fix(npm): address PR review — LICENSE in payloads, dist-tag lookup guard, accurate fallback and platform docs

- Ship the repo LICENSE in both the wrapper and platform payloads (npm only
  includes a license file physically present in the packed directory).
- Guard the platform-publish dist-tag lookup with || true so a transient
  npm view failure (or a bootstrap publish with no latest tag) cannot abort
  the job between the platform and wrapper publishes.
- Gate the wrapper's fallback downloader on the release matrix so platforms
  with no asset (windows-arm64 and unknown combos) skip the spawn instead of
  attempting it every run; keep deliberate retry-on-failure for supported
  platforms and say so honestly in the notice and docs.
- Correct Windows support wording everywhere: x64 only, ARM runs under
  emulation (README, README_ZH, INSTALL.md incl. the release-target list).

.github/workflows/release-artifacts.yml

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

.gitignore

@@ -27,6 +27,10 @@ report.[0-9]_.[0-9]_.[0-9]_.[0-9]_.json
 .cache
 *.tsbuildinfo
 
+# Go build/module caches (GOCACHE/GOMODCACHE pointed here for local runs)
+.gocache/
+.gomodcache/
+
 # IntelliJ based IDEs
 .idea
 
@@ -37,7 +41,10 @@ report.[0-9]_.[0-9]_.[0-9]_.[0-9]_.json
 
 /zero
 /zero.exe
+/zero-windows-command-runner.exe
+/zero-windows-sandbox-setup.exe
 /zero-skills
+/.zero-binary-version
 /snake-game.html
 
 # Superpowers skill artifacts (brainstorm mockups, specs, plans) — local only
@@ -47,3 +54,4 @@ docs/superpowers/
 # Termux/Android build artifacts
 zero-termux-arm64
 zero-linux-sandbox
+zero-seccomp

.release-please-manifest.json

@@ -1 +1 @@
-{".":"0.1.0"}
+{".":"0.3.0"}

AGENTS.md

@@ -132,7 +132,7 @@ description: Run the project's benchmark suite and summarize the deltas.
 3. Report any regression > 5% with the function name and the previous value.
 ```
 
-Only `name` and `description` are recognized in the frontmatter today. The `name` defaults to the directory name. If two skills declare the same name, the one in the lexicographically-first directory wins — duplicates are dropped silently. Plugin-declared skills (section 6) are not yet merged into the loader, so a `skills:` entry inside a plugin's `plugin.json` is not visible to the `skill` tool today.
+Only `name` and `description` are recognized in the frontmatter today. The `name` defaults to the directory name. Within a single skills root, duplicate names are resolved by lexicographic directory order. 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.
 
@@ -235,6 +235,13 @@ Servers are merged from user and project configs (project wins on conflicts). To
 - A `--header` value produced by command substitution (`"Authorization=Bearer $(print-token)"`) in a private shell config that you keep out of git.
 - A secret manager that injects the env var your MCP server reads on its own (the `command` and `args` then run inside that environment).
 
+Project MCP config remains supported for shared team servers, but a project layer
+that changes an existing server's target (`url`, `command`, or `args`) does not
+inherit user credentials. Clear or replace `headers`, `env`, or `oauth` explicitly,
+or use a new server name. OAuth tokens are bound to the resolved server identity,
+so changing a server target can require `zero mcp oauth login <server>` again.
+OAuth server names ending in `.<32 hex chars>` are reserved for token storage.
+
 ### As a server — expose Zero's tools to another agent
 
 ```bash
@@ -282,7 +289,7 @@ A plugin is enabled by being present in the plugins directory and disabled by re
 
 Plugin commands run with the plugin directory as their working directory. Use relative paths; the loader resolves them at activation time.
 
-> **Roadmap.** An in-UI plugins manager (browse, install, enable / disable) is on the backlog. Today you use the `zero plugins` CLI subcommands above. Skills declared inside a plugin's `plugin.json` are not yet merged into the `skill` tool's discovery (see section 3).
+> **Roadmap.** An in-UI plugins manager (browse, install, enable / disable) is on the backlog. Today you use the `zero plugins` CLI subcommands above.
 
 ## 7. Configuration locations
 

CHANGELOG.md

@@ -6,6 +6,111 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 aims to follow [Semantic Versioning](https://semver.org/spec/v2.0.0.html) once the first release is
 tagged. Until then, source builds report the version `dev`.
 
+## [0.3.0](https://github.com/Gitlawb/zero/compare/v0.2.0...v0.3.0) (2026-07-09)
+
+
+### Features
+
+* gate project-scoped hooks, plugins, and MCP servers behind workspace trust ([#529](https://github.com/Gitlawb/zero/issues/529)) ([a880ce8](https://github.com/Gitlawb/zero/commit/a880ce80a6ec72da511fb9bdf6dd69291c72a64b))
+* **modelregistry:** infer reasoning efforts for Hunyuan and vendor-prefixed model ids ([#599](https://github.com/Gitlawb/zero/issues/599)) ([92a92ce](https://github.com/Gitlawb/zero/commit/92a92ceb29dc63f5216ab140ef9a6dd9afe17df8))
+* **providers:** OAuth login profiles and list-first /provider manager ([#560](https://github.com/Gitlawb/zero/issues/560)) ([1655056](https://github.com/Gitlawb/zero/commit/16550569c1be615cfaf244dab25909ec37f6dee6))
+* **tui:** remember recent provider+model selections in /model picker ([#568](https://github.com/Gitlawb/zero/issues/568)) ([d0c4e62](https://github.com/Gitlawb/zero/commit/d0c4e62cd429a0614c15296934c740a08bc0e07b))
+* **tui:** show CLI version on the startup home screen ([#538](https://github.com/Gitlawb/zero/issues/538)) ([fd69233](https://github.com/Gitlawb/zero/commit/fd69233e334f1823a06b5794085a9255b3abdfa8))
+* voice dictation (speech-to-text) ([#557](https://github.com/Gitlawb/zero/issues/557)) ([87158a1](https://github.com/Gitlawb/zero/commit/87158a1c90b4f91fc5f2bb8178ebaf46d7654680))
+
+
+### Bug Fixes
+
+* address bugs found in a multi-agent codebase audit ([#481](https://github.com/Gitlawb/zero/issues/481)) ([008bc9b](https://github.com/Gitlawb/zero/commit/008bc9b3f3ba13c7d4822b9559b020f381ff555b))
+* **agent:** keep tools exposed for max-turn finalization ([#533](https://github.com/Gitlawb/zero/issues/533)) ([3f0503b](https://github.com/Gitlawb/zero/commit/3f0503bc2312ae29d5ade784d8824dc9a3524958))
+* **auth:** persist OpenRouter API key after login ([#595](https://github.com/Gitlawb/zero/issues/595)) ([2a062aa](https://github.com/Gitlawb/zero/commit/2a062aa4014c6cd5e20a57dad4a685f88966f109))
+* bump Go to 1.26.5 for crypto/tls fix (GO-2026-5856) ([#607](https://github.com/Gitlawb/zero/issues/607)) ([a7cfb99](https://github.com/Gitlawb/zero/commit/a7cfb99fed7b88ebc09a2f251cb82864d3c2cade))
+* gitignore Windows sandbox helpers and npm version marker ([#578](https://github.com/Gitlawb/zero/issues/578)) ([25653f6](https://github.com/Gitlawb/zero/commit/25653f686c016f95b81ceb7ff5d5452d37c4d4f3))
+* **mcp:** silence startup warning for unconfigured default servers ([#563](https://github.com/Gitlawb/zero/issues/563)) ([302f58b](https://github.com/Gitlawb/zero/commit/302f58bb5f2a03ec7230354ed4747e4e55c16c50))
+* **mcp:** skip RFC 8414 discovery when OAuth endpoints are preconfigured ([#586](https://github.com/Gitlawb/zero/issues/586)) ([8a52d98](https://github.com/Gitlawb/zero/commit/8a52d98cad7cd0086dee9aede4ce477e432bd385))
+* **modelregistry:** reject oversized models.dev cache responses ([#602](https://github.com/Gitlawb/zero/issues/602)) ([66a6396](https://github.com/Gitlawb/zero/commit/66a63964149fb2f07e646e5f1987627c5cd9ac28))
+* **provider:** stop dropping custom no-auth providers on restart ([#558](https://github.com/Gitlawb/zero/issues/558)) ([ba99fa8](https://github.com/Gitlawb/zero/commit/ba99fa8d487fb28f2700e0ff10b2a25c75303cf7))
+* **sandbox:** Windows-appropriate suggestions when blocking interactive commands ([#414](https://github.com/Gitlawb/zero/issues/414)) ([ba4c007](https://github.com/Gitlawb/zero/commit/ba4c00755dc1c31b3dcca18e50b20f62c6bf5d1f))
+* **tools:** block MSYS and WSL shells under the Windows sandbox ([#587](https://github.com/Gitlawb/zero/issues/587)) ([0666818](https://github.com/Gitlawb/zero/commit/066681855b80e3baf5d07d7397610b25f724e353))
+* **tools:** platform-specific pager suggestions, quote/caret-safe cd detection ([#543](https://github.com/Gitlawb/zero/issues/543)) ([8b248f4](https://github.com/Gitlawb/zero/commit/8b248f4e1198dc86ab332a697fba4cf520823cbd))
+* **tools:** Windows cmd.exe quoting guidance and clipboard escaping fix ([#468](https://github.com/Gitlawb/zero/issues/468)) ([f10ed0c](https://github.com/Gitlawb/zero/commit/f10ed0c893ce6de08923f143d681ba96f0fcfe3a))
+* **tui:** bypass toggleSidebar and toggleMouse global shortcuts when composer is non-empty ([#576](https://github.com/Gitlawb/zero/issues/576)) ([c7346fb](https://github.com/Gitlawb/zero/commit/c7346fbcf01fe7e70ed2fcfccbf9965e5985727b))
+* **tui:** paste protection for Termux char-by-char paste ([#573](https://github.com/Gitlawb/zero/issues/573)) ([8e9149f](https://github.com/Gitlawb/zero/commit/8e9149f1a296bebebf95cae9dd2a7c5156c9dbb6))
+* **tui:** update picker_test for switchProviderModel 4-value return ([#589](https://github.com/Gitlawb/zero/issues/589)) ([8f15650](https://github.com/Gitlawb/zero/commit/8f156506dc92449bd24caa14e36f28477ba00fff))
+* **update:** clearer error on unsupported release platform (android/termux) ([#603](https://github.com/Gitlawb/zero/issues/603)) ([1fc9b2d](https://github.com/Gitlawb/zero/commit/1fc9b2d25c79e089f34cb7b5b6a7f7c7b8233123))
+* **update:** support safe symlink extraction during updates ([#575](https://github.com/Gitlawb/zero/issues/575)) ([ce9cb91](https://github.com/Gitlawb/zero/commit/ce9cb912958a3d9ae0b052bfebb849c28e0e719b))
+* warn about untracked scratch files left behind after a run ([#571](https://github.com/Gitlawb/zero/issues/571)) ([062328b](https://github.com/Gitlawb/zero/commit/062328b632a2d27353a6d47c522bbd22d7282539))
+
+
+### Performance Improvements
+
+* **grep:** stop content scan after head limit ([#601](https://github.com/Gitlawb/zero/issues/601)) ([8a05e64](https://github.com/Gitlawb/zero/commit/8a05e6486f7a63281b869064db03dfc5531e6a04))
+
+## [0.2.0](https://github.com/Gitlawb/zero/compare/v0.1.0...v0.2.0) (2026-07-06)
+
+
+### Features
+
+* add --auto flag for LLM-generated commit messages ([#423](https://github.com/Gitlawb/zero/issues/423)) ([b0abde7](https://github.com/Gitlawb/zero/commit/b0abde7d0697e808480cd59d69a6f4d0c6320475))
+* add zero changes push and pr subcommands, and extra repo-info metrics ([#391](https://github.com/Gitlawb/zero/issues/391)) ([2312abe](https://github.com/Gitlawb/zero/commit/2312abe5ddd95f4c6ef373cfb61cc03092f48cdd))
+* agent quality, caching, retry, and tooling upgrades ([#506](https://github.com/Gitlawb/zero/issues/506)) ([3c81fea](https://github.com/Gitlawb/zero/commit/3c81fea22873ee3df7fc97b10cb4f77792706c4b))
+* **agent:** curb over-engineering the solution in the editing discipline ([#517](https://github.com/Gitlawb/zero/issues/517)) ([f4c998a](https://github.com/Gitlawb/zero/commit/f4c998ac30c4f07ff313a2d706791e857293be49)), closes [#516](https://github.com/Gitlawb/zero/issues/516)
+* **agent:** inject per-user config.UserConfigDir()/zero/ZERO.md guidelines into system prompt ([#475](https://github.com/Gitlawb/zero/issues/475)) ([7b10aab](https://github.com/Gitlawb/zero/commit/7b10aab74bf14a01166b2cea22deab79bba9850b))
+* **openai:** forward prompt_cache_key for server-side prefix cache routing ([#515](https://github.com/Gitlawb/zero/issues/515)) ([87e7e69](https://github.com/Gitlawb/zero/commit/87e7e69afd18b5539579856f3a61c6a95bc445ae))
+* **providers:** add `zero providers models` to discover a provider's models ([#386](https://github.com/Gitlawb/zero/issues/386)) ([0bc8074](https://github.com/Gitlawb/zero/commit/0bc8074c97b0310e4a9d70c3f967003ee5e8a59f))
+* **providers:** add KiloCode and OpenCode provider support ([#388](https://github.com/Gitlawb/zero/issues/388)) ([b1ccb6d](https://github.com/Gitlawb/zero/commit/b1ccb6d9c1875377f5e5ea81a1304edd1e41ab4f))
+* **providers:** add Meituan LongCat catalog preset ([#424](https://github.com/Gitlawb/zero/issues/424)) ([b4275e3](https://github.com/Gitlawb/zero/commit/b4275e350472b2490212bf814709819d354c1216))
+* **providers:** split minimax zai into intl cn ([#398](https://github.com/Gitlawb/zero/issues/398)) ([aaad4d2](https://github.com/Gitlawb/zero/commit/aaad4d271270f41af837b6f3b60ae80beba0c645))
+* require manual approval before npm publish + drop release-as pin ([#369](https://github.com/Gitlawb/zero/issues/369)) ([bd89a1f](https://github.com/Gitlawb/zero/commit/bd89a1f451643c1b65ec803070abc7b116631ebe))
+* **sandbox:** unelevated Windows fallback tier instead of prompts-only degrade ([#427](https://github.com/Gitlawb/zero/issues/427)) ([b9ddd6f](https://github.com/Gitlawb/zero/commit/b9ddd6f42138312a1fee8d8bb67c46c8eb1dea2f))
+* support shift enter for composer newlines ([#462](https://github.com/Gitlawb/zero/issues/462)) ([daf65e0](https://github.com/Gitlawb/zero/commit/daf65e0af9a040314d4ab337b0ad59c55416b7bc))
+* **tui:** /loop — repeat a prompt or command on an interval or self-paced ([#502](https://github.com/Gitlawb/zero/issues/502)) ([387fe67](https://github.com/Gitlawb/zero/commit/387fe67ee7cd81317c9c969f5906a4437080fea3))
+* **tui:** add search/filter to provider picker in setup wizard ([#400](https://github.com/Gitlawb/zero/issues/400)) ([2fcea71](https://github.com/Gitlawb/zero/commit/2fcea71778d23e050c93409c471aef45b68c1621))
+* **update:** add zero upgrade command to apply self-updates ([#461](https://github.com/Gitlawb/zero/issues/461)) ([5f36349](https://github.com/Gitlawb/zero/commit/5f36349c1884e81fa9bc66bb5fe813b627e897b7))
+
+
+### Bug Fixes
+
+* **action:** keep provider key scoped to zero step ([#448](https://github.com/Gitlawb/zero/issues/448)) ([407a927](https://github.com/Gitlawb/zero/commit/407a92739ff508cba32d2c12b3f36f0efcdd54c3))
+* add android platform support for Termux npm install ([#455](https://github.com/Gitlawb/zero/issues/455)) ([9bd93c6](https://github.com/Gitlawb/zero/commit/9bd93c62f8d57fb74057284aa66a1b6e1429dcdd)), closes [#449](https://github.com/Gitlawb/zero/issues/449)
+* **agent:** reject a malformed additional_permissions payload before prompting ([#453](https://github.com/Gitlawb/zero/issues/453)) ([e4f760e](https://github.com/Gitlawb/zero/commit/e4f760ee8bd57299cd2fcb37e8e23130037c2607))
+* allow non-TLS connections to private-network provider endpoints ([#444](https://github.com/Gitlawb/zero/issues/444)) ([1d86384](https://github.com/Gitlawb/zero/commit/1d8638466ca31517eb9db2b9353d3dce1cbeeabc))
+* **auth:** route zero auth login chatgpt to the dedicated ChatGPT flow ([#443](https://github.com/Gitlawb/zero/issues/443)) ([305a62c](https://github.com/Gitlawb/zero/commit/305a62c954ca6cec00bc58d5398f933415156aff))
+* **config:** fall back to a usable saved provider instead of forcing full re-onboarding ([#410](https://github.com/Gitlawb/zero/issues/410)) ([c60ad87](https://github.com/Gitlawb/zero/commit/c60ad8729f79bb841114d352ee2d2fe29d5d0e41))
+* **config:** let a gateway ANTHROPIC_BASE_URL resolve as anthropic-compatible ([#497](https://github.com/Gitlawb/zero/issues/497)) ([30dd7c3](https://github.com/Gitlawb/zero/commit/30dd7c3112ad22d42fa12b5addd4e38f4beda42a)), closes [#479](https://github.com/Gitlawb/zero/issues/479)
+* **config:** unbrick first-run setup — default google/anthropic models, enter setup on fixable config errors ([#385](https://github.com/Gitlawb/zero/issues/385)) ([72eed06](https://github.com/Gitlawb/zero/commit/72eed06b4f94c43d75d31fe54a58d2f566de059e))
+* **config:** use ~/.config on macOS and enter setup when no provider ([#371](https://github.com/Gitlawb/zero/issues/371)) ([#372](https://github.com/Gitlawb/zero/issues/372)) ([027a8f2](https://github.com/Gitlawb/zero/commit/027a8f2768b17b89f5c8270887f156e2ccda69ea))
+* **docs:** rename AGENTS.MD > AGENTS.md ([#438](https://github.com/Gitlawb/zero/issues/438)) ([4266baf](https://github.com/Gitlawb/zero/commit/4266baf222df583ed2078b776687f12d496475b5))
+* **gemini:** strip unsupported JSON Schema fields from tool declarations ([#374](https://github.com/Gitlawb/zero/issues/374)) ([39e7100](https://github.com/Gitlawb/zero/commit/39e7100674150144a1152e3110c64c7cf0321d64)), closes [#373](https://github.com/Gitlawb/zero/issues/373)
+* **install:** persist install dir to user PATH on Windows ([#407](https://github.com/Gitlawb/zero/issues/407)) ([bdb1b0e](https://github.com/Gitlawb/zero/commit/bdb1b0ecd15859b1712a6037d296dace7f9c3c3f))
+* **mcp:** block cross-origin credential redirects ([#396](https://github.com/Gitlawb/zero/issues/396)) ([f915f70](https://github.com/Gitlawb/zero/commit/f915f70e5a3096e2419fa8d961a0f84a626fa4a9))
+* **oauth:** treat Windows ERROR_ACCESS_DENIED as lock contention in createSecretFile ([#445](https://github.com/Gitlawb/zero/issues/445)) ([d05e914](https://github.com/Gitlawb/zero/commit/d05e9148a7f79f67d1d3c31fca2775f21fbd331e))
+* **openai:** handle Ollama reasoning stream deltas ([#486](https://github.com/Gitlawb/zero/issues/486)) ([f6c0606](https://github.com/Gitlawb/zero/commit/f6c060631e18e082dda24cc4dc0903c31c2120d6))
+* preserve conversation context in exec prompts ([#460](https://github.com/Gitlawb/zero/issues/460)) ([949ee43](https://github.com/Gitlawb/zero/commit/949ee43f71e5cb7fab4695c5cb7b442fe4ecfbf7))
+* **provider-wizard:** allow multiple custom OpenAI-compatible providers ([#403](https://github.com/Gitlawb/zero/issues/403)) ([3fbbd28](https://github.com/Gitlawb/zero/commit/3fbbd28e4c586822cc4312c86232d94befe56e87))
+* **sandbox:** fix nested pipe creation under the Windows restricted token ([#456](https://github.com/Gitlawb/zero/issues/456)) ([563a6db](https://github.com/Gitlawb/zero/commit/563a6dbe91e65d5daeefd7626e8a77e30a6d8fb2))
+* **sandbox:** gate /tmp test assertions on GOOS, not path existence ([#426](https://github.com/Gitlawb/zero/issues/426)) ([f653dca](https://github.com/Gitlawb/zero/commit/f653dcac363fb69ad7be5b35e6e0fa6d2bce476d))
+* **sandbox:** self-heal a corrupt unelevated setup marker ([#437](https://github.com/Gitlawb/zero/issues/437)) ([8d0c5fe](https://github.com/Gitlawb/zero/commit/8d0c5feccb8bdbfb015df0508aa6e3bcbd1fd0e8))
+* **specialist:** cap max specialist nesting depth ([#491](https://github.com/Gitlawb/zero/issues/491)) ([177442c](https://github.com/Gitlawb/zero/commit/177442cfe4015bd8df04cc9894f98b468ee796d4))
+* Termux/Android support — PRoot scroll, SIGSYS sandbox, build docs ([#509](https://github.com/Gitlawb/zero/issues/509)) ([0f69d99](https://github.com/Gitlawb/zero/commit/0f69d995e9b586b774f66c066b21abab5e03024a))
+* **tools:** Block MSYS coreutils under Windows sandbox ([#476](https://github.com/Gitlawb/zero/issues/476)) ([81aad58](https://github.com/Gitlawb/zero/commit/81aad58d97839e51c068e2f08907618991fdc3fb))
+* **tools:** CRLF line ending mismatch in edit_file tool on Windows ([#378](https://github.com/Gitlawb/zero/issues/378)) ([33dc7ae](https://github.com/Gitlawb/zero/commit/33dc7ae2cc82c5389675531e1416856dae7151ce))
+* **tools:** fix cmd.exe /S/C corrupting commands with embedded quotes ([#465](https://github.com/Gitlawb/zero/issues/465)) ([190241b](https://github.com/Gitlawb/zero/commit/190241bd593f43211b766e0b13c8e89802d4bb37))
+* **tools:** flag piped POSIX utilities before running on Windows ([#412](https://github.com/Gitlawb/zero/issues/412)) ([5658a36](https://github.com/Gitlawb/zero/commit/5658a366274fc59a9d5336b06a21019c9c25cbf1))
+* **tools:** make grep and glob respect run cancellation ([#464](https://github.com/Gitlawb/zero/issues/464)) ([ba6c026](https://github.com/Gitlawb/zero/commit/ba6c0264697b7d7ed479f6e782fba9700a481e3d))
+* **tools:** require permission before web_search requests ([#382](https://github.com/Gitlawb/zero/issues/382)) ([960db96](https://github.com/Gitlawb/zero/commit/960db9660e4e31dc588fe8f7d6f116ff5e225566))
+* **tui:** compose help overlay through the viewport overlay pipeline ([#421](https://github.com/Gitlawb/zero/issues/421)) ([5b2b4de](https://github.com/Gitlawb/zero/commit/5b2b4dea1aaf9e0f68baa25e97e83296fb17b1a2))
+* **tui:** keep the profile name on /model switch so the stored key resolves ([#441](https://github.com/Gitlawb/zero/issues/441)) ([9134148](https://github.com/Gitlawb/zero/commit/9134148f4df3e4e556fba6c2f8babfdf6fcfeee1)), closes [#440](https://github.com/Gitlawb/zero/issues/440)
+* **tui:** resolve every permission request so the agent can't deadlock ([#397](https://github.com/Gitlawb/zero/issues/397)) ([952788f](https://github.com/Gitlawb/zero/commit/952788f72d32957659fe004521fcc8372b9ba9b4))
+* **tui:** show an M suffix for million-scale token counts ([#457](https://github.com/Gitlawb/zero/issues/457)) ([0562e3b](https://github.com/Gitlawb/zero/commit/0562e3bef7df2328610a48a1e81632a8da4aec64))
+* **tui:** title /model rows by model name, not the catalog description ([#395](https://github.com/Gitlawb/zero/issues/395)) ([cdf9d83](https://github.com/Gitlawb/zero/commit/cdf9d839ae57a729f292f36f7c5b0c67b41b288d))
+
+
+### Performance Improvements
+
+* cache TUI model registry ([#496](https://github.com/Gitlawb/zero/issues/496)) ([e7d88b4](https://github.com/Gitlawb/zero/commit/e7d88b4b518049733da25a8447c00144bd1da518))
+* universal tool-output ceiling with spill + async post-edit diagnostics ([#518](https://github.com/Gitlawb/zero/issues/518)) ([95ccd5b](https://github.com/Gitlawb/zero/commit/95ccd5bc327f6fb464ff0239f7229de789f578dc))
+
 ## 0.1.0 (2026-07-02)
 
 

README.md

@@ -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
 

README_ZH.md

@@ -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
 

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 {

docs/HOW_ZERO_WORKS.md

@@ -0,0 +1,787 @@
+# How Zero Works
+
+This document explains Zero's architecture at a high level: the layers, module
+boundaries, data flow, and runtime concepts behind the interactive TUI and the
+headless `zero exec` mode.
+
+It is written as a progressive map: each high-level section introduces the
+concepts you need for orientation, then points to the later section that explains
+the mechanics in more detail.
+
+## Big Picture
+
+Zero is a Go CLI application with two main user-facing surfaces:
+
+- **Interactive TUI** (`zero`): a Bubble Tea application that renders chat,
+  tool cards, permission prompts, model/provider pickers, and session controls.
+- **Headless exec** (`zero exec ...`): a scriptable command runner that uses the
+  same agent loop and tools, but writes text, JSON, or stream-JSON events.
+
+Both surfaces converge on the same core agent runtime:
+
+```mermaid
+flowchart TD
+    User[User / automation] --> CLI[cmd/zero -> internal/cli]
+
+    CLI --> TUI[Interactive TUI<br/>internal/tui]
+    CLI --> Exec[Headless exec<br/>internal/cli/exec.go]
+    CLI --> ACP[ACP / editor bridge<br/>internal/acp]
+
+    TUI --> Agent[Agent loop<br/>internal/agent]
+    Exec --> Agent
+    ACP --> Agent
+
+    Agent --> Provider[Provider runtime<br/>internal/zeroruntime.Provider]
+    Provider --> APIs[OpenAI / Anthropic / Gemini / compatible APIs]
+
+    Agent --> Registry[Tool registry<br/>internal/tools.Registry]
+    Registry --> CoreTools[Core tools<br/>read/edit/shell/search/plan]
+    Registry --> MCP[MCP tools<br/>internal/mcp]
+    Registry --> Specialists[Specialist tools<br/>internal/specialist]
+    Registry --> Plugins[Plugin tools<br/>internal/plugins]
+    Registry --> LocalControl[Browser / terminal / desktop helpers]
+
+    Agent --> Sandbox[Permission + sandbox engine<br/>internal/sandbox]
+    Agent --> Hooks[Hooks dispatcher<br/>internal/hooks]
+    Agent --> Sessions[Session event store<br/>internal/sessions]
+```
+
+The important split is:
+
+- `internal/tui` is the **presentation layer**. For the mechanics, see
+  [Presentation Layer: Bubble Tea TUI](#presentation-layer-bubble-tea-tui) and
+  [How the TUI and Agent Loop Connect](#how-the-tui-and-agent-loop-connect).
+- `internal/agent` is the **reasoning/tool loop**. For the per-turn mechanics,
+  see [Agent Loop](#agent-loop), [Tool Execution Lifecycle](#tool-execution-lifecycle),
+  and [Context Budget and Compaction](#context-budget-and-compaction).
+- `internal/tools`, `internal/sandbox`, and provider packages are the **runtime
+  capability layer**. For the details, see [Runtime Provider Layer](#runtime-provider-layer),
+  [Tools and Capabilities](#tools-and-capabilities), and
+  [Permission and Sandbox Model](#permission-and-sandbox-model).
+- `internal/sessions` is the **durable event log** used for resume, fork,
+  rewind, sub-agents, and history. See [Session Storage and History](#session-storage-and-history).
+
+## Architecture Spine
+
+The easiest way to understand Zero is not as a set of independent packages, but
+as one spine with several mechanisms attached to it:
+
+```mermaid
+flowchart LR
+    Surface[Surface<br/>TUI / exec / ACP] --> Run[Run request<br/>prompt + runtime options]
+    Run --> Agent[Agent loop<br/>turns, tools, recovery]
+    Agent --> Capabilities[Capabilities<br/>providers + tools + sandbox]
+    Capabilities --> Events[Runtime events<br/>text, tool calls, permissions, usage]
+    Events --> Surface
+    Events --> Sessions[Session log<br/>resume, fork, rewind, history]
+
+    Config[Config / guidelines / skills] --> Run
+    Extensions[MCP / specialists / plugins / hooks] --> Capabilities
+```
+
+The **TUI + agent loop** is the main story because it explains the normal user
+experience: type a prompt, watch the model stream, approve tool use, see file
+changes, and receive a final answer. Other mechanisms are still important, but
+most of them attach to the same spine:
+
+| Mechanism | Where it attaches | What it changes |
+| --- | --- | --- |
+| Config, provider profiles, model registry | Before a run starts | Which provider/model is used and what defaults apply. |
+| Project/user guidelines, skills, specialists context | Prompt construction | What instructions and available capabilities the model sees. |
+| Tool registry, MCP, plugins, local-control helpers | Capability layer | Which actions the model may request. |
+| Permission policy, sandbox, command-prefix grants | Tool execution | Whether a requested action can run, needs approval, or is denied. |
+| Hooks | Tool lifecycle | Optional before/after observers or blockers around tool calls. |
+| Sessions, checkpoints, compaction metadata | Event persistence/context | What can be resumed, forked, rewound, or summarized for later turns. |
+| Stream-JSON | Surface layer | How the same runtime events are emitted for automation instead of rendered in the TUI. |
+
+So this document focuses first on the TUI/agent-loop bridge, then explains the
+supporting mechanisms in terms of where they connect to that bridge.
+
+Use this as the reading path:
+
+| If you see this concept in a high-level flow | Read this detail section |
+| --- | --- |
+| Surface submits a run | [Startup Flow](#startup-flow), then [How the TUI and Agent Loop Connect](#how-the-tui-and-agent-loop-connect) |
+| Agent builds context and advances turns | [Agent Loop](#agent-loop) |
+| “Maybe compact context” | [Context Budget and Compaction](#context-budget-and-compaction) |
+| Model requests a tool | [Tool Execution Lifecycle](#tool-execution-lifecycle), then [Tools and Capabilities](#tools-and-capabilities) |
+| Tool needs approval or sandbox access | [Permission and Sandbox Model](#permission-and-sandbox-model) |
+| Events appear in the UI or stream-JSON | [How the TUI and Agent Loop Connect](#how-the-tui-and-agent-loop-connect), [Stream-JSON Protocol](#stream-json-protocol) |
+| Resume, fork, rewind, or sub-agent history | [Session Storage and History](#session-storage-and-history) |
+
+## Main Packages
+
+The sections above describe the architecture in concepts. If you want to connect
+those concepts to implementation details, use this table as the code map before
+reading the deeper flows below.
+
+| Package | Role |
+| --- | --- |
+| `cmd/zero` | Tiny binary entrypoint; calls `internal/cli.Run`. |
+| `internal/cli` | Command parsing, config resolution, provider creation, registry setup, TUI/exec/ACP launch paths. |
+| `internal/tui` | Bubble Tea model/update/view code, transcript rendering, composer, prompts, slash commands, setup wizard. |
+| `internal/agent` | Core loop that sends messages to the model, streams output, executes tools, handles permissions, compaction, retries, and final answers. |
+| `internal/zeroruntime` | Provider-neutral message, tool-call, stream, usage, and image types. Providers implement `Provider.StreamCompletion`. |
+| `internal/providers` | Factory that turns a resolved provider profile into an OpenAI, Anthropic, Gemini, Codex, or compatible provider adapter. |
+| `internal/tools` | Tool interface, registry, core read/write/shell/network tools, local-control wrappers, tool safety metadata, redaction, output ceilings. |
+| `internal/sandbox` | Permission policy, path/network checks, grant store, command-prefix approvals, and platform sandbox backends. |
+| `internal/sessions` | Local append-only session storage: `metadata.json` + `events.jsonl`. |
+| `internal/config` | User/project config resolution, active provider, preferences, sandbox/tool settings. |
+| `internal/mcp` | MCP server config, client runtime, permission store, and MCP tool registration. |
+| `internal/specialist` / `internal/swarm` | Sub-agent manifests and team/member orchestration exposed as tools. |
+| `internal/localcontrol` / `internal/browser` | Runtime helpers used by config-gated local-control tool wrappers. |
+| `internal/plugins` / `internal/skills` / `internal/hooks` | Extension surfaces loaded into the agent context, tool registry, or tool lifecycle. |
+| `internal/streamjson` | Machine-readable protocol for `zero exec --input-format/--output-format stream-json`. |
+
+## Startup Flow
+
+### Interactive TUI (`zero`)
+
+```mermaid
+sequenceDiagram
+    participant Main as cmd/zero
+    participant CLI as internal/cli
+    participant Config as internal/config
+    participant TUI as internal/tui
+    participant Agent as internal/agent
+
+    Main->>CLI: Run(args)
+    CLI->>Config: resolve config for workspace
+    CLI->>CLI: build provider
+    CLI->>CLI: build tool registry
+    CLI->>CLI: register core, config-gated local-control, specialists, MCP, plugins
+    CLI->>CLI: create sandbox engine + session store
+    CLI->>TUI: run Bubble Tea with Options
+    TUI->>TUI: Init / Update / View event loop
+    TUI->>Agent: agent.Run on submitted prompt
+```
+
+The TUI owns terminal interaction. It tracks composer state, transcript rows,
+scrolling, permission dialogs, model/provider pickers, setup wizard state, and
+visual updates. When the user submits a prompt, the TUI builds `agent.Options`
+with callbacks like `OnText`, `OnToolCall`, `OnToolResult`, `OnPermission`,
+`OnAskUser`, and `OnUsage`. Those callbacks convert runtime events into visible
+transcript rows and session events.
+
+### Headless exec (`zero exec`)
+
+```mermaid
+flowchart TD
+    A[Parse exec flags] --> B[Resolve config + provider]
+    B --> C[Build tool registry]
+    C --> D[Resolve prompt / stream-json input / images]
+    D --> E[Prepare or resume session]
+    E --> F[Create stream-json/text/json writer]
+    F --> G[agent.Run]
+    G --> H[Write final output and exit code]
+```
+
+`zero exec` uses the same `agent.Run` loop as the TUI, but without interactive
+terminal state. It can persist sessions, resume/fork existing sessions, run in
+stream-JSON mode for automation, and always requires a stronger completion
+signal so a no-tool assistant response is not treated as success while work is
+still pending.
+
+## Agent Loop
+
+The core loop lives in `internal/agent.Run`. Conceptually, every turn follows
+this cycle:
+
+```mermaid
+flowchart TD
+    Start[Build system prompt + user prompt] --> Expose[Partition visible tools]
+    Expose --> Compact[Maybe compact context]
+    Compact --> Request[Send CompletionRequest to provider]
+    Request --> Stream[Collect streamed text, reasoning, usage, tool calls]
+
+    Stream --> HasTools{Tool calls?}
+    HasTools -- No --> FinalCheck{Final answer acceptable?}
+    FinalCheck -- Yes --> Done[Return Result]
+    FinalCheck -- No --> Nudge[Append user nudge and continue]
+
+    HasTools -- Yes --> Execute[Execute tool calls]
+    Execute --> Permission[Permission / sandbox / hooks]
+    Permission --> ToolResult[Append tool result messages]
+    ToolResult --> Diagnostics[Post-edit diagnostics / self-correct]
+    Diagnostics --> Expose
+```
+
+Key responsibilities of the loop:
+
+1. **Build prompt state** — combines the system prompt, project/user guidelines,
+   loaded skills/specialist instructions, the current user prompt, and images. The
+   resulting message list is the model-facing context that later tool results are
+   appended to. Resumed session history is replayed by the caller before a new run,
+   not passed as a separate `agent.Run` history field.
+2. **Expose tools** — starts from the registry built by the CLI, then filters
+   tool definitions by permission mode, enabled or disabled filters, and
+   deferred-loading rules. Tool schemas also count toward context budget, so they
+   matter for compaction decisions.
+3. **Call provider** — sends a provider-neutral `zeroruntime.CompletionRequest`.
+4. **Stream events** — forwards text, reasoning, tool-call deltas, and usage to
+   the caller via callbacks.
+5. **Execute tools** — decodes tool arguments, checks filters, applies
+   permissions and sandbox policy, dispatches hooks, runs the registered tool,
+   redacts output, and appends a tool message. See
+   [Tool Execution Lifecycle](#tool-execution-lifecycle).
+6. **Recover and guard** — handles malformed tool calls, repeated failures,
+   empty turns, context pressure, stream reconnect/retry, and max-turn fallback.
+   Context pressure is handled by [Context Budget and Compaction](#context-budget-and-compaction).
+7. **Finalize** — returns a final answer or a structured incomplete/stop reason.
+
+The model never directly mutates the workspace. It asks for tool calls; Zero
+validates and executes those calls through the registry, permission policy, and
+sandbox engine.
+
+## Tool Execution Lifecycle
+
+A tool call is part of the model conversation, not a side channel that bypasses
+it. The model proposes an action, Zero decides whether that action is allowed to
+run, the tool returns a structured result, and that result becomes context for the
+next provider request.
+
+```mermaid
+sequenceDiagram
+    participant Model as LLM / provider
+    participant Agent as agent.Run
+    participant Policy as filters + permissions + sandbox
+    participant Hooks as hooks
+    participant Tool as tool implementation
+    participant Surface as TUI / exec callbacks
+    participant Transcript as model conversation + session events
+
+    Agent->>Model: request with messages + available tool schemas
+    Model-->>Agent: assistant text and/or tool_call(name, args)
+    Agent-->>Surface: stream text / tool-call preview
+    Agent->>Agent: collect and decode tool arguments
+    Agent->>Policy: check tool filters, permission mode, sandbox policy
+
+    alt denied before execution
+        Policy-->>Agent: denied result
+        Agent-->>Surface: tool failed/denied event
+        Agent->>Transcript: append tool result explaining denial
+    else needs user approval
+        Policy-->>Surface: permission request
+        Surface-->>Policy: allow / deny / remember decision
+        Policy-->>Agent: decision
+    end
+
+    alt approved to execute
+        Agent->>Hooks: beforeTool hooks
+        Hooks-->>Agent: allow / block / annotate
+        Agent->>Tool: execute(ctx, args)
+        Tool-->>Agent: output, error, metadata
+        Agent->>Hooks: afterTool hooks
+        Agent->>Agent: redact, truncate, classify success/failure
+        Agent-->>Surface: OnToolResult callback
+        Agent->>Transcript: append tool result message
+    end
+
+    Agent->>Model: next request includes prior assistant tool_call + tool result
+    Model-->>Agent: continue with more tools or final answer
+```
+
+Conceptually, one requested tool goes through these phases:
+
+1. **Proposal** — the provider streams assistant content and may emit one or more
+   tool calls. Zero can show the partial tool arguments in the TUI while they are
+   still streaming, but nothing has executed yet.
+2. **Decode and gate** — the agent validates JSON arguments, checks whether the
+   tool is enabled, applies permission mode, asks the sandbox engine about path,
+   network, shell, or command-prefix constraints, and may invoke permission
+   callbacks.
+3. **Execute** — the selected registry entry runs locally, through MCP, as a
+   specialist/sub-agent, or through a plugin/local-control helper. Hooks can run
+   before or after this step.
+4. **Normalize result** — Zero turns the outcome into an agent `ToolResult`: a
+   success, failure, denial, cancellation, timeout, or validation error with
+   output/metadata that may be redacted or truncated.
+5. **Append to context** — the result is appended as a tool-result message paired
+   with the assistant tool call. The next model request includes that result, so
+   the model can inspect it and decide what to do next.
+6. **Render and persist** — the result is surfaced to the TUI, stream-JSON
+   writer, or ACP client. TUI and exec record rich tool/session events for
+   resume/history; ACP currently persists conversational messages and streams
+   tool/permission updates to the client.
+
+Tool success and failure are both informative context. A successful `read_file`
+result gives the model file contents; a failed `edit_file` result gives the model
+an error such as “old string not found”; a denied shell command tells the model
+that the action is not currently permitted. The model is expected to use that
+feedback to continue, choose another tool, ask the user, or explain the blocker.
+
+The judgement that “information is sufficient” is made by the model, not by the
+TUI. Zero controls the loop mechanics: it keeps appending messages, enforces
+safety, retries or nudges around malformed/empty/incomplete turns, and stops on
+hard limits. But the semantic decision to call another tool versus produce a
+final answer comes from the next provider response. In other words:
+
+```mermaid
+flowchart TD
+    Result[Tool result appended to conversation] --> Next[Next provider request]
+    Next --> Decide{Model response}
+    Decide -- More information/action needed --> MoreTools[Request another tool]
+    MoreTools --> Result
+    Decide -- User input needed --> AskUser[Call ask_user or explain question]
+    AskUser --> Result
+    Decide -- Enough information --> Final[Final assistant answer]
+    Decide -- Invalid / empty / unsafe --> Guard[Agent guardrail, retry, denial, or stop reason]
+    Guard --> Next
+```
+
+This is why tool output is usually more than UI decoration. It is part of the
+working memory for the next model turn, subject to context-window pressure and
+compaction. The TUI may summarize or prettify it for the user, but the agent loop
+also stores the machine-facing result that the model sees.
+
+## Context Budget and Compaction
+
+The `Maybe compact context` step in the agent-loop diagram is Zero's context-window
+pressure valve. It is not a separate memory engine; it rewrites the current
+model-facing conversation so the next provider request fits within the configured
+context window.
+
+Compaction is enabled only when the run knows a positive `ContextWindow`. When it
+is enabled, the loop does this near the top of each turn:
+
+```mermaid
+flowchart TD
+    Tools[Partition visible tools] --> Estimate[Estimate messages + tool-definition tokens]
+    Estimate --> Under{Under threshold?}
+    Under -- Yes --> Keep[Keep context as-is]
+    Under -- No --> Grown{Grown past low-water mark?}
+    Grown -- No --> Keep
+    Grown -- Yes --> Prune[Prune stale large tool outputs]
+    Prune --> Recheck{Back under threshold?}
+    Recheck -- Yes --> KeepPruned[Use pruned context]
+    Recheck -- No --> Summarize[Summarize oldest middle]
+    Summarize --> Replace[system messages + summary + recent suffix]
+```
+
+The trigger is intentionally approximate:
+
+- Zero estimates request size using message text, images, tool-call arguments, and
+  the tool definitions advertised on that turn.
+- The proactive threshold is roughly **two-thirds of the context window**.
+- Provider usage from completed turns is used to calibrate the rough estimate over
+  time, so the budget gauge and compaction decision track the provider more
+  closely after a few turns.
+- A low-water mark prevents the loop from re-summarizing the same already-compacted
+  history on every turn.
+
+When compaction actually runs, it tries to preserve the parts most likely to be
+needed by the model:
+
+- leading system messages stay verbatim;
+- the most recent turns stay verbatim (`CompactionPreserveLast`, defaulting to the
+  agent's normal preserve count);
+- the preserved suffix is widened to a provider-valid turn boundary so tool-result
+  messages are not replayed without their corresponding assistant tool call;
+- the older middle is summarized into one user-role message labeled
+  `[Summary of earlier conversation]`;
+- structured state that should not be paraphrased away, such as active plan/loaded
+  skill state, is preserved alongside the prose summary.
+
+There are two compaction paths:
+
+1. **Proactive compaction** runs before the provider call when the estimated
+   request is over threshold.
+2. **Reactive compaction** runs after a provider returns a context-limit style
+   error. Zero compacts once and retries that same turn; if the provider still
+   rejects the request, the error is surfaced rather than looping forever.
+
+Compaction failures are conservative. If the summarizer fails during proactive
+compaction, Zero keeps the original messages and tries to continue. It does not
+drop conversation history just to make room.
+
+## Presentation Layer: Bubble Tea TUI
+
+`internal/tui` is a Bubble Tea application. The usual Bubble Tea shape applies:
+
+```mermaid
+flowchart LR
+    Init[model.Init] --> Update[model.Update]
+    Update --> View[model.View]
+    View --> Terminal[Terminal frame]
+    Terminal --> Msg[tea.Msg: key, mouse, window, agent event]
+    Msg --> Update
+```
+
+The TUI is intentionally not the agent brain. It is responsible for:
+
+- collecting user input from the composer;
+- rendering assistant text, reasoning, tool cards, permissions, plans, usage,
+  files, and sidebars;
+- handling slash commands and modal surfaces;
+- converting agent callbacks into Bubble Tea messages;
+- persisting visible session events;
+- allowing live provider/model/permission changes.
+
+When a prompt is submitted, the TUI starts a background command that invokes
+`agent.Run`. Streaming callbacks send messages back into the Bubble Tea update
+loop, so rendering remains responsive while the agent is working.
+
+## How the TUI and Agent Loop Connect
+
+A common source of confusion is that Zero has two event loops running at once:
+
+1. **Bubble Tea loop** — owns terminal input, in-memory UI state, and frames.
+2. **Agent loop** — owns model turns, tool execution, permissions, retries, and
+   stop conditions.
+
+They are connected by `tea.Cmd` and `agent.Options` callbacks. The TUI never
+runs provider/tool logic directly inside `Update`; instead, submit handling gates
+and dispatches a prompt, the TUI starts a background command, and that command
+calls `agent.Run` with callbacks that send runtime events back to the Bubble Tea
+program. Function names in this section are illustrative code anchors, not API
+boundaries.
+
+```mermaid
+sequenceDiagram
+    participant User
+    participant TUI as Bubble Tea model<br/>internal/tui
+    participant Cmd as tea.Cmd goroutine<br/>runAgentWithOptions
+    participant Agent as agent.Run<br/>internal/agent
+    participant Provider as Provider.StreamCompletion
+    participant Tools as Tool registry + sandbox
+    participant Store as Session store
+
+    User->>TUI: type prompt / press Enter
+    TUI->>TUI: validate/gate submit and append visible user row
+    TUI->>Cmd: launch tea.Cmd with runID + cancellable context
+    Cmd->>Agent: agent.Run(prompt, provider, options)
+    Agent->>Provider: stream completion request
+    Provider-->>Agent: text / reasoning / tool-call deltas
+    Agent-->>Cmd: OnText / OnReasoning / OnToolCallStart callbacks
+    Cmd-->>TUI: runtimeMessageSink(tea.Msg)
+    TUI->>TUI: Update appends/updates transcript rows
+    TUI-->>User: View renders new frame
+
+    Agent->>Tools: execute collected tool call
+    Tools->>Tools: permission, sandbox, hooks, redaction
+    Tools-->>Agent: tool result
+    Agent-->>Cmd: OnToolCall / OnToolResult / OnPermission / OnUsage
+    Cmd->>Store: batch session events for resume/history
+    Cmd-->>TUI: final agentResponseMsg
+    TUI->>TUI: mark run complete, flush rows/events, unlock composer
+```
+
+In the TUI path, the same event can have three representations:
+
+| Runtime thing | Agent representation | TUI representation | Session representation |
+| --- | --- | --- | --- |
+| Assistant text | streamed provider text forwarded through `OnText` | incremental assistant row text | assistant message event |
+| Reasoning preview | `OnReasoning` callback | collapsible/preview reasoning row | usually not the same as final assistant text |
+| Tool call | `agent.ToolCall` from collected provider stream | tool card, specialist card, or plan panel update | `EventToolCall` |
+| Tool result | `agent.ToolResult` | result card/status, changed files, plan/specialist updates | `EventToolResult` |
+| Permission prompt/audit | `PermissionRequest` / `PermissionEvent` | modal prompt or permission row | permission request/decision events |
+| Usage/context | provider usage + measured context | context gauge / usage metadata | usage event |
+
+### One interactive turn, end to end
+
+For a normal prompt in the interactive TUI:
+
+1. Submit handling validates composer state and command gates, then dispatches
+   the prompt to the launch path.
+2. The launch path appends immediate UI rows, creates a cancellable run context,
+   and starts `runAgent` as a `tea.Cmd`; run bookkeeping assigns the `runID` used
+   to correlate later events.
+3. `runAgentWithOptions` copies the current UI-controlled runtime choices into
+   `agent.Options`: provider/model labels, permission mode, cwd, session ID,
+   images, registry, response style, context window, and self-correction hooks.
+4. The command wraps agent callbacks. For example:
+   - `OnText` sends incremental assistant text into the TUI message sink.
+   - `OnToolCallStart` / `OnToolCallDelta` show live tool-argument streaming.
+   - `OnToolCall` creates visible tool rows and appends session events.
+   - `OnToolResult` creates result rows, updates plan/specialist side panels, and
+     records changed files.
+   - `OnPermissionRequest` sends a prompt to the TUI and blocks the agent loop
+     until the user answers or the run context is cancelled.
+   - `OnAskUser` does the same for model-initiated clarification questions.
+5. `agent.Run` builds provider-neutral messages and tool definitions, streams a
+   provider response, executes any requested tools through the registry/sandbox,
+   appends tool results back into the model conversation, and repeats until it has
+   a final answer or a stop condition.
+6. When `agent.Run` returns, the command emits `agentResponseMsg`. The Bubble Tea
+   loop marks the run complete, flushes queued rows/session events, updates usage
+   and file state, and re-enables the composer.
+
+The important architectural point: **the TUI is stateful and reactive, but the
+agent loop is the source of runtime truth for model/tool progress**. Presentation
+state is derived from callback events and persisted session events; workspace
+changes happen only through tools executed by the agent runtime.
+
+### Feature flows use the same bridge
+
+For one interactive turn, the supporting systems do not run beside the TUI/agent
+loop; they run inside or around it:
+
+```mermaid
+flowchart TD
+    Submit[TUI submit] --> Options[Build agent.Options]
+    Options --> Context[Assemble context<br/>config, guidelines, history, skills]
+    Context --> ProviderCall[Provider stream]
+    ProviderCall --> StreamEvents[Text/reasoning/tool-call events]
+    StreamEvents --> Render[TUI render callbacks]
+
+    ProviderCall --> ToolCall{Tool call?}
+    ToolCall -- No --> Final[Final response]
+    ToolCall -- Yes --> ToolGate[Tool gate<br/>filters, permissions, sandbox]
+    ToolGate --> HooksBefore[beforeTool hooks]
+    HooksBefore --> Execute[Tool / MCP / specialist / plugin]
+    Execute --> HooksAfter[afterTool hooks]
+    HooksAfter --> ToolResult[Tool result message]
+    ToolResult --> Persist[Session events]
+    ToolResult --> Context
+
+    Final --> Persist
+    Persist --> Resume[Resume / fork / rewind / parent metadata]
+```
+
+Most interactive features are variations of that same bridge rather than separate
+architectures:
+
+- **Slash commands** either mutate TUI state directly (`/model`, `/permission`,
+  `/help`) or start short `tea.Cmd` jobs that later return a message to `Update`
+  (`/doctor`, `/search`, `/compact`, session commands).
+- **Permissions** sit between a model-requested tool call and actual execution:
+  the agent blocks on `OnPermissionRequest`, the TUI surfaces a modal and sends
+  the decision back through a channel, and the session log records the
+  request/decision events.
+- **Plans, specialists/sub-agents, and ask-user prompts** all follow the same
+  runtime pattern: the model requests a capability, the agent records the
+  call/result in its conversation, and the TUI mirrors the visible state as plan
+  panels, specialist cards, or question forms. Specialist session metadata can
+  carry parent/root and sub-agent/spec fields, with richer linkage depending on
+  the specialist or swarm path.
+- **Streaming tool writes** start as provider tool-call deltas before execution;
+  the TUI shows the arguments live, then replaces/augments them with the actual
+  `ToolResult` after the tool runs.
+- **Compaction** is an agent-loop context-management step; the TUI may show that
+  it happened, but the purpose is to keep the next provider request within budget.
+- **Stream-JSON** is another surface over the same callbacks: `zero exec`
+  serializes the events as JSON protocol events instead of Bubble Tea views.
+
+So a feature-specific deep dive should usually explain: what user event starts it,
+which `tea.Msg`/callback carries it, which agent/tool event is authoritative, and
+which session events make it resumable.
+
+## Runtime Provider Layer
+
+Providers are hidden behind this interface in `internal/zeroruntime`:
+
+```go
+type Provider interface {
+    StreamCompletion(ctx context.Context, request CompletionRequest) (<-chan StreamEvent, error)
+}
+```
+
+The agent loop only knows about provider-neutral messages, tool definitions,
+stream events, reasoning blocks, images, and token usage. Provider adapters in
+`internal/providers/*` translate those neutral types to provider-specific APIs:
+
+- OpenAI and OpenAI-compatible APIs;
+- Anthropic and Anthropic-compatible APIs;
+- Gemini;
+- Codex / ChatGPT OAuth-backed runtime.
+
+`internal/providers.New` resolves the configured model through the model
+registry/catalog, normalizes API model IDs and base URLs, attaches API keys or
+OAuth resolvers, and returns the concrete adapter.
+
+## Tools and Capabilities
+
+Tools implement a common interface and are registered by name in
+`internal/tools.Registry`.
+
+```mermaid
+flowchart TD
+    AgentToolCall[Tool call from model] --> Decode[Decode JSON args]
+    Decode --> Filter[Enabled/disabled + permission-mode filter]
+    Filter --> Precheck[Pre-permission rejecters]
+    Precheck --> SandboxEval[Sandbox Evaluate]
+    SandboxEval --> Prompt{Needs user approval?}
+    Prompt -- Yes --> Decision[Permission callback]
+    Prompt -- No --> Run[Registry.RunWithOptions]
+    Decision --> Run
+    Run --> Redact[Redact secrets + enforce output ceiling]
+    Redact --> Message[Return tool result to model]
+```
+
+Core tool groups include:
+
+- **Read-only tools**: file reads, directory listing, glob, grep, LSP navigation,
+  skills, `ask_user`, permission requests.
+- **Write tools**: `write_file`, `edit_file`, `apply_patch`, `update_plan`.
+- **Shell tools**: `exec_command`, `write_stdin`, legacy `bash`.
+- **Network tools**: `web_fetch` and optional web search backend.
+- **Local-control tools**: config-gated wrappers for browser, terminal, desktop,
+  and artifact helpers backed by packages such as `internal/localcontrol` and
+  `internal/browser`.
+- **Extension tools**: MCP tools, specialists/sub-agents, swarm tools, plugins.
+
+Tool definitions include safety metadata: side-effect class, permission level,
+reason, and whether prompt-gated tools are advertised in auto-like modes. The
+agent uses this metadata plus per-turn filters and deferred-loading state to
+choose which registered tools the model can see and whether a call can run
+immediately.
+
+## Permission and Sandbox Model
+
+Zero has two related but separate checks:
+
+1. **Tool permission metadata** says whether a tool is allowed, denied, or needs
+   approval.
+2. **Sandbox policy** evaluates the actual request: path scope, network access,
+   destructive shell patterns, explicit escalation, persistent/session grants,
+   and platform isolation availability.
+
+```mermaid
+flowchart TD
+    Call[Tool call] --> Safety[Tool Safety metadata]
+    Call --> Scope[Workspace/additional roots scope]
+    Safety --> Engine[Sandbox Engine Evaluate]
+    Scope --> Engine
+    Engine --> Allow[Allow]
+    Engine --> Prompt[Prompt user]
+    Engine --> Deny[Deny]
+    Prompt --> Grant[Turn/session/persistent grant]
+    Grant --> Execute[Execute tool]
+    Allow --> Execute
+    Deny --> Result[Return denied tool result]
+```
+
+The sandbox engine can:
+
+- auto-allow workspace writes for safe file-edit tools;
+- block reads/writes outside the allowed scope;
+- prompt for recoverable outside-workspace access;
+- deny or prompt for network depending on policy;
+- detect destructive shell commands;
+- run shell commands through a platform backend when available;
+- remember session or persistent grants;
+- support narrow command-prefix approvals.
+
+The user-facing TUI and `zero exec` mode both route permission requests through
+callbacks, but the TUI can show interactive modals while headless mode generally
+returns non-interactive fallback behavior unless a protocol/client handles it.
+
+## Session Storage and History
+
+Sessions live locally under Zero's data directory. A session directory contains:
+
+- `metadata.json` — session ID, title, cwd, provider/model, parent/root IDs,
+  spec/sub-agent fields, timestamps, and event count.
+- `events.jsonl` — append-only event log with messages, tool calls/results,
+  permissions, usage, checkpoints, forks, compaction, and specs. TUI and exec use
+  the richer event stream; ACP currently records conversational message history
+  while streaming tool and permission updates to the client.
+
+```mermaid
+flowchart LR
+    TUI[TUI callbacks] --> Store[internal/sessions.Store]
+    Exec[exec recorder] --> Store
+    Store --> Metadata[metadata.json]
+    Store --> Events[events.jsonl]
+    Events --> Resume[resume]
+    Events --> Fork[fork]
+    Events --> Rewind[rewind/checkpoints]
+    Events --> Children[parent/root metadata]
+```
+
+This is why sessions can be resumed, forked, rewound, and inspected without a
+remote service. Zero stores the local event history; providers only receive the
+messages needed for the current completion request.
+
+## Stream-JSON Protocol
+
+`zero exec --output-format stream-json` emits newline-delimited JSON events for
+integrations and CI. The protocol is implemented in `internal/streamjson` and is
+specified in [STREAM_JSON_PROTOCOL.md](STREAM_JSON_PROTOCOL.md).
+
+Common output event types include:
+
+- `run_start`, `text`, `reasoning`, `tool_call`, `tool_result`;
+- `permission`, `permission_request`, `permission_decision`;
+- `checkpoint`, `restore`, `usage`, `final`, `warning`, `error`, `run_end`.
+
+Input stream-JSON can provide `prompt` or `message` events and image payloads.
+The CLI parses those events, resolves a single prompt, validates/decodes images,
+and passes the result into the same agent loop used everywhere else.
+
+## Extensibility Flow
+
+Zero's extension surfaces are loaded before the agent loop starts:
+
+```mermaid
+flowchart TD
+    Config[Config resolution] --> Registry[Tool registry]
+    Registry --> Core[Core tools]
+    Registry --> MCP[MCP server tools]
+    Registry --> Specialist[Specialist / Task tools]
+    Registry --> Plugin[Plugin tools]
+
+    Config --> Prompt[System prompt context]
+    Prompt --> Guidelines[AGENTS.md / ZERO.md]
+    Prompt --> Skills[Skills list]
+    Prompt --> SpecialistContext[Specialist descriptions]
+
+    ToolLifecycle[Tool lifecycle] --> Hooks[beforeTool / afterTool hooks]
+```
+
+- **Project/user guidelines** become prompt context.
+- **Skills** are instruction packs the model can load on demand.
+- **Specialists** are sub-agents callable through the `Task` tool.
+- **MCP servers** contribute external tools.
+- **Plugins** can add tools, hooks, and skill roots. Bootstrap 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.
+- **Hooks** can observe or block tool lifecycle events.
+
+## End-to-End Data Flow
+
+A typical interactive edit request looks like this:
+
+```mermaid
+sequenceDiagram
+    participant U as User
+    participant T as TUI
+    participant A as Agent loop
+    participant P as Provider
+    participant R as Tool registry
+    participant S as Sandbox
+    participant FS as Workspace/session files
+
+    U->>T: Type prompt and press Enter
+    T->>A: agent.Run(prompt, provider, options)
+    A->>P: CompletionRequest(messages, tool definitions)
+    P-->>A: Stream text + tool call edit_file
+    A->>S: Evaluate edit_file args
+    S-->>A: allow/prompt/deny decision
+    A->>T: permission callback if prompt needed
+    T-->>A: user decision
+    A->>R: Run edit_file
+    R->>FS: Write workspace file
+    R-->>A: Tool result + changed files
+    A->>P: Next CompletionRequest with tool result
+    P-->>A: Final answer
+    A-->>T: Result + callbacks
+    T->>FS: Append session events
+    T-->>U: Render final transcript
+```
+
+The same request in `zero exec` follows the same provider/tool/sandbox path, but
+streaming events are written to stdout/stderr instead of rendered in Bubble Tea.
+
+## Mental Model
+
+When you work on Zero, start by locating the change on the four-layer spine from
+[Architecture Spine](#architecture-spine) — Surface, Agent loop, Capabilities,
+Persistence — then jump to the package that owns that boundary:
+
+- UI behavior? Start in `internal/tui`.
+- Command-line behavior? Start in `internal/cli`.
+- Model/tool-turn behavior? Start in `internal/agent`.
+- A new capability? Register a tool in `internal/tools` or an extension surface.
+- Safety policy? Start in `internal/sandbox` and agent permission wiring.
+- Resume/history behavior? Start in `internal/sessions` and session recorders.

docs/INSTALL.md

@@ -6,8 +6,9 @@ Zero is distributed as:
 - release archives on GitHub Releases
 - source builds with Go 1.25+
 
-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:
@@ -252,8 +226,7 @@ Supported targets:
 - `linux-arm64`
 - `macos-x64`
 - `macos-arm64`
-- `windows-x64`
-- `windows-arm64`
+- `windows-x64` (Windows on ARM runs this build under emulation)
 
 Each archive must have a matching `.sha256` file. The install scripts download
 both files, verify the checksum, and then copy the binary into the install

docs/NPM_PACKAGING.md

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

docs/NPM_WRAPPER_SMOKE.md

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

docs/README.md

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

go.mod

@@ -1,8 +1,6 @@
 module github.com/Gitlawb/zero
 
-go 1.25.0
-
-toolchain go1.26.4
+go 1.26.5
 
 require (
 	charm.land/bubbles/v2 v2.1.0
@@ -14,6 +12,7 @@ require (
 	github.com/charmbracelet/colorprofile v0.4.3
 	github.com/charmbracelet/x/ansi v0.11.7
 	github.com/charmbracelet/x/term v0.2.2
+	github.com/coder/websocket v1.8.15
 	github.com/ledongthuc/pdf v0.0.0-20250511090121-5959a4027728
 	golang.org/x/sys v0.46.0
 	mvdan.cc/sh/v3 v3.13.1

go.sum

@@ -32,6 +32,8 @@ github.com/clipperhouse/displaywidth v0.11.0 h1:lBc6kY44VFw+TDx4I8opi/EtL9m20WSE
 github.com/clipperhouse/displaywidth v0.11.0/go.mod h1:bkrFNkf81G8HyVqmKGxsPufD3JhNl3dSqnGhOoSD/o0=
 github.com/clipperhouse/uax29/v2 v2.7.0 h1:+gs4oBZ2gPfVrKPthwbMzWZDaAFPGYK72F0NJv2v7Vk=
 github.com/clipperhouse/uax29/v2 v2.7.0/go.mod h1:EFJ2TJMRUaplDxHKj1qAEhCtQPW2tJSwu5BF98AuoVM=
+github.com/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/go-quicktest/qt v1.102.0 h1:HSQxCeh5YZH3EL3W39ixjtyaEhcWSXQHtHnMBzSs474=

internal/acp/agent.go

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

internal/acp/agent_test.go

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

internal/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/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

@@ -109,7 +109,7 @@ const escalationFailedNoticePrefix = "Note: could not switch to the requested mo
 // The system prompt (core coding-craft instructions + workspace context + safety
 // confirmation policy) is assembled in system_prompt.go via buildSystemPrompt.
 
-func Run(ctx context.Context, prompt string, provider Provider, options Options) (Result, error) {
+func Run(ctx context.Context, prompt string, provider Provider, options Options) (result Result, err error) {
 	if provider == nil {
 		return Result{}, errors.New("agent provider is required")
 	}
@@ -161,7 +161,11 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 	// doesn't re-run the recursive schema→map conversion for every tool every turn.
 	toolDefCache := map[string]zeroruntime.ToolDefinition{}
 
-	result := Result{Messages: copyMessages(messages)}
+	result = Result{Messages: copyMessages(messages)}
+	dispatchSessionStart(ctx, options)
+	defer func() {
+		dispatchSessionEnd(ctx, options, result, err)
+	}()
 	for turn := 0; turn < maxTurns; turn++ {
 		result.Turns = turn + 1
 
@@ -721,7 +725,8 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 			Content: nudge,
 		})
 	}
-	if answer, finalMessages, finishReason := finalAnswerAfterMaxTurns(ctx, provider, messages, options); strings.TrimSpace(answer) != "" {
+	finalExposed, _ := partitionToolsCached(registry, permissionMode, options, loaded, toolDefCache)
+	if answer, finalMessages, finishReason := finalAnswerAfterMaxTurns(ctx, provider, messages, finalExposed, options); strings.TrimSpace(answer) != "" {
 		result.FinalAnswer = answer
 		result.FinishReason = finishReason
 		result.Messages = copyMessages(finalMessages)
@@ -741,7 +746,7 @@ func Run(ctx context.Context, prompt string, provider Provider, options Options)
 	return result, nil
 }
 
-func finalAnswerAfterMaxTurns(ctx context.Context, provider Provider, messages []zeroruntime.Message, options Options) (string, []zeroruntime.Message, string) {
+func finalAnswerAfterMaxTurns(ctx context.Context, provider Provider, messages []zeroruntime.Message, toolDefs []zeroruntime.ToolDefinition, options Options) (string, []zeroruntime.Message, string) {
 	finalMessages := copyMessages(messages)
 	finalMessages = append(finalMessages, zeroruntime.Message{
 		Role:    zeroruntime.MessageRoleUser,
@@ -752,6 +757,7 @@ func finalAnswerAfterMaxTurns(ctx context.Context, provider Provider, messages [
 	// transient hiccup doesn't drop the final summary (AUDIT-L1).
 	stream, err := streamWithReconnect(ctx, provider, zeroruntime.CompletionRequest{
 		Messages:        copyMessages(finalMessages),
+		Tools:           toolDefs,
 		ReasoningEffort: options.ReasoningEffort,
 		PromptCacheKey:  options.SessionID,
 	}, reconnectNoticeFor(options))
@@ -1568,6 +1574,54 @@ func dispatchAfterTool(ctx context.Context, options Options, call ToolCall, args
 	return strings.TrimSpace(strings.Join(outcome.Messages, "\n"))
 }
 
+// dispatchSessionStart runs configured sessionStart hooks once before the first
+// model turn. Lifecycle hooks are advisory: dispatcher failures are audited but
+// never block the run.
+func dispatchSessionStart(ctx context.Context, options Options) {
+	if options.Hooks == nil {
+		return
+	}
+	options.Hooks.Dispatch(ctx, hooks.DispatchInput{
+		Event: hooks.EventSessionStart,
+		Payload: map[string]any{
+			"event":        string(hooks.EventSessionStart),
+			"sessionId":    options.SessionID,
+			"cwd":          options.Cwd,
+			"provider":     options.ProviderName,
+			"model":        options.Model,
+			"depth":        options.Depth,
+			"tag":          options.Tag,
+			"sessionTitle": options.SessionTitle,
+		},
+	})
+}
+
+// dispatchSessionEnd runs configured sessionEnd hooks once when the agent run
+// exits, including early error returns. Lifecycle hooks are advisory.
+func dispatchSessionEnd(ctx context.Context, options Options, result Result, runErr error) {
+	if options.Hooks == nil {
+		return
+	}
+	payload := map[string]any{
+		"event":        string(hooks.EventSessionEnd),
+		"sessionId":    options.SessionID,
+		"cwd":          options.Cwd,
+		"provider":     options.ProviderName,
+		"model":        options.Model,
+		"turns":        result.Turns,
+		"stopReason":   string(result.StopReason),
+		"finishReason": result.FinishReason,
+		"incomplete":   result.Incomplete,
+	}
+	if runErr != nil {
+		payload["error"] = runErr.Error()
+	}
+	options.Hooks.Dispatch(ctx, hooks.DispatchInput{
+		Event:   hooks.EventSessionEnd,
+		Payload: payload,
+	})
+}
+
 // blockedByHookResult is the tool result for a call vetoed by a beforeTool hook.
 func blockedByHookResult(call ToolCall, outcome hooks.DispatchOutcome) ToolResult {
 	// The hook reason (its stdout/stderr) is model-visible here, an intercepted

internal/agent/loop_test.go

@@ -11,6 +11,7 @@ import (
 	"strings"
 	"testing"
 
+	"github.com/Gitlawb/zero/internal/hooks"
 	"github.com/Gitlawb/zero/internal/sandbox"
 	"github.com/Gitlawb/zero/internal/specmode"
 	"github.com/Gitlawb/zero/internal/tools"
@@ -38,6 +39,59 @@ func (provider *mockProvider) StreamCompletion(ctx context.Context, request zero
 	return ch, nil
 }
 
+func TestRunDispatchesSessionLifecycleHooks(t *testing.T) {
+	audit, err := hooks.NewAuditStore(hooks.AuditStoreOptions{AuditPath: filepath.Join(t.TempDir(), "audit.jsonl")})
+	if err != nil {
+		t.Fatalf("NewAuditStore: %v", err)
+	}
+	dispatcher := hooks.NewDispatcher(hooks.DispatcherOptions{
+		Config: hooks.Config{
+			Enabled: true,
+			Hooks: []hooks.Definition{
+				{ID: "zero.session-start", Event: hooks.EventSessionStart, Command: "zero-missing-hook-command", Enabled: true},
+				{ID: "zero.session-end", Event: hooks.EventSessionEnd, Command: "zero-missing-hook-command", Enabled: true},
+			},
+		},
+		Audit: audit,
+	})
+	provider := &mockProvider{turns: [][]zeroruntime.StreamEvent{{
+		{Type: zeroruntime.StreamEventText, Content: "done"},
+		{Type: zeroruntime.StreamEventDone},
+	}}}
+
+	_, err = Run(context.Background(), "hi", provider, Options{
+		SessionID:    "session-123",
+		Cwd:          t.TempDir(),
+		ProviderName: "test-provider",
+		Model:        "test-model",
+		Hooks:        dispatcher,
+	})
+	if err != nil {
+		t.Fatalf("Run: %v", err)
+	}
+
+	events, err := audit.ReadEvents()
+	if err != nil {
+		t.Fatalf("ReadEvents: %v", err)
+	}
+	started := map[hooks.Event]int{}
+	completed := map[hooks.Event]int{}
+	for _, event := range events {
+		switch event.Type {
+		case "hook_execution_started":
+			started[event.Event]++
+		case "hook_execution_completed":
+			completed[event.Event]++
+		}
+	}
+	if started[hooks.EventSessionStart] != 1 || completed[hooks.EventSessionStart] != 1 {
+		t.Fatalf("sessionStart audit counts started/completed = %d/%d, events=%#v", started[hooks.EventSessionStart], completed[hooks.EventSessionStart], events)
+	}
+	if started[hooks.EventSessionEnd] != 1 || completed[hooks.EventSessionEnd] != 1 {
+		t.Fatalf("sessionEnd audit counts started/completed = %d/%d, events=%#v", started[hooks.EventSessionEnd], completed[hooks.EventSessionEnd], events)
+	}
+}
+
 type recordingWebSearchTool struct {
 	calls []map[string]any
 }
@@ -2768,11 +2822,11 @@ func TestRunRequestsFinalAnswerAfterMaxTurns(t *testing.T) {
 		t.Fatalf("expected final answer from finalization turn, got %q", result.FinalAnswer)
 	}
 	if len(provider.requests) != 2 {
-		t.Fatalf("expected final no-tools request after max turns, got %d requests", len(provider.requests))
+		t.Fatalf("expected finalization request after max turns, got %d requests", len(provider.requests))
 	}
 	finalRequest := provider.requests[1]
-	if len(finalRequest.Tools) != 0 {
-		t.Fatalf("finalization request must not advertise tools, got %#v", finalRequest.Tools)
+	if toolDefinitionByName(finalRequest.Tools, "read_file") == nil {
+		t.Fatalf("finalization request must keep prior tools exposed for strict provider replay, got %#v", finalRequest.Tools)
 	}
 	lastMessage := finalRequest.Messages[len(finalRequest.Messages)-1]
 	if lastMessage.Role != zeroruntime.MessageRoleUser || !strings.Contains(lastMessage.Content, "tool-turn limit") {
@@ -2913,11 +2967,22 @@ func TestBuildSystemPromptInjectsHostShellContext(t *testing.T) {
 		t.Fatalf("expected operating system in environment block, got %q", prompt)
 	}
 	if runtime.GOOS == "windows" {
-		for _, want := range []string{"Windows cmd.exe syntax", "cwd argument", "MSYS binaries", "grep", "require_escalated"} {
+		for _, want := range []string{"Windows cmd.exe syntax", "cwd argument", "MSYS binaries", "grep", "require_escalated", "use double quotes around the value"} {
 			if !strings.Contains(prompt, want) {
 				t.Fatalf("expected Windows shell guidance to mention %q, got %q", want, prompt)
 			}
 		}
+		// The examples must themselves use the safe (double-quoted) form; a
+		// single-quoted example here would teach the model the exact syntax
+		// that fails under cmd.exe.
+		for _, want := range []string{`--jq ".a | b"`, `-run "A|B"`} {
+			if !strings.Contains(prompt, want) {
+				t.Fatalf("expected double-quoted example %q in Windows shell guidance, got %q", want, prompt)
+			}
+		}
+		if strings.Contains(prompt, `'.a | b'`) || strings.Contains(prompt, `'A|B'`) {
+			t.Fatalf("Windows shell guidance must not show the unsafe single-quoted form, got %q", prompt)
+		}
 	} else if !strings.Contains(prompt, "/bin/sh syntax") {
 		t.Fatalf("expected POSIX shell guidance in prompt, got %q", prompt)
 	}

internal/agent/system_prompt.go

@@ -286,7 +286,7 @@ func workspaceContext(cwd string) string {
 	b.WriteString("Working directory: " + cwd + "\n")
 	b.WriteString("Operating system: " + runtime.GOOS + "\n")
 	if runtime.GOOS == "windows" {
-		b.WriteString("Shell syntax: Windows cmd.exe syntax for exec_command/bash tools; prefer the workdir/cwd argument instead of cd when changing directories. Do not pipe to or invoke POSIX coreutils from Git for Windows (usr\\bin head/grep/tail/cat/...): they are MSYS binaries and fail under the write-restricted sandbox; use native Zero tools (grep, read_file, list_directory, glob) or cmd.exe findstr/more instead, or sandbox_permissions require_escalated only when host-level execution is truly required.\n")
+		b.WriteString("Shell syntax: Windows cmd.exe syntax for exec_command/bash tools. To put | & > < etc inside an arg value, use double quotes around the value, not single quotes (single quotes do not protect metachars in cmd.exe): gh --jq \".a | b\", go test -run \"A|B\". Do not pipe to or invoke POSIX coreutils from Git for Windows (usr\\bin head/grep/tail/cat/...): they are MSYS binaries and fail under the write-restricted sandbox; use native Zero tools (grep, read_file, list_directory, glob) or cmd.exe findstr/more instead, or sandbox_permissions require_escalated only when host-level execution is truly required. Prefer the workdir/cwd argument over cd when changing directories.\n")
 	} else {
 		b.WriteString("Shell syntax: /bin/sh syntax for exec_command/bash tools; prefer the workdir/cwd argument instead of cd when changing directories.\n")
 	}

internal/cli/app.go

@@ -25,6 +25,7 @@ import (
 	"github.com/Gitlawb/zero/internal/plugins"
 	"github.com/Gitlawb/zero/internal/providerhealth"
 	"github.com/Gitlawb/zero/internal/providermodeldiscovery"
+	"github.com/Gitlawb/zero/internal/provideroauth"
 	"github.com/Gitlawb/zero/internal/provideronboarding"
 	"github.com/Gitlawb/zero/internal/providers"
 	"github.com/Gitlawb/zero/internal/redaction"
@@ -50,7 +51,7 @@ type appDeps struct {
 	stdin            io.Reader
 	userConfigPath   func() (string, error)
 	resolveConfig    func(workspaceRoot string, overrides config.Overrides) (config.ResolvedConfig, error)
-	resolveMCPConfig func(workspaceRoot string) (config.MCPConfig, error)
+	resolveMCPConfig func(workspaceRoot string, excludeProject bool) (config.MCPConfig, error)
 	newProvider      func(config.ProviderProfile) (zeroruntime.Provider, error)
 	// exportActiveProvider pins spawned children to the run's provider (production:
 	// config.SetActiveProviderEnv, set in defaultAppDeps — deliberately NOT filled
@@ -60,6 +61,7 @@ type appDeps struct {
 	probeProviderHealth    func(context.Context, providerhealth.Options) providerhealth.Result
 	discoverProviderModels func(context.Context, config.ProviderProfile) ([]providermodeldiscovery.Model, error)
 	detectLocalRuntimes    func(context.Context, provideronboarding.LocalDetectOptions) []provideronboarding.DetectedLocalRuntime
+	openRouterLogin        func(context.Context, provideroauth.OpenRouterOptions) (string, error)
 	newSessionStore        func() *sessions.Store
 	loadPlugins            func(plugins.LoadOptions) (plugins.LoadResult, error)
 	loadHooks              func(hooks.LoadOptions) (hooks.LoadResult, error)
@@ -123,11 +125,12 @@ func defaultAppDeps() appDeps {
 			options.Overrides = overrides
 			return config.Resolve(options)
 		},
-		resolveMCPConfig: func(workspaceRoot string) (config.MCPConfig, error) {
+		resolveMCPConfig: func(workspaceRoot string, excludeProject bool) (config.MCPConfig, error) {
 			options, err := config.DefaultResolveOptions(workspaceRoot)
 			if err != nil {
 				return config.MCPConfig{}, err
 			}
+			options.ExcludeProject = excludeProject
 			return config.ResolveMCP(options)
 		},
 		newProvider: func(profile config.ProviderProfile) (zeroruntime.Provider, error) {
@@ -144,6 +147,7 @@ func defaultAppDeps() appDeps {
 		probeProviderHealth:    providerhealth.Probe,
 		discoverProviderModels: defaultDiscoverProviderModels,
 		detectLocalRuntimes:    provideronboarding.DetectLocalRuntimes,
+		openRouterLogin:        provideroauth.OpenRouterLogin,
 		newSessionStore: func() *sessions.Store {
 			return sessions.NewStore(sessions.StoreOptions{})
 		},
@@ -403,6 +407,8 @@ func runWithDeps(args []string, stdout io.Writer, stderr io.Writer, deps appDeps
 		return runWorktrees(args[1:], stdout, stderr, deps)
 	case "verify":
 		return runVerifyCommand(args[1:], stdout, stderr, deps)
+	case "trust":
+		return runTrust(args[1:], stdout, stderr, deps)
 	case "eval":
 		return runAgentEvalCommand(args[1:], stdout, stderr, deps)
 	case "changes", "change":
@@ -465,6 +471,9 @@ func fillAppDeps(deps appDeps) appDeps {
 	if deps.detectLocalRuntimes == nil {
 		deps.detectLocalRuntimes = defaults.detectLocalRuntimes
 	}
+	if deps.openRouterLogin == nil {
+		deps.openRouterLogin = defaults.openRouterLogin
+	}
 	if deps.newSessionStore == nil {
 		deps.newSessionStore = defaults.newSessionStore
 	}
@@ -554,6 +563,11 @@ func fillAppDeps(deps appDeps) appDeps {
 	deps.newProvider = func(profile config.ProviderProfile) (zeroruntime.Provider, error) {
 		return baseNewProvider(applyStoredProviderKeyAt(profile, userConfigPath))
 	}
+	baseProbeProviderHealth := deps.probeProviderHealth
+	deps.probeProviderHealth = func(ctx context.Context, options providerhealth.Options) providerhealth.Result {
+		options.Profile = applyStoredProviderKeyAt(options.Profile, userConfigPath)
+		return baseProbeProviderHealth(ctx, options)
+	}
 	return deps
 }
 
@@ -658,7 +672,17 @@ func runInteractiveTUIWithSetup(stderr io.Writer, deps appDeps, permissionMode a
 		return writeAppError(stderr, "failed to initialize specialist tools: "+err.Error(), 1)
 	}
 	defer closeSpecialistRuntime(stderr, specialistRuntime)
-	mcpConfig, err := deps.resolveMCPConfig(workspaceRoot)
+	// The TUI has no --worktree reassignment, so trustRoot == workspaceRoot here.
+	// Gate the project MCP layer behind the workspace-trust check (fail-closed): an
+	// untrusted workspace must not spawn its ./.zero/config.json stdio MCP servers.
+	// Keep the store-read error so the notice below can distinguish a fail-closed
+	// store error from a clean untrusted verdict.
+	mcpExcludeProject, mcpTrustErrored := resolveTrust(workspaceRoot)
+	mcpSkip := trustSkip{
+		excludedProjectConfig: mcpExcludeProject && projectMCPConfigExists(workspaceRoot),
+		trustCheckErrored:     mcpTrustErrored,
+	}
+	mcpConfig, err := deps.resolveMCPConfig(workspaceRoot, mcpExcludeProject)
 	if err != nil {
 		return writeAppError(stderr, err.Error(), 1)
 	}
@@ -686,15 +710,24 @@ func runInteractiveTUIWithSetup(stderr io.Writer, deps appDeps, permissionMode a
 	defer closeMCPRuntime(stderr, mcpRuntime)
 	// A server that could not be reached or validated is skipped, not fatal (one
 	// bad MCP server must not abort startup) — surface each so a missing tool set is
-	// explained rather than silently absent.
+	// explained rather than silently absent. A built-in default the user never
+	// configured (e.g. keyless Firecrawl with no credentials) is the exception: it
+	// was never asked for, so its failure is not worth a startup warning — only
+	// servers the user actually configured warn on failure (issue #552).
 	for _, skipped := range mcpRuntime.Skipped() {
+		if skipped.UnconfiguredDefault {
+			continue
+		}
 		fmt.Fprintf(stderr, "warning: MCP server %s unavailable, skipped: %s\n", skipped.Name, redaction.ErrorMessage(skipped.Err, redaction.Options{}))
 	}
 	// Make local plugins live: register their declared tools into the registry and
 	// collect their hooks + skill roots for the dispatcher and skill tool below.
 	// Done after specialist + MCP registration so plugin tools are part of the
 	// deferral count, and it fails OPEN — a malformed plugin is warned and skipped.
-	pluginActivation := activatePlugins(workspaceRoot, registry, deps, stderr)
+	// The interactive TUI is not worktree-reassigned, so the trust root is the
+	// launch directory itself.
+	trustRoot := workspaceRoot
+	pluginActivation := activatePlugins(workspaceRoot, registry, deps, stderr, trustRoot)
 	// Ask (not Auto) is the interactive default: in Auto, ToolAdvertised exposes
 	// only PermissionAllow tools, so prompt-gated tools (write_file/edit_file/bash/
 	// apply_patch) would never be offered to the model — the TUI could neither edit
@@ -729,8 +762,26 @@ func runInteractiveTUIWithSetup(stderr io.Writer, deps appDeps, permissionMode a
 		Scope:         scope,
 	})
 	lastKnownMCPConfig := mcpConfig
+	fileTracker := tools.NewFileTracker()
+	var scratchBaseline scratchFileBaseline
+	sttServerManager := newDictationServerManager(resolved.STT)
+	// Keep STT downloads in the SAME config tree the rest of the TUI uses. Deriving
+	// from userConfigPath (rather than config.UserConfigDir()) matters when the config
+	// root is overridden — e.g. in tests or a custom ZERO config dir — so the two
+	// don't diverge. userConfigPath points at .../zero/config.json, so its dir is the
+	// zero config dir.
+	sttDownloadRoot := ""
+	if userConfigPath != "" {
+		sttDownloadRoot = filepath.Join(filepath.Dir(userConfigPath), "stt")
+	}
+	// Build the hooks dispatcher out of the AgentOptions literal so its trust skip
+	// report can be combined with the plugin activation's, and emit at most one
+	// notice when project hooks/plugins were dropped for an untrusted workspace.
+	hookDispatcher, hookSkip := newHookDispatcherWithExtra(workspaceRoot, pluginActivation.hooks, trustRoot)
+	emitTrustNotice(stderr, hookSkip, pluginActivation.trustSkip, mcpSkip)
 	return deps.runTUI(context.Background(), tui.Options{
 		Cwd:                  workspaceRoot,
+		Version:              version,
 		Theme:                theme,
 		SavedTheme:           resolved.Preferences.Theme,
 		UserConfigPath:       userConfigPath,
@@ -741,17 +792,24 @@ func runInteractiveTUIWithSetup(stderr io.Writer, deps appDeps, permissionMode a
 		ProviderProfile:      resolved.Provider,
 		SavedProviders:       usableSavedProviders(resolved.Providers),
 		FavoriteModels:       resolved.Preferences.FavoriteModels,
+		RecentModels:         resolved.Preferences.RecentModels,
 		RecapsEnabled:        resolved.Preferences.RecapsEnabled(),
 		Provider:             provider,
 		NewProvider:          deps.newProvider,
 		ProbeProviderHealth:  deps.probeProviderHealth,
 		UserAgent:            userAgent(),
-		Registry:             registry,
-		SessionStore:         deps.newSessionStore(),
-		SandboxStore:         sandboxStore,
-		MCPConfig:            mcpConfig,
-		MCPPermissionStore:   mcpPermissionStore,
-		MCPTokenStore:        mcpTokenStore,
+		PrepareRunCompletionWarning: func() {
+			scratchBaseline = scratchFileSnapshot(workspaceRoot)
+		},
+		RunCompletionWarning: func() string {
+			return scratchFileWarning(workspaceRoot, scratchBaseline)
+		},
+		Registry:           registry,
+		SessionStore:       deps.newSessionStore(),
+		SandboxStore:       sandboxStore,
+		MCPConfig:          mcpConfig,
+		MCPPermissionStore: mcpPermissionStore,
+		MCPTokenStore:      mcpTokenStore,
 		MCPCommand: func(ctx context.Context, args []string) tui.MCPCommandResult {
 			if ctx == nil {
 				ctx = context.Background()
@@ -759,7 +817,7 @@ func runInteractiveTUIWithSetup(stderr io.Writer, deps appDeps, permissionMode a
 			var stdout, stderr bytes.Buffer
 			exitCode := runMCPWithContext(ctx, args, &stdout, &stderr, deps)
 			nextConfig := lastKnownMCPConfig
-			if refreshed, err := deps.resolveMCPConfig(workspaceRoot); err == nil {
+			if refreshed, err := deps.resolveMCPConfig(workspaceRoot, mcpExcludeProject); err == nil {
 				lastKnownMCPConfig = refreshed
 				nextConfig = refreshed
 			}
@@ -777,8 +835,8 @@ func runInteractiveTUIWithSetup(stderr io.Writer, deps appDeps, permissionMode a
 			PermissionMode: permissionMode,
 			Autonomy:       "low",
 			Sandbox:        sandboxEngine,
-			FileTracker:    tools.NewFileTracker(),
-			Hooks:          newHookDispatcherWithExtra(workspaceRoot, pluginActivation.hooks),
+			FileTracker:    fileTracker,
+			Hooks:          hookDispatcher,
 			DeferThreshold: resolved.Tools.DeferThreshold,
 			Specialists:    specialistRuntime.specialists,
 			Skills:         pluginActivation.skillInfos(deps.skillsDir()),
@@ -791,9 +849,15 @@ func runInteractiveTUIWithSetup(stderr io.Writer, deps appDeps, permissionMode a
 			merged, _ := plugins.MergedSkillsLoaded(deps.skillsDir(), pluginActivation.skillRoots)
 			return merged
 		}),
-		PermissionMode: permissionMode,
-		Notify:         resolved.Notify,
-		KeyBindings:    resolved.KeyBindings,
+		PermissionMode:            permissionMode,
+		Notify:                    resolved.Notify,
+		KeyBindings:               resolved.KeyBindings,
+		STT:                       resolved.STT,
+		BuildDictationTranscriber: newDictationTranscriberFactory(resolved, userConfigPath, sttServerManager),
+		ShutdownDictationServer:   sttServerManager.Shutdown,
+		STTDownloadRoot:           sttDownloadRoot,
+		STTKeyStatus:              newSTTKeyStatus(resolved, userConfigPath),
+		SaveSTTKey:                newSaveSTTKey(userConfigPath),
 		Setup: tui.SetupOptions{
 			Visible:    setupVisible,
 			Required:   needsSetup,
@@ -1283,6 +1347,8 @@ 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

@@ -20,6 +20,7 @@ import (
 	"github.com/Gitlawb/zero/internal/tools"
 	"github.com/Gitlawb/zero/internal/tui"
 	"github.com/Gitlawb/zero/internal/update"
+	"github.com/Gitlawb/zero/internal/workspacetrust"
 	"github.com/Gitlawb/zero/internal/zeroruntime"
 )
 
@@ -158,6 +159,86 @@ func TestRunNoArgsLaunchesSetupTUIWithNilProviderWhenNoProviderConfigured(t *tes
 	assertAgentOptions(t, launchedOptions, 12, agent.PermissionModeAsk)
 }
 
+type fakeMCPRuntimeWithSkips struct {
+	skipped []mcp.SkippedServer
+}
+
+func (r fakeMCPRuntimeWithSkips) Close() error { return nil }
+
+func (r fakeMCPRuntimeWithSkips) Skipped() []mcp.SkippedServer { return r.skipped }
+
+func TestTUIStartupSuppressesWarningForUnconfiguredDefaultServer(t *testing.T) {
+	var stdout bytes.Buffer
+	var stderr bytes.Buffer
+	cwd := t.TempDir()
+	setCLIUserConfigRoot(t)
+
+	exitCode := runWithDeps([]string{}, &stdout, &stderr, appDeps{
+		getwd: func() (string, error) {
+			return cwd, nil
+		},
+		resolveConfig: func(workspaceRoot string, overrides config.Overrides) (config.ResolvedConfig, error) {
+			return config.ResolvedConfig{MaxTurns: 8}, nil
+		},
+		resolveMCPConfig: func(workspaceRoot string, excludeProject bool) (config.MCPConfig, error) {
+			return config.MCPConfig{Servers: map[string]config.MCPServerConfig{
+				"firecrawl": config.DefaultMCPServers()["firecrawl"],
+			}}, nil
+		},
+		registerMCPTools: func(context.Context, *tools.Registry, config.MCPConfig, mcp.RegisterOptions) (mcpToolRuntime, error) {
+			return fakeMCPRuntimeWithSkips{skipped: []mcp.SkippedServer{
+				{Name: "firecrawl", Err: errors.New("returned HTTP 401"), UnconfiguredDefault: true},
+			}}, nil
+		},
+		runTUI: func(ctx context.Context, options tui.Options) int {
+			return 0
+		},
+	})
+
+	if exitCode != 0 {
+		t.Fatalf("exitCode = %d stderr=%s", exitCode, stderr.String())
+	}
+	if strings.Contains(stderr.String(), "firecrawl") {
+		t.Fatalf("stderr = %q, want no warning for an unconfigured default server", stderr.String())
+	}
+}
+
+func TestTUIStartupWarnsForUserConfiguredServerSkip(t *testing.T) {
+	var stdout bytes.Buffer
+	var stderr bytes.Buffer
+	cwd := t.TempDir()
+	setCLIUserConfigRoot(t)
+
+	exitCode := runWithDeps([]string{}, &stdout, &stderr, appDeps{
+		getwd: func() (string, error) {
+			return cwd, nil
+		},
+		resolveConfig: func(workspaceRoot string, overrides config.Overrides) (config.ResolvedConfig, error) {
+			return config.ResolvedConfig{MaxTurns: 8}, nil
+		},
+		resolveMCPConfig: func(workspaceRoot string, excludeProject bool) (config.MCPConfig, error) {
+			return config.MCPConfig{Servers: map[string]config.MCPServerConfig{
+				"custom": {Type: "stdio", Command: "custom-mcp"},
+			}}, nil
+		},
+		registerMCPTools: func(context.Context, *tools.Registry, config.MCPConfig, mcp.RegisterOptions) (mcpToolRuntime, error) {
+			return fakeMCPRuntimeWithSkips{skipped: []mcp.SkippedServer{
+				{Name: "custom", Err: errors.New("connect failed"), UnconfiguredDefault: false},
+			}}, nil
+		},
+		runTUI: func(ctx context.Context, options tui.Options) int {
+			return 0
+		},
+	})
+
+	if exitCode != 0 {
+		t.Fatalf("exitCode = %d stderr=%s", exitCode, stderr.String())
+	}
+	if !strings.Contains(stderr.String(), "warning: MCP server custom unavailable") {
+		t.Fatalf("stderr = %q, want a warning for a user-configured server that failed", stderr.String())
+	}
+}
+
 func TestRunNoArgsEntersSetupWhenResolveReportsNoActiveProvider(t *testing.T) {
 	var stdout bytes.Buffer
 	var stderr bytes.Buffer
@@ -324,7 +405,7 @@ func TestRunNoArgsLaunchesTUIWithMCPState(t *testing.T) {
 		resolveConfig: func(workspaceRoot string, overrides config.Overrides) (config.ResolvedConfig, error) {
 			return config.ResolvedConfig{MaxTurns: 8}, nil
 		},
-		resolveMCPConfig: func(workspaceRoot string) (config.MCPConfig, error) {
+		resolveMCPConfig: func(workspaceRoot string, _ bool) (config.MCPConfig, error) {
 			if workspaceRoot != cwd {
 				t.Fatalf("workspaceRoot = %q, want %q", workspaceRoot, cwd)
 			}
@@ -399,7 +480,7 @@ func TestTUIMCPCommandUsesLastGoodConfigOnRefreshError(t *testing.T) {
 		resolveConfig: func(workspaceRoot string, overrides config.Overrides) (config.ResolvedConfig, error) {
 			return config.ResolvedConfig{MaxTurns: 8}, nil
 		},
-		resolveMCPConfig: func(workspaceRoot string) (config.MCPConfig, error) {
+		resolveMCPConfig: func(workspaceRoot string, _ bool) (config.MCPConfig, error) {
 			resolveCalls++
 			switch resolveCalls {
 			case 1:
@@ -466,7 +547,7 @@ func TestRunNoArgsClosesPartialMCPRuntimeWhenRegistrationFails(t *testing.T) {
 		resolveConfig: func(workspaceRoot string, overrides config.Overrides) (config.ResolvedConfig, error) {
 			return config.ResolvedConfig{MaxTurns: 8}, nil
 		},
-		resolveMCPConfig: func(workspaceRoot string) (config.MCPConfig, error) {
+		resolveMCPConfig: func(workspaceRoot string, _ bool) (config.MCPConfig, error) {
 			return config.MCPConfig{Servers: map[string]config.MCPServerConfig{
 				"docs": {Type: "stdio", Command: "docs-mcp"},
 			}}, nil
@@ -521,7 +602,7 @@ func TestRunNoArgsSoftFailsMCPTokenStoreInit(t *testing.T) {
 		resolveConfig: func(workspaceRoot string, overrides config.Overrides) (config.ResolvedConfig, error) {
 			return config.ResolvedConfig{MaxTurns: 8}, nil
 		},
-		resolveMCPConfig: func(workspaceRoot string) (config.MCPConfig, error) {
+		resolveMCPConfig: func(workspaceRoot string, _ bool) (config.MCPConfig, error) {
 			return config.MCPConfig{}, nil
 		},
 		newMCPStore: func() (*mcp.PermissionStore, error) {
@@ -660,6 +741,52 @@ func TestRunNoArgsLaunchesTUIInAskPermissionMode(t *testing.T) {
 	}
 }
 
+// TestRunInteractiveSurfacesMCPTrustNotice executes the interactive (TUI) path's MCP
+// trust notice (app.go), the third notice site alongside exec and spec-draft. An
+// untrusted repo whose only project config is MCP must print the notice before runTUI;
+// trusting it silences it. runTUI is stubbed so nothing renders; resolveMCPConfig
+// returns no servers so nothing spawns -- the notice depends only on the trust verdict
+// and the real ./.zero/config.json.
+func TestRunInteractiveSurfacesMCPTrustNotice(t *testing.T) {
+	setTrustConfigRoot(t)
+	repo := t.TempDir()
+	if err := os.MkdirAll(filepath.Join(repo, ".zero"), 0o700); err != nil {
+		t.Fatal(err)
+	}
+	body := `{"mcp":{"servers":{"proj":{"type":"stdio","command":"proj-cmd"}}}}`
+	if err := os.WriteFile(filepath.Join(repo, ".zero", "config.json"), []byte(body), 0o600); err != nil {
+		t.Fatal(err)
+	}
+
+	run := func() string {
+		var stdout, stderr bytes.Buffer
+		code := runWithDeps([]string{}, &stdout, &stderr, appDeps{
+			getwd: func() (string, error) { return repo, nil },
+			resolveConfig: func(string, config.Overrides) (config.ResolvedConfig, error) {
+				return config.ResolvedConfig{MaxTurns: 3}, nil
+			},
+			resolveMCPConfig: func(string, bool) (config.MCPConfig, error) { return config.MCPConfig{}, nil },
+			runTUI:           func(context.Context, tui.Options) int { return 0 },
+		})
+		if code != 0 {
+			t.Fatalf("interactive launch exit = %d, stderr=%q", code, stderr.String())
+		}
+		return stderr.String()
+	}
+
+	untrusted := run()
+	if !strings.Contains(untrusted, "MCP servers") || !strings.Contains(untrusted, "zero trust") {
+		t.Fatalf("untrusted interactive launch must surface the MCP trust notice, stderr=%q", untrusted)
+	}
+
+	if err := workspacetrust.Trust(repo); err != nil {
+		t.Fatal(err)
+	}
+	if trusted := run(); strings.Contains(trusted, "ignoring project") {
+		t.Fatalf("trusted interactive launch must not emit a trust notice, stderr=%q", trusted)
+	}
+}
+
 func TestRunSkipPermissionsUnsafeLaunchesTUIInUnsafeMode(t *testing.T) {
 	var stdout bytes.Buffer
 	var stderr bytes.Buffer

internal/cli/auth.go

@@ -6,15 +6,52 @@ import (
 	"io"
 	"net/http"
 	"os"
+	"path/filepath"
 	"strings"
 	"time"
 
 	"github.com/Gitlawb/zero/internal/config"
 	"github.com/Gitlawb/zero/internal/oauth"
+	"github.com/Gitlawb/zero/internal/providercatalog"
 	"github.com/Gitlawb/zero/internal/provideroauth"
 	"github.com/Gitlawb/zero/internal/redaction"
 )
 
+// ensureLoginProviderProfile makes a freshly stored OAuth login visible as a
+// provider: without a profile in config.json the login shows up nowhere — not in
+// `zero providers list`, not in the TUI picker — and `zero providers use <name>`
+// fails "not found", so the user's working active provider looks broken while
+// the new login looks lost. Returns user-facing guidance lines ("" when the
+// login has no catalog entry to scaffold from). Best-effort by design: the
+// token is already stored, so a profile-write failure degrades to a warning
+// line instead of failing a login that succeeded.
+func ensureLoginProviderProfile(deps appDeps, provider string) string {
+	if _, ok := providercatalog.Get(provider); !ok {
+		// Custom OAuth server without a catalog entry — no endpoint/model defaults
+		// to scaffold a profile from; the user wires their own profile.
+		return ""
+	}
+	configPath, err := deps.userConfigPath()
+	if err != nil {
+		return "warning: login saved, but no provider profile was written: " + err.Error()
+	}
+	ensured, err := config.EnsureCatalogProvider(configPath, provider)
+	if err != nil {
+		return "warning: login saved, but no provider profile was written: " + err.Error()
+	}
+	active := strings.EqualFold(strings.TrimSpace(ensured.Active), strings.TrimSpace(ensured.Name))
+	switch {
+	case ensured.Created && active:
+		return fmt.Sprintf("Added provider %q to your config and set it active.", ensured.Name)
+	case ensured.Created:
+		return fmt.Sprintf("Added provider %q to your config; the active provider is still %q.\nSwitch with: zero providers use %s", ensured.Name, ensured.Active, ensured.Name)
+	case active:
+		return fmt.Sprintf("Provider %q is configured and active.", ensured.Name)
+	default:
+		return fmt.Sprintf("Provider %q is already configured.\nSwitch with: zero providers use %s", ensured.Name, ensured.Name)
+	}
+}
+
 // runAuth dispatches `zero auth <command>` for provider OAuth login. It is
 // additive and independent of `zero mcp oauth` (MCP server auth), which is
 // unchanged.
@@ -46,11 +83,11 @@ func runAuth(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) in
 	}
 }
 
-// runAuthOpenRouter runs OpenRouter's browser PKCE login and prints the freshly
-// minted API key. Unlike `auth login` (which stores an OAuth bearer token),
-// OpenRouter's flow mints a normal API key; the setup wizard saves it to a
-// provider profile, while this command prints it for manual configuration.
-func runAuthOpenRouter(args []string, stdout io.Writer, stderr io.Writer, _ appDeps) int {
+// runAuthOpenRouter runs OpenRouter's browser PKCE login and saves the freshly
+// minted API key into the user's provider credential store. Unlike `auth login`
+// (which stores an OAuth bearer token), OpenRouter's flow mints a normal API key;
+// persist it immediately so the provider is usable after the command completes.
+func runAuthOpenRouter(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) int {
 	for _, a := range args {
 		if a == "-h" || a == "--help" || a == "help" {
 			_ = writeAuthHelp(stdout)
@@ -62,7 +99,7 @@ func runAuthOpenRouter(args []string, stdout io.Writer, stderr io.Writer, _ appD
 	if len(args) > 0 {
 		return writeExecUsageError(stderr, fmt.Sprintf("zero auth openrouter takes no arguments (got %q)", args[0]))
 	}
-	key, err := provideroauth.OpenRouterLogin(context.Background(), provideroauth.OpenRouterOptions{
+	key, err := deps.openRouterLogin(context.Background(), provideroauth.OpenRouterOptions{
 		Out:        stdout,
 		HTTPClient: &http.Client{Timeout: 30 * time.Second},
 		// ZERO_OPENROUTER_BASE_URL overrides the endpoint (self-hosted gateway or tests).
@@ -71,12 +108,59 @@ func runAuthOpenRouter(args []string, stdout io.Writer, stderr io.Writer, _ appD
 	if err != nil {
 		return writeAppError(stderr, redaction.ErrorMessage(err, redaction.Options{}), exitCrash)
 	}
-	if _, err := fmt.Fprintf(stdout, "\nOpenRouter login complete — new API key minted.\nUse it with zero, e.g.:\n  export OPENROUTER_API_KEY=%s\n(or add it to a provider profile with catalogID \"openrouter\").\n", key); err != nil {
+	key = strings.TrimSpace(key)
+	line, err := saveOpenRouterProviderKey(deps, key)
+	if err != nil {
+		if _, writeErr := fmt.Fprintf(stdout, "\nOpenRouter login complete — new API key minted, but Zero could not save it: %s\nUse it manually, e.g.:\n  export OPENROUTER_API_KEY=%s\n", err, key); writeErr != nil {
+			return exitCrash
+		}
+		return exitSuccess
+	}
+	if _, err := fmt.Fprintf(stdout, "\nOpenRouter login complete — new API key saved.\n%s\n", line); err != nil {
 		return exitCrash
 	}
 	return exitSuccess
 }
 
+func saveOpenRouterProviderKey(deps appDeps, key string) (string, error) {
+	key = strings.TrimSpace(key)
+	if key == "" {
+		return "", fmt.Errorf("OpenRouter returned an empty API key")
+	}
+	configPath, err := deps.userConfigPath()
+	if err != nil {
+		return "", err
+	}
+	ensured, err := config.EnsureCatalogProvider(configPath, "openrouter")
+	if err != nil {
+		return "", err
+	}
+	store, err := config.ProviderKeyStoreAt(filepath.Dir(configPath))
+	if err != nil {
+		return "", err
+	}
+	if err := store.Set(ensured.Name, key); err != nil {
+		return "", err
+	}
+	if err := config.MarkProviderAPIKeyStored(configPath, ensured.Name); err != nil {
+		// Best-effort rollback: don't leave the key orphaned in the credential
+		// store while config.json still says it isn't there.
+		_, _ = store.Delete(ensured.Name)
+		return "", err
+	}
+	active := strings.EqualFold(strings.TrimSpace(ensured.Active), strings.TrimSpace(ensured.Name))
+	switch {
+	case ensured.Created && active:
+		return fmt.Sprintf("Added provider %q to your config and set it active.", ensured.Name), nil
+	case ensured.Created:
+		return fmt.Sprintf("Added provider %q to your config; the active provider is still %q.\nSwitch with: zero providers use %s", ensured.Name, ensured.Active, ensured.Name), nil
+	case active:
+		return fmt.Sprintf("Provider %q is configured, active, and ready to use.", ensured.Name), nil
+	default:
+		return fmt.Sprintf("Provider %q is configured with the new key.\nSwitch with: zero providers use %s", ensured.Name, ensured.Name), nil
+	}
+}
+
 // runAuthChatGPT runs the ChatGPT (Codex) browser PKCE login, persists the
 // bearer + chatgpt-account-id claim under the "chatgpt" provider key, and
 // prints a status block. The bearer routes to chatgpt.com/backend-api/codex
@@ -138,8 +222,10 @@ func runAuthChatGPT(args []string, stdout io.Writer, stderr io.Writer, deps appD
 	if _, err := fmt.Fprint(stdout, statuses); err != nil {
 		return exitCrash
 	}
-	if _, err := fmt.Fprint(stdout, "\nUse it with zero, e.g.:\n  zero --provider chatgpt --model gpt-5.5\n"); err != nil {
-		return exitCrash
+	if line := ensureLoginProviderProfile(deps, "chatgpt"); line != "" {
+		if _, err := fmt.Fprintf(stdout, "\n%s\n", line); err != nil {
+			return exitCrash
+		}
 	}
 	return exitSuccess
 }
@@ -331,6 +417,11 @@ func runAuthLogin(args []string, stdout io.Writer, stderr io.Writer, deps appDep
 	if _, err := fmt.Fprintf(stdout, "Logged in to %s.\n%s\n", parsed.positional[0], oauth.FormatStatuses([]oauth.Status{status})); err != nil {
 		return exitCrash
 	}
+	if line := ensureLoginProviderProfile(deps, provider); line != "" {
+		if _, err := fmt.Fprintln(stdout, line); err != nil {
+			return exitCrash
+		}
+	}
 	return exitSuccess
 }
 

internal/cli/auth_propagation_test.go

@@ -2,12 +2,14 @@ package cli
 
 import (
 	"bytes"
+	"context"
 	"errors"
 	"os"
 	"path/filepath"
 	"testing"
 
 	"github.com/Gitlawb/zero/internal/config"
+	"github.com/Gitlawb/zero/internal/providerhealth"
 	"github.com/Gitlawb/zero/internal/zeroruntime"
 )
 
@@ -72,6 +74,32 @@ func TestFillAppDepsWrapsNewProviderWithStoredKey(t *testing.T) {
 	}
 }
 
+// fillAppDeps also wraps probeProviderHealth, the code path zero providers
+// check --connectivity, zero doctor --connectivity, and the TUI doctor panel
+// all go through. Passing the raw resolved profile previously sent an
+// unauthenticated (keyless) health probe for apiKeyStored profiles, same
+// regression class as TestFillAppDepsWrapsNewProviderWithStoredKey but for
+// health checks instead of the runtime provider build.
+func TestFillAppDepsWrapsProbeProviderHealthWithStoredKey(t *testing.T) {
+	configPath := seedStoredProviderKey(t, "echo", "sk-stored-health")
+
+	var captured config.ProviderProfile
+	deps := fillAppDeps(appDeps{
+		userConfigPath: func() (string, error) { return configPath, nil },
+		probeProviderHealth: func(_ context.Context, options providerhealth.Options) providerhealth.Result {
+			captured = options.Profile
+			return providerhealth.Result{Status: providerhealth.StatusPass}
+		},
+	})
+
+	deps.probeProviderHealth(context.Background(), providerhealth.Options{
+		Profile: config.ProviderProfile{Name: "echo", APIKeyStored: true},
+	})
+	if captured.APIKey != "sk-stored-health" {
+		t.Fatalf("wrapped probeProviderHealth ran with APIKey = %q, want the stored key (keyless health-check regression)", captured.APIKey)
+	}
+}
+
 // buildProvider (the TUI/exec STARTUP construction site) must export
 // ZERO_PROVIDER so children spawned at any point in the run are pinned to the
 // parent's provider from launch — not only after an in-session switch. Without

internal/cli/auth_test.go

@@ -2,18 +2,26 @@ package cli
 
 import (
 	"bytes"
+	"context"
+	"encoding/json"
+	"os"
 	"path/filepath"
 	"strings"
 	"testing"
 
+	"github.com/Gitlawb/zero/internal/config"
 	"github.com/Gitlawb/zero/internal/oauth"
+	"github.com/Gitlawb/zero/internal/provideroauth"
 )
 
-// withAuthStore points the provider OAuth store at a temp file for the test.
+// withAuthStore points the provider OAuth store at a temp file for the test,
+// pinning the file backend so an inherited ZERO_OAUTH_STORAGE=keyring can't
+// ignore the temp path and hit the OS keychain.
 func withAuthStore(t *testing.T) string {
 	t.Helper()
 	path := filepath.Join(t.TempDir(), "oauth-tokens.json")
 	t.Setenv("ZERO_OAUTH_TOKENS_PATH", path)
+	t.Setenv("ZERO_OAUTH_STORAGE", "file")
 	return path
 }
 
@@ -145,6 +153,44 @@ func TestRunAuthOpenRouterRejectsArgs(t *testing.T) {
 	}
 }
 
+func TestRunAuthOpenRouterSavesMintedKey(t *testing.T) {
+	configPath := filepath.Join(t.TempDir(), "config.json")
+	var stdout, stderr bytes.Buffer
+
+	code := runWithDeps([]string{"auth", "openrouter"}, &stdout, &stderr, appDeps{
+		userConfigPath: func() (string, error) { return configPath, nil },
+		openRouterLogin: func(context.Context, provideroauth.OpenRouterOptions) (string, error) {
+			return "sk-openrouter-test", nil
+		},
+	})
+
+	if code != exitSuccess {
+		t.Fatalf("exit = %d, stderr = %q", code, stderr.String())
+	}
+	if !strings.Contains(stdout.String(), "new API key saved") {
+		t.Fatalf("expected saved-key confirmation, got %q", stdout.String())
+	}
+	cfg := readCLIConfigFixture(t, configPath)
+	if cfg.ActiveProvider != "openrouter" || len(cfg.Providers) != 1 {
+		t.Fatalf("config = %#v", cfg)
+	}
+	profile := cfg.Providers[0]
+	if profile.Name != "openrouter" || profile.CatalogID != "openrouter" || !profile.APIKeyStored || profile.APIKey != "" || profile.APIKeyEnv != "" {
+		t.Fatalf("provider not stored-key sanitized: %#v", profile)
+	}
+	store, err := config.ProviderKeyStoreAt(filepath.Dir(configPath))
+	if err != nil {
+		t.Fatal(err)
+	}
+	key, ok, err := store.Get("openrouter")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !ok || key != "sk-openrouter-test" {
+		t.Fatalf("stored key = %q, %v", key, ok)
+	}
+}
+
 func TestRunAuthHelp(t *testing.T) {
 	var stdout, stderr bytes.Buffer
 	if code := runWithDeps([]string{"auth", "--help"}, &stdout, &stderr, appDeps{}); code != exitSuccess {
@@ -197,3 +243,81 @@ func TestRunAuthLoginChatGPTRejectsScope(t *testing.T) {
 		t.Fatalf("stderr = %q, want the ChatGPT-specific --scope rejection", stderr.String())
 	}
 }
+
+func TestEnsureLoginProviderProfileAddsProviderWithoutStealingActive(t *testing.T) {
+	configPath := filepath.Join(t.TempDir(), "config.json")
+	seed := `{"activeProvider":"opengateway","providers":[{"name":"opengateway","provider_kind":"openai-compatible","baseURL":"https://gateway.example.com/v1","apiKeyStored":true,"model":"some-model"}]}`
+	if err := os.WriteFile(configPath, []byte(seed), 0o600); err != nil {
+		t.Fatalf("seed config: %v", err)
+	}
+	deps := appDeps{userConfigPath: func() (string, error) { return configPath, nil }}
+
+	line := ensureLoginProviderProfile(deps, "chatgpt")
+	if !strings.Contains(line, `Added provider "chatgpt"`) {
+		t.Fatalf("expected added-provider guidance, got %q", line)
+	}
+	if !strings.Contains(line, "zero providers use chatgpt") {
+		t.Fatalf("expected switch hint, got %q", line)
+	}
+
+	cfg := readCLIConfigFixture(t, configPath)
+	if cfg.ActiveProvider != "opengateway" {
+		t.Fatalf("active provider changed to %q", cfg.ActiveProvider)
+	}
+	names := make([]string, 0, len(cfg.Providers))
+	for _, provider := range cfg.Providers {
+		names = append(names, provider.Name)
+	}
+	if len(cfg.Providers) != 2 || cfg.Providers[1].CatalogID != "chatgpt" {
+		t.Fatalf("expected chatgpt profile appended, got %v", names)
+	}
+
+	// A second login must be a no-op with switch guidance, not a duplicate.
+	line = ensureLoginProviderProfile(deps, "chatgpt")
+	if !strings.Contains(line, "already configured") {
+		t.Fatalf("expected already-configured guidance, got %q", line)
+	}
+	cfg = readCLIConfigFixture(t, configPath)
+	if len(cfg.Providers) != 2 {
+		t.Fatalf("repeat login duplicated the profile: %d providers", len(cfg.Providers))
+	}
+}
+
+func TestEnsureLoginProviderProfileActivatesOnFreshConfig(t *testing.T) {
+	configPath := filepath.Join(t.TempDir(), "config.json")
+	deps := appDeps{userConfigPath: func() (string, error) { return configPath, nil }}
+
+	line := ensureLoginProviderProfile(deps, "chatgpt")
+	if !strings.Contains(line, "set it active") {
+		t.Fatalf("fresh config should adopt the login as active, got %q", line)
+	}
+	cfg := readCLIConfigFixture(t, configPath)
+	if cfg.ActiveProvider != "chatgpt" {
+		t.Fatalf("active provider = %q, want chatgpt", cfg.ActiveProvider)
+	}
+}
+
+func TestEnsureLoginProviderProfileSkipsNonCatalogProviders(t *testing.T) {
+	configPath := filepath.Join(t.TempDir(), "config.json")
+	deps := appDeps{userConfigPath: func() (string, error) { return configPath, nil }}
+
+	if line := ensureLoginProviderProfile(deps, "my-custom-oauth-server"); line != "" {
+		t.Fatalf("custom OAuth server must not scaffold a profile, got %q", line)
+	}
+	if _, err := os.Stat(configPath); err == nil {
+		t.Fatalf("config must not be created for a non-catalog login")
+	}
+}
+
+func readCLIConfigFixture(t *testing.T, path string) config.FileConfig {
+	t.Helper()
+	data, err := os.ReadFile(path)
+	if err != nil {
+		t.Fatalf("read config: %v", err)
+	}
+	var cfg config.FileConfig
+	if err := json.Unmarshal(data, &cfg); err != nil {
+		t.Fatalf("decode config: %v", err)
+	}
+	return cfg
+}

internal/cli/backends.go

@@ -125,7 +125,9 @@ func backendLifecycleSnapshot(deps appDeps) (zerocommands.BackendLifecycleSnapsh
 		return zerocommands.BackendLifecycleSnapshot{}, fmt.Errorf("failed to resolve workspace: %w", err)
 	}
 
-	cfg, err := deps.resolveMCPConfig(cwd)
+	// Reporting/enumeration only, never spawns a server, so it is left ungated
+	// (excludeProject=false) to mirror the doctor/status hooks and plugins reports.
+	cfg, err := deps.resolveMCPConfig(cwd, false)
 	if err != nil {
 		return zerocommands.BackendLifecycleSnapshot{}, err
 	}
@@ -152,7 +154,9 @@ func backendDoctorReport(deps appDeps) (zerocommands.BackendDoctorReport, error)
 		return zerocommands.BackendDoctorReport{}, fmt.Errorf("failed to resolve workspace: %w", err)
 	}
 
-	cfg, err := deps.resolveMCPConfig(cwd)
+	// Reporting/enumeration only, never spawns a server, so it is left ungated
+	// (excludeProject=false) to mirror the doctor/status hooks and plugins reports.
+	cfg, err := deps.resolveMCPConfig(cwd, false)
 	if err != nil {
 		return zerocommands.BackendDoctorReport{}, err
 	}

internal/cli/backends_test.go

@@ -26,7 +26,7 @@ func TestRunBackendsJSONUsesLifecycleSnapshotWithoutConnectingMCP(t *testing.T)
 	secret := "sk-proj-" + strings.Repeat("a", 24)
 	deps := appDeps{
 		getwd: func() (string, error) { return cwd, nil },
-		resolveMCPConfig: func(workspaceRoot string) (config.MCPConfig, error) {
+		resolveMCPConfig: func(workspaceRoot string, _ bool) (config.MCPConfig, error) {
 			if workspaceRoot != cwd {
 				t.Fatalf("workspaceRoot = %q, want %q", workspaceRoot, cwd)
 			}
@@ -123,7 +123,7 @@ func TestRunBackendsJSONUsesLifecycleSnapshotWithoutConnectingMCP(t *testing.T)
 func TestRunBackendsTextAndHelp(t *testing.T) {
 	deps := appDeps{
 		getwd: func() (string, error) { return t.TempDir(), nil },
-		resolveMCPConfig: func(string) (config.MCPConfig, error) {
+		resolveMCPConfig: func(string, bool) (config.MCPConfig, error) {
 			return config.MCPConfig{}, nil
 		},
 		loadHooks: func(hooks.LoadOptions) (hooks.LoadResult, error) {
@@ -164,7 +164,7 @@ func TestRunBackendsDoctorJSONAndTextWithoutConnectingMCP(t *testing.T) {
 	secret := "sk-proj-" + strings.Repeat("b", 24)
 	deps := appDeps{
 		getwd: func() (string, error) { return cwd, nil },
-		resolveMCPConfig: func(workspaceRoot string) (config.MCPConfig, error) {
+		resolveMCPConfig: func(workspaceRoot string, _ bool) (config.MCPConfig, error) {
 			if workspaceRoot != cwd {
 				t.Fatalf("workspaceRoot = %q, want %q", workspaceRoot, cwd)
 			}
@@ -301,7 +301,7 @@ func TestRunBackendsDoctorDoesNotConnectOrExecuteConfiguredBackends(t *testing.T
 
 	deps := appDeps{
 		getwd: func() (string, error) { return cwd, nil },
-		resolveMCPConfig: func(string) (config.MCPConfig, error) {
+		resolveMCPConfig: func(string, bool) (config.MCPConfig, error) {
 			return config.MCPConfig{Servers: map[string]config.MCPServerConfig{
 				"remote": {Type: "http", URL: server.URL + "/mcp"},
 			}}, nil

internal/cli/command_center.go

@@ -75,6 +75,12 @@ func runProviders(args []string, stdout io.Writer, stderr io.Writer, deps appDep
 	if command == "use" {
 		return runProvidersUse(args, stdout, stderr, deps)
 	}
+	if command == "remove" || command == "rm" {
+		return runProvidersRemove(args, stdout, stderr, deps)
+	}
+	if command == "rename" {
+		return runProvidersRename(args, stdout, stderr, deps)
+	}
 	if command == "setup" {
 		return runProvidersSetup(args, stdout, stderr, deps)
 	}
@@ -246,6 +252,20 @@ func parseCommandCenterArgs(args []string, allowModelFilters bool, allowProvider
 
 func summarizeConfig(resolved config.ResolvedConfig) configSummary {
 	summary := zerocommands.ConfigSnapshotFromResolved(resolved)
+	// Mark keyless profiles that authenticate via a stored OAuth login (e.g.
+	// ChatGPT) so the list shows "oauth login" instead of "api key: not set" —
+	// the same candidate matching the runtime resolver uses.
+	logins := oauthLoggedInProviders()
+	if len(logins) > 0 {
+		for index := range summary.Providers {
+			for _, profile := range resolved.Providers {
+				if profile.Name == summary.Providers[index].Name {
+					summary.Providers[index].OAuthLogin = providerHasOAuthLogin(profile, logins)
+					break
+				}
+			}
+		}
+	}
 	sort.SliceStable(summary.Providers, func(i int, j int) bool {
 		if summary.Providers[i].Active != summary.Providers[j].Active {
 			return summary.Providers[i].Active
@@ -323,7 +343,7 @@ func formatProviderSummaries(command string, providers []providerSummary) string
 				"model: "+displayCLIValue(provider.Model, "none"),
 				"api model: "+displayCLIValue(provider.APIModel, "unknown"),
 				"base url: "+displayCLIValue(provider.BaseURL, "default"),
-				"api key: "+apiKeyState(provider.APIKeySet),
+				"api key: "+providerCredentialState(provider),
 			)
 			if provider.Message != "" {
 				lines = append(lines, "status: "+provider.Status+" - "+provider.Message)
@@ -340,7 +360,7 @@ func formatProviderLine(provider providerSummary) string {
 	if provider.Active {
 		marker = "*"
 	}
-	line := fmt.Sprintf("%s %s [%s] model=%s apiModel=%s api key: %s", marker, displayCLIValue(provider.Name, "none"), displayCLIValue(provider.ProviderKind, "unknown"), displayCLIValue(provider.Model, "none"), displayCLIValue(provider.APIModel, "unknown"), apiKeyState(provider.APIKeySet))
+	line := fmt.Sprintf("%s %s [%s] model=%s apiModel=%s api key: %s", marker, displayCLIValue(provider.Name, "none"), displayCLIValue(provider.ProviderKind, "unknown"), displayCLIValue(provider.Model, "none"), displayCLIValue(provider.APIModel, "unknown"), providerCredentialState(provider))
 	if provider.Message != "" {
 		line += " (" + provider.Status + ": " + provider.Message + ")"
 	}
@@ -413,6 +433,15 @@ func apiKeyState(set bool) string {
 	return "not set"
 }
 
+// providerCredentialState renders the credential column: a configured key wins,
+// then a stored OAuth login, then "not set".
+func providerCredentialState(provider providerSummary) string {
+	if !provider.APIKeySet && provider.OAuthLogin {
+		return "oauth login"
+	}
+	return apiKeyState(provider.APIKeySet)
+}
+
 func displayCLIValue(value string, fallback string) string {
 	if strings.TrimSpace(value) == "" {
 		return fallback
@@ -466,6 +495,8 @@ func writeProvidersHelp(w io.Writer) error {
   zero providers add <catalog-id> [flags]
   zero providers check [name] [flags]
   zero providers use <name> [flags]
+  zero providers remove <name> [flags]
+  zero providers rename <old> <new> [flags]
   zero providers setup <catalog-id> [flags]
   zero providers detect [flags]
   zero providers models [name] [flags]

internal/cli/command_center_test.go

@@ -11,6 +11,7 @@ import (
 	"testing"
 
 	"github.com/Gitlawb/zero/internal/config"
+	"github.com/Gitlawb/zero/internal/oauth"
 	"github.com/Gitlawb/zero/internal/providerhealth"
 	"github.com/Gitlawb/zero/internal/zerocommands"
 	"github.com/Gitlawb/zero/internal/zeroruntime"
@@ -835,3 +836,57 @@ func (commandCenterProvider) StreamCompletion(context.Context, zeroruntime.Compl
 	close(ch)
 	return ch, nil
 }
+
+func TestProviderCredentialStateShowsOAuthLogin(t *testing.T) {
+	if got := providerCredentialState(providerSummary{APIKeySet: true}); got != "set" {
+		t.Fatalf("keyed provider = %q, want set", got)
+	}
+	if got := providerCredentialState(providerSummary{OAuthLogin: true}); got != "oauth login" {
+		t.Fatalf("token-login provider = %q, want oauth login", got)
+	}
+	if got := providerCredentialState(providerSummary{}); got != "not set" {
+		t.Fatalf("credential-less provider = %q, want not set", got)
+	}
+}
+
+// TestProvidersListMarksOAuthLoginProviders: a keyless profile whose credential
+// is a stored OAuth login (the shape `zero auth chatgpt` now writes) must render
+// as "oauth login", not as a broken "api key: not set" entry.
+func TestProvidersListMarksOAuthLoginProviders(t *testing.T) {
+	tokensPath := filepath.Join(t.TempDir(), "oauth-tokens.json")
+	t.Setenv("ZERO_OAUTH_TOKENS_PATH", tokensPath)
+	store, err := oauth.NewStore(oauth.StoreOptions{})
+	if err != nil {
+		t.Fatalf("oauth store: %v", err)
+	}
+	if err := store.Save(oauth.ProviderKey("chatgpt"), oauth.Token{AccessToken: "bearer-123"}); err != nil {
+		t.Fatalf("save token: %v", err)
+	}
+
+	resolved := config.ResolvedConfig{
+		ActiveProvider: "opengateway",
+		Providers: []config.ProviderProfile{
+			{Name: "opengateway", ProviderKind: config.ProviderKindOpenAICompatible, BaseURL: "https://gateway.example.com/v1", APIKeyStored: true, Model: "some-model"},
+			{Name: "chatgpt", CatalogID: "chatgpt", ProviderKind: config.ProviderKindOpenAICompatible, BaseURL: "https://chatgpt.com/backend-api/codex", Model: "gpt-5.5"},
+		},
+	}
+	summary := summarizeConfig(resolved)
+	byName := map[string]providerSummary{}
+	for _, provider := range summary.Providers {
+		byName[provider.Name] = provider
+	}
+	if !byName["opengateway"].APIKeySet {
+		t.Fatalf("stored-key provider must report a credential: %+v", byName["opengateway"])
+	}
+	if byName["opengateway"].OAuthLogin {
+		t.Fatalf("key-authed provider must not claim the OAuth login: %+v", byName["opengateway"])
+	}
+	if !byName["chatgpt"].OAuthLogin || byName["chatgpt"].APIKeySet {
+		t.Fatalf("chatgpt must be marked oauth-login and keyless: %+v", byName["chatgpt"])
+	}
+
+	rendered := formatProviderSummaries("list", summary.Providers)
+	if !strings.Contains(rendered, "oauth login") {
+		t.Fatalf("list should render the oauth login state, got:\n%s", rendered)
+	}
+}

internal/cli/deferred_wiring_test.go

@@ -142,7 +142,7 @@ func TestRunExecListToolsAdvertisesMCPToolsWithoutToolSearch(t *testing.T) {
 		resolveConfig: func(string, config.Overrides) (config.ResolvedConfig, error) {
 			return execResolvedConfig(), nil
 		},
-		resolveMCPConfig: func(string) (config.MCPConfig, error) {
+		resolveMCPConfig: func(string, bool) (config.MCPConfig, error) {
 			return config.MCPConfig{Servers: map[string]config.MCPServerConfig{
 				"docs": {Type: "stdio", Command: "docs-mcp"},
 			}}, nil
@@ -177,7 +177,7 @@ func TestRunExecListToolsHonorsJSONFormat(t *testing.T) {
 		resolveConfig: func(string, config.Overrides) (config.ResolvedConfig, error) {
 			return execResolvedConfig(), nil
 		},
-		resolveMCPConfig: func(string) (config.MCPConfig, error) { return config.MCPConfig{}, nil },
+		resolveMCPConfig: func(string, bool) (config.MCPConfig, error) { return config.MCPConfig{}, nil },
 		newMCPStore:      func() (*mcp.PermissionStore, error) { return nil, nil },
 		registerMCPTools: func(_ context.Context, _ *tools.Registry, _ config.MCPConfig, _ mcp.RegisterOptions) (mcpToolRuntime, error) {
 			return closeFunc(func() error { return nil }), nil
@@ -234,7 +234,7 @@ func TestTUIRunThreadsDeferThresholdAndRegistersToolSearch(t *testing.T) {
 				Tools: config.ToolsConfig{DeferThreshold: 2},
 			}, nil
 		},
-		resolveMCPConfig: func(string) (config.MCPConfig, error) {
+		resolveMCPConfig: func(string, bool) (config.MCPConfig, error) {
 			return config.MCPConfig{Servers: map[string]config.MCPServerConfig{
 				"docs": {Type: "stdio", Command: "docs-mcp"},
 			}}, nil

internal/cli/dictation.go

@@ -0,0 +1,248 @@
+package cli
+
+import (
+	"fmt"
+	neturl "net/url"
+	"os"
+	"path/filepath"
+	"strings"
+
+	"github.com/Gitlawb/zero/internal/config"
+	"github.com/Gitlawb/zero/internal/dictation"
+	"github.com/Gitlawb/zero/internal/tui"
+)
+
+// Default batch cloud models (§8). Overridable via stt.model.
+const (
+	defaultGroqSTTModel   = "whisper-large-v3-turbo"
+	defaultOpenAISTTModel = "whisper-1"
+	groqSTTBaseURL        = "https://api.groq.com/openai/v1"
+)
+
+// newDictationTranscriberFactory returns the closure the TUI calls to build a
+// transcriber when a recording starts. It lives here, not in the TUI, because it
+// resolves provider API keys (credstore + env + configured providers) — logic
+// that belongs with the rest of the CLI's credential handling.
+//
+// serverMgr is the shared, session-long sherpa-onnx streaming server; it is
+// spawned lazily on the first local-streaming recording and torn down on exit.
+//
+// preferStreaming asks for the streaming backend; the returned bool reports
+// whether streaming was actually built (false → the caller uses the batch
+// pipeline). A misconfigured streaming backend surfaces its own SetupError.
+func newDictationTranscriberFactory(resolved config.ResolvedConfig, userConfigPath string, serverMgr *dictation.ServerManager) func(config.STTConfig, bool) (tui.Transcriber, bool, error) {
+	return func(stt config.STTConfig, preferStreaming bool) (tui.Transcriber, bool, error) {
+		// stt is passed in (not captured) so a mid-session change — e.g. the F9
+		// auto-download writing localModelPath — takes effect on the next call.
+		if serverMgr != nil {
+			serverMgr.SetModelPath(stt.LocalModelPath)
+		}
+		if preferStreaming {
+			t, err := buildStreamingTranscriber(stt, resolved, userConfigPath, serverMgr)
+			switch {
+			case err == nil && t != nil:
+				return t, true, nil
+			case err != nil && stt.STTStreamProvider() != config.STTProviderLocal:
+				// An explicitly-chosen cloud streaming backend that failed to build
+				// (e.g. a missing key) should surface its own error, not silently
+				// fall back to batch — the user asked for that backend specifically.
+				return nil, true, err
+			}
+			// The default local streaming backend isn't set up: fall through to the
+			// batch pipeline so a cloud-batch provider (or a batch-only setup) still
+			// works instead of erroring on a streaming server the user never asked for.
+		}
+		t, err := buildBatchTranscriber(stt, resolved, userConfigPath)
+		return t, false, err
+	}
+}
+
+// buildStreamingTranscriber constructs the configured streaming backend, or
+// returns (nil, nil) when streaming is not applicable so the caller falls back
+// to batch.
+func buildStreamingTranscriber(stt config.STTConfig, resolved config.ResolvedConfig, userConfigPath string, serverMgr *dictation.ServerManager) (dictation.Transcriber, error) {
+	switch stt.STTStreamProvider() {
+	case config.STTProviderDeepgram:
+		key := resolveSTTKey(resolved, userConfigPath, dictation.StreamProviderDeepgram, "DEEPGRAM_API_KEY")
+		return dictation.NewDeepgramTranscriber(dictation.DeepgramConfig{
+			APIKey:   key,
+			Model:    stt.StreamModel,
+			Language: stt.Language,
+		})
+	case config.STTProviderOpenAI:
+		key := resolveSTTKey(resolved, userConfigPath, dictation.ProviderOpenAI, "OPENAI_API_KEY")
+		return dictation.NewOpenAIRealtimeTranscriber(dictation.OpenAIRealtimeConfig{
+			APIKey: key,
+			Model:  stt.StreamModel,
+		})
+	default: // local streaming via the warm sherpa-onnx server
+		// STTStreamProvider() defaults to "local" whenever streamProvider is unset —
+		// including when the user picked a CLOUD batch provider (Groq/OpenAI) and
+		// never touched streaming. Don't silently stream through the local sherpa
+		// server in that case: local streaming is only appropriate when it was
+		// explicitly requested (streamProvider: "local") OR the batch provider is
+		// itself local. Otherwise report "not applicable" so the factory falls back
+		// to the cloud batch pipeline the user actually chose.
+		localWanted := stt.StreamProvider == config.STTProviderLocal || stt.STTProvider() == config.STTProviderLocal
+		if !localWanted {
+			return nil, nil
+		}
+		// The local streaming transcriber validates the model only lazily (at
+		// connect time), so gate on the model path here — with none set, report
+		// "not applicable" (nil, nil) so the factory falls back to the batch
+		// pipeline instead of committing to a server that can't start.
+		if serverMgr == nil || strings.TrimSpace(stt.LocalModelPath) == "" {
+			return nil, nil
+		}
+		return dictation.NewLocalStreamingTranscriber(serverMgr), nil
+	}
+}
+
+// newDictationServerManager builds the warm sherpa-onnx streaming server manager
+// from config. Construction is side-effect free — the server spawns lazily on
+// the first local-streaming recording — so it is always safe to create.
+func newDictationServerManager(stt config.STTConfig) *dictation.ServerManager {
+	return dictation.NewServerManager(dictation.ServerConfig{
+		Binary:     stt.LocalServerBinary,
+		ModelPath:  stt.LocalModelPath,
+		Port:       stt.LocalServerPort,
+		NumThreads: stt.NumThreads,
+	})
+}
+
+// buildBatchTranscriber constructs the configured batch transcriber, resolving
+// credentials for the cloud providers.
+func buildBatchTranscriber(stt config.STTConfig, resolved config.ResolvedConfig, userConfigPath string) (dictation.Transcriber, error) {
+	switch stt.STTProvider() {
+	case config.STTProviderGroq:
+		key := resolveSTTKey(resolved, userConfigPath, dictation.ProviderGroq, "GROQ_API_KEY")
+		model := stt.Model
+		if model == "" {
+			model = defaultGroqSTTModel
+		}
+		return dictation.NewCloudTranscriber(dictation.CloudConfig{
+			Provider: dictation.ProviderGroq,
+			BaseURL:  groqSTTBaseURL,
+			APIKey:   key,
+			Model:    model,
+			Language: stt.Language,
+		})
+	case config.STTProviderOpenAI:
+		key := resolveSTTKey(resolved, userConfigPath, dictation.ProviderOpenAI, "OPENAI_API_KEY")
+		model := stt.Model
+		if model == "" {
+			model = defaultOpenAISTTModel
+		}
+		return dictation.NewCloudTranscriber(dictation.CloudConfig{
+			Provider: dictation.ProviderOpenAI,
+			BaseURL:  config.OpenAIBaseURL,
+			APIKey:   key,
+			Model:    model,
+			Language: stt.Language,
+		})
+	default: // local
+		return dictation.NewLocalTranscriber(dictation.LocalConfig{
+			Binary:     stt.LocalBinary,
+			ModelPath:  stt.LocalModelPath,
+			NumThreads: stt.NumThreads,
+		})
+	}
+}
+
+// sttEnvVar returns the conventional environment variable a cloud STT provider's
+// key is read from.
+func sttEnvVar(provider string) string {
+	switch provider {
+	case dictation.ProviderGroq:
+		return "GROQ_API_KEY"
+	case dictation.ProviderOpenAI: // == StreamProviderOpenAI ("openai")
+		return "OPENAI_API_KEY"
+	case dictation.StreamProviderDeepgram:
+		return "DEEPGRAM_API_KEY"
+	}
+	return ""
+}
+
+// newSTTKeyStatus returns a predicate reporting whether a cloud STT provider
+// already has a resolvable API key (configured provider, credstore, or env).
+func newSTTKeyStatus(resolved config.ResolvedConfig, userConfigPath string) func(string) bool {
+	return func(provider string) bool {
+		return resolveSTTKey(resolved, userConfigPath, provider, sttEnvVar(provider)) != ""
+	}
+}
+
+// newSaveSTTKey returns a function that stores a cloud STT provider's API key in
+// the encrypted credential store, so the inline prompt can persist it.
+func newSaveSTTKey(userConfigPath string) func(string, string) error {
+	return func(provider, key string) error {
+		if strings.TrimSpace(userConfigPath) == "" {
+			return fmt.Errorf("no config path to store the key")
+		}
+		store, err := config.ProviderKeyStoreAt(filepath.Dir(userConfigPath))
+		if err != nil {
+			return err
+		}
+		return store.Set(provider, strings.TrimSpace(key))
+	}
+}
+
+// resolveSTTKey finds an API key for an STT cloud provider, reusing Zero's
+// existing credentials (§2 — "zero new credential UI"). Precedence: a matching
+// configured provider's resolved key, then the encrypted credential store, then
+// the provider's conventional environment variable.
+func resolveSTTKey(resolved config.ResolvedConfig, userConfigPath, provider, envVar string) string {
+	if key := keyFromConfiguredProviders(resolved, provider); key != "" {
+		return key
+	}
+	if userConfigPath != "" {
+		if store, err := config.ProviderKeyStoreAt(filepath.Dir(userConfigPath)); err == nil {
+			if key, ok, _ := store.Get(provider); ok && strings.TrimSpace(key) != "" {
+				return strings.TrimSpace(key)
+			}
+		}
+	}
+	return strings.TrimSpace(os.Getenv(envVar))
+}
+
+// keyFromConfiguredProviders returns the resolved API key of a configured
+// provider that belongs to the given cloud provider (matched by catalog id,
+// provider name/kind, or base-URL host), or "" when none is configured.
+func keyFromConfiguredProviders(resolved config.ResolvedConfig, provider string) string {
+	provider = strings.ToLower(provider)
+	profiles := append([]config.ProviderProfile{resolved.Provider}, resolved.Providers...)
+	for _, p := range profiles {
+		if strings.TrimSpace(p.APIKey) == "" {
+			continue
+		}
+		if providerMatches(p, provider) {
+			return strings.TrimSpace(p.APIKey)
+		}
+	}
+	return ""
+}
+
+func providerMatches(p config.ProviderProfile, provider string) bool {
+	provider = strings.ToLower(provider)
+	// Match on stable identifiers (catalog id, provider name/kind, profile
+	// name) rather than raw BaseURL substring matching: a Groq profile's
+	// BaseURL (api.groq.com/openai/v1) contains "openai", so a substring
+	// check can hand the wrong key to an OpenAI STT lookup and fail auth.
+	for _, f := range []string{
+		strings.ToLower(p.CatalogID),
+		strings.ToLower(p.Provider),
+		strings.ToLower(string(p.ProviderKind)),
+		strings.ToLower(p.Name),
+	} {
+		if f != "" && strings.Contains(f, provider) {
+			return true
+		}
+	}
+	// BaseURL: compare the parsed host only, not a raw substring on the
+	// full URL (which lets "openai" match api.groq.com/openai/v1).
+	if u, err := neturl.Parse(p.BaseURL); err == nil && u.Host != "" {
+		if host := strings.ToLower(u.Host); strings.Contains(host, provider) {
+			return true
+		}
+	}
+	return false
+}

internal/cli/dictation_test.go

@@ -0,0 +1,96 @@
+package cli
+
+import (
+	"errors"
+	"testing"
+
+	"github.com/Gitlawb/zero/internal/config"
+	"github.com/Gitlawb/zero/internal/dictation"
+)
+
+func TestFactoryFallsBackToBatchWhenLocalStreamingNotConfigured(t *testing.T) {
+	t.Setenv("GROQ_API_KEY", "gsk-test")
+	resolved := config.ResolvedConfig{STT: config.STTConfig{
+		Provider: config.STTProviderGroq, // cloud batch
+		// StreamProvider unset → defaults to local, which has no model configured.
+	}}
+	factory := newDictationTranscriberFactory(resolved, "", newDictationServerManager(resolved.STT))
+
+	tr, streaming, err := factory(resolved.STT, true) // prefer streaming
+	if err != nil {
+		t.Fatalf("expected batch fallback, got error: %v", err)
+	}
+	if streaming {
+		t.Error("should have fallen back to batch (streaming=false)")
+	}
+	if tr == nil {
+		t.Error("expected a Groq batch transcriber")
+	}
+}
+
+func TestFactoryCloudProviderIgnoresLeftoverLocalModel(t *testing.T) {
+	// Regression: pick local once (LocalModelPath gets set), then switch to a
+	// cloud batch provider. streamProvider stays unset (defaults to "local"), so
+	// the streaming factory must NOT fire up the sherpa server on the leftover
+	// path — it should fall back to the cloud batch pipeline the user chose.
+	t.Setenv("GROQ_API_KEY", "gsk-test")
+	stt := config.STTConfig{
+		Provider:       config.STTProviderGroq, // cloud batch, chosen now
+		LocalModelPath: "/leftover/model/dir",  // set by an earlier local selection
+	}
+	resolved := config.ResolvedConfig{STT: stt}
+	factory := newDictationTranscriberFactory(resolved, "", newDictationServerManager(stt))
+
+	tr, streaming, err := factory(stt, true) // prefer streaming
+	if err != nil {
+		t.Fatalf("expected cloud batch fallback, got error: %v", err)
+	}
+	if streaming {
+		t.Error("a cloud provider must not stream through the local sherpa server")
+	}
+	if tr == nil {
+		t.Error("expected a Groq batch transcriber")
+	}
+}
+
+func TestFactorySurfacesExplicitCloudStreamingError(t *testing.T) {
+	// Deepgram chosen explicitly but no key resolvable → surface the setup error.
+	resolved := config.ResolvedConfig{STT: config.STTConfig{
+		StreamProvider: config.STTProviderDeepgram,
+	}}
+	factory := newDictationTranscriberFactory(resolved, "", nil)
+
+	_, streaming, err := factory(resolved.STT, true)
+	if err == nil {
+		t.Fatal("expected a setup error for missing Deepgram key")
+	}
+	if !streaming {
+		t.Error("an explicit cloud streaming choice should report streaming=true even on error")
+	}
+	var setupErr *dictation.SetupError
+	if !errors.As(err, &setupErr) {
+		t.Errorf("want *SetupError, got %v", err)
+	}
+}
+
+func TestFactoryLocalBatchSetupErrorWhenNoModel(t *testing.T) {
+	resolved := config.ResolvedConfig{STT: config.STTConfig{Provider: config.STTProviderLocal}}
+	factory := newDictationTranscriberFactory(resolved, "", newDictationServerManager(resolved.STT))
+
+	_, _, err := factory(resolved.STT, true)
+	var setupErr *dictation.SetupError
+	if !errors.As(err, &setupErr) {
+		t.Fatalf("want *SetupError guiding local model setup, got %v", err)
+	}
+}
+
+func TestResolveSTTKeyPrefersConfiguredProvider(t *testing.T) {
+	resolved := config.ResolvedConfig{
+		Providers: []config.ProviderProfile{
+			{Name: "groq", CatalogID: "groq", BaseURL: "https://api.groq.com/openai/v1", APIKey: "gsk-configured"},
+		},
+	}
+	if got := resolveSTTKey(resolved, "", dictation.ProviderGroq, "GROQ_API_KEY"); got != "gsk-configured" {
+		t.Errorf("key = %q, want the configured provider's key", got)
+	}
+}

internal/cli/exec.go

@@ -110,6 +110,16 @@ 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.
@@ -155,6 +165,10 @@ func runExec(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) in
 	if err != nil {
 		return writeExecFormatUsageError(stdout, stderr, options.outputFormat, err.Error())
 	}
+	// trustRoot is the ORIGINAL launch directory, captured before any --worktree
+	// reassignment below, so a worktree of a trusted repo inherits that repo's
+	// trust instead of being seen as a fresh untrusted path.
+	trustRoot := workspaceRoot
 	if options.worktree {
 		preparedWorktree, err := deps.prepareWorktree(context.Background(), worktrees.Options{
 			Cwd:     workspaceRoot,
@@ -201,7 +215,7 @@ func runExec(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) in
 	if options.useSpec {
 		permissionMode = agent.PermissionModeSpecDraft
 	}
-	mcpRuntime, err := registerMCPToolsForWorkspace(context.Background(), workspaceRoot, registry, deps, execMCPAutonomy(options))
+	mcpRuntime, mcpSkip, err := registerMCPToolsForWorkspace(context.Background(), workspaceRoot, registry, deps, execMCPAutonomy(options), trustRoot)
 	if err != nil {
 		return writeExecProviderError(stdout, stderr, options.outputFormat, "mcp_error", err.Error())
 	}
@@ -210,7 +224,7 @@ func runExec(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) in
 	// registry and collect their hooks + skill roots for the dispatcher and skill
 	// tool below. Done before --list-tools and filter validation so plugin tools
 	// are listable and filter-validatable; it fails OPEN (a bad plugin is skipped).
-	pluginActivation := activatePlugins(workspaceRoot, registry, deps, stderr)
+	pluginActivation := activatePlugins(workspaceRoot, registry, deps, stderr, trustRoot)
 	if options.useSpec {
 		specmode.RegisterDraftTools(registry, workspaceRoot, deps.now)
 	}
@@ -415,6 +429,8 @@ func runExec(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) in
 			stderr:             stderr,
 			deps:               deps,
 			workspaceRoot:      workspaceRoot,
+			trustRoot:          trustRoot,
+			mcpSkip:            mcpSkip,
 			registry:           registry,
 			modelRegistry:      modelRegistry,
 			resolved:           resolved,
@@ -510,6 +526,20 @@ func runExec(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) in
 	// lspShutdown tears down any language-server sessions either half spawned.
 	selfCorrector, fileDiagnostics, lspShutdown := newExecSelfCorrector(options.selfCorrect, workspaceRoot, options.autonomy)
 	defer lspShutdown()
+	// Kept as a named variable (rather than inlined into Options below) because
+	// the write tools also use it for conflict checks across a run.
+	fileTracker := tools.NewFileTracker()
+	scratchBaseline := scratchFileSnapshot(workspaceRoot)
+	emitScratchWarning := func() {
+		if notice := scratchFileWarning(workspaceRoot, scratchBaseline); notice != "" {
+			writer.warning(notice)
+		}
+	}
+	// Build the hooks dispatcher out of the struct literal so its trust skip report
+	// can be combined with the plugin activation's, and emit at most one notice when
+	// project hooks/plugins were dropped for an untrusted workspace.
+	hookDispatcher, hookSkip := newHookDispatcherWithExtra(workspaceRoot, pluginActivation.hooks, trustRoot)
+	emitTrustNotice(stderr, hookSkip, pluginActivation.trustSkip, mcpSkip)
 	result, err := agent.Run(runCtx, agentPrompt, provider, agent.Options{
 		MaxTurns:         resolved.MaxTurns,
 		ContextWindow:    resolveAgentContextWindow(runCtx, modelRegistry, resolved.Provider),
@@ -536,11 +566,13 @@ 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:             tools.NewFileTracker(),
-		Hooks:                   newHookDispatcherWithExtra(workspaceRoot, pluginActivation.hooks),
+		FileTracker:             fileTracker,
+		Hooks:                   hookDispatcher,
 		EnabledTools:            options.enabledTools,
 		DisabledTools:           options.disabledTools,
 		OnText:                  writer.text,
@@ -598,6 +630,7 @@ func runExec(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) in
 		return exitCrash
 	}
 	if err != nil {
+		emitScratchWarning()
 		// A Ctrl+C / SIGTERM cancellation is a clean shutdown, not a provider error.
 		if errors.Is(err, context.Canceled) || runCtx.Err() != nil {
 			sessionRecorder.append(sessions.EventError, map[string]any{"message": "interrupted"})
@@ -648,6 +681,10 @@ func runExec(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) in
 	if notice := result.TruncationNotice(); notice != "" {
 		writer.warning(notice)
 	}
+	// Surface new scratch/debug-like files left untracked in git when we finish —
+	// including files created by shell commands that bypass the file tools (issue
+	// #551). Normal new deliverable files stay quiet to avoid warning fatigue.
+	emitScratchWarning()
 	// A headless run the completion gate marked INCOMPLETE (no-tool-call stall,
 	// self-reported non-completion, or max-turns cutoff) must NOT be reported as a
 	// success: exit 4 AND a machine-readable terminal event saying so. Handled per

internal/cli/exec_parse.go

@@ -30,6 +30,8 @@ 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 == "--notify":
 			value, next, err := nextFlagValue(args, index, arg)
 			if err != nil {
@@ -402,6 +404,11 @@ 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."}
 	}

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_protocol_test.go

@@ -5,6 +5,7 @@ import (
 	"context"
 	"encoding/json"
 	"errors"
+	"os"
 	"path/filepath"
 	"slices"
 	"strings"
@@ -149,7 +150,7 @@ func TestRunExecListsMCPToolsWithoutProviderConstruction(t *testing.T) {
 			providerBuilt = true
 			return nil, errors.New("provider should not be constructed for --list-tools")
 		},
-		resolveMCPConfig: func(workspaceRoot string) (config.MCPConfig, error) {
+		resolveMCPConfig: func(workspaceRoot string, _ bool) (config.MCPConfig, error) {
 			if workspaceRoot != cwd {
 				t.Fatalf("workspaceRoot = %q, want %q", workspaceRoot, cwd)
 			}
@@ -192,7 +193,7 @@ func TestRunExecLogsMCPRuntimeCloseError(t *testing.T) {
 		getwd: func() (string, error) {
 			return cwd, nil
 		},
-		resolveMCPConfig: func(workspaceRoot string) (config.MCPConfig, error) {
+		resolveMCPConfig: func(workspaceRoot string, _ bool) (config.MCPConfig, error) {
 			if workspaceRoot != cwd {
 				t.Fatalf("workspaceRoot = %q, want %q", workspaceRoot, cwd)
 			}
@@ -712,6 +713,48 @@ func TestRunExecJSONInterruptedEmitsTerminalEvents(t *testing.T) {
 	}
 }
 
+func TestRunExecWarnsAboutScratchFileWhenProviderErrors(t *testing.T) {
+	dataHome := t.TempDir()
+	t.Setenv("XDG_DATA_HOME", dataHome)
+	cwd := t.TempDir()
+	runGitForScratchTest(t, cwd, "init")
+	if err := os.WriteFile(filepath.Join(cwd, "README.md"), []byte("hi"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	runGitForScratchTest(t, cwd, "add", "README.md")
+	runGitForScratchTest(t, cwd, "commit", "-m", "init")
+	var stdout bytes.Buffer
+	var stderr bytes.Buffer
+
+	exitCode := runWithDeps([]string{"exec", "--output-format", "stream-json", "write scratch then fail"}, &stdout, &stderr, appDeps{
+		getwd: func() (string, error) { return cwd, nil },
+		resolveConfig: func(_ string, _ config.Overrides) (config.ResolvedConfig, error) {
+			return execResolvedConfig(), nil
+		},
+		newProvider: func(config.ProviderProfile) (zeroruntime.Provider, error) {
+			return toolThenErrorExecProvider{
+				toolCallID: "call_scratch",
+				toolName:   "write_file",
+				arguments:  `{"path":"_debug.py","content":"print('debug')"}`,
+				err:        errors.New("provider failed after tool"),
+			}, nil
+		},
+		newSandboxStore: func() (*sandbox.GrantStore, error) {
+			return sandbox.NewGrantStore(sandbox.StoreOptions{FilePath: filepath.Join(t.TempDir(), "sandbox-grants.json")})
+		},
+	})
+
+	if exitCode != exitProvider {
+		t.Fatalf("expected provider exit %d, got %d: stdout=%q stderr=%q", exitProvider, exitCode, stdout.String(), stderr.String())
+	}
+	events := decodeJSONLines(t, stdout.String())
+	warning := findJSONEvent(t, events, "warning")
+	message, _ := warning["message"].(string)
+	if !strings.Contains(message, "_debug.py") {
+		t.Fatalf("expected scratch warning to mention _debug.py, got %#v", warning)
+	}
+}
+
 func TestRunExecReadsStreamJSONPromptFromStdin(t *testing.T) {
 	t.Setenv("XDG_DATA_HOME", t.TempDir())
 	cwd := t.TempDir()
@@ -790,6 +833,13 @@ type toolCallingExecProvider struct {
 	answer     string
 }
 
+type toolThenErrorExecProvider struct {
+	toolCallID string
+	toolName   string
+	arguments  string
+	err        error
+}
+
 type reasoningExecProvider struct{}
 
 func (reasoningExecProvider) StreamCompletion(context.Context, zeroruntime.CompletionRequest) (<-chan zeroruntime.StreamEvent, error) {
@@ -825,6 +875,26 @@ func (provider toolCallingExecProvider) StreamCompletion(ctx context.Context, re
 	return ch, nil
 }
 
+func (provider toolThenErrorExecProvider) StreamCompletion(ctx context.Context, request zeroruntime.CompletionRequest) (<-chan zeroruntime.StreamEvent, error) {
+	for _, message := range request.Messages {
+		if message.Role == zeroruntime.MessageRoleTool {
+			return nil, provider.err
+		}
+	}
+	ch := make(chan zeroruntime.StreamEvent, 4)
+	select {
+	case <-ctx.Done():
+		close(ch)
+		return ch, ctx.Err()
+	case ch <- zeroruntime.StreamEvent{Type: zeroruntime.StreamEventToolCallStart, ToolCallID: provider.toolCallID, ToolName: provider.toolName}:
+	}
+	ch <- zeroruntime.StreamEvent{Type: zeroruntime.StreamEventToolCallDelta, ToolCallID: provider.toolCallID, ArgumentsFragment: provider.arguments}
+	ch <- zeroruntime.StreamEvent{Type: zeroruntime.StreamEventToolCallEnd, ToolCallID: provider.toolCallID}
+	ch <- zeroruntime.StreamEvent{Type: zeroruntime.StreamEventDone}
+	close(ch)
+	return ch, nil
+}
+
 func findJSONEvent(t *testing.T, events []map[string]any, eventType string) map[string]any {
 	t.Helper()
 	_, event := findJSONEventIndex(t, events, eventType)

internal/cli/exec_provider_hint_test.go

@@ -20,8 +20,8 @@ func TestWriteExecProviderErrorAppendsHint(t *testing.T) {
 		if !strings.Contains(out, "auth error:") {
 			t.Fatalf("expected raw message, got %q", out)
 		}
-		if !strings.Contains(out, "zero auth") {
-			t.Fatalf("expected an actionable hint referencing `zero auth`, got %q", out)
+		if !strings.Contains(out, "zero setup") || !strings.Contains(out, "zero auth openrouter") {
+			t.Fatalf("expected an actionable setup/auth hint, got %q", out)
 		}
 	})
 
@@ -43,7 +43,7 @@ func TestWriteExecProviderErrorAppendsHint(t *testing.T) {
 		if stderr.Len() != 0 {
 			t.Fatalf("json mode must not write to stderr, got %q", stderr.String())
 		}
-		if strings.Contains(stdout.String(), "zero auth") {
+		if strings.Contains(stdout.String(), "zero setup") {
 			t.Fatalf("json mode must not inject a hint into the structured payload, got %q", stdout.String())
 		}
 	})

internal/cli/exec_spec.go

@@ -21,11 +21,19 @@ import (
 )
 
 type execSpecDraftRun struct {
-	options            execOptions
-	stdout             io.Writer
-	stderr             io.Writer
-	deps               appDeps
-	workspaceRoot      string
+	options       execOptions
+	stdout        io.Writer
+	stderr        io.Writer
+	deps          appDeps
+	workspaceRoot string
+	// trustRoot is the ORIGINAL launch directory (captured before any --worktree
+	// reassignment), so a --use-spec run inside a --worktree of a trusted repo
+	// still keys the trust check on the source repo, not the generated worktree path.
+	trustRoot string
+	// mcpSkip is the trust verdict from the workspace MCP registration in runExec
+	// (which runs before this spec-draft path), so the spec-draft notice can report a
+	// dropped project MCP layer instead of leaving it silent.
+	mcpSkip            trustSkip
 	registry           *tools.Registry
 	modelRegistry      modelregistry.Registry
 	resolved           config.ResolvedConfig
@@ -96,6 +104,13 @@ func runExecSpecDraft(run execSpecDraftRun) int {
 	var draftInfo execSpecDraftInfo
 	runCtx, stopSignals := signalContext()
 	defer stopSignals()
+	// The spec-draft path activates no plugins, so the plugin skip is omitted. Trust
+	// keys on run.trustRoot (the original launch dir), not run.workspaceRoot, which
+	// may be a --worktree path; this keeps a --use-spec --worktree run of a trusted
+	// repo trusted. Emit at most one notice when project hooks or the project MCP
+	// layer (registered earlier in runExec, carried in run.mcpSkip) were dropped.
+	hookDispatcher, hookSkip := newHookDispatcher(run.workspaceRoot, run.trustRoot)
+	emitTrustNotice(run.stderr, hookSkip, run.mcpSkip)
 	result, err := agent.Run(runCtx, run.prompt, run.provider, agent.Options{
 		MaxTurns:        run.resolved.MaxTurns,
 		ContextWindow:   resolveAgentContextWindow(runCtx, run.modelRegistry, run.resolved.Provider),
@@ -112,7 +127,7 @@ func runExecSpecDraft(run execSpecDraftRun) int {
 		Autonomy:        "low",
 		Sandbox:         run.sandboxEngine,
 		FileTracker:     tools.NewFileTracker(),
-		Hooks:           newHookDispatcher(run.workspaceRoot),
+		Hooks:           hookDispatcher,
 		EnabledTools:    run.options.enabledTools,
 		DisabledTools:   run.options.disabledTools,
 		OnText:          writer.text,

internal/cli/exec_spec_test.go

@@ -51,7 +51,7 @@ func TestRunExecUseSpecCreatesDraftSession(t *testing.T) {
 				MaxTurns: 3,
 			}, nil
 		},
-		resolveMCPConfig: func(string) (config.MCPConfig, error) {
+		resolveMCPConfig: func(string, bool) (config.MCPConfig, error) {
 			return config.MCPConfig{}, nil
 		},
 		newProvider: func(config.ProviderProfile) (zeroruntime.Provider, error) {

internal/cli/extensions.go

@@ -205,7 +205,9 @@ func runMCPLegacyList(args []string, stdout io.Writer, stderr io.Writer, deps ap
 	if err != nil {
 		return writeAppError(stderr, "failed to resolve workspace: "+err.Error(), exitCrash)
 	}
-	cfg, err := deps.resolveMCPConfig(cwd)
+	// Enumeration for the extensions listing, never spawns a server, so it is left
+	// ungated (excludeProject=false) like the other doctor/status report sites.
+	cfg, err := deps.resolveMCPConfig(cwd, false)
 	if err != nil {
 		return writeAppError(stderr, redaction.ErrorMessage(err, redaction.Options{}), exitCrash)
 	}
@@ -286,11 +288,21 @@ func runMCPTools(ctx context.Context, args []string, stdout io.Writer, stderr io
 			return writeAppError(stderr, "failed to resolve workspace: "+err.Error(), exitCrash)
 		}
 		registry := tools.NewRegistry()
-		mcpRuntime, err := registerMCPToolsForWorkspace(ctx, cwd, registry, deps, mcp.AutonomyLow)
+		// `mcp tools list` connects to (spawns) each server to enumerate its live tools,
+		// so it is a spawn site and gates the project layer behind the trust check. No
+		// --worktree reassignment on this command path, so trustRoot == cwd.
+		mcpRuntime, mcpSkip, err := registerMCPToolsForWorkspace(ctx, cwd, registry, deps, mcp.AutonomyLow, cwd)
 		if err != nil {
 			return writeAppError(stderr, redaction.ErrorMessage(err, redaction.Options{}), exitCrash)
 		}
 		defer closeMCPRuntime(stderr, mcpRuntime)
+		// Surface the trust skip on stderr so an empty (or short) list in an untrusted
+		// workspace is explained rather than read as "nothing configured". The notice
+		// goes to stderr, leaving the list (text or --json) on stdout intact.
+		emitTrustNotice(stderr, mcpSkip)
+		// Surface the trust skip on stderr so an empty (or short) list in an untrusted
+		// workspace is explained rather than read as "nothing configured". The notice
+		// goes to stderr, leaving the list (text or --json) on stdout intact.
 		items := mcpToolList(registry)
 		if options.json {
 			payload := struct {

internal/cli/extensions_test.go

@@ -204,7 +204,7 @@ func TestRunMCPToolsListJSONAndText(t *testing.T) {
 	closeCalls := 0
 	deps := appDeps{
 		getwd: func() (string, error) { return cwd, nil },
-		resolveMCPConfig: func(workspaceRoot string) (config.MCPConfig, error) {
+		resolveMCPConfig: func(workspaceRoot string, _ bool) (config.MCPConfig, error) {
 			if workspaceRoot != cwd {
 				t.Fatalf("workspaceRoot = %q, want %q", workspaceRoot, cwd)
 			}
@@ -264,7 +264,7 @@ func TestRunMCPLegacyListAliases(t *testing.T) {
 	closeCalls := 0
 	deps := appDeps{
 		getwd: func() (string, error) { return cwd, nil },
-		resolveMCPConfig: func(workspaceRoot string) (config.MCPConfig, error) {
+		resolveMCPConfig: func(workspaceRoot string, _ bool) (config.MCPConfig, error) {
 			if workspaceRoot != cwd {
 				t.Fatalf("workspaceRoot = %q, want %q", workspaceRoot, cwd)
 			}

internal/cli/hook_dispatch.go

@@ -1,17 +1,63 @@
 package cli
 
 import (
+	"fmt"
+	"io"
+	"os"
+	"path/filepath"
+
 	"github.com/Gitlawb/zero/internal/hooks"
+	"github.com/Gitlawb/zero/internal/workspacetrust"
 )
 
+// trustSkip reports whether a runtime chokepoint dropped the project layer because
+// the workspace was not trusted, so the caller can emit a single combined notice.
+// Both chokepoints (hooks and plugins) return one, and the caller ORs them.
+type trustSkip struct {
+	// excludedProjectConfig is true when the project layer was excluded AND a
+	// project config for this surface actually exists on disk (so a notice is
+	// worth showing). It stays false in a trusted workspace, or when the
+	// workspace has no project config to skip.
+	excludedProjectConfig bool
+	// trustCheckErrored is true when the exclusion was caused by a trust-store
+	// read error rather than a clean untrusted verdict, so the caller can name
+	// the error in its notice.
+	trustCheckErrored bool
+}
+
+// resolveTrust computes the fail-closed trust verdict for a chokepoint. It treats
+// any error OR an empty trustRoot as untrusted, so a forgotten or future call site
+// cannot fail open. It returns whether the project layer should be excluded and
+// whether the decision was driven by a store-read error.
+//
+// Each chokepoint (hooks and plugins) calls this itself rather than the caller
+// resolving trust once and passing a bool down. That is deliberate: keeping the
+// check inside the chokepoint is what makes the gate fail-closed by construction.
+// Do NOT hoist it to the callers behind an excludeProject bool whose zero value
+// includes the project layer, that reintroduces a fail-open default. The extra
+// store read per session is negligible.
+func resolveTrust(trustRoot string) (excludeProject bool, trustCheckErrored bool) {
+	if trustRoot == "" {
+		return true, false
+	}
+	trusted, err := workspacetrust.IsTrusted(trustRoot)
+	if err != nil {
+		return true, true
+	}
+	return !trusted, false
+}
+
 // newHookDispatcher builds the per-session hooks dispatcher for a workspace,
 // merging user + project hooks.json and wiring the audit store. It fails OPEN:
 // any load or setup error yields a nil dispatcher, which Dispatch treats as a
 // no-op, so a malformed hooks config can never wedge tool execution. With no
 // hooks configured the dispatcher selects nothing and runs no commands, so the
 // hot path stays free of overhead until a user opts in via hooks.json.
-func newHookDispatcher(workspaceRoot string) *hooks.Dispatcher {
-	return newHookDispatcherWithExtra(workspaceRoot, nil)
+//
+// trustRoot is the original launch directory (resolved before any --worktree
+// reassignment); the project layer loads only when that root is trusted.
+func newHookDispatcher(workspaceRoot string, trustRoot string) (*hooks.Dispatcher, trustSkip) {
+	return newHookDispatcherWithExtra(workspaceRoot, nil, trustRoot)
 }
 
 // newHookDispatcherWithExtra builds the dispatcher like newHookDispatcher but also
@@ -20,10 +66,22 @@ func newHookDispatcher(workspaceRoot string) *hooks.Dispatcher {
 // appended after the configured hooks; their ids are plugin-namespaced (plugin
 // id + hook name) so they never collide with hooks.json ids. A nil/empty extra
 // slice is byte-equivalent to newHookDispatcher.
-func newHookDispatcherWithExtra(workspaceRoot string, extra []hooks.Definition) *hooks.Dispatcher {
-	loaded, err := hooks.LoadConfig(hooks.LoadOptions{Cwd: workspaceRoot})
+//
+// The trust check lives here, inside the chokepoint, so no caller can bypass it:
+// the project layer is dropped (ExcludeProject) whenever trustRoot is empty, the
+// trust store cannot be read, or the workspace is not trusted (fail-closed). The
+// returned trustSkip lets the caller emit one combined notice; the notice itself
+// is NOT emitted here.
+func newHookDispatcherWithExtra(workspaceRoot string, extra []hooks.Definition, trustRoot string) (*hooks.Dispatcher, trustSkip) {
+	excludeProject, trustCheckErrored := resolveTrust(trustRoot)
+	skip := trustSkip{
+		excludedProjectConfig: excludeProject && projectHooksFileExists(workspaceRoot),
+		trustCheckErrored:     trustCheckErrored,
+	}
+
+	loaded, err := hooks.LoadConfig(hooks.LoadOptions{Cwd: workspaceRoot, ExcludeProject: excludeProject})
 	if err != nil {
-		return nil
+		return nil, skip
 	}
 	var audit *hooks.AuditStore
 	if store, err := hooks.NewAuditStore(hooks.AuditStoreOptions{}); err == nil {
@@ -52,5 +110,42 @@ func newHookDispatcherWithExtra(workspaceRoot string, extra []hooks.Definition)
 		Config: config,
 		Audit:  audit,
 		Cwd:    workspaceRoot,
-	})
+	}), skip
+}
+
+// projectHooksFileExists reports whether a ./.zero/hooks.json is present under
+// workspaceRoot, so the caller only notices about config it actually skipped.
+func projectHooksFileExists(workspaceRoot string) bool {
+	if workspaceRoot == "" {
+		return false
+	}
+	info, err := os.Stat(filepath.Join(workspaceRoot, ".zero", "hooks.json"))
+	return err == nil && !info.IsDir()
+}
+
+// emitTrustNotice writes at most one stderr line summarizing that project-scoped
+// hooks, plugins, and/or MCP servers were skipped in an untrusted workspace. It
+// takes the skip report from each trust-gated surface (hooks, plugins, MCP) and ORs
+// them, so a workspace that only drops one surface still gets the notice. It is
+// computed once per session by the caller (each session-setup site runs once), so it
+// is naturally once-per-process. When any surface's skip was a trust-store read
+// error, the notice names that so a transient config-dir problem is diagnosable.
+func emitTrustNotice(stderr io.Writer, skips ...trustSkip) {
+	if stderr == nil {
+		return
+	}
+	excluded := false
+	storeErrored := false
+	for _, skip := range skips {
+		excluded = excluded || skip.excludedProjectConfig
+		storeErrored = storeErrored || skip.trustCheckErrored
+	}
+	if !excluded {
+		return
+	}
+	if storeErrored {
+		_, _ = fmt.Fprintln(stderr, "zero: the workspace-trust store could not be read; ignoring project hooks/plugins/MCP servers (fail-closed). Run 'zero trust' to enable.")
+		return
+	}
+	_, _ = fmt.Fprintln(stderr, "zero: ignoring project hooks/plugins/MCP servers in an untrusted workspace. Run 'zero trust' to enable.")
 }

internal/cli/mcp_commands_test.go

@@ -519,7 +519,7 @@ func TestRunMCPListRedactsURLCredentialsAndSensitiveQueryParams(t *testing.T) {
 	commandSecret := "sk-proj-" + strings.Repeat("a", 24)
 	deps := appDeps{
 		getwd: func() (string, error) { return cwd, nil },
-		resolveMCPConfig: func(workspaceRoot string) (config.MCPConfig, error) {
+		resolveMCPConfig: func(workspaceRoot string, _ bool) (config.MCPConfig, error) {
 			if workspaceRoot != cwd {
 				t.Fatalf("workspaceRoot = %q, want %q", workspaceRoot, cwd)
 			}
@@ -566,7 +566,7 @@ func TestRunMCPCheckRegistersOnlyRequestedServer(t *testing.T) {
 
 	exitCode := runWithDeps([]string{"mcp", "check", "docs", "--json"}, &stdout, &stderr, appDeps{
 		getwd: func() (string, error) { return cwd, nil },
-		resolveMCPConfig: func(workspaceRoot string) (config.MCPConfig, error) {
+		resolveMCPConfig: func(workspaceRoot string, _ bool) (config.MCPConfig, error) {
 			if workspaceRoot != cwd {
 				t.Fatalf("workspaceRoot = %q, want %q", workspaceRoot, cwd)
 			}
@@ -618,7 +618,7 @@ func TestRunMCPCheckClosesRuntimeReturnedWithError(t *testing.T) {
 
 	exitCode := runWithDeps([]string{"mcp", "check", "docs"}, &stdout, &stderr, appDeps{
 		getwd: func() (string, error) { return cwd, nil },
-		resolveMCPConfig: func(workspaceRoot string) (config.MCPConfig, error) {
+		resolveMCPConfig: func(workspaceRoot string, _ bool) (config.MCPConfig, error) {
 			return config.MCPConfig{Servers: map[string]config.MCPServerConfig{
 				"docs": {Type: "stdio", Command: "docs-mcp"},
 			}}, nil

internal/cli/mcp_config.go

@@ -29,6 +29,27 @@ type mcpWritableConfig struct {
 	serverRaw map[string]json.RawMessage
 }
 
+// projectMCPConfigExists reports whether the workspace's project ./.zero/config.json
+// declares any MCP servers, so the trust notice fires only when project MCP config was
+// actually skipped (mirroring projectHooksFileExists / projectPluginsDirExists). A
+// missing or unparseable file, or one that declares no servers, returns false: there
+// is nothing to notice about. This only reads the file; it never spawns a server, so
+// it is safe to call on an untrusted workspace.
+func projectMCPConfigExists(workspaceRoot string) bool {
+	if workspaceRoot == "" {
+		return false
+	}
+	data, err := os.ReadFile(filepath.Join(workspaceRoot, ".zero", "config.json"))
+	if err != nil {
+		return false
+	}
+	var fc config.FileConfig
+	if err := json.Unmarshal(data, &fc); err != nil {
+		return false
+	}
+	return len(fc.MCP.Servers) > 0
+}
+
 func runMCPAdd(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) int {
 	options, help, err := parseMCPAddArgs(args)
 	if err != nil {
@@ -243,12 +264,25 @@ func runMCPCheck(ctx context.Context, args []string, stdout io.Writer, stderr io
 	if err != nil {
 		return writeAppError(stderr, "failed to resolve workspace: "+err.Error(), exitCrash)
 	}
-	cfg, err := deps.resolveMCPConfig(cwd)
+	// `mcp check` connects to (spawns) the named server below to enumerate its tools,
+	// so it is a spawn site: gate the project layer behind the trust check (fail-closed)
+	// so a cloned repo cannot have `zero mcp check <its-server>` run its command. No
+	// --worktree reassignment on this command path, so trustRoot == cwd.
+	mcpExcludeProject, trustCheckErrored := resolveTrust(cwd)
+	cfg, err := deps.resolveMCPConfig(cwd, mcpExcludeProject)
 	if err != nil {
 		return writeAppError(stderr, redaction.ErrorMessage(err, redaction.Options{}), exitCrash)
 	}
 	raw, ok := cfg.Servers[serverName]
 	if !ok {
+		// The server may be missing because the project layer was gated out in an
+		// untrusted workspace; emit the same one-line notice the other spawn sites use so
+		// a dropped project server reads as "run zero trust", not a bare miss. The notice
+		// self-gates: it stays silent unless a project MCP config was actually excluded.
+		emitTrustNotice(stderr, trustSkip{
+			excludedProjectConfig: mcpExcludeProject && projectMCPConfigExists(cwd),
+			trustCheckErrored:     trustCheckErrored,
+		})
 		return writeAppError(stderr, fmt.Sprintf("MCP server %q is not configured", serverName), exitCrash)
 	}
 	if raw.Disabled {

internal/cli/mcp_oauth.go

@@ -55,8 +55,12 @@ func runMCPOAuthLogin(args []string, stdout io.Writer, stderr io.Writer, deps ap
 	}
 	serverName := positional[0]
 
-	server, err := resolveOAuthServer(deps, serverName)
+	server, skip, err := resolveOAuthServer(deps, serverName)
 	if err != nil {
+		// When the named server was dropped because the workspace is untrusted, explain
+		// that with the same one-line notice the other trust-gated MCP paths emit, so a
+		// gated project OAuth server reads as "run zero trust", not a bare miss.
+		emitTrustNotice(stderr, skip)
 		return writeAppError(stderr, redaction.ErrorMessage(err, redaction.Options{}), exitCrash)
 	}
 
@@ -82,7 +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)
 	}
 
@@ -112,7 +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)
 	}
@@ -193,33 +197,47 @@ func filterTokenStatuses(statuses []mcp.TokenStatus, serverName string) []mcp.To
 }
 
 // resolveOAuthServer loads the workspace MCP config and returns the named server
-// after verifying it declares OAuth authentication.
-func resolveOAuthServer(deps appDeps, serverName string) (mcp.Server, error) {
+// after verifying it declares OAuth authentication. It also returns a trustSkip so the
+// caller can emit the workspace-trust notice when a project server was dropped.
+//
+// Login is gated behind workspace trust (fail-closed) like the stdio spawn paths: it
+// loads a project-scoped MCP server and drives its OAuth flow (endpoint discovery,
+// dynamic client registration, code exchange, token persist) against the server's
+// configured URL. That is a project-scoped server being activated and contacted, so an
+// untrusted clone's ./.zero/config.json OAuth server must be dropped rather than loaded —
+// leaving it ungated would let a cloned repo initiate outbound OAuth I/O and persist a
+// token. There is no --worktree reassignment on this command path, so trustRoot == cwd.
+func resolveOAuthServer(deps appDeps, serverName string) (mcp.Server, trustSkip, error) {
 	if err := mcp.ValidateServerName(serverName); err != nil {
-		return mcp.Server{}, err
+		return mcp.Server{}, trustSkip{}, err
 	}
 	cwd, err := deps.getwd()
 	if err != nil {
-		return mcp.Server{}, fmt.Errorf("failed to resolve workspace: %w", err)
+		return mcp.Server{}, trustSkip{}, fmt.Errorf("failed to resolve workspace: %w", err)
 	}
-	cfg, err := deps.resolveMCPConfig(cwd)
+	excludeProject, trustCheckErrored := resolveTrust(cwd)
+	skip := trustSkip{
+		excludedProjectConfig: excludeProject && projectMCPConfigExists(cwd),
+		trustCheckErrored:     trustCheckErrored,
+	}
+	cfg, err := deps.resolveMCPConfig(cwd, excludeProject)
 	if err != nil {
-		return mcp.Server{}, err
+		return mcp.Server{}, skip, err
 	}
 	servers, err := mcp.NormalizeConfig(cfg)
 	if err != nil {
-		return mcp.Server{}, err
+		return mcp.Server{}, skip, err
 	}
 	for _, server := range servers {
 		if server.Name != serverName {
 			continue
 		}
 		if !strings.EqualFold(server.Auth, mcp.ServerAuthOAuth) {
-			return mcp.Server{}, fmt.Errorf("MCP server %q does not declare auth: \"oauth\"", serverName)
+			return mcp.Server{}, skip, fmt.Errorf("MCP server %q does not declare auth: \"oauth\"", serverName)
 		}
-		return server, nil
+		return server, skip, nil
 	}
-	return mcp.Server{}, fmt.Errorf("MCP server %q is not configured", serverName)
+	return mcp.Server{}, skip, fmt.Errorf("MCP server %q is not configured", serverName)
 }
 
 func oauthConfigForServer(server mcp.Server) mcp.OAuthConfig {

internal/cli/mcp_oauth_test.go

@@ -6,14 +6,17 @@ import (
 	"net/http"
 	"net/http/httptest"
 	"net/url"
+	"os"
 	"path/filepath"
 	"strings"
 	"sync"
+	"sync/atomic"
 	"testing"
 	"time"
 
 	"github.com/Gitlawb/zero/internal/config"
 	"github.com/Gitlawb/zero/internal/mcp"
+	"github.com/Gitlawb/zero/internal/workspacetrust"
 )
 
 // syncBuffer is a goroutine-safe writer used when a background goroutine reads
@@ -85,6 +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
@@ -98,6 +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) {
@@ -119,23 +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) (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,
 	}
@@ -170,9 +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)
@@ -194,7 +208,7 @@ func TestRunMCPOAuthLoginRejectsNonOAuthServer(t *testing.T) {
 	deps := appDeps{
 		getwd:            func() (string, error) { return cwd, nil },
 		newMCPTokenStore: func() (*mcp.TokenStore, error) { return store, nil },
-		resolveMCPConfig: func(workspaceRoot string) (config.MCPConfig, error) {
+		resolveMCPConfig: func(workspaceRoot string, _ bool) (config.MCPConfig, error) {
 			return config.MCPConfig{Servers: map[string]config.MCPServerConfig{
 				"plain": {Type: "http", URL: "https://plain.invalid/mcp"},
 			}}, nil
@@ -210,6 +224,291 @@ func TestRunMCPOAuthLoginRejectsNonOAuthServer(t *testing.T) {
 	}
 }
 
+// TestRunMCPOAuthLoginGatedInUntrustedWorkspace is the load-bearing security test for
+// the OAuth login gate: an OAuth MCP server defined ONLY in an untrusted workspace's
+// ./.zero/config.json must be dropped before login. `zero mcp oauth login` must refuse
+// with "not configured", surface the trust notice, and store NO token — so the
+// discovery/registration/token-exchange flow never fires against the project URL. If the
+// gate is reverted to resolveMCPConfig(cwd, false), the server survives, login proceeds,
+// a token is stored, and every assertion below fails.
+func TestRunMCPOAuthLoginGatedInUntrustedWorkspace(t *testing.T) {
+	setTrustConfigRoot(t) // isolates the trust store; the workspace stays untrusted
+	// A live server standing in for the untrusted project's OAuth server. A gated login
+	// must never reach it, so hits must stay 0 — a direct assertion of no outbound I/O
+	// (discovery/registration/token exchange) against a cloned repo's configured URL. If
+	// the gate leaks, mcp.Login contacts this URL and the hit count rises.
+	var hits atomic.Int64
+	oauthSrv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		hits.Add(1)
+		w.WriteHeader(http.StatusNotFound)
+	}))
+	defer oauthSrv.Close()
+
+	cwd := t.TempDir()
+	// A real project MCP config on disk so projectMCPConfigExists() is true and the
+	// notice fires. resolveMCPConfig is faked, but the notice detector reads disk.
+	if err := os.MkdirAll(filepath.Join(cwd, ".zero"), 0o700); err != nil {
+		t.Fatal(err)
+	}
+	body := `{"mcp":{"servers":{"remote":{"type":"http","url":"` + oauthSrv.URL + `","auth":"oauth"}}}}`
+	if err := os.WriteFile(filepath.Join(cwd, ".zero", "config.json"), []byte(body), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	store, err := mcp.NewTokenStore(mcp.TokenStoreOptions{
+		FilePath: filepath.Join(t.TempDir(), "mcp-oauth-tokens.json"),
+	})
+	if err != nil {
+		t.Fatalf("NewTokenStore() error = %v", err)
+	}
+	// Points every OAuth endpoint at the live server: if the gate leaks and login
+	// proceeds, the flow contacts oauthSrv and hits.Load() becomes non-zero.
+	projectServer := config.MCPServerConfig{
+		Type: "http",
+		URL:  oauthSrv.URL,
+		Auth: "oauth",
+		OAuth: &config.MCPOAuthConfig{
+			ClientID:              "client-123",
+			AuthorizationEndpoint: oauthSrv.URL + "/authorize",
+			TokenEndpoint:         oauthSrv.URL + "/token",
+			Scopes:                []string{"read"},
+		},
+	}
+	deps := appDeps{
+		getwd:            func() (string, error) { return cwd, nil },
+		newMCPTokenStore: func() (*mcp.TokenStore, error) { return store, nil },
+		resolveMCPConfig: func(_ string, excludeProject bool) (config.MCPConfig, error) {
+			servers := map[string]config.MCPServerConfig{}
+			if !excludeProject {
+				servers["remote"] = projectServer
+			}
+			return config.MCPConfig{Servers: servers}, nil
+		},
+		now: time.Now,
+	}
+
+	// The gate must short-circuit BEFORE the interactive OAuth flow. Bound the run so a
+	// regression that drops the gate (login proceeds and blocks on the loopback callback)
+	// fails here in 10s instead of hanging until the package test timeout.
+	stdout := &syncBuffer{}
+	stderr := &syncBuffer{}
+	done := make(chan int, 1)
+	go func() {
+		done <- runWithDeps([]string{"mcp", "oauth", "login", "remote"}, stdout, stderr, deps)
+	}()
+	var exitCode int
+	select {
+	case exitCode = <-done:
+	case <-time.After(10 * time.Second):
+		t.Fatalf("login was not gated; it blocked on the OAuth flow — the untrusted gate is missing. stderr=%q", stderr.String())
+	}
+	if exitCode == exitSuccess {
+		t.Fatalf("login on a project-only OAuth server in an untrusted workspace must fail; stdout=%q stderr=%q", stdout.String(), stderr.String())
+	}
+	if !strings.Contains(stderr.String(), "not configured") {
+		t.Fatalf("stderr must report the server as not configured, got %q", stderr.String())
+	}
+	if !strings.Contains(stderr.String(), "MCP servers") || !strings.Contains(stderr.String(), "zero trust") {
+		t.Fatalf("stderr must surface the workspace-trust notice, got %q", stderr.String())
+	}
+	if _, ok, _ := store.Load("remote"); ok {
+		t.Fatal("no token must be stored when login is gated")
+	}
+	if got := hits.Load(); got != 0 {
+		t.Fatalf("gated login must never contact the project OAuth server URL, got %d hit(s)", got)
+	}
+}
+
+// TestResolveOAuthServerTrustedWorkspaceReturnsServer proves the gate does not
+// over-block: in a TRUSTED workspace the project OAuth server resolves normally with a
+// clean skip, so login proceeds as before. Calls resolveOAuthServer directly to avoid
+// driving the full interactive login flow.
+func TestResolveOAuthServerTrustedWorkspaceReturnsServer(t *testing.T) {
+	setTrustConfigRoot(t)
+	cwd := t.TempDir()
+	if err := workspacetrust.Trust(cwd); err != nil {
+		t.Fatalf("Trust(cwd): %v", err)
+	}
+	var gotExclude bool
+	deps := appDeps{
+		getwd: func() (string, error) { return cwd, nil },
+		resolveMCPConfig: func(_ string, excludeProject bool) (config.MCPConfig, error) {
+			gotExclude = excludeProject
+			servers := map[string]config.MCPServerConfig{}
+			if !excludeProject {
+				servers["remote"] = oauthTestServerConfig()
+			}
+			return config.MCPConfig{Servers: servers}, nil
+		},
+	}
+
+	server, skip, err := resolveOAuthServer(deps, "remote")
+	if err != nil {
+		t.Fatalf("trusted workspace must resolve the project OAuth server, got err %v", err)
+	}
+	if gotExclude {
+		t.Fatalf("trusted workspace must resolve MCP config with excludeProject=false")
+	}
+	if server.Name != "remote" {
+		t.Fatalf("server.Name = %q, want remote", server.Name)
+	}
+	if skip.excludedProjectConfig {
+		t.Fatalf("trusted workspace must not report an excluded project config")
+	}
+}
+
+// TestResolveOAuthServerGatedInUntrustedWorkspace is the fast load-bearing guard for the
+// gate: in an untrusted workspace a project-only OAuth server is dropped, so resolution
+// fails with "not configured" and flags the excluded project config for the notice.
+// Reverting the gate to resolveMCPConfig(cwd, false) makes the server resolve and this
+// test fails immediately (no 10s login-flow timeout needed).
+func TestResolveOAuthServerGatedInUntrustedWorkspace(t *testing.T) {
+	setTrustConfigRoot(t)
+	cwd := t.TempDir()
+	if err := os.MkdirAll(filepath.Join(cwd, ".zero"), 0o700); err != nil {
+		t.Fatal(err)
+	}
+	body := `{"mcp":{"servers":{"remote":{"type":"http","url":"https://remote.invalid/mcp","auth":"oauth"}}}}`
+	if err := os.WriteFile(filepath.Join(cwd, ".zero", "config.json"), []byte(body), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	deps := appDeps{
+		getwd: func() (string, error) { return cwd, nil },
+		resolveMCPConfig: func(_ string, excludeProject bool) (config.MCPConfig, error) {
+			servers := map[string]config.MCPServerConfig{}
+			if !excludeProject {
+				servers["remote"] = oauthTestServerConfig()
+			}
+			return config.MCPConfig{Servers: servers}, nil
+		},
+	}
+
+	_, skip, err := resolveOAuthServer(deps, "remote")
+	if err == nil {
+		t.Fatal("untrusted workspace must not resolve a project-only OAuth server")
+	}
+	if !strings.Contains(err.Error(), "not configured") {
+		t.Fatalf("want not-configured error, got %v", err)
+	}
+	if !skip.excludedProjectConfig {
+		t.Fatal("a dropped project OAuth config must flag excludedProjectConfig for the notice")
+	}
+	if skip.trustCheckErrored {
+		t.Fatal("a clean untrusted verdict is not a store-read error")
+	}
+}
+
+func oauthTestServerConfig() config.MCPServerConfig {
+	return config.MCPServerConfig{
+		Type: "http",
+		URL:  "https://remote.invalid/mcp",
+		Auth: "oauth",
+		OAuth: &config.MCPOAuthConfig{
+			ClientID:              "client-123",
+			AuthorizationEndpoint: "https://remote.invalid/authorize",
+			TokenEndpoint:         "https://remote.invalid/token",
+			Scopes:                []string{"read"},
+		},
+	}
+}
+
+// TestRunMCPOAuthLoginTrustedProjectServerStoresToken proves the gate does not over-block
+// end-to-end: in a TRUSTED workspace, a project-only OAuth server (returned by the fake
+// ONLY when excludeProject is false) is resolved, the full login flow runs, and the token
+// is persisted with exitSuccess. This exercises the trusted-resolve + login-success path
+// as one execution; if trust were dropped, excludeProject would be true, the fake would
+// drop the server, and login would fail with "not configured".
+func TestRunMCPOAuthLoginTrustedProjectServerStoresToken(t *testing.T) {
+	setTrustConfigRoot(t)
+	tokenServer := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		_ = json.NewEncoder(w).Encode(map[string]any{
+			"access_token":  "access-final",
+			"refresh_token": "refresh-final",
+			"token_type":    "Bearer",
+			"expires_in":    3600,
+		})
+	}))
+	defer tokenServer.Close()
+
+	cwd := t.TempDir()
+	if err := workspacetrust.Trust(cwd); err != nil {
+		t.Fatalf("Trust(cwd): %v", err)
+	}
+	store, err := mcp.NewTokenStore(mcp.TokenStoreOptions{
+		FilePath: filepath.Join(t.TempDir(), "mcp-oauth-tokens.json"),
+	})
+	if err != nil {
+		t.Fatalf("NewTokenStore() error = %v", err)
+	}
+	projectServer := config.MCPServerConfig{
+		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"},
+		},
+	}
+	var gotExclude bool
+	deps := appDeps{
+		getwd:            func() (string, error) { return cwd, nil },
+		newMCPTokenStore: func() (*mcp.TokenStore, error) { return store, nil },
+		resolveMCPConfig: func(_ string, excludeProject bool) (config.MCPConfig, error) {
+			gotExclude = excludeProject
+			servers := map[string]config.MCPServerConfig{}
+			if !excludeProject {
+				servers["remote"] = projectServer
+			}
+			return config.MCPConfig{Servers: servers}, nil
+		},
+		now: time.Now,
+	}
+
+	stdout := &syncBuffer{}
+	stderr := &syncBuffer{}
+	go func() {
+		deadline := time.Now().Add(3 * time.Second)
+		for time.Now().Before(deadline) {
+			callbackURL := extractCallbackURL(stdout.String())
+			if callbackURL != "" {
+				_, _ = http.Get(callbackURL)
+				return
+			}
+			time.Sleep(5 * time.Millisecond)
+		}
+	}()
+	done := make(chan int, 1)
+	go func() {
+		done <- runWithDeps([]string{"mcp", "oauth", "login", "remote"}, stdout, stderr, deps)
+	}()
+	var exitCode int
+	select {
+	case exitCode = <-done:
+	case <-time.After(10 * time.Second):
+		t.Fatalf("trusted login did not complete within 10s; stderr=%s stdout=%s", stderr.String(), stdout.String())
+	}
+	if exitCode != exitSuccess {
+		t.Fatalf("trusted login must succeed; exit=%d stderr=%s", exitCode, stderr.String())
+	}
+	if gotExclude {
+		t.Fatal("trusted workspace must resolve MCP config with excludeProject=false")
+	}
+	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)
+	}
+	if token.AccessToken != "access-final" {
+		t.Fatalf("stored token = %#v", token)
+	}
+}
+
 func TestRunMCPOAuthUnknownSubcommand(t *testing.T) {
 	var stdout, stderr bytes.Buffer
 	exitCode := runWithDeps([]string{"mcp", "oauth", "bogus"}, &stdout, &stderr, appDeps{})

internal/cli/mcp_tools.go

@@ -19,22 +19,38 @@ type mcpToolListItem struct {
 	Permission  string `json:"permission"`
 }
 
-func registerMCPToolsForWorkspace(ctx context.Context, workspaceRoot string, registry *tools.Registry, deps appDeps, autonomy mcp.PermissionAutonomy) (mcpToolRuntime, error) {
-	cfg, err := deps.resolveMCPConfig(workspaceRoot)
+// registerMCPToolsForWorkspace resolves and registers the workspace's MCP servers.
+// This spawns stdio servers, so it gates the project config layer behind the
+// workspace-trust check: trustRoot is the ORIGINAL launch directory (resolved before
+// any --worktree reassignment) so a worktree of a trusted repo inherits that trust.
+// resolveTrust fails closed, so an empty trustRoot or a store-read error excludes the
+// project layer and a cloned repo cannot spawn its ./.zero/config.json MCP servers.
+//
+// It returns a trustSkip alongside the runtime so the caller can fold the MCP gate
+// into the one-line trust notice (mirroring the hooks and plugins chokepoints);
+// otherwise a workspace whose only project config is MCP would be gated silently.
+func registerMCPToolsForWorkspace(ctx context.Context, workspaceRoot string, registry *tools.Registry, deps appDeps, autonomy mcp.PermissionAutonomy, trustRoot string) (mcpToolRuntime, trustSkip, error) {
+	excludeProject, trustCheckErrored := resolveTrust(trustRoot)
+	skip := trustSkip{
+		excludedProjectConfig: excludeProject && projectMCPConfigExists(workspaceRoot),
+		trustCheckErrored:     trustCheckErrored,
+	}
+	cfg, err := deps.resolveMCPConfig(workspaceRoot, excludeProject)
 	if err != nil {
-		return nil, err
+		return nil, skip, err
 	}
 	if len(cfg.Servers) == 0 {
-		return noopMCPRuntime{}, nil
+		return noopMCPRuntime{}, skip, nil
 	}
 	store, err := deps.newMCPStore()
 	if err != nil {
-		return nil, err
+		return nil, skip, err
 	}
-	return deps.registerMCPTools(ctx, registry, cfg, mcp.RegisterOptions{
+	runtime, err := deps.registerMCPTools(ctx, registry, cfg, mcp.RegisterOptions{
 		PermissionStore: store,
 		Autonomy:        autonomy,
 	})
+	return runtime, skip, err
 }
 
 func execMCPAutonomy(options execOptions) mcp.PermissionAutonomy {

internal/cli/mcp_trust_edge_test.go

@@ -0,0 +1,221 @@
+package cli
+
+// Edge-case coverage for the MCP trust gate on the oauth-login and check paths:
+// the error-return contract of the gated resolveOAuthServer, the trust-store-read-error
+// notice variant, and the accepted (advisory) notice over-emission on the not-oauth path.
+
+import (
+	"bytes"
+	"errors"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"github.com/Gitlawb/zero/internal/config"
+)
+
+// writeProjectMCPConfig drops a ./.zero/config.json under dir declaring one MCP server,
+// so projectMCPConfigExists() reads it as present.
+func writeProjectMCPConfig(t *testing.T, dir string) {
+	t.Helper()
+	if err := os.MkdirAll(filepath.Join(dir, ".zero"), 0o700); err != nil {
+		t.Fatal(err)
+	}
+	body := `{"mcp":{"servers":{"proj":{"type":"stdio","command":"proj-cmd"}}}}`
+	if err := os.WriteFile(filepath.Join(dir, ".zero", "config.json"), []byte(body), 0o600); err != nil {
+		t.Fatal(err)
+	}
+}
+
+// breakTrustStore makes the trust store unreadable (creates trust.json as a directory),
+// so workspacetrust.IsTrusted returns an error and resolveTrust fails closed with
+// trustCheckErrored=true.
+func breakTrustStore(t *testing.T) {
+	t.Helper()
+	configRoot := setTrustConfigRoot(t)
+	if err := os.MkdirAll(filepath.Join(configRoot, "zero", "trust.json"), 0o700); err != nil {
+		t.Fatal(err)
+	}
+}
+
+// TestResolveOAuthServerValidationErrorReturnsZeroSkip locks the new 3-value contract on
+// the pre-resolve error path: an invalid server name errors before trust is resolved, so
+// the returned skip is zero (nothing was excluded yet).
+func TestResolveOAuthServerValidationErrorReturnsZeroSkip(t *testing.T) {
+	deps := appDeps{getwd: func() (string, error) { return t.TempDir(), nil }}
+	_, skip, err := resolveOAuthServer(deps, "bad name with spaces")
+	if err == nil {
+		t.Fatal("invalid server name must error")
+	}
+	if skip.excludedProjectConfig || skip.trustCheckErrored {
+		t.Fatalf("a validation error must return a zero skip, got %+v", skip)
+	}
+}
+
+// TestResolveOAuthServerGetwdErrorPropagates covers the getwd error path of the gated
+// resolver.
+func TestResolveOAuthServerGetwdErrorPropagates(t *testing.T) {
+	deps := appDeps{getwd: func() (string, error) { return "", errors.New("boom") }}
+	_, _, err := resolveOAuthServer(deps, "remote")
+	if err == nil || !strings.Contains(err.Error(), "failed to resolve workspace") {
+		t.Fatalf("getwd error must propagate, got %v", err)
+	}
+}
+
+// TestRunMCPOAuthLoginStoreErrorNotice proves the store-read-error branch: when the trust
+// store cannot be read, login fails closed AND the notice names the store error (the
+// "could not be read" variant) rather than the plain untrusted wording.
+func TestRunMCPOAuthLoginStoreErrorNotice(t *testing.T) {
+	breakTrustStore(t)
+	cwd := t.TempDir()
+	writeProjectMCPConfig(t, cwd)
+	deps := appDeps{
+		getwd: func() (string, error) { return cwd, nil },
+		resolveMCPConfig: func(_ string, excludeProject bool) (config.MCPConfig, error) {
+			servers := map[string]config.MCPServerConfig{}
+			if !excludeProject {
+				servers["proj"] = config.MCPServerConfig{Type: "http", URL: "https://x.invalid", Auth: "oauth"}
+			}
+			return config.MCPConfig{Servers: servers}, nil
+		},
+	}
+	var out, errBuf bytes.Buffer
+	if code := runWithDeps([]string{"mcp", "oauth", "login", "proj"}, &out, &errBuf, deps); code == exitSuccess {
+		t.Fatal("login must fail closed on a trust-store read error")
+	}
+	if got := errBuf.String(); !strings.Contains(got, "could not be read") || !strings.Contains(got, "not configured") {
+		t.Fatalf("store-error login must emit the store-errored notice AND not-configured, got %q", got)
+	}
+}
+
+// TestRunMCPCheckStoreErrorNotice is the mcp-check sibling of the store-error case.
+func TestRunMCPCheckStoreErrorNotice(t *testing.T) {
+	breakTrustStore(t)
+	cwd := t.TempDir()
+	writeProjectMCPConfig(t, cwd)
+	deps := appDeps{
+		getwd: func() (string, error) { return cwd, nil },
+		resolveMCPConfig: func(_ string, excludeProject bool) (config.MCPConfig, error) {
+			servers := map[string]config.MCPServerConfig{}
+			if !excludeProject {
+				servers["proj"] = config.MCPServerConfig{Type: "stdio", Command: "proj-cmd"}
+			}
+			return config.MCPConfig{Servers: servers}, nil
+		},
+	}
+	var out, errBuf bytes.Buffer
+	if code := runWithDeps([]string{"mcp", "check", "proj"}, &out, &errBuf, deps); code == exitSuccess {
+		t.Fatal("mcp check must fail closed on a trust-store read error")
+	}
+	if got := errBuf.String(); !strings.Contains(got, "could not be read") {
+		t.Fatalf("mcp check store-error must emit the store-errored notice, got %q", got)
+	}
+}
+
+// TestRunMCPOAuthLoginNoticeCoFiresWithNonOAuthError documents accepted behavior: in an
+// untrusted workspace with project MCP config on disk, a login that fails because the
+// named USER-config server is not OAuth still co-emits the trust notice (project MCP was
+// excluded this run, which is true). The notice is advisory and keyed on "project config
+// was dropped", not on "this server was the dropped one" — the same coarse scoping the
+// mcp-check path uses. If this behavior changes intentionally, update this test.
+func TestRunMCPOAuthLoginNoticeCoFiresWithNonOAuthError(t *testing.T) {
+	setTrustConfigRoot(t) // untrusted
+	cwd := t.TempDir()
+	writeProjectMCPConfig(t, cwd)
+	deps := appDeps{
+		getwd: func() (string, error) { return cwd, nil },
+		// "foo" is a user-config server (survives the gate) that does not declare oauth.
+		resolveMCPConfig: func(_ string, _ bool) (config.MCPConfig, error) {
+			return config.MCPConfig{Servers: map[string]config.MCPServerConfig{
+				"foo": {Type: "http", URL: "https://foo.invalid"},
+			}}, nil
+		},
+	}
+	var out, errBuf bytes.Buffer
+	if code := runWithDeps([]string{"mcp", "oauth", "login", "foo"}, &out, &errBuf, deps); code == exitSuccess {
+		t.Fatal("login on a non-oauth server must fail")
+	}
+	got := errBuf.String()
+	if !strings.Contains(got, "oauth") {
+		t.Fatalf("want the not-oauth error, got %q", got)
+	}
+	if !strings.Contains(got, "zero trust") {
+		t.Fatalf("expected the advisory trust notice to co-fire (accepted behavior), got %q", got)
+	}
+}
+
+// TestResolveOAuthServerResolveConfigErrorPropagates covers the resolveMCPConfig-error
+// return of the gated resolver: the error propagates with the computed skip.
+func TestResolveOAuthServerResolveConfigErrorPropagates(t *testing.T) {
+	setTrustConfigRoot(t)
+	deps := appDeps{
+		getwd: func() (string, error) { return t.TempDir(), nil },
+		resolveMCPConfig: func(_ string, _ bool) (config.MCPConfig, error) {
+			return config.MCPConfig{}, errors.New("resolve boom")
+		},
+	}
+	_, _, err := resolveOAuthServer(deps, "remote")
+	if err == nil || !strings.Contains(err.Error(), "resolve boom") {
+		t.Fatalf("resolveMCPConfig error must propagate, got %v", err)
+	}
+}
+
+// TestResolveOAuthServerNormalizeErrorPropagates covers the NormalizeConfig-error return:
+// a malformed server (stdio with no command) fails normalization.
+func TestResolveOAuthServerNormalizeErrorPropagates(t *testing.T) {
+	setTrustConfigRoot(t)
+	deps := appDeps{
+		getwd: func() (string, error) { return t.TempDir(), nil },
+		resolveMCPConfig: func(_ string, _ bool) (config.MCPConfig, error) {
+			return config.MCPConfig{Servers: map[string]config.MCPServerConfig{
+				"remote": {Type: "stdio"}, // no command -> NormalizeConfig rejects
+			}}, nil
+		},
+	}
+	_, _, err := resolveOAuthServer(deps, "remote")
+	if err == nil || !strings.Contains(err.Error(), "requires command") {
+		t.Fatalf("NormalizeConfig error must propagate, got %v", err)
+	}
+}
+
+// TestRunMCPCheckDisabledServer covers the found-but-disabled branch of runMCPCheck.
+func TestRunMCPCheckDisabledServer(t *testing.T) {
+	setTrustConfigRoot(t)
+	deps := appDeps{
+		getwd: func() (string, error) { return t.TempDir(), nil },
+		resolveMCPConfig: func(_ string, _ bool) (config.MCPConfig, error) {
+			return config.MCPConfig{Servers: map[string]config.MCPServerConfig{
+				"proj": {Type: "stdio", Command: "c", Disabled: true},
+			}}, nil
+		},
+	}
+	var out, errBuf bytes.Buffer
+	if code := runWithDeps([]string{"mcp", "check", "proj"}, &out, &errBuf, deps); code == exitSuccess {
+		t.Fatal("mcp check on a disabled server must fail")
+	}
+	if got := errBuf.String(); !strings.Contains(got, "is disabled") {
+		t.Fatalf("want disabled error, got %q", got)
+	}
+}
+
+// TestRunMCPCheckNormalizeError covers the found-but-malformed branch of runMCPCheck: the
+// scoped NormalizeConfig fails before any spawn.
+func TestRunMCPCheckNormalizeError(t *testing.T) {
+	setTrustConfigRoot(t)
+	deps := appDeps{
+		getwd: func() (string, error) { return t.TempDir(), nil },
+		resolveMCPConfig: func(_ string, _ bool) (config.MCPConfig, error) {
+			return config.MCPConfig{Servers: map[string]config.MCPServerConfig{
+				"proj": {Type: "stdio"}, // no command
+			}}, nil
+		},
+	}
+	var out, errBuf bytes.Buffer
+	if code := runWithDeps([]string{"mcp", "check", "proj"}, &out, &errBuf, deps); code == exitSuccess {
+		t.Fatal("mcp check on a malformed server must fail")
+	}
+	if got := errBuf.String(); !strings.Contains(got, "requires command") {
+		t.Fatalf("want normalize error, got %q", got)
+	}
+}

internal/cli/mcp_trust_test.go

@@ -0,0 +1,358 @@
+package cli
+
+import (
+	"bytes"
+	"context"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"github.com/Gitlawb/zero/internal/config"
+	"github.com/Gitlawb/zero/internal/mcp"
+	"github.com/Gitlawb/zero/internal/tools"
+	"github.com/Gitlawb/zero/internal/workspacetrust"
+)
+
+// mcpTrustDeps builds an appDeps whose resolveMCPConfig HONORS excludeProject: it
+// returns a project stdio server only when excludeProject is false, mirroring the
+// real ResolveMCP gate. It records the excludeProject value it was called with and
+// whether registerMCPTools (which spawns servers) actually fired.
+func mcpTrustDeps(gotExclude *bool, spawned *bool) appDeps {
+	return appDeps{
+		resolveMCPConfig: func(_ string, excludeProject bool) (config.MCPConfig, error) {
+			*gotExclude = excludeProject
+			servers := map[string]config.MCPServerConfig{}
+			if !excludeProject {
+				servers["proj-srv"] = config.MCPServerConfig{Type: "stdio", Command: "proj-cmd"}
+			}
+			return config.MCPConfig{Servers: servers}, nil
+		},
+		newMCPStore: func() (*mcp.PermissionStore, error) { return nil, nil },
+		registerMCPTools: func(_ context.Context, _ *tools.Registry, _ config.MCPConfig, _ mcp.RegisterOptions) (mcpToolRuntime, error) {
+			*spawned = true
+			return closeFunc(func() error { return nil }), nil
+		},
+	}
+}
+
+// TestMCPGateUntrustedExcludesProjectServer proves the P0 fix: an untrusted trustRoot
+// makes registerMCPToolsForWorkspace resolve MCP config with excludeProject=true and,
+// because that drops the project server, it never spawns anything. This test is
+// load-bearing: if the gate is removed (a hardcoded excludeProject=false passed to
+// resolveMCPConfig), gotExclude is false, the project server survives, spawned flips
+// true, and every assertion below fails.
+func TestMCPGateUntrustedExcludesProjectServer(t *testing.T) {
+	setTrustConfigRoot(t)
+	repo := t.TempDir() // never trusted
+
+	var gotExclude, spawned bool
+	deps := mcpTrustDeps(&gotExclude, &spawned)
+	registry := tools.NewRegistry()
+
+	runtime, skip, err := registerMCPToolsForWorkspace(context.Background(), repo, registry, deps, mcp.AutonomyLow, repo)
+	if err != nil {
+		t.Fatalf("registerMCPToolsForWorkspace: %v", err)
+	}
+	defer func() { _ = runtime.Close() }()
+
+	if !gotExclude {
+		t.Fatalf("untrusted workspace must resolve MCP config with excludeProject=true")
+	}
+	if spawned {
+		t.Fatalf("untrusted workspace must not spawn the project MCP server")
+	}
+	if _, ok := runtime.(noopMCPRuntime); !ok {
+		t.Fatalf("with the project server dropped, the runtime should be the noop runtime, got %T", runtime)
+	}
+	// This repo has no ./.zero/config.json, so there is no project MCP config to
+	// notice about even though it is untrusted; the skip must stay clean.
+	if skip.excludedProjectConfig {
+		t.Fatalf("no project MCP config on disk, so the skip must not flag an excluded config")
+	}
+	if skip.trustCheckErrored {
+		t.Fatalf("a clean untrusted verdict is not a store-read error")
+	}
+}
+
+// TestMCPGateEmptyTrustRootFailsClosed proves fail-closed-by-construction: a caller
+// that forgot to resolve trustRoot (empty) still excludes the project layer.
+func TestMCPGateEmptyTrustRootFailsClosed(t *testing.T) {
+	setTrustConfigRoot(t)
+	repo := t.TempDir()
+	// Even trusting the repo must not help when the caller passes an empty root.
+	if err := workspacetrust.Trust(repo); err != nil {
+		t.Fatalf("Trust(repo): %v", err)
+	}
+
+	var gotExclude, spawned bool
+	deps := mcpTrustDeps(&gotExclude, &spawned)
+	registry := tools.NewRegistry()
+
+	runtime, skip, err := registerMCPToolsForWorkspace(context.Background(), repo, registry, deps, mcp.AutonomyLow, "")
+	if err != nil {
+		t.Fatalf("registerMCPToolsForWorkspace: %v", err)
+	}
+	defer func() { _ = runtime.Close() }()
+
+	if !gotExclude {
+		t.Fatalf("empty trustRoot must fail closed (excludeProject=true)")
+	}
+	if spawned {
+		t.Fatalf("empty trustRoot must not spawn the project MCP server")
+	}
+	// Empty trustRoot is a clean fail-closed verdict, not a store-read error.
+	if skip.trustCheckErrored {
+		t.Fatalf("empty trustRoot is not a store-read error")
+	}
+}
+
+// TestMCPGateTrustedSpawnsProjectServer proves R3 for MCP: after Trust(repo) the
+// project layer is included (excludeProject=false) and the project server spawns.
+func TestMCPGateTrustedSpawnsProjectServer(t *testing.T) {
+	setTrustConfigRoot(t)
+	repo := t.TempDir()
+	if err := workspacetrust.Trust(repo); err != nil {
+		t.Fatalf("Trust(repo): %v", err)
+	}
+
+	var gotExclude, spawned bool
+	deps := mcpTrustDeps(&gotExclude, &spawned)
+	registry := tools.NewRegistry()
+
+	runtime, skip, err := registerMCPToolsForWorkspace(context.Background(), repo, registry, deps, mcp.AutonomyLow, repo)
+	if err != nil {
+		t.Fatalf("registerMCPToolsForWorkspace: %v", err)
+	}
+	defer func() { _ = runtime.Close() }()
+
+	if gotExclude {
+		t.Fatalf("trusted workspace must resolve MCP config with excludeProject=false")
+	}
+	if !spawned {
+		t.Fatalf("trusted workspace must spawn the project MCP server")
+	}
+	if skip.excludedProjectConfig {
+		t.Fatalf("trusted workspace must not report the project MCP layer excluded")
+	}
+}
+
+// TestMCPToolsListSurfacesTrustNotice proves `zero mcp tools list` no longer drops the
+// project MCP layer silently in an untrusted workspace: the gated skip is surfaced on
+// stderr (the list stays on stdout), so an empty list is explained rather than read as
+// "nothing configured". Trusting the repo silences it.
+func TestMCPToolsListSurfacesTrustNotice(t *testing.T) {
+	setTrustConfigRoot(t)
+	repo := t.TempDir()
+	if err := os.MkdirAll(filepath.Join(repo, ".zero"), 0o700); err != nil {
+		t.Fatal(err)
+	}
+	body := `{"mcp":{"servers":{"proj":{"type":"stdio","command":"proj-cmd"}}}}`
+	if err := os.WriteFile(filepath.Join(repo, ".zero", "config.json"), []byte(body), 0o600); err != nil {
+		t.Fatal(err)
+	}
+
+	run := func() string {
+		var out, errBuf bytes.Buffer
+		deps := appDeps{
+			getwd:            func() (string, error) { return repo, nil },
+			resolveMCPConfig: func(string, bool) (config.MCPConfig, error) { return config.MCPConfig{}, nil },
+		}
+		if code := runWithDeps([]string{"mcp", "tools", "list"}, &out, &errBuf, deps); code != exitSuccess {
+			t.Fatalf("mcp tools list exit = %d, stderr=%q", code, errBuf.String())
+		}
+		return errBuf.String()
+	}
+
+	// Untrusted: the gated project MCP layer must be explained on stderr.
+	if errUntrusted := run(); !strings.Contains(errUntrusted, "MCP servers") || !strings.Contains(errUntrusted, "zero trust") {
+		t.Fatalf("untrusted `mcp tools list` must surface the trust notice on stderr, got %q", errUntrusted)
+	}
+
+	// Trusted: nothing is skipped, so no notice.
+	if err := workspacetrust.Trust(repo); err != nil {
+		t.Fatal(err)
+	}
+	if errTrusted := run(); strings.Contains(errTrusted, "ignoring project") {
+		t.Fatalf("trusted `mcp tools list` must not emit a trust notice, got %q", errTrusted)
+	}
+}
+
+// TestProjectMCPConfigExists exercises every branch of the notice-gating detector:
+// only a ./.zero/config.json that parses AND declares at least one server is true.
+func TestProjectMCPConfigExists(t *testing.T) {
+	writeCfg := func(t *testing.T, dir, body string) {
+		t.Helper()
+		if err := os.MkdirAll(filepath.Join(dir, ".zero"), 0o700); err != nil {
+			t.Fatal(err)
+		}
+		if err := os.WriteFile(filepath.Join(dir, ".zero", "config.json"), []byte(body), 0o600); err != nil {
+			t.Fatal(err)
+		}
+	}
+	if projectMCPConfigExists("") {
+		t.Fatal("empty workspace root must be false")
+	}
+	cases := []struct {
+		name string
+		body string // "" means write no config.json at all
+		want bool
+	}{
+		{"no file", "", false},
+		{"declares a server", `{"mcp":{"servers":{"a":{"type":"stdio","command":"x"}}}}`, true},
+		{"empty servers map", `{"mcp":{"servers":{}}}`, false},
+		{"config without mcp key", `{"model":"x"}`, false},
+		{"unparseable json", `{not valid`, false},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			dir := t.TempDir()
+			if tc.body != "" {
+				writeCfg(t, dir, tc.body)
+			}
+			if got := projectMCPConfigExists(dir); got != tc.want {
+				t.Fatalf("projectMCPConfigExists = %v, want %v", got, tc.want)
+			}
+		})
+	}
+}
+
+// TestMCPGateFailClosedOnStoreError proves the MCP surface reports a store-read error
+// (trust.json created as a directory) as trustCheckErrored, so the caller's notice can
+// name the fail-closed reason -- the same error path the hook and plugin gates cover.
+func TestMCPGateFailClosedOnStoreError(t *testing.T) {
+	configRoot := setTrustConfigRoot(t)
+	trustPath := filepath.Join(configRoot, "zero", "trust.json")
+	if err := os.MkdirAll(trustPath, 0o700); err != nil {
+		t.Fatalf("create trust.json as a directory: %v", err)
+	}
+	repo := t.TempDir()
+	if err := os.MkdirAll(filepath.Join(repo, ".zero"), 0o700); err != nil {
+		t.Fatalf("mkdir project .zero: %v", err)
+	}
+	body := `{"mcp":{"servers":{"proj":{"type":"stdio","command":"proj-cmd"}}}}`
+	if err := os.WriteFile(filepath.Join(repo, ".zero", "config.json"), []byte(body), 0o600); err != nil {
+		t.Fatalf("write project config.json: %v", err)
+	}
+
+	var gotExclude, spawned bool
+	deps := mcpTrustDeps(&gotExclude, &spawned)
+	runtime, skip, err := registerMCPToolsForWorkspace(context.Background(), repo, tools.NewRegistry(), deps, mcp.AutonomyLow, repo)
+	if err != nil {
+		t.Fatalf("registerMCPToolsForWorkspace: %v", err)
+	}
+	defer func() { _ = runtime.Close() }()
+	if !gotExclude {
+		t.Fatalf("a store-read error must fail closed (excludeProject=true)")
+	}
+	if spawned {
+		t.Fatalf("a store-read error must not spawn the project MCP server")
+	}
+	if !skip.trustCheckErrored {
+		t.Fatalf("skip must mark the store-read error so the notice can name it")
+	}
+	if !skip.excludedProjectConfig {
+		t.Fatalf("the project MCP layer must be reported excluded on the error path")
+	}
+}
+
+// TestMCPGateUntrustedNoticesProjectMCPConfig proves the notice-surfacing fix: when an
+// untrusted workspace actually declares project MCP servers in ./.zero/config.json,
+// registerMCPToolsForWorkspace reports excludedProjectConfig=true so the caller can
+// warn (the CodeRabbit finding: project MCP was gated silently). Trusting the repo
+// clears the skip.
+func TestMCPGateUntrustedNoticesProjectMCPConfig(t *testing.T) {
+	setTrustConfigRoot(t)
+	repo := t.TempDir()
+	// A real project MCP config on disk: projectMCPConfigExists reads this file.
+	if err := os.MkdirAll(filepath.Join(repo, ".zero"), 0o700); err != nil {
+		t.Fatalf("mkdir project .zero: %v", err)
+	}
+	body := `{"mcp":{"servers":{"proj":{"type":"stdio","command":"proj-cmd"}}}}`
+	if err := os.WriteFile(filepath.Join(repo, ".zero", "config.json"), []byte(body), 0o600); err != nil {
+		t.Fatalf("write project config.json: %v", err)
+	}
+
+	var gotExclude, spawned bool
+	deps := mcpTrustDeps(&gotExclude, &spawned)
+	registry := tools.NewRegistry()
+
+	// Untrusted: the project MCP layer is dropped AND flagged for the notice.
+	runtime, skip, err := registerMCPToolsForWorkspace(context.Background(), repo, registry, deps, mcp.AutonomyLow, repo)
+	if err != nil {
+		t.Fatalf("registerMCPToolsForWorkspace (untrusted): %v", err)
+	}
+	defer func() { _ = runtime.Close() }()
+	if !skip.excludedProjectConfig {
+		t.Fatalf("untrusted workspace with project MCP config must flag excludedProjectConfig for the notice")
+	}
+	if skip.trustCheckErrored {
+		t.Fatalf("a clean untrusted verdict is not a store-read error")
+	}
+
+	// Trusted: nothing is skipped, so no notice.
+	if err := workspacetrust.Trust(repo); err != nil {
+		t.Fatalf("Trust(repo): %v", err)
+	}
+	_, trustedSkip, err := registerMCPToolsForWorkspace(context.Background(), repo, tools.NewRegistry(), deps, mcp.AutonomyLow, repo)
+	if err != nil {
+		t.Fatalf("registerMCPToolsForWorkspace (trusted): %v", err)
+	}
+	if trustedSkip.excludedProjectConfig {
+		t.Fatalf("trusted workspace must not flag an excluded project MCP layer")
+	}
+}
+
+// TestMCPCheckSurfacesTrustNotice proves the `zero mcp check` notice fix: in an untrusted
+// workspace whose only definition of a server lives in ./.zero/config.json, the gate
+// drops the server and the command must emit the one-line trust notice before the
+// "not configured" error, instead of a bare miss that hides the trust exclusion. The R4
+// half proves the notice self-gates: a genuinely-absent server in a workspace with no
+// project config on disk prints "not configured" with no notice.
+func TestMCPCheckSurfacesTrustNotice(t *testing.T) {
+	setTrustConfigRoot(t)
+
+	// resolveMCPConfig fake that HONORS excludeProject: `proj` survives only when the
+	// project layer is included, mirroring the real gate.
+	dropFake := func(_ string, excludeProject bool) (config.MCPConfig, error) {
+		servers := map[string]config.MCPServerConfig{}
+		if !excludeProject {
+			servers["proj"] = config.MCPServerConfig{Type: "stdio", Command: "proj-cmd"}
+		}
+		return config.MCPConfig{Servers: servers}, nil
+	}
+
+	// Untrusted workspace whose ONLY definition of `proj` is project config on disk.
+	repo := t.TempDir()
+	if err := os.MkdirAll(filepath.Join(repo, ".zero"), 0o700); err != nil {
+		t.Fatal(err)
+	}
+	body := `{"mcp":{"servers":{"proj":{"type":"stdio","command":"proj-cmd"}}}}`
+	if err := os.WriteFile(filepath.Join(repo, ".zero", "config.json"), []byte(body), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	var out, errBuf bytes.Buffer
+	deps := appDeps{getwd: func() (string, error) { return repo, nil }, resolveMCPConfig: dropFake}
+	if code := runWithDeps([]string{"mcp", "check", "proj"}, &out, &errBuf, deps); code == exitSuccess {
+		t.Fatalf("mcp check on a gated project server must fail, got success; stderr=%q", errBuf.String())
+	}
+	if got := errBuf.String(); !strings.Contains(got, "not configured") ||
+		!strings.Contains(got, "MCP servers") || !strings.Contains(got, "zero trust") {
+		t.Fatalf("untrusted mcp check must report not-configured AND surface the trust notice, got %q", got)
+	}
+
+	// R4: an untrusted workspace with NO project MCP config on disk has nothing to
+	// notice about, so an absent server prints "not configured" with no trust notice.
+	bare := t.TempDir()
+	var out2, errBuf2 bytes.Buffer
+	deps2 := appDeps{getwd: func() (string, error) { return bare, nil }, resolveMCPConfig: dropFake}
+	if code := runWithDeps([]string{"mcp", "check", "ghost"}, &out2, &errBuf2, deps2); code == exitSuccess {
+		t.Fatalf("mcp check on an absent server must fail")
+	}
+	if got := errBuf2.String(); !strings.Contains(got, "not configured") {
+		t.Fatalf("stderr must report not configured, got %q", got)
+	}
+	if got := errBuf2.String(); strings.Contains(got, "ignoring project") || strings.Contains(got, "zero trust") {
+		t.Fatalf("no project config on disk means no trust notice, got %q", got)
+	}
+}

internal/cli/plugin_activate.go

@@ -3,6 +3,8 @@ package cli
 import (
 	"fmt"
 	"io"
+	"os"
+	"path/filepath"
 	"strings"
 
 	"github.com/Gitlawb/zero/internal/agent"
@@ -15,9 +17,13 @@ import (
 // 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.
 type pluginActivation struct {
 	hooks      []hooks.Definition
 	skillRoots []string
+	trustSkip
 }
 
 // activatePlugins loads the workspace's plugins and makes their declared
@@ -27,12 +33,26 @@ type pluginActivation struct {
 //
 // It fails OPEN: any load error (or a malformed plugin) is surfaced as a warning
 // on stderr and otherwise skipped, so a broken plugin can never wedge startup —
-// mirroring how newHookDispatcher and skills.Load tolerate bad input.
-func activatePlugins(workspaceRoot string, registry *tools.Registry, deps appDeps, stderr io.Writer) pluginActivation {
-	loaded, err := deps.loadPlugins(plugins.LoadOptions{Cwd: workspaceRoot})
+// mirroring how newHookDispatcher and skills.Load tolerate bad input. That
+// fail-OPEN handling of MALFORMED plugins is separate from the trust gate below.
+//
+// The trust check lives here, inside the chokepoint, so no caller can bypass it:
+// the project plugin root is dropped (ExcludeProject) whenever trustRoot is empty,
+// the trust store cannot be read, or the workspace is not trusted (fail-closed).
+// trustRoot is the original launch directory (resolved before any --worktree
+// reassignment). The returned activation carries the skip report so the caller can
+// emit one combined notice; the notice itself is NOT emitted here.
+func activatePlugins(workspaceRoot string, registry *tools.Registry, deps appDeps, stderr io.Writer, trustRoot string) pluginActivation {
+	excludeProject, trustCheckErrored := resolveTrust(trustRoot)
+	skip := trustSkip{
+		excludedProjectConfig: excludeProject && projectPluginsDirExists(workspaceRoot),
+		trustCheckErrored:     trustCheckErrored,
+	}
+
+	loaded, err := deps.loadPlugins(plugins.LoadOptions{Cwd: workspaceRoot, ExcludeProject: excludeProject})
 	if err != nil {
 		writePluginActivationWarning(stderr, "failed to load plugins: "+err.Error())
-		return pluginActivation{}
+		return pluginActivation{trustSkip: skip}
 	}
 
 	// Load fails OPEN per plugin: a malformed manifest is recorded as a diagnostic
@@ -56,7 +76,18 @@ func activatePlugins(workspaceRoot string, registry *tools.Registry, deps appDep
 		registry.Register(plugins.NewSkillTool(deps.skillsDir(), result.SkillRoots))
 	}
 
-	return pluginActivation{hooks: result.Hooks, skillRoots: result.SkillRoots}
+	return pluginActivation{hooks: result.Hooks, skillRoots: result.SkillRoots, trustSkip: skip}
+}
+
+// projectPluginsDirExists reports whether a ./.zero/plugins directory is present
+// under workspaceRoot, so the caller only notices about config it actually
+// skipped.
+func projectPluginsDirExists(workspaceRoot string) bool {
+	if workspaceRoot == "" {
+		return false
+	}
+	info, err := os.Stat(filepath.Join(workspaceRoot, ".zero", "plugins"))
+	return err == nil && info.IsDir()
 }
 
 // skillInfos resolves the reusable skills the model can load via the skill tool —

internal/cli/plugin_activate_test.go

@@ -51,7 +51,8 @@ func TestActivatePluginsRegistersToolAndCollectsHooks(t *testing.T) {
 
 	registry := tools.NewRegistry()
 	var stderr bytes.Buffer
-	activation := activatePlugins(t.TempDir(), registry, fakePluginDeps(t, loaded), &stderr)
+	workspace := t.TempDir()
+	activation := activatePlugins(workspace, registry, fakePluginDeps(t, loaded), &stderr, workspace)
 
 	if _, ok := registry.Get("demo_lookup"); !ok {
 		t.Fatalf("plugin tool not registered into the bootstrap registry")
@@ -90,7 +91,8 @@ func TestActivatePluginsRegistersPluginSkillTool(t *testing.T) {
 	registry := tools.NewRegistry()
 	registry.Register(tools.NewSkillTool(t.TempDir())) // core skill tool, like the real bootstrap
 	var stderr bytes.Buffer
-	activation := activatePlugins(t.TempDir(), registry, fakePluginDeps(t, loaded), &stderr)
+	workspace := t.TempDir()
+	activation := activatePlugins(workspace, registry, fakePluginDeps(t, loaded), &stderr, workspace)
 
 	if len(activation.skillRoots) != 1 {
 		t.Fatalf("expected one plugin skill root, got %#v", activation.skillRoots)
@@ -124,7 +126,8 @@ func TestActivatePluginsSurfacesLoadDiagnostics(t *testing.T) {
 	}
 	registry := tools.NewRegistry()
 	var stderr bytes.Buffer
-	activatePlugins(t.TempDir(), registry, deps, &stderr)
+	workspace := t.TempDir()
+	activatePlugins(workspace, registry, deps, &stderr, workspace)
 
 	if stderr.Len() == 0 {
 		t.Fatal("a load diagnostic must be surfaced on stderr")
@@ -143,7 +146,8 @@ func TestActivatePluginsFailsOpenOnLoadError(t *testing.T) {
 	}
 	registry := tools.NewRegistry()
 	var stderr bytes.Buffer
-	activation := activatePlugins(t.TempDir(), registry, deps, &stderr)
+	workspace := t.TempDir()
+	activation := activatePlugins(workspace, registry, deps, &stderr, workspace)
 
 	if len(activation.hooks) != 0 || len(activation.skillRoots) != 0 {
 		t.Fatalf("a load error must yield an inert activation, got %#v", activation)
@@ -173,7 +177,8 @@ func TestActivatePluginsWarnsOnMalformedPluginButKeepsGood(t *testing.T) {
 
 	registry := tools.NewRegistry()
 	var stderr bytes.Buffer
-	activatePlugins(t.TempDir(), registry, fakePluginDeps(t, []plugins.LoadedPlugin{bad, good}), &stderr)
+	workspace := t.TempDir()
+	activatePlugins(workspace, registry, fakePluginDeps(t, []plugins.LoadedPlugin{bad, good}), &stderr, workspace)
 
 	if _, ok := registry.Get("good_tool"); !ok {
 		t.Fatal("the good plugin tool must still register despite a malformed sibling")
@@ -198,7 +203,7 @@ func TestNewHookDispatcherWithExtraFoldsPluginHooks(t *testing.T) {
 		Enabled: true,
 	}}
 
-	dispatcher := newHookDispatcherWithExtra(workspace, extra)
+	dispatcher, _ := newHookDispatcherWithExtra(workspace, extra, workspace)
 	if dispatcher == nil {
 		t.Fatal("dispatcher should never be nil for a clean workspace")
 	}

internal/cli/provider_models.go

@@ -8,6 +8,7 @@ import (
 	"strings"
 
 	"github.com/Gitlawb/zero/internal/config"
+	"github.com/Gitlawb/zero/internal/providermodelcatalog"
 	"github.com/Gitlawb/zero/internal/providermodeldiscovery"
 )
 
@@ -51,6 +52,19 @@ func runProvidersModels(args []string, stdout io.Writer, stderr io.Writer, deps
 		return writeAppError(stderr, err.Error(), exitProvider)
 	}
 
+	// Apply provider-specific model filtering for catalog-backed profiles.
+	// For example, opencode-go-anthropic-compatible only permits Qwen and
+	// MiniMax model IDs; the raw /zen/go/v1/models endpoint returns many more.
+	if profile.CatalogID != "" {
+		filtered := make([]providermodeldiscovery.Model, 0, len(models))
+		for _, model := range models {
+			if providermodelcatalog.ModelIDAllowedForProvider(profile.CatalogID, model.ID) {
+				filtered = append(filtered, model)
+			}
+		}
+		models = filtered
+	}
+
 	if options.json {
 		items := make([]map[string]any, 0, len(models))
 		for _, model := range models {

internal/cli/provider_models_test.go

@@ -132,6 +132,62 @@ func TestRunProvidersModelsUnknownProviderFails(t *testing.T) {
 	}
 }
 
+func TestRunProvidersModelsFiltersForCatalogProfile(t *testing.T) {
+	var stdout bytes.Buffer
+	var stderr bytes.Buffer
+	deps := commandCenterDeps(t)
+	// Override the resolved config to include a catalog-backed profile.
+	deps.resolveConfig = func(_ string, _ config.Overrides) (config.ResolvedConfig, error) {
+		profile := config.ProviderProfile{
+			Name:         "opencode-go",
+			ProviderKind: config.ProviderKindAnthropicCompat,
+			CatalogID:    "opencode-go-anthropic-compatible",
+			BaseURL:      "https://api.example.com/zen/go/v1",
+			APIKey:       "sk-test",
+			Model:        "minimax-max",
+		}
+		return config.ResolvedConfig{
+			ActiveProvider: "opencode-go",
+			Providers:      []config.ProviderProfile{profile},
+			Provider:       profile,
+		}, nil
+	}
+	deps.discoverProviderModels = func(_ context.Context, _ config.ProviderProfile) ([]providermodeldiscovery.Model, error) {
+		return []providermodeldiscovery.Model{
+			{ID: "deepseek-chat", Description: "DeepSeek Chat"},
+			{ID: "qwen-max", Description: "Qwen Max"},
+			{ID: "minimax-max-01", Description: "MiniMax Max"},
+			{ID: "kimi-k2", Description: "Kimi K2"},
+			{ID: "glm-4", Description: "GLM-4"},
+		}, nil
+	}
+
+	exitCode := runWithDeps([]string{"providers", "models"}, &stdout, &stderr, deps)
+
+	if exitCode != exitSuccess {
+		t.Fatalf("exit = %d, want %d: %s", exitCode, exitSuccess, stderr.String())
+	}
+	out := stdout.String()
+	// Only Qwen and MiniMax model IDs should survive the filter.
+	if strings.Contains(out, "deepseek-chat") {
+		t.Error("output should not contain deepseek-chat (blocked by opencode-go-anthropic filter)")
+	}
+	if strings.Contains(out, "kimi-k2") {
+		t.Error("output should not contain kimi-k2 (blocked by opencode-go-anthropic filter)")
+	}
+	if strings.Contains(out, "glm-4") {
+		t.Error("output should not contain glm-4 (blocked by opencode-go-anthropic filter)")
+	}
+	for _, want := range []string{"qwen-max", "minimax-max-01"} {
+		if !strings.Contains(out, want) {
+			t.Fatalf("output missing allowed model %q:\n%s", want, out)
+		}
+	}
+	if !strings.Contains(out, "2 models discovered") {
+		t.Fatalf("output should show 2 models:\n%s", out)
+	}
+}
+
 func TestRunProvidersModelsSurfacesDiscoveryError(t *testing.T) {
 	var stdout bytes.Buffer
 	var stderr bytes.Buffer

internal/cli/provider_onboarding.go

@@ -3,6 +3,7 @@ package cli
 import (
 	"fmt"
 	"io"
+	"path/filepath"
 	"strconv"
 	"strings"
 	"unicode"
@@ -306,3 +307,143 @@ func providerCommandArg(value string) string {
 	}
 	return value
 }
+
+type providerNamesOptions struct {
+	names []string
+	json  bool
+}
+
+func parseProviderNamesArgs(args []string, want int, usage string) (providerNamesOptions, bool, error) {
+	options := providerNamesOptions{}
+	for _, arg := range args {
+		switch {
+		case arg == "-h" || arg == "--help" || arg == "help":
+			return options, true, nil
+		case arg == "--json":
+			options.json = true
+		case strings.HasPrefix(arg, "-"):
+			return options, false, execUsageError{fmt.Sprintf("unknown flag %q", arg)}
+		default:
+			options.names = append(options.names, arg)
+		}
+	}
+	if len(options.names) != want {
+		return options, false, execUsageError{usage}
+	}
+	return options, false, nil
+}
+
+// runProvidersRemove deletes a saved provider profile and its stored API key.
+// The OAuth token (if any) is kept — logins outlive profiles so re-adding the
+// provider needs no new browser round-trip; `zero auth logout <name>` removes it.
+func runProvidersRemove(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) int {
+	options, help, err := parseProviderNamesArgs(args, 1, "usage: zero providers remove <name>")
+	if err != nil {
+		return writeExecUsageError(stderr, err.Error())
+	}
+	if help {
+		if err := writeProvidersHelp(stdout); err != nil {
+			return exitCrash
+		}
+		return exitSuccess
+	}
+	configPath, err := deps.userConfigPath()
+	if err != nil {
+		return writeAppError(stderr, err.Error(), exitCrash)
+	}
+	name := options.names[0]
+	cfg, err := config.RemoveProvider(configPath, name)
+	if err != nil {
+		return writeAppError(stderr, err.Error(), exitCrash)
+	}
+	// Delete the key from the store BESIDE the config being edited — the same
+	// store setup/rename write to — not the default-path store, so a
+	// non-default config path cannot leave the encrypted key behind.
+	keyRemoved, keyErr := removeStoredProviderKeyAt(configPath, name)
+	if options.json {
+		payload := map[string]any{
+			"removed":        name,
+			"keyRemoved":     keyRemoved,
+			"activeProvider": cfg.ActiveProvider,
+			"configPath":     configPath,
+		}
+		if keyErr != nil {
+			// A lingering secret must not read as a clean removal.
+			payload["keyError"] = keyErr.Error()
+		}
+		if err := writePrettyJSON(stdout, payload); err != nil {
+			return exitCrash
+		}
+		return exitSuccess
+	}
+	if _, err := fmt.Fprintf(stdout, "Removed provider %s\n", name); err != nil {
+		return exitCrash
+	}
+	if keyErr != nil {
+		if _, err := fmt.Fprintf(stderr, "warning: its stored API key could not be deleted and remains in the credential store: %v\n", keyErr); err != nil {
+			return exitCrash
+		}
+	} else if keyRemoved {
+		if _, err := fmt.Fprintln(stdout, "Deleted its stored API key."); err != nil {
+			return exitCrash
+		}
+	}
+	if active := strings.TrimSpace(cfg.ActiveProvider); active != "" {
+		if _, err := fmt.Fprintf(stdout, "Active provider: %s\n", active); err != nil {
+			return exitCrash
+		}
+	} else {
+		if _, err := fmt.Fprintln(stdout, "No providers remain — run zero setup to add one."); err != nil {
+			return exitCrash
+		}
+	}
+	return exitSuccess
+}
+
+// removeStoredProviderKeyAt deletes a provider's API key from the credential
+// store co-located with configPath (the store SecureProviderProfile captured
+// it into and RenameProvider migrates within).
+func removeStoredProviderKeyAt(configPath string, provider string) (bool, error) {
+	store, err := config.ProviderKeyStoreAt(filepath.Dir(configPath))
+	if err != nil {
+		return false, err
+	}
+	return store.Delete(provider)
+}
+
+// runProvidersRename renames a saved provider profile, migrating its stored
+// API key and the activeProvider pointer along with it (config.RenameProvider).
+func runProvidersRename(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) int {
+	options, help, err := parseProviderNamesArgs(args, 2, "usage: zero providers rename <old> <new>")
+	if err != nil {
+		return writeExecUsageError(stderr, err.Error())
+	}
+	if help {
+		if err := writeProvidersHelp(stdout); err != nil {
+			return exitCrash
+		}
+		return exitSuccess
+	}
+	configPath, err := deps.userConfigPath()
+	if err != nil {
+		return writeAppError(stderr, err.Error(), exitCrash)
+	}
+	cfg, err := config.RenameProvider(configPath, options.names[0], options.names[1])
+	if err != nil {
+		return writeAppError(stderr, err.Error(), exitCrash)
+	}
+	if options.json {
+		if err := writePrettyJSON(stdout, map[string]any{
+			"renamed":        map[string]string{"from": options.names[0], "to": options.names[1]},
+			"activeProvider": cfg.ActiveProvider,
+			"configPath":     configPath,
+		}); err != nil {
+			return exitCrash
+		}
+		return exitSuccess
+	}
+	if _, err := fmt.Fprintf(stdout, "Renamed provider %s to %s\n", options.names[0], options.names[1]); err != nil {
+		return exitCrash
+	}
+	return exitSuccess
+}

internal/cli/provider_onboarding_test.go

@@ -234,3 +234,48 @@ func writeProviderOnboardingConfig(t *testing.T, path string, cfg config.FileCon
 		t.Fatalf("write config: %v", err)
 	}
 }
+
+// TestRunProvidersRemoveDeletesKeyBesideConfig: the stored key must be deleted
+// from the credential store CO-LOCATED with the config being edited (where
+// SecureProviderProfile captured it), not the default-path store.
+func TestRunProvidersRemoveDeletesKeyBesideConfig(t *testing.T) {
+	t.Setenv("ZERO_CRED_STORAGE", "encrypted-file")
+	dir := t.TempDir()
+	configPath := filepath.Join(dir, "config.json")
+	seed := `{"activeProvider":"gw","providers":[{"name":"gw","provider_kind":"openai-compatible","baseURL":"https://gw.example.com/v1","apiKeyStored":true,"model":"m1"},{"name":"other","provider_kind":"openai-compatible","baseURL":"https://o.example.com/v1","model":"m2"}]}`
+	if err := os.WriteFile(configPath, []byte(seed), 0o600); err != nil {
+		t.Fatalf("seed config: %v", err)
+	}
+	store, err := config.ProviderKeyStoreAt(dir)
+	if err != nil {
+		t.Fatalf("open store: %v", err)
+	}
+	if err := store.Set("gw", "sk-secret"); err != nil {
+		t.Fatalf("seed key: %v", err)
+	}
+
+	var stdout, stderr bytes.Buffer
+	deps := appDeps{userConfigPath: func() (string, error) { return configPath, nil }}
+	if code := runWithDeps([]string{"providers", "remove", "gw", "--json"}, &stdout, &stderr, deps); code != exitSuccess {
+		t.Fatalf("remove failed: code=%d stderr=%s", code, stderr.String())
+	}
+
+	var payload struct {
+		Removed        string `json:"removed"`
+		KeyRemoved     bool   `json:"keyRemoved"`
+		KeyError       string `json:"keyError"`
+		ActiveProvider string `json:"activeProvider"`
+	}
+	if err := json.Unmarshal(stdout.Bytes(), &payload); err != nil {
+		t.Fatalf("decode JSON: %v\n%s", err, stdout.String())
+	}
+	if payload.Removed != "gw" || !payload.KeyRemoved || payload.KeyError != "" {
+		t.Fatalf("unexpected payload: %+v", payload)
+	}
+	if payload.ActiveProvider != "other" {
+		t.Fatalf("active must hand off, got %q", payload.ActiveProvider)
+	}
+	if _, ok, _ := store.Get("gw"); ok {
+		t.Fatalf("stored key must be deleted from the store beside the config")
+	}
+}

internal/cli/provider_setup.go

@@ -457,7 +457,9 @@ func selectProviderForCheck(resolved config.ResolvedConfig, name string) (config
 }
 
 func validateProviderRuntimeReady(profile config.ProviderProfile) error {
-	hasCredential := providerProfileHasCredential(profile)
+	// A stored OAuth login (e.g. `zero auth chatgpt`) is a credential too: a
+	// keyless token-login profile must pass the readiness check instead of being
+	// told to set an API key it will never have.
 	if profile.CatalogID != "" {
 		descriptor, err := providercatalog.Require(profile.CatalogID)
 		if err != nil {
@@ -466,23 +468,19 @@ func validateProviderRuntimeReady(profile config.ProviderProfile) error {
 		if !providercatalog.RuntimeSupported(descriptor) {
 			return fmt.Errorf("provider %q uses transport %q: %s", descriptor.ID, descriptor.Transport, providercatalog.RuntimeUnsupportedReason(descriptor))
 		}
-		if descriptor.RequiresAuth && !hasCredential {
-			apiKeyEnv := strings.TrimSpace(profile.APIKeyEnv)
-			if apiKeyEnv == "" && len(descriptor.AuthEnvVars) > 0 {
-				apiKeyEnv = descriptor.AuthEnvVars[0]
-			}
-			if apiKeyEnv != "" {
-				return fmt.Errorf("provider %s requires API key; set %s", profile.Name, apiKeyEnv)
-			}
-			return fmt.Errorf("provider %s requires API key", profile.Name)
-		}
+	}
+	if providerHasOAuthLogin(profile, oauthLoggedInProviders()) {
 		return nil
 	}
-	switch profile.ProviderKind {
-	case config.ProviderKindOpenAI, config.ProviderKindAnthropic, config.ProviderKindGoogle:
-		if !hasCredential {
-			return fmt.Errorf("provider %s requires API key", profile.Name)
+	// Delegates to the single source of truth shared with the TUI and the
+	// setup wizard, so `providers check` never disagrees with them about
+	// whether a saved provider (including a no-auth custom endpoint) is
+	// actually missing a credential.
+	if envVar, missing := profile.MissingCredentialEnv(); missing {
+		if envVar != "" {
+			return fmt.Errorf("provider %s requires API key; set %s", profile.Name, envVar)
 		}
+		return fmt.Errorf("provider %s requires API key", profile.Name)
 	}
 	return nil
 }

internal/cli/provider_setup_test.go

@@ -0,0 +1,68 @@
+package cli
+
+import (
+	"testing"
+
+	"github.com/Gitlawb/zero/internal/config"
+)
+
+// Regression for issue #555's follow-up: `zero providers check` must not
+// error that a no-auth custom endpoint requires an API key, matching what
+// /model and /providers already treat as usable.
+func TestValidateProviderRuntimeReadyCustomEndpoint(t *testing.T) {
+	cases := []struct {
+		name    string
+		profile config.ProviderProfile
+		wantErr bool
+	}{
+		{
+			name: "custom openai compatible with no credential configured",
+			profile: config.ProviderProfile{
+				Name:      "local-llama",
+				CatalogID: "custom-openai-compatible",
+				BaseURL:   "http://192.168.1.50:8080/v1",
+				Model:     "custom-model",
+			},
+			wantErr: false,
+		},
+		{
+			name: "custom openai compatible with stale legacy default env",
+			profile: config.ProviderProfile{
+				Name:      "local-llama",
+				CatalogID: "custom-openai-compatible",
+				BaseURL:   "http://192.168.1.50:8080/v1",
+				APIKeyEnv: "OPENAI_API_KEY",
+				Model:     "custom-model",
+			},
+			wantErr: false,
+		},
+		{
+			name: "custom openai compatible with explicit non-default env still requires it",
+			profile: config.ProviderProfile{
+				Name:      "local-llama",
+				CatalogID: "custom-openai-compatible",
+				BaseURL:   "http://192.168.1.50:8080/v1",
+				APIKeyEnv: "LLAMA_CPP_API_KEY",
+				Model:     "custom-model",
+			},
+			wantErr: true,
+		},
+		{
+			name: "catalog provider missing key still errors",
+			profile: config.ProviderProfile{
+				Name:      "groq",
+				CatalogID: "groq",
+				Model:     "llama-3.3-70b-versatile",
+			},
+			wantErr: true,
+		},
+	}
+	for _, c := range cases {
+		t.Run(c.name, func(t *testing.T) {
+			err := validateProviderRuntimeReady(c.profile)
+			if (err != nil) != c.wantErr {
+				t.Fatalf("validateProviderRuntimeReady() error = %v, wantErr %v", err, c.wantErr)
+			}
+		})
+	}
+}

internal/cli/scratch_files.go

@@ -0,0 +1,114 @@
+package cli
+
+import (
+	"context"
+	"os/exec"
+	"path/filepath"
+	"strconv"
+	"strings"
+	"time"
+)
+
+type scratchFileBaseline map[string]bool
+
+// scratchFileSnapshot records the untracked scratch-like files that already
+// existed before a run starts. Completion warnings compare against this baseline
+// so Zero only reports files newly left behind by the run, including files made
+// by shell commands that bypass the file tools.
+func scratchFileSnapshot(workspaceRoot string) scratchFileBaseline {
+	untracked, ok := gitUntrackedFiles(workspaceRoot)
+	if !ok {
+		return nil
+	}
+	baseline := make(scratchFileBaseline, len(untracked))
+	for _, path := range untracked {
+		if scratchLikePath(path) {
+			baseline[path] = true
+		}
+	}
+	return baseline
+}
+
+// scratchFileWarning builds a completion-time warning listing scratch/debug-like
+// files that became untracked during this run. It deliberately does not warn for
+// every new untracked file: brand-new deliverables are a normal coding-task
+// result, while names such as `_debug.py`, `_fix_test.py`, or `scratch.js` are
+// the class of throwaway artifacts called out in issue #551.
+//
+// Returns "" when there is nothing to report: git is unavailable,
+// workspaceRoot isn't a git repo, no baseline was captured, or no new
+// scratch-like untracked files remain.
+func scratchFileWarning(workspaceRoot string, baseline scratchFileBaseline) string {
+	if baseline == nil {
+		return ""
+	}
+	untracked, ok := gitUntrackedFiles(workspaceRoot)
+	if !ok {
+		return ""
+	}
+
+	var scratchFiles []string
+	for _, path := range untracked {
+		if baseline[path] || !scratchLikePath(path) {
+			continue
+		}
+		baseline[path] = true
+		scratchFiles = append(scratchFiles, path)
+	}
+	if len(scratchFiles) == 0 {
+		return ""
+	}
+
+	displayPaths := make([]string, len(scratchFiles))
+	for i, path := range scratchFiles {
+		displayPaths[i] = filepath.ToSlash(path)
+	}
+	plural := "s"
+	if len(displayPaths) == 1 {
+		plural = ""
+	}
+	return "This run left " + strconv.Itoa(len(displayPaths)) + " new scratch/debug-like file" + plural + " untracked in git: " +
+		strings.Join(displayPaths, ", ") + ". Review before committing/staging."
+}
+
+func gitUntrackedFiles(workspaceRoot string) ([]string, bool) {
+	if _, err := exec.LookPath("git"); err != nil {
+		return nil, false
+	}
+
+	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
+	defer cancel()
+	output, err := exec.CommandContext(ctx, "git", "-C", workspaceRoot, "status", "--porcelain", "-z", "--untracked-files=all").Output()
+	if err != nil {
+		// Not a git repo (or git failed for some other reason) — nothing
+		// reliable to report against, so stay silent rather than guess.
+		return nil, false
+	}
+
+	var untracked []string
+	for _, entry := range strings.Split(string(output), "\x00") {
+		if !strings.HasPrefix(entry, "?? ") {
+			continue
+		}
+		relative := strings.TrimPrefix(entry, "?? ")
+		untracked = append(untracked, relative)
+	}
+	return untracked, true
+}
+
+func scratchLikePath(path string) bool {
+	name := strings.ToLower(filepath.Base(filepath.ToSlash(path)))
+	name = strings.TrimSpace(name)
+	if name == "" {
+		return false
+	}
+	if strings.HasPrefix(name, "scratch") || strings.HasPrefix(name, "debug") || strings.HasPrefix(name, "tmp") || strings.HasPrefix(name, "temp") || strings.HasPrefix(name, "repro") {
+		return true
+	}
+	for _, marker := range []string{"scratch", "debug", "tmp", "temp", "repro", "fix_test"} {
+		if strings.Contains(name, marker) {
+			return true
+		}
+	}
+	return false
+}

internal/cli/scratch_files_test.go

@@ -0,0 +1,174 @@
+package cli
+
+import (
+	"os"
+	"os/exec"
+	"path/filepath"
+	"strings"
+	"testing"
+)
+
+func runGitForScratchTest(t *testing.T, dir string, args ...string) {
+	t.Helper()
+	if _, err := exec.LookPath("git"); err != nil {
+		t.Skip("git not available")
+	}
+	cmd := exec.Command("git", append([]string{"-C", dir}, args...)...)
+	cmd.Env = append(os.Environ(),
+		"GIT_AUTHOR_NAME=test", "[email protected]",
+		"GIT_COMMITTER_NAME=test", "[email protected]",
+	)
+	if output, err := cmd.CombinedOutput(); err != nil {
+		t.Fatalf("git %v: %v\n%s", args, err, output)
+	}
+}
+
+func TestScratchFileWarningFlagsUntrackedCreatedFile(t *testing.T) {
+	root := t.TempDir()
+	runGitForScratchTest(t, root, "init")
+	if err := os.WriteFile(filepath.Join(root, "README.md"), []byte("hi"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	runGitForScratchTest(t, root, "add", "README.md")
+	runGitForScratchTest(t, root, "commit", "-m", "init")
+	baseline := scratchFileSnapshot(root)
+
+	scratchPath := filepath.Join(root, "_debug file.py")
+	if err := os.WriteFile(scratchPath, []byte("print('debug')"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	warning := scratchFileWarning(root, baseline)
+	if warning == "" {
+		t.Fatal("expected a warning about the untracked scratch file")
+	}
+	if !strings.Contains(warning, "_debug file.py") {
+		t.Fatalf("expected warning to mention _debug file.py, got %q", warning)
+	}
+}
+
+func TestScratchFileWarningDetectsShellCreatedScratchFile(t *testing.T) {
+	root := t.TempDir()
+	runGitForScratchTest(t, root, "init")
+	if err := os.WriteFile(filepath.Join(root, "README.md"), []byte("hi"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	runGitForScratchTest(t, root, "add", "README.md")
+	runGitForScratchTest(t, root, "commit", "-m", "init")
+	baseline := scratchFileSnapshot(root)
+
+	// Simulate a shell heredoc/redirect creating a scratch file. The warning is
+	// based on the git before/after state, not only FileTracker-created paths.
+	if err := os.WriteFile(filepath.Join(root, "_fix_test.py"), []byte("print('debug')"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	warning := scratchFileWarning(root, baseline)
+	if !strings.Contains(warning, "_fix_test.py") {
+		t.Fatalf("expected warning to mention shell-created _fix_test.py, got %q", warning)
+	}
+}
+
+func TestScratchFileWarningSilentForNormalDeliverableFile(t *testing.T) {
+	root := t.TempDir()
+	runGitForScratchTest(t, root, "init")
+	if err := os.WriteFile(filepath.Join(root, "README.md"), []byte("hi"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	runGitForScratchTest(t, root, "add", "README.md")
+	runGitForScratchTest(t, root, "commit", "-m", "init")
+	baseline := scratchFileSnapshot(root)
+
+	if err := os.WriteFile(filepath.Join(root, "app.py"), []byte("print('app')"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	if warning := scratchFileWarning(root, baseline); warning != "" {
+		t.Fatalf("expected no warning for a normal new deliverable file, got %q", warning)
+	}
+}
+
+func TestScratchFileWarningSilentForUnderscoreDeliverableFile(t *testing.T) {
+	root := t.TempDir()
+	runGitForScratchTest(t, root, "init")
+	if err := os.WriteFile(filepath.Join(root, "README.md"), []byte("hi"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	runGitForScratchTest(t, root, "add", "README.md")
+	runGitForScratchTest(t, root, "commit", "-m", "init")
+	baseline := scratchFileSnapshot(root)
+
+	if err := os.WriteFile(filepath.Join(root, "_generated.go"), []byte("package main\n"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	if warning := scratchFileWarning(root, baseline); warning != "" {
+		t.Fatalf("expected no warning for an underscore-prefixed deliverable file, got %q", warning)
+	}
+}
+
+func TestScratchFileWarningSilentWhenFileIsTracked(t *testing.T) {
+	root := t.TempDir()
+	runGitForScratchTest(t, root, "init")
+	baseline := scratchFileSnapshot(root)
+	trackedPath := filepath.Join(root, "_debug.py")
+	if err := os.WriteFile(trackedPath, []byte("print('app')"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	runGitForScratchTest(t, root, "add", "_debug.py")
+	runGitForScratchTest(t, root, "commit", "-m", "init")
+
+	// A scratch-like file subsequently staged/committed by the
+	// model itself is no longer a loose scratch file, so no warning.
+	if warning := scratchFileWarning(root, baseline); warning != "" {
+		t.Fatalf("expected no warning for a tracked file, got %q", warning)
+	}
+}
+
+func TestScratchFileWarningSilentWhenNoFilesCreated(t *testing.T) {
+	root := t.TempDir()
+	runGitForScratchTest(t, root, "init")
+	baseline := scratchFileSnapshot(root)
+	if warning := scratchFileWarning(root, baseline); warning != "" {
+		t.Fatalf("expected no warning when nothing was created, got %q", warning)
+	}
+}
+
+func TestScratchFileWarningSilentWhenFileWasRemovedAgain(t *testing.T) {
+	root := t.TempDir()
+	runGitForScratchTest(t, root, "init")
+	baseline := scratchFileSnapshot(root)
+	scratchPath := filepath.Join(root, "_debug.py")
+	if err := os.WriteFile(scratchPath, []byte("print('debug')"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.Remove(scratchPath); err != nil {
+		t.Fatal(err)
+	}
+	// The model created and then cleaned up its own scratch file — nothing to warn about.
+	if warning := scratchFileWarning(root, baseline); warning != "" {
+		t.Fatalf("expected no warning for a removed file, got %q", warning)
+	}
+}
+
+func TestScratchFileWarningSilentForPreexistingScratchFile(t *testing.T) {
+	root := t.TempDir()
+	runGitForScratchTest(t, root, "init")
+	if err := os.WriteFile(filepath.Join(root, "_debug.py"), []byte("print('debug')"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	baseline := scratchFileSnapshot(root)
+
+	if warning := scratchFileWarning(root, baseline); warning != "" {
+		t.Fatalf("expected no warning for a pre-existing scratch file, got %q", warning)
+	}
+}
+
+func TestScratchFileWarningSilentWhenNotAGitRepo(t *testing.T) {
+	root := t.TempDir()
+	scratchPath := filepath.Join(root, "_debug.py")
+	if err := os.WriteFile(scratchPath, []byte("print('debug')"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	if warning := scratchFileWarning(root, scratchFileSnapshot(root)); warning != "" {
+		t.Fatalf("expected no warning outside a git repo, got %q", warning)
+	}
+}

internal/cli/setup.go

@@ -322,7 +322,9 @@ func setupRequired(resolved config.ResolvedConfig) bool {
 // the candidate set is permissive (profile name + catalog ID).
 func providerHasOAuthLogin(profile config.ProviderProfile, oauthLogins map[string]bool) bool {
 	for _, name := range profile.OAuthLoginCandidates() {
-		if oauthLogins[name] {
+		// Store keys are normalized to lower case (oauth.ProviderKey), so the
+		// presence check must compare the same way.
+		if oauthLogins[strings.ToLower(strings.TrimSpace(name))] {
 			return true
 		}
 	}
@@ -344,18 +346,19 @@ func oauthLoggedInProviders() map[string]bool {
 	}
 	for _, status := range statuses {
 		if status.HasToken {
-			out[strings.TrimPrefix(status.Key, oauth.KeyPrefixProvider)] = true
+			out[strings.ToLower(strings.TrimPrefix(status.Key, oauth.KeyPrefixProvider))] = true
 		}
 	}
 	return out
 }
 
 // firstUsableProvider returns the saved provider best suited to run without
-// onboarding: the first usable (inline credential present, or no-auth/local)
-// non-local provider, else the first usable local one. It lets the CLI fall back
-// to an already-configured login when the active provider happens to lack a
-// credential, instead of re-running onboarding every launch.
+// onboarding: the first usable (inline credential, stored OAuth login, or
+// no-auth/local) non-local provider, else the first usable local one. It lets
+// the CLI fall back to an already-configured login when the active provider
+// happens to lack a credential, instead of re-running onboarding every launch.
 func firstUsableProvider(providers []config.ProviderProfile) (config.ProviderProfile, bool) {
+	logins := oauthLoggedInProviders()
 	var localFallback config.ProviderProfile
 	haveLocal := false
 	for _, profile := range providers {
@@ -371,7 +374,11 @@ func firstUsableProvider(providers []config.ProviderProfile) (config.ProviderPro
 				continue
 			}
 		}
-		if _, missing := setupMissingCredentialEnv(profile); missing {
+		// A stored OAuth login (e.g. `zero auth login xai`) is a credential too, even
+		// when the profile has no inline key / env var — mirrors setupRequired and
+		// usableSavedProviders so this fallback doesn't force onboarding for a
+		// provider the user is already authenticated with.
+		if _, missing := setupMissingCredentialEnv(profile); missing && !providerHasOAuthLogin(profile, logins) {
 			continue
 		}
 		if providerProfileIsLocal(profile) {
@@ -496,41 +503,14 @@ func setupTryThisExample(profile config.ProviderProfile) string {
 	return example
 }
 
+// setupMissingCredentialEnv reports the environment variable a saved profile
+// is expected to read an API key from, and whether that expectation is
+// unmet. It delegates to config.ProviderProfile.MissingCredentialEnv, the
+// single source of truth shared with the TUI's provider status/wizard
+// surfaces and the CLI's `providers check` readiness gate, so all three never
+// disagree about whether a saved provider is missing a credential.
 func setupMissingCredentialEnv(profile config.ProviderProfile) (string, bool) {
-	if providerProfileHasCredential(profile) {
-		return "", false
-	}
-	if catalogID := strings.TrimSpace(profile.CatalogID); catalogID != "" {
-		descriptor, err := providercatalog.Require(catalogID)
-		if err != nil || !descriptor.RequiresAuth {
-			return "", false
-		}
-		return firstNonEmptyCLI(profile.APIKeyEnv, setupProviderEnvVar(descriptor)), true
-	}
-
-	switch normalizedSetupProviderKind(profile) {
-	case config.ProviderKindOpenAI, config.ProviderKindOpenAICompatible:
-		return firstNonEmptyCLI(profile.APIKeyEnv, "OPENAI_API_KEY"), true
-	case config.ProviderKindAnthropic, config.ProviderKindAnthropicCompat:
-		return firstNonEmptyCLI(profile.APIKeyEnv, "ANTHROPIC_API_KEY"), true
-	case config.ProviderKindGoogle:
-		return firstNonEmptyCLI(profile.APIKeyEnv, "GEMINI_API_KEY"), true
-	default:
-		if strings.TrimSpace(profile.APIKeyEnv) != "" {
-			return strings.TrimSpace(profile.APIKeyEnv), true
-		}
-		return "", false
-	}
-}
-
-func normalizedSetupProviderKind(profile config.ProviderProfile) config.ProviderKind {
-	if kind := strings.TrimSpace(string(profile.ProviderKind)); kind != "" {
-		return config.ProviderKind(strings.ToLower(kind))
-	}
-	if provider := strings.TrimSpace(profile.Provider); provider != "" {
-		return config.ProviderKind(strings.ToLower(provider))
-	}
-	return ""
+	return profile.MissingCredentialEnv()
 }
 
 func setupCheckCommand(name string) string {

internal/cli/setup_fallback_test.go

@@ -2,11 +2,14 @@ package cli
 
 import (
 	"testing"
+	"time"
 
 	"github.com/Gitlawb/zero/internal/config"
+	"github.com/Gitlawb/zero/internal/oauth"
 )
 
 func TestFirstUsableProviderPrefersRemoteKeyed(t *testing.T) {
+	withAuthStore(t)
 	providers := []config.ProviderProfile{
 		{Name: "ollama", CatalogID: "ollama", BaseURL: "http://localhost:11434/v1", APIKey: "k"},      // usable but local
 		{Name: "moonshot", CatalogID: "moonshot", BaseURL: "https://api.moonshot.ai/v1", APIKey: "k"}, // usable, remote
@@ -19,6 +22,7 @@ func TestFirstUsableProviderPrefersRemoteKeyed(t *testing.T) {
 }
 
 func TestFirstUsableProviderFallsBackToLocal(t *testing.T) {
+	withAuthStore(t)
 	providers := []config.ProviderProfile{
 		{Name: "xai", CatalogID: "xai", APIKeyEnv: "XAI_API_KEY"},                                // not usable
 		{Name: "ollama", CatalogID: "ollama", BaseURL: "http://localhost:11434/v1", APIKey: "k"}, // local, usable
@@ -30,6 +34,7 @@ func TestFirstUsableProviderFallsBackToLocal(t *testing.T) {
 }
 
 func TestFirstUsableProviderNoneUsable(t *testing.T) {
+	withAuthStore(t)
 	providers := []config.ProviderProfile{
 		{Name: "xai", CatalogID: "xai", APIKeyEnv: "XAI_API_KEY"},
 		{Name: "openai", CatalogID: "openai", APIKeyEnv: "OPENAI_API_KEY"},
@@ -42,6 +47,10 @@ func TestFirstUsableProviderNoneUsable(t *testing.T) {
 // A keyless local proxy (chatgpt-proxy, RequiresAuth=false) is usable without a
 // credential, so it can serve as a fallback rather than forcing onboarding.
 func TestFirstUsableProviderAcceptsKeylessLocalProxy(t *testing.T) {
+	// Isolate the OAuth token store: on a developer machine with a real xai
+	// login, the env-only xai profile would count as usable and win over the
+	// local proxy this test is about.
+	withAuthStore(t)
 	providers := []config.ProviderProfile{
 		{Name: "xai", CatalogID: "xai", APIKeyEnv: "XAI_API_KEY"},
 		{Name: "chatgpt", CatalogID: "chatgpt-proxy", BaseURL: "http://localhost:10531/v1"},
@@ -56,6 +65,7 @@ func TestFirstUsableProviderAcceptsKeylessLocalProxy(t *testing.T) {
 // BaseURL has no endpoint, so it must be skipped rather than selected as a
 // fallback that fails at first use. A stale CatalogID with a BaseURL still works.
 func TestFirstUsableProviderSkipsUnresolvableCatalogWithoutBaseURL(t *testing.T) {
+	withAuthStore(t)
 	providers := []config.ProviderProfile{
 		{Name: "ghost", CatalogID: "no-such-catalog-entry", APIKey: "k"}, // unusable: no endpoint
 		{Name: "custom", CatalogID: "no-such-catalog-entry", BaseURL: "https://api.custom.test/v1", APIKey: "k"},
@@ -66,6 +76,29 @@ func TestFirstUsableProviderSkipsUnresolvableCatalogWithoutBaseURL(t *testing.T)
 	}
 }
 
+// An OAuth-only provider (no inline key, no env var) must be selectable as a
+// fallback, matching setupRequired/usableSavedProviders — otherwise a fully
+// authenticated user gets forced back into onboarding when activeProvider
+// goes stale.
+func TestFirstUsableProviderRecognizesOAuthLogin(t *testing.T) {
+	withAuthStore(t)
+	store, err := oauth.NewStore(oauth.StoreOptions{})
+	if err != nil {
+		t.Fatalf("NewStore: %v", err)
+	}
+	if err := store.Save(oauth.ProviderKey("xai"), oauth.Token{AccessToken: "tok", ExpiresAt: time.Now().Add(time.Hour)}); err != nil {
+		t.Fatalf("seed token: %v", err)
+	}
+
+	providers := []config.ProviderProfile{
+		{Name: "xai", CatalogID: "xai", APIKeyEnv: "XAI_API_KEY"}, // no inline key/env, but logged in via OAuth
+	}
+	got, ok := firstUsableProvider(providers)
+	if !ok || got.Name != "xai" {
+		t.Fatalf("want OAuth-logged-in provider (xai), got %q ok=%v", got.Name, ok)
+	}
+}
+
 func TestProviderProfileIsLocal(t *testing.T) {
 	cases := []struct {
 		name    string
@@ -89,3 +122,32 @@ func TestProviderProfileIsLocal(t *testing.T) {
 		})
 	}
 }
+
+// TestFirstUsableProviderAcceptsOAuthLoginProfile: a keyless catalog profile
+// whose credential is a stored OAuth login (the shape the login flows persist)
+// must be usable during active-pointer recovery — the same rule setupRequired
+// and usableSavedProviders apply — instead of falling through to onboarding.
+func TestFirstUsableProviderAcceptsOAuthLoginProfile(t *testing.T) {
+	withAuthStore(t)
+	providers := []config.ProviderProfile{
+		{Name: "chatgpt", CatalogID: "chatgpt", BaseURL: "https://chatgpt.com/backend-api/codex", Model: "gpt-5.5"},
+	}
+
+	// No login stored: still not usable.
+	if got, ok := firstUsableProvider(providers); ok {
+		t.Fatalf("keyless profile without a login must not be usable, got %q", got.Name)
+	}
+
+	store, err := oauth.NewStore(oauth.StoreOptions{})
+	if err != nil {
+		t.Fatalf("oauth store: %v", err)
+	}
+	if err := store.Save(oauth.ProviderKey("chatgpt"), oauth.Token{AccessToken: "bearer-123"}); err != nil {
+		t.Fatalf("save token: %v", err)
+	}
+
+	got, ok := firstUsableProvider(providers)
+	if !ok || got.Name != "chatgpt" {
+		t.Fatalf("want OAuth-login profile to be usable, got %q ok=%v", got.Name, ok)
+	}
+}

internal/cli/setup_test.go

@@ -64,6 +64,54 @@ func TestSetupMissingCredentialEnv(t *testing.T) {
 			},
 			want: false,
 		},
+		{
+			// Regression for issue #555: a custom OpenAI-compatible profile saved
+			// with no credential (e.g. a local llama.cpp server with no auth) must
+			// not be reported as missing one, or it gets filtered out of /model on
+			// the next run even though it was configured exactly as intended.
+			name: "custom openai compatible with no credential configured",
+			profile: config.ProviderProfile{
+				Name:      "local-llama",
+				CatalogID: "custom-openai-compatible",
+				BaseURL:   "http://192.168.1.50:8080/v1",
+			},
+			want: false,
+		},
+		{
+			name: "custom openai compatible with explicit api key env",
+			profile: config.ProviderProfile{
+				Name:      "local-llama",
+				CatalogID: "custom-openai-compatible",
+				BaseURL:   "http://192.168.1.50:8080/v1",
+				APIKeyEnv: "LLAMA_CPP_API_KEY",
+			},
+			wantEnv: "LLAMA_CPP_API_KEY",
+			want:    true,
+		},
+		{
+			// Self-heal regression: profiles saved by the pre-fix wizard were
+			// stamped with the catalog's own guessed default (OPENAI_API_KEY)
+			// even when the endpoint needed no auth. That stale value must be
+			// treated the same as unset so already-broken saved profiles start
+			// working again without a config.json rewrite.
+			name: "custom openai compatible with stale legacy default env",
+			profile: config.ProviderProfile{
+				Name:      "local-llama",
+				CatalogID: "custom-openai-compatible",
+				BaseURL:   "http://192.168.1.50:8080/v1",
+				APIKeyEnv: "OPENAI_API_KEY",
+			},
+			want: false,
+		},
+		{
+			name: "custom anthropic compatible with no credential configured",
+			profile: config.ProviderProfile{
+				Name:      "local-proxy",
+				CatalogID: "custom-anthropic-compatible",
+				BaseURL:   "http://localhost:9000/anthropic",
+			},
+			want: false,
+		},
 	}
 
 	for _, tt := range tests {

internal/cli/trust.go

@@ -0,0 +1,119 @@
+package cli
+
+import (
+	"fmt"
+	"io"
+
+	"github.com/Gitlawb/zero/internal/redaction"
+	"github.com/Gitlawb/zero/internal/workspacetrust"
+)
+
+// runTrust implements `zero trust`, letting the user opt a workspace into running
+// its project-scoped executable config (hooks, plugins, MCP servers). Trust is
+// keyed on the exact normalized working directory, matching the exact-match trust
+// model and cwd-relative project-config discovery.
+//
+//	zero trust                trust the current working directory
+//	zero trust list           print the trusted roots, one per line
+//	zero trust remove [path]  untrust the current directory, or a named path
+func runTrust(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) int {
+	if len(args) == 0 {
+		return trustCurrentDir(stdout, stderr, deps)
+	}
+	switch args[0] {
+	case "list":
+		return trustList(stdout, stderr)
+	case "remove", "rm", "untrust":
+		return trustRemove(args[1:], stdout, stderr, deps)
+	case "-h", "--help", "help":
+		// Explicit help is a success path: write usage to stdout and exit 0, matching
+		// the other subcommands (mcp, sandbox, skills, cron, ...). Only the unknown-
+		// subcommand error path below writes usage to stderr with a usage exit code.
+		writeTrustUsage(stdout)
+		return exitSuccess
+	default:
+		if _, err := fmt.Fprintf(stderr, "zero trust: unknown subcommand %q\n\n", args[0]); err != nil {
+			return exitCrash
+		}
+		writeTrustUsage(stderr)
+		return exitUsage
+	}
+}
+
+// trustCurrentDir trusts the exact current working directory.
+func trustCurrentDir(stdout io.Writer, stderr io.Writer, deps appDeps) int {
+	cwd, err := deps.getwd()
+	if err != nil {
+		return writeAppError(stderr, redaction.ErrorMessage(fmt.Errorf("resolve workspace: %w", err), redaction.Options{}), exitCrash)
+	}
+	if err := workspacetrust.Trust(cwd); err != nil {
+		return writeAppError(stderr, redaction.ErrorMessage(err, redaction.Options{}), exitCrash)
+	}
+	if _, err := fmt.Fprintf(stdout, "Trusted %s\n", cwd); err != nil {
+		return exitCrash
+	}
+	return exitSuccess
+}
+
+// trustList prints each trusted root, one per line, or a friendly line when none
+// are trusted.
+func trustList(stdout io.Writer, stderr io.Writer) int {
+	roots, err := workspacetrust.List()
+	if err != nil {
+		return writeAppError(stderr, redaction.ErrorMessage(err, redaction.Options{}), exitCrash)
+	}
+	if len(roots) == 0 {
+		if _, err := fmt.Fprintln(stdout, "No trusted workspaces."); err != nil {
+			return exitCrash
+		}
+		return exitSuccess
+	}
+	for _, root := range roots {
+		if _, err := fmt.Fprintln(stdout, root); err != nil {
+			return exitCrash
+		}
+	}
+	return exitSuccess
+}
+
+// trustRemove untrusts the current directory, or a named path when one is given.
+// The path argument is passed to workspacetrust.Untrust verbatim, which normalizes
+// it (filepath.Abs + filepath.EvalSymlinks) the same way the store does, so a
+// relative or trailing-slash argument still matches the stored entry.
+func trustRemove(args []string, stdout io.Writer, stderr io.Writer, deps appDeps) int {
+	var target string
+	switch len(args) {
+	case 0:
+		cwd, err := deps.getwd()
+		if err != nil {
+			return writeAppError(stderr, redaction.ErrorMessage(fmt.Errorf("resolve workspace: %w", err), redaction.Options{}), exitCrash)
+		}
+		target = cwd
+	case 1:
+		target = args[0]
+	default:
+		if _, err := fmt.Fprintln(stderr, "usage: zero trust remove [path]"); err != nil {
+			return exitCrash
+		}
+		return exitUsage
+	}
+	if err := workspacetrust.Untrust(target); err != nil {
+		return writeAppError(stderr, redaction.ErrorMessage(err, redaction.Options{}), exitCrash)
+	}
+	if _, err := fmt.Fprintf(stdout, "Untrusted %s\n", target); err != nil {
+		return exitCrash
+	}
+	return exitSuccess
+}
+
+func writeTrustUsage(w io.Writer) {
+	_, _ = fmt.Fprint(w, `Usage:
+  zero trust                Trust the current working directory
+  zero trust list           List trusted workspace roots
+  zero trust remove [path]  Untrust the current directory, or a named path
+
+Trust lets Zero run a workspace's project-scoped hooks, plugins, and MCP
+servers (./.zero/hooks.json, ./.zero/plugins/, project MCP config). Trust is
+exact per directory.
+`)
+}

internal/cli/trust_e2e_test.go

@@ -0,0 +1,363 @@
+package cli
+
+import (
+	"bytes"
+	"context"
+	"os"
+	"path/filepath"
+	"runtime"
+	"strings"
+	"testing"
+
+	"github.com/Gitlawb/zero/internal/agent"
+	"github.com/Gitlawb/zero/internal/config"
+	"github.com/Gitlawb/zero/internal/tools"
+	"github.com/Gitlawb/zero/internal/workspacetrust"
+	"github.com/Gitlawb/zero/internal/worktrees"
+	"github.com/Gitlawb/zero/internal/zeroruntime"
+)
+
+// These tests close the end-to-end gap the chokepoint unit tests leave open:
+// they drive the REAL agent loop (and, for the worktree cases, the real exec
+// entry point) with a fake provider that emits a tool call, so a project
+// beforeTool hook actually fires (or is gated away) through production code, not
+// a direct dispatcher call.
+
+// markerTool is a minimal, always-allowed tool the fake provider "calls" so the
+// agent loop reaches dispatchBeforeTool (which only fires for a registered,
+// permitted tool). Its own Run is a no-op; the observable effect is the project
+// beforeTool hook the gate did or did not load.
+type markerTool struct{}
+
+func (markerTool) Name() string        { return "marker_tool" }
+func (markerTool) Description() string { return "test-only no-op tool" }
+func (markerTool) Parameters() tools.Schema {
+	return tools.Schema{Type: "object", AdditionalProperties: false}
+}
+func (markerTool) Safety() tools.Safety { return tools.Safety{Permission: tools.PermissionAllow} }
+func (markerTool) Run(context.Context, map[string]any) tools.Result {
+	return tools.Result{Status: tools.StatusOK, Output: "ok"}
+}
+
+// toolThenTextProvider calls toolName on the first turn, then answers with text so
+// the loop terminates. It detects "first turn" by the absence of a prior tool
+// result in the message history.
+type toolThenTextProvider struct{ toolName string }
+
+func (p toolThenTextProvider) StreamCompletion(_ context.Context, req zeroruntime.CompletionRequest) (<-chan zeroruntime.StreamEvent, error) {
+	toolAlreadyCalled := false
+	for _, m := range req.Messages {
+		if m.Role == zeroruntime.MessageRoleTool {
+			toolAlreadyCalled = true
+			break
+		}
+	}
+	ch := make(chan zeroruntime.StreamEvent, 8)
+	go func() {
+		defer close(ch)
+		if toolAlreadyCalled {
+			ch <- zeroruntime.StreamEvent{Type: zeroruntime.StreamEventText, Content: "done"}
+			ch <- zeroruntime.StreamEvent{Type: zeroruntime.StreamEventDone}
+			return
+		}
+		ch <- zeroruntime.StreamEvent{Type: zeroruntime.StreamEventToolCallStart, ToolCallID: "c1", ToolName: p.toolName}
+		ch <- zeroruntime.StreamEvent{Type: zeroruntime.StreamEventToolCallDelta, ToolCallID: "c1", ArgumentsFragment: `{"pattern":"*"}`}
+		ch <- zeroruntime.StreamEvent{Type: zeroruntime.StreamEventToolCallEnd, ToolCallID: "c1"}
+		ch <- zeroruntime.StreamEvent{Type: zeroruntime.StreamEventDone}
+	}()
+	return ch, nil
+}
+
+// writeMarkerHook writes a ./.zero/hooks.json under dir whose enabled beforeTool
+// hook touches markerPath when it fires. The hook shells out to /bin/sh, so the
+// trusted-case assertion (the marker must appear) is meaningless on Windows, where
+// that interpreter does not exist; skip there, matching writeMarkerHookScript.
+func writeMarkerHook(t *testing.T, dir, markerPath string) {
+	t.Helper()
+	if runtime.GOOS == "windows" {
+		t.Skip("marker hook is POSIX-shell based (/bin/sh)")
+	}
+	if err := os.MkdirAll(filepath.Join(dir, ".zero"), 0o755); err != nil {
+		t.Fatal(err)
+	}
+	body := `{"enabled":true,"hooks":[{"id":"m","event":"beforeTool","command":"/bin/sh","args":["-c","touch '` + markerPath + `'"],"enabled":true}]}`
+	if err := os.WriteFile(filepath.Join(dir, ".zero", "hooks.json"), []byte(body), 0o644); err != nil {
+		t.Fatal(err)
+	}
+}
+
+// TestTrustGateBlocksToolHookThroughAgentRun closes gap #2: the gate blocks a
+// real tool call's beforeTool hook through the production agent.Run loop, not
+// just a direct dispatcher call. Untrusted => the project hook is not in the
+// dispatcher, so the tool call fires nothing. Trusted => it fires.
+func TestTrustGateBlocksToolHookThroughAgentRun(t *testing.T) {
+	setTrustConfigRoot(t)
+	repo := t.TempDir()
+	marker := filepath.Join(t.TempDir(), "marker")
+	writeMarkerHook(t, repo, marker)
+
+	runOnce := func() {
+		reg := tools.NewRegistry()
+		reg.Register(markerTool{})
+		disp, _ := newHookDispatcherWithExtra(repo, nil, repo)
+		if _, err := agent.Run(context.Background(), "go", toolThenTextProvider{toolName: "marker_tool"}, agent.Options{
+			Registry:       reg,
+			Hooks:          disp,
+			PermissionMode: agent.PermissionModeUnsafe,
+			MaxTurns:       3,
+		}); err != nil {
+			t.Fatalf("agent.Run: %v", err)
+		}
+	}
+
+	// Untrusted: gate excludes the project layer; the beforeTool hook must not run.
+	_ = os.Remove(marker)
+	runOnce()
+	if _, err := os.Stat(marker); err == nil {
+		t.Fatal("UNTRUSTED: project beforeTool hook ran through agent.Run (marker exists) -- gate failed OPEN")
+	}
+
+	// Trusted: the hook is in the dispatcher; the tool call fires it.
+	if err := workspacetrust.Trust(repo); err != nil {
+		t.Fatal(err)
+	}
+	_ = os.Remove(marker)
+	runOnce()
+	if _, err := os.Stat(marker); err != nil {
+		t.Fatalf("TRUSTED: project beforeTool hook did NOT run (marker absent): %v", err)
+	}
+}
+
+// TestTrustGateFiresInDefaultAutoMode is the in-scope proof for the security
+// report: the beforeTool hook fires in the DEFAULT permission mode (auto), NOT
+// only under --skip-permissions-unsafe. marker_tool has PermissionAllow safety,
+// so effectivePermission grants it at loop.go without a prompt regardless of
+// mode; dispatchBeforeTool then runs the project hook. This means the
+// vulnerability is reachable in normal operation with the sandbox ON, so it is
+// not the "requires the user to disable the sandbox" out-of-scope case.
+//
+// Trusted => the hook fires in auto mode without any unsafe flag; untrusted =>
+// the gate blocks it, also in auto mode. No OnPermissionRequest is wired: an
+// auto-allowed tool needs no approval callback, which is the whole point.
+func TestTrustGateFiresInDefaultAutoMode(t *testing.T) {
+	setTrustConfigRoot(t)
+	repo := t.TempDir()
+	marker := filepath.Join(t.TempDir(), "marker")
+	writeMarkerHook(t, repo, marker)
+
+	runOnce := func(mode agent.PermissionMode) {
+		reg := tools.NewRegistry()
+		reg.Register(markerTool{})
+		disp, _ := newHookDispatcherWithExtra(repo, nil, repo)
+		if _, err := agent.Run(context.Background(), "go", toolThenTextProvider{toolName: "marker_tool"}, agent.Options{
+			Registry:       reg,
+			Hooks:          disp,
+			PermissionMode: mode,
+			MaxTurns:       3,
+		}); err != nil {
+			t.Fatalf("agent.Run (%s): %v", mode, err)
+		}
+	}
+
+	// Untrusted, default auto mode: the gate excludes the project layer, so the
+	// hook must NOT fire even though the tool call is permitted.
+	_ = os.Remove(marker)
+	runOnce(agent.PermissionModeAuto)
+	if _, err := os.Stat(marker); err == nil {
+		t.Fatal("UNTRUSTED auto mode: project beforeTool hook ran -- gate failed OPEN in the default mode")
+	}
+
+	// Trusted, default auto mode (sandbox on, NO --skip-permissions-unsafe): the
+	// auto-allowed tool is granted with no prompt and its beforeTool hook fires.
+	if err := workspacetrust.Trust(repo); err != nil {
+		t.Fatal(err)
+	}
+	_ = os.Remove(marker)
+	runOnce(agent.PermissionModeAuto)
+	if _, err := os.Stat(marker); err != nil {
+		t.Fatalf("TRUSTED auto mode: project beforeTool hook did NOT fire without an unsafe flag (marker absent): %v", err)
+	}
+
+	// And it is not an auto-only quirk: ask mode (also sandboxed, non-unsafe)
+	// fires the same way for an auto-allowed tool.
+	_ = os.Remove(marker)
+	runOnce(agent.PermissionModeAsk)
+	if _, err := os.Stat(marker); err != nil {
+		t.Fatalf("TRUSTED ask mode: project beforeTool hook did NOT fire (marker absent): %v", err)
+	}
+}
+
+// TestExecSurfacesMCPTrustNotice proves the exec path actually threads the MCP trust
+// skip into the one-line notice (the CodeRabbit finding), not just that the unit
+// pieces work in isolation: an untrusted repo whose only project config is MCP prints
+// the notice through the real runExec wiring; trusting it silences it. resolveMCPConfig
+// is stubbed to return no servers so nothing spawns -- the notice depends only on the
+// trust verdict and the real ./.zero/config.json that projectMCPConfigExists reads.
+func TestExecSurfacesMCPTrustNotice(t *testing.T) {
+	if runtime.GOOS == "windows" {
+		t.Skip("exec fake-provider harness assumes a POSIX process environment")
+	}
+	setTrustConfigRoot(t)
+	repo := t.TempDir()
+	if err := os.MkdirAll(filepath.Join(repo, ".zero"), 0o700); err != nil {
+		t.Fatal(err)
+	}
+	body := `{"mcp":{"servers":{"proj":{"type":"stdio","command":"proj-cmd"}}}}`
+	if err := os.WriteFile(filepath.Join(repo, ".zero", "config.json"), []byte(body), 0o600); err != nil {
+		t.Fatal(err)
+	}
+
+	run := func() string {
+		var out, errBuf bytes.Buffer
+		code := runWithDeps([]string{"exec", "--skip-permissions-unsafe", "--max-turns", "3", "go"}, &out, &errBuf, appDeps{
+			getwd:         func() (string, error) { return repo, nil },
+			resolveConfig: func(string, config.Overrides) (config.ResolvedConfig, error) { return execResolvedConfig(), nil },
+			newProvider: func(config.ProviderProfile) (zeroruntime.Provider, error) {
+				return toolThenTextProvider{toolName: "glob"}, nil
+			},
+			resolveMCPConfig: func(string, bool) (config.MCPConfig, error) { return config.MCPConfig{}, nil },
+		})
+		if code != exitSuccess {
+			t.Fatalf("exec exit = %d, stderr=%q", code, errBuf.String())
+		}
+		return errBuf.String()
+	}
+
+	// Untrusted: the project MCP layer is dropped, so the notice must name it. The repo
+	// has no project hooks/plugins, so the MCP skip is the ONLY thing that can fire it.
+	untrusted := run()
+	if !strings.Contains(untrusted, "MCP servers") || !strings.Contains(untrusted, "zero trust") {
+		t.Fatalf("untrusted exec must surface the MCP trust notice, stderr=%q", untrusted)
+	}
+
+	// Trusted: nothing is skipped, so no notice at all.
+	if err := workspacetrust.Trust(repo); err != nil {
+		t.Fatal(err)
+	}
+	if trusted := run(); strings.Contains(trusted, "ignoring project") {
+		t.Fatalf("trusted exec must not emit a trust notice, stderr=%q", trusted)
+	}
+}
+
+// TestExecSpecSurfacesMCPTrustNotice is the --use-spec analogue of the test above:
+// the spec-draft path (exec_spec.go) must also thread the MCP skip into its notice.
+// The fake provider never submits a spec, so the run exits non-zero; that is
+// orthogonal to trust, so we assert only the notice, not the exit code.
+func TestExecSpecSurfacesMCPTrustNotice(t *testing.T) {
+	if runtime.GOOS == "windows" {
+		t.Skip("exec fake-provider harness assumes a POSIX process environment")
+	}
+	setTrustConfigRoot(t)
+	repo := t.TempDir()
+	if err := os.MkdirAll(filepath.Join(repo, ".zero"), 0o700); err != nil {
+		t.Fatal(err)
+	}
+	body := `{"mcp":{"servers":{"proj":{"type":"stdio","command":"proj-cmd"}}}}`
+	if err := os.WriteFile(filepath.Join(repo, ".zero", "config.json"), []byte(body), 0o600); err != nil {
+		t.Fatal(err)
+	}
+
+	var out, errBuf bytes.Buffer
+	_ = runWithDeps([]string{"exec", "--use-spec", "--skip-permissions-unsafe", "--max-turns", "3", "go"}, &out, &errBuf, appDeps{
+		getwd:         func() (string, error) { return repo, nil },
+		resolveConfig: func(string, config.Overrides) (config.ResolvedConfig, error) { return execResolvedConfig(), nil },
+		newProvider: func(config.ProviderProfile) (zeroruntime.Provider, error) {
+			return toolThenTextProvider{toolName: "glob"}, nil
+		},
+		resolveMCPConfig: func(string, bool) (config.MCPConfig, error) { return config.MCPConfig{}, nil },
+	})
+	if !strings.Contains(errBuf.String(), "MCP servers") || !strings.Contains(errBuf.String(), "zero trust") {
+		t.Fatalf("untrusted --use-spec exec must surface the MCP trust notice, stderr=%q", errBuf.String())
+	}
+}
+
+// runExecTrust drives the full exec entry point with a fake worktree and a
+// tool-calling provider, returning the exit code. The provider calls the core
+// "glob" tool so dispatchBeforeTool fires inside the real exec-built registry.
+func runExecTrust(t *testing.T, extraArgs []string, launchDir, worktreeDir string) int {
+	t.Helper()
+	t.Setenv("XDG_DATA_HOME", t.TempDir())
+	args := append([]string{"exec", "--worktree", "--skip-permissions-unsafe", "--max-turns", "3"}, extraArgs...)
+	args = append(args, "go")
+	var out, errBuf bytes.Buffer
+	return runWithDeps(args, &out, &errBuf, appDeps{
+		getwd: func() (string, error) { return launchDir, nil },
+		prepareWorktree: func(context.Context, worktrees.Options) (worktrees.Result, error) {
+			return worktrees.Result{Path: worktreeDir}, nil
+		},
+		resolveConfig: func(string, config.Overrides) (config.ResolvedConfig, error) {
+			return execResolvedConfig(), nil
+		},
+		newProvider: func(config.ProviderProfile) (zeroruntime.Provider, error) {
+			return toolThenTextProvider{toolName: "glob"}, nil
+		},
+	})
+}
+
+// TestExecWorktreeInheritsTrustEndToEnd closes gap #1 for the exec path: trust is
+// keyed on the ORIGINAL launch dir, not the generated worktree path. The worktree
+// checkout carries the committed .zero/hooks.json; the gate must load it only when
+// the SOURCE repo (launch dir) is trusted, proving exec.go captures trustRoot
+// before the --worktree reassignment and threads it into the chokepoint.
+func TestExecWorktreeInheritsTrustEndToEnd(t *testing.T) {
+	setTrustConfigRoot(t)
+	repo := t.TempDir()     // original launch dir -- the trust key
+	worktree := t.TempDir() // simulated worktree checkout (workspaceRoot after reassignment)
+	marker := filepath.Join(t.TempDir(), "marker")
+	writeMarkerHook(t, worktree, marker) // the checkout carries the committed hook
+
+	// Untrusted source repo: worktree hooks must NOT run.
+	_ = os.Remove(marker)
+	if code := runExecTrust(t, nil, repo, worktree); code != exitSuccess {
+		t.Fatalf("exec --worktree (untrusted) exit = %d", code)
+	}
+	if _, err := os.Stat(marker); err == nil {
+		t.Fatal("UNTRUSTED worktree: project hook ran -- exec keyed trust on the worktree path or failed open")
+	}
+
+	// Trusted source repo: the worktree inherits its trust, so the hook runs.
+	if err := workspacetrust.Trust(repo); err != nil {
+		t.Fatal(err)
+	}
+	_ = os.Remove(marker)
+	if code := runExecTrust(t, nil, repo, worktree); code != exitSuccess {
+		t.Fatalf("exec --worktree (trusted) exit = %d", code)
+	}
+	if _, err := os.Stat(marker); err != nil {
+		t.Fatalf("TRUSTED worktree: project hook did NOT run (marker absent) -- worktree trust inheritance broken: %v", err)
+	}
+}
+
+// TestExecSpecWorktreeInheritsTrustEndToEnd closes gap #1 for the spec-draft path
+// (the exact --use-spec --worktree combination the review fix addressed): the
+// spec-draft chokepoint must also key trust on the original launch dir.
+func TestExecSpecWorktreeInheritsTrustEndToEnd(t *testing.T) {
+	setTrustConfigRoot(t)
+	repo := t.TempDir()
+	worktree := t.TempDir()
+	marker := filepath.Join(t.TempDir(), "marker")
+	writeMarkerHook(t, worktree, marker)
+
+	// The spec-draft flow itself exits non-zero here (the fake provider does not
+	// submit a real spec), which is orthogonal to trust: the hook dispatcher is
+	// built (keyed on run.trustRoot) before the agent runs, and glob (a read-only
+	// allow tool) is advertised in spec-draft, so beforeTool still fires. We assert
+	// only the marker, the trust behavior, not the spec-flow exit code.
+
+	// Untrusted: spec-draft in a worktree of an untrusted repo runs no project hook.
+	_ = os.Remove(marker)
+	_ = runExecTrust(t, []string{"--use-spec"}, repo, worktree)
+	if _, err := os.Stat(marker); err == nil {
+		t.Fatal("UNTRUSTED spec-draft worktree: project hook ran -- spec-draft keyed trust on the worktree path")
+	}
+
+	// Trusted: the spec-draft path inherits the source repo's trust.
+	if err := workspacetrust.Trust(repo); err != nil {
+		t.Fatal(err)
+	}
+	_ = os.Remove(marker)
+	_ = runExecTrust(t, []string{"--use-spec"}, repo, worktree)
+	if _, err := os.Stat(marker); err != nil {
+		t.Fatalf("TRUSTED spec-draft worktree: project hook did NOT run -- spec-draft trust inheritance broken: %v", err)
+	}
+}

internal/cli/trust_gate_test.go

@@ -0,0 +1,466 @@
+package cli
+
+import (
+	"bytes"
+	"context"
+	"os"
+	"path/filepath"
+	"runtime"
+	"testing"
+
+	"github.com/Gitlawb/zero/internal/hooks"
+	"github.com/Gitlawb/zero/internal/plugins"
+	"github.com/Gitlawb/zero/internal/tools"
+	"github.com/Gitlawb/zero/internal/workspacetrust"
+)
+
+// setTrustConfigRoot redirects both the workspace-trust store and the user-level
+// hooks/plugins config to a fresh temp dir with a GOOS-aware env switch, mirroring
+// setUserConfigRoot in internal/config/paths_test.go, and returns that dir so tests
+// can build the exact paths the code resolves to.
+//
+// XDG_CONFIG_HOME is the lever on macOS as well as Linux: config.UserConfigDir is
+// XDG-first there (it only falls back to $HOME/.config when XDG_CONFIG_HOME is
+// unset), and the hooks loader resolves its user layer from XDG_CONFIG_HOME too. An
+// earlier version set HOME on darwin and returned the raw root; that left the store
+// at <root>/.config/zero while the tests wrote to <root>/zero, so the store-error
+// and user-hook cases silently missed. Only Windows needs its own variable (APPDATA,
+// what os.UserConfigDir consults). XDG_DATA_HOME is redirected so the hook audit
+// store never touches the user's real data dir.
+func setTrustConfigRoot(t *testing.T) string {
+	t.Helper()
+	root := t.TempDir()
+	switch runtime.GOOS {
+	case "windows":
+		t.Setenv("APPDATA", root)
+	default:
+		t.Setenv("XDG_CONFIG_HOME", root)
+	}
+	t.Setenv("XDG_DATA_HOME", t.TempDir())
+	return root
+}
+
+// writeMarkerHookScript writes an executable script that creates markerPath when
+// run, and returns its absolute path. The hook dispatcher runs commands directly
+// (no shell), so a real executable is needed to observe execution.
+func writeMarkerHookScript(t *testing.T, markerPath string) string {
+	t.Helper()
+	if runtime.GOOS == "windows" {
+		t.Skip("marker-hook script is POSIX-shell based")
+	}
+	dir := t.TempDir()
+	script := filepath.Join(dir, "mark.sh")
+	body := "#!/bin/sh\n: > " + shellQuote(markerPath) + "\n"
+	if err := os.WriteFile(script, []byte(body), 0o700); err != nil {
+		t.Fatalf("write marker hook script: %v", err)
+	}
+	return script
+}
+
+func shellQuote(path string) string {
+	return "'" + path + "'"
+}
+
+// writeProjectHooks writes a project ./.zero/hooks.json under repo with the given
+// hooks and top-level enabled flag.
+func writeProjectHooks(t *testing.T, repo string, enabled bool, defs ...hooks.Definition) {
+	t.Helper()
+	path := filepath.Join(repo, ".zero", "hooks.json")
+	if err := os.MkdirAll(filepath.Dir(path), 0o700); err != nil {
+		t.Fatalf("mkdir project .zero: %v", err)
+	}
+	if err := hooks.WriteConfig(path, hooks.Config{Enabled: enabled, Hooks: defs}); err != nil {
+		t.Fatalf("write project hooks.json: %v", err)
+	}
+}
+
+// writeUserHooks writes a user-level hooks.json into the redirected config root.
+func writeUserHooks(t *testing.T, configRoot string, enabled bool, defs ...hooks.Definition) {
+	t.Helper()
+	path := filepath.Join(configRoot, "zero", "hooks.json")
+	if err := os.MkdirAll(filepath.Dir(path), 0o700); err != nil {
+		t.Fatalf("mkdir user config dir: %v", err)
+	}
+	if err := hooks.WriteConfig(path, hooks.Config{Enabled: enabled, Hooks: defs}); err != nil {
+		t.Fatalf("write user hooks.json: %v", err)
+	}
+}
+
+func dispatchBeforeTool(dispatcher *hooks.Dispatcher) hooks.DispatchOutcome {
+	return dispatcher.Dispatch(context.Background(), hooks.DispatchInput{
+		Event:    hooks.EventBeforeTool,
+		ToolName: "bash",
+	})
+}
+
+func fileExists(path string) bool {
+	_, err := os.Stat(path)
+	return err == nil
+}
+
+// TestHookGateUntrustedExcludesProjectLayer proves R1: an untrusted trustRoot with
+// an enabled project beforeTool hook runs nothing and writes no marker.
+func TestHookGateUntrustedExcludesProjectLayer(t *testing.T) {
+	setTrustConfigRoot(t)
+	repo := t.TempDir()
+	marker := filepath.Join(t.TempDir(), "ran")
+	script := writeMarkerHookScript(t, marker)
+	writeProjectHooks(t, repo, true, hooks.Definition{
+		ID: "proj.mark", Event: hooks.EventBeforeTool, Command: script, Enabled: true,
+	})
+
+	dispatcher, skip := newHookDispatcherWithExtra(repo, nil, repo)
+	outcome := dispatchBeforeTool(dispatcher)
+
+	if outcome.Ran != 0 {
+		t.Fatalf("untrusted workspace should run 0 project hooks, Ran=%d", outcome.Ran)
+	}
+	if fileExists(marker) {
+		t.Fatalf("untrusted workspace must not execute the project hook (marker exists)")
+	}
+	if !skip.excludedProjectConfig {
+		t.Fatalf("skip report should flag the excluded project hooks file")
+	}
+	if skip.trustCheckErrored {
+		t.Fatalf("a clean untrusted verdict is not a store-read error")
+	}
+}
+
+// TestHookGateTrustedRunsProjectHook proves R3: after Trust(repo) the project hook
+// runs and the marker appears.
+func TestHookGateTrustedRunsProjectHook(t *testing.T) {
+	setTrustConfigRoot(t)
+	repo := t.TempDir()
+	marker := filepath.Join(t.TempDir(), "ran")
+	script := writeMarkerHookScript(t, marker)
+	writeProjectHooks(t, repo, true, hooks.Definition{
+		ID: "proj.mark", Event: hooks.EventBeforeTool, Command: script, Enabled: true,
+	})
+
+	if err := workspacetrust.Trust(repo); err != nil {
+		t.Fatalf("Trust(repo): %v", err)
+	}
+
+	dispatcher, skip := newHookDispatcherWithExtra(repo, nil, repo)
+	outcome := dispatchBeforeTool(dispatcher)
+
+	if outcome.Ran != 1 {
+		t.Fatalf("trusted workspace should run the project hook, Ran=%d", outcome.Ran)
+	}
+	if !fileExists(marker) {
+		t.Fatalf("trusted workspace must execute the project hook (marker missing)")
+	}
+	if skip.excludedProjectConfig {
+		t.Fatalf("a trusted workspace must not report the project layer as excluded")
+	}
+}
+
+// TestHookGateEmptyTrustRootFailsClosed proves the fail-closed-by-construction
+// guard: a caller that forgot to resolve trustRoot (empty) still excludes the
+// project layer, so no marker appears.
+func TestHookGateEmptyTrustRootFailsClosed(t *testing.T) {
+	setTrustConfigRoot(t)
+	repo := t.TempDir()
+	marker := filepath.Join(t.TempDir(), "ran")
+	script := writeMarkerHookScript(t, marker)
+	writeProjectHooks(t, repo, true, hooks.Definition{
+		ID: "proj.mark", Event: hooks.EventBeforeTool, Command: script, Enabled: true,
+	})
+	// Even trusting the repo must not help when the caller passes an empty root.
+	if err := workspacetrust.Trust(repo); err != nil {
+		t.Fatalf("Trust(repo): %v", err)
+	}
+
+	dispatcher, skip := newHookDispatcherWithExtra(repo, nil, "")
+	outcome := dispatchBeforeTool(dispatcher)
+
+	if outcome.Ran != 0 {
+		t.Fatalf("empty trustRoot must fail closed, Ran=%d", outcome.Ran)
+	}
+	if fileExists(marker) {
+		t.Fatalf("empty trustRoot must not execute the project hook (marker exists)")
+	}
+	if !skip.excludedProjectConfig {
+		t.Fatalf("empty trustRoot should report the project layer excluded")
+	}
+}
+
+// TestHookGateFailClosedOnStoreError proves R5: a real IsTrusted error (trust.json
+// created as a directory, per U1) is treated as untrusted, the project hook does
+// not run, and the skip report marks the store-read error.
+func TestHookGateFailClosedOnStoreError(t *testing.T) {
+	configRoot := setTrustConfigRoot(t)
+	// Create the trust store path as a DIRECTORY so os.ReadFile fails with a
+	// non-ErrNotExist error, forcing IsTrusted to return (false, non-nil).
+	trustPath := filepath.Join(configRoot, "zero", "trust.json")
+	if err := os.MkdirAll(trustPath, 0o700); err != nil {
+		t.Fatalf("create trust.json as a directory: %v", err)
+	}
+
+	repo := t.TempDir()
+	marker := filepath.Join(t.TempDir(), "ran")
+	script := writeMarkerHookScript(t, marker)
+	writeProjectHooks(t, repo, true, hooks.Definition{
+		ID: "proj.mark", Event: hooks.EventBeforeTool, Command: script, Enabled: true,
+	})
+
+	dispatcher, skip := newHookDispatcherWithExtra(repo, nil, repo)
+	outcome := dispatchBeforeTool(dispatcher)
+
+	if outcome.Ran != 0 {
+		t.Fatalf("a store-read error must fail closed, Ran=%d", outcome.Ran)
+	}
+	if fileExists(marker) {
+		t.Fatalf("a store-read error must not execute the project hook (marker exists)")
+	}
+	if !skip.trustCheckErrored {
+		t.Fatalf("skip report must mark the store-read error")
+	}
+	if !skip.excludedProjectConfig {
+		t.Fatalf("the project layer must be reported excluded on the error path")
+	}
+}
+
+// TestHookGateUserHookStillRunsWhenProjectExcluded proves R4/R10: with the project
+// layer excluded, a user-level hook still fires, and a project hooks.json that sets
+// global enabled:false or defines a same-ID hook cannot disable or override it.
+func TestHookGateUserHookStillRunsWhenProjectExcluded(t *testing.T) {
+	configRoot := setTrustConfigRoot(t)
+	repo := t.TempDir()
+	marker := filepath.Join(t.TempDir(), "ran")
+	script := writeMarkerHookScript(t, marker)
+
+	// User hook that writes the marker.
+	writeUserHooks(t, configRoot, true, hooks.Definition{
+		ID: "shared.mark", Event: hooks.EventBeforeTool, Command: script, Enabled: true,
+	})
+	// Project config that would, if loaded, disable the whole surface (enabled:false)
+	// AND override the user hook by ID with a disabled no-op. The gate drops the
+	// whole project layer, so neither takes effect.
+	writeProjectHooks(t, repo, false, hooks.Definition{
+		ID: "shared.mark", Event: hooks.EventBeforeTool, Command: "true", Enabled: false,
+	})
+
+	// Untrusted: the project layer is excluded.
+	dispatcher, _ := newHookDispatcherWithExtra(repo, nil, repo)
+	outcome := dispatchBeforeTool(dispatcher)
+
+	if outcome.Ran != 1 {
+		t.Fatalf("user hook must still run when the project layer is excluded, Ran=%d", outcome.Ran)
+	}
+	if !fileExists(marker) {
+		t.Fatalf("user hook must fire (marker missing); project config must not disable/override it")
+	}
+}
+
+// TestHookGateWorktreeInheritsTrust proves the worktree case: with repo trusted,
+// calling the chokepoint with a different workspaceRoot (the worktree path) but
+// trustRoot=repo loads the project hooks, so trust keys on trustRoot not
+// workspaceRoot.
+func TestHookGateWorktreeInheritsTrust(t *testing.T) {
+	setTrustConfigRoot(t)
+	repo := t.TempDir()
+	worktree := t.TempDir() // a distinct, never-trusted path
+	marker := filepath.Join(t.TempDir(), "ran")
+	script := writeMarkerHookScript(t, marker)
+	// The project hooks live under the worktree (that is where config is read from).
+	writeProjectHooks(t, worktree, true, hooks.Definition{
+		ID: "proj.mark", Event: hooks.EventBeforeTool, Command: script, Enabled: true,
+	})
+
+	if err := workspacetrust.Trust(repo); err != nil {
+		t.Fatalf("Trust(repo): %v", err)
+	}
+
+	// workspaceRoot is the worktree, but trustRoot is the original repo.
+	dispatcher, skip := newHookDispatcherWithExtra(worktree, nil, repo)
+	outcome := dispatchBeforeTool(dispatcher)
+
+	if outcome.Ran != 1 {
+		t.Fatalf("a worktree of a trusted repo should load project hooks, Ran=%d", outcome.Ran)
+	}
+	if !fileExists(marker) {
+		t.Fatalf("worktree run should execute the project hook (marker missing)")
+	}
+	if skip.excludedProjectConfig {
+		t.Fatalf("worktree of a trusted repo must not report the project layer excluded")
+	}
+}
+
+// TestPluginGateUntrustedExcludesProject proves R2/R5: an untrusted trustRoot makes
+// activatePlugins load with ExcludeProject=true and report the skip.
+func TestPluginGateUntrustedExcludesProject(t *testing.T) {
+	setTrustConfigRoot(t)
+	repo := t.TempDir()
+	// A ./.zero/plugins dir so the skip report flags an excluded project config.
+	if err := os.MkdirAll(filepath.Join(repo, ".zero", "plugins"), 0o700); err != nil {
+		t.Fatalf("mkdir project plugins dir: %v", err)
+	}
+
+	var gotExclude bool
+	deps := appDeps{
+		loadPlugins: func(opts plugins.LoadOptions) (plugins.LoadResult, error) {
+			gotExclude = opts.ExcludeProject
+			return plugins.LoadResult{}, nil
+		},
+		skillsDir: func() string { return t.TempDir() },
+	}
+	registry := tools.NewRegistry()
+	var stderr bytes.Buffer
+	activation := activatePlugins(repo, registry, deps, &stderr, repo)
+
+	if !gotExclude {
+		t.Fatalf("untrusted workspace must pass ExcludeProject=true to loadPlugins")
+	}
+	if !activation.excludedProjectConfig {
+		t.Fatalf("skip report should flag the excluded ./.zero/plugins dir")
+	}
+	if activation.trustCheckErrored {
+		t.Fatalf("a clean untrusted verdict is not a store-read error")
+	}
+}
+
+// TestPluginGateTrustedIncludesProject proves R3: after Trust(repo), activatePlugins
+// loads with ExcludeProject=false.
+func TestPluginGateTrustedIncludesProject(t *testing.T) {
+	setTrustConfigRoot(t)
+	repo := t.TempDir()
+	if err := workspacetrust.Trust(repo); err != nil {
+		t.Fatalf("Trust(repo): %v", err)
+	}
+
+	var gotExclude bool
+	deps := appDeps{
+		loadPlugins: func(opts plugins.LoadOptions) (plugins.LoadResult, error) {
+			gotExclude = opts.ExcludeProject
+			return plugins.LoadResult{}, nil
+		},
+		skillsDir: func() string { return t.TempDir() },
+	}
+	registry := tools.NewRegistry()
+	var stderr bytes.Buffer
+	activation := activatePlugins(repo, registry, deps, &stderr, repo)
+
+	if gotExclude {
+		t.Fatalf("trusted workspace must pass ExcludeProject=false to loadPlugins")
+	}
+	if activation.excludedProjectConfig {
+		t.Fatalf("trusted workspace must not report the project config excluded")
+	}
+}
+
+// TestPluginGateFailClosedOnStoreError proves R5 for plugins: a store-read error
+// forces ExcludeProject=true and marks the error.
+func TestPluginGateFailClosedOnStoreError(t *testing.T) {
+	configRoot := setTrustConfigRoot(t)
+	trustPath := filepath.Join(configRoot, "zero", "trust.json")
+	if err := os.MkdirAll(trustPath, 0o700); err != nil {
+		t.Fatalf("create trust.json as a directory: %v", err)
+	}
+	repo := t.TempDir()
+	if err := os.MkdirAll(filepath.Join(repo, ".zero", "plugins"), 0o700); err != nil {
+		t.Fatalf("mkdir project plugins dir: %v", err)
+	}
+
+	var gotExclude bool
+	deps := appDeps{
+		loadPlugins: func(opts plugins.LoadOptions) (plugins.LoadResult, error) {
+			gotExclude = opts.ExcludeProject
+			return plugins.LoadResult{}, nil
+		},
+		skillsDir: func() string { return t.TempDir() },
+	}
+	registry := tools.NewRegistry()
+	var stderr bytes.Buffer
+	activation := activatePlugins(repo, registry, deps, &stderr, repo)
+
+	if !gotExclude {
+		t.Fatalf("a store-read error must fail closed (ExcludeProject=true)")
+	}
+	if !activation.trustCheckErrored {
+		t.Fatalf("skip report must mark the store-read error")
+	}
+}
+
+// TestEmitTrustNoticeOneLineWhenSkipped proves R6: exactly one notice when either
+// surface skips project config, none when trusted, and distinct error text on a
+// store-read failure.
+func TestEmitTrustNoticeOneLineWhenSkipped(t *testing.T) {
+	t.Run("both surfaces skipped yields one line", func(t *testing.T) {
+		var buf bytes.Buffer
+		emitTrustNotice(&buf,
+			trustSkip{excludedProjectConfig: true},
+			trustSkip{excludedProjectConfig: true})
+		lines := nonEmptyLines(buf.String())
+		if len(lines) != 1 {
+			t.Fatalf("expected exactly one notice line, got %d: %q", len(lines), buf.String())
+		}
+		if !bytes.Contains(buf.Bytes(), []byte("zero trust")) {
+			t.Fatalf("notice should point at 'zero trust', got %q", buf.String())
+		}
+	})
+
+	t.Run("trusted yields no notice", func(t *testing.T) {
+		var buf bytes.Buffer
+		emitTrustNotice(&buf, trustSkip{}, trustSkip{})
+		if buf.Len() != 0 {
+			t.Fatalf("a trusted session must emit no notice, got %q", buf.String())
+		}
+	})
+
+	t.Run("store-read error uses distinct text", func(t *testing.T) {
+		var buf bytes.Buffer
+		emitTrustNotice(&buf,
+			trustSkip{excludedProjectConfig: true, trustCheckErrored: true},
+			trustSkip{})
+		lines := nonEmptyLines(buf.String())
+		if len(lines) != 1 {
+			t.Fatalf("expected exactly one notice line, got %d: %q", len(lines), buf.String())
+		}
+		if !bytes.Contains(buf.Bytes(), []byte("could not be read")) {
+			t.Fatalf("error-path notice should name the store read failure, got %q", buf.String())
+		}
+	})
+
+	t.Run("only one surface skipped still yields one line", func(t *testing.T) {
+		var buf bytes.Buffer
+		emitTrustNotice(&buf, trustSkip{excludedProjectConfig: true}, trustSkip{})
+		lines := nonEmptyLines(buf.String())
+		if len(lines) != 1 {
+			t.Fatalf("expected exactly one notice line, got %d: %q", len(lines), buf.String())
+		}
+	})
+
+	t.Run("mcp skip alone yields one line naming mcp", func(t *testing.T) {
+		var buf bytes.Buffer
+		// Hooks and plugins clean; only the MCP surface (third arg) dropped project config.
+		emitTrustNotice(&buf, trustSkip{}, trustSkip{}, trustSkip{excludedProjectConfig: true})
+		lines := nonEmptyLines(buf.String())
+		if len(lines) != 1 {
+			t.Fatalf("expected exactly one notice line, got %d: %q", len(lines), buf.String())
+		}
+		if !bytes.Contains(buf.Bytes(), []byte("MCP")) {
+			t.Fatalf("notice should name MCP when the MCP surface is skipped, got %q", buf.String())
+		}
+	})
+
+	t.Run("store error with nothing excluded yields no notice", func(t *testing.T) {
+		// trustCheckErrored but no surface had project config to skip (excludedProjectConfig
+		// false): there is nothing to warn about, so the excluded-gate wins over the error.
+		var buf bytes.Buffer
+		emitTrustNotice(&buf, trustSkip{trustCheckErrored: true}, trustSkip{trustCheckErrored: true})
+		if buf.Len() != 0 {
+			t.Fatalf("a store error with no skipped project config must emit nothing, got %q", buf.String())
+		}
+	})
+}
+
+func nonEmptyLines(s string) []string {
+	out := []string{}
+	for _, line := range bytes.Split([]byte(s), []byte("\n")) {
+		if len(bytes.TrimSpace(line)) > 0 {
+			out = append(out, string(line))
+		}
+	}
+	return out
+}

internal/cli/trust_test.go

@@ -0,0 +1,227 @@
+package cli
+
+import (
+	"bytes"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"github.com/Gitlawb/zero/internal/workspacetrust"
+)
+
+// trustDeps builds a test appDeps whose getwd returns a fixed directory, so the
+// trust command can be exercised without touching the real cwd.
+func trustDeps(cwd string) appDeps {
+	return appDeps{getwd: func() (string, error) { return cwd, nil }}
+}
+
+// TestRunTrustAddThenList proves `trust` trusts the cwd and `trust list` shows it.
+func TestRunTrustAddThenList(t *testing.T) {
+	setTrustConfigRoot(t)
+	cwd := t.TempDir()
+	deps := trustDeps(cwd)
+
+	var out, errBuf bytes.Buffer
+	if code := runTrust(nil, &out, &errBuf, deps); code != exitSuccess {
+		t.Fatalf("trust returned %d, want %d; stderr=%q", code, exitSuccess, errBuf.String())
+	}
+	if !strings.Contains(out.String(), "Trusted "+cwd) {
+		t.Fatalf("trust stdout = %q, want it to contain %q", out.String(), "Trusted "+cwd)
+	}
+
+	trusted, err := workspacetrust.IsTrusted(cwd)
+	if err != nil {
+		t.Fatalf("IsTrusted after trust: %v", err)
+	}
+	if !trusted {
+		t.Fatalf("cwd should be trusted after `trust`")
+	}
+
+	// The store canonicalizes with filepath.EvalSymlinks, so `list` prints the
+	// resolved path. On platforms where the temp dir is a symlink (macOS /var) or a
+	// Windows 8.3 short name (RUNNER~1 -> runneradmin), that differs from the raw
+	// cwd; compare against the same canonical form the store records.
+	wantListed := cwd
+	if resolved, err := filepath.EvalSymlinks(cwd); err == nil {
+		wantListed = resolved
+	}
+
+	out.Reset()
+	errBuf.Reset()
+	if code := runTrust([]string{"list"}, &out, &errBuf, deps); code != exitSuccess {
+		t.Fatalf("trust list returned %d, want %d; stderr=%q", code, exitSuccess, errBuf.String())
+	}
+	if !strings.Contains(out.String(), wantListed) {
+		t.Fatalf("trust list stdout = %q, want it to contain the trusted root %q", out.String(), wantListed)
+	}
+}
+
+// TestRunTrustListShowsCanonicalPath executes the canonical-path behavior that broke
+// TestRunTrustAddThenList on the Windows runner (an 8.3 short name expanded by
+// EvalSymlinks). It uses a symlinked cwd -- the portable, Linux-runnable analogue of
+// that divergence -- and proves `trust list` prints the resolved path the store
+// records, never the raw pre-resolution path.
+func TestRunTrustListShowsCanonicalPath(t *testing.T) {
+	setTrustConfigRoot(t)
+	realDir := t.TempDir()
+	link := filepath.Join(t.TempDir(), "link")
+	if err := os.Symlink(realDir, link); err != nil {
+		t.Skipf("symlinks unavailable on this platform: %v", err)
+	}
+	resolved, err := filepath.EvalSymlinks(link)
+	if err != nil {
+		t.Fatalf("EvalSymlinks(link): %v", err)
+	}
+	if resolved == link {
+		t.Skip("symlink did not change the path on this platform")
+	}
+
+	deps := trustDeps(link)
+	if code := runTrust(nil, &bytes.Buffer{}, &bytes.Buffer{}, deps); code != exitSuccess {
+		t.Fatalf("trust returned %d", code)
+	}
+
+	var out, errBuf bytes.Buffer
+	if code := runTrust([]string{"list"}, &out, &errBuf, deps); code != exitSuccess {
+		t.Fatalf("trust list returned %d; stderr=%q", code, errBuf.String())
+	}
+	if !strings.Contains(out.String(), resolved) {
+		t.Fatalf("trust list = %q, want the canonical path %q", out.String(), resolved)
+	}
+	if strings.Contains(out.String(), link) {
+		t.Fatalf("trust list = %q, must not contain the raw pre-resolution path %q", out.String(), link)
+	}
+}
+
+// TestRunTrustRemoveCurrentDir proves `trust remove` (no arg) untrusts the cwd and
+// `trust list` then reports none.
+func TestRunTrustRemoveCurrentDir(t *testing.T) {
+	setTrustConfigRoot(t)
+	cwd := t.TempDir()
+	deps := trustDeps(cwd)
+
+	if code := runTrust(nil, &bytes.Buffer{}, &bytes.Buffer{}, deps); code != exitSuccess {
+		t.Fatalf("trust setup returned %d", code)
+	}
+
+	var out, errBuf bytes.Buffer
+	if code := runTrust([]string{"remove"}, &out, &errBuf, deps); code != exitSuccess {
+		t.Fatalf("trust remove returned %d, want %d; stderr=%q", code, exitSuccess, errBuf.String())
+	}
+	if !strings.Contains(out.String(), "Untrusted "+cwd) {
+		t.Fatalf("trust remove stdout = %q, want it to contain %q", out.String(), "Untrusted "+cwd)
+	}
+
+	trusted, err := workspacetrust.IsTrusted(cwd)
+	if err != nil {
+		t.Fatalf("IsTrusted after remove: %v", err)
+	}
+	if trusted {
+		t.Fatalf("cwd should not be trusted after `trust remove`")
+	}
+
+	out.Reset()
+	errBuf.Reset()
+	if code := runTrust([]string{"list"}, &out, &errBuf, deps); code != exitSuccess {
+		t.Fatalf("trust list returned %d, want %d", code, exitSuccess)
+	}
+	if !strings.Contains(out.String(), "No trusted workspaces.") {
+		t.Fatalf("trust list stdout = %q, want %q for an empty store", out.String(), "No trusted workspaces.")
+	}
+}
+
+// TestRunTrustRemoveNamedPathNonCanonical proves `trust remove <path>` untrusts a
+// specific path, and that a non-canonical argument (a trailing slash) still matches
+// the stored normalized entry.
+func TestRunTrustRemoveNamedPathNonCanonical(t *testing.T) {
+	setTrustConfigRoot(t)
+	// The cwd is a different directory than the one we trust-and-remove, so this
+	// exercises the named-path branch, not the bare-cwd branch.
+	cwd := t.TempDir()
+	target := t.TempDir()
+	deps := trustDeps(cwd)
+
+	if err := workspacetrust.Trust(target); err != nil {
+		t.Fatalf("Trust(target): %v", err)
+	}
+	trusted, err := workspacetrust.IsTrusted(target)
+	if err != nil || !trusted {
+		t.Fatalf("target should be trusted before remove (trusted=%v err=%v)", trusted, err)
+	}
+
+	// Pass the target with a trailing slash: normalization must still match the
+	// stored canonical entry.
+	var out, errBuf bytes.Buffer
+	if code := runTrust([]string{"remove", target + "/"}, &out, &errBuf, deps); code != exitSuccess {
+		t.Fatalf("trust remove <path> returned %d, want %d; stderr=%q", code, exitSuccess, errBuf.String())
+	}
+
+	trusted, err = workspacetrust.IsTrusted(target)
+	if err != nil {
+		t.Fatalf("IsTrusted after named remove: %v", err)
+	}
+	if trusted {
+		t.Fatalf("target should be untrusted after `trust remove <path>/` despite the trailing slash")
+	}
+}
+
+// TestRunTrustUnknownSubcommand proves an unknown subcommand returns exit code 2 and
+// writes usage to stderr (not stdout).
+func TestRunTrustUnknownSubcommand(t *testing.T) {
+	setTrustConfigRoot(t)
+	deps := trustDeps(t.TempDir())
+
+	var out, errBuf bytes.Buffer
+	if code := runTrust([]string{"bogus"}, &out, &errBuf, deps); code != exitUsage {
+		t.Fatalf("unknown subcommand returned %d, want %d", code, exitUsage)
+	}
+	if errBuf.Len() == 0 {
+		t.Fatalf("unknown subcommand should write usage to stderr, stderr was empty")
+	}
+	if !strings.Contains(errBuf.String(), "trust") {
+		t.Fatalf("stderr usage = %q, want it to mention `trust`", errBuf.String())
+	}
+	if out.Len() != 0 {
+		t.Fatalf("unknown subcommand should not write to stdout, got %q", out.String())
+	}
+}
+
+// TestRunTrustRemoveTooManyArgs proves `trust remove a b` (more than one path) is a
+// usage error: exit code 2, usage on stderr, nothing on stdout, and no store change.
+func TestRunTrustRemoveTooManyArgs(t *testing.T) {
+	setTrustConfigRoot(t)
+	deps := trustDeps(t.TempDir())
+
+	var out, errBuf bytes.Buffer
+	if code := runTrust([]string{"remove", "a", "b"}, &out, &errBuf, deps); code != exitUsage {
+		t.Fatalf("remove with two args returned %d, want %d", code, exitUsage)
+	}
+	if errBuf.Len() == 0 {
+		t.Fatalf("remove with two args should write usage to stderr, stderr was empty")
+	}
+	if out.Len() != 0 {
+		t.Fatalf("remove with two args should not write to stdout, got %q", out.String())
+	}
+}
+
+// TestRunTrustHelp proves the -h / --help / help subcommands print usage to stdout,
+// nothing to stderr, and return success, matching the other CLI help entrypoints
+// (unknown subcommands, by contrast, write to stderr with the usage exit code).
+func TestRunTrustHelp(t *testing.T) {
+	setTrustConfigRoot(t)
+	deps := trustDeps(t.TempDir())
+
+	for _, flag := range []string{"-h", "--help", "help"} {
+		var out, errBuf bytes.Buffer
+		if code := runTrust([]string{flag}, &out, &errBuf, deps); code != exitSuccess {
+			t.Fatalf("trust %s returned %d, want %d", flag, code, exitSuccess)
+		}
+		if !strings.Contains(out.String(), "Usage") {
+			t.Fatalf("trust %s stdout = %q, want it to contain usage text", flag, out.String())
+		}
+		if errBuf.Len() != 0 {
+			t.Fatalf("trust %s should not write to stderr, got %q", flag, errBuf.String())
+		}
+	}
+}

internal/config/credentials.go

@@ -8,6 +8,7 @@ import (
 	"strings"
 
 	"github.com/Gitlawb/zero/internal/credstore"
+	"github.com/Gitlawb/zero/internal/providercatalog"
 )
 
 // ProviderKeyStoreAt opens the encrypted credential store whose file backend lives
@@ -168,6 +169,86 @@ func (profile ProviderProfile) HasConfiguredCredential() bool {
 		strings.TrimSpace(profile.AuthHeaderValue) != ""
 }
 
+// MissingCredentialEnv reports the environment variable this profile is
+// expected to read an API key from, and whether that expectation is
+// currently unmet. It is the single source of truth for "does this saved
+// provider need a credential it doesn't have" — the CLI (`providers check`,
+// `usableSavedProviders`) and TUI (`/providers` status, provider wizard
+// summary) surfaces all delegate here so they never disagree about a
+// profile's status.
+//
+// For the catalog's two "bring your own endpoint" descriptors
+// (custom-openai-compatible, custom-anthropic-compatible), RequiresAuth is
+// just a wizard template default, not a fact about the user's actual
+// endpoint: a custom target may need no auth at all. So for those two only,
+// a credential is expected exclusively when this profile carries its own
+// explicit APIKeyEnv that differs from the catalog's guessed default — a
+// value matching that default is indistinguishable from the wizard's old
+// auto-stamp bug (issue #555) and is treated the same as unset, which also
+// self-heals profiles saved before that bug was fixed.
+func (profile ProviderProfile) MissingCredentialEnv() (string, bool) {
+	if profile.HasConfiguredCredential() {
+		return "", false
+	}
+	if catalogID := strings.TrimSpace(profile.CatalogID); catalogID != "" {
+		descriptor, err := providercatalog.Require(catalogID)
+		if err != nil {
+			return "", false
+		}
+		defaultEnvVar := firstNonEmptyTrimmed(descriptor.AuthEnvVars...)
+		if descriptor.Custom {
+			envVar := strings.TrimSpace(profile.APIKeyEnv)
+			if envVar == "" || envVar == defaultEnvVar {
+				return "", false
+			}
+			return envVar, true
+		}
+		if !descriptor.RequiresAuth {
+			return "", false
+		}
+		if envVar := strings.TrimSpace(profile.APIKeyEnv); envVar != "" {
+			return envVar, true
+		}
+		return defaultEnvVar, true
+	}
+
+	switch profile.normalizedProviderKind() {
+	case ProviderKindOpenAI, ProviderKindOpenAICompatible:
+		return firstNonEmptyTrimmed(profile.APIKeyEnv, "OPENAI_API_KEY"), true
+	case ProviderKindAnthropic, ProviderKindAnthropicCompat:
+		return firstNonEmptyTrimmed(profile.APIKeyEnv, "ANTHROPIC_API_KEY"), true
+	case ProviderKindGoogle:
+		return firstNonEmptyTrimmed(profile.APIKeyEnv, "GEMINI_API_KEY"), true
+	default:
+		if envVar := strings.TrimSpace(profile.APIKeyEnv); envVar != "" {
+			return envVar, true
+		}
+		return "", false
+	}
+}
+
+// normalizedProviderKind resolves this profile's kind, falling back to the
+// legacy Provider string field (lowercased) when ProviderKind is unset —
+// matching how the resolver (resolver.go) treats the two fields as synonyms.
+func (profile ProviderProfile) normalizedProviderKind() ProviderKind {
+	if kind := strings.TrimSpace(string(profile.ProviderKind)); kind != "" {
+		return ProviderKind(strings.ToLower(kind))
+	}
+	if provider := strings.TrimSpace(profile.Provider); provider != "" {
+		return ProviderKind(strings.ToLower(provider))
+	}
+	return ""
+}
+
+func firstNonEmptyTrimmed(values ...string) string {
+	for _, value := range values {
+		if trimmed := strings.TrimSpace(value); trimmed != "" {
+			return trimmed
+		}
+	}
+	return ""
+}
+
 // OAuthLoginCandidates returns the login names to try, in order, when resolving
 // this profile's OAuth token — used by the runtime bearer resolver, the Codex
 // account-header resolver, and the onboarding presence check so all three agree.

internal/config/credentials_test.go

@@ -238,3 +238,82 @@ func TestOAuthLoginCandidates(t *testing.T) {
 		})
 	}
 }
+
+func TestProviderProfileMissingCredentialEnv(t *testing.T) {
+	cases := []struct {
+		name    string
+		profile ProviderProfile
+		wantEnv string
+		want    bool
+	}{
+		{
+			name:    "catalog provider requiring auth",
+			profile: ProviderProfile{Name: "groq", CatalogID: "groq"},
+			wantEnv: "GROQ_API_KEY",
+			want:    true,
+		},
+		{
+			name:    "local catalog provider",
+			profile: ProviderProfile{Name: "local", CatalogID: "ollama"},
+			want:    false,
+		},
+		{
+			name:    "credential resolved via inline key",
+			profile: ProviderProfile{Name: "openai", ProviderKind: ProviderKindOpenAI, APIKey: "sk-test"},
+			want:    false,
+		},
+		{
+			// issue #555: a custom endpoint left with no credential means "no
+			// auth needed" (e.g. a local llama.cpp server), not "missing".
+			name: "custom openai compatible with no credential configured",
+			profile: ProviderProfile{
+				Name:      "local-llama",
+				CatalogID: "custom-openai-compatible",
+				BaseURL:   "http://192.168.1.50:8080/v1",
+			},
+			want: false,
+		},
+		{
+			name: "custom openai compatible with explicit non-default api key env",
+			profile: ProviderProfile{
+				Name:      "local-llama",
+				CatalogID: "custom-openai-compatible",
+				APIKeyEnv: "LLAMA_CPP_API_KEY",
+			},
+			wantEnv: "LLAMA_CPP_API_KEY",
+			want:    true,
+		},
+		{
+			// Self-heal: a profile stamped by the pre-fix wizard with the
+			// catalog's own guessed default is indistinguishable from that bug
+			// and must not stay filtered out after upgrading.
+			name: "custom openai compatible with stale legacy default env",
+			profile: ProviderProfile{
+				Name:      "local-llama",
+				CatalogID: "custom-openai-compatible",
+				APIKeyEnv: "OPENAI_API_KEY",
+			},
+			want: false,
+		},
+		{
+			name:    "openai compatible without catalog falls back to provider kind",
+			profile: ProviderProfile{Name: "custom", ProviderKind: ProviderKindOpenAICompatible},
+			wantEnv: "OPENAI_API_KEY",
+			want:    true,
+		},
+		{
+			name:    "legacy Provider string field fallback",
+			profile: ProviderProfile{Name: "legacy", Provider: "anthropic"},
+			wantEnv: "ANTHROPIC_API_KEY",
+			want:    true,
+		},
+	}
+	for _, c := range cases {
+		t.Run(c.name, func(t *testing.T) {
+			gotEnv, got := c.profile.MissingCredentialEnv()
+			if got != c.want || gotEnv != c.wantEnv {
+				t.Fatalf("MissingCredentialEnv() = (%q, %v), want (%q, %v)", gotEnv, got, c.wantEnv, c.want)
+			}
+		})
+	}
+}

internal/config/local_control_test.go

@@ -21,7 +21,7 @@ func TestResolveLocalControlFromUserConfig(t *testing.T) {
 		t.Fatalf("write config: %v", err)
 	}
 
-	resolved, err := Resolve(ResolveOptions{UserConfigPath: path})
+	resolved, err := Resolve(ResolveOptions{UserConfigPath: path, Env: map[string]string{}})
 	if err != nil {
 		t.Fatalf("Resolve returned error: %v", err)
 	}
@@ -48,7 +48,7 @@ func TestResolveLocalControlIgnoresProjectOptIn(t *testing.T) {
 		t.Fatalf("write project config: %v", err)
 	}
 
-	resolved, err := Resolve(ResolveOptions{ProjectConfigPath: projectPath})
+	resolved, err := Resolve(ResolveOptions{ProjectConfigPath: projectPath, Env: map[string]string{}})
 	if err != nil {
 		t.Fatalf("Resolve returned error: %v", err)
 	}
@@ -61,7 +61,7 @@ func TestResolveLocalControlIgnoresProjectOptIn(t *testing.T) {
 }
 
 func TestLocalControlDefaultsBrowserAndTerminalOn(t *testing.T) {
-	resolved, err := Resolve(ResolveOptions{})
+	resolved, err := Resolve(ResolveOptions{Env: map[string]string{}})
 	if err != nil {
 		t.Fatalf("Resolve returned error: %v", err)
 	}
@@ -81,7 +81,7 @@ func TestLocalControlExplicitGlobalDisable(t *testing.T) {
 	if err := os.WriteFile(path, []byte(`{"localControl":{"enabled":false}}`), 0o600); err != nil {
 		t.Fatalf("write config: %v", err)
 	}
-	resolved, err := Resolve(ResolveOptions{UserConfigPath: path})
+	resolved, err := Resolve(ResolveOptions{UserConfigPath: path, Env: map[string]string{}})
 	if err != nil {
 		t.Fatalf("Resolve returned error: %v", err)
 	}

internal/config/mcp_defaults.go

@@ -1,6 +1,9 @@
 package config
 
-import "strings"
+import (
+	"reflect"
+	"strings"
+)
 
 // DefaultMCPServers returns the MCP servers Zero ships ENABLED by default so web
 // search and scraping work out of the box with no setup and no API key. They are
@@ -29,3 +32,29 @@ func IsDefaultMCPServer(name string) bool {
 	_, ok := DefaultMCPServers()[strings.TrimSpace(name)]
 	return ok
 }
+
+// IsUnconfiguredDefault reports whether server is one of Zero's built-in
+// defaults that the user never wrote an entry for in their config — i.e. it is
+// running with whatever Zero ships (e.g. keyless Firecrawl, no credentials).
+//
+// Both conditions below must hold:
+//   - !server.configured: the user's JSON never declared an object for this
+//     server key at all (set by MCPServerConfig.UnmarshalJSON only when it
+//     actually ran for this key). Any explicit action — including a
+//     disable/enable toggle like `zero mcp enable firecrawl` that leaves the
+//     resolved value unchanged — sets configured, so it always counts as
+//     user-configured, even though the value comparison below could not tell
+//     the difference on its own.
+//   - reflect.DeepEqual(def, server): the value still matches the default.
+//     This is the fallback for callers that construct MCPServerConfig
+//     directly rather than through the JSON/merge pipeline (server.configured
+//     is then always false) — without it, any hand-built config with
+//     different field values would be misreported as unconfigured.
+//
+// Callers use this to tell "server we turned on for the user" apart from
+// "server the user configured themselves," e.g. to avoid warning loudly when
+// an out-of-the-box default that was never given credentials fails to connect.
+func IsUnconfiguredDefault(name string, server MCPServerConfig) bool {
+	def, ok := DefaultMCPServers()[strings.TrimSpace(name)]
+	return ok && !server.configured && reflect.DeepEqual(def, server)
+}

internal/config/mcp_defaults_test.go

@@ -49,6 +49,69 @@ func TestResolveMCPUserCanDisableDefault(t *testing.T) {
 	}
 }
 
+func TestIsUnconfiguredDefault(t *testing.T) {
+	if !IsUnconfiguredDefault("firecrawl", DefaultMCPServers()["firecrawl"]) {
+		t.Fatal("an untouched firecrawl default should be reported as unconfigured")
+	}
+	if IsUnconfiguredDefault("firecrawl", MCPServerConfig{Type: "http", URL: "http://localhost:3002/mcp"}) {
+		t.Fatal("a server overriding the default URL is no longer unconfigured")
+	}
+	if IsUnconfiguredDefault("firecrawl", MCPServerConfig{Type: "http", URL: "https://mcp.firecrawl.dev/v2/mcp", Auth: "bearer"}) {
+		t.Fatal("a server with credentials added is no longer unconfigured")
+	}
+	if IsUnconfiguredDefault("not-a-default", MCPServerConfig{}) {
+		t.Fatal("a server with no matching default can never be unconfigured-default")
+	}
+}
+
+func TestResolveMCPExplicitReenableIsNotUnconfiguredDefault(t *testing.T) {
+	// `zero mcp enable firecrawl` after a prior disable writes {"disabled":false}
+	// explicitly. The resolved value is identical to the untouched default (both
+	// enabled, no credentials), but the user DID take an explicit action here, so
+	// IsUnconfiguredDefault must not treat it as untouched (issue #563 review).
+	path := filepath.Join(t.TempDir(), "config.json")
+	if err := os.WriteFile(path, []byte(`{"mcp":{"servers":{"firecrawl":{"disabled":false}}}}`), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	cfg, err := ResolveMCP(ResolveOptions{UserConfigPath: path})
+	if err != nil {
+		t.Fatalf("ResolveMCP: %v", err)
+	}
+	firecrawl := cfg.Servers["firecrawl"]
+	if firecrawl.Disabled {
+		t.Fatalf("explicit re-enable should leave the server enabled: %#v", firecrawl)
+	}
+	if IsUnconfiguredDefault("firecrawl", firecrawl) {
+		t.Fatal("an explicit enable/disable toggle must count as user-configured, even though the resolved value matches the default")
+	}
+}
+
+func TestResolveMCPExplicitRedeclareOfDefaultValuesIsNotUnconfiguredDefault(t *testing.T) {
+	// A user who copies firecrawl's exact default type/url into their config
+	// (e.g. from an example file, planning to add credentials later) produces a
+	// resolved value byte-identical to DefaultMCPServers()["firecrawl"] — the
+	// same trap TestResolveMCPExplicitReenableIsNotUnconfiguredDefault covers for
+	// the disabled toggle. IsUnconfiguredDefault must still treat this as
+	// user-configured because the user's JSON declared an entry for it, even
+	// though a plain resolved-value comparison could not tell the difference.
+	path := filepath.Join(t.TempDir(), "config.json")
+	if err := os.WriteFile(path, []byte(`{"mcp":{"servers":{"firecrawl":{"type":"http","url":"https://mcp.firecrawl.dev/v2/mcp"}}}}`), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	cfg, err := ResolveMCP(ResolveOptions{UserConfigPath: path})
+	if err != nil {
+		t.Fatalf("ResolveMCP: %v", err)
+	}
+	firecrawl := cfg.Servers["firecrawl"]
+	want := DefaultMCPServers()["firecrawl"]
+	if firecrawl.Type != want.Type || firecrawl.URL != want.URL || firecrawl.Disabled != want.Disabled {
+		t.Fatalf("expected the resolved value to match the default's fields exactly: %#v", firecrawl)
+	}
+	if IsUnconfiguredDefault("firecrawl", firecrawl) {
+		t.Fatal("redeclaring the default's exact values is still an explicit user configuration, not an untouched default")
+	}
+}
+
 func TestResolveMCPUserCanOverrideDefaultURLKeepingOtherFields(t *testing.T) {
 	path := filepath.Join(t.TempDir(), "config.json")
 	// Point firecrawl at a self-hosted instance; the default's Type must survive.

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/resolve_mcp_exclude_project_test.go

@@ -0,0 +1,61 @@
+package config
+
+import (
+	"os"
+	"path/filepath"
+	"testing"
+)
+
+// TestResolveMCPExcludeProjectDropsProjectServers proves that ExcludeProject drops
+// the project config layer from MCP resolution (fail-closed for an untrusted
+// workspace) while keeping the built-in defaults and the user config server.
+func TestResolveMCPExcludeProjectDropsProjectServers(t *testing.T) {
+	userPath := filepath.Join(t.TempDir(), "config.json")
+	if err := os.WriteFile(userPath, []byte(`{"mcp":{"servers":{"user-srv":{"type":"stdio","command":"user-cmd"}}}}`), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	projectPath := filepath.Join(t.TempDir(), "config.json")
+	if err := os.WriteFile(projectPath, []byte(`{"mcp":{"servers":{"proj-srv":{"type":"stdio","command":"proj-cmd"}}}}`), 0o600); err != nil {
+		t.Fatal(err)
+	}
+
+	t.Run("included when ExcludeProject is false", func(t *testing.T) {
+		cfg, err := ResolveMCP(ResolveOptions{
+			UserConfigPath:    userPath,
+			ProjectConfigPath: projectPath,
+			ExcludeProject:    false,
+		})
+		if err != nil {
+			t.Fatalf("ResolveMCP: %v", err)
+		}
+		if _, ok := cfg.Servers["proj-srv"]; !ok {
+			t.Fatal("a trusted resolve (ExcludeProject=false) must include the project server")
+		}
+		if _, ok := cfg.Servers["user-srv"]; !ok {
+			t.Fatal("the user server must always be present")
+		}
+		if _, ok := cfg.Servers["firecrawl"]; !ok {
+			t.Fatal("the built-in default must always be present")
+		}
+	})
+
+	t.Run("excluded when ExcludeProject is true", func(t *testing.T) {
+		cfg, err := ResolveMCP(ResolveOptions{
+			UserConfigPath:    userPath,
+			ProjectConfigPath: projectPath,
+			ExcludeProject:    true,
+		})
+		if err != nil {
+			t.Fatalf("ResolveMCP: %v", err)
+		}
+		if _, ok := cfg.Servers["proj-srv"]; ok {
+			t.Fatal("an untrusted resolve (ExcludeProject=true) must drop the project server")
+		}
+		if _, ok := cfg.Servers["user-srv"]; !ok {
+			t.Fatal("the user server must survive when the project layer is dropped")
+		}
+		if _, ok := cfg.Servers["firecrawl"]; !ok {
+			t.Fatal("the built-in default must survive when the project layer is dropped")
+		}
+	})
+}

internal/config/resolver.go

@@ -127,6 +127,9 @@ func Resolve(options ResolveOptions) (ResolvedConfig, error) {
 			return ResolvedConfig{}, fmt.Errorf("invalid notify.focusMode %q: expected unfocused, always, or focused", focusMode)
 		}
 	}
+	if err := validateSTTConfig(cfg.STT); err != nil {
+		return ResolvedConfig{}, err
+	}
 
 	providers, active, err := normalizeProviders(cfg.Providers, cfg.ActiveProvider, options.Env)
 	if err != nil {
@@ -150,6 +153,7 @@ func Resolve(options ResolveOptions) (ResolvedConfig, error) {
 		Preferences:    cfg.Preferences,
 		KeyBindings:    cfg.KeyBindings,
 		LocalControl:   cfg.LocalControl,
+		STT:            cfg.STT,
 	}, nil
 }
 
@@ -159,16 +163,25 @@ 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
-		}
-		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
 }
@@ -223,6 +236,9 @@ func mergeConfig(dst *FileConfig, src FileConfig) {
 	if src.Preferences.FavoriteModels != nil {
 		dst.Preferences.FavoriteModels = normalizeFavoriteModels(src.Preferences.FavoriteModels)
 	}
+	if src.Preferences.RecentModels != nil {
+		dst.Preferences.RecentModels = NormalizeRecentModels(src.Preferences.RecentModels)
+	}
 	if src.Preferences.Recaps != nil {
 		dst.Preferences.Recaps = src.Preferences.Recaps
 	}
@@ -231,6 +247,7 @@ func mergeConfig(dst *FileConfig, src FileConfig) {
 	}
 	mergeLocalControlConfig(&dst.LocalControl, src.LocalControl)
 	mergeKeyBindings(&dst.KeyBindings, src.KeyBindings)
+	mergeSTTConfig(&dst.STT, src.STT)
 }
 
 func mergeProjectConfig(dst *FileConfig, src FileConfig) error {
@@ -247,7 +264,9 @@ 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
@@ -681,6 +700,7 @@ func applyOverrides(cfg *FileConfig, overrides Overrides) {
 	}
 	mergeLocalControlConfig(&cfg.LocalControl, overrides.LocalControl)
 	mergeKeyBindings(&cfg.KeyBindings, overrides.KeyBindings)
+	mergeSTTConfig(&cfg.STT, overrides.STT)
 	for _, provider := range overrides.Providers {
 		mergeProvider(cfg, provider)
 	}
@@ -734,58 +754,89 @@ func mergeKeyBindings(dst *KeyBindingsConfig, src KeyBindingsConfig) {
 	}
 }
 
-func mergeMCPConfig(dst *MCPConfig, src MCPConfig) {
-	if len(src.Servers) == 0 {
-		return
+// mergeSTTConfig overlays any set field of src onto dst. Each field carries its
+// own "unset" sentinel (empty string, 0, or nil *bool), matching how the other
+// section mergers detect intent.
+func mergeSTTConfig(dst *STTConfig, src STTConfig) {
+	if src.Provider != "" {
+		dst.Provider = src.Provider
 	}
-	if dst.Servers == nil {
-		dst.Servers = map[string]MCPServerConfig{}
+	if src.StreamProvider != "" {
+		dst.StreamProvider = src.StreamProvider
 	}
-	for name, server := range src.Servers {
-		dst.Servers[name] = mergeMCPServer(dst.Servers[name], server)
+	if src.Streaming != nil {
+		dst.Streaming = src.Streaming
+	}
+	if src.Model != "" {
+		dst.Model = src.Model
+	}
+	if src.StreamModel != "" {
+		dst.StreamModel = src.StreamModel
+	}
+	if src.LocalModelPath != "" {
+		dst.LocalModelPath = src.LocalModelPath
+	}
+	if src.LocalBinary != "" {
+		dst.LocalBinary = src.LocalBinary
+	}
+	if src.LocalServerBinary != "" {
+		dst.LocalServerBinary = src.LocalServerBinary
+	}
+	if src.LocalServerPort != 0 {
+		dst.LocalServerPort = src.LocalServerPort
+	}
+	if src.EngineVersion != "" {
+		dst.EngineVersion = src.EngineVersion
+	}
+	if src.NumThreads != 0 {
+		dst.NumThreads = src.NumThreads
+	}
+	if src.Language != "" {
+		dst.Language = src.Language
+	}
+	if src.MaxDurationSeconds != 0 {
+		dst.MaxDurationSeconds = src.MaxDurationSeconds
+	}
+	if src.SilenceAutoStop != nil {
+		dst.SilenceAutoStop = src.SilenceAutoStop
+	}
+	if src.AutoSubmit != nil {
+		dst.AutoSubmit = src.AutoSubmit
+	}
+	if src.WindowsAudioDevice != "" {
+		dst.WindowsAudioDevice = src.WindowsAudioDevice
 	}
 }
 
-func mergeMCPServer(base MCPServerConfig, next MCPServerConfig) MCPServerConfig {
-	if strings.TrimSpace(next.Type) != "" {
-		base.Type = next.Type
+// validateSTTConfig rejects unknown provider/streamProvider values and a
+// negative maxDurationSeconds at load time — a clear startup error naming the
+// bad value and the valid options, never a silent fallback the user did not ask
+// for (§11a).
+func validateSTTConfig(cfg STTConfig) error {
+	if cfg.Provider != "" {
+		switch cfg.Provider {
+		case STTProviderLocal, STTProviderGroq, STTProviderOpenAI:
+		default:
+			return fmt.Errorf("invalid stt.provider %q: expected local, groq, or openai", cfg.Provider)
+		}
 	}
-	if strings.TrimSpace(next.Command) != "" {
-		base.Command = next.Command
+	if cfg.StreamProvider != "" {
+		switch cfg.StreamProvider {
+		case STTProviderLocal, STTProviderDeepgram, STTProviderOpenAI:
+		default:
+			return fmt.Errorf("invalid stt.streamProvider %q: expected local, deepgram, or openai", cfg.StreamProvider)
+		}
 	}
-	if next.Args != nil {
-		base.Args = append([]string{}, next.Args...)
+	if cfg.MaxDurationSeconds < 0 {
+		return fmt.Errorf("invalid stt.maxDurationSeconds %d: must be >= 0 (0 uses the default)", cfg.MaxDurationSeconds)
 	}
-	if next.Env != nil {
-		base.Env = copyMCPStringMap(next.Env)
+	if cfg.NumThreads < 0 {
+		return fmt.Errorf("invalid stt.numThreads %d: must be >= 0 (0 uses the engine default)", cfg.NumThreads)
 	}
-	if strings.TrimSpace(next.URL) != "" {
-		base.URL = next.URL
+	if cfg.LocalServerPort < 0 || cfg.LocalServerPort > 65535 {
+		return fmt.Errorf("invalid stt.localServerPort %d: must be between 1 and 65535 (0 uses the default)", cfg.LocalServerPort)
 	}
-	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
-	}
-	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
+	return nil
 }
 
 func hasProviderFields(profile ProviderProfile) bool {

internal/config/resolver_test.go

@@ -127,6 +127,79 @@ func TestResolveLoadsFavoriteModelsFromUserConfigOnly(t *testing.T) {
 	}
 }
 
+func TestResolveLoadsRecentModelsFromUserConfigOnly(t *testing.T) {
+	userPath := writeConfig(t, `{
+		"activeProvider": "user",
+		"providers": [{
+			"name": "user",
+			"provider": "openai",
+			"apiKey": "sk-user",
+			"model": "gpt-user"
+		}],
+		"preferences": {
+			"recentModels": [
+				{"provider": "openrouter", "model": " google/gemini-2.5-pro "},
+				{"provider": "openrouter", "model": "minimax/minimax-m2.1"}
+			]
+		}
+	}`)
+	projectPath := writeConfig(t, `{
+		"preferences": {
+			"recentModels": [{"provider": "project", "model": "project-model"}]
+		}
+	}`)
+
+	resolved, err := Resolve(ResolveOptions{
+		UserConfigPath:    userPath,
+		ProjectConfigPath: projectPath,
+		Env:               map[string]string{},
+	})
+	if err != nil {
+		t.Fatalf("Resolve() error = %v", err)
+	}
+
+	want := []RecentModelEntry{
+		{Provider: "openrouter", Model: "google/gemini-2.5-pro"},
+		{Provider: "openrouter", Model: "minimax/minimax-m2.1"},
+	}
+	if !reflect.DeepEqual(resolved.Preferences.RecentModels, want) {
+		t.Fatalf("RecentModels = %#v, want %#v", resolved.Preferences.RecentModels, want)
+	}
+}
+
+// recentModels must never be merged from project config (same posture as
+// favoriteModels): a project setting it with no corresponding user setting
+// should leave the resolved value empty, not pick up the project's list.
+func TestResolveIgnoresProjectOnlyRecentModels(t *testing.T) {
+	userPath := writeConfig(t, `{
+		"activeProvider": "user",
+		"providers": [{
+			"name": "user",
+			"provider": "openai",
+			"apiKey": "sk-user",
+			"model": "gpt-user"
+		}]
+	}`)
+	projectPath := writeConfig(t, `{
+		"preferences": {
+			"recentModels": [{"provider": "project", "model": "project-model"}]
+		}
+	}`)
+
+	resolved, err := Resolve(ResolveOptions{
+		UserConfigPath:    userPath,
+		ProjectConfigPath: projectPath,
+		Env:               map[string]string{},
+	})
+	if err != nil {
+		t.Fatalf("Resolve() error = %v", err)
+	}
+
+	if len(resolved.Preferences.RecentModels) != 0 {
+		t.Fatalf("RecentModels = %#v, want empty (project-only recentModels must not be merged)", resolved.Preferences.RecentModels)
+	}
+}
+
 func TestResolveLoadsProviderCatalogSnakeAndCamelJSONFields(t *testing.T) {
 	path := writeConfig(t, `{
 		"activeProvider": "snake",

internal/config/stt_test.go

@@ -0,0 +1,139 @@
+package config
+
+import (
+	"os"
+	"strings"
+	"testing"
+)
+
+func TestSetSTTLocalEngineResetsStreamProvider(t *testing.T) {
+	dir := t.TempDir()
+	path := dir + "/config.json"
+	// Pre-seed a cloud streaming provider — the situation the fix guards against.
+	if err := os.WriteFile(path, []byte(`{"stt":{"streamProvider":"deepgram"}}`), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	cfg, err := SetSTTLocalEngine(path, "sherpa-onnx-offline", "sherpa-onnx-online-websocket-server", "/models/kroko", true)
+	if err != nil {
+		t.Fatalf("SetSTTLocalEngine: %v", err)
+	}
+	// Switching to a local model must point BOTH pipelines at the local engine, or
+	// the leftover cloud streamProvider keeps the live transcript on the cloud.
+	if cfg.STT.Provider != STTProviderLocal || cfg.STT.StreamProvider != STTProviderLocal {
+		t.Errorf("both providers should be local, got provider=%q streamProvider=%q", cfg.STT.Provider, cfg.STT.StreamProvider)
+	}
+}
+
+func TestValidateSTTConfigRejectsUnknownProvider(t *testing.T) {
+	_, issues := ValidateBytes([]byte(`{"stt":{"provider":"assemblyai"}}`))
+	if len(issues) == 0 {
+		t.Fatal("expected an issue for unknown stt.provider")
+	}
+	if !strings.Contains(issues[0].Message, "assemblyai") || !strings.Contains(issues[0].Message, "local, groq, or openai") {
+		t.Errorf("issue message = %q, want it to name the bad value and valid options", issues[0].Message)
+	}
+}
+
+func TestValidateSTTConfigRejectsUnknownStreamProvider(t *testing.T) {
+	_, issues := ValidateBytes([]byte(`{"stt":{"streamProvider":"groq"}}`))
+	if len(issues) == 0 {
+		t.Fatal("expected an issue: groq has no streaming product, not a valid streamProvider")
+	}
+	if !strings.Contains(issues[0].Message, "deepgram") {
+		t.Errorf("issue message = %q", issues[0].Message)
+	}
+}
+
+func TestValidateSTTConfigAcceptsValid(t *testing.T) {
+	_, issues := ValidateBytes([]byte(`{"stt":{"provider":"groq","streamProvider":"deepgram","maxDurationSeconds":120}}`))
+	if len(issues) != 0 {
+		t.Fatalf("unexpected issues: %+v", issues)
+	}
+}
+
+func TestValidateSTTConfigRejectsNegativeDuration(t *testing.T) {
+	_, issues := ValidateBytes([]byte(`{"stt":{"maxDurationSeconds":-5}}`))
+	if len(issues) == 0 {
+		t.Fatal("expected an issue for negative maxDurationSeconds")
+	}
+}
+
+func TestSTTConfigDefaults(t *testing.T) {
+	cfg := STTConfig{}
+	if cfg.STTProvider() != STTProviderLocal {
+		t.Errorf("default provider = %q, want local", cfg.STTProvider())
+	}
+	if cfg.STTStreamProvider() != STTProviderLocal {
+		t.Errorf("default streamProvider = %q, want local", cfg.STTStreamProvider())
+	}
+	if !cfg.StreamingEnabled() {
+		t.Error("streaming should default on")
+	}
+	if !cfg.SilenceAutoStopEnabled() {
+		t.Error("silence auto-stop should default on")
+	}
+	if cfg.AutoSubmitEnabled() {
+		t.Error("auto-submit should default off (insert-for-review is the safety net)")
+	}
+	if !cfg.Empty() {
+		t.Error("zero-value STTConfig should be Empty")
+	}
+}
+
+func TestSTTConfigTriStateBools(t *testing.T) {
+	cfg, issues := ValidateBytes([]byte(`{"stt":{"autoSubmit":true,"streaming":false,"silenceAutoStop":false}}`))
+	if len(issues) != 0 {
+		t.Fatalf("unexpected issues: %+v", issues)
+	}
+	if !cfg.STT.AutoSubmitEnabled() {
+		t.Error("autoSubmit:true should be enabled")
+	}
+	if cfg.STT.StreamingEnabled() {
+		t.Error("streaming:false should be disabled")
+	}
+	if cfg.STT.SilenceAutoStopEnabled() {
+		t.Error("silenceAutoStop:false should be disabled")
+	}
+}
+
+func TestSTTConfigRoundTripsThroughMarshal(t *testing.T) {
+	cfg, _ := ValidateBytes([]byte(`{"stt":{"provider":"groq","model":"whisper-large-v3-turbo","localServerPort":6007}}`))
+	data, err := cfg.MarshalJSON()
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !strings.Contains(string(data), `"stt"`) {
+		t.Errorf("marshaled config dropped stt section: %s", data)
+	}
+	reparsed, issues := ValidateBytes(data)
+	if len(issues) != 0 {
+		t.Fatalf("re-parse issues: %+v", issues)
+	}
+	if reparsed.STT.Provider != STTProviderGroq || reparsed.STT.Model != "whisper-large-v3-turbo" || reparsed.STT.LocalServerPort != 6007 {
+		t.Errorf("round-trip lost fields: %+v", reparsed.STT)
+	}
+}
+
+func TestSetSTTModelPersists(t *testing.T) {
+	dir := t.TempDir()
+	path := dir + "/config.json"
+	// Cloud provider stores under stt.model.
+	if _, err := SetSTTModel(path, STTProviderGroq, "whisper-large-v3-turbo"); err != nil {
+		t.Fatalf("SetSTTModel: %v", err)
+	}
+	cfg, issues := ValidateFile(path)
+	if len(issues) != 0 {
+		t.Fatalf("issues: %+v", issues)
+	}
+	if cfg.STT.Provider != STTProviderGroq || cfg.STT.Model != "whisper-large-v3-turbo" {
+		t.Errorf("persisted STT = %+v", cfg.STT)
+	}
+	// Local provider stores under stt.localModelPath.
+	if _, err := SetSTTModel(path, STTProviderLocal, "/models/moonshine"); err != nil {
+		t.Fatalf("SetSTTModel local: %v", err)
+	}
+	cfg, _ = ValidateFile(path)
+	if cfg.STT.LocalModelPath != "/models/moonshine" {
+		t.Errorf("local model path = %q", cfg.STT.LocalModelPath)
+	}
+}

internal/config/types.go

@@ -97,6 +97,11 @@ type ToolsConfig struct {
 
 type PreferencesConfig struct {
 	FavoriteModels []string `json:"favoriteModels,omitempty"`
+	// RecentModels is the short automatic history of provider+model pairs the
+	// user has switched to via /model, newest first. Unlike FavoriteModels
+	// (manual pins), this list is maintained automatically on every switch and
+	// capped to MaxRecentModels entries. See RecentModelEntry.
+	RecentModels []RecentModelEntry `json:"recentModels,omitempty"`
 	// Theme is the persisted TUI palette preference — "auto" or a registered theme
 	// name (e.g. "dracula"). Applied at startup below the --theme flag and
 	// ZERO_THEME, so a /theme choice survives restart. Empty = unset (defaults auto).
@@ -107,6 +112,18 @@ type PreferencesConfig struct {
 	Recaps *bool `json:"recaps,omitempty"`
 }
 
+// RecentModelEntry is one provider-qualified model selection recorded in
+// Preferences.RecentModels. Provider is the saved provider profile's Name (not
+// a display label), so a recent entry can be resolved back to a concrete
+// profile the same way the /model picker's cross-provider rows are.
+type RecentModelEntry struct {
+	Provider string `json:"provider"`
+	Model    string `json:"model"`
+}
+
+// MaxRecentModels caps the persisted/displayed recent-selection history.
+const MaxRecentModels = 5
+
 // RecapsEnabled reports whether post-turn recaps are on. Unset defaults to ON.
 func (p PreferencesConfig) RecapsEnabled() bool {
 	return p.Recaps == nil || *p.Recaps
@@ -131,6 +148,115 @@ type KeyBindingsConfig struct {
 	ToggleSidebar KeyBindingDef `json:"toggleSidebar,omitempty"`
 }
 
+// STTProviderKind is the batch (or streaming) transcription backend selector.
+// Validated at config load time against the known set, so a typo fails loudly
+// with the valid options rather than silently falling back to a default.
+type STTProviderKind string
+
+const (
+	STTProviderLocal    STTProviderKind = "local"
+	STTProviderGroq     STTProviderKind = "groq"
+	STTProviderOpenAI   STTProviderKind = "openai"
+	STTProviderDeepgram STTProviderKind = "deepgram"
+)
+
+// STTConfig configures speech-to-text dictation. All fields are optional; empty
+// values take the documented defaults. Booleans that need a real tri-state
+// (distinguishing "unset" from "false") use *bool, matching PreferencesConfig.Recaps.
+type STTConfig struct {
+	// Provider is the batch transcription backend: "local" (sherpa-onnx-offline,
+	// the default), "groq", or "openai". Termux and the fallback everywhere use
+	// this path.
+	Provider STTProviderKind `json:"provider,omitempty"`
+	// StreamProvider is the streaming backend on desktop: "local" (sherpa-onnx
+	// websocket server, default), "deepgram", or "openai" (Realtime).
+	StreamProvider STTProviderKind `json:"streamProvider,omitempty"`
+	// Streaming enables the live-transcript pipeline on platforms that support
+	// it (desktop). Defaults to on; set false to always use the batch pipeline.
+	Streaming *bool `json:"streaming,omitempty"`
+	// Model overrides the cloud batch model (e.g. "whisper-large-v3-turbo" for
+	// Groq, "whisper-1" for OpenAI). Empty uses the provider default.
+	Model string `json:"model,omitempty"`
+	// StreamModel overrides the cloud streaming model (Deepgram/OpenAI Realtime).
+	StreamModel string `json:"streamModel,omitempty"`
+	// LocalModelPath is the sherpa-onnx model directory for local transcription
+	// (batch and streaming). Required to use a local provider.
+	LocalModelPath string `json:"localModelPath,omitempty"`
+	// LocalBinary overrides the offline binary name/path (default
+	// "sherpa-onnx-offline"); LocalServerBinary overrides the streaming server
+	// (default "sherpa-onnx-online-websocket-server"). Both looked up on PATH.
+	LocalBinary       string `json:"localBinary,omitempty"`
+	LocalServerBinary string `json:"localServerBinary,omitempty"`
+	// LocalServerPort is the localhost port for the sherpa-onnx websocket server
+	// (default 6006).
+	LocalServerPort int `json:"localServerPort,omitempty"`
+	// EngineVersion selects the sherpa-onnx release the auto-download fetches
+	// ("" → a pinned known-good default; "latest" or any release tag also works,
+	// so a newer engine needs no Zero update). Verified against the release's
+	// published SHA256 digest either way.
+	EngineVersion string `json:"engineVersion,omitempty"`
+	// NumThreads sets the local engine's thread count (0 = engine default).
+	NumThreads int `json:"numThreads,omitempty"`
+	// Language optionally constrains recognition (ISO-639-1, e.g. "en"). Empty =
+	// auto-detect.
+	Language string `json:"language,omitempty"`
+	// MaxDurationSeconds is the hard runaway-recording cap on every platform
+	// (0 = default 300s).
+	MaxDurationSeconds int `json:"maxDurationSeconds,omitempty"`
+	// SilenceAutoStop ends a recording ~2s after the signal goes quiet, as a
+	// backstop for a forgotten manual stop (not VAD-triggered start). Defaults on.
+	SilenceAutoStop *bool `json:"silenceAutoStop,omitempty"`
+	// AutoSubmit fires the transcript at the agent instead of inserting it for
+	// review. Defaults OFF — insert-for-review is the safety net for a misheard
+	// prompt (§3).
+	AutoSubmit *bool `json:"autoSubmit,omitempty"`
+	// WindowsAudioDevice names the dshow capture device (auto-detected when empty).
+	WindowsAudioDevice string `json:"windowsAudioDevice,omitempty"`
+}
+
+// STTProvider returns the configured batch provider, defaulting to local.
+func (c STTConfig) STTProvider() STTProviderKind {
+	if c.Provider == "" {
+		return STTProviderLocal
+	}
+	return c.Provider
+}
+
+// STTStreamProvider returns the configured streaming provider, defaulting to local.
+func (c STTConfig) STTStreamProvider() STTProviderKind {
+	if c.StreamProvider == "" {
+		return STTProviderLocal
+	}
+	return c.StreamProvider
+}
+
+// StreamingEnabled reports whether the live pipeline is on. Unset defaults ON.
+func (c STTConfig) StreamingEnabled() bool {
+	return c.Streaming == nil || *c.Streaming
+}
+
+// SilenceAutoStopEnabled reports whether trailing-silence auto-stop is on.
+// Unset defaults ON.
+func (c STTConfig) SilenceAutoStopEnabled() bool {
+	return c.SilenceAutoStop == nil || *c.SilenceAutoStop
+}
+
+// AutoSubmitEnabled reports whether transcripts are auto-fired. Unset defaults OFF.
+func (c STTConfig) AutoSubmitEnabled() bool {
+	return c.AutoSubmit != nil && *c.AutoSubmit
+}
+
+// Empty reports whether the STT config carries no user-set values (so it can be
+// omitted from a marshaled config).
+func (c STTConfig) Empty() bool {
+	return c.Provider == "" && c.StreamProvider == "" && c.Streaming == nil &&
+		c.Model == "" && c.StreamModel == "" && c.LocalModelPath == "" &&
+		c.LocalBinary == "" && c.LocalServerBinary == "" && c.LocalServerPort == 0 &&
+		c.EngineVersion == "" && c.NumThreads == 0 && c.Language == "" &&
+		c.MaxDurationSeconds == 0 && c.SilenceAutoStop == nil && c.AutoSubmit == nil &&
+		c.WindowsAudioDevice == ""
+}
+
 // LocalControlConfig controls local browser/desktop/terminal automation helpers.
 // Helpers are discovered lazily by the tool that needs them; no setup command or
 // background probe is required during startup.
@@ -227,6 +353,7 @@ type FileConfig struct {
 	Preferences    PreferencesConfig  `json:"preferences,omitempty"`
 	KeyBindings    KeyBindingsConfig  `json:"keybindings,omitempty"`
 	LocalControl   LocalControlConfig `json:"localControl,omitempty"`
+	STT            STTConfig          `json:"stt,omitempty"`
 }
 
 func (cfg FileConfig) MarshalJSON() ([]byte, error) {
@@ -242,6 +369,7 @@ func (cfg FileConfig) MarshalJSON() ([]byte, error) {
 		Preferences    PreferencesConfig   `json:"preferences,omitempty"`
 		KeyBindings    KeyBindingsConfig   `json:"keybindings,omitempty"`
 		LocalControl   *LocalControlConfig `json:"localControl,omitempty"`
+		STT            *STTConfig          `json:"stt,omitempty"`
 	}
 	raw := rawConfig{
 		ActiveProvider: cfg.ActiveProvider,
@@ -258,6 +386,9 @@ func (cfg FileConfig) MarshalJSON() ([]byte, error) {
 	if !cfg.LocalControl.Empty() {
 		raw.LocalControl = &cfg.LocalControl
 	}
+	if !cfg.STT.Empty() {
+		raw.STT = &cfg.STT
+	}
 	return json.Marshal(raw)
 }
 
@@ -267,6 +398,11 @@ type ResolveOptions struct {
 	ProviderCommand   string
 	Env               map[string]string
 	Overrides         Overrides
+	// ExcludeProject drops the project config layer (ProjectConfigPath) from MCP
+	// resolution when the workspace is untrusted, so a cloned repo's ./.zero/config.json
+	// cannot spawn stdio MCP servers. It is fail-closed: only a trusted workspace sets
+	// it false. Mirrors the ExcludeProject option hooks and plugins already honor.
+	ExcludeProject bool
 }
 
 type Overrides struct {
@@ -280,6 +416,7 @@ type Overrides struct {
 	Tools          ToolsConfig
 	KeyBindings    KeyBindingsConfig
 	LocalControl   LocalControlConfig
+	STT            STTConfig
 }
 
 type ResolvedConfig struct {
@@ -295,6 +432,7 @@ type ResolvedConfig struct {
 	Preferences    PreferencesConfig
 	KeyBindings    KeyBindingsConfig
 	LocalControl   LocalControlConfig
+	STT            STTConfig
 }
 
 type MCPConfig struct {
@@ -302,16 +440,29 @@ 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
+	// DefaultMCPServers() is never unmarshaled from JSON, so it starts false;
+	// any explicit entry in the user/project file — even one that happens to
+	// repeat a default's exact field values (e.g. re-declaring firecrawl's
+	// default URL) — sets it true. IsUnconfiguredDefault checks this alongside
+	// a resolved-value comparison, so redeclaring default values verbatim still
+	// counts as user-configured.
+	configured bool
 }
 
 // MCPOAuthConfig describes how to authenticate to a remote MCP server using an
@@ -342,6 +493,7 @@ func (cfg *FileConfig) UnmarshalJSON(data []byte) error {
 		Preferences     PreferencesConfig          `json:"preferences"`
 		KeyBindings     KeyBindingsConfig          `json:"keybindings"`
 		LocalControl    LocalControlConfig         `json:"localControl"`
+		STT             STTConfig                  `json:"stt"`
 		MCPServers      map[string]MCPServerConfig `json:"mcpServers"`
 		MCPServersSnake map[string]MCPServerConfig `json:"mcp_servers"`
 	}
@@ -369,6 +521,7 @@ func (cfg *FileConfig) UnmarshalJSON(data []byte) error {
 	cfg.Preferences = raw.Preferences
 	cfg.KeyBindings = raw.KeyBindings
 	cfg.LocalControl = raw.LocalControl
+	cfg.STT = raw.STT
 	if cfg.MCP.Servers == nil && (len(raw.MCPServers) > 0 || len(raw.MCPServersSnake) > 0) {
 		cfg.MCP.Servers = map[string]MCPServerConfig{}
 	}
@@ -503,6 +656,10 @@ func (server *MCPServerConfig) UnmarshalJSON(data []byte) error {
 		server.Disabled = *raw.Disabled
 		server.disabledSet = true
 	}
+	// This method only runs when the user's JSON actually has an object for this
+	// server key (built-in defaults are seeded as Go struct literals, never
+	// unmarshaled), so reaching here at all means the user configured it.
+	server.configured = true
 	return nil
 }
 

internal/config/validate.go

@@ -54,5 +54,8 @@ func validateSemantics(cfg FileConfig) []Issue {
 		// normalizeProviders already redacts secrets via providerError.
 		return []Issue{{FieldPath: "providers", Message: err.Error()}}
 	}
+	if err := validateSTTConfig(cfg.STT); err != nil {
+		return []Issue{{FieldPath: "stt", Message: err.Error()}}
+	}
 	return nil
 }

internal/config/writer.go

@@ -7,6 +7,8 @@ import (
 	"path/filepath"
 	"sort"
 	"strings"
+
+	"github.com/Gitlawb/zero/internal/providercatalog"
 )
 
 func UpsertProvider(path string, profile ProviderProfile, setActive bool) (FileConfig, error) {
@@ -29,6 +31,20 @@ func UpsertProvider(path string, profile ProviderProfile, setActive bool) (FileC
 	}
 
 	mergeProvider(&cfg, profile)
+	// mergeProfile deliberately ignores APIKeyStored — during resolve-time
+	// layering a project config must not be able to claim the user's stored
+	// keys. This user-config WRITE path re-applies the marker: capturing a key
+	// via SecureProviderProfile onto a previously env/no-key profile must
+	// persist apiKeyStored, or the secret sits in the credential store while
+	// every ApplyStoredAPIKey gate skips it (PR #560 review).
+	if profile.APIKeyStored {
+		for index := range cfg.Providers {
+			if cfg.Providers[index].Name == profile.Name {
+				cfg.Providers[index].APIKeyStored = true
+				break
+			}
+		}
+	}
 	if setActive || strings.TrimSpace(cfg.ActiveProvider) == "" {
 		cfg.ActiveProvider = profile.Name
 	}
@@ -39,6 +55,95 @@ func UpsertProvider(path string, profile ProviderProfile, setActive bool) (FileC
 	return cfg, nil
 }
 
+// EnsuredProvider reports the outcome of EnsureCatalogProvider: the profile name
+// that serves the catalog entry, whether it was newly created, and which provider
+// is active after the call (unchanged unless it was blank).
+type EnsuredProvider struct {
+	Name    string
+	Created bool
+	Active  string
+}
+
+// EnsureCatalogProvider guarantees a provider profile exists in the config at
+// path for the given catalog entry. OAuth login flows call this right after
+// storing a token: a login is only reachable from the provider list and
+// `zero providers use` when a profile exists, but a login must never replace or
+// deactivate the user's current active provider — so an existing profile whose
+// Name or CatalogID already matches is left completely untouched (its name,
+// credentials, and model are the user's), and a created profile is NOT marked
+// active unless no provider was active at all.
+func EnsureCatalogProvider(path string, catalogID string) (EnsuredProvider, error) {
+	path = strings.TrimSpace(path)
+	if path == "" {
+		return EnsuredProvider{}, fmt.Errorf("config path is required")
+	}
+	descriptor, err := providercatalog.Require(catalogID)
+	if err != nil {
+		return EnsuredProvider{}, err
+	}
+
+	cfg := FileConfig{}
+	if data, err := os.ReadFile(path); err == nil {
+		if err := json.Unmarshal(data, &cfg); err != nil {
+			return EnsuredProvider{}, fmt.Errorf("invalid config JSON %s: %w", path, err)
+		}
+	} else if !os.IsNotExist(err) {
+		return EnsuredProvider{}, fmt.Errorf("read config %s: %w", path, err)
+	}
+	for _, provider := range cfg.Providers {
+		if strings.EqualFold(strings.TrimSpace(provider.CatalogID), descriptor.ID) ||
+			strings.EqualFold(strings.TrimSpace(provider.Name), descriptor.ID) {
+			return EnsuredProvider{Name: provider.Name, Active: cfg.ActiveProvider}, nil
+		}
+	}
+
+	profile := ProviderProfile{
+		Name:         descriptor.ID,
+		ProviderKind: providerKindForCatalogTransport(descriptor.Transport),
+		CatalogID:    descriptor.ID,
+		BaseURL:      descriptor.DefaultBaseURL,
+		Model:        descriptor.DefaultModel,
+	}
+	written, err := UpsertProvider(path, profile, false)
+	if err != nil {
+		return EnsuredProvider{}, err
+	}
+	return EnsuredProvider{Name: profile.Name, Created: true, Active: written.ActiveProvider}, nil
+}
+
+// MarkProviderAPIKeyStored records that provider's API key now lives in the
+// credential store. It also clears inline/env key fields so the stored key is the
+// runtime credential; an old apiKeyEnv value must not keep overriding a freshly
+// captured key from `zero auth openrouter` or provider setup.
+func MarkProviderAPIKeyStored(path string, provider string) error {
+	path = strings.TrimSpace(path)
+	if path == "" {
+		return fmt.Errorf("config path is required")
+	}
+	provider = strings.TrimSpace(provider)
+	if provider == "" {
+		return fmt.Errorf("provider name is required")
+	}
+
+	data, err := os.ReadFile(path)
+	if err != nil {
+		return fmt.Errorf("read config %s: %w", path, err)
+	}
+	var cfg FileConfig
+	if err := json.Unmarshal(data, &cfg); err != nil {
+		return fmt.Errorf("invalid config JSON %s: %w", path, err)
+	}
+	for index := range cfg.Providers {
+		if strings.EqualFold(strings.TrimSpace(cfg.Providers[index].Name), provider) {
+			cfg.Providers[index].APIKey = ""
+			cfg.Providers[index].APIKeyEnv = ""
+			cfg.Providers[index].APIKeyStored = true
+			return writeConfigFile(path, cfg)
+		}
+	}
+	return fmt.Errorf("provider %q not found", provider)
+}
+
 func SetActiveProvider(path string, name string) (FileConfig, error) {
 	path = strings.TrimSpace(path)
 	if path == "" {
@@ -72,6 +177,296 @@ func SetActiveProvider(path string, name string) (FileConfig, error) {
 	return FileConfig{}, fmt.Errorf("provider %q not found", name)
 }
 
+// RemoveProvider deletes the named provider profile from the config at path.
+// When the removed profile was active, activeProvider hands off to the first
+// remaining provider (or clears when none remain) so the config never points at
+// a profile that no longer exists. The caller owns cleaning up the credential
+// store entry — config stays pure of secret I/O on the read path, and remove
+// keeps that symmetry by only touching config.json.
+func RemoveProvider(path string, name string) (FileConfig, error) {
+	path = strings.TrimSpace(path)
+	if path == "" {
+		return FileConfig{}, fmt.Errorf("config path is required")
+	}
+	name = strings.TrimSpace(name)
+	if name == "" {
+		return FileConfig{}, fmt.Errorf("provider name is required")
+	}
+
+	data, err := os.ReadFile(path)
+	if err != nil {
+		return FileConfig{}, fmt.Errorf("read config %s: %w", path, err)
+	}
+	cfg := FileConfig{}
+	if err := json.Unmarshal(data, &cfg); err != nil {
+		return FileConfig{}, fmt.Errorf("invalid config JSON %s: %w", path, err)
+	}
+
+	index := -1
+	for i, provider := range cfg.Providers {
+		if strings.EqualFold(strings.TrimSpace(provider.Name), name) {
+			index = i
+			break
+		}
+	}
+	if index < 0 {
+		return FileConfig{}, fmt.Errorf("provider %q not found", name)
+	}
+	removed := cfg.Providers[index]
+	cfg.Providers = append(cfg.Providers[:index], cfg.Providers[index+1:]...)
+	if strings.EqualFold(strings.TrimSpace(cfg.ActiveProvider), strings.TrimSpace(removed.Name)) {
+		cfg.ActiveProvider = ""
+		if len(cfg.Providers) > 0 {
+			cfg.ActiveProvider = cfg.Providers[0].Name
+		}
+	}
+	if err := writeConfigFile(path, cfg); err != nil {
+		return FileConfig{}, err
+	}
+	return cfg, nil
+}
+
+// RenameProvider renames a provider profile, keeping everything keyed by the
+// profile name consistent: the activeProvider pointer follows the rename, and a
+// key in the encrypted credential store (APIKeyStored) is migrated to the new
+// name BEFORE the config is rewritten — the store write must succeed first so a
+// failed migration never strands the config pointing at a key that no longer
+// resolves. OAuth tokens are deliberately not migrated: the runtime's login
+// candidates fall back to the profile's CatalogID, which every OAuth-capable
+// catalog profile carries, so a rename keeps the login reachable.
+func RenameProvider(path string, oldName string, newName string) (FileConfig, error) {
+	path = strings.TrimSpace(path)
+	if path == "" {
+		return FileConfig{}, fmt.Errorf("config path is required")
+	}
+	oldName = strings.TrimSpace(oldName)
+	newName = strings.TrimSpace(newName)
+	if oldName == "" || newName == "" {
+		return FileConfig{}, fmt.Errorf("provider names are required")
+	}
+
+	data, err := os.ReadFile(path)
+	if err != nil {
+		return FileConfig{}, fmt.Errorf("read config %s: %w", path, err)
+	}
+	cfg := FileConfig{}
+	if err := json.Unmarshal(data, &cfg); err != nil {
+		return FileConfig{}, fmt.Errorf("invalid config JSON %s: %w", path, err)
+	}
+
+	index := -1
+	for i, provider := range cfg.Providers {
+		providerName := strings.TrimSpace(provider.Name)
+		if strings.EqualFold(providerName, oldName) {
+			index = i
+			continue
+		}
+		if strings.EqualFold(providerName, newName) {
+			return FileConfig{}, fmt.Errorf("provider %q already exists", newName)
+		}
+	}
+	if index < 0 {
+		return FileConfig{}, fmt.Errorf("provider %q not found", oldName)
+	}
+	if strings.EqualFold(oldName, newName) && cfg.Providers[index].Name == newName {
+		return cfg, nil
+	}
+
+	previousName := cfg.Providers[index].Name
+	keyMigrated := false
+	if cfg.Providers[index].APIKeyStored {
+		if err := migrateStoredProviderKey(path, previousName, newName); err != nil {
+			return FileConfig{}, fmt.Errorf("migrate stored key for %q: %w", oldName, err)
+		}
+		keyMigrated = true
+	}
+	if strings.EqualFold(strings.TrimSpace(cfg.ActiveProvider), strings.TrimSpace(previousName)) {
+		cfg.ActiveProvider = newName
+	}
+	cfg.Providers[index].Name = newName
+	if err := writeConfigFile(path, cfg); err != nil {
+		if keyMigrated {
+			// Compensate best-effort: config.json still names the OLD profile, so
+			// move the key back where that config can find it — otherwise a failed
+			// rewrite strands the key under a name no profile carries.
+			_ = migrateStoredProviderKey(path, newName, previousName)
+		}
+		return FileConfig{}, err
+	}
+	return cfg, nil
+}
+
+// ProviderEdit is a field-level edit of one saved provider, applied by
+// EditProvider in a single atomic write. Name is the CURRENT profile name
+// (matched case-insensitively); NewName renames (case-only renames included).
+// Empty BaseURL/Model/APIKey mean "leave unchanged"; Description is applied
+// VERBATIM (the editor always knows the full desired text, so clearing works).
+type ProviderEdit struct {
+	Name         string
+	NewName      string
+	BaseURL      string
+	Model        string
+	APIKey       string
+	APIKeyStored bool
+	Description  string
+}
+
+// EditProvider applies a provider edit in ONE read-modify-write: rename
+// (activeProvider follows; a stored key migrates, with a best-effort rollback
+// if the config write fails), field updates, the stored-key marker, and the
+// verbatim description. A single write keeps the operation atomic — the
+// previous rename+upsert+describe sequence could fail halfway and leave
+// config.json renamed while every in-memory consumer still held the old name —
+// and, unlike UpsertProvider's exact-name merge, the case-insensitive match
+// here makes a case-only rename (groq -> Groq) an in-place update instead of
+// an appended duplicate profile.
+func EditProvider(path string, edit ProviderEdit) (FileConfig, error) {
+	path = strings.TrimSpace(path)
+	if path == "" {
+		return FileConfig{}, fmt.Errorf("config path is required")
+	}
+	oldName := strings.TrimSpace(edit.Name)
+	if oldName == "" {
+		return FileConfig{}, fmt.Errorf("provider name is required")
+	}
+	newName := strings.TrimSpace(edit.NewName)
+	if newName == "" {
+		newName = oldName
+	}
+
+	data, err := os.ReadFile(path)
+	if err != nil {
+		return FileConfig{}, fmt.Errorf("read config %s: %w", path, err)
+	}
+	cfg := FileConfig{}
+	if err := json.Unmarshal(data, &cfg); err != nil {
+		return FileConfig{}, fmt.Errorf("invalid config JSON %s: %w", path, err)
+	}
+
+	index := -1
+	for i, provider := range cfg.Providers {
+		providerName := strings.TrimSpace(provider.Name)
+		if strings.EqualFold(providerName, oldName) {
+			index = i
+			continue
+		}
+		if strings.EqualFold(providerName, newName) {
+			return FileConfig{}, fmt.Errorf("provider %q already exists", newName)
+		}
+	}
+	if index < 0 {
+		return FileConfig{}, fmt.Errorf("provider %q not found", oldName)
+	}
+
+	previousName := cfg.Providers[index].Name
+	renamed := previousName != newName
+	keyMigrated := false
+	// A rename moves the stored key along: either the profile's existing entry,
+	// or a replacement key the caller just captured — the contract is that a
+	// captured key is stored under the CURRENT name before EditProvider runs, so
+	// one migration covers both. migrateStoredProviderKey no-ops on case-only
+	// renames (the store normalizes names), so it cannot delete the key it just
+	// moved.
+	if renamed && (cfg.Providers[index].APIKeyStored || edit.APIKeyStored) {
+		if err := migrateStoredProviderKey(path, previousName, newName); err != nil {
+			return FileConfig{}, fmt.Errorf("migrate stored key for %q: %w", oldName, err)
+		}
+		keyMigrated = true
+	}
+	if renamed && strings.EqualFold(strings.TrimSpace(cfg.ActiveProvider), strings.TrimSpace(previousName)) {
+		cfg.ActiveProvider = newName
+	}
+
+	profile := &cfg.Providers[index]
+	profile.Name = newName
+	if baseURL := strings.TrimSpace(edit.BaseURL); baseURL != "" {
+		profile.BaseURL = baseURL
+	}
+	if model := strings.TrimSpace(edit.Model); model != "" {
+		profile.Model = model
+	}
+	if apiKey := strings.TrimSpace(edit.APIKey); apiKey != "" {
+		profile.APIKey = apiKey
+	}
+	if edit.APIKeyStored {
+		profile.APIKeyStored = true
+	}
+	profile.Description = strings.TrimSpace(edit.Description)
+
+	if err := writeConfigFile(path, cfg); err != nil {
+		if keyMigrated {
+			// Compensate best-effort: config.json still names the OLD profile, so
+			// move the key back where that config can find it.
+			_ = migrateStoredProviderKey(path, newName, previousName)
+		}
+		return FileConfig{}, err
+	}
+	return cfg, nil
+}
+
+// SetProviderDescription sets a provider's description VERBATIM — including to
+// empty. The generic UpsertProvider merge treats empty fields as "leave
+// unchanged", so clearing a description needs this dedicated setter.
+func SetProviderDescription(path string, name string, description string) (FileConfig, error) {
+	path = strings.TrimSpace(path)
+	if path == "" {
+		return FileConfig{}, fmt.Errorf("config path is required")
+	}
+	name = strings.TrimSpace(name)
+	if name == "" {
+		return FileConfig{}, fmt.Errorf("provider name is required")
+	}
+
+	data, err := os.ReadFile(path)
+	if err != nil {
+		return FileConfig{}, fmt.Errorf("read config %s: %w", path, err)
+	}
+	cfg := FileConfig{}
+	if err := json.Unmarshal(data, &cfg); err != nil {
+		return FileConfig{}, fmt.Errorf("invalid config JSON %s: %w", path, err)
+	}
+	for index := range cfg.Providers {
+		if strings.EqualFold(strings.TrimSpace(cfg.Providers[index].Name), name) {
+			cfg.Providers[index].Description = strings.TrimSpace(description)
+			if err := writeConfigFile(path, cfg); err != nil {
+				return FileConfig{}, err
+			}
+			return cfg, nil
+		}
+	}
+	return FileConfig{}, fmt.Errorf("provider %q not found", name)
+}
+
+// migrateStoredProviderKey moves a credential-store entry to a new provider
+// name: write-new-then-delete-old, so an interruption can leave a duplicate but
+// never a missing key. A missing source entry is a no-op (the marker may be
+// stale); only a failed WRITE aborts the rename.
+func migrateStoredProviderKey(configPath string, oldName string, newName string) error {
+	// The store normalizes names case-insensitively, so a case-only rename
+	// (groq -> Groq) targets ONE entry: Set(new) rewrites it in place and
+	// Delete(old) would then remove the key that was just "moved". Nothing to
+	// migrate — the existing entry already serves the new name.
+	if strings.EqualFold(strings.TrimSpace(oldName), strings.TrimSpace(newName)) {
+		return nil
+	}
+	store, err := ProviderKeyStoreAt(filepath.Dir(configPath))
+	if err != nil {
+		return err
+	}
+	key, ok, err := store.Get(oldName)
+	if err != nil {
+		return err
+	}
+	if !ok || strings.TrimSpace(key) == "" {
+		return nil
+	}
+	if err := store.Set(newName, key); err != nil {
+		return err
+	}
+	_, _ = store.Delete(oldName)
+	return nil
+}
+
 func SetProviderModel(path string, name string, model string) (FileConfig, error) {
 	path = strings.TrimSpace(path)
 	if path == "" {
@@ -131,6 +526,32 @@ func SetFavoriteModels(path string, models []string) (FileConfig, error) {
 	return cfg, nil
 }
 
+// SetRecentModels persists the automatic recent-model-switch history,
+// mirroring SetFavoriteModels (read-modify-atomic-write). Unlike favorites,
+// order is preserved (newest first) rather than sorted, since it reflects
+// switch recency, not an alphabetical preference list.
+func SetRecentModels(path string, entries []RecentModelEntry) (FileConfig, error) {
+	path = strings.TrimSpace(path)
+	if path == "" {
+		return FileConfig{}, fmt.Errorf("config path is required")
+	}
+
+	cfg := FileConfig{}
+	if data, err := os.ReadFile(path); err == nil {
+		if err := json.Unmarshal(data, &cfg); err != nil {
+			return FileConfig{}, fmt.Errorf("invalid config JSON %s: %w", path, err)
+		}
+	} else if !os.IsNotExist(err) {
+		return FileConfig{}, fmt.Errorf("read config %s: %w", path, err)
+	}
+
+	cfg.Preferences.RecentModels = NormalizeRecentModels(entries)
+	if err := writeConfigFile(path, cfg); err != nil {
+		return FileConfig{}, err
+	}
+	return cfg, nil
+}
+
 // SetRecapsEnabled persists the post-turn recap preference, mirroring
 // SetFavoriteModels (read-modify-atomic-write).
 func SetRecapsEnabled(path string, enabled bool) (FileConfig, error) {
@@ -176,6 +597,106 @@ func SetTheme(path string, theme string) (FileConfig, error) {
 	return cfg, nil
 }
 
+// SetSTTModel persists the dictation model and its provider, mirroring
+// SetTheme (read-modify-atomic-write). provider must be one of the known STT
+// provider kinds; a local provider stores the model as stt.localModelPath,
+// otherwise as stt.model. A blank model clears the stored value for that slot.
+func SetSTTModel(path string, provider STTProviderKind, model string) (FileConfig, error) {
+	path = strings.TrimSpace(path)
+	if path == "" {
+		return FileConfig{}, fmt.Errorf("config path is required")
+	}
+	cfg := FileConfig{}
+	if data, err := os.ReadFile(path); err == nil {
+		if err := json.Unmarshal(data, &cfg); err != nil {
+			return FileConfig{}, fmt.Errorf("invalid config JSON %s: %w", path, err)
+		}
+	} else if !os.IsNotExist(err) {
+		return FileConfig{}, fmt.Errorf("read config %s: %w", path, err)
+	}
+	if provider != "" {
+		cfg.STT.Provider = provider
+	}
+	model = strings.TrimSpace(model)
+	if cfg.STT.STTProvider() == STTProviderLocal {
+		cfg.STT.LocalModelPath = model
+	} else {
+		cfg.STT.Model = model
+	}
+	if err := validateSTTConfig(cfg.STT); err != nil {
+		return FileConfig{}, err
+	}
+	if err := writeConfigFile(path, cfg); err != nil {
+		return FileConfig{}, err
+	}
+	return cfg, nil
+}
+
+// SetSTTLocalEngine persists the paths of an auto-downloaded local engine +
+// model and switches dictation to the local provider, mirroring SetTheme
+// (read-modify-atomic-write). streaming selects the pipeline matching the
+// downloaded model (a streaming transducer vs a batch model). Called after a
+// download completes.
+func SetSTTLocalEngine(path, binary, serverBinary, modelPath string, streaming bool) (FileConfig, error) {
+	path = strings.TrimSpace(path)
+	if path == "" {
+		return FileConfig{}, fmt.Errorf("config path is required")
+	}
+	cfg := FileConfig{}
+	if data, err := os.ReadFile(path); err == nil {
+		if err := json.Unmarshal(data, &cfg); err != nil {
+			return FileConfig{}, fmt.Errorf("invalid config JSON %s: %w", path, err)
+		}
+	} else if !os.IsNotExist(err) {
+		return FileConfig{}, fmt.Errorf("read config %s: %w", path, err)
+	}
+	cfg.STT.Provider = STTProviderLocal
+	// Point BOTH pipelines at the local engine. Without resetting StreamProvider, a
+	// previously-chosen cloud value (deepgram/openai) would still win in
+	// buildStreamingTranscriber, so the live transcript would keep hitting the cloud
+	// after the user switched to a downloaded local model.
+	cfg.STT.StreamProvider = STTProviderLocal
+	cfg.STT.LocalBinary = strings.TrimSpace(binary)
+	cfg.STT.LocalServerBinary = strings.TrimSpace(serverBinary)
+	cfg.STT.LocalModelPath = strings.TrimSpace(modelPath)
+	// Match the pipeline to the downloaded model: a streaming transducer drives
+	// the websocket server for a live transcript; a batch model uses the offline
+	// binary.
+	s := streaming
+	cfg.STT.Streaming = &s
+	if err := validateSTTConfig(cfg.STT); err != nil {
+		return FileConfig{}, err
+	}
+	if err := writeConfigFile(path, cfg); err != nil {
+		return FileConfig{}, err
+	}
+	return cfg, nil
+}
+
+// SetSTTProvider persists just the dictation batch provider, mirroring SetTheme.
+func SetSTTProvider(path string, provider STTProviderKind) (FileConfig, error) {
+	path = strings.TrimSpace(path)
+	if path == "" {
+		return FileConfig{}, fmt.Errorf("config path is required")
+	}
+	cfg := FileConfig{}
+	if data, err := os.ReadFile(path); err == nil {
+		if err := json.Unmarshal(data, &cfg); err != nil {
+			return FileConfig{}, fmt.Errorf("invalid config JSON %s: %w", path, err)
+		}
+	} else if !os.IsNotExist(err) {
+		return FileConfig{}, fmt.Errorf("read config %s: %w", path, err)
+	}
+	cfg.STT.Provider = provider
+	if err := validateSTTConfig(cfg.STT); err != nil {
+		return FileConfig{}, err
+	}
+	if err := writeConfigFile(path, cfg); err != nil {
+		return FileConfig{}, err
+	}
+	return cfg, nil
+}
+
 func normalizeFavoriteModels(models []string) []string {
 	seen := map[string]bool{}
 	favorites := make([]string, 0, len(models))
@@ -191,6 +712,35 @@ func normalizeFavoriteModels(models []string) []string {
 	return favorites
 }
 
+// NormalizeRecentModels trims, drops entries with no model id, de-duplicates
+// by provider+model pair (keeping the first/newest occurrence), and caps the
+// result to MaxRecentModels. Order is preserved — the caller is responsible
+// for passing entries newest-first. Exported so callers outside this package
+// (e.g. the TUI, which keeps its own in-memory copy of recent history) apply
+// the exact same normalization rules as the persisted config, instead of
+// maintaining a second, independently-drifting copy of this logic.
+func NormalizeRecentModels(entries []RecentModelEntry) []RecentModelEntry {
+	seen := map[string]bool{}
+	recent := make([]RecentModelEntry, 0, len(entries))
+	for _, entry := range entries {
+		provider := strings.TrimSpace(entry.Provider)
+		model := strings.TrimSpace(entry.Model)
+		if model == "" {
+			continue
+		}
+		key := strings.ToLower(provider) + "\x00" + model
+		if seen[key] {
+			continue
+		}
+		seen[key] = true
+		recent = append(recent, RecentModelEntry{Provider: provider, Model: model})
+		if len(recent) >= MaxRecentModels {
+			break
+		}
+	}
+	return recent
+}
+
 func writeConfigFile(path string, cfg FileConfig) error {
 	dir := filepath.Dir(path)
 	if dir != "." && dir != "" {

internal/config/writer_test.go

@@ -5,6 +5,7 @@ import (
 	"errors"
 	"io/fs"
 	"os"
+	"os/exec"
 	"path/filepath"
 	"reflect"
 	"runtime"
@@ -272,6 +273,69 @@ func TestSetFavoriteModelsPersistsUserPreferences(t *testing.T) {
 	}
 }
 
+func TestSetRecentModelsPersistsOrderDedupesAndCaps(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "zero.json")
+	writeConfigFixture(t, path, FileConfig{
+		ActiveProvider: "openai",
+		Providers: []ProviderProfile{
+			{Name: "openai", ProviderKind: ProviderKindOpenAI, Model: "gpt-4.1"},
+		},
+	}, 0o600)
+
+	cfg, err := SetRecentModels(path, []RecentModelEntry{
+		{Provider: " openrouter ", Model: " google/gemini-2.5-pro "},
+		{Provider: "openrouter", Model: "minimax/minimax-m2.1"},
+		{Provider: "openrouter", Model: "google/gemini-2.5-pro"}, // duplicate pair, older: dropped
+		{Provider: "", Model: ""},                                // blank model: dropped
+		{Provider: "openrouter", Model: "a"},
+		{Provider: "openrouter", Model: "b"},
+		{Provider: "openrouter", Model: "c"},
+		{Provider: "openrouter", Model: "d"}, // beyond MaxRecentModels (5): dropped
+	})
+	if err != nil {
+		t.Fatalf("SetRecentModels() error = %v", err)
+	}
+
+	want := []RecentModelEntry{
+		{Provider: "openrouter", Model: "google/gemini-2.5-pro"},
+		{Provider: "openrouter", Model: "minimax/minimax-m2.1"},
+		{Provider: "openrouter", Model: "a"},
+		{Provider: "openrouter", Model: "b"},
+		{Provider: "openrouter", Model: "c"},
+	}
+	if !reflect.DeepEqual(cfg.Preferences.RecentModels, want) {
+		t.Fatalf("RecentModels = %#v, want %#v", cfg.Preferences.RecentModels, want)
+	}
+	persisted := readConfigFixture(t, path)
+	if !reflect.DeepEqual(persisted.Preferences.RecentModels, want) {
+		t.Fatalf("persisted RecentModels = %#v, want %#v", persisted.Preferences.RecentModels, want)
+	}
+	if persisted.ActiveProvider != "openai" || len(persisted.Providers) != 1 {
+		t.Fatalf("provider config was not preserved: %#v", persisted)
+	}
+}
+
+// Two providers offering the same model id must both survive normalization —
+// recent history de-duplicates by provider+model pair, not model id alone.
+func TestSetRecentModelsDedupesByProviderAndModelPair(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "zero.json")
+
+	cfg, err := SetRecentModels(path, []RecentModelEntry{
+		{Provider: "provider-a", Model: "shared-model"},
+		{Provider: "provider-b", Model: "shared-model"},
+	})
+	if err != nil {
+		t.Fatalf("SetRecentModels() error = %v", err)
+	}
+	want := []RecentModelEntry{
+		{Provider: "provider-a", Model: "shared-model"},
+		{Provider: "provider-b", Model: "shared-model"},
+	}
+	if !reflect.DeepEqual(cfg.Preferences.RecentModels, want) {
+		t.Fatalf("RecentModels = %#v, want both providers preserved: %#v", cfg.Preferences.RecentModels, want)
+	}
+}
+
 func TestSetThemePersistsUserPreference(t *testing.T) {
 	path := filepath.Join(t.TempDir(), "zero.json")
 	writeConfigFixture(t, path, FileConfig{
@@ -385,3 +449,575 @@ func readConfigFixture(t *testing.T, path string) FileConfig {
 	}
 	return cfg
 }
+
+func TestEnsureCatalogProviderCreatesProfileWithoutStealingActive(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "zero.json")
+	writeConfigFixture(t, path, FileConfig{
+		ActiveProvider: "opengateway",
+		Providers: []ProviderProfile{
+			{
+				Name:         "opengateway",
+				ProviderKind: ProviderKindOpenAICompatible,
+				BaseURL:      "https://gateway.example.com/v1",
+				APIKeyStored: true,
+				Model:        "some-model",
+			},
+		},
+	}, 0o600)
+
+	ensured, err := EnsureCatalogProvider(path, "chatgpt")
+	if err != nil {
+		t.Fatalf("EnsureCatalogProvider: %v", err)
+	}
+	if !ensured.Created {
+		t.Fatalf("expected profile to be created")
+	}
+	if ensured.Name != "chatgpt" {
+		t.Fatalf("expected profile name chatgpt, got %q", ensured.Name)
+	}
+	if ensured.Active != "opengateway" {
+		t.Fatalf("active provider must stay opengateway, got %q", ensured.Active)
+	}
+
+	cfg := readConfigFixture(t, path)
+	if cfg.ActiveProvider != "opengateway" {
+		t.Fatalf("persisted active provider changed to %q", cfg.ActiveProvider)
+	}
+	if len(cfg.Providers) != 2 {
+		t.Fatalf("expected 2 providers, got %d", len(cfg.Providers))
+	}
+	chatgpt := cfg.Providers[1]
+	if chatgpt.Name != "chatgpt" || chatgpt.CatalogID != "chatgpt" {
+		t.Fatalf("unexpected created profile: %+v", chatgpt)
+	}
+	if chatgpt.Model == "" || chatgpt.BaseURL == "" {
+		t.Fatalf("created profile must carry catalog defaults, got %+v", chatgpt)
+	}
+	if chatgpt.APIKey != "" || chatgpt.APIKeyStored {
+		t.Fatalf("OAuth-created profile must stay keyless, got %+v", chatgpt)
+	}
+}
+
+func TestEnsureCatalogProviderLeavesExistingProfileUntouched(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "zero.json")
+	original := FileConfig{
+		ActiveProvider: "opengateway",
+		Providers: []ProviderProfile{
+			{Name: "opengateway", ProviderKind: ProviderKindOpenAICompatible, BaseURL: "https://gateway.example.com/v1", Model: "some-model"},
+			// Renamed profile that already serves the chatgpt catalog entry.
+			{Name: "codex", CatalogID: "chatgpt", Model: "gpt-5.5"},
+		},
+	}
+	data := writeConfigFixture(t, path, original, 0o600)
+
+	ensured, err := EnsureCatalogProvider(path, "chatgpt")
+	if err != nil {
+		t.Fatalf("EnsureCatalogProvider: %v", err)
+	}
+	if ensured.Created {
+		t.Fatalf("existing profile must not be recreated")
+	}
+	if ensured.Name != "codex" {
+		t.Fatalf("expected existing profile name codex, got %q", ensured.Name)
+	}
+	after, err := os.ReadFile(path)
+	if err != nil {
+		t.Fatalf("read config: %v", err)
+	}
+	if string(after) != string(data) {
+		t.Fatalf("config rewritten for a no-op ensure:\nbefore: %s\nafter: %s", data, after)
+	}
+}
+
+func TestEnsureCatalogProviderActivatesOnEmptyConfig(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "zero.json")
+
+	ensured, err := EnsureCatalogProvider(path, "chatgpt")
+	if err != nil {
+		t.Fatalf("EnsureCatalogProvider: %v", err)
+	}
+	if !ensured.Created {
+		t.Fatalf("expected profile to be created")
+	}
+	if ensured.Active != "chatgpt" {
+		t.Fatalf("blank active must adopt the new provider, got %q", ensured.Active)
+	}
+}
+
+func TestEnsureCatalogProviderRejectsUnknownCatalogID(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "zero.json")
+	if _, err := EnsureCatalogProvider(path, "no-such-provider"); err == nil {
+		t.Fatalf("expected unknown catalog id to error")
+	}
+	if _, err := os.Stat(path); !errors.Is(err, fs.ErrNotExist) {
+		t.Fatalf("config must not be written for an unknown catalog id")
+	}
+}
+
+func TestMarkProviderAPIKeyStoredClearsInlineAndEnvKey(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "zero.json")
+	writeConfigFixture(t, path, FileConfig{
+		ActiveProvider: "openrouter",
+		Providers: []ProviderProfile{{
+			Name:      "openrouter",
+			CatalogID: "openrouter",
+			APIKey:    "sk-inline",
+			APIKeyEnv: "OPENROUTER_API_KEY",
+			Model:     "openai/gpt-4.1",
+		}},
+	}, 0o600)
+
+	if err := MarkProviderAPIKeyStored(path, "openrouter"); err != nil {
+		t.Fatalf("MarkProviderAPIKeyStored: %v", err)
+	}
+
+	cfg := readConfigFixture(t, path)
+	if len(cfg.Providers) != 1 {
+		t.Fatalf("providers = %#v", cfg.Providers)
+	}
+	profile := cfg.Providers[0]
+	if !profile.APIKeyStored || profile.APIKey != "" || profile.APIKeyEnv != "" {
+		t.Fatalf("provider credential fields = %#v", profile)
+	}
+	if profile.Model != "openai/gpt-4.1" || cfg.ActiveProvider != "openrouter" {
+		t.Fatalf("unrelated config changed: %#v", cfg)
+	}
+}
+
+func TestRemoveProviderDeletesAndHandsOffActive(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "zero.json")
+	writeConfigFixture(t, path, FileConfig{
+		ActiveProvider: "beta",
+		Providers: []ProviderProfile{
+			{Name: "alpha", ProviderKind: ProviderKindOpenAICompatible, BaseURL: "https://a.example.com/v1", Model: "m1"},
+			{Name: "beta", ProviderKind: ProviderKindOpenAICompatible, BaseURL: "https://b.example.com/v1", Model: "m2"},
+		},
+	}, 0o600)
+
+	cfg, err := RemoveProvider(path, " BETA ")
+	if err != nil {
+		t.Fatalf("RemoveProvider() error = %v", err)
+	}
+	if len(cfg.Providers) != 1 || cfg.Providers[0].Name != "alpha" {
+		t.Fatalf("expected only alpha to remain, got %+v", cfg.Providers)
+	}
+	if cfg.ActiveProvider != "alpha" {
+		t.Fatalf("active must hand off to the first remaining provider, got %q", cfg.ActiveProvider)
+	}
+
+	persisted := readConfigFixture(t, path)
+	if len(persisted.Providers) != 1 || persisted.ActiveProvider != "alpha" {
+		t.Fatalf("persisted config wrong: %+v", persisted)
+	}
+
+	// Removing the last provider clears the active pointer entirely.
+	cfg, err = RemoveProvider(path, "alpha")
+	if err != nil {
+		t.Fatalf("RemoveProvider(last) error = %v", err)
+	}
+	if len(cfg.Providers) != 0 || cfg.ActiveProvider != "" {
+		t.Fatalf("expected empty providers and no active, got %+v", cfg)
+	}
+}
+
+func TestRemoveProviderKeepsActiveWhenOtherRemoved(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "zero.json")
+	writeConfigFixture(t, path, FileConfig{
+		ActiveProvider: "alpha",
+		Providers: []ProviderProfile{
+			{Name: "alpha", ProviderKind: ProviderKindOpenAICompatible, BaseURL: "https://a.example.com/v1", Model: "m1"},
+			{Name: "beta", ProviderKind: ProviderKindOpenAICompatible, BaseURL: "https://b.example.com/v1", Model: "m2"},
+		},
+	}, 0o600)
+
+	cfg, err := RemoveProvider(path, "beta")
+	if err != nil {
+		t.Fatalf("RemoveProvider() error = %v", err)
+	}
+	if cfg.ActiveProvider != "alpha" {
+		t.Fatalf("active provider must be untouched, got %q", cfg.ActiveProvider)
+	}
+}
+
+func TestRemoveProviderRejectsUnknownWithoutRewriting(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "zero.json")
+	before := writeConfigFixture(t, path, FileConfig{
+		ActiveProvider: "alpha",
+		Providers:      []ProviderProfile{{Name: "alpha", ProviderKind: ProviderKindOpenAICompatible, BaseURL: "https://a.example.com/v1", Model: "m1"}},
+	}, 0o600)
+
+	if _, err := RemoveProvider(path, "ghost"); err == nil {
+		t.Fatal("RemoveProvider() error = nil, want not-found error")
+	}
+	after, err := os.ReadFile(path)
+	if err != nil {
+		t.Fatalf("read config: %v", err)
+	}
+	if string(after) != string(before) {
+		t.Fatalf("config was rewritten for unknown provider")
+	}
+}
+
+func TestRenameProviderFollowsActiveAndMigratesStoredKey(t *testing.T) {
+	dir := t.TempDir()
+	// Force the file credential backend so the test never touches the real OS
+	// keyring regardless of platform.
+	t.Setenv("ZERO_CRED_STORAGE", "encrypted-file")
+	path := filepath.Join(dir, "config.json")
+	writeConfigFixture(t, path, FileConfig{
+		ActiveProvider: "oldname",
+		Providers: []ProviderProfile{
+			{Name: "oldname", ProviderKind: ProviderKindOpenAICompatible, BaseURL: "https://a.example.com/v1", APIKeyStored: true, Model: "m1"},
+			{Name: "other", ProviderKind: ProviderKindOpenAICompatible, BaseURL: "https://b.example.com/v1", Model: "m2"},
+		},
+	}, 0o600)
+	store, err := ProviderKeyStoreAt(dir)
+	if err != nil {
+		t.Fatalf("open store: %v", err)
+	}
+	if err := store.Set("oldname", "sk-secret"); err != nil {
+		t.Fatalf("seed store: %v", err)
+	}
+
+	cfg, err := RenameProvider(path, "oldname", "newname")
+	if err != nil {
+		t.Fatalf("RenameProvider() error = %v", err)
+	}
+	if cfg.ActiveProvider != "newname" {
+		t.Fatalf("active must follow the rename, got %q", cfg.ActiveProvider)
+	}
+	if cfg.Providers[0].Name != "newname" || !cfg.Providers[0].APIKeyStored {
+		t.Fatalf("renamed profile wrong: %+v", cfg.Providers[0])
+	}
+	if cfg.Providers[1].Name != "other" {
+		t.Fatalf("unrelated profile changed: %+v", cfg.Providers[1])
+	}
+
+	key, ok, err := store.Get("newname")
+	if err != nil || !ok || key != "sk-secret" {
+		t.Fatalf("stored key must migrate to the new name, got key=%q ok=%v err=%v", key, ok, err)
+	}
+	if _, ok, _ := store.Get("oldname"); ok {
+		t.Fatalf("old credential entry must be deleted after migration")
+	}
+}
+
+func TestRenameProviderRejectsCollisionAndUnknown(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "zero.json")
+	before := writeConfigFixture(t, path, FileConfig{
+		ActiveProvider: "alpha",
+		Providers: []ProviderProfile{
+			{Name: "alpha", ProviderKind: ProviderKindOpenAICompatible, BaseURL: "https://a.example.com/v1", Model: "m1"},
+			{Name: "beta", ProviderKind: ProviderKindOpenAICompatible, BaseURL: "https://b.example.com/v1", Model: "m2"},
+		},
+	}, 0o600)
+
+	if _, err := RenameProvider(path, "alpha", "BETA"); err == nil {
+		t.Fatal("rename onto an existing name must fail")
+	}
+	if _, err := RenameProvider(path, "ghost", "gamma"); err == nil {
+		t.Fatal("renaming an unknown provider must fail")
+	}
+	after, err := os.ReadFile(path)
+	if err != nil {
+		t.Fatalf("read config: %v", err)
+	}
+	if string(after) != string(before) {
+		t.Fatalf("config was rewritten by a rejected rename")
+	}
+}
+
+func TestUpsertProviderPreservesStoredKeyMarkerOnExistingProfile(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "zero.json")
+	// An env-keyed profile with NO stored-key marker — the shape a provider has
+	// before its key is captured into the credential store.
+	writeConfigFixture(t, path, FileConfig{
+		ActiveProvider: "groq",
+		Providers: []ProviderProfile{
+			{Name: "groq", ProviderKind: ProviderKindOpenAICompatible, BaseURL: "https://api.groq.com/openai/v1", APIKeyEnv: "GROQ_API_KEY", Model: "m1"},
+		},
+	}, 0o600)
+
+	// The manager/setup edit paths persist a SecureProviderProfile-shaped
+	// profile: key already in the store, marker set, inline key cleared.
+	cfg, err := UpsertProvider(path, ProviderProfile{Name: "groq", APIKeyStored: true}, false)
+	if err != nil {
+		t.Fatalf("UpsertProvider() error = %v", err)
+	}
+	if !cfg.Providers[0].APIKeyStored {
+		t.Fatalf("APIKeyStored marker must survive the merge, got %+v", cfg.Providers[0])
+	}
+	if cfg.Providers[0].APIKeyEnv != "GROQ_API_KEY" || cfg.Providers[0].BaseURL == "" {
+		t.Fatalf("other fields must be preserved: %+v", cfg.Providers[0])
+	}
+	persisted := readConfigFixture(t, path)
+	if !persisted.Providers[0].APIKeyStored {
+		t.Fatalf("marker not persisted to disk: %+v", persisted.Providers[0])
+	}
+}
+
+func TestSetProviderDescriptionSetsAndClears(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "zero.json")
+	writeConfigFixture(t, path, FileConfig{
+		ActiveProvider: "alpha",
+		Providers: []ProviderProfile{
+			{Name: "alpha", ProviderKind: ProviderKindOpenAICompatible, BaseURL: "https://a.example.com/v1", Model: "m1", Description: "old text"},
+		},
+	}, 0o600)
+
+	cfg, err := SetProviderDescription(path, " ALPHA ", "new text")
+	if err != nil {
+		t.Fatalf("SetProviderDescription() error = %v", err)
+	}
+	if cfg.Providers[0].Description != "new text" {
+		t.Fatalf("description not set: %+v", cfg.Providers[0])
+	}
+
+	// Clearing must persist too — the reason this setter exists (UpsertProvider's
+	// merge treats an empty description as "leave unchanged").
+	cfg, err = SetProviderDescription(path, "alpha", "  ")
+	if err != nil {
+		t.Fatalf("SetProviderDescription(clear) error = %v", err)
+	}
+	if cfg.Providers[0].Description != "" {
+		t.Fatalf("description not cleared: %+v", cfg.Providers[0])
+	}
+	persisted := readConfigFixture(t, path)
+	if persisted.Providers[0].Description != "" {
+		t.Fatalf("cleared description not persisted: %+v", persisted.Providers[0])
+	}
+
+	if _, err := SetProviderDescription(path, "ghost", "x"); err == nil {
+		t.Fatal("unknown provider must error")
+	}
+}
+
+func TestRenameProviderRollsBackKeyMigrationWhenConfigWriteFails(t *testing.T) {
+	if runtime.GOOS != "darwin" {
+		t.Skip("uses chflags uchg to force the config write to fail; macOS only")
+	}
+	dir := t.TempDir()
+	t.Setenv("ZERO_CRED_STORAGE", "encrypted-file")
+	path := filepath.Join(dir, "config.json")
+	writeConfigFixture(t, path, FileConfig{
+		ActiveProvider: "oldname",
+		Providers: []ProviderProfile{
+			{Name: "oldname", ProviderKind: ProviderKindOpenAICompatible, BaseURL: "https://a.example.com/v1", APIKeyStored: true, Model: "m1"},
+		},
+	}, 0o600)
+	store, err := ProviderKeyStoreAt(dir)
+	if err != nil {
+		t.Fatalf("open store: %v", err)
+	}
+	if err := store.Set("oldname", "sk-secret"); err != nil {
+		t.Fatalf("seed store: %v", err)
+	}
+
+	// Immutable flag: temp-file creation and store writes in the directory keep
+	// working, but the final rename over config.json fails — the exact window
+	// where a migrated key would otherwise strand under the new name.
+	if out, err := exec.Command("chflags", "uchg", path).CombinedOutput(); err != nil {
+		t.Skipf("chflags uchg unavailable: %v (%s)", err, out)
+	}
+	t.Cleanup(func() { _ = exec.Command("chflags", "nouchg", path).Run() })
+
+	if _, err := RenameProvider(path, "oldname", "newname"); err == nil {
+		t.Fatal("expected the config write to fail under the immutable flag")
+	}
+
+	key, ok, err := store.Get("oldname")
+	if err != nil || !ok || key != "sk-secret" {
+		t.Fatalf("key must be rolled back to the old name, got key=%q ok=%v err=%v", key, ok, err)
+	}
+	if _, ok, _ := store.Get("newname"); ok {
+		t.Fatalf("rolled-back migration must not leave a key under the new name")
+	}
+}
+
+// TestRenameProviderCaseOnlyKeepsStoredKey: the credential store normalizes
+// names case-insensitively, so a case-only rename targets ONE store entry —
+// migrating would Set and then Delete the same key, losing it (PR #560 review).
+func TestRenameProviderCaseOnlyKeepsStoredKey(t *testing.T) {
+	dir := t.TempDir()
+	t.Setenv("ZERO_CRED_STORAGE", "encrypted-file")
+	path := filepath.Join(dir, "config.json")
+	writeConfigFixture(t, path, FileConfig{
+		ActiveProvider: "groq",
+		Providers: []ProviderProfile{
+			{Name: "groq", ProviderKind: ProviderKindOpenAICompatible, BaseURL: "https://api.groq.com/openai/v1", APIKeyStored: true, Model: "m1"},
+		},
+	}, 0o600)
+	store, err := ProviderKeyStoreAt(dir)
+	if err != nil {
+		t.Fatalf("open store: %v", err)
+	}
+	if err := store.Set("groq", "sk-secret"); err != nil {
+		t.Fatalf("seed store: %v", err)
+	}
+
+	cfg, err := RenameProvider(path, "groq", "Groq")
+	if err != nil {
+		t.Fatalf("RenameProvider() error = %v", err)
+	}
+	if cfg.Providers[0].Name != "Groq" || cfg.ActiveProvider != "Groq" {
+		t.Fatalf("case-only rename must still apply to config: %+v", cfg)
+	}
+	// The store is case-insensitive: the key must remain retrievable under the
+	// new casing (same entry), not deleted by a same-entry "migration".
+	if key, ok, err := store.Get("Groq"); err != nil || !ok || key != "sk-secret" {
+		t.Fatalf("stored key lost on case-only rename: key=%q ok=%v err=%v", key, ok, err)
+	}
+}
+
+func TestEditProviderAppliesRenameFieldsAndDescriptionAtomically(t *testing.T) {
+	dir := t.TempDir()
+	t.Setenv("ZERO_CRED_STORAGE", "encrypted-file")
+	path := filepath.Join(dir, "config.json")
+	writeConfigFixture(t, path, FileConfig{
+		ActiveProvider: "groq",
+		Providers: []ProviderProfile{
+			{Name: "groq", ProviderKind: ProviderKindOpenAICompatible, BaseURL: "https://api.groq.com/openai/v1", APIKeyStored: true, Model: "m1", Description: "old text"},
+			{Name: "other", ProviderKind: ProviderKindOpenAICompatible, BaseURL: "https://o.example.com/v1", Model: "m2"},
+		},
+	}, 0o600)
+	store, err := ProviderKeyStoreAt(dir)
+	if err != nil {
+		t.Fatalf("open store: %v", err)
+	}
+	if err := store.Set("groq", "sk-old"); err != nil {
+		t.Fatalf("seed store: %v", err)
+	}
+
+	cfg, err := EditProvider(path, ProviderEdit{
+		Name:        "groq",
+		NewName:     "grok-main",
+		Model:       "m1-pro",
+		Description: "", // cleared — must persist verbatim
+	})
+	if err != nil {
+		t.Fatalf("EditProvider() error = %v", err)
+	}
+	edited := cfg.Providers[0]
+	if edited.Name != "grok-main" || edited.Model != "m1-pro" || edited.Description != "" {
+		t.Fatalf("edit not applied: %+v", edited)
+	}
+	if edited.BaseURL != "https://api.groq.com/openai/v1" || !edited.APIKeyStored {
+		t.Fatalf("untouched fields must survive: %+v", edited)
+	}
+	if cfg.ActiveProvider != "grok-main" {
+		t.Fatalf("active must follow the rename, got %q", cfg.ActiveProvider)
+	}
+	if key, ok, _ := store.Get("grok-main"); !ok || key != "sk-old" {
+		t.Fatalf("stored key must migrate with the rename, got %q ok=%v", key, ok)
+	}
+	if len(cfg.Providers) != 2 || cfg.Providers[1].Name != "other" {
+		t.Fatalf("unrelated profile changed: %+v", cfg.Providers)
+	}
+}
+
+// TestEditProviderCaseOnlyRenameUpdatesInPlace: the manager previously skipped
+// RenameProvider on case-insensitively-equal names and fell into UpsertProvider,
+// whose case-SENSITIVE merge appended a duplicate profile. EditProvider matches
+// case-insensitively, so a case-only rename is an in-place update and the store
+// entry (case-normalized) survives.
+func TestEditProviderCaseOnlyRenameUpdatesInPlace(t *testing.T) {
+	dir := t.TempDir()
+	t.Setenv("ZERO_CRED_STORAGE", "encrypted-file")
+	path := filepath.Join(dir, "config.json")
+	writeConfigFixture(t, path, FileConfig{
+		ActiveProvider: "groq",
+		Providers: []ProviderProfile{
+			{Name: "groq", ProviderKind: ProviderKindOpenAICompatible, BaseURL: "https://api.groq.com/openai/v1", APIKeyStored: true, Model: "m1"},
+		},
+	}, 0o600)
+	store, err := ProviderKeyStoreAt(dir)
+	if err != nil {
+		t.Fatalf("open store: %v", err)
+	}
+	if err := store.Set("groq", "sk-secret"); err != nil {
+		t.Fatalf("seed store: %v", err)
+	}
+
+	cfg, err := EditProvider(path, ProviderEdit{Name: "groq", NewName: "Groq"})
+	if err != nil {
+		t.Fatalf("EditProvider() error = %v", err)
+	}
+	if len(cfg.Providers) != 1 {
+		t.Fatalf("case-only rename must not duplicate the profile: %+v", cfg.Providers)
+	}
+	if cfg.Providers[0].Name != "Groq" || !cfg.Providers[0].APIKeyStored {
+		t.Fatalf("in-place update wrong: %+v", cfg.Providers[0])
+	}
+	if cfg.Providers[0].BaseURL != "https://api.groq.com/openai/v1" {
+		t.Fatalf("fields must survive a case-only rename: %+v", cfg.Providers[0])
+	}
+	if cfg.ActiveProvider != "Groq" {
+		t.Fatalf("active must follow, got %q", cfg.ActiveProvider)
+	}
+	if key, ok, _ := store.Get("Groq"); !ok || key != "sk-secret" {
+		t.Fatalf("stored key lost on case-only rename: %q ok=%v", key, ok)
+	}
+}
+
+// TestEditProviderRenameMigratesFreshlyCapturedKey: replacing the key AND
+// renaming in one edit — the caller captures under the CURRENT name and the
+// rename migration moves it, so the new key lands under the new name.
+func TestEditProviderRenameMigratesFreshlyCapturedKey(t *testing.T) {
+	dir := t.TempDir()
+	t.Setenv("ZERO_CRED_STORAGE", "encrypted-file")
+	path := filepath.Join(dir, "config.json")
+	writeConfigFixture(t, path, FileConfig{
+		ActiveProvider: "gw",
+		Providers: []ProviderProfile{
+			{Name: "gw", ProviderKind: ProviderKindOpenAICompatible, BaseURL: "https://gw.example.com/v1", APIKeyEnv: "GW_KEY", Model: "m1"},
+		},
+	}, 0o600)
+	store, err := ProviderKeyStoreAt(dir)
+	if err != nil {
+		t.Fatalf("open store: %v", err)
+	}
+	// The caller's contract: a replacement key is stored under the CURRENT name
+	// before EditProvider runs (what SecureProviderProfile does).
+	if err := store.Set("gw", "sk-new"); err != nil {
+		t.Fatalf("seed store: %v", err)
+	}
+
+	cfg, err := EditProvider(path, ProviderEdit{Name: "gw", NewName: "gateway", APIKeyStored: true})
+	if err != nil {
+		t.Fatalf("EditProvider() error = %v", err)
+	}
+	if !cfg.Providers[0].APIKeyStored || cfg.Providers[0].Name != "gateway" {
+		t.Fatalf("marker/rename wrong: %+v", cfg.Providers[0])
+	}
+	if key, ok, _ := store.Get("gateway"); !ok || key != "sk-new" {
+		t.Fatalf("captured key must migrate to the new name, got %q ok=%v", key, ok)
+	}
+	if _, ok, _ := store.Get("gw"); ok {
+		t.Fatalf("old entry must be cleaned up after migration")
+	}
+}
+
+func TestEditProviderRejectsCollisionAndUnknown(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "config.json")
+	before := writeConfigFixture(t, path, FileConfig{
+		ActiveProvider: "alpha",
+		Providers: []ProviderProfile{
+			{Name: "alpha", ProviderKind: ProviderKindOpenAICompatible, BaseURL: "https://a.example.com/v1", Model: "m1"},
+			{Name: "beta", ProviderKind: ProviderKindOpenAICompatible, BaseURL: "https://b.example.com/v1", Model: "m2"},
+		},
+	}, 0o600)
+
+	if _, err := EditProvider(path, ProviderEdit{Name: "alpha", NewName: "BETA"}); err == nil {
+		t.Fatal("rename onto an existing name must fail")
+	}
+	if _, err := EditProvider(path, ProviderEdit{Name: "ghost", Model: "x"}); err == nil {
+		t.Fatal("editing an unknown provider must fail")
+	}
+	after, err := os.ReadFile(path)
+	if err != nil {
+		t.Fatalf("read config: %v", err)
+	}
+	if string(after) != string(before) {
+		t.Fatalf("config was rewritten by a rejected edit")
+	}
+}

internal/dictation/dictation.go

@@ -0,0 +1,104 @@
+// Package dictation implements speech-to-text capture and transcription for
+// the composer: press a key, talk, and the transcript lands in the input for
+// review — never auto-fired at the agent.
+//
+// Two pipelines share a keybinding and a composer-insertion target:
+//
+//   - Batch: record to a temp WAV file (Recorder.Start/Stop), then transcribe
+//     in one shot (Transcriber.Transcribe) — a local sherpa-onnx-offline exec
+//     or a single HTTPS POST to Groq/OpenAI. The only pipeline on Termux.
+//   - Streaming: capture raw PCM from the mic tool's stdout
+//     (Recorder.StartStreaming), feed 50ms chunks to a websocket
+//     (Transcriber.StreamTranscribe), and render partial transcripts live.
+//
+// Every OS integration shells out via exec.Command with discrete argv — never
+// a shell string, never CGO — matching internal/imageinput/clipboard.go. The
+// external binaries (arecord/sox/ffmpeg/termux-microphone-record, sherpa-onnx)
+// are looked up on PATH and degrade to a clear setup message when missing,
+// matching internal/lsp's philosophy for language servers.
+package dictation
+
+import (
+	"bytes"
+	"fmt"
+)
+
+// Batch transcription providers (config value stt.provider).
+const (
+	ProviderLocal  = "local"
+	ProviderGroq   = "groq"
+	ProviderOpenAI = "openai"
+)
+
+// Streaming transcription providers (config value stt.streamProvider).
+const (
+	StreamProviderLocal    = "local"
+	StreamProviderDeepgram = "deepgram"
+	StreamProviderOpenAI   = "openai"
+)
+
+// BatchProviders lists valid stt.provider values, for config validation.
+func BatchProviders() []string {
+	return []string{ProviderLocal, ProviderGroq, ProviderOpenAI}
+}
+
+// StreamProviders lists valid stt.streamProvider values, for config validation.
+func StreamProviders() []string {
+	return []string{StreamProviderLocal, StreamProviderDeepgram, StreamProviderOpenAI}
+}
+
+// AudioFormat identifies a recorded container format, sniffed from the bytes —
+// desktop recorders produce WAV, Termux's recorder produces M4A/AAC.
+type AudioFormat string
+
+const (
+	FormatWAV     AudioFormat = "wav"
+	FormatM4A     AudioFormat = "m4a"
+	FormatUnknown AudioFormat = ""
+)
+
+// SniffFormat detects the audio container from magic bytes. WAV is
+// "RIFF....WAVE"; MP4-family (Termux's .m4a) has "ftyp" at offset 4.
+func SniffFormat(data []byte) AudioFormat {
+	if len(data) >= 12 && bytes.Equal(data[:4], []byte("RIFF")) && bytes.Equal(data[8:12], []byte("WAVE")) {
+		return FormatWAV
+	}
+	if len(data) >= 8 && bytes.Equal(data[4:8], []byte("ftyp")) {
+		return FormatM4A
+	}
+	return FormatUnknown
+}
+
+// FileName returns a transcription-upload filename for the format.
+func (f AudioFormat) FileName() string {
+	switch f {
+	case FormatM4A:
+		return "audio.m4a"
+	default:
+		return "audio.wav"
+	}
+}
+
+// SetupError reports a missing external tool with install guidance. Missing
+// binaries are an expected condition (the user hasn't set dictation up yet),
+// not a bug — callers show the hint instead of a raw exec error.
+type SetupError struct {
+	Tool string
+	Hint string
+}
+
+func (e *SetupError) Error() string {
+	return fmt.Sprintf("%s not found — %s", e.Tool, e.Hint)
+}
+
+// AuthError reports that a cloud provider rejected a transcription request for
+// authentication reasons (HTTP 401/403) — a missing or invalid API key. It is a
+// fixable credential problem, so callers (the TUI) can offer an inline key
+// prompt for Provider instead of a dead-end error. Message is the already-
+// classified, secret-redacted user-facing text.
+type AuthError struct {
+	Provider string
+	Message  string
+}
+
+func (e *AuthError) Error() string { return e.Message }

internal/dictation/download.go

@@ -0,0 +1,893 @@
+package dictation
+
+import (
+	"archive/tar"
+	"compress/bzip2"
+	"context"
+	"crypto/sha256"
+	"encoding/hex"
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+	"os"
+	"path/filepath"
+	"runtime"
+	"strings"
+)
+
+// Auto-download of the local engine + a default model (opt-in, behind a confirm
+// in the F9 flow). Assets come from the official k2-fsa/sherpa-onnx GitHub
+// releases and are verified against a SHA256 digest before use — the same trust
+// boundary as installing them yourself, automated. Auto-downloading a native
+// executable is a real trust decision, which is why it is opt-in and verified.
+//
+// Rather than freeze a single hardcoded version, the downloader resolves the
+// release ASSETS via the GitHub API at download time and verifies each download
+// against the API's per-asset digest — so a newer sherpa-onnx release (or
+// "latest") works with no code change. A known-good version is still the DEFAULT
+// (reproducible, smoke-tested), and for that default the resolved digest is
+// cross-checked against a pinned value so a silently-changed release is caught.
+
+// DefaultSherpaVersion is the pinned, smoke-tested release used unless
+// stt.engineVersion overrides it. "latest" or any tag also works.
+const DefaultSherpaVersion = "v1.13.3"
+
+const (
+	sherpaRepo      = "k2-fsa/sherpa-onnx"
+	modelReleaseTag = "asr-models" // rolling release the ASR models live in
+	modelAssetName  = "sherpa-onnx-moonshine-tiny-en-int8.tar.bz2"
+	defaultAPIBase  = "https://api.github.com"
+)
+
+// engineAssetSuffix maps GOOS-GOARCH to the version-independent suffix of the
+// sherpa-onnx executable bundle for that platform. Matching by suffix (not a
+// full version-embedded name) is what lets any release resolve. The bundles are
+// self-contained (RPATH $ORIGIN/../lib), so the extracted bin/ executables run
+// with no library-path setup.
+var engineAssetSuffix = map[string]string{
+	"linux-amd64":   "linux-x64-shared-no-tts.tar.bz2",
+	"darwin-arm64":  "osx-arm64-shared-no-tts.tar.bz2",
+	"darwin-amd64":  "osx-x64-shared-no-tts.tar.bz2",
+	"windows-amd64": "win-x64-shared-MT-Release-no-tts.tar.bz2",
+}
+
+// pinnedEngineDigest cross-checks the DefaultSherpaVersion engine assets (a
+// silently-changed pinned release is refused). Empty for non-default versions,
+// which are verified against the API digest only.
+var pinnedEngineDigest = map[string]string{
+	"linux-amd64":   "e5cbfd96f3666c25d2272d468e337764fce6a4d50ed074e4b8e5021bfd7fefc1",
+	"darwin-arm64":  "ea73233bc250019ccf1f8a16feb29460efcd3425b50cac16980bcb76281d60bf",
+	"darwin-amd64":  "563423887d63d8a831c473e0ab6815139fe2e7a3482936716c2a128071dd217f",
+	"windows-amd64": "343c6c39cb4fe1880a9fb1abc8401d5174629ccbb147fca1fd53c440b2664570",
+}
+
+// pinnedModelDigest cross-checks the default Moonshine-tiny int8 model.
+const pinnedModelDigest = "d5fe6ec4334fef36255b2a4010412cad4c007e33103fec62fb5d17cad88086f2"
+
+// ModelVariant is a downloadable model the /stt-model picker offers, with a
+// pinned SHA256 (the model release predates GitHub's per-asset digests, so these
+// must be pinned). Batch variants use the offline binary; the streaming variant
+// (a transducer) drives the websocket server for a live transcript.
+type ModelVariant struct {
+	ID          string
+	Label       string
+	Description string
+	Bytes       int64
+	AssetName   string
+	Digest      string
+	DirName     string
+	Streaming   bool
+	// Recommended marks a curated, checksum-pinned, validated model (shown first).
+	Recommended bool
+}
+
+// ModelVariants returns the curated English models the download picker offers.
+// This is a hand-picked shortlist (each pinned by SHA256 since the model release
+// predates GitHub's per-asset digests), NOT the whole sherpa-onnx model zoo —
+// many more models (other languages, sizes, and streaming/batch families) work
+// via a manual stt.localModelPath. Streaming (live) options come first.
+func ModelVariants() []ModelVariant {
+	return []ModelVariant{
+		// Live / streaming — transcript builds up as you speak.
+		// DirName MUST match what ListModels derives from the asset ("model-"+id, id =
+		// asset minus "sherpa-onnx-" prefix and ".tar.bz2" suffix, dates included) so a
+		// model extracts to the same directory whether picked from the curated shortlist
+		// or the full browse list — otherwise installed-detection disagrees between the
+		// two views and the same model can be downloaded twice. Recommended matches the
+		// full list too (recommendedModelIDs marks every curated variant), so the ★ and
+		// ordering don't change when the background fetch swaps the curated list out.
+		{
+			ID: "streaming-zipformer-en-kroko", Label: "Kroko (newest, fastest)",
+			Description: "smallest live model", Bytes: 54 << 20,
+			AssetName:   "sherpa-onnx-streaming-zipformer-en-kroko-2025-08-06.tar.bz2",
+			Digest:      "c8676e5ff9ac2a85296e53ee0fd4d5fb1db6770e7a7647166eeafe349ade6834",
+			DirName:     "model-streaming-zipformer-en-kroko-2025-08-06",
+			Streaming:   true,
+			Recommended: true,
+		},
+		{
+			ID: "streaming-zipformer-20m", Label: "Zipformer 20M",
+			Description: "small, well-tested", Bytes: 121 << 20,
+			AssetName: "sherpa-onnx-streaming-zipformer-en-20M-2023-02-17.tar.bz2",
+			Digest:    "9c559283e8498d3fe95913c79ca1cb454bb26281ac2b102b41306c7d752765d9",
+			DirName:   "model-streaming-zipformer-en-20M-2023-02-17",
+			Streaming: true,
+		},
+		{
+			ID: "streaming-zipformer", Label: "Zipformer",
+			Description: "larger, more accurate", Bytes: 297 << 20,
+			AssetName: "sherpa-onnx-streaming-zipformer-en-2023-06-26.tar.bz2",
+			Digest:    "639e25b578e9e997131402199419c13a941f8e4e198e2da1ce57dbf5cf401282",
+			DirName:   "model-streaming-zipformer-en-2023-06-26",
+			Streaming: true,
+		},
+		{
+			ID: "nemo-streaming-fast-conformer-transducer-en-80ms-int8", Label: "NeMo streaming (fast)",
+			Description: "low-latency English", Bytes: 98 << 20,
+			AssetName:   "sherpa-onnx-nemo-streaming-fast-conformer-transducer-en-80ms-int8.tar.bz2",
+			Digest:      "7bd33a914e93370a1ba9c2066d9e841bdcad8613fa2a00537c1ae15d851a14d8",
+			DirName:     "model-nemo-streaming-fast-conformer-transducer-en-80ms-int8",
+			Streaming:   true,
+			Recommended: true,
+		},
+		// Batch — transcribes once when you stop.
+		{
+			ID: "moonshine-tiny", Label: "Moonshine tiny",
+			Description: "fast, smallest", Bytes: 103 << 20,
+			AssetName: "sherpa-onnx-moonshine-tiny-en-int8.tar.bz2",
+			Digest:    pinnedModelDigest,
+			DirName:   "model-moonshine-tiny-en-int8",
+		},
+		{
+			ID: "moonshine-base", Label: "Moonshine base",
+			Description: "more accurate, larger", Bytes: 239 << 20,
+			AssetName:   "sherpa-onnx-moonshine-base-en-int8.tar.bz2",
+			Digest:      "21870cecaa2e44e4e2bf63e02d1072bed183ccd10284871353bd9d24dad14e5e",
+			DirName:     "model-moonshine-base-en-int8",
+			Recommended: true,
+		},
+		{
+			ID: "sense-voice-zh-en-ja-ko-yue-int8-2025-09-09", Label: "SenseVoice",
+			Description: "multilingual, accurate", Bytes: 158 << 20,
+			AssetName:   "sherpa-onnx-sense-voice-zh-en-ja-ko-yue-int8-2025-09-09.tar.bz2",
+			Digest:      "7305f7905bfcf77fa0b39388a313f3da35c68d971661a65475b56fb2162c8e63",
+			DirName:     "model-sense-voice-zh-en-ja-ko-yue-int8-2025-09-09",
+			Recommended: true,
+		},
+		{
+			// Heavy (~1 GB) but the best transcription quality — recommended when
+			// accuracy matters more than speed/size. The asr-models release publishes
+			// no digest for it, so (like every browse-listed model) it downloads with
+			// TLS-only verification.
+			ID: "whisper-large-v3", Label: "Whisper large v3",
+			Description: "best accuracy, heavy", Bytes: 1019 << 20,
+			AssetName:   "sherpa-onnx-whisper-large-v3.tar.bz2",
+			DirName:     "model-whisper-large-v3",
+			Recommended: true,
+		},
+	}
+}
+
+// recommendedModelIDs marks the curated variants flagged Recommended so the full
+// browse list stars (and surfaces) exactly the ones the curated shortlist does —
+// keyed off the same Recommended flag so the two views never disagree.
+var recommendedModelIDs = func() map[string]bool {
+	m := map[string]bool{}
+	for _, v := range ModelVariants() {
+		if v.Recommended {
+			m[v.AssetName] = true
+		}
+	}
+	return m
+}()
+
+// pinnedModelDigests maps a model asset name to its pinned SHA256 (the curated
+// variants), so browse-listed models we happen to know get verified too.
+var pinnedModelDigests = func() map[string]string {
+	m := map[string]string{}
+	for _, v := range ModelVariants() {
+		m[v.AssetName] = v.Digest
+	}
+	return m
+}()
+
+// curatedLabels maps a curated model's asset name to its friendly label, so the
+// recommended rows read nicely in the full browse list too.
+var curatedLabels = func() map[string]string {
+	m := map[string]string{}
+	for _, v := range ModelVariants() {
+		m[v.AssetName] = v.Label
+	}
+	return m
+}()
+
+// browseLabel turns a raw asset id into a shorter display label: it drops the
+// redundant "-int8"/quantization noise and the "streaming-" prefix (the group
+// header already says live vs batch).
+func browseLabel(id string) string {
+	label := strings.TrimPrefix(id, "streaming-")
+	for _, drop := range []string{"-int8", ".int8", "-quantized", "-fp16"} {
+		label = strings.ReplaceAll(label, drop, "")
+	}
+	return label
+}
+
+// ListModels fetches every transcription model in the sherpa-onnx asr-models
+// release and returns them as variants (the curated ones flagged Recommended and
+// carrying a pinned digest; the rest with an empty digest, downloaded with
+// TLS-only verification since that release predates GitHub's per-asset digests).
+// Non-ASR assets (VAD, TTS, punctuation, speaker/keyword/…): filtered out.
+func ListModels(ctx context.Context, client *http.Client, apiBase string) ([]ModelVariant, error) {
+	if client == nil {
+		client = http.DefaultClient
+	}
+	if apiBase == "" {
+		apiBase = defaultAPIBase
+	}
+	// Resolve the release id, then page its assets (a release can have hundreds).
+	relURL := fmt.Sprintf("%s/repos/%s/releases/tags/%s", strings.TrimRight(apiBase, "/"), sherpaRepo, modelReleaseTag)
+	var rel struct {
+		ID     int64 `json:"id"`
+		Assets []githubAsset
+	}
+	if err := getJSON(ctx, client, relURL, &rel); err != nil {
+		return nil, fmt.Errorf("listing dictation models: %w", err)
+	}
+	assets := rel.Assets
+	// If the inline assets look capped, page the assets endpoint for the rest.
+	for page := 1; len(assets)%100 == 0 && len(assets) > 0; page++ {
+		var more []githubAsset
+		u := fmt.Sprintf("%s/repos/%s/releases/%d/assets?per_page=100&page=%d", strings.TrimRight(apiBase, "/"), sherpaRepo, rel.ID, page+1)
+		if err := getJSON(ctx, client, u, &more); err != nil || len(more) == 0 {
+			break
+		}
+		assets = append(assets, more...)
+		if len(more) < 100 {
+			break
+		}
+	}
+
+	seen := map[string]bool{}
+	var out []ModelVariant
+	for _, a := range assets {
+		if seen[a.Name] || !isTranscriptionModelAsset(a.Name) {
+			continue
+		}
+		seen[a.Name] = true
+		id := strings.TrimSuffix(strings.TrimPrefix(a.Name, "sherpa-onnx-"), ".tar.bz2")
+		digest := strings.TrimPrefix(a.Digest, "sha256:")
+		if digest == "" {
+			digest = pinnedModelDigests[a.Name] // may still be empty (browse/unverified)
+		}
+		label := browseLabel(id)
+		desc := modelFamily(a.Name) // readable language, e.g. "French"
+		if nice, ok := curatedLabels[a.Name]; ok {
+			label = nice
+			desc = curatedDescriptions[a.Name] // curated blurb ("fast, smallest")
+		}
+		out = append(out, ModelVariant{
+			ID:          id,
+			Label:       label,
+			Description: desc,
+			Bytes:       a.Size,
+			AssetName:   a.Name,
+			Digest:      digest,
+			DirName:     "model-" + id,
+			Streaming:   strings.Contains(strings.ToLower(a.Name), "streaming-"),
+			Recommended: recommendedModelIDs[a.Name],
+		})
+	}
+	return out, nil
+}
+
+type githubAsset struct {
+	Name   string `json:"name"`
+	Digest string `json:"digest"`
+	Size   int64  `json:"size"`
+}
+
+func getJSON(ctx context.Context, client *http.Client, url string, v any) error {
+	req, err := http.NewRequestWithContext(ctx, http.MethodGet, url, nil)
+	if err != nil {
+		return err
+	}
+	req.Header.Set("Accept", "application/vnd.github+json")
+	resp, err := client.Do(req)
+	if err != nil {
+		return err
+	}
+	defer resp.Body.Close()
+	if resp.StatusCode != http.StatusOK {
+		return fmt.Errorf("GitHub API HTTP %d", resp.StatusCode)
+	}
+	return json.NewDecoder(resp.Body).Decode(v)
+}
+
+// isTranscriptionModelAsset reports whether an asset is a speech-to-text model
+// (as opposed to a VAD/TTS/punctuation/speaker/keyword/enhancement tool).
+func isTranscriptionModelAsset(name string) bool {
+	n := strings.ToLower(name)
+	if !strings.HasPrefix(n, "sherpa-onnx-") || !strings.HasSuffix(n, ".tar.bz2") {
+		return false
+	}
+	for _, bad := range []string{
+		"-vad", "vad-", "tts", "punct", "speaker", "diariz", "kws", "keyword",
+		"audio-tagging", "denois", "source-separation", "speech-enhancement",
+		"spoken-language", "language-identification", "-lid", "melo", "kokoro",
+		"vits", "matcha", "piper", "wespeaker",
+		// Rockchip NPU (RKNN) builds for specific ARM boards — not CPU/ONNX, so
+		// the standard sherpa-onnx binary can't run them.
+		"rknn", "rk3566", "rk3568", "rk3562", "rk3588", "rk3576", "rk3528", "rk3399",
+	} {
+		if strings.Contains(n, bad) {
+			return false
+		}
+	}
+	for _, fam := range []string{
+		"zipformer", "moonshine", "whisper", "paraformer", "sense-voice",
+		"sensevoice", "conformer", "nemo", "canary", "parakeet", "fire-red",
+		"fireredasr", "dolphin", "telespeech", "wenet", "omnilingual", "funasr",
+		"medasr", "cohere", "qwen", "citrinet", "transducer", "-ctc-", "zipvoice",
+	} {
+		if strings.Contains(n, fam) {
+			return true
+		}
+	}
+	return false
+}
+
+// langNames maps ISO-ish language tokens in a model name to readable names.
+var langNames = map[string]string{
+	"en": "English", "zh": "Chinese", "yue": "Cantonese", "fr": "French",
+	"de": "German", "es": "Spanish", "ru": "Russian", "ja": "Japanese",
+	"ko": "Korean", "bn": "Bengali", "ar": "Arabic", "hi": "Hindi",
+	"pt": "Portuguese", "nl": "Dutch", "vi": "Vietnamese", "th": "Thai",
+	"uk": "Ukrainian", "fa": "Persian", "pl": "Polish", "tr": "Turkish",
+	"cs": "Czech", "sv": "Swedish", "fil": "Filipino", "tl": "Tagalog",
+	"fi": "Finnish", "el": "Greek", "he": "Hebrew", "ro": "Romanian",
+}
+
+// modelFamily is a short readable descriptor for a model, favoring the language
+// (the most useful thing not already obvious from the group header) and falling
+// back to the family. E.g. "French", "Chinese", "Multilingual".
+func modelFamily(name string) string {
+	n := strings.ToLower(name)
+	if strings.Contains(n, "multi") || strings.Contains(n, "bilingual") ||
+		strings.Contains(n, "trilingual") || strings.Contains(n, "polyglot") ||
+		strings.Contains(n, "omnilingual") {
+		return "Multilingual"
+	}
+	var langs []string
+	seen := map[string]bool{}
+	add := func(name string) {
+		if name != "" && !seen[name] {
+			seen[name] = true
+			langs = append(langs, name)
+		}
+	}
+	for _, tok := range strings.Split(strings.TrimSuffix(n, ".tar.bz2"), "-") {
+		add(langNames[tok])
+	}
+	// Full language words spelled out in some names (e.g. "-cantonese-").
+	for word, name := range map[string]string{
+		"cantonese": "Cantonese", "mandarin": "Chinese", "english": "English",
+		"chinese": "Chinese", "korean": "Korean", "japanese": "Japanese",
+		"russian": "Russian", "arabic": "Arabic", "spanish": "Spanish",
+		"french": "French", "german": "German", "vietnamese": "Vietnamese",
+		"thai": "Thai", "farsi": "Persian", "persian": "Persian",
+		"portuguese": "Portuguese", "italian": "Italian", "dutch": "Dutch",
+		"ukrainian": "Ukrainian", "polish": "Polish", "turkish": "Turkish",
+	} {
+		if strings.Contains(n, word) {
+			add(name)
+		}
+	}
+	switch {
+	case len(langs) == 0:
+		return ""
+	case len(langs) > 2:
+		return "Multilingual"
+	default:
+		return strings.Join(langs, " + ")
+	}
+}
+
+// curatedDescriptions maps a curated model's asset name to its quality blurb.
+var curatedDescriptions = func() map[string]string {
+	m := map[string]string{}
+	for _, v := range ModelVariants() {
+		m[v.AssetName] = v.Description
+	}
+	return m
+}()
+
+// EngineComponents identifies the resolved local-engine paths.
+type EngineComponents struct {
+	BinaryPath string // extracted sherpa-onnx-offline
+	ServerPath string // extracted sherpa-onnx-online-websocket-server
+	ModelPath  string // extracted model directory
+}
+
+// DownloadOptions configures EnsureLocalEngine.
+type DownloadOptions struct {
+	// DestRoot is where extracted engines/models live (e.g. ~/.config/zero/stt).
+	DestRoot string
+	// EngineVersion selects the sherpa-onnx release tag ("" → DefaultSherpaVersion;
+	// "latest" and any tag also work).
+	EngineVersion string
+	// HTTPClient is injectable for tests; nil uses a plain client.
+	HTTPClient *http.Client
+	// APIBase overrides the GitHub API base (tests). "" → api.github.com.
+	APIBase string
+	// Model selection (default: the Moonshine-tiny int8 batch model). Set these
+	// from a ModelVariant to download a different model.
+	ModelAssetName    string
+	ModelPinnedDigest string
+	ModelDirName      string
+	// ModelLabel is the friendly model name shown in progress ("Zipformer 20M").
+	ModelLabel string
+	// Progress receives live status strings for the UI, including a percentage
+	// while downloading (e.g. "Engine 45% · 57/126 MB").
+	Progress func(status string)
+	// platformKey overrides runtime detection (tests).
+	platformKey string
+	// skipPinned disables the pinned-digest cross-check (tests, which use fixture
+	// assets); the API-digest verification still runs.
+	skipPinned bool
+}
+
+// AutoDownloadSupported reports whether the current platform has a known engine
+// asset (so the F9 flow can offer the download). Termux/Android and unusual
+// arches install manually or use a cloud provider.
+func AutoDownloadSupported() bool {
+	_, ok := engineAssetSuffix[platformKey()]
+	return ok
+}
+
+func platformKey() string { return runtime.GOOS + "-" + runtime.GOARCH }
+
+// EnsureLocalEngine downloads (if absent) and verifies the sherpa-onnx engine and
+// the default model into DestRoot, returning the resolved paths. Idempotent: an
+// already-extracted, still-present engine/model is reused without re-downloading.
+// Every archive is SHA256-verified against the GitHub API digest before extraction.
+func EnsureLocalEngine(ctx context.Context, opts DownloadOptions) (EngineComponents, error) {
+	if strings.TrimSpace(opts.DestRoot) == "" {
+		return EngineComponents{}, fmt.Errorf("dictation download: DestRoot is required")
+	}
+	key := opts.platformKey
+	if key == "" {
+		key = platformKey()
+	}
+	suffix, ok := engineAssetSuffix[key]
+	if !ok {
+		return EngineComponents{}, &SetupError{
+			Tool: "sherpa-onnx auto-download",
+			Hint: fmt.Sprintf("no prebuilt engine is known for %s — install sherpa-onnx manually or use a cloud provider (see docs/dictation.md)", key),
+		}
+	}
+	version := strings.TrimSpace(opts.EngineVersion)
+	if version == "" {
+		version = DefaultSherpaVersion
+	}
+	progress := opts.Progress
+	if progress == nil {
+		progress = func(string) {}
+	}
+	client := opts.HTTPClient
+	if client == nil {
+		client = http.DefaultClient
+	}
+	apiBase := opts.APIBase
+	if apiBase == "" {
+		apiBase = defaultAPIBase
+	}
+
+	engineDir := filepath.Join(opts.DestRoot, "engine-"+version+"-"+key)
+	targetWindows := strings.HasPrefix(key, "windows-")
+	// Resolve through the tarball's flattened subdir so an ALREADY-extracted
+	// engine is found and not needlessly re-downloaded (the idempotency check).
+	binPath, serverPath := resolveEnginePaths(engineDir, targetWindows)
+	if !fileExists(binPath) {
+		pinned := ""
+		if version == DefaultSherpaVersion && !opts.skipPinned {
+			pinned = pinnedEngineDigest[key]
+		}
+		asset, err := resolveAsset(ctx, client, apiBase, version, "sherpa-onnx-", suffix)
+		if err != nil {
+			return EngineComponents{}, err
+		}
+		if err := downloadVerifyExtract(ctx, client, asset, pinned, false, "Engine", engineDir, progress); err != nil {
+			return EngineComponents{}, err
+		}
+		binPath, serverPath = resolveEnginePaths(engineDir, targetWindows)
+	}
+	if !fileExists(binPath) {
+		return EngineComponents{}, fmt.Errorf("dictation download: engine binary not found after extraction in %s", engineDir)
+	}
+
+	modelName := opts.ModelAssetName
+	if modelName == "" {
+		modelName = modelAssetName
+	}
+	modelDirName := opts.ModelDirName
+	if modelDirName == "" {
+		modelDirName = "model-moonshine-tiny-en-int8"
+	}
+	modelDir := filepath.Join(opts.DestRoot, modelDirName)
+	if !dirHasModel(modelDir) {
+		asset, err := resolveAsset(ctx, client, apiBase, modelReleaseTag, modelName, "")
+		if err != nil {
+			return EngineComponents{}, err
+		}
+		modelPinned := opts.ModelPinnedDigest
+		// Only fall back to the built-in digest for the DEFAULT model — a different
+		// (browse-listed) model must never be checked against the default's digest.
+		if modelPinned == "" && opts.ModelAssetName == "" {
+			modelPinned = pinnedModelDigest
+		}
+		if opts.skipPinned {
+			modelPinned = ""
+		}
+		modelLabel := opts.ModelLabel
+		if modelLabel == "" {
+			modelLabel = "Model"
+		}
+		// Models are data files from the official release: verify against a digest
+		// when we have one (curated/pinned or API-provided), else allow a TLS-only
+		// download (the model release predates GitHub's per-asset digests). The
+		// engine BINARY above never allows this — a native executable is always
+		// digest-verified.
+		if err := downloadVerifyExtract(ctx, client, asset, modelPinned, true, modelLabel, modelDir, progress); err != nil {
+			return EngineComponents{}, err
+		}
+	}
+	resolvedModel := modelDir
+	if !hasTokensFile(resolvedModel) {
+		if child, err := flattenSingleChild(modelDir); err == nil {
+			resolvedModel = child
+		}
+	}
+	if !hasTokensFile(resolvedModel) {
+		return EngineComponents{}, fmt.Errorf("dictation download: model files not found after extraction in %s", modelDir)
+	}
+	return EngineComponents{BinaryPath: binPath, ServerPath: serverPath, ModelPath: resolvedModel}, nil
+}
+
+type resolvedAsset struct {
+	url    string
+	sha256 string
+	size   int64
+	name   string
+}
+
+// resolveAsset queries the GitHub release for tag and returns the asset whose
+// name has the given prefix and suffix, plus its SHA256 digest and size.
+func resolveAsset(ctx context.Context, client *http.Client, apiBase, tag, namePrefix, nameSuffix string) (resolvedAsset, error) {
+	tagPath := "tags/" + tag
+	if tag == "" || tag == "latest" {
+		tagPath = "latest"
+	}
+	url := fmt.Sprintf("%s/repos/%s/releases/%s", strings.TrimRight(apiBase, "/"), sherpaRepo, tagPath)
+	req, err := http.NewRequestWithContext(ctx, http.MethodGet, url, nil)
+	if err != nil {
+		return resolvedAsset{}, err
+	}
+	req.Header.Set("Accept", "application/vnd.github+json")
+	resp, err := client.Do(req)
+	if err != nil {
+		return resolvedAsset{}, fmt.Errorf("resolving sherpa-onnx release %s: %w", tag, err)
+	}
+	defer resp.Body.Close()
+	if resp.StatusCode != http.StatusOK {
+		return resolvedAsset{}, fmt.Errorf("resolving sherpa-onnx release %s: GitHub API HTTP %d", tag, resp.StatusCode)
+	}
+	var release struct {
+		TagName string `json:"tag_name"`
+		Assets  []struct {
+			Name        string `json:"name"`
+			DownloadURL string `json:"browser_download_url"`
+			Digest      string `json:"digest"`
+			Size        int64  `json:"size"`
+		} `json:"assets"`
+	}
+	if err := json.NewDecoder(resp.Body).Decode(&release); err != nil {
+		return resolvedAsset{}, fmt.Errorf("parsing sherpa-onnx release %s: %w", tag, err)
+	}
+	for _, a := range release.Assets {
+		if !strings.HasPrefix(a.Name, namePrefix) || !strings.HasSuffix(a.Name, nameSuffix) {
+			continue
+		}
+		// The API digest may be empty for older assets (GitHub only populates it for
+		// assets uploaded after ~2024 — the ASR models predate that). An empty
+		// digest is fine here; downloadVerifyExtract requires the API digest OR a
+		// pinned one and refuses only when neither exists.
+		digest := strings.TrimPrefix(a.Digest, "sha256:")
+		return resolvedAsset{url: a.DownloadURL, sha256: digest, size: a.Size, name: a.Name}, nil
+	}
+	return resolvedAsset{}, &SetupError{
+		Tool: "sherpa-onnx auto-download",
+		Hint: fmt.Sprintf("release %s has no asset matching %s*%s — try a different stt.engineVersion or install manually", tag, namePrefix, nameSuffix),
+	}
+}
+
+// downloadVerifyExtract fetches the asset to a temp file, verifies its SHA256
+// (against the API digest, and — when pinned is set — a cross-check that the
+// resolved digest equals the audited value), and extracts the tar.bz2. A
+// mismatch aborts before anything is extracted or run.
+func downloadVerifyExtract(ctx context.Context, client *http.Client, asset resolvedAsset, pinned string, allowUnverified bool, label, destDir string, progress func(string)) error {
+	// Determine the digest to verify against: the API digest and the pinned digest
+	// must agree when both exist; at least one must exist unless allowUnverified
+	// (models are data files from the official release — TLS-only is acceptable;
+	// a native executable never sets allowUnverified).
+	expected := asset.sha256
+	if pinned != "" {
+		if expected != "" && !strings.EqualFold(expected, pinned) {
+			return fmt.Errorf("dictation download: %s digest %s does not match the pinned known-good value %s — refusing (the release may have changed)", asset.name, asset.sha256, pinned)
+		}
+		expected = pinned
+	}
+	if expected == "" && !allowUnverified {
+		return fmt.Errorf("dictation download: no SHA256 available for %s (neither the GitHub API nor a pinned value) — refusing an unverifiable download", asset.name)
+	}
+	// Ensure the parent of destDir exists; the final destDir is created by
+	// the atomic rename at the end of the function.
+	if err := os.MkdirAll(filepath.Dir(destDir), 0o755); err != nil {
+		return err
+	}
+	// Stage the extraction into a sibling temp dir and only swap it into
+	// destDir on success. An interrupted run that left a truncated engine
+	// binary in destDir would otherwise pass fileExists() on the next start
+	// and break the local recorder silently.
+	stageDir, err := os.MkdirTemp(filepath.Dir(destDir), filepath.Base(destDir)+".stage-*")
+	if err != nil {
+		return err
+	}
+	cleanupStage := true
+	defer func() {
+		if cleanupStage {
+			_ = os.RemoveAll(stageDir)
+		}
+	}()
+
+	tmp, err := os.CreateTemp(filepath.Dir(destDir), "stt-dl-*.tar.bz2")
+	if err != nil {
+		return err
+	}
+	tmpPath := tmp.Name()
+	defer os.Remove(tmpPath)
+
+	req, err := http.NewRequestWithContext(ctx, http.MethodGet, asset.url, nil)
+	if err != nil {
+		tmp.Close()
+		return err
+	}
+	resp, err := client.Do(req)
+	if err != nil {
+		tmp.Close()
+		return fmt.Errorf("downloading %s: %w", asset.name, err)
+	}
+	defer resp.Body.Close()
+	if resp.StatusCode != http.StatusOK {
+		tmp.Close()
+		return fmt.Errorf("downloading %s: HTTP %d", asset.name, resp.StatusCode)
+	}
+	hasher := sha256.New()
+	total := asset.size
+	if total <= 0 {
+		total = resp.ContentLength
+	}
+	src := io.Reader(resp.Body)
+	if total > 0 {
+		src = &progressReader{r: resp.Body, total: total, label: label, report: progress, lastPct: -1}
+	}
+	if _, err := io.Copy(io.MultiWriter(tmp, hasher), src); err != nil {
+		tmp.Close()
+		return fmt.Errorf("downloading %s: %w", asset.name, err)
+	}
+	if err := tmp.Close(); err != nil {
+		return err
+	}
+	got := hex.EncodeToString(hasher.Sum(nil))
+	if expected != "" && !strings.EqualFold(got, expected) {
+		return fmt.Errorf("dictation download: checksum mismatch for %s (want %s, got %s) — refusing to extract", asset.name, expected, got)
+	}
+	if err := extractTarBz2(tmpPath, stageDir, "Extracting "+label, progress); err != nil {
+		return err
+	}
+	// Remove any pre-existing destDir so Rename is atomic on the same
+	// filesystem (os.Rename refuses to overwrite a non-empty dir).
+	if err := os.RemoveAll(destDir); err != nil {
+		return fmt.Errorf("clearing previous %s install: %w", label, err)
+	}
+	if err := os.Rename(stageDir, destDir); err != nil {
+		return fmt.Errorf("promoting staged %s: %w", label, err)
+	}
+	cleanupStage = false
+	return nil
+}
+
+// progressReader reports download progress as a percentage on each whole-percent
+// change, formatted "<label> NN% · X/Y MB".
+type progressReader struct {
+	r       io.Reader
+	total   int64
+	read    int64
+	label   string
+	report  func(string)
+	lastPct int
+}
+
+func (p *progressReader) Read(b []byte) (int, error) {
+	n, err := p.r.Read(b)
+	p.read += int64(n)
+	if p.total > 0 {
+		pct := int(p.read * 100 / p.total)
+		if pct != p.lastPct {
+			p.lastPct = pct
+			p.report(fmt.Sprintf("%s %d%% · %d/%d MB", p.label, pct, p.read>>20, p.total>>20))
+		}
+	}
+	return n, err
+}
+
+// enginePaths builds the engine executable paths. The ".exe" suffix follows the
+// TARGET platform (the one the engine was downloaded for), not the host — they
+// match in production (you download your own platform), but keying off the target
+// keeps resolution correct when the two differ (e.g. a cross-platform test).
+func enginePaths(engineDir string, targetWindows bool) (bin, server string) {
+	ext := ""
+	if targetWindows {
+		ext = ".exe"
+	}
+	return filepath.Join(engineDir, "bin", "sherpa-onnx-offline"+ext),
+		filepath.Join(engineDir, "bin", "sherpa-onnx-online-websocket-server"+ext)
+}
+
+// resolveEnginePaths finds the executables, flattening the tarball's single
+// top-level directory when present.
+func resolveEnginePaths(engineDir string, targetWindows bool) (bin, server string) {
+	bin, server = enginePaths(engineDir, targetWindows)
+	if fileExists(bin) {
+		return bin, server
+	}
+	if child, err := flattenSingleChild(engineDir); err == nil && child != engineDir {
+		return enginePaths(child, targetWindows)
+	}
+	return bin, server
+}
+
+// extractTarBz2 unpacks a bzip2-compressed tar into destDir, guarding against
+// path-traversal entries. It reports extraction progress by how much of the
+// (compressed) archive it has consumed — the same MB scale as the download, so
+// "Extracting … 50% · 520/1041 MB" lines up with the download that preceded it.
+func extractTarBz2(archivePath, destDir, label string, progress func(string)) error {
+	f, err := os.Open(archivePath)
+	if err != nil {
+		return err
+	}
+	defer f.Close()
+	var src io.Reader = f
+	if progress != nil {
+		if info, statErr := f.Stat(); statErr == nil && info.Size() > 0 {
+			src = &progressReader{r: f, total: info.Size(), label: label, report: progress, lastPct: -1}
+		}
+	}
+	tr := tar.NewReader(bzip2.NewReader(src))
+	destAbs, err := filepath.Abs(destDir)
+	if err != nil {
+		return err
+	}
+	for {
+		hdr, err := tr.Next()
+		if err == io.EOF {
+			return nil
+		}
+		if err != nil {
+			return err
+		}
+		target := filepath.Join(destDir, hdr.Name)
+		absTarget, err := filepath.Abs(target)
+		if err != nil {
+			return err
+		}
+		if absTarget != destAbs && !strings.HasPrefix(absTarget, destAbs+string(os.PathSeparator)) {
+			return fmt.Errorf("dictation download: archive entry %q escapes destination", hdr.Name)
+		}
+		switch hdr.Typeflag {
+		case tar.TypeDir:
+			if err := os.MkdirAll(target, 0o755); err != nil {
+				return err
+			}
+		case tar.TypeReg:
+			if err := os.MkdirAll(filepath.Dir(target), 0o755); err != nil {
+				return err
+			}
+			out, err := os.OpenFile(target, os.O_CREATE|os.O_TRUNC|os.O_WRONLY, os.FileMode(hdr.Mode)&0o777)
+			if err != nil {
+				return err
+			}
+			if _, err := io.Copy(out, tr); err != nil {
+				out.Close()
+				return err
+			}
+			if err := out.Close(); err != nil {
+				return err
+			}
+		}
+	}
+}
+
+// flattenSingleChild returns the sole subdirectory of dir (the tarball's
+// top-level folder) when dir contains exactly one directory entry.
+func flattenSingleChild(dir string) (string, error) {
+	entries, err := os.ReadDir(dir)
+	if err != nil {
+		return "", err
+	}
+	var dirs []string
+	for _, e := range entries {
+		if e.IsDir() {
+			dirs = append(dirs, e.Name())
+		}
+	}
+	if len(dirs) == 1 {
+		return filepath.Join(dir, dirs[0]), nil
+	}
+	return dir, nil
+}
+
+// ModelDownloaded reports whether a model variant is already extracted under
+// destRoot (so the picker can mark it as installed). dirName is the variant's
+// DirName.
+func ModelDownloaded(destRoot, dirName string) bool {
+	if destRoot == "" || dirName == "" {
+		return false
+	}
+	return dirHasModel(filepath.Join(destRoot, dirName))
+}
+
+// EngineDownloaded reports whether the shared sherpa-onnx engine is already on
+// disk for this platform (one engine serves every model), so the picker can drop
+// the engine's size from a model's download total once it's been fetched.
+func EngineDownloaded(destRoot, version string) bool {
+	if destRoot == "" {
+		return false
+	}
+	if version == "" {
+		version = DefaultSherpaVersion
+	}
+	engineDir := filepath.Join(destRoot, "engine-"+version+"-"+platformKey())
+	bin, _ := resolveEnginePaths(engineDir, runtime.GOOS == "windows")
+	return fileExists(bin)
+}
+
+func dirHasModel(dir string) bool {
+	if hasTokensFile(dir) {
+		return true
+	}
+	if child, err := flattenSingleChild(dir); err == nil && child != dir {
+		return hasTokensFile(child)
+	}
+	return false
+}
+
+// hasTokensFile reports whether dir holds a sherpa tokens file — usually
+// "tokens.txt", but some families (Whisper) prefix it as "<model>-tokens.txt".
+func hasTokensFile(dir string) bool {
+	entries, err := os.ReadDir(dir)
+	if err != nil {
+		return false
+	}
+	for _, e := range entries {
+		if !e.IsDir() && strings.HasSuffix(e.Name(), "tokens.txt") {
+			return true
+		}
+	}
+	return false
+}
+
+func fileExists(path string) bool {
+	info, err := os.Stat(path)
+	return err == nil && !info.IsDir()
+}

internal/dictation/download_smoke_test.go

@@ -0,0 +1,52 @@
+package dictation
+
+import (
+	"context"
+	"os"
+	"path/filepath"
+	"testing"
+)
+
+// TestEnsureLocalEngineRealDownload exercises the full auto-download against the
+// REAL k2-fsa/sherpa-onnx GitHub API (resolve → verify digest → extract → run).
+// It downloads ~130MB, so it is opt-in via ZERO_STT_DOWNLOAD_TEST=1.
+//
+//	ZERO_STT_DOWNLOAD_TEST=1 go test ./internal/dictation/ -run RealDownload -v -timeout 10m
+func TestEnsureLocalEngineRealDownload(t *testing.T) {
+	if os.Getenv("ZERO_STT_DOWNLOAD_TEST") == "" {
+		t.Skip("set ZERO_STT_DOWNLOAD_TEST=1 to run the real ~130MB download smoke test")
+	}
+	dest := t.TempDir()
+	comp, err := EnsureLocalEngine(context.Background(), DownloadOptions{
+		DestRoot: dest,
+		Progress: func(s string) { t.Logf("stage: %s", s) },
+	})
+	if err != nil {
+		t.Fatalf("EnsureLocalEngine: %v", err)
+	}
+	if !fileExists(comp.BinaryPath) || !fileExists(filepath.Join(comp.ModelPath, "tokens.txt")) {
+		t.Fatalf("missing components: %+v", comp)
+	}
+
+	// The downloaded engine + model must actually transcribe.
+	wav := filepath.Join(comp.ModelPath, "test_wavs", "0.wav")
+	if !fileExists(wav) {
+		t.Skipf("no bundled test wav at %s", wav)
+	}
+	audio, err := os.ReadFile(wav)
+	if err != nil {
+		t.Fatal(err)
+	}
+	tr, err := NewLocalTranscriber(LocalConfig{Binary: comp.BinaryPath, ModelPath: comp.ModelPath})
+	if err != nil {
+		t.Fatal(err)
+	}
+	text, err := tr.Transcribe(context.Background(), audio)
+	if err != nil {
+		t.Fatalf("Transcribe with downloaded engine: %v", err)
+	}
+	if text == "" {
+		t.Fatal("downloaded engine returned empty transcript")
+	}
+	t.Logf("downloaded-engine transcript: %q", text)
+}

internal/dictation/download_test.go

@@ -0,0 +1,245 @@
+package dictation
+
+import (
+	"context"
+	"encoding/base64"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"os"
+	"path/filepath"
+	"runtime"
+	"strings"
+	"testing"
+)
+
+// Tiny tar.bz2 fixtures (generated in-repo): the engine has top/bin/sherpa-onnx-
+// offline + server; the model has top/tokens.txt.
+const (
+	engineTarB64 = "QlpoOTFBWSZTWW+fEfAAAW1/hMGRAIBAA/+AAgAgRH9v3+AAAQCoMAEYAYA00NGjEZANABoYYA00NGjEZANABoYIpE0TRoTGkA2poGgxqepl1H6w5anW3tY2OggUi1AW/d29pO0GLUoOtnQsrGFNowIEh0cImtQ9gyHqYH7unm+/wZSLXsBuBDDCfHiahCsGEgGL9/pUG+RQZo76xlaDIGKWDErPKuyhXYOlzJMG3iNa56kiluctzUqgZ2A1AGwgcB3jtB5PB7pYYlDHH8OvOznM6B6MHPRqa7sTE1vwDBc0LjmG4ah5F+YYaXHTZpGmcemWW1yMh1adhdrB6A3nAQP8XckU4UJBvnxHwA=="
+	engineSHA    = "97b5cb6a417705860f8940e1bf9ffec8cc60e8666a0d228cd60f1de7607d503d"
+	modelTarB64  = "QlpoOTFBWSZTWTtAoX4AANp/hMOQBIBAAf+AAAIQBHoJ3mAAAQAIMAC5sG1SNMhNqGmQwIyeKEagam1DQ0AAABFRT0T0mg0aeoDQHpH7Nzrs24wsbVW8hMQHClNI0LuLhKMGY4sEQi6N2qQWP1id8o1KtprGd4hll4uQkMGA1nFMXAwohEZOnIMIPQKQl2k0GbSGwkC0lJb2IWOQ57ETYbmxFGspaRp1QyT6q4sKj855st7GzOGj5eIwmWt6wRajEQl/F3JFOFCQO0Chfg=="
+	modelSHA     = "16c6b1dcb40f6b21c0de70bc1a8edab95375e4de90dac7c076fd2cc7a34fc59a"
+)
+
+// fakeReleaseServer serves the GitHub release API + asset bytes for the engine
+// and model, exercising resolveAsset + downloadVerifyExtract against a real HTTP
+// round-trip (with the real coder/websocket-free download path).
+func fakeReleaseServer(t *testing.T, engineDigest, modelDigest string) *httptest.Server {
+	t.Helper()
+	engineBytes, _ := base64.StdEncoding.DecodeString(engineTarB64)
+	modelBytes, _ := base64.StdEncoding.DecodeString(modelTarB64)
+	mux := http.NewServeMux()
+	base := ""
+	release := func(assetName, digest, dlPath string) string {
+		body, _ := json.Marshal(map[string]any{
+			"tag_name": "test",
+			"assets": []map[string]any{{
+				"name":                 assetName,
+				"browser_download_url": base + dlPath,
+				"digest":               "sha256:" + digest,
+				"size":                 123,
+			}},
+		})
+		return string(body)
+	}
+	var srv *httptest.Server
+	// Any release tag under this prefix resolves: asr-models → the model, every
+	// other tag → the engine (so the default-version pinned-check test works too).
+	mux.HandleFunc("/repos/k2-fsa/sherpa-onnx/releases/tags/", func(w http.ResponseWriter, r *http.Request) {
+		base = srv.URL
+		if strings.HasSuffix(r.URL.Path, "/asr-models") {
+			fmt.Fprint(w, release(modelAssetName, modelDigest, "/model.tar.bz2"))
+			return
+		}
+		fmt.Fprint(w, release("sherpa-onnx-test-linux-x64-shared-no-tts.tar.bz2", engineDigest, "/engine.tar.bz2"))
+	})
+	mux.HandleFunc("/engine.tar.bz2", func(w http.ResponseWriter, r *http.Request) { w.Write(engineBytes) })
+	mux.HandleFunc("/model.tar.bz2", func(w http.ResponseWriter, r *http.Request) { w.Write(modelBytes) })
+	srv = httptest.NewServer(mux)
+	base = srv.URL
+	t.Cleanup(srv.Close)
+	return srv
+}
+
+func TestResolveEnginePathsHonorsTargetPlatform(t *testing.T) {
+	// The engine binary name follows the TARGET platform, not the host: a Windows
+	// engine has ".exe", a Linux one doesn't. Resolution must find each regardless
+	// of the OS the test runs on (this is what makes the cross-platform download
+	// test pass on Windows CI).
+	for _, tc := range []struct {
+		name    string
+		target  bool
+		binName string
+	}{
+		{"linux target", false, "sherpa-onnx-offline"},
+		{"windows target", true, "sherpa-onnx-offline.exe"},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			dir := t.TempDir()
+			binDir := filepath.Join(dir, "bin")
+			if err := os.MkdirAll(binDir, 0o755); err != nil {
+				t.Fatal(err)
+			}
+			if err := os.WriteFile(filepath.Join(binDir, tc.binName), []byte("x"), 0o755); err != nil {
+				t.Fatal(err)
+			}
+			bin, _ := resolveEnginePaths(dir, tc.target)
+			if !fileExists(bin) {
+				t.Errorf("resolveEnginePaths(target-windows=%v) = %q; not found", tc.target, bin)
+			}
+			if !strings.HasSuffix(bin, tc.binName) {
+				t.Errorf("bin = %q, want suffix %q", bin, tc.binName)
+			}
+		})
+	}
+}
+
+func TestEnsureLocalEngineDownloadsAndExtracts(t *testing.T) {
+	srv := fakeReleaseServer(t, engineSHA, modelSHA)
+	dest := t.TempDir()
+	var stages []string
+	comp, err := EnsureLocalEngine(context.Background(), DownloadOptions{
+		DestRoot:      dest,
+		EngineVersion: "test",
+		APIBase:       srv.URL,
+		platformKey:   "linux-amd64",
+		skipPinned:    true,
+		Progress:      func(s string) { stages = append(stages, s) },
+	})
+	if err != nil {
+		t.Fatalf("EnsureLocalEngine: %v", err)
+	}
+	if !fileExists(comp.BinaryPath) || !strings.HasSuffix(comp.BinaryPath, "sherpa-onnx-offline") {
+		t.Errorf("binary path wrong: %q", comp.BinaryPath)
+	}
+	if !fileExists(comp.ServerPath) {
+		t.Errorf("server path missing: %q", comp.ServerPath)
+	}
+	if !fileExists(filepath.Join(comp.ModelPath, "tokens.txt")) {
+		t.Errorf("model tokens.txt missing under %q", comp.ModelPath)
+	}
+	// The extracted binary keeps its exec bit — Unix only; Windows has no
+	// executable permission bit, so os.FileMode never reports 0o100 there.
+	if runtime.GOOS != "windows" {
+		if info, err := os.Stat(comp.BinaryPath); err == nil && info.Mode()&0o100 == 0 {
+			t.Error("engine binary should be executable")
+		}
+	}
+	if len(stages) == 0 {
+		t.Error("expected progress stages")
+	}
+
+	// Idempotent: a second call reuses the extracted engine + model and downloads
+	// NOTHING (no progress stages) — the fix for re-downloading an already-present
+	// engine whose binary lives in the tarball's flattened subdir.
+	var stages2 []string
+	comp2, err := EnsureLocalEngine(context.Background(), DownloadOptions{
+		DestRoot: dest, EngineVersion: "test", APIBase: srv.URL, platformKey: "linux-amd64", skipPinned: true,
+		Progress: func(s string) { stages2 = append(stages2, s) },
+	})
+	if err != nil || comp2.BinaryPath != comp.BinaryPath {
+		t.Errorf("second call should reuse: %v / %q vs %q", err, comp2.BinaryPath, comp.BinaryPath)
+	}
+	if len(stages2) != 0 {
+		t.Errorf("a fully-cached second call must not download/extract anything, got stages: %v", stages2)
+	}
+}
+
+func TestEnsureLocalEngineRejectsChecksumMismatch(t *testing.T) {
+	// Serve the engine with a wrong digest → verification must refuse to extract.
+	srv := fakeReleaseServer(t, strings.Repeat("0", 64), modelSHA)
+	_, err := EnsureLocalEngine(context.Background(), DownloadOptions{
+		DestRoot: t.TempDir(), EngineVersion: "test", APIBase: srv.URL, platformKey: "linux-amd64", skipPinned: true,
+	})
+	if err == nil || !strings.Contains(err.Error(), "checksum mismatch") {
+		t.Fatalf("expected checksum-mismatch error, got %v", err)
+	}
+}
+
+func TestEnsureLocalEnginePinnedCrossCheckRefusesChangedRelease(t *testing.T) {
+	// Default version + a digest that doesn't match the pinned value → refuse,
+	// even though the download itself would verify against the API digest.
+	srv := fakeReleaseServer(t, engineSHA, modelSHA)
+	_, err := EnsureLocalEngine(context.Background(), DownloadOptions{
+		DestRoot:      t.TempDir(),
+		EngineVersion: DefaultSherpaVersion, // triggers the pinned cross-check
+		APIBase:       srv.URL,
+		platformKey:   "linux-amd64",
+		// skipPinned false: the pinned digest for linux-amd64 != engineSHA fixture.
+	})
+	if err == nil || !strings.Contains(err.Error(), "pinned known-good") {
+		t.Fatalf("expected pinned cross-check refusal, got %v", err)
+	}
+}
+
+func TestResolveAssetMissingDigestAllowed(t *testing.T) {
+	// An older asset with no API digest resolves fine (verification against a
+	// pinned value happens later); resolveAsset no longer refuses here.
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		json.NewEncoder(w).Encode(map[string]any{
+			"assets": []map[string]any{{"name": "x-linux-x64-shared-no-tts.tar.bz2", "browser_download_url": "http://x", "digest": ""}},
+		})
+	}))
+	defer srv.Close()
+	asset, err := resolveAsset(context.Background(), http.DefaultClient, srv.URL, "test", "x-", "linux-x64-shared-no-tts.tar.bz2")
+	if err != nil {
+		t.Fatalf("resolveAsset should tolerate a missing digest, got %v", err)
+	}
+	if asset.sha256 != "" {
+		t.Errorf("expected empty digest, got %q", asset.sha256)
+	}
+}
+
+func TestUnverifiableDownloadRefused(t *testing.T) {
+	// No API digest AND no pinned digest → refuse to download/extract.
+	srv := fakeReleaseServer(t, "", "") // both assets served with empty digest
+	_, err := EnsureLocalEngine(context.Background(), DownloadOptions{
+		DestRoot: t.TempDir(), EngineVersion: "test", APIBase: srv.URL, platformKey: "linux-amd64", skipPinned: true,
+	})
+	if err == nil || !strings.Contains(err.Error(), "unverifiable") {
+		t.Fatalf("expected refusal of an unverifiable download, got %v", err)
+	}
+}
+
+func TestProgressReaderReportsPercent(t *testing.T) {
+	data := make([]byte, 1000)
+	var last string
+	pr := &progressReader{r: bytesReader(data), total: 1000, label: "Engine", report: func(s string) { last = s }, lastPct: -1}
+	buf := make([]byte, 100)
+	for {
+		if _, err := pr.Read(buf); err != nil {
+			break
+		}
+	}
+	if !strings.Contains(last, "Engine") || !strings.Contains(last, "100%") {
+		t.Errorf("expected a final 'Engine 100%%' report, got %q", last)
+	}
+}
+
+func bytesReader(b []byte) *sliceReader { return &sliceReader{b: b} }
+
+type sliceReader struct {
+	b   []byte
+	pos int
+}
+
+func (s *sliceReader) Read(p []byte) (int, error) {
+	if s.pos >= len(s.b) {
+		return 0, io.EOF
+	}
+	n := copy(p, s.b[s.pos:])
+	s.pos += n
+	return n, nil
+}
+
+func TestUnsupportedPlatformSetupError(t *testing.T) {
+	_, err := EnsureLocalEngine(context.Background(), DownloadOptions{DestRoot: t.TempDir(), platformKey: "plan9-mips"})
+	var setupErr *SetupError
+	if !errors.As(err, &setupErr) {
+		t.Fatalf("want *SetupError for unsupported platform, got %v", err)
+	}
+}

internal/dictation/listmodels_test.go

@@ -0,0 +1,67 @@
+package dictation
+
+import (
+	"context"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+)
+
+func TestListModelsFiltersAndFlags(t *testing.T) {
+	assets := []map[string]any{
+		{"name": "sherpa-onnx-streaming-zipformer-en-2023-06-26.tar.bz2", "size": 1 << 20, "digest": "sha256:abc"},
+		{"name": "sherpa-onnx-whisper-tiny.en.tar.bz2", "size": 2 << 20},
+		{"name": "sherpa-onnx-moonshine-tiny-en-int8.tar.bz2", "size": 3 << 20},                  // curated but not flagged recommended
+		{"name": "sherpa-onnx-streaming-zipformer-en-kroko-2025-08-06.tar.bz2", "size": 1 << 20}, // curated → recommended
+		{"name": "sherpa-onnx-paraformer-zh-2024-03-09.tar.bz2", "size": 4 << 20},
+		// Non-ASR tools that MUST be filtered out:
+		{"name": "sherpa-onnx-vad-silero.tar.bz2", "size": 1},
+		{"name": "vits-piper-en_US-amy.tar.bz2", "size": 1},
+		{"name": "sherpa-onnx-punct-en.tar.bz2", "size": 1},
+		{"name": "sherpa-onnx-kws-zipformer.tar.bz2", "size": 1},
+	}
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		json.NewEncoder(w).Encode(map[string]any{"id": 1, "assets": assets})
+	}))
+	defer srv.Close()
+
+	vs, err := ListModels(context.Background(), http.DefaultClient, srv.URL)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(vs) != 5 {
+		t.Fatalf("expected 5 transcription models (tools filtered), got %d", len(vs))
+	}
+	byID := map[string]ModelVariant{}
+	for _, v := range vs {
+		byID[v.ID] = v
+	}
+	if !byID["streaming-zipformer-en-2023-06-26"].Streaming {
+		t.Error("streaming model should be flagged Streaming")
+	}
+	if byID["whisper-tiny.en"].Streaming {
+		t.Error("whisper is batch, not streaming")
+	}
+	if !byID["streaming-zipformer-en-kroko-2025-08-06"].Recommended {
+		t.Error("a curated model flagged Recommended should be Recommended")
+	}
+	if byID["moonshine-tiny-en-int8"].Recommended {
+		t.Error("moonshine tiny is curated but no longer flagged Recommended (no star)")
+	}
+	if byID["paraformer-zh-2024-03-09"].Recommended {
+		t.Error("an uncurated model should not be Recommended")
+	}
+	// A curated model with no API digest still gets its pinned digest.
+	if byID["moonshine-tiny-en-int8"].Digest == "" {
+		t.Error("curated model should carry a pinned digest")
+	}
+	// An uncurated model with no API digest has an empty digest (TLS-only download).
+	if byID["whisper-tiny.en"].Digest != "" {
+		t.Error("uncurated model without an API digest should have an empty digest")
+	}
+	// The API-provided digest is used when present.
+	if byID["streaming-zipformer-en-2023-06-26"].Digest != "abc" {
+		t.Errorf("expected the API digest, got %q", byID["streaming-zipformer-en-2023-06-26"].Digest)
+	}
+}

internal/dictation/local_smoke_test.go

@@ -0,0 +1,40 @@
+package dictation
+
+import (
+	"context"
+	"os"
+	"testing"
+)
+
+// TestLocalTranscribeRealBinary exercises the full local batch path against a
+// REAL sherpa-onnx-offline binary and model — the manual smoke test §14 calls
+// out as not CI-able. Skipped unless both env vars point at a real install:
+//
+//	ZERO_STT_TEST_BINARY=/path/to/sherpa-onnx-offline \
+//	ZERO_STT_TEST_MODEL=/path/to/model-dir \
+//	ZERO_STT_TEST_WAV=/path/to/16k-mono.wav \
+//	go test ./internal/dictation/ -run TestLocalTranscribeRealBinary -v
+func TestLocalTranscribeRealBinary(t *testing.T) {
+	binary := os.Getenv("ZERO_STT_TEST_BINARY")
+	model := os.Getenv("ZERO_STT_TEST_MODEL")
+	wav := os.Getenv("ZERO_STT_TEST_WAV")
+	if binary == "" || model == "" || wav == "" {
+		t.Skip("set ZERO_STT_TEST_BINARY, ZERO_STT_TEST_MODEL, ZERO_STT_TEST_WAV to run the real-engine smoke test")
+	}
+	audio, err := os.ReadFile(wav)
+	if err != nil {
+		t.Fatalf("read wav: %v", err)
+	}
+	tr, err := NewLocalTranscriber(LocalConfig{Binary: binary, ModelPath: model})
+	if err != nil {
+		t.Fatalf("NewLocalTranscriber: %v", err)
+	}
+	text, err := tr.Transcribe(context.Background(), audio)
+	if err != nil {
+		t.Fatalf("Transcribe: %v", err)
+	}
+	if text == "" {
+		t.Fatal("real engine returned an empty transcript")
+	}
+	t.Logf("transcript: %q", text)
+}

internal/dictation/owner_unix.go

@@ -0,0 +1,24 @@
+//go:build !windows
+
+package dictation
+
+import (
+	"fmt"
+	"os"
+	"syscall"
+)
+
+// checkAudioDirOwner rejects a dictation temp directory not owned by the
+// current user — on a shared /tmp another user could have pre-created the
+// path and would then control its lifetime. Mirrors
+// internal/tools/spill_owner_unix.go.
+func checkAudioDirOwner(info os.FileInfo) error {
+	stat, ok := info.Sys().(*syscall.Stat_t)
+	if !ok {
+		return nil
+	}
+	if int(stat.Uid) != os.Geteuid() {
+		return fmt.Errorf("dictation temp directory is owned by uid %d, not the current user", stat.Uid)
+	}
+	return nil
+}

internal/dictation/owner_windows.go

@@ -0,0 +1,12 @@
+//go:build windows
+
+package dictation
+
+import "os"
+
+// checkAudioDirOwner is a no-op on Windows: %TEMP% is per-user by default and
+// there is no portable uid to compare. The Lstat symlink/dir check in
+// audioSpillDir still applies. Mirrors internal/tools/spill_owner_windows.go.
+func checkAudioDirOwner(os.FileInfo) error {
+	return nil
+}

internal/dictation/recorder.go

@@ -0,0 +1,581 @@
+package dictation
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"math"
+	"os"
+	"os/exec"
+	"runtime"
+	"strconv"
+	"strings"
+	"sync"
+	"time"
+)
+
+// Recorder captures microphone audio. Start/Stop is the batch path (used on
+// Termux, and as the fallback everywhere); StartStreaming is the desktop-only
+// continuous-PCM path.
+type Recorder interface {
+	Start() error
+	// Stop ends a batch recording and returns the recorded audio bytes (WAV on
+	// desktop, M4A on Termux — SniffFormat distinguishes them).
+	Stop() ([]byte, error)
+	// StartStreaming begins continuous capture of raw little-endian PCM16 mono
+	// at the configured sample rate. Chunks arrive on the channel in ~50ms
+	// units; the channel closes when capture ends (stop called, max duration
+	// hit, or trailing-silence auto-stop). stop is idempotent.
+	StartStreaming() (chunks <-chan []byte, stop func() error, err error)
+}
+
+// Platform selects the audio-capture toolchain (§9 of the design doc).
+type Platform string
+
+const (
+	PlatformLinux       Platform = "linux"
+	PlatformDarwin      Platform = "darwin"
+	PlatformWindows     Platform = "windows"
+	PlatformTermux      Platform = "termux"
+	PlatformUnsupported Platform = ""
+)
+
+// DetectPlatform picks the capture toolchain for the current host. Termux is
+// detected by its environment (it reports GOOS=linux/android): TERMUX_VERSION
+// is exported to every Termux shell, and PREFIX carries the app package path.
+func DetectPlatform() Platform {
+	if os.Getenv("TERMUX_VERSION") != "" || strings.Contains(os.Getenv("PREFIX"), "com.termux") {
+		return PlatformTermux
+	}
+	switch runtime.GOOS {
+	case "linux":
+		return PlatformLinux
+	case "darwin":
+		return PlatformDarwin
+	case "windows":
+		return PlatformWindows
+	}
+	return PlatformUnsupported
+}
+
+// StreamingSupported reports whether the platform's capture tool can write raw
+// PCM to stdout. Termux's termux-microphone-record requires a regular output
+// file — a hard tool limitation, so Termux is batch-only permanently.
+func (p Platform) StreamingSupported() bool {
+	switch p {
+	case PlatformLinux, PlatformDarwin, PlatformWindows:
+		return true
+	}
+	return false
+}
+
+const (
+	// DefaultSampleRate is what the local engine and Deepgram consume.
+	// OpenAI Realtime's pcm16 requires 24kHz; its transcriber constructs the
+	// recorder accordingly.
+	DefaultSampleRate = 16000
+	// DefaultMaxDuration is the hard runaway-recording cap (§12), applied on
+	// every platform, batch and streaming. Overridable via
+	// stt.maxDurationSeconds.
+	DefaultMaxDuration = 5 * time.Minute
+
+	// Trailing-silence auto-stop (§9): stop ~2s after the signal falls below a
+	// 3% amplitude threshold — SoX's own `silence 1 0.1 3% 1 2.0 3%` numbers.
+	silenceThresholdPct = 3
+	silenceHold         = 2 * time.Second
+
+	chunkInterval = 50 * time.Millisecond
+	stopGrace     = 3 * time.Second
+)
+
+// RecorderOptions configures a Recorder. Zero values pick sensible defaults.
+type RecorderOptions struct {
+	Platform        Platform
+	SampleRate      int
+	MaxDuration     time.Duration
+	SilenceAutoStop bool
+	// WindowsAudioDevice is the dshow capture device name; auto-detected from
+	// `ffmpeg -list_devices` when empty.
+	WindowsAudioDevice string
+
+	// Test seams (unexported): fake the OS boundary, mirroring how
+	// internal/lsp/manager_test.go stubs serverStarter.
+	starter   processStarter
+	runOutput commandOutputRunner
+	tempDir   func() (string, error)
+	now       func() time.Time
+}
+
+func (o RecorderOptions) withDefaults() RecorderOptions {
+	if o.Platform == PlatformUnsupported {
+		o.Platform = DetectPlatform()
+	}
+	if o.SampleRate <= 0 {
+		o.SampleRate = DefaultSampleRate
+	}
+	if o.MaxDuration <= 0 {
+		o.MaxDuration = DefaultMaxDuration
+	}
+	if o.starter == nil {
+		o.starter = startProcess
+	}
+	if o.runOutput == nil {
+		o.runOutput = runCommandOutput
+	}
+	if o.tempDir == nil {
+		o.tempDir = audioSpillDir
+	}
+	if o.now == nil {
+		o.now = time.Now
+	}
+	return o
+}
+
+// NewRecorder builds the platform recorder. Construction is cheap and
+// side-effect free; tools are only exec'd on Start/StartStreaming.
+func NewRecorder(opts RecorderOptions) Recorder {
+	o := opts.withDefaults()
+	return &recorder{opts: o}
+}
+
+type recorder struct {
+	opts RecorderOptions
+
+	mu        sync.Mutex
+	proc      processHandle
+	filePath  string
+	recording bool
+}
+
+var errAlreadyRecording = errors.New("a recording is already in progress")
+
+// Start begins a batch (record-to-file) capture.
+func (r *recorder) Start() error {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	if r.recording {
+		return errAlreadyRecording
+	}
+	if r.opts.Platform == PlatformUnsupported {
+		return fmt.Errorf("dictation is not supported on %s", runtime.GOOS)
+	}
+
+	dir, err := r.opts.tempDir()
+	if err != nil {
+		return fmt.Errorf("preparing dictation temp dir: %w", err)
+	}
+	ext := ".wav"
+	if r.opts.Platform == PlatformTermux {
+		ext = ".m4a"
+	}
+	tmp, err := os.CreateTemp(dir, "rec-*"+ext)
+	if err != nil {
+		return fmt.Errorf("creating dictation temp file: %w", err)
+	}
+	path := tmp.Name()
+	_ = tmp.Close()
+
+	if r.opts.Platform == PlatformTermux {
+		// termux-microphone-record starts the recording in the Termux:API
+		// companion app and exits immediately; there is no long-lived process
+		// to hold. Stop is a separate `-q` invocation.
+		ctx, cancel := context.WithTimeout(context.Background(), 10*time.Second)
+		defer cancel()
+		out, err := r.opts.runOutput(ctx, "termux-microphone-record",
+			"-l", strconv.Itoa(maxSeconds(r.opts.MaxDuration)), "-f", path)
+		if err != nil {
+			_ = os.Remove(path)
+			return startError("termux-microphone-record", r.opts.Platform, err, out)
+		}
+		r.filePath = path
+		r.recording = true
+		return nil
+	}
+
+	spec, err := batchCommand(r.opts, path)
+	if err != nil {
+		_ = os.Remove(path)
+		return err
+	}
+	proc, _, err := r.opts.starter(spec)
+	if err != nil {
+		_ = os.Remove(path)
+		return startError(spec.name, r.opts.Platform, err, nil)
+	}
+	r.proc = proc
+	r.filePath = path
+	r.recording = true
+	return nil
+}
+
+// Stop ends a batch capture, reads the recorded file, deletes it, and returns
+// the audio bytes.
+func (r *recorder) Stop() ([]byte, error) {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	if !r.recording {
+		return nil, errors.New("no recording in progress")
+	}
+	proc, path := r.proc, r.filePath
+	// Termux: Start's termux-microphone-record branch leaves r.proc nil because
+	// there's no long-lived process to hold (Stop is a separate `-q` invocation
+	// via runOutput). Treat that as legitimate; only reject a nil proc that
+	// came from StartStreaming (which owns its process via the returned
+	// closure).
+	if proc == nil && r.opts.Platform != PlatformTermux {
+		return nil, errors.New("recording was started with StartStreaming; call the stop closure returned from StartStreaming instead of Stop()")
+	}
+	r.proc, r.filePath, r.recording = nil, "", false
+	defer os.Remove(path)
+
+	if r.opts.Platform == PlatformTermux {
+		ctx, cancel := context.WithTimeout(context.Background(), 10*time.Second)
+		defer cancel()
+		if out, err := r.opts.runOutput(ctx, "termux-microphone-record", "-q"); err != nil {
+			return nil, fmt.Errorf("stopping termux recording: %w (%s)", err, strings.TrimSpace(string(out)))
+		}
+		return readRecordingWithRetry(path, r.opts.now)
+	}
+
+	if err := proc.StopGracefully(); err != nil {
+		_ = proc.Kill()
+	}
+	// A signal-terminated capture tool reports a non-zero exit — expected, not
+	// an error. The recorded file's contents are the real success signal.
+	_ = waitWithTimeout(proc, stopGrace)
+	data, err := os.ReadFile(path)
+	if err != nil {
+		return nil, fmt.Errorf("reading recorded audio: %w", err)
+	}
+	if len(data) == 0 {
+		return nil, errors.New("no audio was captured (is the microphone available?)")
+	}
+	return data, nil
+}
+
+// readRecordingWithRetry reads the Termux recording, retrying briefly: the
+// companion app finalizes the file asynchronously after `-q` returns.
+func readRecordingWithRetry(path string, now func() time.Time) ([]byte, error) {
+	deadline := now().Add(2 * time.Second)
+	for {
+		data, err := os.ReadFile(path)
+		if err == nil && len(data) > 0 {
+			return data, nil
+		}
+		if now().After(deadline) {
+			if err != nil {
+				return nil, fmt.Errorf("reading recorded audio: %w", err)
+			}
+			return nil, errors.New("no audio was captured (check Termux:API app and microphone permission)")
+		}
+		time.Sleep(50 * time.Millisecond)
+	}
+}
+
+// batchCommand builds the record-to-file argv for desktop platforms (§9).
+// The max-duration cap is enforced by the tool itself so a runaway recording
+// ends even if Zero never calls Stop.
+func batchCommand(opts RecorderOptions, path string) (commandSpec, error) {
+	rate := strconv.Itoa(opts.SampleRate)
+	maxSec := strconv.Itoa(maxSeconds(opts.MaxDuration))
+	switch opts.Platform {
+	case PlatformLinux:
+		return commandSpec{
+			name: "arecord",
+			args: []string{"-q", "-f", "S16_LE", "-r", rate, "-c", "1", "-t", "wav", "-d", maxSec, path},
+		}, nil
+	case PlatformDarwin:
+		args := []string{"-q", "-d", "-r", rate, "-c", "1", "-b", "16", "-e", "signed-integer", path, "trim", "0", maxSec}
+		if opts.SilenceAutoStop {
+			// SoX's built-in trailing-silence auto-stop: end ~2s after the
+			// signal falls below 3% amplitude. A backstop for a forgotten
+			// manual stop, not VAD — recording still starts on an explicit key.
+			args = append(args, "silence", "1", "0.1", "3%", "1", "2.0", "3%")
+		}
+		return commandSpec{name: "sox", args: args}, nil
+	case PlatformWindows:
+		device := opts.WindowsAudioDevice
+		if device == "" {
+			detected, err := detectWindowsAudioDevice(opts.runOutput)
+			if err != nil {
+				return commandSpec{}, err
+			}
+			device = detected
+		}
+		return commandSpec{
+			name: "ffmpeg",
+			args: []string{"-hide_banner", "-loglevel", "error", "-f", "dshow", "-i", "audio=" + device,
+				"-ac", "1", "-ar", rate, "-c:a", "pcm_s16le", "-t", maxSec, "-y", path},
+			stopViaStdin: "q",
+		}, nil
+	}
+	return commandSpec{}, fmt.Errorf("no batch capture tool for platform %q", opts.Platform)
+}
+
+// streamCommand builds the raw-PCM-to-stdout argv for desktop platforms (§9).
+// Same tools as batch, invoked with their stdout flag — Termux has no stdout
+// mode, which is why it is batch-only.
+func streamCommand(opts RecorderOptions) (commandSpec, error) {
+	rate := strconv.Itoa(opts.SampleRate)
+	switch opts.Platform {
+	case PlatformLinux:
+		return commandSpec{
+			name:       "arecord",
+			args:       []string{"-q", "-t", "raw", "-f", "S16_LE", "-r", rate, "-c", "1", "-"},
+			wantStdout: true,
+		}, nil
+	case PlatformDarwin:
+		return commandSpec{
+			name:       "sox",
+			args:       []string{"-q", "-d", "-t", "raw", "-r", rate, "-e", "signed", "-b", "16", "-c", "1", "-"},
+			wantStdout: true,
+		}, nil
+	case PlatformWindows:
+		device := opts.WindowsAudioDevice
+		if device == "" {
+			detected, err := detectWindowsAudioDevice(opts.runOutput)
+			if err != nil {
+				return commandSpec{}, err
+			}
+			device = detected
+		}
+		return commandSpec{
+			name: "ffmpeg",
+			args: []string{"-hide_banner", "-loglevel", "error", "-f", "dshow", "-i", "audio=" + device,
+				"-ac", "1", "-ar", rate, "-f", "s16le", "-"},
+			wantStdout:   true,
+			stopViaStdin: "q",
+		}, nil
+	}
+	return commandSpec{}, fmt.Errorf("streaming capture is not supported on platform %q", opts.Platform)
+}
+
+// StartStreaming begins continuous PCM capture with the §12 safety nets: a
+// hard max-duration cap and Go-side trailing-silence auto-stop (the streaming
+// analog of SoX's silence effect — we see the samples, so we measure them).
+func (r *recorder) StartStreaming() (<-chan []byte, func() error, error) {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	if r.recording {
+		return nil, nil, errAlreadyRecording
+	}
+	spec, err := streamCommand(r.opts)
+	if err != nil {
+		return nil, nil, err
+	}
+	proc, stdout, err := r.opts.starter(spec)
+	if err != nil {
+		return nil, nil, startError(spec.name, r.opts.Platform, err, nil)
+	}
+	if stdout == nil {
+		_ = proc.Kill()
+		return nil, nil, errors.New("capture process has no stdout pipe")
+	}
+	r.recording = true
+
+	chunkBytes := chunkSizeBytes(r.opts.SampleRate)
+	chunks := make(chan []byte, 8)
+	done := make(chan struct{})
+	var stopOnce sync.Once
+	stop := func() error {
+		stopOnce.Do(func() {
+			close(done)
+			if err := proc.StopGracefully(); err != nil {
+				_ = proc.Kill()
+			}
+			_ = waitWithTimeout(proc, stopGrace)
+			_ = stdout.Close()
+			r.mu.Lock()
+			r.recording = false
+			r.mu.Unlock()
+		})
+		return nil
+	}
+
+	// Max-duration cap: streaming has no tool-side -d/-t equivalent that also
+	// flushes stdout cleanly everywhere, so the cap runs on our side of the pipe.
+	capTimer := time.AfterFunc(r.opts.MaxDuration, func() { _ = stop() })
+
+	go func() {
+		defer close(chunks)
+		defer capTimer.Stop()
+		// Tear down on ANY exit — including the capture process ending on its own
+		// (EOF/crash), which the loop below returns on without going through an
+		// explicit stop(). stop() is idempotent (sync.Once), so this is safe even
+		// when the session was already stopped externally; it clears r.recording and
+		// closes proc/stdout so a self-terminating device doesn't wedge the recorder.
+		defer func() { _ = stop() }()
+		threshold := int16(32767 * silenceThresholdPct / 100)
+		silentChunks, speechSeen := 0, false
+		silenceLimit := int(silenceHold / chunkInterval)
+		for {
+			buf := make([]byte, chunkBytes)
+			n, err := readFullOrEOF(stdout, buf)
+			if n > 0 {
+				chunk := buf[:n]
+				if r.opts.SilenceAutoStop {
+					if chunkPeak(chunk) >= threshold {
+						speechSeen = true
+						silentChunks = 0
+					} else if speechSeen {
+						silentChunks++
+						if silentChunks >= silenceLimit {
+							go func() { _ = stop() }()
+						}
+					}
+				}
+				select {
+				case chunks <- chunk:
+				case <-done:
+					return
+				}
+			}
+			if err != nil {
+				return
+			}
+		}
+	}()
+	return chunks, stop, nil
+}
+
+// chunkSizeBytes is 50ms of PCM16 mono at the given rate — small chunks mean
+// faster time-to-first-partial (§6a).
+func chunkSizeBytes(sampleRate int) int {
+	return sampleRate * 2 / int(time.Second/chunkInterval)
+}
+
+// readFullOrEOF fills buf from r, returning what it read and any terminal
+// error. A short final read (pipe closing mid-chunk) still returns its bytes.
+func readFullOrEOF(r interface{ Read([]byte) (int, error) }, buf []byte) (int, error) {
+	total := 0
+	for total < len(buf) {
+		n, err := r.Read(buf[total:])
+		total += n
+		if err != nil {
+			return total, err
+		}
+	}
+	return total, nil
+}
+
+// chunkPeak returns the loudest absolute sample in a little-endian PCM16 chunk.
+func chunkPeak(chunk []byte) int16 {
+	var peak int16
+	for i := 0; i+1 < len(chunk); i += 2 {
+		s := int16(uint16(chunk[i]) | uint16(chunk[i+1])<<8)
+		if s == -32768 {
+			return 32767
+		}
+		if s < 0 {
+			s = -s
+		}
+		if s > peak {
+			peak = s
+		}
+	}
+	return peak
+}
+
+// levelFloorDB and levelCeilDB bound the RMS→meter mapping. Speech RMS is small
+// (roughly -45 dBFS quiet to -15 dBFS loud), so a LINEAR scale barely moves the
+// bars — a dB (logarithmic) scale, like real audio meters, is what makes normal
+// speech span the meter.
+const (
+	levelFloorDB = -50.0 // ≈ near-silence → empty
+	levelCeilDB  = -15.0 // loud speech → full
+)
+
+// ChunkLevel returns a perceptual 0..1 loudness for a little-endian PCM16 chunk,
+// driving the live recording waveform from the actual microphone input. It uses
+// an RMS→dBFS→linear mapping over [levelFloorDB, levelCeilDB].
+func ChunkLevel(chunk []byte) float64 {
+	var sumSq float64
+	n := 0
+	for i := 0; i+1 < len(chunk); i += 2 {
+		s := int16(uint16(chunk[i]) | uint16(chunk[i+1])<<8)
+		f := float64(s) / 32768.0
+		sumSq += f * f
+		n++
+	}
+	if n == 0 {
+		return 0
+	}
+	rms := math.Sqrt(sumSq / float64(n))
+	if rms <= 0 {
+		return 0
+	}
+	db := 20 * math.Log10(rms) // dBFS (negative)
+	level := (db - levelFloorDB) / (levelCeilDB - levelFloorDB)
+	if level < 0 {
+		level = 0
+	}
+	if level > 1 {
+		level = 1
+	}
+	return level
+}
+
+func maxSeconds(d time.Duration) int {
+	sec := int(d / time.Second)
+	if sec < 1 {
+		sec = 1
+	}
+	return sec
+}
+
+// detectWindowsAudioDevice picks the first dshow audio capture device from
+// `ffmpeg -list_devices`. dshow has no "default" alias, so a concrete device
+// name is required; stt.windowsAudioDevice overrides the auto-pick.
+func detectWindowsAudioDevice(runOutput commandOutputRunner) (string, error) {
+	ctx, cancel := context.WithTimeout(context.Background(), 10*time.Second)
+	defer cancel()
+	out, _ := runOutput(ctx, "ffmpeg", "-hide_banner", "-list_devices", "true", "-f", "dshow", "-i", "dummy")
+	// Listing always "fails" (dummy isn't a real input); the device table is
+	// in the output regardless. Lines look like: [dshow @ ...] "Mic name" (audio)
+	for _, line := range strings.Split(string(out), "\n") {
+		if !strings.Contains(line, "(audio)") {
+			continue
+		}
+		start := strings.Index(line, `"`)
+		if start < 0 {
+			continue
+		}
+		rest := line[start+1:]
+		end := strings.Index(rest, `"`)
+		if end <= 0 {
+			continue
+		}
+		return rest[:end], nil
+	}
+	return "", &SetupError{
+		Tool: "ffmpeg",
+		Hint: "no dshow audio capture device found — set stt.windowsAudioDevice to your microphone's name (list with: ffmpeg -list_devices true -f dshow -i dummy)",
+	}
+}
+
+// startError maps a failed capture-tool launch to setup guidance: a missing
+// binary is an expected unconfigured state (matching internal/lsp's
+// missing-server philosophy), not an internal error.
+func startError(tool string, platform Platform, err error, output []byte) error {
+	if errors.Is(err, exec.ErrNotFound) {
+		return &SetupError{Tool: tool, Hint: installHint(platform)}
+	}
+	if len(output) > 0 {
+		return fmt.Errorf("starting %s: %w (%s)", tool, err, strings.TrimSpace(string(output)))
+	}
+	return fmt.Errorf("starting %s: %w", tool, err)
+}
+
+func installHint(platform Platform) string {
+	switch platform {
+	case PlatformLinux:
+		return "install alsa-utils to enable dictation (e.g. sudo dnf install alsa-utils / sudo apt install alsa-utils)"
+	case PlatformDarwin:
+		return "install sox to enable dictation (brew install sox)"
+	case PlatformWindows:
+		return "install ffmpeg and ensure it is on PATH to enable dictation"
+	case PlatformTermux:
+		return "run `pkg install termux-api` and install the Termux:API companion app to enable dictation"
+	}
+	return "dictation is not supported on this platform"
+}

internal/dictation/recorder_test.go

@@ -0,0 +1,501 @@
+package dictation
+
+import (
+	"bytes"
+	"context"
+	"encoding/binary"
+	"errors"
+	"io"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"sync"
+	"testing"
+	"time"
+)
+
+// fakeProcess is a processHandle whose stdout is driven from an in-memory
+// buffer, mirroring how internal/lsp/manager_test.go stubs its subprocess over
+// in-memory pipes rather than spawning a real one.
+type fakeProcess struct {
+	stopped   bool
+	killed    bool
+	waitErr   error
+	stopErr   error
+	closeOnce sync.Once
+	done      chan struct{}
+}
+
+func newFakeProcess() *fakeProcess { return &fakeProcess{done: make(chan struct{})} }
+
+func (f *fakeProcess) StopGracefully() error {
+	f.stopped = true
+	f.closeOnce.Do(func() { close(f.done) })
+	return f.stopErr
+}
+func (f *fakeProcess) Wait() error {
+	<-f.done
+	return f.waitErr
+}
+func (f *fakeProcess) Kill() error {
+	f.killed = true
+	f.closeOnce.Do(func() { close(f.done) })
+	return nil
+}
+
+// captured records the last spec a starter was asked to launch, so tests can
+// assert the exact discrete argv per platform.
+type captured struct {
+	spec commandSpec
+}
+
+// fileWritingStarter simulates a record-to-file batch tool: on Start it writes
+// fixture bytes to the output path (last positional arg for arecord/sox, or the
+// -y target for ffmpeg) and returns a process that "records" until stopped.
+func fileWritingStarter(t *testing.T, cap *captured, fixture []byte) processStarter {
+	t.Helper()
+	return func(spec commandSpec) (processHandle, io.ReadCloser, error) {
+		cap.spec = spec
+		path := outputPath(spec)
+		if path == "" {
+			t.Fatalf("could not determine output path from argv: %v", spec.args)
+		}
+		if err := os.WriteFile(path, fixture, 0o600); err != nil {
+			t.Fatalf("fixture write: %v", err)
+		}
+		return newFakeProcess(), nil, nil
+	}
+}
+
+func outputPath(spec commandSpec) string {
+	for i, a := range spec.args {
+		if a == "-y" && i+1 < len(spec.args) {
+			return spec.args[i+1]
+		}
+	}
+	// arecord/sox: the path is the last arg (sox appends trim/silence effects
+	// after it, so scan from the front for the first existing-dir target).
+	last := spec.args[len(spec.args)-1]
+	if filepath.IsAbs(last) {
+		return last
+	}
+	for _, a := range spec.args {
+		if filepath.IsAbs(a) && filepath.Ext(a) != "" {
+			return a
+		}
+	}
+	return ""
+}
+
+func wavFixture() []byte {
+	// Minimal valid RIFF/WAVE header + a little data, enough for SniffFormat.
+	var b bytes.Buffer
+	b.WriteString("RIFF")
+	binary.Write(&b, binary.LittleEndian, uint32(36))
+	b.WriteString("WAVE")
+	b.WriteString("fmt ")
+	binary.Write(&b, binary.LittleEndian, uint32(16))
+	b.Write(make([]byte, 16))
+	b.WriteString("data")
+	binary.Write(&b, binary.LittleEndian, uint32(4))
+	b.Write([]byte{1, 2, 3, 4})
+	return b.Bytes()
+}
+
+func testTempDir(t *testing.T) func() (string, error) {
+	dir := t.TempDir()
+	return func() (string, error) { return dir, nil }
+}
+
+func TestBatchRecordLinuxArgv(t *testing.T) {
+	cap := &captured{}
+	fixture := wavFixture()
+	r := NewRecorder(RecorderOptions{
+		Platform: PlatformLinux,
+		starter:  fileWritingStarter(t, cap, fixture),
+		tempDir:  testTempDir(t),
+	})
+	if err := r.Start(); err != nil {
+		t.Fatalf("Start: %v", err)
+	}
+	got, err := r.Stop()
+	if err != nil {
+		t.Fatalf("Stop: %v", err)
+	}
+	if !bytes.Equal(got, fixture) {
+		t.Fatalf("audio mismatch: got %d bytes", len(got))
+	}
+	if cap.spec.name != "arecord" {
+		t.Errorf("tool = %q, want arecord", cap.spec.name)
+	}
+	assertArgs(t, cap.spec.args, "-f", "S16_LE")
+	assertArgs(t, cap.spec.args, "-r", "16000")
+	assertArgs(t, cap.spec.args, "-c", "1")
+}
+
+func TestBatchRecordDarwinArgv(t *testing.T) {
+	cap := &captured{}
+	r := NewRecorder(RecorderOptions{
+		Platform:        PlatformDarwin,
+		SilenceAutoStop: true,
+		starter:         fileWritingStarter(t, cap, wavFixture()),
+		tempDir:         testTempDir(t),
+	})
+	if err := r.Start(); err != nil {
+		t.Fatalf("Start: %v", err)
+	}
+	if _, err := r.Stop(); err != nil {
+		t.Fatalf("Stop: %v", err)
+	}
+	if cap.spec.name != "sox" {
+		t.Errorf("tool = %q, want sox", cap.spec.name)
+	}
+	if !containsSubseq(cap.spec.args, []string{"silence", "1", "0.1", "3%", "1", "2.0", "3%"}) {
+		t.Errorf("expected SoX silence auto-stop effect in argv: %v", cap.spec.args)
+	}
+}
+
+func TestBatchRecordWindowsArgv(t *testing.T) {
+	cap := &captured{}
+	r := NewRecorder(RecorderOptions{
+		Platform:           PlatformWindows,
+		WindowsAudioDevice: "Microphone (Realtek)",
+		starter:            fileWritingStarter(t, cap, wavFixture()),
+		tempDir:            testTempDir(t),
+	})
+	if err := r.Start(); err != nil {
+		t.Fatalf("Start: %v", err)
+	}
+	if _, err := r.Stop(); err != nil {
+		t.Fatalf("Stop: %v", err)
+	}
+	if cap.spec.name != "ffmpeg" {
+		t.Errorf("tool = %q, want ffmpeg", cap.spec.name)
+	}
+	assertArgs(t, cap.spec.args, "-i", "audio=Microphone (Realtek)")
+	if cap.spec.stopViaStdin != "q" {
+		t.Errorf("ffmpeg should stop via stdin 'q', got %q", cap.spec.stopViaStdin)
+	}
+}
+
+func TestBatchRecordTermuxUsesCommandRunner(t *testing.T) {
+	dir := t.TempDir()
+	var calls [][]string
+	runOutput := func(_ context.Context, name string, args ...string) ([]byte, error) {
+		calls = append(calls, append([]string{name}, args...))
+		if len(args) > 0 && args[0] == "-l" {
+			// The -f target is the last arg; write the fixture there.
+			path := args[len(args)-1]
+			_ = os.WriteFile(path, m4aFixture(), 0o600)
+		}
+		return nil, nil
+	}
+	r := NewRecorder(RecorderOptions{
+		Platform: PlatformTermux,
+		starter: func(commandSpec) (processHandle, io.ReadCloser, error) {
+			t.Fatal("termux must not use the process starter")
+			return nil, nil, nil
+		},
+		runOutput: runOutput,
+		tempDir:   func() (string, error) { return dir, nil },
+	})
+	if err := r.Start(); err != nil {
+		t.Fatalf("Start: %v", err)
+	}
+	got, err := r.Stop()
+	if err != nil {
+		t.Fatalf("Stop: %v", err)
+	}
+	if SniffFormat(got) != FormatM4A {
+		t.Errorf("expected m4a audio, got format %q", SniffFormat(got))
+	}
+	if len(calls) != 2 || calls[0][0] != "termux-microphone-record" || calls[1][len(calls[1])-1] != "-q" {
+		t.Errorf("unexpected termux calls: %v", calls)
+	}
+}
+
+func TestStartMissingBinaryReturnsSetupError(t *testing.T) {
+	r := NewRecorder(RecorderOptions{
+		Platform: PlatformLinux,
+		starter: func(commandSpec) (processHandle, io.ReadCloser, error) {
+			return nil, nil, exec.ErrNotFound
+		},
+		tempDir: testTempDir(t),
+	})
+	err := r.Start()
+	var setupErr *SetupError
+	if !errors.As(err, &setupErr) {
+		t.Fatalf("want *SetupError for missing binary, got %v", err)
+	}
+	if setupErr.Tool != "arecord" {
+		t.Errorf("SetupError.Tool = %q, want arecord", setupErr.Tool)
+	}
+}
+
+func TestDoubleStartRejected(t *testing.T) {
+	r := NewRecorder(RecorderOptions{
+		Platform: PlatformLinux,
+		starter:  fileWritingStarter(t, &captured{}, wavFixture()),
+		tempDir:  testTempDir(t),
+	})
+	if err := r.Start(); err != nil {
+		t.Fatalf("Start: %v", err)
+	}
+	if err := r.Start(); !errors.Is(err, errAlreadyRecording) {
+		t.Errorf("second Start should be rejected, got %v", err)
+	}
+}
+
+func TestStreamingChunksAndStop(t *testing.T) {
+	// 300ms of PCM16 @16kHz = 6 chunks of 50ms. Feed a fixed buffer, then EOF.
+	pcm := make([]byte, chunkSizeBytes(16000)*6)
+	starter := func(spec commandSpec) (processHandle, io.ReadCloser, error) {
+		if !spec.wantStdout {
+			t.Fatal("streaming spec must request stdout")
+		}
+		return newFakeProcess(), io.NopCloser(bytes.NewReader(pcm)), nil
+	}
+	r := NewRecorder(RecorderOptions{
+		Platform: PlatformLinux,
+		starter:  starter,
+		tempDir:  testTempDir(t),
+	})
+	chunks, stop, err := r.StartStreaming()
+	if err != nil {
+		t.Fatalf("StartStreaming: %v", err)
+	}
+	var total int
+	for c := range chunks {
+		total += len(c)
+	}
+	_ = stop()
+	if total != len(pcm) {
+		t.Errorf("streamed %d bytes, want %d", total, len(pcm))
+	}
+}
+
+func TestStreamingStopIdempotent(t *testing.T) {
+	// A blocking reader so the capture goroutine parks until we stop it.
+	pr, pw := io.Pipe()
+	defer pw.Close()
+	starter := func(commandSpec) (processHandle, io.ReadCloser, error) {
+		return newFakeProcess(), pr, nil
+	}
+	r := NewRecorder(RecorderOptions{Platform: PlatformLinux, starter: starter, tempDir: testTempDir(t)})
+	chunks, stop, err := r.StartStreaming()
+	if err != nil {
+		t.Fatalf("StartStreaming: %v", err)
+	}
+	if err := stop(); err != nil {
+		t.Fatalf("stop: %v", err)
+	}
+	if err := stop(); err != nil {
+		t.Fatalf("second stop should be a no-op, got %v", err)
+	}
+	// Draining the channel must terminate now that we've stopped.
+	drained := make(chan struct{})
+	go func() {
+		for range chunks {
+		}
+		close(drained)
+	}()
+	select {
+	case <-drained:
+	case <-time.After(2 * time.Second):
+		t.Fatal("chunks channel did not close after stop")
+	}
+}
+
+func TestBatchMaxDurationCapInArgv(t *testing.T) {
+	cap := &captured{}
+	r := NewRecorder(RecorderOptions{
+		Platform:    PlatformLinux,
+		MaxDuration: 90 * time.Second,
+		starter:     fileWritingStarter(t, cap, wavFixture()),
+		tempDir:     testTempDir(t),
+	})
+	if err := r.Start(); err != nil {
+		t.Fatalf("Start: %v", err)
+	}
+	if _, err := r.Stop(); err != nil {
+		t.Fatalf("Stop: %v", err)
+	}
+	// arecord's -d bounds a runaway recording even if Stop is never called.
+	assertArgs(t, cap.spec.args, "-d", "90")
+}
+
+func TestRecordedTempFileDeletedAfterStop(t *testing.T) {
+	dir := t.TempDir()
+	cap := &captured{}
+	r := NewRecorder(RecorderOptions{
+		Platform: PlatformLinux,
+		starter:  fileWritingStarter(t, cap, wavFixture()),
+		tempDir:  func() (string, error) { return dir, nil },
+	})
+	if err := r.Start(); err != nil {
+		t.Fatalf("Start: %v", err)
+	}
+	if _, err := r.Stop(); err != nil {
+		t.Fatalf("Stop: %v", err)
+	}
+	entries, _ := os.ReadDir(dir)
+	if len(entries) != 0 {
+		t.Errorf("recorded audio should be deleted after Stop; found %d files", len(entries))
+	}
+}
+
+func TestStreamingSilenceAutoStop(t *testing.T) {
+	// Speech (loud) then a run of silence longer than silenceHold must auto-stop.
+	rate := 16000
+	chunk := chunkSizeBytes(rate)
+	loud := make([]byte, chunk)
+	for i := 0; i+1 < len(loud); i += 2 {
+		loud[i], loud[i+1] = 0x00, 0x40 // +16384, well above threshold
+	}
+	silence := make([]byte, chunk) // zeros
+	var buf []byte
+	buf = append(buf, loud...)
+	silentChunks := int(silenceHold/chunkInterval) + 3
+	for i := 0; i < silentChunks; i++ {
+		buf = append(buf, silence...)
+	}
+	starter := func(commandSpec) (processHandle, io.ReadCloser, error) {
+		return newFakeProcess(), io.NopCloser(bytes.NewReader(buf)), nil
+	}
+	r := NewRecorder(RecorderOptions{
+		Platform:        PlatformLinux,
+		SampleRate:      rate,
+		SilenceAutoStop: true,
+		starter:         starter,
+		tempDir:         testTempDir(t),
+	})
+	chunks, stop, err := r.StartStreaming()
+	if err != nil {
+		t.Fatalf("StartStreaming: %v", err)
+	}
+	defer stop()
+	// Draining must terminate on its own once silence auto-stop fires.
+	drained := make(chan struct{})
+	go func() {
+		for range chunks {
+		}
+		close(drained)
+	}()
+	select {
+	case <-drained:
+	case <-time.After(3 * time.Second):
+		t.Fatal("silence auto-stop did not end the stream")
+	}
+}
+
+func TestChunkPeak(t *testing.T) {
+	loud := []byte{0x00, 0x40} // +16384
+	if got := chunkPeak(loud); got != 16384 {
+		t.Errorf("peak = %d, want 16384", got)
+	}
+	quiet := []byte{0x10, 0x00} // +16
+	if got := chunkPeak(quiet); got != 16 {
+		t.Errorf("peak = %d, want 16", got)
+	}
+	min := []byte{0x00, 0x80} // -32768 -> clamped to 32767
+	if got := chunkPeak(min); got != 32767 {
+		t.Errorf("peak = %d, want 32767 (clamp)", got)
+	}
+}
+
+func TestDetectWindowsAudioDevice(t *testing.T) {
+	listing := `[dshow @ 0000] DirectShow audio devices
+[dshow @ 0000]  "Microphone (USB Audio)" (audio)
+[dshow @ 0000]     Alternative name "@device_cm_..."`
+	runOutput := func(context.Context, string, ...string) ([]byte, error) { return []byte(listing), nil }
+	got, err := detectWindowsAudioDevice(runOutput)
+	if err != nil {
+		t.Fatalf("detect: %v", err)
+	}
+	if got != "Microphone (USB Audio)" {
+		t.Errorf("device = %q", got)
+	}
+}
+
+func TestSniffFormat(t *testing.T) {
+	if SniffFormat(wavFixture()) != FormatWAV {
+		t.Error("wav not detected")
+	}
+	if SniffFormat(m4aFixture()) != FormatM4A {
+		t.Error("m4a not detected")
+	}
+	if SniffFormat([]byte("nope")) != FormatUnknown {
+		t.Error("unknown not detected")
+	}
+}
+
+func m4aFixture() []byte {
+	b := make([]byte, 16)
+	copy(b[4:], []byte("ftyp"))
+	copy(b[8:], []byte("M4A "))
+	return b
+}
+
+func assertArgs(t *testing.T, args []string, want ...string) {
+	t.Helper()
+	if !containsSubseq(args, want) {
+		t.Errorf("argv %v missing subsequence %v", args, want)
+	}
+}
+
+func containsSubseq(hay, needle []string) bool {
+	if len(needle) == 0 {
+		return true
+	}
+	for i := 0; i+len(needle) <= len(hay); i++ {
+		if equalSlice(hay[i:i+len(needle)], needle) {
+			return true
+		}
+	}
+	return false
+}
+
+func equalSlice(a, b []string) bool {
+	if len(a) != len(b) {
+		return false
+	}
+	for i := range a {
+		if a[i] != b[i] {
+			return false
+		}
+	}
+	return true
+}
+
+func TestChunkLevel(t *testing.T) {
+	// Silence → ~0.
+	silence := make([]byte, 640)
+	if lvl := ChunkLevel(silence); lvl > 0.01 {
+		t.Errorf("silence level = %f, want ~0", lvl)
+	}
+	// A loud full-scale tone → clamped to 1.
+	loud := make([]byte, 640)
+	for i := 0; i+1 < len(loud); i += 2 {
+		loud[i], loud[i+1] = 0xFF, 0x7F // +32767
+	}
+	if lvl := ChunkLevel(loud); lvl < 0.99 {
+		t.Errorf("full-scale level = %f, want ~1", lvl)
+	}
+	// A normal speech-level signal (~-30 dBFS) should land mid-meter, not pinned
+	// to the bottom (the whole point of the dB mapping).
+	mid := make([]byte, 640)
+	for i := 0; i+1 < len(mid); i += 2 {
+		mid[i], mid[i+1] = byte(1000&0xFF), byte(1000>>8) // ≈ -30 dBFS
+	}
+	if lvl := ChunkLevel(mid); lvl < 0.4 || lvl > 0.75 {
+		t.Errorf("normal-speech level = %f, want a visible mid-range value", lvl)
+	}
+	// A quiet signal (~-45 dBFS) should be low but non-zero.
+	quiet := make([]byte, 640)
+	for i := 0; i+1 < len(quiet); i += 2 {
+		quiet[i], quiet[i+1] = byte(180&0xFF), byte(180>>8)
+	}
+	if lvl := ChunkLevel(quiet); lvl <= 0 || lvl > 0.35 {
+		t.Errorf("quiet level = %f, want a small non-zero value", lvl)
+	}
+}

internal/dictation/runner.go

@@ -0,0 +1,128 @@
+package dictation
+
+import (
+	"bytes"
+	"context"
+	"io"
+	"os"
+	"os/exec"
+	"time"
+)
+
+// commandSpec describes one capture-process invocation. Argv is always
+// discrete — never a shell string (the same invariant documented in
+// internal/imageinput/clipboard.go).
+type commandSpec struct {
+	name string
+	args []string
+	// wantStdout wires up a stdout pipe (streaming capture reads raw PCM).
+	wantStdout bool
+	// stopViaStdin, when non-empty, makes graceful stop write this to the
+	// process's stdin instead of signaling — ffmpeg finalizes cleanly on "q",
+	// and Windows has no os.Interrupt delivery anyway.
+	stopViaStdin string
+}
+
+// processHandle is a started capture process. Injectable via processStarter so
+// unit tests fake OS-boundary behavior, mirroring internal/lsp's serverStarter
+// seam (this package's clipboard.go-style shell-outs get the test seam that
+// file lacks).
+type processHandle interface {
+	// StopGracefully asks the tool to finish and flush output (SIGINT, or the
+	// spec's stdin stop text). The recorder falls back to Kill on timeout.
+	StopGracefully() error
+	Wait() error
+	Kill() error
+}
+
+// processStarter launches a capture process, returning its handle and — when
+// the spec asks for it — its stdout pipe.
+type processStarter func(spec commandSpec) (processHandle, io.ReadCloser, error)
+
+// commandOutputRunner runs a short one-shot command to completion and returns
+// its combined output. Used for helper invocations that aren't recordings
+// (Termux start/stop, Windows ffmpeg device listing) and the offline
+// transcription exec — ctx lets a caller cancel/timeout a wedged process.
+type commandOutputRunner func(ctx context.Context, name string, args ...string) ([]byte, error)
+
+func startProcess(spec commandSpec) (processHandle, io.ReadCloser, error) {
+	cmd := exec.Command(spec.name, spec.args...)
+	var stdout io.ReadCloser
+	if spec.wantStdout {
+		pipe, err := cmd.StdoutPipe()
+		if err != nil {
+			return nil, nil, err
+		}
+		stdout = pipe
+	}
+	var stdin io.WriteCloser
+	if spec.stopViaStdin != "" {
+		pipe, err := cmd.StdinPipe()
+		if err != nil {
+			return nil, nil, err
+		}
+		stdin = pipe
+	}
+	if err := cmd.Start(); err != nil {
+		// StdoutPipe/StdinPipe allocate fds the runtime only reclaims via
+		// Wait; on Start failure the process is never established, so close
+		// any pipes we already opened to avoid leaking them.
+		if stdout != nil {
+			_ = stdout.Close()
+		}
+		if stdin != nil {
+			_ = stdin.Close()
+		}
+		return nil, nil, err
+	}
+	return &realProcess{cmd: cmd, stdin: stdin, stopText: spec.stopViaStdin}, stdout, nil
+}
+
+type realProcess struct {
+	cmd      *exec.Cmd
+	stdin    io.WriteCloser
+	stopText string
+}
+
+func (p *realProcess) StopGracefully() error {
+	if p.stopText != "" && p.stdin != nil {
+		_, err := io.WriteString(p.stdin, p.stopText)
+		closeErr := p.stdin.Close()
+		if err != nil {
+			return err
+		}
+		return closeErr
+	}
+	// os.Interrupt is how arecord/sox finalize their WAV header. Unsupported
+	// on Windows, but the only Windows tool (ffmpeg) stops via stdin above.
+	return p.cmd.Process.Signal(os.Interrupt)
+}
+
+func (p *realProcess) Wait() error { return p.cmd.Wait() }
+
+func (p *realProcess) Kill() error { return p.cmd.Process.Kill() }
+
+func runCommandOutput(ctx context.Context, name string, args ...string) ([]byte, error) {
+	cmd := exec.CommandContext(ctx, name, args...)
+	var out bytes.Buffer
+	cmd.Stdout = &out
+	cmd.Stderr = &out
+	err := cmd.Run()
+	return out.Bytes(), err
+}
+
+// waitWithTimeout waits for the process, killing it if it ignores the graceful
+// stop for longer than the grace window. Capture tools exit on SIGINT/"q"
+// almost instantly; a hang here means a wedged device driver, and a leaked
+// recording process holds the microphone open.
+func waitWithTimeout(proc processHandle, grace time.Duration) error {
+	done := make(chan error, 1)
+	go func() { done <- proc.Wait() }()
+	select {
+	case err := <-done:
+		return err
+	case <-time.After(grace):
+		_ = proc.Kill()
+		return <-done
+	}
+}

internal/dictation/server.go

@@ -0,0 +1,255 @@
+package dictation
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"net"
+	"os"
+	"os/exec"
+	"strconv"
+	"strings"
+	"sync"
+	"time"
+)
+
+// DefaultServerPort is the localhost port the sherpa-onnx streaming server binds.
+const DefaultServerPort = 6006
+
+// ServerConfig configures the warm sherpa-onnx streaming server (§6a).
+type ServerConfig struct {
+	// Binary is the streaming server executable (default
+	// "sherpa-onnx-online-websocket-server"); looked up on PATH.
+	Binary string
+	// ModelPath is the sherpa-onnx streaming model directory (transducer:
+	// encoder/decoder/joiner + tokens).
+	ModelPath  string
+	Port       int
+	NumThreads int
+
+	// Test seams.
+	starter     processStarter
+	healthCheck func(ctx context.Context, port int) error
+	aliveCheck  func(port int) bool
+	now         func() time.Time
+}
+
+// ServerManager owns one long-lived sherpa-onnx streaming server, spawned lazily
+// on first streaming use and kept warm for the session — a websocket server with
+// ~1-2s startup latency can't be respawned per utterance (§6a). Its lifecycle
+// mirrors internal/lsp's Manager: lazy start, health-checked reuse, one restart
+// on a crashed process, torn down on exit. Safe for concurrent use.
+type ServerManager struct {
+	cfg ServerConfig
+
+	mu   sync.Mutex
+	proc processHandle
+	url  string
+}
+
+// NewServerManager builds a manager. Construction is side-effect free; the
+// server is only spawned on the first EnsureRunning.
+func NewServerManager(cfg ServerConfig) *ServerManager {
+	if cfg.Binary == "" {
+		cfg.Binary = "sherpa-onnx-online-websocket-server"
+	}
+	if cfg.Port == 0 {
+		cfg.Port = DefaultServerPort
+	}
+	if cfg.starter == nil {
+		cfg.starter = startProcess
+	}
+	if cfg.healthCheck == nil {
+		cfg.healthCheck = dialHealthCheck
+	}
+	if cfg.aliveCheck == nil {
+		cfg.aliveCheck = func(port int) bool { return portListening(port) }
+	}
+	if cfg.now == nil {
+		cfg.now = time.Now
+	}
+	return &ServerManager{cfg: cfg}
+}
+
+// EnsureRunning returns the websocket URL of a live server, starting one if
+// needed. A previously-started server whose process has since died is reaped and
+// replaced (mirroring the LSP manager's crashed-session eviction).
+func (m *ServerManager) EnsureRunning(ctx context.Context) (string, error) {
+	m.mu.Lock()
+	defer m.mu.Unlock()
+
+	if m.proc != nil {
+		if m.processAlive() {
+			return m.url, nil
+		}
+		// The cached server died; reap it so a fresh one starts below.
+		_ = m.proc.Kill()
+		m.proc = nil
+		m.url = ""
+	}
+
+	spec, err := onlineServerCommand(m.cfg)
+	if err != nil {
+		return "", err
+	}
+	proc, _, err := m.cfg.starter(spec)
+	if err != nil {
+		if errors.Is(err, exec.ErrNotFound) {
+			return "", &SetupError{
+				Tool: m.cfg.Binary,
+				Hint: "install sherpa-onnx and put its online-websocket-server binary on PATH (see docs/dictation.md)",
+			}
+		}
+		return "", fmt.Errorf("starting sherpa-onnx server: %w", err)
+	}
+
+	// Health-check the port before handing back a URL — the server takes ~1-2s to
+	// bind, and a client connecting too early would fail (§6a).
+	if err := m.cfg.healthCheck(ctx, m.cfg.Port); err != nil {
+		_ = proc.Kill()
+		return "", fmt.Errorf("sherpa-onnx server did not become ready: %w", err)
+	}
+	m.proc = proc
+	m.url = fmt.Sprintf("ws://127.0.0.1:%d", m.cfg.Port)
+	return m.url, nil
+}
+
+// processAlive reports whether the tracked server is still reachable. Probing
+// the port (rather than the process handle) reflects real reachability and is
+// the seam tests replace.
+func (m *ServerManager) processAlive() bool {
+	return m.cfg.aliveCheck(m.cfg.Port)
+}
+
+// portListening reports whether something accepts TCP connections on the port.
+func portListening(port int) bool {
+	conn, err := net.DialTimeout("tcp", net.JoinHostPort("127.0.0.1", strconv.Itoa(port)), 200*time.Millisecond)
+	if err != nil {
+		return false
+	}
+	_ = conn.Close()
+	return true
+}
+
+// Shutdown stops the running server. Called alongside the LSP manager's shutdown
+// on exit — a leaked warm server holds a port and CPU.
+func (m *ServerManager) Shutdown(ctx context.Context) error {
+	m.mu.Lock()
+	proc := m.proc
+	m.proc = nil
+	m.url = ""
+	m.mu.Unlock()
+	if proc == nil {
+		return nil
+	}
+	if err := proc.StopGracefully(); err != nil {
+		_ = proc.Kill()
+	}
+	done := make(chan error, 1)
+	go func() { done <- waitWithTimeout(proc, stopGrace) }()
+	select {
+	case err := <-done:
+		return err
+	case <-ctx.Done():
+		_ = proc.Kill()
+		return ctx.Err()
+	}
+}
+
+// URL returns the current server URL without starting one ("" when not running).
+func (m *ServerManager) URL() string {
+	m.mu.Lock()
+	defer m.mu.Unlock()
+	return m.url
+}
+
+// SetModelPath updates the model directory the server launches with, for a
+// mid-session config change (e.g. an F9 auto-download). It takes effect the next
+// time the server (re)starts; a running server is left as-is.
+func (m *ServerManager) SetModelPath(path string) {
+	m.mu.Lock()
+	defer m.mu.Unlock()
+	if path != "" {
+		m.cfg.ModelPath = path
+	}
+}
+
+// onlineServerCommand builds the streaming-server argv from the model directory.
+// The online server uses a streaming transducer (encoder/decoder/joiner) plus
+// tokens — the shape the sherpa-onnx docs' example uses (§6a).
+func onlineServerCommand(cfg ServerConfig) (commandSpec, error) {
+	if strings.TrimSpace(cfg.ModelPath) == "" {
+		return commandSpec{}, &SetupError{
+			Tool: "local STT streaming model",
+			Hint: "set stt.localModelPath to a sherpa-onnx streaming model directory (see docs/dictation.md)",
+		}
+	}
+	files, err := onnxFilesIn(cfg.ModelPath)
+	if err != nil {
+		return commandSpec{}, &SetupError{Tool: "local STT streaming model", Hint: err.Error()}
+	}
+	tokens := findTokensFile(files)
+	// A streaming model ships encoder/decoder/joiner, often BOTH an int8 and an
+	// fp32 copy of each. Pick a CONSISTENT set — all int8 or all fp32 — because
+	// mixing (int8 encoder + fp32 decoder) is invalid. Prefer int8 (smaller,
+	// faster) when a full int8 set is present.
+	set := consistentQuantizedSet(files, "encoder", "decoder", "joiner")
+	var encoder, decoder, joiner string
+	if set != nil {
+		encoder, decoder, joiner = set[0], set[1], set[2]
+	}
+	if tokens == "" || encoder == "" || decoder == "" || joiner == "" {
+		return commandSpec{}, &SetupError{
+			Tool: "local STT streaming model",
+			Hint: fmt.Sprintf("streaming needs a transducer model (encoder/decoder/joiner .onnx + tokens.txt) in %q; see docs/dictation.md", cfg.ModelPath),
+		}
+	}
+	args := []string{
+		"--port=" + strconv.Itoa(cfg.Port),
+		"--tokens=" + tokens,
+		"--encoder=" + encoder,
+		"--decoder=" + decoder,
+		"--joiner=" + joiner,
+		"--decoding-method=greedy_search",
+	}
+	if cfg.NumThreads > 0 {
+		args = append(args, "--num-threads="+strconv.Itoa(cfg.NumThreads))
+	}
+	return commandSpec{name: cfg.Binary, args: args}, nil
+}
+
+func onnxFilesIn(dir string) (map[string]string, error) {
+	entries, err := os.ReadDir(dir)
+	if err != nil {
+		return nil, fmt.Errorf("stt.localModelPath %q is not readable: %v", dir, err)
+	}
+	files := map[string]string{}
+	for _, e := range entries {
+		if e.IsDir() {
+			continue
+		}
+		files[strings.ToLower(e.Name())] = dir + string(os.PathSeparator) + e.Name()
+	}
+	return files, nil
+}
+
+// dialHealthCheck polls the server's TCP port until it accepts a connection or
+// the deadline passes.
+func dialHealthCheck(ctx context.Context, port int) error {
+	addr := net.JoinHostPort("127.0.0.1", strconv.Itoa(port))
+	deadline := time.Now().Add(5 * time.Second)
+	for {
+		if ctx.Err() != nil {
+			return ctx.Err()
+		}
+		conn, err := net.DialTimeout("tcp", addr, 500*time.Millisecond)
+		if err == nil {
+			_ = conn.Close()
+			return nil
+		}
+		if time.Now().After(deadline) {
+			return fmt.Errorf("port %d not listening after 5s", port)
+		}
+		time.Sleep(100 * time.Millisecond)
+	}
+}

internal/dictation/server_test.go

@@ -0,0 +1,189 @@
+package dictation
+
+import (
+	"context"
+	"encoding/binary"
+	"errors"
+	"io"
+	"math"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"sync"
+	"testing"
+)
+
+func streamingModelDir(t *testing.T) string {
+	t.Helper()
+	dir := t.TempDir()
+	for _, f := range []string{"encoder.onnx", "decoder.onnx", "joiner.onnx", "tokens.txt"} {
+		if err := os.WriteFile(filepath.Join(dir, f), []byte("x"), 0o600); err != nil {
+			t.Fatal(err)
+		}
+	}
+	return dir
+}
+
+func TestServerManagerLazyStartAndReuse(t *testing.T) {
+	var startCount int
+	var mu sync.Mutex
+	starter := func(spec commandSpec) (processHandle, io.ReadCloser, error) {
+		mu.Lock()
+		startCount++
+		mu.Unlock()
+		if spec.name != "sherpa-onnx-online-websocket-server" {
+			t.Errorf("binary = %q", spec.name)
+		}
+		return newFakeProcess(), nil, nil
+	}
+	alive := true
+	m := NewServerManager(ServerConfig{
+		ModelPath:   streamingModelDir(t),
+		Port:        6006,
+		starter:     starter,
+		healthCheck: func(context.Context, int) error { return nil },
+		aliveCheck:  func(int) bool { return alive },
+	})
+
+	url, err := m.EnsureRunning(context.Background())
+	if err != nil {
+		t.Fatalf("EnsureRunning: %v", err)
+	}
+	if url != "ws://127.0.0.1:6006" {
+		t.Errorf("url = %q", url)
+	}
+	// Second call reuses the warm server (alive) — no new start.
+	if _, err := m.EnsureRunning(context.Background()); err != nil {
+		t.Fatalf("EnsureRunning reuse: %v", err)
+	}
+	if startCount != 1 {
+		t.Errorf("expected 1 start (warm reuse), got %d", startCount)
+	}
+}
+
+func TestServerManagerRestartsCrashedServer(t *testing.T) {
+	var startCount int
+	starter := func(commandSpec) (processHandle, io.ReadCloser, error) {
+		startCount++
+		return newFakeProcess(), nil, nil
+	}
+	alive := true
+	m := NewServerManager(ServerConfig{
+		ModelPath:   streamingModelDir(t),
+		starter:     starter,
+		healthCheck: func(context.Context, int) error { return nil },
+		aliveCheck:  func(int) bool { return alive },
+	})
+	if _, err := m.EnsureRunning(context.Background()); err != nil {
+		t.Fatal(err)
+	}
+	alive = false // simulate a crashed server
+	if _, err := m.EnsureRunning(context.Background()); err != nil {
+		t.Fatal(err)
+	}
+	if startCount != 2 {
+		t.Errorf("expected a restart (2 starts), got %d", startCount)
+	}
+}
+
+func TestServerManagerMissingBinaryIsSetupError(t *testing.T) {
+	m := NewServerManager(ServerConfig{
+		ModelPath: streamingModelDir(t),
+		starter:   func(commandSpec) (processHandle, io.ReadCloser, error) { return nil, nil, exec.ErrNotFound },
+	})
+	_, err := m.EnsureRunning(context.Background())
+	var setupErr *SetupError
+	if !errors.As(err, &setupErr) {
+		t.Fatalf("want *SetupError, got %v", err)
+	}
+}
+
+func TestServerManagerMissingModelIsSetupError(t *testing.T) {
+	m := NewServerManager(ServerConfig{
+		ModelPath: t.TempDir(), // empty dir, no onnx/tokens
+		starter:   func(commandSpec) (processHandle, io.ReadCloser, error) { return newFakeProcess(), nil, nil },
+	})
+	_, err := m.EnsureRunning(context.Background())
+	var setupErr *SetupError
+	if !errors.As(err, &setupErr) {
+		t.Fatalf("want *SetupError for missing streaming model, got %v", err)
+	}
+}
+
+func TestServerManagerShutdown(t *testing.T) {
+	proc := newFakeProcess()
+	m := NewServerManager(ServerConfig{
+		ModelPath:   streamingModelDir(t),
+		starter:     func(commandSpec) (processHandle, io.ReadCloser, error) { return proc, nil, nil },
+		healthCheck: func(context.Context, int) error { return nil },
+		aliveCheck:  func(int) bool { return true },
+	})
+	if _, err := m.EnsureRunning(context.Background()); err != nil {
+		t.Fatal(err)
+	}
+	if err := m.Shutdown(context.Background()); err != nil {
+		t.Fatalf("Shutdown: %v", err)
+	}
+	if !proc.stopped && !proc.killed {
+		t.Error("Shutdown should stop the server process")
+	}
+	if m.URL() != "" {
+		t.Error("URL should be empty after shutdown")
+	}
+}
+
+func TestOnlineServerCommandBuildsTransducerArgs(t *testing.T) {
+	spec, err := onlineServerCommand(ServerConfig{Binary: "sherpa-onnx-online-websocket-server", ModelPath: streamingModelDir(t), Port: 6007})
+	if err != nil {
+		t.Fatal(err)
+	}
+	joined := ""
+	for _, a := range spec.args {
+		joined += a + " "
+	}
+	for _, want := range []string{"--port=6007", "--tokens=", "--encoder=", "--decoder=", "--joiner=", "--decoding-method=greedy_search"} {
+		if !contains(joined, want) {
+			t.Errorf("missing %q in args: %s", want, joined)
+		}
+	}
+}
+
+func TestPCM16ToFloat32LE(t *testing.T) {
+	// int16 max (32767) and min (-32768), little-endian.
+	pcm := []byte{0xFF, 0x7F, 0x00, 0x80}
+	out := pcm16ToFloat32LE(pcm)
+	if len(out) != 8 {
+		t.Fatalf("expected 8 bytes (2 float32), got %d", len(out))
+	}
+	f0 := math.Float32frombits(binary.LittleEndian.Uint32(out[0:4]))
+	f1 := math.Float32frombits(binary.LittleEndian.Uint32(out[4:8]))
+	if f0 < 0.99 || f0 > 1.0 {
+		t.Errorf("32767 → %f, want ~1.0", f0)
+	}
+	if f1 != -1.0 {
+		t.Errorf("-32768 → %f, want -1.0", f1)
+	}
+}
+
+func TestParseSherpaStreamResult(t *testing.T) {
+	text, segment, final, ok := parseSherpaStreamResult([]byte(`{"text":" hello world ","segment":2,"is_final":true}`))
+	if !ok || text != "hello world" || segment != 2 || !final {
+		t.Errorf("parsed = (%q, %d, %v, %v)", text, segment, final, ok)
+	}
+	if _, _, _, ok := parseSherpaStreamResult([]byte("not json")); ok {
+		t.Error("non-JSON should not parse")
+	}
+}
+
+func contains(hay, needle string) bool {
+	return len(needle) == 0 || (len(hay) >= len(needle) && indexOf(hay, needle) >= 0)
+}
+
+func indexOf(hay, needle string) int {
+	for i := 0; i+len(needle) <= len(hay); i++ {
+		if hay[i:i+len(needle)] == needle {
+			return i
+		}
+	}
+	return -1
+}

internal/dictation/spill.go

@@ -0,0 +1,51 @@
+package dictation
+
+import (
+	"fmt"
+	"os"
+	"path/filepath"
+)
+
+// Recorded audio is staged in a hardened per-user temp directory, mirroring
+// internal/tools/spill.go (PR #506): uid-suffixed name so users on a shared
+// /tmp cannot collide, 0700 perms, and a pre-existing path is only accepted
+// when it is a real directory (not a symlink that would redirect recordings)
+// owned by the current user. Files are removed as soon as their bytes are
+// read back after Stop — audio never lingers on disk.
+
+func audioSpillRoot() string {
+	name := "zero-dictation"
+	if uid := os.Getuid(); uid >= 0 {
+		name = fmt.Sprintf("zero-dictation-%d", uid)
+	}
+	return filepath.Join(os.TempDir(), name)
+}
+
+func audioSpillDir() (string, error) {
+	dir := audioSpillRoot()
+	if err := os.MkdirAll(dir, 0o700); err != nil {
+		return "", err
+	}
+	// MkdirAll follows symlinks and leaves an existing directory untouched, so
+	// verify what is actually at the path.
+	info, err := os.Lstat(dir)
+	if err != nil {
+		return "", err
+	}
+	if !info.IsDir() {
+		return "", fmt.Errorf("dictation temp path %s is not a directory", dir)
+	}
+	// MkdirAll only applies 0o700 to dirs it creates; an existing directory
+	// (e.g. one created under a loose umask in a prior run) can be group- or
+	// world-readable. Tighten it so staged audio (PII) is never visible to
+	// other users on a shared /tmp.
+	if info.Mode().Perm() != 0o700 {
+		if err := os.Chmod(dir, 0o700); err != nil {
+			return "", fmt.Errorf("tightening dictation temp dir permissions: %w", err)
+		}
+	}
+	if err := checkAudioDirOwner(info); err != nil {
+		return "", err
+	}
+	return dir, nil
+}

internal/dictation/stream_smoke_test.go

@@ -0,0 +1,67 @@
+package dictation
+
+import (
+	"context"
+	"os"
+	"testing"
+	"time"
+)
+
+// TestLocalStreamingRealServer drives the full local streaming path against a
+// REAL sherpa-onnx-online-websocket-server + streaming model: spawn the server,
+// connect, feed a wav's PCM in 50ms chunks, and check the transcript. Opt-in:
+//
+//	ZERO_STT_STREAM_SERVER=/path/to/sherpa-onnx-online-websocket-server \
+//	ZERO_STT_STREAM_MODEL=/path/to/streaming-model-dir \
+//	ZERO_STT_STREAM_WAV=/path/to/16k-mono.wav \
+//	go test ./internal/dictation/ -run RealServer -v -timeout 5m
+func TestLocalStreamingRealServer(t *testing.T) {
+	server := os.Getenv("ZERO_STT_STREAM_SERVER")
+	model := os.Getenv("ZERO_STT_STREAM_MODEL")
+	wav := os.Getenv("ZERO_STT_STREAM_WAV")
+	if server == "" || model == "" || wav == "" {
+		t.Skip("set ZERO_STT_STREAM_SERVER, ZERO_STT_STREAM_MODEL, ZERO_STT_STREAM_WAV to run the streaming smoke test")
+	}
+	raw, err := os.ReadFile(wav)
+	if err != nil {
+		t.Fatal(err)
+	}
+	pcm := stripWavHeader(raw)
+
+	mgr := NewServerManager(ServerConfig{Binary: server, ModelPath: model, Port: 6100})
+	defer mgr.Shutdown(context.Background())
+	tr := NewLocalStreamingTranscriber(mgr)
+
+	chunks := make(chan []byte)
+	go func() {
+		defer close(chunks)
+		size := chunkSizeBytes(DefaultSampleRate) // 50ms
+		for off := 0; off < len(pcm); off += size {
+			end := off + size
+			if end > len(pcm) {
+				end = len(pcm)
+			}
+			chunks <- pcm[off:end]
+			time.Sleep(10 * time.Millisecond) // pace it a little, like a mic
+		}
+	}()
+
+	var partials int
+	final, err := tr.StreamTranscribe(context.Background(), chunks, func(string, bool) { partials++ })
+	if err != nil {
+		t.Fatalf("StreamTranscribe: %v", err)
+	}
+	if final == "" {
+		t.Fatal("streaming returned empty transcript")
+	}
+	t.Logf("streaming transcript (%d partials): %q", partials, final)
+}
+
+// stripWavHeader drops a standard 44-byte PCM WAV header, returning the raw
+// int16 samples. Good enough for the canonical 16kHz/mono/16-bit test wavs.
+func stripWavHeader(b []byte) []byte {
+	if len(b) > 44 && string(b[:4]) == "RIFF" {
+		return b[44:]
+	}
+	return b
+}

internal/dictation/transcriber.go

@@ -0,0 +1,45 @@
+package dictation
+
+import "context"
+
+// Transcriber turns captured audio into text. Batch (Transcribe) and streaming
+// (StreamTranscribe) are one interface but two independent code paths — a
+// batch-only provider (Groq/OpenAI, sherpa-onnx-offline) leaves
+// StreamTranscribe returning ErrStreamingUnsupported, and a streaming provider
+// still implements Transcribe for the batch/Termux fallback.
+type Transcriber interface {
+	// Transcribe returns the final text for one recorded clip.
+	Transcribe(ctx context.Context, audio []byte) (string, error)
+
+	// StreamTranscribe consumes PCM16 chunks and reports incremental results
+	// via onPartial(text, final) as decoding proceeds, returning the complete
+	// final transcript when the chunk channel closes.
+	//
+	// onPartial's text is the CURRENT best transcript of the utterance so far
+	// (cumulative, not a delta) so the composer can replace the live region
+	// wholesale; final marks a segment the provider considers settled.
+	//
+	// On a mid-stream failure (network drop for cloud, subprocess crash for
+	// local) StreamTranscribe returns the best-effort text accumulated from
+	// onPartial calls made so far PLUS a non-nil error — already-streamed text
+	// is never discarded, only the untranscribed tail after the failure is lost.
+	StreamTranscribe(ctx context.Context, chunks <-chan []byte, onPartial func(text string, final bool)) (string, error)
+}
+
+// SampleRate lets a Transcriber declare the capture rate its wire format
+// expects, so the recorder can be built to match (OpenAI Realtime wants 24kHz;
+// everything else uses 16kHz). A Transcriber that does not implement this uses
+// DefaultSampleRate.
+type SampleRate interface {
+	SampleRate() int
+}
+
+// RequiredSampleRate returns the capture rate a Transcriber needs.
+func RequiredSampleRate(t Transcriber) int {
+	if sr, ok := t.(SampleRate); ok {
+		if rate := sr.SampleRate(); rate > 0 {
+			return rate
+		}
+	}
+	return DefaultSampleRate
+}

internal/dictation/transcriber_cloud.go

@@ -0,0 +1,144 @@
+package dictation
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"io"
+	"mime/multipart"
+	"net/http"
+	"strings"
+
+	"github.com/Gitlawb/zero/internal/providers/providerio"
+)
+
+// CloudConfig configures a batch cloud transcriber. Groq and OpenAI share the
+// OpenAI-compatible /audio/transcriptions multipart endpoint — the design's
+// "each owns its endpoint/model specifics while sharing the request shape"
+// (§6). Key and base URL are resolved by the caller (TUI layer) from
+// config/credstore, keeping this type decoupled and unit-testable.
+type CloudConfig struct {
+	Provider string // ProviderGroq or ProviderOpenAI, for error messages
+	BaseURL  string // e.g. https://api.groq.com/openai/v1
+	APIKey   string
+	Model    string // e.g. whisper-large-v3-turbo, whisper-1
+	// Language optionally constrains recognition (ISO-639-1). Empty = auto.
+	Language string
+
+	// HTTPClient is injectable for tests; nil uses providerio's shared,
+	// stall-hardened client.
+	HTTPClient *http.Client
+}
+
+type cloudTranscriber struct {
+	cfg CloudConfig
+}
+
+// NewCloudTranscriber builds a batch transcriber for Groq or OpenAI.
+func NewCloudTranscriber(cfg CloudConfig) (Transcriber, error) {
+	if strings.TrimSpace(cfg.APIKey) == "" {
+		return nil, &SetupError{
+			Tool: cfg.Provider + " API key",
+			Hint: fmt.Sprintf("set a %s API key (run `zero auth`) to use cloud dictation", cfg.Provider),
+		}
+	}
+	if cfg.Model == "" {
+		return nil, fmt.Errorf("dictation: no model configured for provider %q", cfg.Provider)
+	}
+	base, err := providerio.NormalizeBaseURL(cfg.BaseURL, cfg.BaseURL, cfg.Provider)
+	if err != nil {
+		return nil, err
+	}
+	cfg.BaseURL = base
+	return &cloudTranscriber{cfg: cfg}, nil
+}
+
+func (c *cloudTranscriber) Transcribe(ctx context.Context, audio []byte) (string, error) {
+	if len(audio) == 0 {
+		return "", errors.New("no audio to transcribe")
+	}
+	body, contentType, err := c.buildMultipart(audio)
+	if err != nil {
+		return "", err
+	}
+	endpoint := c.cfg.BaseURL + "/audio/transcriptions"
+	req, err := http.NewRequestWithContext(ctx, http.MethodPost, endpoint, bytes.NewReader(body))
+	if err != nil {
+		return "", err
+	}
+	req.Header.Set("Content-Type", contentType)
+	req.Header.Set("Authorization", "Bearer "+c.cfg.APIKey)
+
+	resp, err := providerio.HTTPClient(c.cfg.HTTPClient).Do(req)
+	if err != nil {
+		return "", fmt.Errorf("%s transcription request failed: %w", c.cfg.Provider, err)
+	}
+	defer resp.Body.Close()
+	respBody, _ := io.ReadAll(io.LimitReader(resp.Body, 1<<20))
+	if resp.StatusCode != http.StatusOK {
+		msg := providerio.ClassifiedError(resp.StatusCode, strings.TrimSpace(string(respBody)), c.cfg.APIKey)
+		// A 401/403 is a fixable credential problem: surface it as a typed AuthError
+		// so the TUI can offer an inline key prompt instead of a dead-end message.
+		if resp.StatusCode == http.StatusUnauthorized || resp.StatusCode == http.StatusForbidden {
+			return "", &AuthError{Provider: c.cfg.Provider, Message: msg}
+		}
+		return "", errors.New(msg)
+	}
+	var parsed struct {
+		Text  string `json:"text"`
+		Error struct {
+			Message string `json:"message"`
+		} `json:"error"`
+	}
+	if err := json.Unmarshal(respBody, &parsed); err != nil {
+		return "", fmt.Errorf("%s transcription: could not parse response: %w", c.cfg.Provider, err)
+	}
+	if parsed.Error.Message != "" {
+		return "", fmt.Errorf("%s transcription error: %s", c.cfg.Provider, providerio.Redact(parsed.Error.Message, c.cfg.APIKey))
+	}
+	return strings.TrimSpace(parsed.Text), nil
+}
+
+// buildMultipart assembles the transcription upload. The file field name and
+// the "model"/"response_format" fields match the OpenAI audio API that Groq
+// mirrors. The filename's extension tells the server the container type —
+// SniffFormat keeps WAV/M4A honest so Termux's AAC uploads work.
+func (c *cloudTranscriber) buildMultipart(audio []byte) ([]byte, string, error) {
+	var buf bytes.Buffer
+	w := multipart.NewWriter(&buf)
+	fw, err := w.CreateFormFile("file", SniffFormat(audio).FileName())
+	if err != nil {
+		return nil, "", err
+	}
+	if _, err := fw.Write(audio); err != nil {
+		return nil, "", err
+	}
+	if err := w.WriteField("model", c.cfg.Model); err != nil {
+		return nil, "", err
+	}
+	if err := w.WriteField("response_format", "json"); err != nil {
+		return nil, "", err
+	}
+	if c.cfg.Language != "" {
+		if err := w.WriteField("language", c.cfg.Language); err != nil {
+			return nil, "", err
+		}
+	}
+	if err := w.Close(); err != nil {
+		return nil, "", err
+	}
+	return buf.Bytes(), w.FormDataContentType(), nil
+}
+
+// StreamTranscribe is not supported for the batch cloud providers (Groq has no
+// streaming product; OpenAI batch is one-shot). Streaming uses the dedicated
+// Deepgram/OpenAI-Realtime transcribers instead.
+func (c *cloudTranscriber) StreamTranscribe(context.Context, <-chan []byte, func(string, bool)) (string, error) {
+	return "", ErrStreamingUnsupported
+}
+
+// ErrStreamingUnsupported reports that a batch-only Transcriber cannot stream;
+// the caller falls back to the batch pipeline.
+var ErrStreamingUnsupported = errors.New("this transcription provider does not support streaming")