mixdog 0.9.150 → 0.9.152

package/README.md

@@ -1,20 +1,35 @@
-# mixdog
+# Mixdog
 
 [![npm](https://img.shields.io/npm/v/mixdog)](https://www.npmjs.com/package/mixdog)
 ![node](https://img.shields.io/badge/node-%3E%3D22-brightgreen)
 ![license](https://img.shields.io/badge/license-MIT-blue)
 
-Mixdog is a standalone coding agent for orchestrated, multi-provider workflows.
-Use it from a full-screen terminal UI or the Windows desktop app.
+## Better results. Less cost. More work.
+
+Mixdog is an efficiency-first AI coding harness designed to deliver equal or
+better performance with less context, time, and cost—so you can complete more
+work within the same API budget or subscription quota.
+
+With simple setup and an intuitive UX, Mixdog makes powerful orchestration,
+parallel tasks, and seamless work across terminal, desktop, and web accessible
+to everyone—from beginners to experts.
+
+**The easiest way to get more out of every coding model.**
 
 ## Get started
 
-### Windows desktop
+### Desktop
 
-[Download Mixdog Desktop for Windows (x64)](https://github.com/tribgames/mixdog/releases/latest/download/mixdog-desktop-win-x64.exe)
+| Platform | Download |
+| --- | --- |
+| Windows x64 | [Installer](https://github.com/tribgames/mixdog/releases/latest/download/mixdog-desktop-win-x64.exe) |
+| macOS Apple silicon | [DMG](https://github.com/tribgames/mixdog/releases/latest/download/mixdog-desktop-mac-arm64.dmg) |
+| macOS Intel | [DMG](https://github.com/tribgames/mixdog/releases/latest/download/mixdog-desktop-mac-x64.dmg) |
+| Linux x86_64 | [AppImage](https://github.com/tribgames/mixdog/releases/latest/download/mixdog-desktop-linux-x86_64.AppImage) |
+| Linux arm64 | [AppImage](https://github.com/tribgames/mixdog/releases/latest/download/mixdog-desktop-linux-arm64.AppImage) |
 
-The desktop package is currently unsigned, so Windows SmartScreen may show a
-security warning during installation.
+Desktop packages are currently unsigned, so Windows SmartScreen or macOS
+Gatekeeper may show a security warning.
 
 ### CLI
 
@@ -28,28 +43,68 @@ mixdog
 First run guides you through provider authentication, model selection, and
 workflow setup.
 
+## Benchmarks
+
+Terminal-Bench 2.1 — same model, same 89 tasks, same official verifier, with
+only the harness changed. Against the native CLI of each model family, Mixdog
+scores higher while spending less to get there.
+
+### GPT-5.6 Sol xhigh — Mixdog vs Codex CLI
+
+![Terminal-Bench 2.1: Mixdog with GPT-5.6 Sol xhigh versus Codex CLI](https://raw.githubusercontent.com/tribgames/mixdog/main/benchmarks/terminal-bench-2.1/tb21-sol-vs-codex.svg)
+
+- **86.5%** (385/445) vs Codex CLI's **84.3%** (75/89)
+- **42%** lower priced cost — $0.641 vs $1.096 per trial
+- **45%** smaller median final context — 18.5k vs 33.5k tokens
+- **1.11×** faster — 339s vs 378s per trial
+
+### Claude Opus 5 — Mixdog vs Claude Code
+
+![Terminal-Bench 2.1: Mixdog with Claude Opus 5 versus Claude Code](https://raw.githubusercontent.com/tribgames/mixdog/main/benchmarks/terminal-bench-2.1/tb21-opus-vs-claude-code.svg)
+
+- **79/89** vs Claude Code's **77/89**
+- **19%** lower priced cost — $104.29 vs $129.21 per run
+- **28%** smaller median final context — 27.6k vs 38.2k tokens
+- **1.15×** faster
+
+Every run uses the official Harbor verifier, fast mode off, a 272k context
+window, and zero retries. The Mixdog Sol run follows the protocol the official
+Terminal-Bench leaderboard requires — all 89 tasks repeated five times (`k=5`,
+445 trials); the Codex CLI baseline and both Opus-side runs are single passes
+(`k=1`, 89 trials each).
+
+The leaderboard is not accepting community submissions, so every run here ships
+its raw artifacts instead — Harbor verdicts, official verifier output, pinned
+task checksums, and the usage snapshots behind every cost figure — alongside
+the harness, presets, and metric scripts that recompute each number above:
+[`benchmarks/terminal-bench-2.1/`](benchmarks/terminal-bench-2.1/).
+
 ## Highlights
 
 - **Multi-provider routing** — assign different providers and models by role.
+- **Shared live sessions** — move between the TUI, desktop windows, and paired
+  browsers without starting a second copy of the session.
 - **Efficient context** — cache-aware prompts, compaction, resumable sessions,
   and focused repo-native tools.
 - **Complete coding surface** — read, search, edit, test, review, web search,
   MCP, skills, hooks, and plugins.
 - **Local memory** — semantic and lexical recall with project-scoped context
   and multilingual retrieval.
-- **Remote workflows** — optional web relay, Discord, Telegram, voice, and cron
-  schedules.
-- **Windows desktop app** — agent panes, Monaco editor, git, terminal, file
-  explorer, Studio, automation, and settings in one workbench.
+- **Encrypted remote access** — pair the installable web app with Desktop and
+  use Mixdog from a browser or phone over authenticated E2EE.
+- **Desktop coding app** — agent panes, Monaco editor, Git, terminals, file
+  explorer, Studio, automation, voice input, and settings in one app.
 
 ## Providers
 
 Mixdog supports subscription OAuth and API-key routes, including:
 
-- Anthropic and Claude accounts
-- OpenAI and ChatGPT/Codex accounts
-- Google Gemini
-- xAI Grok
+- Anthropic API keys and Claude account OAuth
+- OpenAI API keys and ChatGPT/Codex account OAuth
+- Google Gemini and Antigravity OAuth
+- xAI API keys and Grok account OAuth
+- OpenRouter API keys and its unified model catalog
+- Experimental Cursor account OAuth
 - DeepSeek and OpenCode Go
 - OpenAI-compatible APIs
 - Ollama and LM Studio
@@ -107,40 +162,57 @@ stdout; diagnostics remain on stderr.
 ## TUI commands
 
 ```text
-/providers   configure provider authentication and local endpoints
-/model       choose the main provider and model
-/workflow    choose the active workflow
-/agents      inspect agents and model overrides
-/project     switch the current project
-/resume      resume a saved session
-/memory      inspect and edit core memory
-/mcp         manage MCP servers and tools
-/skills      select a skill
-/channels    manage remote channels
-/compact     compact older context
-/setting     open settings
-/update      check for updates
-/doctor      diagnose installation health
+/clear        start a fresh chat
+/project      switch the current project
+/resume       resume a saved chat
+/compact      compact older conversation context
+/autoclear    manage idle-time context clearing
+/context      inspect the current context surface
+/usage        show provider quota and balance
+/providers    configure authentication and local endpoints
+/model        choose the main provider and model
+/websearch    choose the web search route
+/workflow     choose the active workflow
+/agents       inspect agents and model overrides
+/effort       set reasoning effort
+/fast         toggle supported model fast mode
+/OutputStyle  choose the Lead response style
+/theme        change the TUI color theme
+/memory       inspect and edit core memory
+/mcp          manage MCP servers and tools
+/skills       choose a skill for the next request
+/plugins      manage local plugin integrations
+/hooks        manage before-tool hooks and events
+/setting      open runtime settings
+/profile      set your title and response language
+/update       check for updates
+/doctor       diagnose installation health
+/quit         quit the TUI
 ```
 
 Workflows and agents are Markdown definition packs (`WORKFLOW.md`, `AGENT.md`).
 Built-in packs ship with Mixdog; custom packs live under the Mixdog data
 directory.
 
-## Windows desktop app
+## Desktop app
 
-Mixdog Desktop runs the same agent runtime as the CLI in an Electron
-workbench:
+Mixdog Desktop runs the same agent runtime as the CLI:
 
 - Split panes for parallel, independently routed agent sessions
+- Live session handoff between the TUI, desktop windows, and paired browsers
 - Monaco editor, LSP integration, diffs, and turn-by-turn edit review
 - Git staging, commits, branches, and generated commit messages
-- Windows file explorer with previews, thumbnails, and drag-and-drop
-- Integrated PowerShell and ConPTY terminal tabs
+- File explorer with previews, thumbnails, search, and drag-and-drop
+- Integrated terminal tabs using the local system shell
 - Image and video generation Studio with a persistent local gallery
-- Visual workflow, agent, and schedule editors
+- Visual workflow, agent, schedule, and webhook editors
+- Voice dictation with an optional local transcription runtime
 - Provider setup, usage, git identity, and remote pairing settings
 
+The paired remote web app is installable on desktop and mobile browsers. It
+uses an authenticated end-to-end encrypted connection before session state,
+terminal data, files, or operation requests cross the relay.
+
 For desktop development:
 
 ```bash
@@ -148,20 +220,6 @@ cd apps/desktop
 npm run dev
 ```
 
-## Terminal-Bench 2.1
-
-Controlled single-model runs on the same 89 tasks produced:
-
-- **82/89** with Claude Opus 5 vs Claude Code's **77/89**
-- **79/89** with GPT-5.6 Sol xhigh vs Codex CLI's **75/89**
-- **1.21×** faster vs Claude Code and **1.15×** faster vs Codex CLI
-- **31–47%** smaller median final context
-- **16%** lower priced cost vs Claude Code and **41%** lower vs Codex CLI
-
-These are self-reported single runs (`k=1`, 2026-08-23), not leaderboard
-submissions. Raw artifacts, commands, comparison charts, and metric scripts
-live under [`benchmarks/terminal-bench-2.1/`](benchmarks/terminal-bench-2.1/).
-
 ## Data and configuration
 
 Mixdog uses `~/.mixdog` as its home root and `~/.mixdog/data` for runtime data
@@ -196,10 +254,11 @@ Main directories:
 
 ```text
 src/            CLI, TUI, runtime, workflows, agents, and rules
-apps/desktop/   Windows desktop workbench
-apps/relay/     remote web relay
+apps/desktop/   cross-platform desktop app
+apps/relay/     remote web app and relay
 native/         native process, search, patch, and support binaries
 scripts/        tests, diagnostics, benchmarks, and build scripts
+benchmarks/     reproducible benchmark harnesses, results, and raw artifacts
 vendor/         vendored runtime components
 ```
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "mixdog",
-  "version": "0.9.150",
+  "version": "0.9.152",
   "private": false,
   "type": "module",
   "description": "Standalone mixdog coding-agent CLI/TUI workspace.",
@@ -66,7 +66,7 @@
     "test:native-edit-wire": "node scripts/native-edit-wire-test.mjs",
     "test:release-critical": "npm run test:release-assets && npm run smoke:patch && npm run test:providers",
     "test:spec-advisory": "node scripts/advisory-spec-test.mjs && npm run test:spec-advisory --prefix apps/desktop",
-    "test:session-transport": "node --test scripts/session-transport-test.mjs scripts/daemon-bootstrap-test.mjs",
+    "test:session-transport": "node --test scripts/session-transport-test.mjs scripts/daemon-bootstrap-test.mjs src/runtime/shared/child-spawn-remote.test.mjs src/standalone/session-runtime-agent-control-client.test.mjs src/standalone/session-runtime-dispatch-cancel.test.mjs src/standalone/session-runtime-host-factory.test.mjs src/standalone/session-runtime-inline-host.test.mjs src/standalone/session-runtime-provider-cooldown.test.mjs src/standalone/session-runtime-shard-host.test.mjs src/standalone/session-runtime-shard-router.test.mjs",
     "test:session": "node --test scripts/runtime-turn-contract-test.mjs scripts/session-save-fault-store-test.mjs scripts/turn-review-revert-test.mjs src/runtime/shared/tool-surface.test.mjs",
     "test:media": "node --test src/runtime/media/store.test.mjs src/runtime/media/renditions.test.mjs src/runtime/media/adapters/codex-image.test.mjs",
     "failures": "node scripts/tool-failures.mjs",
@@ -139,6 +139,7 @@
     "undici": "^8.5.0",
     "wrap-ansi": "^10.0.0",
     "ws": "^8.21.0",
+    "yaml": "^2.9.0",
     "zod": "^3.25.76"
   },
   "overrides": {

package/scripts/tool-stress.mjs

@@ -17,7 +17,6 @@ import { executePatchTool } from '../src/runtime/agent/orchestrator/tools/patch.
 import { normalizeToolEnvelope } from '../src/runtime/agent/orchestrator/session/tool-envelope.mjs';
 import { warmNativeSpawnServer } from '../src/runtime/agent/orchestrator/tools/lib/native-spawn-client.mjs';
 import { warmNativeSearchServer } from '../src/runtime/agent/orchestrator/tools/builtin/native-search-client.mjs';
-import { prewarmFindEnumeration } from '../src/runtime/agent/orchestrator/tools/builtin/list-tool.mjs';
 
 if (!process.argv.includes('--unsafe-live')) {
   console.error('Refusing high-impact tool stress without --unsafe-live; run `node scripts/tool-search-bench.mjs` for safe exploration diagnostics.');
@@ -66,7 +65,6 @@ try {
   await Promise.all([
     warmNativeSpawnServer(),
     warmNativeSearchServer(),
-    prewarmFindEnumeration(root),
   ]);
   prewarmMs = Date.now() - prewarmStarted;
   // ── Phase A+C: concurrent multi-session waves (search/read/graph/shell +

package/src/defaults/skills/setup/SKILL.md

@@ -15,7 +15,7 @@ description: Use this skill only to inspect or modify a Mixdog user's persisted
 
 - `<mixdogData>` 해석 순서: `MIXDOG_DATA_DIR` → `<MIXDOG_HOME|~/.mixdog>/data` (`src/runtime/shared/plugin-paths.mjs`)
 - 통합 config: `<mixdogData>/mixdog-config.json`
-- 프로젝트 config: `<cwd>/.mcp.json`, `<cwd>/.mixdog/skills/`, `<cwd>/.mixdog/hooks.json`
+- 프로젝트 config: `<cwd>/.mixdog/hooks.json`
 - TUI 명령 목록: `src/tui/app/slash-commands.mjs`
 - TUI 설정 허브: `/setting` (별칭 `/settings`, `/config`) → `src/tui/app/settings-picker.mjs`
 - Desktop 설정: `Ctrl+,` → General / Context / Output style / Providers / Git / Skills / MCP / Plugins / Hooks / Connection / System / Shortcuts / About
@@ -111,7 +111,8 @@ API 키·토큰·OAuth 자격은 config에 쓰지 않는다.
 1. **확인**: Desktop **Settings → General → Web search**, TUI `/setting → Web search`.
 2. **변경**: toggle → `setWebSearchEnabled()`.
 3. **저장**: `agent.modules.webSearch.enabled`.
-4. **검증**: 다음 세션의 tool surface.
+4. **검색 범위**: 전역 `<mixdogData>/skills/`와 활성 Plugin skill만 사용한다. Project `.mixdog/skills/`는 읽지 않는다.
+5. **검증**: 열린 runtime과 다음 세션의 tool surface.
 
 ### reasoning effort / Fast
 
@@ -263,18 +264,15 @@ TUI·Desktop interval 편집 UI는 없다.
 
 ### MCP 추가
 
-UI에는 서버 추가·삭제·전체 재연결 액션이 없다.
-
-1. **확인**: `/mcp` 또는 Desktop **Settings → MCP**, `mcpStatus()`, `<cwd>/.mcp.json`.
+1. **확인**: `/mcp` 또는 Desktop **Settings → MCP**, `mcpStatus()`, `agent.mcpServers`.
 2. **변경**:
-   - 전역: 디스크 `agent.mcpServers.<name>` 편집
-   - 프로젝트: `<cwd>/.mcp.json`의 `mcpServers` wrapper 또는 bare name map 편집
+   - Desktop: **Settings → MCP → Add**
+   - 디스크: `agent.mcpServers.<name>` 편집
    - API 자동화: `addMcpServer()`, `removeMcpServer()`, `reconnectMcp()`
 3. **transport**:
    - stdio: `type`, `command`, `args`, `cwd`, `env`
    - URL: `http`, `sse`, `ws`와 `url`, 선택 `headers`
-4. **제약**: API로 stdio를 추가할 때 `cwd`는 현재 프로젝트 아래여야 한다.
-5. **검증**: `connected`, `toolCount`, `source`, `transport`, `error`.
+4. **검증**: `connected`, `toolCount`, `transport`, `error`.
 
 파일을 직접 편집했다면 mixdog 재시작이 기본 적용 경로다.
 
@@ -282,17 +280,15 @@ UI에는 서버 추가·삭제·전체 재연결 액션이 없다.
 
 TUI `/mcp`와 Desktop MCP 모두 서버별 toggle을 지원하며 live connection과 세션 tool surface를 재동기화한다.
 
-- `source: project`: `<cwd>/.mcp.json` 항목의 `enabled`를 직접 저장
-- `source: config`: 서버 정의는 유지하고 현재 프로젝트용 override를 `agent.mcpProjectOverrides[normalizedCwd][name].enabled`에 저장
-- 즉, 전역 config 서버를 toggle해도 전역 `agent.mcpServers.<name>.enabled`를 바꾸지 않는다.
+- 서버 정의는 유지하고 `agent.mcpServers.<name>.enabled`를 전역으로 변경한다.
+- Project `.mcp.json`과 Project별 override는 읽지 않는다.
 - turn 실행 중 toggle은 turn 종료 경계에서 세션을 재생성한다.
 
 ### MCP 진단
 
-1. `/mcp`의 source/transport/error를 먼저 읽는다.
-2. 프로젝트와 전역 이름 충돌 시 `<cwd>/.mcp.json`이 우선한다.
-3. stdio는 command·args·cwd·자식 env, URL transport는 scheme·endpoint·headers·방화벽을 확인한다.
-4. `connected:true`와 기대 tool 노출로 검증한다.
+1. `/mcp`의 transport/error를 먼저 읽는다.
+2. stdio는 command·args·cwd·자식 env, URL transport는 scheme·endpoint·headers·방화벽을 확인한다.
+3. `connected:true`와 기대 tool 노출로 검증한다.
 
 ### Skills enable / disable
 
@@ -317,11 +313,11 @@ UI 생성 액션은 없다.
 
 1. **확인**: `/plugins` 또는 Desktop **Settings → Plugins**.
 2. **추가**: Git URL, `owner/repo`, 기존 local path → `addPlugin()`.
-3. **관리**: update/metadata refresh, plugin MCP enable/reconfigure, root/MCP name 복사, uninstall.
+3. **관리**: 전역 enable/disable, update/metadata refresh, plugin MCP enable/reconfigure, root/MCP name 복사, uninstall.
 4. **저장**: `<mixdogData>/plugins/registry.json`; managed Git checkout은 `<mixdogData>/plugins/installed/`.
 5. **검증**: `pluginsStatus()`, plugin skill 수, MCP server 노출.
 
-Plugins 화면에는 일반적인 “plugin 활성/비활성” toggle이 없다. MCP와 skill 활성 상태는 각각 MCP/Skills 화면에서 관리한다.
+Plugin toggle은 해당 Plugin의 Skills와 MCP를 전역으로 함께 활성화/비활성화한다. 개별 MCP와 Skill은 각 화면에서도 관리할 수 있다.
 
 ### Hooks
 
@@ -380,8 +376,8 @@ Discord/Telegram messaging은 제거됐다. schedules/webhooks 자동화와 voic
 
 1. **확인**: 상태줄 cwd, `/project`, Desktop rail **Projects**.
 2. **변경**: `/project [path]` 또는 picker. TUI picker는 등록·생성·rename도 지원한다.
-3. **적용**: cwd 변경 시 프로젝트 `.mcp.json`과 skills를 다시 읽고 MCP를 재연결한다.
-4. **검증**: cwd, `/mcp` source, `/skills` project 항목.
+3. **적용**: cwd 변경은 실행 경로만 바꾸며 전역 Skills·Plugins·MCP 상태는 유지한다.
+4. **검증**: cwd와 실행 경로.
 
 ### Git (Desktop 전용)
 
@@ -438,11 +434,6 @@ Desktop과 현재 runtime에서 소비되지 않는 아래 key는 사용자 옵
         "cwd": "<project-subdir>",
         "env": {}
       }
-    },
-    "mcpProjectOverrides": {
-      "<normalized-cwd>": {
-        "<name>": { "enabled": false }
-      }
     }
   }
 }
@@ -466,8 +457,8 @@ URL transport는 `type` + `url` + 선택 `headers`를 사용한다.
 
 ### 우선순위와 금지
 
-- MCP 이름 충돌: project `.mcp.json` > `agent.mcpServers`
-- Skill 이름 충돌: project > global > plugin
+- MCP는 전역 `agent.mcpServers`만 사용한다.
+- Skill 이름 충돌: global > plugin
 - `Mixdog.md` 자동 프롬프트 로드는 없다. skill/core memory를 사용한다.
 - 확인되지 않은 key는 추측하지 말고 TODO로 남긴다.
 - UI에 없는 runtime API를 사용자 UI처럼 설명하지 않는다.

package/src/headless-exec.mjs

@@ -14,6 +14,7 @@ import {
 import { hasActiveBackgroundTasks } from './runtime/shared/background-tasks.mjs';
 import { installProcessSignalCleanup } from './runtime/shared/process-shutdown.mjs';
 import { stopStandaloneMemoryRuntimesForProcess } from './standalone/memory-runtime-proxy.mjs';
+import { shutdownDaemonForRuntimeRoot } from './standalone/session-client.mjs';
 import { applyUsageDelta, createSessionStats } from './ui/session-stats.mjs';
 
 function clean(value) {
@@ -22,17 +23,10 @@ function clean(value) {
 
 export async function prewarmHeadlessSearch(cwd, {
   loadNativeSearch = () => import('./runtime/agent/orchestrator/tools/builtin/native-search-client.mjs'),
-  loadListTool = () => import('./runtime/agent/orchestrator/tools/builtin/list-tool.mjs'),
 } = {}) {
-  const root = clean(cwd) || process.cwd();
-  const [nativeSearch, listTool] = await Promise.all([
-    loadNativeSearch(),
-    loadListTool(),
-  ]);
-  await Promise.all([
-    nativeSearch.warmNativeSearchServer(),
-    listTool.prewarmFindEnumeration(root),
-  ]);
+  void cwd;
+  const nativeSearch = await loadNativeSearch();
+  await nativeSearch.warmNativeSearchServer();
 }
 
 function nonNegativeNumber(value) {
@@ -558,6 +552,7 @@ export async function runHeadlessExec({
   boundaryFactory = createPristineExecutionBoundary,
   runtimeFactory = null,
   memoryRuntimeCleanup = stopStandaloneMemoryRuntimesForProcess,
+  daemonRuntimeCleanup = shutdownDaemonForRuntimeRoot,
   hasActiveTasks = hasActiveBackgroundTasks,
   installSignalCleanupFn = installProcessSignalCleanup,
 } = {}) {
@@ -616,17 +611,28 @@ export async function runHeadlessExec({
       } catch (error) {
         errors.push(error);
       }
-      let memoryCleanupFailed = false;
+      let resourceCleanupFailed = false;
+      if (boundary?.runtimeRoot) {
+        try {
+          await daemonRuntimeCleanup(boundary.runtimeRoot, {
+            waitForExit: true,
+            timeoutMs: 8_000,
+          });
+        } catch (error) {
+          resourceCleanupFailed = true;
+          errors.push(error);
+        }
+      }
       if (boundary) {
         try {
           await memoryRuntimeCleanup({ waitForExit: true, timeoutMs: 10_000 });
         } catch (error) {
-          memoryCleanupFailed = true;
+          resourceCleanupFailed = true;
           errors.push(error);
         }
       }
       try {
-        const cleanupResult = boundary?.cleanup(memoryCleanupFailed
+        const cleanupResult = boundary?.cleanup(resourceCleanupFailed
           ? { preserveRoot: true }
           : { tolerateRootRemovalFailure: true });
         if (cleanupResult?.rootRemovalError) {

package/src/headless-exec.test.mjs

@@ -9,7 +9,7 @@ import { prewarmHeadlessSearch, runHeadlessExec } from './headless-exec.mjs';
 import { resolveCursorOAuthAccessToken } from './runtime/agent/orchestrator/providers/cursor-auth.mjs';
 import { createPristineExecutionBoundary } from './runtime/shared/pristine-execution.mjs';
 
-test('headless search prewarm starts the native server and Project inventory together', async () => {
+test('headless search prewarm starts only the canonical native server', async () => {
   const calls = [];
   await prewarmHeadlessSearch('/app/project', {
     loadNativeSearch: async () => ({
@@ -18,13 +18,8 @@ test('headless search prewarm starts the native server and Project inventory tog
         return true;
       },
     }),
-    loadListTool: async () => ({
-      async prewarmFindEnumeration(root) {
-        calls.push(`index:${root}`);
-      },
-    }),
   });
-  assert.deepEqual(calls.sort(), ['index:/app/project', 'server']);
+  assert.deepEqual(calls, ['server']);
 });
 
 test('pristine headless execution binds Cursor OAuth credentials in process', async () => {
@@ -62,6 +57,7 @@ test('headless exec runs one implicit-approval session and waits for tracked tas
   let activeChecks = 0;
   let boundaryCleaned = false;
   let runtimeClosed = false;
+  const daemonCleanupCalls = [];
   const cleanupOrder = [];
   try {
     const code = await runHeadlessExec({
@@ -75,6 +71,7 @@ test('headless exec runs one implicit-approval session and waits for tracked tas
       write: (text) => output.push(text),
       writeErr: (text) => errors.push(text),
       boundaryFactory: () => ({
+        runtimeRoot: join(root, 'runtime-root'),
         loadConfig: () => ({ providers: { 'openai-oauth': { enabled: true } } }),
         cleanup: () => {
           boundaryCleaned = true;
@@ -106,6 +103,10 @@ test('headless exec runs one implicit-approval session and waits for tracked tas
       memoryRuntimeCleanup: async () => {
         cleanupOrder.push('memory');
       },
+      daemonRuntimeCleanup: async (runtimeRoot, options) => {
+        daemonCleanupCalls.push({ runtimeRoot, options });
+        cleanupOrder.push('daemon');
+      },
       hasActiveTasks: (scope) => {
         activeScopes.push(scope);
         activeChecks += 1;
@@ -128,7 +129,11 @@ test('headless exec runs one implicit-approval session and waits for tracked tas
     });
     assert.equal(boundaryCleaned, true);
     assert.equal(runtimeClosed, true);
-    assert.deepEqual(cleanupOrder, ['runtime', 'memory', 'boundary']);
+    assert.deepEqual(cleanupOrder, ['runtime', 'daemon', 'memory', 'boundary']);
+    assert.deepEqual(daemonCleanupCalls, [{
+      runtimeRoot: join(root, 'runtime-root'),
+      options: { waitForExit: true, timeoutMs: 8_000 },
+    }]);
     const usage = JSON.parse(readFileSync(usageLogPath, 'utf8'));
     assert.deepEqual(usage.sessions[0].models, ['gpt-test', 'gpt-fallback']);
     assert.deepEqual(usage.totals, {
@@ -143,6 +148,39 @@ test('headless exec runs one implicit-approval session and waits for tracked tas
   }
 });
 
+test('headless exec preserves the pristine root when isolated daemon shutdown fails', async () => {
+  const errors = [];
+  let cleanupOptions = null;
+  const code = await runHeadlessExec({
+    message: 'done',
+    provider: 'openai-oauth',
+    model: 'gpt-test',
+    usageLogPath: '',
+    write() {},
+    writeErr: (text) => errors.push(text),
+    boundaryFactory: () => ({
+      runtimeRoot: '/isolated/runtime',
+      loadConfig: () => ({ providers: { 'openai-oauth': { enabled: true } } }),
+      cleanup: (options) => { cleanupOptions = options; },
+    }),
+    runtimeFactory: async () => ({
+      id: 'sess_cleanup_failure',
+      model: 'gpt-test',
+      clientHostPid: 123,
+      async ask() { return { result: { content: 'done' } }; },
+      async close() {},
+    }),
+    daemonRuntimeCleanup: async () => { throw new Error('daemon stuck'); },
+    memoryRuntimeCleanup: async () => {},
+    hasActiveTasks: () => false,
+    installSignalCleanupFn: () => ({ uninstall() {} }),
+  });
+
+  assert.equal(code, 1);
+  assert.deepEqual(cleanupOptions, { preserveRoot: true });
+  assert.ok(errors.some((line) => line.includes('shutdown failed: daemon stuck')));
+});
+
 test('headless exec flushes the usage snapshot mid-session, before any exit path', async () => {
   const root = mkdtempSync(join(tmpdir(), 'mixdog-headless-usage-flush-test-'));
   const usageLogPath = join(root, 'usage.json');

package/src/rules/shared/30-exploration.md

@@ -27,7 +27,8 @@
   locator searches. Within the current project, pass project-relative paths and
   omit optional scopes equal to its root; explicit paths may be outside cwd
   only for targets outside the project.
-- Inspect source content only when its format is required and unknown.
+- Before deciding how to parse, count, transform, or summarize files whose
+  format has not been inspected, inspect the original content itself.
 - Returned declarations, bodies, usages, relations, and contextual spans from
   any tool — not only `read` — are source context; `read` covers only omitted
   lines or missing anchored ranges.