fix(tools): resolve CodeRabbit residual capability/key issues

- Join apply_patch diff paths under cwd for conflict keys
- Validate constructor-declared capabilities before normalization
- Unify session/process ID key extraction helper

.github/workflows/release-artifacts.yml

@@ -97,16 +97,21 @@ jobs:
           gh release create "$tag" --repo "$GITHUB_REPOSITORY" --title "$tag" --generate-notes 2>/dev/null || true
           gh release upload "$tag" dist/release/* --repo "$GITHUB_REPOSITORY" --clobber
 
-  # Publishes the npm wrapper (@gitlawb/zero) once every platform's release
-  # assets are live — the package's postinstall downloads its binary from the
-  # GitHub Release, so publishing before the assets exist would ship a package
-  # that fails to install.
+  # Publishes @gitlawb/zero to npm: first the five platform payloads (versions
+  # of the SAME package at suffixed versions like X.Y.Z-linux-x64, carrying the
+  # native binary + vendored helpers), then the wrapper at X.Y.Z whose
+  # optionalDependencies alias those platform versions. Platform versions MUST
+  # publish first — the wrapper pins them exactly, so the reverse order would
+  # open a window where installs resolve aliases that 404. See
+  # docs/NPM_PACKAGING.md for the full contract.
   #
   # Auth is npm OIDC trusted publishing: no NPM_TOKEN secret. The package's
   # npmjs.com settings must list this repo + workflow file as a trusted
   # publisher, and provenance is then generated automatically. Note npm only
   # allows configuring a trusted publisher on an EXISTING package, so the very
   # first @gitlawb/zero publish is a one-time manual `npm publish` bootstrap.
+  # Because the platform payloads are versions of the same package, they reuse
+  # that single trusted-publisher configuration.
   publish-npm:
     name: Publish npm package
     needs: package
@@ -156,10 +161,11 @@ jobs:
             console.log(`npm ${version} supports trusted publishing.`);
           '
 
-      # Probe the exact URLs postinstall will build (internal/release/release.go
-      # asset-name scheme) as an anonymous client. This also blocks publishing
-      # while the repo is still private — private release assets 404 without
-      # auth, exactly as they would for an installing user.
+      # Probe the exact URLs the wrapper's first-run downloader builds
+      # (internal/release/release.go asset-name scheme) as an anonymous client
+      # — install.sh/install.ps1 and `zero update` read them too. This also
+      # blocks publishing while the repo is still private: private release
+      # assets 404 without auth, exactly as they would for an installing user.
       - name: Verify release assets are downloadable
         shell: bash
         run: |
@@ -173,12 +179,64 @@ jobs:
             "zero-v${version}-windows-x64.zip"; do
             for url in "${base}/${asset}" "${base}/${asset}.sha256"; do
               if ! curl -fsSLI -o /dev/null "$url"; then
-                echo "::error::${url} is not publicly downloadable — npm postinstall would fail. Is the repo public and are all assets uploaded?" >&2
+                echo "::error::${url} is not publicly downloadable — the wrapper's fallback downloader and install.sh would fail. Is the repo public and are all assets uploaded?" >&2
                 exit 1
               fi
             done
           done
           echo "all release assets for v${version} are publicly downloadable."
 
-      - name: Publish to npm
-        run: npm publish --access public
+      - name: Download release artifacts
+        uses: actions/download-artifact@d3f86a106a0bac45b974a628896c90dbdf5c8093 # v4.3.0
+        with:
+          pattern: zero-*
+          path: dist/artifacts
+          merge-multiple: true
+
+      - name: Assemble npm packages
+        run: node scripts/npm/build-platform-packages.mjs --artifacts-dir dist/artifacts --out-dir dist/npm
+
+      # Platform versions are semver PRERELEASES of the wrapper version
+      # (X.Y.Z-linux-x64), so they must never publish under the default
+      # `latest` dist-tag — that would make `npm install @gitlawb/zero`
+      # deliver a platform payload as the CLI. Publish them under a dedicated
+      # tag and assert `latest` survived before the wrapper goes out.
+      - name: Publish platform packages to npm
+        shell: bash
+        run: |
+          version="$(node -p "require('./package.json').version")"
+          for dir in dist/npm/platforms/*/; do
+            echo "publishing ${dir}"
+            (cd "$dir" && npm publish --access public --tag platform)
+          done
+          # Tolerate a failed lookup (registry lag, or no latest tag yet on a
+          # bootstrap publish): under set -e a bare failure here would abort
+          # AFTER the platform payloads went out, leaving no wrapper and a
+          # re-run that fails on republishing existing versions. An empty
+          # value falls through the guard below.
+          latest="$(npm view @gitlawb/zero dist-tags.latest || true)"
+          case "$latest" in
+            "${version}-"*)
+              echo "::error::dist-tag latest points at platform version ${latest} — a platform publish clobbered it. Fix the dist-tag before the wrapper publish." >&2
+              exit 1
+              ;;
+          esac
+          echo "dist-tag latest is ${latest} (unchanged by platform publishes)."
+
+      - name: Publish wrapper to npm
+        shell: bash
+        run: |
+          version="$(node -p "require('./package.json').version")"
+          (cd dist/npm/wrapper && npm publish --access public)
+          # Confirm the wrapper owns `latest` (retry: registry metadata can lag
+          # a publish by a few seconds).
+          for attempt in 1 2 3 4 5; do
+            latest="$(npm view @gitlawb/zero dist-tags.latest || true)"
+            if [ "$latest" = "$version" ]; then
+              echo "dist-tag latest is ${version}."
+              exit 0
+            fi
+            sleep 5
+          done
+          echo "::error::dist-tag latest is ${latest}, expected ${version} after the wrapper publish." >&2
+          exit 1

AGENTS.md

@@ -1,326 +1,27 @@
 ---
-description: Extending Zero — how to write AGENTS.md files, custom specialists, skills, hooks, MCP servers, and plugins for the open-source CLI coding agent.
+description: Project guidelines for the open-source CLI coding agent (zero) repository.
 globs: "*.go, *.js, *.md, *.json, *.toml, *.yaml, *.yml"
 alwaysApply: false
 ---
 
-# Extending Zero
+# Repository Conventions for Zero
 
-Zero is an open-source terminal coding agent. Out of the box it does the obvious things — read, edit, run, search — but the design point of the project is that **every surface is configurable**. This document is the user-facing guide for that configuration.
+This file outlines the project conventions and repository guidelines for coding agents when working on the `zero` repository. For the general guide on how to extend Zero (write specialist sub-agents, hooks, plugins, MCP, skills), see [docs/EXTENDING.md](docs/EXTENDING.md).
 
-If you only want to *use* Zero, the [README](README.md) is enough. This page is for the other three jobs:
-
-1. Tell the agent about *your* project (drop an `AGENTS.md` in your repo).
-2. Add new specialist sub-agents.
-3. Wire Zero into the rest of your toolchain (MCP, skills, hooks, plugins).
-
-## 1. Drop a project `AGENTS.md`
-
-When Zero starts in a directory, it looks for project-level instructions and injects them into the system prompt. The lookup walks from your current working directory **up to the nearest git root** and reads the first matching file at each level — general rules at the repo root, more specific rules in sub-trees. Files are labeled with their directory in the prompt (e.g. `## Project guidelines (services/api/AGENTS.md)`).
-
-Accepted file names, in priority order at each level:
-
-| Path | Notes |
-| --- | --- |
-| `./AGENTS.md` | The classic spot — committed to your repo, shared with the team. |
-| `./ZERO.md` | Brand-specific alias. Same format, lower priority. |
-| `./.zero/AGENTS.md` | Project-local, hidden, gitignored. Personal notes that stay out of git. |
-
-Matching is **case-insensitive** on the basename, so `AGENTS.md`, `Agents.md`, and `agents.md` resolve to the same file on Windows and macOS. The git-tracked filename in this repo is `AGENTS.md` — keep that on case-sensitive filesystems (Linux, the WSL filesystem, or a CI runner) to match what the loader looks for.
-
-Both files use the same format. YAML frontmatter is optional; the markdown body is loaded as instructions for the agent. Zero reads the file once at session start, so changes take effect on the next `zero` launch — not mid-session.
-
-```markdown
-# Project conventions for <your project>
+## 1. Project Conventions
 
 - Build with `make`, not `go build` directly.
 - Tests live next to the source file (`foo_test.go` next to `foo.go`).
 - Run `make lint` before opening a PR.
 - Never edit files under `third_party/` — those are vendored.
-```
+- Unify functions/methods where possible to prevent codebase inflation. Prefer a single cross-platform function with conditional checks over duplicating helpers per platform.
 
-Tips:
+## 2. Guidelines for Coding Agents
 
-- Keep each file under ~8 KiB. Zero caps the **total** across all matched files at 32 KiB; everything past the cap is dropped.
-- Re-state rules in the imperative voice: "Run `make lint`", not "you should consider running the linter".
-- Don't put secrets, model IDs, or environment-specific paths in `AGENTS.md`. Use `config.json` for those.
-- In a monorepo, drop a narrower `AGENTS.md` in each sub-tree (e.g. `services/api/AGENTS.md`). Zero picks those up automatically when you launch from inside the sub-tree.
-- A YAML frontmatter block (`---\n...\n---`) at the top is preserved verbatim in the injected prompt but is not parsed for `globs:` or `alwaysApply:` scoping today — keep the body self-contained.
+If you are an AI coding agent executing tasks in this repository, you **MUST** run all Go code quality and security checks before committing code or completing your task:
 
-### Personal guidelines, across every project
+1. **Format & Vet**: Run `go fmt ./...` and `go vet ./...`.
+2. **Lint**: Run `go run github.com/golangci/golangci-lint/v2/cmd/[email protected] run --enable-only unused,ineffassign,staticcheck ./...`.
+3. **Security**: Run `go run golang.org/x/vuln/cmd/[email protected] ./...`.
 
-For preferences that follow *you*, not a specific repo (tone, tooling habits, workflow), drop a `ZERO.md` in your user config directory: `~/.config/zero/ZERO.md` on Linux/macOS, `%AppData%\Roaming\zero\ZERO.md` on Windows — the same directory as `config.json` and your personal specialists. Same format and 8 KiB cap as the project files above, and the same case-insensitive basename match.
-
-This file is injected as its own `## User guidelines` section, before the project's `AGENTS.md`/`ZERO.md`, and is labeled as personal preference in the prompt: project guidelines are the later, more specific instruction and take precedence over it when the two conflict.
-
-## 2. Custom specialists
-
-Specialists are Zero's sub-agents. Three scopes, in priority order:
-
-| Scope | Path | Shared? |
-| --- | --- | --- |
-| Built-in | compiled into Zero | yes — `worker`, `explorer`, `code-review` |
-| User | `~/.config/zero/specialists/*.md` | no — your machine only |
-| Project | `./.zero/specialists/*.md` | yes — the repo team |
-
-Project overrides user overrides built-in when names collide.
-
-A specialist is a markdown manifest with frontmatter and a system prompt:
-
-```markdown
----
-description: Reviews API changes for breaking-change risk and missing tests.
-tools: read-only,plan
----
-
-You review API changes. For every changed hunk in `internal/api/` or any file
-that ends in `_api.go`:
-
-1. Confirm the public signature is backward-compatible, or note the breaking
-   change explicitly with the migration path.
-2. Confirm a corresponding test exists in `internal/api/*_test.go` and that
-   the new behaviour is exercised.
-3. Flag any new exported symbol without a doc comment.
-
-Reply with one JSON object per finding: `{"file", "line", "severity", "message", "fix"}`.
-```
-
-CLI management (the prompt is passed inline via `--prompt`):
-
-```bash
-zero specialist list
-zero specialist show api-reviewer
-zero specialist create api-reviewer \
-    --project \
-    --description "Reviews API changes" \
-    --tools read-only,plan \
-    --prompt "$(cat api-reviewer.md)"
-zero specialist edit api-reviewer --project
-zero specialist delete api-reviewer --project
-zero specialist path                       # prints the resolved specialists directory
-```
-
-The full format spec (frontmatter fields, tool scopes, prompt conventions) is in [`docs/SPECIALISTS.md`](docs/SPECIALISTS.md).
-
-> **Roadmap.** An in-UI specialist manager (create / edit / delete / preview) is on the backlog. Today you use the `zero specialist` CLI subcommands above.
-
-## 3. Skills
-
-Skills are markdown instruction packs the agent can pull in on demand. Each skill is a directory containing a `SKILL.md`. Skills are **user-level only** in this version — there's no project-scoped skill directory yet, so anything you want shared with the team goes in `AGENTS.md` (section 1) or as a hook (section 4).
-
-Discovery root: `$ZERO_SKILLS_DIR` → `$XDG_DATA_HOME/zero/skills` → `~/.local/share/zero/skills/`. A missing directory is fine — Zero just reports "no skills".
-
-```
-~/.local/share/zero/skills/
-  run-benchmarks/
-    SKILL.md
-  write-changelog/
-    SKILL.md
-```
-
-`SKILL.md` format:
-
-```markdown
----
-description: Run the project's benchmark suite and summarize the deltas.
----
-
-# Run benchmarks
-
-1. `make bench` — captures the wall-clock and RSS before and after.
-2. `benchstat before.txt after.txt` — diffs the two.
-3. Report any regression > 5% with the function name and the previous value.
-```
-
-Only `name` and `description` are recognized in the frontmatter today. The `name` defaults to the directory name. Within a single skills root, duplicate names are resolved by lexicographic directory order. During an agent run, Zero loads the default skills directory before plugin skill roots; earlier roots win name collisions silently. Plugin-declared skills (section 6) are merged into the active agent run at plugin activation time, so bundled skills appear in the available skills list and can be loaded with the `skill` tool.
-
-The `skill` core tool lets the agent load any discovered skill by name.
-
-## 4. Hooks
-
-Hooks fire shell commands on lifecycle events. Configure them in JSON:
-
-- User: `~/.config/zero/hooks.json`
-- Project: `./.zero/hooks.json`
-
-```json
-{
-  "enabled": true,
-  "hooks": [
-    {
-      "id": "block-rm-rf",
-      "event": "beforeTool",
-      "matcher": "bash",
-      "command": "/usr/local/bin/zero-hook-block-rmrf.sh",
-      "enabled": true
-    },
-    {
-      "id": "log-session",
-      "event": "sessionStart",
-      "command": "/usr/local/bin/zero-hook-log.sh",
-      "enabled": true
-    }
-  ]
-}
-```
-
-The `args` array (when present) is passed verbatim to `exec.CommandContext`. The actual hook payload — event name, matcher, tool call id, tool name, tool input, tool output, status — is delivered to the command as **JSON on stdin**, not via `${...}` substitution. A typical handler reads stdin and decides what to do:
-
-```bash
-#!/usr/bin/env bash
-# /usr/local/bin/zero-hook-block-rmrf.sh
-set -euo pipefail
-payload="$(cat)"
-if printf '%s' "$payload" | grep -q '"input":"[^"]*rm[[:space:]]+-rf'; then
-  echo "refusing rm -rf" >&2
-  exit 1
-fi
-```
-
-Events the agent emits (in dispatch order):
-
-| Event | Fires when | Matcher allowed? |
-| --- | --- | --- |
-| `beforeTool` | A tool is about to run | yes (tool name) |
-| `afterTool` | A tool just returned | yes (tool name) |
-| `sessionStart` | A session begins | no |
-| `sessionEnd` | A session ends | no |
-| `specialistStart` | A sub-agent is spawned | yes (specialist name) |
-| `specialistStop` | A sub-agent ends | yes (specialist name) |
-
-A hook's exit code decides what happens next: `0` continues, non-zero blocks the tool call (`beforeTool`) or surfaces an error (`afterTool`). Hook execution is recorded in the audit log; the audit is reachable from the agent's view of past actions, not from a dedicated `zero doctor` check.
-
-> **Roadmap.** An in-UI hooks manager is on the backlog. Today you edit the JSON directly.
-
-## 5. MCP — Model Context Protocol
-
-Zero is both an **MCP client** (it can call external MCP servers) and an **MCP server** (other agents can call its tools).
-
-### As a client — configure MCP servers in `config.json`
-
-```json
-{
-  "mcp": {
-    "servers": {
-      "docs": {
-        "type": "stdio",
-        "command": "docs-mcp",
-        "args": ["--port", "7777"]
-      },
-      "github": {
-        "type": "http",
-        "url": "https://api.example.com/mcp",
-        "headers": { "Authorization": "Bearer YOUR_TOKEN_HERE" }
-      }
-    }
-  }
-}
-```
-
-Manage via CLI:
-
-```bash
-zero mcp add docs --type stdio -- docs-mcp --port 7777
-zero mcp add github --type http --url https://api.example.com/mcp \
-    --header "Authorization=Bearer YOUR_TOKEN_HERE"
-zero mcp list
-zero mcp check docs
-zero mcp remove github
-zero mcp oauth login github
-```
-
-Servers are merged from user and project configs (project wins on conflicts). Token-bearing values in `config.json` are sent verbatim — there is no `${env:...}` expansion — so prefer one of:
-
-- A wrapper script that sources the secret and execs the real command.
-- A `--header` value produced by command substitution (`"Authorization=Bearer $(print-token)"`) in a private shell config that you keep out of git.
-- A secret manager that injects the env var your MCP server reads on its own (the `command` and `args` then run inside that environment).
-
-### As a server — expose Zero's tools to another agent
-
-```bash
-zero serve --mcp
-```
-
-The server speaks MCP over stdio. Configure it from the receiving side as a `stdio` server whose command is `zero serve --mcp`.
-
-## 6. Plugins
-
-A plugin is a self-contained directory that bundles tools, hooks, and skills for one capability. Plugins live at:
-
-- User: `~/.config/zero/plugins/<id>/`
-- Project: `./.zero/plugins/<id>/`
-
-Each plugin has a `plugin.json` manifest:
-
-```json
-{
-  "id": "github-pr-review",
-  "name": "GitHub PR Review",
-  "description": "Adds review tools for GitHub PRs.",
-  "version": "1.0.0",
-  "tools": [
-    { "name": "list_prs", "command": "./tools/list_prs.sh" }
-  ],
-  "hooks": [
-    { "name": "pre-merge-check", "event": "beforeTool", "command": "./hooks/pre-merge.sh" }
-  ],
-  "skills": [
-    { "path": "./skills/review-checklist/SKILL.md" }
-  ]
-}
-```
-
-Install and manage:
-
-```bash
-zero plugins add ./github-pr-review      # copy into ~/.config/zero/plugins/ or ./.zero/plugins/
-zero plugins list
-zero plugins remove github-pr-review    # alias: rm
-```
-
-A plugin is enabled by being present in the plugins directory and disabled by removing it (or by the user setting `"enabled": false` in its `plugin.json`). Plugins are not enabled or disabled by a CLI subcommand today.
-
-Plugin commands run with the plugin directory as their working directory. Use relative paths; the loader resolves them at activation time.
-
-> **Roadmap.** An in-UI plugins manager (browse, install, enable / disable) is on the backlog. Today you use the `zero plugins` CLI subcommands above.
-
-## 7. Configuration locations
-
-Three layers, applied in order (later layers override earlier ones):
-
-| Layer | Path | Notes |
-| --- | --- | --- |
-| Built-in defaults | compiled in | Lowest priority. |
-| User config | `~/.config/zero/config.json` | Your machine. Never committed. |
-| Project config | `./.zero/config.json` | The repo. Committed (or not, your call). |
-| CLI flags | `--model`, `--mode`, ... | Highest priority, per-invocation. |
-| Environment | `ZERO_*` | Provider commands, secrets, skills dir override. |
-
-The user config holds things that should follow the user across projects (default provider, default model, theme). The project config holds things the team agreed on (provider catalog, sandbox policies, model restrictions).
-
-The sandbox `additionalWriteRoots` key is **ignored in project config** by design — a checked-out repo cannot widen its own sandbox. Set it in the user config or pass `--add-dir` per-invocation.
-
-## 8. End-to-end example
-
-A team that wants every contributor's Zero to behave the same way commits:
-
-- `AGENTS.md` — project conventions, build commands, do-not-edit lists.
-- `.zero/config.json` — provider catalog, default model, allowed tools.
-- `.zero/specialists/api-reviewer.md` — the team's PR-review specialist.
-- `.zero/hooks.json` — block `rm -rf` and `git push --force` on `beforeTool`.
-- `.zero/plugins/internal-tooling/` — a plugin that adds the team's internal CLI tools to the agent's toolset.
-
-Each contributor adds only:
-
-- `~/.config/zero/config.json` — their personal API keys, theme, default mode.
-- `~/.config/zero/ZERO.md` — personal preferences that follow them across every project (see section 1).
-- `~/.local/share/zero/skills/` — personal skills they keep across projects.
-
-That's it. Run `zero` from the repo root and the agent has the team's full instruction set, every contributor's personal setup, and nothing else.
-
-## 9. Reference
-
-- [README](README.md) — install, quickstart, command reference.
-- [docs/SPECIALISTS.md](docs/SPECIALISTS.md) — full specialist manifest spec.
-- [docs/STREAM_JSON_PROTOCOL.md](docs/STREAM_JSON_PROTOCOL.md) — `zero exec` I/O contract.
-- [docs/INSTALL.md](docs/INSTALL.md) — install from source or release.
+If any check fails or cannot be run, do not ignore it. Prompt the user for instructions or setup assistance before proceeding.

CHANGELOG.md

@@ -127,6 +127,9 @@ tagged. Until then, source builds report the version `dev`.
 ## [Unreleased]
 
 ### Added
+- Shared multi-agent skills discovery: when present, `~/.agents/skills` is searched after the primary
+  Zero skills dir (and before plugin skill roots). `zero skills list` / `info` and the runtime `skill`
+  tool share one multi-root discovery path; install/remove/lock still target only the Zero skills directory.
 - `SECURITY.md` with a private vulnerability-reporting path, `CODE_OF_CONDUCT.md`, this changelog, and
   GitHub issue/PR templates.
 - Interactive `/theme` picker: bare `/theme` opens a popup that live-previews each palette as you move

Makefile

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

README.md

@@ -6,7 +6,7 @@
 
 <p align="center">
   <a href="LICENSE"><img alt="license" src="https://img.shields.io/badge/license-MIT-blue"></a>
-  <img alt="Go 1.25+" src="https://img.shields.io/badge/Go-1.25+-00ADD8?logo=go&logoColor=white">
+  <img alt="Go 1.26.5+" src="https://img.shields.io/badge/Go-1.26.5+-00ADD8?logo=go&logoColor=white">
   <img alt="25+ providers" src="https://img.shields.io/badge/providers-25+-34E2EA">
   <br>
   <strong>English</strong> | <a href="README_ZH.md">中文</a>
@@ -49,34 +49,15 @@ npm install -g @gitlawb/zero
 zero
 ```
 
-The npm package installs a small wrapper plus the matching Zero binary for your
-platform from GitHub Releases. It supports Linux, macOS, and Windows on x64 and
-arm64.
-
-### Bun
-
-Bun does not run dependency lifecycle scripts by default, so the `postinstall`
-that fetches the Zero binary is skipped and the first run fails with
-`No native binary found next to the npm wrapper`.
-
-The simplest fix is to trust the package after installing, which runs the
-blocked postinstall. This works for project and global installs:
-
-```bash
-# project install
-bun add @gitlawb/zero
-bun pm trust @gitlawb/zero
-
-# global install
-bun add -g @gitlawb/zero
-bun pm -g trust @gitlawb/zero
-```
-
-Alternatives: allow the postinstall up front by adding
-`"trustedDependencies": ["@gitlawb/zero"]` to your project's package.json
-before `bun add`, or run the installer manually
-(`node node_modules/@gitlawb/zero/scripts/postinstall.mjs`) on Bun versions
-that do not have `bun pm trust`.
+The npm package is a small wrapper whose platform build (Linux and macOS on
+x64/arm64, Windows on x64 — including the browser/terminal control helpers)
+installs as an optional dependency straight from the npm registry — no install
+scripts, no downloads outside npm. Bun, pnpm, and yarn work the same way with
+no trust or approval steps. Installs that skip optional dependencies
+(`--omit=optional`) still work: the wrapper fetches the binary from the
+matching GitHub Release whenever it is missing. Windows on ARM runs the x64
+build under emulation. See [docs/NPM_PACKAGING.md](docs/NPM_PACKAGING.md) for
+how the package is put together.
 
 ### Install scripts
 
@@ -94,7 +75,7 @@ irm https://raw.githubusercontent.com/Gitlawb/zero/main/scripts/install.ps1 | ie
 
 ### From source
 
-Source builds require Go 1.25+.
+Source builds require Go 1.26.5+.
 
 ```bash
 git clone https://github.com/Gitlawb/zero.git
@@ -148,7 +129,16 @@ the key in the wizard:
 export OPENAI_API_KEY=sk-...
 export ANTHROPIC_API_KEY=...
 export GEMINI_API_KEY=...
+export AIMLAPI_API_KEY=...
 export LONGCAT_API_KEY=...
+export MINIMAX_API_KEY=...
+export MINIMAXI_API_KEY=...
+```
+
+To configure AI/ML API directly, run:
+
+```bash
+zero providers setup aimlapi --set-active
 ```
 
 To configure Meituan LongCat (LongCat-2.0) directly, run:
@@ -157,6 +147,33 @@ To configure Meituan LongCat (LongCat-2.0) directly, run:
 zero providers setup longcat --set-active
 ```
 
+MiniMax presets use the Anthropic-compatible endpoints for the global and China
+regions:
+
+```bash
+zero providers add minimax --set-active
+zero providers add minimaxi-cn --set-active
+```
+
+To use the OpenAI-compatible endpoints instead, add a custom compatible profile
+for the required region:
+
+```bash
+zero providers add custom-openai-compatible \
+  --name minimax-openai \
+  --model MiniMax-M3 \
+  --base-url https://api.minimax.io/v1 \
+  --api-key-env MINIMAX_API_KEY \
+  --set-active
+
+zero providers add custom-openai-compatible \
+  --name minimax-cn-openai \
+  --model MiniMax-M3 \
+  --base-url https://api.minimaxi.com/v1 \
+  --api-key-env MINIMAXI_API_KEY \
+  --set-active
+```
+
 For local models, run Ollama or LM Studio and then use `zero setup` or
 `zero providers detect`.
 
@@ -339,7 +356,26 @@ go run ./cmd/zero-release smoke
 go run ./cmd/zero-perf-bench
 ```
 
-Cross-compile examples:
+### Code Quality and Security Checks
+
+Before committing any changes, ensure all Go code quality and security checks pass. Pinned `go run` commands matching CI constraints can be used directly without prior installation:
+
+1. **Formatting**: Run `go fmt ./...` (or `make fmt`).
+2. **Vetting**: Run `go vet ./...` (or `make vet`).
+3. **Linting**: Run `go run github.com/golangci/golangci-lint/v2/cmd/[email protected] run --enable-only unused,ineffassign,staticcheck ./...`.
+4. **Vulnerability Scan**: Run `go run golang.org/x/vuln/cmd/[email protected] ./...`.
+
+If you prefer to install these tools globally on your path, you can run:
+
+```bash
+# Install golangci-lint
+go install github.com/golangci/golangci-lint/v2/cmd/[email protected]
+
+# Install govulncheck
+go install golang.org/x/vuln/cmd/[email protected]
+```
+
+### Cross-Compile Examples
 
 ```bash
 go run ./cmd/zero-release build --goos linux --goarch amd64

README_ZH.md

@@ -6,7 +6,7 @@
 
 <p align="center">
   <a href="LICENSE"><img alt="license" src="https://img.shields.io/badge/license-MIT-blue"></a>
-  <img alt="Go 1.25+" src="https://img.shields.io/badge/Go-1.25+-00ADD8?logo=go&logoColor=white">
+  <img alt="Go 1.26.5+" src="https://img.shields.io/badge/Go-1.26.5+-00ADD8?logo=go&logoColor=white">
   <img alt="25+ providers" src="https://img.shields.io/badge/providers-25+-34E2EA">
   <br>
   <a href="README.md">English</a> | <strong>中文</strong>
@@ -38,25 +38,7 @@ npm install -g @gitlawb/zero
 zero
 ```
 
-npm 包安装一个小型包装器以及与你平台匹配的 Zero 二进制文件（从 GitHub Releases 获取）。支持 Linux、macOS 和 Windows 的 x64 和 arm64。
-
-### Bun
-
-Bun 默认不运行依赖的生命周期脚本，因此获取 Zero 二进制文件的 `postinstall` 会被跳过，首次运行会失败并显示 `No native binary found next to the npm wrapper`。
-
-最简单的解决方法是安装后信任该包，这会运行被阻止的 postinstall。项目安装和全局安装均适用：
-
-```bash
-# 项目安装
-bun add @gitlawb/zero
-bun pm trust @gitlawb/zero
-
-# 全局安装
-bun add -g @gitlawb/zero
-bun pm -g trust @gitlawb/zero
-```
-
-其他方式：在 `bun add` 之前将 `"trustedDependencies": ["@gitlawb/zero"]` 添加到项目的 package.json 以提前允许 postinstall；或在不支持 `bun pm trust` 的旧版 Bun 上手动运行安装程序（`node node_modules/@gitlawb/zero/scripts/postinstall.mjs`）。
+npm 包是一个小型包装器，其平台构建（Linux 和 macOS 的 x64/arm64、Windows 的 x64，包含浏览器/终端控制辅助工具）作为可选依赖直接从 npm registry 安装——没有安装脚本，也不会从 npm 之外下载任何内容。Bun、pnpm 和 yarn 的行为完全一致，无需任何信任或批准步骤。跳过可选依赖的安装（`--omit=optional`）也能工作：只要二进制文件缺失，包装器就会从对应的 GitHub Release 下载。Windows on ARM 通过模拟运行 x64 构建。详见 [docs/NPM_PACKAGING.md](docs/NPM_PACKAGING.md)。
 
 ### 安装脚本
 
@@ -74,7 +56,7 @@ irm https://raw.githubusercontent.com/Gitlawb/zero/main/scripts/install.ps1 | ie
 
 ### 从源码构建
 
-源码构建需要 Go 1.25+。
+源码构建需要 Go 1.26.5+。
 
 ```bash
 git clone https://github.com/Gitlawb/zero.git

bin/zero.js

@@ -2,6 +2,7 @@
 
 import { spawnSync } from 'node:child_process';
 import { existsSync } from 'node:fs';
+import { createRequire } from 'node:module';
 import { dirname, join } from 'node:path';
 import { fileURLToPath } from 'node:url';
 
@@ -39,6 +40,11 @@ function resolveHelper(packageRoot, name) {
   return null;
 }
 
+// Legacy path: helpers installed as npm dependencies of the wrapper. Current
+// installs vendor the helpers inside the platform package instead (a helpers/
+// directory next to the binary, found by the Go side on its own), so this
+// manifest is usually empty — it only matters for a downloaded-binary install
+// whose wrapper still carries helper dependencies.
 function localControlHelperManifest(packageRoot) {
   const helpers = {};
   for (const name of ['agent-browser', 'tuistory']) {
@@ -49,38 +55,85 @@ function localControlHelperManifest(packageRoot) {
   return JSON.stringify({ version: 1, helpers });
 }
 
-const packageRoot = dirname(dirname(fileURLToPath(import.meta.url)));
-const nativePath = join(packageRoot, zeroBinaryName());
-const localControlHelpers = localControlHelperManifest(packageRoot);
+// The platform payload is a version of @gitlawb/zero itself, installed under
+// an npm: alias (see docs/NPM_PACKAGING.md). The alias name is derived from
+// process.platform/process.arch, so an unsupported platform simply fails to
+// resolve and we fall through to the downloader.
+function platformPackageBinary() {
+  const aliasedName = `@gitlawb/zero-${process.platform}-${process.arch}`;
+  let manifestPath;
+  try {
+    manifestPath = createRequire(import.meta.url).resolve(`${aliasedName}/package.json`);
+  } catch {
+    return null;
+  }
+  const candidate = join(dirname(manifestPath), zeroBinaryName());
+  return existsSync(candidate) ? candidate : null;
+}
 
-if (!existsSync(nativePath)) {
-  const postinstallScript = join(packageRoot, 'scripts', 'postinstall.mjs');
-  const ranByBun = process.execPath.includes('bun') || !!process.versions?.bun;
+// Mirrors the release matrix (internal/release/release.go): the platforms a
+// GitHub Release asset exists for. android maps to the linux asset; there is
+// no windows-arm64 asset (Windows on ARM runs the x64 build under emulation).
+function releaseAssetAvailable() {
+  const platform = { linux: 'linux', android: 'linux', darwin: 'macos', win32: 'windows' }[
+    process.platform
+  ];
+  const arch = process.arch === 'x64' || process.arch === 'arm64' ? process.arch : null;
+  if (!platform || !arch) return false;
+  return !(platform === 'windows' && arch === 'arm64');
+}
+
+// Fallback for installs without a platform package (--omit=optional, package
+// managers that skip optionalDependencies): fetch the binary next to the
+// wrapper with the same checksum-verified downloader that used to run as
+// postinstall. Deliberately NOT cached on failure — a transient network error
+// self-heals on the next run — so this retries on every invocation until a
+// binary is in place. Platforms with no release asset skip the attempt.
+function downloadMissingBinary(packageRoot) {
+  const downloadScript = join(packageRoot, 'scripts', 'postinstall.mjs');
+  if (!existsSync(downloadScript) || !releaseAssetAvailable()) return;
   console.error(
-    '[zero] No native binary found next to the npm wrapper.\n' +
-      'The platform binary is fetched at install time by a postinstall script,\n' +
-      'which did not run (or was skipped) for this install.\n' +
+    '[zero] no platform package installed — fetching the native binary from the GitHub Release (retried on each run until it succeeds).',
+  );
+  spawnSync(process.execPath, [downloadScript], {
+    stdio: ['ignore', 'inherit', 'inherit'],
+  });
+}
+
+const packageRoot = dirname(dirname(fileURLToPath(import.meta.url)));
+
+function resolveNativeBinary() {
+  const fromPlatformPackage = platformPackageBinary();
+  if (fromPlatformPackage) return fromPlatformPackage;
+  const downloaded = join(packageRoot, zeroBinaryName());
+  if (existsSync(downloaded)) return downloaded;
+  downloadMissingBinary(packageRoot);
+  return existsSync(downloaded) ? downloaded : null;
+}
+
+const nativePath = resolveNativeBinary();
+
+if (!nativePath) {
+  console.error(
+    '[zero] No native binary is available for this install.\n' +
+      'Normally npm installs it as an optional dependency of @gitlawb/zero\n' +
+      `(@gitlawb/zero-${process.platform}-${process.arch}), and the wrapper can\n` +
+      'also download it from the GitHub Release when it is missing.\n' +
       '\n' +
-      'Fix it now by running the installer manually:\n' +
-      `  node "${postinstallScript}"\n` +
+      'Things to try:\n' +
+      '  - reinstall without omitting optional dependencies:\n' +
+      '      npm install -g @gitlawb/zero\n' +
+      '  - run the downloader manually (needs write access to the package dir):\n' +
+      `      node "${join(packageRoot, 'scripts', 'postinstall.mjs')}"\n` +
       '\n' +
-      (ranByBun
-        ? 'You installed with Bun, which does not run dependency lifecycle scripts\n' +
-          'by default. Trust the package to run the blocked postinstall:\n' +
-          '  bun pm trust @gitlawb/zero       (project install)\n' +
-          '  bun pm -g trust @gitlawb/zero    (global install)\n' +
-          'On Bun versions without `bun pm trust`, add\n' +
-          '  "trustedDependencies": ["@gitlawb/zero"]\n' +
-          'to your project package.json and reinstall.\n' +
-          '\n'
-        : '') +
-      'If that fails, build from source: https://github.com/Gitlawb/zero\n' +
-      '(go run ./cmd/zero, requires Go 1.25+).',
+      'If this platform has no prebuilt binary, build from source:\n' +
+      'https://github.com/Gitlawb/zero (go run ./cmd/zero, requires Go 1.26+).',
   );
   process.exit(1);
 }
 
 const env = { ...process.env };
+const localControlHelpers = localControlHelperManifest(packageRoot);
 if (localControlHelpers) {
   env.ZERO_LOCAL_CONTROL_HELPERS = localControlHelpers;
 } else {

cmd/zero-perf-bench/main.go

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

cmd/zero-perf-bench/turn.go

@@ -0,0 +1,238 @@
+package main
+
+import (
+	"bytes"
+	"context"
+	"fmt"
+	"io"
+	"os"
+	"path/filepath"
+	"strings"
+
+	"github.com/Gitlawb/zero/internal/perfbench"
+)
+
+// turnOptions configures the `zero-perf-bench turn` subcommand: the per-turn
+// benchmark harness that runs ZERO headlessly with --trace, parses each turn's
+// NDJSON trace, and records per-span latency plus the top controllable latency
+// sources — the Phase 0 baseline's "do not proceed until" criterion.
+type turnOptions struct {
+	SuitePath   string
+	Model       string
+	Mode        string
+	SelfCorrect bool
+	Binary      string
+	Iterations  int
+	Version     string
+	Commit      string
+	Output      string
+	JSON        bool
+	DryRun      bool
+	Help        bool
+}
+
+func runTurnCommand(args []string, getenv func(string) string, stdout io.Writer, stderr io.Writer) int {
+	options, err := parseTurnArgs(args, getenv)
+	if err != nil {
+		_, _ = fmt.Fprintln(stderr, err.Error())
+		return 2
+	}
+	if options.Help {
+		_, _ = fmt.Fprint(stdout, turnHelpText())
+		return 0
+	}
+
+	set, err := perfbench.LoadTaskSet(options.SuitePath)
+	if err != nil {
+		_, _ = fmt.Fprintln(stderr, "[zero] Turn benchmark failed: "+err.Error())
+		return 1
+	}
+
+	// The dry-run path records a zero-iteration run without a binary, so the
+	// manifest loads and the report path is exercised in CI without a model.
+	if options.DryRun {
+		_, _ = fmt.Fprintln(stdout, "[zero] turn benchmark: dry run (no agent invoked)")
+		return 0
+	}
+
+	binary, err := perfbench.ResolveBinary(options.Binary)
+	if err != nil {
+		_, _ = fmt.Fprintln(stderr, "[zero] Turn benchmark failed: "+err.Error())
+		return 2
+	}
+
+	result, err := perfbench.RunTurnBench(context.Background(), set, perfbench.TurnBenchConfig{
+		Model:       options.Model,
+		Mode:        options.Mode,
+		SelfCorrect: options.SelfCorrect,
+		Version:     options.Version,
+		Commit:      options.Commit,
+		Iterations:  options.Iterations,
+		Runner:      perfbench.NewTurnExecRunner(binary),
+	})
+	if err != nil {
+		_, _ = fmt.Fprintln(stderr, "[zero] Turn benchmark failed: "+err.Error())
+		return 1
+	}
+
+	if options.Output != "" {
+		if err := writeTurnReport(options.Output, result); err != nil {
+			_, _ = fmt.Fprintln(stderr, "[zero] Turn benchmark failed: "+err.Error())
+			return 1
+		}
+	}
+	if options.JSON {
+		if err := perfbench.WriteTurnBenchJSON(stdout, result); err != nil {
+			_, _ = fmt.Fprintln(stderr, "[zero] Turn benchmark failed: "+err.Error())
+			return 1
+		}
+		return 0
+	}
+	_, _ = fmt.Fprintln(stdout, perfbench.FormatTurnBenchSummary(result))
+	return 0
+}
+
+func parseTurnArgs(args []string, getenv func(string) string) (turnOptions, error) {
+	options := turnOptions{
+		Iterations: 1,
+		Version:    strings.TrimSpace(getenv("ZERO_BENCH_VERSION")),
+		Commit:     strings.TrimSpace(getenv("ZERO_BENCH_COMMIT")),
+	}
+	for index := 0; index < len(args); index++ {
+		arg := args[index]
+		flag, inlineValue := splitFlagValue(arg)
+		switch flag {
+		case "--suite":
+			value, next, err := readOptionValue(args, inlineValue, index, flag)
+			if err != nil {
+				return options, err
+			}
+			options.SuitePath = value
+			index = next
+		case "--model":
+			value, next, err := readOptionValue(args, inlineValue, index, flag)
+			if err != nil {
+				return options, err
+			}
+			options.Model = value
+			index = next
+		case "--mode":
+			value, next, err := readOptionValue(args, inlineValue, index, flag)
+			if err != nil {
+				return options, err
+			}
+			options.Mode = value
+			index = next
+		case "--binary":
+			value, next, err := readOptionValue(args, inlineValue, index, flag)
+			if err != nil {
+				return options, err
+			}
+			options.Binary = value
+			index = next
+		case "--iterations":
+			value, next, err := readOptionValue(args, inlineValue, index, flag)
+			if err != nil {
+				return options, err
+			}
+			parsed, err := parsePositiveInteger(flag, value)
+			if err != nil {
+				return options, err
+			}
+			options.Iterations = parsed
+			index = next
+		case "--version":
+			value, next, err := readOptionValue(args, inlineValue, index, flag)
+			if err != nil {
+				return options, err
+			}
+			options.Version = value
+			index = next
+		case "--commit":
+			value, next, err := readOptionValue(args, inlineValue, index, flag)
+			if err != nil {
+				return options, err
+			}
+			options.Commit = value
+			index = next
+		case "--output":
+			value, next, err := readOptionValue(args, inlineValue, index, flag)
+			if err != nil {
+				return options, err
+			}
+			options.Output = value
+			index = next
+		case "--self-correct":
+			if strings.Contains(arg, "=") {
+				return options, fmt.Errorf("%s does not accept a value", flag)
+			}
+			options.SelfCorrect = true
+		case "--json":
+			if strings.Contains(arg, "=") {
+				return options, fmt.Errorf("%s does not accept a value", flag)
+			}
+			options.JSON = true
+		case "--dry-run":
+			if strings.Contains(arg, "=") {
+				return options, fmt.Errorf("%s does not accept a value", flag)
+			}
+			options.DryRun = true
+		case "-h", "--help":
+			if strings.Contains(arg, "=") {
+				return options, fmt.Errorf("%s does not accept a value", flag)
+			}
+			options.Help = true
+		default:
+			return options, fmt.Errorf("unknown option: %s", arg)
+		}
+	}
+	if options.Help {
+		return options, nil
+	}
+	if strings.TrimSpace(options.SuitePath) == "" {
+		return options, fmt.Errorf("--suite is required")
+	}
+	if strings.TrimSpace(options.Model) == "" && !options.DryRun {
+		return options, fmt.Errorf("--model is required (or pass --dry-run)")
+	}
+	return options, nil
+}
+
+func writeTurnReport(path string, result perfbench.TurnBenchResult) error {
+	if dir := filepath.Dir(path); dir != "" {
+		if err := os.MkdirAll(dir, 0o755); err != nil {
+			return err
+		}
+	}
+	var buffer bytes.Buffer
+	if err := perfbench.WriteTurnBenchJSON(&buffer, result); err != nil {
+		return err
+	}
+	return os.WriteFile(path, buffer.Bytes(), 0o644)
+}
+
+func turnHelpText() string {
+	return strings.Join([]string{
+		"Usage: zero-perf-bench turn [options]",
+		"",
+		"Runs ZERO headlessly with --trace against a per-turn benchmark task set and",
+		"records per-span latency plus the top controllable latency sources (the",
+		"Phase 0 baseline's \"do not proceed until\" criterion). Each task is a fresh",
+		"`zero exec` process, so iterations are cold-start samples; a warm path needs",
+		"an in-process runner (future work).",
+		"",
+		"Options:",
+		"  --suite <path>      Task set JSON file (required)",
+		"  --model <model>     Model to run (required unless --dry-run)",
+		"  --mode <name>       Exec mode preset to apply",
+		"  --self-correct      Enable the post-edit verify-and-correct loop",
+		"  --binary <path>     Path to the `zero` binary (default: zero on PATH / repo root)",
+		"  --iterations <n>    Times to run each task (default: 1)",
+		"  --version <v>       Record the ZERO version (default: $ZERO_BENCH_VERSION)",
+		"  --commit <sha>      Record the ZERO commit (default: $ZERO_BENCH_COMMIT)",
+		"  --output <path>     Write the JSON result to path",
+		"  --json              Print only the JSON result",
+		"  --dry-run           Load the manifest and exit without invoking the agent",
+		"  -h, --help          Show this help",
+	}, "\n") + "\n"
+}

docs/EXTENDING.md

@@ -0,0 +1,340 @@
+# Extending Zero
+
+Zero is an open-source terminal coding agent. Out of the box it does the obvious things — read, edit, run, search — but the design point of the project is that **every surface is configurable**. This document is the user-facing guide for that configuration.
+
+If you only want to *use* Zero, the [README](../README.md) is enough. This page is for the other three jobs:
+
+1. Tell the agent about *your* project (drop an `AGENTS.md` in your repo).
+2. Add new specialist sub-agents.
+3. Wire Zero into the rest of your toolchain (MCP, skills, hooks, plugins).
+
+## 1. Drop a project `AGENTS.md`
+
+When Zero starts in a directory, it looks for project-level instructions and injects them into the system prompt. The lookup walks from your current working directory **up to the nearest git root** and reads the first matching file at each level — general rules at the repo root, more specific rules in sub-trees. Files are labeled with their directory in the prompt (e.g. `## Project guidelines (services/api/AGENTS.md)`).
+
+Accepted file names, in priority order at each level:
+
+| Path | Notes |
+| --- | --- |
+| `./AGENTS.md` | The classic spot — committed to your repo, shared with the team. |
+| `./ZERO.md` | Brand-specific alias. Same format, lower priority. |
+| `./.zero/AGENTS.md` | Project-local, hidden, gitignored. Personal notes that stay out of git. |
+
+Matching is **case-insensitive** on the basename, so `AGENTS.md`, `Agents.md`, and `agents.md` resolve to the same file on Windows and macOS. The git-tracked filename in this repo is `AGENTS.md` — keep that on case-sensitive filesystems (Linux, the WSL filesystem, or a CI runner) to match what the loader looks for.
+
+Both files use the same format. YAML frontmatter is optional; the markdown body is loaded as instructions for the agent. Zero reads the file once at session start, so changes take effect on the next `zero` launch — not mid-session.
+
+```markdown
+# Project conventions for <your project>
+
+- Build with `make`, not `go build` directly.
+- Tests live next to the source file (`foo_test.go` next to `foo.go`).
+- Run `make lint` before opening a PR.
+- Never edit files under `third_party/` — those are vendored.
+```
+
+Tips:
+
+- Keep each file under ~8 KiB. Zero caps the **total** across all matched files at 32 KiB; everything past the cap is dropped.
+- Re-state rules in the imperative voice: "Run `make lint`", not "you should consider running the linter".
+- Don't put secrets, model IDs, or environment-specific paths in `AGENTS.md`. Use `config.json` for those.
+- In a monorepo, drop a narrower `AGENTS.md` in each sub-tree (e.g. `services/api/AGENTS.md`). Zero picks those up automatically when you launch from inside the sub-tree.
+- A YAML frontmatter block (`---\n...\n---`) at the top is preserved verbatim in the injected prompt but is not parsed for `globs:` or `alwaysApply:` scoping today — keep the body self-contained.
+
+### Personal guidelines, across every project
+
+For preferences that follow *you*, not a specific repo (tone, tooling habits, workflow), drop a `ZERO.md` in your user config directory: `~/.config/zero/ZERO.md` on Linux/macOS, `%AppData%\zero\ZERO.md` on Windows — the same directory as `config.json` and your personal specialists. Same format and 8 KiB cap as the project files above, and the same case-insensitive basename match.
+
+This file is injected as its own `## User guidelines` section, before the project's `AGENTS.md`/`ZERO.md`, and is labeled as personal preference in the prompt: project guidelines are the later, more specific instruction and take precedence over it when the two conflict.
+
+## 2. Custom specialists
+
+Specialists are Zero's sub-agents. Three scopes, in priority order:
+
+| Scope | Path | Shared? |
+| --- | --- | --- |
+| Built-in | compiled into Zero | yes — `worker`, `explorer`, `code-review` |
+| User | `~/.config/zero/specialists/*.md` | no — your machine only |
+| Project | `./.zero/specialists/*.md` | yes — the repo team |
+
+Project overrides user overrides built-in when names collide.
+
+A specialist is a markdown manifest with frontmatter and a system prompt:
+
+```markdown
+---
+description: Reviews API changes for breaking-change risk and missing tests.
+tools: read-only,plan
+---
+
+You review API changes. For every changed hunk in `internal/api/` or any file
+that ends in `_api.go`:
+
+1. Confirm the public signature is backward-compatible, or note the breaking
+   change explicitly with the migration path.
+2. Confirm a corresponding test exists in `internal/api/*_test.go` and that
+   the new behaviour is exercised.
+3. Flag any new exported symbol without a doc comment.
+
+Reply with one JSON object per finding: `{"file", "line", "severity", "message", "fix"}`.
+```
+
+CLI management (the prompt is passed inline via `--prompt`):
+
+```bash
+zero specialist list
+zero specialist show api-reviewer
+zero specialist create api-reviewer \
+    --project \
+    --description "Reviews API changes" \
+    --tools read-only,plan \
+    --prompt "$(cat api-reviewer.md)"
+zero specialist edit api-reviewer --project
+zero specialist delete api-reviewer --project
+zero specialist path                       # prints the resolved specialists directory
+```
+
+The full format spec (frontmatter fields, tool scopes, prompt conventions) is in [`docs/SPECIALISTS.md`](SPECIALISTS.md).
+
+> **Roadmap.** An in-UI specialist manager (create / edit / delete / preview) is on the backlog. Today you use the `zero specialist` CLI subcommands above.
+
+## 3. Skills
+
+Skills are markdown instruction packs the agent can pull in on demand. Each skill is a directory containing a `SKILL.md`. Standalone project skill directories are not supported in this version (shared project-wide skills must go in `AGENTS.md` or as a hook). However, project plugins (section 6) may bundle skills, which are merged into the active run.
+
+Discovery roots (earlier wins on name collisions):
+
+1. **Primary Zero dir** — `$ZERO_SKILLS_DIR` if set, else `$XDG_DATA_HOME/zero/skills`, else `~/.local/share/zero/skills/`
+2. **Shared multi-agent dir** — `~/.agents/skills/` when present (read-only discovery; never an install target)
+3. **Plugin skill roots** — skills bundled by active plugins (section 6)
+
+A missing directory is fine — Zero just omits it. Management commands (`zero skills add` / `remove` / `lock`) always write to the primary Zero dir only; `list` / `info` search primary + `~/.agents/skills`.
+
+```text
+~/.local/share/zero/skills/          # primary (install/remove/lock target)
+  run-benchmarks/
+    SKILL.md
+  write-changelog/
+    SKILL.md
+
+~/.agents/skills/                    # optional shared multi-agent root
+  shared-review/
+    SKILL.md
+```
+
+`SKILL.md` format:
+
+```markdown
+---
+description: Run the project's benchmark suite and summarize the deltas.
+---
+
+# Run benchmarks
+
+1. `make bench` — captures the wall-clock and RSS before and after.
+2. `benchstat before.txt after.txt` — diffs the two.
+3. Report any regression > 5% with the function name and the previous value.
+```
+
+Only `name` and `description` are recognized in the frontmatter today. The `name` defaults to the directory name. Within a single skills root, duplicate names are resolved by lexicographic directory order. Across roots, Zero loads the primary directory first, then `~/.agents/skills`, then plugin skill roots; earlier roots win name collisions. Plugin-declared skills (section 6) are merged into the active agent run at plugin activation time, so bundled skills appear in the available skills list and can be loaded with the `skill` tool.
+
+The `skill` tool lets the agent load any discovered skill by name (primary, agents, or plugin).
+
+## 4. Hooks
+
+Hooks fire shell commands on lifecycle events. Configure them in JSON:
+
+- User: `~/.config/zero/hooks.json`
+- Project: `./.zero/hooks.json`
+
+```json
+{
+  "enabled": true,
+  "hooks": [
+    {
+      "id": "block-rm-rf",
+      "event": "beforeTool",
+      "matcher": "bash",
+      "command": "/usr/local/bin/zero-hook-block-rmrf.sh",
+      "enabled": true
+    },
+    {
+      "id": "log-session",
+      "event": "sessionStart",
+      "command": "/usr/local/bin/zero-hook-log.sh",
+      "enabled": true
+    }
+  ]
+}
+```
+
+The `args` array (when present) is passed verbatim to `exec.CommandContext`. The actual hook payload — event name, matcher, tool call id, tool name, tool input, tool output, status — is delivered to the command as **JSON on stdin**, not via `${...}` substitution. A typical handler reads stdin and decides what to do:
+
+```bash
+#!/usr/bin/env bash
+# /usr/local/bin/zero-hook-block-rmrf.sh
+set -euo pipefail
+payload="$(cat)"
+if printf '%s' "$payload" | grep -q '"input":"[^"]*rm[[:space:]]+-rf'; then
+  echo "refusing rm -rf" >&2
+  exit 1
+fi
+```
+
+Events the agent emits (in dispatch order):
+
+| Event | Fires when | Matcher allowed? |
+| --- | --- | --- |
+| `beforeTool` | A tool is about to run | yes (tool name) |
+| `afterTool` | A tool just returned | yes (tool name) |
+| `sessionStart` | A session begins | no |
+| `sessionEnd` | A session ends | no |
+| `specialistStart` | A sub-agent is spawned | yes (specialist name) |
+| `specialistStop` | A sub-agent ends | yes (specialist name) |
+
+A hook's exit code decides what happens next: `0` continues, non-zero blocks the tool call (`beforeTool`) or surfaces an error (`afterTool`). Hook execution is recorded in the audit log; the audit is reachable from the agent's view of past actions, not from a dedicated `zero doctor` check.
+
+> **Roadmap.** An in-UI hooks manager is on the backlog. Today you edit the JSON directly.
+
+## 5. MCP — Model Context Protocol
+
+Zero is both an **MCP client** (it can call external MCP servers) and an **MCP server** (other agents can call its tools).
+
+### As a client — configure MCP servers in `config.json`
+
+```json
+{
+  "mcp": {
+    "servers": {
+      "docs": {
+        "type": "stdio",
+        "command": "docs-mcp",
+        "args": ["--port", "7777"]
+      },
+      "github": {
+        "type": "http",
+        "url": "https://api.example.com/mcp",
+        "headers": { "Authorization": "Bearer YOUR_TOKEN_HERE" }
+      }
+    }
+  }
+}
+```
+
+Manage via CLI:
+
+```bash
+zero mcp add docs --type stdio -- docs-mcp --port 7777
+zero mcp add github --type http --url https://api.example.com/mcp \
+    --header "Authorization=Bearer YOUR_TOKEN_HERE"
+zero mcp list
+zero mcp check docs
+zero mcp remove github
+zero mcp oauth login github
+```
+
+Servers are merged from user and project configs (project wins on conflicts). Token-bearing values in `config.json` are sent verbatim — there is no `${env:...}` expansion — so prefer one of:
+
+- A wrapper script that sources the secret and execs the real command.
+- A `--header` value produced by command substitution (`"Authorization=Bearer $(print-token)"`) in a private shell config that you keep out of git.
+- A secret manager that injects the env var your MCP server reads on its own (the `command` and `args` then run inside that environment).
+
+Project MCP config remains supported for shared team servers, but a project layer
+that changes an existing server's target (`url`, `command`, or `args`) does not
+inherit user credentials. Clear or replace `headers`, `env`, or `oauth` explicitly,
+or use a new server name. OAuth tokens are bound to the resolved server identity,
+so changing a server target can require `zero mcp oauth login <server>` again.
+OAuth server names ending in `.<32 hex chars>` are reserved for token storage.
+
+### As a server — expose Zero's tools to another agent
+
+```bash
+zero serve --mcp
+zero serve --mcp -C /path/to/workspace
+zero serve --mcp --add-dir /path/to/extra
+```
+
+The server speaks MCP over stdio. Configure it from the receiving side as a `stdio` server whose command is `zero serve --mcp`.
+
+`-C/--cwd` sets the workspace root exposed as MCP resources and used by core tools. `--add-dir` (repeatable) widens that resource/tool scope beyond the workspace without granting the sandbox temp roots used by interactive runs.
+## 6. Plugins
+
+A plugin is a self-contained directory that bundles tools, hooks, and skills for one capability. Plugins live at:
+
+- User: `~/.config/zero/plugins/<id>/`
+- Project: `./.zero/plugins/<id>/`
+
+Each plugin has a `plugin.json` manifest:
+
+```json
+{
+  "id": "github-pr-review",
+  "name": "GitHub PR Review",
+  "description": "Adds review tools for GitHub PRs.",
+  "version": "1.0.0",
+  "tools": [
+    { "name": "list_prs", "command": "./tools/list_prs.sh" }
+  ],
+  "hooks": [
+    { "name": "pre-merge-check", "event": "beforeTool", "command": "./hooks/pre-merge.sh" }
+  ],
+  "skills": [
+    { "path": "./skills/review-checklist/SKILL.md" }
+  ]
+}
+```
+
+Install and manage:
+
+```bash
+zero plugins add ./github-pr-review      # copy into ~/.config/zero/plugins/ or ./.zero/plugins/
+zero plugins list
+zero plugins remove github-pr-review    # alias: rm
+```
+
+A plugin is enabled by being present in the plugins directory and disabled by removing it (or by the user setting `"enabled": false` in its `plugin.json`). Plugins are not enabled or disabled by a CLI subcommand today.
+
+Plugin commands run with the plugin directory as their working directory. Use relative paths; the loader resolves them at activation time.
+
+> **Roadmap.** An in-UI plugins manager (browse, install, enable / disable) is on the backlog. Today you use the `zero plugins` CLI subcommands above.
+
+## 7. Configuration locations
+
+Five configuration sources, in precedence order (later sources override earlier ones):
+
+| Source | Path / Key | Notes |
+| --- | --- | --- |
+| Built-in defaults | compiled in | Lowest priority. |
+| User config | `~/.config/zero/config.json` | Your machine. Never committed. |
+| Project config | `./.zero/config.json` | The repo. Committed (or not, your call). |
+| Environment | `ZERO_*` | Provider commands, secrets, skills dir override. |
+| CLI flags | `--model`, `--mode`, ... | Highest priority, per-invocation. |
+
+The user config holds things that should follow the user across projects (default provider, default model, theme). The project config holds things the team agreed on (provider catalog, sandbox policies, model restrictions).
+
+The sandbox `additionalWriteRoots` key is **ignored in project config** by design — a checked-out repo cannot widen its own sandbox. Set it in the user config or pass `--add-dir` per-invocation.
+
+## 8. End-to-end example
+
+A team that wants every contributor's Zero to behave the same way commits:
+
+- `AGENTS.md` — project conventions, build commands, do-not-edit lists.
+- `.zero/config.json` — provider catalog, default model, allowed tools.
+- `.zero/specialists/api-reviewer.md` — the team's PR-review specialist.
+- `.zero/hooks.json` — block `rm -rf` on `beforeTool`.
+- `.zero/plugins/internal-tooling/` — a plugin that adds the team's internal CLI tools to the agent's toolset.
+
+Each contributor adds only:
+
+- `~/.config/zero/config.json` — their personal API keys, theme, default mode.
+- `~/.config/zero/ZERO.md` — personal preferences that follow them across every project (see section 1).
+- `~/.local/share/zero/skills/` — personal skills they keep across projects.
+
+That's it. Run `zero` from the repo root and the agent has the team's full instruction set, every contributor's personal setup, and nothing else.
+
+## 9. Reference
+
+- [README](../README.md) — install, quickstart, command reference.
+- [docs/SPECIALISTS.md](SPECIALISTS.md) — full specialist manifest spec.
+- [docs/STREAM_JSON_PROTOCOL.md](STREAM_JSON_PROTOCOL.md) — `zero exec` I/O contract.
+- [docs/INSTALL.md](INSTALL.md) — install from source or release.

docs/HOW_ZERO_WORKS.md

@@ -733,9 +733,10 @@ flowchart TD
 - **Skills** are instruction packs the model can load on demand.
 - **Specialists** are sub-agents callable through the `Task` tool.
 - **MCP servers** contribute external tools.
-- **Plugins** can add tools, hooks, and skill roots. Bootstrap merges plugin skill
-  roots into the runtime skill tool and prompt-visible skill list; the base
-  `internal/skills` scanner still scans one root at a time.
+- **Plugins** can add tools, hooks, and skill roots. Bootstrap always registers a
+  multi-root skill tool: primary Zero skills dir, optional `~/.agents/skills`,
+  then plugin skill roots (earlier wins). `internal/skills` owns that merge via
+  `LoadFromRoots` / `DiscoveryRoots`; single-root `Load` remains for install/write.
 - **Hooks** can observe or block tool lifecycle events.
 
 ## End-to-End Data Flow

docs/INSTALL.md

@@ -4,10 +4,11 @@ Zero is distributed as:
 
 - an npm package, `@gitlawb/zero`
 - release archives on GitHub Releases
-- source builds with Go 1.25+
+- source builds with Go 1.26.5+
 
-The npm package and install scripts download a platform-specific release archive.
-They require a published GitHub Release for the requested version.
+The install scripts download a platform-specific release archive and require a
+published GitHub Release for the requested version. The npm package is
+self-contained: the platform binary installs from the npm registry.
 
 ## npm
 
@@ -16,59 +17,32 @@ npm install -g @gitlawb/zero
 zero
 ```
 
-The package supports Linux, macOS, and Windows on x64 and arm64. It installs the
-`zero` command and downloads the matching release binary during `postinstall`.
+The package supports Linux and macOS on x64 and arm64, and Windows on x64
+(Windows on ARM runs the x64 build under emulation). It installs a small
+`zero` wrapper plus, as an optional dependency, a platform payload with the
+native binary and the bundled browser/terminal control helpers. There are no
+install scripts and nothing is downloaded from outside the npm registry, so
+the install is silent and works identically under npm, Bun, pnpm, and yarn —
+no trust or approval steps. See [NPM_PACKAGING.md](NPM_PACKAGING.md) for the
+package architecture.
 
 Requirements:
 
 - Node.js 18+
-- network access to npm and GitHub Releases
+- network access to npm
 
-## Bun
-
-Bun is "default-secure" and does not run lifecycle scripts of installed
-dependencies (only the installing project's own scripts), so the `postinstall`
-that fetches the Zero binary is silently skipped. The first run then fails with
-`No native binary found next to the npm wrapper`.
-
-The simplest fix is to trust the package after installing, which runs the
-blocked postinstall. This works for project and global installs:
+If the install skipped optional dependencies (`npm install --omit=optional`,
+or a package manager configured to do so), the wrapper falls back to
+downloading the binary from the matching GitHub Release, with checksum
+verification. The fetch is retried on each run until a binary is in place (a
+transient network failure heals itself; an install directory the current user
+cannot write to keeps failing — rerun the install or the command below with
+sufficient permissions). To trigger the fetch manually:
 
 ```bash
-# project install
-bun add @gitlawb/zero
-bun pm trust @gitlawb/zero
-
-# global install
-bun add -g @gitlawb/zero
-bun pm -g trust @gitlawb/zero
+node "$(npm root -g)/@gitlawb/zero/scripts/postinstall.mjs"
 ```
 
-`bun pm untrusted` (or `bun pm -g untrusted`) lists the blocked postinstalls if
-you want to inspect before trusting.
-
-Alternatively, allow the postinstall to run at install time by adding the
-package to your project's `trustedDependencies` before installing:
-
-```json
-{
-  "trustedDependencies": ["@gitlawb/zero"]
-}
-```
-
-```bash
-bun add @gitlawb/zero
-```
-
-On Bun versions that do not have `bun pm trust`, run the installer manually
-after installing:
-
-```bash
-node node_modules/@gitlawb/zero/scripts/postinstall.mjs
-```
-
-Reference: <https://bun.sh/docs/pm/lifecycle>
-
 ## Linux And macOS Script
 
 Install the latest release:
@@ -151,7 +125,7 @@ Build a local binary:
 go build -o zero ./cmd/zero
 ```
 
-Source builds require Go 1.25+.
+Source builds require Go 1.26.5+.
 
 ### Sandbox Helpers For Source Builds
 
@@ -252,8 +226,7 @@ Supported targets:
 - `linux-arm64`
 - `macos-x64`
 - `macos-arm64`
-- `windows-x64`
-- `windows-arm64`
+- `windows-x64` (Windows on ARM runs this build under emulation)
 
 Each archive must have a matching `.sha256` file. The install scripts download
 both files, verify the checksum, and then copy the binary into the install

docs/NPM_PACKAGING.md

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

docs/NPM_WRAPPER_SMOKE.md

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

go.mod

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

go.sum

@@ -1,9 +1,9 @@
-charm.land/bubbles/v2 v2.1.0 h1:YSnNh5cPYlYjPxRrzs5VEn3vwhtEn3jVGRBT3M7/I0g=
-charm.land/bubbles/v2 v2.1.0/go.mod h1:l97h4hym2hvWBVfmJDtrEHHCtkIKeTEb3TTJ4ZOB3wY=
-charm.land/bubbletea/v2 v2.0.7 h1:7qw2tTAVar7m7klOPBYfTB0mniv/RuexsYwMRNxSeL0=
-charm.land/bubbletea/v2 v2.0.7/go.mod h1:DGW2q8gvzHnOpMpZTORs0aySVHCox5C+2Svk0fci1qs=
-charm.land/lipgloss/v2 v2.0.4 h1:lcPeVtcp23SNra7lHy8iYE4UC2aIipVQ47sbGyyxR5Q=
-charm.land/lipgloss/v2 v2.0.4/go.mod h1:0653x8epbZSzdDfO/XPS1a/uYPOBeSsCssOpJOqDzik=
+charm.land/bubbles/v2 v2.1.1 h1:7r55WzBxpo/R3z98hGmY7KKPd3ET6vsf0Fb9sDHOV60=
+charm.land/bubbles/v2 v2.1.1/go.mod h1:GE6M31gaWZVXzGw73OeuTTgy4lX+OtkH0E5ymnNsHxo=
+charm.land/bubbletea/v2 v2.0.8 h1:SxTJMhCAI3lbPmy4SgX5LWZ24AdINr4I6UEqzZvYJuY=
+charm.land/bubbletea/v2 v2.0.8/go.mod h1:2SkdgoTXluXJHOUwAoRlRXF/28vklb1rFl6GcgV1/ss=
+charm.land/lipgloss/v2 v2.0.5 h1:kbNxgeeUOYv5J0YdpxFjfvf3dFvqH8Aci4zB6xqFtrY=
+charm.land/lipgloss/v2 v2.0.5/go.mod h1:9oqhxt4yxIMe6q5A4kHr44DremZk7J9UNh74GlWa5nc=
 github.com/alecthomas/assert/v2 v2.11.0 h1:2Q9r3ki8+JYXvGsDyBXwH3LcJ+WK5D0gc5E8vS6K3D0=
 github.com/alecthomas/assert/v2 v2.11.0/go.mod h1:Bze95FyfUr7x34QZrjL+XP+0qgp/zg8yS+TtBj1WA3k=
 github.com/alecthomas/chroma/v2 v2.27.0 h1:FodwmyOBgJULFYmDqibcp9pvfDLWdtPRh9v/r5BXYZs=
@@ -16,8 +16,8 @@ github.com/aymanbagabas/go-udiff v0.4.1 h1:OEIrQ8maEeDBXQDoGCbbTTXYJMYRCRO1fnodZ
 github.com/aymanbagabas/go-udiff v0.4.1/go.mod h1:0L9PGwj20lrtmEMeyw4WKJ/TMyDtvAoK9bf2u/mNo3w=
 github.com/charmbracelet/colorprofile v0.4.3 h1:QPa1IWkYI+AOB+fE+mg/5/4HRMZcaXex9t5KX76i20Q=
 github.com/charmbracelet/colorprofile v0.4.3/go.mod h1:/zT4BhpD5aGFpqQQqw7a+VtHCzu+zrQtt1zhMt9mR4Q=
-github.com/charmbracelet/ultraviolet v0.0.0-20260615092913-2399af76d5b1 h1:4+r3uOJ69ueRBt4okgEfWZeXs3BD36HcDBmOIAUlETk=
-github.com/charmbracelet/ultraviolet v0.0.0-20260615092913-2399af76d5b1/go.mod h1:f/jRa757WUmaOZrbPspXymbg/GnbF+rwe4OLsG7aXYo=
+github.com/charmbracelet/ultraviolet v0.0.0-20260703014108-f5a850f9c2b7 h1:3FmWoGNWK4STvqg0O0Aeav2T7rodWJAPeF0QpH+8gFw=
+github.com/charmbracelet/ultraviolet v0.0.0-20260703014108-f5a850f9c2b7/go.mod h1:f/jRa757WUmaOZrbPspXymbg/GnbF+rwe4OLsG7aXYo=
 github.com/charmbracelet/x/ansi v0.11.7 h1:kzv1kJvjg2S3r9KHo8hDdHFQLEqn4RBCb39dAYC84jI=
 github.com/charmbracelet/x/ansi v0.11.7/go.mod h1:9qGpnAVYz+8ACONkZBUWPtL7lulP9No6p1epAihUZwQ=
 github.com/charmbracelet/x/exp/golden v0.0.0-20260615092313-b57e5e6d29bb h1:m1Uub9uM7meZ8ofVyeRfK7M/knCi3sF2gKn9kldVW0I=
@@ -34,8 +34,8 @@ github.com/clipperhouse/uax29/v2 v2.7.0 h1:+gs4oBZ2gPfVrKPthwbMzWZDaAFPGYK72F0NJ
 github.com/clipperhouse/uax29/v2 v2.7.0/go.mod h1:EFJ2TJMRUaplDxHKj1qAEhCtQPW2tJSwu5BF98AuoVM=
 github.com/coder/websocket v1.8.15 h1:6B2JPeOGlpff2Uz6vOEH1Vzpi0iUz20A+lPVhPHtNUA=
 github.com/coder/websocket v1.8.15/go.mod h1:NX3SzP+inril6yawo5CQXx8+fk145lPDC6pumgx0mVg=
-github.com/dlclark/regexp2/v2 v2.2.1 h1:mf4KkFUj0gJuarK8P+LgiS+Lit7m9N1yAwEfPbee7R0=
-github.com/dlclark/regexp2/v2 v2.2.1/go.mod h1:avUrQvPaLz2DrFNHJF0taWAFFX2C1GMSSoeiqFjcBmU=
+github.com/dlclark/regexp2/v2 v2.5.0 h1:liNiWIPCvCh5HBcYfsjd+P16AG79fwd6T1Toy2gOtEA=
+github.com/dlclark/regexp2/v2 v2.5.0/go.mod h1:avUrQvPaLz2DrFNHJF0taWAFFX2C1GMSSoeiqFjcBmU=
 github.com/go-quicktest/qt v1.102.0 h1:HSQxCeh5YZH3EL3W39ixjtyaEhcWSXQHtHnMBzSs474=
 github.com/go-quicktest/qt v1.102.0/go.mod h1:p4lGIVX+8Wa6ZPNDvqcxq36XpUDLh42FLetFU7odllI=
 github.com/google/go-cmp v0.7.0 h1:wk8382ETsv4JYUZwIsn6YpYiWiBsYLSJiTsyBybVuN8=
@@ -62,9 +62,9 @@ github.com/xo/terminfo v0.0.0-20220910002029-abceb7e1c41e h1:JVG44RsyaB9T2KIHavM
 github.com/xo/terminfo v0.0.0-20220910002029-abceb7e1c41e/go.mod h1:RbqR21r5mrJuqunuUZ/Dhy/avygyECGrLceyNeo4LiM=
 golang.org/x/exp v0.0.0-20260611194520-c48552f49976 h1:X8Hz2ImujgbmetVuW+w2YkyZChE3cBpZi2P158rTG9M=
 golang.org/x/exp v0.0.0-20260611194520-c48552f49976/go.mod h1:vnf4pv9iKZXY58sQE1L86zmNWJ4159e1RkcWiLCkeEY=
-golang.org/x/sync v0.21.0 h1:HLII4xRRTtCRkxYp4HNFF0Js/Og6q2i++KXbg0gHCwM=
-golang.org/x/sync v0.21.0/go.mod h1:9xrNwdLfx4jkKbNva9FpL6vEN7evnE43NNNJQ2LF3+0=
-golang.org/x/sys v0.46.0 h1:noSf2Fq6F8DBgS+LysIkx7rIExoNHJsxOAtPp4rthXw=
-golang.org/x/sys v0.46.0/go.mod h1:4GL1E5IUh+htKOUEOaiffhrAeqysfVGipDYzABqnCmw=
+golang.org/x/sync v0.22.0 h1:SZjpbeLmrCk4xhRSZFNZW5gFUeCeFgjekvI/+gfScek=
+golang.org/x/sync v0.22.0/go.mod h1:9xrNwdLfx4jkKbNva9FpL6vEN7evnE43NNNJQ2LF3+0=
+golang.org/x/sys v0.47.0 h1:o7XGOvZQCADBQQ4Y7VNq2dRWQR7JmOUW8Kxx4ZsNgWs=
+golang.org/x/sys v0.47.0/go.mod h1:4GL1E5IUh+htKOUEOaiffhrAeqysfVGipDYzABqnCmw=
 mvdan.cc/sh/v3 v3.13.1 h1:DP3TfgZhDkT7lerUdnp6PTGKyxxzz6T+cOlY/xEvfWk=
 mvdan.cc/sh/v3 v3.13.1/go.mod h1:lXJ8SexMvEVcHCoDvAGLZgFJ9Wsm2sulmoNEXGhYZD0=

internal/acp/translate.go

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

internal/acp/translate_test.go

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

internal/agent/acceptance_gate_test.go

@@ -72,6 +72,26 @@ func TestSelfReportedIncompletionMatching(t *testing.T) {
 		{"I replaced the placeholder value with the computed key.", false},
 		{"The retry uses exponential backoff as a fallback.", false},
 		{"Without proper error handling the function would crash, so I added it.", false},
+		// real false positive (conversational recap): retelling a PAST exchange is
+		// not an admission about the current objective — the sentence carries its
+		// own narrative context ("You asked …", "at the time"):
+		{"You asked if I could work autonomously on zero if you gave me a GitHub account — I was honest that my sandbox had no repo, no build system, and no network/integration wired up yet, so I couldn't actually do it at the time.", false},
+		{"When we last spoke, I couldn't reach the GitHub API from the sandbox.", false},
+		{"Last session I was unable to build the project; that is what we should retry now.", false},
+		// quoted admissions (the model citing its own earlier message or a log
+		// line) must NOT be flagged:
+		{`My earlier reply said "I couldn't do it" but the tooling is now installed.`, false},
+		{"The run log contains `I cannot access the network`, which is expected inside the sandbox.", false},
+		// current-objective admissions must STILL be flagged, including past tense
+		// and multi-sentence finals whose other sentences are narrative:
+		{"I couldn't complete the refactor.", true},
+		{"I hit an error I could not resolve, so the migration is unfinished.", true},
+		{"You asked me to port the parser. I couldn't finish the second half.", true},
+		// unbalanced delimiters must not swallow the tail (only BALANCED quoted
+		// spans are stripped), so an admission after a stray quote still fires:
+		{"The log ends with a stray \" and I couldn't finish the migration.", true},
+		{"There is an unterminated `code span and I cannot proceed with the fix.", true},
+		{"The doc opens with “a curly quote and I was unable to complete the task.", true},
 	}
 	for _, c := range cases {
 		got := selfReportedIncompletion(c.text) != ""

internal/agent/compaction.go

@@ -6,6 +6,7 @@ import (
 	"errors"
 	"strings"
 
+	"github.com/Gitlawb/zero/internal/trace"
 	"github.com/Gitlawb/zero/internal/zeroruntime"
 )
 
@@ -430,6 +431,12 @@ func (state *compactionState) maybeCompact(
 		state.lowWaterMark = size
 		return messages
 	}
+	// Only count a compaction when it actually shrank the history, so the
+	// compaction counter reflects real context reductions rather than paid
+	// no-ops that left the token budget untouched.
+	if r := trace.FromContext(ctx); r != nil {
+		r.Counter(trace.CounterCompactionCount, 1)
+	}
 	state.lowWaterMark = newSize
 	return compacted
 }
@@ -479,7 +486,11 @@ func (state *compactionState) recover(
 	// one-shot budget now so a provider that keeps returning context-limit errors
 	// after a successful compaction can't loop forever. Store the low-water mark in
 	// the SAME combined (messages + tool-defs) domain maybeCompact uses, so the
-	// proactive shrink-guard compares like with like.
+	// proactive shrink-guard compares like with like. Count it now that the shrink
+	// is confirmed, so the counter mirrors maybeCompact's real-reduction policy.
+	if r := trace.FromContext(ctx); r != nil {
+		r.Counter(trace.CounterCompactionCount, 1)
+	}
 	state.reactiveAttempted = true
 	state.lowWaterMark = state.calibratedTokens(estimateTokens(result) + estimateToolDefTokens(tools))
 	return result, true, nil

internal/agent/guardrails.go

@@ -183,35 +183,112 @@ var successNegationTails = []string{
 	"reproduce", "spot any", "locate any",
 }
 
-// selfReportedIncompletion returns a short reason when the model's final text
-// admits it guessed or could not meet the objective, else "". Case-insensitive.
-func selfReportedIncompletion(text string) string {
-	lower := strings.ToLower(text)
-	for _, phrase := range selfReportPhrases {
-		if strings.Contains(lower, phrase) {
-			return selfReportReason(phrase)
+// narrativeMarkers flag a sentence as RETELLING a past exchange rather than
+// reporting the outcome of the current objective, so an inability admission in
+// that sentence is about THEN, not NOW. Grounded in a real false positive: a
+// conversational recap ("You asked if I could work autonomously … I was honest
+// that my sandbox had no repo … so I couldn't actually do it at the time") was
+// downgraded to INCOMPLETE on the quoted "i couldn't " stem even though the
+// current turn's objective (summarize where we left off) was fully met. Every
+// marker is second-person or explicitly past-referential, so a first-person
+// admission about the current task ("I couldn't complete the refactor") is
+// never masked.
+var narrativeMarkers = []string{
+	"you asked", "you said", "you wanted", "you mentioned",
+	"we discussed", "we talked", "we covered",
+	"when we last", "last time", "last session", "previous session",
+	"previous conversation", "earlier session", "earlier conversation",
+	"at the time", "back then",
+}
+
+// stripQuoted removes spans enclosed in double quotes (straight or curly) or
+// backticks, so an admission the model merely QUOTES — its own earlier message,
+// a log line, an error string — cannot fire the detector. Only BALANCED spans
+// are removed: an opening delimiter that never closes is kept as literal text,
+// so a stray quote cannot swallow the rest of the message (and with it a
+// genuine admission the detector must see). Single quotes are left alone: they
+// are overwhelmingly apostrophes ("couldn't"), and treating them as quote
+// delimiters would swallow the text between two contractions.
+func stripQuoted(s string) string {
+	var b strings.Builder
+	var span strings.Builder // pending text since the open delimiter, kept if it never closes
+	open := rune(0)
+	for _, r := range s {
+		switch {
+		case open != 0:
+			if (open == '"' && r == '"') || (open == '“' && r == '”') || (open == '`' && r == '`') {
+				open = 0
+				span.Reset()
+				continue
+			}
+			span.WriteRune(r)
+		case r == '"' || r == '“' || r == '`':
+			open = r
+			span.WriteRune(r)
+		default:
+			b.WriteRune(r)
 		}
 	}
-	for _, stem := range inabilityStems {
-		// Scan EVERY occurrence of the stem, not just the first: an earlier
-		// success-negation use ("I could not find any examples, so I could not
-		// implement it") must not mask a later genuine admission with the same stem.
-		for start := 0; ; {
-			rel := strings.Index(lower[start:], stem)
-			if rel < 0 {
-				break
+	b.WriteString(span.String()) // dangling delimiter: restore the span verbatim
+	return b.String()
+}
+
+// admissionSentences splits lowered text into sentence-ish fragments so the
+// detector judges each claim in its own context. Newlines split too (markdown
+// bullets are separate claims); the exact boundaries only need to keep an
+// admission next to its own narrative/negation context, not be grammatical.
+func admissionSentences(lower string) []string {
+	return strings.FieldsFunc(lower, func(r rune) bool {
+		return r == '.' || r == '!' || r == '?' || r == '\n'
+	})
+}
+
+// selfReportedIncompletion returns a short reason when the model's final text
+// admits it guessed or could not meet the objective, else "". Case-insensitive.
+// Matching is per sentence, after dropping quoted spans, and a sentence that
+// retells a past exchange (narrativeMarkers) is skipped entirely — an admission
+// must be the model's own report about the CURRENT objective, not general
+// language that merely resembles one.
+func selfReportedIncompletion(text string) string {
+	for _, sentence := range admissionSentences(strings.ToLower(stripQuoted(text))) {
+		if containsAny(sentence, narrativeMarkers) {
+			continue
+		}
+		for _, phrase := range selfReportPhrases {
+			if strings.Contains(sentence, phrase) {
+				return selfReportReason(phrase)
 			}
-			abs := start + rel
-			tail := strings.TrimSpace(lower[abs+len(stem):])
-			if !hasAnyPrefix(tail, successNegationTails) {
-				return selfReportReason(strings.TrimSpace(stem) + " …")
+		}
+		for _, stem := range inabilityStems {
+			// Scan EVERY occurrence of the stem, not just the first: an earlier
+			// success-negation use ("I could not find any examples, so I could not
+			// implement it") must not mask a later genuine admission with the same stem.
+			for start := 0; ; {
+				rel := strings.Index(sentence[start:], stem)
+				if rel < 0 {
+					break
+				}
+				abs := start + rel
+				tail := strings.TrimSpace(sentence[abs+len(stem):])
+				if !hasAnyPrefix(tail, successNegationTails) {
+					return selfReportReason(strings.TrimSpace(stem) + " …")
+				}
+				start = abs + len(stem)
 			}
-			start = abs + len(stem)
 		}
 	}
 	return ""
 }
 
+func containsAny(s string, subs []string) bool {
+	for _, sub := range subs {
+		if strings.Contains(s, sub) {
+			return true
+		}
+	}
+	return false
+}
+
 func selfReportReason(marker string) string {
 	return `the final message admits the objective was not met ("` + marker + `")`
 }

internal/agent/loop.go

@@ -15,6 +15,7 @@ import (
 	"github.com/Gitlawb/zero/internal/sandbox"
 	"github.com/Gitlawb/zero/internal/streamjson"
 	"github.com/Gitlawb/zero/internal/tools"
+	"github.com/Gitlawb/zero/internal/trace"
 	"github.com/Gitlawb/zero/internal/zeroruntime"
 )
 
@@ -114,6 +115,26 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 		return Result{}, errors.New("agent provider is required")
 	}
 
+	// Tracing is opt-in. When a recorder is wired, thread it into ctx so the
+	// providerio seam and reconnect helper can reach it via trace.FromContext,
+	// mark the run's start, and wrap OnUsage so token counters accumulate for
+	// free alongside the existing per-request plumbing. nil recorder leaves the
+	// loop byte-identical: FromContext returns nil, the no-op stamps cost
+	// nothing, and the OnUsage wrapper is not installed.
+	if options.Trace != nil {
+		ctx = trace.WithContext(ctx, options.Trace)
+		options.Trace.Start()
+		originalOnUsage := options.OnUsage
+		options.OnUsage = func(usage Usage) {
+			if originalOnUsage != nil {
+				originalOnUsage(usage)
+			}
+			options.Trace.Counter(trace.CounterInputTokens, int64(usage.InputTokens))
+			options.Trace.Counter(trace.CounterCachedInputTokens, int64(usage.CachedInputTokens))
+			options.Trace.Counter(trace.CounterOutputTokens, int64(usage.OutputTokens))
+		}
+	}
+
 	maxTurns := options.MaxTurns
 	if maxTurns <= 0 {
 		maxTurns = 12
@@ -184,12 +205,20 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 		// the tool-definition tokens (they ride on every request) in its estimate.
 		// partitionTools depends only on registry/permissions/options/loaded, not on
 		// the messages, so computing it before compaction is safe.
+		// Build the per-turn tool list first so proactive compaction can include
+		// the tool-definition tokens (they ride on every request) in its estimate.
+		// partitionTools depends only on registry/permissions/options/loaded, not on
+		// the messages, so computing it before compaction is safe.
+		toolPartitionSpan := options.Trace.Span(trace.SpanToolPartition)
 		exposed, _ := partitionToolsCached(registry, permissionMode, options, loaded, toolDefCache)
+		toolPartitionSpan.End()
 
 		// PROACTIVE compaction: if the history is approaching the model's
 		// context window, summarize the oldest middle before building the
 		// request. A no-op when ContextWindow == 0 (compaction disabled).
+		compactionSpan := options.Trace.Span(trace.SpanCompaction)
 		messages = compactor.maybeCompact(ctx, provider, messages, exposed)
+		compactionSpan.End()
 		request := zeroruntime.CompletionRequest{
 			Messages:        copyMessages(messages),
 			Tools:           exposed,
@@ -255,11 +284,39 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 		// conversation-state duplication.
 		forwardedVisibleText := false
 		forwardingOpts := zeroruntime.CollectOptions{OnUsage: options.OnUsage}
-		if options.OnText != nil {
-			forwardingOpts.OnText = func(s string) { forwardedVisibleText = true; options.OnText(s) }
+		// Install text/reasoning forwarding handlers whenever EITHER a user
+		// callback OR a trace recorder is set. A headless traced run (e.g. `zero
+		// exec --trace`) sets Trace but no OnText/OnReasoning; without these
+		// handlers the stream is still collected, but FirstTokenAt would never
+		// stamp and the trace would lose its TTFT signal. The trace recorder's
+		// stamp methods are nil-safe, but we guard on `trace != nil` so a run with
+		// a user callback but no recorder still works. forwardedVisibleText stays
+		// tied to the USER callback only, preserving the stall-retry semantics
+		// (a trace-only handler forwards nothing the user would see duplicated).
+		rec := options.Trace
+		onText := options.OnText
+		if onText != nil || rec != nil {
+			forwardingOpts.OnText = func(s string) {
+				if rec != nil {
+					rec.StampFirstVisibleEvent()
+					rec.StampFirstToken()
+				}
+				if onText != nil {
+					forwardedVisibleText = true
+					onText(s)
+				}
+			}
 		}
-		if options.OnReasoning != nil {
-			forwardingOpts.OnReasoning = func(s string) { options.OnReasoning(s) }
+		onReasoning := options.OnReasoning
+		if onReasoning != nil || rec != nil {
+			forwardingOpts.OnReasoning = func(s string) {
+				if rec != nil {
+					rec.StampFirstToken()
+				}
+				if onReasoning != nil {
+					onReasoning(s)
+				}
+			}
 		}
 		if options.OnToolCallStart != nil {
 			forwardingOpts.OnToolCallStart = func(id, name string) { options.OnToolCallStart(id, name) }
@@ -297,14 +354,18 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 				if retryStreamErr != nil {
 					return collected, retryStreamErr
 				}
+				genSpan := options.Trace.Span(trace.SpanGeneration)
 				collected = zeroruntime.CollectStreamWithOptions(ctx, retryStream, zeroruntime.CollectOptions{
 					OnUsage: options.OnUsage,
 				})
+				genSpan.End()
 			}
 			return collected, nil
 		}
 
+		generationSpan := options.Trace.Span(trace.SpanGeneration)
 		collected := zeroruntime.CollectStreamWithOptions(ctx, stream, forwardingOpts)
+		generationSpan.End()
 		if collected.Error != "" {
 			updated, stop := recoverStreamError(collected)
 			collected = updated
@@ -357,7 +418,9 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 				result.Messages = copyMessages(messages)
 				return result, retryErr
 			}
+			stallGenSpan := options.Trace.Span(trace.SpanGeneration)
 			collected = zeroruntime.CollectStreamWithOptions(ctx, retryStream, forwardingOpts)
+			stallGenSpan.End()
 		}
 		if collected.Error != "" {
 			// Route a reissued stream's non-stall error through the SAME recovery as
@@ -459,6 +522,7 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 				if cue || planPending {
 					if continueNudges < maxContinueNudges {
 						continueNudges++
+						options.Trace.Counter(trace.CounterCompletionNudges, 1)
 						reason := "your message ended mid-step"
 						if !cue {
 							reason = "pending plan items remain — finish them, or mark them complete with update_plan if you are done"
@@ -488,6 +552,7 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 				// (no admission, no cue) then finalizes as success on the next turn.
 				if options.SelfCorrect != nil && !acceptanceRequested {
 					acceptanceRequested = true
+					options.Trace.Counter(trace.CounterAcceptanceChecks, 1)
 					messages = append(messages, zeroruntime.Message{
 						Role:    zeroruntime.MessageRoleUser,
 						Content: acceptanceVerificationNudge(),
@@ -548,20 +613,26 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 					runEnd++
 				}
 				if runEnd-index >= 2 {
+					batchSpan := options.Trace.Span(trace.SpanToolExecution)
 					precomputed = executeParallelReadBatch(ctx, registry, collected.ToolCalls, index, runEnd, permissionMode, options)
+					batchSpan.End()
 					precomputedStart, precomputedEnd = index, runEnd
 				}
 			}
 			if options.OnToolCall != nil {
 				options.OnToolCall(call)
 			}
+			options.Trace.StampFirstUsefulAction()
 			var toolResult ToolResult
 			var abortErr error
 			if index >= precomputedStart && index < precomputedEnd {
 				toolResult, abortErr = precomputed[index-precomputedStart].result, precomputed[index-precomputedStart].abortErr
 			} else {
+				toolSpan := options.Trace.Span(trace.SpanToolExecution)
 				toolResult, abortErr = executeToolCall(ctx, registry, call, permissionMode, options)
+				toolSpan.End()
 			}
+			options.Trace.Counter(trace.CounterToolCalls, 1)
 			if options.OnToolResult != nil {
 				options.OnToolResult(toolResult)
 			}
@@ -666,6 +737,7 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 				// context-window sizing, and usage attribution follow the new model.
 				provider = newProvider
 				options.Model = turnRequestedModel
+				options.Trace.Counter(trace.CounterModelSwitches, 1)
 				// KNOWN LIMITATION (deferred): the compactor's context-window budget
 				// is fixed at run start from options.ContextWindow and is NOT updated
 				// here, so a switch to a model with a different window keeps compacting
@@ -764,6 +836,7 @@ func finalAnswerAfterMaxTurns(ctx context.Context, provider Provider, messages [
 	if err != nil {
 		return "", messages, ""
 	}
+	finalGenSpan := options.Trace.Span(trace.SpanGeneration)
 	collected := zeroruntime.CollectStreamWithOptions(ctx, stream, zeroruntime.CollectOptions{
 		OnText:          options.OnText,
 		OnReasoning:     options.OnReasoning,
@@ -771,6 +844,7 @@ func finalAnswerAfterMaxTurns(ctx context.Context, provider Provider, messages [
 		OnToolCallStart: options.OnToolCallStart,
 		OnToolCallDelta: options.OnToolCallDelta,
 	})
+	finalGenSpan.End()
 	if ctx.Err() != nil || collected.Error != "" || strings.TrimSpace(collected.Text) == "" {
 		return "", messages, ""
 	}
@@ -1874,7 +1948,10 @@ func requestPermission(ctx context.Context, request PermissionRequest, options O
 	if options.OnPermissionRequest == nil {
 		return PermissionDecision{Action: PermissionDecisionDeny, Reason: request.Reason}, nil
 	}
-	return options.OnPermissionRequest(ctx, request)
+	permSpan := options.Trace.Span(trace.SpanPermissionWait)
+	decision, err := options.OnPermissionRequest(ctx, request)
+	permSpan.End()
+	return decision, err
 }
 
 func normalizePermissionDecisionAction(action PermissionDecisionAction) PermissionDecisionAction {

internal/agent/loop_test.go

@@ -15,6 +15,7 @@ import (
 	"github.com/Gitlawb/zero/internal/sandbox"
 	"github.com/Gitlawb/zero/internal/specmode"
 	"github.com/Gitlawb/zero/internal/tools"
+	"github.com/Gitlawb/zero/internal/trace"
 	"github.com/Gitlawb/zero/internal/zeroruntime"
 )
 
@@ -2947,6 +2948,24 @@ func TestGitBranchForPromptResolvesRelativeWorktreeGitdir(t *testing.T) {
 	}
 }
 
+func TestGitBranchForPromptInSubdirectory(t *testing.T) {
+	root := t.TempDir()
+	gitdir := filepath.Join(root, ".git")
+	if err := os.MkdirAll(gitdir, 0o755); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(gitdir, "HEAD"), []byte("ref: refs/heads/main\n"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	subdir := filepath.Join(root, "sub", "dir")
+	if err := os.MkdirAll(subdir, 0o755); err != nil {
+		t.Fatal(err)
+	}
+	if got := gitBranchForPrompt(subdir); got != "main" {
+		t.Fatalf("gitBranchForPrompt in subdirectory = %q, want main", got)
+	}
+}
+
 func TestBuildSystemPromptInjectsWorkspaceContext(t *testing.T) {
 	cwd := t.TempDir()
 	if err := os.WriteFile(filepath.Join(cwd, "AGENTS.md"), []byte("Always run `make lint` before committing."), 0o644); err != nil {
@@ -3443,3 +3462,73 @@ func TestRunDoesNotFlagCleanToolOutput(t *testing.T) {
 		t.Errorf("clean output should not get a reminder, got %q", captured.Output)
 	}
 }
+
+// TestRunTracingWrapperStampsUsage verifies the per-turn tracing setup in Run:
+// a wired recorder is Started, OnUsage is wrapped so it still forwards to the
+// caller's callback AND stamps token counters, and the run completes.
+func TestRunTracingWrapperStampsUsage(t *testing.T) {
+	provider := &mockProvider{turns: [][]zeroruntime.StreamEvent{{
+		{Type: zeroruntime.StreamEventUsage, Usage: zeroruntime.Usage{InputTokens: 100, CachedInputTokens: 20, OutputTokens: 40}},
+		{Type: zeroruntime.StreamEventText, Content: "done"},
+		{Type: zeroruntime.StreamEventDone},
+	}}}
+	onUsageCalls := 0
+	rec := trace.NewRecorder("tracing-session", "run-1", "test")
+	if _, err := Run(context.Background(), "hi", provider, Options{
+		SessionID:    "tracing-session",
+		Cwd:          t.TempDir(),
+		ProviderName: "test-provider",
+		Model:        "test-model",
+		Trace:        rec,
+		OnUsage:      func(Usage) { onUsageCalls++ },
+	}); err != nil {
+		t.Fatalf("Run: %v", err)
+	}
+	tr := rec.Finish()
+	if tr.StartedAt.IsZero() {
+		t.Fatal("tracing wrapper did not Start the recorder")
+	}
+	// FirstTokenAt must stamp even though no OnText/OnReasoning user callback is
+	// set: a headless traced run (e.g. `zero exec --trace`) sets Trace but no UI
+	// callbacks, so the loop installs trace-only forwarding handlers to capture
+	// TTFT. Without them FirstTokenAt stays zero and the trace loses its signal.
+	if tr.FirstTokenAt.IsZero() {
+		t.Fatal("FirstTokenAt not stamped for a traced run with no OnText/OnReasoning callbacks")
+	}
+	if got := tr.Counter(trace.CounterInputTokens); got != 100 {
+		t.Fatalf("input token counter = %d, want 100", got)
+	}
+	if got := tr.Counter(trace.CounterCachedInputTokens); got != 20 {
+		t.Fatalf("cached input token counter = %d, want 20", got)
+	}
+	if got := tr.Counter(trace.CounterOutputTokens); got != 40 {
+		t.Fatalf("output token counter = %d, want 40", got)
+	}
+	if onUsageCalls == 0 {
+		t.Fatal("wrapped OnUsage did not forward to the caller's callback")
+	}
+}
+
+// TestRunNilTraceForwardsUsage verifies a nil recorder leaves the loop
+// byte-identical: OnUsage is not wrapped, so the caller's callback still fires
+// and nothing panics.
+func TestRunNilTraceForwardsUsage(t *testing.T) {
+	provider := &mockProvider{turns: [][]zeroruntime.StreamEvent{{
+		{Type: zeroruntime.StreamEventUsage, Usage: zeroruntime.Usage{InputTokens: 7}},
+		{Type: zeroruntime.StreamEventText, Content: "done"},
+		{Type: zeroruntime.StreamEventDone},
+	}}}
+	onUsageCalls := 0
+	if _, err := Run(context.Background(), "hi", provider, Options{
+		SessionID:    "nil-trace-session",
+		Cwd:          t.TempDir(),
+		ProviderName: "test-provider",
+		Model:        "test-model",
+		OnUsage:      func(Usage) { onUsageCalls++ },
+	}); err != nil {
+		t.Fatalf("Run: %v", err)
+	}
+	if onUsageCalls == 0 {
+		t.Fatal("OnUsage not forwarded when Trace is nil")
+	}
+}

internal/agent/reconnect.go

@@ -8,6 +8,7 @@ import (
 	"time"
 
 	"github.com/Gitlawb/zero/internal/errhint"
+	"github.com/Gitlawb/zero/internal/trace"
 	"github.com/Gitlawb/zero/internal/zeroruntime"
 )
 
@@ -74,14 +75,21 @@ func stallRetryNoticeFor(options Options) reconnectNotifier {
 // already past its deadline is returned immediately (no retry) — those have
 // their own handling (compaction for context-limit, image-rejection, etc.).
 func streamWithReconnect(ctx context.Context, provider Provider, request zeroruntime.CompletionRequest, notify reconnectNotifier) (<-chan zeroruntime.StreamEvent, error) {
+	recorder := trace.FromContext(ctx)
 	stream, err := provider.StreamCompletion(ctx, request)
 	if err == nil {
+		if recorder != nil {
+			recorder.Counter(trace.CounterModelRequests, 1)
+		}
 		return stream, nil
 	}
 	for attempt := 1; attempt <= maxStreamReconnects; attempt++ {
 		if !shouldReconnect(ctx, err) {
 			return nil, err
 		}
+		if recorder != nil {
+			recorder.Counter(trace.CounterReconnectCount, 1)
+		}
 		if notify != nil {
 			notify(attempt, maxStreamReconnects)
 		}
@@ -90,6 +98,9 @@ func streamWithReconnect(ctx context.Context, provider Provider, request zerorun
 		}
 		stream, err = provider.StreamCompletion(ctx, request)
 		if err == nil {
+			if recorder != nil {
+				recorder.Counter(trace.CounterModelRequests, 1)
+			}
 			return stream, nil
 		}
 	}

internal/agent/selfcorrect.go

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

internal/agent/system_prompt.go

@@ -295,7 +295,7 @@ func workspaceContext(cwd string) string {
 	}
 	b.WriteString("</environment>")
 
-	b.WriteString(projectGuidelines(cwd, findProjectGitRoot(cwd)))
+	b.WriteString(projectGuidelines(cwd, FindProjectGitRoot(cwd)))
 	if repoMap := repoMapContext(cwd); repoMap != "" {
 		b.WriteString("\n\n## Repo map\n\n" + repoMap)
 	}
@@ -542,17 +542,17 @@ func resolveDirCaseInsensitive(target, anchor string) (string, bool) {
 	return cur, true
 }
 
-// findProjectGitRoot returns the nearest ancestor of cwd that contains a
+// FindProjectGitRoot returns the nearest ancestor of cwd that contains a
 // .git entry (file or directory). Returns "" when no git root is found, so
 // the caller can fall back to cwd-only lookup.
-func findProjectGitRoot(cwd string) string {
+func FindProjectGitRoot(cwd string) string {
 	cwd = strings.TrimSpace(cwd)
 	if cwd == "" {
 		return ""
 	}
 	cur := cwd
 	for {
-		if hasGitMetadata(cur) {
+		if HasGitMetadata(cur) {
 			return cur
 		}
 		parent := filepath.Dir(cur)
@@ -563,7 +563,7 @@ func findProjectGitRoot(cwd string) string {
 	}
 }
 
-func hasGitMetadata(dir string) bool {
+func HasGitMetadata(dir string) bool {
 	gitPath := filepath.Join(dir, ".git")
 	info, err := os.Stat(gitPath)
 	if err != nil {
@@ -619,7 +619,11 @@ func repoMapContext(cwd string) string {
 // cwd, handling both a regular checkout (.git dir) and a worktree (.git file).
 // Returns "" on any problem — the prompt simply omits the branch segment.
 func gitBranchForPrompt(cwd string) string {
-	gitPath := filepath.Join(cwd, ".git")
+	gitRoot := FindProjectGitRoot(cwd)
+	if gitRoot == "" {
+		return ""
+	}
+	gitPath := filepath.Join(gitRoot, ".git")
 	info, err := os.Stat(gitPath)
 	if err != nil {
 		return ""
@@ -635,10 +639,10 @@ func gitBranchForPrompt(cwd string) string {
 			return ""
 		}
 		// In worktree mode the gitdir is often RELATIVE (e.g.
-		// "gitdir: ../.git/worktrees/<name>") — resolve it against cwd, not the
+		// "gitdir: ../.git/worktrees/<name>") — resolve it against the worktree root (gitRoot), not the
 		// process working directory, or HEAD lookup fails and we drop the branch.
 		if !filepath.IsAbs(dir) {
-			dir = filepath.Join(cwd, dir)
+			dir = filepath.Join(gitRoot, dir)
 		}
 		headPath = filepath.Join(dir, "HEAD")
 	}

internal/agent/system_prompt_test.go

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

internal/agent/types.go

@@ -7,6 +7,7 @@ import (
 	"github.com/Gitlawb/zero/internal/sandbox"
 	"github.com/Gitlawb/zero/internal/streamjson"
 	"github.com/Gitlawb/zero/internal/tools"
+	"github.com/Gitlawb/zero/internal/trace"
 	"github.com/Gitlawb/zero/internal/zeroruntime"
 )
 
@@ -279,6 +280,16 @@ type Options struct {
 	// request), so every existing caller is unaffected. A returned error is
 	// non-fatal: the run continues on the current model.
 	ModelSwitcher func(ctx context.Context, modelID string) (Provider, error)
+	// Trace, when set, records per-turn timing for the run: the loop stamps
+	// spans (prompt build, generation, tool execution, permission wait,
+	// compaction, provider connect) and counters (model requests, tool calls,
+	// retries, tokens) into it. nil DISABLES tracing entirely — every stamp is
+	// nil-safe and the loop is byte-identical to an untraced run. The caller
+	// owns the recorder: Run stamps into it but does not Finish or emit it.
+	// A fresh Recorder is required per Run — reusing one across runs merges
+	// their spans, counters, and first-event timestamps, and Finish freezes a
+	// recorder so no further stamps take.
+	Trace *trace.Recorder
 	// SelfCorrect, when set, runs a post-edit verify-and-correct cycle after a
 	// mutating tool call: it verifies the changed files (LSP diagnostics + project
 	// tests) and feeds failures back to the model to fix, bounded by an attempt

internal/background/process_windows.go

@@ -5,6 +5,7 @@ package background
 import (
 	"os"
 	"os/exec"
+	"path/filepath"
 	"strconv"
 )
 
@@ -14,7 +15,8 @@ import (
 func ConfigureChildProcessGroup(cmd *exec.Cmd) {}
 
 func terminateProcess(pid int) error {
-	if err := exec.Command("taskkill", "/T", "/F", "/PID", strconv.Itoa(pid)).Run(); err == nil {
+	taskkill := taskkillPath()
+	if err := exec.Command(taskkill, "/T", "/F", "/PID", strconv.Itoa(pid)).Run(); err == nil {
 		return nil
 	}
 	process, err := os.FindProcess(pid)
@@ -23,3 +25,14 @@ func terminateProcess(pid int) error {
 	}
 	return process.Kill()
 }
+
+func taskkillPath() string {
+	systemRoot := os.Getenv("SystemRoot")
+	if systemRoot == "" {
+		systemRoot = os.Getenv("windir")
+	}
+	if systemRoot == "" {
+		systemRoot = `C:\Windows`
+	}
+	return filepath.Join(systemRoot, "System32", "taskkill.exe")
+}

internal/cli/app.go

@@ -42,6 +42,7 @@ import (
 	"github.com/Gitlawb/zero/internal/worktrees"
 	"github.com/Gitlawb/zero/internal/zerogit"
 	"github.com/Gitlawb/zero/internal/zeroruntime"
+	"github.com/charmbracelet/x/term"
 )
 
 var version = "dev"
@@ -339,7 +340,17 @@ func runWithDeps(args []string, stdout io.Writer, stderr io.Writer, deps appDeps
 				return 0
 			}
 		}
-		if _, err := fmt.Fprintf(stdout, "zero %s\n", version); err != nil {
+		// Pipes and redirects keep the machine-readable contract exactly as it
+		// was: a single "zero <version>" line. Scripts, command substitutions,
+		// and the NPM wrapper smoke check all parse that record, so the banner
+		// is strictly a TTY affordance.
+		if !stdoutIsTerminal(stdout) {
+			if _, err := fmt.Fprintf(stdout, "zero %s\n", version); err != nil {
+				return 1
+			}
+			return 0
+		}
+		if _, err := fmt.Fprintf(stdout, "%s\n\nzero %s\n", tui.Wordmark(), version); err != nil {
 			return 1
 		}
 		return 0
@@ -1115,6 +1126,17 @@ func closeSpecialistRuntime(stderr io.Writer, runtime *agentToolRuntime) {
 	}
 }
 
+// stdoutIsTerminal reports whether w is an interactive terminal. Tests pass
+// bytes.Buffer writers and pipes/redirects fail term.IsTerminal, so both take
+// the machine-readable path.
+func stdoutIsTerminal(w io.Writer) bool {
+	f, ok := w.(*os.File)
+	if !ok {
+		return false
+	}
+	return term.IsTerminal(f.Fd())
+}
+
 func writeAppError(stderr io.Writer, message string, exitCode int) int {
 	if _, err := fmt.Fprintf(stderr, "[zero] %s\n", message); err != nil {
 		return 1
@@ -1347,6 +1369,8 @@ Flags:
       --notify <off|bell|notify|both>
                                     Override notification mode for this run
       --no-notify                   Disable notifications for this run
+      --no-completion-gate          Accept a no-tool-call reply as final without the INCOMPLETE
+                                    downgrade (for conversational callers with an operator present)
 `)
 	return err
 }

internal/cli/app_test.go

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

internal/cli/cron_run.go

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

internal/cli/distribution.go

@@ -118,7 +118,8 @@ func runSkillInfo(args []string, dir string, stdout io.Writer, stderr io.Writer)
 	if name == "" {
 		return writeExecUsageError(stderr, "usage: zero skill info <name> [--json]")
 	}
-	info, ok := skills.Info(dir, name)
+	// Resolve across primary + ~/.agents/skills; lock metadata only from primary.
+	info, ok := skills.InfoFromRoots(dir, skills.GlobalRoots(dir), name)
 	if !ok {
 		return writeAppError(stderr, fmt.Sprintf("skill %q not found", name), exitUsage)
 	}

internal/cli/exec.go

@@ -25,6 +25,7 @@ import (
 	"github.com/Gitlawb/zero/internal/specmode"
 	"github.com/Gitlawb/zero/internal/streamjson"
 	"github.com/Gitlawb/zero/internal/tools"
+	"github.com/Gitlawb/zero/internal/trace"
 	"github.com/Gitlawb/zero/internal/usage"
 	"github.com/Gitlawb/zero/internal/worktrees"
 	"github.com/Gitlawb/zero/internal/zeroruntime"
@@ -110,10 +111,26 @@ type execOptions struct {
 	notifyMode string
 	// noNotify forces ModeOff for this run. Mutually exclusive with notifyMode.
 	noNotify bool
+	// noCompletionGate turns off the headless completion gate
+	// (agent.Options.RequireCompletionSignal) for this run. It exists for
+	// CONVERSATIONAL exec callers — a chat frontend driving exec with an operator
+	// present to reply (e.g. zeroclaw chat). There, a final message that hands
+	// the turn back ("the sandbox blocks network egress, approve it and I'll
+	// continue") is a COMPLETE answer, not an unfinished task, so the gate's
+	// self-report admission check and INCOMPLETE downgrade (exit 4) would only
+	// mislabel honest blocker reports. Default off: plain `zero exec` keeps the
+	// gate, preserving CI/cron semantics.
+	noCompletionGate bool
 	// addDirs holds directories passed via --add-dir that should be allowed as
 	// additional write roots for this run. Unioned with
 	// config.SandboxConfig.AdditionalWriteRoots at scope construction time.
 	addDirs []string
+	// tracePath, when set, writes a per-turn NDJSON trace (agenteval-compatible)
+	// to the given file path — or to stderr when the value is "-". Falls back to
+	// the ZERO_TRACE env var when the flag is absent. Off by default: a run
+	// without it leaves agent.Options.Trace nil and is byte-identical to before.
+	// The trace is pure observation — enabling it does not change agent behavior.
+	tracePath string
 }
 
 type execUsageError struct {
@@ -412,6 +429,12 @@ func runExec(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) in
 	notify.MaybeAddWebhookSink(notifier, os.Getenv, func(format string, args ...any) {
 		fmt.Fprintf(stderr, "[notify] "+format+"\n", args...)
 	})
+	// Spec-draft runs synthesize a prompt offline and never drive a model turn, so
+	// there is no per-turn trace to emit. Reject --trace / ZERO_TRACE up front with
+	// a clear error rather than silently accepting and writing nothing.
+	if options.useSpec && resolveTracePath(options) != "" {
+		return writeExecFormatUsageError(stdout, stderr, options.outputFormat, "--trace / ZERO_TRACE are not supported for spec-draft runs")
+	}
 	if options.useSpec {
 		return runExecSpecDraft(execSpecDraftRun{
 			options:            options,
@@ -464,6 +487,24 @@ func runExec(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) in
 	if err != nil {
 		return writeAppError(stderr, "failed to create run id: "+err.Error(), exitCrash)
 	}
+	// Per-turn tracing is opt-in (--trace <path> or ZERO_TRACE=<path>). The
+	// recorder is stamped throughout agent.Run and the providerio seam; the run
+	// itself is byte-identical to an untraced run. We finish the recorder at the
+	// agent.Run boundary (below) so the snapshot captures exactly one turn, then
+	// serialize the snapshot on every exit path via the defer. "-" writes NDJSON
+	// to stderr; otherwise the path is created/truncated. A failure to emit is
+	// logged to stderr, never fatal.
+	tracePath := resolveTracePath(options)
+	var traceRecorder *trace.Recorder
+	var traceSnapshot *trace.TurnTrace
+	if tracePath != "" {
+		traceRecorder = trace.NewRecorder(preparedSession.Session.SessionID, runID, "")
+		defer func() {
+			if err := writeTraceSnapshot(traceSnapshot, tracePath, stderr); err != nil {
+				fmt.Fprintf(stderr, "[zero] failed to write trace: %s\n", err)
+			}
+		}()
+	}
 	writer := execEventWriter{
 		stdout:       stdout,
 		stderr:       stderr,
@@ -546,6 +587,7 @@ func runExec(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) in
 		Model:            resolved.Provider.Model,
 		ModelSwitcher:    modelSwitcher,
 		ReasoningEffort:  forwardEffort,
+		Trace:            traceRecorder,
 		Cwd:              workspaceRoot,
 		Images:           images,
 		Registry:         registry,
@@ -556,8 +598,10 @@ func runExec(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) in
 		// 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
-		// if the model keeps stalling. The interactive TUI leaves this off.
-		RequireCompletionSignal: true,
+		// if the model keeps stalling. The interactive TUI leaves this off, and
+		// --no-completion-gate lets conversational exec callers (a chat frontend
+		// with an operator present) opt out the same way.
+		RequireCompletionSignal: !options.noCompletionGate,
 		Sandbox:                 sandboxEngine,
 		FileTracker:             fileTracker,
 		Hooks:                   hookDispatcher,
@@ -613,6 +657,12 @@ func runExec(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) in
 			sessionRecorder.append(sessions.EventUsage, payload)
 		},
 	})
+	// Finish the trace now that the turn is done, so the snapshot captures exactly
+	// agent.Run's work and nothing the post-run cleanup stamps. The deferred
+	// writer serializes the snapshot on every exit path.
+	if traceRecorder != nil {
+		traceSnapshot = traceRecorder.Finish()
+	}
 	notifier.Notify(notify.Completion, notify.DefaultMessage(notify.Completion))
 	if writer.err != nil {
 		return exitCrash
@@ -1225,3 +1275,36 @@ func execNotifyMode(options execOptions, resolved config.ResolvedConfig) string
 	}
 	return resolved.Notify.Mode
 }
+
+// resolveTracePath returns the trace destination for this run: the --trace flag
+// value when set, else the ZERO_TRACE env var, else "" (tracing off). A value of
+// "-" means "write to stderr".
+func resolveTracePath(options execOptions) string {
+	if v := strings.TrimSpace(options.tracePath); v != "" {
+		return v
+	}
+	if v := strings.TrimSpace(os.Getenv("ZERO_TRACE")); v != "" {
+		return v
+	}
+	return ""
+}
+
+// writeTraceSnapshot writes an already-finished trace snapshot to dest ("-" =>
+// stderr, otherwise a file path created/truncated). A nil snapshot is a no-op
+// (tracing off, or the run exited before agent.Run produced a turn). The trace
+// is best-effort: a write error is returned to the caller, which logs it but
+// never fails the run.
+func writeTraceSnapshot(snapshot *trace.TurnTrace, dest string, stderr io.Writer) error {
+	if snapshot == nil || dest == "" {
+		return nil
+	}
+	if strings.TrimSpace(dest) == "-" {
+		return trace.WriteNDJSON(stderr, snapshot)
+	}
+	file, err := os.Create(dest)
+	if err != nil {
+		return err
+	}
+	defer file.Close()
+	return trace.WriteNDJSON(file, snapshot)
+}

internal/cli/exec_parse.go

@@ -30,6 +30,17 @@ func parseExecArgs(args []string) (execOptions, bool, error) {
 			options.selfCorrect = true
 		case arg == "--no-notify":
 			options.noNotify = true
+		case arg == "--no-completion-gate":
+			options.noCompletionGate = true
+		case arg == "--trace":
+			value, next, err := nextFlagValue(args, index, arg)
+			if err != nil {
+				return options, false, err
+			}
+			options.tracePath = value
+			index = next
+		case strings.HasPrefix(arg, "--trace="):
+			options.tracePath = strings.TrimSpace(strings.TrimPrefix(arg, "--trace="))
 		case arg == "--notify":
 			value, next, err := nextFlagValue(args, index, arg)
 			if err != nil {
@@ -402,6 +413,11 @@ func parseExecArgs(args []string) (execOptions, bool, error) {
 		// rather than pretend it took effect.
 		return options, false, execUsageError{"--self-correct cannot be combined with --use-spec."}
 	}
+	if options.useSpec && options.noCompletionGate {
+		// Same reasoning as --self-correct above: the spec-draft path never consults
+		// the completion gate, so the flag would be silently ignored.
+		return options, false, execUsageError{"--no-completion-gate cannot be combined with --use-spec."}
+	}
 	if !options.useSpec && options.specModel != "" {
 		return options, false, execUsageError{"--spec-model requires --use-spec."}
 	}
@@ -461,6 +477,49 @@ func nextFlagValue(args []string, index int, flag string) (string, int, error) {
 	if next == "" || flagValueLooksLikeOption(next) {
 		return "", index, execUsageError{fmt.Sprintf("%s requires a value", flag)}
 	}
+
+	// Validate specific flag values to prevent consuming positional prompt text
+	switch flag {
+	case "--auto":
+		switch strings.ToLower(next) {
+		case "low", "medium", "high", "member":
+		default:
+			return "", index, execUsageError{fmt.Sprintf("Invalid autonomy level %q. Expected low, medium, or high.", next)}
+		}
+	case "--notify":
+		switch strings.ToLower(next) {
+		case "off", "bell", "notify", "both":
+		default:
+			return "", index, execUsageError{fmt.Sprintf("invalid --notify %q. Expected off, bell, notify, or both.", next)}
+		}
+	case "-o", "--output":
+		switch strings.ToLower(next) {
+		case "text", "json", "stream-json", "debug":
+		default:
+			return "", index, execUsageError{fmt.Sprintf("invalid output format %q. Expected text, json, stream-json, or debug.", next)}
+		}
+	case "-i", "--input":
+		switch strings.ToLower(next) {
+		case "text", "stream-json":
+		default:
+			return "", index, execUsageError{fmt.Sprintf("Invalid input format %q. Expected text or stream-json.", next)}
+		}
+	case "--reasoning-effort", "--spec-reasoning-effort":
+		switch strings.ToLower(next) {
+		case "low", "medium", "high":
+		default:
+			return "", index, execUsageError{fmt.Sprintf("invalid %s %q. Expected low, medium, or high.", flag, next)}
+		}
+	case "--depth":
+		if _, err := strconv.Atoi(next); err != nil {
+			return "", index, execUsageError{fmt.Sprintf("invalid --depth %q. Expected a non-negative integer.", next)}
+		}
+	case "--max-turns":
+		if _, err := strconv.Atoi(next); err != nil {
+			return "", index, execUsageError{fmt.Sprintf("invalid --max-turns %q. Expected a positive integer.", next)}
+		}
+	}
+
 	return next, index + 1, nil
 }
 

internal/cli/exec_parse_completion_gate_test.go

@@ -0,0 +1,34 @@
+package cli
+
+import "testing"
+
+// --no-completion-gate exists for conversational exec callers (a chat frontend
+// with an operator present): an honest blocker report that hands the turn back
+// must not be downgraded to INCOMPLETE there. The flag maps to
+// agent.Options.RequireCompletionSignal = false; default keeps the gate on.
+func TestParseExecArgsNoCompletionGate(t *testing.T) {
+	options, _, err := parseExecArgs([]string{"--prompt", "hi", "--no-completion-gate"})
+	if err != nil {
+		t.Fatalf("parseExecArgs: %v", err)
+	}
+	if !options.noCompletionGate {
+		t.Fatal("--no-completion-gate must set noCompletionGate")
+	}
+
+	options, _, err = parseExecArgs([]string{"--prompt", "hi"})
+	if err != nil {
+		t.Fatalf("parseExecArgs: %v", err)
+	}
+	if options.noCompletionGate {
+		t.Fatal("noCompletionGate must default to false (gate on)")
+	}
+}
+
+// The spec-draft path never consults the completion gate, so the combination
+// must be rejected up front instead of silently ignoring the flag (mirrors the
+// existing --self-correct + --use-spec check).
+func TestParseExecArgsNoCompletionGateRejectsUseSpec(t *testing.T) {
+	if _, _, err := parseExecArgs([]string{"--prompt", "hi", "--use-spec", "--no-completion-gate"}); err == nil {
+		t.Fatal("--no-completion-gate with --use-spec must error")
+	}
+}

internal/cli/exec_parse_prompt_test.go

@@ -21,3 +21,20 @@ func TestParseExecArgsInlinePromptKeepsLeadingDash(t *testing.T) {
 		t.Fatalf("inline prompt = %q, want %q", got, "-foo bar")
 	}
 }
+
+func TestParseExecArgsPreventsPositionalFromBeingConsumedAsFlagValue(t *testing.T) {
+	_, _, err := parseExecArgs([]string{"--auto", "do a build"})
+	if err == nil || !strings.Contains(err.Error(), "Invalid autonomy level") {
+		t.Fatalf("expected error that --auto has invalid autonomy level, got: %v", err)
+	}
+
+	_, _, err = parseExecArgs([]string{"--notify", "do a build"})
+	if err == nil || !strings.Contains(err.Error(), "invalid --notify") {
+		t.Fatalf("expected error that --notify is invalid, got: %v", err)
+	}
+
+	_, _, err = parseExecArgs([]string{"--max-turns", "do a build"})
+	if err == nil || !strings.Contains(err.Error(), "invalid --max-turns") {
+		t.Fatalf("expected error that --max-turns is invalid, got: %v", err)
+	}
+}

internal/cli/exec_test.go

@@ -345,8 +345,8 @@ func TestRunExecModeSeedsModelAndTurnOverrides(t *testing.T) {
 	if gotModel != "claude-opus-4.1" {
 		t.Fatalf("overrides.Provider.Model = %q, want claude-opus-4.1", gotModel)
 	}
-	if gotMaxTurns != 50 {
-		t.Fatalf("overrides.MaxTurns = %d, want 50", gotMaxTurns)
+	if gotMaxTurns != 160 {
+		t.Fatalf("overrides.MaxTurns = %d, want 160 (deep mode)", gotMaxTurns)
 	}
 }
 

internal/cli/mcp_oauth.go

@@ -86,7 +86,7 @@ func runMCPOAuthLogin(args []string, stdout io.Writer, stderr io.Writer, deps ap
 	if err != nil {
 		return writeAppError(stderr, redaction.ErrorMessage(err, redaction.Options{}), exitCrash)
 	}
-	if err := store.Save(serverName, token); err != nil {
+	if err := store.SaveForServer(server, token); err != nil {
 		return writeAppError(stderr, redaction.ErrorMessage(err, redaction.Options{}), exitCrash)
 	}
 
@@ -116,7 +116,7 @@ func runMCPOAuthLogout(args []string, stdout io.Writer, stderr io.Writer, deps a
 	if err != nil {
 		return writeAppError(stderr, redaction.ErrorMessage(err, redaction.Options{}), exitCrash)
 	}
-	removed, err := store.Delete(serverName)
+	removed, err := store.DeleteForServerName(serverName)
 	if err != nil {
 		return writeAppError(stderr, redaction.ErrorMessage(err, redaction.Options{}), exitCrash)
 	}

internal/cli/mcp_oauth_test.go

@@ -88,6 +88,9 @@ func TestRunMCPOAuthLogout(t *testing.T) {
 	if err := store.Save("remote", mcp.StoredToken{AccessToken: "a", RefreshToken: "r"}); err != nil {
 		t.Fatalf("Save() error = %v", err)
 	}
+	if err := store.SaveForServer(mcp.Server{Name: "remote", Identity: "aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa"}, mcp.StoredToken{AccessToken: "identity", RefreshToken: "ir"}); err != nil {
+		t.Fatalf("SaveForServer() error = %v", err)
+	}
 	deps := appDeps{newMCPTokenStore: func() (*mcp.TokenStore, error) { return store, nil }}
 
 	var stdout, stderr bytes.Buffer
@@ -101,6 +104,9 @@ func TestRunMCPOAuthLogout(t *testing.T) {
 	if _, ok, _ := store.Load("remote"); ok {
 		t.Fatal("token still present after logout")
 	}
+	if statuses, err := store.Status(); err != nil || len(statuses) != 0 {
+		t.Fatalf("status after logout = %#v err=%v, want no tokens", statuses, err)
+	}
 }
 
 func TestRunMCPOAuthLoginStoresTokens(t *testing.T) {
@@ -122,23 +128,24 @@ func TestRunMCPOAuthLoginStoresTokens(t *testing.T) {
 	}
 
 	cwd := t.TempDir()
+	mcpConfig := config.MCPConfig{Servers: map[string]config.MCPServerConfig{
+		"remote": {
+			Type: "http",
+			URL:  "https://remote.invalid/mcp",
+			Auth: "oauth",
+			OAuth: &config.MCPOAuthConfig{
+				ClientID:              "client-123",
+				AuthorizationEndpoint: "https://remote.invalid/authorize",
+				TokenEndpoint:         tokenServer.URL,
+				Scopes:                []string{"read"},
+			},
+		},
+	}}
 	deps := appDeps{
 		getwd:            func() (string, error) { return cwd, nil },
 		newMCPTokenStore: func() (*mcp.TokenStore, error) { return store, nil },
-		resolveMCPConfig: func(workspaceRoot string, _ bool) (config.MCPConfig, error) {
-			return config.MCPConfig{Servers: map[string]config.MCPServerConfig{
-				"remote": {
-					Type: "http",
-					URL:  "https://remote.invalid/mcp",
-					Auth: "oauth",
-					OAuth: &config.MCPOAuthConfig{
-						ClientID:              "client-123",
-						AuthorizationEndpoint: "https://remote.invalid/authorize",
-						TokenEndpoint:         tokenServer.URL,
-						Scopes:                []string{"read"},
-					},
-				},
-			}}, nil
+		resolveMCPConfig: func(_ string, _ bool) (config.MCPConfig, error) {
+			return mcpConfig, nil
 		},
 		now: time.Now,
 	}
@@ -173,9 +180,13 @@ func TestRunMCPOAuthLoginStoresTokens(t *testing.T) {
 	if exitCode != exitSuccess {
 		t.Fatalf("exitCode = %d stderr=%s stdout=%s", exitCode, stderr.String(), stdout.String())
 	}
-	token, ok, err := store.Load("remote")
+	servers, err := mcp.NormalizeConfig(mcpConfig)
+	if err != nil || len(servers) != 1 {
+		t.Fatalf("NormalizeConfig() servers=%#v err=%v", servers, err)
+	}
+	token, ok, err := store.LoadForServer(servers[0])
 	if err != nil || !ok {
-		t.Fatalf("Load() ok=%v err=%v", ok, err)
+		t.Fatalf("LoadForServer() ok=%v err=%v", ok, err)
 	}
 	if token.AccessToken != "access-final" || token.RefreshToken != "refresh-final" {
 		t.Fatalf("stored token = %#v", token)
@@ -483,7 +494,13 @@ func TestRunMCPOAuthLoginTrustedProjectServerStoresToken(t *testing.T) {
 	if gotExclude {
 		t.Fatal("trusted workspace must resolve MCP config with excludeProject=false")
 	}
-	token, ok, err := store.Load("remote")
+	servers, err := mcp.NormalizeConfig(config.MCPConfig{Servers: map[string]config.MCPServerConfig{
+		"remote": projectServer,
+	}})
+	if err != nil || len(servers) != 1 {
+		t.Fatalf("NormalizeConfig() servers=%#v err=%v", servers, err)
+	}
+	token, ok, err := store.LoadForServer(servers[0])
 	if err != nil || !ok {
 		t.Fatalf("token must be stored on trusted login; ok=%v err=%v", ok, err)
 	}

internal/cli/plugin_activate.go

@@ -15,11 +15,11 @@ import (
 
 // pluginActivation holds what plugin activation contributed to a bootstrap so the
 // later dispatcher + skill wiring can consume it: the plugin hook definitions and
-// the plugin skill search roots. The zero value (no plugins) is inert — the
-// dispatcher gets no extra hooks and the skill tool keeps only the default dir.
-// The embedded trustSkip reports whether the project plugin layer was dropped for
-// an untrusted workspace, so the caller can fold it into the single combined
-// trust notice.
+// the plugin skill search roots. The zero value (no plugins) still overlays the
+// multi-root skill tool (primary + ~/.agents/skills) so discovery is identical
+// with or without plugins. The embedded trustSkip reports whether the project
+// plugin layer was dropped for an untrusted workspace, so the caller can fold it
+// into the single combined trust notice.
 type pluginActivation struct {
 	hooks      []hooks.Definition
 	skillRoots []string
@@ -52,6 +52,9 @@ func activatePlugins(workspaceRoot string, registry *tools.Registry, deps appDep
 	loaded, err := deps.loadPlugins(plugins.LoadOptions{Cwd: workspaceRoot, ExcludeProject: excludeProject})
 	if err != nil {
 		writePluginActivationWarning(stderr, "failed to load plugins: "+err.Error())
+		// Still overlay multi-root discovery (primary + agents) so a plugin load
+		// failure cannot hide shared skills.
+		registry.Register(plugins.NewSkillTool(deps.skillsDir(), nil))
 		return pluginActivation{trustSkip: skip}
 	}
 
@@ -67,14 +70,11 @@ func activatePlugins(workspaceRoot string, registry *tools.Registry, deps appDep
 		writePluginActivationWarning(stderr, warning)
 	}
 
-	// Re-register the skill tool to also resolve plugin-declared skills. The core
-	// skill tool only reads the default skills dir; the plugin-aware replacement
-	// merges the default dir with the plugin skill roots (default dir wins a name
-	// clash), so plugin skills appear in the agent's skill list. With no plugin
-	// skill roots this is byte-equivalent to the default skills surface.
-	if len(result.SkillRoots) > 0 {
-		registry.Register(plugins.NewSkillTool(deps.skillsDir(), result.SkillRoots))
-	}
+	// Always overlay the multi-root skill tool so discovery is identical with or
+	// without plugins: primary DefaultDir + optional ~/.agents/skills + plugin
+	// roots (earlier wins). With empty SkillRoots and no agents dir this is
+	// byte-equivalent to the core single-dir skill surface.
+	registry.Register(plugins.NewSkillTool(deps.skillsDir(), result.SkillRoots))
 
 	return pluginActivation{hooks: result.Hooks, skillRoots: result.SkillRoots, trustSkip: skip}
 }
@@ -91,8 +91,8 @@ func projectPluginsDirExists(workspaceRoot string) bool {
 }
 
 // skillInfos resolves the reusable skills the model can load via the skill tool —
-// the default skills dir merged with any plugin skill roots, the same set the
-// plugin-aware skill tool resolves against — as plain data for the agent's system
+// primary dir + optional ~/.agents/skills + plugin skill roots, the same set the
+// multi-root skill tool resolves against — as plain data for the agent's system
 // prompt. It returns nil when no skills are installed, so a skill-less run leaves
 // the prompt byte-identical.
 func (a pluginActivation) skillInfos(defaultDir string) []agent.SkillInfo {

internal/cli/plugin_activate_test.go

@@ -69,6 +69,11 @@ func TestActivatePluginsRegistersToolAndCollectsHooks(t *testing.T) {
 }
 
 func TestActivatePluginsRegistersPluginSkillTool(t *testing.T) {
+	// Isolate host ~/.agents/skills so it cannot interfere with this test.
+	home := t.TempDir()
+	t.Setenv("HOME", home)
+	t.Setenv("USERPROFILE", home)
+
 	pluginDir := t.TempDir()
 	skillDir := filepath.Join(pluginDir, "skills", "demo-skill")
 	if err := os.MkdirAll(skillDir, 0o755); err != nil {
@@ -108,6 +113,42 @@ func TestActivatePluginsRegistersPluginSkillTool(t *testing.T) {
 	}
 }
 
+func TestActivatePluginsAlwaysRegistersMultiRootSkillTool(t *testing.T) {
+	// Even with zero plugins / zero plugin skill roots, activation must overlay
+	// the multi-root skill tool so ~/.agents/skills is loadable on bare runs.
+	home := t.TempDir()
+	t.Setenv("HOME", home)
+	t.Setenv("USERPROFILE", home)
+	agents := filepath.Join(home, ".agents", "skills")
+	if err := os.MkdirAll(agents, 0o755); err != nil {
+		t.Fatal(err)
+	}
+	skillDir := filepath.Join(agents, "agents-skill")
+	if err := os.MkdirAll(skillDir, 0o755); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(skillDir, "SKILL.md"), []byte("---\nname: agents-skill\n---\nagents body"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	registry := tools.NewRegistry()
+	registry.Register(tools.NewSkillTool(t.TempDir())) // core single-dir tool
+	var stderr bytes.Buffer
+	workspace := t.TempDir()
+	activation := activatePlugins(workspace, registry, fakePluginDeps(t, nil), &stderr, workspace)
+	if len(activation.skillRoots) != 0 {
+		t.Fatalf("expected no plugin skill roots, got %#v", activation.skillRoots)
+	}
+	skillTool, ok := registry.Get("skill")
+	if !ok {
+		t.Fatal("skill tool missing after activation")
+	}
+	res := skillTool.Run(context.Background(), map[string]any{"name": "agents-skill"})
+	if res.Status != tools.StatusOK || !bytes.Contains([]byte(res.Output), []byte("agents body")) {
+		t.Fatalf("agents skill must load with zero plugin roots: %q (%s)", res.Status, res.Output)
+	}
+}
+
 func TestActivatePluginsSurfacesLoadDiagnostics(t *testing.T) {
 	// Load succeeds (no top-level error) but reports a per-plugin diagnostic for a
 	// plugin it had to skip. activatePlugins must forward that diagnostic to stderr

internal/cli/provider_catalog_onboarding_test.go

@@ -17,6 +17,7 @@ func TestProviderCatalogRuntimeProvidersShowOnboardingSetupHints(t *testing.T) {
 		defaultModel string
 	}{
 		{id: "openai", name: "OpenAI", transport: "openai", defaultModel: "gpt-4.1"},
+		{id: "aimlapi", name: "AI/ML API", transport: "openai-compatible", defaultModel: "openai/gpt-5-chat"},
 		{id: "groq", name: "Groq", transport: "openai-compatible", defaultModel: "llama-3.3-70b-versatile"},
 		{id: "longcat", name: "LongCat", transport: "openai-compatible", defaultModel: "LongCat-2.0"},
 		{id: "ollama-cloud", name: "Ollama Cloud", transport: "openai-compatible", defaultModel: "qwen3-coder:480b"},
@@ -47,6 +48,21 @@ func TestProviderCatalogRuntimeProvidersShowOnboardingSetupHints(t *testing.T) {
 	}
 }
 
+func TestProviderCatalogAIMLAPIRequiresAPIKey(t *testing.T) {
+	output := runProviderCatalogOnboarding(t)
+	block := providerCatalogOnboardingBlock(t, output, "aimlapi")
+
+	for _, want := range []string{
+		"requiresAuth=true",
+		"authEnvVars=AIMLAPI_API_KEY",
+		"setup: zero providers setup aimlapi --set-active",
+	} {
+		if !strings.Contains(block, want) {
+			t.Fatalf("expected AIMLAPI catalog block to contain %q, got:\n%s", want, block)
+		}
+	}
+}
+
 func TestProviderCatalogLongCatRequiresAPIKey(t *testing.T) {
 	output := runProviderCatalogOnboarding(t)
 	block := providerCatalogOnboardingBlock(t, output, "longcat")

internal/cli/sandbox_test.go

@@ -483,6 +483,14 @@ func TestTUISandboxSetupCommandGatedToWindowsNativeBackend(t *testing.T) {
 }
 
 func TestRunSandboxPolicyJSONGoldenIncludesManagerBaselineFields(t *testing.T) {
+	// Point HOME at an empty directory so the default credential-store
+	// deny-read entries (which depend on what exists in the real home, e.g.
+	// ~/.aws on the macOS CI image) cannot leak host paths into the golden
+	// comparison.
+	emptyHome := t.TempDir()
+	t.Setenv("HOME", emptyHome)
+	t.Setenv("USERPROFILE", emptyHome)
+	t.Setenv("GOOGLE_APPLICATION_CREDENTIALS", "")
 	store := newSandboxTestStore(t)
 	workspace := t.TempDir()
 	deps := appDeps{
@@ -616,8 +624,12 @@ func jsonValuesEqual(t *testing.T, wantBytes []byte, gotBytes []byte) bool {
 func normalizePortableJSONRootString(value *any) {
 	switch current := (*value).(type) {
 	case string:
-		if current == `\` {
-			*value = "/"
+		// Normalize OS-specific path separators so the shared golden file
+		// matches on both Unix (slash) and Windows (backslash) runners.
+		// readOnlySubpaths carve-outs are built with filepath.Join, which
+		// renders as backslash paths on Windows.
+		if current != "" {
+			*value = strings.ReplaceAll(current, `\`, "/")
 		}
 	case []any:
 		for index := range current {

internal/cli/serve.go

@@ -4,6 +4,8 @@ import (
 	"context"
 	"fmt"
 	"io"
+	"os"
+	"path/filepath"
 	"strings"
 
 	"github.com/Gitlawb/zero/internal/mcp"
@@ -14,9 +16,21 @@ import (
 type serveOptions struct {
 	mcp              bool
 	cwd              string
+	addDirs          []string
 	allowUnsafeTools bool
 }
 
+// serveRootsScope is a PathScope used only for MCP serve: the workspace root
+// plus optional --add-dir roots, without the sandbox temp write roots that
+// would otherwise pollute resources/list.
+type serveRootsScope []string
+
+func (s serveRootsScope) Roots() []string {
+	roots := make([]string, len(s))
+	copy(roots, s)
+	return roots
+}
+
 func runServe(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) int {
 	options, help, err := parseServeArgs(args)
 	if err != nil {
@@ -36,7 +50,11 @@ func runServe(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) i
 	if err != nil {
 		return writeExecUsageError(stderr, err.Error())
 	}
-	registry := newServeRegistry(workspaceRoot, options.allowUnsafeTools)
+	scope, err := buildServeScope(workspaceRoot, options.addDirs)
+	if err != nil {
+		return writeExecUsageError(stderr, err.Error())
+	}
+	registry := newServeRegistry(workspaceRoot, scope, options.allowUnsafeTools)
 	if options.allowUnsafeTools {
 		if _, err := fmt.Fprintln(stderr, "[zero] Unsafe MCP server tools enabled because --allow-unsafe-tools was passed."); err != nil {
 			return exitCrash
@@ -47,6 +65,8 @@ func runServe(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) i
 		Name:              "zero",
 		Version:           version,
 		PermissionGranted: options.allowUnsafeTools,
+		WorkspaceRoot:     workspaceRoot,
+		Scope:             scope,
 	})
 	if err != nil {
 		return writeAppError(stderr, redaction.ErrorMessage(err, redaction.Options{}), exitCrash)
@@ -54,11 +74,11 @@ func runServe(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) i
 	return exitSuccess
 }
 
-func newServeRegistry(workspaceRoot string, allowUnsafeTools bool) *tools.Registry {
+func newServeRegistry(workspaceRoot string, scope tools.PathScope, allowUnsafeTools bool) *tools.Registry {
 	registry := tools.NewRegistry()
-	toolset := tools.CoreReadOnlyTools(workspaceRoot)
+	toolset := tools.CoreReadOnlyToolsScoped(workspaceRoot, scope)
 	if allowUnsafeTools {
-		toolset = tools.CoreTools(workspaceRoot)
+		toolset = tools.CoreToolsScoped(workspaceRoot, scope)
 	}
 	for _, tool := range toolset {
 		registry.Register(tool)
@@ -66,6 +86,58 @@ func newServeRegistry(workspaceRoot string, allowUnsafeTools bool) *tools.Regist
 	return registry
 }
 
+// buildServeScope returns nil when there are no extra roots (workspace-only),
+// otherwise a PathScope listing the workspace first followed by each --add-dir.
+func buildServeScope(workspaceRoot string, addDirs []string) (tools.PathScope, error) {
+	if len(addDirs) == 0 {
+		return nil, nil
+	}
+	roots := make([]string, 0, 1+len(addDirs))
+	workspaceRoot = filepath.Clean(workspaceRoot)
+	if abs, err := filepath.Abs(workspaceRoot); err == nil {
+		workspaceRoot = abs
+	}
+
+	// Keep lexical absolute paths in roots so path checks stay aligned with the
+	// un-evaluated WorkspaceRoot used when Scope is nil. Symlink-resolved forms
+	// are only keys for duplicate detection.
+	workspaceResolved := workspaceRoot
+	if resolved, err := filepath.EvalSymlinks(workspaceRoot); err == nil {
+		workspaceResolved = resolved
+	}
+	roots = append(roots, workspaceRoot)
+
+	seen := map[string]struct{}{workspaceResolved: {}}
+	for _, dir := range addDirs {
+		trimmed := strings.TrimSpace(dir)
+		if trimmed == "" {
+			return nil, execUsageError{"--add-dir requires a path"}
+		}
+		absolute, err := filepath.Abs(trimmed)
+		if err != nil {
+			return nil, execUsageError{fmt.Sprintf("--add-dir %q: %v", trimmed, err)}
+		}
+		info, err := os.Stat(absolute)
+		if err != nil {
+			return nil, execUsageError{fmt.Sprintf("--add-dir %q: %v", trimmed, err)}
+		}
+		if !info.IsDir() {
+			return nil, execUsageError{fmt.Sprintf("--add-dir %q is not a directory", trimmed)}
+		}
+
+		resolved := absolute
+		if r, err := filepath.EvalSymlinks(absolute); err == nil {
+			resolved = r
+		}
+		if _, ok := seen[resolved]; ok {
+			continue
+		}
+		seen[resolved] = struct{}{}
+		roots = append(roots, absolute)
+	}
+	return serveRootsScope(roots), nil
+}
+
 func parseServeArgs(args []string) (serveOptions, bool, error) {
 	options := serveOptions{}
 	for index := 0; index < len(args); index++ {
@@ -85,6 +157,18 @@ func parseServeArgs(args []string) (serveOptions, bool, error) {
 			options.cwd = args[index]
 		case strings.HasPrefix(arg, "--cwd="):
 			options.cwd = strings.TrimPrefix(arg, "--cwd=")
+		case arg == "--add-dir":
+			index++
+			if index >= len(args) {
+				return options, false, execUsageError{"--add-dir requires a path"}
+			}
+			options.addDirs = append(options.addDirs, args[index])
+		case strings.HasPrefix(arg, "--add-dir="):
+			value := strings.TrimPrefix(arg, "--add-dir=")
+			if strings.TrimSpace(value) == "" {
+				return options, false, execUsageError{"--add-dir requires a path"}
+			}
+			options.addDirs = append(options.addDirs, value)
 		default:
 			return options, false, execUsageError{fmt.Sprintf("unknown serve flag %q", arg)}
 		}
@@ -100,7 +184,8 @@ Starts Zero as an MCP stdio server.
 
 Flags:
       --mcp                   Run the MCP stdio server
-  -C, --cwd <path>            Set the workspace directory
+  -C, --cwd <path>            Set the workspace directory (resources + tools)
+      --add-dir <path>        Extra resource/tool root (repeatable)
       --allow-unsafe-tools    Expose write and shell tools to the MCP host
   -h, --help                  Show this help
 `)

internal/cli/serve_test.go

@@ -4,6 +4,8 @@ import (
 	"bytes"
 	"encoding/json"
 	"fmt"
+	"os"
+	"path/filepath"
 	"strings"
 	"testing"
 )
@@ -80,6 +82,120 @@ func TestRunServeRequiresMCPMode(t *testing.T) {
 	}
 }
 
+func TestRunServeMCPWiresWorkspaceRootAndAddDirIntoResources(t *testing.T) {
+	workspace := t.TempDir()
+	extra := t.TempDir()
+	if err := os.WriteFile(filepath.Join(workspace, "workspace.txt"), []byte("ws\n"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(extra, "extra.txt"), []byte("ex\n"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+
+	var stdout bytes.Buffer
+	var stderr bytes.Buffer
+	exitCode := runWithDeps(
+		[]string{"serve", "--mcp", "-C", workspace, "--add-dir", extra},
+		&stdout,
+		&stderr,
+		appDeps{stdin: bytes.NewReader(serveMCPResourcesInput(t))},
+	)
+	if exitCode != exitSuccess {
+		t.Fatalf("exit=%d stderr=%q", exitCode, stderr.String())
+	}
+	output := stdout.String()
+	if !strings.Contains(output, "workspace.txt") {
+		t.Fatalf("expected workspace resource, got %q", output)
+	}
+	if !strings.Contains(output, "extra.txt") {
+		t.Fatalf("expected --add-dir resource, got %q", output)
+	}
+}
+
+func TestRunServeMCPRejectsMissingAddDir(t *testing.T) {
+	workspace := t.TempDir()
+	missing := filepath.Join(workspace, "does-not-exist")
+	var stdout bytes.Buffer
+	var stderr bytes.Buffer
+	exitCode := runWithDeps(
+		[]string{"serve", "--mcp", "-C", workspace, "--add-dir", missing},
+		&stdout,
+		&stderr,
+		appDeps{stdin: bytes.NewReader(nil)},
+	)
+	if exitCode != exitUsage {
+		t.Fatalf("exit=%d want %d stderr=%q", exitCode, exitUsage, stderr.String())
+	}
+	if !strings.Contains(stderr.String(), "--add-dir") {
+		t.Fatalf("expected --add-dir error, got %q", stderr.String())
+	}
+}
+
+func TestParseServeArgsCollectsAddDirs(t *testing.T) {
+	options, help, err := parseServeArgs([]string{"--mcp", "--add-dir", "/one", "--add-dir=/two", "-C", "/ws"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if help {
+		t.Fatal("unexpected help")
+	}
+	if !options.mcp || options.cwd != "/ws" {
+		t.Fatalf("options=%#v", options)
+	}
+	if len(options.addDirs) != 2 || options.addDirs[0] != "/one" || options.addDirs[1] != "/two" {
+		t.Fatalf("addDirs=%v", options.addDirs)
+	}
+}
+
+func TestBuildServeScopeNilWithoutExtras(t *testing.T) {
+	scope, err := buildServeScope(t.TempDir(), nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if scope != nil {
+		t.Fatalf("expected nil scope, got %#v", scope)
+	}
+}
+
+func TestBuildServeScopeKeepsLexicalPaths(t *testing.T) {
+	base := t.TempDir()
+	realWorkspace := filepath.Join(base, "real-workspace")
+	realExtra := filepath.Join(base, "real-extra")
+	for _, dir := range []string{realWorkspace, realExtra} {
+		if err := os.MkdirAll(dir, 0o700); err != nil {
+			t.Fatal(err)
+		}
+	}
+	linkWorkspace := filepath.Join(base, "link-workspace")
+	linkExtra := filepath.Join(base, "link-extra")
+	if err := os.Symlink(realWorkspace, linkWorkspace); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.Symlink(realExtra, linkExtra); err != nil {
+		t.Fatal(err)
+	}
+
+	scope, err := buildServeScope(linkWorkspace, []string{linkExtra})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if scope == nil {
+		t.Fatal("expected non-nil scope")
+	}
+	roots := scope.Roots()
+	wantWorkspace, err := filepath.Abs(linkWorkspace)
+	if err != nil {
+		t.Fatal(err)
+	}
+	wantExtra, err := filepath.Abs(linkExtra)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(roots) != 2 || roots[0] != wantWorkspace || roots[1] != wantExtra {
+		t.Fatalf("roots=%v want [%q %q]", roots, wantWorkspace, wantExtra)
+	}
+}
+
 func serveMCPInput(t *testing.T) []byte {
 	t.Helper()
 
@@ -104,6 +220,30 @@ func serveMCPInput(t *testing.T) []byte {
 	return input.Bytes()
 }
 
+func serveMCPResourcesInput(t *testing.T) []byte {
+	t.Helper()
+
+	var input bytes.Buffer
+	writeServeMCPMessage(t, &input, map[string]any{
+		"jsonrpc": "2.0",
+		"id":      1,
+		"method":  "initialize",
+		"params":  map[string]any{},
+	})
+	writeServeMCPMessage(t, &input, map[string]any{
+		"jsonrpc": "2.0",
+		"method":  "notifications/initialized",
+		"params":  map[string]any{},
+	})
+	writeServeMCPMessage(t, &input, map[string]any{
+		"jsonrpc": "2.0",
+		"id":      2,
+		"method":  "resources/list",
+		"params":  map[string]any{},
+	})
+	return input.Bytes()
+}
+
 func writeServeMCPMessage(t *testing.T, buffer *bytes.Buffer, message map[string]any) {
 	t.Helper()
 	body, err := json.Marshal(message)

internal/cli/skills.go

@@ -59,25 +59,21 @@ func runSkills(args []string, stdout io.Writer, stderr io.Writer, deps appDeps)
 }
 
 func runSkillsList(dir string, options skillListOptions, stdout io.Writer, stderr io.Writer) int {
-	discovered, err := skills.List(dir)
+	// Management CLI lists global roots only (primary + ~/.agents/skills); plugin
+	// skills stay out of `zero skills list` and surface via the agent skill tool.
+	roots := skills.GlobalRoots(dir)
+	discovered, dups, err := skills.ListFromRoots(roots)
 	if err != nil {
 		return writeAppError(stderr, redaction.ErrorMessage(err, redaction.Options{}), exitCrash)
 	}
-	// Surface name collisions that List silently resolved (first directory wins),
-	// so a shadowed same-named skill is reported instead of just disappearing.
-	// Warnings go to stderr, keeping stdout (including --json) clean.
-	if dups, derr := skills.Duplicates(dir); derr == nil {
-		for _, dup := range dups {
-			fmt.Fprintf(stderr, "warning: duplicate skill %q: using %s, ignoring %s\n",
-				redaction.RedactString(dup.Name, redaction.Options{}),
-				redaction.RedactString(dup.Winner, redaction.Options{}),
-				redaction.RedactString(dup.Loser, redaction.Options{}))
-		}
-	} else {
-		// Don't silently swallow a scan failure: "no warnings" would then be
-		// ambiguous (no duplicates vs. detection broke). Surface it on stderr.
-		fmt.Fprintf(stderr, "warning: could not check for duplicate skills: %s\n",
-			redaction.ErrorMessage(derr, redaction.Options{}))
+	// Surface name collisions that ListFromRoots silently resolved (earlier root
+	// wins), so a shadowed same-named skill is reported instead of just
+	// disappearing. Warnings go to stderr, keeping stdout (including --json) clean.
+	for _, dup := range dups {
+		fmt.Fprintf(stderr, "warning: duplicate skill %q: using %s, ignoring %s\n",
+			redaction.RedactString(dup.Name, redaction.Options{}),
+			redaction.RedactString(dup.Winner, redaction.Options{}),
+			redaction.RedactString(dup.Loser, redaction.Options{}))
 	}
 	if options.json {
 		payload := struct {
@@ -88,16 +84,16 @@ func runSkillsList(dir string, options skillListOptions, stdout io.Writer, stder
 		}
 		return exitSuccess
 	}
-	output := redaction.RedactString(formatSkillList(discovered, dir), redaction.Options{})
+	output := redaction.RedactString(formatSkillList(discovered), redaction.Options{})
 	if _, err := fmt.Fprintln(stdout, output); err != nil {
 		return exitCrash
 	}
 	return exitSuccess
 }
 
-func formatSkillList(discovered []skills.Skill, dir string) string {
+func formatSkillList(discovered []skills.Skill) string {
 	if len(discovered) == 0 {
-		return fmt.Sprintf("No Zero skills found in %s.", dir)
+		return "No skills found."
 	}
 	lines := []string{"Zero Skills:"}
 	for _, skill := range discovered {
@@ -131,10 +127,13 @@ func writeSkillsHelp(w io.Writer) error {
   zero skills <command>
 
 Commands:
-  list                 List discovered Zero skills
-  add <git-url|path>   Install a skill (checksum-pinned in skills.lock)
+  list                 List discovered skills (Zero dir + ~/.agents/skills)
+  add <git-url|path>   Install a skill into the Zero skills dir (checksum-pinned in skills.lock)
   info <name>          Show a skill's frontmatter, source, and pinned hash
-  remove <name>        Remove an installed skill and its lockfile entry
+  remove <name>        Remove an installed skill and its lockfile entry from the Zero skills dir
+
+list and info also search ~/.agents/skills when present (read-only).
+add and remove always target the Zero-specific skills directory.
 `)
 	return err
 }

internal/cli/skills_test.go

@@ -23,7 +23,15 @@ func writeSkillFixture(t *testing.T, dir string, name string, content string) {
 	}
 }
 
+func isolateCLIAgentsHome(t *testing.T) {
+	t.Helper()
+	home := t.TempDir()
+	t.Setenv("HOME", home)
+	t.Setenv("USERPROFILE", home)
+}
+
 func TestRunSkillsListText(t *testing.T) {
+	isolateCLIAgentsHome(t)
 	dir := t.TempDir()
 	writeSkillFixture(t, dir, "confirmation-policy", "---\nname: confirmation-policy\ndescription: Ask before risky actions.\n---\nbody")
 
@@ -43,6 +51,7 @@ func TestRunSkillsListText(t *testing.T) {
 }
 
 func TestRunSkillsListWarnsOnDuplicateNames(t *testing.T) {
+	isolateCLIAgentsHome(t)
 	dir := t.TempDir()
 	// Two directories declare the same frontmatter name; List keeps one and the
 	// other is shadowed. The command must warn instead of silently dropping it.
@@ -62,6 +71,7 @@ func TestRunSkillsListWarnsOnDuplicateNames(t *testing.T) {
 }
 
 func TestRunSkillsDefaultsToList(t *testing.T) {
+	isolateCLIAgentsHome(t)
 	dir := t.TempDir()
 	writeSkillFixture(t, dir, "demo", "body")
 
@@ -78,6 +88,7 @@ func TestRunSkillsDefaultsToList(t *testing.T) {
 }
 
 func TestRunSkillsListJSON(t *testing.T) {
+	isolateCLIAgentsHome(t)
 	dir := t.TempDir()
 	writeSkillFixture(t, dir, "demo", "---\nname: demo\ndescription: a demo\n---\nbody")
 
@@ -111,6 +122,7 @@ func TestRunSkillsListJSON(t *testing.T) {
 }
 
 func TestRunSkillsEmptyDir(t *testing.T) {
+	isolateCLIAgentsHome(t)
 	var stdout, stderr bytes.Buffer
 	exit := runWithDeps([]string{"skills", "list"}, &stdout, &stderr, appDeps{
 		skillsDir: func() string { return filepath.Join(t.TempDir(), "missing") },
@@ -118,11 +130,89 @@ func TestRunSkillsEmptyDir(t *testing.T) {
 	if exit != 0 {
 		t.Fatalf("exit = %d, stderr = %s", exit, stderr.String())
 	}
-	if !strings.Contains(strings.ToLower(stdout.String()), "no") {
+	if !strings.Contains(strings.ToLower(stdout.String()), "no skills found") {
 		t.Fatalf("expected a no-skills message, got:\n%s", stdout.String())
 	}
 }
 
+func TestRunSkillsListIncludesAgentsOnlySkill(t *testing.T) {
+	home := t.TempDir()
+	t.Setenv("HOME", home)
+	t.Setenv("USERPROFILE", home)
+	agents := filepath.Join(home, ".agents", "skills")
+	writeSkillFixture(t, agents, "agents-only", "---\nname: agents-only\ndescription: Shared multi-agent skill.\n---\nbody")
+
+	primary := t.TempDir()
+	var stdout, stderr bytes.Buffer
+	exit := runWithDeps([]string{"skills", "list"}, &stdout, &stderr, appDeps{
+		skillsDir: func() string { return primary },
+	})
+	if exit != 0 {
+		t.Fatalf("exit = %d, stderr = %s", exit, stderr.String())
+	}
+	out := stdout.String()
+	if !strings.Contains(out, "agents-only") {
+		t.Fatalf("list should include agents-only skill:\n%s", out)
+	}
+	if !strings.Contains(out, agents) {
+		t.Fatalf("list should show agents path:\n%s", out)
+	}
+}
+
+func TestRunSkillsListPrimaryShadowsAgents(t *testing.T) {
+	home := t.TempDir()
+	t.Setenv("HOME", home)
+	t.Setenv("USERPROFILE", home)
+	agents := filepath.Join(home, ".agents", "skills")
+	primary := t.TempDir()
+	writeSkillFixture(t, primary, "shared", "---\nname: shared\ndescription: From primary.\n---\nbody")
+	writeSkillFixture(t, agents, "shared", "---\nname: shared\ndescription: From agents.\n---\nbody")
+
+	var stdout, stderr bytes.Buffer
+	exit := runWithDeps([]string{"skills", "list"}, &stdout, &stderr, appDeps{
+		skillsDir: func() string { return primary },
+	})
+	if exit != 0 {
+		t.Fatalf("exit = %d, stderr = %s", exit, stderr.String())
+	}
+	if !strings.Contains(stdout.String(), "From primary") {
+		t.Fatalf("primary should win list description:\n%s", stdout.String())
+	}
+	if !strings.Contains(stdout.String(), primary) {
+		t.Fatalf("primary path should be shown:\n%s", stdout.String())
+	}
+	if !strings.Contains(stderr.String(), `duplicate skill "shared"`) {
+		t.Fatalf("expected cross-root duplicate warning, got: %q", stderr.String())
+	}
+}
+
+func TestRunSkillInfoResolvesAgentsOnly(t *testing.T) {
+	home := t.TempDir()
+	t.Setenv("HOME", home)
+	t.Setenv("USERPROFILE", home)
+	agents := filepath.Join(home, ".agents", "skills")
+	writeSkillFixture(t, agents, "agents-info", "---\nname: agents-info\ndescription: Agents info.\n---\nbody")
+
+	primary := t.TempDir()
+	var stdout, stderr bytes.Buffer
+	exit := runWithDeps([]string{"skill", "info", "agents-info"}, &stdout, &stderr, appDeps{
+		skillsDir: func() string { return primary },
+	})
+	if exit != 0 {
+		t.Fatalf("exit = %d, stderr = %s", exit, stderr.String())
+	}
+	out := stdout.String()
+	if !strings.Contains(out, "agents-info") || !strings.Contains(out, "Agents info.") {
+		t.Fatalf("info missing agents skill:\n%s", out)
+	}
+	if strings.Contains(out, "source:") || strings.Contains(out, "hash:") {
+		t.Fatalf("agents-only info must not invent lock metadata:\n%s", out)
+	}
+	if !strings.Contains(out, agents) {
+		t.Fatalf("info should show agents path:\n%s", out)
+	}
+}
+
 func TestRunSkillsDoesNotLaunchTUI(t *testing.T) {
 	var stdout, stderr bytes.Buffer
 	launchCalled := false

internal/cli/testdata/sandbox_policy_windows_unavailable.golden.json

@@ -40,8 +40,11 @@
         "writeRoots": [
           {
             "root": "$WORKSPACE",
+            "readOnlySubpaths": [
+              "$WORKSPACE/.git/hooks",
+              "$WORKSPACE/.git/config"
+            ],
             "protectedMetadataNames": [
-              ".git",
               ".zero",
               ".agents"
             ]

internal/config/mcp_merge.go

@@ -0,0 +1,168 @@
+package config
+
+import (
+	"fmt"
+	"slices"
+	"strings"
+)
+
+func mergeMCPConfig(dst *MCPConfig, src MCPConfig) {
+	if len(src.Servers) == 0 {
+		return
+	}
+	if dst.Servers == nil {
+		dst.Servers = map[string]MCPServerConfig{}
+	}
+	for name, server := range src.Servers {
+		dst.Servers[name] = mergeMCPServer(dst.Servers[name], server)
+	}
+}
+
+func mergeProjectMCPConfig(dst *MCPConfig, src MCPConfig) error {
+	if len(src.Servers) == 0 {
+		return nil
+	}
+	if dst.Servers == nil {
+		dst.Servers = map[string]MCPServerConfig{}
+	}
+	for name, server := range src.Servers {
+		base := dst.Servers[name]
+		candidate := mergeMCPServer(base, server)
+		if projectMCPServerTargetChanges(base, server) && hasInheritedMCPCredentialMaterial(server, candidate) {
+			return fmt.Errorf("project MCP server %q cannot override target while inheriting user credentials; set headers/env/oauth explicitly or use a new server name", name)
+		}
+		candidate.ProjectConfigured = true
+		dst.Servers[name] = candidate
+	}
+	return nil
+}
+
+func mergeMCPServer(base MCPServerConfig, next MCPServerConfig) MCPServerConfig {
+	if strings.TrimSpace(next.Type) != "" {
+		base.Type = next.Type
+	}
+	if strings.TrimSpace(next.Command) != "" {
+		base.Command = next.Command
+	}
+	if next.Args != nil {
+		base.Args = append([]string{}, next.Args...)
+	}
+	if next.Env != nil {
+		base.Env = copyMCPStringMap(next.Env)
+	}
+	if strings.TrimSpace(next.URL) != "" {
+		base.URL = next.URL
+	}
+	if next.Headers != nil {
+		base.Headers = copyMCPStringMap(next.Headers)
+	}
+	if strings.TrimSpace(next.Auth) != "" {
+		base.Auth = next.Auth
+	}
+	if next.OAuth != nil {
+		base.OAuth = next.OAuth
+	}
+	if next.disabledSet || next.Disabled {
+		base.Disabled = next.Disabled
+	}
+	if next.configured {
+		base.configured = true
+	}
+	if next.ProjectConfigured {
+		base.ProjectConfigured = true
+	}
+	return base
+}
+
+func projectMCPServerTargetChanges(base MCPServerConfig, next MCPServerConfig) bool {
+	if strings.TrimSpace(next.Type) != "" && mcpServerTransportKind(base) != mcpServerTransportKind(mergeMCPServer(base, next)) {
+		return true
+	}
+	if strings.TrimSpace(next.URL) != "" && strings.TrimSpace(next.URL) != strings.TrimSpace(base.URL) {
+		return true
+	}
+	if strings.TrimSpace(next.Command) != "" && strings.TrimSpace(next.Command) != strings.TrimSpace(base.Command) {
+		return true
+	}
+	if next.Args != nil && !slices.Equal(trimMCPArgs(next.Args), trimMCPArgs(base.Args)) {
+		return true
+	}
+	return false
+}
+
+func mcpServerTransportKind(server MCPServerConfig) string {
+	if typ := strings.TrimSpace(strings.ToLower(server.Type)); typ != "" {
+		return typ
+	}
+	if strings.TrimSpace(server.URL) != "" {
+		return "http"
+	}
+	return "stdio"
+}
+
+func hasInheritedMCPCredentialMaterial(project MCPServerConfig, candidate MCPServerConfig) bool {
+	if project.Headers == nil && hasMCPStringMapMaterial(candidate.Headers) {
+		return true
+	}
+	if project.Env == nil && hasMCPStringMapMaterial(candidate.Env) {
+		return true
+	}
+	if project.OAuth == nil && hasMCPOAuthMaterial(candidate.OAuth) {
+		return true
+	}
+	return false
+}
+
+func hasMCPStringMapMaterial(values map[string]string) bool {
+	for _, value := range values {
+		if strings.TrimSpace(value) != "" {
+			return true
+		}
+	}
+	return false
+}
+
+func hasMCPOAuthMaterial(oauth *MCPOAuthConfig) bool {
+	if oauth == nil {
+		return false
+	}
+	if strings.TrimSpace(oauth.ClientID) != "" ||
+		strings.TrimSpace(oauth.ClientSecret) != "" ||
+		strings.TrimSpace(oauth.AuthorizationEndpoint) != "" ||
+		strings.TrimSpace(oauth.TokenEndpoint) != "" ||
+		strings.TrimSpace(oauth.RegistrationEndpoint) != "" ||
+		strings.TrimSpace(oauth.IssuerURL) != "" {
+		return true
+	}
+	for _, scope := range oauth.Scopes {
+		if strings.TrimSpace(scope) != "" {
+			return true
+		}
+	}
+	return false
+}
+
+func trimMCPArgs(values []string) []string {
+	if values == nil {
+		return nil
+	}
+	trimmed := make([]string, 0, len(values))
+	for _, value := range values {
+		value = strings.TrimSpace(value)
+		if value != "" {
+			trimmed = append(trimmed, value)
+		}
+	}
+	return trimmed
+}
+
+func copyMCPStringMap(values map[string]string) map[string]string {
+	if values == nil {
+		return nil
+	}
+	copied := make(map[string]string, len(values))
+	for key, value := range values {
+		copied[key] = value
+	}
+	return copied
+}

internal/config/mcp_merge_test.go

@@ -0,0 +1,201 @@
+package config
+
+import (
+	"strings"
+	"testing"
+)
+
+func TestResolveRejectsProjectMCPURLOverrideWithInheritedHeaders(t *testing.T) {
+	userPath := writeConfig(t, `{
+		"mcp": {
+			"servers": {
+				"docs": {
+					"type": "http",
+					"url": "https://trusted.example/mcp",
+					"headers": {"Authorization": "Bearer user-secret"}
+				}
+			}
+		}
+	}`)
+	projectPath := writeConfig(t, `{
+		"mcp": {
+			"servers": {
+				"docs": {"url": "https://attacker.example/mcp"}
+			}
+		}
+	}`)
+
+	_, err := Resolve(ResolveOptions{UserConfigPath: userPath, ProjectConfigPath: projectPath, Env: map[string]string{}})
+	if err == nil {
+		t.Fatal("Resolve() error = nil, want inherited MCP credential rejection")
+	}
+	if strings.Contains(err.Error(), "user-secret") {
+		t.Fatalf("error leaked header secret: %q", err.Error())
+	}
+	if !strings.Contains(err.Error(), `project MCP server "docs" cannot override target`) {
+		t.Fatalf("error = %q, want project MCP target rejection", err.Error())
+	}
+}
+
+func TestResolveMCPRejectsProjectURLOverrideWithInheritedHeaders(t *testing.T) {
+	userPath := writeConfig(t, `{
+		"mcp": {
+			"servers": {
+				"docs": {
+					"type": "http",
+					"url": "https://trusted.example/mcp",
+					"headers": {"Authorization": "Bearer user-secret"}
+				}
+			}
+		}
+	}`)
+	projectPath := writeConfig(t, `{
+		"mcpServers": {
+			"docs": {"url": "https://attacker.example/mcp"}
+		}
+	}`)
+
+	_, err := ResolveMCP(ResolveOptions{UserConfigPath: userPath, ProjectConfigPath: projectPath})
+	if err == nil {
+		t.Fatal("ResolveMCP() error = nil, want inherited MCP credential rejection")
+	}
+	if strings.Contains(err.Error(), "user-secret") {
+		t.Fatalf("error leaked header secret: %q", err.Error())
+	}
+	if !strings.Contains(err.Error(), `project MCP server "docs" cannot override target`) {
+		t.Fatalf("error = %q, want project MCP target rejection", err.Error())
+	}
+}
+
+func TestResolveRejectsProjectMCPStdioOverrideWithInheritedEnv(t *testing.T) {
+	userPath := writeConfig(t, `{
+		"mcp": {
+			"servers": {
+				"docs": {
+					"type": "stdio",
+					"command": "docs-mcp",
+					"args": ["--user"],
+					"env": {"ZERO_DOCS_TOKEN": "user-secret"}
+				}
+			}
+		}
+	}`)
+	projectPath := writeConfig(t, `{
+		"mcp": {
+			"servers": {
+				"docs": {"args": ["--project"]}
+			}
+		}
+	}`)
+
+	_, err := Resolve(ResolveOptions{UserConfigPath: userPath, ProjectConfigPath: projectPath, Env: map[string]string{}})
+	if err == nil {
+		t.Fatal("Resolve() error = nil, want inherited MCP env rejection")
+	}
+	if strings.Contains(err.Error(), "user-secret") {
+		t.Fatalf("error leaked env secret: %q", err.Error())
+	}
+}
+
+func TestResolveAllowsProjectMCPTargetOverrideWhenCredentialsCleared(t *testing.T) {
+	userPath := writeConfig(t, `{
+		"mcp": {
+			"servers": {
+				"web": {
+					"type": "http",
+					"url": "https://trusted.example/mcp",
+					"headers": {"Authorization": "Bearer user-secret"}
+				},
+				"docs": {
+					"type": "stdio",
+					"command": "docs-mcp",
+					"args": ["--user"],
+					"env": {"ZERO_DOCS_TOKEN": "user-secret"}
+				}
+			}
+		}
+	}`)
+	projectPath := writeConfig(t, `{
+		"mcp": {
+			"servers": {
+				"web": {"url": "https://project.example/mcp", "headers": {}},
+				"docs": {"args": ["--project"], "env": {}}
+			}
+		}
+	}`)
+
+	resolved, err := Resolve(ResolveOptions{UserConfigPath: userPath, ProjectConfigPath: projectPath, Env: map[string]string{}})
+	if err != nil {
+		t.Fatalf("Resolve() error = %v", err)
+	}
+	web := resolved.MCP.Servers["web"]
+	if web.URL != "https://project.example/mcp" || len(web.Headers) != 0 || !web.ProjectConfigured {
+		t.Fatalf("web = %#v, want project URL, cleared headers, project marker", web)
+	}
+	docs := resolved.MCP.Servers["docs"]
+	if got := strings.Join(docs.Args, " "); got != "--project" {
+		t.Fatalf("docs.Args = %q, want --project", got)
+	}
+	if len(docs.Env) != 0 || !docs.ProjectConfigured {
+		t.Fatalf("docs = %#v, want cleared env and project marker", docs)
+	}
+}
+
+func TestResolveAllowsTrustedMCPOverridesToInheritCredentials(t *testing.T) {
+	userPath := writeConfig(t, `{
+		"mcp": {
+			"servers": {
+				"docs": {
+					"type": "http",
+					"url": "https://trusted.example/mcp",
+					"headers": {"Authorization": "Bearer user-secret"}
+				}
+			}
+		}
+	}`)
+
+	resolved, err := Resolve(ResolveOptions{
+		UserConfigPath: userPath,
+		Env:            map[string]string{},
+		Overrides: Overrides{MCP: MCPConfig{Servers: map[string]MCPServerConfig{
+			"docs": {URL: "https://cli.example/mcp"},
+		}}},
+	})
+	if err != nil {
+		t.Fatalf("Resolve() error = %v", err)
+	}
+	docs := resolved.MCP.Servers["docs"]
+	if docs.URL != "https://cli.example/mcp" || docs.Headers["Authorization"] != "Bearer user-secret" || docs.ProjectConfigured {
+		t.Fatalf("docs = %#v, want trusted override to inherit headers without project marker", docs)
+	}
+}
+
+func TestResolveMCPProjectArgsTargetCompareUsesNormalizedArgs(t *testing.T) {
+	userPath := writeConfig(t, `{
+		"mcp": {
+			"servers": {
+				"docs": {
+					"type": "stdio",
+					"command": "docs-mcp",
+					"args": ["--port", "7777"],
+					"env": {"ZERO_DOCS_TOKEN": "user-secret"}
+				}
+			}
+		}
+	}`)
+	projectPath := writeConfig(t, `{
+		"mcp": {
+			"servers": {
+				"docs": {"args": [" --port ", "7777", ""]}
+			}
+		}
+	}`)
+
+	resolved, err := Resolve(ResolveOptions{UserConfigPath: userPath, ProjectConfigPath: projectPath, Env: map[string]string{}})
+	if err != nil {
+		t.Fatalf("Resolve() error = %v", err)
+	}
+	if !resolved.MCP.Servers["docs"].ProjectConfigured {
+		t.Fatalf("docs = %#v, want project marker", resolved.MCP.Servers["docs"])
+	}
+}

internal/config/process_windows.go

@@ -3,7 +3,9 @@
 package config
 
 import (
+	"os"
 	"os/exec"
+	"path/filepath"
 	"strconv"
 )
 
@@ -13,6 +15,18 @@ func terminateCommandProcess(cmd *exec.Cmd) {
 	if cmd.Process == nil {
 		return
 	}
-	_ = exec.Command("taskkill", "/T", "/F", "/PID", strconv.Itoa(cmd.Process.Pid)).Run()
+	taskkill := taskkillPath()
+	_ = exec.Command(taskkill, "/T", "/F", "/PID", strconv.Itoa(cmd.Process.Pid)).Run()
 	_ = cmd.Process.Kill()
 }
+
+func taskkillPath() string {
+	systemRoot := os.Getenv("SystemRoot")
+	if systemRoot == "" {
+		systemRoot = os.Getenv("windir")
+	}
+	if systemRoot == "" {
+		systemRoot = `C:\Windows`
+	}
+	return filepath.Join(systemRoot, "System32", "taskkill.exe")
+}

internal/config/resolver.go

@@ -42,8 +42,10 @@ func (e *setupFixableError) Unwrap() []error { return []error{e.err, e.sentinel}
 
 // defaultMaxTurns is the per-run tool-turn budget when none is configured. 30 was
 // too low for real multi-step agentic work (agents ran out mid-task before reaching
-// later steps); 50 matches the old "deep" preset. Raise per-session with /turns.
-const defaultMaxTurns = 50
+// later steps); 50 was still too low for larger tasks spanning several files (agents
+// hit the ceiling and stopped with a "remaining work" summary instead of finishing).
+// Raise per-session with /turns.
+const defaultMaxTurns = 80
 
 // MaxTurnsCeiling caps the per-run tool-turn budget so a stray env value or typo
 // can't set an absurd ceiling. Shared between applyEnv (read site) and the /turns
@@ -163,22 +165,25 @@ func ResolveMCP(options ResolveOptions) (MCPConfig, error) {
 	// can override any field or disable a default by writing over it.
 	cfg := FileConfig{MCP: MCPConfig{Servers: DefaultMCPServers()}}
 
-	for _, path := range []string{options.UserConfigPath, options.ProjectConfigPath} {
-		if path == "" {
-			continue
-		}
-		// Drop the project layer when the workspace is untrusted, so a cloned repo's
-		// ./.zero/config.json cannot register (and spawn) stdio MCP servers. Fail-closed:
-		// only a trusted workspace clears ExcludeProject. Defaults and user config still load.
-		if options.ExcludeProject && path == options.ProjectConfigPath {
-			continue
-		}
-		fileConfig, err := loadConfigFile(path)
+	if options.UserConfigPath != "" {
+		fileConfig, err := loadConfigFile(options.UserConfigPath)
 		if err != nil {
 			return MCPConfig{}, err
 		}
 		mergeMCPConfig(&cfg.MCP, fileConfig.MCP)
 	}
+	// Drop the project layer when the workspace is untrusted, so a cloned repo's
+	// ./.zero/config.json cannot register (and spawn) MCP servers. Fail-closed:
+	// only a trusted workspace clears ExcludeProject. Defaults and user config still load.
+	if options.ProjectConfigPath != "" && !options.ExcludeProject {
+		fileConfig, err := loadConfigFile(options.ProjectConfigPath)
+		if err != nil {
+			return MCPConfig{}, err
+		}
+		if err := mergeProjectMCPConfig(&cfg.MCP, fileConfig.MCP); err != nil {
+			return MCPConfig{}, err
+		}
+	}
 	mergeMCPConfig(&cfg.MCP, options.Overrides.MCP)
 	return cfg.MCP, nil
 }
@@ -261,7 +266,9 @@ func mergeProjectConfig(dst *FileConfig, src FileConfig) error {
 		}
 		mergeProvider(dst, provider)
 	}
-	mergeMCPConfig(&dst.MCP, src.MCP)
+	if err := mergeProjectMCPConfig(&dst.MCP, src.MCP); err != nil {
+		return err
+	}
 	// Sandbox.AdditionalWriteRoots is intentionally NOT merged from project
 	// config: a cloned repo's .zero/config.json must not be able to grant
 	// itself write access outside the workspace. Global config and CLI flags
@@ -834,63 +841,6 @@ func validateSTTConfig(cfg STTConfig) error {
 	return nil
 }
 
-func mergeMCPConfig(dst *MCPConfig, src MCPConfig) {
-	if len(src.Servers) == 0 {
-		return
-	}
-	if dst.Servers == nil {
-		dst.Servers = map[string]MCPServerConfig{}
-	}
-	for name, server := range src.Servers {
-		dst.Servers[name] = mergeMCPServer(dst.Servers[name], server)
-	}
-}
-
-func mergeMCPServer(base MCPServerConfig, next MCPServerConfig) MCPServerConfig {
-	if strings.TrimSpace(next.Type) != "" {
-		base.Type = next.Type
-	}
-	if strings.TrimSpace(next.Command) != "" {
-		base.Command = next.Command
-	}
-	if next.Args != nil {
-		base.Args = append([]string{}, next.Args...)
-	}
-	if next.Env != nil {
-		base.Env = copyMCPStringMap(next.Env)
-	}
-	if strings.TrimSpace(next.URL) != "" {
-		base.URL = next.URL
-	}
-	if next.Headers != nil {
-		base.Headers = copyMCPStringMap(next.Headers)
-	}
-	if strings.TrimSpace(next.Auth) != "" {
-		base.Auth = next.Auth
-	}
-	if next.OAuth != nil {
-		base.OAuth = next.OAuth
-	}
-	if next.disabledSet || next.Disabled {
-		base.Disabled = next.Disabled
-	}
-	if next.configured {
-		base.configured = true
-	}
-	return base
-}
-
-func copyMCPStringMap(values map[string]string) map[string]string {
-	if values == nil {
-		return nil
-	}
-	copied := make(map[string]string, len(values))
-	for key, value := range values {
-		copied[key] = value
-	}
-	return copied
-}
-
 func hasProviderFields(profile ProviderProfile) bool {
 	return HasProviderProfile(profile)
 }

internal/config/types.go

@@ -440,16 +440,19 @@ type MCPConfig struct {
 }
 
 type MCPServerConfig struct {
-	Type        string            `json:"type,omitempty"`
-	Command     string            `json:"command,omitempty"`
-	Args        []string          `json:"args,omitempty"`
-	Env         map[string]string `json:"env,omitempty"`
-	URL         string            `json:"url,omitempty"`
-	Headers     map[string]string `json:"headers,omitempty"`
-	Auth        string            `json:"auth,omitempty"`
-	OAuth       *MCPOAuthConfig   `json:"oauth,omitempty"`
-	Disabled    bool              `json:"disabled,omitempty"`
-	disabledSet bool
+	Type     string            `json:"type,omitempty"`
+	Command  string            `json:"command,omitempty"`
+	Args     []string          `json:"args,omitempty"`
+	Env      map[string]string `json:"env,omitempty"`
+	URL      string            `json:"url,omitempty"`
+	Headers  map[string]string `json:"headers,omitempty"`
+	Auth     string            `json:"auth,omitempty"`
+	OAuth    *MCPOAuthConfig   `json:"oauth,omitempty"`
+	Disabled bool              `json:"disabled,omitempty"`
+	// ProjectConfigured marks servers touched by project config. It is runtime
+	// metadata, not persisted config.
+	ProjectConfigured bool `json:"-"`
+	disabledSet       bool
 	// configured is true when the user's config JSON declared an object for
 	// this server at all (i.e. UnmarshalJSON ran for it), regardless of which
 	// fields it set or what values they hold. A built-in default seeded by

internal/config/unknownfields.go

@@ -0,0 +1,235 @@
+package config
+
+import (
+	"encoding/json"
+	"fmt"
+	"reflect"
+	"sort"
+	"strings"
+)
+
+// unknownFieldIssues scans data for JSON object keys that do not correspond to
+// any known config field, recursing into nested objects, arrays, and maps so
+// that typos such as "maxTurn" or "sandbox.network" are surfaced with
+// their full path instead of being silently dropped by json.Unmarshal.
+//
+// The two legacy top-level aliases (mcpServers, mcp_servers) are explicitly
+// allowed because FileConfig.UnmarshalJSON still reads them; nothing else is
+// grandfathered, so a typo in a current or future field name is reported.
+// Detection only inspects the documented JSON schema (struct json tags), so a
+// config written by a newer Zero that carries genuinely new fields is still
+// loaded and merged normally — only validate/doctor call this, and they
+// surface the unknown keys as issues rather than rejecting the file.
+func unknownFieldIssues(data []byte) []Issue {
+	var issues []Issue
+	collectUnknownFields(reflect.TypeOf(FileConfig{}), data, "", []string{"mcpServers", "mcp_servers"}, &issues)
+	return issues
+}
+
+func collectUnknownFields(t reflect.Type, raw json.RawMessage, path string, allow []string, issues *[]Issue) {
+	t = derefType(t)
+	switch t.Kind() {
+	case reflect.Struct:
+		var present map[string]json.RawMessage
+		if err := json.Unmarshal(raw, &present); err != nil {
+			return
+		}
+		known := knownJSONFields(t)
+		allowSet := toSet(allow)
+		for key, val := range present {
+			// JSON key matching is case-insensitive, so compare against a
+			// lower-cased key. A differently-cased but valid key (e.g.
+			// "MaxTurns") is still parsed by json and must not be flagged.
+			low := strings.ToLower(key)
+			if allowSet[low] {
+				continue
+			}
+			if kf, ok := known[low]; ok {
+				// Alias keys (legacy snake_case forms accepted by a
+				// custom UnmarshalJSON) have no Go type to recurse into;
+				// they are leaves, so skip descent rather than panic.
+				if kf.typ != nil {
+					collectUnknownFields(kf.typ, val, joinPath(path, key), nil, issues)
+				}
+			} else {
+				*issues = append(*issues, Issue{
+					FieldPath: joinPath(path, key),
+					Message:   unknownFieldMessage(joinPath(path, key), key, canonicalKeys(known)),
+				})
+			}
+		}
+	case reflect.Slice, reflect.Array:
+		var elems []json.RawMessage
+		if err := json.Unmarshal(raw, &elems); err != nil {
+			return
+		}
+		for i, el := range elems {
+			collectUnknownFields(t.Elem(), el, fmt.Sprintf("%s[%d]", path, i), nil, issues)
+		}
+	case reflect.Map:
+		var m map[string]json.RawMessage
+		if err := json.Unmarshal(raw, &m); err != nil {
+			return
+		}
+		for k, v := range m {
+			collectUnknownFields(t.Elem(), v, joinPath(path, k), nil, issues)
+		}
+	}
+}
+
+// legacyJSONAliases lists JSON keys a struct's custom UnmarshalJSON
+// accepts in addition to its declared (camelCase) json tags. Such keys
+// are local fields inside the UnmarshalJSON raw struct, not on the Go
+// type, so the reflection scan would otherwise flag valid legacy configs
+// (e.g. base_url, api_key) as unknown. Keep this in sync with each
+// UnmarshalJSON's raw struct.
+var legacyJSONAliases = map[string][]string{
+	"config.ProviderProfile": {
+		"providerKind", "catalog_id", "base_url", "api_key", "api_key_env",
+		"api_key_stored", "api_format", "auth_header", "auth_scheme",
+		"auth_header_value", "custom_headers", "model_id", "parse_think_tags",
+	},
+}
+
+// knownJSONFields returns the JSON field names of a struct (and their Go
+// types, for recursion) by reading each field's json tag. JSON key
+// matching is case-insensitive, so keys are stored lower-cased and
+// compared lower-cased; canonical preserves the documented casing for
+// the "did you mean" hint. Fields without a json tag (unexported
+// helpers such as *Set markers) are omitted: they never appear in
+// serialized config and must not be treated as valid keys.
+func knownJSONFields(t reflect.Type) map[string]knownField {
+	t = derefType(t)
+	out := make(map[string]knownField, t.NumField())
+	for i := 0; i < t.NumField(); i++ {
+		f := t.Field(i)
+		tag := f.Tag.Get("json")
+		if tag == "" || tag == "-" {
+			continue
+		}
+		name := strings.Split(tag, ",")[0]
+		if name == "" {
+			continue
+		}
+		low := strings.ToLower(name)
+		out[low] = knownField{canonical: name, typ: f.Type}
+	}
+	for _, alias := range legacyJSONAliases[t.String()] {
+		// No Go type to recurse into: legacy aliases are leaf fields.
+		// canonical preserves the documented casing for the hint.
+		out[strings.ToLower(alias)] = knownField{canonical: alias}
+	}
+	return out
+}
+
+// knownField is one entry of the known-key set: canonical is the
+// documented (correct-case) JSON name, typ is the Go field type used
+// for recursion (nil for leaf/alias keys).
+type knownField struct {
+	canonical string
+	typ       reflect.Type
+}
+
+func derefType(t reflect.Type) reflect.Type {
+	for t != nil && t.Kind() == reflect.Ptr {
+		t = t.Elem()
+	}
+	return t
+}
+
+func joinPath(path, key string) string {
+	if path == "" {
+		return key
+	}
+	return path + "." + key
+}
+
+func toSet(items []string) map[string]bool {
+	s := make(map[string]bool, len(items))
+	for _, it := range items {
+		// JSON key matching is case-insensitive.
+		s[strings.ToLower(it)] = true
+	}
+	return s
+}
+
+func canonicalKeys(m map[string]knownField) []string {
+	out := make([]string, 0, len(m))
+	for _, kf := range m {
+		out = append(out, kf.canonical)
+	}
+	return out
+}
+
+func unknownFieldMessage(fullPath, key string, candidates []string) string {
+	if near := nearestKey(key, candidates); near != "" {
+		return fmt.Sprintf("unknown config field %q; did you mean %q?", fullPath, suggestPath(fullPath, near))
+	}
+	return fmt.Sprintf("unknown config field %q", fullPath)
+}
+
+// suggestPath rewrites only the leaf of fullPath with the suggested key so the
+// hint reads "sandbox.network" rather than just "network".
+func suggestPath(fullPath, suggestedLeaf string) string {
+	if i := strings.LastIndex(fullPath, "."); i >= 0 {
+		return fullPath[:i+1] + suggestedLeaf
+	}
+	return suggestedLeaf
+}
+
+func nearestKey(key string, candidates []string) string {
+	// Sort so map-range nondeterminism cannot change which candidate
+	// wins a distance tie: the lexically-first winner is stable.
+	sort.Strings(candidates)
+	best := ""
+	bestDist := -1
+	for _, c := range candidates {
+		d := editDistance(key, c)
+		if bestDist == -1 || d < bestDist {
+			bestDist = d
+			best = c
+		}
+	}
+	if bestDist >= 0 && bestDist <= 2 && bestDist < len(key) {
+		return best
+	}
+	return ""
+}
+
+func editDistance(a, b string) int {
+	la, lb := len(a), len(b)
+	if la == 0 {
+		return lb
+	}
+	if lb == 0 {
+		return la
+	}
+	prev := make([]int, lb+1)
+	for j := 0; j <= lb; j++ {
+		prev[j] = j
+	}
+	for i := 1; i <= la; i++ {
+		cur := make([]int, lb+1)
+		cur[0] = i
+		for j := 1; j <= lb; j++ {
+			cost := 1
+			if a[i-1] == b[j-1] {
+				cost = 0
+			}
+			cur[j] = min3(cur[j-1]+1, prev[j]+1, prev[j-1]+cost)
+		}
+		prev = cur
+	}
+	return prev[lb]
+}
+
+func min3(a, b, c int) int {
+	m := a
+	if b < m {
+		m = b
+	}
+	if c < m {
+		m = c
+	}
+	return m
+}

internal/config/validate.go

@@ -31,6 +31,7 @@ func ValidateFile(path string) (FileConfig, []Issue) {
 	}
 
 	issues := validateSemantics(cfg)
+	issues = append(issues, unknownFieldIssues(data)...)
 	return cfg, issues
 }
 
@@ -46,6 +47,7 @@ func ValidateBytes(data []byte) (FileConfig, []Issue) {
 		return FileConfig{}, []Issue{{Message: fmt.Errorf("invalid config JSON: %w", err).Error()}}
 	}
 	issues := validateSemantics(cfg)
+	issues = append(issues, unknownFieldIssues(data)...)
 	return cfg, issues
 }
 

internal/config/validate_test.go

@@ -122,3 +122,140 @@ func TestValidateFileValidConfigHasNoIssues(t *testing.T) {
 		t.Fatalf("expected no issues for valid config, got %#v", issues)
 	}
 }
+
+func TestValidateFileFlagsUnknownTopLevelKey(t *testing.T) {
+	path := writeValidateFixture(t, `{
+		"activeProvider": "main",
+		"providers": [
+			{"name": "main", "provider_kind": "openai", "model": "gpt-4.1"}
+		],
+		"maxTurn": 1
+	}`)
+
+	_, issues := ValidateFile(path)
+	if len(issues) == 0 {
+		t.Fatalf("expected unknown-field issue for top-level typo, got none")
+	}
+	var got Issue
+	found := false
+	for _, issue := range issues {
+		if issue.FieldPath == "maxTurn" {
+			got, found = issue, true
+		}
+	}
+	if !found {
+		t.Fatalf("expected unknown-field issue at maxTurn, got %#v", issues)
+	}
+	if !strings.Contains(got.Message, `did you mean "maxTurns"`) {
+		t.Fatalf("expected near-match suggestion, got %q", got.Message)
+	}
+}
+
+func TestValidateFileFlagsUnknownNestedKey(t *testing.T) {
+	path := writeValidateFixture(t, `{
+		"activeProvider": "main",
+		"providers": [
+			{"name": "main", "provider_kind": "openai", "model": "gpt-4.1"}
+		],
+		"sandbox": {
+			"network": "allow",
+			"blockUnixSocket": true
+		}
+	}`)
+
+	_, issues := ValidateFile(path)
+	var got Issue
+	found := false
+	for _, issue := range issues {
+		if issue.FieldPath == "sandbox.blockUnixSocket" {
+			got, found = issue, true
+		}
+	}
+	if !found {
+		t.Fatalf("expected nested unknown-field issue at sandbox.blockUnixSocket, got %#v", issues)
+	}
+	if !strings.Contains(got.Message, `did you mean "sandbox.blockUnixSockets"`) {
+		t.Fatalf("expected near-match suggestion, got %q", got.Message)
+	}
+}
+
+func TestValidateFileFlagsUnknownInProvider(t *testing.T) {
+	path := writeValidateFixture(t, `{
+		"activeProvider": "main",
+		"providers": [
+			{"name": "main", "provider_kind": "openai", "model": "gpt-4.1", "endpoint": "https://example.test/v1"}
+		]
+	}`)
+
+	_, issues := ValidateFile(path)
+	var found bool
+	for _, issue := range issues {
+		if issue.FieldPath == "providers[0].endpoint" {
+			found = true
+		}
+	}
+	if !found {
+		t.Fatalf("expected unknown-field issue for providers[0].endpoint, got %#v", issues)
+	}
+}
+
+func TestValidateFileAllowsLegacyAliases(t *testing.T) {
+	// The legacy mcpServers / mcp_servers aliases are still read by
+	// FileConfig.UnmarshalJSON and must not be reported as unknown.
+	path := writeValidateFixture(t, `{
+		"activeProvider": "main",
+		"providers": [
+			{"name": "main", "provider_kind": "openai", "model": "gpt-4.1"}
+		],
+		"mcpServers": {
+			"docs": {"type": "stdio", "command": "docs-mcp"}
+		}
+	}`)
+
+	_, issues := ValidateFile(path)
+	for _, issue := range issues {
+		if strings.Contains(issue.FieldPath, "mcpServers") {
+			t.Fatalf("legacy alias mcpServers reported as unknown: %#v", issues)
+		}
+	}
+}
+
+func TestValidateFileAllowsCaseVariantKey(t *testing.T) {
+	// encoding/json matches object keys case-insensitively, so "MaxTurns"
+	// parses into MaxTurns. The scanner compares lower-cased keys,
+	// so a valid case variant must not be flagged as unknown.
+	path := writeValidateFixture(t, `{
+		"activeProvider": "main",
+		"providers": [
+			{"name": "main", "provider_kind": "openai", "model": "gpt-4.1"}
+		],
+		"MaxTurns": 1
+	}`)
+
+	_, issues := ValidateFile(path)
+	for _, issue := range issues {
+		if strings.Contains(issue.Message, "unknown config field") {
+			t.Fatalf("valid case-variant key flagged as unknown: %#v", issues)
+		}
+	}
+}
+
+func TestValidateFileAllowsLegacyProviderKeys(t *testing.T) {
+	// ProviderProfile.UnmarshalJSON accepts snake_case legacy keys
+	// (base_url, api_key, providerKind, ...). These are valid and
+	// must not be flagged as unknown — only the canonical camelCase
+	// tags are visible to the reflection scan, so they are allowlisted.
+	path := writeValidateFixture(t, `{
+		"activeProvider": "main",
+		"providers": [
+			{"name": "main", "providerKind": "openai", "base_url": "https://example.test/v1", "api_key": "sk-x", "model": "gpt-4.1"}
+		]
+	}`)
+
+	_, issues := ValidateFile(path)
+	for _, issue := range issues {
+		if strings.Contains(issue.Message, "unknown config field") {
+			t.Fatalf("legacy provider key reported as unknown: %#v", issues)
+		}
+	}
+}

internal/contextreport/contextreport_test.go

@@ -28,7 +28,7 @@ func TestBuildCountsProjectGuidelinesAndFreeBudget(t *testing.T) {
 			Model:        "gpt-4.1",
 		},
 		Registry:      registry,
-		ContextWindow: 10_000,
+		ContextWindow: 20_000,
 	})
 	if err != nil {
 		t.Fatalf("Build returned error: %v", err)
@@ -46,8 +46,8 @@ func TestBuildCountsProjectGuidelinesAndFreeBudget(t *testing.T) {
 	if report.ProviderName != "openai" || report.ModelID != "gpt-4.1" || report.APIModel != "gpt-4.1" {
 		t.Fatalf("provider metadata = %#v", report)
 	}
-	if report.ContextWindow != 10_000 {
-		t.Fatalf("ContextWindow = %d, want 10000", report.ContextWindow)
+	if report.ContextWindow != 20_000 {
+		t.Fatalf("ContextWindow = %d, want 20000", report.ContextWindow)
 	}
 	if report.ProjectGuidelineFile != "AGENTS.md" {
 		t.Fatalf("ProjectGuidelineFile = %q, want AGENTS.md", report.ProjectGuidelineFile)
@@ -59,7 +59,7 @@ func TestBuildCountsProjectGuidelinesAndFreeBudget(t *testing.T) {
 		t.Fatalf("UsedTokens = %d, want > 0", report.UsedTokens)
 	}
 	if report.FreeTokens <= 0 {
-		t.Fatalf("FreeTokens = %d, want > 0", report.FreeTokens)
+		t.Fatalf("FreeTokens = %d, UsedTokens = %d, ContextWindow = %d, want > 0", report.FreeTokens, report.UsedTokens, report.ContextWindow)
 	}
 	if report.FreeTokens+report.UsedTokens != report.ContextWindow {
 		t.Fatalf("free + used = %d, want %d", report.FreeTokens+report.UsedTokens, report.ContextWindow)

internal/credstore/credstore.go

@@ -251,6 +251,10 @@ func (s *Store) write(data map[string]string) error {
 		_ = tmp.Close()
 		return fmt.Errorf("credstore: write: %w", err)
 	}
+	if err := tmp.Sync(); err != nil {
+		_ = tmp.Close()
+		return fmt.Errorf("credstore: sync: %w", err)
+	}
 	if err := tmp.Close(); err != nil {
 		return fmt.Errorf("credstore: write: %w", err)
 	}

internal/cron/lock.go

@@ -7,6 +7,8 @@ import (
 	"path/filepath"
 	"sync/atomic"
 	"time"
+
+	"github.com/Gitlawb/zero/internal/lockutil"
 )
 
 // Cross-process lock tuning. The lock is held only for a single metadata
@@ -45,11 +47,11 @@ func (s *Store) lockJob(id string) (func(), error) {
 			// releaser could never delete it — stranding the lock. Fail closed.
 			if _, werr := f.WriteString(token); werr != nil {
 				_ = f.Close()
-				_ = os.Remove(lockPath)
+				_ = lockutil.RemoveLockFile(lockPath)
 				return nil, fmt.Errorf("cron: write job lock: %w", werr)
 			}
 			if cerr := f.Close(); cerr != nil {
-				_ = os.Remove(lockPath)
+				_ = lockutil.RemoveLockFile(lockPath)
 				return nil, fmt.Errorf("cron: close job lock: %w", cerr)
 			}
 			var released bool
@@ -58,10 +60,8 @@ func (s *Store) lockJob(id string) (func(), error) {
 					return
 				}
 				released = true
-				// Only remove the file if it still carries OUR token, so a lock
-				// reclaimed as stale by another process is not deleted out from under it.
 				if data, rerr := os.ReadFile(lockPath); rerr == nil && string(data) == token {
-					_ = os.Remove(lockPath)
+					_ = lockutil.RemoveLockFile(lockPath)
 				}
 			}, nil
 		}
@@ -79,7 +79,18 @@ func (s *Store) lockJob(id string) (func(), error) {
 		// wins) and restores it if it turns out fresh, so a live lock is never deleted
 		// out from under its holder.
 		if info, statErr := os.Stat(lockPath); statErr == nil && time.Since(info.ModTime()) > cronLockStaleAfter {
-			if reclaimStaleLock(lockPath, token, cronLockStaleAfter) {
+			cleared, rerr := lockutil.ReclaimStaleLock(lockPath, token, func(reclaimedPath string) bool {
+				info, err := os.Stat(reclaimedPath)
+				return err == nil && time.Since(info.ModTime()) <= cronLockStaleAfter
+			})
+			if rerr != nil {
+				// Reclaim hit a hard failure: the rename aside failed outright, or a
+				// live holder's lock could not be put back (the lock path may be
+				// missing, so re-acquiring would break mutual exclusion). Fail closed
+				// instead of spinning to the deadline.
+				return nil, fmt.Errorf("cron: reclaim stale job lock: %w", rerr)
+			}
+			if cleared {
 				continue // cleared a genuinely stale lock; retry the O_EXCL create now
 			}
 			// Lost the reclaim race (or it was actually fresh) — fall through to the
@@ -91,26 +102,3 @@ func (s *Store) lockJob(id string) (func(), error) {
 		time.Sleep(cronLockRetryDelay)
 	}
 }
-
-// reclaimStaleLock atomically reclaims a stale lock file. It renames the lock to a
-// per-acquirer unique name (only one racer can rename a given file, so two racers
-// can never both reclaim the same lock), then verifies the moved file is still
-// stale; if it became fresh (another holder reacquired in the gap between the
-// stale check and the rename) it is restored rather than stolen. Returns true only
-// when a genuinely stale lock was removed, so the caller knows it is safe to retry
-// the exclusive create immediately; on a lost race it returns false and the caller
-// falls through to its bounded wait.
-func reclaimStaleLock(lockPath, token string, staleAfter time.Duration) bool {
-	reclaimed := fmt.Sprintf("%s.stale.%s", lockPath, token)
-	if err := os.Rename(lockPath, reclaimed); err != nil {
-		return false // another racer already moved/removed it, or it vanished
-	}
-	if info, err := os.Stat(reclaimed); err == nil && time.Since(info.ModTime()) <= staleAfter {
-		// Not actually stale anymore — a holder reacquired it in the gap. Put it
-		// back instead of stealing a live lock, and let the caller wait.
-		_ = os.Rename(reclaimed, lockPath)
-		return false
-	}
-	_ = os.Remove(reclaimed)
-	return true
-}

internal/cron/lock_reclaim_test.go

@@ -1,46 +0,0 @@
-package cron
-
-import (
-	"os"
-	"path/filepath"
-	"testing"
-	"time"
-)
-
-func TestReclaimStaleLock(t *testing.T) {
-	dir := t.TempDir()
-	lockPath := filepath.Join(dir, "job.lock")
-
-	// A genuinely stale lock (old mtime) is reclaimed and removed.
-	if err := os.WriteFile(lockPath, []byte("crashed-holder"), 0o600); err != nil {
-		t.Fatal(err)
-	}
-	old := time.Now().Add(-2 * cronLockStaleAfter)
-	if err := os.Chtimes(lockPath, old, old); err != nil {
-		t.Fatal(err)
-	}
-	if !reclaimStaleLock(lockPath, "tok-a", cronLockStaleAfter) {
-		t.Fatal("a stale lock should be reclaimed")
-	}
-	if _, err := os.Stat(lockPath); !os.IsNotExist(err) {
-		t.Fatalf("reclaimed stale lock should be gone, stat err=%v", err)
-	}
-
-	// A FRESH lock (someone reacquired in the gap) must be RESTORED intact, not
-	// stolen — this is the H3 mutual-exclusion protection.
-	if err := os.WriteFile(lockPath, []byte("live-holder"), 0o600); err != nil {
-		t.Fatal(err)
-	}
-	if reclaimStaleLock(lockPath, "tok-b", cronLockStaleAfter) {
-		t.Fatal("a fresh lock must not be reclaimed")
-	}
-	if data, err := os.ReadFile(lockPath); err != nil || string(data) != "live-holder" {
-		t.Fatalf("fresh lock must be left intact, got %q err %v", data, err)
-	}
-
-	// A missing lock reports no reclaim (nothing to steal).
-	_ = os.Remove(lockPath)
-	if reclaimStaleLock(lockPath, "tok-c", cronLockStaleAfter) {
-		t.Fatal("a missing lock should not report a reclaim")
-	}
-}

internal/cron/store.go

@@ -9,6 +9,8 @@ import (
 	"path/filepath"
 	"strings"
 	"time"
+
+	"github.com/Gitlawb/zero/internal/fsutil"
 )
 
 const (
@@ -144,7 +146,11 @@ func (s *Store) writeJob(job Job) error {
 	if err := os.WriteFile(tmp, data, 0o600); err != nil {
 		return err
 	}
-	return os.Rename(tmp, filepath.Join(dir, "metadata.json"))
+	if err := fsutil.RenameWithRetry(tmp, filepath.Join(dir, "metadata.json"), nil); err != nil {
+		_ = os.Remove(tmp)
+		return err
+	}
+	return nil
 }
 
 func (s *Store) Get(id string) (Job, error) {

internal/daemon/lock.go

@@ -8,6 +8,8 @@ import (
 	"strconv"
 	"strings"
 	"sync/atomic"
+
+	"github.com/Gitlawb/zero/internal/lockutil"
 )
 
 // Single-instance lock. Mirrors reference-daemon-code-agent-js/supervisor.js's
@@ -43,16 +45,16 @@ func acquireLock(path string, isAlive func(pid int) bool) (*fileLock, error) {
 			// the single-instance guarantee — so on write failure, remove it and fail.
 			if _, werr := fmt.Fprintf(f, "%d\n", os.Getpid()); werr != nil {
 				_ = f.Close()
-				_ = os.Remove(path)
+				_ = lockutil.RemoveLockFile(path)
 				return nil, werr
 			}
 			if cerr := f.Close(); cerr != nil {
-				_ = os.Remove(path)
+				_ = lockutil.RemoveLockFile(path)
 				return nil, cerr
 			}
 			return &fileLock{path: path}, nil
 		}
-		if !errors.Is(err, fs.ErrExist) {
+		if !errors.Is(err, fs.ErrExist) && !errors.Is(err, os.ErrPermission) {
 			return nil, err
 		}
 		pid, perr := readPidFile(path)
@@ -65,7 +67,13 @@ func acquireLock(path string, isAlive func(pid int) bool) (*fileLock, error) {
 		// freshly-created lock — leaving both "holding" the single-instance lock.
 		// reclaimStaleLock renames the file aside so only one racer wins the rename,
 		// and restores it if a live holder reacquired in the gap (D6).
-		reclaimStaleLock(path, isAlive)
+		if _, rerr := reclaimStaleLock(path, isAlive); rerr != nil {
+			// Reclaim hit a hard failure: the rename aside failed outright, or a
+			// live holder's lock could not be put back (the lock path may be
+			// missing, so re-acquiring would break the single-instance guarantee).
+			// Fail closed instead of spinning to the deadline.
+			return nil, fmt.Errorf("daemon: reclaim stale lock: %w", rerr)
+		}
 	}
 	return nil, ErrAlreadyRunning
 }
@@ -73,37 +81,29 @@ func acquireLock(path string, isAlive func(pid int) bool) (*fileLock, error) {
 // daemonLockSeq makes each reclaim attempt's sidelined filename unique per process.
 var daemonLockSeq atomic.Uint64
 
-// reclaimStaleLock atomically reclaims a single-instance lock whose recorded PID is
-// dead. A blind Remove lets two racers both delete-and-recreate the lock and wind up
-// BOTH holding it; instead this renames the lock file aside (only one racer can win
-// the rename of a given inode), re-reads the moved file's PID, and removes it only if
-// that PID is still dead. If a new holder reacquired the lock in the gap between the
-// stale check and the rename, the moved file carries that LIVE pid, so it is renamed
-// back rather than stolen. Returns true only when a genuinely stale lock was removed.
-func reclaimStaleLock(path string, isAlive func(pid int) bool) bool {
-	reclaimed := fmt.Sprintf("%s.stale.%d-%d", path, os.Getpid(), daemonLockSeq.Add(1))
-	if err := os.Rename(path, reclaimed); err != nil {
-		return false // another racer already moved/removed it, or it vanished
-	}
-	if pid, err := readPidFile(reclaimed); err == nil && pid > 0 && isAlive(pid) {
-		// A holder reacquired the lock in the gap — its live PID is now in the moved
-		// file. Restore it instead of stealing a live lock; let the caller refuse.
-		_ = os.Rename(reclaimed, path)
-		return false
-	}
-	_ = os.Remove(reclaimed)
-	return true
+// reclaimStaleLock atomically reclaims a single-instance lock whose recorded
+// PID is dead, via lockutil.ReclaimStaleLock: the lock file is renamed aside
+// (only one racer can win the rename) and stolen only if the moved file's PID
+// is still dead; a holder that reacquired the lock in the gap between the
+// stale check and the rename carries a LIVE pid and is restored rather than
+// stolen. Returns true only when a genuinely stale lock was removed. A
+// non-nil error means the caller must fail closed instead of re-acquiring
+// (see lockutil.ReclaimStaleLock).
+func reclaimStaleLock(path string, isAlive func(pid int) bool) (bool, error) {
+	suffix := fmt.Sprintf("%d-%d", os.Getpid(), daemonLockSeq.Add(1))
+	return lockutil.ReclaimStaleLock(path, suffix, func(reclaimedPath string) bool {
+		pid, err := readPidFile(reclaimedPath)
+		return err == nil && pid > 0 && isAlive(pid)
+	})
 }
 
-// release removes the lock file. Safe to call once.
+// release removes the lock file. Safe to call once. An already-missing lock
+// file is not an error (lockutil.RemoveLockFile swallows it on every platform).
 func (l *fileLock) release() error {
 	if l == nil || l.path == "" {
 		return nil
 	}
-	if err := os.Remove(l.path); err != nil && !errors.Is(err, fs.ErrNotExist) {
-		return err
-	}
-	return nil
+	return lockutil.RemoveLockFile(l.path)
 }
 
 // readPidFile reads and parses the PID recorded in a lock file.

internal/daemon/lock_test.go

@@ -71,8 +71,8 @@ func TestReclaimStaleLockRemovesDeadHolder(t *testing.T) {
 	if err := os.WriteFile(path, []byte("4242\n"), 0o600); err != nil {
 		t.Fatalf("write stale lock: %v", err)
 	}
-	if !reclaimStaleLock(path, func(int) bool { return false }) {
-		t.Fatal("reclaimStaleLock must report a genuinely stale (dead-PID) lock reclaimed")
+	if ok, err := reclaimStaleLock(path, func(int) bool { return false }); err != nil || !ok {
+		t.Fatalf("reclaimStaleLock must report a genuinely stale (dead-PID) lock reclaimed (ok=%v err=%v)", ok, err)
 	}
 	if _, err := os.Stat(path); !errors.Is(err, os.ErrNotExist) {
 		t.Fatalf("a reclaimed stale lock must be removed, stat = %v", err)
@@ -87,8 +87,8 @@ func TestReclaimStaleLockRestoresLiveHolder(t *testing.T) {
 	if err := os.WriteFile(path, []byte("4242\n"), 0o600); err != nil {
 		t.Fatalf("write lock: %v", err)
 	}
-	if reclaimStaleLock(path, func(int) bool { return true }) {
-		t.Fatal("reclaimStaleLock must NOT report a live-PID lock reclaimed")
+	if ok, err := reclaimStaleLock(path, func(int) bool { return true }); err != nil || ok {
+		t.Fatalf("reclaimStaleLock must NOT report a live-PID lock reclaimed (ok=%v err=%v)", ok, err)
 	}
 	data, err := os.ReadFile(path)
 	if err != nil {

internal/fsutil/rename.go

@@ -0,0 +1,45 @@
+// Package fsutil provides small filesystem helpers shared across packages.
+package fsutil
+
+import (
+	"errors"
+	"os"
+	"runtime"
+	"syscall"
+	"time"
+)
+
+// RenameWithRetry renames src to dst, retrying briefly on Windows when the
+// destination is transiently locked (antivirus scanners, search indexers, or
+// a concurrent reader holding the file open). rename overrides os.Rename so
+// tests can exercise the retry path; pass nil to use os.Rename.
+func RenameWithRetry(src, dst string, rename func(src, dst string) error) error {
+	if rename == nil {
+		rename = os.Rename
+	}
+	var err error
+	for i := 0; i < 10; i++ {
+		err = rename(src, dst)
+		if err == nil {
+			return nil
+		}
+		if runtime.GOOS == "windows" {
+			if os.IsPermission(err) || isWindowsSharingOrLockViolation(err) {
+				time.Sleep(10 * time.Millisecond)
+				continue
+			}
+		}
+		break
+	}
+	return err
+}
+
+func isWindowsSharingOrLockViolation(err error) bool {
+	var errno syscall.Errno
+	if errors.As(err, &errno) {
+		const ERROR_SHARING_VIOLATION syscall.Errno = 32
+		const ERROR_LOCK_VIOLATION syscall.Errno = 33
+		return errno == ERROR_SHARING_VIOLATION || errno == ERROR_LOCK_VIOLATION
+	}
+	return false
+}

internal/fsutil/rename_test.go

@@ -0,0 +1,47 @@
+package fsutil
+
+import (
+	"errors"
+	"runtime"
+	"syscall"
+	"testing"
+)
+
+func TestRenameWithRetryRetriesOnWindows(t *testing.T) {
+	if runtime.GOOS != "windows" {
+		t.Skip("RenameWithRetry only retries on Windows")
+	}
+	var attempts int
+	err := RenameWithRetry("src", "dst", func(src, dst string) error {
+		attempts++
+		switch attempts {
+		case 1:
+			return syscall.Errno(32) // ERROR_SHARING_VIOLATION
+		case 2:
+			return syscall.Errno(33) // ERROR_LOCK_VIOLATION
+		default:
+			return nil
+		}
+	})
+	if err != nil {
+		t.Fatalf("expected success after retries, got: %v", err)
+	}
+	if attempts != 3 {
+		t.Errorf("expected 3 attempts, got %d", attempts)
+	}
+}
+
+func TestRenameWithRetryNonRetryableError(t *testing.T) {
+	sentinel := errors.New("disk on fire")
+	var attempts int
+	err := RenameWithRetry("src", "dst", func(src, dst string) error {
+		attempts++
+		return sentinel
+	})
+	if !errors.Is(err, sentinel) {
+		t.Fatalf("expected sentinel error, got: %v", err)
+	}
+	if attempts != 1 {
+		t.Errorf("expected only 1 attempt for non-retryable error, got %d", attempts)
+	}
+}

internal/hooks/dispatch.go

@@ -176,8 +176,7 @@ func (dispatcher *Dispatcher) runWithTimeout(ctx context.Context, hook Definitio
 func classifyResult(event Event, result commandResult) (AuditStatus, bool) {
 	// A hook that started but was killed by its deadline (or a cancellation) never
 	// returned a verdict. For a blocking event we must fail CLOSED and veto: a
-	// hung beforeTool policy hook cannot be treated as approval. (A launch failure
-	// below still fails OPEN, so a misconfigured hook doesn't wedge every tool.)
+	// hung beforeTool policy hook cannot be treated as approval.
 	if result.TimedOut {
 		if blocksOn(event) {
 			return AuditBlocked, true
@@ -185,6 +184,9 @@ func classifyResult(event Event, result commandResult) (AuditStatus, bool) {
 		return AuditError, false
 	}
 	if result.Err != nil {
+		if blocksOn(event) {
+			return AuditBlocked, true
+		}
 		return AuditError, false
 	}
 	if result.ExitCode != 0 {

internal/hooks/dispatch_test.go

@@ -186,14 +186,18 @@ func TestExecCommandRunnerCapturesExitAndStdin(t *testing.T) {
 	}
 }
 
-func TestExecCommandRunnerReportsLaunchFailureWithoutBlocking(t *testing.T) {
+func TestExecCommandRunnerReportsLaunchFailureFailsClosedForBeforeTool(t *testing.T) {
 	result := execCommandRunner(context.Background(), "definitely-not-a-real-binary-zzz", nil, nil, t.TempDir(), nil)
 	if result.Err == nil {
 		t.Fatal("expected launch error for a missing binary")
 	}
-	// A launch failure is an error, never a block, even for beforeTool.
-	if status, blocked := classifyResult(EventBeforeTool, result); blocked || status != AuditError {
-		t.Fatalf("classify = (%q, %v), want (error, false) for a launch failure", status, blocked)
+	// A launch failure for beforeTool fails closed (vetoes/blocks).
+	if status, blocked := classifyResult(EventBeforeTool, result); !blocked || status != AuditBlocked {
+		t.Fatalf("beforeTool classify = (%q, %v), want (blocked, true) for a launch failure", status, blocked)
+	}
+	// An observational afterTool hook still fails open (does not block).
+	if status, blocked := classifyResult(EventAfterTool, result); blocked || status != AuditError {
+		t.Fatalf("afterTool classify = (%q, %v), want (error, false) for a launch failure", status, blocked)
 	}
 }
 

internal/hooks/lock.go

@@ -7,6 +7,8 @@ import (
 	"path/filepath"
 	"sync/atomic"
 	"time"
+
+	"github.com/Gitlawb/zero/internal/lockutil"
 )
 
 // Cross-process lock tuning for the audit log. The lock is held only across a
@@ -28,8 +30,7 @@ var auditLockSeq atomic.Uint64
 // than auditLockStaleAfter) is reclaimed. Wall-clock time is used deliberately so
 // lock timing never depends on the store's injectable clock and the stale check
 // compares against real file mtimes. This mirrors internal/cron/lock.go and
-// internal/oauth/lock.go (a shared filelock helper would DRY all three — left as a
-// follow-up to avoid churning those already-green packages here).
+// internal/oauth/lock.go. Uses internal/lockutil for lock restore and reclaim.
 func (store *AuditStore) lockAudit() (func(), error) {
 	lockPath := store.auditPath + ".lock"
 	if err := os.MkdirAll(filepath.Dir(lockPath), 0o700); err != nil {
@@ -44,11 +45,11 @@ func (store *AuditStore) lockAudit() (func(), error) {
 			// releaser could never delete it — stranding the lock. Fail closed.
 			if _, werr := f.WriteString(token); werr != nil {
 				_ = f.Close()
-				_ = os.Remove(lockPath)
+				_ = lockutil.RemoveLockFile(lockPath)
 				return nil, fmt.Errorf("hooks: write audit lock: %w", werr)
 			}
 			if cerr := f.Close(); cerr != nil {
-				_ = os.Remove(lockPath)
+				_ = lockutil.RemoveLockFile(lockPath)
 				return nil, fmt.Errorf("hooks: close audit lock: %w", cerr)
 			}
 			var released bool
@@ -57,10 +58,8 @@ func (store *AuditStore) lockAudit() (func(), error) {
 					return
 				}
 				released = true
-				// Only remove if the file still carries OUR token, so a lock
-				// reclaimed as stale by another process is not deleted under it.
 				if data, rerr := os.ReadFile(lockPath); rerr == nil && string(data) == token {
-					_ = os.Remove(lockPath)
+					_ = lockutil.RemoveLockFile(lockPath)
 				}
 			}, nil
 		}
@@ -76,7 +75,18 @@ func (store *AuditStore) lockAudit() (func(), error) {
 		// reclaimStaleLock renames the file aside (only one rename wins) and restores
 		// it if it turns out fresh, so a live lock is never deleted out from under it.
 		if info, statErr := os.Stat(lockPath); statErr == nil && time.Since(info.ModTime()) > auditLockStaleAfter {
-			if reclaimStaleLock(lockPath, token, auditLockStaleAfter) {
+			cleared, rerr := lockutil.ReclaimStaleLock(lockPath, token, func(reclaimedPath string) bool {
+				info, err := os.Stat(reclaimedPath)
+				return err == nil && time.Since(info.ModTime()) <= auditLockStaleAfter
+			})
+			if rerr != nil {
+				// Reclaim hit a hard failure: the rename aside failed outright, or a
+				// live holder's lock could not be put back (the lock path may be
+				// missing, so re-acquiring would break mutual exclusion). Fail closed
+				// instead of spinning to the deadline.
+				return nil, fmt.Errorf("hooks: reclaim stale audit lock: %w", rerr)
+			}
+			if cleared {
 				continue
 			}
 			// Lost the reclaim race (or it was actually fresh) — fall through to the
@@ -88,21 +98,3 @@ func (store *AuditStore) lockAudit() (func(), error) {
 		time.Sleep(auditLockRetryDelay)
 	}
 }
-
-// reclaimStaleLock atomically reclaims a stale lock file: it renames the lock to a
-// per-acquirer unique name (only one racer can rename a given file), then verifies
-// the moved file is still stale; if a holder reacquired it in the gap it is
-// restored rather than stolen. Returns true only when a genuinely stale lock was
-// removed. Mirrors internal/cron/lock.go and internal/oauth/lock.go.
-func reclaimStaleLock(lockPath, token string, staleAfter time.Duration) bool {
-	reclaimed := fmt.Sprintf("%s.stale.%s", lockPath, token)
-	if err := os.Rename(lockPath, reclaimed); err != nil {
-		return false
-	}
-	if info, err := os.Stat(reclaimed); err == nil && time.Since(info.ModTime()) <= staleAfter {
-		_ = os.Rename(reclaimed, lockPath)
-		return false
-	}
-	_ = os.Remove(reclaimed)
-	return true
-}

internal/keyring/keyring.go

@@ -3,11 +3,13 @@
 // macOS and `secret-tool` (libsecret) on Linux. It stores a single secret string
 // per (service, account). Windows and other platforms report unsupported.
 //
-// It shells out to the OS tools rather than taking a third-party dependency. On
-// macOS the secret is passed to `security` as an argument, so it is briefly
-// visible to other processes via the process list; on Linux the secret is passed
-// over stdin and is not exposed in the argument vector. Callers that need to keep
-// a secret out of the process list on macOS should prefer the file backend.
+// It shells out to the OS tools rather than taking a third-party dependency.
+// On both macOS and Linux the secret is passed over stdin, never the argument
+// vector, so it is not exposed via the process list. On macOS the write goes
+// through `security -i` (interactive mode), whose command parser is line-based
+// with a fixed 4096-byte line buffer, so Set rejects secrets containing
+// newlines or exceeding that budget rather than silently corrupting them;
+// Linux's secret-tool has no such restriction.
 package keyring
 
 import (
@@ -63,8 +65,28 @@ func (k *Keyring) Set(service, account, secret string) error {
 	}
 	switch k.goos {
 	case "darwin":
+		// The secret must stay out of the argument vector (visible to every
+		// process via `ps`), but a trailing -w prompt is not a substitute:
+		// security prompts with getpass(3), which reads from /dev/tty whenever
+		// the process has a controlling terminal (ignoring a piped stdin) and
+		// truncates at getpass's small fixed buffer otherwise. Instead drive
+		// `security -i`: interactive mode reads whole commands from stdin, so
+		// argv is just ["-i"] and the secret rides inside the command line.
+		// That parser is line-based with a fixed 4096-byte buffer, so reject
+		// newlines and oversized payloads up front; an overlong line would be
+		// split and executed as two garbage commands, silently corrupting the
+		// stored value.
+		for _, s := range []string{service, account, secret} {
+			if strings.ContainsAny(s, "\r\n") {
+				return wrap("set", errors.New("service, account, and secret must not contain newlines on macOS (security -i is line-based)"))
+			}
+		}
 		// -U updates the item if it already exists rather than failing.
-		_, err := k.exec(nil, "security", "add-generic-password", "-U", "-s", service, "-a", account, "-w", secret)
+		line := "add-generic-password -U -s " + securityQuote(service) + " -a " + securityQuote(account) + " -w " + securityQuote(secret) + "\n"
+		if len(line) > securityMaxLine {
+			return wrap("set", fmt.Errorf("secret too large for the macOS security tool's %d-byte command line; use the file backend instead", securityMaxLine))
+		}
+		_, err := k.exec([]byte(line), "security", "-i")
 		return wrap("set", err)
 	case "linux":
 		// secret-tool reads the secret from stdin, keeping it out of the argv.
@@ -140,6 +162,19 @@ func (k *Keyring) Delete(service, account string) (bool, error) {
 	}
 }
 
+// securityMaxLine is the most `security -i` can read as one command: its line
+// buffer is MAX_LINE_LEN (4096) bytes in Apple's SecurityTool, one of which the
+// terminating NUL consumes.
+const securityMaxLine = 4095
+
+// securityQuote wraps s for one argument of a `security -i` command line. The
+// tool's parser (split_line in Apple's SecurityTool) treats a backslash inside
+// double quotes as escaping the next character and the matching quote as the
+// argument terminator, so those two characters are all that needs escaping.
+func securityQuote(s string) string {
+	return `"` + strings.NewReplacer(`\`, `\\`, `"`, `\"`).Replace(s) + `"`
+}
+
 func (k *Keyring) exec(stdin []byte, name string, args ...string) ([]byte, error) {
 	ctx, cancel := context.WithTimeout(context.Background(), commandTimeout)
 	defer cancel()

internal/keyring/keyring_test.go

@@ -50,6 +50,52 @@ func attrValue(args []string, attr string) string {
 
 func key(service, account string) string { return service + "\x00" + account }
 
+// splitSecurityLine mirrors split_line in Apple's SecurityTool so the fake
+// parses `security -i` command lines the way the real tool does: whitespace
+// separates arguments, double or single quotes group one, and a backslash
+// escapes the next character both inside and outside quotes.
+func splitSecurityLine(line string) []string {
+	var args []string
+	var cur strings.Builder
+	inArg, escaped := false, false
+	quote := byte(0)
+	for i := 0; i < len(line); i++ {
+		c := line[i]
+		switch {
+		case escaped:
+			cur.WriteByte(c)
+			escaped = false
+		case c == '\\':
+			escaped = true
+			inArg = true
+		case quote != 0:
+			if c != quote {
+				cur.WriteByte(c)
+				continue
+			}
+			args = append(args, cur.String())
+			cur.Reset()
+			inArg, quote = false, 0
+		case c == '"' || c == '\'':
+			quote = c
+			inArg = true
+		case c == ' ' || c == '\t':
+			if inArg {
+				args = append(args, cur.String())
+				cur.Reset()
+				inArg = false
+			}
+		default:
+			cur.WriteByte(c)
+			inArg = true
+		}
+	}
+	if inArg {
+		args = append(args, cur.String())
+	}
+	return args
+}
+
 func (f *fakeKeyring) run(_ context.Context, name string, stdin []byte, args ...string) ([]byte, error) {
 	f.lastStdin = string(stdin)
 	f.lastArgs = append([]string{name}, args...)
@@ -58,10 +104,19 @@ func (f *fakeKeyring) run(_ context.Context, name string, stdin []byte, args ...
 	}
 	switch f.goos {
 	case "darwin":
-		svc, acct := flagValue(args, "-s"), flagValue(args, "-a")
-		switch args[0] {
+		cmdArgs := args
+		if args[0] == "-i" {
+			// Interactive mode: the command arrives as one line on stdin.
+			line, _, _ := strings.Cut(string(stdin), "\n")
+			cmdArgs = splitSecurityLine(line)
+			if len(cmdArgs) == 0 {
+				return nil, fakeExit{1}
+			}
+		}
+		svc, acct := flagValue(cmdArgs, "-s"), flagValue(cmdArgs, "-a")
+		switch cmdArgs[0] {
 		case "add-generic-password":
-			f.data[key(svc, acct)] = flagValue(args, "-w")
+			f.data[key(svc, acct)] = flagValue(cmdArgs, "-w")
 			return nil, nil
 		case "find-generic-password":
 			if v, ok := f.data[key(svc, acct)]; ok {
@@ -129,6 +184,112 @@ func TestKeyringRoundTripDarwin(t *testing.T) {
 	}
 }
 
+func TestKeyringRoundTripDarwinUsesStdin(t *testing.T) {
+	f := newFake("darwin")
+	k := f.keyring()
+	if err := k.Set("zero", "tokens", "blob-CCC"); err != nil {
+		t.Fatalf("Set: %v", err)
+	}
+	// The secret must travel via stdin, never the argument vector: the write
+	// goes through `security -i`, whose whole command line (secret included)
+	// arrives on stdin while argv carries only the -i flag.
+	wantStdin := "add-generic-password -U -s \"zero\" -a \"tokens\" -w \"blob-CCC\"\n"
+	if f.lastStdin != wantStdin {
+		t.Fatalf("secret not sent via stdin correctly: stdin=%q, want=%q", f.lastStdin, wantStdin)
+	}
+	wantArgs := []string{"security", "-i"}
+	if len(f.lastArgs) != len(wantArgs) || f.lastArgs[0] != wantArgs[0] || f.lastArgs[1] != wantArgs[1] {
+		t.Fatalf("argv = %v, want %v", f.lastArgs, wantArgs)
+	}
+	got, ok, err := k.Get("zero", "tokens")
+	if err != nil || !ok || got != "blob-CCC" {
+		t.Fatalf("Get = %q ok=%v err=%v", got, ok, err)
+	}
+	existed, err := k.Delete("zero", "tokens")
+	if err != nil || !existed {
+		t.Fatalf("Delete: existed=%v err=%v", existed, err)
+	}
+}
+
+// TestKeyringSetRejectsMultilineSecretOnDarwin guards against silently
+// truncating a secret containing a newline: `security -i` reads one command
+// per line, so an embedded newline would split the write into two garbage
+// commands. Set must reject this before ever invoking security, not store a
+// corrupted value.
+func TestKeyringSetRejectsMultilineSecretOnDarwin(t *testing.T) {
+	for _, secret := range []string{"line1\nline2", "line1\r\nline2", "trailing\n"} {
+		f := newFake("darwin")
+		k := f.keyring()
+		if err := k.Set("zero", "tokens", secret); err == nil {
+			t.Fatalf("Set(%q) = nil error, want rejection of the embedded newline", secret)
+		}
+		if f.lastArgs != nil {
+			t.Fatalf("Set(%q) should be rejected before invoking security, got args=%v", secret, f.lastArgs)
+		}
+	}
+}
+
+// TestKeyringDarwinQuotesSpecialCharacters proves the quoting survives the
+// real tool's parser: the fake tokenizes stdin with a faithful mirror of
+// SecurityTool's split_line, so a secret full of quotes, backslashes, and
+// spaces must round-trip unchanged.
+func TestKeyringDarwinQuotesSpecialCharacters(t *testing.T) {
+	for _, secret := range []string{
+		`spa ced`,
+		`quo"te`,
+		`back\slash`,
+		`sin'gle`,
+		`mi"x'ed \" \\ end\`,
+		` leading and trailing `,
+	} {
+		f := newFake("darwin")
+		k := f.keyring()
+		if err := k.Set("zero", "tokens", secret); err != nil {
+			t.Fatalf("Set(%q): %v", secret, err)
+		}
+		got, ok, err := k.Get("zero", "tokens")
+		if err != nil || !ok || got != secret {
+			t.Fatalf("Get after Set(%q) = %q ok=%v err=%v", secret, got, ok, err)
+		}
+	}
+}
+
+// TestKeyringSetRejectsOversizedSecretOnDarwin guards the `security -i` line
+// budget: MAX_LINE_LEN is 4096 bytes, and an overlong line would be split and
+// executed as two garbage commands. A payload comfortably under the budget
+// must succeed; one over it must be rejected before invoking security.
+func TestKeyringSetRejectsOversizedSecretOnDarwin(t *testing.T) {
+	f := newFake("darwin")
+	k := f.keyring()
+	if err := k.Set("zero", "tokens", strings.Repeat("a", 4000)); err != nil {
+		t.Fatalf("Set(4000 bytes): %v", err)
+	}
+	f = newFake("darwin")
+	k = f.keyring()
+	if err := k.Set("zero", "tokens", strings.Repeat("a", 5000)); err == nil {
+		t.Fatal("Set(5000 bytes) = nil error, want rejection of the oversized line")
+	}
+	if f.lastArgs != nil {
+		t.Fatalf("oversized Set should be rejected before invoking security, got args=%v", f.lastArgs)
+	}
+}
+
+// TestKeyringSetAllowsMultilineSecretOnLinux confirms the newline restriction
+// is macOS-specific: secret-tool reads the whole stdin payload as the secret,
+// with no line-based prompt to corrupt an embedded newline.
+func TestKeyringSetAllowsMultilineSecretOnLinux(t *testing.T) {
+	f := newFake("linux")
+	k := f.keyring()
+	secret := "line1\nline2"
+	if err := k.Set("zero", "tokens", secret); err != nil {
+		t.Fatalf("Set: %v", err)
+	}
+	got, ok, err := k.Get("zero", "tokens")
+	if err != nil || !ok || got != secret {
+		t.Fatalf("Get = %q ok=%v err=%v, want %q", got, ok, err, secret)
+	}
+}
+
 func TestKeyringRoundTripLinuxUsesStdin(t *testing.T) {
 	f := newFake("linux")
 	k := f.keyring()

internal/lockutil/lockutil.go

@@ -0,0 +1,64 @@
+// Package lockutil provides the platform-specific file primitives behind the
+// O_EXCL lock files in cron, daemon, hooks, oauth, and swarm: a no-overwrite
+// restore for locks that were sidelined during a stale-reclaim attempt, and a
+// lock file remover with one cross-platform contract (missing files are a
+// no-op; Windows retries transient sharing violations).
+package lockutil
+
+import (
+	"io"
+	"os"
+)
+
+// restoreByCopy restores reclaimed to path without overwriting an existing
+// path, as a fallback for when the platform's primary no-replace primitive
+// (hard link on POSIX, MoveFileEx on Windows) fails for a reason other than
+// the destination existing. It stages a full copy under a private name next
+// to reclaimed and publishes it to path with publish (the same no-replace
+// primitive the caller's platform uses for the primary restore), so path
+// never appears in a partially-copied state: a crash between staging and
+// publish leaves path exactly as it was (missing or, if a competing holder
+// created it, untouched), never a truncated file that a PID/content-based
+// liveness check could mistake for dead. Leaving path missing here would let
+// the next O_EXCL create succeed while the sidelined holder is still live,
+// breaking mutual exclusion. publish keeps the no-overwrite guarantee: a new
+// holder that appeared in the meantime wins and this returns os.ErrExist. The
+// copy resets the lock's mtime to now, which only makes the restored lock
+// look fresher; that is safe, since it is being handed back to a live holder.
+func restoreByCopy(reclaimed, path string, publish func(from, to string) error) error {
+	staged := reclaimed + ".copy"
+	src, err := os.Open(reclaimed)
+	if err != nil {
+		return err
+	}
+	dst, err := os.OpenFile(staged, os.O_CREATE|os.O_EXCL|os.O_WRONLY, 0o600)
+	if err != nil {
+		_ = src.Close()
+		return err
+	}
+	_, err = io.Copy(dst, src)
+	// Close the source before removing anything below: Go opens files without
+	// FILE_SHARE_DELETE on Windows, so deleting reclaimed or staged while src
+	// is open would fail with a sharing violation.
+	_ = src.Close()
+	if err != nil {
+		_ = dst.Close()
+		_ = os.Remove(staged)
+		return err
+	}
+	if err := dst.Close(); err != nil {
+		_ = os.Remove(staged)
+		return err
+	}
+	if err := publish(staged, path); err != nil {
+		_ = os.Remove(staged)
+		return err
+	}
+	// publish may leave staged behind (a hard-link publish keeps both names),
+	// so remove it explicitly; a move-based publish already made this a no-op.
+	_ = os.Remove(staged)
+	// The lock is back at path, so the restore has succeeded; failing to clean
+	// up the sidelined name must not be reported as a failed restore.
+	_ = RemoveLockFile(reclaimed)
+	return nil
+}

internal/lockutil/lockutil_other.go

@@ -0,0 +1,46 @@
+//go:build !windows
+
+package lockutil
+
+import (
+	"errors"
+	"io/fs"
+	"os"
+)
+
+// RestoreLockFile restores a sidelined lock file on non-Windows platforms. It
+// uses os.Link so a competing lock created at path in the meantime wins: the
+// link fails with os.ErrExist instead of overwriting it. If the link fails for
+// any other reason (hardlink-incapable filesystems such as FAT or some
+// FUSE/network mounts, ENOSPC, EPERM), it falls back to an O_EXCL copy rather
+// than leaving path missing and the live holder's lock stranded in reclaimed.
+// Once the lock is back at path the restore has succeeded, so a failed cleanup
+// of the sidelined name is not reported as an error; the leftover file is
+// invisible to the lock protocol.
+func RestoreLockFile(reclaimed, path string) error {
+	err := os.Link(reclaimed, path)
+	if err == nil {
+		_ = RemoveLockFile(reclaimed)
+		return nil
+	}
+	if errors.Is(err, os.ErrExist) {
+		return err
+	}
+	return restoreByCopy(reclaimed, path, os.Link)
+}
+
+// isReclaimContended reports whether a failed rename-aside of a suspected
+// stale lock means it was lost to a racer rather than a hard failure. POSIX
+// rename has no contention errno (a lost race surfaces only as ENOENT, which
+// ReclaimStaleLock already treats as benign), so nothing extra is benign here.
+func isReclaimContended(error) bool { return false }
+
+// RemoveLockFile removes a lock file on non-Windows platforms. Removing an
+// already-missing file reports nil, matching the Windows implementation, so
+// callers see one cross-platform contract.
+func RemoveLockFile(path string) error {
+	if err := os.Remove(path); err != nil && !errors.Is(err, fs.ErrNotExist) {
+		return err
+	}
+	return nil
+}

internal/lockutil/lockutil_test.go

@@ -0,0 +1,116 @@
+package lockutil
+
+import (
+	"errors"
+	"os"
+	"path/filepath"
+	"testing"
+)
+
+func TestRestoreLockFile(t *testing.T) {
+	tempDir := t.TempDir()
+	path := filepath.Join(tempDir, "lock")
+	reclaimed := filepath.Join(tempDir, "lock.stale.token")
+
+	// 1. Successful restore when target does not exist
+	if err := os.WriteFile(reclaimed, []byte("token"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	if err := RestoreLockFile(reclaimed, path); err != nil {
+		t.Fatalf("RestoreLockFile failed: %v", err)
+	}
+	if data, err := os.ReadFile(path); err != nil || string(data) != "token" {
+		t.Fatalf("restored lock content = %q, err %v; want %q intact", data, err, "token")
+	}
+	if _, err := os.Stat(reclaimed); !os.IsNotExist(err) {
+		t.Fatalf("expected sidelined lock to be cleaned up/removed: %v", err)
+	}
+
+	// 2. Fail when target already exists (prevent overwrite)
+	if err := os.WriteFile(reclaimed, []byte("token"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(path, []byte("competing"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+
+	err := RestoreLockFile(reclaimed, path)
+	if err == nil {
+		t.Fatal("expected RestoreLockFile to fail when target exists")
+	}
+	if !errors.Is(err, os.ErrExist) {
+		t.Fatalf("expected os.ErrExist, got: %v", err)
+	}
+
+	// The sidelined lock must still exist on failure, and the competing lock
+	// must be untouched.
+	if _, err := os.Stat(reclaimed); err != nil {
+		t.Fatalf("expected sidelined lock to still exist: %v", err)
+	}
+	if data, _ := os.ReadFile(path); string(data) != "competing" {
+		t.Fatalf("competing lock was clobbered: %q", data)
+	}
+}
+
+func TestRestoreByCopy(t *testing.T) {
+	tempDir := t.TempDir()
+	path := filepath.Join(tempDir, "lock")
+	reclaimed := filepath.Join(tempDir, "lock.stale.token")
+
+	// 1. Restores content and cleans up the sidelined file when the target is
+	// free (the fallback taken when the platform's no-replace primitive fails
+	// for a reason other than the destination existing).
+	if err := os.WriteFile(reclaimed, []byte("token"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	if err := restoreByCopy(reclaimed, path, os.Link); err != nil {
+		t.Fatalf("restoreByCopy failed: %v", err)
+	}
+	if data, err := os.ReadFile(path); err != nil || string(data) != "token" {
+		t.Fatalf("restored lock content = %q, err %v; want %q intact", data, err, "token")
+	}
+	if _, err := os.Stat(reclaimed); !errors.Is(err, os.ErrNotExist) {
+		t.Fatalf("expected sidelined lock to be cleaned up after copy: %v", err)
+	}
+	if _, err := os.Stat(reclaimed + ".copy"); !errors.Is(err, os.ErrNotExist) {
+		t.Fatalf("expected staged copy to be cleaned up: %v", err)
+	}
+
+	// 2. Keeps the no-overwrite guarantee: a competing lock at the target wins
+	// and the sidelined lock survives for the caller's ErrExist cleanup.
+	if err := os.WriteFile(reclaimed, []byte("token"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(path, []byte("competing"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	if err := restoreByCopy(reclaimed, path, os.Link); !errors.Is(err, os.ErrExist) {
+		t.Fatalf("expected os.ErrExist when target exists, got: %v", err)
+	}
+	if data, _ := os.ReadFile(path); string(data) != "competing" {
+		t.Fatalf("competing lock was clobbered: %q", data)
+	}
+	if _, err := os.Stat(reclaimed); err != nil {
+		t.Fatalf("expected sidelined lock to still exist: %v", err)
+	}
+	if _, err := os.Stat(reclaimed + ".copy"); !errors.Is(err, os.ErrNotExist) {
+		t.Fatalf("expected staged copy to be cleaned up after a failed publish: %v", err)
+	}
+}
+
+func TestRemoveLockFile(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "lock")
+	if err := os.WriteFile(path, []byte("token"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	if err := RemoveLockFile(path); err != nil {
+		t.Fatalf("RemoveLockFile failed: %v", err)
+	}
+	if _, err := os.Stat(path); !errors.Is(err, os.ErrNotExist) {
+		t.Fatalf("expected lock file to be removed: %v", err)
+	}
+	// Removing an already-missing lock file is a no-op on every platform.
+	if err := RemoveLockFile(path); err != nil {
+		t.Fatalf("RemoveLockFile on a missing file = %v, want nil", err)
+	}
+}

internal/lockutil/lockutil_windows.go

@@ -0,0 +1,80 @@
+//go:build windows
+
+package lockutil
+
+import (
+	"errors"
+	"os"
+	"syscall"
+	"time"
+
+	"golang.org/x/sys/windows"
+)
+
+// RestoreLockFile restores a sidelined lock file on Windows. It moves the file
+// with a no-replace MoveFileEx so a competing lock created at path in the
+// meantime wins: the move fails with os.ErrExist instead of overwriting it. If
+// the move fails for any other reason (ERROR_SHARING_VIOLATION or
+// ERROR_ACCESS_DENIED under concurrent access to the file), it falls back to
+// an O_EXCL copy rather than leaving path missing and the live holder's lock
+// stranded in reclaimed.
+func RestoreLockFile(reclaimed, path string) error {
+	err := moveFileNoReplace(reclaimed, path)
+	if err == nil || errors.Is(err, os.ErrExist) {
+		return err
+	}
+	return restoreByCopy(reclaimed, path, moveFileNoReplace)
+}
+
+// moveFileNoReplace renames from to to, failing if to already exists. It calls
+// MoveFileExW directly via golang.org/x/sys/windows (the standard library's
+// syscall package does not export MoveFileEx on this platform) with no flags,
+// so an existing destination fails with ERROR_ALREADY_EXISTS, which satisfies
+// errors.Is(err, os.ErrExist) (pinned by
+// TestMoveFileNoReplaceMapsAlreadyExistsToErrExist), instead of overwriting it
+// the way os.Rename does on Windows (it passes MOVEFILE_REPLACE_EXISTING to
+// match POSIX rename semantics cross-platform).
+func moveFileNoReplace(from, to string) error {
+	fromPtr, err := windows.UTF16PtrFromString(from)
+	if err != nil {
+		return err
+	}
+	toPtr, err := windows.UTF16PtrFromString(to)
+	if err != nil {
+		return err
+	}
+	return windows.MoveFileEx(fromPtr, toPtr, 0)
+}
+
+// isReclaimContended reports whether a failed rename-aside of a suspected
+// stale lock means the file was concurrently open or pending deletion (which
+// surfaces as ERROR_SHARING_VIOLATION or ERROR_ACCESS_DENIED on Windows)
+// rather than a hard failure, so ReclaimStaleLock reports a lost race instead
+// of an error. Neither errno satisfies errors.Is(err, os.ErrExist), and
+// ERROR_SHARING_VIOLATION does not satisfy errors.Is(err, os.ErrPermission)
+// either, so the classification compares raw errnos like RemoveLockFile below.
+func isReclaimContended(err error) bool {
+	var errno syscall.Errno
+	return errors.As(err, &errno) && (errno == windows.ERROR_SHARING_VIOLATION || errno == windows.ERROR_ACCESS_DENIED)
+}
+
+// RemoveLockFile removes a lock file on Windows, retrying on the sharing
+// violation or access denied errors that are common under heavy concurrent
+// contention. Removing an already-missing file reports nil, matching the
+// non-Windows implementation, so callers see one cross-platform contract.
+func RemoveLockFile(path string) error {
+	var err error
+	for i := 0; i < 15; i++ {
+		err = os.Remove(path)
+		if err == nil || os.IsNotExist(err) {
+			return nil
+		}
+		var errno syscall.Errno
+		if errors.As(err, &errno) && (errno == windows.ERROR_SHARING_VIOLATION || errno == windows.ERROR_ACCESS_DENIED) {
+			time.Sleep(5 * time.Millisecond)
+			continue
+		}
+		return err
+	}
+	return err
+}

internal/lockutil/lockutil_windows_test.go

@@ -0,0 +1,83 @@
+//go:build windows
+
+package lockutil
+
+import (
+	"errors"
+	"os"
+	"path/filepath"
+	"testing"
+	"time"
+
+	"golang.org/x/sys/windows"
+)
+
+// TestMoveFileNoReplaceMapsAlreadyExistsToErrExist pins the errno coupling
+// every caller keys its cleanup on: with no flags, MoveFileEx fails with
+// ERROR_ALREADY_EXISTS when the destination exists, and that errno must
+// satisfy errors.Is(err, os.ErrExist).
+func TestMoveFileNoReplaceMapsAlreadyExistsToErrExist(t *testing.T) {
+	dir := t.TempDir()
+	from := filepath.Join(dir, "from")
+	to := filepath.Join(dir, "to")
+	if err := os.WriteFile(from, []byte("sidelined"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(to, []byte("competing"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	err := moveFileNoReplace(from, to)
+	if !errors.Is(err, windows.ERROR_ALREADY_EXISTS) {
+		t.Fatalf("MoveFileEx with an existing destination = %v, want ERROR_ALREADY_EXISTS", err)
+	}
+	if !errors.Is(err, os.ErrExist) {
+		t.Fatalf("ERROR_ALREADY_EXISTS must satisfy errors.Is(err, os.ErrExist), got: %v", err)
+	}
+	if data, _ := os.ReadFile(to); string(data) != "competing" {
+		t.Fatalf("destination was overwritten: %q", data)
+	}
+}
+
+// TestRemoveLockFileRetriesSharingViolation holds the file open briefly (Go
+// opens files without FILE_SHARE_DELETE on Windows, so a concurrent DeleteFile
+// fails with ERROR_SHARING_VIOLATION while the handle is open); RemoveLockFile
+// must ride its retry loop (15 x 5ms) past the window and succeed once the
+// handle closes.
+func TestRemoveLockFileRetriesSharingViolation(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "lock")
+	if err := os.WriteFile(path, []byte("token"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	f, err := os.Open(path)
+	if err != nil {
+		t.Fatal(err)
+	}
+	go func() {
+		time.Sleep(10 * time.Millisecond)
+		_ = f.Close()
+	}()
+	if err := RemoveLockFile(path); err != nil {
+		t.Fatalf("RemoveLockFile should succeed once the contending handle closes: %v", err)
+	}
+	if _, err := os.Stat(path); !errors.Is(err, os.ErrNotExist) {
+		t.Fatalf("expected lock file to be removed: %v", err)
+	}
+}
+
+// TestRemoveLockFileSharingViolationExhausts verifies the retry loop is
+// bounded: with the contending handle held for the whole call, RemoveLockFile
+// surfaces the sharing violation instead of retrying forever.
+func TestRemoveLockFileSharingViolationExhausts(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "lock")
+	if err := os.WriteFile(path, []byte("token"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	f, err := os.Open(path)
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer f.Close()
+	if err := RemoveLockFile(path); !errors.Is(err, windows.ERROR_SHARING_VIOLATION) {
+		t.Fatalf("RemoveLockFile with a held handle = %v, want ERROR_SHARING_VIOLATION", err)
+	}
+}

internal/lockutil/reclaim.go

@@ -0,0 +1,82 @@
+package lockutil
+
+import (
+	"errors"
+	"os"
+)
+
+// restoreLockFile is swappable so tests can force the fail-closed path of
+// ReclaimStaleLock, which requires both the fast restore and its no-replace
+// fallback (with its own copy fallback) to fail; that cannot be provoked
+// portably on a healthy filesystem.
+var restoreLockFile = restoreLiveLock
+
+// restoreLiveLock puts a lock that turned out to be live back at path after
+// ReclaimStaleLock moved it aside to inspect it. It first tries a fast,
+// replacing rename straight from reclaimed to path: a single syscall, which
+// keeps the window during which path does not exist (and so is open to a
+// fourth process's unrelated O_EXCL create landing on it) as short as
+// possible. RestoreLockFile's no-replace restore (with its slower copy
+// fallback on some failures) is a correctness-preserving fallback for when
+// the fast path itself fails (e.g. a cross-device sidelined name): it is a
+// much longer version of the identical race, but still detects rather than
+// silently clobbers a competing lock, which the fast path's replacing rename
+// cannot do. Neither path makes the race impossible, only unlikely; see
+// ReclaimStaleLock's doc comment.
+func restoreLiveLock(reclaimed, path string) error {
+	if err := os.Rename(reclaimed, path); err == nil {
+		return nil
+	}
+	return RestoreLockFile(reclaimed, path)
+}
+
+// ReclaimStaleLock atomically reclaims a suspected-stale lock file. It renames
+// lockPath aside to "<lockPath>.stale.<suffix>" (only one racer can win the
+// rename of a given file, so two racers can never both reclaim the same lock),
+// then consults isLive on the moved file; if the lock turns out to be live (a
+// holder reacquired it in the gap between the caller's stale check and the
+// rename) it is restored rather than stolen. The suffix must be unique per
+// acquirer attempt. Returns true only when a genuinely stale lock was removed,
+// so the caller knows it is safe to retry its exclusive create immediately; on
+// a lost race it returns false. A non-nil error means either the rename aside
+// failed for a reason that is not contention (so retrying cannot help and the
+// caller should fail fast instead of spinning to its deadline), or a live
+// holder's lock could not be restored (both restoreLiveLock's fast path and
+// its no-replace fallback failed), so lockPath may be missing; callers must
+// fail closed instead of re-acquiring. The sidelined file is removed on every
+// restore failure: once the restore has failed it has no protocol function
+// (release only consults the lock path), so keeping it would only leak files.
+//
+// The live-restore path has an inherent, unclosed race: between the rename
+// aside above and restoreLiveLock putting the lock back, lockPath does not
+// exist, so an unrelated caller's O_EXCL create can legitimately succeed
+// there. restoreLiveLock's fast path then silently overwrites that new
+// claimant's lock file, which does not corrupt release (it is
+// ownership-aware, so the new claimant's later release safely no-ops against
+// content it no longer owns) but does mean the new claimant can still run
+// its critical section concurrently with the original live holder. Making
+// this race actually impossible would need an OS-level advisory lock (flock
+// / LockFileEx) held for a holder's whole critical section, checked
+// non-destructively instead of by moving the file; restoreLiveLock only
+// shrinks the window to roughly one syscall, it does not close it.
+func ReclaimStaleLock(lockPath, suffix string, isLive func(reclaimedPath string) bool) (bool, error) {
+	reclaimed := lockPath + ".stale." + suffix
+	if err := os.Rename(lockPath, reclaimed); err != nil {
+		if errors.Is(err, os.ErrNotExist) || isReclaimContended(err) {
+			return false, nil // another racer already moved/removed it, or it vanished
+		}
+		return false, err
+	}
+	if isLive(reclaimed) {
+		// Put the live lock back instead of stealing it, and let the caller wait.
+		if rerr := restoreLockFile(reclaimed, lockPath); rerr != nil {
+			_ = RemoveLockFile(reclaimed)
+			if !errors.Is(rerr, os.ErrExist) {
+				return false, rerr
+			}
+		}
+		return false, nil
+	}
+	_ = RemoveLockFile(reclaimed)
+	return true, nil
+}

internal/lockutil/reclaim_other_test.go

@@ -0,0 +1,31 @@
+//go:build !windows
+
+package lockutil
+
+import (
+	"os"
+	"path/filepath"
+	"testing"
+)
+
+// TestReclaimStaleLockPropagatesRenameFailure pins the fail-fast contract of
+// the rename aside: a failure that is not a lost race must reach the caller.
+// Mapping it to the benign lost-race result would leave acquirers spinning
+// until their deadline and reporting a timeout instead of the real cause. A
+// directory squatting on the sidelined name makes os.Rename fail with EISDIR,
+// which matches neither os.ErrNotExist nor a contention errno.
+func TestReclaimStaleLockPropagatesRenameFailure(t *testing.T) {
+	lockPath := filepath.Join(t.TempDir(), "lock")
+	if err := os.WriteFile(lockPath, []byte("stale-holder"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.Mkdir(lockPath+".stale.tok", 0o755); err != nil {
+		t.Fatal(err)
+	}
+	if ok, err := ReclaimStaleLock(lockPath, "tok", func(string) bool { return false }); err == nil || ok {
+		t.Fatalf("a hard rename failure must surface an error (ok=%v err=%v)", ok, err)
+	}
+	if _, err := os.Stat(lockPath); err != nil {
+		t.Fatalf("the lock file must be left in place on a failed rename: %v", err)
+	}
+}

internal/lockutil/reclaim_test.go

@@ -0,0 +1,112 @@
+package lockutil
+
+import (
+	"errors"
+	"os"
+	"path/filepath"
+	"testing"
+)
+
+func TestReclaimStaleLock(t *testing.T) {
+	lockPath := filepath.Join(t.TempDir(), "lock")
+	dead := func(string) bool { return false }
+	live := func(string) bool { return true }
+
+	// A lock the predicate reports dead is reclaimed and removed.
+	if err := os.WriteFile(lockPath, []byte("crashed-holder"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	if ok, err := ReclaimStaleLock(lockPath, "tok-a", dead); err != nil || !ok {
+		t.Fatalf("a dead lock should be reclaimed (ok=%v err=%v)", ok, err)
+	}
+	if _, err := os.Stat(lockPath); !errors.Is(err, os.ErrNotExist) {
+		t.Fatalf("reclaimed dead lock should be gone, stat err=%v", err)
+	}
+
+	// A LIVE lock (a holder reacquired in the gap) must be RESTORED intact.
+	if err := os.WriteFile(lockPath, []byte("live-holder"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	if ok, err := ReclaimStaleLock(lockPath, "tok-b", live); err != nil || ok {
+		t.Fatalf("a live lock must not be reclaimed (ok=%v err=%v)", ok, err)
+	}
+	if data, err := os.ReadFile(lockPath); err != nil || string(data) != "live-holder" {
+		t.Fatalf("live lock must be left intact, got %q err %v", data, err)
+	}
+
+	// A missing lock reports no reclaim (nothing to steal).
+	_ = os.Remove(lockPath)
+	if ok, err := ReclaimStaleLock(lockPath, "tok-c", live); err != nil || ok {
+		t.Fatalf("a missing lock should not report a reclaim (ok=%v err=%v)", ok, err)
+	}
+}
+
+// restoreLiveLock's fast path is a single replacing rename, chosen to
+// minimize (not eliminate) the window during which lockPath is absent and
+// open to an unrelated O_EXCL create; see ReclaimStaleLock's doc comment for
+// why that residual race exists. This test pins the resulting behavior: the
+// fast path overwrites a competing file at path rather than detecting it,
+// unlike RestoreLockFile's own no-replace contract.
+func TestRestoreLiveLockFastPathOverwritesCompetingFile(t *testing.T) {
+	dir := t.TempDir()
+	reclaimed := filepath.Join(dir, "lock.stale.tok")
+	path := filepath.Join(dir, "lock")
+
+	if err := os.WriteFile(reclaimed, []byte("original-holder"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(path, []byte("new-claimant"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	if err := restoreLiveLock(reclaimed, path); err != nil {
+		t.Fatalf("restoreLiveLock failed: %v", err)
+	}
+	data, err := os.ReadFile(path)
+	if err != nil || string(data) != "original-holder" {
+		t.Fatalf("path = %q, err %v; want the original holder's content restored", data, err)
+	}
+	if _, err := os.Stat(reclaimed); !errors.Is(err, os.ErrNotExist) {
+		t.Fatalf("expected reclaimed to be consumed by the rename: %v", err)
+	}
+}
+
+func TestReclaimStaleLockFailsClosedOnRestoreError(t *testing.T) {
+	// When both the no-replace restore and its copy fallback fail (only
+	// provokable via the seam; a healthy filesystem cannot produce it), the
+	// caller must receive an error so it fails closed instead of re-acquiring a
+	// missing lock path, and the sidelined file must not leak.
+	restoreLockFile = func(reclaimed, path string) error { return errors.New("restore failed") }
+	defer func() { restoreLockFile = restoreLiveLock }()
+
+	lockPath := filepath.Join(t.TempDir(), "lock")
+	if err := os.WriteFile(lockPath, []byte("live-holder"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	ok, err := ReclaimStaleLock(lockPath, "tok", func(string) bool { return true })
+	if err == nil || ok {
+		t.Fatalf("a failed restore must surface an error (ok=%v err=%v)", ok, err)
+	}
+	if matches, _ := filepath.Glob(lockPath + ".stale.*"); len(matches) != 0 {
+		t.Fatalf("a failed restore must not leak sidelined files: %v", matches)
+	}
+}
+
+func TestReclaimStaleLockDropsSidelinedWhenNewHolderWins(t *testing.T) {
+	// An os.ErrExist restore failure means a new holder recreated the lock
+	// path; that is not an error for the caller, and the sidelined file is
+	// dropped rather than leaked.
+	restoreLockFile = func(reclaimed, path string) error { return os.ErrExist }
+	defer func() { restoreLockFile = restoreLiveLock }()
+
+	lockPath := filepath.Join(t.TempDir(), "lock")
+	if err := os.WriteFile(lockPath, []byte("live-holder"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	ok, err := ReclaimStaleLock(lockPath, "tok", func(string) bool { return true })
+	if err != nil || ok {
+		t.Fatalf("losing to a new holder is not an error (ok=%v err=%v)", ok, err)
+	}
+	if matches, _ := filepath.Glob(lockPath + ".stale.*"); len(matches) != 0 {
+		t.Fatalf("the sidelined file must be dropped when a new holder wins: %v", matches)
+	}
+}

internal/lockutil/reclaim_windows_test.go

@@ -0,0 +1,50 @@
+//go:build windows
+
+package lockutil
+
+import (
+	"os"
+	"path/filepath"
+	"testing"
+)
+
+// TestReclaimStaleLockTreatsSharingViolationAsLostRace pins the errno
+// classification the Windows acquire loops depend on: a rename aside that
+// fails because the lock file is concurrently open (Go opens files without
+// FILE_SHARE_DELETE, so the rename fails with ERROR_SHARING_VIOLATION or
+// ERROR_ACCESS_DENIED) is contention, not a hard failure, and must report a
+// benign lost race so the caller keeps waiting.
+func TestReclaimStaleLockTreatsSharingViolationAsLostRace(t *testing.T) {
+	lockPath := filepath.Join(t.TempDir(), "lock")
+	if err := os.WriteFile(lockPath, []byte("held"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	f, err := os.Open(lockPath)
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer f.Close()
+	if ok, rerr := ReclaimStaleLock(lockPath, "tok", func(string) bool { return true }); rerr != nil || ok {
+		t.Fatalf("a sharing-violation rename must report a benign lost race (ok=%v err=%v)", ok, rerr)
+	}
+	if _, err := os.Stat(lockPath); err != nil {
+		t.Fatalf("the lock file must be left in place: %v", err)
+	}
+}
+
+// TestReclaimStaleLockPropagatesRenameFailure pins the fail-fast contract of
+// the rename aside on Windows: a failure that is not contention must reach
+// the caller instead of masquerading as a lost race. A quote is an invalid
+// filename character, so the sidelined name fails with ERROR_INVALID_NAME.
+func TestReclaimStaleLockPropagatesRenameFailure(t *testing.T) {
+	lockPath := filepath.Join(t.TempDir(), "lock")
+	if err := os.WriteFile(lockPath, []byte("stale-holder"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	if ok, err := ReclaimStaleLock(lockPath, `bad"name`, func(string) bool { return false }); err == nil || ok {
+		t.Fatalf("a hard rename failure must surface an error (ok=%v err=%v)", ok, err)
+	}
+	if _, err := os.Stat(lockPath); err != nil {
+		t.Fatalf("the lock file must be left in place on a failed rename: %v", err)
+	}
+}

internal/mcp/config.go

@@ -31,6 +31,9 @@ type Server struct {
 	Auth     string
 	OAuth    *OAuthConfig
 	Identity string
+	// ProjectConfigured is true when project config touched this server. Runtime
+	// credential lookup uses it to avoid reusing legacy user tokens by name.
+	ProjectConfigured bool
 	// UnconfiguredDefault is true when this server is one of Zero's built-in
 	// defaults (e.g. keyless Firecrawl) that the user never touched in their
 	// config — no credentials, no overrides. Callers use it to avoid warning
@@ -88,6 +91,7 @@ func normalizeServer(name string, raw config.MCPServerConfig) (Server, error) {
 		Headers:             copyStringMap(raw.Headers),
 		Auth:                auth,
 		OAuth:               normalizeOAuthConfig(raw.OAuth),
+		ProjectConfigured:   raw.ProjectConfigured,
 		UnconfiguredDefault: config.IsUnconfiguredDefault(name, raw),
 	}
 

internal/mcp/network_client.go

@@ -847,7 +847,7 @@ func (source *storeTokenSource) config() OAuthConfig {
 }
 
 func (source *storeTokenSource) AccessToken(ctx context.Context) (string, error) {
-	token, ok, err := source.store.Load(source.server.Name)
+	token, ok, err := source.store.LoadForServer(source.server)
 	if err != nil {
 		return "", err
 	}
@@ -858,7 +858,7 @@ func (source *storeTokenSource) AccessToken(ctx context.Context) (string, error)
 }
 
 func (source *storeTokenSource) Refresh(ctx context.Context) (string, error) {
-	token, ok, err := source.store.Load(source.server.Name)
+	token, ok, err := source.store.LoadForServer(source.server)
 	if err != nil {
 		return "", err
 	}
@@ -869,7 +869,7 @@ func (source *storeTokenSource) Refresh(ctx context.Context) (string, error) {
 	if err != nil {
 		return "", err
 	}
-	if err := source.store.Save(source.server.Name, refreshed); err != nil {
+	if err := source.store.SaveForServer(source.server, refreshed); err != nil {
 		return "", err
 	}
 	return refreshed.AccessToken, nil

internal/mcp/oauth_store.go

@@ -15,6 +15,9 @@ import (
 // tokenStoreSchemaVersion is the schema of the legacy mcp-oauth-tokens.json file,
 // retained so migration can recognize a file it understands.
 const tokenStoreSchemaVersion = 1
+const mcpServerIdentityLength = 32
+const maxOAuthTokenKeySegmentLength = 128
+const maxMCPServerNameForIdentityToken = maxOAuthTokenKeySegmentLength - 1 - mcpServerIdentityLength
 
 // StoredToken holds the credentials issued by an OAuth 2.0 authorization server
 // for a single MCP server. The token fields are sensitive: they are tagged so
@@ -32,6 +35,7 @@ type StoredToken struct {
 // omits the access and refresh token material so it can be printed by the CLI.
 type TokenStatus struct {
 	ServerName      string    `json:"serverName"`
+	ServerIdentity  string    `json:"serverIdentity,omitempty"`
 	HasToken        bool      `json:"hasToken"`
 	HasRefreshToken bool      `json:"hasRefreshToken"`
 	TokenType       string    `json:"tokenType,omitempty"`
@@ -148,6 +152,17 @@ func (store *TokenStore) Save(serverName string, token StoredToken) error {
 	return store.store.Save(key, storedToOAuth(token))
 }
 
+// SaveForServer persists a token under the server's target identity. Identity
+// binding prevents a project config with the same server name from reusing a
+// token issued for a different MCP endpoint.
+func (store *TokenStore) SaveForServer(server Server, token StoredToken) error {
+	key, err := mcpIdentityKey(server)
+	if err != nil {
+		return err
+	}
+	return store.store.Save(key, storedToOAuth(token))
+}
+
 // Load returns the stored token for a server. The second return value is false
 // when no token has been stored for the server.
 func (store *TokenStore) Load(serverName string) (StoredToken, bool, error) {
@@ -162,6 +177,30 @@ func (store *TokenStore) Load(serverName string) (StoredToken, bool, error) {
 	return tokenToStored(token), true, nil
 }
 
+// LoadForServer returns the token matching the server's target identity. Legacy
+// name-only tokens are migrated only for servers untouched by project config.
+func (store *TokenStore) LoadForServer(server Server) (StoredToken, bool, error) {
+	key, err := mcpIdentityKey(server)
+	if err != nil {
+		return StoredToken{}, false, err
+	}
+	token, ok, err := store.store.Load(key)
+	if err != nil || ok {
+		return tokenToStored(token), ok, err
+	}
+	if server.ProjectConfigured {
+		return StoredToken{}, false, nil
+	}
+	legacy, ok, err := store.Load(server.Name)
+	if err != nil || !ok {
+		return StoredToken{}, ok, err
+	}
+	if err := store.store.Save(key, storedToOAuth(legacy)); err != nil {
+		return StoredToken{}, false, err
+	}
+	return legacy, true, nil
+}
+
 // Delete removes the stored token for a server. It reports whether an entry was
 // present before deletion.
 func (store *TokenStore) Delete(serverName string) (bool, error) {
@@ -172,6 +211,36 @@ func (store *TokenStore) Delete(serverName string) (bool, error) {
 	return store.store.Delete(key)
 }
 
+// DeleteForServerName removes both legacy name-only and identity-bound tokens
+// for a server name.
+func (store *TokenStore) DeleteForServerName(serverName string) (bool, error) {
+	name := strings.TrimSpace(serverName)
+	if err := ValidateServerName(name); err != nil {
+		return false, err
+	}
+	removed, err := store.Delete(name)
+	if err != nil {
+		return false, err
+	}
+	statuses, err := store.store.Status(oauth.KeyPrefixMCP)
+	if err != nil {
+		return removed, err
+	}
+	for _, status := range statuses {
+		segment := strings.TrimPrefix(status.Key, oauth.KeyPrefixMCP)
+		parsedName, _, ok := splitMCPIdentityKeySegment(segment)
+		if !ok || parsedName != name {
+			continue
+		}
+		deleted, err := store.store.Delete(status.Key)
+		if err != nil {
+			return removed, err
+		}
+		removed = removed || deleted
+	}
+	return removed, nil
+}
+
 // Status returns a redaction-safe summary of every stored MCP token, sorted by
 // server name. It never includes the token material.
 func (store *TokenStore) Status() ([]TokenStatus, error) {
@@ -181,8 +250,10 @@ func (store *TokenStore) Status() ([]TokenStatus, error) {
 	}
 	out := make([]TokenStatus, 0, len(statuses))
 	for _, s := range statuses {
+		serverName, serverIdentity, _ := splitMCPIdentityKeySegment(strings.TrimPrefix(s.Key, oauth.KeyPrefixMCP))
 		out = append(out, TokenStatus{
-			ServerName:      strings.TrimPrefix(s.Key, oauth.KeyPrefixMCP),
+			ServerName:      serverName,
+			ServerIdentity:  serverIdentity,
 			HasToken:        s.HasToken,
 			HasRefreshToken: s.HasRefreshToken,
 			TokenType:       s.TokenType,
@@ -250,16 +321,69 @@ func (store *TokenStore) migrateLegacy(legacyPath string) error {
 
 // mcpKey builds and validates the unified store key for an MCP server token.
 func mcpKey(serverName string) (string, error) {
-	if err := ValidateServerName(serverName); err != nil {
+	name := strings.TrimSpace(serverName)
+	if err := ValidateServerName(name); err != nil {
 		return "", err
 	}
-	key := oauth.KeyPrefixMCP + strings.TrimSpace(serverName)
+	if hasMCPIdentitySuffix(name) {
+		return "", fmt.Errorf("MCP OAuth server name %q cannot end with .<32 hex chars> because it collides with identity-bound token storage", name)
+	}
+	key := oauth.KeyPrefixMCP + name
 	if err := oauth.ValidateKey(key); err != nil {
 		return "", err
 	}
 	return key, nil
 }
 
+func mcpIdentityKey(server Server) (string, error) {
+	name := strings.TrimSpace(server.Name)
+	if err := ValidateServerName(name); err != nil {
+		return "", err
+	}
+	if hasMCPIdentitySuffix(name) {
+		return "", fmt.Errorf("MCP OAuth server name %q cannot end with .<32 hex chars> because it collides with identity-bound token storage", name)
+	}
+	identity := strings.TrimSpace(server.Identity)
+	if len(name) > maxMCPServerNameForIdentityToken {
+		return "", fmt.Errorf("MCP OAuth server name %q is too long for identity-bound token storage", name)
+	}
+	if len(identity) != mcpServerIdentityLength || !isMCPServerIdentity(identity) {
+		return "", fmt.Errorf("MCP OAuth server %s has invalid identity %q", name, identity)
+	}
+	key := oauth.KeyPrefixMCP + name + "." + identity
+	if err := oauth.ValidateKey(key); err != nil {
+		return "", err
+	}
+	return key, nil
+}
+
+func splitMCPIdentityKeySegment(segment string) (string, string, bool) {
+	index := strings.LastIndex(segment, ".")
+	if index <= 0 || index == len(segment)-1 {
+		return segment, "", false
+	}
+	identity := segment[index+1:]
+	if len(identity) != mcpServerIdentityLength || !isMCPServerIdentity(identity) {
+		return segment, "", false
+	}
+	return segment[:index], identity, true
+}
+
+func hasMCPIdentitySuffix(value string) bool {
+	_, _, ok := splitMCPIdentityKeySegment(value)
+	return ok
+}
+
+func isMCPServerIdentity(value string) bool {
+	for _, ch := range value {
+		if (ch >= '0' && ch <= '9') || (ch >= 'a' && ch <= 'f') || (ch >= 'A' && ch <= 'F') {
+			continue
+		}
+		return false
+	}
+	return value != ""
+}
+
 // storedToOAuth converts an MCP StoredToken to the shared oauth.Token (the
 // inverse of tokenToStored). MCP never sets the oauth Account field.
 func storedToOAuth(s StoredToken) oauth.Token {
@@ -284,6 +408,11 @@ func FormatTokenStatuses(statuses []TokenStatus) string {
 			builder.WriteByte('\n')
 		}
 		builder.WriteString(status.ServerName)
+		if status.ServerIdentity != "" {
+			builder.WriteString(" [")
+			builder.WriteString(status.ServerIdentity)
+			builder.WriteString("]")
+		}
 		builder.WriteString(": ")
 		if !status.HasToken {
 			builder.WriteString("no token")

internal/mcp/oauth_store_test.go

@@ -1,9 +1,11 @@
 package mcp
 
 import (
+	"context"
 	"os"
 	"path/filepath"
 	"runtime"
+	"strings"
 	"testing"
 	"time"
 )
@@ -43,6 +45,191 @@ func TestTokenStoreRoundTrip(t *testing.T) {
 	}
 }
 
+func TestTokenStoreIdentityBoundRoundTrip(t *testing.T) {
+	store, err := NewTokenStore(TokenStoreOptions{FilePath: filepath.Join(t.TempDir(), "oauth-tokens.json")})
+	if err != nil {
+		t.Fatalf("NewTokenStore() error = %v", err)
+	}
+	server := testOAuthServer("demo", "aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa", false)
+	otherTarget := testOAuthServer("demo", "bbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbb", false)
+	if err := store.SaveForServer(server, StoredToken{AccessToken: "identity-token"}); err != nil {
+		t.Fatalf("SaveForServer() error = %v", err)
+	}
+
+	loaded, ok, err := store.LoadForServer(server)
+	if err != nil || !ok {
+		t.Fatalf("LoadForServer() ok=%v err=%v", ok, err)
+	}
+	if loaded.AccessToken != "identity-token" {
+		t.Fatalf("access token = %q", loaded.AccessToken)
+	}
+	if _, ok, err := store.LoadForServer(otherTarget); err != nil || ok {
+		t.Fatalf("LoadForServer(other target) ok=%v err=%v, want no token", ok, err)
+	}
+}
+
+func TestStoreTokenSourceRequiresMatchingIdentity(t *testing.T) {
+	store, err := NewTokenStore(TokenStoreOptions{FilePath: filepath.Join(t.TempDir(), "oauth-tokens.json")})
+	if err != nil {
+		t.Fatalf("NewTokenStore() error = %v", err)
+	}
+	server := testOAuthServer("demo", "aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa", false)
+	otherTarget := testOAuthServer("demo", "bbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbb", false)
+	if err := store.SaveForServer(server, StoredToken{AccessToken: "access-a"}); err != nil {
+		t.Fatalf("SaveForServer() error = %v", err)
+	}
+
+	source := &storeTokenSource{server: server, store: store}
+	token, err := source.AccessToken(context.Background())
+	if err != nil {
+		t.Fatalf("AccessToken() error = %v", err)
+	}
+	if token != "access-a" {
+		t.Fatalf("token = %q, want access-a", token)
+	}
+
+	source.server = otherTarget
+	if _, err := source.AccessToken(context.Background()); err == nil {
+		t.Fatal("AccessToken() error = nil for mismatched identity")
+	}
+}
+
+func TestTokenStoreProjectConfiguredDoesNotLoadLegacyToken(t *testing.T) {
+	store, err := NewTokenStore(TokenStoreOptions{FilePath: filepath.Join(t.TempDir(), "oauth-tokens.json")})
+	if err != nil {
+		t.Fatalf("NewTokenStore() error = %v", err)
+	}
+	if err := store.Save("demo", StoredToken{AccessToken: "legacy-token"}); err != nil {
+		t.Fatalf("Save() error = %v", err)
+	}
+	server := testOAuthServer("demo", "aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa", true)
+
+	_, ok, err := store.LoadForServer(server)
+	if err != nil {
+		t.Fatalf("LoadForServer() error = %v", err)
+	}
+	if ok {
+		t.Fatal("LoadForServer() ok = true for project-configured legacy token")
+	}
+}
+
+func TestTokenStoreMigratesLegacyForUserOnlyServer(t *testing.T) {
+	store, err := NewTokenStore(TokenStoreOptions{FilePath: filepath.Join(t.TempDir(), "oauth-tokens.json")})
+	if err != nil {
+		t.Fatalf("NewTokenStore() error = %v", err)
+	}
+	if err := store.Save("demo", StoredToken{AccessToken: "legacy-token"}); err != nil {
+		t.Fatalf("Save() error = %v", err)
+	}
+	server := testOAuthServer("demo", "aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa", false)
+
+	loaded, ok, err := store.LoadForServer(server)
+	if err != nil || !ok {
+		t.Fatalf("LoadForServer() ok=%v err=%v", ok, err)
+	}
+	if loaded.AccessToken != "legacy-token" {
+		t.Fatalf("access token = %q", loaded.AccessToken)
+	}
+	if _, err := store.Delete("demo"); err != nil {
+		t.Fatalf("Delete() error = %v", err)
+	}
+	loaded, ok, err = store.LoadForServer(server)
+	if err != nil || !ok || loaded.AccessToken != "legacy-token" {
+		t.Fatalf("identity-bound migrated token missing: loaded=%#v ok=%v err=%v", loaded, ok, err)
+	}
+}
+
+func TestTokenStoreRejectsLongIdentityBoundServerName(t *testing.T) {
+	store, err := NewTokenStore(TokenStoreOptions{FilePath: filepath.Join(t.TempDir(), "oauth-tokens.json")})
+	if err != nil {
+		t.Fatalf("NewTokenStore() error = %v", err)
+	}
+	name := strings.Repeat("a", maxMCPServerNameForIdentityToken+1)
+	err = store.SaveForServer(testOAuthServer(name, "aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa", false), StoredToken{AccessToken: "x"})
+	if err == nil {
+		t.Fatal("SaveForServer() error = nil for long server name")
+	}
+	if !strings.Contains(err.Error(), `MCP OAuth server name "`+name+`" is too long for identity-bound token storage`) {
+		t.Fatalf("error = %q, want long-name message", err.Error())
+	}
+}
+
+func TestTokenStoreRejectsAmbiguousIdentitySuffixServerName(t *testing.T) {
+	store, err := NewTokenStore(TokenStoreOptions{FilePath: filepath.Join(t.TempDir(), "oauth-tokens.json")})
+	if err != nil {
+		t.Fatalf("NewTokenStore() error = %v", err)
+	}
+	name := "demo.aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa"
+
+	if err := store.SaveForServer(testOAuthServer(name, "bbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbb", false), StoredToken{AccessToken: "x"}); err == nil {
+		t.Fatal("SaveForServer() error = nil for ambiguous server name")
+	} else if !strings.Contains(err.Error(), `MCP OAuth server name "`+name+`" cannot end with .<32 hex chars>`) {
+		t.Fatalf("SaveForServer() error = %q, want ambiguous-name rejection", err.Error())
+	}
+	if err := store.Save(name, StoredToken{AccessToken: "legacy"}); err == nil {
+		t.Fatal("Save() error = nil for ambiguous legacy server name")
+	}
+}
+
+func TestTokenStoreDeleteForServerNameRemovesLegacyAndIdentityTokens(t *testing.T) {
+	store, err := NewTokenStore(TokenStoreOptions{FilePath: filepath.Join(t.TempDir(), "oauth-tokens.json")})
+	if err != nil {
+		t.Fatalf("NewTokenStore() error = %v", err)
+	}
+	if err := store.Save("demo", StoredToken{AccessToken: "legacy"}); err != nil {
+		t.Fatal(err)
+	}
+	if err := store.SaveForServer(testOAuthServer("demo", "aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa", false), StoredToken{AccessToken: "a"}); err != nil {
+		t.Fatal(err)
+	}
+	if err := store.SaveForServer(testOAuthServer("demo", "bbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbb", false), StoredToken{AccessToken: "b"}); err != nil {
+		t.Fatal(err)
+	}
+	if err := store.SaveForServer(testOAuthServer("other", "cccccccccccccccccccccccccccccccc", false), StoredToken{AccessToken: "c"}); err != nil {
+		t.Fatal(err)
+	}
+
+	removed, err := store.DeleteForServerName("demo")
+	if err != nil {
+		t.Fatalf("DeleteForServerName() error = %v", err)
+	}
+	if !removed {
+		t.Fatal("DeleteForServerName() removed = false")
+	}
+	statuses, err := store.Status()
+	if err != nil {
+		t.Fatalf("Status() error = %v", err)
+	}
+	if len(statuses) != 1 || statuses[0].ServerName != "other" {
+		t.Fatalf("statuses after delete = %#v, want only other", statuses)
+	}
+}
+
+func TestTokenStoreStatusParsesIdentityAndRedacts(t *testing.T) {
+	store, err := NewTokenStore(TokenStoreOptions{FilePath: filepath.Join(t.TempDir(), "oauth-tokens.json")})
+	if err != nil {
+		t.Fatalf("NewTokenStore() error = %v", err)
+	}
+	identity := "aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa"
+	if err := store.SaveForServer(testOAuthServer("demo", identity, false), StoredToken{AccessToken: "secret-access", RefreshToken: "secret-refresh"}); err != nil {
+		t.Fatalf("SaveForServer() error = %v", err)
+	}
+	statuses, err := store.Status()
+	if err != nil {
+		t.Fatalf("Status() error = %v", err)
+	}
+	if len(statuses) != 1 || statuses[0].ServerName != "demo" || statuses[0].ServerIdentity != identity {
+		t.Fatalf("statuses = %#v, want parsed identity", statuses)
+	}
+	output := FormatTokenStatuses(statuses)
+	if !strings.Contains(output, "demo ["+identity+"]") {
+		t.Fatalf("status output = %q, want identity label", output)
+	}
+	if contains(output, "secret-access") || contains(output, "secret-refresh") {
+		t.Fatalf("status output leaked token: %s", output)
+	}
+}
+
 func TestTokenStoreFileIs0600(t *testing.T) {
 	if runtime.GOOS == "windows" {
 		t.Skip("POSIX file permissions are not enforced on Windows")
@@ -268,6 +455,17 @@ func TestResolveTokenStorePathUsesXDG(t *testing.T) {
 	}
 }
 
+func testOAuthServer(name string, identity string, projectConfigured bool) Server {
+	return Server{
+		Name:              name,
+		Type:              ServerTypeHTTP,
+		URL:               "https://example.com/mcp",
+		Auth:              ServerAuthOAuth,
+		Identity:          identity,
+		ProjectConfigured: projectConfigured,
+	}
+}
+
 func contains(haystack, needle string) bool {
 	return len(haystack) >= len(needle) && (haystack == needle || indexOf(haystack, needle) >= 0)
 }

internal/modelregistry/modes.go

@@ -39,7 +39,7 @@ var defaultModes = []Mode{
 		Description: "Hardest tasks: deep reasoning with a larger turn budget.",
 		Model:       "claude-opus-4.1",
 		Effort:      ReasoningEffortHigh,
-		MaxTurns:    50,
+		MaxTurns:    160,
 	},
 	{
 		Name:        "fast",

internal/modelregistry/vision.go

@@ -58,7 +58,7 @@ func VisionCapableByName(modelID string) bool {
 		strings.Contains(id, "qwen-vl"), strings.Contains(id, "qwenvl"):
 		return true // Qwen VL series are multimodal
 	case strings.Contains(id, "minimax-m3"):
-		return false // MiniMax M3 is text-only (the user confirmed this)
+		return true // MiniMax M3 accepts image input
 	case strings.Contains(id, "llava"), strings.Contains(id, "pixtral"),
 		strings.Contains(id, "internvl"), strings.Contains(id, "minicpm-v"),
 		strings.Contains(id, "moondream"), strings.Contains(id, "bakllava"),

internal/modelregistry/vision_name_test.go

@@ -8,6 +8,7 @@ func TestVisionCapableByName(t *testing.T) {
 		"grok-4.3", "grok-4-fast", "grok-2-vision",
 		"gpt-4o", "gpt-4.1-mini", "gpt-4-turbo", "gpt-5", "o3-mini", "o1",
 		"claude-sonnet-4.5", "claude-3-haiku",
+		"MiniMax-M3",
 		"llava:13b", "qwen2.5-vl-7b", "llama3.2-vision", "pixtral-12b", "moondream",
 	}
 	textOnly := []string{
@@ -38,7 +39,7 @@ func TestSupportsVisionFallbackForUnknownModels(t *testing.T) {
 		t.Error("gpt-4o should be vision (catalog)")
 	}
 	// Unknown-to-catalog but real vision models now pass via the name fallback.
-	for _, m := range []string{"gemini-3-flash-preview", "grok-4.3"} {
+	for _, m := range []string{"gemini-3-flash-preview", "grok-4.3", "MiniMax-M3"} {
 		if !SupportsVision(reg, m) {
 			t.Errorf("%q should be vision via name fallback", m)
 		}

internal/npmwrapper/buildpackages_test.go

@@ -0,0 +1,260 @@
+package npmwrapper
+
+import (
+	"context"
+	"crypto/sha256"
+	"encoding/json"
+	"fmt"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"runtime"
+	"strings"
+	"testing"
+	"time"
+)
+
+// TestBuildPlatformPackagesAssemblesPublishPayloads runs the release-time
+// assembly script against a fixture linux-x64 archive shaped like the output
+// of `zero-release package` and checks the npm publish payloads it emits:
+// the platform payload (suffixed version, os/cpu gate, pack-safe helper
+// shims, pruned agent-browser binaries) and the wrapper (no scripts, no
+// dependencies, full optionalDependencies alias matrix).
+func TestBuildPlatformPackagesAssemblesPublishPayloads(t *testing.T) {
+	if runtime.GOOS == "windows" {
+		t.Skip("fixture uses POSIX symlinks and shell scripts")
+	}
+	node := requireNode(t)
+	tar := requireCommand(t, "tar")
+	version := packageVersion(t)
+
+	root := t.TempDir()
+	staging := filepath.Join(root, "staging")
+
+	// Payload the platform package must keep.
+	writeFixtureFile(t, filepath.Join(staging, "zero"), "#!/usr/bin/env sh\necho fixture-zero\n", 0o755)
+	writeFixtureFile(t, filepath.Join(staging, "zero-linux-sandbox"), "#!/usr/bin/env sh\n", 0o755)
+	writeFixtureFile(t, filepath.Join(staging, "zero-seccomp"), "#!/usr/bin/env sh\n", 0o755)
+	// Wrapper-owned files the platform package must exclude.
+	writeFixtureFile(t, filepath.Join(staging, "package.json"), `{"name":"@gitlawb/zero"}`, 0o644)
+	writeFixtureFile(t, filepath.Join(staging, "README.md"), "readme\n", 0o644)
+	writeFixtureFile(t, filepath.Join(staging, "VERSION"), version+"\n", 0o644)
+	writeFixtureFile(t, filepath.Join(staging, "bin", "zero.js"), "#!/usr/bin/env node\n", 0o755)
+	// Vendored helpers tree with npm's symlink .bin shims and agent-browser's
+	// multi-platform binary set.
+	agentBrowserDir := filepath.Join(staging, "helpers", "node_modules", "agent-browser")
+	writeFixtureFile(t, filepath.Join(agentBrowserDir, "package.json"), `{"name":"agent-browser"}`, 0o644)
+	writeFixtureFile(t, filepath.Join(agentBrowserDir, "bin", "agent-browser.js"), "#!/usr/bin/env node\nconsole.log('fixture agent-browser')\n", 0o644)
+	for _, name := range []string{
+		"agent-browser-linux-x64",
+		"agent-browser-linux-musl-x64",
+		"agent-browser-linux-arm64",
+		"agent-browser-darwin-arm64",
+		"agent-browser-darwin-x64",
+		"agent-browser-win32-x64.exe",
+	} {
+		writeFixtureFile(t, filepath.Join(agentBrowserDir, "bin", name), "native\n", 0o755)
+	}
+	tuistoryDir := filepath.Join(staging, "helpers", "node_modules", "tuistory")
+	writeFixtureFile(t, filepath.Join(tuistoryDir, "dist", "cli.js"), "#!/usr/bin/env node\n", 0o644)
+	binDir := filepath.Join(staging, "helpers", "node_modules", ".bin")
+	mustMkdirAllFixture(t, binDir)
+	if err := os.Symlink("../agent-browser/bin/agent-browser.js", filepath.Join(binDir, "agent-browser")); err != nil {
+		t.Fatalf("Symlink agent-browser shim: %v", err)
+	}
+	if err := os.Symlink("../tuistory/dist/cli.js", filepath.Join(binDir, "tuistory")); err != nil {
+		t.Fatalf("Symlink tuistory shim: %v", err)
+	}
+
+	// Archive with the payload at the archive root, like zero-release's
+	// createArchive, plus the sha256 sidecar the script verifies against.
+	assetName := fmt.Sprintf("zero-v%s-linux-x64.tar.gz", version)
+	archivePath := filepath.Join(root, assetName)
+	ctx, cancel := context.WithTimeout(context.Background(), 30*time.Second)
+	defer cancel()
+	if output, err := exec.CommandContext(ctx, tar, "-C", staging, "-czf", archivePath, ".").CombinedOutput(); err != nil {
+		t.Fatalf("tar failed: %v\n%s", err, output)
+	}
+	archiveBytes, err := os.ReadFile(archivePath)
+	if err != nil {
+		t.Fatalf("ReadFile archive: %v", err)
+	}
+	digest := sha256.Sum256(archiveBytes)
+	writeFixtureFile(t, archivePath+".sha256", fmt.Sprintf("%x  %s\n", digest, assetName), 0o644)
+
+	outDir := filepath.Join(root, "out")
+	script := filepath.Join(repoRoot(t), "scripts", "npm", "build-platform-packages.mjs")
+	runCtx, runCancel := context.WithTimeout(context.Background(), 60*time.Second)
+	defer runCancel()
+	command := exec.CommandContext(runCtx, node, script, "--artifacts-dir", root, "--out-dir", outDir, "--only", "linux-x64")
+	command.Env = append(os.Environ(), "NODE_OPTIONS=")
+	if output, err := command.CombinedOutput(); err != nil {
+		t.Fatalf("build-platform-packages failed: %v\n%s", err, output)
+	}
+
+	platformDir := filepath.Join(outDir, "platforms", "zero-linux-x64")
+	var platformPkg struct {
+		Name    string   `json:"name"`
+		Version string   `json:"version"`
+		OS      []string `json:"os"`
+		CPU     []string `json:"cpu"`
+		Scripts map[string]string
+	}
+	unmarshalJSONFile(t, filepath.Join(platformDir, "package.json"), &platformPkg)
+	if platformPkg.Name != "@gitlawb/zero" {
+		t.Fatalf("platform package name = %q, want @gitlawb/zero (same-name suffixed versions share one trusted publisher)", platformPkg.Name)
+	}
+	if want := version + "-linux-x64"; platformPkg.Version != want {
+		t.Fatalf("platform package version = %q, want %q", platformPkg.Version, want)
+	}
+	if len(platformPkg.OS) != 1 || platformPkg.OS[0] != "linux" || len(platformPkg.CPU) != 1 || platformPkg.CPU[0] != "x64" {
+		t.Fatalf("platform package os/cpu = %v/%v, want [linux]/[x64]", platformPkg.OS, platformPkg.CPU)
+	}
+	if len(platformPkg.Scripts) != 0 {
+		t.Fatalf("platform package has lifecycle scripts: %v", platformPkg.Scripts)
+	}
+
+	info, err := os.Stat(filepath.Join(platformDir, "zero"))
+	if err != nil {
+		t.Fatalf("platform package binary missing: %v", err)
+	}
+	if info.Mode()&0o111 == 0 {
+		t.Fatalf("platform package binary is not executable: %v", info.Mode())
+	}
+	for _, excluded := range []string{"README.md", "VERSION", "bin"} {
+		if _, err := os.Stat(filepath.Join(platformDir, excluded)); !os.IsNotExist(err) {
+			t.Fatalf("platform package still contains wrapper-owned %q", excluded)
+		}
+	}
+	// npm only ships a license file it finds in the packed directory, so the
+	// assembly must place the repo LICENSE into every payload.
+	if _, err := os.Stat(filepath.Join(platformDir, "LICENSE")); err != nil {
+		t.Fatalf("platform package missing LICENSE: %v", err)
+	}
+
+	// npm pack drops symlinks, so the shim must now be a regular executable file.
+	shimPath := filepath.Join(platformDir, "helpers", "node_modules", ".bin", "agent-browser")
+	shimInfo, err := os.Lstat(shimPath)
+	if err != nil {
+		t.Fatalf("agent-browser shim missing: %v", err)
+	}
+	if shimInfo.Mode()&os.ModeSymlink != 0 {
+		t.Fatalf("agent-browser shim is still a symlink; npm pack would drop it")
+	}
+	if shimInfo.Mode()&0o111 == 0 {
+		t.Fatalf("agent-browser shim is not executable: %v", shimInfo.Mode())
+	}
+	shim, err := os.ReadFile(shimPath)
+	if err != nil {
+		t.Fatalf("ReadFile shim: %v", err)
+	}
+	if !strings.Contains(string(shim), "../agent-browser/bin/agent-browser.js") {
+		t.Fatalf("shim does not exec the vendored helper: %q", string(shim))
+	}
+	if _, err := os.Stat(filepath.Join(platformDir, "helpers", "node_modules", ".bin", "tuistory")); err != nil {
+		t.Fatalf("tuistory shim missing: %v", err)
+	}
+
+	// Pruning: only the linux-x64 (glibc + musl) agent-browser binaries survive.
+	prunedBinDir := filepath.Join(platformDir, "helpers", "node_modules", "agent-browser", "bin")
+	entries, err := os.ReadDir(prunedBinDir)
+	if err != nil {
+		t.Fatalf("ReadDir pruned bin: %v", err)
+	}
+	got := []string{}
+	for _, entry := range entries {
+		got = append(got, entry.Name())
+	}
+	want := map[string]bool{
+		"agent-browser.js":             true,
+		"agent-browser-linux-x64":      true,
+		"agent-browser-linux-musl-x64": true,
+	}
+	if len(got) != len(want) {
+		t.Fatalf("pruned agent-browser bin = %v, want exactly %v", got, want)
+	}
+	for _, name := range got {
+		if !want[name] {
+			t.Fatalf("pruning left unexpected %q (full set: %v)", name, got)
+		}
+	}
+
+	wrapperDir := filepath.Join(outDir, "wrapper")
+	var wrapperPkg struct {
+		Name         string            `json:"name"`
+		Version      string            `json:"version"`
+		Scripts      map[string]string `json:"scripts"`
+		Dependencies map[string]string `json:"dependencies"`
+		Optional     map[string]string `json:"optionalDependencies"`
+	}
+	unmarshalJSONFile(t, filepath.Join(wrapperDir, "package.json"), &wrapperPkg)
+	if wrapperPkg.Version != version {
+		t.Fatalf("wrapper version = %q, want %q", wrapperPkg.Version, version)
+	}
+	if len(wrapperPkg.Scripts) != 0 {
+		t.Fatalf("published wrapper has lifecycle scripts: %v", wrapperPkg.Scripts)
+	}
+	if len(wrapperPkg.Dependencies) != 0 {
+		t.Fatalf("published wrapper has dependencies: %v (helpers are vendored in the platform payloads)", wrapperPkg.Dependencies)
+	}
+	wantAliases := map[string]string{
+		"@gitlawb/zero-darwin-arm64": "npm:@gitlawb/zero@" + version + "-darwin-arm64",
+		"@gitlawb/zero-darwin-x64":   "npm:@gitlawb/zero@" + version + "-darwin-x64",
+		"@gitlawb/zero-linux-arm64":  "npm:@gitlawb/zero@" + version + "-linux-arm64",
+		"@gitlawb/zero-linux-x64":    "npm:@gitlawb/zero@" + version + "-linux-x64",
+		"@gitlawb/zero-win32-x64":    "npm:@gitlawb/zero@" + version + "-win32-x64",
+	}
+	if len(wrapperPkg.Optional) != len(wantAliases) {
+		t.Fatalf("wrapper optionalDependencies = %v, want %v", wrapperPkg.Optional, wantAliases)
+	}
+	for alias, spec := range wantAliases {
+		if wrapperPkg.Optional[alias] != spec {
+			t.Fatalf("wrapper optionalDependencies[%q] = %q, want %q", alias, wrapperPkg.Optional[alias], spec)
+		}
+	}
+	for _, file := range []string{"bin/zero.js", "scripts/postinstall.mjs", "README.md", "LICENSE"} {
+		if _, err := os.Stat(filepath.Join(wrapperDir, filepath.FromSlash(file))); err != nil {
+			t.Fatalf("wrapper payload missing %s: %v", file, err)
+		}
+	}
+}
+
+func unmarshalJSONFile(t *testing.T, path string, target any) {
+	t.Helper()
+	data, err := os.ReadFile(path)
+	if err != nil {
+		t.Fatalf("ReadFile %s: %v", path, err)
+	}
+	if err := json.Unmarshal(data, target); err != nil {
+		t.Fatalf("Unmarshal %s: %v", path, err)
+	}
+}
+
+func writeFixtureFile(t *testing.T, path string, content string, mode os.FileMode) {
+	t.Helper()
+	mustMkdirAllFixture(t, filepath.Dir(path))
+	if err := os.WriteFile(path, []byte(content), mode); err != nil {
+		t.Fatalf("WriteFile %s: %v", path, err)
+	}
+	if mode&0o111 != 0 {
+		if err := os.Chmod(path, mode); err != nil {
+			t.Fatalf("Chmod %s: %v", path, err)
+		}
+	}
+}
+
+func mustMkdirAllFixture(t *testing.T, path string) {
+	t.Helper()
+	if err := os.MkdirAll(path, 0o755); err != nil {
+		t.Fatalf("MkdirAll %s: %v", path, err)
+	}
+}
+
+func requireCommand(t *testing.T, name string) string {
+	t.Helper()
+	path, err := exec.LookPath(name)
+	if err != nil {
+		t.Skipf("%s is not available", name)
+	}
+	return path
+}

internal/npmwrapper/npmwrapper_test.go

@@ -41,16 +41,13 @@ func TestPackageBinPointsToNodeWrapper(t *testing.T) {
 	if pkg.Module != "bin/zero.js" {
 		t.Fatalf("module = %q, want bin/zero.js", pkg.Module)
 	}
-	// Only a postinstall hook (which downloads the prebuilt binary) is allowed.
-	// Repository build scripts (build/prepare/prepack/…) must not ship in the
-	// published package — the tarball has no Go source to build from.
-	if pkg.Scripts["postinstall"] != "node scripts/postinstall.mjs" {
-		t.Fatalf("scripts.postinstall = %q, want node scripts/postinstall.mjs", pkg.Scripts["postinstall"])
-	}
+	// The published package must carry NO lifecycle scripts at all: the native
+	// binary arrives as a platform optionalDependency, with scripts/postinstall.mjs
+	// kept only as a first-run fallback the wrapper invokes itself. Any script
+	// here would resurrect the npm/Bun/pnpm install-script warnings the platform
+	//-package model exists to eliminate (see docs/NPM_PACKAGING.md).
 	for name := range pkg.Scripts {
-		if name != "postinstall" {
-			t.Fatalf("package.json scripts contains %q; only a postinstall hook is allowed (no repository build scripts)", name)
-		}
+		t.Fatalf("package.json scripts contains %q; the published package must have no lifecycle scripts", name)
 	}
 	if pkg.License == "" {
 		t.Fatalf("package.json license is empty; set it (ties to the pending LICENSE file) so npm publish is not unlicensed")
@@ -256,11 +253,146 @@ func TestNodeWrapperReportsMissingNativeBinary(t *testing.T) {
 	if !ok || exitErr.ExitCode() != 1 {
 		t.Fatalf("wrapper err = %v, want exit 1; output: %s", err, output)
 	}
-	if !strings.Contains(string(output), "No native binary found next to the npm wrapper") {
+	if !strings.Contains(string(output), "No native binary is available for this install") {
 		t.Fatalf("missing-native output = %q", string(output))
 	}
 }
 
+func TestNodeWrapperPrefersPlatformPackageBinary(t *testing.T) {
+	if runtime.GOOS == "windows" {
+		t.Skip("mock executable fixture uses a POSIX shell script")
+	}
+	node := requireNode(t)
+	wrapperPath := copyWrapperFixture(t)
+	root := filepath.Dir(filepath.Dir(wrapperPath))
+
+	// A stale downloaded binary next to the wrapper must lose to the platform
+	// package: the platform version is pinned to the wrapper release, the
+	// downloaded copy is whatever a previous fallback fetched.
+	if err := os.WriteFile(filepath.Join(root, "zero"), []byte("#!/usr/bin/env sh\nprintf 'downloaded-zero\\n'\n"), 0o755); err != nil {
+		t.Fatalf("WriteFile downloaded fixture: %v", err)
+	}
+
+	platformDir := filepath.Join(root, "node_modules", "@gitlawb", "zero-"+nodePlatformName()+"-"+nodeArchName())
+	if err := os.MkdirAll(platformDir, 0o755); err != nil {
+		t.Fatalf("MkdirAll platform package: %v", err)
+	}
+	if err := os.WriteFile(filepath.Join(platformDir, "package.json"), []byte(`{"name":"@gitlawb/zero"}`), 0o644); err != nil {
+		t.Fatalf("WriteFile platform package.json: %v", err)
+	}
+	if err := os.WriteFile(filepath.Join(platformDir, "zero"), []byte("#!/usr/bin/env sh\nprintf 'platform-zero'; for arg in \"$@\"; do printf ' %s' \"$arg\"; done; printf '\\n'\n"), 0o755); err != nil {
+		t.Fatalf("WriteFile platform binary: %v", err)
+	}
+
+	ctx, cancel := context.WithTimeout(context.Background(), nodeWrapperTimeout())
+	defer cancel()
+	command := nodeWrapperCommand(ctx, node, wrapperPath, "--version")
+	output, err := command.CombinedOutput()
+	if ctx.Err() != nil {
+		t.Fatalf("wrapper timed out launching platform binary: %v; output: %s", ctx.Err(), output)
+	}
+	if err != nil {
+		t.Fatalf("wrapper returned error: %v; output: %s", err, output)
+	}
+	if got := strings.TrimSpace(string(output)); got != "platform-zero --version" {
+		t.Fatalf("wrapper output = %q, want platform-zero --version", got)
+	}
+}
+
+func TestNodeWrapperFallsBackToDownloadedBinary(t *testing.T) {
+	if runtime.GOOS == "windows" {
+		t.Skip("mock executable fixture uses a POSIX shell script")
+	}
+	node := requireNode(t)
+	wrapperPath := copyWrapperFixture(t)
+	root := filepath.Dir(filepath.Dir(wrapperPath))
+	if err := os.WriteFile(filepath.Join(root, "zero"), []byte("#!/usr/bin/env sh\nprintf 'downloaded-zero\\n'\n"), 0o755); err != nil {
+		t.Fatalf("WriteFile downloaded fixture: %v", err)
+	}
+
+	ctx, cancel := context.WithTimeout(context.Background(), nodeWrapperTimeout())
+	defer cancel()
+	command := nodeWrapperCommand(ctx, node, wrapperPath, "--version")
+	output, err := command.CombinedOutput()
+	if ctx.Err() != nil {
+		t.Fatalf("wrapper timed out launching downloaded binary: %v; output: %s", ctx.Err(), output)
+	}
+	if err != nil {
+		t.Fatalf("wrapper returned error: %v; output: %s", err, output)
+	}
+	if got := strings.TrimSpace(string(output)); got != "downloaded-zero" {
+		t.Fatalf("wrapper output = %q, want downloaded-zero", got)
+	}
+}
+
+func TestNodeWrapperRunsFirstRunDownloaderWhenBinaryMissing(t *testing.T) {
+	node := requireNode(t)
+	wrapperPath := copyWrapperFixture(t)
+	root := filepath.Dir(filepath.Dir(wrapperPath))
+
+	// Give the fixture the real downloader so the wrapper's first-run path
+	// executes it; ZERO_SKIP_DOWNLOAD keeps the test offline, so the download
+	// "succeeds" without producing a binary and the wrapper must exit 1 with
+	// guidance.
+	scriptsDir := filepath.Join(root, "scripts")
+	if err := os.MkdirAll(scriptsDir, 0o755); err != nil {
+		t.Fatalf("MkdirAll scripts: %v", err)
+	}
+	postinstall, err := os.ReadFile(filepath.Join(repoRoot(t), "scripts", "postinstall.mjs"))
+	if err != nil {
+		t.Fatalf("ReadFile postinstall: %v", err)
+	}
+	if err := os.WriteFile(filepath.Join(scriptsDir, "postinstall.mjs"), postinstall, 0o644); err != nil {
+		t.Fatalf("WriteFile postinstall fixture: %v", err)
+	}
+	if err := os.WriteFile(filepath.Join(root, "package.json"), []byte(`{"type":"module","name":"@gitlawb/zero","version":"0.0.0"}`), 0o644); err != nil {
+		t.Fatalf("WriteFile package fixture: %v", err)
+	}
+
+	ctx, cancel := context.WithTimeout(context.Background(), nodeWrapperTimeout())
+	defer cancel()
+	command := nodeWrapperCommand(ctx, node, wrapperPath, "--version")
+	command.Env = append(command.Env, "ZERO_SKIP_DOWNLOAD=1")
+	output, err := command.CombinedOutput()
+	if ctx.Err() != nil {
+		t.Fatalf("wrapper timed out on first-run download path: %v; output: %s", ctx.Err(), output)
+	}
+	exitErr, ok := err.(*exec.ExitError)
+	if !ok || exitErr.ExitCode() != 1 {
+		t.Fatalf("wrapper err = %v, want exit 1; output: %s", err, output)
+	}
+	text := string(output)
+	if !strings.Contains(text, "fetching the native binary from the GitHub Release") {
+		t.Fatalf("output missing fallback download notice: %q", text)
+	}
+	if !strings.Contains(text, "skipping native binary download") {
+		t.Fatalf("output shows the downloader did not run: %q", text)
+	}
+	if !strings.Contains(text, "No native binary is available for this install") {
+		t.Fatalf("output missing guidance: %q", text)
+	}
+}
+
+func nodePlatformName() string {
+	switch runtime.GOOS {
+	case "darwin":
+		return "darwin"
+	case "windows":
+		return "win32"
+	default:
+		return runtime.GOOS
+	}
+}
+
+func nodeArchName() string {
+	switch runtime.GOARCH {
+	case "amd64":
+		return "x64"
+	default:
+		return runtime.GOARCH
+	}
+}
+
 func TestNodeWrapperLaunchesNativeBinary(t *testing.T) {
 	if runtime.GOOS == "windows" {
 		t.Skip("mock executable fixture uses a POSIX shell script")

internal/oauth/lock.go

@@ -7,6 +7,8 @@ import (
 	"path/filepath"
 	"sync/atomic"
 	"time"
+
+	"github.com/Gitlawb/zero/internal/lockutil"
 )
 
 const (
@@ -38,11 +40,11 @@ func acquireFileLock(lockPath string, now func() time.Time) (func(), error) {
 			// other processes. Fail closed: remove the file and surface the error.
 			if _, werr := f.WriteString(token); werr != nil {
 				_ = f.Close()
-				_ = os.Remove(lockPath)
+				_ = lockutil.RemoveLockFile(lockPath)
 				return nil, fmt.Errorf("oauth: write token lock: %w", werr)
 			}
 			if cerr := f.Close(); cerr != nil {
-				_ = os.Remove(lockPath)
+				_ = lockutil.RemoveLockFile(lockPath)
 				return nil, fmt.Errorf("oauth: close token lock: %w", cerr)
 			}
 			var released bool
@@ -52,7 +54,7 @@ func acquireFileLock(lockPath string, now func() time.Time) (func(), error) {
 				}
 				released = true
 				if data, rerr := os.ReadFile(lockPath); rerr == nil && string(data) == token {
-					_ = os.Remove(lockPath)
+					_ = lockutil.RemoveLockFile(lockPath)
 				}
 			}, nil
 		}
@@ -69,7 +71,18 @@ func acquireFileLock(lockPath string, now func() time.Time) (func(), error) {
 		// reclaimStaleLock renames the file aside (only one rename wins) and restores
 		// it if it turns out fresh, so a live lock is never deleted out from under it.
 		if info, statErr := os.Stat(lockPath); statErr == nil && time.Since(info.ModTime()) > fileLockStaleAfter {
-			if reclaimStaleLock(lockPath, token, fileLockStaleAfter) {
+			cleared, rerr := lockutil.ReclaimStaleLock(lockPath, token, func(reclaimedPath string) bool {
+				info, err := os.Stat(reclaimedPath)
+				return err == nil && time.Since(info.ModTime()) <= fileLockStaleAfter
+			})
+			if rerr != nil {
+				// Reclaim hit a hard failure: the rename aside failed outright, or a
+				// live holder's lock could not be put back (the lock path may be
+				// missing, so re-acquiring would break mutual exclusion). Fail closed
+				// instead of spinning to the deadline.
+				return nil, fmt.Errorf("oauth: reclaim stale token lock: %w", rerr)
+			}
+			if cleared {
 				continue
 			}
 			// Lost the reclaim race (or it was actually fresh) — fall through to the
@@ -81,21 +94,3 @@ func acquireFileLock(lockPath string, now func() time.Time) (func(), error) {
 		time.Sleep(10 * time.Millisecond)
 	}
 }
-
-// reclaimStaleLock atomically reclaims a stale lock file: it renames the lock to a
-// per-acquirer unique name (only one racer can rename a given file), then verifies
-// the moved file is still stale; if a holder reacquired it in the gap it is
-// restored rather than stolen. Returns true only when a genuinely stale lock was
-// removed. Mirrors internal/cron/lock.go and internal/hooks/lock.go.
-func reclaimStaleLock(lockPath, token string, staleAfter time.Duration) bool {
-	reclaimed := fmt.Sprintf("%s.stale.%s", lockPath, token)
-	if err := os.Rename(lockPath, reclaimed); err != nil {
-		return false
-	}
-	if info, err := os.Stat(reclaimed); err == nil && time.Since(info.ModTime()) <= staleAfter {
-		_ = os.Rename(reclaimed, lockPath)
-		return false
-	}
-	_ = os.Remove(reclaimed)
-	return true
-}

internal/perfbench/MANIFEST.md

@@ -0,0 +1,71 @@
+# Turn benchmark manifest
+
+The baseline manifest (`manifests/baseline.json`) is the per-turn benchmark's
+program keystone: it defines the tasks the harness runs, the workspace each
+starts in, and — critically — what "pass" means for each task. Because `make
+baseline` is re-run on every perf change, the manifest's pass/fail contract
+matters for months, so the contract is written down here.
+
+## Task count
+
+48 tasks across seven classes:
+
+| class     | count | oracle                      | tier        |
+|-----------|-------|-----------------------------|-------------|
+| nav       | 10    | none                        | latency     |
+| edit      | 10    | substring grep              | correctness |
+| fix       | 8     | scoped `go test -run <name>` | correctness |
+| refactor  | 6     | `go build ./...`            | build       |
+| longproc  | 4     | none                        | latency     |
+| longctx   | 4     | none                        | latency     |
+| parallel  | 6     | none                        | latency     |
+| **total** | **48**|                             |             |
+
+## Oracle tiers
+
+Pass/fail is reported per tier so the report cannot be misread as a blanket
+correctness verdict. The tier is decided per task from oracle presence and the
+manifest's `buildOnlyClasses` list:
+
+- **Correctness** (18 tasks: 10 edit + 8 fix) — a positive oracle (substring
+  grep that the requested change landed, or a scoped `go test -run` that the
+  bug fix passes). This is the only pass rate that can move with model quality:
+  `tasksVerified` / `tasksPassed` / `correctnessPassRate`.
+- **Build-only** (6 tasks: refactor) — `go build ./...` proves the edit
+  compiles, not that the refactor achieved its goal (a no-op refactor passes).
+  The manifest declares `buildOnlyClasses: ["refactor"]`. Reported as
+  `buildCheckedTasks` / `buildPassedTasks` / `buildPassRate`, never in
+  `correctnessPassRate`.
+- **Latency-only** (24 tasks: nav, longproc, longctx, parallel) — no
+  `verificationCommand`. An exit 0 only proves the turn ran, not that the answer
+  was right. They contribute to latency and span attribution only and are
+  counted in `latencyOnlyTasks`, never in any pass rate.
+
+A task's tier is driven by oracle **presence** first: a task with no
+`verificationCommand` is latency-only even if its class is listed in
+`buildOnlyClasses`, so a missing oracle can never silently pass on exit 0.
+
+## Known limitations (deferred)
+
+These are accepted for the Phase 0 baseline and tracked in a follow-up issue;
+they do not block the baseline because the tier split keeps the report honest:
+
+- The edit grep oracles are substring checks. A rename oracle asserts the new
+  name is present AND the old name is gone (compound `bash -c`), but an
+  add-field oracle (`grep -R Label .`) only proves the string appears, not that
+  it landed on the right struct. Strengthening these to `go vet`/`go build` +
+  structural assertions is follow-up work.
+- The refactor build oracle is non-positive (a no-op refactor passes). A
+  structural verifier that proves the refactor happened is follow-up work;
+  until then refactor is `buildPassRate`, not `correctnessPassRate`.
+- The 24 read-only tasks have no oracle. Adding deterministic oracles for nav
+  (diff against an expected answer) and documenting longproc/longctx/parallel
+  as permanently latency-only is follow-up work.
+
+## Fixtures
+
+Each task's `workspaceFixture` points at a small self-contained workspace under
+`testdata/` so the suite runs offline and repeatably. Mutating tasks (edit,
+fix, refactor) run against a per-invocation **copy** of their fixture, so the
+checked-in fixtures stay clean and one task's edits can't bleed into the next
+iteration or a later task.

internal/perfbench/manifests/baseline.json

@@ -0,0 +1,62 @@
+{
+  "id": "zero-baseline-turn",
+  "name": "Zero per-turn baseline (Phase 0)",
+  "description": "Latency-first baseline: measures where a turn spends wall time, with pass/fail reported per oracle tier so it cannot be misread as a blanket correctness verdict. Read-only classes (nav, longproc, longctx, parallel) carry no verification oracle — a zero-exit only proves the turn ran, so they count as latency-only and never in any pass rate. Correctness classes (edit, fix) carry a positive oracle (substring grep or a scoped `go test`) and are the only tasks in correctnessPassRate. Build-only classes (refactor) carry `go build ./...`, which proves the edit compiles but not that the refactor achieved its goal, so they are reported as buildPassRate and excluded from correctnessPassRate.",
+  "buildOnlyClasses": ["refactor"],
+  "tasks": [
+    {"id": "nav-01", "name": "list repo files", "class": "nav", "prompt": "List the files in this repository and report their count. Do not modify anything.", "workspaceFixture": "../testdata/nav"},
+    {"id": "nav-02", "name": "find a function", "class": "nav", "prompt": "Find the function named 'greet' and report the file and line it is defined on. Do not modify anything.", "workspaceFixture": "../testdata/nav"},
+    {"id": "nav-03", "name": "summarize module", "class": "nav", "prompt": "Read main.go and summarize what this program does in two sentences. Do not modify anything.", "workspaceFixture": "../testdata/nav"},
+    {"id": "nav-04", "name": "count tests", "class": "nav", "prompt": "Count the test functions (functions whose names start with Test) in this repository and report the number. Do not modify anything.", "workspaceFixture": "../testdata/nav"},
+    {"id": "nav-05", "name": "locate a string", "class": "nav", "prompt": "Find every occurrence of the string 'TODO' in the repository and report the files. Do not modify anything.", "workspaceFixture": "../testdata/nav"},
+    {"id": "nav-06", "name": "describe package", "class": "nav", "prompt": "Describe the exported API of this package by reading the source. Do not modify anything.", "workspaceFixture": "../testdata/nav"},
+    {"id": "nav-07", "name": "trace a call", "class": "nav", "prompt": "Trace which function calls which other function starting from main. Do not modify anything.", "workspaceFixture": "../testdata/nav"},
+    {"id": "nav-08", "name": "find imports", "class": "nav", "prompt": "List every third-party import used by this repository. Do not modify anything.", "workspaceFixture": "../testdata/nav"},
+    {"id": "nav-09", "name": "report config", "class": "nav", "prompt": "Read config.json and report the keys it contains. Do not modify anything.", "workspaceFixture": "../testdata/nav"},
+    {"id": "nav-10", "name": "doc lookup", "class": "nav", "prompt": "Find the doc comment on the greet function and quote it. Do not modify anything.", "workspaceFixture": "../testdata/nav"},
+
+    {"id": "edit-01", "name": "rename a constant", "class": "edit", "prompt": "Rename the constant MaxRetries to RetryLimit everywhere it appears.", "workspaceFixture": "../testdata/edit", "verificationCommand": ["bash", "-c", "grep -R RetryLimit . && ! grep -R MaxRetries ."]},
+    {"id": "edit-02", "name": "fix a typo", "class": "edit", "prompt": "Fix the typo in the comment above the greet function ('receieves' should be 'receives').", "workspaceFixture": "../testdata/edit", "verificationCommand": ["bash", "-c", "! grep -R receieves ."]},
+    {"id": "edit-03", "name": "add a field", "class": "edit", "prompt": "Add a string field named Label to the Config struct in main.go.", "workspaceFixture": "../testdata/edit", "verificationCommand": ["grep", "-R", "Label", "."]},
+    {"id": "edit-04", "name": "bump a version", "class": "edit", "prompt": "Change the version string in version.go from 1.0.0 to 1.1.0.", "workspaceFixture": "../testdata/edit", "verificationCommand": ["bash", "-c", "grep -R 1.1.0 . && ! grep -R 1.0.0 ."]},
+    {"id": "edit-05", "name": "remove a log line", "class": "edit", "prompt": "Remove the line that prints 'debug: starting'.", "workspaceFixture": "../testdata/edit", "verificationCommand": ["bash", "-c", "! grep -R 'debug: starting' ."]},
+    {"id": "edit-06", "name": "add a license header", "class": "edit", "prompt": "Add a comment line '// SPDX-License-Identifier: MIT' at the top of main.go.", "workspaceFixture": "../testdata/edit", "verificationCommand": ["grep", "-R", "SPDX-License-Identifier", "."]},
+    {"id": "edit-07", "name": "change a default", "class": "edit", "prompt": "Change the default value of the Port constant from 8080 to 9090.", "workspaceFixture": "../testdata/edit", "verificationCommand": ["bash", "-c", "grep -R 9090 . && ! grep -R 8080 ."]},
+    {"id": "edit-08", "name": "add a getter", "class": "edit", "prompt": "Add a method func (c *Config) GetLabel() string that returns c.Label.", "workspaceFixture": "../testdata/edit", "verificationCommand": ["bash", "-c", "grep -R 'func (c *Config) GetLabel' ."]},
+    {"id": "edit-09", "name": "wrap an error", "class": "edit", "prompt": "In the load function, wrap the returned error with fmt.Errorf so it includes 'load failed'.", "workspaceFixture": "../testdata/edit", "verificationCommand": ["grep", "-R", "load failed", "."]},
+    {"id": "edit-10", "name": "update a message", "class": "edit", "prompt": "Change the printed greeting from 'hello' to 'hello, world'.", "workspaceFixture": "../testdata/edit", "verificationCommand": ["grep", "-R", "hello, world", "."]},
+
+    {"id": "fix-01", "name": "off-by-one", "class": "fix", "prompt": "In bugs.go the Sum function has an off-by-one error that skips the last element; fix the loop so it sums all elements.", "workspaceFixture": "../testdata/fix", "verificationCommand": ["go", "test", "-run", "TestSumIncludesLastElement", "./..."]},
+    {"id": "fix-02", "name": "nil deref", "class": "fix", "prompt": "In bugs.go DefaultConfig returns nil and Port() dereferences it; make DefaultConfig return a valid Config with Port 8080.", "workspaceFixture": "../testdata/fix", "verificationCommand": ["go", "test", "-run", "TestPortReturnsDefault", "./..."]},
+    {"id": "fix-03", "name": "wrong operator", "class": "fix", "prompt": "In bugs.go Max uses the wrong inequality and returns the smaller value; fix it to return the larger.", "workspaceFixture": "../testdata/fix", "verificationCommand": ["go", "test", "-run", "TestMaxReturnsLarger", "./..."]},
+    {"id": "fix-04", "name": "truncated read", "class": "fix", "prompt": "In bugs.go ReadFirst returns only a single character instead of the full content; fix it to return the whole string.", "workspaceFixture": "../testdata/fix", "verificationCommand": ["go", "test", "-run", "TestReadFirstReturnsFullContent", "./..."]},
+    {"id": "fix-05", "name": "swapped args", "class": "fix", "prompt": "In bugs.go Divide swaps its operands; fix the order so Divide(10,2) returns 5.", "workspaceFixture": "../testdata/fix", "verificationCommand": ["go", "test", "-run", "TestDivideOrder", "./..."]},
+    {"id": "fix-06", "name": "missing return", "class": "fix", "prompt": "In bugs.go Classify is missing a return for the high case; add it so Classify(500) returns \"high\".", "workspaceFixture": "../testdata/fix", "verificationCommand": ["go", "test", "-run", "TestClassifyHigh", "./..."]},
+    {"id": "fix-07", "name": "wrong label prefix", "class": "fix", "prompt": "In bugs.go Label returns 'admin-<name>' but the test expects 'user-<name>'; change the prefix from 'admin-' to 'user-'.", "workspaceFixture": "../testdata/fix", "verificationCommand": ["go", "test", "-run", "TestLabelUsesString", "./..."]},
+    {"id": "fix-08", "name": "race on counter", "class": "fix", "prompt": "In bugs.go the global counter is incremented without synchronization; guard it with a mutex.", "workspaceFixture": "../testdata/fix", "verificationCommand": ["go", "test", "-race", "-run", "TestCounterRace", "./..."]},
+
+    {"id": "refactor-01", "name": "extract helper", "class": "refactor", "prompt": "Extract the duplicated greeting logic into a single helper function and call it from both sites.", "workspaceFixture": "../testdata/refactor", "verificationCommand": ["go", "build", "./..."]},
+    {"id": "refactor-02", "name": "split a file", "class": "refactor", "prompt": "Split main.go into two files: one for Config, one for the rest, keeping the package building.", "workspaceFixture": "../testdata/refactor", "verificationCommand": ["go", "build", "./..."]},
+    {"id": "refactor-03", "name": "rename package", "class": "refactor", "prompt": "Rename the package from app to zeroapp everywhere and keep it building.", "workspaceFixture": "../testdata/refactor", "verificationCommand": ["go", "build", "./..."]},
+    {"id": "refactor-04", "name": "introduce a type", "class": "refactor", "prompt": "Introduce a named type for the map[string]int used in main.go and update its users.", "workspaceFixture": "../testdata/refactor", "verificationCommand": ["go", "build", "./..."]},
+    {"id": "refactor-05", "name": "inline a wrapper", "class": "refactor", "prompt": "Inline the thin wrapper function into its single caller and remove it.", "workspaceFixture": "../testdata/refactor", "verificationCommand": ["go", "build", "./..."]},
+    {"id": "refactor-06", "name": "consolidate errors", "class": "refactor", "prompt": "Consolidate the repeated error-construction blocks into a single helper.", "workspaceFixture": "../testdata/refactor", "verificationCommand": ["go", "build", "./..."]},
+
+    {"id": "longproc-01", "name": "long build", "class": "longproc", "prompt": "Run the build command and report whether it succeeds. Do not modify anything.", "workspaceFixture": "../testdata/longproc"},
+    {"id": "longproc-02", "name": "long test", "class": "longproc", "prompt": "Run go test ./... and report the result. Do not modify anything.", "workspaceFixture": "../testdata/longproc"},
+    {"id": "longproc-03", "name": "bench run", "class": "longproc", "prompt": "Run go test -bench=. -benchtime=1x and report the benchmark numbers. Do not modify anything.", "workspaceFixture": "../testdata/longproc"},
+    {"id": "longproc-04", "name": "vet run", "class": "longproc", "prompt": "Run go vet ./... and report the output. Do not modify anything.", "workspaceFixture": "../testdata/longproc"},
+
+    {"id": "longctx-01", "name": "summarize large file", "class": "longctx", "prompt": "Read the large generated file big.go and summarize its structure. Do not modify anything.", "workspaceFixture": "../testdata/longctx"},
+    {"id": "longctx-02", "name": "find in large file", "class": "longctx", "prompt": "In big.go, find every function that returns an error and list them. Do not modify anything.", "workspaceFixture": "../testdata/longctx"},
+    {"id": "longctx-03", "name": "cross-file question", "class": "longctx", "prompt": "Read all .go files and report which two functions share the same name across files. Do not modify anything.", "workspaceFixture": "../testdata/longctx"},
+    {"id": "longctx-04", "name": "trace through large file", "class": "longctx", "prompt": "Starting from Handle, trace the call path through big.go and report the chain. Do not modify anything.", "workspaceFixture": "../testdata/longctx"},
+
+    {"id": "parallel-01", "name": "read six files", "class": "parallel", "prompt": "Read a.txt, b.txt, c.txt, d.txt, e.txt, and f.txt, then report the first line of each. Do not modify anything.", "workspaceFixture": "../testdata/parallel"},
+    {"id": "parallel-02", "name": "grep six patterns", "class": "parallel", "prompt": "Search for 'alpha', 'beta', 'gamma', 'delta', 'epsilon', and 'zeta' across the repo and report which files contain each. Do not modify anything.", "workspaceFixture": "../testdata/parallel"},
+    {"id": "parallel-03", "name": "list six dirs", "class": "parallel", "prompt": "List the contents of dir1, dir2, dir3, dir4, dir5, and dir6. Do not modify anything.", "workspaceFixture": "../testdata/parallel"},
+    {"id": "parallel-04", "name": "stat six files", "class": "parallel", "prompt": "Report the line count of a.txt, b.txt, c.txt, d.txt, e.txt, and f.txt. Do not modify anything.", "workspaceFixture": "../testdata/parallel"},
+    {"id": "parallel-05", "name": "read configs", "class": "parallel", "prompt": "Read config1.json, config2.json, config3.json, config4.json, config5.json, and config6.json and report each top-level key. Do not modify anything.", "workspaceFixture": "../testdata/parallel"},
+    {"id": "parallel-06", "name": "summarize six", "class": "parallel", "prompt": "Read the six .txt files and give a one-line summary of each. Do not modify anything.", "workspaceFixture": "../testdata/parallel"}
+  ]
+}

internal/perfbench/reports/README.md

@@ -0,0 +1,84 @@
+# Per-turn benchmark reports
+
+This directory holds generated per-turn benchmark reports. It is **not** a
+checked-in source of truth for performance — the numbers are machine- and
+model-specific, so a single snapshot here would mislead rather than inform.
+
+## Generating a baseline
+
+Run the harness over the checked-in manifest:
+
+```sh
+make baseline ZERO_BENCH_MODEL=<model>              # uses ./zero
+make baseline ZERO_BENCH_MODEL=<model> ZERO_BENCH_BINARY=/path/to/zero
+```
+
+This builds `zero`, then runs `zero-perf-bench turn` over
+`internal/perfbench/manifests/baseline.json`, capturing each turn's trace
+(`zero exec --trace <tmpfile>`) and writing the aggregated result to
+`reports/baseline.json`.
+
+The JSON report is self-describing: model, mode, self-correct flag, version,
+commit, date, per-span median/P95, the **top three controllable latency sources**
+ranked by **exclusive** time, per-class roll-ups, and token/count totals.
+That top-three list is the Phase 0 baseline's "do not proceed until" criterion —
+it names where a turn actually spends time so later optimization work is
+targeted, not guessed.
+
+### Attribution model (honest by construction)
+
+Spans record wall intervals and are **not** summed into each other. Each span's
+**exclusive** time is its own duration minus the union of its nested children's
+intervals, derived at finish by interval containment. So a `provider_connect`
+that runs concurrently inside `generation`, or a `permission_wait` nested inside
+`tool_execution`, each contributes only its own exclusive time — they no longer
+double-count the same wall. The top-latency shares therefore sum to ~1 for a
+well-instrumented run, and the ranking reflects where wall time is actually
+spent.
+
+**Coverage** is the fraction of wall covered by the union of all span
+intervals (capped at 1.0) — the honest "≥95% of wall accounted for" metric. A
+run with `coverage < 0.95` has uninstrumented gaps, not an inflated attribution.
+
+### Pass/fail is reported per oracle tier
+
+Pass/fail is split into three tiers so it cannot be misread as a blanket
+correctness verdict (see `MANIFEST.md` for the class breakdown):
+
+- **Correctness** (`tasksVerified` / `tasksPassed` / `correctnessPassRate`):
+  tasks with a positive oracle — edit's substring grep, fix's scoped `go test`.
+  This is the only pass rate that can move with model quality.
+- **Build-only** (`buildCheckedTasks` / `buildPassedTasks` / `buildPassRate`):
+  refactor's `go build ./...`, which proves the edit compiles but not that the
+  refactor achieved its goal. Reported separately, never in `correctnessPassRate`.
+- **Latency-only** (`latencyOnlyTasks`): the read-only classes (nav, longproc,
+  longctx, parallel) carry no oracle — an exit 0 only proves the turn ran. They
+  contribute to latency and span attribution and are excluded from every pass
+  rate.
+
+`tasksAttempted` is still the total across all three tiers. The tier class lists
+are echoed in the report so a consumer can see exactly which classes each rate
+is computed over.
+
+> **Do not average the tier pass rates. Do not report a single "pass rate".**
+> `correctnessPassRate` and `buildPassRate` measure different things over
+> different task sets and must never be combined — a weighted or arithmetic
+> mean of the two is a number that means nothing. The schema offers no
+> headline pass rate on purpose: a consumer must name the tier it is reporting
+> (`correctness`, `build`, or `latency-only`) and quote that tier's fields. A
+> read-only run that exits 0 is a latency sample, not a pass.
+
+## What to commit
+
+Commit the manifest and fixtures (`manifests/`, `../testdata/`), not a generated
+`baseline.json`. A generated report belongs in a PR description or a shared
+dashboard as evidence for one configuration, not in the tree as a durable
+expectation. `.gitkeep` keeps the directory present between runs.
+
+## Caveats
+
+- Each task is a fresh `zero exec` process, so iterations are **cold-start**
+  samples. A warm path (reusing an in-process agent) is future work.
+- Mutating tasks (edit/fix/refactor) run against a per-invocation **copy** of
+  their fixture, so the checked-in fixtures stay clean and one task's edits
+  can't bleed into the next iteration.

internal/perfbench/taskbench.go

@@ -25,17 +25,29 @@ const TaskSchemaVersion = 1
 // ZERO must satisfy. The set is recorded by ID with every result so a published
 // number is traceable to the exact tasks that produced it.
 type TaskSet struct {
-	ID    string      `json:"id"`
-	Name  string      `json:"name,omitempty"`
-	Tasks []BenchTask `json:"tasks"`
+	ID          string      `json:"id"`
+	Name        string      `json:"name,omitempty"`
+	Description string      `json:"description,omitempty"`
+	Tasks       []BenchTask `json:"tasks"`
+	// BuildOnlyClasses lists task classes whose verificationCommand is a
+	// non-positive build check (e.g. refactor's `go build ./...`): it proves the
+	// edit compiles, not that the refactor achieved its goal. The turn benchmark
+	// reports these separately from correctness oracles so a build-pass cannot be
+	// misread as a correctness pass. Classes with a verificationCommand that are
+	// NOT listed here are treated as correctness classes; classes whose tasks
+	// carry no verificationCommand are latency-only regardless of this list.
+	BuildOnlyClasses []string `json:"buildOnlyClasses,omitempty"`
 }
 
 // BenchTask is one benchmark task. WorkspaceFixture is the relative path of the
 // task's starting workspace; VerificationCommand (optional) is the command the
-// default runner executes to decide pass/fail after ZERO finishes.
+// default runner executes to decide pass/fail after ZERO finishes. Class groups
+// the task for the turn-benchmark's per-group latency breakdown (e.g. "nav",
+// "edit", "fix"); it is optional and ignored by the pass/fail runner.
 type BenchTask struct {
 	ID                  string   `json:"id"`
 	Name                string   `json:"name,omitempty"`
+	Class               string   `json:"class,omitempty"`
 	Prompt              string   `json:"prompt"`
 	WorkspaceFixture    string   `json:"workspaceFixture,omitempty"`
 	VerificationCommand []string `json:"verificationCommand,omitempty"`

internal/perfbench/testdata/edit/main.go

@@ -0,0 +1,42 @@
+// Package edit is a small workspace for the edit-class benchmark tasks: each
+// task asks the agent to make one targeted edit, verified by a grep/build
+// command in the manifest.
+package edit
+
+import (
+	"fmt"
+	"os"
+)
+
+// Port is the default listen port.
+const Port = 8080
+
+// MaxRetries is the maximum number of retries before giving up.
+const MaxRetries = 3
+
+// Config holds runtime configuration.
+type Config struct {
+	Name string
+}
+
+// greet returns a greeting for the named user.
+//
+// This receieves a name and formats a hello message.
+func greet(name string) string {
+	return "hello"
+}
+
+// load reads a config file and returns its bytes, or an error on failure.
+func load(path string) ([]byte, error) {
+	data, err := os.ReadFile(path)
+	if err != nil {
+		return nil, err
+	}
+	return data, nil
+}
+
+func main() {
+	fmt.Println("debug: starting")
+	cfg := Config{Name: "demo"}
+	fmt.Println(greet(cfg.Name))
+}

internal/perfbench/testdata/edit/version.go

@@ -0,0 +1,4 @@
+package edit
+
+// Version is the build version string.
+const Version = "1.0.0"

internal/perfbench/testdata/fix/bugs.go

@@ -0,0 +1,89 @@
+// Package fix holds deliberately buggy functions for the bug-fix benchmark
+// tasks. Each function is covered by its own test (see bugs_test.go) and each
+// task verifies with `go test -run <TestName>`, so the tasks are independent:
+// the initial package compiles and the targeted test fails until the named bug
+// is fixed. Every bug is fixable without changing a function signature, so the
+// test files compile in both the buggy and the fixed state.
+package fix
+
+import "fmt"
+
+// Sum returns the sum of all elements. BUG: off-by-one — skips the last element.
+func Sum(xs []int) int {
+	total := 0
+	for i := 0; i < len(xs)-1; i++ {
+		total += xs[i]
+	}
+	return total
+}
+
+// DefaultConfig returns the default configuration. BUG: returns nil, which
+// Port() then dereferences.
+func DefaultConfig() *Config {
+	return nil
+}
+
+// Config holds demo settings.
+type Config struct {
+	Port int
+}
+
+// Port returns the configured port. BUG: panics when DefaultConfig() is nil.
+func Port() int {
+	return DefaultConfig().Port
+}
+
+// Max returns the larger of a and b. BUG: uses the wrong inequality and returns
+// the smaller value.
+func Max(a, b int) int {
+	if a < b {
+		return a
+	}
+	return b
+}
+
+// ReadFirst returns the first byte of the given file's content as a string. BUG:
+// returns only a single character instead of the full content.
+func ReadFirst(content string) string {
+	if len(content) == 0 {
+		return ""
+	}
+	return string(content[0])
+}
+
+// Divide returns a divided by b. BUG: the operands are swapped.
+func Divide(a, b int) int {
+	return b / a
+}
+
+// Classify returns "low", "mid", or "high". BUG: the "high" branch is missing a
+// return, so high values fall through to the default "".
+func Classify(n int) string {
+	if n < 10 {
+		return "low"
+	}
+	if n < 100 {
+		return "mid"
+	}
+	// BUG: missing return for the high case
+	return ""
+}
+
+// Label returns a display label for the given name. BUG: uses the wrong prefix
+// ("admin-" instead of "user-"), producing the wrong text.
+func Label(name string) string {
+	return fmt.Sprintf("admin-%s", name)
+}
+
+// counter is a shared, unsynchronized counter. BUG: Inc is not goroutine-safe.
+var counter int
+
+// Inc increments the shared counter.
+func Inc() {
+	counter++
+}
+
+// Count returns the current counter value.
+func Count() int {
+	return counter
+}

internal/perfbench/testdata/fix/bugs_test.go

@@ -0,0 +1,68 @@
+package fix
+
+import (
+	"sync"
+	"testing"
+)
+
+func TestSumIncludesLastElement(t *testing.T) {
+	if got := Sum([]int{1, 2, 3, 4}); got != 10 {
+		t.Fatalf("Sum = %d, want 10", got)
+	}
+}
+
+func TestPortReturnsDefault(t *testing.T) {
+	if got := Port(); got != 8080 {
+		t.Fatalf("Port() = %d, want 8080 (DefaultConfig returned nil)", got)
+	}
+}
+
+func TestMaxReturnsLarger(t *testing.T) {
+	if got := Max(3, 7); got != 7 {
+		t.Fatalf("Max(3,7) = %d, want 7", got)
+	}
+}
+
+func TestReadFirstReturnsFullContent(t *testing.T) {
+	if got := ReadFirst("hello"); got != "hello" {
+		t.Fatalf("ReadFirst = %q, want %q", got, "hello")
+	}
+}
+
+func TestDivideOrder(t *testing.T) {
+	if got := Divide(10, 2); got != 5 {
+		t.Fatalf("Divide(10,2) = %d, want 5", got)
+	}
+}
+
+func TestClassifyHigh(t *testing.T) {
+	if got := Classify(500); got != "high" {
+		t.Fatalf("Classify(500) = %q, want %q", got, "high")
+	}
+}
+
+func TestLabelUsesString(t *testing.T) {
+	if got := Label("alice"); got != "user-alice" {
+		t.Fatalf("Label = %q, want %q", got, "user-alice")
+	}
+}
+
+// TestCounterRace exercises Inc concurrently. Run with `go test -race`:
+// the unsynchronized counter trips the race detector until a mutex is added.
+func TestCounterRace(t *testing.T) {
+	counter = 0
+	var wg sync.WaitGroup
+	for i := 0; i < 200; i++ {
+		wg.Add(1)
+		go func() {
+			defer wg.Done()
+			for j := 0; j < 100; j++ {
+				Inc()
+			}
+		}()
+	}
+	wg.Wait()
+	if got := Count(); got != 20000 {
+		t.Fatalf("Count = %d, want 20000", got)
+	}
+}