place, so the previous
+binary is moved aside to `<binary>.zero-update-<random>.old` first. The updater
+records that exact file — bound to its filesystem identity, in per-user state
+outside the installation directory — and deletes only that recorded copy after
+the next update is verified in place. Backups it did not create are never
+removed, so a file such as `zero.exe.before-manual-patch.old` is left alone. A
+recorded copy that something else holds open (an on-access scanner, an editor)
+stays recorded and is removed by a later update instead.
+
+An update **refuses to run** while unresolved recovery state exists beside the
+binary:
+
+| State on disk | Meaning |
+|---|---|
+| `<binary>.old` (or `<binary>.<suffix>.old`) plus a `.keep` marker, or a per-user update-recovery record naming that copy | A previous update could not restore the original binary. The `.old` file may be the last binary the updater verified; the installed one may be unverified. The `.keep` marker is in the installation directory and can be deleted; the identity-bound record is not, so deleting the marker alone does not end the refusal. |
+| `<binary>.…old.<suffix>.recovery` | A previous update could not even write the marker, so it moved the last verified binary to that name. |
+| The binary is missing and one or more `*.old` files exist | The previous attempt was interrupted between moves. |
+
+The refusal names the paths involved and the two moves that end the state:
+either move the recovery binary back over the executable path, or — if the
+installed binary is the one you want — delete the recovery copy itself (not
+only the `.keep` marker) after verifying the installed binary, then update
+again. Deleting the marker alone is not enough: a failed restore also records
+that copy in per-user state outside the installation directory, and the next
+update consults that record even if the marker is gone.
+
+That trusted record lives under `%AppData%\zero\update-recovery` (per-user).
+The recovery copy and its `.keep` marker sit in the installation directory
+(per-machine). A second Windows account on the same machine does not see the
+first account's unresolved record; the sibling marker is the machine-visible
+signal for every account. There is no separate machine-wide store: writing one
+would need a shared writable location and ACL rules the updater does not own.
+
+This is deliberately fail-closed and differs from older releases, which deleted
+`<binary>.old` and proceeded. The trade-off is that anyone who can write in the
+installation directory can plant `<binary>.old` and `<binary>.old.keep` there
+and make every subsequent update refuse until an operator clears them. That is
+the safe direction: the alternative is an update that overwrites the only
+verified copy of the previous binary. If updates start refusing, inspect those
+files before removing them — an installation directory that a lower-privileged
+account can write to is itself worth fixing.

go.mod

@@ -1,26 +1,28 @@
 module github.com/Gitlawb/zero
 
-go 1.26.5
+go 1.26.6
 
 require (
 	charm.land/bubbles/v2 v2.1.1
-	charm.land/bubbletea/v2 v2.0.8
+	charm.land/bubbletea/v2 v2.0.9
 	charm.land/lipgloss/v2 v2.0.5
 	github.com/Microsoft/go-winio v0.6.2
 	github.com/alecthomas/chroma/v2 v2.27.0
 	github.com/atotto/clipboard v0.1.4
 	github.com/aymanbagabas/go-udiff v0.4.1
 	github.com/charmbracelet/colorprofile v0.4.3
+	github.com/charmbracelet/ultraviolet v0.0.0-20260703014108-f5a850f9c2b7
 	github.com/charmbracelet/x/ansi v0.11.7
 	github.com/charmbracelet/x/term v0.2.2
 	github.com/coder/websocket v1.8.15
 	github.com/ledongthuc/pdf v0.0.0-20250511090121-5959a4027728
+	golang.org/x/image v0.45.0
 	golang.org/x/sys v0.47.0
+	gopkg.in/yaml.v3 v3.0.1
 	mvdan.cc/sh/v3 v3.13.1
 )
 
 require (
-	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

go.sum

@@ -1,7 +1,7 @@
 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/bubbletea/v2 v2.0.9 h1:DpJCMWKgzQK8SJv4zbKKFHAI10ymWy/evClPFk0k0f8=
+charm.land/bubbletea/v2 v2.0.9/go.mod h1:2SkdgoTXluXJHOUwAoRlRXF/28vklb1rFl6GcgV1/ss=
 charm.land/lipgloss/v2 v2.0.5 h1:kbNxgeeUOYv5J0YdpxFjfvf3dFvqH8Aci4zB6xqFtrY=
 charm.land/lipgloss/v2 v2.0.5/go.mod h1:9oqhxt4yxIMe6q5A4kHr44DremZk7J9UNh74GlWa5nc=
 github.com/Microsoft/go-winio v0.6.2 h1:F2VQgta7ecxGYO8k3ZZz3RS8fVIXVxONVUPlNERoyfY=
@@ -64,9 +64,15 @@ 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/image v0.45.0 h1:FMb1nTbH5H9vF55SriQHgFw5GnNL9Jg6L25BwXKzhB0=
+golang.org/x/image v0.45.0/go.mod h1:n62x/7RqlwXDvGsSU4u6IUTUf6KghUZ9Bt7cG/T9Fx4=
 golang.org/x/sync v0.22.0 h1:SZjpbeLmrCk4xhRSZFNZW5gFUeCeFgjekvI/+gfScek=
 golang.org/x/sync v0.22.0/go.mod h1:9xrNwdLfx4jkKbNva9FpL6vEN7evnE43NNNJQ2LF3+0=
 golang.org/x/sys v0.47.0 h1:o7XGOvZQCADBQQ4Y7VNq2dRWQR7JmOUW8Kxx4ZsNgWs=
 golang.org/x/sys v0.47.0/go.mod h1:4GL1E5IUh+htKOUEOaiffhrAeqysfVGipDYzABqnCmw=
+gopkg.in/check.v1 v0.0.0-20161208181325-20d25e280405 h1:yhCVgyC4o1eVCa2tZl7eS0r+SDo693bJlVdllGtEeKM=
+gopkg.in/check.v1 v0.0.0-20161208181325-20d25e280405/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0=
+gopkg.in/yaml.v3 v3.0.1 h1:fxVm/GzAzEWqLHuvctI91KS9hhNmmWOoWu0XTYJS7CA=
+gopkg.in/yaml.v3 v3.0.1/go.mod h1:K4uyk7z7BCEPqu6E+C64Yfv1cQ7kz7rIZviUmN+EgEM=
 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

@@ -14,6 +14,7 @@ import (
 	"github.com/Gitlawb/zero/internal/providercatalog"
 	"github.com/Gitlawb/zero/internal/providermodelcatalog"
 	"github.com/Gitlawb/zero/internal/providermodeldiscovery"
+	"github.com/Gitlawb/zero/internal/redaction"
 	"github.com/Gitlawb/zero/internal/sandbox"
 	"github.com/Gitlawb/zero/internal/sessions"
 	"github.com/Gitlawb/zero/internal/tools"
@@ -30,10 +31,12 @@ type Deps struct {
 	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.
-	BuildWorkspace func(workspaceRoot string, resolved config.ResolvedConfig) (*tools.Registry, *sandbox.Engine, error)
+	// BuildWorkspace creates the per-turn scoped tool workspace for a validated
+	// workspace root. Its Close method releases any resources the registry owns
+	// (notably MCP server connections) after the turn completes or is cancelled.
+	// Shell tools remain confined exactly like the exec surface — never unconfined
+	// on the host.
+	BuildWorkspace func(ctx context.Context, workspaceRoot string, resolved config.ResolvedConfig, mode agent.PermissionMode) (*Workspace, error)
 	// ResolveWorkspaceRoot validates + normalizes a client-supplied cwd (must be an
 	// existing directory; never the bare root). It is the file-tool confinement root.
 	ResolveWorkspaceRoot func(cwd string) (string, error)
@@ -41,6 +44,27 @@ type Deps struct {
 	AgentInfo            Implementation
 }
 
+// Workspace is the per-turn execution environment passed to the agent. ACP does
+// not retain it across turns because MCP connections belong to the registry that
+// advertised their tools; keeping a stale registry after a cancelled turn would
+// leak its server process and its permission state into the next turn.
+type Workspace struct {
+	Registry       *tools.Registry
+	Sandbox        *sandbox.Engine
+	DeferThreshold int
+	Notices        []string
+	Cleanup        func() error
+}
+
+// Close releases resources created with the workspace. A nil cleanup is valid
+// for core-only workspaces.
+func (w *Workspace) Close() error {
+	if w == nil || w.Cleanup == nil {
+		return nil
+	}
+	return w.Cleanup()
+}
+
 // Agent is the ACP agent server bound to one JSON-RPC connection (one editor).
 type Agent struct {
 	conn *Conn
@@ -241,13 +265,34 @@ func (a *Agent) runTurn(ctx context.Context, sess *acpSession, userText string,
 	if err != nil {
 		return "", RPCError(codeInternalError, "provider: "+err.Error())
 	}
+	mode := sess.currentMode()
 	// Build the SCOPED registry + sandbox engine for this session's workspace so
 	// shell/file tools are confined to the workspace exactly like the exec surface.
-	registry, sandboxEngine, err := a.deps.BuildWorkspace(sess.cwd, resolved)
+	// The workspace can also own MCP connections, which must be closed after every
+	// turn even if the agent run returns an error or the client cancels it.
+	workspace, err := a.deps.BuildWorkspace(ctx, sess.cwd, resolved, mode)
 	if err != nil {
 		return "", RPCError(codeInternalError, "workspace: "+err.Error())
 	}
+	if workspace == nil || workspace.Registry == nil {
+		if workspace != nil {
+			_ = workspace.Close()
+		}
+		return "", RPCError(codeInternalError, "workspace: missing tool registry")
+	}
+	defer func() {
+		if err := workspace.Close(); err != nil {
+			log.Printf("acp: close workspace: %v", err)
+		}
+	}()
+	registry := workspace.Registry
+	sandboxEngine := workspace.Sandbox
 	note := &notifier{conn: a.conn, sessionID: sess.id}
+	for _, notice := range workspace.Notices {
+		if text := strings.TrimSpace(redaction.RedactString(notice, redaction.Options{})); text != "" {
+			note.text("\n\n[zero warning] " + text + "\n")
+		}
+	}
 
 	opts := agent.Options{
 		Cwd:            sess.cwd,
@@ -256,7 +301,8 @@ func (a *Agent) runTurn(ctx context.Context, sess *acpSession, userText string,
 		Model:          resolved.Provider.Model,
 		Registry:       registry,
 		Sandbox:        sandboxEngine,
-		PermissionMode: sess.currentMode(),
+		PermissionMode: mode,
+		DeferThreshold: workspace.DeferThreshold,
 		MaxTurns:       resolved.MaxTurns,
 		Images:         images,
 		OnText:         note.text,
@@ -296,6 +342,16 @@ func stopReasonFor(result agent.Result, err error) (string, error) {
 		if errors.Is(err, context.Canceled) {
 			return StopCancelled, nil
 		}
+		// A CLIENT THAT CANCELLED A PERMISSION PROMPT DID NOT FAIL. Only
+		// context.Canceled was matched here, so dismissing a permission dialog —
+		// a deliberate action, and for apply_patch the only refusal a client is
+		// offered — came back as JSON-RPC -32603 carrying the internal sentinel
+		// text. Editors and the desktop app both render that as a crashed turn,
+		// so declining a tool looked like ZERO falling over. It is a
+		// cancellation, and StopCancelled is what ACP has for saying so.
+		if errors.Is(err, agent.ErrPermissionApprovalCanceled) {
+			return StopCancelled, nil
+		}
 		return "", err
 	}
 	if result.FinishReason == "length" {
@@ -311,10 +367,14 @@ func stopReasonFor(result agent.Result, err error) (string, error) {
 // session/request_permission request and maps the outcome back. Failure to reach
 // the client fails closed to deny.
 func (a *Agent) requestPermission(ctx context.Context, sessionID string, req agent.PermissionRequest) (agent.PermissionDecision, error) {
+	// Built ONCE and used for both halves. Sending one list and validating the
+	// reply against another is the defect this whole path had; see
+	// decisionFromOutcome.
+	options := buildPermissionOptions(req)
 	params := RequestPermissionParams{
 		SessionID: sessionID,
 		ToolCall:  permissionToolCall(req),
-		Options:   buildPermissionOptions(req),
+		Options:   options,
 	}
 	var result RequestPermissionResult
 	if err := a.conn.Call(ctx, MethodSessionRequestPerm, params, &result); err != nil {
@@ -323,7 +383,7 @@ func (a *Agent) requestPermission(ctx context.Context, sessionID string, req age
 		}
 		return agent.PermissionDecision{Action: agent.PermissionDecisionDeny, Reason: "permission request failed: " + err.Error()}, nil
 	}
-	return decisionFromOutcome(result.Outcome, req.AvailableDecisions), nil
+	return decisionFromOutcome(result.Outcome, options), nil
 }
 
 func (a *Agent) emitPlan(registry *tools.Registry, note *notifier) {

internal/acp/agent_test.go

@@ -8,6 +8,7 @@ import (
 	"io"
 	"os"
 	"path/filepath"
+	"slices"
 	"strings"
 	"testing"
 	"time"
@@ -35,6 +36,35 @@ func (f fakeProvider) StreamCompletion(_ context.Context, _ zeroruntime.Completi
 	return ch, nil
 }
 
+type acpDeferredPromptTool struct{ name string }
+
+func (t acpDeferredPromptTool) Name() string           { return t.name }
+func (acpDeferredPromptTool) Description() string      { return "deferred MCP test tool" }
+func (acpDeferredPromptTool) Parameters() tools.Schema { return tools.Schema{Type: "object"} }
+func (acpDeferredPromptTool) Deferred() bool           { return true }
+func (acpDeferredPromptTool) Run(context.Context, map[string]any) tools.Result {
+	return tools.Result{Status: tools.StatusOK, Output: "ok"}
+}
+func (acpDeferredPromptTool) Safety() tools.Safety {
+	return tools.Safety{
+		SideEffect: tools.SideEffectNetwork, Permission: tools.PermissionPrompt,
+		AdvertiseInAuto: true,
+	}
+}
+
+type captureToolsProvider struct {
+	requests chan zeroruntime.CompletionRequest
+}
+
+func (p captureToolsProvider) StreamCompletion(_ context.Context, request zeroruntime.CompletionRequest) (<-chan zeroruntime.StreamEvent, error) {
+	p.requests <- request
+	ch := make(chan zeroruntime.StreamEvent, 2)
+	ch <- zeroruntime.StreamEvent{Type: zeroruntime.StreamEventText, Content: "done"}
+	ch <- zeroruntime.StreamEvent{Type: zeroruntime.StreamEventDone}
+	close(ch)
+	return ch, nil
+}
+
 func testDeps(t *testing.T) Deps {
 	t.Helper()
 	store := sessions.NewStore(sessions.StoreOptions{RootDir: t.TempDir()})
@@ -53,10 +83,10 @@ func testDeps(t *testing.T) Deps {
 			return fakeProvider{text: "Hello from ZERO"}, nil
 		},
 		RunAgent: agent.Run,
-		BuildWorkspace: func(string, config.ResolvedConfig) (*tools.Registry, *sandbox.Engine, error) {
+		BuildWorkspace: func(context.Context, string, config.ResolvedConfig, agent.PermissionMode) (*Workspace, error) {
 			r := tools.NewRegistry()
 			r.Register(tools.NewUpdatePlanTool())
-			return r, nil, nil
+			return &Workspace{Registry: r}, nil
 		},
 		ResolveWorkspaceRoot: func(cwd string) (string, error) { return cwd, nil },
 		Store:                store,
@@ -468,8 +498,8 @@ func TestACPRunTurnWiresSandboxAndScopedRegistry(t *testing.T) {
 	reg := tools.NewRegistry()
 	reg.Register(tools.NewUpdatePlanTool())
 	engine := sandbox.NewEngine(sandbox.EngineOptions{WorkspaceRoot: t.TempDir()})
-	deps.BuildWorkspace = func(string, config.ResolvedConfig) (*tools.Registry, *sandbox.Engine, error) {
-		return reg, engine, nil
+	deps.BuildWorkspace = func(context.Context, string, config.ResolvedConfig, agent.PermissionMode) (*Workspace, error) {
+		return &Workspace{Registry: reg, Sandbox: engine}, nil
 	}
 	var captured agent.Options
 	deps.RunAgent = func(_ context.Context, _ string, _ zeroruntime.Provider, opts agent.Options) (agent.Result, error) {
@@ -497,6 +527,124 @@ func TestACPRunTurnWiresSandboxAndScopedRegistry(t *testing.T) {
 	}
 }
 
+func TestACPRunTurnUsesWorkspaceDeferThresholdAtProviderBoundary(t *testing.T) {
+	for _, tc := range []struct {
+		name      string
+		threshold int
+		want      []string
+	}{
+		{name: "explicit zero disables", threshold: 0, want: []string{"mcp_docs_alpha", "mcp_docs_beta"}},
+		{name: "below threshold stays eager", threshold: 3, want: []string{"mcp_docs_alpha", "mcp_docs_beta"}},
+		{name: "at threshold defers", threshold: 2, want: []string{tools.ToolSearchToolName}},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			requests := make(chan zeroruntime.CompletionRequest, 4)
+			deps := testDeps(t)
+			deps.NewProvider = func(config.ProviderProfile) (zeroruntime.Provider, error) {
+				return captureToolsProvider{requests: requests}, nil
+			}
+			deps.BuildWorkspace = func(context.Context, string, config.ResolvedConfig, agent.PermissionMode) (*Workspace, error) {
+				registry := tools.NewRegistry()
+				registry.Register(acpDeferredPromptTool{name: "mcp_docs_alpha"})
+				registry.Register(acpDeferredPromptTool{name: "mcp_docs_beta"})
+				registry.Register(tools.NewToolSearchTool(registry))
+				return &Workspace{Registry: registry, DeferThreshold: tc.threshold}, nil
+			}
+
+			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.Fatal(err)
+			}
+			if err := h.client.Call(ctx, MethodSessionPrompt, PromptParams{SessionID: created.SessionID, Prompt: []ContentBlock{TextBlock("hi")}}, &PromptResult{}); err != nil {
+				t.Fatal(err)
+			}
+			request := <-requests
+			got := make([]string, 0, len(request.Tools))
+			for _, definition := range request.Tools {
+				got = append(got, definition.Name)
+			}
+			if !slices.Equal(got, tc.want) {
+				t.Fatalf("provider tools = %#v, want %#v", got, tc.want)
+			}
+		})
+	}
+}
+
+func TestACPRunTurnEmitsWorkspaceSetupNotices(t *testing.T) {
+	deps := testDeps(t)
+	deps.BuildWorkspace = func(context.Context, string, config.ResolvedConfig, agent.PermissionMode) (*Workspace, error) {
+		registry := tools.NewRegistry()
+		registry.Register(tools.NewUpdatePlanTool())
+		return &Workspace{Registry: registry, Notices: []string{"MCP server docs unavailable, skipped: timeout"}}, nil
+	}
+	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.Fatal(err)
+	}
+	if err := h.client.Call(ctx, MethodSessionPrompt, PromptParams{SessionID: created.SessionID, Prompt: []ContentBlock{TextBlock("hi")}}, &PromptResult{}); err != nil {
+		t.Fatal(err)
+	}
+	got := drainText(t, h.updates)
+	if !strings.Contains(got, "[zero warning] MCP server docs unavailable, skipped: timeout") || !strings.Contains(got, "Hello from ZERO") {
+		t.Fatalf("streamed text = %q", got)
+	}
+}
+
+// TestACPRunTurnClosesWorkspace proves a per-turn workspace never outlives the
+// agent run. This is particularly important for MCP: its registry owns live
+// client connections and, for stdio servers, child processes.
+func TestACPRunTurnClosesWorkspaceAfterSuccessAndFailure(t *testing.T) {
+	for _, tc := range []struct {
+		name   string
+		runErr error
+	}{
+		{name: "success"},
+		{name: "agent failure", runErr: errors.New("provider interrupted")},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			deps := testDeps(t)
+			closed := 0
+			deps.BuildWorkspace = func(context.Context, string, config.ResolvedConfig, agent.PermissionMode) (*Workspace, error) {
+				registry := tools.NewRegistry()
+				registry.Register(tools.NewUpdatePlanTool())
+				return &Workspace{Registry: registry, Cleanup: func() error {
+					closed++
+					return nil
+				}}, nil
+			}
+			deps.RunAgent = func(context.Context, string, zeroruntime.Provider, agent.Options) (agent.Result, error) {
+				return agent.Result{FinalAnswer: "ok"}, tc.runErr
+			}
+			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)
+			}
+			err := h.client.Call(ctx, MethodSessionPrompt, PromptParams{SessionID: created.SessionID, Prompt: []ContentBlock{TextBlock("hello")}}, &PromptResult{})
+			if tc.runErr == nil && err != nil {
+				t.Fatalf("session/prompt: %v", err)
+			}
+			if tc.runErr != nil && err == nil {
+				t.Fatal("session/prompt succeeded after agent failure")
+			}
+			if closed != 1 {
+				t.Fatalf("workspace cleanup calls = %d, want 1", closed)
+			}
+		})
+	}
+}
+
 // TestACPRejectsInvalidCwd confirms session/new fails when the workspace root
 // resolver rejects the client cwd (e.g. filesystem root).
 func TestACPRejectsInvalidCwd(t *testing.T) {

internal/acp/desktopinterop_test.go

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

internal/acp/permission.go

@@ -15,14 +15,7 @@ import (
 // PermissionOptions. Only the actions ZERO actually presented (AvailableDecisions)
 // are surfaced; the optionId is the ZERO action string for a clean round-trip.
 func buildPermissionOptions(req agent.PermissionRequest) []PermissionOption {
-	actions := req.AvailableDecisions
-	if len(actions) == 0 {
-		// Sensible default if ZERO didn't enumerate: allow once / reject.
-		actions = []agent.PermissionDecisionAction{
-			agent.PermissionDecisionAllow,
-			agent.PermissionDecisionDeny,
-		}
-	}
+	actions := offeredDecisions(req)
 	options := make([]PermissionOption, 0, len(actions))
 	for _, action := range actions {
 		kind, name := optionKindFor(action, req.PrefixApprovalEscalates)
@@ -38,6 +31,32 @@ func buildPermissionOptions(req agent.PermissionRequest) []PermissionOption {
 	return options
 }
 
+// offeredDecisions resolves what this request actually offers the client.
+//
+// ONE RESOLVER, USED BY BOTH SIDES, and that is the entire point of it existing.
+// The fallback below used to live inside buildPermissionOptions while
+// requestPermission validated the client's answer against the raw
+// req.AvailableDecisions. When ZERO did not enumerate — which is every
+// permission event that is not a prompt, and includes a shell call whose sandbox
+// decision came back deny — the two disagreed completely: the client was sent
+// "Allow" and "Reject", and validation was performed against an EMPTY list, so
+// every option the client could possibly click failed closed to deny with
+// "permission option was not offered".
+//
+// The user clicked Allow and ZERO recorded a denial. Nothing surfaced it,
+// because a denial is a legitimate answer — the tool was simply refused and the
+// turn carried on as though the user had rejected it.
+func offeredDecisions(req agent.PermissionRequest) []agent.PermissionDecisionAction {
+	if len(req.AvailableDecisions) > 0 {
+		return req.AvailableDecisions
+	}
+	// Sensible default if ZERO didn't enumerate: allow once / reject.
+	return []agent.PermissionDecisionAction{
+		agent.PermissionDecisionAllow,
+		agent.PermissionDecisionDeny,
+	}
+}
+
 // optionKindFor maps a ZERO decision action to an ACP PermissionOptionKind and a
 // human label. Returns an empty kind for actions that ACP expresses through the
 // outcome rather than an option (cancel).
@@ -49,8 +68,16 @@ func buildPermissionOptions(req agent.PermissionRequest) []PermissionOption {
 // disclosure cannot be TUI-only.
 func optionKindFor(action agent.PermissionDecisionAction, escalates bool) (kind, name string) {
 	switch action {
-	case agent.PermissionDecisionAllow, agent.PermissionDecisionAllowStrict:
+	case agent.PermissionDecisionAllow:
 		return PermAllowOnce, "Allow"
+	case agent.PermissionDecisionAllowStrict:
+		// SPLIT FROM PLAIN ALLOW because the label is the only thing an ACP
+		// client shows. Both were labelled "Allow", and request_permissions
+		// offers them together — so the panel presented two identical buttons,
+		// and whichever the user picked looked the same while one of them
+		// silently turned on strict auto-review of what was granted. The option
+		// id carries no room for the distinction, so it has to be in the name.
+		return PermAllowOnce, "Allow with strict review"
 	case agent.PermissionDecisionAllowForSession:
 		return PermAllowAlways, "Allow for this session"
 	case agent.PermissionDecisionAllowPrefix:
@@ -70,19 +97,31 @@ func optionKindFor(action agent.PermissionDecisionAction, escalates bool) (kind,
 
 // decisionFromOutcome maps the client's permission outcome back to a ZERO
 // decision. A cancelled outcome cancels the run; a selected option id is the ZERO
-// action verbatim (validated against what was offered); anything unrecognized
-// fails closed to deny.
-func decisionFromOutcome(outcome RequestPermissionOutcome, offered []agent.PermissionDecisionAction) agent.PermissionDecision {
+// action verbatim (validated against what was actually sent); anything
+// unrecognized fails closed to deny.
+//
+// IT TAKES THE OPTIONS THAT WERE SENT, not the decisions the request offered,
+// and the two are not the same list. Some actions never become options —
+// PermissionDecisionCancel is dropped by optionKindFor because ACP expresses
+// cancellation through the outcome instead — yet it IS enumerated in
+// AvailableDecisions for shell commands and apply_patch, the two commonest
+// prompts. Validating against the decisions therefore accepted
+// {"outcome":"selected","optionId":"cancel"}: an identifier no client was ever
+// shown, aborting the whole turn.
+//
+// Passing the sent options makes "what we offered" and "what we accept" the
+// same slice rather than two things kept in step, so anything optionKindFor
+// declines to render is excluded structurally instead of by name.
+func decisionFromOutcome(outcome RequestPermissionOutcome, offered []PermissionOption) agent.PermissionDecision {
 	switch outcome.Outcome {
 	case OutcomeCancelled:
 		return agent.PermissionDecision{Action: agent.PermissionDecisionCancel, Reason: "client cancelled"}
 	case OutcomeSelected:
-		action := agent.PermissionDecisionAction(outcome.OptionID)
 		// Bind to what was actually offered for THIS call: a client must not be able
 		// to return a broader grant (always_allow / allow_for_session) that wasn't
 		// presented. Anything not offered fails closed to deny.
-		if actionOffered(action, offered) {
-			return agent.PermissionDecision{Action: action}
+		if actionOffered(outcome.OptionID, offered) {
+			return agent.PermissionDecision{Action: agent.PermissionDecisionAction(outcome.OptionID)}
 		}
 		return agent.PermissionDecision{Action: agent.PermissionDecisionDeny, Reason: "permission option was not offered"}
 	default:
@@ -90,9 +129,9 @@ func decisionFromOutcome(outcome RequestPermissionOutcome, offered []agent.Permi
 	}
 }
 
-func actionOffered(action agent.PermissionDecisionAction, offered []agent.PermissionDecisionAction) bool {
-	for _, a := range offered {
-		if a == action {
+func actionOffered(optionID string, offered []PermissionOption) bool {
+	for _, option := range offered {
+		if option.OptionID == optionID {
 			return true
 		}
 	}
@@ -103,13 +142,15 @@ func actionOffered(action agent.PermissionDecisionAction, offered []agent.Permis
 // session/request_permission request from a ZERO permission request.
 func permissionToolCall(req agent.PermissionRequest) ToolCallUpdate {
 	args := marshalArgs(req.Args)
-	return ToolCallUpdate{
+	upd := ToolCallUpdate{
 		ToolCallID: req.ToolCallID,
 		Title:      toolTitle(req.ToolName, string(args)),
 		Kind:       toolKindFor(req.ToolName),
 		Status:     ToolStatusPending,
 		RawInput:   rawInputBytes(args),
 	}
+	attachBrowserToolDetails(&upd, req.ToolName)
+	return upd
 }
 
 func marshalArgs(args map[string]any) []byte {

internal/acp/permission_escalation_test.go

@@ -56,10 +56,11 @@ func TestACPPrefixOptionsStaySilentWithoutEscalation(t *testing.T) {
 // what was offered, so decorating it would fail every prefix approval closed.
 func TestACPDisclosureDoesNotDisturbTheOptionIDRoundTrip(t *testing.T) {
 	request := escalationRequest(true)
-	for _, option := range buildPermissionOptions(request) {
+	options := buildPermissionOptions(request)
+	for _, option := range options {
 		decision := decisionFromOutcome(
 			RequestPermissionOutcome{Outcome: OutcomeSelected, OptionID: option.OptionID},
-			request.AvailableDecisions,
+			options,
 		)
 		if string(decision.Action) != option.OptionID {
 			t.Errorf("option id %q round-tripped to %q (%s)", option.OptionID, decision.Action, decision.Reason)

internal/acp/permission_test.go

@@ -39,11 +39,14 @@ func TestBuildPermissionOptionsDefault(t *testing.T) {
 }
 
 func TestDecisionFromOutcome(t *testing.T) {
-	offered := []agent.PermissionDecisionAction{
-		agent.PermissionDecisionAllow,
-		agent.PermissionDecisionAlwaysAllow,
-		agent.PermissionDecisionDeny,
-	}
+	offered := buildPermissionOptions(agent.PermissionRequest{
+		ToolName: "bash",
+		AvailableDecisions: []agent.PermissionDecisionAction{
+			agent.PermissionDecisionAllow,
+			agent.PermissionDecisionAlwaysAllow,
+			agent.PermissionDecisionDeny,
+		},
+	})
 	if d := decisionFromOutcome(RequestPermissionOutcome{Outcome: OutcomeCancelled}, offered); d.Action != agent.PermissionDecisionCancel {
 		t.Errorf("cancelled -> %q, want cancel", d.Action)
 	}
@@ -79,3 +82,17 @@ func TestPermissionToolCall(t *testing.T) {
 		t.Error("expected rawInput from args")
 	}
 }
+
+func TestPermissionToolCallKeepsTheBrowserDescriptor(t *testing.T) {
+	call := permissionToolCall(agent.PermissionRequest{
+		ToolCallID: "browser-1",
+		ToolName:   "browser_connect",
+		Args:       map[string]any{"target": "127.0.0.1:9222"},
+	})
+	if got := browserDescriptor(t, call); got != (BrowserToolDetails{Version: 1, Command: "connect"}) {
+		t.Fatalf("browser descriptor = %#v", got)
+	}
+	if call.Title != "browser connect" {
+		t.Fatalf("title = %q", call.Title)
+	}
+}

internal/acp/translate.go

@@ -2,7 +2,9 @@ package acp
 
 import (
 	"encoding/json"
+	"net/url"
 	"strings"
+	"unicode"
 	"unicode/utf8"
 
 	"github.com/Gitlawb/zero/internal/agent"
@@ -44,12 +46,125 @@ func toolKindFor(name string) string {
 
 // toolTitle builds a concise human title, e.g. "read_file src/main.go".
 func toolTitle(name, rawArgs string) string {
+	if browser, ok := browserToolDetails(name); ok {
+		return browserToolTitle(browser.Command, rawArgs)
+	}
 	if hint := primaryArgHint(rawArgs); hint != "" {
 		return name + " " + hint
 	}
 	return name
 }
 
+// browserToolDetails identifies ZERO's local browser helpers without treating
+// similarly named MCP tools as browser automation. The descriptor intentionally
+// contains no request data: ACP tool input is already protocol-visible, but a
+// durable UI must not need to retain text, local CDP targets, or full URLs just
+// to recognise the browser operation.
+func browserToolDetails(name string) (*BrowserToolDetails, bool) {
+	const prefix = "browser_"
+	command, ok := strings.CutPrefix(name, prefix)
+	if !ok {
+		return nil, false
+	}
+	switch command {
+	case "install", "launch", "connect", "open", "snapshot", "click", "type", "press", "action":
+		return &BrowserToolDetails{Version: 1, Command: command}, true
+	default:
+		return nil, false
+	}
+}
+
+const zeroBrowserMetaKey = "github.com/Gitlawb/zero/browser"
+
+// attachBrowserToolDetails stores ZERO's browser descriptor in ACP's reserved
+// extension channel. Keeping this in one helper prevents start, result, and
+// permission payloads from drifting onto different wire shapes.
+func attachBrowserToolDetails(update *ToolCallUpdate, name string) {
+	browser, ok := browserToolDetails(name)
+	if !ok {
+		return
+	}
+	raw, err := json.Marshal(browser)
+	if err != nil {
+		return
+	}
+	update.Meta = map[string]json.RawMessage{zeroBrowserMetaKey: raw}
+}
+
+// browserToolTitle avoids putting browser_type text, an attached DevTools
+// endpoint, or a URL query/fragment in a tool-card title. Those values can
+// carry credentials or session data; the UI only needs the operation and, for
+// navigation, a human-recognisable origin.
+func browserToolTitle(command, rawArgs string) string {
+	switch command {
+	case "action":
+		action, ok := exactJSONStringArg(rawArgs, "command")
+		if !ok {
+			return "browser action"
+		}
+		if action, ok := tools.NormalizedBrowserActionCommand(action); ok {
+			return "browser action " + action
+		}
+		return "browser action"
+	case "open":
+		rawURL, ok := exactJSONStringArg(rawArgs, "url")
+		if !ok {
+			return "browser open"
+		}
+		normalized, err := tools.NormalizeBrowserOpenURL(rawURL)
+		if err != nil {
+			return "browser open"
+		}
+		u, err := url.Parse(normalized)
+		if err != nil || u.Scheme == "" || u.Host == "" {
+			return "browser open"
+		}
+		origin := u.Scheme + "://" + u.Host
+		if !browserTitleTextSafe(origin) {
+			return "browser open"
+		}
+		return "browser open " + truncateHint(origin)
+	default:
+		return "browser " + command
+	}
+}
+
+// browserTitleTextSafe validates text after URL parsing has decoded escaped
+// UTF-8 in the host. Valid UTF-8 alone is not presentation-safe: control,
+// format/bidi, and line/paragraph separator runes can reorder or split the
+// permission label shown to a user. The execution URL remains unchanged.
+func browserTitleTextSafe(text string) bool {
+	if !utf8.ValidString(text) {
+		return false
+	}
+	for _, r := range text {
+		if unicode.IsControl(r) || unicode.Is(unicode.Cf, r) || unicode.Is(unicode.Zl, r) || unicode.Is(unicode.Zp, r) {
+			return false
+		}
+	}
+	return true
+}
+
+// exactJSONStringArg mirrors ZERO's map-based tool argument decoding: only the
+// exact JSON key is considered, and a non-string value is invalid. In
+// particular, an incidental "URL" key must not change a permission title when
+// browser_open will only read "url".
+func exactJSONStringArg(rawArgs, key string) (string, bool) {
+	var args map[string]json.RawMessage
+	if json.Unmarshal([]byte(rawArgs), &args) != nil {
+		return "", false
+	}
+	raw, ok := args[key]
+	if !ok {
+		return "", false
+	}
+	var value string
+	if json.Unmarshal(raw, &value) != nil {
+		return "", false
+	}
+	return value, true
+}
+
 // primaryArgHint extracts the most relevant argument (path/pattern/command) from
 // raw JSON arguments. Best-effort; returns "" when it can't parse.
 func primaryArgHint(rawArgs string) string {
@@ -89,7 +204,7 @@ func rawInput(args string) json.RawMessage {
 // toolCallStart maps an advertised ZERO tool call to the initial ACP "tool_call"
 // update (status in_progress — ZERO executes immediately after advertising).
 func toolCallStart(call agent.ToolCall) ToolCallUpdate {
-	return ToolCallUpdate{
+	upd := ToolCallUpdate{
 		SessionUpdate: UpdateToolCall,
 		ToolCallID:    call.ID,
 		Title:         toolTitle(call.Name, call.Arguments),
@@ -97,6 +212,8 @@ func toolCallStart(call agent.ToolCall) ToolCallUpdate {
 		Status:        ToolStatusInProgress,
 		RawInput:      rawInput(call.Arguments),
 	}
+	attachBrowserToolDetails(&upd, call.Name)
+	return upd
 }
 
 // toolCallResult maps a finished ZERO tool result to a "tool_call_update".
@@ -116,6 +233,7 @@ func toolCallResult(result agent.ToolResult) ToolCallUpdate {
 	if locs := toolResultLocations(result); len(locs) > 0 {
 		upd.Locations = locs
 	}
+	attachBrowserToolDetails(&upd, result.Name)
 	return upd
 }
 

internal/acp/translate_test.go

@@ -1,14 +1,29 @@
 package acp
 
 import (
+	"encoding/json"
 	"strings"
 	"testing"
+	"unicode"
 	"unicode/utf8"
 
 	"github.com/Gitlawb/zero/internal/agent"
 	"github.com/Gitlawb/zero/internal/tools"
 )
 
+func browserDescriptor(t *testing.T, update ToolCallUpdate) BrowserToolDetails {
+	t.Helper()
+	raw, ok := update.Meta[zeroBrowserMetaKey]
+	if !ok {
+		t.Fatalf("browser metadata = %#v, want %q", update.Meta, zeroBrowserMetaKey)
+	}
+	var details BrowserToolDetails
+	if err := json.Unmarshal(raw, &details); err != nil {
+		t.Fatalf("decode browser metadata: %v", err)
+	}
+	return details
+}
+
 func TestAgentMessageAndThoughtChunks(t *testing.T) {
 	m := agentMessageChunk("hello")
 	if m.SessionUpdate != UpdateAgentMessageChunk || m.Content.Type != "text" || m.Content.Text != "hello" {
@@ -56,6 +71,217 @@ func TestToolTitleAndHint(t *testing.T) {
 	}
 }
 
+func TestBrowserToolUpdatesAreStructuredAndPresentationSafe(t *testing.T) {
+	start := toolCallStart(agent.ToolCall{
+		ID:        "browser-1",
+		Name:      "browser_open",
+		Arguments: `{"url":"https://example.com/settings?token=not-for-a-title#account"}`,
+	})
+	if got := browserDescriptor(t, start); got != (BrowserToolDetails{Version: 1, Command: "open"}) {
+		t.Fatalf("browser descriptor = %#v, want open", got)
+	}
+	if start.Title != "browser open https://example.com" {
+		t.Fatalf("browser title = %q", start.Title)
+	}
+	if strings.Contains(start.Title, "token=") || strings.Contains(start.Title, "#account") {
+		t.Fatalf("browser title leaked URL-sensitive data: %q", start.Title)
+	}
+	encoded, err := json.Marshal(start)
+	if err != nil {
+		t.Fatal(err)
+	}
+	var wire struct {
+		Meta map[string]json.RawMessage `json:"_meta"`
+	}
+	if err := json.Unmarshal(encoded, &wire); err != nil {
+		t.Fatal(err)
+	}
+	if _, ok := wire.Meta[zeroBrowserMetaKey]; !ok {
+		t.Fatalf("browser wire metadata = %#v", wire.Meta)
+	}
+
+	typed := toolCallStart(agent.ToolCall{
+		ID:        "browser-2",
+		Name:      "browser_type",
+		Arguments: `{"ref":"email","text":"[email protected]"}`,
+	})
+	if got := browserDescriptor(t, typed); got.Command != "type" {
+		t.Fatalf("browser type descriptor = %#v", got)
+	}
+	if typed.Title != "browser type" || strings.Contains(typed.Title, "[email protected]") {
+		t.Fatalf("browser type title = %q", typed.Title)
+	}
+
+	action := toolCallStart(agent.ToolCall{
+		ID:        "browser-3",
+		Name:      "browser_action",
+		Arguments: `{"command":"keyboard_insert_text","args":["[email protected]"]}`,
+	})
+	if action.Title != "browser action keyboard_insert_text" {
+		t.Fatalf("browser action title = %q", action.Title)
+	}
+
+	result := toolCallResult(agent.ToolResult{
+		ToolCallID: "browser-2",
+		Name:       "browser_type",
+		Status:     tools.StatusOK,
+	})
+	if got := browserDescriptor(t, result); got.Command != "type" {
+		t.Fatalf("browser result descriptor = %#v", got)
+	}
+}
+
+func TestBrowserDescriptorSurvivesProtocolShapedRoundTrip(t *testing.T) {
+	updates := []ToolCallUpdate{
+		toolCallStart(agent.ToolCall{
+			ID:        "start",
+			Name:      "browser_open",
+			Arguments: `{"url":"https://user:[email protected]/private?token=secret#fragment"}`,
+		}),
+		toolCallResult(agent.ToolResult{
+			ToolCallID: "result",
+			Name:       "browser_type",
+			Status:     tools.StatusOK,
+		}),
+		permissionToolCall(agent.PermissionRequest{
+			ToolCallID: "permission",
+			ToolName:   "browser_connect",
+			Args:       map[string]any{"target": "127.0.0.1:9222"},
+		}),
+	}
+
+	type protocolToolCallUpdate struct {
+		SessionUpdate string                     `json:"sessionUpdate,omitempty"`
+		ToolCallID    string                     `json:"toolCallId"`
+		Title         string                     `json:"title,omitempty"`
+		Kind          string                     `json:"kind,omitempty"`
+		Status        string                     `json:"status,omitempty"`
+		RawInput      json.RawMessage            `json:"rawInput,omitempty"`
+		Content       []ToolCallContent          `json:"content,omitempty"`
+		Locations     []ToolCallLocation         `json:"locations,omitempty"`
+		Meta          map[string]json.RawMessage `json:"_meta,omitempty"`
+	}
+
+	for _, update := range updates {
+		encoded, err := json.Marshal(update)
+		if err != nil {
+			t.Fatal(err)
+		}
+		var root map[string]json.RawMessage
+		if err := json.Unmarshal(encoded, &root); err != nil {
+			t.Fatal(err)
+		}
+		if _, ok := root["browser"]; ok {
+			t.Fatalf("browser descriptor escaped ACP _meta: %s", encoded)
+		}
+
+		var protocol protocolToolCallUpdate
+		if err := json.Unmarshal(encoded, &protocol); err != nil {
+			t.Fatal(err)
+		}
+		forwarded, err := json.Marshal(protocol)
+		if err != nil {
+			t.Fatal(err)
+		}
+		var roundTripped ToolCallUpdate
+		if err := json.Unmarshal(forwarded, &roundTripped); err != nil {
+			t.Fatal(err)
+		}
+		details := browserDescriptor(t, roundTripped)
+		if details.Version != 1 || details.Command == "" {
+			t.Fatalf("round-tripped browser descriptor = %#v", details)
+		}
+		descriptorJSON := string(roundTripped.Meta[zeroBrowserMetaKey])
+		for _, secret := range []string{"password", "private", "token", "fragment", "127.0.0.1", "9222"} {
+			if strings.Contains(descriptorJSON, secret) {
+				t.Fatalf("browser descriptor leaked %q: %s", secret, descriptorJSON)
+			}
+		}
+	}
+}
+
+func TestBrowserPermissionTitlesMirrorSafeToolArguments(t *testing.T) {
+	if got := browserToolTitle("open", `{"url":"evil.example.test/pay?token=hidden#fragment"}`); got != "browser open https://evil.example.test" {
+		t.Fatalf("bare-host title = %q", got)
+	}
+	if got := browserToolTitle("open", `{"URL":"https://different.example.test"}`); got != "browser open" {
+		t.Fatalf("case-variant URL title = %q", got)
+	}
+	if got := browserToolTitle("open", `{"URL":"https://different.example.test","url":"https://actual.example.test/path"}`); got != "browser open https://actual.example.test" {
+		t.Fatalf("exact URL key title = %q", got)
+	}
+	if got := browserToolTitle("action", `{"command":"not an action"}`); got != "browser action" {
+		t.Fatalf("unknown browser action title = %q", got)
+	}
+
+	longHost := "https://" + strings.Repeat("a", 200) + ".example.test/path?token=hidden"
+	title := browserToolTitle("open", `{"url":"`+longHost+`"}`)
+	if !utf8.ValidString(title) || utf8.RuneCountInString(title) > len("browser open ")+61 || strings.Contains(title, "token=") {
+		t.Fatalf("bounded browser origin title = %q", title)
+	}
+}
+
+func TestBrowserOpenTitlesRejectDecodedUnicodePresentationControls(t *testing.T) {
+	for _, rawURL := range []string{
+		"https://safe.example%E2%80%AEevil.test/path",
+		"https://safe.example%E2%81%A6evil.test/path",
+		"https://safe.example%C2%85evil.test/path",
+		"https://safe.example%E2%80%A8evil.test/path",
+		"https://safe.example%E2%80%A9evil.test/path",
+	} {
+		t.Run(rawURL, func(t *testing.T) {
+			normalized, err := tools.NormalizeBrowserOpenURL(rawURL)
+			if err != nil {
+				t.Fatalf("execution URL rejected: %v", err)
+			}
+			if normalized != rawURL {
+				t.Fatalf("execution URL = %q, want unchanged %q", normalized, rawURL)
+			}
+
+			args, err := json.Marshal(map[string]any{"url": rawURL})
+			if err != nil {
+				t.Fatal(err)
+			}
+			updates := []ToolCallUpdate{
+				toolCallStart(agent.ToolCall{ID: "start", Name: "browser_open", Arguments: string(args)}),
+				permissionToolCall(agent.PermissionRequest{ToolCallID: "permission", ToolName: "browser_open", Args: map[string]any{"url": rawURL}}),
+			}
+			for _, update := range updates {
+				encoded, err := json.Marshal(update)
+				if err != nil {
+					t.Fatal(err)
+				}
+				var decoded ToolCallUpdate
+				if err := json.Unmarshal(encoded, &decoded); err != nil {
+					t.Fatal(err)
+				}
+				if decoded.Title != "browser open" {
+					t.Fatalf("unsafe browser title survived wire round trip: %q", decoded.Title)
+				}
+				for _, r := range decoded.Title {
+					if unicode.IsControl(r) || unicode.Is(unicode.Cf, r) || unicode.Is(unicode.Zl, r) || unicode.Is(unicode.Zp, r) {
+						t.Fatalf("browser title contains unsafe presentation rune %U: %q", r, decoded.Title)
+					}
+				}
+			}
+		})
+	}
+}
+
+func TestBrowserDescriptorDoesNotClaimSimilarlyNamedMCPTools(t *testing.T) {
+	start := toolCallStart(agent.ToolCall{ID: "mcp-1", Name: "browser_plugin_open", Arguments: `{}`})
+	if len(start.Meta) != 0 {
+		t.Fatalf("MCP-like tool received built-in browser metadata: %#v", start.Meta)
+	}
+	encoded, err := json.Marshal(start)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if strings.Contains(string(encoded), `"browser"`) {
+		t.Fatalf("non-browser tool encoded browser field: %s", encoded)
+	}
+}
+
 func TestToolCallStart(t *testing.T) {
 	upd := toolCallStart(agent.ToolCall{ID: "tc1", Name: "read_file", Arguments: `{"path":"a.go"}`})
 	if upd.SessionUpdate != UpdateToolCall {

internal/acp/types.go

@@ -210,6 +210,21 @@ type ToolCallUpdate struct {
 	RawInput      json.RawMessage    `json:"rawInput,omitempty"`
 	Content       []ToolCallContent  `json:"content,omitempty"`
 	Locations     []ToolCallLocation `json:"locations,omitempty"`
+	// Meta is ACP's extension channel. ZERO-owned values must remain beneath a
+	// namespaced key so protocol-shaped clients can preserve them while decoding
+	// and re-encoding a tool call.
+	Meta map[string]json.RawMessage `json:"_meta,omitempty"`
+}
+
+// BrowserToolDetails identifies the browser helper operation behind a tool
+// call. Version is the schema version for this optional ZERO extension;
+// Command is one of install, launch, connect, open, snapshot, click, type,
+// press, or action. Future fields must remain display-safe and must not
+// include browser profile data, cookies, typed text, URL paths/queries, or
+// DevTools endpoints.
+type BrowserToolDetails struct {
+	Version int    `json:"version"`
+	Command string `json:"command"`
 }
 
 // ToolCallContent is a tool call's rendered output. ZERO emits "content" (a

internal/agent/command_prefix_test.go

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

internal/agent/compaction.go

@@ -4,7 +4,9 @@ import (
 	"context"
 	"encoding/json"
 	"errors"
+	"strconv"
 	"strings"
+	"unicode/utf8"
 
 	"github.com/Gitlawb/zero/internal/trace"
 	"github.com/Gitlawb/zero/internal/zeroruntime"
@@ -144,11 +146,17 @@ func estimateToolDefTokens(tools []zeroruntime.ToolDefinition) int {
 	for _, tool := range tools {
 		total += ApproxTextTokens(tool.Name)
 		total += ApproxTextTokens(tool.Description)
+		total += ApproxTextTokens(string(tool.Type))
 		if len(tool.Parameters) > 0 {
 			if encoded, err := json.Marshal(tool.Parameters); err == nil {
 				total += ApproxTextTokens(string(encoded))
 			}
 		}
+		if tool.Format != nil {
+			total += ApproxTextTokens(tool.Format.Type)
+			total += ApproxTextTokens(tool.Format.Syntax)
+			total += ApproxTextTokens(tool.Format.Definition)
+		}
 		total += 4 // per-tool overhead
 	}
 	return total
@@ -390,6 +398,27 @@ type compactionState struct {
 	// starts at 1.0 and converges via an EMA as each turn reports actual usage, so
 	// later turns compact nearer to real capacity. Zero is treated as 1.0.
 	calibrationRatio float64
+
+	// summarizerFactory lazily builds the dedicated (cheap) summarizer provider
+	// (Options.Summarizer) for the current main model. summarizerProvider
+	// memoizes the built provider and summarizerResolved records that the
+	// factory ran (it may legitimately return no provider: the main model is
+	// already the cheap one). summarizerBroken is the sticky fallback: once a
+	// build or call fails, the run stays on the main provider until the next
+	// model switch — a misconfigured cheap model must never break (or
+	// repeatedly delay) compaction.
+	summarizerFactory  func(ctx context.Context, mainModelID string) (Provider, error)
+	summarizerProvider Provider
+	summarizerResolved bool
+	summarizerBroken   bool
+	// window is the context window the threshold was derived from. The
+	// reactive free-prune stage compares against it (the limit the provider
+	// enforces), not against the 70% proactive threshold — a pruned history
+	// between the two would still succeed on retry.
+	window int
+	// mainModel is the model the run currently streams on; the summarizer
+	// choice is made against it and remade after a mid-run switch.
+	mainModel string
 }
 
 // calibrate folds one turn's (rawEstimate, actualPromptTokens) sample into the
@@ -424,17 +453,88 @@ func (state *compactionState) calibratedTokens(raw int) int {
 
 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,
-		planner:      newContextPlanner(contextPlannerConfig{contextWindow: options.ContextWindow}),
-		trace:        options.Trace,
+		enabled:           options.ContextWindow > 0,
+		threshold:         compactionThreshold(options.ContextWindow),
+		window:            options.ContextWindow,
+		mainModel:         options.Model,
+		preserveLast:      options.CompactionPreserveLast,
+		onUsage:           options.OnUsage,
+		task:              task,
+		planner:           newContextPlanner(contextPlannerConfig{contextWindow: options.ContextWindow, serviceTier: options.ServiceTier}),
+		trace:             options.Trace,
+		summarizerFactory: options.Summarizer,
 	}
 	return state
 }
 
+// updateWindow re-derives the compaction threshold after a mid-run model
+// switch. The low-water mark is reset: it was measured against the old
+// threshold, and keeping it could suppress a compaction the smaller window
+// now requires. <= 0 (unknown model) leaves everything unchanged.
+func (state *compactionState) updateWindow(window int) {
+	if window <= 0 {
+		return
+	}
+	state.enabled = true
+	state.window = window
+	state.threshold = compactionThreshold(window)
+	state.lowWaterMark = 0
+	if state.planner != nil {
+		state.planner.config.contextWindow = window
+	}
+}
+
+// switchModel records a mid-run model switch: the window is re-derived
+// (window <= 0 keeps the old one) and the summarizer choice is forgotten so
+// the next compaction re-resolves it against the new main model. A run that
+// started on the cheap model (no dedicated summarizer) and escalated to an
+// expensive one therefore regains cheap summaries — the exact workflow a
+// dedicated summarizer exists for.
+func (state *compactionState) switchModel(modelID string, window int) {
+	state.mainModel = modelID
+	state.updateWindow(window)
+	state.summarizerProvider = nil
+	state.summarizerResolved = false
+	state.summarizerBroken = false
+}
+
+// summarizeProvider returns the provider summarization calls should use and
+// whether it is the dedicated summarizer: that one when it applies to the
+// current main model and is healthy, else the main provider. The factory runs
+// at most once per main model; a build error marks the summarizer broken for
+// that model. Callers branch on the boolean rather than comparing Provider
+// values — interface comparison panics on a non-comparable dynamic type.
+func (state *compactionState) summarizeProvider(ctx context.Context, main Provider) (Provider, bool) {
+	if state.summarizerFactory == nil || state.summarizerBroken {
+		return main, false
+	}
+	if !state.summarizerResolved {
+		state.summarizerResolved = true
+		built, err := state.summarizerFactory(ctx, state.mainModel)
+		if err != nil {
+			state.summarizerBroken = true
+			return main, false
+		}
+		state.summarizerProvider = built
+	}
+	if state.summarizerProvider == nil {
+		return main, false
+	}
+	return state.summarizerProvider, true
+}
+
+// fitsWindow reports whether a calibrated size is safely under the limit the
+// provider enforces. The proactive threshold (70% of the window) is a trigger
+// for the paid summarizer, not the hard limit; the reactive path compares
+// against the window itself, with a small margin for estimate error, so a
+// history pruned to, say, 85% of the window takes the free retry.
+func (state *compactionState) fitsWindow(size int) bool {
+	if state.window <= 0 {
+		return size <= state.threshold
+	}
+	return size <= state.window-state.window/20
+}
+
 // maybeCompact runs proactive compaction at the top of a turn. It returns the
 // (possibly compacted) message slice. It is a no-op when compaction is disabled,
 // when the history is under threshold, or when the history has not grown past
@@ -528,10 +628,34 @@ func (state *compactionState) recover(
 		// context-limit error never triggers an unexpected summarization call.
 		return messages, false, nil
 	}
-	if state.reactiveAttempted {
+	if !isContextLimitError(errorMessage) {
 		return messages, false, nil
 	}
-	if !isContextLimitError(errorMessage) {
+
+	// CHEAP FIRST STAGE (mirrors maybeCompact): a context-limit failure can
+	// often be fixed for free by dropping superseded reads and pruning old
+	// tool-result bodies. It runs regardless of the one-shot reactive budget
+	// (it costs nothing and cannot loop: a second pass finds nothing left to
+	// prune). The free retry is taken only when the pruned history fits the
+	// window the provider enforces — otherwise a retry would fail the same
+	// way, so fall through to the paid summarizer in this same call.
+	toolTokens := estimateToolDefTokens(tools)
+	prunedAny := false
+	if pruned, reclaimed := pruneSupersededReadResults(messages); reclaimed > 0 {
+		messages, prunedAny = pruned, true
+	}
+	if pruned, reclaimed := pruneStaleToolOutput(messages, state.preserveLast); reclaimed > 0 {
+		messages, prunedAny = pruned, true
+	}
+	if prunedAny {
+		size := state.calibratedTokens(estimateTokens(messages) + toolTokens)
+		if state.fitsWindow(size) {
+			state.lowWaterMark = size
+			return messages, true, nil
+		}
+	}
+
+	if state.reactiveAttempted {
 		return messages, false, nil
 	}
 
@@ -571,11 +695,21 @@ func (state *compactionState) recover(
 // provider call. The summary stream intentionally does NOT forward OnText (so
 // compaction stays invisible on the user-facing surface), but it DOES forward
 // OnUsage so the summarizer's token cost is still counted by usage/budgeting.
+// The call goes to the dedicated (cheap) summarizer when one applies; if that
+// call fails the summarizer is marked broken for the current model and the
+// same slice is retried once on the main provider, so the caller sees at most
+// one (main-provider-quality) error.
 func (state *compactionState) summarizeClosure(ctx context.Context, provider Provider) func([]zeroruntime.Message) (string, error) {
 	if state.planner == nil {
 		state.planner = newContextPlanner(contextPlannerConfig{})
 	}
 	return func(toSummarize []zeroruntime.Message) (string, error) {
+		chosen, dedicated := state.summarizeProvider(ctx, provider)
+		summary, err := summarizeWithPlanner(ctx, chosen, toSummarize, state.onUsage, state.planner, state.trace)
+		if err == nil || !dedicated {
+			return summary, err
+		}
+		state.summarizerBroken = true
 		return summarizeWithPlanner(ctx, provider, toSummarize, state.onUsage, state.planner, state.trace)
 	}
 }
@@ -659,8 +793,41 @@ func summarizeMessagesOnce(ctx context.Context, provider Provider, messages []ze
 	return summary, nil
 }
 
+// summaryToolResultClamp bounds each tool-result body in the summarizer input.
+// Tool outputs dominate the elided middle's bulk (build logs, file reads,
+// search results), yet the summary instructions only need their gist — the
+// model that ACTED on the full output already turned it into decisions the
+// surrounding messages record. 2,000 chars (~500 tokens) keeps the useful
+// head+tail of a result while cutting the summarizer input by 5-10x on
+// tool-heavy sessions; the summarizer's own cost is the price being cut.
+const summaryToolResultClamp = 2000
+
+// summaryToolArgsClamp bounds each tool-call argument blob in the transcript:
+// a write_file call carries the whole file body in its arguments.
+const summaryToolArgsClamp = 500
+
+// clampForSummary head+tail-truncates text for the summarizer transcript,
+// snapping to rune boundaries so the clamp never splits a UTF-8 sequence.
+func clampForSummary(text string, max int) string {
+	if max <= 0 || len(text) <= max {
+		return text
+	}
+	head := max * 3 / 4
+	tail := max - head
+	for head > 0 && !utf8.RuneStart(text[head]) {
+		head--
+	}
+	start := len(text) - tail
+	for start < len(text) && !utf8.RuneStart(text[start]) {
+		start++
+	}
+	return text[:head] + "\n…[" + strconv.Itoa(len(text)-head-(len(text)-start)) + " chars omitted for summarization]…\n" + text[start:]
+}
+
 // renderTranscript flattens messages into a plain-text transcript for the
 // summarizer. Secret scrubbing already happened upstream at the tool boundary.
+// Tool-result bodies and tool-call arguments are clamped (see the constants
+// above): the summary needs what was done and decided, not the raw payloads.
 func renderTranscript(messages []zeroruntime.Message) string {
 	lines := make([]string, 0, len(messages))
 	for _, message := range messages {
@@ -670,13 +837,13 @@ func renderTranscript(messages []zeroruntime.Message) string {
 			if len(message.ToolCalls) > 0 {
 				calls := make([]string, 0, len(message.ToolCalls))
 				for _, call := range message.ToolCalls {
-					calls = append(calls, call.Name+"("+call.Arguments+")")
+					calls = append(calls, call.Name+"("+clampForSummary(call.Arguments, summaryToolArgsClamp)+")")
 				}
 				line += "\n[tool calls: " + strings.Join(calls, "; ") + "]"
 			}
 			lines = append(lines, line)
 		case zeroruntime.MessageRoleTool:
-			lines = append(lines, "tool result: "+message.Content)
+			lines = append(lines, "tool result: "+clampForSummary(message.Content, summaryToolResultClamp))
 		default:
 			lines = append(lines, string(message.Role)+": "+message.Content)
 		}

internal/agent/compaction_cheap_test.go

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

internal/agent/compaction_projection.go

@@ -126,8 +126,11 @@ func capCompactionBrief(brief string) (string, bool) {
 
 func compactionToolCallLine(call zeroruntime.ToolCall) string {
 	detail := ""
+	if call.Freeform && call.Name == "apply_patch" {
+		detail = strings.Join(patchPaths(call.Arguments), ", ")
+	}
 	var arguments map[string]any
-	if json.Unmarshal([]byte(strings.TrimSpace(call.Arguments)), &arguments) == nil {
+	if detail == "" && json.Unmarshal([]byte(strings.TrimSpace(call.Arguments)), &arguments) == nil {
 		for _, key := range []string{"file_path", "path", "url", "query", "pattern", "prompt", "description", "question"} {
 			if value, ok := arguments[key].(string); ok && strings.TrimSpace(value) != "" {
 				detail = value

internal/agent/compaction_projection_test.go

@@ -73,6 +73,7 @@ func TestProjectCompactionInputRetainsAskUserExchangeAndSemanticArguments(t *tes
 		{Role: zeroruntime.MessageRoleTool, ToolCallID: "ask", Content: "Which database should remain?: Postgres only; do not touch MySQL."},
 		{Role: zeroruntime.MessageRoleAssistant, ToolCalls: []zeroruntime.ToolCall{
 			{ID: "patch", Name: "apply_patch", Arguments: `{"patch":"*** Begin Patch\n*** Update File: internal/db.go\n*** End Patch"}`},
+			{ID: "raw-patch", Name: "apply_patch", Freeform: true, Arguments: "*** Begin Patch\n*** Update File: internal/api.go\n*** Add File: internal/config.go\n*** End Patch\n"},
 			{ID: "fetch", Name: "web_fetch", Arguments: `{"url":"https://example.com/spec"}`},
 			{ID: "task", Name: "task", Arguments: `{"prompt":"Inspect the Windows implementation"}`},
 		}},
@@ -80,7 +81,7 @@ func TestProjectCompactionInputRetainsAskUserExchangeAndSemanticArguments(t *tes
 	}
 
 	content := projectCompactionInput(messages).messages[0].Content
-	for _, want := range []string{"Which database should remain?", "Postgres only; do not touch MySQL", "internal/db.go", "https://example.com/spec", "Inspect the Windows implementation"} {
+	for _, want := range []string{"Which database should remain?", "Postgres only; do not touch MySQL", "internal/db.go", "internal/api.go", "internal/config.go", "https://example.com/spec", "Inspect the Windows implementation"} {
 		if !strings.Contains(content, want) {
 			t.Fatalf("projection missing %q: %q", want, content)
 		}

internal/agent/compaction_test.go

@@ -275,6 +275,27 @@ func TestEstimateToolDefTokensCountsDefinitions(t *testing.T) {
 	if estimateToolDefTokens(one) <= 0 {
 		t.Fatal("a tool definition (name + description + schema) should estimate > 0 tokens")
 	}
+	typed := append([]zeroruntime.ToolDefinition(nil), one...)
+	typed[0].Type = zeroruntime.ToolDefinitionFreeform
+	if estimateToolDefTokens(typed) <= estimateToolDefTokens(one) {
+		t.Fatal("a non-empty tool type must increase the estimated request cost")
+	}
+	for _, test := range []struct {
+		name   string
+		format zeroruntime.ToolDefinitionFormat
+	}{
+		{name: "type", format: zeroruntime.ToolDefinitionFormat{Type: "grammar"}},
+		{name: "syntax", format: zeroruntime.ToolDefinitionFormat{Syntax: "lark"}},
+		{name: "definition", format: zeroruntime.ToolDefinitionFormat{Definition: "start: WORD"}},
+	} {
+		t.Run("format "+test.name, func(t *testing.T) {
+			formatted := append([]zeroruntime.ToolDefinition(nil), one...)
+			formatted[0].Format = &test.format
+			if estimateToolDefTokens(formatted) <= estimateToolDefTokens(one) {
+				t.Fatalf("a non-empty format %s must increase the estimated request cost", test.name)
+			}
+		})
+	}
 	two := append(append([]zeroruntime.ToolDefinition{}, one...), zeroruntime.ToolDefinition{
 		Name:        "write_file",
 		Description: "Write contents to a file in the workspace.",

internal/agent/context_planner.go

@@ -9,6 +9,7 @@ import (
 type contextPlannerConfig struct {
 	contextWindow  int
 	promptCacheKey string
+	serviceTier    string
 	promptParts    systemPromptParts
 }
 
@@ -62,6 +63,7 @@ func (planner *contextPlanner) plan(messages []zeroruntime.Message, toolDefs []z
 		Messages:        requestMessages,
 		Tools:           planner.snapshotTools(toolDefs, fingerprint),
 		ReasoningEffort: reasoningEffort,
+		ServiceTier:     planner.config.serviceTier,
 		PromptCacheKey:  planner.config.promptCacheKey,
 	}
 	breakdown := MeasureContext(request.Messages, request.Tools, planner.config.contextWindow)
@@ -100,6 +102,10 @@ func copyToolDefinitions(toolDefs []zeroruntime.ToolDefinition) []zeroruntime.To
 	for index, definition := range toolDefs {
 		copied[index] = definition
 		copied[index].Parameters = copySchemaMap(definition.Parameters)
+		if definition.Format != nil {
+			format := *definition.Format
+			copied[index].Format = &format
+		}
 	}
 	return copied
 }

internal/agent/context_planner_test.go

@@ -15,6 +15,7 @@ func TestContextPlannerPreservesProviderRequest(t *testing.T) {
 	}
 	toolDefs := []zeroruntime.ToolDefinition{{
 		Name: "read_file", Description: "Read a file",
+		Format: &zeroruntime.ToolDefinitionFormat{Type: "grammar", Syntax: "lark", Definition: "start: PATH"},
 		Parameters: map[string]any{
 			"type":       "object",
 			"required":   []string{"path"},
@@ -25,6 +26,7 @@ func TestContextPlannerPreservesProviderRequest(t *testing.T) {
 	planner := newContextPlanner(contextPlannerConfig{
 		contextWindow:  128_000,
 		promptCacheKey: "session-1",
+		serviceTier:    "priority",
 		promptParts:    systemPromptParts{prompt: "system", baseInstructions: "system"},
 	})
 
@@ -33,6 +35,7 @@ func TestContextPlannerPreservesProviderRequest(t *testing.T) {
 		Messages:        copyMessages(messages),
 		Tools:           toolDefs,
 		ReasoningEffort: "medium",
+		ServiceTier:     "priority",
 		PromptCacheKey:  "session-1",
 	}
 	if !reflect.DeepEqual(plan.Request, want) {
@@ -51,10 +54,12 @@ func TestContextPlannerPreservesProviderRequest(t *testing.T) {
 	toolDefs[0].Parameters["required"].([]string)[0] = "other"
 	toolDefs[0].Parameters["properties"].(map[string]any)["path"].(map[string]any)["type"] = "integer"
 	toolDefs[0].Parameters["anyOf"].([]any)[0].(map[string]any)["type"] = "array"
+	toolDefs[0].Format.Definition = "changed"
 	if plan.Request.Messages[1].Content != "inspect this" || plan.Request.Messages[1].Images[0].Data[0] != 1 {
 		t.Fatalf("planned messages alias caller state: %#v", plan.Request.Messages[1])
 	}
 	if plan.Request.Tools[0].Name != "read_file" ||
+		plan.Request.Tools[0].Format == nil || plan.Request.Tools[0].Format.Definition != "start: PATH" ||
 		plan.Request.Tools[0].Parameters["required"].([]string)[0] != "path" ||
 		plan.Request.Tools[0].Parameters["properties"].(map[string]any)["path"].(map[string]any)["type"] != "string" ||
 		plan.Request.Tools[0].Parameters["anyOf"].([]any)[0].(map[string]any)["type"] != "object" {

internal/agent/escalate_loop_test.go

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

internal/agent/freeform_tool_test.go

@@ -0,0 +1,378 @@
+package agent
+
+import (
+	"context"
+	"os"
+	"path/filepath"
+	"runtime"
+	"strings"
+	"testing"
+
+	"github.com/Gitlawb/zero/internal/sandbox"
+	"github.com/Gitlawb/zero/internal/tools"
+	"github.com/Gitlawb/zero/internal/zeroruntime"
+)
+
+type customApplyPatchTool struct {
+	calls int
+}
+
+type fixedPathScope struct {
+	roots []string
+}
+
+func (scope fixedPathScope) Roots() []string {
+	return append([]string(nil), scope.roots...)
+}
+
+func newFixedPathScope(t *testing.T, roots ...string) fixedPathScope {
+	t.Helper()
+	resolved := make([]string, 0, len(roots))
+	for _, root := range roots {
+		absolute, err := filepath.Abs(root)
+		if err != nil {
+			t.Fatalf("Abs(%q): %v", root, err)
+		}
+		if physical, err := filepath.EvalSymlinks(absolute); err == nil {
+			absolute = physical
+		}
+		resolved = append(resolved, absolute)
+	}
+	return fixedPathScope{roots: resolved}
+}
+
+func (*customApplyPatchTool) Name() string        { return "apply_patch" }
+func (*customApplyPatchTool) Description() string { return "custom JSON tool" }
+func (*customApplyPatchTool) Parameters() tools.Schema {
+	return tools.Schema{Type: "object", Properties: map[string]tools.PropertySchema{"value": {Type: "string"}}}
+}
+func (*customApplyPatchTool) Safety() tools.Safety {
+	return tools.Safety{Permission: tools.PermissionAllow, SideEffect: tools.SideEffectRead}
+}
+func (tool *customApplyPatchTool) Run(context.Context, map[string]any) tools.Result {
+	tool.calls++
+	return tools.Result{Status: tools.StatusOK}
+}
+
+func TestFreeformApplyPatchUsesExistingToolExecutionPath(t *testing.T) {
+	root := t.TempDir()
+	registry := tools.NewRegistry()
+	registry.Register(tools.NewScopedApplyPatchTool(root, nil))
+	patch := "*** Begin Patch\n*** Add File: hello.txt\n+hello\n*** End Patch\n"
+
+	result, err := executeToolCall(context.Background(), registry, zeroruntime.ToolCall{
+		ID:        "call-1",
+		Name:      "apply_patch",
+		Arguments: patch,
+		Freeform:  true,
+	}, PermissionModeUnsafe, Options{Cwd: root, FileTracker: tools.NewFileTracker()})
+	if err != nil {
+		t.Fatalf("executeToolCall: %v", err)
+	}
+	if result.Status != tools.StatusOK {
+		t.Fatalf("freeform apply_patch status = %s: %s", result.Status, result.Output)
+	}
+	content, err := os.ReadFile(filepath.Join(root, "hello.txt"))
+	if err != nil {
+		t.Fatalf("read patched file: %v", err)
+	}
+	if string(content) != "hello\n" {
+		t.Fatalf("patched content = %q, want hello newline", content)
+	}
+}
+
+func TestFreeformApplyPatchSupportsGrantedExtraRoot(t *testing.T) {
+	root := t.TempDir()
+	extra := t.TempDir()
+	scope, err := sandbox.NewScope(root, []string{extra})
+	if err != nil {
+		t.Fatalf("NewScope: %v", err)
+	}
+	registry := tools.NewRegistry()
+	registry.Register(tools.NewScopedApplyPatchTool(root, scope))
+	target := filepath.Join(extra, "hello.txt")
+	patch := "*** Begin Patch\n*** Add File: " + target + "\n+hello\n*** End Patch\n"
+
+	result, err := executeToolCall(context.Background(), registry, zeroruntime.ToolCall{
+		ID: "call-1", Name: "apply_patch", Arguments: patch, Freeform: true,
+	}, PermissionModeUnsafe, Options{Cwd: root, FileTracker: tools.NewFileTracker()})
+	if err != nil {
+		t.Fatalf("executeToolCall: %v", err)
+	}
+	if result.Status != tools.StatusOK {
+		t.Fatalf("freeform extra-root apply_patch status = %s: %s", result.Status, result.Output)
+	}
+	content, err := os.ReadFile(target)
+	if err != nil {
+		t.Fatalf("read extra-root patched file: %v", err)
+	}
+	if string(content) != "hello\n" {
+		t.Fatalf("patched content = %q, want hello newline", content)
+	}
+}
+
+func TestFreeformApplyPatchRejectsMixedWorkspaceAndExtraRoots(t *testing.T) {
+	for _, extraHasMatchingPath := range []bool{false, true} {
+		name := "extra target missing"
+		if extraHasMatchingPath {
+			name = "extra target exists"
+		}
+		t.Run(name, func(t *testing.T) {
+			workspace := t.TempDir()
+			extra := t.TempDir()
+			if err := os.MkdirAll(filepath.Join(workspace, "src"), 0o755); err != nil {
+				t.Fatal(err)
+			}
+			workspaceTarget := filepath.Join(workspace, "src", "a.go")
+			if err := os.WriteFile(workspaceTarget, []byte("old\n"), 0o644); err != nil {
+				t.Fatal(err)
+			}
+			extraTarget := filepath.Join(extra, "src", "a.go")
+			if extraHasMatchingPath {
+				if err := os.MkdirAll(filepath.Dir(extraTarget), 0o755); err != nil {
+					t.Fatal(err)
+				}
+				if err := os.WriteFile(extraTarget, []byte("old\n"), 0o644); err != nil {
+					t.Fatal(err)
+				}
+			}
+			scope, err := sandbox.NewScope(workspace, []string{extra})
+			if err != nil {
+				t.Fatal(err)
+			}
+			registry := tools.NewRegistry()
+			registry.Register(tools.NewScopedApplyPatchTool(workspace, scope))
+			extraCreated := filepath.Join(extra, "b.go")
+			patch := strings.Join([]string{
+				"*** Begin Patch",
+				"*** Update File: src/a.go",
+				"@@",
+				"-old",
+				"+new",
+				"*** Add File: " + extraCreated,
+				"+extra",
+				"*** End Patch",
+			}, "\n")
+
+			result, err := executeToolCall(context.Background(), registry, zeroruntime.ToolCall{
+				ID: "call-1", Name: "apply_patch", Arguments: patch, Freeform: true,
+			}, PermissionModeUnsafe, Options{Cwd: workspace, FileTracker: tools.NewFileTracker()})
+			if err != nil {
+				t.Fatalf("executeToolCall: %v", err)
+			}
+			if result.Status != tools.StatusError || !strings.Contains(result.Output, "mixes workspace-relative paths with an extra write root") {
+				t.Fatalf("mixed-root patch result = %s: %s", result.Status, result.Output)
+			}
+			if content, err := os.ReadFile(workspaceTarget); err != nil || string(content) != "old\n" {
+				t.Fatalf("workspace target changed: content=%q err=%v", content, err)
+			}
+			if extraHasMatchingPath {
+				if content, err := os.ReadFile(extraTarget); err != nil || string(content) != "old\n" {
+					t.Fatalf("extra-root target changed: content=%q err=%v", content, err)
+				}
+			}
+			if _, err := os.Stat(extraCreated); !os.IsNotExist(err) {
+				t.Fatalf("mixed-root patch created extra file: %v", err)
+			}
+		})
+	}
+}
+
+func TestFreeformApplyPatchAcceptsOneSemanticRoot(t *testing.T) {
+	for _, test := range []struct {
+		name         string
+		useExtraRoot bool
+	}{
+		{name: "workspace"},
+		{name: "extra root", useExtraRoot: true},
+	} {
+		t.Run(test.name, func(t *testing.T) {
+			workspace := t.TempDir()
+			extra := t.TempDir()
+			scope, err := sandbox.NewScope(workspace, []string{extra})
+			if err != nil {
+				t.Fatal(err)
+			}
+			root := workspace
+			if test.useExtraRoot {
+				root = extra
+			}
+			if err := os.MkdirAll(filepath.Join(root, "src"), 0o755); err != nil {
+				t.Fatal(err)
+			}
+			target := filepath.Join(root, "src", "a.go")
+			if err := os.WriteFile(target, []byte("old\n"), 0o644); err != nil {
+				t.Fatal(err)
+			}
+			updatePath := target
+			if !test.useExtraRoot {
+				updatePath = filepath.Join("src", "a.go")
+			}
+			created := filepath.Join(root, "b.go")
+			patch := strings.Join([]string{
+				"*** Begin Patch",
+				"*** Update File: " + updatePath,
+				"@@",
+				"-old",
+				"+new",
+				"*** Add File: " + created,
+				"+created",
+				"*** End Patch",
+			}, "\n")
+			registry := tools.NewRegistry()
+			registry.Register(tools.NewScopedApplyPatchTool(workspace, scope))
+			result, err := executeToolCall(context.Background(), registry, zeroruntime.ToolCall{
+				ID: "call-1", Name: "apply_patch", Arguments: patch, Freeform: true,
+			}, PermissionModeUnsafe, Options{Cwd: workspace, FileTracker: tools.NewFileTracker()})
+			if err != nil {
+				t.Fatalf("executeToolCall: %v", err)
+			}
+			if result.Status != tools.StatusOK {
+				t.Fatalf("single-root patch status = %s: %s", result.Status, result.Output)
+			}
+			if content, err := os.ReadFile(target); err != nil || string(content) != "new\n" {
+				t.Fatalf("updated content=%q err=%v", content, err)
+			}
+			if content, err := os.ReadFile(created); err != nil || string(content) != "created\n" {
+				t.Fatalf("created content=%q err=%v", content, err)
+			}
+		})
+	}
+}
+
+func TestFreeformApplyPatchRejectsAbsolutePathOutsideGrantedRoots(t *testing.T) {
+	root := t.TempDir()
+	granted := t.TempDir()
+	outside := t.TempDir()
+	scope := newFixedPathScope(t, root, granted)
+	registry := tools.NewRegistry()
+	registry.Register(tools.NewScopedApplyPatchTool(root, scope))
+	target := filepath.Join(outside, "blocked.txt")
+	patch := "*** Begin Patch\n*** Add File: " + target + "\n+blocked\n*** End Patch\n"
+
+	result, err := executeToolCall(context.Background(), registry, zeroruntime.ToolCall{
+		ID: "call-1", Name: "apply_patch", Arguments: patch, Freeform: true,
+	}, PermissionModeUnsafe, Options{Cwd: root})
+	if err != nil {
+		t.Fatalf("executeToolCall: %v", err)
+	}
+	if result.Status != tools.StatusError {
+		t.Fatalf("ungranted absolute patch status = %s, want error: %s", result.Status, result.Output)
+	}
+	if _, err := os.Stat(target); !os.IsNotExist(err) {
+		t.Fatalf("ungranted absolute patch created target: %v", err)
+	}
+}
+
+func TestPreparedFreeformPatchRejectsIntermediateSymlinkSwap(t *testing.T) {
+	if runtime.GOOS == "windows" {
+		t.Skip("creating directory symlinks requires privileges on Windows")
+	}
+	root := t.TempDir()
+	extra := t.TempDir()
+	outside := t.TempDir()
+	intermediate := filepath.Join(extra, "nested")
+	if err := os.Mkdir(intermediate, 0o755); err != nil {
+		t.Fatalf("Mkdir: %v", err)
+	}
+	scope := newFixedPathScope(t, root, extra)
+	tool := tools.NewScopedApplyPatchTool(root, scope)
+	target := filepath.Join(intermediate, "blocked.txt")
+	patch := "*** Begin Patch\n*** Add File: " + target + "\n+blocked\n*** End Patch\n"
+	args, err := tools.PrepareFreeformApplyPatchArguments(tool, patch)
+	if err != nil {
+		t.Fatalf("PrepareFreeformApplyPatchArguments: %v", err)
+	}
+
+	if err := os.Remove(intermediate); err != nil {
+		t.Fatalf("Remove intermediate: %v", err)
+	}
+	if err := os.Symlink(outside, intermediate); err != nil {
+		t.Fatalf("Symlink: %v", err)
+	}
+	result := tool.Run(context.Background(), args)
+	if result.Status != tools.StatusError {
+		t.Fatalf("symlink-swapped patch status = %s, want error: %s", result.Status, result.Output)
+	}
+	if _, err := os.Stat(filepath.Join(outside, "blocked.txt")); !os.IsNotExist(err) {
+		t.Fatalf("symlink-swapped patch wrote outside root: %v", err)
+	}
+}
+
+func TestUnknownFreeformToolFailsClosed(t *testing.T) {
+	result, err := executeToolCall(context.Background(), tools.NewRegistry(), zeroruntime.ToolCall{
+		ID: "call-1", Name: "unknown", Arguments: "raw", Freeform: true,
+	}, PermissionModeUnsafe, Options{})
+	if err != nil {
+		t.Fatalf("executeToolCall: %v", err)
+	}
+	if result.Status != tools.StatusError {
+		t.Fatalf("unknown freeform status = %s, want error", result.Status)
+	}
+}
+
+func TestCustomApplyPatchCannotUseFreeformContract(t *testing.T) {
+	custom := &customApplyPatchTool{}
+	registry := tools.NewRegistry()
+	registry.Register(custom)
+
+	result, err := executeToolCall(context.Background(), registry, zeroruntime.ToolCall{
+		ID: "call-1", Name: "apply_patch", Arguments: "raw patch", Freeform: true,
+	}, PermissionModeUnsafe, Options{})
+	if err != nil {
+		t.Fatalf("executeToolCall: %v", err)
+	}
+	if result.Status != tools.StatusError || !strings.Contains(result.Output, "Unsupported freeform tool call") {
+		t.Fatalf("custom freeform apply_patch result = %#v, want unsupported error", result)
+	}
+	if custom.calls != 0 {
+		t.Fatalf("custom apply_patch executed %d times", custom.calls)
+	}
+}
+
+func TestDeniedFreeformApplyPatchDoesNotWrite(t *testing.T) {
+	root := t.TempDir()
+	registry := tools.NewRegistry()
+	registry.Register(tools.NewScopedApplyPatchTool(root, nil))
+	patch := "*** Begin Patch\n*** Add File: denied.txt\n+must not exist\n*** End Patch\n"
+
+	result, err := executeToolCall(context.Background(), registry, zeroruntime.ToolCall{
+		ID:        "call-1",
+		Name:      "apply_patch",
+		Arguments: patch,
+		Freeform:  true,
+	}, PermissionModeAsk, Options{
+		Cwd: root,
+		OnPermissionRequest: func(context.Context, PermissionRequest) (PermissionDecision, error) {
+			return PermissionDecision{Action: PermissionDecisionDeny, Reason: "test denial"}, nil
+		},
+	})
+	if err != nil {
+		t.Fatalf("executeToolCall: %v", err)
+	}
+	if result.Status != tools.StatusError {
+		t.Fatalf("denied freeform status = %s, want error", result.Status)
+	}
+	if _, err := os.Stat(filepath.Join(root, "denied.txt")); !os.IsNotExist(err) {
+		t.Fatalf("denied freeform patch changed the workspace: %v", err)
+	}
+}
+
+func TestHistorySafeToolCallsPreservesRawFreeformArguments(t *testing.T) {
+	raw := "*** Begin Patch\n*** Update File: internal/main.go\n@@\n-old\n+new\n*** End Patch\n"
+	safe := historySafeToolCalls([]ToolCall{{
+		ID: "call-1", ProviderCallID: "provider-1", Name: "apply_patch", Arguments: raw, Freeform: true,
+	}})
+	if len(safe) != 1 || safe[0].Arguments != raw || safe[0].ProviderCallID != "provider-1" || !safe[0].Freeform {
+		t.Fatalf("freeform history changed raw call: %#v", safe)
+	}
+}
+
+func TestAbortedToolResultPreservesProviderCallID(t *testing.T) {
+	messages := appendAbortedToolResults(nil, []ToolCall{{
+		ID: "call-1", ProviderCallID: "provider-1", Name: "read_file",
+	}})
+	if len(messages) != 1 || messages[0].ToolCallID != "call-1" || messages[0].ToolCallProviderID != "provider-1" || !messages[0].IsError {
+		t.Fatalf("aborted tool result lost provider identity: %#v", messages)
+	}
+}

internal/agent/guardrails.go

@@ -33,11 +33,11 @@ const (
 	// knows before spending more tool turns.
 	toolOnlyProgressReminderAt = 6
 
-	// planReminderTurn is the turn (1-based) by the end of which a multi-step
-	// task should have called update_plan; if it hasn't, a one-time reminder is
-	// injected. Set to 3 (not 2) so short, legitimate two-step tasks finish
-	// without a spurious planning nag.
-	planReminderTurn = 3
+	// planReminderTurn and planReminderToolThreshold keep bounded implementation
+	// work plan-free while still nudging a genuinely sustained run that has not
+	// established any structured progress. Both thresholds must be met.
+	planReminderTurn          = 7
+	planReminderToolThreshold = 7
 
 	// planToolName is the planning tool the loop watches for by name.
 	planToolName = "update_plan"
@@ -591,13 +591,12 @@ func (state *guardState) planReminder(turn int) string {
 		return planStaleReminder(state.toolCallsSincePlanUpdate)
 	}
 
-	// NOT-CALLED reminder: by the end of planReminderTurn the model should have
-	// called update_plan if it's doing a multi-step task (>=1 other tool call).
-	// One-shot for the whole run.
+	// NOT-CALLED reminder: once both the turn and tool-call thresholds are met,
+	// sustained work should have called update_plan. One-shot for the whole run.
 	if !state.notCalledReminderSent &&
 		!state.planEverCalled &&
 		turn >= planReminderTurn &&
-		state.totalToolCalls >= 1 {
+		state.totalToolCalls >= planReminderToolThreshold {
 		state.notCalledReminderSent = true
 		return planNotCalledReminder()
 	}

internal/agent/guardrails_test.go

@@ -257,19 +257,16 @@ func TestRunDoesNotCountDroppedToolCallTurnsAsEmpty(t *testing.T) {
 }
 
 func TestRunInjectsPlanNotCalledReminderForMultiStepTask(t *testing.T) {
+	const expectedThreshold = 7
+	if planReminderToolThreshold != expectedThreshold {
+		t.Fatalf("plan reminder tool threshold = %d, want %d", planReminderToolThreshold, expectedThreshold)
+	}
 	root := t.TempDir()
 	writeAgentTestFile(t, root+"/notes.txt", "alpha")
 	registry := tools.NewRegistry()
 	registry.Register(tools.NewScopedReadFileTool(root, nil))
 
-	provider := &mockProvider{
-		turns: [][]zeroruntime.StreamEvent{
-			toolTurn("call-1", "read_file", `{"path":"notes.txt"}`), // turn 1: other tool call
-			toolTurn("call-2", "read_file", `{"path":"notes.txt"}`), // turn 2: still no update_plan
-			toolTurn("call-3", "read_file", `{"path":"notes.txt"}`), // turn 3: reminder fires here
-			textTurn("done"),
-		},
-	}
+	provider := &mockProvider{turns: append(repeatedReadTurns(expectedThreshold), textTurn("done"))}
 
 	result, err := Run(context.Background(), "go", provider, Options{
 		Registry: registry,
@@ -287,6 +284,44 @@ func TestRunInjectsPlanNotCalledReminderForMultiStepTask(t *testing.T) {
 	}
 }
 
+func TestRunDoesNotInjectPlanReminderForBoundedToolSequence(t *testing.T) {
+	const expectedThreshold = 7
+	root := t.TempDir()
+	writeAgentTestFile(t, root+"/notes.txt", "alpha")
+	registry := tools.NewRegistry()
+	registry.Register(tools.NewScopedReadFileTool(root, nil))
+
+	provider := &mockProvider{turns: append(repeatedReadTurns(expectedThreshold-1), textTurn("done"))}
+	result, err := Run(context.Background(), "go", provider, Options{Registry: registry, MaxTurns: 12})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if result.FinalAnswer != "done" {
+		t.Fatalf("expected final answer, got %q", result.FinalAnswer)
+	}
+	if count := countUserMessagesContaining(result.Messages, planNotCalledReminderMarker); count != 0 {
+		t.Fatalf("bounded tool sequence must not draw a plan reminder, got %d", count)
+	}
+}
+
+func TestPlanReminderWaitsForTurnThresholdAfterEnoughToolCalls(t *testing.T) {
+	state := guardState{totalToolCalls: 7}
+	if got := state.planReminder(6); got != "" {
+		t.Fatalf("planReminder(6) = %q, want no reminder before turn threshold", got)
+	}
+	if state.notCalledReminderSent {
+		t.Fatal("early turn marked the not-called reminder as sent")
+	}
+}
+
+func repeatedReadTurns(count int) [][]zeroruntime.StreamEvent {
+	turns := make([][]zeroruntime.StreamEvent, 0, count)
+	for range count {
+		turns = append(turns, toolTurn("call", "read_file", `{"path":"notes.txt"}`))
+	}
+	return turns
+}
+
 func TestRunDoesNotInjectPlanReminderForTrivialTask(t *testing.T) {
 	root := t.TempDir()
 	writeAgentTestFile(t, root+"/notes.txt", "alpha")

internal/agent/loop.go

@@ -40,7 +40,15 @@ const (
 	toolResultControlSpecReview = "spec_review_required"
 )
 
-var errPermissionApprovalCanceled = errors.New("permission approval cancelled")
+// ErrPermissionApprovalCanceled ends a run because the caller cancelled a
+// permission prompt, rather than because anything failed.
+//
+// EXPORTED FOR THE SURFACES. A cancel is a normal ending and each surface says
+// so its own way — the TUI unwinds, and ACP has a "cancelled" stop reason for
+// exactly this. While it was unexported ACP could not distinguish it from a
+// fault and reported it as an internal error, so declining a tool looked like
+// the agent crashing.
+var ErrPermissionApprovalCanceled = errors.New("permission approval cancelled")
 
 // isImageRejectionError reports whether err is a provider 400 that rejects
 // image/multimodal content. This is checked BEFORE the compaction-retry path
@@ -133,6 +141,7 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 			}
 			options.Trace.Counter(trace.CounterInputTokens, int64(usage.InputTokens))
 			options.Trace.Counter(trace.CounterCachedInputTokens, int64(usage.CachedInputTokens))
+			options.Trace.Counter(trace.CounterCacheWriteTokens, int64(usage.CacheWriteTokens))
 			options.Trace.Counter(trace.CounterOutputTokens, int64(usage.OutputTokens))
 		}
 	}
@@ -186,6 +195,7 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 	planner := newContextPlanner(contextPlannerConfig{
 		contextWindow:  options.ContextWindow,
 		promptCacheKey: options.SessionID,
+		serviceTier:    options.ServiceTier,
 		promptParts:    promptParts,
 	})
 	messages := zeroruntime.SeedMessagesWithImages(promptParts.prompt, prompt, options.Images)
@@ -688,11 +698,12 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 				turnRequestedModel = toolResult.RequestedModel
 			}
 			messages = append(messages, zeroruntime.Message{
-				Role:         zeroruntime.MessageRoleTool,
-				Content:      toolResult.ModelOutput(),
-				ToolCallID:   toolResult.ToolCallID,
-				IsError:      toolResult.Status == tools.StatusError,
-				ChangedFiles: append([]string(nil), toolResult.ChangedFiles...),
+				Role:               zeroruntime.MessageRoleTool,
+				Content:            toolResult.ModelOutput(),
+				ToolCallID:         toolResult.ToolCallID,
+				ToolCallProviderID: call.ProviderCallID,
+				IsError:            toolResult.Status == tools.StatusError,
+				ChangedFiles:       append([]string(nil), toolResult.ChangedFiles...),
 			})
 			// Images ride a following USER message rather than the tool result
 			// above. Every provider drops images on a tool-role message —
@@ -829,12 +840,27 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 					_ = session.Prewarm(ctx)
 					provider = sessionProvider{session: session}
 					options.Model = turnRequestedModel
+					// Service tiers are model capabilities. A successful escalation may
+					// target a model that does not support the tier selected before the
+					// switch, so let the destination use its default.
+					options.ServiceTier = ""
+					planner.config.serviceTier = ""
 					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.
+					// Re-derive the context window for the new model so compaction and
+					// the OnContext budget report track the model actually in force —
+					// without this a switch to a smaller-window model can overflow the
+					// target, and a switch to a larger one over-compacts. An unknown
+					// model (<= 0) keeps the original window. The compactor also
+					// re-resolves its dedicated summarizer against the new model.
+					window := 0
+					if options.ContextWindowFor != nil {
+						window = options.ContextWindowFor(turnRequestedModel)
+					}
+					if window > 0 {
+						options.ContextWindow = window
+						planner.config.contextWindow = window
+					}
+					compactor.switchModel(turnRequestedModel, window)
 				}
 			}
 		}
@@ -1021,6 +1047,9 @@ func historySafeToolCalls(calls []ToolCall) []ToolCall {
 	safe := make([]ToolCall, len(calls))
 	for i, call := range calls {
 		safe[i] = call
+		if call.Freeform {
+			continue
+		}
 		args := strings.TrimSpace(call.Arguments)
 		switch {
 		case args == "":
@@ -1104,8 +1133,28 @@ func decodeToolArguments(arguments string, v any) error {
 }
 
 func executeToolCall(ctx context.Context, registry *tools.Registry, call ToolCall, permissionMode PermissionMode, options Options) (ToolResult, error) {
+	tool, toolFound := registry.Get(call.Name)
 	args := map[string]any{}
-	if call.Arguments != "" {
+	if call.Freeform {
+		if !toolFound || !tools.IsBuiltInApplyPatch(tool) {
+			return ToolResult{
+				ToolCallID: call.ID,
+				Name:       call.Name,
+				Status:     tools.StatusError,
+				Output:     "Error: Unsupported freeform tool call for " + call.Name + ".",
+			}, nil
+		}
+		prepared, err := tools.PrepareFreeformApplyPatchArguments(tool, call.Arguments)
+		if err != nil {
+			return ToolResult{
+				ToolCallID: call.ID,
+				Name:       call.Name,
+				Status:     tools.StatusError,
+				Output:     "Error: Invalid freeform apply_patch input: " + err.Error(),
+			}, nil
+		}
+		args = prepared
+	} else if call.Arguments != "" {
 		if err := decodeToolArguments(call.Arguments, &args); err != nil {
 			return ToolResult{
 				ToolCallID: call.ID,
@@ -1133,7 +1182,6 @@ func executeToolCall(ctx context.Context, registry *tools.Registry, call ToolCal
 			DenialReason: DenialFiltered,
 		}, nil
 	}
-	tool, toolFound := registry.Get(call.Name)
 	if (permissionMode == PermissionModeSpecDraft || permissionMode == PermissionModePlan) && toolFound && !ToolAdvertised(tool, permissionMode) {
 		modeName := string(permissionMode)
 		return ToolResult{
@@ -1358,7 +1406,7 @@ func executeToolCall(ctx context.Context, registry *tools.Registry, call ToolCal
 			}
 		case PermissionDecisionCancel:
 			emitCanceledPermission(options, call, requestEvent, decisionReason)
-			return canceledPermissionResult(call, decisionReason, requestEvent), fmt.Errorf("%w for %s", errPermissionApprovalCanceled, call.Name)
+			return canceledPermissionResult(call, decisionReason, requestEvent), fmt.Errorf("%w for %s", ErrPermissionApprovalCanceled, call.Name)
 		default:
 			emitDeniedPermission(options, call, requestEvent, decisionReason)
 			return deniedPermissionResult(call, decisionReason, requestEvent), nil
@@ -1542,7 +1590,7 @@ func maybeRetryUnsandboxedAfterSandboxRestriction(ctx context.Context, registry
 	case PermissionDecisionCancel:
 		emitCanceledPermission(options, call, requestEvent, reason)
 		canceled := canceledPermissionResult(call, reason, requestEvent)
-		return result, &canceled, true, decision.Action, reason, nil, fmt.Errorf("%w for %s", errPermissionApprovalCanceled, call.Name)
+		return result, &canceled, true, decision.Action, reason, nil, fmt.Errorf("%w for %s", ErrPermissionApprovalCanceled, call.Name)
 	default:
 		emitDeniedPermission(options, call, requestEvent, reason)
 		denied := deniedPermissionResult(call, reason, requestEvent)
@@ -1590,7 +1638,7 @@ func maybeRetryWithNetworkAfterSandboxDenial(ctx context.Context, registry *tool
 	case PermissionDecisionCancel:
 		emitCanceledPermission(options, call, requestEvent, reason)
 		canceled := canceledPermissionResult(call, reason, requestEvent)
-		return result, &canceled, true, decision.Action, reason, fmt.Errorf("%w for %s", errPermissionApprovalCanceled, call.Name)
+		return result, &canceled, true, decision.Action, reason, fmt.Errorf("%w for %s", ErrPermissionApprovalCanceled, call.Name)
 	default:
 		emitDeniedPermission(options, call, requestEvent, reason)
 		denied := deniedPermissionResult(call, reason, requestEvent)
@@ -3081,13 +3129,15 @@ func partitionToolsCached(registry *tools.Registry, permissionMode PermissionMod
 	// Deferral may activate only when tool_search is actually runnable; otherwise
 	// the loop would hide deferred tools behind a loader the dispatch gate rejects
 	// — an inescapable dead-end. "Runnable" mirrors executeToolCall's gate:
-	// registered, not in DisabledTools, and advertised in the current permission
-	// mode (e.g. not spec-draft, where tool_search is not advertised). The
+	// registered, model-visible, not in DisabledTools, and advertised in the
+	// current permission mode (e.g. not spec-draft, where tool_search is not
+	// advertised). The
 	// EnabledTools allowlist is intentionally NOT checked here — tool_search is
 	// exempt from the allowlist at dispatch, so an allowlist that omits it must
 	// not disable deferral.
 	loader, loaderFound := registry.Get(tools.ToolSearchToolName)
 	loaderUsable := loaderFound &&
+		tools.ModelVisible(loader) &&
 		!containsToolName(options.DisabledTools, tools.ToolSearchToolName) &&
 		ToolAdvertised(loader, permissionMode)
 
@@ -3194,15 +3244,21 @@ func cachedRuntimeToolDefinition(defCache map[string]zeroruntime.ToolDefinition,
 // runtimeToolDefinition renders a tool's advertised definition (name, description,
 // JSON-schema parameters) as sent to the provider.
 func runtimeToolDefinition(tool tools.Tool) zeroruntime.ToolDefinition {
-	return zeroruntime.ToolDefinition{
+	definition := zeroruntime.ToolDefinition{
 		Name:        tool.Name(),
 		Description: tool.Description(),
 		Parameters:  schemaToRuntimeMap(tool.Parameters()),
 	}
+	if tools.IsBuiltInApplyPatch(tool) {
+		definition.Type = zeroruntime.ToolDefinitionFreeform
+	}
+	return definition
 }
 
 func ToolVisible(tool tools.Tool, permissionMode PermissionMode, enabledTools []string, disabledTools []string) bool {
-	return ToolAllowedByFilters(tool.Name(), enabledTools, disabledTools) && ToolAdvertised(tool, permissionMode)
+	return tools.ModelVisible(tool) &&
+		ToolAllowedByFilters(tool.Name(), enabledTools, disabledTools) &&
+		ToolAdvertised(tool, permissionMode)
 }
 
 func ToolAllowedByFilters(name string, enabledTools []string, disabledTools []string) bool {
@@ -3351,10 +3407,11 @@ func loadedToolsFromResult(meta map[string]string) []string {
 func appendAbortedToolResults(messages []Message, remaining []ToolCall) []Message {
 	for _, call := range remaining {
 		messages = append(messages, zeroruntime.Message{
-			Role:       zeroruntime.MessageRoleTool,
-			Content:    abortedToolResultNotice,
-			ToolCallID: call.ID,
-			IsError:    true,
+			Role:               zeroruntime.MessageRoleTool,
+			Content:            abortedToolResultNotice,
+			ToolCallID:         call.ID,
+			ToolCallProviderID: call.ProviderCallID,
+			IsError:            true,
 		})
 	}
 	return messages

internal/agent/loop_test.go

@@ -1853,7 +1853,7 @@ func TestRunAbortsWhenPermissionRequestCanceled(t *testing.T) {
 		},
 	})
 
-	if !errors.Is(err, errPermissionApprovalCanceled) {
+	if !errors.Is(err, ErrPermissionApprovalCanceled) {
 		t.Fatalf("expected permission approval cancel error, got %v", err)
 	}
 	if result.FinalAnswer != "" {
@@ -4096,7 +4096,7 @@ func TestRunDoesNotFlagCleanToolOutput(t *testing.T) {
 // 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.StreamEventUsage, Usage: zeroruntime.Usage{PromptTokens: 100, CachedInputTokens: 20, CacheWriteTokens: 10, CompletionTokens: 40}},
 		{Type: zeroruntime.StreamEventText, Content: "done"},
 		{Type: zeroruntime.StreamEventDone},
 	}}}
@@ -4135,6 +4135,9 @@ func TestRunTracingWrapperStampsUsage(t *testing.T) {
 	if got := tr.Counter(trace.CounterCachedInputTokens); got != 20 {
 		t.Fatalf("cached input token counter = %d, want 20", got)
 	}
+	if got := tr.Counter(trace.CounterCacheWriteTokens); got != 10 {
+		t.Fatalf("cache-write token counter = %d, want 10", got)
+	}
 	if got := tr.Counter(trace.CounterOutputTokens); got != 40 {
 		t.Fatalf("output token counter = %d, want 40", got)
 	}
@@ -4340,3 +4343,95 @@ func TestPlanModeHonorsBeforeToolVeto(t *testing.T) {
 		t.Fatalf("expected tool result to mention the beforeTool veto, got %q", combined)
 	}
 }
+
+// A cancelled permission on either sandbox-retry path must abort the run with a
+// recognisable cancellation, and must not run the escalated retry.
+//
+// BOTH PATHS, because both wrap ErrPermissionApprovalCanceled and each is
+// reached by a different kind of denial: one by an unrestricted platform denial
+// asking for approval, the other by an external-network denial. Surfaces map
+// this sentinel onto their own idea of a cancel — ACP returns stopReason
+// "cancelled" for it — so a path that produced a bare error instead would show
+// the user a crash where they had simply declined.
+func TestCancellingASandboxRetryAbortsWithoutRetrying(t *testing.T) {
+	denial := func(capability execution.CapabilityKind, scope string) tools.Result {
+		return tools.Result{
+			Status: tools.StatusError,
+			ExecutionOutcome: &execution.Outcome{
+				State: execution.StateDenied,
+				Kind:  execution.OutcomeEnforcementDenied,
+				Denial: &execution.Denial{
+					Capability:  execution.Capability{Kind: capability, Scope: scope},
+					Source:      execution.DenialSourcePlatformSandbox,
+					Reason:      "denied by the sandbox",
+					Recoverable: true,
+					NextAction:  execution.DenialNextActionRequestApproval,
+				},
+			},
+		}
+	}
+
+	for _, testCase := range []struct {
+		name   string
+		result tools.Result
+		run    func(context.Context, *tools.Registry, ToolCall, tools.Tool, map[string]any, tools.Result, Options) error
+	}{
+		{
+			name:   "unsandboxed retry",
+			result: denial(execution.CapabilityUnrestricted, "host"),
+			run: func(ctx context.Context, registry *tools.Registry, call ToolCall, tool tools.Tool, args map[string]any, result tools.Result, options Options) error {
+				_, _, _, _, _, _, abortErr := maybeRetryUnsandboxedAfterSandboxRestriction(
+					ctx, registry, call, tool, args, result, PermissionModeAsk, options, nil)
+				return abortErr
+			},
+		},
+		{
+			name:   "network retry",
+			result: denial(execution.CapabilityExternalNetwork, ""),
+			run: func(ctx context.Context, registry *tools.Registry, call ToolCall, tool tools.Tool, args map[string]any, result tools.Result, options Options) error {
+				_, _, _, _, _, abortErr := maybeRetryWithNetworkAfterSandboxDenial(
+					ctx, registry, call, tool, args, result, PermissionModeAsk, options, nil)
+				return abortErr
+			},
+		},
+	} {
+		t.Run(testCase.name, func(t *testing.T) {
+			retryTool := &sandboxNamespaceLimitedRetryTool{}
+			registry := tools.NewRegistry()
+			registry.Register(retryTool)
+
+			asked := 0
+			options := Options{
+				Registry:       registry,
+				PermissionMode: PermissionModeAsk,
+				Autonomy:       "medium",
+				Sandbox: sandbox.NewEngine(sandbox.EngineOptions{
+					WorkspaceRoot: t.TempDir(),
+					Policy:        sandbox.DefaultPolicy(),
+				}),
+				OnPermissionRequest: func(context.Context, PermissionRequest) (PermissionDecision, error) {
+					asked++
+					return PermissionDecision{Action: PermissionDecisionCancel, Reason: "client cancelled"}, nil
+				},
+			}
+
+			call := ToolCall{ID: "call-1", Name: "bash", Arguments: `{"command":"curl example.com"}`}
+			args := map[string]any{"command": "curl example.com"}
+
+			abortErr := testCase.run(context.Background(), registry, call, retryTool, args, testCase.result, options)
+
+			if asked != 1 {
+				t.Fatalf("permission was requested %d times, want exactly 1", asked)
+			}
+			if !errors.Is(abortErr, ErrPermissionApprovalCanceled) {
+				t.Fatalf("abort error = %v, want it to match ErrPermissionApprovalCanceled", abortErr)
+			}
+			// The escalated retry is the thing the permission was gating. A
+			// cancelled prompt that still ran it would have run the command the
+			// user just declined.
+			if len(retryTool.calls) != 0 {
+				t.Fatalf("the escalated retry ran %d times after a cancel: %+v", len(retryTool.calls), retryTool.calls)
+			}
+		})
+	}
+}

internal/agent/partition_cache_test.go

@@ -31,6 +31,19 @@ func (t countingSchemaTool) Run(_ context.Context, _ map[string]any) tools.Resul
 	return tools.Result{Status: tools.StatusOK}
 }
 
+func TestRuntimeToolDefinitionMarksOnlyBuiltInApplyPatchFreeform(t *testing.T) {
+	builtIn := runtimeToolDefinition(tools.NewScopedApplyPatchTool(t.TempDir(), nil))
+	if builtIn.Type != zeroruntime.ToolDefinitionFreeform {
+		t.Fatalf("built-in apply_patch type = %q, want freeform", builtIn.Type)
+	}
+
+	calls := 0
+	custom := runtimeToolDefinition(countingSchemaTool{name: "apply_patch", calls: &calls})
+	if custom.Type != "" {
+		t.Fatalf("custom apply_patch type = %q, want legacy function shape", custom.Type)
+	}
+}
+
 // The cache renders a tool's schema once and reuses it across calls, and its
 // output is identical to the uncached path.
 func TestPartitionToolsCacheRendersOnceAndMatchesUncached(t *testing.T) {

internal/agent/prompt_budget_test.go

@@ -1,9 +1,11 @@
 package agent
 
 import (
+	"strings"
 	"testing"
 
 	"github.com/Gitlawb/zero/internal/tools"
+	"github.com/Gitlawb/zero/internal/zeroruntime"
 )
 
 // These ceilings are a deliberate ratchet on Zero's fixed per-turn overhead — the
@@ -33,7 +35,24 @@ const (
 	// The ratchet did its job: it caught the creep and forced a decision instead
 	// of a drift. That is the point of it, so keep raising it deliberately rather
 	// than reflexively.
+	//
+	// THIS COUNTS THE SCHEMAS ONLY, NOT THE HOST'S SHELL GUIDANCE. Until
+	// 2026-09-07 it counted both, and that made it a different measurement on
+	// every machine: exec_command appends shell guidance whose size depends on
+	// which shell was detected, so the same tree measured 3498 on Linux, ~3600 on
+	// a Windows host with pwsh 7, and 3653 on Windows PowerShell 5.1, which is
+	// stock Windows. The 5.1 case has been over this ceiling since view_image
+	// landed, so `go test ./...` failed for anyone on a stock Windows box while
+	// CI stayed green, because the runners have pwsh 7 and take the cheaper
+	// branch. A budget that passes or fails on which PowerShell happens to be
+	// installed is not ratcheting anything.
 	maxEagerToolSchemaTokens = 3650
+	// The host-specific half, ratcheted on its own so it cannot creep either.
+	// Measured 2026-09-07: Windows PowerShell 5.1 is the most expensive host at
+	// 155 tokens (722 chars), pwsh 7 is 110, cmd.exe is 87, and every POSIX host
+	// is 0 because no guidance is appended there. Set above the worst case, not
+	// above whatever this machine happens to be.
+	maxExecCommandShellGuidanceTokens = 200
 )
 
 func TestSystemPromptTokenBudget(t *testing.T) {
@@ -56,8 +75,35 @@ func TestEagerToolSchemaTokenBudget(t *testing.T) {
 	// Options{} keeps DeferThreshold at 0, so deferral is inactive and every core
 	// tool is exposed eagerly — exactly what a plugin-free session sends each turn.
 	exposed, _ := partitionTools(registry, PermissionModeAuto, Options{}, map[string]bool{})
-	got := estimateToolDefTokens(exposed)
-	t.Logf("eager core tool schemas: %d tokens across %d tools", got, len(exposed))
+	total := estimateToolDefTokens(exposed)
+	// exec_command's description carries this host's shell guidance, and that
+	// guidance is the one part of the eager set that is not the same everywhere.
+	// Charged separately below rather than folded in here, or this ceiling would
+	// mean a different thing on every machine. See the constants above.
+	//
+	// Removed from the description BEFORE estimating rather than subtracted from
+	// the total afterwards. ApproxTextTokens is an integer division, so
+	// floor((schema+guidance)/4) - floor(guidance/4) comes out one token high
+	// whenever the two remainders add past four, and which side of that line a
+	// host lands on depends on the length of its guidance: exactly the
+	// host-dependence this number is supposed to be rid of.
+	guidance := tools.HostExecCommandShellGuidance()
+	schemas := make([]zeroruntime.ToolDefinition, len(exposed))
+	copy(schemas, exposed)
+	stripped := 0
+	for index := range schemas {
+		without := strings.TrimSuffix(schemas[index].Description, guidance)
+		if without != schemas[index].Description {
+			stripped++
+		}
+		schemas[index].Description = without
+	}
+	if guidance != "" && stripped != 1 {
+		t.Fatalf("SETUP INVALID: this host's shell guidance came off %d descriptions, want exactly exec_command's; internal/tools pins it as that description's tail", stripped)
+	}
+	got := estimateToolDefTokens(schemas)
+	guidanceTokens := ApproxTextTokens(guidance)
+	t.Logf("eager core tool schemas: %d tokens across %d tools (%d total, %d host shell guidance)", got, len(exposed), total, guidanceTokens)
 	if got > maxEagerToolSchemaTokens {
 		// NOT "defer a tool": this test pins DeferThreshold at 0, so marking a
 		// tool deferred leaves it exposed here and changes nothing. Deferral is
@@ -66,4 +112,25 @@ func TestEagerToolSchemaTokenBudget(t *testing.T) {
 		// deliberate raise.
 		t.Fatalf("eager tool schemas are %d tokens, over the %d ceiling — trim a schema, drop a core tool, or raise the ceiling deliberately (deferring will NOT help: this test disables deferral)", got, maxEagerToolSchemaTokens)
 	}
+	if guidanceTokens > maxExecCommandShellGuidanceTokens {
+		t.Fatalf("this host's exec_command shell guidance is %d tokens, over the %d ceiling — trim the guidance in internal/tools/shell_runtime.go or raise the ceiling deliberately", guidanceTokens, maxExecCommandShellGuidanceTokens)
+	}
+}
+
+func TestAgentAdvertisesBothEditTools(t *testing.T) {
+	registry := tools.NewRegistry()
+	for _, tool := range tools.CoreToolsScoped(t.TempDir(), nil) {
+		registry.Register(tool)
+	}
+	exposed, _ := partitionTools(registry, PermissionModeAsk, Options{}, map[string]bool{})
+	names := make(map[string]bool, len(exposed))
+	for _, definition := range exposed {
+		names[definition.Name] = true
+	}
+	if !names["apply_patch"] {
+		t.Fatal("agent must retain apply_patch for multi-hunk changes")
+	}
+	if !names["edit_file"] {
+		t.Fatal("agent must receive edit_file for targeted exact replacements")
+	}
 }

internal/agent/prompt_fingerprint.go

@@ -58,12 +58,18 @@ func toolSubstrings(exposed []zeroruntime.ToolDefinition) (toolsSubstr, schemaSu
 	for _, def := range exposed {
 		writeLengthPrefixed(&toolsSB, def.Name)
 		writeLengthPrefixed(&toolsSB, def.Description)
+		writeLengthPrefixed(&toolsSB, string(def.Type))
 		// 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))
+		if def.Format != nil {
+			writeLengthPrefixed(&schemaSB, def.Format.Type)
+			writeLengthPrefixed(&schemaSB, def.Format.Syntax)
+			writeLengthPrefixed(&schemaSB, def.Format.Definition)
+		}
 	}
 	return toolsSB.String(), schemaSB.String()
 }

internal/agent/prompt_fingerprint_test.go

@@ -73,6 +73,38 @@ func TestComputePrefixFingerprintSchemaChangeChangesSchemaHash(t *testing.T) {
 	}
 }
 
+func TestComputePrefixFingerprintIncludesToolTypeAndFormat(t *testing.T) {
+	opts := Options{Cwd: t.TempDir(), SystemPrompt: "core"}
+	base := zeroruntime.ToolDefinition{Name: "apply_patch", Description: "Patch files"}
+	baseFingerprint := ComputePrefixFingerprint(opts, []zeroruntime.ToolDefinition{base})
+
+	typed := base
+	typed.Type = zeroruntime.ToolDefinitionFreeform
+	if got := ComputePrefixFingerprint(opts, []zeroruntime.ToolDefinition{typed}); got.ToolsHash == baseFingerprint.ToolsHash {
+		t.Fatal("ToolsHash must change when tool Type changes")
+	}
+
+	formatCases := []struct {
+		name   string
+		format zeroruntime.ToolDefinitionFormat
+	}{
+		{name: "type", format: zeroruntime.ToolDefinitionFormat{Type: "grammar"}},
+		{name: "syntax", format: zeroruntime.ToolDefinitionFormat{Syntax: "lark"}},
+		{name: "definition", format: zeroruntime.ToolDefinitionFormat{Definition: "start: PATCH"}},
+		{name: "empty", format: zeroruntime.ToolDefinitionFormat{}},
+	}
+	for _, test := range formatCases {
+		t.Run(test.name, func(t *testing.T) {
+			withFormat := base
+			withFormat.Format = &test.format
+			got := ComputePrefixFingerprint(opts, []zeroruntime.ToolDefinition{withFormat})
+			if got.SchemaHash == baseFingerprint.SchemaHash {
+				t.Fatalf("SchemaHash must change for %s format input", test.name)
+			}
+		})
+	}
+}
+
 // 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

internal/agent/system_prompt.md

@@ -30,13 +30,10 @@ work.
    read-before-edit discipline: inspect the target file and nearby callers,
    tests, or config before you modify behavior. Never edit a file you have not
    read.
-2. **Plan.** Use update_plan for implementation or investigation that has at
-   least three meaningful dependent steps, then keep it current at milestones.
-   Mark completed work and the next in-progress step together when practical;
-   do not spend separate turns updating it after every file or command. Keep at
-   most one item in_progress. Skip plans for simple lookups, explanations, code
-   navigation, and other short tasks. Never create a plan after the work is
-   already complete merely to describe what you did.
+2. **Plan.** Use update_plan for multi-component, sequencing-uncertain, or long
+   work. Skip it for bounded changes in one component. Update only when the
+   overall phase changes, keep at most one item in_progress, and never create a
+   plan after the work is complete.
 3. **Implement.** Make focused changes that match the surrounding code's style,
    naming, and conventions. Prefer the smallest change that fully solves the
    problem. Avoid broad refactors, unrelated rewrites, dependency churn, and
@@ -52,37 +49,48 @@ work.
 
 - Choose the narrowest tool that safely accomplishes the step. Prefer native
   file tools - read_file, read_minified_file, list_directory, glob, grep,
-  write_file, edit_file, apply_patch - over shelling out to
+  edit_file, apply_patch, write_file - over shelling out to
   cat/sed/awk/python for file operations.
   They are safer, reviewable, and produce clean diffs.
-- Prefer read_minified_file when initially exploring source code; it preserves
-  code while removing comments and redundant whitespace. Use read_file when
-  exact text, comments, or line numbers are needed.
+- Use read_minified_file only to explore large or unfamiliar source. Use
+  read_file for small or likely edit targets and exact text, comments, or line
+  numbers. Do not read both unless a new need for exact content appears.
 - Keep edits focused and reviewable. A single patch may update several related
   files when they form one coherent change; do not hide unrelated edits in a
   bulk shell or script rewrite.
-- For edits to existing files, prefer edit_file or apply_patch with minimal,
-  targeted diffs. Match the existing indentation, imports, and idioms. Match the
-  file's comment density: do not add explanatory comments unless the user asks or
-  the code is already comment-dense.
+- For edits to existing files, use edit_file for a targeted change (old_string
+  must match the file exactly and be unique, so include a few surrounding
+  lines) and apply_patch when one coherent change spans several hunks or
+  files. Use write_file only to create a file or when most of it changes; do
+  not rewrite a whole file to change a few lines. Match the existing
+  indentation, imports, and idioms. Match the file's comment density: do not
+  add explanatory comments unless the user asks or the code is already
+  comment-dense.
+- A successful edit result already confirms the change; do not re-read a file
+  just to verify an edit that succeeded. If an edit fails, read the error, fix
+  the old_string or hunk, and retry the same tool rather than switching to a
+  full rewrite.
 - Solve the problem as posed, not a more general version of it. Add no
   speculative abstraction, configurability, or handling for cases that cannot
   occur, and nothing the user did not ask for. A small diff can still be
-  over-built; if a 200-line solution could be 50, rewrite it.
+  over-built; if a 200-line solution could be 50, rewrite it. Once the required
+  behavior passes validation, stop instead of adding optional cleanup.
 - Preserve behavior you were not asked to change. Do not delete or rewrite code
   you did not author unless the task requires it; if you must, say so.
 
 ## Testing gate (mandatory)
 
-- After any change to code, verify after edits by running the project's
-  validators before you summarize or commit: tests, type-checks, linters, and/or
-  the build, as appropriate. Scope them to the change while iterating; reserve
-  full-suite runs for milestones.
-- If you are unsure which validators apply, search the repo (Makefile, package
-  manifests, CI config) to find them.
+- After code changes, run the project's documented appropriate validators.
+  Scope iteration checks, combine compatible commands when diagnostics stay
+  useful, and run the final set once after the last edit. Rerun only after a
+  change or failure.
+- If unsure which entry points apply, inspect the Makefile, manifests, or CI;
+  do not invent a parallel flow.
 - Never claim a task is done, and never commit, while validators are failing. If
   they fail, fix the cause and rerun; do not paper over it. If you could not run
   a validator, say so explicitly rather than implying success.
+- A successful native patch confirms the edit. Reread only when the next change
+  needs exact content or the result was ambiguous.
 
 ## Tool use
 
@@ -104,7 +112,7 @@ work.
   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.
+  with exec_command, or use edit_file/apply_patch/write_file 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
@@ -140,6 +148,7 @@ work.
 - Default to concise, skimmable output. Lead with the answer or the result.
 - Use GitHub-flavored Markdown: headings to structure longer replies, fenced
   code blocks for code, and `inline code` for file paths, commands, symbols, and
-  short snippets. Reference code as `file:line` so it is clickable.
+  short snippets. Reference code with file paths and include line numbers when
+  already known; do not look them up solely for the final summary.
 - Report outcomes faithfully: if tests failed, show it; if a step was skipped,
   say so; when something is done and verified, state it plainly without hedging.

internal/agent/system_prompt_models.go

@@ -60,8 +60,11 @@ const openAIPromptAddendum = `<model_guidance>
   longer answers, fenced code blocks for code, and ` + "`inline code`" + ` for paths,
   commands, and symbols.
 - Strongly prefer the native file tools (read_file, list_directory, grep, glob,
-  write_file, edit_file, apply_patch) over shelling out to cat/sed/awk/python for
-  file work. Make one tool call per file; do not batch file writes into a script.
+  edit_file, apply_patch, write_file) over shelling out to cat/sed/awk/python for
+  file work. Use them directly; never run apply_patch through exec_command.
+  read_file, edit_file and write_file take one file per call, while one coherent
+  apply_patch may span several files. Keep independent native calls in the same
+  turn, and do not batch file writes into a script.
 - Persist until the task is fully handled this turn: gather context, implement,
   run the validators, and report — do not stop at a partial result.
 </model_guidance>`

internal/agent/system_prompt_models_test.go

@@ -29,6 +29,9 @@ func TestBuildSystemPromptAppendsModelAddendum(t *testing.T) {
 	if got := buildSystemPrompt(Options{Model: "gpt-5"}); !strings.Contains(got, openAIPromptAddendum) {
 		t.Fatalf("expected the OpenAI addendum in the gpt-5 prompt")
 	}
+	if got := strings.Join(strings.Fields(strings.ToLower(buildSystemPrompt(Options{Model: "gpt-5"}))), " "); !strings.Contains(got, "never run apply_patch through exec_command") {
+		t.Fatalf("OpenAI prompt must route edits through the native patch tool:\n%s", got)
+	}
 	// Claude is aligned with the core prompt and gets no family addendum now that
 	// comment discipline is universal; it must not pick up another family's block.
 	claude := buildSystemPrompt(Options{Model: "claude-opus-4-6"})

internal/agent/system_prompt_test.go

@@ -29,14 +29,17 @@ func TestMain(m *testing.M) {
 }
 
 func TestCoreSystemPromptIncludesCodingQualityRules(t *testing.T) {
-	prompt := strings.ToLower(buildSystemPrompt(Options{}))
+	prompt := strings.Join(strings.Fields(strings.ToLower(buildSystemPrompt(Options{}))), " ")
+	if strings.Contains(prompt, "clickable file paths") {
+		t.Fatalf("system prompt must not claim inline-code file paths are clickable: %s", prompt)
+	}
 
 	for _, want := range []string{
 		"read-before-edit",
 		"inspect the target file",
 		"plan then act",
 		"choose the narrowest tool",
-		"prefer edit_file or apply_patch",
+		"for edits to existing files, use edit_file",
 		"verify after edits",
 		"honor the active permission mode",
 		"avoid broad refactors",
@@ -44,6 +47,13 @@ func TestCoreSystemPromptIncludesCodingQualityRules(t *testing.T) {
 		"do not recognize",
 		"scaled to the work",
 		"comment density",
+		"skip it for bounded changes in one component",
+		"do not read both unless a new need for exact content appears",
+		"combine compatible commands when diagnostics stay useful",
+		"do not invent a parallel flow",
+		"a successful native patch confirms the edit",
+		"once the required behavior passes validation, stop",
+		"do not look them up solely for the final summary",
 	} {
 		if !strings.Contains(prompt, want) {
 			t.Fatalf("expected core system prompt to include %q, got:\n%s", want, buildSystemPrompt(Options{}))

internal/agent/types.go

@@ -325,6 +325,7 @@ type Options struct {
 	ProviderName     string
 	Model            string
 	ReasoningEffort  string
+	ServiceTier      string
 	Cwd              string
 	SystemPrompt     string
 	// TransientSystemPrompt adds trusted runtime guidance for this run only.
@@ -347,10 +348,26 @@ type Options struct {
 	// CompactionPreserveLast is how many trailing messages compaction keeps
 	// verbatim. <= 0 falls back to defaultCompactionPreserveLast.
 	CompactionPreserveLast int
-	Registry               *tools.Registry
-	PermissionMode         PermissionMode
-	Autonomy               string
-	Sandbox                *sandbox.Engine
+	// Summarizer, when set, lazily builds the provider used for compaction
+	// summarization calls — typically a cheap/fast model, since summaries at
+	// main-model prices are the single most expensive recurring event in a
+	// long run. It receives the main model currently in force and may return
+	// (nil, nil) when no dedicated summarizer applies to it (the main model is
+	// already the cheap one). Built on the first paid compaction and again
+	// after a mid-run model switch. Any failure (build or call) falls back to
+	// the run's main provider until the next switch, so a misconfigured
+	// summarizer can never break compaction. nil keeps today's behavior.
+	Summarizer func(ctx context.Context, mainModelID string) (Provider, error)
+	// ContextWindowFor, when set, resolves a model ID to its context window so
+	// the compactor can re-derive its threshold after a mid-run model switch
+	// (escalate_model). Without it a switch keeps compacting against the
+	// original model's window — overflowing a smaller target or over-compacting
+	// a larger one. Return <= 0 when the model is unknown (window unchanged).
+	ContextWindowFor func(modelID string) int
+	Registry         *tools.Registry
+	PermissionMode   PermissionMode
+	Autonomy         string
+	Sandbox          *sandbox.Engine
 	// FileTracker records per-session file read/write versions so the write tools
 	// can detect a file changed on disk outside Zero since it was last read. nil
 	// disables the check. Created once per session and threaded into every tool run.

internal/agenteval/trace.go

@@ -29,10 +29,6 @@ func ParseTraceEventKeys(stdout string) []string {
 	return keys
 }
 
-func TraceEventKeys(stdout string) []string {
-	return ParseTraceEventKeys(stdout)
-}
-
 func MissingTraceEvents(required []string, stdout string) []string {
 	seen := map[string]bool{}
 	for _, key := range ParseTraceEventKeys(stdout) {

internal/background/manager.go

@@ -453,7 +453,12 @@ func (manager *Manager) normalizeLoadedTask(fileTaskID string, task Task) (Task,
 	}
 	outputFile, err := manager.outputFile(task.ID, task.OutputFile)
 	if err != nil {
-		return Task{}, false, err
+		var migrated bool
+		outputFile, migrated = manager.migrateLegacyOutputFile(task.ID, task.OutputFile)
+		if !migrated {
+			return Task{}, false, err
+		}
+		changed = true
 	}
 	if outputFile != task.OutputFile {
 		changed = true
@@ -562,6 +567,32 @@ func (manager *Manager) outputFile(taskID string, requested string) (string, err
 	return path, nil
 }
 
+// migrateLegacyOutputFile recovers task metadata written under a previous home
+// or XDG data root. It accepts only the canonical per-task output
+// name in the historic zero/background layout, and only when the corresponding
+// regular file already exists inside the active manager root.
+func (manager *Manager) migrateLegacyOutputFile(taskID string, requested string) (string, bool) {
+	requested = strings.TrimSpace(requested)
+	if !filepath.IsAbs(requested) {
+		return "", false
+	}
+	requested = filepath.Clean(requested)
+	expectedName := taskID + ".ndjson"
+	if filepath.Base(requested) != expectedName {
+		return "", false
+	}
+	legacyRoot := filepath.Dir(requested)
+	if filepath.Base(legacyRoot) != "background" || filepath.Base(filepath.Dir(legacyRoot)) != "zero" {
+		return "", false
+	}
+	canonical := filepath.Join(manager.rootDir, expectedName)
+	info, err := os.Lstat(canonical)
+	if err != nil || !info.Mode().IsRegular() {
+		return "", false
+	}
+	return canonical, true
+}
+
 func validTaskID(taskID string) bool {
 	return taskIDPattern.MatchString(strings.TrimSpace(taskID))
 }

internal/background/manager_test.go

@@ -1,6 +1,7 @@
 package background
 
 import (
+	"encoding/json"
 	"errors"
 	"os"
 	"path/filepath"
@@ -191,6 +192,88 @@ func TestManagerSkipsInvalidPersistedMetadata(t *testing.T) {
 	}
 }
 
+func TestManagerMigratesLegacyOutputPathWhenCanonicalOutputExists(t *testing.T) {
+	root := t.TempDir()
+	const taskID = "legacy_task"
+	canonicalOutput := filepath.Join(root, taskID+".ndjson")
+	if err := os.WriteFile(canonicalOutput, []byte("legacy output\n"), 0o600); err != nil {
+		t.Fatalf("write canonical output: %v", err)
+	}
+	legacyOutput := filepath.Join(t.TempDir(), "zero", "background", taskID+".ndjson")
+	metadata, err := json.Marshal(Task{
+		ID:         taskID,
+		Type:       "specialist",
+		Status:     StatusCompleted,
+		OutputFile: legacyOutput,
+		StartedAt:  time.Date(2026, 8, 1, 10, 0, 0, 0, time.UTC),
+	})
+	if err != nil {
+		t.Fatalf("marshal legacy metadata: %v", err)
+	}
+	if err := os.WriteFile(filepath.Join(root, taskID+".json"), append(metadata, '\n'), 0o600); err != nil {
+		t.Fatalf("write legacy metadata: %v", err)
+	}
+
+	manager, err := NewManager(root)
+	if err != nil {
+		t.Fatalf("NewManager returned error: %v", err)
+	}
+	task, ok := manager.Get(taskID)
+	if !ok {
+		t.Fatal("manager did not load migrated task")
+	}
+	if task.OutputFile != canonicalOutput {
+		t.Fatalf("migrated output file = %q, want %q", task.OutputFile, canonicalOutput)
+	}
+	if warnings := manager.LoadWarnings(); len(warnings) != 0 {
+		t.Fatalf("migration produced load warnings: %#v", warnings)
+	}
+
+	persisted, err := os.ReadFile(filepath.Join(root, taskID+".json"))
+	if err != nil {
+		t.Fatalf("read migrated metadata: %v", err)
+	}
+	var saved Task
+	if err := json.Unmarshal(persisted, &saved); err != nil {
+		t.Fatalf("decode migrated metadata: %v", err)
+	}
+	if saved.OutputFile != canonicalOutput {
+		t.Fatalf("persisted output file = %q, want %q", saved.OutputFile, canonicalOutput)
+	}
+}
+
+func TestManagerRejectsUnrelatedAbsoluteOutputPathDuringMigration(t *testing.T) {
+	root := t.TempDir()
+	const taskID = "legacy_task"
+	if err := os.WriteFile(filepath.Join(root, taskID+".ndjson"), []byte("output\n"), 0o600); err != nil {
+		t.Fatalf("write canonical output: %v", err)
+	}
+	metadata, err := json.Marshal(Task{
+		ID:         taskID,
+		Type:       "specialist",
+		Status:     StatusCompleted,
+		OutputFile: filepath.Join(t.TempDir(), taskID+".ndjson"),
+		StartedAt:  time.Date(2026, 8, 1, 10, 0, 0, 0, time.UTC),
+	})
+	if err != nil {
+		t.Fatalf("marshal metadata: %v", err)
+	}
+	if err := os.WriteFile(filepath.Join(root, taskID+".json"), append(metadata, '\n'), 0o600); err != nil {
+		t.Fatalf("write metadata: %v", err)
+	}
+
+	manager, err := NewManager(root)
+	if err != nil {
+		t.Fatalf("NewManager returned error: %v", err)
+	}
+	if _, ok := manager.Get(taskID); ok {
+		t.Fatal("manager loaded task with an unrelated absolute output path")
+	}
+	if warnings := strings.Join(manager.LoadWarnings(), "\n"); !strings.Contains(warnings, "output file must be inside") {
+		t.Fatalf("load warnings = %q, want containment rejection", warnings)
+	}
+}
+
 func TestManagerMarksReloadedRunningTaskNonKillable(t *testing.T) {
 	root := t.TempDir()
 	manager, err := NewManager(root)

internal/background/process_posix.go

@@ -3,7 +3,9 @@
 package background
 
 import (
+	"errors"
 	"os/exec"
+	"syscall"
 	"time"
 
 	"github.com/Gitlawb/zero/internal/execution"
@@ -39,3 +41,43 @@ func ConfigureChildProcessGroup(cmd *exec.Cmd) {
 func terminateProcess(pid int) error {
 	return execution.TerminateProcessTree(pid, terminationGracePeriod, terminationPollInterval)
 }
+
+// terminateOwnedProcess terminates cmd's process group directly when its launch
+// configuration proves it was made a group leader. This avoids fragile Getpgid
+// rediscovery after an owned leader exits. Ordinary commands fall back to the
+// safe PID/tree path rather than assuming their PID is also a process-group ID.
+//
+// The bool result is whether the leader had already exited before termination
+// was attempted — the POSIX analogue of Windows GetExitCodeProcess returning a
+// value other than STILL_ACTIVE. A waitable zombie is already-exited: Wait can
+// collect it, but kill(0) still succeeds, and when ps is unavailable
+// signalTargetRunning conservatively treats that as "still running".
+// TerminateCommand uses the flag to discard the resulting spurious
+// SIGKILL-timeout once the reap has independently succeeded. The flag is about
+// the leader only; group/tree termination is still invoked so live descendants
+// are signalled.
+//
+// The Setpgid && Pgid == 0 guard only recognizes ConfigureChildProcessGroup's
+// own convention. A command made its own session leader via Setsid (as opposed
+// to Setpgid) is also its own process-group leader in practice, but takes the
+// slower rediscovery path here since Setsid isn't checked. Harmless today
+// because TerminateCommand has exactly one caller in this codebase; worth
+// covering explicitly if Setsid-configured commands start using this path too.
+func terminateOwnedProcess(cmd *exec.Cmd) (bool, error) {
+	alreadyExited := leaderWaitableExited(cmd.Process.Pid)
+	if cmd.SysProcAttr != nil && cmd.SysProcAttr.Setpgid && cmd.SysProcAttr.Pgid == 0 {
+		return alreadyExited, execution.TerminateProcessGroup(cmd.Process.Pid, terminationGracePeriod, terminationPollInterval)
+	}
+	return alreadyExited, terminateProcess(cmd.Process.Pid)
+}
+
+// terminationTargetGoneAfterReap independently checks the exact signal target
+// used for an owned command. In particular, reaping a zombie group leader does
+// not make this return true while any descendant still occupies that group.
+func terminationTargetGoneAfterReap(cmd *exec.Cmd) bool {
+	target := cmd.Process.Pid
+	if cmd.SysProcAttr != nil && cmd.SysProcAttr.Setpgid && cmd.SysProcAttr.Pgid == 0 {
+		target = -target
+	}
+	return errors.Is(syscall.Kill(target, syscall.Signal(0)), syscall.ESRCH)
+}

internal/background/process_posix_test.go

@@ -6,6 +6,7 @@ import (
 	"bufio"
 	"errors"
 	"os/exec"
+	"runtime"
 	"strconv"
 	"strings"
 	"syscall"
@@ -147,11 +148,12 @@ func TestTerminateProcessKillsForkedChildren(t *testing.T) {
 		t.Fatalf("terminateProcess: %v", err)
 	}
 
-	// The forked child must be gone too (poll until reaped by init).
+	// The forked child must no longer be running. An orphaned zombie is already
+	// dead but may remain visible briefly until the platform's init reaps it.
 	deadline := time.Now().Add(2 * time.Second)
-	// 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) {
+	// processStopped treats only ESRCH or an observed zombie state as stopped;
+	// other errors (for example EPERM) remain live and keep polling.
+	for !processStopped(childPID) {
 		if time.Now().After(deadline) {
 			_ = syscall.Kill(childPID, syscall.SIGKILL)
 			t.Fatalf("forked child %d survived terminateProcess — group kill failed", childPID)
@@ -159,3 +161,176 @@ func TestTerminateProcessKillsForkedChildren(t *testing.T) {
 		time.Sleep(20 * time.Millisecond)
 	}
 }
+
+func TestTerminateCommandStopsUnconfiguredCommand(t *testing.T) {
+	cmd := exec.Command("sleep", "30")
+	if err := cmd.Start(); err != nil {
+		t.Fatalf("start: %v", err)
+	}
+	t.Cleanup(func() {
+		if cmd.ProcessState == nil {
+			_ = cmd.Process.Kill()
+		}
+	})
+
+	if err := TerminateCommand(cmd); err != nil {
+		t.Fatalf("TerminateCommand: %v", err)
+	}
+	if cmd.ProcessState == nil {
+		t.Fatal("unconfigured command was not reaped")
+	}
+}
+
+func TestTerminateCommandKillsChildAfterLeaderExits(t *testing.T) {
+	grace, poll := terminationGracePeriod, terminationPollInterval
+	terminationGracePeriod, terminationPollInterval = 2*time.Second, 20*time.Millisecond
+	t.Cleanup(func() { terminationGracePeriod, terminationPollInterval = grace, poll })
+
+	// The leader exits immediately after launching the child. TerminateCommand
+	// must capture and signal the process group before Wait reaps the leader and
+	// makes its group identity unavailable.
+	cmd := exec.Command("sh", "-c", "sleep 300 & echo $!; exit 0")
+	ConfigureChildProcessGroup(cmd)
+	stdout, err := cmd.StdoutPipe()
+	if err != nil {
+		t.Fatalf("stdout pipe: %v", err)
+	}
+	if err := cmd.Start(); err != nil {
+		t.Fatalf("start: %v", err)
+	}
+	line, err := bufio.NewReader(stdout).ReadString('\n')
+	if err != nil {
+		t.Fatalf("read forked child pid: %v", err)
+	}
+	childPID, err := strconv.Atoi(strings.TrimSpace(line))
+	if err != nil {
+		t.Fatalf("parse forked child pid %q: %v", line, err)
+	}
+
+	if err := TerminateCommand(cmd); err != nil {
+		t.Fatalf("TerminateCommand: %v", err)
+	}
+
+	deadline := time.Now().Add(2 * time.Second)
+	for !processStopped(childPID) {
+		if time.Now().After(deadline) {
+			_ = syscall.Kill(childPID, syscall.SIGKILL)
+			t.Fatalf("forked child %d survived TerminateCommand", childPID)
+		}
+		time.Sleep(20 * time.Millisecond)
+	}
+}
+
+func TestTerminateCommandZombieLeaderWithoutPS(t *testing.T) {
+	if runtime.GOOS != "linux" && runtime.GOOS != "darwin" {
+		t.Skip("waitable-exited leader probe is implemented on linux and darwin")
+	}
+
+	grace, poll := terminationGracePeriod, terminationPollInterval
+	terminationGracePeriod, terminationPollInterval = 150*time.Millisecond, 10*time.Millisecond
+	t.Cleanup(func() { terminationGracePeriod, terminationPollInterval = grace, poll })
+
+	// Issue #862: a zombie group leader with ps unresolvable used to burn both
+	// grace periods (kill(0) succeeds against a zombie) and return
+	// "did not exit after SIGKILL" even though cmd.Wait reaped exit status 0.
+	// The already-exited probe must answer without ps so TerminateCommand can
+	// discard that spurious timeout after a successful reap.
+	cmd := exec.Command("sh", "-c", "exit 0")
+	ConfigureChildProcessGroup(cmd)
+	if err := cmd.Start(); err != nil {
+		t.Fatalf("start: %v", err)
+	}
+	pid := cmd.Process.Pid
+	t.Cleanup(func() {
+		if cmd.ProcessState == nil {
+			_ = cmd.Process.Kill()
+			_ = cmd.Wait()
+		}
+	})
+
+	deadline := time.Now().Add(5 * time.Second)
+	for !leaderWaitableExited(pid) {
+		if time.Now().After(deadline) {
+			t.Fatalf("leader %d did not become waitable-exited within 5s", pid)
+		}
+		time.Sleep(10 * time.Millisecond)
+	}
+
+	t.Setenv("PATH", "")
+	if err := TerminateCommand(cmd); err != nil {
+		t.Fatalf("TerminateCommand: %v, want nil (zombie leader with ps unavailable must not report SIGKILL timeout)", err)
+	}
+	if cmd.ProcessState == nil {
+		t.Fatal("zombie leader was not reaped")
+	}
+}
+
+func TestTerminateCommandPreservesGroupFailureWithExitedLeaderAndLiveChild(t *testing.T) {
+	if runtime.GOOS != "linux" && runtime.GOOS != "darwin" {
+		t.Skip("waitable-exited leader probe is implemented on linux and darwin")
+	}
+
+	cmd := exec.Command("sh", "-c", "sleep 300 & echo $!; exit 0")
+	ConfigureChildProcessGroup(cmd)
+	stdout, err := cmd.StdoutPipe()
+	if err != nil {
+		t.Fatalf("stdout pipe: %v", err)
+	}
+	if err := cmd.Start(); err != nil {
+		t.Fatalf("start: %v", err)
+	}
+	line, err := bufio.NewReader(stdout).ReadString('\n')
+	if err != nil {
+		t.Fatalf("read forked child pid: %v", err)
+	}
+	childPID, err := strconv.Atoi(strings.TrimSpace(line))
+	if err != nil {
+		t.Fatalf("parse forked child pid %q: %v", line, err)
+	}
+	t.Cleanup(func() {
+		_ = syscall.Kill(childPID, syscall.SIGKILL)
+		if cmd.ProcessState == nil {
+			_ = cmd.Wait()
+		}
+	})
+
+	deadline := time.Now().Add(5 * time.Second)
+	for !leaderWaitableExited(cmd.Process.Pid) {
+		if time.Now().After(deadline) {
+			t.Fatalf("leader %d did not become waitable-exited within 5s", cmd.Process.Pid)
+		}
+		time.Sleep(10 * time.Millisecond)
+	}
+
+	treeErr := errors.New("process group could not be terminated")
+	originalTerminateOwnedProcess := terminateOwnedProcessForTest
+	terminateOwnedProcessForTest = func(got *exec.Cmd) (bool, error) {
+		if got != cmd {
+			t.Fatalf("terminate called with %p, want %p", got, cmd)
+		}
+		return true, treeErr
+	}
+	t.Cleanup(func() { terminateOwnedProcessForTest = originalTerminateOwnedProcess })
+
+	err = TerminateCommand(cmd)
+	if !errors.Is(err, treeErr) {
+		t.Fatalf("TerminateCommand error = %v, want live-group termination failure", err)
+	}
+	if cmd.ProcessState == nil {
+		t.Fatal("exited leader was not reaped")
+	}
+	if processStopped(childPID) {
+		t.Fatalf("child %d unexpectedly stopped; failure boundary needs a live group member", childPID)
+	}
+}
+
+func processStopped(pid int) bool {
+	if errors.Is(syscall.Kill(pid, syscall.Signal(0)), syscall.ESRCH) {
+		return true
+	}
+	state, err := exec.Command("ps", "-o", "stat=", "-p", strconv.Itoa(pid)).Output()
+	if err != nil {
+		return errors.Is(syscall.Kill(pid, syscall.Signal(0)), syscall.ESRCH)
+	}
+	return strings.HasPrefix(strings.TrimSpace(string(state)), "Z")
+}

internal/background/process_waitable_darwin.go

@@ -0,0 +1,28 @@
+//go:build darwin
+
+package background
+
+import "golang.org/x/sys/unix"
+
+// szomb is sys/proc.h SZOMB: the process has exited and is waiting to be reaped.
+const szomb int8 = 5
+
+// leaderWaitableExited reports whether pid has already exited and is waiting
+// to be reaped. This is the POSIX analogue of Windows GetExitCodeProcess
+// returning a value other than STILL_ACTIVE: the leader is a zombie, waitable
+// as exited, without collecting it. A true result does not mean the process
+// group is empty — descendants may still be running.
+//
+// Darwin has no /proc and x/sys/unix does not wrap waitid here, so the probe
+// is kern.proc.pid via sysctl. It does not use ps, which is the #862 failure
+// mode when PATH cannot resolve ps.
+func leaderWaitableExited(pid int) bool {
+	if pid <= 1 {
+		return false
+	}
+	kinfo, err := unix.SysctlKinfoProc("kern.proc.pid", pid)
+	if err != nil {
+		return false
+	}
+	return kinfo.Proc.P_stat == szomb
+}

internal/background/process_waitable_linux.go

@@ -0,0 +1,68 @@
+//go:build linux
+
+package background
+
+import (
+	"bytes"
+	"os"
+	"strconv"
+
+	"golang.org/x/sys/unix"
+)
+
+// POSIX waitid si_code values for an already-exited child (bits/waitflags.h).
+const (
+	cldExited int32 = 1
+	cldKilled int32 = 2
+	cldDumped int32 = 3
+)
+
+// leaderWaitableExited reports whether pid has already exited and is waiting
+// to be reaped. This is the POSIX analogue of Windows GetExitCodeProcess
+// returning a value other than STILL_ACTIVE: the leader is waitable as exited
+// (typically a zombie) without collecting it. A true result does not mean the
+// process group is empty — descendants may still be running.
+//
+// Detection does not use ps. waitid(WNOHANG|WNOWAIT) asks the kernel whether
+// an exit status is already available. /proc/<pid>/stat state Z is the
+// fallback when waitid cannot answer (for example if procfs is the only
+// usable source). Either path is independent of PATH, which is the #862
+// failure mode: empty PATH makes ps unresolvable and kill(0) still succeeds
+// against a zombie.
+func leaderWaitableExited(pid int) bool {
+	if pid <= 1 {
+		return false
+	}
+	if waitidAlreadyExited(pid) {
+		return true
+	}
+	return procIsZombie(pid)
+}
+
+func waitidAlreadyExited(pid int) bool {
+	var info unix.Siginfo
+	err := unix.Waitid(unix.P_PID, pid, &info, unix.WEXITED|unix.WNOHANG|unix.WNOWAIT, nil)
+	if err != nil {
+		return false
+	}
+	switch info.Code {
+	case cldExited, cldKilled, cldDumped:
+		return true
+	}
+	return false
+}
+
+func procIsZombie(pid int) bool {
+	data, err := os.ReadFile("/proc/" + strconv.Itoa(pid) + "/stat")
+	if err != nil {
+		return false
+	}
+	// /proc/<pid>/stat: pid (comm) state ... — comm may contain spaces or
+	// parentheses, so the state is the first field after the last ')'.
+	i := bytes.LastIndexByte(data, ')')
+	if i < 0 || i+1 >= len(data) {
+		return false
+	}
+	rest := bytes.TrimSpace(data[i+1:])
+	return len(rest) > 0 && rest[0] == 'Z'
+}

internal/background/process_waitable_other.go

@@ -0,0 +1,10 @@
+//go:build !windows && !linux && !darwin
+
+package background
+
+// leaderWaitableExited cannot positively identify a waitable-exited leader on
+// this platform without ps. Returning false keeps TerminateCommand's existing
+// conservative behaviour: a termination error is not discarded after reap.
+func leaderWaitableExited(pid int) bool {
+	return false
+}

internal/background/process_windows.go

@@ -3,11 +3,20 @@
 package background
 
 import (
+	"fmt"
 	"os/exec"
 
 	"github.com/Gitlawb/zero/internal/execution"
+	"golang.org/x/sys/windows"
 )
 
+// stillActiveExitCode is the documented GetExitCodeProcess value for a process
+// that has not terminated. x/sys/windows does not expose Windows' STILL_ACTIVE
+// constant.
+const stillActiveExitCode uint32 = 259
+
+var terminateProcessForTest = terminateProcess
+
 // 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).
@@ -16,3 +25,33 @@ func ConfigureChildProcessGroup(cmd *exec.Cmd) { execution.ConfigureProcessGroup
 func terminateProcess(pid int) error {
 	return execution.TerminateProcessTree(pid, 0, 0)
 }
+
+// terminateOwnedProcess asks taskkill /T to terminate the tree rooted at cmd,
+// with KillProcessTree's direct Process.Kill fallback if taskkill fails. taskkill
+// can discover descendants only while the root PID still exists; unlike a POSIX
+// process group, a Windows tree has no independently addressable identity after
+// its root exits.
+func terminateOwnedProcess(cmd *exec.Cmd) (bool, error) {
+	var (
+		alreadyExited bool
+		terminateErr  error
+	)
+	err := cmd.Process.WithHandle(func(handle uintptr) {
+		var exitCode uint32
+		if windows.GetExitCodeProcess(windows.Handle(handle), &exitCode) == nil {
+			alreadyExited = exitCode != stillActiveExitCode
+		}
+		// Keep the exact process identity pinned while the PID-based taskkill
+		// operation runs, preventing the PID from being recycled underneath it.
+		terminateErr = terminateProcessForTest(cmd.Process.Pid)
+	})
+	if err != nil {
+		return false, fmt.Errorf("pin process %d for termination: %w", cmd.Process.Pid, err)
+	}
+	return alreadyExited, terminateErr
+}
+
+// Windows has no persistent process-group identity to query after a dead root
+// is reaped. Preserve TerminateCommand's existing, documented dead-root
+// behavior; the stronger independent target check is available on POSIX.
+func terminationTargetGoneAfterReap(*exec.Cmd) bool { return true }

internal/background/process_windows_test.go

@@ -0,0 +1,106 @@
+//go:build windows
+
+package background
+
+import (
+	"bufio"
+	"errors"
+	"os/exec"
+	"strconv"
+	"strings"
+	"testing"
+	"time"
+)
+
+func TestTerminateCommandReturnsTreeFailureAfterRootFallbackReaps(t *testing.T) {
+	cmd := exec.Command("powershell.exe", "-NoProfile", "-NonInteractive", "-Command", "Start-Sleep -Seconds 300")
+	if err := cmd.Start(); err != nil {
+		t.Fatalf("start: %v", err)
+	}
+	treeErr := errors.New("taskkill /T failed")
+	originalTerminateProcess := terminateProcessForTest
+	terminateProcessForTest = func(pid int) error {
+		if err := cmd.Process.Kill(); err != nil {
+			return errors.Join(treeErr, err)
+		}
+		return treeErr
+	}
+	t.Cleanup(func() { terminateProcessForTest = originalTerminateProcess })
+
+	err := TerminateCommand(cmd)
+	if !errors.Is(err, treeErr) {
+		t.Fatalf("TerminateCommand error = %v, want tree failure", err)
+	}
+	if cmd.ProcessState == nil {
+		t.Fatal("root process was not reaped")
+	}
+}
+
+func TestTerminateCommandReapsExitedLeader(t *testing.T) {
+	cmd := exec.Command("cmd.exe", "/c", "exit", "0")
+	if err := cmd.Start(); err != nil {
+		t.Fatalf("start: %v", err)
+	}
+	// Let the child exit without calling Wait so cleanup exercises the Windows
+	// taskkill/TerminateProcess race while it still owns the process handle.
+	time.Sleep(500 * time.Millisecond)
+
+	// jatmn's #774 finding: taskkill /T racing an already-exited PID (and the
+	// Process.Kill fallback) both fail here, but the reap below still succeeds —
+	// the process-handle exit-code check proves the leader was already gone, so
+	// TerminateCommand must not surface the expected kill-attempt error.
+	if err := TerminateCommand(cmd); err != nil {
+		t.Fatalf("TerminateCommand: %v, want nil (leader was already gone and got reaped)", err)
+	}
+	if cmd.ProcessState == nil {
+		t.Fatal("exited process was not reaped")
+	}
+}
+
+func TestTerminateCommandDeadLeaderCannotIdentifyDescendantTree(t *testing.T) {
+	cmd := exec.Command("powershell.exe", "-NoProfile", "-NonInteractive", "-Command",
+		`$p = Start-Process powershell.exe -ArgumentList '-NoProfile','-Command','Start-Sleep -Seconds 300' -PassThru; $p.Id`)
+	stdout, err := cmd.StdoutPipe()
+	if err != nil {
+		t.Fatalf("stdout pipe: %v", err)
+	}
+	if err := cmd.Start(); err != nil {
+		t.Fatalf("start: %v", err)
+	}
+	line, err := bufio.NewReader(stdout).ReadString('\n')
+	if err != nil {
+		t.Fatalf("read descendant pid: %v", err)
+	}
+	descendantPID, err := strconv.Atoi(strings.TrimSpace(line))
+	if err != nil {
+		t.Fatalf("parse descendant pid %q: %v", line, err)
+	}
+	t.Cleanup(func() {
+		_ = exec.Command("taskkill.exe", "/F", "/PID", strconv.Itoa(descendantPID)).Run()
+	})
+
+	// Wait until the leader is observably exited but deliberately leave it
+	// unreaped, matching the POSIX regression's ordering.
+	deadline := time.Now().Add(5 * time.Second)
+	for processExistsOnWindows(cmd.Process.Pid) {
+		if time.Now().After(deadline) {
+			t.Fatal("leader did not exit")
+		}
+		time.Sleep(20 * time.Millisecond)
+	}
+	if err := TerminateCommand(cmd); err != nil {
+		t.Fatalf("TerminateCommand: %v", err)
+	}
+
+	// Windows has no persistent group ID analogous to POSIX: once the leader is
+	// gone, taskkill /T cannot rediscover this descendant. Assert that platform
+	// distinction rather than claiming TerminateCommand can provide it here.
+	if !processExistsOnWindows(descendantPID) {
+		t.Fatal("descendant unexpectedly stopped after its root exited")
+	}
+}
+
+func processExistsOnWindows(pid int) bool {
+	return exec.Command("powershell.exe", "-NoProfile", "-NonInteractive", "-Command",
+		"Get-Process -Id "+strconv.Itoa(pid)+" -ErrorAction Stop | Out-Null").Run() == nil
+}

internal/background/terminate.go

@@ -1,5 +1,18 @@
 package background
 
+import (
+	"errors"
+	"fmt"
+	"os/exec"
+	"time"
+)
+
+const commandReapTimeout = 3 * time.Second
+
+// terminateOwnedProcessForTest is a seam for the failure-boundary tests. Tests
+// replace it only while no other test in this package is running in parallel.
+var terminateOwnedProcessForTest = terminateOwnedProcess
+
 // TerminateProcess stops a background process by PID — on Windows its process
 // tree; on POSIX its whole process group when the PID leads its own group (the
 // invariant ConfigureChildProcessGroup establishes for processes started through
@@ -12,3 +25,77 @@ package background
 func TerminateProcess(pid int) error {
 	return terminateProcess(pid)
 }
+
+// TerminateCommand stops a started command and reaps its leader. The caller must
+// have exclusive ownership of cmd: it must not have previously called Wait or
+// Process.Release, and no goroutine may call either concurrently. On POSIX it
+// stops the whole group when the command was configured as its leader; ordinary
+// commands safely fall back to PID/tree discovery. When POSIX observes a
+// waitable-zombie leader before signalling, a termination error is discarded
+// after reap only if the launch-time signal target is independently gone; the
+// leader observation alone says nothing about descendants. Windows retains its
+// existing behavior of ignoring the expected kill-attempt error when
+// GetExitCodeProcess shows that the root was already dead. Windows cannot
+// rediscover a tree from a dead root, so descendants may survive there. `zero
+// daemon start` needs this operation when readiness times out: it launched the
+// child, so it must both stop the tree and collect the leader.
+//
+// The order matters: the tree is signalled first, because Wait releases the
+// leader's PID and a later group lookup could then resolve to nothing (or, worse,
+// to a recycled PID). Termination stays in the shared execution lifecycle
+// primitives rather than introducing a second platform-specific killer.
+//
+// If the bounded reap times out, the Wait call already started by this function
+// continues in its goroutine. From that point onward the caller must not access
+// cmd at all, including ProcessState, because that would race with Wait.
+func TerminateCommand(cmd *exec.Cmd) error {
+	if cmd == nil || cmd.Process == nil {
+		return errors.New("terminate command: process was never started")
+	}
+	leaderAlreadyExited, terminateErr := terminateOwnedProcessForTest(cmd)
+	reapErr := waitForTerminatedCommandWithin(cmd, commandReapTimeout)
+	if reapErr != nil {
+		if terminateErr != nil {
+			return fmt.Errorf("%v (reap failed: %w)", terminateErr, reapErr)
+		}
+		return reapErr
+	}
+	if terminateErr != nil {
+		if !leaderAlreadyExited || !terminationTargetGoneAfterReap(cmd) {
+			return terminateErr
+		}
+	}
+	// On POSIX, discard a termination error only when the leader was already
+	// exited and the launch-time signal target is independently gone after reap.
+	// A successful leader reap alone says nothing about live descendants.
+	// Windows retains its existing dead-root semantics because it has no
+	// persistent group identity to query after the root exits.
+	return nil
+}
+
+// classifyWaitError treats a non-zero exit status as success: a process being
+// terminated is expected to report one (or a signal), so the only interesting
+// failure is Wait itself not working.
+func classifyWaitError(waitErr error) error {
+	var exitErr *exec.ExitError
+	if waitErr != nil && !errors.As(waitErr, &exitErr) {
+		return fmt.Errorf("reap process: %w", waitErr)
+	}
+	return nil
+}
+
+// waitForTerminatedCommandWithin bounds the reap so a child that somehow survives
+// termination cannot hang the caller — the daemon-start cleanup path runs while a
+// user waits at the CLI. On timeout its Wait goroutine is intentionally left to
+// finish: exec.Cmd permits only one Wait call, and abandoning it without a reaper
+// would leak the child once it eventually exits.
+func waitForTerminatedCommandWithin(cmd *exec.Cmd, timeout time.Duration) error {
+	waitDone := make(chan error, 1)
+	go func() { waitDone <- cmd.Wait() }()
+	select {
+	case waitErr := <-waitDone:
+		return classifyWaitError(waitErr)
+	case <-time.After(timeout):
+		return fmt.Errorf("process %d did not reap after termination", cmd.Process.Pid)
+	}
+}

internal/cli/acp.go

@@ -10,9 +10,11 @@ import (
 	"github.com/Gitlawb/zero/internal/acp"
 	"github.com/Gitlawb/zero/internal/agent"
 	"github.com/Gitlawb/zero/internal/config"
+	"github.com/Gitlawb/zero/internal/execution"
+	"github.com/Gitlawb/zero/internal/mcp"
 	"github.com/Gitlawb/zero/internal/providermodeldiscovery"
+	"github.com/Gitlawb/zero/internal/redaction"
 	"github.com/Gitlawb/zero/internal/sandbox"
-	"github.com/Gitlawb/zero/internal/tools"
 )
 
 const acpUsage = `zero acp — serve the Agent Client Protocol (ACP) over stdio
@@ -57,20 +59,8 @@ func runACP(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) int
 		// surface — no ACP-specific credential handling needed.
 		NewProvider: deps.newProvider,
 		RunAgent:    agent.Run,
-		// Build the SCOPED registry + sandbox engine per workspace, exactly like the
-		// exec surface, so ACP shell/file tools are confined — never run unconfined.
-		BuildWorkspace: func(workspaceRoot string, resolved config.ResolvedConfig) (*tools.Registry, *sandbox.Engine, error) {
-			scope, err := sandbox.NewScope(workspaceRoot, resolved.Sandbox.AdditionalWriteRoots)
-			if err != nil {
-				return nil, nil, err
-			}
-			engine, err := buildExecSandboxEngine(workspaceRoot, resolved, deps, scope)
-			if err != nil {
-				return nil, nil, err
-			}
-			registry := newCoreRegistryScoped(workspaceRoot, scope)
-			registerLocalControlTools(registry, workspaceRoot, resolved.LocalControl)
-			return registry, engine, nil
+		BuildWorkspace: func(ctx context.Context, workspaceRoot string, resolved config.ResolvedConfig, mode agent.PermissionMode) (*acp.Workspace, error) {
+			return buildACPWorkspace(ctx, workspaceRoot, resolved, mode, deps)
 		},
 		ResolveWorkspaceRoot: acpWorkspaceRootResolver(deps),
 		Store:                deps.newSessionStore(),
@@ -85,6 +75,76 @@ func runACP(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) int
 	return exitSuccess
 }
 
+// buildACPWorkspace matches exec's registry construction for one ACP turn. MCP
+// servers use the same sandbox-prepared execution runner, project configuration is
+// gated by the validated workspace's trust state, and their runtime is released
+// when acp.Agent completes the turn. ACP deliberately stays at low autonomy: an
+// editor connection never upgrades MCP permissions beyond an interactive prompt.
+func buildACPWorkspace(ctx context.Context, workspaceRoot string, resolved config.ResolvedConfig, mode agent.PermissionMode, deps appDeps) (*acp.Workspace, error) {
+	scope, err := sandbox.NewScope(workspaceRoot, resolved.Sandbox.AdditionalWriteRoots)
+	if err != nil {
+		return nil, err
+	}
+	engine, err := buildExecSandboxEngine(workspaceRoot, resolved, deps, scope)
+	if err != nil {
+		return nil, err
+	}
+	registry := newCoreRegistryScoped(workspaceRoot, scope)
+
+	workspace := &acp.Workspace{
+		Registry: registry, Sandbox: engine,
+		DeferThreshold: resolved.Tools.DeferThreshold,
+	}
+	if mode != agent.PermissionModePlan {
+		// MCP stdio servers are subprocesses. Passing the engine to their runner
+		// keeps them inside the exact sandbox / lifecycle path used by zero exec.
+		runtime, trustSkip, err := registerMCPToolsForWorkspaceWithOptions(
+			ctx, workspaceRoot, registry, deps, mcp.AutonomyLow, workspaceRoot,
+			mcp.RegisterOptions{Execution: execution.NewRunner(engine), AdvertiseInAuto: true},
+		)
+		if err != nil {
+			// RegisterTools may have connected an earlier server before reporting a
+			// later failure, so do not orphan a partial runtime on this error path.
+			if runtime != nil {
+				_ = runtime.Close()
+			}
+			return nil, err
+		}
+		workspace.Cleanup = runtime.Close
+		workspace.Notices = acpMCPSetupNotices(trustSkip, runtime)
+	}
+	registerLocalControlTools(registry, workspaceRoot, resolved.LocalControl)
+	// MCP tools are deferred-eligible. Register their loader only after every
+	// ACP-visible tool is present, using the same mode the agent receives.
+	registerToolSearchIfEligible(registry, resolved.Tools.DeferThreshold, mode, nil, nil)
+	return workspace, nil
+}
+
+func acpMCPSetupNotices(skip trustSkip, runtime mcpToolRuntime) []string {
+	var notices []string
+	if skip.excludedProjectConfig {
+		if skip.trustCheckErrored {
+			notices = append(notices, "The workspace-trust store could not be read; project MCP servers were ignored (fail-closed). Run 'zero trust' to enable them.")
+		} else {
+			notices = append(notices, "Project MCP servers were ignored in this untrusted workspace. Run 'zero trust' to enable them.")
+		}
+	}
+	if runtime == nil {
+		return notices
+	}
+	for _, skipped := range runtime.Skipped() {
+		if skipped.UnconfiguredDefault {
+			continue
+		}
+		message := fmt.Sprintf("MCP server %s unavailable, skipped", skipped.Name)
+		if skipped.Err != nil {
+			message += ": " + redaction.ErrorMessage(skipped.Err, redaction.Options{})
+		}
+		notices = append(notices, redaction.RedactString(message, redaction.Options{}))
+	}
+	return notices
+}
+
 // acpWorkspaceRootResolver validates a client-supplied cwd into a confinement
 // root. It reuses exec's resolveWorkspaceRoot (abs+clean, must be an existing
 // dir) and additionally rejects the filesystem root and the home directory — an

internal/cli/acp_test.go

@@ -5,10 +5,19 @@ import (
 	"context"
 	"errors"
 	"io"
+	"os"
+	"path/filepath"
 	"strings"
 	"sync"
 	"testing"
 	"time"
+
+	"github.com/Gitlawb/zero/internal/agent"
+	"github.com/Gitlawb/zero/internal/config"
+	"github.com/Gitlawb/zero/internal/mcp"
+	"github.com/Gitlawb/zero/internal/sandbox"
+	"github.com/Gitlawb/zero/internal/tools"
+	"github.com/Gitlawb/zero/internal/workspacetrust"
 )
 
 type acpTestReader func([]byte) (int, error)
@@ -108,6 +117,198 @@ func TestRunACPIdleCancellationExitsCleanly(t *testing.T) {
 	}
 }
 
+func TestBuildACPWorkspaceRegistersSandboxedMCPToolsAndClosesThem(t *testing.T) {
+	setTrustConfigRoot(t)
+	workspaceRoot := t.TempDir()
+	if err := workspacetrust.Trust(workspaceRoot); err != nil {
+		t.Fatalf("trust workspace: %v", err)
+	}
+	grantStore, err := sandbox.NewGrantStore(sandbox.StoreOptions{FilePath: t.TempDir() + "/grants.json"})
+	if err != nil {
+		t.Fatalf("new grant store: %v", err)
+	}
+
+	var gotExclude, registered, closed bool
+	deps := fillAppDeps(appDeps{
+		resolveMCPConfig: func(root string, excludeProject bool) (config.MCPConfig, error) {
+			if root != workspaceRoot {
+				t.Fatalf("MCP workspace root = %q, want %q", root, workspaceRoot)
+			}
+			gotExclude = excludeProject
+			return config.MCPConfig{Servers: map[string]config.MCPServerConfig{
+				"docs": {Type: "stdio", Command: "fake-docs"},
+			}}, nil
+		},
+		newMCPStore:     func() (*mcp.PermissionStore, error) { return nil, nil },
+		newSandboxStore: func() (*sandbox.GrantStore, error) { return grantStore, nil },
+		registerMCPTools: func(ctx context.Context, registry *tools.Registry, cfg config.MCPConfig, options mcp.RegisterOptions) (mcpToolRuntime, error) {
+			if ctx == nil {
+				t.Fatal("MCP registration received nil context")
+			}
+			if len(cfg.Servers) != 1 || cfg.Servers["docs"].Command != "fake-docs" {
+				t.Fatalf("MCP config = %+v", cfg)
+			}
+			if options.Autonomy != mcp.AutonomyLow {
+				t.Fatalf("MCP autonomy = %q, want low", options.Autonomy)
+			}
+			if options.Execution == nil {
+				t.Fatal("MCP stdio server was not given the sandbox execution runner")
+			}
+			if options.WorkspaceRoot != workspaceRoot {
+				t.Fatalf("MCP execution workspace = %q, want %q", options.WorkspaceRoot, workspaceRoot)
+			}
+			if !options.AdvertiseInAuto {
+				t.Fatal("ACP MCP tools must be advertised in auto mode without being pre-approved")
+			}
+			registered = true
+			registry.Register(cliFakeDeferredTool{
+				name: "mcp_docs_search", permission: tools.PermissionPrompt,
+				advertiseInAuto: options.AdvertiseInAuto,
+			})
+			return closeFunc(func() error {
+				closed = true
+				return nil
+			}), nil
+		},
+	})
+
+	workspace, err := buildACPWorkspace(context.Background(), workspaceRoot, config.ResolvedConfig{
+		Tools: config.ToolsConfig{DeferThreshold: 1},
+	}, agent.PermissionModeAuto, deps)
+	if err != nil {
+		t.Fatalf("buildACPWorkspace: %v", err)
+	}
+	if gotExclude {
+		t.Fatal("trusted ACP workspace excluded project MCP configuration")
+	}
+	if !registered {
+		t.Fatal("ACP workspace did not register configured MCP tools")
+	}
+	if _, ok := workspace.Registry.Get("mcp_docs_search"); !ok {
+		t.Fatal("MCP tool is absent from ACP agent registry")
+	}
+	if _, ok := workspace.Registry.Get(tools.ToolSearchToolName); !ok {
+		t.Fatal("ACP registry is missing tool_search for deferred MCP tools")
+	}
+	if workspace.DeferThreshold != 1 {
+		t.Fatalf("workspace defer threshold = %d, want 1", workspace.DeferThreshold)
+	}
+	if err := workspace.Close(); err != nil {
+		t.Fatalf("close ACP workspace: %v", err)
+	}
+	if !closed {
+		t.Fatal("ACP workspace did not close its MCP runtime")
+	}
+}
+
+func TestBuildACPWorkspacePreservesRedactedMCPSetupNotices(t *testing.T) {
+	setTrustConfigRoot(t)
+	workspaceRoot := t.TempDir()
+	if err := os.MkdirAll(filepath.Join(workspaceRoot, ".zero"), 0o755); err != nil {
+		t.Fatal(err)
+	}
+	projectConfig := `{"mcp":{"servers":{"project-docs":{"type":"stdio","command":"project-docs"}}}}`
+	if err := os.WriteFile(filepath.Join(workspaceRoot, ".zero", "config.json"), []byte(projectConfig), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	secret := "ghp_ABCDEFGHIJKLMNOPQRSTUVWXYZ0123456789"
+	grantStore, err := sandbox.NewGrantStore(sandbox.StoreOptions{FilePath: filepath.Join(t.TempDir(), "grants.json")})
+	if err != nil {
+		t.Fatal(err)
+	}
+	deps := fillAppDeps(appDeps{
+		resolveMCPConfig: func(string, bool) (config.MCPConfig, error) {
+			return config.MCPConfig{Servers: map[string]config.MCPServerConfig{
+				"docs": {Type: "stdio", Command: "fake-docs"},
+			}}, nil
+		},
+		newMCPStore:     func() (*mcp.PermissionStore, error) { return nil, nil },
+		newSandboxStore: func() (*sandbox.GrantStore, error) { return grantStore, nil },
+		registerMCPTools: func(context.Context, *tools.Registry, config.MCPConfig, mcp.RegisterOptions) (mcpToolRuntime, error) {
+			return fakeMCPRuntimeWithSkips{skipped: []mcp.SkippedServer{{
+				Name: "docs", Err: errors.New("token=" + secret),
+			}}}, nil
+		},
+	})
+
+	workspace, err := buildACPWorkspace(context.Background(), workspaceRoot, config.ResolvedConfig{}, agent.PermissionModeAuto, deps)
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer workspace.Close()
+	joined := strings.Join(workspace.Notices, "\n")
+	if !strings.Contains(joined, "untrusted workspace") || !strings.Contains(joined, "MCP server docs unavailable") {
+		t.Fatalf("workspace notices = %#v", workspace.Notices)
+	}
+	if strings.Contains(joined, secret) || !strings.Contains(joined, "[REDACTED]") {
+		t.Fatalf("workspace notices leaked secret: %q", joined)
+	}
+}
+
+func TestBuildACPWorkspaceSkipsMCPInPlanMode(t *testing.T) {
+	grantStore, err := sandbox.NewGrantStore(sandbox.StoreOptions{FilePath: t.TempDir() + "/grants.json"})
+	if err != nil {
+		t.Fatalf("new grant store: %v", err)
+	}
+	called := false
+	deps := fillAppDeps(appDeps{
+		resolveMCPConfig: func(string, bool) (config.MCPConfig, error) {
+			called = true
+			return config.MCPConfig{}, nil
+		},
+		newSandboxStore: func() (*sandbox.GrantStore, error) { return grantStore, nil },
+		registerMCPTools: func(context.Context, *tools.Registry, config.MCPConfig, mcp.RegisterOptions) (mcpToolRuntime, error) {
+			called = true
+			return nil, nil
+		},
+	})
+
+	workspace, err := buildACPWorkspace(context.Background(), t.TempDir(), config.ResolvedConfig{}, agent.PermissionModePlan, deps)
+	if err != nil {
+		t.Fatalf("buildACPWorkspace: %v", err)
+	}
+	if called {
+		t.Fatal("plan-mode ACP workspace must not resolve or start MCP servers")
+	}
+	if err := workspace.Close(); err != nil {
+		t.Fatalf("close plan workspace: %v", err)
+	}
+}
+
+func TestBuildACPWorkspaceClosesPartialMCPRuntimeOnRegistrationError(t *testing.T) {
+	grantStore, err := sandbox.NewGrantStore(sandbox.StoreOptions{FilePath: t.TempDir() + "/grants.json"})
+	if err != nil {
+		t.Fatalf("new grant store: %v", err)
+	}
+	closed := false
+	deps := fillAppDeps(appDeps{
+		resolveMCPConfig: func(string, bool) (config.MCPConfig, error) {
+			return config.MCPConfig{Servers: map[string]config.MCPServerConfig{
+				"partial": {Type: "stdio", Command: "fake-partial"},
+			}}, nil
+		},
+		newMCPStore:     func() (*mcp.PermissionStore, error) { return nil, nil },
+		newSandboxStore: func() (*sandbox.GrantStore, error) { return grantStore, nil },
+		registerMCPTools: func(context.Context, *tools.Registry, config.MCPConfig, mcp.RegisterOptions) (mcpToolRuntime, error) {
+			return closeFunc(func() error {
+				closed = true
+				return nil
+			}), errors.New("second MCP server failed")
+		},
+	})
+
+	workspace, err := buildACPWorkspace(context.Background(), t.TempDir(), config.ResolvedConfig{}, agent.PermissionModeAuto, deps)
+	if err == nil {
+		t.Fatal("buildACPWorkspace succeeded after MCP registration error")
+	}
+	if workspace != nil {
+		t.Fatal("buildACPWorkspace returned a workspace after MCP registration error")
+	}
+	if !closed {
+		t.Fatal("partial MCP runtime was not closed after registration error")
+	}
+}
+
 type acpNotifyingReadCloser struct {
 	io.ReadCloser
 	readStarted chan<- struct{}

internal/cli/app.go

@@ -64,38 +64,53 @@ type appDeps struct {
 	// 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
-	openRouterLogin        func(context.Context, provideroauth.OpenRouterOptions) (string, error)
-	newSessionStore        func() *sessions.Store
-	loadPlugins            func(plugins.LoadOptions) (plugins.LoadResult, error)
-	loadHooks              func(hooks.LoadOptions) (hooks.LoadResult, error)
-	skillsDir              func() string
-	pluginsDir             func() string
-	toolsDir               func() string
-	newMCPStore            func() (*mcp.PermissionStore, error)
-	newMCPTokenStore       func() (*mcp.TokenStore, error)
-	newSandboxStore        func() (*sandbox.GrantStore, error)
-	selectSandboxBackend   func(sandbox.BackendOptions) sandbox.Backend
-	runSandboxSetupHelper  func(path string, args []string, stdout io.Writer, stderr io.Writer) error
-	registerMCPTools       func(context.Context, *tools.Registry, config.MCPConfig, mcp.RegisterOptions) (mcpToolRuntime, error)
-	prepareWorktree        func(context.Context, worktrees.Options) (worktrees.Result, error)
-	releaseWorktree        func(context.Context, worktrees.Options, string) error
-	detectVerifyPlan       func(string) (verify.Plan, error)
-	runVerify              func(context.Context, verify.Plan, verify.RunOptions) verify.Report
-	runSelfVerify          func(context.Context, verify.Plan, selfverify.Options) selfverify.Report
-	runAgentEval           func(context.Context, agentEvalOptions) (agentEvalReport, error)
-	inspectChanges         func(context.Context, zerogit.InspectOptions) (zerogit.ChangeSummary, error)
-	commitChanges          func(context.Context, zerogit.CommitOptions) (zerogit.CommitResult, error)
-	pushChanges            func(context.Context, zerogit.PushOptions) (zerogit.PushResult, error)
-	createPR               func(context.Context, zerogit.PROptions) (zerogit.PRResult, error)
-	runTUI                 func(context.Context, tui.Options) int
-	runEditor              func(string) error
-	checkUpdate            func(context.Context, update.Options) (update.Result, error)
-	applyUpdate            func(context.Context, update.Options) (update.ApplyResult, error)
-	now                    func() time.Time
+	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
+	openRouterLogin              func(context.Context, provideroauth.OpenRouterOptions) (string, error)
+	newSessionStore              func() *sessions.Store
+	loadPlugins                  func(plugins.LoadOptions) (plugins.LoadResult, error)
+	loadHooks                    func(hooks.LoadOptions) (hooks.LoadResult, error)
+	skillsDir                    func() string
+	pluginsDir                   func() string
+	toolsDir                     func() string
+	newMCPStore                  func() (*mcp.PermissionStore, error)
+	newMCPTokenStore             func() (*mcp.TokenStore, error)
+	newSandboxStore              func() (*sandbox.GrantStore, error)
+	selectSandboxBackend         func(sandbox.BackendOptions) sandbox.Backend
+	runSandboxSetupHelper        func(path string, args []string, stdout io.Writer, stderr io.Writer) error
+	registerMCPTools             func(context.Context, *tools.Registry, config.MCPConfig, mcp.RegisterOptions) (mcpToolRuntime, error)
+	prepareWorktree              func(context.Context, worktrees.Options) (worktrees.Result, error)
+	releaseWorktree              func(context.Context, worktrees.Options, string) error
+	detectVerifyPlan             func(string) (verify.Plan, error)
+	runVerify                    func(context.Context, verify.Plan, verify.RunOptions) verify.Report
+	runSelfVerify                func(context.Context, verify.Plan, selfverify.Options) selfverify.Report
+	runAgentEval                 func(context.Context, agentEvalOptions) (agentEvalReport, error)
+	inspectChanges               func(context.Context, zerogit.InspectOptions) (zerogit.ChangeSummary, error)
+	commitChanges                func(context.Context, zerogit.CommitOptions) (zerogit.CommitResult, error)
+	pushChanges                  func(context.Context, zerogit.PushOptions) (zerogit.PushResult, error)
+	createPR                     func(context.Context, zerogit.PROptions) (zerogit.PRResult, error)
+	createBranch                 func(context.Context, zerogit.BranchOptions) (zerogit.BranchResult, error)
+	isDefaultBranch              func(context.Context, zerogit.DefaultBranchOptions) (bool, string, string, error)
+	currentGitUser               func(context.Context, string) string
+	headCommitSubject            func(context.Context, string) string
+	commitsAhead                 func(context.Context, string, string, string) (int, error)
+	isUnbornRemote               func(context.Context, string, string) (bool, error)
+	refreshTrackingRef           func(context.Context, string, string, string) error
+	branchUpstreamRemote         func(context.Context, string, string) string
+	branchUpstreamRemoteAndMerge func(context.Context, string, string) (string, string)
+	resolveRemoteBranchTip       func(context.Context, string, string, string) (string, error)
+	remoteHasBranch              func(context.Context, string, string, string) (bool, error)
+	currentGitBranch             func(context.Context, string) string
+	currentBranchTip             func(context.Context, string) string
+	deleteBranch                 func(context.Context, string, string, string) error
+	resetBranchRef               func(context.Context, string, string, string, string) error
+	runTUI                       func(context.Context, tui.Options) int
+	runEditor                    func(string) error
+	checkUpdate                  func(context.Context, update.Options) (update.Result, error)
+	applyUpdate                  func(context.Context, update.Options) (update.ApplyResult, error)
+	now                          func() time.Time
 }
 
 type mcpToolRuntime interface {
@@ -198,11 +213,52 @@ func defaultAppDeps() appDeps {
 		commitChanges:    zerogit.Commit,
 		pushChanges:      zerogit.Push,
 		createPR:         zerogit.CreatePR,
-		runTUI:           tui.Run,
-		runEditor:        openEditor,
-		checkUpdate:      update.Check,
-		applyUpdate:      update.Apply,
-		now:              time.Now,
+		createBranch:     zerogit.CreateBranch,
+		isDefaultBranch:  zerogit.IsDefaultBranch,
+		currentGitUser: func(ctx context.Context, cwd string) string {
+			return zerogit.CurrentGitUser(ctx, cwd, nil)
+		},
+		headCommitSubject: func(ctx context.Context, cwd string) string {
+			return zerogit.HeadCommitSubject(ctx, cwd, nil)
+		},
+		commitsAhead: func(ctx context.Context, cwd, remote, branch string) (int, error) {
+			return zerogit.CommitsAhead(ctx, cwd, remote, branch, nil)
+		},
+		isUnbornRemote: func(ctx context.Context, cwd, remote string) (bool, error) {
+			return zerogit.IsUnbornRemote(ctx, cwd, remote, nil)
+		},
+		refreshTrackingRef: func(ctx context.Context, cwd, remote, branch string) error {
+			return zerogit.RefreshTrackingRef(ctx, cwd, remote, branch, nil)
+		},
+		branchUpstreamRemote: func(ctx context.Context, cwd, branch string) string {
+			return zerogit.UpstreamRemote(ctx, cwd, branch, nil)
+		},
+		branchUpstreamRemoteAndMerge: func(ctx context.Context, cwd, branch string) (string, string) {
+			return zerogit.UpstreamRemoteAndMergeBranch(ctx, cwd, branch, nil)
+		},
+		resolveRemoteBranchTip: func(ctx context.Context, cwd, remote, branch string) (string, error) {
+			return zerogit.ResolveRemoteBranchTip(ctx, cwd, remote, branch, nil)
+		},
+		remoteHasBranch: func(ctx context.Context, cwd, remote, branch string) (bool, error) {
+			return zerogit.RemoteHasBranch(ctx, cwd, remote, branch, nil)
+		},
+		currentGitBranch: func(ctx context.Context, cwd string) string {
+			return zerogit.CurrentBranch(ctx, cwd, nil)
+		},
+		currentBranchTip: func(ctx context.Context, cwd string) string {
+			return zerogit.CurrentBranchTip(ctx, cwd, nil)
+		},
+		deleteBranch: func(ctx context.Context, cwd, fallbackBranch, branchToDelete string) error {
+			return zerogit.DeleteBranch(ctx, cwd, fallbackBranch, branchToDelete, nil)
+		},
+		resetBranchRef: func(ctx context.Context, cwd, branch, newTip, expectedOld string) error {
+			return zerogit.ResetBranchRef(ctx, cwd, branch, newTip, nil, expectedOld)
+		},
+		runTUI:      tui.Run,
+		runEditor:   openEditor,
+		checkUpdate: update.Check,
+		applyUpdate: update.Apply,
+		now:         time.Now,
 	}
 }
 
@@ -255,40 +311,35 @@ func runWithDeps(args []string, stdout io.Writer, stderr io.Writer, deps appDeps
 	// cache file on the machine. The refresh itself is fired in exec/TUI startup.
 	modelregistry.EnableModelsDevOverlay()
 
-	addDirs, args, err := splitLeadingAddDirFlags(args)
+	// --add-dir, --theme and --allow-escalation may be written in any order;
+	// see splitLeadingRootFlags for why they are split together.
+	root, args, err := splitLeadingRootFlags(args)
 	if err != nil {
 		return writeAppError(stderr, err.Error(), 1)
 	}
-	// --theme <name> selects the TUI palette non-interactively (auto or any registered
-	// theme; populates tui.Options.Theme, which resolveThemeMode prefers over
-	// ZERO_THEME). Re-split --add-dir afterward so it may appear on either side of --theme.
-	theme, args, err := splitLeadingThemeFlag(args)
-	if err != nil {
-		return writeAppError(stderr, err.Error(), 1)
-	}
-	moreDirs, args, err := splitLeadingAddDirFlags(args)
-	if err != nil {
-		return writeAppError(stderr, err.Error(), 1)
-	}
-	addDirs = append(addDirs, moreDirs...)
+	addDirs, theme, allowEscalation := root.addDirs, root.theme, root.allowEscalation
 
 	if len(args) == 0 {
-		return runInteractiveTUI(stderr, deps, agent.PermissionModeAsk, addDirs, theme)
+		return runInteractiveTUI(stderr, deps, agent.PermissionModeAsk, addDirs, theme, allowEscalation)
 	}
 
-	// --add-dir grants an extra write root, and only the interactive TUI and
-	// exec dispatch paths consume one. Fail loud everywhere else rather than
-	// silently discarding an explicit grant — including help/version, which
-	// run no agent and could only ignore it. The allowlist names exactly the
-	// cases below that forward addDirs; a future subcommand is rejected by
-	// default until it opts in here.
-	if len(addDirs) > 0 {
-		switch args[0] {
-		case "--skip-permissions-unsafe", "-p", "--prompt", "exec":
-			// Forwarded by the matching case below.
-		default:
-			return writeAppError(stderr, "--add-dir is only supported for the interactive TUI and exec", 1)
-		}
+	// --add-dir grants an extra write root and --allow-escalation opts a run
+	// into mid-run model escalation; only the interactive TUI and exec consume
+	// either. Both are forwarded to exec below and rejected loudly everywhere
+	// else rather than silently discarded, including help/version, which run
+	// no agent and could only ignore them. The allowlist names exactly the
+	// cases below that forward; a future subcommand is rejected by default
+	// until it opts in here.
+	forwardsRootFlags := false
+	switch args[0] {
+	case "--skip-permissions-unsafe", "-p", "--prompt", "exec":
+		forwardsRootFlags = true
+	}
+	if len(addDirs) > 0 && !forwardsRootFlags {
+		return writeAppError(stderr, "--add-dir is only supported for the interactive TUI and exec", 1)
+	}
+	if allowEscalation && !forwardsRootFlags {
+		return writeAppError(stderr, "--allow-escalation is only supported for the interactive TUI and exec", 1)
 	}
 
 	switch args[0] {
@@ -298,26 +349,24 @@ func runWithDeps(args []string, stdout io.Writer, stderr io.Writer, deps appDeps
 		// reach unsafe mode in the shell — and the "!" shell escape (which is
 		// gated behind unsafe) was therefore unreachable.
 		//
-		// --add-dir may legally appear on either side of the flag, so re-split
-		// the remaining args and merge with the dirs already collected. Any
-		// trailing non-flag args were ignored on this path before --add-dir
-		// existed and still are — but an --add-dir hidden BEHIND one would be
-		// silently dropped with them, so reject that misplacement loudly.
-		moreDirs, rest, err := splitLeadingAddDirFlags(args[1:])
+		// The root flags may legally appear on either side of this flag, so
+		// split them again from what follows it and merge with what the root
+		// already took. Any trailing non-flag args were ignored on this path
+		// before --add-dir existed and still are, but an --add-dir hidden
+		// BEHIND one would be silently dropped with them, so reject that
+		// misplacement loudly below.
+		more, rest, err := splitLeadingRootFlags(args[1:])
 		if err != nil {
 			return writeAppError(stderr, err.Error(), 1)
 		}
-		// --theme may appear here too; extract it before the stray-arg checks so it is
-		// not rejected as an unexpected positional, then re-split --add-dir after it.
-		skipTheme, rest, err := splitLeadingThemeFlag(rest)
-		if err != nil {
-			return writeAppError(stderr, err.Error(), 1)
+		moreDirs := more.addDirs
+		// A --theme written before the flag was taken at the root and used to be
+		// dropped here; one written after it wins, as the last occurrence does.
+		skipTheme := theme
+		if more.theme != "" {
+			skipTheme = more.theme
 		}
-		evenMoreDirs, rest, err := splitLeadingAddDirFlags(rest)
-		if err != nil {
-			return writeAppError(stderr, err.Error(), 1)
-		}
-		moreDirs = append(moreDirs, evenMoreDirs...)
+		skipAllowEscalation := more.allowEscalation
 		// A misplaced --add-dir anywhere in the remainder is the more specific error,
 		// so check for it across all of rest before rejecting stray args.
 		for _, arg := range rest {
@@ -334,7 +383,7 @@ func runWithDeps(args []string, stdout io.Writer, stderr io.Writer, deps appDeps
 				return writeAppError(stderr, "--skip-permissions-unsafe launches the interactive TUI and takes no prompt or subcommand; for a one-shot unsafe run use `zero exec --skip-permissions-unsafe -p \"...\"`", 1)
 			}
 		}
-		return runInteractiveTUI(stderr, deps, agent.PermissionModeUnsafe, append(append([]string{}, addDirs...), moreDirs...), skipTheme)
+		return runInteractiveTUI(stderr, deps, agent.PermissionModeUnsafe, append(append([]string{}, addDirs...), moreDirs...), skipTheme, allowEscalation || skipAllowEscalation)
 	case "-h", "--help", "help":
 		if err := writeHelp(stdout); err != nil {
 			return 1
@@ -367,16 +416,16 @@ func runWithDeps(args []string, stdout io.Writer, stderr io.Writer, deps appDeps
 		if len(args) < 2 {
 			return writePromptRequired(stderr)
 		}
-		// Forward leading --add-dir occurrences so exec's own parser collects them.
+		// Forward the root flags exec consumes so its own parser collects them.
 		// Use the inline --prompt=<value> form so a prompt whose first character is a
 		// dash (e.g. `zero -p "-foo"`) is taken verbatim instead of being mistaken for
 		// a flag and rejected with "--prompt requires a value" (matches the cron path).
-		execArgs := append(addDirFlagArgs(addDirs), "--prompt="+args[1])
+		execArgs := append(rootFlagArgs(addDirs, allowEscalation), "--prompt="+args[1])
 		execArgs = append(execArgs, args[2:]...)
 		return runExec(execArgs, stdout, stderr, deps)
 	case "exec":
-		// Forward leading --add-dir occurrences so exec's own parser collects them.
-		return runExec(append(addDirFlagArgs(addDirs), args[1:]...), stdout, stderr, deps)
+		// Forward the root flags exec consumes so its own parser collects them.
+		return runExec(append(rootFlagArgs(addDirs, allowEscalation), args[1:]...), stdout, stderr, deps)
 	case "completions":
 		return runCompletions(args[1:], stdout, stderr)
 	case "daemon":
@@ -562,6 +611,45 @@ func fillAppDeps(deps appDeps) appDeps {
 	if deps.createPR == nil {
 		deps.createPR = defaults.createPR
 	}
+	if deps.createBranch == nil {
+		deps.createBranch = defaults.createBranch
+	}
+	if deps.isDefaultBranch == nil {
+		deps.isDefaultBranch = defaults.isDefaultBranch
+	}
+	if deps.currentGitUser == nil {
+		deps.currentGitUser = defaults.currentGitUser
+	}
+	if deps.headCommitSubject == nil {
+		deps.headCommitSubject = defaults.headCommitSubject
+	}
+	if deps.commitsAhead == nil {
+		deps.commitsAhead = defaults.commitsAhead
+	}
+	if deps.isUnbornRemote == nil {
+		deps.isUnbornRemote = defaults.isUnbornRemote
+	}
+	if deps.refreshTrackingRef == nil {
+		deps.refreshTrackingRef = defaults.refreshTrackingRef
+	}
+	if deps.branchUpstreamRemote == nil {
+		deps.branchUpstreamRemote = defaults.branchUpstreamRemote
+	}
+	if deps.branchUpstreamRemoteAndMerge == nil {
+		deps.branchUpstreamRemoteAndMerge = defaults.branchUpstreamRemoteAndMerge
+	}
+	if deps.remoteHasBranch == nil {
+		deps.remoteHasBranch = defaults.remoteHasBranch
+	}
+	if deps.currentGitBranch == nil {
+		deps.currentGitBranch = defaults.currentGitBranch
+	}
+	if deps.currentBranchTip == nil {
+		deps.currentBranchTip = defaults.currentBranchTip
+	}
+	// resolveRemoteBranchTip, deleteBranch, and resetBranchRef stay nil when
+	// unset so unit tests that mock createBranch without a real git tree do not
+	// hit real git restore/delete commands.
 	if deps.runTUI == nil {
 		deps.runTUI = defaults.runTUI
 	}
@@ -599,11 +687,11 @@ func fillAppDeps(deps appDeps) appDeps {
 	return deps
 }
 
-func runInteractiveTUI(stderr io.Writer, deps appDeps, permissionMode agent.PermissionMode, addDirs []string, theme string) int {
-	return runInteractiveTUIWithSetup(stderr, deps, permissionMode, addDirs, theme, false)
+func runInteractiveTUI(stderr io.Writer, deps appDeps, permissionMode agent.PermissionMode, addDirs []string, theme string, allowEscalation bool) int {
+	return runInteractiveTUIWithSetup(stderr, deps, permissionMode, addDirs, theme, false, allowEscalation)
 }
 
-func runInteractiveTUIWithSetup(stderr io.Writer, deps appDeps, permissionMode agent.PermissionMode, addDirs []string, theme string, forceSetup bool) int {
+func runInteractiveTUIWithSetup(stderr io.Writer, deps appDeps, permissionMode agent.PermissionMode, addDirs []string, theme string, forceSetup bool, allowEscalation bool) int {
 	// Refresh the models.dev pricing/limits cache in the background when stale;
 	// the overlay is read at registry construction from the cache file, so this
 	// benefits the next run and never blocks or fails this one.
@@ -695,6 +783,14 @@ func runInteractiveTUIWithSetup(stderr io.Writer, deps appDeps, permissionMode a
 
 	registry := newCoreRegistryScoped(workspaceRoot, scope)
 	registerLocalControlTools(registry, workspaceRoot, resolved.LocalControl)
+	// Mid-run model escalation is opt-in on this surface too. The tool is present
+	// only when the operator asked for it with --allow-escalation, and the
+	// switchers that make it do anything ride on the same flag through
+	// Options.AllowEscalation below. Registering one without the other ships a
+	// tool the loop will never act on, which looks like a feature and is not.
+	if allowEscalation {
+		registry.Register(tools.NewEscalateModelTool())
+	}
 	executionRunner := execution.NewRunner(nil)
 	sandboxStore, err := deps.newSandboxStore()
 	if err != nil {
@@ -744,24 +840,26 @@ func runInteractiveTUIWithSetup(stderr io.Writer, deps appDeps, permissionMode a
 		mcpTokenStore = nil
 		err = nil
 	}
+	criticalMCPConfig, optionalMCPConfig := splitMCPStartupConfig(mcpConfig)
 	mcpRuntime := mcpToolRuntime(noopMCPRuntime{})
-	if len(mcpConfig.Servers) > 0 {
-		mcpRuntime, err = deps.registerMCPTools(context.Background(), registry, mcpConfig, mcp.RegisterOptions{
+	if len(criticalMCPConfig.Servers) > 0 {
+		runtime, registerErr := deps.registerMCPTools(context.Background(), registry, criticalMCPConfig, mcp.RegisterOptions{
 			PermissionStore: mcpPermissionStore,
 			Autonomy:        mcp.AutonomyLow,
 			Execution:       executionRunner,
 			WorkspaceRoot:   workspaceRoot,
 		})
-	}
-	if err != nil {
-		closeMCPRuntime(stderr, mcpRuntime)
-		return writeAppError(stderr, err.Error(), 1)
+		if registerErr != nil {
+			closeMCPRuntime(stderr, runtime)
+			return writeAppError(stderr, redaction.ErrorMessage(registerErr, redaction.Options{}), 1)
+		}
+		mcpRuntime = runtime
 	}
 	defer closeMCPRuntime(stderr, mcpRuntime)
 	// A server that could not be reached or validated is skipped, not fatal (one
 	// bad MCP server must not abort startup) — surface each so a missing tool set is
 	// explained rather than silently absent. A built-in default the user never
-	// configured (e.g. keyless Firecrawl with no credentials) is the exception: it
+	// configured (e.g. keyless Exa with no credentials) is the exception: it
 	// was never asked for, so its failure is not worth a startup warning — only
 	// servers the user actually configured warn on failure (issue #552).
 	for _, skipped := range mcpRuntime.Skipped() {
@@ -814,11 +912,35 @@ func runInteractiveTUIWithSetup(stderr io.Writer, deps appDeps, permissionMode a
 	}
 	// Activate deferred MCP-tool loading for the interactive run only when the
 	// VISIBLE deferred-eligible count meets the resolved threshold, matching exec.
-	// The registry is complete (core + specialist + MCP + plugins) here, so the
-	// count is accurate; below threshold this is a no-op and the surface is
-	// unchanged. The interactive surface applies no operator tool filters, so
-	// enabled/disabled are nil — matching the AgentOptions below.
+	// This first pass covers every prompt-critical tool. Optional built-in MCP
+	// defaults re-run the same gate after publishing their atomic batch, so a
+	// later catalog generation cannot miss its loader. 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)
+	var optionalMCPRuntime *optionalMCPStartup
+	var awaitToolReadiness func(context.Context)
+	if len(optionalMCPConfig.Servers) > 0 {
+		optionalMCPRuntime = startOptionalMCP(
+			context.Background(),
+			registry,
+			optionalMCPConfig,
+			mcp.RegisterOptions{
+				PermissionStore: mcpPermissionStore,
+				Autonomy:        mcp.AutonomyLow,
+				Execution:       executionRunner,
+				WorkspaceRoot:   workspaceRoot,
+			},
+			deps.registerMCPTools,
+			func() {
+				registerToolSearchIfEligible(registry, resolved.Tools.DeferThreshold, permissionMode, nil, nil)
+			},
+		)
+		defer closeMCPRuntime(stderr, optionalMCPRuntime)
+		awaitToolReadiness = func(ctx context.Context) {
+			optionalMCPRuntime.Await(ctx, optionalMCPPromptGrace)
+		}
+	}
 	lastKnownMCPConfig := mcpConfig
 	fileTracker := tools.NewFileTracker()
 	var scratchBaseline scratchFileBaseline
@@ -842,6 +964,7 @@ func runInteractiveTUIWithSetup(stderr io.Writer, deps appDeps, permissionMode a
 		Version:              version,
 		Theme:                theme,
 		SavedTheme:           resolved.Preferences.Theme,
+		SavedPet:             resolved.Preferences.Pet,
 		UserConfigPath:       userConfigPath,
 		DoctorUserConfigPath: doctorUserConfigPath,
 		ProjectConfigPath:    projectConfigPath,
@@ -852,10 +975,20 @@ func runInteractiveTUIWithSetup(stderr io.Writer, deps appDeps, permissionMode a
 		FavoriteModels:       resolved.Preferences.FavoriteModels,
 		RecentModels:         resolved.Preferences.RecentModels,
 		RecapsEnabled:        resolved.Preferences.RecapsEnabled(),
+		CompactionModel:      resolved.Preferences.CompactionModel,
 		Provider:             provider,
 		NewProvider:          deps.newProvider,
-		ProbeProviderHealth:  deps.probeProviderHealth,
-		UserAgent:            userAgent(),
+		NewTurnSessionProvider: func(profile config.ProviderProfile, provider zeroruntime.Provider) zeroruntime.TurnSessionProvider {
+			if provider == nil {
+				return nil
+			}
+			if optimized, ok := providers.OptimizedTurnSessions(profile, provider, providers.Options{}); ok {
+				return optimized
+			}
+			return providers.DefaultTurnSessions(profile, provider, providers.Options{})
+		},
+		ProbeProviderHealth: deps.probeProviderHealth,
+		UserAgent:           userAgent(),
 		PrepareRunCompletionWarning: func() {
 			scratchBaseline = scratchFileSnapshot(workspaceRoot)
 		},
@@ -863,6 +996,7 @@ func runInteractiveTUIWithSetup(stderr io.Writer, deps appDeps, permissionMode a
 			return scratchFileWarning(workspaceRoot, scratchBaseline)
 		},
 		Registry:           registry,
+		AwaitToolReadiness: awaitToolReadiness,
 		SessionStore:       sessionStore,
 		PeerService:        peerService,
 		SandboxStore:       sandboxStore,
@@ -900,6 +1034,7 @@ func runInteractiveTUIWithSetup(stderr io.Writer, deps appDeps, permissionMode a
 			Specialists:    specialistRuntime.specialists,
 			Skills:         pluginActivation.skillInfos(deps.skillsDir()),
 		},
+		AllowEscalation: allowEscalation,
 		// LoadSkills backs /skills and direct /<skill-name> invocation in the TUI.
 		// It resolves against the same merged set (default dir + plugin skill
 		// roots) as the skill tool and the system-prompt list, re-read per use so
@@ -1245,6 +1380,7 @@ Flags:
   -v, --version                  Print version
   -p, --prompt                   Run a one-shot prompt
       --add-dir <path>           Allow writes in an extra directory (repeatable)
+      --allow-escalation         Let the agent escalate to a stronger model mid-run via escalate_model
       --skip-permissions-unsafe  Launch the interactive shell in unsafe mode (enables the ! shell escape)
 `)
 	return err
@@ -1261,6 +1397,17 @@ func addDirFlagArgs(addDirs []string) []string {
 	return flags
 }
 
+// rootFlagArgs re-synthesises the root flags exec consumes, in the spelling
+// its own parser accepts, so a flag written before the subcommand reaches
+// the run exactly as one written after it would.
+func rootFlagArgs(addDirs []string, allowEscalation bool) []string {
+	flags := addDirFlagArgs(addDirs)
+	if allowEscalation {
+		flags = append(flags, "--allow-escalation")
+	}
+	return flags
+}
+
 // splitLeadingAddDirFlags strips leading --add-dir flags from the root
 // argument list (zero --add-dir <path> [--add-dir <path>] [subcommand …]).
 // Subcommands like exec parse their own --add-dir occurrences.
@@ -1301,6 +1448,83 @@ func splitLeadingAddDirFlags(args []string) ([]string, []string, error) {
 	return addDirs, args, nil
 }
 
+// rootFlags is what the leading root flags amount to once every one of them
+// has been stripped from the front of the argument list.
+type rootFlags struct {
+	addDirs         []string
+	theme           string
+	allowEscalation bool
+}
+
+// splitLeadingRootFlags strips --add-dir, --theme and --allow-escalation from
+// the front of args in whatever order they were written, stopping at the
+// first token none of them claims.
+//
+// EACH SPLITTER STOPS AT THE FIRST TOKEN IT DOES NOT OWN, so running them once
+// in a fixed sequence made the order the operator wrote them in load-bearing:
+// a flag handled late in the sequence stranded every flag written after it as
+// an unknown command, and `zero --allow-escalation --theme auto` exited with
+// an argument error instead of launching. Running the sequence until it makes
+// no progress accepts every ordering, including a flag repeated on both sides
+// of another.
+func splitLeadingRootFlags(args []string) (rootFlags, []string, error) {
+	var flags rootFlags
+	for {
+		before := len(args)
+		addDirs, rest, err := splitLeadingAddDirFlags(args)
+		if err != nil {
+			return rootFlags{}, nil, err
+		}
+		flags.addDirs = append(flags.addDirs, addDirs...)
+		theme, rest, err := splitLeadingThemeFlag(rest)
+		if err != nil {
+			return rootFlags{}, nil, err
+		}
+		if theme != "" {
+			// The last occurrence wins across passes, as it does within one.
+			flags.theme = theme
+		}
+		allowEscalation, rest, err := splitLeadingAllowEscalationFlag(rest)
+		if err != nil {
+			return rootFlags{}, nil, err
+		}
+		flags.allowEscalation = flags.allowEscalation || allowEscalation
+		args = rest
+		if len(args) == before {
+			return flags, args, nil
+		}
+	}
+}
+
+// splitLeadingAllowEscalationFlag strips a leading --allow-escalation from the
+// root argument list, opting the interactive session into mid-run model
+// escalation.
+//
+// OPT-IN, THE SAME WAY exec IS. Escalation moves a run onto a different model,
+// which changes what the run costs and which provider sees the conversation, so
+// it is a decision the operator makes rather than a default. The exec flag
+// already answers this conservatively and the interactive surface should not
+// answer it differently.
+//
+// Bare flag only: repeating it is harmless, and an =value form is rejected so a
+// mistyped --allow-escalation=false is a loud error instead of silently enabling
+// the thing it was trying to turn off.
+func splitLeadingAllowEscalationFlag(args []string) (bool, []string, error) {
+	allow := false
+	for len(args) > 0 {
+		switch {
+		case args[0] == "--allow-escalation":
+			allow = true
+			args = args[1:]
+		case strings.HasPrefix(args[0], "--allow-escalation="):
+			return false, nil, errors.New("--allow-escalation takes no value; pass it bare to enable mid-run model escalation, or omit it")
+		default:
+			return allow, args, nil
+		}
+	}
+	return allow, args, nil
+}
+
 // splitLeadingThemeFlag strips a leading --theme <auto|theme-name> (space or =form)
 // from the root argument list and validates it against the registered themes. The
 // last occurrence wins. A value outside the allowed set is a loud error rather than

internal/cli/app_test.go

@@ -17,6 +17,7 @@ import (
 	"github.com/Gitlawb/zero/internal/agent"
 	"github.com/Gitlawb/zero/internal/config"
 	"github.com/Gitlawb/zero/internal/mcp"
+	"github.com/Gitlawb/zero/internal/redaction"
 	"github.com/Gitlawb/zero/internal/tools"
 	"github.com/Gitlawb/zero/internal/tui"
 	"github.com/Gitlawb/zero/internal/update"
@@ -211,12 +212,12 @@ func TestTUIStartupSuppressesWarningForUnconfiguredDefaultServer(t *testing.T) {
 		},
 		resolveMCPConfig: func(workspaceRoot string, excludeProject bool) (config.MCPConfig, error) {
 			return config.MCPConfig{Servers: map[string]config.MCPServerConfig{
-				"firecrawl": config.DefaultMCPServers()["firecrawl"],
+				"exa": config.DefaultMCPServers()["exa"],
 			}}, nil
 		},
 		registerMCPTools: func(context.Context, *tools.Registry, config.MCPConfig, mcp.RegisterOptions) (mcpToolRuntime, error) {
 			return fakeMCPRuntimeWithSkips{skipped: []mcp.SkippedServer{
-				{Name: "firecrawl", Err: errors.New("returned HTTP 401"), UnconfiguredDefault: true},
+				{Name: "exa", Err: errors.New("returned HTTP 401"), UnconfiguredDefault: true},
 			}}, nil
 		},
 		runTUI: func(ctx context.Context, options tui.Options) int {
@@ -227,7 +228,7 @@ func TestTUIStartupSuppressesWarningForUnconfiguredDefaultServer(t *testing.T) {
 	if exitCode != 0 {
 		t.Fatalf("exitCode = %d stderr=%s", exitCode, stderr.String())
 	}
-	if strings.Contains(stderr.String(), "firecrawl") {
+	if strings.Contains(stderr.String(), "exa") {
 		t.Fatalf("stderr = %q, want no warning for an unconfigured default server", stderr.String())
 	}
 }
@@ -486,6 +487,74 @@ func TestRunNoArgsLaunchesTUIWithMCPState(t *testing.T) {
 	}
 }
 
+func TestRunNoArgsPaintsBeforeOptionalDefaultMCPIsReady(t *testing.T) {
+	var stdout bytes.Buffer
+	var stderr bytes.Buffer
+	cwd := t.TempDir()
+	userConfigPath := filepath.Join(t.TempDir(), "zero", "config.json")
+	permissionStore, err := mcp.NewPermissionStore(mcp.StoreOptions{FilePath: filepath.Join(t.TempDir(), "mcp-permissions.json")})
+	if err != nil {
+		t.Fatalf("NewPermissionStore() error = %v", err)
+	}
+	tokenStore, err := mcp.NewTokenStore(mcp.TokenStoreOptions{FilePath: filepath.Join(t.TempDir(), "mcp-oauth.json")})
+	if err != nil {
+		t.Fatalf("NewTokenStore() error = %v", err)
+	}
+	started := make(chan struct{})
+	release := make(chan struct{})
+	configResult := make(chan error, 1)
+
+	exitCode := runWithDeps([]string{}, &stdout, &stderr, appDeps{
+		getwd:          func() (string, error) { return cwd, nil },
+		userConfigPath: func() (string, error) { return userConfigPath, nil },
+		resolveConfig: func(string, config.Overrides) (config.ResolvedConfig, error) {
+			return config.ResolvedConfig{MaxTurns: 8}, nil
+		},
+		resolveMCPConfig: func(string, bool) (config.MCPConfig, error) {
+			return config.MCPConfig{Servers: config.DefaultMCPServers()}, nil
+		},
+		newMCPStore:      func() (*mcp.PermissionStore, error) { return permissionStore, nil },
+		newMCPTokenStore: func() (*mcp.TokenStore, error) { return tokenStore, nil },
+		registerMCPTools: func(_ context.Context, registry *tools.Registry, cfg config.MCPConfig, _ mcp.RegisterOptions) (mcpToolRuntime, error) {
+			var configErr error
+			if _, ok := cfg.Servers["exa"]; !ok || len(cfg.Servers) != 1 {
+				configErr = fmt.Errorf("optional MCP config = %#v, want only exa", cfg.Servers)
+			}
+			configResult <- configErr
+			close(started)
+			<-release
+			registry.Register(cliFakeMCPRegistryTool{})
+			return noopMCPRuntime{}, nil
+		},
+		runTUI: func(ctx context.Context, options tui.Options) int {
+			select {
+			case <-started:
+			case <-time.After(time.Second):
+				t.Fatal("optional MCP initialization did not start")
+			}
+			if options.AwaitToolReadiness == nil {
+				t.Fatal("TUI did not receive optional tool readiness barrier")
+			}
+			if _, ok := options.Registry.Get("mcp_docs_lookup"); ok {
+				t.Fatal("optional MCP tool was published before registration completed")
+			}
+			close(release)
+			options.AwaitToolReadiness(ctx)
+			if _, ok := options.Registry.Get("mcp_docs_lookup"); !ok {
+				t.Fatal("optional MCP tool was not visible after readiness completed")
+			}
+			return 0
+		},
+	})
+
+	if exitCode != 0 {
+		t.Fatalf("exitCode = %d stderr=%s", exitCode, stderr.String())
+	}
+	if configErr := <-configResult; configErr != nil {
+		t.Fatal(configErr)
+	}
+}
+
 func TestTUIMCPCommandUsesLastGoodConfigOnRefreshError(t *testing.T) {
 	var stdout bytes.Buffer
 	var stderr bytes.Buffer
@@ -565,6 +634,7 @@ func TestRunNoArgsClosesPartialMCPRuntimeWhenRegistrationFails(t *testing.T) {
 		t.Fatalf("NewTokenStore() error = %v", err)
 	}
 	runtimeClosed := false
+	secret := "sk-proj-" + strings.Repeat("a", 23) + "0"
 
 	exitCode := runWithDeps([]string{}, &stdout, &stderr, appDeps{
 		getwd: func() (string, error) {
@@ -591,7 +661,7 @@ func TestRunNoArgsClosesPartialMCPRuntimeWhenRegistrationFails(t *testing.T) {
 			return closeFunc(func() error {
 				runtimeClosed = true
 				return nil
-			}), errors.New("register mcp tools failed")
+			}), fmt.Errorf("register mcp tools failed: %s", secret)
 		},
 		runTUI: func(ctx context.Context, options tui.Options) int {
 			t.Fatal("TUI should not launch after MCP registration fails")
@@ -605,6 +675,12 @@ func TestRunNoArgsClosesPartialMCPRuntimeWhenRegistrationFails(t *testing.T) {
 	if !strings.Contains(stderr.String(), "register mcp tools failed") {
 		t.Fatalf("stderr missing registration error: %s", stderr.String())
 	}
+	if strings.Contains(stderr.String(), secret) {
+		t.Fatalf("stderr leaked MCP registration secret: %s", stderr.String())
+	}
+	if !strings.Contains(stderr.String(), redaction.RedactedSecret) {
+		t.Fatalf("stderr missing redaction marker: %s", stderr.String())
+	}
 	if !runtimeClosed {
 		t.Fatal("partial MCP runtime was not closed after registration error")
 	}

internal/cli/auth.go

@@ -74,6 +74,8 @@ func runAuth(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) in
 		return runAuthStatus(args[1:], stdout, stderr, deps)
 	case "refresh":
 		return runAuthRefresh(args[1:], stdout, stderr, deps)
+	case "reset":
+		return runAuthReset(args[1:], stdout, stderr, deps)
 	case "openrouter":
 		return runAuthOpenRouter(args[1:], stdout, stderr, deps)
 	case "chatgpt":
@@ -261,6 +263,7 @@ type authArgs struct {
 	json       bool
 	device     bool
 	watch      bool
+	confirm    bool
 	scopes     []string
 	help       bool
 }
@@ -286,6 +289,8 @@ func parseAuthArgs(sub string, args []string) (authArgs, error) {
 			parsed.device = true
 		case arg == "--watch":
 			parsed.watch = true
+		case arg == "--confirm":
+			parsed.confirm = true
 		case arg == "--scope":
 			if i+1 >= len(args) {
 				return authArgs{}, fmt.Errorf("--scope requires a value")
@@ -321,6 +326,7 @@ func validateAuthFlags(sub string, a authArgs) error {
 		"logout":  {"json": true},
 		"status":  {"json": true},
 		"refresh": {"watch": true},
+		"reset":   {"json": true, "confirm": true},
 	}[sub]
 	bad := func(name string) error { return fmt.Errorf("zero auth %s does not accept %s", sub, name) }
 	if a.json && !allowed["json"] {
@@ -332,6 +338,9 @@ func validateAuthFlags(sub string, a authArgs) error {
 	if a.watch && !allowed["watch"] {
 		return bad("--watch")
 	}
+	if a.confirm && !allowed["confirm"] {
+		return bad("--confirm")
+	}
 	if len(a.scopes) > 0 && !allowed["scope"] {
 		return bad("--scope")
 	}
@@ -569,6 +578,43 @@ func runAuthRefreshWatch(manager *oauth.Manager, key, provider string, stdout io
 	return exitSuccess
 }
 
+func runAuthReset(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) int {
+	parsed, err := parseAuthArgs("reset", args)
+	if err != nil {
+		return writeExecUsageError(stderr, err.Error())
+	}
+	if parsed.help {
+		_ = writeAuthHelp(stdout)
+		return exitSuccess
+	}
+	if len(parsed.positional) > 0 {
+		return writeExecUsageError(stderr, fmt.Sprintf("zero auth reset takes no arguments (got %q)", parsed.positional[0]))
+	}
+	if !parsed.confirm {
+		return writeExecUsageError(stderr, "zero auth reset clears all stored OAuth logins and shared MCP tokens; rerun with --confirm to reset")
+	}
+	manager, err := newAuthManager(deps, stdout)
+	if err != nil {
+		return writeAppError(stderr, redaction.ErrorMessage(err, redaction.Options{}), exitCrash)
+	}
+	if err := manager.Reset(); err != nil {
+		return writeAppError(stderr, redaction.ErrorMessage(err, redaction.Options{}), exitCrash)
+	}
+	if parsed.json {
+		payload := struct {
+			Reset bool `json:"reset"`
+		}{Reset: true}
+		if err := writePrettyJSON(stdout, payload); err != nil {
+			return exitCrash
+		}
+		return exitSuccess
+	}
+	if _, err := fmt.Fprintln(stdout, "Reset OAuth token store."); err != nil {
+		return exitCrash
+	}
+	return exitSuccess
+}
+
 func filterAuthStatuses(statuses []oauth.Status, provider string) []oauth.Status {
 	want := oauth.ProviderKey(provider)
 	filtered := make([]oauth.Status, 0, 1)
@@ -589,6 +635,7 @@ Commands:
   logout <provider>                               Delete a provider's stored login
   status [provider]                               Show login presence/expiry (never the token)
   refresh <provider> [--watch]                    Force a token refresh (--watch keeps it fresh)
+  reset --confirm [--json]                        Clear all OAuth logins and shared MCP tokens (also repairs corrupted stores)
   openrouter                                      Log in to OpenRouter in the browser; mints an API key
   chatgpt                                         Log in to ChatGPT in the browser (Codex backend, ChatGPT Plus/Pro)
 
@@ -615,7 +662,7 @@ Flags:
       --device   Use the device-code flow (headless/SSH; no browser)
       --scope    Add an OAuth scope (repeatable)
       --watch    Keep the token fresh in the foreground (refresh only)
-      --json     Print result as JSON (status/logout)
+      --json     Print result as JSON (status/logout/reset)
   -h, --help     Show this help
 `)
 	return err

internal/cli/auth_test.go

@@ -129,6 +129,8 @@ func TestRunAuthRejectsWrongFlags(t *testing.T) {
 		{"auth", "logout", "demo", "--scope", "x"}, // scope is login-only
 		{"auth", "refresh", "demo", "--json"},      // json not for refresh
 		{"auth", "login", "demo", "--scope", ""},   // empty scope rejected
+		{"auth", "status", "--confirm"},            // confirm is reset-only
+		{"auth", "reset", "--confirm", "--watch"},  // confirmation does not allow unrelated flags
 	}
 	for _, args := range cases {
 		var stdout, stderr bytes.Buffer
@@ -196,13 +198,102 @@ func TestRunAuthHelp(t *testing.T) {
 	if code := runWithDeps([]string{"auth", "--help"}, &stdout, &stderr, appDeps{}); code != exitSuccess {
 		t.Fatalf("exit = %d", code)
 	}
-	for _, want := range []string{"zero auth", "login", "logout", "status", "refresh", "--device"} {
+	for _, want := range []string{"zero auth", "login", "logout", "status", "refresh", "reset --confirm", "--device"} {
 		if !strings.Contains(stdout.String(), want) {
 			t.Fatalf("help missing %q:\n%s", want, stdout.String())
 		}
 	}
 }
 
+func TestRunAuthResetRequiresConfirmation(t *testing.T) {
+	for _, args := range [][]string{{}, {"--json"}} {
+		t.Run(strings.Join(args, " "), func(t *testing.T) {
+			path := withAuthStore(t)
+			store, err := oauth.NewStore(oauth.StoreOptions{FilePath: path})
+			if err != nil {
+				t.Fatal(err)
+			}
+			if err := store.Save(oauth.ProviderKey("demo"), oauth.Token{AccessToken: "synthetic-reset-test-token"}); err != nil {
+				t.Fatal(err)
+			}
+			before, err := os.ReadFile(path)
+			if err != nil {
+				t.Fatal(err)
+			}
+			var stdout, stderr bytes.Buffer
+			code := runAuthReset(args, &stdout, &stderr, appDeps{})
+			after, readErr := os.ReadFile(path)
+			if readErr != nil || !bytes.Equal(before, after) {
+				t.Fatalf("unconfirmed reset changed the store: %v", readErr)
+			}
+			if code == exitSuccess || !strings.Contains(stderr.String(), "--confirm") || stdout.Len() != 0 {
+				t.Fatalf("unconfirmed reset: exit=%d stdout=%q stderr=%q", code, stdout.String(), stderr.String())
+			}
+		})
+	}
+}
+
+func TestRunAuthResetJSON(t *testing.T) {
+	path := withAuthStore(t)
+	store, err := oauth.NewStore(oauth.StoreOptions{FilePath: path})
+	if err != nil {
+		t.Fatalf("NewStore: %v", err)
+	}
+	if err := store.Save(oauth.ProviderKey("demo"), oauth.Token{AccessToken: "synthetic-reset-test-token"}); err != nil {
+		t.Fatalf("Save: %v", err)
+	}
+	var stdout, stderr bytes.Buffer
+	if code := runWithDeps([]string{"auth", "reset", "--confirm", "--json"}, &stdout, &stderr, appDeps{}); code != exitSuccess {
+		t.Fatalf("exit = %d stderr=%s", code, stderr.String())
+	}
+	var payload struct {
+		Reset bool `json:"reset"`
+	}
+	if err := json.Unmarshal(stdout.Bytes(), &payload); err != nil {
+		t.Fatalf("decode JSON: %v (stdout=%q)", err, stdout.String())
+	}
+	if !payload.Reset {
+		t.Fatalf("payload = %+v, want reset=true", payload)
+	}
+	stdout.Reset()
+	stderr.Reset()
+	if code := runWithDeps([]string{"auth", "status"}, &stdout, &stderr, appDeps{}); code != exitSuccess {
+		t.Fatalf("status exit = %d stderr=%s", code, stderr.String())
+	}
+	if !strings.Contains(stdout.String(), "No OAuth provider logins are stored.") {
+		t.Fatalf("expected empty store after JSON reset, got: %q", stdout.String())
+	}
+}
+
+func TestRunAuthReset(t *testing.T) {
+	path := withAuthStore(t)
+	store, err := oauth.NewStore(oauth.StoreOptions{FilePath: path})
+	if err != nil {
+		t.Fatalf("NewStore: %v", err)
+	}
+	if err := store.Save(oauth.ProviderKey("demo"), oauth.Token{AccessToken: "secret"}); err != nil {
+		t.Fatalf("Save: %v", err)
+	}
+
+	var stdout, stderr bytes.Buffer
+	if code := runWithDeps([]string{"auth", "reset", "--confirm"}, &stdout, &stderr, appDeps{}); code != exitSuccess {
+		t.Fatalf("exit = %d stderr=%s", code, stderr.String())
+	}
+	if !strings.Contains(stdout.String(), "Reset OAuth token store") {
+		t.Fatalf("stdout = %q", stdout.String())
+	}
+
+	// Verify store is now empty.
+	stdout.Reset()
+	stderr.Reset()
+	if code := runWithDeps([]string{"auth", "status"}, &stdout, &stderr, appDeps{}); code != exitSuccess {
+		t.Fatalf("status exit = %d", code)
+	}
+	if !strings.Contains(stdout.String(), "No OAuth provider logins are stored.") {
+		t.Fatalf("expected empty store after reset, got: %q", stdout.String())
+	}
+}
+
 // TestRunAuthLoginChatGPTRoutesToDedicatedFlow verifies `zero auth login
 // chatgpt` reaches the dedicated ChatGPT login (fixed-port loopback + mandatory
 // authorize params), not the generic manager path. The generic login accepts

internal/cli/command_center.go

@@ -407,7 +407,11 @@ func formatProviderCatalogLine(provider providerCatalogSummary) string {
 		provider.RuntimeSupported,
 	))
 	if provider.RuntimeSupported {
-		lines = append(lines, "    setup: zero providers setup "+displayCLIValue(provider.ID, "unknown")+" --set-active")
+		setup := "    setup: zero providers setup " + displayCLIValue(provider.ID, "unknown") + " --set-active"
+		if provider.ID == "atomic-chat-local" {
+			setup = "    setup: run zero setup to select a loaded model, or zero providers detect to get an add command"
+		}
+		lines = append(lines, setup)
 	} else {
 		lines = append(lines, "    unsupported: "+displayCLIValue(provider.RuntimeUnsupportedReason, "unknown"))
 	}
@@ -502,7 +506,7 @@ func writeProvidersHelp(w io.Writer) error {
   zero providers models [name] [flags]
 
 Inspects resolved provider profiles and provider catalog descriptors without printing secrets.
-Detect probes for running local runtimes (Ollama, LM Studio) and prints adopt commands plus per-provider next steps.
+Detect probes for running local runtimes (Ollama, LM Studio, Atomic Chat) and prints adopt commands plus per-provider next steps.
 Models probes a provider's live model-listing endpoint (e.g. an OpenAI-compatible /v1/models) and lists the models it serves — including custom OpenAI-/Anthropic-compatible endpoints — so a self-hosted provider needs no per-model config.
 
 Flags:

internal/cli/command_center_test.go

@@ -967,3 +967,10 @@ func TestProvidersListMarksOAuthLoginProviders(t *testing.T) {
 		t.Fatalf("list should render the oauth login state, got:\n%s", rendered)
 	}
 }
+
+func TestAtomicCatalogSetupRequiresLoadedModel(t *testing.T) {
+	out := formatProviderCatalogLine(providerCatalogSummary{ID: "atomic-chat-local", RuntimeSupported: true})
+	if strings.Contains(out, "zero providers setup atomic-chat-local --set-active") || !strings.Contains(out, "zero setup") {
+		t.Fatalf("catalog advertises a failing setup command: %s", out)
+	}
+}

internal/cli/completions.go

@@ -20,7 +20,7 @@ type completionContext struct {
 }
 
 var completionRoot = completionNode{
-	flags: []string{"-h", "--help", "-v", "--version", "-p", "--prompt", "--add-dir", "--theme", "--skip-permissions-unsafe"},
+	flags: []string{"-h", "--help", "-v", "--version", "-p", "--prompt", "--add-dir", "--allow-escalation", "--theme", "--skip-permissions-unsafe"},
 	children: []completionNode{
 		{names: []string{"exec"}, flags: []string{
 			"-h", "--help", "-f", "--file", "--image", "--add-dir", "--mode", "-m", "--model",
@@ -78,7 +78,8 @@ var completionRoot = completionNode{
 			{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{"auth"}, children: append(leafNodes("openrouter", "chatgpt", "login", "logout", "status", "refresh"),
+			completionNode{names: []string{"reset"}, flags: []string{"--confirm", "--json", "--help"}})},
 		{names: []string{"sandbox"}, children: []completionNode{
 			{names: []string{"policy"}}, {names: []string{"setup"}}, {names: []string{"check"}},
 			{names: []string{"grants"}, children: leafNodes("list", "allow", "deny", "revoke", "clear")},

internal/cli/completions_test.go

@@ -159,6 +159,8 @@ func TestCompletionTreeCoversAliasesNestingAndCommonFlags(t *testing.T) {
 	assertCandidates(t, byPath["worktree"], "prepare", "release")
 	assertCandidates(t, byPath["daemon"], "start", "stop", "status", "run", "attach")
 	assertCandidates(t, byPath["mcp oauth"], "login", "logout", "status")
+	assertCandidates(t, byPath["auth"], "openrouter", "chatgpt", "login", "logout", "status", "refresh", "reset")
+	assertCandidates(t, byPath["auth reset"], "--confirm", "--json", "--help")
 	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")

internal/cli/daemon.go

@@ -1,6 +1,7 @@
 package cli
 
 import (
+	"errors"
 	"fmt"
 	"io"
 	"os"
@@ -160,20 +161,61 @@ func runDaemonStartDetached(paths daemon.Paths, stdout io.Writer, stderr io.Writ
 	cmd.Stdout = logFile
 	cmd.Stderr = logFile
 	background.ConfigureChildProcessGroup(cmd) // own process group: outlives this shell
-	if err := cmd.Start(); err != nil {
+	if err := startAndAwaitDaemonProcess(cmd, func() bool { return daemonReachable(paths) }, 5*time.Second, 25*time.Millisecond); err != nil {
+		if errors.Is(err, errDaemonStartTimeout) {
+			return writeAppError(stderr, err.Error()+"; see "+logPath, exitCrash)
+		}
 		return writeAppError(stderr, err.Error(), exitCrash)
 	}
-	_ = cmd.Process.Release()
+	fmt.Fprintf(stdout, "zero daemon started (socket %s)\n", paths.Socket)
+	return exitSuccess
+}
 
-	deadline := time.Now().Add(5 * time.Second)
-	for time.Now().Before(deadline) {
-		if daemonReachable(paths) {
-			fmt.Fprintf(stdout, "zero daemon started (socket %s)\n", paths.Socket)
-			return exitSuccess
-		}
-		time.Sleep(25 * time.Millisecond)
+var errDaemonStartTimeout = errors.New("daemon did not come up within timeout")
+
+func startAndAwaitDaemonProcess(cmd *exec.Cmd, reachable func() bool, timeout time.Duration, pollInterval time.Duration) error {
+	if err := cmd.Start(); err != nil {
+		return err
 	}
-	return writeAppError(stderr, "daemon did not come up within timeout; see "+logPath, exitCrash)
+	if waitForDaemonReadiness(reachable, timeout, pollInterval) {
+		if err := cmd.Process.Release(); err != nil {
+			cleanupErr := terminateAndReapDaemonProcess(cmd)
+			if cleanupErr != nil {
+				return fmt.Errorf("release daemon process: %w (cleanup failed: %v)", err, cleanupErr)
+			}
+			return fmt.Errorf("release daemon process: %w", err)
+		}
+		return nil
+	}
+	if err := terminateAndReapDaemonProcess(cmd); err != nil {
+		return fmt.Errorf("%w; cleanup failed: %v", errDaemonStartTimeout, err)
+	}
+	return errDaemonStartTimeout
+}
+
+func waitForDaemonReadiness(reachable func() bool, timeout time.Duration, pollInterval time.Duration) bool {
+	deadline := time.Now().Add(timeout)
+	for {
+		if reachable() {
+			return true
+		}
+		remaining := time.Until(deadline)
+		if remaining <= 0 {
+			break
+		}
+		time.Sleep(min(pollInterval, remaining))
+	}
+	// The loop's last reachable() call already ran immediately before this one
+	// (right after the final Sleep, with the same deadline check separating
+	// them), so this adds no extra waiting time and is not a grace window for a
+	// slow-starting daemon — it only re-checks the same instant in case
+	// reachable() itself is non-deterministic. To tolerate a genuinely slow
+	// start, raise the timeout passed to startAndAwaitDaemonProcess instead.
+	return reachable()
+}
+
+func terminateAndReapDaemonProcess(cmd *exec.Cmd) error {
+	return background.TerminateCommand(cmd)
 }
 
 func runDaemonStop(args []string, stdout io.Writer, stderr io.Writer) int {

internal/cli/daemon_test.go

@@ -2,8 +2,15 @@ package cli
 
 import (
 	"bytes"
+	"errors"
+	"os"
+	"os/exec"
+	"path/filepath"
 	"strings"
 	"testing"
+	"time"
+
+	"github.com/Gitlawb/zero/internal/background"
 )
 
 // isolateDaemonPaths points DefaultPaths at a temp dir so the test never touches
@@ -121,3 +128,84 @@ func TestDaemonSubcommandsRejectExtraArgs(t *testing.T) {
 		}
 	}
 }
+
+func TestWaitForDaemonReadinessChecksTimeoutBoundary(t *testing.T) {
+	checks := 0
+	ready := waitForDaemonReadiness(func() bool {
+		checks++
+		return checks == 2
+	}, 0, time.Millisecond)
+	if !ready {
+		t.Fatal("daemon that became reachable at the timeout boundary was not detected")
+	}
+	if checks != 2 {
+		t.Fatalf("reachability checks = %d, want initial and final checks", checks)
+	}
+}
+
+func TestTerminateAndReapDaemonProcess(t *testing.T) {
+	cmd := daemonDetachedChildCommand("hang")
+	if err := cmd.Start(); err != nil {
+		t.Fatalf("start helper process: %v", err)
+	}
+
+	if err := terminateAndReapDaemonProcess(cmd); err != nil {
+		t.Fatalf("terminate and reap helper process: %v", err)
+	}
+	if cmd.ProcessState == nil {
+		t.Fatalf("helper process was not reaped: state=%v", cmd.ProcessState)
+	}
+	if cmd.ProcessState.Success() {
+		t.Fatalf("helper process unexpectedly exited successfully: %v", cmd.ProcessState)
+	}
+}
+
+func TestStartAndAwaitDaemonProcessTimeoutTerminatesAndReaps(t *testing.T) {
+	cmd := daemonDetachedChildCommand("hang")
+	if err := startAndAwaitDaemonProcess(cmd, func() bool { return false }, 0, time.Millisecond); !errors.Is(err, errDaemonStartTimeout) {
+		t.Fatalf("start and await error = %v, want timeout", err)
+	}
+	if cmd.ProcessState == nil {
+		t.Fatal("timed-out helper process was not reaped")
+	}
+}
+
+func TestStartAndAwaitDaemonProcessReleasesReadyChild(t *testing.T) {
+	marker := filepath.Join(t.TempDir(), "child-finished")
+	cmd := daemonDetachedChildCommand("mark", "ZERO_TEST_DAEMON_CHILD_MARKER="+marker)
+	if err := startAndAwaitDaemonProcess(cmd, func() bool { return true }, time.Second, time.Millisecond); err != nil {
+		t.Fatalf("start and await ready helper: %v", err)
+	}
+
+	deadline := time.Now().Add(5 * time.Second)
+	for time.Now().Before(deadline) {
+		if _, err := os.Stat(marker); err == nil {
+			return
+		}
+		time.Sleep(10 * time.Millisecond)
+	}
+	t.Fatal("released helper process did not remain alive to write its marker")
+}
+
+func daemonDetachedChildCommand(mode string, extraEnv ...string) *exec.Cmd {
+	cmd := exec.Command(os.Args[0], "-test.run=TestDaemonDetachedChildProcess")
+	cmd.Env = append(os.Environ(), "ZERO_TEST_DAEMON_DETACHED_CHILD="+mode)
+	cmd.Env = append(cmd.Env, extraEnv...)
+	background.ConfigureChildProcessGroup(cmd)
+	return cmd
+}
+
+func TestDaemonDetachedChildProcess(t *testing.T) {
+	switch os.Getenv("ZERO_TEST_DAEMON_DETACHED_CHILD") {
+	case "":
+		return
+	case "mark":
+		if err := os.WriteFile(os.Getenv("ZERO_TEST_DAEMON_CHILD_MARKER"), []byte("ready"), 0o600); err != nil {
+			os.Exit(2)
+		}
+		os.Exit(0)
+	}
+	for {
+		time.Sleep(time.Hour)
+	}
+}

internal/cli/deferred_wiring_test.go

@@ -18,14 +18,20 @@ import (
 // cliFakeDeferredTool is deferred-eligible (implements Deferred() bool), mirroring
 // an MCP registry tool, so it counts toward the deferral threshold.
 type cliFakeDeferredTool struct {
-	name string
+	name            string
+	permission      tools.Permission
+	advertiseInAuto bool
 }
 
 func (t cliFakeDeferredTool) Name() string             { return t.name }
 func (t cliFakeDeferredTool) Description() string      { return "fake deferred tool" }
 func (t cliFakeDeferredTool) Parameters() tools.Schema { return tools.Schema{Type: "object"} }
 func (t cliFakeDeferredTool) Safety() tools.Safety {
-	return tools.Safety{SideEffect: tools.SideEffectNetwork, Permission: tools.PermissionAllow}
+	permission := t.permission
+	if permission == "" {
+		permission = tools.PermissionAllow
+	}
+	return tools.Safety{SideEffect: tools.SideEffectNetwork, Permission: permission, AdvertiseInAuto: t.advertiseInAuto}
 }
 func (t cliFakeDeferredTool) Run(context.Context, map[string]any) tools.Result {
 	return tools.Result{Status: tools.StatusOK, Output: "ok"}

internal/cli/exec.go

@@ -442,57 +442,22 @@ func runExec(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) in
 	// the resolved model and is reassigned by the model switcher on a mid-run
 	// escalation so post-switch turns are attributed to the escalated model.
 	currentModel := resolved.Provider.Model
-	var modelSwitcher func(context.Context, string) (agent.Provider, error)
-	if options.allowEscalation {
-		modelSwitcher = func(_ context.Context, modelID string) (agent.Provider, error) {
-			// deps.newProvider is wrapped (fillAppDeps) to apply the stored key, so
-			// the escalated provider is authenticated even though resolved.Provider
-			// is the pure profile — no per-site key handling here.
-			switchedProfile := resolved.Provider
-			switchedProfile.Model = modelID
-			switchedProvider, err := deps.newProvider(switchedProfile)
-			if err != nil {
-				return nil, err
-			}
-			// Mirror the agent loop's switch guard (it only reassigns the provider
-			// when newProvider != nil). Updating currentModel only on a non-nil
-			// provider keeps usage attribution consistent with whether the loop
-			// actually switched — a (nil, nil) return leaves both untouched.
-			if switchedProvider != nil {
-				currentModel = modelID
-			}
-			return switchedProvider, nil
-		}
-	}
-
-	// Optimized OpenAI turn sessions (ZERO_OPENAI_TURN_SESSION, default off).
-	// nil when gated off or the profile is ineligible: agent.Run then wraps the
-	// provider in its default adapter — the exact code path of today. The
-	// session switcher is installed only when the run START is optimized, so
-	// the legacy ModelSwitcher path above stays untouched otherwise.
+	// Optimized OpenAI turn sessions (ZERO_OPENAI_TURN_SESSION, default off). nil
+	// when gated off or the profile is ineligible: agent.Run then wraps the
+	// provider in its default adapter. This is the run's STARTING session provider
+	// and is used whether or not escalation is enabled.
 	turnSessions, _ := providers.OptimizedTurnSessions(resolved.Provider, provider, providers.Options{})
+	// Both switchers come from one shared builder so exec and the interactive TUI
+	// cannot drift on the nil contracts the agent loop depends on. The session
+	// switcher is nil unless this run STARTED optimized, which is what keeps a
+	// default-adapter run on the default adapter.
+	var modelSwitcher func(context.Context, string) (agent.Provider, error)
 	var modelSessionSwitcher func(context.Context, string) (zeroruntime.TurnSessionProvider, error)
-	if options.allowEscalation && turnSessions != nil {
-		modelSessionSwitcher = func(_ context.Context, modelID string) (zeroruntime.TurnSessionProvider, error) {
-			switchedProfile := resolved.Provider
-			switchedProfile.Model = modelID
-			switchedProvider, err := deps.newProvider(switchedProfile)
-			if err != nil {
-				return nil, err
-			}
-			if switchedProvider == nil {
-				// The loop treats a nil session source as "no swap" — mirror the
-				// legacy closure's (nil, nil) contract.
-				return nil, nil
-			}
-			currentModel = modelID
-			if optimized, ok := providers.OptimizedTurnSessions(switchedProfile, switchedProvider, providers.Options{}); ok {
-				return optimized, nil
-			}
-			// Ineligible switch target: default adapter, but with the switched
-			// model's resolved capability projection preserved.
-			return providers.DefaultTurnSessions(switchedProfile, switchedProvider, providers.Options{}), nil
-		}
+	if options.allowEscalation {
+		modelSwitcher, modelSessionSwitcher = providers.EscalationSwitchers(
+			resolved.Provider, provider, deps.newProvider,
+			func(modelID string) { currentModel = modelID },
+		)
 	}
 
 	runMetadata, err := resolveExecRunMetadata(resolved.Provider)
@@ -682,16 +647,20 @@ func runExec(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) in
 		ModelSwitcher:        modelSwitcher,
 		TurnSessionProvider:  turnSessions,
 		ModelSessionSwitcher: modelSessionSwitcher,
-		ReasoningEffort:      forwardEffort,
-		Trace:                traceRecorder,
-		Cwd:                  workspaceRoot,
-		Images:               images,
-		Registry:             registry,
-		PermissionMode:       permissionMode,
-		Autonomy:             options.autonomy,
-		SelfCorrect:          selfCorrector,
-		FileDiagnostics:      fileDiagnostics,
-		Profile:              execProfile.Policy(displacedMaxTurns, execProfileFilledEffort),
+		Summarizer:           summarizerFactory(resolved, deps.newProvider),
+		ContextWindowFor: func(modelID string) int {
+			return modelregistry.AgentContextWindow(modelContextWindow(modelRegistry, modelID))
+		},
+		ReasoningEffort: forwardEffort,
+		Trace:           traceRecorder,
+		Cwd:             workspaceRoot,
+		Images:          images,
+		Registry:        registry,
+		PermissionMode:  permissionMode,
+		Autonomy:        options.autonomy,
+		SelfCorrect:     selfCorrector,
+		FileDiagnostics: fileDiagnostics,
+		Profile:         execProfile.Policy(displacedMaxTurns, execProfileFilledEffort),
 		// Headless exec: don't accept a no-tool-call turn as "done" while work
 		// clearly remains (pending plan items / a mid-step continuation cue) —
 		// nudge to continue, and finalize as INCOMPLETE rather than false success
@@ -939,6 +908,9 @@ func registerToolSearchIfEligible(registry *tools.Registry, deferThreshold int,
 	if deferThreshold <= 0 {
 		return
 	}
+	if _, registered := registry.Get(tools.ToolSearchToolName); registered {
+		return
+	}
 	if deferredEligibleCount(registry, permissionMode, enabledTools, disabledTools) < deferThreshold {
 		return
 	}

internal/cli/exec_writer.go

@@ -267,23 +267,29 @@ func streamJSONPermissionEventType(event agent.PermissionEvent) streamjson.Event
 func (writer *execEventWriter) usage(usage agent.Usage) {
 	if writer.format == execOutputJSON {
 		writer.writeJSON(map[string]any{
-			"type":              "usage",
-			"prompt_tokens":     usage.PromptTokens,
-			"completion_tokens": usage.CompletionTokens,
-			"total_tokens":      usage.TotalTokens(),
+			"type":                "usage",
+			"prompt_tokens":       usage.PromptTokens,
+			"completion_tokens":   usage.CompletionTokens,
+			"cached_input_tokens": usage.CachedInputTokens,
+			"cache_write_tokens":  usage.CacheWriteTokens,
+			"total_tokens":        usage.TotalTokens(),
 		})
 		return
 	}
 	if writer.format == execOutputStreamJSON {
 		promptTokens := usage.EffectiveInputTokens()
 		completionTokens := usage.EffectiveOutputTokens()
+		cachedInputTokens := usage.CachedInputTokens
+		cacheWriteTokens := usage.CacheWriteTokens
 		totalTokens := usage.TotalTokens()
 		writer.writeStreamJSON(streamjson.Event{
-			Type:             streamjson.EventUsage,
-			RunID:            writer.runID,
-			PromptTokens:     &promptTokens,
-			CompletionTokens: &completionTokens,
-			TotalTokens:      &totalTokens,
+			Type:              streamjson.EventUsage,
+			RunID:             writer.runID,
+			PromptTokens:      &promptTokens,
+			CompletionTokens:  &completionTokens,
+			CachedInputTokens: &cachedInputTokens,
+			CacheWriteTokens:  &cacheWriteTokens,
+			TotalTokens:       &totalTokens,
 		})
 	}
 }

internal/cli/exec_writer_test.go

@@ -52,3 +52,39 @@ func TestExecWriterPropagatesToolResultTruncation(t *testing.T) {
 		})
 	}
 }
+
+func TestExecWriterPropagatesCacheUsage(t *testing.T) {
+	for _, format := range []execOutputFormat{execOutputJSON, execOutputStreamJSON} {
+		t.Run(string(format), func(t *testing.T) {
+			var stdout bytes.Buffer
+			writer := execEventWriter{
+				stdout:       &stdout,
+				format:       format,
+				runID:        "run_usage",
+				streamedText: &strings.Builder{},
+			}
+			writer.usage(agent.Usage{
+				InputTokens:       100,
+				OutputTokens:      20,
+				PromptTokens:      100,
+				CompletionTokens:  20,
+				CachedInputTokens: 60,
+				CacheWriteTokens:  10,
+			})
+			if writer.err != nil {
+				t.Fatalf("usage: %v", writer.err)
+			}
+			var payload map[string]any
+			if err := json.Unmarshal(bytes.TrimSpace(stdout.Bytes()), &payload); err != nil {
+				t.Fatalf("decode output %q: %v", stdout.String(), err)
+			}
+			cachedKey, writeKey := "cached_input_tokens", "cache_write_tokens"
+			if format == execOutputStreamJSON {
+				cachedKey, writeKey = "cachedInputTokens", "cacheWriteTokens"
+			}
+			if payload[cachedKey] != float64(60) || payload[writeKey] != float64(10) {
+				t.Fatalf("cache usage missing from payload: %#v", payload)
+			}
+		})
+	}
+}

internal/cli/mcp_config.go

@@ -29,6 +29,8 @@ type mcpWritableConfig struct {
 	serverRaw map[string]json.RawMessage
 }
 
+var lockMCPConfigFile = config.LockFile
+
 // projectMCPConfigExists reports whether the workspace's project ./.zero/config.json
 // declares any MCP servers, so the trust notice fires only when project MCP config was
 // actually skipped (mirroring projectHooksFileExists / projectPluginsDirExists). A
@@ -50,7 +52,7 @@ func projectMCPConfigExists(workspaceRoot string) bool {
 	return len(fc.MCP.Servers) > 0
 }
 
-func runMCPAdd(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) int {
+func runMCPAdd(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) (exitCode int) {
 	options, help, err := parseMCPAddArgs(args)
 	if err != nil {
 		return writeExecUsageError(stderr, err.Error())
@@ -66,6 +68,26 @@ func runMCPAdd(args []string, stdout io.Writer, stderr io.Writer, deps appDeps)
 	if err != nil {
 		return writeAppError(stderr, "failed to resolve user config: "+err.Error(), exitCrash)
 	}
+	// This edits the same user config document the config package mutates, with
+	// the same read-modify-write + rename shape, so it takes the same
+	// cross-process lock. Without it, an MCP config edit racing a provider or
+	// preference write would silently drop whichever landed first (issue #832).
+	unlock, err := lockMCPConfigFile(configPath)
+	if err != nil {
+		return writeAppError(stderr, redaction.ErrorMessage(err, redaction.Options{}), exitCrash)
+	}
+	released := false
+	defer func() {
+		if released {
+			return
+		}
+		// A failed release leaves the lock held for the rest of the process, so
+		// exiting success here would claim a state the next config write cannot
+		// reproduce. It must not mask a failure this command already reported.
+		if releaseErr := unlock(); releaseErr != nil && exitCode == exitSuccess {
+			exitCode = writeAppError(stderr, redaction.ErrorMessage(releaseErr, redaction.Options{}), exitCrash)
+		}
+	}()
 	cfg, err := readMCPWritableConfig(configPath)
 	if err != nil {
 		return writeAppError(stderr, redaction.ErrorMessage(err, redaction.Options{}), exitCrash)
@@ -80,6 +102,10 @@ func runMCPAdd(args []string, stdout io.Writer, stderr io.Writer, deps appDeps)
 	if err := writeMCPWritableConfig(configPath, cfg); err != nil {
 		return writeAppError(stderr, redaction.ErrorMessage(err, redaction.Options{}), exitCrash)
 	}
+	if err := unlock(); err != nil {
+		return writeAppError(stderr, redaction.ErrorMessage(err, redaction.Options{}), exitCrash)
+	}
+	released = true
 
 	if options.json {
 		payload := struct {
@@ -109,7 +135,7 @@ func runMCPAdd(args []string, stdout io.Writer, stderr io.Writer, deps appDeps)
 	return exitSuccess
 }
 
-func runMCPRemove(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) int {
+func runMCPRemove(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) (exitCode int) {
 	options, positional, help, err := parseMCPConfigPositionalCommand(args, "remove")
 	if err != nil {
 		return writeExecUsageError(stderr, err.Error())
@@ -132,6 +158,26 @@ func runMCPRemove(args []string, stdout io.Writer, stderr io.Writer, deps appDep
 	if err != nil {
 		return writeAppError(stderr, "failed to resolve user config: "+err.Error(), exitCrash)
 	}
+	// This edits the same user config document the config package mutates, with
+	// the same read-modify-write + rename shape, so it takes the same
+	// cross-process lock. Without it, an MCP config edit racing a provider or
+	// preference write would silently drop whichever landed first (issue #832).
+	unlock, err := lockMCPConfigFile(configPath)
+	if err != nil {
+		return writeAppError(stderr, redaction.ErrorMessage(err, redaction.Options{}), exitCrash)
+	}
+	released := false
+	defer func() {
+		if released {
+			return
+		}
+		// A failed release leaves the lock held for the rest of the process, so
+		// exiting success here would claim a state the next config write cannot
+		// reproduce. It must not mask a failure this command already reported.
+		if releaseErr := unlock(); releaseErr != nil && exitCode == exitSuccess {
+			exitCode = writeAppError(stderr, redaction.ErrorMessage(releaseErr, redaction.Options{}), exitCrash)
+		}
+	}()
 	cfg, err := readMCPWritableConfig(configPath)
 	if err != nil {
 		return writeAppError(stderr, redaction.ErrorMessage(err, redaction.Options{}), exitCrash)
@@ -145,6 +191,10 @@ func runMCPRemove(args []string, stdout io.Writer, stderr io.Writer, deps appDep
 			return writeAppError(stderr, redaction.ErrorMessage(err, redaction.Options{}), exitCrash)
 		}
 	}
+	if err := unlock(); err != nil {
+		return writeAppError(stderr, redaction.ErrorMessage(err, redaction.Options{}), exitCrash)
+	}
+	released = true
 
 	if options.json {
 		payload := struct {
@@ -167,7 +217,7 @@ func runMCPRemove(args []string, stdout io.Writer, stderr io.Writer, deps appDep
 	return exitSuccess
 }
 
-func runMCPToggle(args []string, stdout io.Writer, stderr io.Writer, deps appDeps, disabled bool) int {
+func runMCPToggle(args []string, stdout io.Writer, stderr io.Writer, deps appDeps, disabled bool) (exitCode int) {
 	commandName := "enable"
 	if disabled {
 		commandName = "disable"
@@ -194,6 +244,26 @@ func runMCPToggle(args []string, stdout io.Writer, stderr io.Writer, deps appDep
 	if err != nil {
 		return writeAppError(stderr, "failed to resolve user config: "+err.Error(), exitCrash)
 	}
+	// This edits the same user config document the config package mutates, with
+	// the same read-modify-write + rename shape, so it takes the same
+	// cross-process lock. Without it, an MCP config edit racing a provider or
+	// preference write would silently drop whichever landed first (issue #832).
+	unlock, err := lockMCPConfigFile(configPath)
+	if err != nil {
+		return writeAppError(stderr, redaction.ErrorMessage(err, redaction.Options{}), exitCrash)
+	}
+	released := false
+	defer func() {
+		if released {
+			return
+		}
+		// A failed release leaves the lock held for the rest of the process, so
+		// exiting success here would claim a state the next config write cannot
+		// reproduce. It must not mask a failure this command already reported.
+		if releaseErr := unlock(); releaseErr != nil && exitCode == exitSuccess {
+			exitCode = writeAppError(stderr, redaction.ErrorMessage(releaseErr, redaction.Options{}), exitCrash)
+		}
+	}()
 	cfg, err := readMCPWritableConfig(configPath)
 	if err != nil {
 		return writeAppError(stderr, redaction.ErrorMessage(err, redaction.Options{}), exitCrash)
@@ -210,6 +280,10 @@ func runMCPToggle(args []string, stdout io.Writer, stderr io.Writer, deps appDep
 			return writeAppError(stderr, redaction.ErrorMessage(err, redaction.Options{}), exitCrash)
 		}
 	}
+	if err := unlock(); err != nil {
+		return writeAppError(stderr, redaction.ErrorMessage(err, redaction.Options{}), exitCrash)
+	}
+	released = true
 
 	if options.json {
 		payload := struct {

internal/cli/mcp_config_lock_test.go

@@ -0,0 +1,221 @@
+package cli
+
+import (
+	"bytes"
+	"errors"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/Gitlawb/zero/internal/config"
+)
+
+type mcpConfigLockCase struct {
+	name     string
+	args     []string
+	disabled bool
+	server   bool
+}
+
+func mcpConfigLockCases() []mcpConfigLockCase {
+	return []mcpConfigLockCase{
+		{name: "add", args: []string{"mcp", "add", "docs", "--", "docs-mcp"}},
+		{name: "remove", args: []string{"mcp", "remove", "docs"}, server: true},
+		{name: "enable", args: []string{"mcp", "enable", "docs"}, server: true, disabled: true},
+		{name: "disable", args: []string{"mcp", "disable", "docs"}, server: true},
+	}
+}
+
+func seedMCPConfigLockCase(t *testing.T, path string, testCase mcpConfigLockCase) {
+	t.Helper()
+	servers := map[string]config.MCPServerConfig{}
+	if testCase.server {
+		servers["docs"] = config.MCPServerConfig{Type: "stdio", Command: "docs-mcp", Disabled: testCase.disabled}
+	}
+	writeMCPCommandConfig(t, path, config.FileConfig{
+		ActiveProvider: "seed",
+		Providers:      []config.ProviderProfile{{Name: "seed", Model: "seed-model"}},
+		MCP:            config.MCPConfig{Servers: servers},
+	})
+}
+
+func assertMCPConfigLockMutation(t *testing.T, path string, testCase mcpConfigLockCase) {
+	t.Helper()
+	cfg := readMCPCommandConfig(t, path)
+	server, exists := cfg.MCP.Servers["docs"]
+	switch testCase.name {
+	case "add":
+		if !exists {
+			t.Fatal("mcp add did not persist the server")
+		}
+	case "remove":
+		if exists {
+			t.Fatal("mcp remove left the server configured")
+		}
+	case "enable":
+		if !exists || server.Disabled {
+			t.Fatalf("mcp enable result = %+v, exists=%v", server, exists)
+		}
+	case "disable":
+		if !exists || !server.Disabled {
+			t.Fatalf("mcp disable result = %+v, exists=%v", server, exists)
+		}
+	}
+	if cfg.Preferences.Theme != "dracula" {
+		t.Errorf("theme update was lost: theme = %q, want dracula", cfg.Preferences.Theme)
+	}
+	if cfg.ActiveProvider != "seed" || len(cfg.Providers) != 1 || cfg.Providers[0].Name != "seed" {
+		t.Errorf("seeded provider was lost: active=%q providers=%#v", cfg.ActiveProvider, cfg.Providers)
+	}
+}
+
+// TestRunMCPConfigCommandsParticipateInConfigLock covers the half of issue #832
+// that lives outside internal/config. MCP config commands read the SAME user
+// config document, edit it, and republish it with the same temp-file+rename
+// shape as the config package's mutators. Locking only the config package would
+// leave these writers free to clobber a concurrent provider or preference
+// update, and be clobbered by one, with the file still valid JSON afterwards.
+//
+// Racing the two writers and hoping to observe a lost update is unreliable —
+// the interleaving that loses one is narrow, and the test passed consistently
+// against the unlocked code. So this asserts the property that actually matters
+// and is deterministic: while the config lock is held elsewhere, the MCP
+// writer's update CANNOT land. Once released it completes, and both updates
+// survive.
+func TestRunMCPConfigCommandsParticipateInConfigLock(t *testing.T) {
+	for _, testCase := range mcpConfigLockCases() {
+		t.Run(testCase.name, func(t *testing.T) {
+			configPath := filepath.Join(t.TempDir(), "zero", "config.json")
+			seedMCPConfigLockCase(t, configPath, testCase)
+			before, err := os.ReadFile(configPath)
+			if err != nil {
+				t.Fatal(err)
+			}
+
+			unlock, err := config.LockFile(configPath)
+			if err != nil {
+				t.Fatalf("acquire config lock: %v", err)
+			}
+			released := false
+			release := func() {
+				if !released {
+					released = true
+					if err := unlock(); err != nil {
+						t.Fatalf("release config lock: %v", err)
+					}
+				}
+			}
+			defer release()
+
+			original := lockMCPConfigFile
+			lockAttempted := make(chan struct{})
+			lockMCPConfigFile = func(path string) (func() error, error) {
+				close(lockAttempted)
+				return original(path)
+			}
+			t.Cleanup(func() { lockMCPConfigFile = original })
+
+			var stdout, stderr bytes.Buffer
+			done := make(chan int, 1)
+			go func() {
+				done <- runWithDeps(testCase.args, &stdout, &stderr, appDeps{
+					userConfigPath: func() (string, error) { return configPath, nil },
+				})
+			}()
+
+			select {
+			case <-lockAttempted:
+			case <-time.After(30 * time.Second):
+				t.Fatalf("mcp %s did not attempt config lock acquisition", testCase.name)
+			}
+
+			for range 20 {
+				after, err := os.ReadFile(configPath)
+				if err != nil {
+					t.Fatal(err)
+				}
+				if !bytes.Equal(after, before) {
+					t.Fatalf("mcp %s changed config while its lock was held", testCase.name)
+				}
+				select {
+				case exitCode := <-done:
+					t.Fatalf("mcp %s completed (exit %d) while its lock was held", testCase.name, exitCode)
+				default:
+				}
+				time.Sleep(5 * time.Millisecond)
+			}
+
+			release()
+			select {
+			case exitCode := <-done:
+				if exitCode != exitSuccess {
+					t.Fatalf("mcp %s exitCode = %d stderr=%s", testCase.name, exitCode, stderr.String())
+				}
+			case <-time.After(30 * time.Second):
+				t.Fatalf("mcp %s did not finish after the config lock was released", testCase.name)
+			}
+
+			if _, err := config.SetTheme(configPath, "dracula"); err != nil {
+				t.Fatalf("SetTheme: %v", err)
+			}
+			assertMCPConfigLockMutation(t, configPath, testCase)
+		})
+	}
+}
+
+func TestRunMCPConfigCommandsReportLockAcquisitionFailure(t *testing.T) {
+	sentinel := errors.New("injected lock acquisition failure")
+	original := lockMCPConfigFile
+	lockMCPConfigFile = func(string) (func() error, error) { return nil, sentinel }
+	t.Cleanup(func() { lockMCPConfigFile = original })
+
+	for _, testCase := range mcpConfigLockCases() {
+		t.Run(testCase.name, func(t *testing.T) {
+			configPath := filepath.Join(t.TempDir(), "zero", "config.json")
+			seedMCPConfigLockCase(t, configPath, testCase)
+			before, err := os.ReadFile(configPath)
+			if err != nil {
+				t.Fatal(err)
+			}
+			var stdout, stderr bytes.Buffer
+			exitCode := runWithDeps(testCase.args, &stdout, &stderr, appDeps{
+				userConfigPath: func() (string, error) { return configPath, nil },
+			})
+			if exitCode != exitCrash || stdout.Len() != 0 || !strings.Contains(stderr.String(), sentinel.Error()) {
+				t.Fatalf("exit=%d stdout=%q stderr=%q, want lock error without success output", exitCode, stdout.String(), stderr.String())
+			}
+			after, err := os.ReadFile(configPath)
+			if err != nil {
+				t.Fatal(err)
+			}
+			if !bytes.Equal(after, before) {
+				t.Fatal("command changed config after lock acquisition failed")
+			}
+		})
+	}
+}
+
+func TestRunMCPConfigCommandsReleaseBeforeSuccessOutput(t *testing.T) {
+	sentinel := errors.New("injected lock release failure")
+	original := lockMCPConfigFile
+	lockMCPConfigFile = func(string) (func() error, error) {
+		return func() error { return sentinel }, nil
+	}
+	t.Cleanup(func() { lockMCPConfigFile = original })
+
+	for _, testCase := range mcpConfigLockCases() {
+		t.Run(testCase.name, func(t *testing.T) {
+			configPath := filepath.Join(t.TempDir(), "zero", "config.json")
+			seedMCPConfigLockCase(t, configPath, testCase)
+			var stdout, stderr bytes.Buffer
+			exitCode := runWithDeps(testCase.args, &stdout, &stderr, appDeps{
+				userConfigPath: func() (string, error) { return configPath, nil },
+			})
+			if exitCode != exitCrash || stdout.Len() != 0 || !strings.Contains(stderr.String(), sentinel.Error()) {
+				t.Fatalf("exit=%d stdout=%q stderr=%q, want unlock error before success output", exitCode, stdout.String(), stderr.String())
+			}
+		})
+	}
+}

internal/cli/mcp_default_test.go

@@ -9,15 +9,15 @@ import (
 	"github.com/Gitlawb/zero/internal/config"
 )
 
-// `zero mcp disable firecrawl` must work even though firecrawl is a built-in
+// `zero mcp disable exa` must work even though exa is a built-in
 // default that is not written to the user's config file until overridden.
-func TestRunMCPDisableSeededFirecrawlDefault(t *testing.T) {
+func TestRunMCPDisableSeededExaDefault(t *testing.T) {
 	configPath := filepath.Join(t.TempDir(), "zero", "config.json")
-	// A config with no firecrawl entry — the default lives in code, not the file.
+	// A config with no Exa entry — the default lives in code, not the file.
 	writeMCPCommandRawConfig(t, configPath, `{"activeProvider":"fast"}`)
 
 	var out, errBuf bytes.Buffer
-	code := runWithDeps([]string{"mcp", "disable", "firecrawl", "--json"}, &out, &errBuf, appDeps{
+	code := runWithDeps([]string{"mcp", "disable", "exa", "--json"}, &out, &errBuf, appDeps{
 		userConfigPath: func() (string, error) { return configPath, nil },
 	})
 	if code != exitSuccess {
@@ -31,8 +31,8 @@ func TestRunMCPDisableSeededFirecrawlDefault(t *testing.T) {
 	if err := json.Unmarshal(out.Bytes(), &payload); err != nil {
 		t.Fatalf("decode disable JSON: %v\n%s", err, out.String())
 	}
-	if payload.ServerName != "firecrawl" || !payload.Disabled || !payload.Changed {
-		t.Fatalf("disable payload = %#v, want firecrawl disabled+changed", payload)
+	if payload.ServerName != "exa" || !payload.Disabled || !payload.Changed {
+		t.Fatalf("disable payload = %#v, want exa disabled+changed", payload)
 	}
 
 	// End-to-end: resolving that config now turns the default off.
@@ -40,7 +40,7 @@ func TestRunMCPDisableSeededFirecrawlDefault(t *testing.T) {
 	if err != nil {
 		t.Fatalf("ResolveMCP: %v", err)
 	}
-	if !cfg.Servers["firecrawl"].Disabled {
-		t.Fatal("expected `mcp disable firecrawl` to turn the seeded default off in the resolved config")
+	if !cfg.Servers["exa"].Disabled {
+		t.Fatal("expected `mcp disable exa` to turn the seeded default off in the resolved config")
 	}
 }

internal/cli/mcp_startup.go

@@ -0,0 +1,147 @@
+package cli
+
+import (
+	"context"
+	"sync"
+	"time"
+
+	"github.com/Gitlawb/zero/internal/config"
+	"github.com/Gitlawb/zero/internal/mcp"
+	"github.com/Gitlawb/zero/internal/tools"
+)
+
+const (
+	optionalMCPPromptGrace = time.Second
+	optionalMCPCloseGrace  = 250 * time.Millisecond
+)
+
+// splitMCPStartupConfig keeps explicitly configured servers on the
+// prompt-critical path. Only unchanged built-in defaults may initialize after
+// first paint; their failure was already intentionally silent.
+func splitMCPStartupConfig(cfg config.MCPConfig) (critical config.MCPConfig, optional config.MCPConfig) {
+	critical.Servers = map[string]config.MCPServerConfig{}
+	optional.Servers = map[string]config.MCPServerConfig{}
+	for name, server := range cfg.Servers {
+		if server.Disabled {
+			continue
+		}
+		if config.IsUnconfiguredDefault(name, server) {
+			optional.Servers[name] = server
+			continue
+		}
+		critical.Servers[name] = server
+	}
+	return critical, optional
+}
+
+type optionalMCPStartup struct {
+	cancel context.CancelFunc
+	done   chan struct{}
+
+	mu      sync.Mutex
+	runtime mcpToolRuntime
+
+	closeOnce        sync.Once
+	runtimeCloseOnce sync.Once
+	closeErr         error
+}
+
+func startOptionalMCP(
+	parent context.Context,
+	registry *tools.Registry,
+	cfg config.MCPConfig,
+	options mcp.RegisterOptions,
+	register func(context.Context, *tools.Registry, config.MCPConfig, mcp.RegisterOptions) (mcpToolRuntime, error),
+	onReady func(),
+) *optionalMCPStartup {
+	ctx, cancel := context.WithCancel(parent)
+	startup := &optionalMCPStartup{cancel: cancel, done: make(chan struct{})}
+	go func() {
+		runtime, err := register(ctx, registry, cfg, options)
+		if err == nil && onReady != nil {
+			onReady()
+		}
+		startup.mu.Lock()
+		startup.runtime = runtime
+		startup.mu.Unlock()
+		close(startup.done)
+	}()
+	return startup
+}
+
+// Await gives optional tools one bounded opportunity to join the next turn's
+// immutable registry snapshot. A timeout leaves this turn on the existing core
+// tool set; a later turn observes the completed generation.
+func (startup *optionalMCPStartup) Await(ctx context.Context, timeout time.Duration) bool {
+	if startup == nil {
+		return true
+	}
+	select {
+	case <-startup.done:
+		return true
+	default:
+	}
+	if timeout <= 0 {
+		select {
+		case <-startup.done:
+			return true
+		case <-ctx.Done():
+			return false
+		}
+	}
+	timer := time.NewTimer(timeout)
+	defer timer.Stop()
+	select {
+	case <-startup.done:
+		return true
+	case <-timer.C:
+		return false
+	case <-ctx.Done():
+		return false
+	}
+}
+
+func (startup *optionalMCPStartup) Close() error {
+	if startup == nil {
+		return nil
+	}
+	startup.closeOnce.Do(func() {
+		startup.cancel()
+		reaped := make(chan struct{})
+		go func() {
+			defer close(reaped)
+			<-startup.done
+			startup.closeRuntime()
+		}()
+		timer := time.NewTimer(optionalMCPCloseGrace)
+		defer timer.Stop()
+		select {
+		case <-reaped:
+		case <-timer.C:
+		}
+	})
+	startup.mu.Lock()
+	defer startup.mu.Unlock()
+	return startup.closeErr
+}
+
+func (startup *optionalMCPStartup) closeRuntime() {
+	startup.runtimeCloseOnce.Do(func() {
+		startup.mu.Lock()
+		runtime := startup.runtime
+		startup.mu.Unlock()
+		if runtime == nil {
+			return
+		}
+		err := runtime.Close()
+		startup.mu.Lock()
+		if err != nil && startup.closeErr == nil {
+			startup.closeErr = err
+		}
+		startup.mu.Unlock()
+	})
+}
+
+func (startup *optionalMCPStartup) Skipped() []mcp.SkippedServer {
+	return nil
+}

internal/cli/mcp_startup_test.go

@@ -0,0 +1,187 @@
+package cli
+
+import (
+	"context"
+	"errors"
+	"testing"
+	"time"
+
+	"github.com/Gitlawb/zero/internal/config"
+	"github.com/Gitlawb/zero/internal/mcp"
+	"github.com/Gitlawb/zero/internal/tools"
+)
+
+type closeTrackingMCPRuntime struct {
+	closed chan struct{}
+	err    error
+}
+
+func (runtime *closeTrackingMCPRuntime) Close() error {
+	close(runtime.closed)
+	return runtime.err
+}
+
+func (*closeTrackingMCPRuntime) Skipped() []mcp.SkippedServer { return nil }
+
+func TestSplitMCPStartupConfigDefersOnlyUnconfiguredDefaults(t *testing.T) {
+	defaults := config.DefaultMCPServers()
+	cfg := config.MCPConfig{Servers: map[string]config.MCPServerConfig{
+		"exa":      defaults["exa"],
+		"docs":     {Type: "http", URL: "https://docs.example.test/mcp"},
+		"disabled": {Type: "http", URL: "https://disabled.example.test/mcp", Disabled: true},
+	}}
+
+	critical, optional := splitMCPStartupConfig(cfg)
+	if _, ok := critical.Servers["docs"]; !ok || len(critical.Servers) != 1 {
+		t.Fatalf("critical servers = %#v, want only docs", critical.Servers)
+	}
+	if _, ok := optional.Servers["exa"]; !ok || len(optional.Servers) != 1 {
+		t.Fatalf("optional servers = %#v, want only exa", optional.Servers)
+	}
+}
+
+func TestOptionalMCPAwaitTimesOutThenObservesPublishedTools(t *testing.T) {
+	registry := tools.NewRegistry()
+	release := make(chan struct{})
+	started := make(chan struct{})
+	startup := startOptionalMCP(
+		context.Background(),
+		registry,
+		config.MCPConfig{Servers: config.DefaultMCPServers()},
+		mcp.RegisterOptions{},
+		func(ctx context.Context, registry *tools.Registry, _ config.MCPConfig, _ mcp.RegisterOptions) (mcpToolRuntime, error) {
+			close(started)
+			select {
+			case <-release:
+				registry.RegisterBatch([]tools.Tool{tools.NewScopedReadFileTool(t.TempDir(), nil)})
+				return noopMCPRuntime{}, nil
+			case <-ctx.Done():
+				return noopMCPRuntime{}, ctx.Err()
+			}
+		},
+		nil,
+	)
+	defer func() { _ = startup.Close() }()
+	<-started
+
+	if startup.Await(t.Context(), time.Millisecond) {
+		t.Fatal("optional startup unexpectedly completed before release")
+	}
+	if got := len(registry.All()); got != 0 {
+		t.Fatalf("registry published tools before readiness: %d", got)
+	}
+	close(release)
+	if !startup.Await(t.Context(), time.Second) {
+		t.Fatal("optional startup did not complete after release")
+	}
+	if _, ok := registry.Get("read_file"); !ok {
+		t.Fatal("completed optional startup did not publish its tool batch")
+	}
+}
+
+func TestOptionalMCPFailedRegistrationSkipsReadinessCallback(t *testing.T) {
+	readyCalls := 0
+	startup := startOptionalMCP(
+		context.Background(),
+		tools.NewRegistry(),
+		config.MCPConfig{Servers: config.DefaultMCPServers()},
+		mcp.RegisterOptions{},
+		func(context.Context, *tools.Registry, config.MCPConfig, mcp.RegisterOptions) (mcpToolRuntime, error) {
+			return noopMCPRuntime{}, errors.New("connect failed")
+		},
+		func() { readyCalls++ },
+	)
+	defer func() { _ = startup.Close() }()
+
+	if !startup.Await(t.Context(), time.Second) {
+		t.Fatal("failed optional startup did not settle")
+	}
+	if readyCalls != 0 {
+		t.Fatalf("readiness callback ran after failed registration: %d", readyCalls)
+	}
+}
+
+func TestOptionalMCPCloseCancelsBlockedStartup(t *testing.T) {
+	started := make(chan struct{})
+	startup := startOptionalMCP(
+		context.Background(),
+		tools.NewRegistry(),
+		config.MCPConfig{Servers: config.DefaultMCPServers()},
+		mcp.RegisterOptions{},
+		func(ctx context.Context, _ *tools.Registry, _ config.MCPConfig, _ mcp.RegisterOptions) (mcpToolRuntime, error) {
+			close(started)
+			<-ctx.Done()
+			return noopMCPRuntime{}, nil
+		},
+		nil,
+	)
+	<-started
+	if err := startup.Close(); err != nil {
+		t.Fatalf("Close() error = %v", err)
+	}
+}
+
+func TestOptionalMCPCloseReturnsRuntimeErrorWithinGrace(t *testing.T) {
+	closeErr := errors.New("close failed")
+	runtime := &closeTrackingMCPRuntime{closed: make(chan struct{}), err: closeErr}
+	startup := startOptionalMCP(
+		context.Background(),
+		tools.NewRegistry(),
+		config.MCPConfig{Servers: config.DefaultMCPServers()},
+		mcp.RegisterOptions{},
+		func(context.Context, *tools.Registry, config.MCPConfig, mcp.RegisterOptions) (mcpToolRuntime, error) {
+			return runtime, nil
+		},
+		nil,
+	)
+	if !startup.Await(t.Context(), time.Second) {
+		t.Fatal("optional startup did not settle")
+	}
+	if err := startup.Close(); !errors.Is(err, closeErr) {
+		t.Fatalf("Close() error = %v, want %v", err, closeErr)
+	}
+}
+
+func TestOptionalMCPCloseBoundsRegistrationThatIgnoresCancellation(t *testing.T) {
+	started := make(chan struct{})
+	release := make(chan struct{})
+	released := false
+	defer func() {
+		if !released {
+			close(release)
+		}
+	}()
+	runtime := &closeTrackingMCPRuntime{closed: make(chan struct{})}
+	startup := startOptionalMCP(
+		context.Background(),
+		tools.NewRegistry(),
+		config.MCPConfig{Servers: config.DefaultMCPServers()},
+		mcp.RegisterOptions{},
+		func(context.Context, *tools.Registry, config.MCPConfig, mcp.RegisterOptions) (mcpToolRuntime, error) {
+			close(started)
+			<-release
+			return runtime, nil
+		},
+		nil,
+	)
+	<-started
+
+	closeResult := make(chan error, 1)
+	go func() { closeResult <- startup.Close() }()
+	select {
+	case err := <-closeResult:
+		if err != nil {
+			t.Fatalf("Close() error = %v", err)
+		}
+	case <-time.After(time.Second):
+		t.Fatal("Close blocked on registration after cancellation")
+	}
+
+	close(release)
+	released = true
+	select {
+	case <-runtime.closed:
+	case <-time.After(time.Second):
+		t.Fatal("late optional MCP runtime was not closed")
+	}
+}

internal/cli/mcp_tools.go

@@ -31,6 +31,14 @@ type mcpToolListItem struct {
 // into the one-line trust notice (mirroring the hooks and plugins chokepoints);
 // otherwise a workspace whose only project config is MCP would be gated silently.
 func registerMCPToolsForWorkspace(ctx context.Context, workspaceRoot string, registry *tools.Registry, deps appDeps, autonomy mcp.PermissionAutonomy, trustRoot string, runners ...*execution.Runner) (mcpToolRuntime, trustSkip, error) {
+	options := mcp.RegisterOptions{}
+	if len(runners) > 0 {
+		options.Execution = runners[0]
+	}
+	return registerMCPToolsForWorkspaceWithOptions(ctx, workspaceRoot, registry, deps, autonomy, trustRoot, options)
+}
+
+func registerMCPToolsForWorkspaceWithOptions(ctx context.Context, workspaceRoot string, registry *tools.Registry, deps appDeps, autonomy mcp.PermissionAutonomy, trustRoot string, options mcp.RegisterOptions) (mcpToolRuntime, trustSkip, error) {
 	excludeProject, trustCheckErrored := resolveTrust(trustRoot)
 	skip := trustSkip{
 		excludedProjectConfig: excludeProject && projectMCPConfigExists(workspaceRoot),
@@ -47,16 +55,10 @@ func registerMCPToolsForWorkspace(ctx context.Context, workspaceRoot string, reg
 	if err != nil {
 		return nil, skip, err
 	}
-	var runner *execution.Runner
-	if len(runners) > 0 {
-		runner = runners[0]
-	}
-	runtime, err := deps.registerMCPTools(ctx, registry, cfg, mcp.RegisterOptions{
-		PermissionStore: store,
-		Autonomy:        autonomy,
-		Execution:       runner,
-		WorkspaceRoot:   workspaceRoot,
-	})
+	options.PermissionStore = store
+	options.Autonomy = autonomy
+	options.WorkspaceRoot = workspaceRoot
+	runtime, err := deps.registerMCPTools(ctx, registry, cfg, options)
 	return runtime, skip, err
 }
 

internal/cli/oauth_provider.go

@@ -1,13 +1,8 @@
 package cli
 
 import (
-	"context"
-	"errors"
-	"net/http"
-	"time"
-
 	"github.com/Gitlawb/zero/internal/config"
-	"github.com/Gitlawb/zero/internal/oauth"
+	"github.com/Gitlawb/zero/internal/providers"
 	"github.com/Gitlawb/zero/internal/providers/providerio"
 )
 
@@ -28,46 +23,5 @@ import (
 // ProviderProfile.OAuthLoginCandidates so the runtime resolver, the Codex account
 // resolver, and the onboarding presence check never diverge.
 func oauthLoginForProfile(profile config.ProviderProfile) (providerio.TokenResolver, string) {
-	candidates := profile.OAuthLoginCandidates()
-	if len(candidates) == 0 {
-		return nil, ""
-	}
-	store, err := oauth.NewStore(oauth.StoreOptions{})
-	if err != nil {
-		return nil, ""
-	}
-	_, key, ok := oauth.FirstStored(store, candidates)
-	if !ok {
-		// No login under any candidate (or unreadable/invalid keys) → API-key
-		// auth, no resolver.
-		return nil, ""
-	}
-	manager, err := oauth.NewManager(oauth.ManagerOptions{
-		Store:      store,
-		HTTPClient: &http.Client{Timeout: 30 * time.Second},
-		// Refreshing a token the user logged into (possibly a preset provider like
-		// xAI) re-resolves that provider's OAuth config, which needs the preset.
-		AllowPresets: true,
-	})
-	if err != nil {
-		return nil, ""
-	}
-	resolver := func(ctx context.Context, forceRefresh bool) (string, string, bool, error) {
-		var token string
-		var rerr error
-		if forceRefresh {
-			token, rerr = manager.Handle401(ctx, key)
-		} else {
-			token, rerr = manager.GetFresh(ctx, key)
-		}
-		if errors.Is(rerr, oauth.ErrNoToken) {
-			// The login was removed since construction → fall back to the API key.
-			return "", "", false, nil
-		}
-		if rerr != nil {
-			return "", "", false, rerr
-		}
-		return "Authorization", "Bearer " + token, true, nil
-	}
-	return resolver, key
+	return providers.OAuthLoginForProfile(profile)
 }

internal/cli/provider_adoption_test.go

@@ -0,0 +1,117 @@
+package cli
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"net/http"
+	"net/http/httptest"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/Gitlawb/zero/internal/provideronboarding"
+	"mvdan.cc/sh/v3/shell"
+)
+
+func TestLocalAdoptionSavesAndReloadsEligibleModel(t *testing.T) {
+	const wantModel = "-loaded chat model"
+	requests := make(chan string, 4)
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		switch r.URL.Path {
+		case "/v1/models":
+			w.Header().Set("Content-Type", "application/json")
+			_, _ = fmt.Fprintf(w, `{"data":[{"id":"text-embedding-local"},{"id":%q}]}`, wantModel)
+		case "/v1/chat/completions":
+			var body struct {
+				Model string `json:"model"`
+			}
+			if err := json.NewDecoder(r.Body).Decode(&body); err != nil {
+				t.Errorf("decode completion request: %v", err)
+				w.WriteHeader(http.StatusBadRequest)
+				return
+			}
+			if r.Header.Get("Authorization") != "" {
+				t.Error("local adoption unexpectedly sent authorization")
+			}
+			select {
+			case requests <- body.Model:
+			default:
+				t.Error("unexpected extra completion requests")
+			}
+			w.Header().Set("Content-Type", "text/event-stream")
+			_, _ = fmt.Fprint(w, "data: {\"choices\":[{\"delta\":{\"content\":\"adoption ok\"},\"finish_reason\":\"stop\"}]}\n\ndata: [DONE]\n\n")
+		default:
+			w.WriteHeader(http.StatusNotFound)
+		}
+	}))
+	defer server.Close()
+	detected := provideronboarding.DetectLocalRuntimes(context.Background(), provideronboarding.LocalDetectOptions{
+		HTTPClient: server.Client(),
+		Candidates: []provideronboarding.LocalRuntime{{CatalogID: "atomic-chat-local", Name: "Atomic Chat Local", DefaultModel: "local-model", BaseURL: server.URL + "/v1"}},
+	})
+	if len(detected) != 1 {
+		t.Fatalf("expected one detected runtime: %+v", detected)
+	}
+	args, err := shell.Fields(detected[0].SetupAction().Command, func(string) string { return "" })
+	if err != nil || len(args) < 4 {
+		t.Fatalf("invalid adoption command: args=%q err=%v", args, err)
+	}
+	root := t.TempDir()
+	run := func(args ...string) string {
+		t.Helper()
+		ctx, cancel := context.WithTimeout(context.Background(), 30*time.Second)
+		defer cancel()
+		childArgs := append([]string{"-test.run=^TestLocalAdoptionCLIProcess$", "--"}, args...)
+		cmd := exec.CommandContext(ctx, os.Args[0], childArgs...)
+		cmd.Dir = root
+		cmd.Env = []string{
+			"ZERO_TEST_LOCAL_ADOPTION=1", "ZERO_CRED_STORAGE=encrypted-file",
+			"HOME=" + root, "USERPROFILE=" + root,
+			"XDG_CONFIG_HOME=" + filepath.Join(root, "config"),
+			"XDG_CACHE_HOME=" + filepath.Join(root, "cache"),
+			"XDG_STATE_HOME=" + filepath.Join(root, "state"),
+			"APPDATA=" + filepath.Join(root, "config"),
+			"LOCALAPPDATA=" + filepath.Join(root, "cache"),
+		}
+		for _, key := range []string{"PATH", "SystemRoot", "WINDIR", "TMPDIR", "TEMP", "TMP"} {
+			if value, ok := os.LookupEnv(key); ok {
+				cmd.Env = append(cmd.Env, key+"="+value)
+			}
+		}
+		output, err := cmd.CombinedOutput()
+		if err != nil {
+			t.Fatalf("CLI %q failed: %v\n%s", args, err, output)
+		}
+		return string(output)
+	}
+	// The mock uses a random port; only override the endpoint, retaining all
+	// generated arguments for the real add parser and persistence path.
+	run(append(args[1:], "--base-url", server.URL+"/v1")...)
+	if output := run("exec", "--cwd", root, "Reply with adoption ok"); !strings.Contains(output, "adoption ok") {
+		t.Fatalf("completion output = %q", output)
+	}
+	select {
+	case model := <-requests:
+		if model != wantModel {
+			t.Fatalf("fresh process sent model %q, want eligible model %q", model, wantModel)
+		}
+	default:
+		t.Fatal("fresh process did not send a completion request")
+	}
+}
+
+func TestLocalAdoptionCLIProcess(t *testing.T) {
+	if os.Getenv("ZERO_TEST_LOCAL_ADOPTION") != "1" {
+		return
+	}
+	for i, arg := range os.Args {
+		if arg == "--" {
+			os.Exit(Run(os.Args[i+1:], os.Stdout, os.Stderr))
+		}
+	}
+	t.Fatal("missing helper argument separator")
+}

internal/cli/provider_detect.go

@@ -40,7 +40,7 @@ type providerDetectReport struct {
 }
 
 // runProvidersDetect probes the machine for running local, OpenAI-compatible
-// model runtimes (Ollama, LM Studio) and prints a no-key adopt command for each
+// model runtimes (Ollama, LM Studio, Atomic Chat) and prints a no-key adopt command for each
 // one it finds, followed by the next-step actions for every already-configured
 // provider. It is the onboarding-advice surface — "what can I do right now?" —
 // and never errors on a machine with nothing running locally (it just reports an
@@ -146,6 +146,8 @@ func formatProviderDetectReport(report providerDetectReport) string {
 		}
 		if command := strings.TrimSpace(runtime.Action.Command); command != "" {
 			lines = append(lines, "    "+runtime.Action.Label+": "+command)
+		} else if detail := strings.TrimSpace(runtime.Action.Detail); detail != "" {
+			lines = append(lines, "    "+runtime.Action.Label+": "+detail)
 		}
 	}
 

internal/cli/provider_detect_test.go

@@ -105,3 +105,13 @@ func TestRunProvidersDetectJSONNoRuntimesActiveProvider(t *testing.T) {
 		t.Fatalf("expected only a Check action for an active keyed provider, got %#v", payload.Providers[0].Actions)
 	}
 }
+
+func TestProviderDetectShowsGuidanceWithoutCommand(t *testing.T) {
+	for _, models := range [][]string{nil, {"local-model"}, {"x&calc"}} {
+		runtime := provideronboarding.DetectedLocalRuntime{LocalRuntime: provideronboarding.LocalRuntime{CatalogID: "atomic-chat-local", Name: "Atomic Chat Local", DefaultModel: "local-model"}, Models: models}
+		out := formatProviderDetectReport(buildProviderDetectReport(config.ResolvedConfig{}, []provideronboarding.DetectedLocalRuntime{runtime}))
+		if strings.Contains(out, "providers add") || !strings.Contains(out, runtime.SetupAction().Detail) {
+			t.Fatalf("missing actionable guidance: %s", out)
+		}
+	}
+}

internal/cli/provider_models.go

@@ -73,6 +73,18 @@ func runProvidersModels(args []string, stdout io.Writer, stderr io.Writer, deps
 			if description := strings.TrimSpace(model.Description); description != "" {
 				entry["description"] = description
 			}
+			if len(model.ReasoningEfforts) > 0 {
+				entry["reasoning_efforts"] = append([]string{}, model.ReasoningEfforts...)
+			}
+			if effort := strings.TrimSpace(model.DefaultReasoningEffort); effort != "" {
+				entry["default_reasoning_effort"] = effort
+			}
+			if len(model.ServiceTiers) > 0 {
+				entry["service_tiers"] = append([]string{}, model.ServiceTiers...)
+			}
+			if tier := strings.TrimSpace(model.DefaultServiceTier); tier != "" {
+				entry["default_service_tier"] = tier
+			}
 			items = append(items, entry)
 		}
 		payload := map[string]any{

internal/cli/provider_setup.go

@@ -434,6 +434,14 @@ func providerProfileForAdd(options providerAddOptions) (config.ProviderProfile,
 			catalogHeaders = aimlapi.WithResolvedPartnerHeader(catalogHeaders)
 		}
 	}
+	// Discovery fallback can pass the nonempty catalog placeholder through
+	// either CLI setup path. Never persist it as an Atomic Chat model.
+	if descriptor.ID == "atomic-chat-local" {
+		model := strings.TrimSpace(options.model)
+		if model == "" || model == "local-model" {
+			return config.ProviderProfile{}, fmt.Errorf("provider %q serves a locally loaded model; pass --model <id> (run `zero providers detect` to see the served model)", descriptor.ID)
+		}
+	}
 	profile := config.ProviderProfile{
 		Name:            name,
 		ProviderKind:    providerKindForDescriptor(descriptor),

internal/cli/provider_setup_test.go

@@ -1,9 +1,12 @@
 package cli
 
 import (
+	"strings"
 	"testing"
 
 	"github.com/Gitlawb/zero/internal/config"
+	"github.com/Gitlawb/zero/internal/provideronboarding"
+	"mvdan.cc/sh/v3/shell"
 )
 
 // Regression for issue #555's follow-up: `zero providers check` must not
@@ -66,3 +69,58 @@ func TestValidateProviderRuntimeReadyCustomEndpoint(t *testing.T) {
 		})
 	}
 }
+
+// atomic-chat-local without --model would persist the catalog placeholder
+// "local-model", which the Atomic Chat server never serves, so the first
+// completion fails. Adding it must require a real model instead.
+func TestProviderProfileForAddRequiresModelForAtomicChatLocal(t *testing.T) {
+	if _, err := providerProfileForAdd(providerAddOptions{catalogID: "atomic-chat-local"}); err == nil {
+		t.Fatalf("providerProfileForAdd(atomic-chat-local, no --model) = nil error, want a require-model error")
+	} else if !strings.Contains(err.Error(), "--model") {
+		t.Fatalf("error should tell the user to pass --model, got %v", err)
+	}
+
+	// The interactive wizards and the no-id detect fallback resolve the model to
+	// the catalog DefaultModel and pass it through as a non-empty value, so the
+	// placeholder itself must be rejected, not just an empty --model.
+	if _, err := providerProfileForAdd(providerAddOptions{catalogID: "atomic-chat-local", model: "local-model"}); err == nil {
+		t.Fatalf("providerProfileForAdd(atomic-chat-local, --model local-model) = nil error, want reject of the catalog placeholder")
+	}
+
+	profile, err := providerProfileForAdd(providerAddOptions{catalogID: "atomic-chat-local", model: "unsloth/gemma-4-E2B-it-GGUF"})
+	if err != nil {
+		t.Fatalf("providerProfileForAdd(atomic-chat-local, --model) returned error: %v", err)
+	}
+	if profile.Model != "unsloth/gemma-4-E2B-it-GGUF" {
+		t.Fatalf("profile.Model = %q, want the explicit model", profile.Model)
+	}
+	if profile.Model == "local-model" {
+		t.Fatalf("profile persisted the catalog placeholder")
+	}
+}
+
+func TestDetectedModelActionSurvivesAddParser(t *testing.T) {
+	for _, catalogID := range []string{"atomic-chat-local", "lmstudio", "ollama"} {
+		for _, modelID := range []string{"-loaded-model", "--set-active", "--model", "-loaded model", "ordinary/model", "model with spaces"} {
+			t.Run(catalogID+"/"+modelID, func(t *testing.T) {
+				detected := provideronboarding.DetectedLocalRuntime{
+					LocalRuntime: provideronboarding.LocalRuntime{CatalogID: catalogID, Name: "Local Runtime", DefaultModel: "local-model"},
+					Models:       []string{modelID},
+				}
+				command := detected.SetupAction().Command
+				args, err := shell.Fields(command, func(string) string { return "" })
+				if err != nil || len(args) < 4 {
+					t.Fatalf("invalid adoption command %q: %v", command, err)
+				}
+				options, help, err := parseProviderAddArgs(args[3:])
+				if err != nil || help {
+					t.Fatalf("generated adoption command rejected by add parser: %q: %v", command, err)
+				}
+				profile, err := providerProfileForAdd(options)
+				if err != nil || profile.Model != modelID || !options.setActive || profile.Name != "Local Runtime" {
+					t.Fatalf("adoption changed model or options: model=%q name=%q active=%v err=%v", profile.Model, profile.Name, options.setActive, err)
+				}
+			})
+		}
+	}
+}

internal/cli/root_flag_order_test.go

@@ -0,0 +1,211 @@
+package cli
+
+import (
+	"bytes"
+	"context"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"github.com/Gitlawb/zero/internal/agent"
+	"github.com/Gitlawb/zero/internal/config"
+	"github.com/Gitlawb/zero/internal/sandbox"
+	"github.com/Gitlawb/zero/internal/tui"
+	"github.com/Gitlawb/zero/internal/zeroruntime"
+)
+
+// rootFlagOrderings is every ordering of the given flag groups, each group
+// kept intact so a flag stays next to its value.
+func rootFlagOrderings(groups [][]string) [][]string {
+	if len(groups) == 0 {
+		return [][]string{{}}
+	}
+	var orderings [][]string
+	for index, group := range groups {
+		rest := append(append([][]string{}, groups[:index]...), groups[index+1:]...)
+		for _, tail := range rootFlagOrderings(rest) {
+			orderings = append(orderings, append(append([]string{}, group...), tail...))
+		}
+	}
+	return orderings
+}
+
+// ROOT FLAGS COMPOSE IN ANY ORDER. Each leading-flag splitter stops at the first
+// token it does not own, so running them once in a fixed sequence made the
+// order the operator wrote them in load-bearing: `zero --allow-escalation
+// --theme auto` stranded `--theme auto` as an unknown command and exited with
+// an argument error instead of launching. Every ordering of the three root
+// flags, with --skip-permissions-unsafe absent or at any position, has to reach
+// the TUI with all three applied.
+func TestRootFlagsComposeInAnyOrder(t *testing.T) {
+	extra := t.TempDir()
+	resolvedExtra, err := filepath.EvalSymlinks(extra)
+	if err != nil {
+		t.Fatal(err)
+	}
+	groups := [][]string{{"--add-dir", extra}, {"--theme", "auto"}, {"--allow-escalation"}}
+	cases := rootFlagOrderings(groups)
+	cases = append(cases, rootFlagOrderings(append(groups, []string{"--skip-permissions-unsafe"}))...)
+	if len(cases) != 6+24 {
+		t.Fatalf("SETUP INVALID: %d orderings, want 30", len(cases))
+	}
+	for _, args := range cases {
+		t.Run(strings.Join(args, " "), func(t *testing.T) {
+			options := captureTUIOptions(t, args...)
+			wantMode := agent.PermissionModeAsk
+			for _, arg := range args {
+				if arg == "--skip-permissions-unsafe" {
+					wantMode = agent.PermissionModeUnsafe
+				}
+			}
+			if options.PermissionMode != wantMode {
+				t.Errorf("PermissionMode = %q, want %q", options.PermissionMode, wantMode)
+			}
+			if !options.AllowEscalation {
+				t.Error("escalation not enabled")
+			}
+			if options.Theme != "auto" {
+				t.Errorf("Theme = %q, want auto", options.Theme)
+			}
+			if options.AgentOptions.Sandbox == nil {
+				t.Fatal("no sandbox engine on the launched options")
+			}
+			roots := options.AgentOptions.Sandbox.Scope().Roots()
+			found := false
+			for _, root := range roots {
+				if root == resolvedExtra {
+					found = true
+				}
+			}
+			if !found {
+				t.Errorf("scope roots = %v, want the --add-dir root %q", roots, resolvedExtra)
+			}
+		})
+	}
+}
+
+// The last --theme wins even with another root flag between two of them, and a
+// --theme written before --skip-permissions-unsafe reaches the TUI: the unsafe
+// path used to launch with only the theme written after the flag, dropping one
+// written before it.
+func TestRootThemeLastOccurrenceWinsAcrossOtherFlags(t *testing.T) {
+	for _, testCase := range []struct {
+		args []string
+		want string
+	}{
+		{[]string{"--theme", "light", "--allow-escalation", "--theme", "auto"}, "auto"},
+		{[]string{"--theme", "auto", "--skip-permissions-unsafe"}, "auto"},
+		{[]string{"--theme", "light", "--skip-permissions-unsafe", "--theme", "auto"}, "auto"},
+	} {
+		t.Run(strings.Join(testCase.args, " "), func(t *testing.T) {
+			options := captureTUIOptions(t, testCase.args...)
+			if options.Theme != testCase.want {
+				t.Fatalf("Theme = %q, want %q", options.Theme, testCase.want)
+			}
+		})
+	}
+}
+
+// THE ROOT FLAG REACHES EXEC. A --allow-escalation written before the
+// subcommand is forwarded the way --add-dir is, so `zero --allow-escalation -p
+// "..."` runs with the opt-in the help text promises instead of silently
+// without it. The negative control is the same run without the flag, which
+// must not advertise escalate_model.
+func TestRootAllowEscalationForwardsIntoExec(t *testing.T) {
+	for _, testCase := range []struct {
+		name string
+		args []string
+	}{
+		{"exec subcommand", []string{"--allow-escalation", "exec", "say hi"}},
+		{"prompt flag", []string{"--allow-escalation", "-p", "say hi"}},
+	} {
+		t.Run(testCase.name, func(t *testing.T) {
+			if !execAdvertisesEscalateModel(t, testCase.args) {
+				t.Fatalf("%v ran without escalate_model advertised: the root flag was dropped before exec", testCase.args)
+			}
+			if execAdvertisesEscalateModel(t, testCase.args[1:]) {
+				t.Fatalf("%v advertised escalate_model without the flag", testCase.args[1:])
+			}
+		})
+	}
+}
+
+// Everywhere else the flag is rejected loudly rather than discarded, the way
+// --add-dir is: help and version run no agent and could only ignore it.
+func TestRootAllowEscalationIsRejectedWhereNothingConsumesIt(t *testing.T) {
+	for _, args := range [][]string{
+		{"--allow-escalation", "version"},
+		{"--allow-escalation", "help"},
+		{"--allow-escalation", "completions", "bash"},
+	} {
+		t.Run(strings.Join(args, " "), func(t *testing.T) {
+			var stdout, stderr bytes.Buffer
+			launched := false
+			exitCode := runWithDeps(args, &stdout, &stderr, appDeps{
+				getwd:  func() (string, error) { return t.TempDir(), nil },
+				runTUI: func(context.Context, tui.Options) int { launched = true; return exitSuccess },
+			})
+			if exitCode == exitSuccess || launched {
+				t.Fatalf("exit = %d launched = %v: the flag was discarded instead of rejected", exitCode, launched)
+			}
+			if !strings.Contains(stderr.String(), "--allow-escalation is only supported for the interactive TUI and exec") {
+				t.Fatalf("stderr does not name the flag: %s", stderr.String())
+			}
+		})
+	}
+}
+
+// execAdvertisesEscalateModel dispatches args through the root command with a
+// provider that records the tools advertised on the first request, and reports
+// whether escalate_model was among them.
+func execAdvertisesEscalateModel(t *testing.T, args []string) bool {
+	t.Helper()
+	isolateCLIUserState(t)
+	provider := &toolListingProvider{}
+	var stdout, stderr bytes.Buffer
+	exitCode := runWithDeps(args, &stdout, &stderr, appDeps{
+		getwd: func() (string, error) { return t.TempDir(), nil },
+		resolveConfig: func(string, config.Overrides) (config.ResolvedConfig, error) {
+			return execResolvedConfig(), nil
+		},
+		newProvider: func(config.ProviderProfile) (zeroruntime.Provider, error) {
+			return provider, nil
+		},
+		newSandboxStore: func() (*sandbox.GrantStore, error) {
+			return sandbox.NewGrantStore(sandbox.StoreOptions{FilePath: filepath.Join(t.TempDir(), "sandbox-grants.json")})
+		},
+	})
+	if exitCode != exitSuccess {
+		t.Fatalf("%v: exit = %d, stderr = %s", args, exitCode, stderr.String())
+	}
+	if provider.requests == 0 {
+		t.Fatalf("SETUP INVALID: %v never reached the provider", args)
+	}
+	for _, name := range provider.toolNames {
+		if name == "escalate_model" {
+			return true
+		}
+	}
+	return false
+}
+
+// toolListingProvider answers every request with a one-line text reply and
+// keeps the tool names advertised on the first one.
+type toolListingProvider struct {
+	requests  int
+	toolNames []string
+}
+
+func (provider *toolListingProvider) StreamCompletion(_ context.Context, request zeroruntime.CompletionRequest) (<-chan zeroruntime.StreamEvent, error) {
+	provider.requests++
+	if provider.requests == 1 {
+		for _, tool := range request.Tools {
+			provider.toolNames = append(provider.toolNames, tool.Name)
+		}
+	}
+	ch := make(chan zeroruntime.StreamEvent, 2)
+	ch <- zeroruntime.StreamEvent{Type: zeroruntime.StreamEventText, Content: "hi"}
+	ch <- zeroruntime.StreamEvent{Type: zeroruntime.StreamEventDone}
+	close(ch)
+	return ch, nil
+}

internal/cli/serve_test.go

@@ -168,11 +168,18 @@ func TestBuildServeScopeKeepsLexicalPaths(t *testing.T) {
 	}
 	linkWorkspace := filepath.Join(base, "link-workspace")
 	linkExtra := filepath.Join(base, "link-extra")
+	// SKIPPED, NOT FAILED, WHEN THE PLATFORM WILL NOT MAKE THE LINK. Creating a
+	// symlink on Windows needs SeCreateSymbolicLinkPrivilege, which an ordinary
+	// developer account does not hold, so this reported a failure with nothing
+	// wrong in the code: "A required privilege is not held by the client". CI
+	// runners are elevated, so it only ever showed up locally, where it cost a
+	// round of triage before being ruled environmental. The sibling tests that
+	// need a link already answer this way.
 	if err := os.Symlink(realWorkspace, linkWorkspace); err != nil {
-		t.Fatal(err)
+		t.Skipf("symlink unavailable: %v", err)
 	}
 	if err := os.Symlink(realExtra, linkExtra); err != nil {
-		t.Fatal(err)
+		t.Skipf("symlink unavailable: %v", err)
 	}
 
 	scope, err := buildServeScope(linkWorkspace, []string{linkExtra})

internal/cli/setup.go

@@ -42,7 +42,7 @@ func runSetup(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) i
 		return exitSuccess
 	}
 	if strings.TrimSpace(options.catalogID) == "" {
-		return runInteractiveTUIWithSetup(stderr, deps, "", nil, "", true)
+		return runInteractiveTUIWithSetup(stderr, deps, "", nil, "", true, false)
 	}
 
 	result, err := saveSetupProvider(deps, tui.SetupSelection{

internal/cli/setup_test.go

@@ -367,3 +367,14 @@ func TestVerifySetupProviderDistinguishesMissingFromRejectedKey(t *testing.T) {
 		t.Fatal("a keyless local provider should still be probed")
 	}
 }
+
+func TestSaveSetupProviderRejectsAtomicPlaceholderBeforeConfigAccess(t *testing.T) {
+	for _, model := range []string{"", "local-model", " local-model "} {
+		accessed := false
+		deps := appDeps{userConfigPath: func() (string, error) { accessed = true; return filepath.Join(t.TempDir(), "config.json"), nil }}
+		_, err := saveSetupProvider(deps, tui.SetupSelection{CatalogID: "atomic-chat-local", Model: model}, setupSaveOptions{})
+		if err == nil || !strings.Contains(err.Error(), "--model") || accessed {
+			t.Fatalf("invalid model %q reached config access: accessed=%v err=%v", model, accessed, err)
+		}
+	}
+}

internal/cli/summarizer.go

@@ -0,0 +1,17 @@
+package cli
+
+import (
+	"context"
+
+	"github.com/Gitlawb/zero/internal/agent"
+	"github.com/Gitlawb/zero/internal/config"
+	"github.com/Gitlawb/zero/internal/providers"
+	"github.com/Gitlawb/zero/internal/zeroruntime"
+)
+
+// summarizerFactory adapts the resolved profile and the authenticated provider
+// builder into agent.Options.Summarizer; the selection rules live in
+// providers.CompactionSummarizerFactory, shared with the TUI.
+func summarizerFactory(resolved config.ResolvedConfig, newProvider func(config.ProviderProfile) (zeroruntime.Provider, error)) func(context.Context, string) (agent.Provider, error) {
+	return providers.CompactionSummarizerFactory(resolved.Provider, resolved.Preferences.CompactionModel, newProvider)
+}

internal/cli/tui_escalation_test.go

@@ -0,0 +1,126 @@
+package cli
+
+import (
+	"bytes"
+	"context"
+	"strings"
+	"testing"
+
+	"github.com/Gitlawb/zero/internal/tui"
+)
+
+// captureTUIOptions runs the root command with the TUI launch intercepted, and
+// returns the options the interactive session would have started with.
+func captureTUIOptions(t *testing.T, args ...string) tui.Options {
+	t.Helper()
+	isolateCLIUserState(t)
+	var captured tui.Options
+	var stdout, stderr bytes.Buffer
+	workspace := t.TempDir()
+	exitCode := runWithDeps(args, &stdout, &stderr, appDeps{
+		getwd: func() (string, error) { return workspace, nil },
+		runTUI: func(_ context.Context, options tui.Options) int {
+			captured = options
+			return exitSuccess
+		},
+	})
+	if exitCode != exitSuccess {
+		t.Fatalf("exitCode = %d stdout=%s stderr=%s", exitCode, stdout.String(), stderr.String())
+	}
+	return captured
+}
+
+func hasEscalateModel(options tui.Options) bool {
+	if options.AgentOptions.Registry == nil {
+		return false
+	}
+	_, registered := options.AgentOptions.Registry.Get("escalate_model")
+	return registered
+}
+
+// ESCALATION IS OFF UNLESS THE OPERATOR ASKS FOR IT.
+//
+// Escalation moves a run onto a different model, changing what it costs and which
+// provider sees the conversation, so the interactive surface answers this the same
+// conservative way `zero exec --allow-escalation` already does.
+func TestInteractiveTUIRegistersEscalateModelOnlyWithTheFlag(t *testing.T) {
+	if hasEscalateModel(captureTUIOptions(t)) {
+		t.Error("escalate_model was registered without --allow-escalation")
+	}
+	if !hasEscalateModel(captureTUIOptions(t, "--allow-escalation")) {
+		t.Error("--allow-escalation did not register escalate_model")
+	}
+}
+
+// THE TOOL AND THE SWITCHERS ARE ONE FEATURE, NOT TWO.
+//
+// This is the failure the whole change exists to avoid. The agent loop performs a
+// switch only when a switcher is wired, so registering escalate_model without
+// wiring one ships a tool the model can call and the loop will silently ignore:
+// the run reports an escalation that never happened. The reverse is merely dead
+// code. Both halves ride on the same flag and this asserts they cannot be
+// separated, whichever half a future change touches.
+func TestInteractiveTUIEscalationToolAndSwitchersAreWiredTogether(t *testing.T) {
+	for _, testCase := range []struct {
+		name string
+		args []string
+		want bool
+	}{
+		{name: "default", args: nil, want: false},
+		{name: "flagged", args: []string{"--allow-escalation"}, want: true},
+	} {
+		t.Run(testCase.name, func(t *testing.T) {
+			options := captureTUIOptions(t, testCase.args...)
+			if got := hasEscalateModel(options); got != testCase.want {
+				t.Fatalf("escalate_model registered = %v, want %v", got, testCase.want)
+			}
+			if options.AllowEscalation != testCase.want {
+				t.Errorf("AllowEscalation = %v, want %v: the tool is registered on one gate and the switchers on another, so they can drift apart",
+					options.AllowEscalation, testCase.want)
+			}
+		})
+	}
+}
+
+// The flag is accepted on either side of --skip-permissions-unsafe, like the
+// other root flags that may appear there.
+func TestInteractiveTUIAcceptsAllowEscalationAroundUnsafe(t *testing.T) {
+	for _, args := range [][]string{
+		{"--allow-escalation", "--skip-permissions-unsafe"},
+		{"--skip-permissions-unsafe", "--allow-escalation"},
+	} {
+		options := captureTUIOptions(t, args...)
+		if !options.AllowEscalation {
+			t.Errorf("%v did not enable escalation", args)
+		}
+		if !hasEscalateModel(options) {
+			t.Errorf("%v did not register escalate_model", args)
+		}
+	}
+}
+
+// An =value form is a loud error rather than a silent enable, so a mistyped
+// --allow-escalation=false cannot turn the feature ON.
+func TestAllowEscalationRejectsAValue(t *testing.T) {
+	var stdout, stderr bytes.Buffer
+	exitCode := runWithDeps([]string{"--allow-escalation=false"}, &stdout, &stderr, appDeps{
+		getwd:  func() (string, error) { return t.TempDir(), nil },
+		runTUI: func(context.Context, tui.Options) int { return exitSuccess },
+	})
+	if exitCode == exitSuccess {
+		t.Fatal("--allow-escalation=false was accepted; a mistyped disable must not silently enable escalation")
+	}
+	if !strings.Contains(stderr.String(), "--allow-escalation takes no value") {
+		t.Errorf("stderr does not explain the flag: %s", stderr.String())
+	}
+}
+
+func TestRootHelpDocumentsAllowEscalation(t *testing.T) {
+	var stdout, stderr bytes.Buffer
+	if exitCode := runWithDeps([]string{"--help"}, &stdout, &stderr, appDeps{}); exitCode != exitSuccess {
+		t.Fatalf("exitCode = %d stderr=%s", exitCode, stderr.String())
+	}
+	if !strings.Contains(stdout.String(), "--allow-escalation") {
+		t.Error("the root help does not mention --allow-escalation, so the flag is undiscoverable")
+	}
+}

internal/cli/user_state_isolation_test.go

@@ -0,0 +1,139 @@
+package cli
+
+import (
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"github.com/Gitlawb/zero/internal/config"
+)
+
+// userStateEnvNames is every per-user base directory the root command can
+// resolve against, across platforms. One throwaway root behind all of them
+// lands the resolution inside it whichever branch it takes.
+var userStateEnvNames = []string{
+	"HOME", "USERPROFILE", "APPDATA", "LOCALAPPDATA",
+	"XDG_CONFIG_HOME", "XDG_DATA_HOME", "XDG_CACHE_HOME", "XDG_STATE_HOME",
+}
+
+// isolateCLIUserState points every user-scoped path the root command touches at
+// throwaway directories, and keeps credential work off the host keyring.
+//
+// A TEMPORARY WORKING DIRECTORY IS NOT ISOLATION. runWithDeps fills the
+// dependencies a test leaves out with the production ones, and the launch paths
+// read user config, open stores, refresh the models.dev cache and, for an
+// interactive run, migrate any inline plaintext API key into the credential
+// store before they ever reach an injected callback. Without this, running these
+// tests on a developer machine rewrites that developer's config.json and moves
+// their key into their keychain, makes a real network call, and reports results
+// that depend on whatever providers, MCP servers and plugins that machine has.
+func isolateCLIUserState(t *testing.T) {
+	t.Helper()
+	root := t.TempDir()
+	for _, name := range userStateEnvNames {
+		t.Setenv(name, root)
+	}
+	// The models.dev overlay cache, and the background refresh that writes it:
+	// a wiring test has no business making a network call or leaving a cache
+	// file behind.
+	t.Setenv("ZERO_MODELS_CACHE_PATH", filepath.Join(root, "modelsdev.json"))
+	t.Setenv("ZERO_DISABLE_MODELS_FETCH", "1")
+	// Credentials resolve keyring-first. This keeps a migrated key in a file
+	// under the fixture root rather than in the developer's OS keychain.
+	t.Setenv("ZERO_CRED_STORAGE", "encrypted-file")
+}
+
+// AND THE ISOLATION IS ITSELF PINNED, ONE HELPER AT A TIME. A seeded config
+// outside the helper fixture roots stands in for a developer's real one: the
+// launch helper must leave it exactly as it found it, inline API key included,
+// must not write a credential file beside it, and must have moved the per-user
+// paths off that root before running anything.
+//
+// EACH HELPER GETS ITS OWN SUBTEST because t.Setenv restores at the end of the
+// test, not when the helper returns: exercising both in one test would let the
+// first helper isolation cover for a second helper that had none.
+func TestCLILaunchHelpersLeaveUserStateAlone(t *testing.T) {
+	for _, testCase := range []struct {
+		name     string
+		exercise func(*testing.T)
+	}{
+		{"captureTUIOptions", func(t *testing.T) { _ = captureTUIOptions(t, "--allow-escalation") }},
+		{"execAdvertisesEscalateModel", func(t *testing.T) {
+			_ = execAdvertisesEscalateModel(t, []string{"--allow-escalation", "exec", "say hi"})
+		}},
+	} {
+		t.Run(testCase.name, func(t *testing.T) {
+			seedRoot, configPath, seeded, before := seedEmulatedUserConfig(t)
+
+			testCase.exercise(t)
+
+			after, err := os.ReadFile(configPath)
+			if err != nil {
+				t.Fatal(err)
+			}
+			if string(after) != string(seeded) {
+				t.Errorf("the seeded user config was rewritten:\n before: %s\n  after: %s", seeded, after)
+			}
+			entries, err := os.ReadDir(filepath.Dir(configPath))
+			if err != nil {
+				t.Fatal(err)
+			}
+			if len(entries) != len(before) {
+				names := []string{}
+				for _, entry := range entries {
+					names = append(names, entry.Name())
+				}
+				t.Errorf("the user config directory gained entries: %v", names)
+			}
+
+			// The checks above catch what this platform happens to write. This
+			// one catches the missing isolation itself, so a startup path that
+			// writes somewhere else, or only sometimes, is covered too.
+			resolved, err := config.DefaultUserConfigPath()
+			if err != nil {
+				t.Fatal(err)
+			}
+			if strings.HasPrefix(resolved, seedRoot) {
+				t.Errorf("user config still resolves inside the seeded root (%s): the helper installed no isolation", resolved)
+			}
+			if os.Getenv("ZERO_DISABLE_MODELS_FETCH") == "" {
+				t.Error("the helper left the background models.dev refresh enabled, so this test can make a network call and write a cache file")
+			}
+			if os.Getenv("ZERO_CRED_STORAGE") == "" {
+				t.Error("the helper left credential storage resolving keyring-first, so a migrated key can reach the host keychain")
+			}
+		})
+	}
+}
+
+// seedEmulatedUserConfig points the per-user base directories at a throwaway
+// root and writes a config carrying an inline plaintext API key there, which is
+// what the startup migration rewrites. It returns that root, the config path,
+// its bytes, and the directory listing to compare against.
+func seedEmulatedUserConfig(t *testing.T) (string, string, []byte, []os.DirEntry) {
+	t.Helper()
+	seedRoot := t.TempDir()
+	for _, name := range userStateEnvNames {
+		t.Setenv(name, seedRoot)
+	}
+	configPath, err := config.DefaultUserConfigPath()
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !strings.HasPrefix(configPath, seedRoot) {
+		t.Fatalf("SETUP INVALID: user config resolved to %q, outside the seeded root %q", configPath, seedRoot)
+	}
+	if err := os.MkdirAll(filepath.Dir(configPath), 0o755); err != nil {
+		t.Fatal(err)
+	}
+	seeded := []byte(`{"providers":[{"name":"anthropic","providerKind":"anthropic","model":"claude-haiku-4.5","apiKey":"sk-seeded-plaintext-key"}]}`)
+	if err := os.WriteFile(configPath, seeded, 0o600); err != nil {
+		t.Fatal(err)
+	}
+	before, err := os.ReadDir(filepath.Dir(configPath))
+	if err != nil {
+		t.Fatal(err)
+	}
+	return seedRoot, configPath, seeded, before
+}

internal/cli/workflows.go

@@ -2,9 +2,12 @@ package cli
 
 import (
 	"context"
+	"errors"
 	"fmt"
 	"io"
+	"path"
 	"path/filepath"
+	"regexp"
 	"strconv"
 	"strings"
 	"time"
@@ -602,18 +605,20 @@ func parseChangesArgs(args []string, command string) (changesCommandOptions, boo
 			return options, false, execUsageError{fmt.Sprintf("unexpected changes argument %q", arg)}
 		}
 	}
-	if command != "commit" && options.message != "" {
+	if command != "commit" && options.hasMessage {
 		return options, false, execUsageError{"--message is only valid with `zero changes commit`"}
 	}
-	if command != "commit" && (options.hasMessage || options.dryRun || options.auto) {
-		return options, false, execUsageError{"--message, --dry-run, and --auto are only valid with `zero changes commit`"}
-	}
 	if command == "commit" && options.hasMessage && options.auto {
 		return options, false, execUsageError{"cannot specify both --message and --auto"}
 	}
 	if command != "commit" && command != "push" && options.dryRun {
 		return options, false, execUsageError{"--dry-run is only valid with commit or push"}
 	}
+	// --auto on push/pr is the explicit opt-in for LLM branch naming (see
+	// ensureFeatureBranch); on commit it opts into the LLM commit message.
+	if command != "commit" && command != "push" && command != "pr" && options.auto {
+		return options, false, execUsageError{"--auto is only valid with commit, push, or pr"}
+	}
 	if command != "inspect" && options.baseRef != "" {
 		return options, false, execUsageError{"--base is only valid with `zero changes inspect`"}
 	}
@@ -934,8 +939,7 @@ Flags:
       --fill              Automatically populate PR title and body from commits
       --draft             Create PR as a draft
       --yes               Confirm pushing to a default/protected branch
-  -a, --auto              Auto-generate commit message using LLM (use --dry-run to preview)
-      --dry-run           Preview commit metadata without mutating git state
+  -a, --auto              Use the LLM: commit generates the message, push/pr name the auto-created branch (sends the diff to the provider)
       --json              Print JSON output
   -h, --help              Show this help
 `)
@@ -958,11 +962,31 @@ func runChangesPush(args []string, stdout io.Writer, stderr io.Writer, deps appD
 		return writeExecUsageError(stderr, err.Error())
 	}
 
+	ctx, cancel := context.WithTimeout(context.Background(), 2*time.Minute)
+	defer cancel()
+
+	branch, remote, created, err := ensureFeatureBranch(ctx, stdout, workspaceRoot, options.remote, featureBranchOptions{
+		JSONMode:           options.json,
+		AllowDefaultBranch: options.yes,
+		DryRun:             options.dryRun,
+		AutoNaming:         options.auto,
+		MaxDiffBytes:       options.maxDiffBytes,
+	}, deps)
+	if err != nil {
+		return writeExecUsageError(stderr, err.Error())
+	}
+
+	// context.Background() is deliberate: the 2-minute ctx above is the
+	// feature-branch preflight (inspection, remote probes, branch creation).
+	// The push itself must not be cancelled mid-transfer just because the
+	// preflight budget elapsed; a large push can legitimately take longer.
 	result, err := deps.pushChanges(context.Background(), zerogit.PushOptions{
 		Cwd:                    workspaceRoot,
-		Remote:                 options.remote,
+		Remote:                 firstNonEmptyString(options.remote, remote),
+		Branch:                 branch,
 		Force:                  options.force,
 		DryRun:                 options.dryRun,
+		RequireNewRemoteBranch: created,
 		AllowPushDefaultBranch: options.yes,
 	})
 	if err != nil {
@@ -1010,15 +1034,62 @@ func runChangesPR(args []string, stdout io.Writer, stderr io.Writer, deps appDep
 		return writeExecUsageError(stderr, err.Error())
 	}
 
+	ctx, cancel := context.WithTimeout(context.Background(), 2*time.Minute)
+	defer cancel()
+
+	branch, remote, created, err := ensureFeatureBranch(ctx, stdout, workspaceRoot, options.remote, featureBranchOptions{
+		JSONMode:           options.json,
+		AllowDefaultBranch: options.yes,
+		DryRun:             false,
+		AutoNaming:         options.auto,
+		MaxDiffBytes:       options.maxDiffBytes,
+	}, deps)
+	if err != nil {
+		return writeExecUsageError(stderr, err.Error())
+	}
+
+	// Resolve the target remote for the unborn-remote check. Explicit --remote wins,
+	// followed by the remote resolved by ensureFeatureBranch. When --yes is set,
+	// fall back to the branch upstream or "origin".
+	targetRemote := strings.TrimSpace(options.remote)
+	if targetRemote == "" {
+		targetRemote = remote
+	}
+	if targetRemote == "" {
+		currentBranch := ""
+		if deps.currentGitBranch != nil {
+			currentBranch = deps.currentGitBranch(context.Background(), workspaceRoot)
+		}
+		if deps.branchUpstreamRemote != nil && currentBranch != "" {
+			targetRemote = deps.branchUpstreamRemote(context.Background(), workspaceRoot, currentBranch)
+		}
+		if targetRemote == "" {
+			targetRemote = "origin"
+		}
+	}
+
+	// Probe errors are intentionally ignored so the flow can continue; if the
+	// probe fails, the subsequent push or PR operation provides fail-closed behavior.
+	if deps.isUnbornRemote != nil {
+		if unborn, unbornErr := deps.isUnbornRemote(context.Background(), workspaceRoot, targetRemote); unbornErr == nil && unborn {
+			return writeExecUsageError(stderr, fmt.Sprintf("cannot create pull request on unborn remote %s: push the initial default branch first", targetRemote))
+		}
+	}
+
 	if !options.json {
 		if _, err := fmt.Fprintln(stdout, "Pushing current branch to set upstream..."); err != nil {
 			return exitCrash
 		}
 	}
+	// context.Background() is deliberate here too: runChangesPR reuses the
+	// 2-minute ctx for preflight (branch creation and remote probes), but the
+	// push must not be cancelled mid-transfer if that budget elapsed.
 	pushResult, err := deps.pushChanges(context.Background(), zerogit.PushOptions{
 		Cwd:                    workspaceRoot,
-		Remote:                 options.remote,
+		Remote:                 firstNonEmptyString(options.remote, remote),
+		Branch:                 branch,
 		Force:                  options.force,
+		RequireNewRemoteBranch: created,
 		AllowPushDefaultBranch: options.yes,
 	})
 	if err != nil {
@@ -1100,3 +1171,462 @@ func generateAutoCommitMessage(ctx context.Context, provider zeroruntime.Provide
 	}
 	return msg, nil
 }
+
+type featureBranchOptions struct {
+	JSONMode           bool
+	AllowDefaultBranch bool
+	DryRun             bool
+	AutoNaming         bool
+	MaxDiffBytes       int
+}
+
+// branchPushPlan holds the resolved parameters for a push or PR workflow
+// before any branches are created or refs mutated. Separating plan resolution
+// from mutation allows tests to assert the resolution matrix directly without
+// touching a git repository.
+type branchPushPlan struct {
+	ShouldBranch           bool
+	CurrentBranch          string
+	PushRemote             string
+	RequireNewRemoteBranch bool
+	TrackingRemote         string
+	TrackingBranch         string
+	RestoreTip             string
+	PublishabilityBase     string
+	Summary                zerogit.ChangeSummary
+	Slug                   string
+}
+
+// resolveBranchPushPlan resolves the branch, remotes, restore tip, and naming
+// slug for changes push / pr workflows. It performs all checks (clean working
+// tree, unborn remote detection, ahead count, restore tip resolution) but
+// executes no mutation (no branch creation, ref resets, or deletions).
+func resolveBranchPushPlan(ctx context.Context, stdout io.Writer, workspaceRoot string, requestedRemote string, opts featureBranchOptions, deps appDeps) (branchPushPlan, error) {
+	if opts.AllowDefaultBranch || opts.DryRun {
+		return branchPushPlan{
+			ShouldBranch:           false,
+			PushRemote:             strings.TrimSpace(requestedRemote),
+			RequireNewRemoteBranch: false,
+		}, nil
+	}
+
+	if deps.isDefaultBranch == nil {
+		return branchPushPlan{}, fmt.Errorf("isDefaultBranch dependency missing")
+	}
+	isDefault, currentBranch, remote, err := deps.isDefaultBranch(ctx, zerogit.DefaultBranchOptions{Cwd: workspaceRoot, Remote: requestedRemote})
+	if err != nil {
+		return branchPushPlan{}, err
+	}
+	if !isDefault {
+		// Whether this push needs the nonexistence lease is decided by one
+		// live check against the destination remote, not by inference from
+		// local git state (upstream config, a generated-branch marker):
+		// every local signal that stood in for that check turned out to be
+		// wrong in some ordering (inherited config from checkout -b, a
+		// config write that failed after a successful push, a push aimed at
+		// a different remote than the one local config describes). A branch
+		// missing on the remote gets the nonexistence lease, protecting a
+		// concurrent creator of the same name; a branch already there gets a
+		// plain push, and git's own fast-forward check is the safety net —
+		// if the remote history isn't ours, the push fails with a clear
+		// non-fast-forward error instead of an attempted silent recovery.
+		if deps.remoteHasBranch == nil {
+			return branchPushPlan{
+				ShouldBranch:           false,
+				CurrentBranch:          currentBranch,
+				PushRemote:             remote,
+				RequireNewRemoteBranch: true,
+			}, nil
+		}
+		targetRemote := firstNonEmptyString(requestedRemote, remote)
+		exists, existsErr := deps.remoteHasBranch(ctx, workspaceRoot, targetRemote, currentBranch)
+		if existsErr != nil {
+			return branchPushPlan{}, fmt.Errorf("cannot check whether %s already exists on remote %s: %w", currentBranch, targetRemote, existsErr)
+		}
+		return branchPushPlan{
+			ShouldBranch:           false,
+			CurrentBranch:          currentBranch,
+			PushRemote:             remote,
+			RequireNewRemoteBranch: !exists,
+		}, nil
+	}
+
+	// CreateBranch and Push publish commits only. A dirty working tree would
+	// leave uncommitted edits behind under a branch/PR that does not include
+	// them, so refuse until the tree is clean (commit or stash first).
+	workingTree, err := deps.inspectChanges(ctx, zerogit.InspectOptions{Cwd: workspaceRoot})
+	if err != nil {
+		return branchPushPlan{}, fmt.Errorf("failed to inspect working tree: %w", err)
+	}
+	if !workingTree.Clean {
+		return branchPushPlan{}, fmt.Errorf("working tree has uncommitted changes; commit or stash them before pushing from the default branch")
+	}
+
+	// Capture the source default branch's own upstream before refreshing or
+	// leaving it. --remote selects the push destination; restore must not
+	// rewrite local main to upstream/main when main still tracks origin/main
+	// (fork setups), and refreshing must update the exact tracking ref that
+	// will supply both the ahead count and the restore target.
+	trackingRemote := remote
+	trackingBranch := currentBranch
+	if deps.branchUpstreamRemoteAndMerge != nil {
+		if r, m := deps.branchUpstreamRemoteAndMerge(ctx, workspaceRoot, currentBranch); r != "" && m != "" {
+			trackingRemote = r
+			trackingBranch = m
+		}
+	}
+
+	// Refresh the local remote-tracking ref before trusting it: IsDefaultBranch
+	// already contacted the remote for its symref check, but the tracking ref
+	// commitsAhead reads from is only a local cache (written at clone or the
+	// last fetch) and can sit behind the remote's live tip. Left stale, a real
+	// publishable range could look like zero commits ahead, and the diff
+	// derived below (which reuses this same ref as its base) would be stale
+	// too. A failure here (offline, or a genuinely unborn remote with no such
+	// ref to fetch yet) is folded into the same fail-closed path as an
+	// unresolvable ahead count below.
+	var fetchErr error
+	if deps.refreshTrackingRef != nil {
+		fetchErr = deps.refreshTrackingRef(ctx, workspaceRoot, trackingRemote, trackingBranch)
+	}
+
+	// The ahead check gates on the source upstream (trackingRemote), but the
+	// feature branch is pushed to `pushRemote`, which can be a different
+	// destination (--remote on a fork). Refuse to auto-branch onto an unborn
+	// destination: without an initial default branch the new feature branch
+	// becomes the destination's first ref and its HEAD is left pointing at it.
+	// This is checked once, here, against the push destination, before resolving
+	// commit tips or creating any branches.
+	pushRemote := firstNonEmptyString(requestedRemote, remote)
+	if deps.isUnbornRemote != nil {
+		if unbornDest, unbornDestErr := deps.isUnbornRemote(ctx, workspaceRoot, pushRemote); unbornDestErr == nil && unbornDest {
+			return branchPushPlan{}, fmt.Errorf("remote %s has no branches yet; push the initial default branch first with --yes (`zero changes push --yes`), then use auto-branching for subsequent work", pushRemote)
+		}
+	}
+
+	// Branching off the default branch only makes sense when HEAD carries a
+	// commit that is not already on the remote default branch. A clean,
+	// up-to-date default branch would otherwise publish a feature branch at the
+	// exact default tip. If the ahead count cannot be determined (for example
+	// the remote-tracking ref was never fetched), fail rather than guess. When
+	// the remote is confirmed unborn, refuse auto-branching entirely: the
+	// initial default branch must be established first (changes push --yes)
+	// so the remote has a real HEAD before any feature branch is published.
+	ahead, aheadErr := deps.commitsAhead(ctx, workspaceRoot, trackingRemote, trackingBranch)
+	if fetchErr != nil || aheadErr != nil {
+		var unbornRemote bool
+		var unbornErr error
+		if deps.isUnbornRemote != nil {
+			unbornRemote, unbornErr = deps.isUnbornRemote(ctx, workspaceRoot, trackingRemote)
+		}
+		if unbornErr == nil && unbornRemote {
+			return branchPushPlan{}, fmt.Errorf("remote %s has no branches yet; push the initial default branch first with --yes (`zero changes push --yes`), then use auto-branching for subsequent work", trackingRemote)
+		}
+		cause := aheadErr
+		if cause == nil {
+			cause = fetchErr
+		}
+		return branchPushPlan{}, fmt.Errorf("cannot determine whether HEAD is ahead of %s/%s: %w; fetch the remote tracking branch first", trackingRemote, trackingBranch, cause)
+	}
+	if ahead == 0 {
+		return branchPushPlan{}, fmt.Errorf("no changes to publish: HEAD is not ahead of %s/%s; commit your work before pushing", trackingRemote, trackingBranch)
+	}
+
+	// Resolve the restore tip to a concrete commit before creating any branch.
+	// It must never be a bare local branch name: resolving "main" yields the
+	// very commit the feature branch is about to own, so restore becomes a
+	// silent no-op and the publishable commits stay on the default branch. A
+	// named remote resolves through its just-refreshed remote-tracking ref; a
+	// direct URL or path asks the remote itself. Fail closed rather than
+	// substitute a local branch name.
+	restoreTip := ""
+	if deps.resolveRemoteBranchTip != nil {
+		tip, tipErr := deps.resolveRemoteBranchTip(ctx, workspaceRoot, trackingRemote, trackingBranch)
+		if tipErr != nil {
+			return branchPushPlan{}, fmt.Errorf("cannot resolve restore tip for %s/%s: %w", trackingRemote, trackingBranch, tipErr)
+		}
+		restoreTip = tip
+	}
+	if restoreTip == "" {
+		restoreTip = trackingRemote + "/" + trackingBranch
+	}
+
+	// Name the branch (and, with --auto, send the provider) from what HEAD is
+	// actually ahead of the resolved remote branch by, using the same ref
+	// commitsAhead just checked. A working-tree snapshot can describe edits a
+	// commit-only push will never include.
+	baseRef := restoreTip
+	summary, err := deps.inspectChanges(ctx, zerogit.InspectOptions{Cwd: workspaceRoot, BaseRef: baseRef, MaxDiffBytes: opts.MaxDiffBytes})
+	if err != nil {
+		return branchPushPlan{}, fmt.Errorf("failed to inspect changes: %w", err)
+	}
+
+	slug := fallbackBranchSlug(summary)
+	if len(summary.Files) == 0 {
+		// An empty commit (or one whose only content the diff omits) leaves
+		// nothing to derive a slug from; name the branch from the commit
+		// subject instead.
+		if subject := deps.headCommitSubject(ctx, workspaceRoot); subject != "" {
+			slug = zerogit.SlugifyBranchComponent(subject)
+		}
+	}
+	if opts.AutoNaming && strings.TrimSpace(summary.Diff) != "" {
+		llmSlug := ""
+		if resolved, cfgErr := deps.resolveConfig(workspaceRoot, config.Overrides{}); cfgErr == nil && config.HasProviderProfile(resolved.Provider) {
+			if provider, provErr := deps.newProvider(resolved.Provider); provErr == nil {
+				if !opts.JSONMode {
+					fmt.Fprintln(stdout, "Generating branch name using LLM...")
+				}
+				genCtx, cancel := context.WithTimeout(ctx, 60*time.Second)
+				generated, genErr := generateAutoBranchSlug(genCtx, provider, resolved.Provider.Model, redactChangeSummary(summary))
+				cancel()
+				if genErr == nil && generated != "" {
+					llmSlug = generated
+				}
+			}
+		}
+		if llmSlug != "" {
+			slug = llmSlug
+		} else if !opts.JSONMode {
+			// Cover resolveConfig failures, missing provider profiles,
+			// newProvider errors, generation errors, and empty results so the
+			// user is not left guessing after "Generating..." or silence.
+			fmt.Fprintln(stdout, "LLM branch naming unavailable; using deterministic name.")
+		}
+	}
+
+	return branchPushPlan{
+		ShouldBranch:           true,
+		CurrentBranch:          currentBranch,
+		PushRemote:             pushRemote,
+		RequireNewRemoteBranch: true,
+		TrackingRemote:         trackingRemote,
+		TrackingBranch:         trackingBranch,
+		RestoreTip:             restoreTip,
+		PublishabilityBase:     baseRef,
+		Summary:                summary,
+		Slug:                   slug,
+	}, nil
+}
+
+// ensureFeatureBranch is the branch-naming step `zero changes push`/`pr` run
+// before pushing: pushing straight to the default branch is refused deeper in
+// zerogit.Push, so rather than surface that as a dead end, create and switch
+// to a conventionally named "<user>/<slug>" branch first. It returns the
+// branch push/pr should target, or "" to mean "current HEAD branch, unchanged"
+// (zerogit.Push already treats an empty Branch that way), plus the remote the
+// preflight resolved (requestedRemote, then the original branch's configured
+// upstream, then "origin"), plus whether Push should require that branch not
+// already exist on that remote. Callers must pass the remote to Push: a
+// freshly created branch has no tracking configuration, so Push's own
+// fallback would silently retarget "origin" even when the work came from a
+// branch tracking a different remote. Callers must also pass that third value
+// through as Push's RequireNewRemoteBranch: CreateBranch's own
+// remote-collision probe runs before this returns, and closing that race
+// requires Push's push itself to assert the destination is still new. That
+// requirement also carries across a retry: if the current branch is already
+// non-default (this call didn't create it) but has not been successfully
+// published to the exact target remote under the same branch name, keep the
+// nonexistence lease. branch.<name>.remote alone is not enough: with
+// branch.autoSetupMerge=inherit, checkout -b copies remote=origin from the
+// source branch before any push. After a successful create, the original
+// default branch ref is moved back to its own upstream tip (not necessarily
+// the push destination) so the feature branch exclusively owns the
+// publishable commits; a failed restore rolls the feature branch back.
+// allowDefaultBranch (the --yes flag) and dryRun both opt out via the ""
+// branch / false return, leaving Push's own guard/preview behavior on the
+// default branch unaffected.
+//
+// autoNaming gates the LLM naming path (--auto): these commands were
+// git-only, and sending the change diff to a configured provider on every
+// default-branch push would silently export source code nobody asked to
+// share. Without the opt-in the name comes from deterministic local
+// information only. maxDiffBytes caps the committed-range diff Inspect
+// returns, so a user who passed --diff-bytes to bound the proprietary source
+// sent for LLM naming has that cap honored here just as the commit path does.
+// The working tree must be clean and HEAD must be ahead of the resolved
+// remote branch; both are verified before any branch is created. The
+// inspectChanges, commitsAhead, headCommitSubject, currentGitUser, and
+// createBranch dependencies are required on the default-branch path and are
+// called without nil guards; fillAppDeps populates all of them. The
+// refreshTrackingRef, resolveRemoteBranchTip, isUnbornRemote, resetBranchRef,
+// and deleteBranch dependencies are optional and are skipped when nil.
+func ensureFeatureBranch(ctx context.Context, stdout io.Writer, workspaceRoot string, requestedRemote string, opts featureBranchOptions, deps appDeps) (string, string, bool, error) {
+	plan, err := resolveBranchPushPlan(ctx, stdout, workspaceRoot, requestedRemote, opts, deps)
+	if err != nil {
+		return "", "", false, err
+	}
+	if !plan.ShouldBranch {
+		return plan.CurrentBranch, plan.PushRemote, plan.RequireNewRemoteBranch, nil
+	}
+
+	expectedRestoreTip := ""
+	if deps.currentBranchTip != nil {
+		expectedRestoreTip = deps.currentBranchTip(ctx, workspaceRoot)
+	}
+
+	name := zerogit.BuildBranchName(deps.currentGitUser(ctx, workspaceRoot), plan.Slug)
+	result, err := deps.createBranch(ctx, zerogit.BranchOptions{Cwd: workspaceRoot, Name: name, Remote: plan.PushRemote})
+	if err != nil {
+		return "", "", false, fmt.Errorf("failed to create branch: %w", err)
+	}
+	// The normal flow commits on the default branch first, then creates the
+	// feature branch at the same HEAD. Without moving the original default
+	// ref back, local main keeps the pre-squash commits and diverges from
+	// its upstream after a squash-merge (and a later pull/push can re-publish
+	// them). Once the feature branch owns those commits, restore the default
+	// branch to the source branch's own upstream tip captured above.
+	// Failure-safe: if the restore fails, delete the feature branch and
+	// check the default branch back out so the tree is not left half-moved.
+	if deps.resetBranchRef != nil {
+		if err := deps.resetBranchRef(ctx, workspaceRoot, plan.CurrentBranch, plan.RestoreTip, expectedRestoreTip); err != nil {
+			// A compare-and-swap conflict means the default branch moved
+			// while we were branching, so restoreTip is no longer the tip we
+			// meant to restore. The rollback below deletes the branch that
+			// now exclusively owns the user's commits, which would destroy
+			// work to recover from a race we can simply report. Keep the
+			// branch and let the user reconcile the default branch.
+			if errors.Is(err, zerogit.ErrCompareAndSwapConflict) {
+				return "", "", false, fmt.Errorf("failed to restore default branch %s to %s after auto-branching: %w; generated branch %s was preserved because the default branch changed concurrently", plan.CurrentBranch, plan.RestoreTip, err, result.Branch)
+			}
+			var rollbackErr error
+			if deps.deleteBranch != nil {
+				rollbackErr = deps.deleteBranch(ctx, workspaceRoot, plan.CurrentBranch, result.Branch)
+			}
+			if rollbackErr != nil {
+				return "", "", false, fmt.Errorf("failed to restore default branch %s to %s after auto-branching: %w; rollback of %s also failed: %v", plan.CurrentBranch, plan.RestoreTip, err, result.Branch, rollbackErr)
+			}
+			return "", "", false, fmt.Errorf("failed to restore default branch %s to %s after auto-branching: %w", plan.CurrentBranch, plan.RestoreTip, err)
+		}
+	}
+	if !opts.JSONMode {
+		fmt.Fprintf(stdout, "Created branch %s (was on %s)\n", result.Branch, plan.CurrentBranch)
+	}
+	return result.Branch, plan.PushRemote, true, nil
+}
+
+// fallbackBranchSlug derives a deterministic branch-name slug from a change
+// summary without calling an LLM, so ensureFeatureBranch still works when no
+// provider is configured.
+func fallbackBranchSlug(summary zerogit.ChangeSummary) string {
+	switch len(summary.Files) {
+	case 0:
+		return "changes"
+	case 1:
+		// Git paths are always slash-separated regardless of platform, so
+		// path.Base (not filepath.Base) extracts the filename portably.
+		return zerogit.SlugifyBranchComponent(path.Base(summary.Files[0].Path))
+	default:
+		return fmt.Sprintf("update-%d-files", len(summary.Files))
+	}
+}
+
+// generateAutoBranchSlug asks the model for a short kebab-case slug
+// describing the diff, mirroring generateAutoCommitMessage's prompt shape.
+func generateAutoBranchSlug(ctx context.Context, provider zeroruntime.Provider, model string, summary zerogit.ChangeSummary) (string, error) {
+	var promptBuilder strings.Builder
+	promptBuilder.WriteString("Analyze the following git diff and generate a short git branch name slug for it.\n")
+	promptBuilder.WriteString("The slug must be 2 to 5 lowercase words separated by hyphens (kebab-case), using only letters, digits, and hyphens, with no prefix like \"feature/\" or \"fix/\" and no surrounding quotes.\n")
+	promptBuilder.WriteString("Output ONLY the raw slug text, nothing else.\n\n")
+	promptBuilder.WriteString("Git Diff:\n")
+	promptBuilder.WriteString(summary.Diff)
+
+	request := zeroruntime.CompletionRequest{
+		Messages: []zeroruntime.Message{
+			{Role: zeroruntime.MessageRoleUser, Content: promptBuilder.String()},
+		},
+	}
+	stream, err := provider.StreamCompletion(ctx, request)
+	if err != nil {
+		return "", err
+	}
+	collected := zeroruntime.CollectStream(ctx, stream)
+	if collected.Error != "" {
+		return "", fmt.Errorf("%s", collected.Error)
+	}
+
+	slug := zerogit.SlugifyBranchComponent(extractBranchSlug(collected.Text))
+	if slug == "" {
+		return "", fmt.Errorf("provider returned empty branch slug")
+	}
+	return slug, nil
+}
+
+// slugLineRe matches a line that already reads as a kebab-case slug (letters,
+// digits, and internal single hyphens only). extractBranchSlug prefers such a
+// line so a preamble sentence is never mistaken for the slug.
+var slugLineRe = regexp.MustCompile(`^[a-zA-Z0-9]+(-[a-zA-Z0-9]+)*$`)
+
+func isPreambleText(s string) bool {
+	lower := strings.ToLower(strings.TrimSpace(s))
+	if strings.HasSuffix(lower, ":") || strings.HasSuffix(lower, ".") || strings.HasSuffix(lower, "!") || strings.HasSuffix(lower, "?") {
+		return true
+	}
+	preambles := []string{
+		"here is", "here's", "below is", "below are", "sure", "certainly",
+		"i suggest", "suggested branch", "branch name", "recommended branch",
+		"how about", "you could use",
+	}
+	for _, p := range preambles {
+		if strings.HasPrefix(lower, p) {
+			return true
+		}
+	}
+	return false
+}
+
+// extractBranchSlug pulls the intended slug out of a model response that didn't
+// follow the "output only the raw slug" instruction exactly. It drops Markdown
+// code-fence lines, then prefers a line that already looks like a kebab-case
+// slug: that skips a leading preamble such as "Here is a suggested branch
+// name:" in favor of the "add-login-page" line that follows it. One-word
+// acknowledgements that happen to match the slug regex ("Sure", "Certainly")
+// are filtered as preamble before the early return, so "Sure\nadd login page"
+// yields the real suggestion. Inline labeled forms ("Branch name: add-login-page")
+// still work: the preamble check runs on the extracted value after the label
+// is stripped. If no line is strictly kebab-cased, it prefers a plausible
+// non-preamble line so plain multi-word replies still slugify correctly.
+func extractBranchSlug(text string) string {
+	var firstLine string
+
+	for _, line := range strings.Split(text, "\n") {
+		line = strings.TrimSpace(line)
+		if strings.HasPrefix(line, "```") {
+			continue
+		}
+		line = strings.TrimSpace(strings.Trim(line, `"'`))
+		if line == "" {
+			continue
+		}
+
+		candidate := line
+		if idx := strings.Index(line, ":"); idx != -1 {
+			after := strings.TrimSpace(strings.Trim(line[idx+1:], `"'`))
+			if after != "" {
+				candidate = after
+			}
+		}
+		// Models often terminate a suggestion with sentence punctuation;
+		// strip it so "add login page." is recovered as the slug rather than
+		// rejected as preamble (isPreambleText treats a trailing '.' as a
+		// full sentence). True preamble lines still match the prefix list.
+		candidate = strings.TrimSpace(strings.TrimRight(candidate, ".!?"))
+		if candidate == "" {
+			continue
+		}
+
+		// Preamble classification runs before accepting a slug-shaped line so
+		// "Sure" does not win over a later "add-login-page". The check is on
+		// candidate (post label-strip) so "Branch name: add-login-page" still
+		// returns the extracted slug.
+		if slugLineRe.MatchString(candidate) && !isPreambleText(candidate) {
+			return candidate
+		}
+
+		if firstLine == "" && !isPreambleText(candidate) {
+			firstLine = candidate
+		}
+	}
+
+	return firstLine
+}

internal/config/command_test.go

@@ -105,7 +105,17 @@ func TestLoadProviderCommandTimeout(t *testing.T) {
 //
 // This is a hang-breaker, not a tuned margin. Nothing in a healthy run waits on
 // it, since both tests return about one WaitDelay after the fixture exits.
-const providerCommandTestBudget = 60 * time.Second
+const (
+	providerCommandTestBudget        = 90 * time.Second
+	providerCommandReadyWaitAttempts = 45
+)
+
+func TestProviderCommandBudgetLeavesCleanupMargin(t *testing.T) {
+	readyWaitCeiling := time.Duration(providerCommandReadyWaitAttempts) * time.Second
+	if providerCommandTestBudget < 2*readyWaitCeiling {
+		t.Fatalf("provider command test budget %s leaves less than one readiness window for cleanup after %s", providerCommandTestBudget, readyWaitCeiling)
+	}
+}
 
 // backgroundChildLifetime must exceed providerCommandTestBudget so that a child
 // dying of old age is unobservable: any run where runProviderCommand has not
@@ -341,15 +351,15 @@ func writeCommand(t *testing.T, script commandScript) string {
 			// assertion fail is far more diagnosable than blocking until the
 			// timeout fires and reporting an unrelated error. Each iteration is
 			// about a second, since `ping -n 2` sleeps between its two echoes,
-			// so the ceiling is roughly 15s: far longer than any cold PowerShell
-			// start seen on a runner, and far shorter than
-			// providerCommandTestBudget, so giving up surfaces a legible
+			// so the ceiling is roughly 45s: enough for a heavily loaded Windows
+			// runner while leaving the other half of providerCommandTestBudget for
+			// process exit, PID-file verification, and cleanup. Giving up surfaces a legible
 			// PID-file failure rather than a timeout that hides the real cause.
 			lines = append(lines,
 				":zeroWaitReady",
 				"if exist \""+readyFile+"\" goto zeroReady",
 				"set /a ZERO_READY_TRIES+=1",
-				"if %ZERO_READY_TRIES% GEQ 15 goto zeroReady",
+				"if %ZERO_READY_TRIES% GEQ "+itoa(providerCommandReadyWaitAttempts)+" goto zeroReady",
 				"ping -n 2 127.0.0.1 >nul",
 				"goto zeroWaitReady",
 				":zeroReady",

internal/config/concurrent_writer_test.go

@@ -0,0 +1,396 @@
+package config
+
+import (
+	"encoding/json"
+	"errors"
+	"fmt"
+	"io"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"strings"
+	"sync"
+	"testing"
+	"time"
+)
+
+func readTestConfig(t *testing.T, path string) FileConfig {
+	t.Helper()
+	data, err := os.ReadFile(path)
+	if err != nil {
+		t.Fatalf("read config: %v", err)
+	}
+	cfg := FileConfig{}
+	if err := json.Unmarshal(data, &cfg); err != nil {
+		t.Fatalf("invalid config JSON after concurrent writes: %v\n%s", err, data)
+	}
+	return cfg
+}
+
+func seedTestConfig(t *testing.T, path string) {
+	t.Helper()
+	if _, err := UpsertProvider(path, ProviderProfile{Name: "seed", Model: "seed-model"}, true); err != nil {
+		t.Fatalf("seed config: %v", err)
+	}
+}
+
+// TestConcurrentMutationsDoNotLoseUpdates is the issue #832 regression. Each
+// mutator loads the whole document, edits one field, and publishes a complete
+// replacement by rename. Without a lock spanning load-through-publish, two
+// writers that loaded the same revision both write a full document and the
+// second rename silently discards the first one's acknowledged update.
+//
+// Every mutation below touches an INDEPENDENT field, so a correct
+// implementation ends with all of them present. A lost update shows up as a
+// zero value in exactly one of the fields, with valid JSON either way — which
+// is what made the bug silent.
+func TestConcurrentMutationsDoNotLoseUpdates(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "config.json")
+	seedTestConfig(t, path)
+
+	mutations := []struct {
+		name  string
+		apply func() error
+		check func(FileConfig) error
+	}{
+		{
+			name:  "theme",
+			apply: func() error { _, err := SetTheme(path, "dracula"); return err },
+			check: func(cfg FileConfig) error {
+				if cfg.Preferences.Theme != "dracula" {
+					return fmt.Errorf("theme = %q, want dracula", cfg.Preferences.Theme)
+				}
+				return nil
+			},
+		},
+		{
+			name:  "pet",
+			apply: func() error { _, err := SetPet(path, "otter"); return err },
+			check: func(cfg FileConfig) error {
+				if cfg.Preferences.Pet != "otter" {
+					return fmt.Errorf("pet = %q, want otter", cfg.Preferences.Pet)
+				}
+				return nil
+			},
+		},
+		{
+			name:  "recaps",
+			apply: func() error { _, err := SetRecapsEnabled(path, true); return err },
+			check: func(cfg FileConfig) error {
+				if cfg.Preferences.Recaps == nil || !*cfg.Preferences.Recaps {
+					return fmt.Errorf("recaps = %v, want true", cfg.Preferences.Recaps)
+				}
+				return nil
+			},
+		},
+		{
+			name:  "favorites",
+			apply: func() error { _, err := SetFavoriteModels(path, []string{"fav-model"}); return err },
+			check: func(cfg FileConfig) error {
+				if len(cfg.Preferences.FavoriteModels) != 1 || cfg.Preferences.FavoriteModels[0] != "fav-model" {
+					return fmt.Errorf("favorites = %v, want [fav-model]", cfg.Preferences.FavoriteModels)
+				}
+				return nil
+			},
+		},
+		{
+			name: "provider",
+			apply: func() error {
+				_, err := UpsertProvider(path, ProviderProfile{Name: "added", Model: "added-model"}, false)
+				return err
+			},
+			check: func(cfg FileConfig) error {
+				for _, provider := range cfg.Providers {
+					if provider.Name == "added" {
+						return nil
+					}
+				}
+				return fmt.Errorf("provider %q missing from %d providers", "added", len(cfg.Providers))
+			},
+		},
+	}
+
+	// SetPet edits the raw bytes rather than round-tripping the struct, so it is
+	// included deliberately: it must take the same lock as the struct writers or
+	// it would clobber, and be clobbered by, everything else.
+	errs := make([]error, len(mutations))
+	start := make(chan struct{})
+	var wg sync.WaitGroup
+	for index, mutation := range mutations {
+		wg.Add(1)
+		go func() {
+			defer wg.Done()
+			<-start
+			errs[index] = mutation.apply()
+		}()
+	}
+	close(start)
+	wg.Wait()
+
+	for index, err := range errs {
+		if err != nil {
+			t.Fatalf("%s mutation failed: %v", mutations[index].name, err)
+		}
+	}
+	cfg := readTestConfig(t, path)
+	for _, mutation := range mutations {
+		if err := mutation.check(cfg); err != nil {
+			t.Errorf("%s update was lost: %v", mutation.name, err)
+		}
+	}
+	if cfg.ActiveProvider != "seed" {
+		t.Errorf("activeProvider = %q, want the seeded value preserved", cfg.ActiveProvider)
+	}
+}
+
+// TestConcurrentSameFieldMutationsSerialize proves the lock serializes writers
+// that contend on ONE field: every call must succeed and the file must end with
+// exactly one of the written values, never a merged or truncated document.
+func TestConcurrentSameFieldMutationsSerialize(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "config.json")
+	seedTestConfig(t, path)
+
+	const writers = 24
+	errs := make([]error, writers)
+	start := make(chan struct{})
+	var wg sync.WaitGroup
+	for index := range writers {
+		wg.Add(1)
+		go func() {
+			defer wg.Done()
+			<-start
+			_, errs[index] = SetTheme(path, fmt.Sprintf("theme-%02d", index))
+		}()
+	}
+	close(start)
+	wg.Wait()
+
+	for index, err := range errs {
+		if err != nil {
+			t.Fatalf("writer %d failed: %v", index, err)
+		}
+	}
+	cfg := readTestConfig(t, path)
+	if !strings.HasPrefix(cfg.Preferences.Theme, "theme-") {
+		t.Fatalf("theme = %q, want one of the written values", cfg.Preferences.Theme)
+	}
+}
+
+// TestConcurrentProviderUpsertsAllSurvive is the shape the issue describes most
+// directly: N distinct providers added at once. Each add is an independent
+// successful mutation, so all N must be present afterwards.
+func TestConcurrentProviderUpsertsAllSurvive(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "config.json")
+	seedTestConfig(t, path)
+
+	const providers = 16
+	errs := make([]error, providers)
+	start := make(chan struct{})
+	var wg sync.WaitGroup
+	for index := range providers {
+		wg.Add(1)
+		go func() {
+			defer wg.Done()
+			<-start
+			_, errs[index] = UpsertProvider(path, ProviderProfile{
+				Name:  fmt.Sprintf("provider-%02d", index),
+				Model: fmt.Sprintf("model-%02d", index),
+			}, false)
+		}()
+	}
+	close(start)
+	wg.Wait()
+
+	for index, err := range errs {
+		if err != nil {
+			t.Fatalf("upsert %d failed: %v", index, err)
+		}
+	}
+	cfg := readTestConfig(t, path)
+	present := map[string]string{}
+	for _, provider := range cfg.Providers {
+		present[provider.Name] = provider.Model
+	}
+	for index := range providers {
+		name := fmt.Sprintf("provider-%02d", index)
+		want := fmt.Sprintf("model-%02d", index)
+		if got, ok := present[name]; !ok {
+			t.Errorf("provider %s was lost (config holds %d providers)", name, len(cfg.Providers))
+		} else if got != want {
+			t.Errorf("provider %s model = %q, want %q", name, got, want)
+		}
+	}
+}
+
+// TestCrossProcessMutationExcludesAndPreserves is the coordinated two-process
+// test the issue asks for. Goroutines share this process's descriptors, so only
+// a second OS process shows the lock is held by the kernel rather than by
+// in-process state.
+//
+// The child announces itself and then mutates; the parent holds the lock across
+// that window and asserts the child's write CANNOT land, then does its own
+// mutation and releases. Both updates must survive. The exclusion window is
+// what makes this deterministic in the passing direction: with the lock held,
+// the child is blocked in its retry loop and the theme provably stays unset.
+func TestCrossProcessMutationExcludesAndPreserves(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "config.json")
+	seedTestConfig(t, path)
+
+	unlock, err := lockConfigFile(path)
+	if err != nil {
+		t.Fatalf("acquire lock: %v", err)
+	}
+	released := false
+	release := func() {
+		if !released {
+			released = true
+			unlock()
+		}
+	}
+	defer release()
+
+	child := exec.Command(os.Args[0], "-test.run=^TestConfigConcurrentHelperProcess$")
+	child.Env = append(os.Environ(),
+		"ZERO_CONFIG_CONCURRENT_HELPER=1",
+		"ZERO_CONFIG_CONCURRENT_PATH="+path,
+	)
+	stdout, err := child.StdoutPipe()
+	if err != nil {
+		t.Fatalf("stdout pipe: %v", err)
+	}
+	if err := child.Start(); err != nil {
+		t.Fatalf("start helper: %v", err)
+	}
+	defer func() {
+		if child.ProcessState == nil {
+			_ = child.Process.Kill()
+			_ = child.Wait()
+		}
+	}()
+
+	ready := make([]byte, len("ready\n"))
+	if _, err := io.ReadFull(stdout, ready); err != nil {
+		t.Fatalf("helper never signalled ready: %v", err)
+	}
+
+	// The child is now live and contending for a lock this process holds. Its
+	// write must not appear while we hold it.
+	for range 20 {
+		if theme := readTestConfig(t, path).Preferences.Theme; theme != "" {
+			t.Fatalf("child wrote %q while the lock was held elsewhere; mutations are not excluded", theme)
+		}
+		time.Sleep(5 * time.Millisecond)
+	}
+
+	if _, err := upsertProviderLocked(path, ProviderProfile{Name: "parent", Model: "parent-model"}, false); err != nil {
+		t.Fatalf("parent mutation: %v", err)
+	}
+	release()
+
+	if err := child.Wait(); err != nil {
+		t.Fatalf("helper process failed: %v", err)
+	}
+
+	cfg := readTestConfig(t, path)
+	if cfg.Preferences.Theme != "child-theme" {
+		t.Errorf("child update was lost: theme = %q, want child-theme", cfg.Preferences.Theme)
+	}
+	found := false
+	for _, provider := range cfg.Providers {
+		if provider.Name == "parent" {
+			found = true
+		}
+	}
+	if !found {
+		t.Errorf("parent update was lost: provider %q missing from %d providers", "parent", len(cfg.Providers))
+	}
+}
+
+// TestConfigConcurrentHelperProcess is the child half of the cross-process
+// test. It is skipped unless the parent selected it through the environment.
+func TestConfigConcurrentHelperProcess(t *testing.T) {
+	if os.Getenv("ZERO_CONFIG_CONCURRENT_HELPER") == "" {
+		t.Skip("helper process for TestCrossProcessMutationExcludesAndPreserves")
+	}
+	path := os.Getenv("ZERO_CONFIG_CONCURRENT_PATH")
+	if path == "" {
+		t.Fatal("ZERO_CONFIG_CONCURRENT_PATH is required")
+	}
+	// Announce BEFORE contending, so the parent's exclusion window starts with
+	// this process already running rather than still being forked.
+	fmt.Println("ready")
+	if _, err := SetTheme(path, "child-theme"); err != nil {
+		t.Fatalf("child mutation: %v", err)
+	}
+}
+
+// TestMutationReportsUnlockFailure pins the AGENTS.md rule that a mutator must
+// never report success when unlock failed. A failed release can leave the lock
+// held for the rest of the process, so returning (cfg, nil) after one would
+// claim a state the next mutation cannot reproduce.
+//
+// lockutil.Release is idempotent and reports nil once released, so a real
+// release failure cannot be provoked through the public API — hence the seam.
+// The mutation itself must still have been published: the release failure
+// annotates the result, it does not undo the write.
+func TestMutationReportsUnlockFailure(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "config.json")
+	seedTestConfig(t, path)
+
+	releaseErr := errors.New("release exploded")
+	original := lockConfigFileFn
+	t.Cleanup(func() { lockConfigFileFn = original })
+	lockConfigFileFn = func(p string) (func() error, error) {
+		unlock, err := original(p)
+		if err != nil {
+			return nil, err
+		}
+		return func() error {
+			_ = unlock()
+			return releaseErr
+		}, nil
+	}
+
+	cfg, err := SetTheme(path, "dracula")
+	if !errors.Is(err, releaseErr) {
+		t.Fatalf("SetTheme err = %v, want it to carry the release failure", err)
+	}
+	if cfg.Preferences.Theme != "dracula" {
+		t.Errorf("returned config lost the mutation: theme = %q", cfg.Preferences.Theme)
+	}
+	if got := readTestConfig(t, path).Preferences.Theme; got != "dracula" {
+		t.Errorf("mutation was not published despite only the release failing: theme = %q", got)
+	}
+}
+
+// TestMutationErrorSurvivesUnlockFailure proves the two errors are joined
+// rather than chosen between: a release failure must not mask the mutation
+// error that actually explains what went wrong.
+func TestMutationErrorSurvivesUnlockFailure(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "config.json")
+	seedTestConfig(t, path)
+
+	releaseErr := errors.New("release exploded")
+	original := lockConfigFileFn
+	t.Cleanup(func() { lockConfigFileFn = original })
+	lockConfigFileFn = func(p string) (func() error, error) {
+		unlock, err := original(p)
+		if err != nil {
+			return nil, err
+		}
+		return func() error {
+			_ = unlock()
+			return releaseErr
+		}, nil
+	}
+
+	// An unknown provider is the mutation error; the release failure is joined
+	// onto it, and neither one disappears.
+	_, err := SetActiveProvider(path, "definitely-not-configured")
+	if !errors.Is(err, releaseErr) {
+		t.Fatalf("err = %v, want it to carry the release failure", err)
+	}
+	if !strings.Contains(err.Error(), "definitely-not-configured") {
+		t.Fatalf("err = %v, want it to still name the missing provider", err)
+	}
+}

internal/config/credentials.go

@@ -2,6 +2,7 @@ package config
 
 import (
 	"encoding/json"
+	"errors"
 	"fmt"
 	"os"
 	"path/filepath"
@@ -80,12 +81,21 @@ func ForgetProviderKey(provider string) (bool, error) {
 // config at path, so credential checks no longer claim a stored key after one is
 // removed. No-op when path/provider is empty, the config is absent, or the marker
 // is already unset.
-func ClearProviderKeyStored(path, provider string) (bool, error) {
+func ClearProviderKeyStored(path, provider string) (cleared bool, err error) {
 	path = strings.TrimSpace(path)
 	provider = strings.TrimSpace(provider)
 	if path == "" || provider == "" {
 		return false, nil
 	}
+	unlock, err := lockConfigFileFn(path)
+	if err != nil {
+		return false, err
+	}
+	// Joined, not chosen between: a release failure annotates the result
+	// instead of masking the mutation error that actually explains what went
+	// wrong. Reporting success after a failed unlock would claim a state the
+	// next mutation cannot reproduce.
+	defer func() { err = errors.Join(err, unlock()) }()
 	data, err := os.ReadFile(path)
 	if err != nil {
 		if os.IsNotExist(err) {
@@ -116,11 +126,22 @@ func ClearProviderKeyStored(path, provider string) (bool, error) {
 // leaves the plaintext key in place and never strands a credential. Returns how
 // many keys were migrated; a no-op (0, nil) when path is empty/absent or nothing
 // needs migrating. Safe to run on every startup (idempotent).
-func MigratePlaintextProviderKeys(path string, store APIKeySetter) (int, error) {
+func MigratePlaintextProviderKeys(path string, store APIKeySetter) (count int, err error) {
 	path = strings.TrimSpace(path)
 	if path == "" || store == nil {
 		return 0, nil
 	}
+	// This runs on every startup, so it is the most likely writer to be racing
+	// an interactive mutation in another Zero process.
+	unlock, err := lockConfigFileFn(path)
+	if err != nil {
+		return 0, err
+	}
+	// Joined, not chosen between: a release failure annotates the result
+	// instead of masking the mutation error that actually explains what went
+	// wrong. Reporting success after a failed unlock would claim a state the
+	// next mutation cannot reproduce.
+	defer func() { err = errors.Join(err, unlock()) }()
 	data, err := os.ReadFile(path)
 	if err != nil {
 		if os.IsNotExist(err) {