🐛 fix(desktop): spawn the detector-resolved Codex path, not the bare command

Adding the bundle paths to the detector only made preflight report available; spawn still ran the bare `session.command` ('codex'). A user with only Codex.app (no `codex` on PATH) passed preflight but then hit ENOENT at spawn('codex'). Thread the absolute path the detector resolved out of preflight and feed it to resolveCliSpawnPlan/spawn. This also hardens the existing login-shell-PATH-only case, where spawn runs against a leaner env than detection did. Skipped on Windows, where resolveCliSpawnPlan does its own .cmd/.exe shim resolution from the bare command that we must not pre-empt. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
💄 style(desktop): hint to install gh CLI when PR status unavailable
2026-06-21 06:29:59 +00:00 · 2026-06-13 12:56:17 +08:00 · 2026-06-13 12:53:10 +08:00 · 2026-06-13 11:49:23 +08:00 · 2026-06-13 11:46:16 +08:00 · 2026-06-13 11:21:25 +08:00
708 changed files with 24148 additions and 7298 deletions
@@ -0,0 +1,217 @@
+---
+name: agent-testing
+description: >
+  Agentic end-to-end testing for LobeHub: backend verification via the CLI,
+  frontend verification via agent-browser (Electron), full-stack verification in
+  the browser, and bot-channel verification via osascript. Local-first today,
+  designed to extend to cloud automation. Triggers on 'cli test', 'test with cli',
+  'verify with cli', 'backend test with cli', 'local test', 'test in electron',
+  'test desktop', 'test bot', 'bot test', 'test in discord', 'test in telegram',
+  'test in slack', 'test in wechat', 'test in weixin', 'test in lark', 'test in feishu',
+  'test in qq', 'manual test', 'osascript', 'test report', or any local
+  end-to-end verification task.
+---
+
+# Agent Testing (Agentic End-to-End Verification)
+
+One skill for all agentic end-to-end testing — local-first today, designed to
+also run as full cloud automation. Every test session follows the same
+four-step contract:
+
+```
+Step 0: Env + Auth  →  Step 1: Pick surface  →  Step 2: Run  →  Step 3: Structured report
+```
+
+## Step 0 — Environment setup + auth check (mandatory)
+
+Step 0 is about getting the environment ready: **dependencies are healthy**
+and **auth is green**. A test run that dies halfway on a missing dependency or
+a login wall wastes the whole session — clear both gates BEFORE writing a
+single test step.
+
+### 0.1 Dependencies are installed — root AND standalone apps
+
+The root pnpm workspace does **NOT** cover every app: `pnpm-workspace.yaml`
+lists `packages/**`, `e2e`, `apps/server`, and only `apps/desktop/src/main` —
+**`apps/desktop` and `apps/cli` are standalone**, each keeping its own
+`node_modules` with its own links into `packages/`. A root install does not
+refresh them, so install in every app the test will touch:
+
+```bash
+pnpm install                    # root workspace
+cd apps/desktop && pnpm install # Electron surface
+cd apps/cli && pnpm install     # CLI surface
+```
+
+Symptom of a stale standalone install: the build/launch fails to resolve a
+recently added workspace package — `Rolldown failed to resolve import
+"@lobechat/<pkg>"` (Electron) or `Cannot find module '@lobechat/<pkg>'` (CLI).
+
+### 0.2 Run scripts from the repo root
+
+All paths in this skill (`./.agents/skills/agent-testing/...`) are
+repo-root-relative, and background commands inherit the current working
+directory — a script launched while `cwd` is `apps/desktop` fails with
+`No such file or directory`. Verify `pwd` is the repo root before launching
+long-running scripts.
+
+### 0.3 Auth is green
+
+**Auth is the gate for all automated testing.**
+
+```bash
+./.agents/skills/agent-testing/scripts/setup-auth.sh status
+```
+
+| Surface  | Mechanism                                         | One-key path                   | Standard check             |
+| -------- | ------------------------------------------------- | ------------------------------ | -------------------------- |
+| CLI      | OIDC Device Code Flow (`apps/cli/.lobehub-dev`)   | `setup-auth.sh cli`            | `setup-auth.sh status`     |
+| Web      | better-auth cookie injection into `agent-browser` | `pbpaste \| setup-auth.sh web` | `setup-auth.sh web-verify` |
+| Electron | App's own persistent login state                  | Log in once in the app         | `app-probe.sh auth`        |
+| Bot      | Native apps already logged in                     | —                              | per-platform screenshot    |
+
+Login-state checks are standardized — do NOT hand-roll `window.__LOBE_STORES`
+eval snippets; use `scripts/app-probe.sh auth` (returns `{ isSignedIn, userId }`,
+works for Electron CDP and web sessions via `AB_TARGET`).
+
+If `status` is not all green, fix auth first (the steps that need a human must be
+requested from the user explicitly). Full background and failure modes:
+[references/auth.md](./references/auth.md).
+
+## Step 1 — Pick the surface by change scope
+
+| Change scope                                            | Default surface                      | Why                                                               | Guide                              |
+| ------------------------------------------------------- | ------------------------------------ | ----------------------------------------------------------------- | ---------------------------------- |
+| **Backend** (TRPC router / service / model / migration) | **CLI**                              | Fastest loop, text-assertable output, zero UI flakiness           | [cli/index.md](./cli/index.md)     |
+| **Pure frontend** (components, store, styles, UX)       | **Electron** (agent-browser + CDP)   | Primary product shape; `__LOBE_STORES` state introspection        | [ui/electron.md](./ui/electron.md) |
+| **Full-stack** (new API + UI consuming it)              | **Web** (browser + local dev server) | One surface where network requests and UI are observable together | [ui/web.md](./ui/web.md)           |
+| **Bot channels** (Discord / WeChat / Lark / …)          | Native app via osascript / bridge    | Only way to exercise the real channel end-to-end                  | `bot/<platform>/index.md`          |
+
+Escalate, don't duplicate: verify a backend change with the CLI first; only add
+a UI pass when the change actually affects the UI.
+
+### Environment support (local macOS vs cloud Linux)
+
+The decisive constraint per surface is **how evidence (screenshots) is
+captured**: CDP-based capture (`agent-browser screenshot`) renders from the
+browser engine and needs no real display; OS-level capture (`screencapture`,
+osascript) is macOS-only.
+
+| Surface  | macOS (local) | Linux / cloud (headless)                                  | Screenshot mechanism                                   |
+| -------- | ------------- | --------------------------------------------------------- | ------------------------------------------------------ |
+| CLI      | ✅            | ✅                                                        | n/a — text output                                      |
+| Web      | ✅            | ✅ headless Chromium works natively                       | CDP — no display needed                                |
+| Electron | ✅            | ⚠️ runs, but needs a display server: wrap with `xvfb-run` | CDP works under Xvfb; `capture-app-window.sh` does NOT |
+| Bot      | ✅            | ❌ osascript + native apps are macOS-only                 | macOS `screencapture` only                             |
+
+When a test must stay cloud-portable, prefer CDP-based evidence over
+OS-level capture wherever both exist.
+
+### Bot platforms
+
+| Platform      | Guide                                            | Quick switcher        |
+| ------------- | ------------------------------------------------ | --------------------- |
+| Discord       | [bot/discord/index.md](./bot/discord/index.md)   | `Cmd+K`               |
+| Slack         | [bot/slack/index.md](./bot/slack/index.md)       | `Cmd+K`               |
+| Telegram      | [bot/telegram/index.md](./bot/telegram/index.md) | `Cmd+F`               |
+| WeChat / 微信 | [bot/wechat/index.md](./bot/wechat/index.md)     | `Cmd+F`               |
+| Lark / 飞书   | [bot/lark/index.md](./bot/lark/index.md)         | `Cmd+K`               |
+| QQ            | [bot/qq/index.md](./bot/qq/index.md)             | `Cmd+F`               |
+| iMessage      | [bot/imessage/index.md](./bot/imessage/index.md) | bridge (no osascript) |
+
+Each platform folder contains an `index.md` (activation, navigation,
+send-message, verification snippets) and a `test-<platform>-bot.sh` script
+sharing the interface:
+
+```bash
+./.agents/skills/agent-testing/bot/<platform>/test-<platform>-bot.sh <channel_or_contact> <message> [wait_seconds] [screenshot_path]
+```
+
+New to osascript automation? Read
+[references/osascript.md](./references/osascript.md) first — it is a general
+macOS-automation asset (activate, type, paste, screenshot, accessibility reads,
+gotchas), not bot-specific.
+
+## Step 2 — Run
+
+Surface guides above carry the detailed workflows. Shared infrastructure:
+
+| Need                                 | Where                                                                |
+| ------------------------------------ | -------------------------------------------------------------------- |
+| Start / restart the local dev server | [references/dev-server.md](./references/dev-server.md)               |
+| `agent-browser` command reference    | [references/agent-browser.md](./references/agent-browser.md)         |
+| osascript patterns (general macOS)   | [references/osascript.md](./references/osascript.md)                 |
+| Agent gateway probing                | [references/agent-gateway.md](./references/agent-gateway.md)         |
+| Screen recording                     | [references/record-app-screen.md](./references/record-app-screen.md) |
+
+### Scripts
+
+All under `.agents/skills/agent-testing/scripts/`:
+
+| Script                    | Usage                                                                        |
+| ------------------------- | ---------------------------------------------------------------------------- |
+| `setup-auth.sh`           | One-stop auth setup & status check (`status` / `cli` / `web`)                |
+| `app-probe.sh`            | LobeHub app probes: `auth` / `route` / `ops` / `goto <path>` / `errors`      |
+| `record-gif.sh`           | Frame-sequence → GIF for time-based behavior (streaming, timers, animations) |
+| `report-init.sh`          | Scaffold a structured test report (Step 3)                                   |
+| `electron-dev.sh`         | Manage Electron dev env (start/stop/status/restart, CDP 9222)                |
+| `capture-app-window.sh`   | Screenshot a specific app window (general; used by bot tests)                |
+| `record-app-screen.sh`    | Record app screen (video + periodic screenshots)                             |
+| `record-electron-demo.sh` | Record Electron app demo with ffmpeg                                         |
+| `agent-gateway/`          | Gateway probe / dump / analyze tools                                         |
+
+`app-probe.sh` is the LobeHub-specific fast path into app state — auth check,
+current route, running operations, and `goto <path>` quick navigation
+(`/agent/<agentId>/<topicId>`, `/task/<taskId>`, `/settings`, …) so a test can
+jump straight to the state under test instead of clicking through the UI. See
+[ui/electron.md](./ui/electron.md#lobehub-probes--quick-navigation) for usage.
+
+## Step 3 — Structured report (mandatory deliverable)
+
+Every automated test session ends with a structured, evidence-backed report —
+not a chat-only summary. Scaffold it up front and fill it as you test:
+
+```bash
+DIR=$(./.agents/skills/agent-testing/scripts/report-init.sh my-feature "Verify my feature")
+# ... test, saving screenshots / CLI transcripts into $DIR/assets/ ...
+# fill $DIR/report.md (scope, case table with inline evidence, verdict, score) and $DIR/result.json
+```
+
+Reports live in `.records/reports/<timestamp>-<slug>/` (gitignored): `report.md`
+(human-readable, with screenshots/GIFs embedded directly in the case table),
+`result.json` (machine-readable pass/fail + score), `assets/` (evidence).
+Format spec and evidence rules:
+[references/report.md](./references/report.md).
+
+Two hard rules worth front-loading:
+
+- **Report language = the user's conversation language.** Write the ENTIRE
+  `report.md` (headings included) in the language the user is conversing in —
+  no mixed English. `result.json` keys/status values stay English.
+- **The case table is the main reading surface.** Prefer the compact
+  `# | case | result | key observation | evidence` shape and embed the
+  screenshot/GIF in the evidence cell. Use separate evidence sections only for
+  long CLI transcripts, HAR summaries, or supplemental detail.
+- **Time-based behavior needs a GIF, not a screenshot.** If a case asserts
+  change over time (streaming output, a ticking timer, loading states,
+  animations), record it with `scripts/record-gif.sh` and embed the GIF —
+  a static screenshot cannot prove the behavior.
+
+## Directory map
+
+```
+agent-testing/
+├── SKILL.md            # this router
+├── cli/index.md        # backend verification via the LobeHub CLI
+├── ui/electron.md      # pure-frontend verification in the desktop app
+├── ui/web.md           # full-stack verification in the browser
+├── bot/<platform>/     # bot-channel verification (osascript / bridge)
+├── references/         # shared knowledge: auth, dev-server, agent-browser, osascript, report
+└── scripts/            # setup-auth, report-init, electron-dev, capture, recording, gateway
+```
+
+## Gotchas
+
+- agent-browser: see [references/agent-browser.md](./references/agent-browser.md#gotchas)
+- Electron: see [ui/electron.md](./ui/electron.md#electron-gotchas)
+- osascript: see [references/osascript.md](./references/osascript.md#gotchas)
@@ -2,7 +2,7 @@

 **App name:** `Discord` | **Process name:** `Discord`

-See [osascript-common.md](../osascript-common.md) for shared patterns.
+See [references/osascript.md](../../references/osascript.md) for shared patterns.

 ## Activate & Navigate

@@ -92,6 +92,6 @@ echo "Screenshot saved to /tmp/discord-test-result.png"
 ## Script

 ```bash
-./.agents/skills/local-testing/bot/discord/test-discord-bot.sh "bot-testing" "!ping"
-./.agents/skills/local-testing/bot/discord/test-discord-bot.sh "bot-testing" "/ask Tell me a joke" 30
+./.agents/skills/agent-testing/bot/discord/test-discord-bot.sh "bot-testing" "!ping"
+./.agents/skills/agent-testing/bot/discord/test-discord-bot.sh "bot-testing" "/ask Tell me a joke" 30
 ```
@@ -60,5 +60,5 @@ echo "[$APP] Waiting ${WAIT}s for bot response..."
 sleep "$WAIT"

 echo "[$APP] Capturing screenshot..."
-"$SCRIPT_DIR/../capture-app-window.sh" "$APP" "$SCREENSHOT"
+"$SCRIPT_DIR/../../scripts/capture-app-window.sh" "$APP" "$SCREENSHOT"
 echo "[$APP] Done! Screenshot saved to $SCREENSHOT"
@@ -21,7 +21,7 @@ So the test surface is three layers:
  curl -sS -m4 -o /dev/null -w '%{http_code}\n' \
    "http://127.0.0.1:1234/api/v1/server/info?password=<PW>" # expect 200
  ```
- **Electron dev running with CDP**: `./.agents/skills/local-testing/scripts/electron-dev.sh start`
+- **Electron dev running with CDP**: `./.agents/skills/agent-testing/scripts/electron-dev.sh start`
 - The **iMessage Desktop branch** checked out (the `imessageBridge` IPC group
  and `@lobechat/chat-adapter-imessage` must be compiled into the main bundle).
  Run `pnpm install --ignore-scripts` at the repo root **and** in `apps/desktop/`
@@ -31,7 +31,7 @@ So the test surface is three layers:
 ## Fast path: automated script

 ```bash
-./.agents/skills/local-testing/bot/imessage/test-imessage-bridge.sh '<bluebubbles_password>' [bb_url] [cdp_port]
+./.agents/skills/agent-testing/bot/imessage/test-imessage-bridge.sh '<bluebubbles_password>' [bb_url] [cdp_port]
 ```

 Asserts the whole flow and self-cleans (unique `applicationId` per run, removes
@@ -136,7 +136,7 @@ Verifies the leg the bridge uses to _reply_: `BlueBubblesApiClient.sendText`
 → `POST /api/v1/message/text`. Run the helper against your own number:

 ```bash
-./.agents/skills/local-testing/bot/imessage/send-imessage-test.sh '<bb_password>' '+<E164>' # e.g. +15551234567
+./.agents/skills/agent-testing/bot/imessage/send-imessage-test.sh '<bb_password>' '+<E164>' # e.g. +15551234567
 ```

 **Gotcha that bites everyone:** with `method=apple-script` and a _new_
@@ -2,7 +2,7 @@

 **App name:** `Lark` or `飞书` | **Process name:** `Lark` or `飞书`

-See [osascript-common.md](../osascript-common.md) for shared patterns.
+See [references/osascript.md](../../references/osascript.md) for shared patterns.

 ## Activate & Navigate

@@ -56,6 +56,6 @@ screencapture /tmp/lark-bot-response.png
 ## Script

 ```bash
-./.agents/skills/local-testing/bot/lark/test-lark-bot.sh "bot-testing" "@MyBot hello"
-./.agents/skills/local-testing/bot/lark/test-lark-bot.sh "bot-testing" "Help me with this" 30
+./.agents/skills/agent-testing/bot/lark/test-lark-bot.sh "bot-testing" "@MyBot hello"
+./.agents/skills/agent-testing/bot/lark/test-lark-bot.sh "bot-testing" "Help me with this" 30
 ```
@@ -80,5 +80,5 @@ echo "[$APP] Waiting ${WAIT}s for bot response..."
 sleep "$WAIT"

 echo "[$APP] Capturing screenshot..."
-"$SCRIPT_DIR/../capture-app-window.sh" "$APP" "$SCREENSHOT"
+"$SCRIPT_DIR/../../scripts/capture-app-window.sh" "$APP" "$SCREENSHOT"
 echo "[$APP] Done! Screenshot saved to $SCREENSHOT"
@@ -2,7 +2,7 @@

 **App name:** `QQ` | **Process name:** `QQ`

-See [osascript-common.md](../osascript-common.md) for shared patterns.
+See [references/osascript.md](../../references/osascript.md) for shared patterns.

 ## Activate & Navigate

@@ -57,6 +57,6 @@ screencapture /tmp/qq-bot-response.png
 ## Script

 ```bash
-./.agents/skills/local-testing/bot/qq/test-qq-bot.sh "bot-testing" "Hello bot" 15
-./.agents/skills/local-testing/bot/qq/test-qq-bot.sh "MyBot" "/help" 10
+./.agents/skills/agent-testing/bot/qq/test-qq-bot.sh "bot-testing" "Hello bot" 15
+./.agents/skills/agent-testing/bot/qq/test-qq-bot.sh "MyBot" "/help" 10
 ```
@@ -72,5 +72,5 @@ echo "[$APP] Waiting ${WAIT}s for bot response..."
 sleep "$WAIT"

 echo "[$APP] Capturing screenshot..."
-"$SCRIPT_DIR/../capture-app-window.sh" "$APP" "$SCREENSHOT"
+"$SCRIPT_DIR/../../scripts/capture-app-window.sh" "$APP" "$SCREENSHOT"
 echo "[$APP] Done! Screenshot saved to $SCREENSHOT"
@@ -2,7 +2,7 @@

 **App name:** `Slack` | **Process name:** `Slack`

-See [osascript-common.md](../osascript-common.md) for shared patterns.
+See [references/osascript.md](../../references/osascript.md) for shared patterns.

 ## Activate & Navigate

@@ -68,6 +68,6 @@ screencapture /tmp/slack-bot-response.png
 ## Script

 ```bash
-./.agents/skills/local-testing/bot/slack/test-slack-bot.sh "bot-testing" "@mybot hello"
-./.agents/skills/local-testing/bot/slack/test-slack-bot.sh "bot-testing" "/ask What is 2+2?" 20
+./.agents/skills/agent-testing/bot/slack/test-slack-bot.sh "bot-testing" "@mybot hello"
+./.agents/skills/agent-testing/bot/slack/test-slack-bot.sh "bot-testing" "/ask What is 2+2?" 20
 ```
@@ -60,5 +60,5 @@ echo "[$APP] Waiting ${WAIT}s for bot response..."
 sleep "$WAIT"

 echo "[$APP] Capturing screenshot..."
-"$SCRIPT_DIR/../capture-app-window.sh" "$APP" "$SCREENSHOT"
+"$SCRIPT_DIR/../../scripts/capture-app-window.sh" "$APP" "$SCREENSHOT"
 echo "[$APP] Done! Screenshot saved to $SCREENSHOT"
@@ -2,7 +2,7 @@

 **App name:** `Telegram` | **Process name:** `Telegram`

-See [osascript-common.md](../osascript-common.md) for shared patterns.
+See [references/osascript.md](../../references/osascript.md) for shared patterns.

 ## Activate & Navigate

@@ -75,6 +75,6 @@ curl -s "https://api.telegram.org/bot$TELEGRAM_BOT_TOKEN/getUpdates?limit=5" | j
 ## Script

 ```bash
-./.agents/skills/local-testing/bot/telegram/test-telegram-bot.sh "MyTestBot" "/start"
-./.agents/skills/local-testing/bot/telegram/test-telegram-bot.sh "GPTBot" "Hello" 60
+./.agents/skills/agent-testing/bot/telegram/test-telegram-bot.sh "MyTestBot" "/start"
+./.agents/skills/agent-testing/bot/telegram/test-telegram-bot.sh "GPTBot" "Hello" 60
 ```
@@ -75,5 +75,5 @@ echo "[$APP] Waiting ${WAIT}s for bot response..."
 sleep "$WAIT"

 echo "[$APP] Capturing screenshot..."
-"$SCRIPT_DIR/../capture-app-window.sh" "$APP" "$SCREENSHOT"
+"$SCRIPT_DIR/../../scripts/capture-app-window.sh" "$APP" "$SCREENSHOT"
 echo "[$APP] Done! Screenshot saved to $SCREENSHOT"
@@ -2,7 +2,7 @@

 **App name:** `微信` or `WeChat` | **Process name:** `WeChat`

-See [osascript-common.md](../osascript-common.md) for shared patterns.
+See [references/osascript.md](../../references/osascript.md) for shared patterns.

 ## Activate & Navigate

@@ -76,6 +76,6 @@ screencapture /tmp/wechat-bot-response.png
 ## Script

 ```bash
-./.agents/skills/local-testing/bot/wechat/test-wechat-bot.sh "文件传输助手" "test message" 5
-./.agents/skills/local-testing/bot/wechat/test-wechat-bot.sh "MyBot" "Tell me a joke" 30
+./.agents/skills/agent-testing/bot/wechat/test-wechat-bot.sh "文件传输助手" "test message" 5
+./.agents/skills/agent-testing/bot/wechat/test-wechat-bot.sh "MyBot" "Tell me a joke" 30
 ```
@@ -81,5 +81,5 @@ echo "[$APP] Waiting ${WAIT}s for bot response..."
 sleep "$WAIT"

 echo "[$APP] Capturing screenshot..."
-"$SCRIPT_DIR/../capture-app-window.sh" "$APP" "$SCREENSHOT"
+"$SCRIPT_DIR/../../scripts/capture-app-window.sh" "$APP" "$SCREENSHOT"
 echo "[$APP] Done! Screenshot saved to $SCREENSHOT"
@@ -0,0 +1,142 @@
+# CLI Backend Verification
+
+Default surface for verifying **backend changes** (TRPC routers, services,
+models, migrations) end-to-end: fastest loop, text-assertable output, zero UI
+flakiness.
+
+## When to use
+
+- Verifying TRPC router / service / model changes end-to-end
+- Testing new API fields or response structure changes
+- Validating CLI command output after backend modifications
+- Debugging data flow issues between server and CLI
+
+## Prerequisites
+
+| Requirement  | Details                                                                           |
+| ------------ | --------------------------------------------------------------------------------- |
+| Dev server   | `localhost:3010` — see [../references/dev-server.md](../references/dev-server.md) |
+| CLI source   | `apps/cli/` — runs from source, no rebuild; standalone `node_modules` — run `pnpm install` inside `apps/cli/` (root install does not cover it) |
+| CLI dev mode | `LOBEHUB_CLI_HOME=.lobehub-dev` for isolated credentials                          |
+| Auth         | Device Code Flow login — see [../references/auth.md](../references/auth.md)       |
+
+All CLI dev commands run from `apps/cli/`. Subsequent examples use `$CLI`:
+
+```bash
+CLI="LOBEHUB_CLI_HOME=.lobehub-dev bun src/index.ts"
+```
+
+## Workflow
+
+### Step 1 — Server up?
+
+See [../references/dev-server.md](../references/dev-server.md) for the health
+check, start, and restart commands. Server-side code changes require a restart.
+
+### Step 2 — Auth ready?
+
+```bash
+./.agents/skills/agent-testing/scripts/setup-auth.sh status
+```
+
+If the CLI is not logged in, **the user must run the login themselves**
+(interactive browser authorization):
+
+```bash
+cd apps/cli && LOBEHUB_CLI_HOME=.lobehub-dev bun src/index.ts login --server http://localhost:3010
+```
+
+Credentials persist in `apps/cli/.lobehub-dev/`. Details:
+[../references/auth.md](../references/auth.md).
+
+### Step 3 — Test with CLI commands
+
+CLI runs from source, so CLI-side code changes take effect immediately without
+rebuilding:
+
+```bash
+cd apps/cli
+$CLI <command>
+```
+
+Capture output for the report as you go (e.g. `$CLI task list | tee "$DIR/assets/task-list.txt"`).
+
+### Step 4 — Clean up test data
+
+```bash
+$CLI task delete < id > -y
+$CLI agent delete < id > -y
+```
+
+### Step 5 — Report
+
+Finish with a structured report —
+[../references/report.md](../references/report.md). CLI evidence = exact
+command + trimmed output.
+
+## Common testing patterns
+
+### Task system
+
+```bash
+$CLI task list
+$CLI task create -n "Root Task" -i "Test instruction"
+$CLI task create -n "Child Task" -i "Sub instruction" --parent T-1
+$CLI task view T-1
+$CLI task tree T-1
+$CLI task edit T-1 --status running
+$CLI task comment T-1 -m "Test comment"
+$CLI task delete T-1 -y
+```
+
+### Agent system
+
+```bash
+$CLI agent list
+$CLI agent view <agent-id>
+$CLI agent run <agent-id> -m "Test prompt"
+```
+
+### Document & knowledge base
+
+```bash
+$CLI doc list
+$CLI doc create -t "Test Doc" -c "Content here"
+$CLI doc view <doc-id>
+$CLI kb list
+$CLI kb tree <kb-id>
+```
+
+### Model & provider
+
+```bash
+$CLI model list
+$CLI provider list
+$CLI provider test <provider-id>
+```
+
+## Dev-test cycle
+
+```
+1. Make code changes (service/model/router/type)
+         |
+2. Run unit tests (fast feedback)
+   bunx vitest run --silent='passed-only' '<test-file>'
+         |
+3. Restart dev server (if server-side changes — see dev-server.md)
+         |
+4. CLI verification (end-to-end)
+   $CLI <command>
+         |
+5. Clean up test data + write the report
+```
+
+## Troubleshooting
+
+| Issue                       | Solution                                        |
+| --------------------------- | ----------------------------------------------- |
+| `No authentication found`   | Run `login --server http://localhost:3010`      |
+| `UNAUTHORIZED` on API calls | Token expired; re-run login                     |
+| `ECONNREFUSED`              | Dev server not running — see dev-server.md      |
+| CLI shows old data/behavior | Server needs restart to pick up code changes    |
+| Login opens wrong server    | Must use `--server` flag (env var doesn't work) |
@@ -0,0 +1,257 @@
+# agent-browser CLI Reference
+
+Generic reference for the `agent-browser` CLI — automate Chromium-based apps (Electron, Chrome, web) via Chrome DevTools Protocol. LobeHub-specific patterns live in [../ui/electron.md](../ui/electron.md) and [../ui/web.md](../ui/web.md); authentication recipes live in [auth.md](./auth.md).
+
+Use `agent-browser` to automate Chromium-based apps via Chrome DevTools Protocol.
+
+Install via `npm i -g agent-browser`, `brew install agent-browser`, or `cargo install agent-browser`. Run `agent-browser install` to download Chrome. Run `agent-browser upgrade` to update.
+
+## Core Workflow
+
+Every browser automation follows this pattern:
+
+1. **Navigate**: `agent-browser open <url>`
+2. **Snapshot**: `agent-browser snapshot -i` (get element refs like `@e1`, `@e2`)
+3. **Interact**: Use refs to click, fill, select
+4. **Re-snapshot**: After navigation or DOM changes, get fresh refs
+
+```bash
+agent-browser open https://example.com/form
+agent-browser snapshot -i
+# Output: @e1 [input type="email"], @e2 [input type="password"], @e3 [button] "Submit"
+
+agent-browser fill @e1 "user@example.com"
+agent-browser fill @e2 "password123"
+agent-browser click @e3
+agent-browser wait --load networkidle
+agent-browser snapshot -i # Check result
+```
+
+## Command Chaining
+
+```bash
+# Chain open + wait + snapshot in one call
+agent-browser open https://example.com && agent-browser wait --load networkidle && agent-browser snapshot -i
+```
+
+Use `&&` when you don't need to read intermediate output. Run commands separately when you need to parse output first (e.g., snapshot to discover refs, then interact).
+
+## Essential Commands
+
+```bash
+# Navigation
+agent-browser open <url>              # Navigate (aliases: goto, navigate)
+agent-browser close                   # Close browser
+agent-browser close --all             # Close all active sessions
+
+# Snapshot
+agent-browser snapshot -i             # Interactive elements with refs (recommended)
+agent-browser snapshot -s "#selector" # Scope to CSS selector
+
+# Interaction (use @refs from snapshot)
+agent-browser click @e1               # Click element
+agent-browser click @e1 --new-tab     # Click and open in new tab
+agent-browser fill @e2 "text"         # Clear and type text
+agent-browser type @e2 "text"         # Type without clearing
+agent-browser select @e1 "option"     # Select dropdown option
+agent-browser check @e1               # Check checkbox
+agent-browser press Enter             # Press key
+agent-browser keyboard type "text"    # Type at current focus (no selector)
+agent-browser keyboard inserttext "text"  # Insert without key events
+agent-browser scroll down 500         # Scroll page
+agent-browser scroll down 500 --selector "div.content"  # Scroll within container
+
+# Get information
+agent-browser get text @e1            # Get element text
+agent-browser get url                 # Get current URL
+agent-browser get title               # Get page title
+agent-browser get cdp-url             # Get CDP WebSocket URL
+
+# Wait
+agent-browser wait @e1                # Wait for element
+agent-browser wait --load networkidle # Wait for network idle
+agent-browser wait --url "**/page"    # Wait for URL pattern
+agent-browser wait 2000               # Wait milliseconds
+agent-browser wait --text "Welcome"   # Wait for text to appear
+agent-browser wait --fn "!document.body.innerText.includes('Loading...')"  # Wait for text to disappear
+agent-browser wait "#spinner" --state hidden  # Wait for element to disappear
+
+# Downloads
+agent-browser download @e1 ./file.pdf          # Click element to trigger download
+agent-browser wait --download ./output.zip     # Wait for any download to complete
+
+# Network
+agent-browser network requests                 # Inspect tracked requests
+agent-browser network requests --type xhr,fetch  # Filter by resource type
+agent-browser network requests --method POST   # Filter by HTTP method
+agent-browser network route "**/api/*" --abort # Block matching requests
+agent-browser network har start                # Start HAR recording
+agent-browser network har stop ./capture.har   # Stop and save HAR file
+
+# Viewport & Device Emulation
+agent-browser set viewport 1920 1080          # Set viewport size (default: 1280x720)
+agent-browser set viewport 1920 1080 2        # 2x retina
+agent-browser set device "iPhone 14"          # Emulate device (viewport + user agent)
+
+# Capture
+agent-browser screenshot              # Screenshot to temp dir
+agent-browser screenshot --full       # Full page screenshot
+agent-browser screenshot --annotate   # Annotated screenshot with numbered element labels
+agent-browser pdf output.pdf          # Save as PDF
+
+# Clipboard
+agent-browser clipboard read          # Read text from clipboard
+agent-browser clipboard write "text"  # Write text to clipboard
+agent-browser clipboard copy          # Copy current selection
+agent-browser clipboard paste         # Paste from clipboard
+
+# Dialogs (alert, confirm, prompt, beforeunload)
+agent-browser dialog accept           # Accept dialog
+agent-browser dialog accept "input"   # Accept prompt dialog with text
+agent-browser dialog dismiss          # Dismiss/cancel dialog
+agent-browser dialog status           # Check if dialog is open
+
+# Diff (compare page states)
+agent-browser diff snapshot                        # Compare current vs last snapshot
+agent-browser diff screenshot --baseline before.png  # Visual pixel diff
+agent-browser diff url <url1> <url2>               # Compare two pages
+
+# Streaming
+agent-browser stream enable           # Start WebSocket streaming
+agent-browser stream status           # Inspect streaming state
+agent-browser stream disable          # Stop streaming
+```
+
+## Batch Execution
+
+```bash
+echo '[
+  ["open", "https://example.com"],
+  ["snapshot", "-i"],
+  ["click", "@e1"],
+  ["screenshot", "result.png"]
+]' | agent-browser batch --json
+```
+
+## Authentication
+
+```bash
+# Option 1: Auth vault (credentials stored encrypted)
+echo "$PASSWORD" | agent-browser auth save myapp --url https://app.example.com/login --username user --password-stdin
+agent-browser auth login myapp
+
+# Option 2: Session name (auto-save/restore cookies + localStorage)
+agent-browser --session-name myapp open https://app.example.com/login
+agent-browser close                                                       # State auto-saved
+agent-browser --session-name myapp open https://app.example.com/dashboard # Auto-restored
+
+# Option 3: Persistent profile
+agent-browser --profile ~/.myapp open https://app.example.com/login
+
+# Option 4: State file
+agent-browser state save auth.json
+agent-browser state load auth.json
+```
+
+### LobeHub dev server — inject better-auth cookie
+
+`agent-browser --headed` on macOS can create an off-screen Chromium window, blocking manual login. For a local LobeHub dev server (e.g. `localhost:3010`), copy the `better-auth.session_token` cookie out of a **Network request** in the user's own Chrome DevTools and load it via `state load`. See [auth.md](./auth.md) for the full recipe.
+
+## Semantic Locators (Alternative to Refs)
+
+```bash
+agent-browser find text "Sign In" click
+agent-browser find label "Email" fill "user@test.com"
+agent-browser find role button click --name "Submit"
+agent-browser find placeholder "Search" type "query"
+agent-browser find testid "submit-btn" click
+```
+
+## JavaScript Evaluation (eval)
+
+```bash
+# Simple expressions
+agent-browser eval 'document.title'
+
+# Complex JS: use --stdin with heredoc (RECOMMENDED)
+agent-browser eval --stdin << 'EVALEOF'
+JSON.stringify(
+  Array.from(document.querySelectorAll("img"))
+    .filter(i => !i.alt)
+    .map(i => ({ src: i.src.split("/").pop(), width: i.width }))
+)
+EVALEOF
+
+# Base64 encoding (avoids all shell escaping issues)
+agent-browser eval -b "$(echo -n 'document.title' | base64)"
+```
+
+## Ref Lifecycle
+
+Refs (`@e1`, `@e2`, etc.) are invalidated when the page changes. Always re-snapshot after clicking links/buttons that navigate, form submissions, or dynamic content loading.
+
+## Annotated Screenshots (Vision Mode)
+
+```bash
+agent-browser screenshot --annotate
+# Output includes the image path and a legend:
+#   [1] @e1 button "Submit"
+#   [2] @e2 link "Home"
+agent-browser click @e2 # Click using ref from annotated screenshot
+```
+
+## Parallel Sessions
+
+```bash
+agent-browser --session site1 open https://site-a.com
+agent-browser --session site2 open https://site-b.com
+agent-browser session list
+```
+
+## Connect to Existing Chrome
+
+```bash
+agent-browser --auto-connect snapshot # Auto-discover running Chrome
+agent-browser --cdp 9222 snapshot     # Explicit CDP port
+```
+
+## iOS Simulator (Mobile Safari)
+
+```bash
+agent-browser device list
+agent-browser -p ios --device "iPhone 16 Pro" open https://example.com
+agent-browser -p ios snapshot -i
+agent-browser -p ios tap @e1
+agent-browser -p ios swipe up
+agent-browser -p ios screenshot mobile.png
+agent-browser -p ios close
+```
+
+## Observability Dashboard
+
+```bash
+agent-browser dashboard install
+agent-browser dashboard start # Background server on port 4848
+agent-browser dashboard stop
+```
+
+## Cloud Providers
+
+Use `-p <provider>` to run against cloud browsers: `agentcore`, `browserbase`, `browserless`, `browseruse`, `kernel`.
+
+## Browser Engine Selection
+
+```bash
+agent-browser --engine lightpanda open example.com # 10x faster, 10x less memory
+```
+
+## Gotchas
+
+- **Daemon can get stuck** — if commands hang, `agent-browser close --all` or `pkill -f agent-browser` to reset
+- **HMR invalidates everything** — after code changes, refs break. Re-snapshot or restart
+- **`snapshot -i` doesn't find contenteditable** — use `snapshot -i -C` for rich text editors
+- **`fill` doesn't work on contenteditable** — use `type` for chat inputs
+- **Screenshots go to `~/.agent-browser/tmp/screenshots/`** — read them with the `Read` tool
+- **Dialogs block all commands** — if commands time out, check `agent-browser dialog status`
+- **Default timeout is 25s** — override with `AGENT_BROWSER_DEFAULT_TIMEOUT` (ms) or use explicit waits
+- **Shell quoting corrupts eval** — use `eval --stdin <<'EVALEOF'` for complex JS
@@ -19,13 +19,13 @@ works for any LobeHub streaming session.

 ```bash
 # 1. Start Electron with CDP
-./.agents/skills/local-testing/scripts/electron-dev.sh start
+./.agents/skills/agent-testing/scripts/electron-dev.sh start

 # 2. Navigate to a chat, switch runtime to Cloud Sandbox (gateway mode)

 # 3. Install the probe + helpers
 agent-browser --cdp 9222 eval --stdin \
-  < .agents/skills/local-testing/scripts/agent-gateway/probe.js
+  < .agents/skills/agent-testing/scripts/agent-gateway/probe.js

 # 4. Send a tool-call message — manually or via type+press
 agent-browser --cdp 9222 eval "window.__PROBE_EVENT('SENT')"
@@ -34,15 +34,15 @@ agent-browser --cdp 9222 eval "window.__PROBE_EVENT('SENT')"
 #    rightmost inactive tab as AWAY — edit ROUND_TRIPS / DWELL_MS in the
 #    file if you want different timing)
 agent-browser --cdp 9222 eval --stdin \
-  < .agents/skills/local-testing/scripts/agent-gateway/tab-switch.js
+  < .agents/skills/agent-testing/scripts/agent-gateway/tab-switch.js

 # 6. Wait for streaming to finish, then dump
 agent-browser --cdp 9222 eval --stdin \
-  < .agents/skills/local-testing/scripts/agent-gateway/probe-dump.js \
+  < .agents/skills/agent-testing/scripts/agent-gateway/probe-dump.js \
  > /tmp/probe.json

 # 7. Analyze
-node .agents/skills/local-testing/scripts/agent-gateway/analyze.mjs /tmp/probe.json
+node .agents/skills/agent-testing/scripts/agent-gateway/analyze.mjs /tmp/probe.json
 ```

 The analyzer prints three sections: EVENTS, TIMELINE, REGRESSIONS. If
@@ -0,0 +1,123 @@
+# Auth Setup for Local Agent Testing
+
+**Auth is the gate for all automated testing.** Prepare and verify it before
+writing any test step. The one-stop entry point is:
+
+```bash
+SCRIPT=".agents/skills/agent-testing/scripts/setup-auth.sh"
+
+$SCRIPT status        # check server + CLI + web auth readiness
+$SCRIPT cli           # interactive CLI device-code login (must be run by the user)
+pbpaste | $SCRIPT web # inject a copied Cookie header into the agent-browser session
+$SCRIPT web-verify    # live-check that the agent-browser session is authenticated
+```
+
+`SERVER_URL` defaults to `http://localhost:3010` (this repo's `dev:next` port).
+Override it when testing against another server (e.g. `SERVER_URL=http://localhost:3011`
+in the cloud repo).
+
+## Per-surface overview
+
+| Surface  | Mechanism                                | Persistence                                                       | Human interaction                               |
+| -------- | ---------------------------------------- | ----------------------------------------------------------------- | ----------------------------------------------- |
+| CLI      | OIDC Device Code Flow                    | `apps/cli/.lobehub-dev/settings.json`                             | Yes — browser authorization, every token expiry |
+| Web      | better-auth cookie injection             | `~/.lobehub-agent-testing/web-state.json` + agent-browser session | Copy the Cookie header once per token rotation  |
+| Electron | App's own login state                    | Electron user-data dir                                            | Log in once manually in the app                 |
+| Bot      | Native apps (Discord/WeChat/…) logged in | Each app's own session                                            | Once per app                                    |
+
+## CLI — Device Code Flow
+
+Credentials are isolated from the user's real CLI config via
+`LOBEHUB_CLI_HOME=.lobehub-dev` (kept inside `apps/cli/`, gitignored).
+
+Login requires interactive browser authorization, so **the user must run it
+themselves** (e.g. via the `!` prefix in Claude Code):
+
+```bash
+cd apps/cli && LOBEHUB_CLI_HOME=.lobehub-dev bun src/index.ts login --server http://localhost:3010
+```
+
+- The `--server` flag is required — an env var does NOT work and login will hit
+  the wrong server without it.
+- Check state without logging in: `setup-auth.sh status` (verifies
+  `settings.json` exists and `serverUrl` matches).
+- `UNAUTHORIZED` on API calls means the token expired — re-run login.
+
+## Web — better-auth cookie injection (agent-browser)
+
+`agent-browser --headed` on macOS often creates the Chromium window off-screen —
+the user can't see or interact with it, so manual login inside the agent-browser
+session fails. Instead, copy the **better-auth session cookie** out of the
+user's own logged-in Chrome and inject it as a Playwright-style state file.
+
+Do **not** use this on production URLs — only local dev. Treat the cookie as a
+secret: don't paste it into shared logs, PRs, or commit it anywhere.
+
+### One-key path
+
+1. Ask the user to copy the Cookie header **from a Network request, NOT
+   `document.cookie`** (`document.cookie` cannot see HttpOnly cookies, which is
+   exactly where better-auth puts its session):
+   - Open the logged-in tab (`http://localhost:<port>/…`) in Chrome.
+   - `Cmd+Option+I` → **Network** tab → refresh → click any same-origin request.
+   - Under **Request Headers**, right-click the `Cookie:` line → **Copy value**.
+2. Inject and verify in one shot:
+
+```bash
+pbpaste | ./.agents/skills/agent-testing/scripts/setup-auth.sh web
+```
+
+The script filters the header down to the better-auth cookies
+(`better-auth.session_token`, `better-auth.state`), builds the Playwright
+`storageState` JSON, loads it into the `agent-browser` session (default name
+`lobehub-dev`), opens `SERVER_URL`, and asserts the URL is not `/signin`.
+
+### Using the authenticated session
+
+```bash
+agent-browser --session lobehub-dev open "http://localhost:3010/"
+agent-browser --session lobehub-dev snapshot -i | head -20
+# Look for the user's avatar/name in the sidebar, or absence of the signin form.
+```
+
+### Notes
+
+- `storageState` doesn't enforce the HttpOnly flag on load — the script stores
+  cookies with `httpOnly: false`, which is fine for local dev and sidesteps a
+  CDP-context quirk where HttpOnly cookies sometimes fail to attach.
+- The state file is kept at `~/.lobehub-agent-testing/web-state.json` so
+  `setup-auth.sh status` can report web-auth readiness across sessions.
+
+### Common failure modes
+
+| Symptom                                       | Cause                                                                     | Fix                                               |
+| --------------------------------------------- | ------------------------------------------------------------------------- | ------------------------------------------------- |
+| Still redirects to `/signin` after injection  | User pasted from `document.cookie` → missed HttpOnly session              | Re-pull from Network request Headers, not console |
+| Script reports `no better-auth cookies found` | Separator wrong, or user pasted URL-decoded value                         | Keep the raw `Cookie:` header as-is               |
+| Login works briefly then expires              | `better-auth.session_token` rotated (user logged out / signed in again)   | Re-copy and re-inject                             |
+| Domain mismatch                               | Cookie domain must be `localhost` literally, no leading dot for local dev | —                                                 |
+
+## Electron
+
+The desktop app keeps its own persistent login state in its user-data
+directory — log in once manually inside the app and it survives restarts of
+`electron-dev.sh`. No injection needed. The standard check (do NOT hand-roll a
+store eval) once Electron is up with CDP:
+
+```bash
+./.agents/skills/agent-testing/scripts/app-probe.sh auth
+# → {"ok":true,"isSignedIn":true,"userId":"user_xxx"}
+```
+
+`setup-auth.sh status` runs this probe automatically when CDP 9222 is
+reachable.
+
+## Scope
+
+These recipes only cover **local dev** authentication. They do not:
+
+- Work for production — production cookies are `Secure; HttpOnly; Domain=.lobehub.com`
+  and must be delivered over HTTPS.
+- Replace real OAuth flows — tests that must exercise the login UI itself need a
+  real Chromium with `--remote-debugging-port` or a bot account.
+- Flow cookies back to the user's Chrome — injection is one-way.
@@ -0,0 +1,55 @@
+# Local Dev Server
+
+Single source of truth for starting / restarting the backend that all test
+surfaces (CLI, Electron, Web) hit.
+
+## Ports & modes
+
+| Command             | What it runs                                              | Port                              |
+| ------------------- | --------------------------------------------------------- | --------------------------------- |
+| `pnpm run dev:next` | Next.js backend (API + auth)                              | `3010`                            |
+| `bun run dev`       | Full-stack (Next.js + Vite SPA, via `devStartupSequence`) | `3010` (API) + SPA                |
+| `bun run dev:spa`   | Vite SPA only, proxies API to `3010`                      | `9876` (prints a Debug Proxy URL) |
+
+In the **cloud repo** (where this repo is the `lobehub/` submodule) the dev
+server conventionally runs on `3011` — set `SERVER_URL=http://localhost:3011`
+for the scripts in this skill when testing there.
+
+## Health check
+
+```bash
+curl -s -o /dev/null -w '%{http_code}' http://localhost:3010/
+```
+
+## Start / restart
+
+```bash
+# Start (from repo root)
+pnpm run dev:next
+
+# Restart — required to pick up server-side code changes
+lsof -ti:3010 | xargs kill
+pnpm run dev:next
+```
+
+## When a server restart is needed
+
+Next.js hot-reload may not pick up changes in workspace packages — restart when
+in doubt.
+
+| Change location                                 | Restart? |
+| ----------------------------------------------- | -------- |
+| `apps/server/src/` (routers, services, modules) | Yes      |
+| `src/server/` (agent-hono, workflows-hono)      | Yes      |
+| `packages/database/` (models)                   | Yes      |
+| `packages/types/`                               | Yes      |
+| `packages/prompts/`                             | Yes      |
+| `apps/cli/` (CLI runs from source)              | No       |
+
+## Troubleshooting
+
+| Issue                     | Solution                                                |
+| ------------------------- | ------------------------------------------------------- |
+| `ECONNREFUSED`            | Server not running — start it                           |
+| `EADDRINUSE` on the port  | Already running — `lsof -ti:<port> \| xargs kill` first |
+| Stale data / old behavior | Server needs a restart to pick up code changes          |
@@ -12,13 +12,13 @@ General-purpose screen recording tool for the Electron app. Captures CDP screens

 ```bash
 # Start recording (Electron must be running with CDP)
-.agents/skills/local-testing/scripts/record-app-screen.sh start [output_name]
+.agents/skills/agent-testing/scripts/record-app-screen.sh start [output_name]

 # Stop recording and assemble video
-.agents/skills/local-testing/scripts/record-app-screen.sh stop
+.agents/skills/agent-testing/scripts/record-app-screen.sh stop

 # Check if recording is active
-.agents/skills/local-testing/scripts/record-app-screen.sh status
+.agents/skills/agent-testing/scripts/record-app-screen.sh status
 ```

 ### Arguments
@@ -74,10 +74,10 @@ The `.records/` directory is at the project root and is gitignored.

 ```bash
 # Start Electron
-.agents/skills/local-testing/scripts/electron-dev.sh start
+.agents/skills/agent-testing/scripts/electron-dev.sh start

 # Start recording
-.agents/skills/local-testing/scripts/record-app-screen.sh start my-test
+.agents/skills/agent-testing/scripts/record-app-screen.sh start my-test

 # Run automation
 agent-browser --cdp 9222 click @e61
@@ -86,14 +86,14 @@ agent-browser --cdp 9222 press Enter
 sleep 10

 # Stop and get results
-.agents/skills/local-testing/scripts/record-app-screen.sh stop
+.agents/skills/agent-testing/scripts/record-app-screen.sh stop
 # → .records/my-test.mp4 + .records/my-test/*.png
 ```

 ### Gateway Streaming Demo

 ```bash
-.agents/skills/local-testing/scripts/electron-dev.sh start
+.agents/skills/agent-testing/scripts/electron-dev.sh start

 # Inject gateway URL
 agent-browser --cdp 9222 eval --stdin << 'EOF'
@@ -106,19 +106,19 @@ agent-browser --cdp 9222 eval --stdin << 'EOF'
 EOF

 # Record
-.agents/skills/local-testing/scripts/record-app-screen.sh start gateway-demo
+.agents/skills/agent-testing/scripts/record-app-screen.sh start gateway-demo

 # Navigate to agent, send message, wait for completion...
 # (automation commands here)

-.agents/skills/local-testing/scripts/record-app-screen.sh stop
+.agents/skills/agent-testing/scripts/record-app-screen.sh stop
 open .records/gateway-demo.mp4
 ```

 ### Check Active Recording

 ```bash
-.agents/skills/local-testing/scripts/record-app-screen.sh status
+.agents/skills/agent-testing/scripts/record-app-screen.sh status
 # [record] Active recording
 #   Frames:      42 captured (running: yes)
 #   Screenshots: 14 captured (running: yes)
@@ -0,0 +1,156 @@
+# Structured Test Reports
+
+Every automated test session ends with a structured, evidence-backed report.
+A chat-only summary is not an acceptable deliverable: the report is what the
+user (or a reviewer, or a later agent) audits without replaying the session.
+
+## Location & layout
+
+Reports live under `.records/reports/` (gitignored, like all `.records/`
+output):
+
+```
+.records/reports/<YYYYMMDD-HHMMSS>-<slug>/
+├── report.md      # human-readable report (case table with inline screenshots, verdict)
+├── result.json    # machine-readable results (pass/fail counts, score)
+└── assets/        # evidence: screenshots, HAR files, CLI transcripts
+```
+
+## Workflow
+
+1. **Scaffold up front** — before running the first test step:
+
+   ```bash
+   DIR=$(./.agents/skills/agent-testing/scripts/report-init.sh < slug > "<title>")
+   ```
+
+   The script creates the directory, pre-fills branch / commit / date in both
+   files, and prints the directory path. The scaffold uses the compact report
+   shape below; translate its headings and table labels to the user's language
+   before delivery if needed.
+
+2. **Collect evidence as you test** — every asserted behavior gets one evidence
+   item in `$DIR/assets/`:
+   - UI (static state): `agent-browser screenshot` or `capture-app-window.sh`,
+     then **verify the screenshot with the Read tool before citing it** —
+     never cite an image you haven't looked at.
+   - UI (time-based behavior): **screenshot vs GIF is a judgment you must
+     make per case.** If the assertion is about change over time — streaming
+     output, a ticking timer, loading/progress states, animations,
+     appear/disappear transitions — a static screenshot cannot prove it.
+     Record a frame sequence and synthesize a GIF:
+
+     ```bash
+     # start recording (background), trigger the behavior, wait for it to finish
+     ../scripts/record-gif.sh "$DIR/assets/case2-streaming.gif" 12 2 &
+     GIF_PID=$!
+     # ... drive the scenario ...
+     wait $GIF_PID
+     ```
+
+     Embed it like an image: `![case 2](assets/case2-streaming.gif)`. Verify
+     at least the first/last frames visually (Read the GIF) before citing.
+
+   - CLI: exact command + trimmed output (`$CLI task list | tee "$DIR/assets/task-list.txt"`).
+   - Network: `agent-browser network requests` dumps or HAR files.
+
+3. **Fill `report.md` as you go** — don't reconstruct from memory at the end.
+   The primary evidence belongs in the case table itself: each row should pair
+   the assertion with the screenshot/GIF/link that proves it, so readers can
+   scan the result without jumping between sections.
+
+4. **Set the verdict** in both `report.md` and `result.json`, then link the
+   report directory in your final answer to the user.
+
+## Report language (hard rule)
+
+**`report.md` MUST be written in the language the user is conversing in** —
+the whole file, headings included. If the conversation is in Chinese, the
+report is in Chinese; do not mix English prose into it. The scaffold headings
+are placeholders — translate them when filling if the user is not conversing in
+the scaffold language. Exceptions that stay as-is: code/commands, identifiers,
+log excerpts, and `result.json` (its keys and status values are machine-read
+and stay English; the `title` and case `name` fields follow the user's
+language).
+
+## report.md sections
+
+Default report shape:
+
+| Section          | Content                                                                                      |
+| ---------------- | -------------------------------------------------------------------------------------------- |
+| **Scope**        | What changed / what is being verified; branch, commit, date, surface, entry URL/page, focus  |
+| **Cases**        | Compact table: `# \| Case \| Result \| Key observation \| Evidence`                          |
+| **Verdict**      | Overall verdict first (`pass` / `partial` / `fail`), then the concise reasons and follow-ups |
+| **Verification** | Commands or automated checks run in this session, with trimmed results                       |
+| **Score**        | Pass/fail/blocked counts, optional 0–100 score                                               |
+
+The case table is the main reading surface. Prefer one clear row per user
+scenario or regression assertion, and put the screenshot/GIF directly in the
+`Evidence` cell:
+
+```markdown
+| #   | Case                     | Result | Key observation                                                   | Evidence                                         |
+| --- | ------------------------ | ------ | ----------------------------------------------------------------- | ------------------------------------------------ |
+| 1   | Create a new page        | pass   | Title and body persisted after refresh                            | ![created page](assets/new-page-created.png)     |
+| 2   | Respect requested length | fail   | Requested about 600 Chinese characters; final body was about 1286 | ![final article](assets/write-article-final.png) |
+```
+
+Avoid the old wide table with separate `steps`, `expected`, and `actual`
+columns unless the test is purely non-visual and truly needs that breakdown.
+For UI reports, those columns make screenshot-backed reading harder. Put
+procedural detail in the row's key observation only when it changes the
+interpretation of the result.
+
+Use an extra evidence/detail section only when the inline table cannot carry
+the material cleanly, such as long CLI transcripts, HAR summaries, or multiple
+screenshots for one case. In that situation, keep the table evidence cell as a
+short inline proof or link, then put the longer material under `Verification`
+or a brief `Additional Evidence` section.
+
+Status values: `pass` / `fail` / `blocked` (couldn't run — e.g. auth or env
+missing; a blocked case is not a pass).
+
+## result.json schema
+
+```json
+{
+  "branch": "feat/task-tree",
+  "cases": [
+    {
+      "id": "1",
+      "name": "task tree returns nested children",
+      "surface": "cli",
+      "status": "pass",
+      "evidence": ["assets/task-tree.txt"]
+    }
+  ],
+  "commit": "abc1234",
+  "createdAt": "2026-06-11T15:30:00+08:00",
+  "summary": {
+    "total": 1,
+    "passed": 1,
+    "failed": 0,
+    "blocked": 0,
+    "score": 100,
+    "verdict": "pass"
+  },
+  "surfaces": ["cli"],
+  "title": "Verify task tree API"
+}
+```
+
+`score` is optional — use it when the verdict has a subjective component (UI
+polish, copy quality); omit it for purely binary runs. `verdict` is the single
+word the user reads first: `pass`, `fail`, or `partial`.
+
+## Rules
+
+- **No evidence, no claim** — every `pass`/`fail` in the case table must link
+  at least one asset.
+- **Screenshots must be visually verified** with the Read tool before being
+  cited.
+- **Report failures faithfully** — a failing case with clear evidence is a good
+  report; a vague green one is not.
+- If coverage was cut (cases skipped, surfaces not exercised), say so in the
+  Verdict section — silent truncation reads as "covered everything".
@@ -11,7 +11,7 @@
 //   6. ROLLBACKS — msgN / childN / role drops in the active-topic timeline
 //
 // Usage:
-//   bun run .agents/skills/local-testing/scripts/agent-gateway/analyze-events.ts <dump.json>
+//   bun run .agents/skills/agent-testing/scripts/agent-gateway/analyze-events.ts <dump.json>

 import { readFileSync } from 'node:fs';

@@ -5,16 +5,16 @@
 // streaming-replay test fixtures.
 //
 // Commands:
-//   bun run .agents/skills/local-testing/scripts/agent-gateway/run.ts install
+//   bun run .agents/skills/agent-testing/scripts/agent-gateway/run.ts install
 //       Bundle probe-events.ts and inject into the CDP-attached browser.
 //       Re-installing clears all buffers and re-patches WebSocket / fetch.
 //
-//   bun run .agents/skills/local-testing/scripts/agent-gateway/run.ts dump [name]
+//   bun run .agents/skills/agent-testing/scripts/agent-gateway/run.ts dump [name]
 //       Stop the timeline timer, fetch the capture as JSON, write it to
 //       `.agent-gateway/<name>-<YYYYMMDD-HHmmss>.json`. `name` defaults to
 //       `dump`. Prints the absolute path written.
 //
-//   bun run .agents/skills/local-testing/scripts/agent-gateway/run.ts analyze [path]
+//   bun run .agents/skills/agent-testing/scripts/agent-gateway/run.ts analyze [path]
 //       Run analyze-events.ts on the dump. `path` defaults to the most
 //       recently modified file in `.agent-gateway/`.
 //
@@ -28,7 +28,7 @@ import path from 'node:path';
 import { fileURLToPath } from 'node:url';

 const SCRIPT_DIR = path.dirname(fileURLToPath(import.meta.url));
-// .agents/skills/local-testing/scripts/agent-gateway/ → 5 levels up
+// .agents/skills/agent-testing/scripts/agent-gateway/ → 5 levels up
 const PROJECT_ROOT = path.resolve(SCRIPT_DIR, '../../../../..');
 const DUMP_DIR = path.join(PROJECT_ROOT, '.agent-gateway');

@@ -0,0 +1,95 @@
+#!/usr/bin/env bash
+# app-probe.sh — standardized probes for a running LobeHub app (Electron via
+# CDP, or a web agent-browser session). Use these instead of hand-rolling
+# `window.__LOBE_STORES` eval snippets — especially the auth check.
+#
+# Usage:
+#   app-probe.sh auth              # { isSignedIn, userId } from the user store
+#   app-probe.sh route             # current SPA route
+#   app-probe.sh ops               # running chat operations (type / status / startTime)
+#   app-probe.sh goto <path>       # navigate the SPA to a route (full reload), e.g. goto /agent/agt_xxx
+#   app-probe.sh errors-install    # install a console.error interceptor
+#   app-probe.sh errors            # dump errors captured since errors-install
+#
+# Target selection (default: Electron over CDP 9222):
+#   AB_TARGET="--cdp 9222"             # Electron (default; CDP_PORT also honored)
+#   AB_TARGET="--session lobehub-dev"  # web agent-browser session
+#
+# Common routes (desktop SPA): /  /agent/<agentId>  /agent/<agentId>/<topicId>
+#   /task  /task/<taskId>  /page  /settings  /community
+
+set -euo pipefail
+
+AB_TARGET="${AB_TARGET:---cdp ${CDP_PORT:-9222}}"
+
+run_eval() {
+  # shellcheck disable=SC2086
+  agent-browser $AB_TARGET eval --stdin
+}
+
+case "${1:-}" in
+  auth)
+    run_eval << 'EVALEOF'
+(function () {
+  var stores = window.__LOBE_STORES;
+  if (!stores || !stores.user) return JSON.stringify({ ok: false, reason: 'no user store — app not loaded yet?' });
+  var u = stores.user();
+  return JSON.stringify({ ok: !!u.isSignedIn, isSignedIn: !!u.isSignedIn, userId: (u.user && u.user.id) || null });
+})()
+EVALEOF
+    ;;
+  route)
+    run_eval << 'EVALEOF'
+location.pathname + location.search + location.hash
+EVALEOF
+    ;;
+  ops)
+    run_eval << 'EVALEOF'
+(function () {
+  var stores = window.__LOBE_STORES;
+  if (!stores || !stores.chat) return JSON.stringify({ ok: false, reason: 'no chat store — open a conversation first' });
+  var ops = Object.values(stores.chat().operations || {});
+  var running = ops.filter(function (o) { return o.status === 'running'; });
+  return JSON.stringify({
+    ok: true,
+    running: running.map(function (o) { return { startTime: o.metadata && o.metadata.startTime, type: o.type }; }),
+    runningCount: running.length,
+    total: ops.length,
+  });
+})()
+EVALEOF
+    ;;
+  goto)
+    TARGET_PATH="${2:?Usage: app-probe.sh goto <path>}"
+    # shellcheck disable=SC2086
+    agent-browser $AB_TARGET eval "location.href = '$TARGET_PATH'" > /dev/null
+    sleep 2
+    bash "${BASH_SOURCE[0]}" route
+    ;;
+  errors-install)
+    run_eval << 'EVALEOF'
+(function () {
+  window.__CAPTURED_ERRORS = [];
+  var orig = console.error;
+  console.error = function () {
+    var msg = Array.from(arguments).map(function (a) {
+      if (a instanceof Error) return a.message;
+      return typeof a === 'object' ? JSON.stringify(a) : String(a);
+    }).join(' ');
+    window.__CAPTURED_ERRORS.push(msg);
+    orig.apply(console, arguments);
+  };
+  return 'installed';
+})()
+EVALEOF
+    ;;
+  errors)
+    run_eval << 'EVALEOF'
+JSON.stringify(window.__CAPTURED_ERRORS || 'interceptor not installed — run errors-install first')
+EVALEOF
+    ;;
+  *)
+    echo "Usage: $0 {auth|route|ops|goto <path>|errors-install|errors}" >&2
+    exit 2
+    ;;
+esac
@@ -0,0 +1,61 @@
+#!/usr/bin/env bash
+# record-gif.sh — capture a frame sequence via agent-browser (CDP) and
+# synthesize a GIF for embedding in a test report.
+#
+# Use this whenever the asserted behavior is about CHANGE OVER TIME —
+# streaming output, a ticking timer, loading states, animations. A static
+# screenshot cannot prove those; a GIF can. Cloud-portable: frames come from
+# CDP rendering, no OS-level screen capture.
+#
+# Usage:
+#   record-gif.sh <output.gif> <duration_seconds> [fps]
+#
+#   AB_TARGET="--cdp 9222"             # Electron (default; CDP_PORT honored)
+#   AB_TARGET="--session lobehub-dev"  # web agent-browser session
+#   GIF_WIDTH=960                      # output width (px), default 960
+#
+# Requires ffmpeg (`brew install ffmpeg`). Effective fps is capped by
+# screenshot latency (~0.3-0.5s per frame); 1-2 fps is the realistic range.
+#
+# Example — record a 12s run and embed it in the report:
+#   ./record-gif.sh "$DIR/assets/case2-tray-running.gif" 12 2 &
+#   GIF_PID=$!
+#   # ... trigger the streaming behavior ...
+#   wait $GIF_PID
+
+set -euo pipefail
+
+OUT="${1:?Usage: record-gif.sh <output.gif> <duration_seconds> [fps]}"
+DUR="${2:?Usage: record-gif.sh <output.gif> <duration_seconds> [fps]}"
+FPS="${3:-2}"
+AB_TARGET="${AB_TARGET:---cdp ${CDP_PORT:-9222}}"
+GIF_WIDTH="${GIF_WIDTH:-960}"
+
+command -v ffmpeg > /dev/null || {
+  echo "ffmpeg not found — install with: brew install ffmpeg" >&2
+  exit 1
+}
+
+TMP=$(mktemp -d)
+trap 'rm -rf "$TMP"' EXIT
+
+FRAMES=$((DUR * FPS))
+INTERVAL=$(python3 -c "print(1 / $FPS)")
+
+for i in $(seq -f '%04g' 1 "$FRAMES"); do
+  # shellcheck disable=SC2086
+  agent-browser $AB_TARGET screenshot "$TMP/frame-$i.png" > /dev/null 2>&1 || true
+  sleep "$INTERVAL"
+done
+
+CAPTURED=$(find "$TMP" -name 'frame-*.png' | wc -l | tr -d ' ')
+[ "$CAPTURED" -gt 0 ] || {
+  echo "no frames captured — is the app reachable via $AB_TARGET?" >&2
+  exit 1
+}
+
+ffmpeg -y -loglevel error -framerate "$FPS" -pattern_type glob -i "$TMP/frame-*.png" \
+  -vf "fps=$FPS,scale=$GIF_WIDTH:-1:flags=lanczos,split[s0][s1];[s0]palettegen[p];[s1][p]paletteuse" \
+  "$OUT"
+
+echo "$OUT ($CAPTURED frames @ ${FPS}fps)"
@@ -0,0 +1,88 @@
+#!/usr/bin/env bash
+# report-init.sh — scaffold a structured test report under .records/reports/.
+#
+# Format spec and evidence rules: ../references/report.md
+#
+# Usage:
+#   report-init.sh <slug> [title]
+#
+# Prints the report directory path (capture it: DIR=$(report-init.sh my-test)).
+
+set -euo pipefail
+
+SLUG="${1:?Usage: report-init.sh <slug> [title]}"
+TITLE="${2:-$SLUG}"
+
+REPO_ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/../../../.." && pwd)"
+TS="$(date +%Y%m%d-%H%M%S)"
+DIR="$REPO_ROOT/.records/reports/$TS-$SLUG"
+mkdir -p "$DIR/assets"
+
+BRANCH=$(git -C "$REPO_ROOT" branch --show-current 2> /dev/null || echo "unknown")
+COMMIT=$(git -C "$REPO_ROOT" rev-parse --short HEAD 2> /dev/null || echo "unknown")
+DATE_HUMAN=$(date '+%Y-%m-%d %H:%M')
+DATE_ISO=$(date '+%Y-%m-%dT%H:%M:%S%z')
+
+cat > "$DIR/report.md" << EOF
+# 测试报告：$TITLE
+
+## 范围
+
+<!-- 测试目标 / 变更范围 / 重点风险 -->
+
+- 分支：\`$BRANCH\`
+- 当前提交：\`$COMMIT\`
+- 日期：$DATE_HUMAN
+- 表面：<!-- CLI / Electron + CDP / Web / Bot:<platform> -->
+- 测试页 / 入口：<!-- e.g. /settings or http://localhost:3010 -->
+- 重点：<!-- 本轮最关心的体验、功能或回归点 -->
+
+## 用例
+
+| # | 用例 | 结果 | 关键现象 | 证据 |
+| - | ---- | ---- | -------- | ---- |
+| 1 |      | 待测 |          | ![用例 1](assets/case1.png) |
+
+## 结论
+
+整体结论：\`pending\`。
+
+<!-- 用 1-2 段概括用户最需要知道的结果；失败和阻塞必须明确说明影响。 -->
+
+仍需处理 / 跟进：
+
+- <!-- TODO -->
+
+## 本轮验证
+
+<!-- 如有自动化或命令行验证，保留精简命令与结果；没有则写“未运行额外自动化验证”。 -->
+
+\`\`\`bash
+# command
+\`\`\`
+
+结果：
+
+- <!-- TODO -->
+
+## 评分
+
+- 通过：0
+- 失败：0
+- 阻塞：0
+- 评分：— / 100
+EOF
+
+cat > "$DIR/result.json" << EOF
+{
+  "title": "$TITLE",
+  "createdAt": "$DATE_ISO",
+  "branch": "$BRANCH",
+  "commit": "$COMMIT",
+  "surfaces": [],
+  "cases": [],
+  "summary": { "total": 0, "passed": 0, "failed": 0, "blocked": 0, "verdict": "pending" }
+}
+EOF
+
+echo "$DIR"
@@ -0,0 +1,174 @@
+#!/usr/bin/env bash
+# setup-auth.sh — one-stop auth setup & check for local agent testing.
+#
+# Auth is the gate for all automated testing: prepare it BEFORE writing any
+# test step. Background and failure modes: ../references/auth.md
+#
+# Usage:
+#   setup-auth.sh status        # check server + CLI + web auth readiness
+#   setup-auth.sh cli           # interactive CLI device-code login (run by a human)
+#   setup-auth.sh web           # stdin = Cookie header -> inject into agent-browser session
+#   setup-auth.sh web-verify    # live-check the agent-browser session is authenticated
+#
+# Env:
+#   SERVER_URL  (default http://localhost:3010)   dev server under test
+#   SESSION     (default lobehub-dev)             agent-browser session name
+#   AUTH_DIR    (default ~/.lobehub-agent-testing) where web state is persisted
+
+set -euo pipefail
+
+SERVER_URL="${SERVER_URL:-http://localhost:3010}"
+SESSION="${SESSION:-lobehub-dev}"
+AUTH_DIR="${AUTH_DIR:-$HOME/.lobehub-agent-testing}"
+STATE_FILE="$AUTH_DIR/web-state.json"
+REPO_ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/../../../.." && pwd)"
+CLI_HOME="$REPO_ROOT/apps/cli/.lobehub-dev"
+
+ok()   { printf '  \033[32m✔\033[0m %s\n' "$1"; }
+bad()  { printf '  \033[31m✘\033[0m %s\n' "$1"; }
+note() { printf '      %s\n' "$1"; }
+
+check_server() {
+  local code
+  code=$(curl -s -o /dev/null -w '%{http_code}' "$SERVER_URL/" 2> /dev/null || true)
+  if [[ "$code" =~ ^[23] ]]; then
+    ok "dev server reachable at $SERVER_URL"
+  else
+    bad "dev server NOT reachable at $SERVER_URL (http_code='$code')"
+    note "start it: pnpm run dev:next  (see references/dev-server.md)"
+    return 1
+  fi
+}
+
+check_cli() {
+  if [[ -f "$CLI_HOME/settings.json" ]] && grep -q "$SERVER_URL" "$CLI_HOME/settings.json"; then
+    ok "CLI logged in to $SERVER_URL (creds: apps/cli/.lobehub-dev)"
+  else
+    bad "CLI not logged in to $SERVER_URL"
+    note "ask the user to run:"
+    note "cd apps/cli && LOBEHUB_CLI_HOME=.lobehub-dev bun src/index.ts login --server $SERVER_URL"
+    return 1
+  fi
+}
+
+check_web() {
+  if [[ -f "$STATE_FILE" ]]; then
+    ok "web auth state saved ($STATE_FILE)"
+    note "live-verify: $0 web-verify"
+  else
+    bad "no web auth state for agent-browser"
+    note "copy the Cookie header from Chrome DevTools (Network tab), then:"
+    note "pbpaste | $0 web   (see references/auth.md)"
+    return 1
+  fi
+}
+
+check_electron() {
+  local cdp_port="${CDP_PORT:-9222}"
+  if ! curl -s -o /dev/null --max-time 2 "http://localhost:$cdp_port/json/version" 2> /dev/null; then
+    note "electron: not running (CDP $cdp_port unreachable) — start with electron-dev.sh; check skipped"
+    return 0
+  fi
+  local probe result
+  probe="$(dirname "${BASH_SOURCE[0]}")/app-probe.sh"
+  result=$(bash "$probe" auth 2> /dev/null || true)
+  # agent-browser eval returns the JSON string with escaped quotes — normalize.
+  result="${result//\\/}"
+  if [[ "$result" == *'"isSignedIn":true'* ]]; then
+    ok "electron app signed in ($result)"
+  else
+    bad "electron app NOT signed in ($result)"
+    note "log in once manually inside the app (state persists across restarts)"
+    return 1
+  fi
+}
+
+cmd_status() {
+  echo "agent-testing auth status (SERVER_URL=$SERVER_URL):"
+  local rc=0
+  check_server || rc=1
+  check_cli || rc=1
+  check_web || rc=1
+  check_electron || rc=1
+  if [[ $rc -eq 0 ]]; then
+    echo "all green — safe to start automated testing."
+  else
+    echo "auth NOT ready — fix the ✘ items before writing any test step."
+  fi
+  return $rc
+}
+
+cmd_cli() {
+  echo "Starting CLI device-code login against $SERVER_URL ..."
+  echo "(opens a browser authorization — must be run by a human in a terminal)"
+  cd "$REPO_ROOT/apps/cli"
+  LOBEHUB_CLI_HOME=.lobehub-dev bun src/index.ts login --server "$SERVER_URL"
+}
+
+# Build a Playwright storageState file from a raw Cookie header on stdin,
+# keeping only the better-auth cookies. See references/auth.md for why the
+# header must come from a Network request (HttpOnly) and why httpOnly=false.
+cmd_web() {
+  mkdir -p "$AUTH_DIR"
+  python3 - "$STATE_FILE" << 'PY'
+import json, sys, time
+
+raw = sys.stdin.read().strip()
+if raw.lower().startswith("cookie:"):
+    raw = raw.split(":", 1)[1].strip()
+
+WANTED = {"better-auth.session_token", "better-auth.state"}
+exp = int(time.time()) + 30 * 24 * 3600  # 30 days
+
+cookies = []
+for pair in raw.split("; "):
+    if "=" not in pair:
+        continue
+    name, _, value = pair.partition("=")
+    if name not in WANTED:
+        continue
+    cookies.append({
+        "name": name,
+        "value": value,
+        "domain": "localhost",
+        "path": "/",
+        "expires": exp,
+        "httpOnly": False,
+        "secure": False,
+        "sameSite": "Lax",
+    })
+
+if not cookies:
+    sys.stderr.write("no better-auth cookies found in input — paste the raw Cookie header from a Network request\n")
+    sys.exit(1)
+
+with open(sys.argv[1], "w") as f:
+    json.dump({"cookies": cookies, "origins": []}, f, indent=2)
+print(f"wrote {len(cookies)} cookie(s) to {sys.argv[1]}")
+PY
+  agent-browser --session "$SESSION" state load "$STATE_FILE"
+  cmd_web_verify
+}
+
+cmd_web_verify() {
+  agent-browser --session "$SESSION" open "$SERVER_URL/" > /dev/null
+  local url
+  url=$(agent-browser --session "$SESSION" get url)
+  if [[ "$url" == *"/signin"* || "$url" == *"/login"* ]]; then
+    bad "agent-browser session '$SESSION' NOT authenticated (landed on $url)"
+    note "re-copy the Cookie header and re-run: pbpaste | $0 web"
+    return 1
+  fi
+  ok "agent-browser session '$SESSION' authenticated (at $url)"
+}
+
+case "${1:-status}" in
+  status) cmd_status ;;
+  cli) cmd_cli ;;
+  web) cmd_web ;;
+  web-verify) cmd_web_verify ;;
+  *)
+    echo "Usage: $0 {status|cli|web|web-verify}" >&2
+    exit 2
+    ;;
+esac
@@ -0,0 +1,154 @@
+# Electron (LobeHub Desktop) UI Testing
+
+Default surface for verifying **pure frontend changes** (components, store logic, styles, interactions) in the primary product shape. Drives the Electron renderer over CDP with `agent-browser` — see [../references/agent-browser.md](../references/agent-browser.md) for the full command reference.
+
+**Auth**: the Electron app keeps its own persistent login state — log in once manually in the app; sessions survive restarts. Run `../scripts/setup-auth.sh status` before testing (see [../references/auth.md](../references/auth.md)).
+
+**Linux / headless (cloud)**: Electron itself runs on Linux, but it has no true headless mode — it needs a display server. In a headless environment wrap the launch with `xvfb-run` (virtual framebuffer). Everything CDP-based keeps working under Xvfb: the `agent-browser --cdp 9222` connection, snapshots, eval, and `agent-browser screenshot` (captured from the renderer via CDP, not the OS screen). What does NOT work on Linux: `capture-app-window.sh` (macOS `screencapture`), osascript, and the ffmpeg recording scripts in their current form.
+
+### Setup / Teardown
+
+Use the `electron-dev.sh` script to manage the Electron dev environment. It handles process lifecycle, waits for SPA readiness, and reliably kills all child processes (main + helpers + vite).
+
+```bash
+SCRIPT=".agents/skills/agent-testing/scripts/electron-dev.sh"
+
+# Start Electron dev with CDP (idempotent — skips if already running)
+$SCRIPT start
+
+# Check if Electron is running and CDP is reachable
+$SCRIPT status
+
+# Kill all Electron-related processes (main + helper + vite)
+$SCRIPT stop
+
+# Force fresh restart
+$SCRIPT restart
+```
+
+After `start` succeeds, connect with: `agent-browser --cdp 9222 snapshot -i`
+
+**Always run `$SCRIPT stop` when done testing** — `pkill -f "Electron"` alone won't catch all helper processes.
+
+#### Environment Variables
+
+| Variable          | Default                 | Description                              |
+| ----------------- | ----------------------- | ---------------------------------------- |
+| `CDP_PORT`        | `9222`                  | Chrome DevTools Protocol port            |
+| `ELECTRON_LOG`    | `/tmp/electron-dev.log` | Electron process log                     |
+| `ELECTRON_WAIT_S` | `60`                    | Max seconds to wait for Electron process |
+| `RENDERER_WAIT_S` | `60`                    | Max seconds to wait for SPA to load      |
+
+### LobeHub Probes & Quick Navigation
+
+`scripts/app-probe.sh` is the standard fast path into app state — **use it
+instead of hand-rolling `__LOBE_STORES` eval snippets** for these common needs:
+
+```bash
+PROBE=".agents/skills/agent-testing/scripts/app-probe.sh"
+
+$PROBE auth              # login check (Step 0.3) → { isSignedIn, userId }
+$PROBE route             # current SPA route
+$PROBE ops               # running chat operations (type / startTime)
+$PROBE goto /settings    # jump the SPA straight to a route (full reload)
+$PROBE errors-install    # install console.error interceptor
+$PROBE errors            # dump captured errors
+```
+
+`goto` lets a test enter the state under test directly instead of clicking
+through the UI. Common desktop routes:
+
+| Route                         | Where it lands                       |
+| ----------------------------- | ------------------------------------ |
+| `/`                           | Home (has a chat input)              |
+| `/agent/<agentId>`            | Agent conversation (latest topic)    |
+| `/agent/<agentId>/<topicId>`  | Specific topic in a conversation     |
+| `/task` · `/task/<taskId>`    | Task list / task detail              |
+| `/page`                       | Documents (文稿)                     |
+| `/settings`                   | Settings                             |
+| `/community`                  | Discover / community                 |
+
+Targets default to Electron (`--cdp 9222`); set `AB_TARGET="--session <name>"`
+for web sessions. For deeper or one-off state inspection, fall back to raw
+eval below.
+
+### LobeHub-Specific Patterns
+
+#### Access Zustand Store State
+
+```bash
+agent-browser --cdp 9222 eval --stdin << 'EVALEOF'
+(function() {
+  var chat = window.__LOBE_STORES.chat();
+  var ops = Object.values(chat.operations);
+  return JSON.stringify({
+    ops: ops.map(function(o) { return { type: o.type, status: o.status }; }),
+    activeAgent: chat.activeAgentId,
+    activeTopic: chat.activeTopicId,
+  });
+})()
+EVALEOF
+```
+
+#### Find and Use the Chat Input
+
+```bash
+# The chat input is contenteditable — must use -C flag
+agent-browser --cdp 9222 snapshot -i -C 2>&1 | grep "editable"
+
+agent-browser --cdp 9222 click @e48
+agent-browser --cdp 9222 type @e48 "Hello world"
+agent-browser --cdp 9222 press Enter
+```
+
+#### Wait for Agent to Complete
+
+```bash
+agent-browser --cdp 9222 eval --stdin << 'EVALEOF'
+(function() {
+  var chat = window.__LOBE_STORES.chat();
+  var ops = Object.values(chat.operations);
+  var running = ops.filter(function(o) { return o.status === 'running'; });
+  return running.length === 0 ? 'done' : 'running: ' + running.length;
+})()
+EVALEOF
+```
+
+#### Install Error Interceptor
+
+```bash
+agent-browser --cdp 9222 eval --stdin << 'EVALEOF'
+(function() {
+  window.__CAPTURED_ERRORS = [];
+  var orig = console.error;
+  console.error = function() {
+    var msg = Array.from(arguments).map(function(a) {
+      if (a instanceof Error) return a.message;
+      return typeof a === 'object' ? JSON.stringify(a) : String(a);
+    }).join(' ');
+    window.__CAPTURED_ERRORS.push(msg);
+    orig.apply(console, arguments);
+  };
+  return 'installed';
+})()
+EVALEOF
+
+# Later, check captured errors:
+agent-browser --cdp 9222 eval "JSON.stringify(window.__CAPTURED_ERRORS)"
+```
+
+## Electron Gotchas
+
+- **Always use `electron-dev.sh stop` to clean up** — `pkill -f "Electron"` only kills the main process; helper processes (GPU, renderer, network) survive. The script finds and kills all of them via PID matching against the project's electron binary path.
+- **`npx electron-vite dev` must run from `apps/desktop/`** — running from project root fails silently. The `electron-dev.sh` script handles this automatically.
+- **Dev build auto-opens DevTools, which hijacks the CDP target** — `agent-browser --cdp 9222` may attach to the DevTools page (`devtools://…`) instead of the app (`app://renderer/`). Symptom: `get url` returns a `devtools://` URL. Fix: close the DevTools target and reconnect:
+
+  ```bash
+  DT_ID=$(curl -s http://localhost:9222/json/list | python3 -c "import json,sys; ts=json.load(sys.stdin); print(next(t['id'] for t in ts if t['type']=='page' and t['url'].startswith('devtools://')))")
+  curl -s "http://localhost:9222/json/close/$DT_ID" > /dev/null
+  agent-browser close --all && agent-browser --cdp 9222 get url   # expect app://renderer/
+  ```
+
+- **Don't resize the Electron window after load** — resizing triggers full SPA reload
+- **Store is at `window.__LOBE_STORES`** not `window.__ZUSTAND_STORES__`
+- **Streaming / ticking UI needs GIF evidence** — see `scripts/record-gif.sh`; a static screenshot cannot prove time-based behavior.
@@ -0,0 +1,69 @@
+# Web (Full-Stack) Testing
+
+Default surface for **full-stack changes** — a new/changed API plus the UI that
+consumes it. The browser is the one surface where network requests and UI state
+are observable together, so you can assert both sides of the contract in a
+single run.
+
+For pure-frontend changes prefer [electron.md](./electron.md); for
+backend-only changes prefer [../cli/index.md](../cli/index.md).
+
+## Prerequisites
+
+- Local dev server running — [../references/dev-server.md](../references/dev-server.md)
+- Web auth injected into agent-browser — [../references/auth.md](../references/auth.md):
+
+```bash
+pbpaste | ./.agents/skills/agent-testing/scripts/setup-auth.sh web # after copying the Cookie header
+```
+
+## Option A — agent-browser with injected auth (recommended)
+
+```bash
+SESSION=lobehub-dev
+
+agent-browser --session $SESSION open "http://localhost:3010/"
+agent-browser --session $SESSION snapshot -i
+# interact via refs — full command reference: ../references/agent-browser.md
+```
+
+### Watch the API while driving the UI
+
+```bash
+# After triggering the UI action under test:
+agent-browser --session $SESSION network requests --type xhr,fetch
+agent-browser --session $SESSION network requests --method POST
+
+# Record a full HAR for the report
+agent-browser --session $SESSION network har start
+# ... drive the scenario ...
+agent-browser --session $SESSION network har stop ./capture.har
+```
+
+Assert both layers: the request/response shape (network) and the rendered
+result (snapshot/screenshot). Both belong in the report as evidence.
+
+## Option B — real Chrome with remote debugging
+
+For flows that need a real, visible browser (e.g. exercising the login UI
+itself):
+
+```bash
+/Applications/Google\ Chrome.app/Contents/MacOS/Google\ Chrome \
+  --remote-debugging-port=9222 \
+  --user-data-dir=/tmp/chrome-test-profile \
+  "<URL>" &
+sleep 5
+agent-browser --cdp 9222 snapshot -i
+
+# Or auto-discover running Chrome with remote debugging
+agent-browser --auto-connect snapshot -i
+```
+
+## Option C — Debug Proxy (local frontend, production backend)
+
+`bun run dev:spa` prints a **Debug Proxy** URL
+(`https://app.lobehub.com/_dangerous_local_dev_proxy?debug-host=…`) that loads
+your local Vite SPA inside the online environment — HMR against real server
+config. Useful for verifying frontend behavior against production data, **not**
+for testing backend changes (the backend is production, not your branch).
@@ -1,172 +0,0 @@
---
-name: cli-backend-testing
-description: >
-  CLI + Backend integration testing workflow. Use when verifying backend API changes
-  (TRPC routers, services, models) via the LobeHub CLI against a local dev server.
-  Triggers on 'cli test', 'test with cli', 'verify with cli', 'local cli test',
-  'backend test with cli', or when needing to validate server-side changes end-to-end.
---
-
-# CLI + Backend Integration Testing
-
-Standard workflow for verifying backend changes using the LobeHub CLI (`lh`) against a local dev server.
-
-## When to Use
-
- Verifying TRPC router / service / model changes end-to-end
- Testing new API fields or response structure changes
- Validating CLI command output after backend modifications
- Debugging data flow issues between server and CLI
-
-## Prerequisites
-
-| Requirement  | Details                                                       |
-| ------------ | ------------------------------------------------------------- |
-| Dev server   | `localhost:3011` (Next.js)                                    |
-| CLI source   | `lobehub/apps/cli/`                                           |
-| CLI dev mode | Uses `LOBEHUB_CLI_HOME=.lobehub-dev` for isolated credentials |
-| Auth         | Device Code Flow login to local server                        |
-
-## Quick Reference
-
-All CLI dev commands run from `lobehub/apps/cli/`. Subsequent examples use `$CLI`:
-
-```bash
-CLI="LOBEHUB_CLI_HOME=.lobehub-dev bun src/index.ts"
-```
-
-## Workflow
-
-### Step 1: Ensure Dev Server is Running
-
-```bash
-curl -s -o /dev/null -w '%{http_code}' http://localhost:3011/ 2> /dev/null
-```
-
- **If reachable**: skip to Step 2.
- **If unreachable**: start from cloud repo root:
-
-```bash
-pnpm run dev:next
-```
-
-To **restart** (pick up server-side code changes):
-
-```bash
-lsof -ti:3011 | xargs kill
-pnpm run dev:next
-```
-
-**Important:** Server-side code changes in the submodule (`lobehub/apps/server/src/`, `lobehub/src/server/`, `lobehub/packages/`) require a server restart. Next.js hot-reload may not pick up changes in submodule packages.
-
-### Step 2: Check CLI Authentication
-
-```bash
-cat lobehub/apps/cli/.lobehub-dev/settings.json 2> /dev/null
-```
-
- **If file exists and contains `"serverUrl": "http://localhost:3011"`**: skip to Step 3.
- **If missing or wrong server**: ask the user to run:
-
-```bash
-! cd lobehub/apps/cli && LOBEHUB_CLI_HOME=.lobehub-dev bun src/index.ts login --server http://localhost:3011
-```
-
-> Login requires interactive browser authorization (OIDC Device Code Flow), so the user must run it themselves via `!` prefix. Credentials persist in `lobehub/apps/cli/.lobehub-dev/`.
-
-### Step 3: Test with CLI Commands
-
-CLI runs from source, so CLI-side code changes take effect immediately without rebuilding.
-
-```bash
-cd lobehub/apps/cli
-$CLI <command>
-```
-
-### Step 4: Clean Up Test Data
-
-```bash
-$CLI task delete < id > -y
-$CLI agent delete < id > -y
-```
-
-## Common Testing Patterns
-
-### Task System
-
-```bash
-$CLI task list
-$CLI task create -n "Root Task" -i "Test instruction"
-$CLI task create -n "Child Task" -i "Sub instruction" --parent T-1
-$CLI task view T-1
-$CLI task tree T-1
-$CLI task edit T-1 --status running
-$CLI task comment T-1 -m "Test comment"
-$CLI task delete T-1 -y
-```
-
-### Agent System
-
-```bash
-$CLI agent list
-$CLI agent view <agent-id>
-$CLI agent run <agent-id> -m "Test prompt"
-```
-
-### Document & Knowledge Base
-
-```bash
-$CLI doc list
-$CLI doc create -t "Test Doc" -c "Content here"
-$CLI doc view <doc-id>
-$CLI kb list
-$CLI kb tree <kb-id>
-```
-
-### Model & Provider
-
-```bash
-$CLI model list
-$CLI provider list
-$CLI provider test <provider-id>
-```
-
-## Dev-Test Cycle
-
-```
-1. Make code changes (service/model/router/type)
-         |
-2. Run unit tests (fast feedback)
-   bunx vitest run --silent='passed-only' '<test-file>'
-         |
-3. Restart dev server (if server-side changes)
-   lsof -ti:3011 | xargs kill && pnpm run dev:next
-         |
-4. CLI verification (end-to-end)
-   $CLI <command>
-         |
-5. Clean up test data
-```
-
-### When Server Restart is Needed
-
-| Change Location                                         | Restart? |
-| ------------------------------------------------------- | -------- |
-| `lobehub/apps/server/src/` (routers, services, modules) | Yes      |
-| `lobehub/src/server/` (agent-hono, workflows-hono)      | Yes      |
-| `lobehub/packages/database/` (models)                   | Yes      |
-| `lobehub/packages/types/`                               | Yes      |
-| `lobehub/packages/prompts/`                             | Yes      |
-| `lobehub/apps/cli/` (CLI code)                          | No       |
-| `src/` (cloud overrides)                                | Yes      |
-
-## Troubleshooting
-
-| Issue                       | Solution                                                              |
-| --------------------------- | --------------------------------------------------------------------- |
-| `No authentication found`   | Run `login --server http://localhost:3011`                            |
-| `UNAUTHORIZED` on API calls | Token expired; re-run login                                           |
-| `ECONNREFUSED`              | Dev server not running; start with `pnpm run dev:next`                |
-| CLI shows old data/behavior | Server needs restart to pick up code changes                          |
-| `EADDRINUSE` on port 3011   | Server already running; kill with `lsof -ti:3011 \| xargs kill`       |
-| Login opens wrong server    | Must use `--server http://localhost:3011` flag (env var doesn't work) |
@@ -241,6 +241,6 @@ When the bug comes from a real trace, distill it into the closest existing test
 3. Add or update the narrowest failing test near the broken layer.
 4. Fix the smallest layer that can explain the symptom.
 5. Re-run focused tests.
-6. Only then do an Electron smoke test with the `local-testing` skill if UI confirmation is still needed.
+6. Only then do an Electron smoke test with the `agent-testing` skill if UI confirmation is still needed.

 Do not start with a broad Electron repro if a raw trace or adapter test can prove the fault zone faster.
@@ -1,561 +0,0 @@
---
-name: local-testing
-description: >
-  Local app and bot testing. Uses agent-browser CLI for Electron/web app UI testing,
-  and osascript (AppleScript) for controlling native macOS apps (WeChat, Discord, Telegram, Slack, Lark/飞书, QQ)
-  to test bots. Triggers on 'local test', 'test in electron', 'test desktop', 'test bot',
-  'bot test', 'test in discord', 'test in telegram', 'test in slack', 'test in weixin',
-  'test in wechat', 'test in lark', 'test in feishu', 'test in qq',
-  'manual test', 'osascript', or UI/bot verification tasks.
---
-
-# Local App & Bot Testing
-
-Two approaches for local testing on macOS:
-
-| Approach                    | Tool                | Best For                                             |
-| --------------------------- | ------------------- | ---------------------------------------------------- |
-| **agent-browser + CDP**     | `agent-browser` CLI | Electron apps, web apps (DOM access, JS eval)        |
-| **osascript (AppleScript)** | `osascript -e`      | Native macOS apps (WeChat, Discord, Telegram, Slack) |
-
---
-
-# Part 1: agent-browser (Electron / Web Apps)
-
-Use `agent-browser` to automate Chromium-based apps via Chrome DevTools Protocol.
-
-Install via `npm i -g agent-browser`, `brew install agent-browser`, or `cargo install agent-browser`. Run `agent-browser install` to download Chrome. Run `agent-browser upgrade` to update.
-
-## Core Workflow
-
-Every browser automation follows this pattern:
-
-1. **Navigate**: `agent-browser open <url>`
-2. **Snapshot**: `agent-browser snapshot -i` (get element refs like `@e1`, `@e2`)
-3. **Interact**: Use refs to click, fill, select
-4. **Re-snapshot**: After navigation or DOM changes, get fresh refs
-
-```bash
-agent-browser open https://example.com/form
-agent-browser snapshot -i
-# Output: @e1 [input type="email"], @e2 [input type="password"], @e3 [button] "Submit"
-
-agent-browser fill @e1 "user@example.com"
-agent-browser fill @e2 "password123"
-agent-browser click @e3
-agent-browser wait --load networkidle
-agent-browser snapshot -i # Check result
-```
-
-## Command Chaining
-
-```bash
-# Chain open + wait + snapshot in one call
-agent-browser open https://example.com && agent-browser wait --load networkidle && agent-browser snapshot -i
-```
-
-Use `&&` when you don't need to read intermediate output. Run commands separately when you need to parse output first (e.g., snapshot to discover refs, then interact).
-
-## Essential Commands
-
-```bash
-# Navigation
-agent-browser open <url>              # Navigate (aliases: goto, navigate)
-agent-browser close                   # Close browser
-agent-browser close --all             # Close all active sessions
-
-# Snapshot
-agent-browser snapshot -i             # Interactive elements with refs (recommended)
-agent-browser snapshot -s "#selector" # Scope to CSS selector
-
-# Interaction (use @refs from snapshot)
-agent-browser click @e1               # Click element
-agent-browser click @e1 --new-tab     # Click and open in new tab
-agent-browser fill @e2 "text"         # Clear and type text
-agent-browser type @e2 "text"         # Type without clearing
-agent-browser select @e1 "option"     # Select dropdown option
-agent-browser check @e1               # Check checkbox
-agent-browser press Enter             # Press key
-agent-browser keyboard type "text"    # Type at current focus (no selector)
-agent-browser keyboard inserttext "text"  # Insert without key events
-agent-browser scroll down 500         # Scroll page
-agent-browser scroll down 500 --selector "div.content"  # Scroll within container
-
-# Get information
-agent-browser get text @e1            # Get element text
-agent-browser get url                 # Get current URL
-agent-browser get title               # Get page title
-agent-browser get cdp-url             # Get CDP WebSocket URL
-
-# Wait
-agent-browser wait @e1                # Wait for element
-agent-browser wait --load networkidle # Wait for network idle
-agent-browser wait --url "**/page"    # Wait for URL pattern
-agent-browser wait 2000               # Wait milliseconds
-agent-browser wait --text "Welcome"   # Wait for text to appear
-agent-browser wait --fn "!document.body.innerText.includes('Loading...')"  # Wait for text to disappear
-agent-browser wait "#spinner" --state hidden  # Wait for element to disappear
-
-# Downloads
-agent-browser download @e1 ./file.pdf          # Click element to trigger download
-agent-browser wait --download ./output.zip     # Wait for any download to complete
-
-# Network
-agent-browser network requests                 # Inspect tracked requests
-agent-browser network requests --type xhr,fetch  # Filter by resource type
-agent-browser network requests --method POST   # Filter by HTTP method
-agent-browser network route "**/api/*" --abort # Block matching requests
-agent-browser network har start                # Start HAR recording
-agent-browser network har stop ./capture.har   # Stop and save HAR file
-
-# Viewport & Device Emulation
-agent-browser set viewport 1920 1080          # Set viewport size (default: 1280x720)
-agent-browser set viewport 1920 1080 2        # 2x retina
-agent-browser set device "iPhone 14"          # Emulate device (viewport + user agent)
-
-# Capture
-agent-browser screenshot              # Screenshot to temp dir
-agent-browser screenshot --full       # Full page screenshot
-agent-browser screenshot --annotate   # Annotated screenshot with numbered element labels
-agent-browser pdf output.pdf          # Save as PDF
-
-# Clipboard
-agent-browser clipboard read          # Read text from clipboard
-agent-browser clipboard write "text"  # Write text to clipboard
-agent-browser clipboard copy          # Copy current selection
-agent-browser clipboard paste         # Paste from clipboard
-
-# Dialogs (alert, confirm, prompt, beforeunload)
-agent-browser dialog accept           # Accept dialog
-agent-browser dialog accept "input"   # Accept prompt dialog with text
-agent-browser dialog dismiss          # Dismiss/cancel dialog
-agent-browser dialog status           # Check if dialog is open
-
-# Diff (compare page states)
-agent-browser diff snapshot                        # Compare current vs last snapshot
-agent-browser diff screenshot --baseline before.png  # Visual pixel diff
-agent-browser diff url <url1> <url2>               # Compare two pages
-
-# Streaming
-agent-browser stream enable           # Start WebSocket streaming
-agent-browser stream status           # Inspect streaming state
-agent-browser stream disable          # Stop streaming
-```
-
-## Batch Execution
-
-```bash
-echo '[
-  ["open", "https://example.com"],
-  ["snapshot", "-i"],
-  ["click", "@e1"],
-  ["screenshot", "result.png"]
-]' | agent-browser batch --json
-```
-
-## Authentication
-
-```bash
-# Option 1: Auth vault (credentials stored encrypted)
-echo "$PASSWORD" | agent-browser auth save myapp --url https://app.example.com/login --username user --password-stdin
-agent-browser auth login myapp
-
-# Option 2: Session name (auto-save/restore cookies + localStorage)
-agent-browser --session-name myapp open https://app.example.com/login
-agent-browser close                                                       # State auto-saved
-agent-browser --session-name myapp open https://app.example.com/dashboard # Auto-restored
-
-# Option 3: Persistent profile
-agent-browser --profile ~/.myapp open https://app.example.com/login
-
-# Option 4: State file
-agent-browser state save auth.json
-agent-browser state load auth.json
-```
-
-### LobeHub dev server — inject better-auth cookie
-
-`agent-browser --headed` on macOS can create an off-screen Chromium window, blocking manual login. For a local LobeHub dev server (e.g. `localhost:3011`), copy the `better-auth.session_token` cookie out of a **Network request** in the user's own Chrome DevTools and load it via `state load`. See [references/agent-browser-login.md](./references/agent-browser-login.md) for the full recipe.
-
-## Semantic Locators (Alternative to Refs)
-
-```bash
-agent-browser find text "Sign In" click
-agent-browser find label "Email" fill "user@test.com"
-agent-browser find role button click --name "Submit"
-agent-browser find placeholder "Search" type "query"
-agent-browser find testid "submit-btn" click
-```
-
-## JavaScript Evaluation (eval)
-
-```bash
-# Simple expressions
-agent-browser eval 'document.title'
-
-# Complex JS: use --stdin with heredoc (RECOMMENDED)
-agent-browser eval --stdin << 'EVALEOF'
-JSON.stringify(
-  Array.from(document.querySelectorAll("img"))
-    .filter(i => !i.alt)
-    .map(i => ({ src: i.src.split("/").pop(), width: i.width }))
-)
-EVALEOF
-
-# Base64 encoding (avoids all shell escaping issues)
-agent-browser eval -b "$(echo -n 'document.title' | base64)"
-```
-
-## Ref Lifecycle
-
-Refs (`@e1`, `@e2`, etc.) are invalidated when the page changes. Always re-snapshot after clicking links/buttons that navigate, form submissions, or dynamic content loading.
-
-## Annotated Screenshots (Vision Mode)
-
-```bash
-agent-browser screenshot --annotate
-# Output includes the image path and a legend:
-#   [1] @e1 button "Submit"
-#   [2] @e2 link "Home"
-agent-browser click @e2 # Click using ref from annotated screenshot
-```
-
-## Parallel Sessions
-
-```bash
-agent-browser --session site1 open https://site-a.com
-agent-browser --session site2 open https://site-b.com
-agent-browser session list
-```
-
-## Connect to Existing Chrome
-
-```bash
-agent-browser --auto-connect snapshot # Auto-discover running Chrome
-agent-browser --cdp 9222 snapshot     # Explicit CDP port
-```
-
-## iOS Simulator (Mobile Safari)
-
-```bash
-agent-browser device list
-agent-browser -p ios --device "iPhone 16 Pro" open https://example.com
-agent-browser -p ios snapshot -i
-agent-browser -p ios tap @e1
-agent-browser -p ios swipe up
-agent-browser -p ios screenshot mobile.png
-agent-browser -p ios close
-```
-
-## Observability Dashboard
-
-```bash
-agent-browser dashboard install
-agent-browser dashboard start # Background server on port 4848
-agent-browser dashboard stop
-```
-
-## Cloud Providers
-
-Use `-p <provider>` to run against cloud browsers: `agentcore`, `browserbase`, `browserless`, `browseruse`, `kernel`.
-
-## Browser Engine Selection
-
-```bash
-agent-browser --engine lightpanda open example.com # 10x faster, 10x less memory
-```
-
-## Electron (LobeHub Desktop)
-
-### Setup / Teardown
-
-Use the `electron-dev.sh` script to manage the Electron dev environment. It handles process lifecycle, waits for SPA readiness, and reliably kills all child processes (main + helpers + vite).
-
-```bash
-SCRIPT=".agents/skills/local-testing/scripts/electron-dev.sh"
-
-# Start Electron dev with CDP (idempotent — skips if already running)
-$SCRIPT start
-
-# Check if Electron is running and CDP is reachable
-$SCRIPT status
-
-# Kill all Electron-related processes (main + helper + vite)
-$SCRIPT stop
-
-# Force fresh restart
-$SCRIPT restart
-```
-
-After `start` succeeds, connect with: `agent-browser --cdp 9222 snapshot -i`
-
-**Always run `$SCRIPT stop` when done testing** — `pkill -f "Electron"` alone won't catch all helper processes.
-
-#### Environment Variables
-
-| Variable          | Default                 | Description                              |
-| ----------------- | ----------------------- | ---------------------------------------- |
-| `CDP_PORT`        | `9222`                  | Chrome DevTools Protocol port            |
-| `ELECTRON_LOG`    | `/tmp/electron-dev.log` | Electron process log                     |
-| `ELECTRON_WAIT_S` | `60`                    | Max seconds to wait for Electron process |
-| `RENDERER_WAIT_S` | `60`                    | Max seconds to wait for SPA to load      |
-
-### LobeHub-Specific Patterns
-
-#### Access Zustand Store State
-
-```bash
-agent-browser --cdp 9222 eval --stdin << 'EVALEOF'
-(function() {
-  var chat = window.__LOBE_STORES.chat();
-  var ops = Object.values(chat.operations);
-  return JSON.stringify({
-    ops: ops.map(function(o) { return { type: o.type, status: o.status }; }),
-    activeAgent: chat.activeAgentId,
-    activeTopic: chat.activeTopicId,
-  });
-})()
-EVALEOF
-```
-
-#### Find and Use the Chat Input
-
-```bash
-# The chat input is contenteditable — must use -C flag
-agent-browser --cdp 9222 snapshot -i -C 2>&1 | grep "editable"
-
-agent-browser --cdp 9222 click @e48
-agent-browser --cdp 9222 type @e48 "Hello world"
-agent-browser --cdp 9222 press Enter
-```
-
-#### Wait for Agent to Complete
-
-```bash
-agent-browser --cdp 9222 eval --stdin << 'EVALEOF'
-(function() {
-  var chat = window.__LOBE_STORES.chat();
-  var ops = Object.values(chat.operations);
-  var running = ops.filter(function(o) { return o.status === 'running'; });
-  return running.length === 0 ? 'done' : 'running: ' + running.length;
-})()
-EVALEOF
-```
-
-#### Install Error Interceptor
-
-```bash
-agent-browser --cdp 9222 eval --stdin << 'EVALEOF'
-(function() {
-  window.__CAPTURED_ERRORS = [];
-  var orig = console.error;
-  console.error = function() {
-    var msg = Array.from(arguments).map(function(a) {
-      if (a instanceof Error) return a.message;
-      return typeof a === 'object' ? JSON.stringify(a) : String(a);
-    }).join(' ');
-    window.__CAPTURED_ERRORS.push(msg);
-    orig.apply(console, arguments);
-  };
-  return 'installed';
-})()
-EVALEOF
-
-# Later, check captured errors:
-agent-browser --cdp 9222 eval "JSON.stringify(window.__CAPTURED_ERRORS)"
-```
-
-## Chrome / Web Apps
-
-```bash
-/Applications/Google\ Chrome.app/Contents/MacOS/Google\ Chrome \
-  --remote-debugging-port=9222 \
-  --user-data-dir=/tmp/chrome-test-profile \
-  "<URL>" &
-sleep 5
-agent-browser --cdp 9222 snapshot -i
-
-# Or auto-discover running Chrome with remote debugging
-agent-browser --auto-connect snapshot -i
-```
-
---
-
-# Part 2: osascript (Native macOS App Bot Testing)
-
-Use AppleScript via `osascript` to control native macOS desktop apps for bot testing. Works with any app that supports macOS Accessibility, no CDP or Chromium needed.
-
-The pattern is the same for every platform:
-
-1. **Activate** the app (`tell application "X" to activate`)
-2. **Navigate** to a channel/chat (Quick Switcher `Cmd+K` or Search `Cmd+F`)
-3. **Send** a message (clipboard paste `Cmd+V` + Enter)
-4. **Wait** for the bot response
-5. **Screenshot** for verification (`screencapture` + `Read` tool)
-
-## Per-Platform References
-
-Pick the file for your target platform — each contains activation, navigation, send-message, and verification snippets specific to that app:
-
-Each channel has its own folder under `bot/<channel>/` containing an `index.md`
-(activation, navigation, send-message, and verification snippets specific to
-that app) and its test script:
-
-| Platform      | Reference                                        | Quick switcher |
-| ------------- | ------------------------------------------------ | -------------- |
-| Discord       | [bot/discord/index.md](./bot/discord/index.md)   | `Cmd+K`        |
-| Slack         | [bot/slack/index.md](./bot/slack/index.md)       | `Cmd+K`        |
-| Telegram      | [bot/telegram/index.md](./bot/telegram/index.md) | `Cmd+F`        |
-| WeChat / 微信 | [bot/wechat/index.md](./bot/wechat/index.md)     | `Cmd+F`        |
-| Lark / 飞书   | [bot/lark/index.md](./bot/lark/index.md)         | `Cmd+K`        |
-| QQ            | [bot/qq/index.md](./bot/qq/index.md)             | `Cmd+F`        |
-
-For **shared osascript patterns** (activate, type, paste, screenshot, read accessibility, common workflow template, gotchas), see [bot/osascript-common.md](./bot/osascript-common.md). Read this first if you're new to osascript automation.
-
-## Bridge-based channels (no native app)
-
-Some channels have no native app to drive with osascript — they connect through
-a local bridge inside the Desktop app. These are tested with agent-browser
-(IPC + UI) plus the bridge's own HTTP/REST endpoints, not osascript:
-
-| Channel  | Reference                                        | What it drives                                           |
-| -------- | ------------------------------------------------ | -------------------------------------------------------- |
-| iMessage | [bot/imessage/index.md](./bot/imessage/index.md) | `imessageBridge.*` IPC + local bridge + BlueBubbles REST |
-
-For iMessage there is a one-shot regression script — see `test-imessage-bridge.sh` below.
-
---
-
-# Scripts
-
-**App / recording scripts** in `.agents/skills/local-testing/scripts/`:
-
-| Script                    | Usage                                               |
-| ------------------------- | --------------------------------------------------- |
-| `electron-dev.sh`         | Manage Electron dev env (start/stop/status/restart) |
-| `record-electron-demo.sh` | Record Electron app demo with ffmpeg                |
-| `record-app-screen.sh`    | Record app screen (video + screenshots, start/stop) |
-
-**Bot scripts** live under `.agents/skills/local-testing/bot/`, one folder per
-channel (alongside that channel's `index.md`). The shared
-`capture-app-window.sh` sits at the `bot/` root:
-
-| Script                             | Usage                                                               |
-| ---------------------------------- | ------------------------------------------------------------------- |
-| `capture-app-window.sh`            | Capture screenshot of a specific app window (used by bot tests)     |
-| `discord/test-discord-bot.sh`      | Send message to Discord bot via osascript                           |
-| `slack/test-slack-bot.sh`          | Send message to Slack bot via osascript                             |
-| `telegram/test-telegram-bot.sh`    | Send message to Telegram bot via osascript                          |
-| `wechat/test-wechat-bot.sh`        | Send message to WeChat bot via osascript                            |
-| `lark/test-lark-bot.sh`            | Send message to Lark / 飞书 bot via osascript                       |
-| `qq/test-qq-bot.sh`                | Send message to QQ bot via osascript                                |
-| `imessage/test-imessage-bridge.sh` | Regression-test the iMessage BlueBubbles bridge (IPC + HTTP)        |
-| `imessage/send-imessage-test.sh`   | Send one real iMessage (desktop → BB → iMessage) and verify it sent |
-
-### Window Screenshot Utility
-
-`capture-app-window.sh` captures a screenshot of a specific app window using `screencapture -l <windowID>`. It uses Swift + CGWindowList to find the window by process name, so screenshots work correctly even when the window is on an external monitor or behind other windows.
-
-```bash
-# Standalone usage
-./.agents/skills/local-testing/bot/capture-app-window.sh "Discord" /tmp/discord.png
-./.agents/skills/local-testing/bot/capture-app-window.sh "Slack" /tmp/slack.png
-./.agents/skills/local-testing/bot/capture-app-window.sh "WeChat" /tmp/wechat.png
-```
-
-All bot test scripts use this utility automatically for their screenshots.
-
-### Bot Test Scripts
-
-All bot test scripts share the same interface:
-
-```bash
-./scripts/test-<platform>-bot.sh <channel_or_contact> <message> [wait_seconds] [screenshot_path]
-```
-
-Examples:
-
-```bash
-# Discord — test a bot in #bot-testing channel
-./.agents/skills/local-testing/bot/discord/test-discord-bot.sh "bot-testing" "!ping"
-./.agents/skills/local-testing/bot/discord/test-discord-bot.sh "bot-testing" "/ask Tell me a joke" 30
-
-# Slack — test a bot in #bot-testing channel
-./.agents/skills/local-testing/bot/slack/test-slack-bot.sh "bot-testing" "@mybot hello"
-./.agents/skills/local-testing/bot/slack/test-slack-bot.sh "bot-testing" "/ask What is 2+2?" 20
-
-# Telegram — test a bot by username
-./.agents/skills/local-testing/bot/telegram/test-telegram-bot.sh "MyTestBot" "/start"
-./.agents/skills/local-testing/bot/telegram/test-telegram-bot.sh "GPTBot" "Hello" 60
-
-# WeChat — test a bot or send to a contact
-./.agents/skills/local-testing/bot/wechat/test-wechat-bot.sh "文件传输助手" "test message" 5
-./.agents/skills/local-testing/bot/wechat/test-wechat-bot.sh "MyBot" "Tell me a joke" 30
-
-# Lark/飞书 — test a bot in a group chat
-./.agents/skills/local-testing/bot/lark/test-lark-bot.sh "bot-testing" "@MyBot hello"
-./.agents/skills/local-testing/bot/lark/test-lark-bot.sh "bot-testing" "Help me with this" 30
-
-# QQ — test a bot in a group or direct chat
-./.agents/skills/local-testing/bot/qq/test-qq-bot.sh "bot-testing" "Hello bot" 15
-./.agents/skills/local-testing/bot/qq/test-qq-bot.sh "MyBot" "/help" 10
-```
-
-Each script: activates the app, navigates to the channel/contact, pastes the message via clipboard, sends, waits, and takes a screenshot. Use the `Read` tool on the screenshot for visual verification.
-
-### iMessage bridge regression script
-
-`test-imessage-bridge.sh` does **not** follow the osascript bot interface — it
-drives the Desktop bridge's IPC + HTTP layers and asserts the result, then
-self-cleans. Needs BlueBubbles running and Electron up with CDP.
-
-```bash
-./.agents/skills/local-testing/bot/imessage/test-imessage-bridge.sh '<bluebubbles_password>' [bb_url] [cdp_port]
-# defaults: bb_url=http://127.0.0.1:1234  cdp_port=9222 — exit 0 = all green
-```
-
-It guards the connect/configure flow (testConfig happy + reject paths, first-time
-`upsertConfig` save, bridge running + webhook registered, local-server secret
-enforcement). See [bot/imessage/index.md](./bot/imessage/index.md)
-for the full manual UI flow and known bugs.
-
---
-
-# Screen Recording
-
-Record automated demos using `record-app-screen.sh` (start/stop lifecycle, CDP screenshots + ffmpeg assembly). See [references/record-app-screen.md](references/record-app-screen.md) for full documentation.
-
-```bash
-./.agents/skills/local-testing/scripts/electron-dev.sh start
-./.agents/skills/local-testing/scripts/record-app-screen.sh start my-demo
-# ... run automation ...
-./.agents/skills/local-testing/scripts/record-app-screen.sh stop
-```
-
-Outputs to `.records/` directory (gitignored): `<name>.mp4` (video) + `<name>/` (screenshots every 3s).
-
---
-
-# Gotchas
-
-### agent-browser
-
- **Daemon can get stuck** — if commands hang, `agent-browser close --all` or `pkill -f agent-browser` to reset
- **HMR invalidates everything** — after code changes, refs break. Re-snapshot or restart
- **`snapshot -i` doesn't find contenteditable** — use `snapshot -i -C` for rich text editors
- **`fill` doesn't work on contenteditable** — use `type` for chat inputs
- **Screenshots go to `~/.agent-browser/tmp/screenshots/`** — read them with the `Read` tool
- **Dialogs block all commands** — if commands time out, check `agent-browser dialog status`
- **Default timeout is 25s** — override with `AGENT_BROWSER_DEFAULT_TIMEOUT` (ms) or use explicit waits
- **Shell quoting corrupts eval** — use `eval --stdin <<'EVALEOF'` for complex JS
-
-### Electron-specific
-
- **Always use `electron-dev.sh stop` to clean up** — `pkill -f "Electron"` only kills the main process; helper processes (GPU, renderer, network) survive. The script finds and kills all of them via PID matching against the project's electron binary path.
- **`npx electron-vite dev` must run from `apps/desktop/`** — running from project root fails silently. The `electron-dev.sh` script handles this automatically.
- **Don't resize the Electron window after load** — resizing triggers full SPA reload
- **Store is at `window.__LOBE_STORES`** not `window.__ZUSTAND_STORES__`
-
-### osascript
-
-See [bot/osascript-common.md](./bot/osascript-common.md#gotchas) for the full osascript gotchas list (accessibility permissions, `keystroke` non-ASCII issues, locale-specific app names, rate limiting, etc.).
@@ -1,110 +0,0 @@
-# Log `agent-browser` into a local LobeHub dev server
-
-`agent-browser --headed` on macOS often creates the Chromium window off-screen — the user can't see or interact with it, so manual login inside the agent-browser session fails. Instead of sharing the user's real Chrome profile, copy the **better-auth session cookie** out of a request in DevTools and inject it into the agent-browser session as a Playwright-style state file.
-
-## When to use
-
- You need `agent-browser` to reach an authenticated page on `http://localhost:<port>` (e.g. `localhost:3011`).
- The user already has a logged-in tab of the same dev server in their own Chrome.
- Spawning a headed Chromium to let the user log in manually is unreliable (window off-screen, no interaction).
-
-Do **not** use this on production URLs — only local dev. Treat the cookie as a secret: don't paste it into shared logs, PRs, or commit it anywhere.
-
-## Step 1 — Ask the user to copy the cookie from a Network request, NOT `document.cookie`
-
-`document.cookie` will not return HttpOnly cookies, which is exactly where better-auth puts its session. Instruct the user:
-
-1. Open the logged-in tab (`http://localhost:<port>/…`) in their own Chrome.
-2. `Cmd+Option+I` → **Network** tab.
-3. Refresh, click any same-origin request (e.g. the top-level document request).
-4. In the right pane under **Request Headers**, right-click the `Cookie:` line → **Copy value** (or copy the entire header).
-5. Paste the string into chat.
-
-You only need the better-auth pieces. Everything else (Clerk, `LOBE_LOCALE`, HMR hash, theme vars) is noise and can stay. The minimum viable set is:
-
-```
-better-auth.session_token=<value>; better-auth.state=<value>
-```
-
-## Step 2 — Build a Playwright-style state file
-
-`agent-browser state load` expects Playwright's `storageState` format: a JSON with a `cookies` array and an `origins` array.
-
-```bash
-cat > /tmp/mkstate.py << 'PY'
-import json, sys, time
-
-# Read the Cookie header from stdin (allows optional "Cookie: " prefix).
-raw = sys.stdin.read().strip()
-if raw.lower().startswith("cookie:"):
-    raw = raw.split(":", 1)[1].strip()
-
-# Keep only better-auth cookies. Extend this set if the app genuinely needs more.
-WANTED = {"better-auth.session_token", "better-auth.state"}
-
-cookies = []
-exp = int(time.time()) + 30 * 24 * 3600  # 30 days
-for pair in raw.split("; "):
-    if "=" not in pair:
-        continue
-    name, _, value = pair.partition("=")
-    if name not in WANTED:
-        continue
-    cookies.append({
-        "name": name,
-        "value": value,
-        "domain": "localhost",
-        "path": "/",
-        "expires": exp,
-        "httpOnly": False,
-        "secure": False,
-        "sameSite": "Lax",
-    })
-
-if not cookies:
-    sys.stderr.write("no better-auth cookies found in input\n")
-    sys.exit(1)
-
-print(json.dumps({"cookies": cookies, "origins": []}, indent=2))
-PY
-
-# Feed the copied Cookie header in via env var or heredoc.
-printf '%s' "$COOKIE_HEADER" | python3 /tmp/mkstate.py > /tmp/state.json
-```
-
-**Note on `httpOnly`**: the real cookie in the user's browser is HttpOnly, but `storageState` doesn't enforce the flag on load — it just attaches the value. Storing with `httpOnly: false` is fine for local dev and sidesteps a CDP-context quirk where HttpOnly cookies sometimes fail to attach.
-
-## Step 3 — Load state and navigate
-
-```bash
-SESSION="my-test" # any stable session name
-
-agent-browser --session "$SESSION" state load /tmp/state.json
-agent-browser --session "$SESSION" open "http://localhost:3011/"
-agent-browser --session "$SESSION" get url
-# Expect NOT /signin?callbackUrl=… — if you still see signin, cookie didn't apply.
-```
-
-## Step 4 — Verify
-
-```bash
-agent-browser --session "$SESSION" snapshot -i | head -20
-# Look for the user's avatar/name in the sidebar, or absence of the signin form.
-```
-
-## Common failure modes
-
-| Symptom                                         | Cause                                                                   | Fix                                                  |
-| ----------------------------------------------- | ----------------------------------------------------------------------- | ---------------------------------------------------- |
-| Still redirects to `/signin` after `state load` | User pasted from `document.cookie` → missed HttpOnly session            | Re-pull from Network request Headers, not console    |
-| `state load` reports 0 cookies                  | Separator wrong, or user pasted URL-decoded value                       | Keep the raw `Cookie:` header as-is; split on `"; "` |
-| Login works briefly then expires                | `better-auth.session_token` rotated (user logged out / signed in again) | Re-copy and re-load                                  |
-| Domain mismatch                                 | Use `domain: "localhost"` literally, no leading dot for local dev       | —                                                    |
-
-## Scope
-
-Only covers authenticating an **agent-browser** session into a **local** LobeHub dev server. It does not:
-
- Work for production — production cookies are `Secure; HttpOnly; Domain=.lobehub.com` and must be delivered over HTTPS.
- Replace real OAuth flows — tests that must exercise the login UI need a real Chromium with `--remote-debugging-port` or a bot account.
- Flow cookies back to the user's Chrome — injection is one-way (into agent-browser only).
@@ -0,0 +1,69 @@
+---
+name: model-bank-metadata
+description: 'Backfill and maintain model-bank metadata (knowledgeCutoff, family, generation). Use when adding models, fixing cutoff/family data, running a metadata sweep across aiModels providers, or researching official knowledge cutoffs.'
+user-invocable: false
+---
+
+# Model-Bank Metadata (knowledgeCutoff / family / generation)
+
+How to populate and maintain the three structured metadata fields on `packages/model-bank/src/aiModels/*.ts` model cards, at single-model scale (new model PR) or repo-wide scale (sweep across \~80 provider files / \~1900 entries).
+
+## Field semantics
+
+| Field             | Format                                                                              | Meaning                                                                                                                                                                                 |
+| ----------------- | ----------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| `knowledgeCutoff` | `'YYYY-MM'` (or `'YYYY'` if only the year is published)                             | World-knowledge cutoff. When a vendor distinguishes a **"reliable knowledge cutoff"** from the broader training-data cutoff (Anthropic does), always use the **reliable** one.          |
+| `family`          | lowercase slug (`claude`, `gpt`, `o-series`, `qwen`, `deepseek`, `llama`, `glm`, …) | Model lineage, finer than `organization`. Lets the UI group models and match the same model across aggregator providers.                                                                |
+| `generation`      | family slug + version (`claude-4.6`, `gpt-5.2`, `qwen3.5`, `llama-3.1`)             | Generation within the family. Only set when confidently derivable from the model line's naming. Rolling aliases (`qwen-max`, `deepseek-chat`, `gemini-flash-latest`) get `family` only. |
+
+All three are optional. **The cardinal rule: only fill what an authoritative source states or naming rules derive — never guess.** An empty field is correct for vendors that publish nothing.
+
+No DB migration is ever needed for these: builtin models are merged from model-bank at read time (`repositories/aiInfra/index.ts` spreads the whole card), so new card fields flow to the client automatically.
+
+## Sourcing rules for knowledgeCutoff
+
+Accept only:
+
+- Vendor official docs (platform.openai.com / developers.openai.com, docs.x.ai, ai.google.dev, docs.anthropic.com / platform.claude.com)
+- Official Hugging Face org model cards (huggingface.co/meta-llama/..., etc.)
+- Official tech reports / system cards / launch blog posts
+
+Reject:
+
+- **Third-party aggregator sites** (aiknowledgecutoff.com and similar) — proven to copy one model's value across a whole family. A Cohere sweep once claimed `2024-06` for four distinct base models; none of the cited Cohere pages said that, and the only cutoff Cohere actually publishes is Feb 2023 for the 08-2024 Command R/R+ refresh.
+- **AWS Bedrock model cards as sole source** — proven to conflate launch date with knowledge cutoff (DeepSeek R1's card lists both as "Jan 2025"). If Bedrock is the only place a value appears, leave the field empty.
+- Inference from `releasedAt` — a release date is not a cutoff.
+
+Variant inheritance: dated snapshots (`-2024-08-06`), speed/price tiers of the same checkpoint, quantizations (`-fp8`, `-awq`), context-length variants (`-32k`), ollama `:NNb` tags, and cloud-prefixed ids (`anthropic.`/`us.`/`global.` Bedrock ids) share their base model's cutoff. **Distills do not inherit** from teacher or base — use the distill's own published value or leave empty. **Sizes within one generation can genuinely differ**: Llama 3 8B is Mar 2023 while 70B is Dec 2023 (per Meta's own card) — don't "fix" that to one family-wide value.
+
+Vendors that publish no cutoffs (leave empty, don't chase): Qwen, DeepSeek, GLM/Zhipu, ERNIE, Doubao, Hunyuan, SenseNova, Spark, MiniMax, StepFun, Yi (mostly), Moonshot.
+
+Known per-vendor footguns:
+
+- **Anthropic**: Opus 4.6 reliable cutoff is `2025-05`, Sonnet 4.6 is `2025-08` — easy to swap. Claude 3.7 is `2024-10` (system card: trained through Nov 2024, knowledge cutoff end of Oct 2024). Cite system cards / the models overview, not the Help Center article (a living page that drops retired models — citation rot).
+- **xAI**: docs.x.ai has one blanket sentence covering grok-3/grok-4; mini variants are not named there. Grok 4.20/4.3 have no official cutoff anywhere.
+- **OpenAI**: per-model docs pages (developers.openai.com/api/docs/models/<id>) state cutoffs explicitly, including snapshot differences (gpt-4-1106-preview `2023-04` vs gpt-4-0125-preview `2023-12`).
+
+## family/generation derivation
+
+Rule-based, no research needed: `scripts/derive-family.ts` holds the per-family regex rules. Traps already encoded there — keep them when extending:
+
+- Date suffixes are not versions: `claude-sonnet-4-20250514` is generation `claude-4`, not `claude-4.2`.
+- Size suffixes are not versions: `llama-3-8b` → `llama-3` (not `llama-3.8`); `gemma-7b-it` is **gemma-1** (not gemma-7).
+- Vendor spelling variants: `qwen2p5` = qwen2.5, `llama-v3p1` = llama-3.1, ollama `:NNb` tags, Bedrock `us.`/`global.`/`anthropic.` prefixes.
+- `claude-X.0` normalizes to `claude-X`.
+- Fable/Mythos-class ids (`claude-fable-5`) don't match the opus/sonnet/haiku regex — they are the Mythos class — `family: 'claude-mythos'`, `generation: 'mythos-5'` (set manually; the launch page calls Fable 5 "the generally available Mythos-class model").
+
+## Repo-wide sweep workflow
+
+1. **Extract ids**: `bun .agents/skills/model-bank-metadata/scripts/extract-model-ids.ts` → unique normalized chat-model ids (normalization = last path segment, lowercased). Non-chat types (image/video/embedding/tts) have no knowledge cutoff — skip them.
+2. **Research (multi-agent)**: chunk ids by family (≤50 per chunk) and fan out one research agent per chunk (Workflow tool), each returning `{id, cutoff, source}` with the sourcing rules above baked into the prompt, **plus** one adversarial verify agent per chunk that re-fetches cited sources and refutes unsupported claims. The verify pass is load-bearing: it caught the Cohere aggregator copy-paste and the AWS launch-date conflation.
+3. **Policy filter**: before applying, drop entries whose only source is a rejected category (check the returned `sources` map — e.g. drop everything sourced to aws.amazon.com).
+4. **Apply**: `bun scripts/apply-cutoffs.ts <map.json>` and `bun scripts/apply-family.ts <map.json>` (run from repo root). Both are idempotent codemods keyed on normalized id — aggregator providers get the same values automatically; entries that already have the field are skipped. They rely on the uniform prettier formatting of the data files (entries start `  {` / end `  },`, fields at 4-space indent).
+5. **Verify**: `cd packages/model-bank && bunx vitest run src/aiModels/__tests__/index.test.ts && bunx tsc --noEmit`.
+
+## Maintenance rules
+
+- **New model PRs** should fill all three fields inline, citing the official source in the PR body (see the Anthropic entries in `anthropic.ts` for reference values).
+- **After resolving merge conflicts** in model-bank data files, sanity-check that metadata didn't vanish: `git grep -c knowledgeCutoff -- 'packages/model-bank/src/aiModels/*.ts'` before vs after. A three-way stack of model PRs once silently dropped all 10 Anthropic cutoffs during conflict resolution.
+- Dirty ids exist in aggregator data (a sambanova id once carried a trailing tab). The codemods match ids verbatim — if a map key won't apply, check for invisible characters before assuming the model is missing.
@@ -0,0 +1,73 @@
+/**
+ * One-off codemod: apply a canonical { normalizedModelId: 'YYYY-MM' } map onto
+ * packages/model-bank/src/aiModels/*.ts, inserting `knowledgeCutoff` after the
+ * `id:` line of every chat-model entry that matches and doesn't already have one.
+ *
+ * Relies on the uniform prettier formatting of these files:
+ *   - each model entry starts with `  {` and ends with `  },` (2-space indent)
+ *   - fields are at 4-space indent: `    id: '...'`, `    type: 'chat'`
+ *
+ * Usage: bun /tmp/apply-cutoffs.ts /tmp/cutoff-map.json
+ */
+import { readdirSync, readFileSync, writeFileSync } from 'node:fs';
+import { join } from 'node:path';
+
+const mapPath = process.argv[2];
+if (!mapPath) throw new Error('usage: bun apply-cutoffs.ts <map.json>');
+const map: Record<string, string> = JSON.parse(readFileSync(mapPath, 'utf8'));
+
+const dir = 'packages/model-bank/src/aiModels';
+const normalize = (id: string) => id.split('/').pop()!.toLowerCase();
+
+let touchedFiles = 0;
+let inserted = 0;
+const matchedIds = new Set<string>();
+
+for (const file of readdirSync(dir).filter((f) => f.endsWith('.ts'))) {
+  const path = join(dir, file);
+  const lines = readFileSync(path, 'utf8').split('\n');
+  const out: string[] = [];
+  let changed = false;
+
+  let i = 0;
+  while (i < lines.length) {
+    if (lines[i] !== '  {') {
+      out.push(lines[i]);
+      i++;
+      continue;
+    }
+    // collect one model entry block
+    const start = i;
+    let end = i;
+    while (end < lines.length && lines[end] !== '  },') end++;
+    const block = lines.slice(start, end + 1);
+
+    const idLineIdx = block.findIndex((l) => /^ {4}id: '/.test(l));
+    const isChat = block.some((l) => /^ {4}type: 'chat',?$/.test(l));
+    const hasCutoff = block.some((l) => /^ {4}knowledgeCutoff:/.test(l));
+
+    if (idLineIdx >= 0 && isChat && !hasCutoff) {
+      const rawId = block[idLineIdx].match(/^ {4}id: '(.+)',$/)?.[1];
+      const norm = rawId ? normalize(rawId) : undefined;
+      const cutoff = norm ? map[norm] : undefined;
+      if (cutoff && /^\d{4}(?:-\d{2})?$/.test(cutoff)) {
+        block.splice(idLineIdx + 1, 0, `    knowledgeCutoff: '${cutoff}',`);
+        inserted++;
+        changed = true;
+        matchedIds.add(norm!);
+      }
+    }
+    out.push(...block);
+    i = end + 1;
+  }
+
+  if (changed) {
+    writeFileSync(path, out.join('\n'));
+    touchedFiles++;
+  }
+}
+
+console.log(`inserted ${inserted} knowledgeCutoff fields across ${touchedFiles} files`);
+console.log(`map ids used: ${matchedIds.size}/${Object.keys(map).length}`);
+const unused = Object.keys(map).filter((k) => !matchedIds.has(k));
+if (unused.length) console.log('unused map keys (first 20):', unused.slice(0, 20));
@@ -0,0 +1,49 @@
+import { readdirSync, readFileSync, writeFileSync } from 'node:fs';
+import { join } from 'node:path';
+
+const map: Record<string, { family: string; generation?: string }> = JSON.parse(
+  readFileSync('/tmp/family-map.json', 'utf8'),
+);
+const dir = 'packages/model-bank/src/aiModels';
+const normalize = (id: string) => id.split('/').pop()!.toLowerCase();
+
+let inserted = 0;
+let touchedFiles = 0;
+for (const file of readdirSync(dir).filter((f) => f.endsWith('.ts'))) {
+  const path = join(dir, file);
+  const lines = readFileSync(path, 'utf8').split('\n');
+  const out: string[] = [];
+  let changed = false;
+  let i = 0;
+  while (i < lines.length) {
+    if (lines[i] !== '  {') {
+      out.push(lines[i]);
+      i++;
+      continue;
+    }
+    let end = i;
+    while (end < lines.length && lines[end] !== '  },') end++;
+    const block = lines.slice(i, end + 1);
+    const idLineIdx = block.findIndex((l) => /^ {4}id: '/.test(l));
+    const isChat = block.some((l) => /^ {4}type: 'chat',?$/.test(l));
+    const hasFamily = block.some((l) => /^ {4}family:/.test(l));
+    if (idLineIdx >= 0 && isChat && !hasFamily) {
+      const rawId = block[idLineIdx].match(/^ {4}id: '(.+)',$/)?.[1];
+      const r = rawId ? map[normalize(rawId)] : undefined;
+      if (r) {
+        const add = [`    family: '${r.family}',`];
+        if (r.generation) add.push(`    generation: '${r.generation}',`);
+        block.splice(idLineIdx, 0, ...add);
+        inserted++;
+        changed = true;
+      }
+    }
+    out.push(...block);
+    i = end + 1;
+  }
+  if (changed) {
+    writeFileSync(path, out.join('\n'));
+    touchedFiles++;
+  }
+}
+console.log(`annotated ${inserted} model entries across ${touchedFiles} files`);
@@ -0,0 +1,237 @@
+/* eslint-disable regexp/no-unused-capturing-group */
+/**
+ * Rule-based derivation of { family, generation } from normalized model ids.
+ * Principle: only fill what is confidently derivable; otherwise omit.
+ *
+ * Usage: bun /tmp/derive-family.ts            # print distinct pairs for review
+ *        bun /tmp/derive-family.ts --emit     # write /tmp/family-map.json
+ */
+import { readFileSync, writeFileSync } from 'node:fs';
+
+const ids: string[] = JSON.parse(readFileSync('/tmp/model-ids.json', 'utf8'));
+
+type R = { family: string; generation?: string };
+
+const derive = (id: string): R | undefined => {
+  // strip cloud/bedrock prefixes for matching
+  const m = id.replace(/^(us\.|global\.|eu\.|apac\.)?(anthropic\.|meta\.|cohere\.|azure-)/, '');
+
+  // ---- anthropic ----
+  if (m.startsWith('claude')) {
+    // family = product-line tier (claude-opus/sonnet/haiku/instant); bare claude-2.x has no tier
+    const tier = m.match(/(opus|sonnet|haiku|instant)/)?.[1];
+    const family = tier ? `claude-${tier}` : 'claude';
+    let g = m.match(/^claude-(?:opus|sonnet|haiku)-(\d)[.-](\d)(?!\d)/); // claude-opus-4-8 / claude-haiku-4.5
+    if (g) return { family, generation: `claude-${g[1]}.${g[2]}` };
+    g = m.match(/^claude-(?:opus|sonnet|haiku)-(\d)(?!\d)/); // claude-opus-4
+    if (g) return { family, generation: `claude-${g[1]}` };
+    g = m.match(/^claude-(\d)[.-](\d)(?!\d)/); // claude-3-5-haiku / claude-3.7-sonnet / claude-2.1
+    if (g) return { family, generation: g[2] === '0' ? `claude-${g[1]}` : `claude-${g[1]}.${g[2]}` };
+    g = m.match(/^claude-(\d)(?!\d)/); // claude-3-haiku
+    if (g) return { family, generation: `claude-${g[1]}` };
+    if (m.startsWith('claude-instant')) return { family: 'claude-instant' };
+    if (/^claude-v?2/.test(m)) return { family: 'claude', generation: 'claude-2' };
+    return { family };
+  }
+
+  // ---- openai ----
+  if (/^(gpt-oss|gpt_oss)/.test(m) || m.startsWith('gpt-oss:'))
+    return { family: 'gpt-oss', generation: 'gpt-oss' };
+  if (/^(chatgpt-4o|gpt-4o)/.test(m)) return { family: 'gpt', generation: 'gpt-4o' };
+  if (/^gpt-(3\.5|35)/.test(m)) return { family: 'gpt', generation: 'gpt-3.5' };
+  if (m.startsWith('gpt-audio')) return { family: 'gpt', generation: 'gpt-audio' };
+  {
+    const g = m.match(/^gpt-(\d)\.(\d)/); // gpt-4.1 / gpt-5.2
+    if (g) return { family: 'gpt', generation: `gpt-${g[1]}.${g[2]}` };
+    const g2 = m.match(/^gpt-(\d)(?!\d)/); // gpt-4 / gpt-5
+    if (g2) return { family: 'gpt', generation: `gpt-${g2[1]}` };
+  }
+  {
+    const g = m.match(/^o([134])(-|$)/); // o1 / o3 / o4
+    if (g) return { family: 'o-series', generation: `o${g[1]}` };
+  }
+  if (/^(codex|computer-use-preview)/.test(m)) return { family: 'gpt' };
+
+  // ---- google ----
+  {
+    const g = m.match(/^gemini-(\d+(?:\.\d+)?)/);
+    if (g) return { family: 'gemini', generation: `gemini-${g[1]}` };
+    if (/^gemini-(pro|flash)/.test(m)) return { family: 'gemini' }; // rolling aliases
+    if (m.startsWith('gemma')) {
+      if (/^gemma-?\db/.test(m)) return { family: 'gemma', generation: 'gemma-1' };
+      const v = m.match(/^gemma-?(\d)(?!b)/);
+      return { family: 'gemma', generation: v ? `gemma-${v[1]}` : undefined };
+    }
+    if (/^(codegemma|learnlm|palm)/.test(m)) return { family: m.match(/^[a-z]+/)![0] };
+  }
+
+  // ---- qwen ----
+  if (m.startsWith('qwq')) return { family: 'qwen', generation: 'qwq' };
+  if (m.startsWith('qvq')) return { family: 'qwen', generation: 'qvq' };
+  if (m.startsWith('codeqwen')) return { family: 'qwen' };
+  if (m.startsWith('qwen')) {
+    const g =
+      m.match(/^qwen-?([123](?:\.\d+)?)(?![0-9b])/) || // qwen3.5-plus / qwen-3-14b / qwen2-7b / qwen1.5
+      m.match(/^qwen([23](?:\.\d+)?):/) || // qwen2.5:72b
+      m.match(/^qwen([23])p(\d)/); // qwen2p5 -> handled below
+    if (/^qwen(\d)p(\d)/.test(m)) {
+      const p = m.match(/^qwen(\d)p(\d)/)!;
+      return { family: 'qwen', generation: `qwen${p[1]}.${p[2]}` };
+    }
+    if (g) return { family: 'qwen', generation: `qwen${g[1]}` };
+    return { family: 'qwen' }; // qwen-max/plus/turbo/vl rolling aliases
+  }
+
+  // ---- deepseek ----
+  if (/^(deepseek|azure-deepseek|pro-deepseek)/.test(m) || m.startsWith('deepseek_')) {
+    const s = m.replace(/^pro-/, '').replaceAll('_', '-');
+    if (s.startsWith('deepseek-r1-distill'))
+      return { family: 'deepseek', generation: 'deepseek-r1-distill' };
+    if (s.startsWith('deepseek-r1')) return { family: 'deepseek', generation: 'deepseek-r1' };
+    const g = s.match(/^deepseek-(?:chat-)?v(\d(?:\.\d)?)/);
+    if (g) return { family: 'deepseek', generation: `deepseek-v${g[1]}` };
+    if (/^deepseek-(coder-v2|coder)/.test(s))
+      return { family: 'deepseek', generation: 'deepseek-coder' };
+    return { family: 'deepseek' }; // deepseek-chat / reasoner rolling aliases
+  }
+
+  // ---- meta llama ----
+  if (m.startsWith('codellama')) return { family: 'llama', generation: 'codellama' };
+  if (/^(meta-)?llama|^l3(\d)?-|^llava/.test(m)) {
+    if (m.startsWith('llava')) return { family: 'llava' };
+    const s = m.replace(/^meta-/, '');
+    const g =
+      s.match(/^llama-?([234])(?:[.-](\d))?(?![0-9b])/) || // llama-3.1 / llama3.3 / llama-4
+      s.match(/^llama-?v([234])p?(\d)?/) || // llama-v3p1
+      s.match(/^llama([234])[.:-](\d)?/);
+    if (g) {
+      const gen = g[2] ? `llama-${g[1]}.${g[2]}` : `llama-${g[1]}`;
+      return { family: 'llama', generation: gen };
+    }
+    if (m.startsWith('l3-')) return { family: 'llama', generation: 'llama-3' };
+    if (m.startsWith('l31-')) return { family: 'llama', generation: 'llama-3.1' };
+    return { family: 'llama' };
+  }
+
+  // ---- zhipu ----
+  if (/^(zai-)?glm/.test(m)) {
+    const s = m.replace(/^zai-/, '');
+    if (s.startsWith('glm-z1')) return { family: 'glm', generation: 'glm-z1' };
+    if (s.startsWith('glm-zero')) return { family: 'glm', generation: 'glm-zero' };
+    const g = s.match(/^glm-(\d(?:\.\d)?)/);
+    if (g) return { family: 'glm', generation: `glm-${g[1]}` };
+    return { family: 'glm' };
+  }
+  if (/^(charglm|codegeex|emohaa)/.test(m)) return { family: m.match(/^[a-z]+/)![0] };
+
+  // ---- mistral ----
+  if (
+    /^(open-)?(mistral|mixtral|ministral|codestral|devstral|magistral|pixtral|mathstral|labs-devstral|labs-leanstral|open-codestral)/.test(
+      m,
+    )
+  ) {
+    const fam = m.replace(/^(open-|labs-)/, '').match(/^[a-z]+/)![0];
+    return { family: fam };
+  }
+
+  // ---- xai ----
+  if (m.startsWith('grok')) {
+    const g = m.match(/^grok-(\d(?:\.\d+)?)/);
+    return { family: 'grok', generation: g ? `grok-${g[1]}` : undefined };
+  }
+
+  // ---- moonshot ----
+  if (m.startsWith('kimi')) {
+    const g = m.match(/^kimi-k(\d(?:\.\d)?)/);
+    return { family: 'kimi', generation: g ? `kimi-k${g[1]}` : undefined };
+  }
+  if (m.startsWith('moonshot-kimi-k2')) return { family: 'kimi', generation: 'kimi-k2' };
+  if (m.startsWith('moonshot-v1')) return { family: 'kimi', generation: 'moonshot-v1' };
+
+  // ---- minimax ----
+  if (m.startsWith('minimax')) {
+    if (m.startsWith('minimax-text')) return { family: 'minimax', generation: 'minimax-text-01' };
+    const g = m.match(/^minimax-m(\d(?:\.\d)?)/);
+    return { family: 'minimax', generation: g ? `minimax-m${g[1]}` : undefined };
+  }
+  if (m.startsWith('abab')) return { family: 'minimax', generation: 'abab' };
+
+  // ---- baidu ----
+  if (m.startsWith('ernie')) {
+    if (m.startsWith('ernie-x1')) return { family: 'ernie', generation: 'ernie-x1' };
+    const g = m.match(/^ernie-(\d\.\d)/);
+    return { family: 'ernie', generation: g ? `ernie-${g[1]}` : undefined };
+  }
+  if (m.startsWith('qianfan')) return { family: 'qianfan' };
+
+  // ---- bytedance ----
+  if (m.startsWith('doubao')) {
+    const g = m.match(/^doubao-seed-(\d[.-]\d|\d)/) || m.match(/^doubao-(\d\.\d)/);
+    return { family: 'doubao', generation: g ? `doubao-${g[1].replace('-', '.')}` : undefined };
+  }
+  if (/^(seed-oss|skylark)/.test(m)) return { family: m.startsWith('seed') ? 'doubao' : 'skylark' };
+
+  // ---- tencent ----
+  if (m.startsWith('hunyuan')) {
+    const g = m.match(/^hunyuan-(\d\.\d)/);
+    return { family: 'hunyuan', generation: g ? `hunyuan-${g[1]}` : undefined };
+  }
+  if (m.startsWith('hy3')) return { family: 'hunyuan', generation: 'hunyuan-3' };
+
+  // ---- others (family only / simple version) ----
+  if (m.startsWith('yi-')) return { family: 'yi' };
+  if (/^(command|c4ai-command)/.test(m)) return { family: 'command' };
+  if (/^(aya|c4ai-aya)/.test(m)) return { family: 'aya' };
+  if (/^phi-?(\d)?/.test(m) && m.startsWith('phi')) {
+    const g = m.match(/^phi-?(\d(?:\.\d)?)/);
+    return { family: 'phi', generation: g ? `phi-${g[1]}` : undefined };
+  }
+  if (m.startsWith('wizardlm')) return { family: 'wizardlm' };
+  if (m.startsWith('step-')) {
+    const g = m.match(/^step-(?:r1|(\d(?:\.\d)?))/);
+    return { family: 'step', generation: g?.[1] ? `step-${g[1]}` : undefined };
+  }
+  if (/^(internlm|intern-)/.test(m)) return { family: 'intern' };
+  if (m.startsWith('internvl')) return { family: 'internvl' };
+  if (m.startsWith('baichuan')) {
+    const g = m.match(/^baichuan-?(m?\d)/);
+    return { family: 'baichuan', generation: g ? `baichuan-${g[1]}` : undefined };
+  }
+  if (/^(sensechat|sensenova)/.test(m)) return { family: 'sensenova' };
+  if (/^(spark|generalv|4\.0ultra)/.test(m)) return { family: 'spark' };
+  if (/^(360gpt|360zhinao)/.test(m)) return { family: '360zhinao' };
+  if (/^(jamba|ai21-jamba)/.test(m)) return { family: 'jamba' };
+  if (m.startsWith('sonar')) return { family: 'sonar' };
+  if (/^(nova-lite|nova-micro|nova-pro)/.test(m)) return { family: 'nova' };
+  if (/^(ling|ring)-/.test(m)) return { family: m.match(/^[a-z]+/)![0] };
+  if (m.startsWith('longcat')) return { family: 'longcat' };
+  if (m.startsWith('mimo')) return { family: 'mimo' };
+  if (m.startsWith('taichu')) return { family: 'taichu' };
+  if (/^(hermes|nous-hermes)/.test(m)) return { family: 'hermes' };
+  if (m.startsWith('solar')) return { family: 'solar' };
+  if (m.startsWith('kat-coder')) return { family: 'kat-coder' };
+  if (m.startsWith('dbrx')) return { family: 'dbrx' };
+  if (m.startsWith('morph')) return { family: 'morph' };
+
+  return undefined;
+};
+
+const map: Record<string, R> = {};
+const pairs = new Map<string, number>();
+let derived = 0;
+for (const id of ids) {
+  const r = derive(id);
+  if (!r) continue;
+  derived++;
+  map[id] = r;
+  const key = `${r.family} :: ${r.generation ?? '—'}`;
+  pairs.set(key, (pairs.get(key) || 0) + 1);
+}
+
+console.log(`derived ${derived}/${ids.length}`);
+for (const [k, n] of [...pairs.entries()].sort()) console.log(String(n).padStart(4), k);
+
+if (process.argv.includes('--emit')) {
+  writeFileSync('/tmp/family-map.json', JSON.stringify(map, null, 1));
+  console.log('\nwritten /tmp/family-map.json');
+}
@@ -0,0 +1,23 @@
+/**
+ * Extract unique normalized chat-model ids from packages/model-bank/src/aiModels/*.ts.
+ * Normalization: last path segment, lowercased (matches the apply codemods).
+ *
+ * Usage (repo root): bun .agents/skills/model-bank-metadata/scripts/extract-model-ids.ts [out.json]
+ * Default output: /tmp/model-ids.json
+ */
+import { readdirSync, writeFileSync } from 'node:fs';
+import { join, resolve } from 'node:path';
+
+const dir = resolve('packages/model-bank/src/aiModels');
+const out = process.argv[2] || '/tmp/model-ids.json';
+
+const ids = new Set<string>();
+for (const f of readdirSync(dir).filter((f) => f.endsWith('.ts'))) {
+  const mod = await import(join(dir, f));
+  for (const m of mod.default || []) {
+    if (!m?.id || m.type !== 'chat') continue;
+    ids.add(m.id.split('/').pop()!.toLowerCase());
+  }
+}
+writeFileSync(out, JSON.stringify([...ids].sort(), null, 1));
+console.log(`${ids.size} unique normalized chat ids -> ${out}`);
@@ -50,7 +50,7 @@ Common false positives (do NOT merge):
 - `db-migrations` vs `drizzle` — distinct workflows (migration files vs schema authoring).
 - `microcopy` vs `i18n` — content vs mechanics.
 - `agent-runtime-hooks` vs `agent-tracing` vs `agent-signal` — different surfaces of the agent system.
- `testing` vs `local-testing` vs `cli-backend-testing` — different test types.
+- `testing` vs `agent-testing` — different test types.

 ### 4 — Description format consistency

@@ -5,6 +5,18 @@ inputs:
  node-version:
    description: Node.js version
    required: true
+  cloud-repository:
+    description: Cloud repository to overlay for commercial desktop builds
+    required: false
+    default: lobehub/lobehub-cloud
+  cloud-ref:
+    description: Optional Cloud repository ref
+    required: false
+    default: ''
+  cloud-token:
+    description: GitHub token with permission to read the Cloud repository
+    required: false
+    default: ''

 runs:
  using: composite
@@ -14,9 +26,77 @@ runs:
      with:
        node-version: ${{ inputs.node-version }}

+    - name: Overlay Cloud repository for desktop build
+      if: inputs.cloud-token != ''
+      shell: bash
+      env:
+        CLOUD_CHECKOUT: ${{ runner.temp }}/lobehub-cloud
+        CLOUD_REF: ${{ inputs.cloud-ref }}
+        CLOUD_REPOSITORY: ${{ inputs.cloud-repository }}
+        CLOUD_ROOT: ${{ github.workspace }}/..
+        CLOUD_TOKEN: ${{ inputs.cloud-token }}
+      run: |
+        set -euo pipefail
+
+        cloud_root="$(cd "$GITHUB_WORKSPACE/.." && pwd)"
+        cloud_checkout="$RUNNER_TEMP/lobehub-cloud"
+
+        rm -rf "$cloud_checkout"
+
+        clone_args=(--depth 1)
+        if [ -n "$CLOUD_REF" ]; then
+          clone_args+=(--branch "$CLOUD_REF")
+        fi
+
+        git clone "${clone_args[@]}" "https://x-access-token:${CLOUD_TOKEN}@github.com/${CLOUD_REPOSITORY}.git" "$cloud_checkout"
+
+        node <<'NODE'
+        const fs = require('node:fs');
+        const path = require('node:path');
+
+        const source = process.env.CLOUD_CHECKOUT;
+        const target = process.env.CLOUD_ROOT;
+        const skip = new Set(['.git', 'lobehub', 'node_modules']);
+
+        const copy = (from, to) => {
+          const stat = fs.lstatSync(from);
+          if (stat.isSymbolicLink()) {
+            const link = fs.readlinkSync(from);
+            fs.rmSync(to, { force: true, recursive: true });
+            fs.symlinkSync(link, to);
+            return;
+          }
+
+          if (stat.isDirectory()) {
+            fs.mkdirSync(to, { recursive: true });
+            for (const entry of fs.readdirSync(from)) {
+              if (skip.has(entry)) continue;
+              copy(path.join(from, entry), path.join(to, entry));
+            }
+            return;
+          }
+
+          fs.mkdirSync(path.dirname(to), { recursive: true });
+          fs.copyFileSync(from, to);
+        };
+
+        for (const entry of fs.readdirSync(source)) {
+          if (skip.has(entry)) continue;
+          copy(path.join(source, entry), path.join(target, entry));
+        }
+        NODE
+
+        echo "CLOUD_DESKTOP=1" >> "$GITHUB_ENV"
+        echo "✅ Cloud repository overlaid at $cloud_root"
+
    - name: Install dependencies
      shell: bash
-      run: pnpm install --node-linker=hoisted
+      run: |
+        set -euo pipefail
+        if [ "${CLOUD_DESKTOP:-}" = "1" ]; then
+          cd ..
+        fi
+        pnpm install --node-linker=hoisted

    # 移除国内 electron 镜像配置，GitHub Actions 使用官方源更快
    - name: Remove China electron mirror from .npmrc
@@ -31,4 +111,11 @@ runs:

    - name: Install deps on Desktop
      shell: bash
-      run: npm run install-isolated --prefix=./apps/desktop
+      run: |
+        set -euo pipefail
+        if [ "${CLOUD_DESKTOP:-}" = "1" ]; then
+          cd ..
+          npm run install-isolated --prefix=./lobehub/apps/desktop
+        else
+          npm run install-isolated --prefix=./apps/desktop
+        fi
@@ -30,7 +30,7 @@ jobs:
            This issue is closed, If you have any questions, you can comment and reply.
      - name: Checkout repository
        if: github.event_name == 'pull_request_target' && github.event.pull_request.merged == true
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6

      - name: Check if PR author is maintainer
        if: github.event.pull_request.merged == true
@@ -104,6 +104,7 @@ jobs:
      - name: Setup build environment
        uses: ./.github/actions/desktop-build-setup
        with:
+          cloud-token: ${{ secrets.LOBEHUB_CLOUD_TOKEN }}
          node-version: ${{ env.NODE_VERSION }}

      - name: Set package version
@@ -172,6 +173,7 @@ jobs:
      - name: Setup build environment
        uses: ./.github/actions/desktop-build-setup
        with:
+          cloud-token: ${{ secrets.LOBEHUB_CLOUD_TOKEN }}
          node-version: ${{ env.NODE_VERSION }}

      - name: Set package version
@@ -216,6 +218,7 @@ jobs:
      - name: Setup build environment
        uses: ./.github/actions/desktop-build-setup
        with:
+          cloud-token: ${{ secrets.LOBEHUB_CLOUD_TOKEN }}
          node-version: ${{ env.NODE_VERSION }}

      - name: Set package version
@@ -54,7 +54,7 @@ jobs:
      - name: Setup Node.js
        uses: actions/setup-node@v6
        with:
-          node-version: 24.11.1
+          node-version: 24.16.0
          package-manager-cache: false

      # 主要逻辑：确定构建版本号
@@ -92,6 +92,7 @@ jobs:
      - name: Setup build environment
        uses: ./.github/actions/desktop-build-setup
        with:
+          cloud-token: ${{ secrets.LOBEHUB_CLOUD_TOKEN }}
          node-version: 24.11.1

      # 设置 package.json 的版本号
@@ -87,6 +87,7 @@ jobs:
      - name: Setup build environment
        uses: ./.github/actions/desktop-build-setup
        with:
+          cloud-token: ${{ secrets.LOBEHUB_CLOUD_TOKEN }}
          node-version: ${{ env.NODE_VERSION }}

      - name: Set package version
@@ -223,6 +223,7 @@ jobs:
      - name: Setup build environment
        uses: ./.github/actions/desktop-build-setup
        with:
+          cloud-token: ${{ secrets.LOBEHUB_CLOUD_TOKEN }}
          node-version: ${{ env.NODE_VERSION }}

      - name: Set package version
@@ -409,7 +410,7 @@ jobs:
      - uses: actions/checkout@v6

      - name: Delete old canary GitHub releases
-        uses: actions/github-script@v7
+        uses: actions/github-script@v8
        with:
          script: |
            const { data: releases } = await github.rest.repos.listReleases({
@@ -180,6 +180,7 @@ jobs:
      - name: Setup build environment
        uses: ./.github/actions/desktop-build-setup
        with:
+          cloud-token: ${{ secrets.LOBEHUB_CLOUD_TOKEN }}
          node-version: ${{ env.NODE_VERSION }}

      - name: Set package version
@@ -28,7 +28,7 @@ jobs:
      - name: Setup Node.js
        uses: actions/setup-node@v6
        with:
-          node-version: 24.11.1
+          node-version: 24.16.0

      - name: Setup pnpm
        uses: pnpm/action-setup@v4
@@ -51,7 +51,7 @@ jobs:
      - name: Setup Node.js
        uses: actions/setup-node@v6
        with:
-          node-version: 24.11.1
+          node-version: 24.16.0
          registry-url: https://registry.npmjs.org

      - name: Setup pnpm
@@ -19,12 +19,6 @@ jobs:
    steps:
      - uses: actions/checkout@v6

-      - name: Clean issue notice
-        uses: actions-cool/issues-helper@e361abf610221f09495ad510cb1e69328d839e1c # v3.7.6
-        with:
-          actions: 'close-issues'
-          labels: '🚨 Sync Fail'
-
      - name: Sync upstream changes
        id: sync
        uses: aormsby/Fork-Sync-With-Upstream-action@v3.4
@@ -33,22 +27,4 @@ jobs:
          upstream_sync_branch: main
          target_sync_branch: main
          target_repo_token: ${{ secrets.GITHUB_TOKEN }} # automatically generated, no need to set
-          test_mode: false
-
-      - name: Sync check
-        if: failure()
-        uses: actions-cool/issues-helper@e361abf610221f09495ad510cb1e69328d839e1c # v3.7.6
-        with:
-          actions: 'create-issue'
-          title: '🚨 同步失败 | Sync Fail'
-          labels: '🚨 Sync Fail'
-          body: |
-            Due to a change in the workflow file of the [LobeChat][lobechat] upstream repository, GitHub has automatically suspended the scheduled automatic update. You need to manually sync your fork. Please refer to the detailed [Tutorial][tutorial-en-US] for instructions.
-
-            由于 [LobeChat][lobechat] 上游仓库的 workflow 文件变更，导致 GitHub 自动暂停了本次自动更新，你需要手动 Sync Fork 一次，请查看 [详细教程][tutorial-zh-CN]
-
-            ![](https://github-production-user-asset-6210df.s3.amazonaws.com/17870709/273954625-df80c890-0822-4ac2-95e6-c990785cbed5.png)
-
-            [lobechat]: https://github.com/lobehub/lobe-chat
-            [tutorial-zh-CN]: https://lobehub.com/zh/docs/self-hosting/advanced/upstream-sync
-            [tutorial-en-US]: https://lobehub.com/docs/self-hosting/advanced/upstream-sync
+          test_mode: false
@@ -32,7 +32,7 @@ jobs:
    runs-on: ubuntu-latest
    name: Test Packages
    env:
-      PACKAGES: '@lobechat/file-loaders @lobechat/prompts @lobechat/model-runtime @lobechat/web-crawler @lobechat/electron-server-ipc @lobechat/utils @lobechat/python-interpreter @lobechat/context-engine @lobechat/agent-runtime @lobechat/conversation-flow @lobechat/ssrf-safe-fetch @lobechat/memory-user-memory @lobechat/types @lobechat/trpc @lobechat/app-config @lobechat/locales @lobechat/env @lobechat/builtin-tool-lobe-agent model-bank @lobechat/agent-gateway-client @lobechat/agent-manager-runtime @lobechat/device-gateway-client @lobechat/device-identity @lobechat/eval-dataset-parser @lobechat/eval-rubric @lobechat/fetch-sse @lobechat/heterogeneous-agents'
+      PACKAGES: '@lobechat/file-loaders @lobechat/prompts @lobechat/model-runtime @lobechat/web-crawler @lobechat/electron-server-ipc @lobechat/utils @lobechat/context-engine @lobechat/agent-runtime @lobechat/conversation-flow @lobechat/ssrf-safe-fetch @lobechat/memory-user-memory @lobechat/types @lobechat/trpc @lobechat/app-config @lobechat/locales @lobechat/env @lobechat/builtin-tool-lobe-agent model-bank @lobechat/agent-gateway-client @lobechat/agent-manager-runtime @lobechat/device-gateway-client @lobechat/device-identity @lobechat/eval-dataset-parser @lobechat/eval-rubric @lobechat/fetch-sse @lobechat/heterogeneous-agents'

    steps:
      - name: Checkout
@@ -90,11 +90,23 @@ jobs:
          for package in $PACKAGES; do
            dir="${package#@lobechat/}"
            if [ -f "./packages/$dir/coverage/lcov.info" ]; then
-              echo "Uploading coverage for $dir..."
+              flag="packages/$dir"
+
+              case "$dir" in
+                builtin-tool-*)
+                  flag="builtin-tools"
+                  ;;
+                locales|env|device-gateway-client)
+                  echo "Skipping Codecov upload for $dir."
+                  continue
+                  ;;
+              esac
+
+              echo "Uploading coverage for $dir as $flag..."
              ./codecov upload-coverage \
                $COMMON_ARGS \
                --file ./packages/$dir/coverage/lcov.info \
-                --flag packages/$dir \
+                --flag "$flag" \
                --disable-search
            fi
          done
@@ -105,8 +117,8 @@ jobs:
    if: needs.check-duplicate-run.outputs.should_skip != 'true'
    strategy:
      matrix:
-        shard: [1, 2, 3]
-    name: Test App (shard ${{ matrix.shard }}/3)
+        shard: [1, 2]
+    name: Test App (shard ${{ matrix.shard }}/2)
    runs-on: ubuntu-latest
    steps:
      - name: Checkout
@@ -126,7 +138,7 @@ jobs:
        run: pnpm install

      - name: Run tests
-        run: bunx vitest --coverage --silent='passed-only' --reporter=default --reporter=blob --shard=${{ matrix.shard }}/3
+        run: bunx vitest --coverage --silent='passed-only' --reporter=default --reporter=blob --shard=${{ matrix.shard }}/2 --exclude '**/apps/server/**'

      - name: Upload blob report
        if: ${{ !cancelled() }}
@@ -219,6 +231,40 @@ jobs:
          files: ./apps/desktop/coverage/lcov.info
          flags: desktop

+  test-server:
+    needs: check-duplicate-run
+    if: needs.check-duplicate-run.outputs.should_skip != 'true'
+    name: Test Server
+
+    runs-on: ubuntu-latest
+
+    steps:
+      - name: Checkout
+        env:
+          REF_SHA: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+          REPOSITORY: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.repo.full_name || github.repository }}
+        run: |
+          git init .
+          git remote add origin "https://github.com/${REPOSITORY}.git"
+          git fetch --no-tags --depth=1 origin "${REF_SHA}"
+          git checkout --force FETCH_HEAD
+
+      - name: Setup environment
+        uses: ./.github/actions/setup-env
+
+      - name: Install deps
+        run: pnpm install
+
+      - name: Test Server Coverage
+        run: bunx vitest --coverage --silent='passed-only' --reporter=default --coverage.reportsDirectory=./apps/server/coverage --dir apps/server
+
+      - name: Upload Server coverage to Codecov
+        uses: codecov/codecov-action@v5
+        with:
+          token: ${{ secrets.CODECOV_TOKEN }}
+          files: ./apps/server/coverage/lcov.info
+          flags: server
+
  test-databsae:
    needs: check-duplicate-run
    if: needs.check-duplicate-run.outputs.should_skip != 'true'
@@ -93,9 +93,6 @@ public/swe-worker*
 # Generated files
 src/app/spa/[variants]/[[...path]]/spaHtmlTemplates.ts
 public/*.js
-public/sitemap.xml
-public/sitemap-index.xml
-sitemap*.xml
 robots.txt

 # Git hooks
@@ -35,7 +35,7 @@
    "@lobechat/local-file-shell": "workspace:*",
    "@lobechat/tool-runtime": "workspace:*",
    "@trpc/client": "^11.8.1",
-    "@types/node": "^22.13.5",
+    "@types/node": "^24.13.2",
    "@types/ws": "^8.18.1",
    "commander": "^13.1.0",
    "dayjs": "^1.11.19",
@@ -302,6 +302,7 @@ async function runConnect(options: ConnectOptions, isDaemonChild: boolean) {
        {
          agentType: request.agentType,
          cwd: request.cwd,
+          imageList: request.imageList,
          jwt: request.jwt,
          operationId: request.operationId,
          prompt: request.prompt,
@@ -650,7 +650,7 @@ describe('hetero exec command', () => {
  });

  it('resets the per-message text accumulator at message boundaries (no cross-message duplication)', async () => {
-    // LOBE-10157 Bug 3: the `replace` snapshot accumulator must not span
+    // The `replace` snapshot accumulator must not span
    // message boundaries. Two assistant messages separated by a
    // stream_end/stream_start boundary must each snapshot only their OWN
    // text — otherwise the second message re-emits the first's text verbatim.
@@ -261,7 +261,7 @@ class SerialServerIngester {
    // adapter's `openMainMessage`) must reset it — otherwise it spans the
    // whole run and every later message's snapshot re-emits all prior
    // messages' text verbatim, which the server then persists into the new
-    // DB message (LOBE-10157 Bug 3: cross-message text duplication). Reset
+    // DB message: cross-message text duplication. Reset
    // AFTER flushing the just-ended message's pending snapshot above.
    if (event.type === 'stream_start' || event.type === 'stream_end') {
      this.accumulatedText = '';
@@ -122,4 +122,24 @@ describe('spawnHeteroAgentRun', () => {
      ]),
    );
  });
+
+  it('appends image blocks to stdin when imageList is provided', async () => {
+    const child = makeFakeChild();
+    spawnMock.mockReturnValue(child);
+
+    const ackPromise = spawnHeteroAgentRun({
+      ...baseParams,
+      imageList: [{ id: 'file-1', url: 'https://signed/a.png' }],
+      prompt: 'look at this',
+    });
+    child.emit('spawn');
+    await ackPromise;
+
+    expect(child.stdin.write).toHaveBeenCalledWith(
+      JSON.stringify([
+        { text: 'look at this', type: 'text' },
+        { source: { id: 'file-1', type: 'url', url: 'https://signed/a.png' }, type: 'image' },
+      ]),
+    );
+  });
 });
@@ -1,8 +1,15 @@
 import { spawn } from 'node:child_process';

+import {
+  buildHeteroExecStdinPayload,
+  type HeteroExecImageRef,
+} from '@lobechat/heterogeneous-agents/protocol';
+
 export interface SpawnHeteroAgentRunParams {
  agentType: string;
  cwd?: string;
+  /** Image attachments (signed URLs) appended as image content blocks. */
+  imageList?: HeteroExecImageRef[];
  jwt: string;
  operationId: string;
  prompt: string;
@@ -46,6 +53,7 @@ export function spawnHeteroAgentRun(
  const {
    agentType,
    cwd,
+    imageList,
    jwt,
    operationId,
    prompt,
@@ -77,15 +85,11 @@ export function spawnHeteroAgentRun(
    ...(resumeSessionId ? ['--resume', resumeSessionId] : []),
  ];

-  // With systemContext, send a content-block array so the agent sees the
-  // context block first, then the user's actual prompt — mirrors the desktop
-  // path. `lh hetero exec` coerces both shapes via coerceJsonPrompt.
-  const stdinPayload = systemContext
-    ? JSON.stringify([
-        { text: systemContext, type: 'text' },
-        { text: prompt, type: 'text' },
-      ])
-    : JSON.stringify(prompt);
+  // systemContext / image attachments turn the payload into a content-block
+  // array: context block first, then the user's prompt, then images — mirrors
+  // the desktop path. `lh hetero exec` coerces both shapes via
+  // coerceJsonPrompt.
+  const stdinPayload = buildHeteroExecStdinPayload({ imageList, prompt, systemContext });

  return new Promise<AgentRunAckResult>((resolve) => {
    let settled = false;
@@ -6,6 +6,7 @@ import dotenv from 'dotenv';
 import { defineConfig } from 'electron-vite';
 import type { PluginOption, ViteDevServer } from 'vite';
 import { loadEnv } from 'vite';
+import tsconfigPaths from 'vite-tsconfig-paths';

 import {
  sharedOptimizeDeps,
@@ -88,10 +89,112 @@ function electronDesktopHtmlPlugin(): PluginOption {
  };
 }

+const CLOUD_DESKTOP_BUSINESS_FEATURES_FLAG = '__LOBECLOUD_DESKTOP_BUSINESS_FEATURES__';
+const BUSINESS_CONST_MODULE_ID = '@lobechat/business-const';
+const CLOUD_BUSINESS_CONST_MODULE_ID = '@cloud/business-const';
+const DYNAMIC_BUSINESS_CONST_QUERY = '?lobe-cloud-desktop-business-const';
+
+const createBusinessFeaturesBootstrapScript = () =>
+  `globalThis[${JSON.stringify(CLOUD_DESKTOP_BUSINESS_FEATURES_FLAG)}] = true;`;
+
+const replaceBusinessFlagExport = (code: string, name: string, initializer: string) => {
+  const pattern = new RegExp(`export\\s+(?:const|let|var)\\s+${name}\\s*=\\s*[\\s\\S]*?;`);
+
+  return {
+    code: code.replace(pattern, `export let ${name} = ${initializer};`),
+    replaced: pattern.test(code),
+  };
+};
+
+const injectDynamicBusinessFeatureFlag = (code: string) => {
+  const businessFlag = replaceBusinessFlagExport(
+    code,
+    'ENABLE_BUSINESS_FEATURES',
+    `Boolean(globalThis['${CLOUD_DESKTOP_BUSINESS_FEATURES_FLAG}'])`,
+  );
+  const topicLinkFlag = replaceBusinessFlagExport(
+    businessFlag.code,
+    'ENABLE_TOPIC_LINK_SHARE',
+    'ENABLE_BUSINESS_FEATURES',
+  );
+
+  if (!businessFlag.replaced) {
+    throw new Error('Cannot find ENABLE_BUSINESS_FEATURES export in @cloud/business-const');
+  }
+
+  const topicLinkAssignment = topicLinkFlag.replaced
+    ? '\n  ENABLE_TOPIC_LINK_SHARE = enabled;'
+    : '';
+
+  return `${topicLinkFlag.code}
+
+const __lobeCloudDesktopBusinessFeaturesFlagKey = '${CLOUD_DESKTOP_BUSINESS_FEATURES_FLAG}';
+const __lobeCloudDesktopApplyBusinessFeaturesFlag = (value) => {
+  const enabled = Boolean(value);
+  ENABLE_BUSINESS_FEATURES = enabled;${topicLinkAssignment}
+  return enabled;
+};
+
+const __lobeCloudDesktopExistingDescriptor = Object.getOwnPropertyDescriptor(
+  globalThis,
+  __lobeCloudDesktopBusinessFeaturesFlagKey,
+);
+const __lobeCloudDesktopInitialValue = __lobeCloudDesktopExistingDescriptor?.get
+  ? __lobeCloudDesktopExistingDescriptor.get.call(globalThis)
+  : globalThis[__lobeCloudDesktopBusinessFeaturesFlagKey];
+
+Object.defineProperty(globalThis, __lobeCloudDesktopBusinessFeaturesFlagKey, {
+  configurable: true,
+  get() {
+    return ENABLE_BUSINESS_FEATURES;
+  },
+  set(value) {
+    __lobeCloudDesktopApplyBusinessFeaturesFlag(value);
+  },
+});
+
+__lobeCloudDesktopApplyBusinessFeaturesFlag(__lobeCloudDesktopInitialValue);
+`;
+};
+
+function cloudDesktopBusinessConstPlugin(): PluginOption {
+  return {
+    enforce: 'pre',
+    async resolveId(id, importer) {
+      if (id !== BUSINESS_CONST_MODULE_ID) return;
+
+      const resolved = await this.resolve(CLOUD_BUSINESS_CONST_MODULE_ID, importer, {
+        skipSelf: true,
+      });
+      if (!resolved) throw new Error(`Cannot resolve ${CLOUD_BUSINESS_CONST_MODULE_ID}`);
+
+      return `${resolved.id}${DYNAMIC_BUSINESS_CONST_QUERY}`;
+    },
+    load(id) {
+      if (!id.endsWith(DYNAMIC_BUSINESS_CONST_QUERY)) return;
+
+      const sourcePath = id.slice(0, -DYNAMIC_BUSINESS_CONST_QUERY.length);
+      return injectDynamicBusinessFeatureFlag(readFileSync(sourcePath, 'utf8'));
+    },
+    name: 'lobe-cloud-desktop-business-const',
+    transformIndexHtml() {
+      return [
+        {
+          children: createBusinessFeaturesBootstrapScript(),
+          injectTo: 'head-prepend',
+          tag: 'script',
+        },
+      ];
+    },
+  };
+}
+
 dotenv.config();

 const isDev = process.env.NODE_ENV === 'development';
 const ROOT_DIR = path.resolve(__dirname, '../..');
+const CLOUD_ROOT_DIR = path.resolve(__dirname, '../../..');
+const isCloudDesktopBuild = process.env.CLOUD_DESKTOP === '1';
 const mode = process.env.NODE_ENV === 'production' ? 'production' : 'development';

 Object.assign(process.env, loadEnv(mode, ROOT_DIR, ''));
@@ -105,8 +208,17 @@ const mainProcessRuntimeExternals = [
  ...externalRuntimeModules,
  'node-mac-permissions',
 ];
+const externalNavigationHosts =
+  process.env.DESKTOP_EXTERNAL_NAVIGATION_HOSTS ?? (isCloudDesktopBuild ? 'stripe.com' : '');

 console.info(`[electron-vite.config.ts] Detected UPDATE_CHANNEL: ${updateChannel}`);
+console.info(`[electron-vite.config.ts] Cloud desktop build: ${isCloudDesktopBuild}`);
+
+const cloudTsconfigPathsPlugin = () =>
+  ({
+    ...tsconfigPaths({ projects: [path.resolve(CLOUD_ROOT_DIR, 'tsconfig.json')] }),
+    name: 'lobe-cloud-desktop-tsconfig-paths',
+  }) satisfies PluginOption;

 export default defineConfig({
  main: {
@@ -169,6 +281,7 @@ export default defineConfig({
      sourcemap: isDev ? 'inline' : false,
    },
    define: {
+      'process.env.DESKTOP_EXTERNAL_NAVIGATION_HOSTS': JSON.stringify(externalNavigationHosts),
      'process.env.UPDATE_CHANNEL': JSON.stringify(process.env.UPDATE_CHANNEL),
      'process.env.UPDATE_SERVER_URL': JSON.stringify(process.env.UPDATE_SERVER_URL),
    },
@@ -214,6 +327,8 @@ export default defineConfig({
    },
    optimizeDeps: sharedOptimizeDeps,
    plugins: [
+      isCloudDesktopBuild && cloudTsconfigPathsPlugin(),
+      isCloudDesktopBuild && cloudDesktopBusinessConstPlugin(),
      forceAbsoluteBasePlugin(),
      electronDesktopHtmlPlugin(),
      vanillaExtractPlugin(),
@@ -221,7 +336,7 @@ export default defineConfig({
    ],
    resolve: {
      dedupe: ['react', 'react-dom'],
-      tsconfigPaths: true,
+      tsconfigPaths: !isCloudDesktopBuild,
    },
    // In dev the BrowserWindow loads `app://renderer/` and the Electron main process
    // proxies non-backend requests to this Vite dev server via `net.fetch`. The HMR
@@ -111,7 +111,7 @@
    "undici": "^7.16.0",
    "uuid": "^14.0.0",
    "vite": "8.0.14",
-    "vitest": "3.2.4",
+    "vitest": "3.2.6",
    "zod": "^3.25.76"
  },
  "optionalDependencies": {
@@ -128,7 +128,7 @@
      "node-gyp": "^12.4.0",
      "react": "19.2.4",
      "react-dom": "19.2.4",
-      "vitest": "3.2.4"
+      "vitest": "3.2.6"
    }
  }
 }
@@ -7,6 +7,7 @@ import { getDesktopEnv } from '@/env';
 export const isDev = electronIs.dev();

 export const OFFICIAL_CLOUD_SERVER = getDesktopEnv().OFFICIAL_CLOUD_SERVER;
+export const DESKTOP_EXTERNAL_NAVIGATION_HOSTS = getDesktopEnv().DESKTOP_EXTERNAL_NAVIGATION_HOSTS;

 export const isMac = electronIs.macOS();
 export const isWindows = electronIs.windows();
@@ -91,6 +91,13 @@ export default class BrowserWindowsCtr extends ControllerModule {
    });
  }

+  @IpcMethod()
+  isWindowFullScreen() {
+    return this.withSenderIdentifier((identifier) => {
+      return this.app.browserManager.isWindowFullScreen(identifier);
+    });
+  }
+
  @IpcMethod()
  setWindowAlwaysOnTop(flag: boolean) {
    this.withSenderIdentifier((identifier) => {
@@ -17,6 +17,7 @@ import type {
  KillCommandParams,
  ListLocalFileParams,
  ListProjectSkillsParams,
+  LocalFilePreviewUrlParams,
  LocalReadFileParams,
  LocalReadFilesParams,
  LocalSearchFilesParams,
@@ -300,6 +301,7 @@ export default class GatewayConnectionCtr extends ControllerModule {
      this.heterogeneousAgentCtr.spawnLhHeteroExec({
        agentType: request.agentType,
        cwd: request.cwd,
+        imageList: request.imageList,
        jwt: request.jwt,
        operationId: request.operationId,
        prompt: request.prompt,
@@ -408,6 +410,10 @@ export default class GatewayConnectionCtr extends ControllerModule {
        return this.localFileCtr.getProjectFileIndex(params as { scope?: string });
      }

+      case 'getLocalFilePreview': {
+        return this.localFileCtr.getLocalFilePreview(params as LocalFilePreviewUrlParams);
+      }
+
      case 'listProjectSkills': {
        return this.workspaceCtr.listProjectSkills(params as ListProjectSkillsParams);
      }
@@ -18,13 +18,20 @@ import {
 } from '@lobechat/electron-client-ipc';
 import type { AskUserBridge } from '@lobechat/heterogeneous-agents/askUser';
 import { AskUserMcpServer } from '@lobechat/heterogeneous-agents/askUser';
-import type { AgentContentBlock } from '@lobechat/heterogeneous-agents/spawn';
+import type {
+  AgentContentBlock,
+  HeteroExecImageRef,
+} from '@lobechat/heterogeneous-agents/protocol';
+import { buildHeteroExecStdinPayload } from '@lobechat/heterogeneous-agents/protocol';
+import type { AgentStreamEvent } from '@lobechat/heterogeneous-agents/spawn';
 import {
  AgentStreamPipeline,
  buildAgentInput,
  materializeImageToPath,
  normalizeImage,
+  readCodexSessionModel,
  resolveCliSpawnPlan,
+  resolveCodexInitialModel,
 } from '@lobechat/heterogeneous-agents/spawn';
 import { app as electronApp, BrowserWindow } from 'electron';

@@ -176,13 +183,32 @@ interface AgentSession {
  command: string;
  cwd?: string;
  env?: Record<string, string>;
+  model?: string;
+  modelSource?: string;
+  modelVerificationLastAttemptAt?: number;
+  modelVerificationLastAttemptSessionId?: string;
  process?: ChildProcess;
  resumeSessionId?: string;
  sessionId: string;
+  verifiedModel?: string;
+  verifiedModelContextWindow?: number;
+  verifiedModelProvider?: string;
+  verifiedModelSessionId?: string;
+  verifiedModelSourceFile?: string;
 }

 type SessionErrorPayload = HeterogeneousAgentSessionError | string;

+interface SpawnPreflightResult {
+  /** Set when the CLI is missing/unavailable — spawning must be aborted. */
+  error?: HeterogeneousAgentSessionError;
+  /**
+   * Absolute path the detector resolved the CLI to, to spawn instead of the
+   * bare `session.command`. Undefined when the bare command should be used.
+   */
+  resolvedCommand?: string;
+}
+
 interface CliTraceSession {
  dir: string;
  writeQueue: Promise<void>;
@@ -435,16 +461,14 @@ export default class HeterogeneousAgentCtr extends ControllerModule {
    return relevantStderr || `Agent exited with code ${code}`;
  }

-  private async getSpawnPreflightError(
-    session: AgentSession,
-  ): Promise<HeterogeneousAgentSessionError | undefined> {
+  private async runSpawnPreflight(session: AgentSession): Promise<SpawnPreflightResult> {
    const defaultCommand =
      session.agentType === 'claude-code'
        ? 'claude'
        : session.agentType === 'codex'
          ? 'codex'
          : undefined;
-    if (!defaultCommand) return;
+    if (!defaultCommand) return {};

    const command = this.resolveSessionCommand(session);
    const status =
@@ -456,9 +480,22 @@ export default class HeterogeneousAgentCtr extends ControllerModule {
          );
    const cliMissingError = this.buildCliMissingError(session);

-    if (!status || status.available || !cliMissingError) return;
+    if (!status || status.available || !cliMissingError) {
+      // The detector resolved the binary to an absolute path (e.g. the Codex.app
+      // bundle, or a login-shell-PATH-only install). spawn() runs against a leaner
+      // env than detection did, so a bare `codex` could still ENOENT even though
+      // preflight passed — feed the resolved absolute path through to spawn.
+      //
+      // Skip on Windows: there `resolveCliSpawnPlan` performs its own `.cmd`/`.exe`
+      // shim resolution from the bare command, which we must not pre-empt.
+      const resolvedCommand =
+        process.platform !== 'win32' && status?.available && status.path?.trim()
+          ? status.path
+          : undefined;
+      return { resolvedCommand };
+    }

-    return cliMissingError;
+    return { error: cliMissingError };
  }

  private get shouldTraceCliOutput(): boolean {
@@ -581,12 +618,19 @@ export default class HeterogeneousAgentCtr extends ControllerModule {
            createdAt: createdAt.toISOString(),
            cwd,
            envKeys: session.env ? Object.keys(session.env).sort() : [],
+            model: session.model,
+            modelSource: session.modelSource,
            resumeSessionId: session.resumeSessionId,
            sessionId: session.sessionId,
            stdinBytes: stdinPayload === undefined ? 0 : Buffer.byteLength(stdinPayload),
            stdinFile: stdinPayload === undefined ? undefined : 'stdin.txt',
            stderrFile: 'stderr.log',
            stdoutFile: 'stdout.jsonl',
+            verifiedModel: session.verifiedModel,
+            verifiedModelContextWindow: session.verifiedModelContextWindow,
+            verifiedModelProvider: session.verifiedModelProvider,
+            verifiedModelSessionId: session.verifiedModelSessionId,
+            verifiedModelSourceFile: session.verifiedModelSourceFile,
          },
          null,
          2,
@@ -865,13 +909,13 @@ export default class HeterogeneousAgentCtr extends ControllerModule {
    const session = this.sessions.get(params.sessionId);
    if (!session) throw new Error(`Session not found: ${params.sessionId}`);

-    const preflightError = await this.getSpawnPreflightError(session);
-    if (preflightError) {
+    const preflight = await this.runSpawnPreflight(session);
+    if (preflight.error) {
      this.broadcast('heteroAgentSessionError', {
-        error: preflightError,
+        error: preflight.error,
        sessionId: session.sessionId,
      });
-      throw new Error(preflightError.message);
+      throw new Error(preflight.error.message);
    }

    // Stand up the AskUserQuestion MCP bridge for claude-code prompts BEFORE
@@ -888,6 +932,7 @@ export default class HeterogeneousAgentCtr extends ControllerModule {
    let spawnPlan;
    let traceSession;
    let cwd: string;
+    let spawnEnv: NodeJS.ProcessEnv;
    try {
      const driver = getHeterogeneousAgentDriver(session.agentType);
      spawnPlan = await driver.buildSpawnPlan({
@@ -906,6 +951,23 @@ export default class HeterogeneousAgentCtr extends ControllerModule {
      // Fall back to the user's Desktop so the process never inherits
      // the Electron parent's cwd (which is `/` when launched from Finder).
      cwd = session.cwd || electronApp.getPath('desktop');
+
+      // Forward the user's proxy settings to the CLI. The main-process undici
+      // dispatcher doesn't reach child processes — they need env vars.
+      const proxyEnv = buildProxyEnv(this.app.storeManager.get('networkProxy'));
+      spawnEnv = { ...buildInheritedSpawnEnv(), ...proxyEnv, ...session.env };
+
+      if (session.agentType === 'codex') {
+        const initialModel = await resolveCodexInitialModel({
+          args: spawnPlan.args,
+          env: spawnEnv,
+        });
+        if (initialModel?.model) {
+          session.model = initialModel.model;
+          session.modelSource = initialModel.source;
+        }
+      }
+
      traceSession = await this.createCliTraceSession({
        cliArgs: spawnPlan.args,
        cwd,
@@ -925,7 +987,10 @@ export default class HeterogeneousAgentCtr extends ControllerModule {
    }
    const useStdin = spawnPlan.stdinPayload !== undefined;
    const cliArgs = spawnPlan.args;
-    const resolvedCliSpawnPlan = await resolveCliSpawnPlan(session.command, cliArgs);
+    const resolvedCliSpawnPlan = await resolveCliSpawnPlan(
+      preflight.resolvedCommand || session.command,
+      cliArgs,
+    );

    logger.info(
      'Spawning agent:',
@@ -940,29 +1005,27 @@ export default class HeterogeneousAgentCtr extends ControllerModule {
    // the claude binary can leave bash/grep/etc. tool children running and
    // the CLI hung waiting on them. Windows has different semantics — use
    // taskkill /T /F there; no detached flag needed.
-    // Forward the user's proxy settings to the CLI. The main-process undici
-    // dispatcher doesn't reach child processes — they need env vars.
-    const proxyEnv = buildProxyEnv(this.app.storeManager.get('networkProxy'));
-
    const spawnOptions = {
      cwd,
      detached: process.platform !== 'win32',
      // Strip host Anthropic creds from the inherited env so a developer's
      // shell `ANTHROPIC_API_KEY` can't hijack the CLI's own auth. `session.env`
      // is spread last, so an agent that explicitly configures a key still wins.
-      env: { ...buildInheritedSpawnEnv(), ...proxyEnv, ...session.env },
+      env: spawnEnv,
      stdio: [useStdin ? 'pipe' : 'ignore', 'pipe', 'pipe'] as ['pipe' | 'ignore', 'pipe', 'pipe'],
    };

    return new Promise<void>((resolve, reject) => {
      const proc = spawn(resolvedCliSpawnPlan.command, resolvedCliSpawnPlan.args, spawnOptions);
      this.handleSpawnedAgentProcess({
+        cwd,
        intervention,
        params,
        proc,
        reject,
        resolve,
        session,
+        spawnEnv,
        traceSession,
        useStdin,
        spawnPlan,
@@ -970,23 +1033,86 @@ export default class HeterogeneousAgentCtr extends ControllerModule {
    });
  }

+  private async verifyCodexSessionModel({
+    env,
+    pipeline,
+    session,
+    traceSession,
+  }: {
+    env: NodeJS.ProcessEnv;
+    pipeline: AgentStreamPipeline;
+    session: AgentSession;
+    traceSession: CliTraceSession | undefined;
+  }): Promise<AgentStreamEvent[]> {
+    if (
+      session.agentType !== 'codex' ||
+      !pipeline.sessionId ||
+      session.verifiedModelSessionId === pipeline.sessionId
+    ) {
+      return [];
+    }
+
+    const now = Date.now();
+    if (
+      session.modelVerificationLastAttemptSessionId === pipeline.sessionId &&
+      session.modelVerificationLastAttemptAt &&
+      now - session.modelVerificationLastAttemptAt < 1000
+    ) {
+      return [];
+    }
+    session.modelVerificationLastAttemptSessionId = pipeline.sessionId;
+    session.modelVerificationLastAttemptAt = now;
+
+    const sessionModel = await readCodexSessionModel(pipeline.sessionId, { env });
+    if (!sessionModel?.model) return [];
+
+    const previousModel = session.model;
+    session.verifiedModel = sessionModel.model;
+    session.verifiedModelContextWindow = sessionModel.contextWindow;
+    session.verifiedModelProvider = sessionModel.provider;
+    session.verifiedModelSessionId = pipeline.sessionId;
+    session.verifiedModelSourceFile = sessionModel.sourceFile;
+
+    void this.writeCliTraceJson(traceSession, 'model.json', {
+      initialModel: previousModel,
+      initialModelSource: session.modelSource,
+      sessionId: pipeline.sessionId,
+      verifiedAt: new Date().toISOString(),
+      verifiedContextWindow: sessionModel.contextWindow,
+      verifiedLine: sessionModel.line,
+      verifiedModel: sessionModel.model,
+      verifiedModelProvider: sessionModel.provider,
+      verifiedSourceFile: sessionModel.sourceFile,
+    });
+
+    if (previousModel === sessionModel.model) return [];
+
+    session.model = sessionModel.model;
+    session.modelSource = 'codex-session';
+    return pipeline.configureSession({ model: sessionModel.model });
+  }
+
  private handleSpawnedAgentProcess({
+    cwd,
    intervention,
    params,
    proc,
    reject,
    resolve,
    session,
+    spawnEnv,
    spawnPlan,
    traceSession,
    useStdin,
  }: {
+    cwd: string;
    intervention?: Awaited<ReturnType<HeterogeneousAgentCtr['setupInterventionForOp']>>;
    params: SendPromptParams;
    proc: ChildProcess;
    reject: (reason?: unknown) => void;
    resolve: () => void;
    session: AgentSession;
+    spawnEnv: NodeJS.ProcessEnv;
    spawnPlan: HeterogeneousAgentBuildPlan;
    traceSession: CliTraceSession | undefined;
    useStdin: boolean;
@@ -1021,10 +1147,12 @@ export default class HeterogeneousAgentCtr extends ControllerModule {
    // toStreamEvent all run inside the shared pipeline, so renderer + future
    // server `heteroIngest` see the same `AgentStreamEvent` wire shape with
    // no per-consumer adapter. The pipeline auto-wires the Codex
-    // file-change line-stat tracker when `agentType === 'codex'`, so this
+    // file-change diff/stat tracker when `agentType === 'codex'`, so this
    // controller stays agent-agnostic.
    const pipeline = new AgentStreamPipeline({
      agentType: session.agentType,
+      cwd,
+      initialModel: session.model,
      operationId: params.operationId,
    });
    let stdoutBroadcastQueue: Promise<void> = Promise.resolve();
@@ -1039,6 +1167,14 @@ export default class HeterogeneousAgentCtr extends ControllerModule {
          if (pipeline.sessionId && pipeline.sessionId !== session.agentSessionId) {
            session.agentSessionId = pipeline.sessionId;
          }
+          events.push(
+            ...(await this.verifyCodexSessionModel({
+              env: spawnEnv,
+              pipeline,
+              session,
+              traceSession,
+            })),
+          );
          for (const event of events) {
            this.broadcast('heteroAgentEvent', {
              event,
@@ -1317,6 +1453,8 @@ export default class HeterogeneousAgentCtr extends ControllerModule {
  spawnLhHeteroExec(params: {
    agentType: string;
    cwd?: string;
+    /** Image attachments (signed URLs) appended as image content blocks. */
+    imageList?: HeteroExecImageRef[];
    jwt: string;
    operationId: string;
    prompt: string;
@@ -1328,6 +1466,7 @@ export default class HeterogeneousAgentCtr extends ControllerModule {
    const {
      agentType,
      cwd,
+      imageList,
      jwt,
      operationId,
      prompt,
@@ -1380,16 +1519,11 @@ export default class HeterogeneousAgentCtr extends ControllerModule {
      stdio: ['pipe', 'inherit', 'inherit'],
    });

-    // When systemContext is provided, send a content-block array so CC sees the
-    // context block first, then the user's actual message — mirrors
-    // spawnHeteroSandbox. lh handles JSON arrays via coerceJsonPrompt, so no lh
-    // changes are required.
-    const stdinPayload = systemContext
-      ? JSON.stringify([
-          { text: systemContext, type: 'text' },
-          { text: prompt, type: 'text' },
-        ])
-      : JSON.stringify(prompt);
+    // systemContext / image attachments turn the payload into a content-block
+    // array so CC sees the context block first, then the user's message, then
+    // the images — mirrors spawnHeteroSandbox. lh handles both shapes via
+    // coerceJsonPrompt, so no lh changes are required.
+    const stdinPayload = buildHeteroExecStdinPayload({ imageList, prompt, systemContext });
    child.stdin.write(stdinPayload);
    child.stdin.end();

@@ -12,6 +12,7 @@ import {
  type GrepContentParams,
  type GrepContentResult,
  type ListLocalFileParams,
+  type LocalFilePreviewResult,
  type LocalFilePreviewUrlParams,
  type LocalFilePreviewUrlResult,
  type LocalMoveFilesResultItem,
@@ -65,6 +66,19 @@ const logger = createLogger('controllers:LocalFileCtr');

 const SAFE_PATH_PREFIXES = ['/tmp', '/var/tmp'] as const;

+const TEXT_PREVIEW_MIME_TYPES = new Set([
+  'application/graphql',
+  'application/javascript',
+  'application/json',
+  'application/markdown',
+  'application/toml',
+  'application/xml',
+  'application/yaml',
+  'text/markdown',
+  'text/mdx',
+  'text/x-markdown',
+]);
+
 const normalizeAbsolutePath = (inputPath: string): string =>
  path.normalize(path.isAbsolute(inputPath) ? inputPath : `/${inputPath}`);

@@ -91,6 +105,48 @@ const resolveNearestExistingRealPath = async (targetPath: string): Promise<strin

 const toPosixRelativePath = (filePath: string) => filePath.split(path.sep).join('/');

+const normalizeContentType = (contentType: string): string =>
+  contentType.split(';')[0].trim().toLowerCase();
+
+const isTextPreviewMimeType = (mimeType: string): boolean =>
+  mimeType.startsWith('text/') || TEXT_PREVIEW_MIME_TYPES.has(mimeType);
+
+const serializePreviewFile = ({
+  buffer,
+  contentType,
+}: {
+  buffer: Buffer;
+  contentType: string;
+}): NonNullable<LocalFilePreviewResult['preview']> => {
+  const normalizedContentType = normalizeContentType(contentType);
+
+  if (normalizedContentType.startsWith('image/')) {
+    return {
+      base64: buffer.toString('base64'),
+      contentType: normalizedContentType,
+      type: 'image',
+    };
+  }
+
+  if (isTextPreviewMimeType(normalizedContentType)) {
+    return {
+      content: buffer.toString('utf8'),
+      contentType: normalizedContentType,
+      type: 'text',
+    };
+  }
+
+  if (normalizedContentType === 'application/pdf') {
+    return { contentType: normalizedContentType, type: 'pdf' };
+  }
+
+  if (normalizedContentType.startsWith('video/')) {
+    return { contentType: normalizedContentType, type: 'video' };
+  }
+
+  return { contentType: normalizedContentType, type: 'binary' };
+};
+
 const createProjectFileEntry = (
  root: string,
  absolutePath: string,
@@ -381,11 +437,13 @@ export default class LocalFileCtr extends ControllerModule {

  @IpcMethod()
  async getLocalFilePreviewUrl({
+    accept,
    path: filePath,
    workingDirectory,
  }: LocalFilePreviewUrlParams): Promise<LocalFilePreviewUrlResult> {
    try {
      const url = await this.app.localFileProtocolManager.createPreviewUrl({
+        accept,
        filePath,
        workspaceRoot: workingDirectory,
      });
@@ -401,6 +459,33 @@ export default class LocalFileCtr extends ControllerModule {
    }
  }

+  @IpcMethod()
+  async getLocalFilePreview({
+    accept,
+    path: filePath,
+    workingDirectory,
+  }: LocalFilePreviewUrlParams): Promise<LocalFilePreviewResult> {
+    try {
+      const preview = await this.app.localFileProtocolManager.readPreviewFile({
+        accept,
+        filePath,
+        workspaceRoot: workingDirectory,
+      });
+
+      if (!preview) {
+        return { error: 'File is outside the approved workspace', success: false };
+      }
+
+      return {
+        preview: serializePreviewFile(preview),
+        success: true,
+      };
+    } catch (error) {
+      logger.error('Failed to read local file preview:', error);
+      return { error: (error as Error).message, success: false };
+    }
+  }
+
  @IpcMethod()
  async handlePrepareSkillDirectory({
    forceRefresh,
@@ -29,6 +29,7 @@ const mockCloseWindow = vi.fn();
 const mockMinimizeWindow = vi.fn();
 const mockMaximizeWindow = vi.fn();
 const mockIsWindowMaximized = vi.fn();
+const mockIsWindowFullScreen = vi.fn();
 const mockRetrieveByIdentifier = vi.fn();
 const mockStartSession = vi.fn();
 const testSenderIdentifierString: string = 'test-window-event-id';
@@ -58,6 +59,7 @@ const mockApp = {
    minimizeWindow: mockMinimizeWindow,
    maximizeWindow: mockMaximizeWindow,
    isWindowMaximized: mockIsWindowMaximized,
+    isWindowFullScreen: mockIsWindowFullScreen,
    retrieveByIdentifier: mockRetrieveByIdentifier.mockImplementation(
      (identifier: AppBrowsersIdentifiers | string) => {
        if (identifier === 'some-other-window') {
@@ -166,6 +168,20 @@ describe('BrowserWindowsCtr', () => {
    });
  });

+  describe('isWindowFullScreen', () => {
+    it('should return fullscreen state for the sender window', () => {
+      mockIsWindowFullScreen.mockReturnValueOnce(true);
+
+      const sender = {} as any;
+      const context = { sender, event: { sender } as any } as IpcContext;
+      const result = runWithIpcContext(context, () => browserWindowsCtr.isWindowFullScreen());
+
+      expect(mockGetIdentifierByWebContents).toHaveBeenCalledWith(context.sender);
+      expect(mockIsWindowFullScreen).toHaveBeenCalledWith(testSenderIdentifierString);
+      expect(result).toBe(true);
+    });
+  });
+
  describe('interceptRoute', () => {
    const baseParams = { source: 'link-click' as const };

@@ -442,6 +442,28 @@ describe('HeterogeneousAgentCtr', () => {
      expect(spawnCalls).toHaveLength(0);
    });

+    it('spawns the detector-resolved absolute path (e.g. Codex.app bundle) instead of the bare command', async () => {
+      // A user with only the Codex desktop app has no `codex` on PATH; the
+      // detector resolves the bundled binary. Preflight passing isn't enough —
+      // spawn must target that absolute path or it ENOENTs on the bare command.
+      const bundlePath = '/Applications/Codex.app/Contents/Resources/codex';
+      const detect = vi.fn().mockResolvedValue({ available: true, path: bundlePath });
+      const { proc } = createFakeProc({ stdoutLines: [] });
+      nextFakeProc = proc;
+
+      const ctr = new HeterogeneousAgentCtr({
+        appStoragePath,
+        storeManager: { get: vi.fn() },
+        toolDetectorManager: { detect },
+      } as any);
+      const { sessionId } = await ctr.startSession({ agentType: 'codex', command: 'codex' });
+      await ctr.sendPrompt({ operationId: 'op-test', prompt: 'hello', sessionId });
+
+      expect(detect).toHaveBeenCalledWith('codex', true);
+      expect(spawnCalls).toHaveLength(1);
+      expect(spawnCalls[0].command).toBe(bundlePath);
+    });
+
    it('fails fast when a customized Claude command is unavailable instead of checking the default detector', async () => {
      execFileMock.mockImplementation(
        (
@@ -1,3 +1,5 @@
+import path from 'node:path';
+
 import { zipSync } from 'fflate';
 import { beforeEach, describe, expect, it, vi } from 'vitest';

@@ -88,6 +90,7 @@ const mockLocalFileProtocolManager = {
  approveIndexedProjectRoot: vi.fn(),
  approveProjectRootFromScope: vi.fn(),
  createPreviewUrl: vi.fn(),
+  readPreviewFile: vi.fn(),
 };

 // Mock makeSureDirExist
@@ -146,7 +149,6 @@ describe('LocalFileCtr', () => {

    it('should expand a leading ~ to the user home directory', async () => {
      const os = await import('node:os');
-      const path = await import('node:path');
      vi.mocked(mockShell.openPath).mockResolvedValue('');

      const result = await localFileCtr.handleOpenLocalFile({ path: '~/git/work/file.txt' });
@@ -171,7 +173,6 @@ describe('LocalFileCtr', () => {

    it('should expand a leading ~ when opening a directory', async () => {
      const os = await import('node:os');
-      const path = await import('node:path');
      vi.mocked(mockShell.openPath).mockResolvedValue('');

      const result = await localFileCtr.handleOpenLocalFolder({
@@ -224,6 +225,7 @@ describe('LocalFileCtr', () => {
      });

      expect(mockLocalFileProtocolManager.createPreviewUrl).toHaveBeenCalledWith({
+        accept: undefined,
        filePath: '/workspace/app.ts',
        workspaceRoot: '/workspace',
      });
@@ -246,6 +248,99 @@ describe('LocalFileCtr', () => {
        success: false,
      });
    });
+
+    it('should forward image-only preview URL constraints', async () => {
+      mockLocalFileProtocolManager.createPreviewUrl.mockResolvedValue(
+        'localfile://file/workspace/image.png?token=abc',
+      );
+
+      const result = await localFileCtr.getLocalFilePreviewUrl({
+        accept: 'image',
+        path: '/workspace/image.png',
+        workingDirectory: '/workspace',
+      });
+
+      expect(mockLocalFileProtocolManager.createPreviewUrl).toHaveBeenCalledWith({
+        accept: 'image',
+        filePath: '/workspace/image.png',
+        workspaceRoot: '/workspace',
+      });
+      expect(result).toEqual({
+        success: true,
+        url: 'localfile://file/workspace/image.png?token=abc',
+      });
+    });
+  });
+
+  describe('getLocalFilePreview', () => {
+    it('should return text preview content for an approved workspace file', async () => {
+      mockLocalFileProtocolManager.readPreviewFile.mockResolvedValue({
+        buffer: Buffer.from('const value = 1;'),
+        contentType: 'text/plain; charset=utf-8',
+        realPath: '/workspace/app.ts',
+      });
+
+      const result = await localFileCtr.getLocalFilePreview({
+        path: '/workspace/app.ts',
+        workingDirectory: '/workspace',
+      });
+
+      expect(mockLocalFileProtocolManager.readPreviewFile).toHaveBeenCalledWith({
+        accept: undefined,
+        filePath: '/workspace/app.ts',
+        workspaceRoot: '/workspace',
+      });
+      expect(result).toEqual({
+        preview: {
+          content: 'const value = 1;',
+          contentType: 'text/plain',
+          type: 'text',
+        },
+        success: true,
+      });
+    });
+
+    it('should reject preview payload creation outside an approved workspace', async () => {
+      mockLocalFileProtocolManager.readPreviewFile.mockResolvedValue(null);
+
+      const result = await localFileCtr.getLocalFilePreview({
+        path: '/Users/alice/.ssh/id_rsa',
+        workingDirectory: '/workspace',
+      });
+
+      expect(result).toEqual({
+        error: 'File is outside the approved workspace',
+        success: false,
+      });
+    });
+
+    it('should forward image-only preview read constraints', async () => {
+      mockLocalFileProtocolManager.readPreviewFile.mockResolvedValue({
+        buffer: Buffer.from('image-bytes'),
+        contentType: 'image/png',
+        realPath: '/workspace/image.png',
+      });
+
+      const result = await localFileCtr.getLocalFilePreview({
+        accept: 'image',
+        path: '/workspace/image.png',
+        workingDirectory: '/workspace',
+      });
+
+      expect(mockLocalFileProtocolManager.readPreviewFile).toHaveBeenCalledWith({
+        accept: 'image',
+        filePath: '/workspace/image.png',
+        workspaceRoot: '/workspace',
+      });
+      expect(result).toEqual({
+        preview: {
+          base64: Buffer.from('image-bytes').toString('base64'),
+          contentType: 'image/png',
+          type: 'image',
+        },
+        success: true,
+      });
+    });
  });

  describe('handleWriteFile', () => {
@@ -7,7 +7,7 @@ import type { BrowserWindowConstructorOptions } from 'electron';
 import { app, BrowserWindow, ipcMain, screen, session as electronSession, shell } from 'electron';

 import { preloadDir, resourcesDir } from '@/const/dir';
-import { isMac } from '@/const/env';
+import { DESKTOP_EXTERNAL_NAVIGATION_HOSTS, isMac } from '@/const/env';
 import RemoteServerConfigCtr from '@/controllers/RemoteServerConfigCtr';
 import { backendProxyProtocolManager } from '@/core/infrastructure/BackendProxyProtocolManager';
 import { appendVercelCookie, setResponseHeader } from '@/utils/http-headers';
@@ -19,6 +19,31 @@ import { WindowThemeManager } from './WindowThemeManager';

 const logger = createLogger('core:Browser');

+const getExternalNavigationHosts = () =>
+  DESKTOP_EXTERNAL_NAVIGATION_HOSTS.split(',')
+    .map((host) => host.trim().toLowerCase())
+    .filter(Boolean);
+
+const shouldOpenTopLevelNavigationExternally = (rawUrl: string) => {
+  const externalNavigationHosts = getExternalNavigationHosts();
+  if (externalNavigationHosts.length === 0) return false;
+
+  let url: URL;
+  try {
+    url = new URL(rawUrl);
+  } catch {
+    return false;
+  }
+
+  if (url.protocol !== 'http:' && url.protocol !== 'https:') return false;
+
+  const hostname = url.hostname.toLowerCase();
+
+  return externalNavigationHosts.some(
+    (externalHost) => hostname === externalHost || hostname.endsWith(`.${externalHost}`),
+  );
+};
+
 // ==================== Types ====================

 export interface BrowserWindowOpts extends BrowserWindowConstructorOptions {
@@ -194,10 +219,27 @@ export default class Browser {
    this.setupReadyToShowListener(browserWindow);
    this.setupCloseListener(browserWindow);
    this.setupFocusListener(browserWindow);
+    this.setupFullscreenListener(browserWindow);
+    this.setupTopLevelNavigationListener(browserWindow);
    this.setupWillPreventUnloadListener(browserWindow);
    this.setupContextMenu(browserWindow);
  }

+  private setupTopLevelNavigationListener(browserWindow: BrowserWindow): void {
+    logger.debug(`[${this.identifier}] Setting up top-level navigation listener.`);
+
+    browserWindow.webContents.on('will-navigate', (event, url) => {
+      if (!shouldOpenTopLevelNavigationExternally(url)) return;
+
+      logger.info(`[${this.identifier}] Opening top-level navigation externally: ${url}`);
+      event.preventDefault();
+
+      shell.openExternal(url).catch((error) => {
+        logger.error(`[${this.identifier}] Failed to open external navigation URL: ${url}`, error);
+      });
+    });
+  }
+
  /**
   * Setup window open handler to intercept external links
   * Prevents opening new windows in renderer and uses system browser instead
@@ -268,6 +310,18 @@ export default class Browser {
    });
  }

+  private setupFullscreenListener(browserWindow: BrowserWindow): void {
+    logger.debug(`[${this.identifier}] Setting up fullscreen event listeners.`);
+
+    browserWindow.on('enter-full-screen', () => {
+      this.broadcast('windowFullscreenChanged', { isFullScreen: true });
+    });
+
+    browserWindow.on('leave-full-screen', () => {
+      this.broadcast('windowFullscreenChanged', { isFullScreen: false });
+    });
+  }
+
  /**
   * Setup context menu with platform-specific features
   * Delegates to MenuManager for consistent platform behavior
@@ -368,6 +368,11 @@ export class BrowserManager {
    return browser?.browserWindow.isMaximized() ?? false;
  }

+  isWindowFullScreen(identifier: string) {
+    const browser = this.browsers.get(identifier);
+    return browser?.browserWindow.isFullScreen() ?? false;
+  }
+
  setWindowSize(identifier: string, size: { height?: number; width?: number }) {
    const browser = this.browsers.get(identifier);
    browser?.setWindowSize(size);
@@ -9,6 +9,7 @@ const {
  mockBrowserWindow,
  mockNativeTheme,
  mockIpcMain,
+  mockShell,
  mockScreen,
  MockBrowserWindow,
  mockEnv,
@@ -64,6 +65,7 @@ const {
    MockBrowserWindow: vi.fn().mockImplementation(() => mockBrowserWindow),
    mockBrowserWindow,
    mockEnv: {
+      externalNavigationHosts: '',
      isDev: false,
      isLinux: false,
      isMac: false,
@@ -91,6 +93,9 @@ const {
        workArea: { height: 1080, width: 1920, x: 0, y: 0 },
      }),
    },
+    mockShell: {
+      openExternal: vi.fn().mockResolvedValue(undefined),
+    },
  };
 });

@@ -101,6 +106,7 @@ vi.mock('electron', () => ({
  ipcMain: mockIpcMain,
  nativeTheme: mockNativeTheme,
  screen: mockScreen,
+  shell: mockShell,
 }));

 // Mock logger
@@ -121,6 +127,9 @@ vi.mock('@/const/dir', () => ({
 }));

 vi.mock('@/const/env', () => ({
+  get DESKTOP_EXTERNAL_NAVIGATION_HOSTS() {
+    return mockEnv.externalNavigationHosts;
+  },
  get isDev() {
    return mockEnv.isDev;
  },
@@ -182,6 +191,7 @@ describe('Browser', () => {
    mockEnv.isMac = false;
    mockEnv.isMacTahoe = false;
    mockEnv.isWindows = true;
+    mockEnv.externalNavigationHosts = '';

    // Create mock App
    mockStoreManagerGet = vi.fn().mockReturnValue(undefined);
@@ -531,6 +541,30 @@ describe('Browser', () => {
      });
    });

+    describe('fullscreen events', () => {
+      it('should broadcast fullscreen state changes', () => {
+        const enterHandler = mockBrowserWindow.on.mock.calls.find(
+          (call) => call[0] === 'enter-full-screen',
+        )?.[1];
+        const leaveHandler = mockBrowserWindow.on.mock.calls.find(
+          (call) => call[0] === 'leave-full-screen',
+        )?.[1];
+
+        expect(enterHandler).toBeDefined();
+        expect(leaveHandler).toBeDefined();
+
+        enterHandler();
+        expect(mockBrowserWindow.webContents.send).toHaveBeenCalledWith('windowFullscreenChanged', {
+          isFullScreen: true,
+        });
+
+        leaveHandler();
+        expect(mockBrowserWindow.webContents.send).toHaveBeenCalledWith('windowFullscreenChanged', {
+          isFullScreen: false,
+        });
+      });
+    });
+
    describe('close', () => {
      it('should close window', () => {
        browser.close();
@@ -730,4 +764,38 @@ describe('Browser', () => {
      expect(mockEvent.preventDefault).not.toHaveBeenCalled();
    });
  });
+
+  describe('top-level navigation handling', () => {
+    let willNavigateHandler: (event: any, url: string) => void;
+
+    beforeEach(() => {
+      willNavigateHandler = mockBrowserWindow.webContents.on.mock.calls.find(
+        (call) => call[0] === 'will-navigate',
+      )?.[1];
+    });
+
+    it('should open configured external navigation hosts in system browser', () => {
+      mockEnv.externalNavigationHosts = 'stripe.com';
+      const mockEvent = { preventDefault: vi.fn() };
+
+      expect(willNavigateHandler).toBeDefined();
+      willNavigateHandler(mockEvent, 'https://checkout.stripe.com/c/pay/session_id');
+
+      expect(mockEvent.preventDefault).toHaveBeenCalled();
+      expect(mockShell.openExternal).toHaveBeenCalledWith(
+        'https://checkout.stripe.com/c/pay/session_id',
+      );
+    });
+
+    it('should allow internal result routes in the app window', () => {
+      mockEnv.externalNavigationHosts = 'stripe.com';
+      const mockEvent = { preventDefault: vi.fn() };
+
+      expect(willNavigateHandler).toBeDefined();
+      willNavigateHandler(mockEvent, 'http://localhost:3000/payment/upgrade-success');
+
+      expect(mockEvent.preventDefault).not.toHaveBeenCalled();
+      expect(mockShell.openExternal).not.toHaveBeenCalled();
+    });
+  });
 });
@@ -48,6 +48,27 @@ interface PreviewTokenRecord {
  realPath: string;
 }

+export interface PreviewFileReadResult {
+  buffer: Buffer;
+  contentType: string;
+  realPath: string;
+}
+
+type PreviewFileAccept = 'image';
+
+const normalizeContentType = (contentType: string): string =>
+  contentType.split(';')[0].trim().toLowerCase();
+
+const isAcceptedPreviewContentType = (
+  contentType: string,
+  accept?: PreviewFileAccept,
+): boolean => {
+  if (!accept) return true;
+
+  const normalizedContentType = normalizeContentType(contentType);
+  return accept === 'image' && normalizedContentType.startsWith('image/');
+};
+
 /**
 * Custom `localfile://` protocol for project file previews.
 *
@@ -207,43 +228,65 @@ export class LocalFileProtocolManager {
  }

  async createPreviewUrl({
+    accept,
    filePath,
    workspaceRoot,
  }: {
+    accept?: PreviewFileAccept;
    filePath: string;
    workspaceRoot: string;
  }): Promise<string | null> {
    const normalizedFilePath = normalizeAbsolutePath(filePath);
-    const normalizedWorkspaceRoot = normalizeAbsolutePath(workspaceRoot);
-    if (!normalizedFilePath || !normalizedWorkspaceRoot) return null;
+    if (!normalizedFilePath) return null;

-    const [realFilePath, realWorkspaceRoot] = await Promise.all([
-      realpath(normalizedFilePath),
-      realpath(normalizedWorkspaceRoot),
-    ]);
-    const normalizedRealFilePath = normalizeAbsolutePath(realFilePath);
-    const normalizedRealWorkspaceRoot = normalizeAbsolutePath(realWorkspaceRoot);
-
-    if (!normalizedRealFilePath || !normalizedRealWorkspaceRoot) return null;
-    if (
-      !this.approvedWorkspaceRoots.has(normalizedRealWorkspaceRoot) &&
-      !this.indexedProjectRoots.has(normalizedRealWorkspaceRoot)
-    ) {
-      return null;
-    }
-    if (!isPathWithinRoot(normalizedRealFilePath, normalizedRealWorkspaceRoot)) return null;
+    const realFilePath = accept
+      ? (
+          await this.readPreviewFile({
+            accept,
+            filePath,
+            workspaceRoot,
+          })
+        )?.realPath
+      : await this.resolveApprovedPreviewPath({ filePath, workspaceRoot });
+    if (!realFilePath) return null;

    this.cleanupExpiredTokens();

    const token = randomUUID();
    this.previewTokens.set(token, {
      expiresAt: Date.now() + PREVIEW_TOKEN_TTL_MS,
-      realPath: normalizedRealFilePath,
+      realPath: realFilePath,
    });

    return buildLocalFileUrl(normalizedFilePath, token);
  }

+  async readPreviewFile({
+    accept,
+    filePath,
+    workspaceRoot,
+  }: {
+    accept?: PreviewFileAccept;
+    filePath: string;
+    workspaceRoot: string;
+  }): Promise<PreviewFileReadResult | null> {
+    const realFilePath = await this.resolveApprovedPreviewPath({ filePath, workspaceRoot });
+    if (!realFilePath) return null;
+
+    const fileStat = await stat(realFilePath);
+    if (!fileStat.isFile()) return null;
+
+    const buffer = await readFile(realFilePath);
+    const contentType = resolveLocalFileMimeType(realFilePath, buffer);
+    if (!isAcceptedPreviewContentType(contentType, accept)) return null;
+
+    return {
+      buffer,
+      contentType,
+      realPath: realFilePath,
+    };
+  }
+
  /**
   * Decode the URL pathname back into an absolute filesystem path.
   *
@@ -283,6 +326,36 @@ export class LocalFileProtocolManager {
    return normalized;
  }

+  private async resolveApprovedPreviewPath({
+    filePath,
+    workspaceRoot,
+  }: {
+    filePath: string;
+    workspaceRoot: string;
+  }): Promise<string | null> {
+    const normalizedFilePath = normalizeAbsolutePath(filePath);
+    const normalizedWorkspaceRoot = normalizeAbsolutePath(workspaceRoot);
+    if (!normalizedFilePath || !normalizedWorkspaceRoot) return null;
+
+    const [realFilePath, realWorkspaceRoot] = await Promise.all([
+      realpath(normalizedFilePath),
+      realpath(normalizedWorkspaceRoot),
+    ]);
+    const normalizedRealFilePath = normalizeAbsolutePath(realFilePath);
+    const normalizedRealWorkspaceRoot = normalizeAbsolutePath(realWorkspaceRoot);
+
+    if (!normalizedRealFilePath || !normalizedRealWorkspaceRoot) return null;
+    if (
+      !this.approvedWorkspaceRoots.has(normalizedRealWorkspaceRoot) &&
+      !this.indexedProjectRoots.has(normalizedRealWorkspaceRoot)
+    ) {
+      return null;
+    }
+    if (!isPathWithinRoot(normalizedRealFilePath, normalizedRealWorkspaceRoot)) return null;
+
+    return normalizedRealFilePath;
+  }
+
  private cleanupExpiredTokens() {
    const now = Date.now();
    for (const [token, record] of this.previewTokens) {
@@ -119,6 +119,21 @@ describe('LocalFileProtocolManager', () => {
    expect(response.headers.get('Content-Type')).toBe('text/plain; charset=utf-8');
  });

+  it('does not mint image-only preview URLs for text files', async () => {
+    const manager = new LocalFileProtocolManager();
+    await manager.approveWorkspaceRoot('/Users/alice/project');
+    mockReadFile.mockResolvedValue(Buffer.from('const value = 1;'));
+
+    const url = await manager.createPreviewUrl({
+      accept: 'image',
+      filePath: '/Users/alice/project/App.tsx',
+      workspaceRoot: '/Users/alice/project',
+    });
+
+    expect(url).toBeNull();
+    expect(mockReadFile).toHaveBeenCalledWith('/Users/alice/project/App.tsx');
+  });
+
  it('decodes percent-encoded characters in the path', async () => {
    const manager = new LocalFileProtocolManager();
    manager.registerHandler();
@@ -278,6 +293,52 @@ describe('LocalFileProtocolManager', () => {
    expect(url).toContain('token=');
  });

+  it('reads preview payloads only from approved project roots', async () => {
+    const manager = new LocalFileProtocolManager();
+    await manager.approveIndexedProjectRoot('/Users/alice/project');
+    mockReadFile.mockResolvedValue(Buffer.from('const value = 1;'));
+
+    const result = await manager.readPreviewFile({
+      filePath: '/Users/alice/project/App.tsx',
+      workspaceRoot: '/Users/alice/project',
+    });
+
+    expect(result).toEqual({
+      buffer: Buffer.from('const value = 1;'),
+      contentType: 'text/plain; charset=utf-8',
+      realPath: '/Users/alice/project/App.tsx',
+    });
+    expect(mockReadFile).toHaveBeenCalledWith('/Users/alice/project/App.tsx');
+  });
+
+  it('does not return text payloads for image-only preview reads', async () => {
+    const manager = new LocalFileProtocolManager();
+    await manager.approveIndexedProjectRoot('/Users/alice/project');
+    mockReadFile.mockResolvedValue(Buffer.from('SECRET=value'));
+
+    const result = await manager.readPreviewFile({
+      accept: 'image',
+      filePath: '/Users/alice/project/.env',
+      workspaceRoot: '/Users/alice/project',
+    });
+
+    expect(result).toBeNull();
+    expect(mockReadFile).toHaveBeenCalledWith('/Users/alice/project/.env');
+  });
+
+  it('does not read preview payloads outside the approved workspace root', async () => {
+    const manager = new LocalFileProtocolManager();
+    await manager.approveIndexedProjectRoot('/Users/alice/project');
+
+    const result = await manager.readPreviewFile({
+      filePath: '/Users/alice/.ssh/id_rsa',
+      workspaceRoot: '/Users/alice/project',
+    });
+
+    expect(result).toBeNull();
+    expect(mockReadFile).not.toHaveBeenCalled();
+  });
+
  it('defers registration until app ready when not yet ready', async () => {
    mockApp.isReady.mockReturnValue(false);
    let resolveReady: () => void = () => undefined;
@@ -50,6 +50,13 @@ const envNumber = (defaultValue: number) =>
    }, z.number().optional())
    .default(defaultValue);

+const getRuntimeEnv = () => ({
+  ...process.env,
+  DESKTOP_EXTERNAL_NAVIGATION_HOSTS: process.env.DESKTOP_EXTERNAL_NAVIGATION_HOSTS,
+  UPDATE_CHANNEL: process.env.UPDATE_CHANNEL,
+  UPDATE_SERVER_URL: process.env.UPDATE_SERVER_URL,
+});
+
 /**
 * Desktop (Electron main process) runtime env access.
 *
@@ -63,13 +70,15 @@ export const getDesktopEnv = memoize(() =>
    clientPrefix: 'PUBLIC_',
    emptyStringAsUndefined: true,
    isServer: true,
-    runtimeEnv: process.env,
+    runtimeEnv: getRuntimeEnv(),
    server: {
      DEBUG_VERBOSE: envBoolean(false),

      // escape hatch: allow testing static renderer in dev via env
      DESKTOP_RENDERER_STATIC: envBoolean(false),

+      DESKTOP_EXTERNAL_NAVIGATION_HOSTS: z.string().optional().default(''),
+
      // device gateway url override (dev: point at a local `wrangler dev` instance,
      // e.g. http://localhost:8787). Falls back to the stored value, then the
      // production gateway.
@@ -182,6 +182,42 @@ describe('cliAgentDetectors', () => {
      expect(execFileMock).toHaveBeenCalledTimes(2);
    });

+    it('falls back to the bundled Codex.app binary when `codex` is not on PATH', async () => {
+      // OpenAI's Codex desktop app ships the real CLI inside the .app but does
+      // not symlink it onto PATH. A user with only the desktop app installed
+      // has no `codex` on PATH, so detection must fall back to the bundle.
+      const originalShell = process.env.SHELL;
+      const originalPath = process.env.PATH;
+      // Unset SHELL and pin an already-normalized PATH so `resolveCommandPath`
+      // makes exactly one `which` attempt (no login-shell / normalized-PATH
+      // fallback retry), leaving the bundle as the only viable candidate.
+      delete process.env.SHELL;
+      process.env.PATH = '/usr/bin:/bin';
+
+      try {
+        // 1) `which codex` → not found
+        callExecFileError(new Error('not found'));
+        // 2) absolute bundle path resolves directly (no `which`); validate it
+        callExecFile('codex-cli 0.138.0-alpha.7');
+
+        const { codexDetector } = await import('../cliAgentDetectors');
+        const status = await codexDetector.detect();
+
+        expect(status.available).toBe(true);
+        expect(status.path).toBe('/Applications/Codex.app/Contents/Resources/codex');
+        expect(status.version).toBe('codex-cli 0.138.0-alpha.7');
+
+        // The validation call must target the bundled binary directly via execFile.
+        expect(execMock).not.toHaveBeenCalled();
+        const validateCall = execFileMock.mock.calls.at(-1)!;
+        expect(validateCall[0]).toBe('/Applications/Codex.app/Contents/Resources/codex');
+        expect(validateCall[1]).toEqual(['--version']);
+      } finally {
+        process.env.SHELL = originalShell;
+        process.env.PATH = originalPath;
+      }
+    });
+
    it('falls back to the login shell PATH for tools installed by shell setup', async () => {
      const originalPath = process.env.PATH;
      const originalShell = process.env.SHELL;
@@ -1,5 +1,5 @@
 import { exec, execFile } from 'node:child_process';
-import { platform } from 'node:os';
+import { homedir, platform } from 'node:os';
 import path from 'node:path';
 import { promisify } from 'node:util';

@@ -258,12 +258,34 @@ export const claudeCodeDetector: IToolDetector = createValidatedDetector({
  validateKeywords: ['claude code'],
 });

+/**
+ * OpenAI's Codex desktop app bundles the real `codex` CLI inside its `.app`
+ * but does not symlink it onto PATH. A user who installed only the desktop app
+ * would therefore fail PATH-based detection, so we probe the bundled binary as
+ * a fallback (tried after the PATH lookup, so a user's own install still wins).
+ * Both the system (`/Applications`) and per-user (`~/Applications`) install
+ * locations are covered. macOS only for now.
+ */
+const getCodexCandidates = (): string[] => {
+  const candidates = ['codex'];
+
+  if (platform() === 'darwin') {
+    const bundleRelativePath = 'Codex.app/Contents/Resources/codex';
+    candidates.push(
+      path.join('/Applications', bundleRelativePath),
+      path.join(homedir(), 'Applications', bundleRelativePath),
+    );
+  }
+
+  return candidates;
+};
+
 /**
 * OpenAI Codex CLI
 * @see https://github.com/openai/codex
 */
 export const codexDetector: IToolDetector = createValidatedDetector({
-  candidates: ['codex'],
+  candidates: getCodexCandidates(),
  description: 'Codex - OpenAI agentic coding CLI',
  name: 'codex',
  priority: 2,
@@ -17,24 +17,23 @@ const log = debug('lobe-server:agent-runtime:coordinator');
 * decision) starts, but that resume runs under a **new** operationId with
 * its own event stream. For the paused operationId no further events will
 * arrive, so clients should stop waiting the same way they do on done.
+ *
+ * `waiting_for_async_tool` is different: deferred tools such as server
+ * sub-agents resume the SAME operationId after the out-of-band result is
+ * backfilled. Ending the stream at park time makes the client mark the turn
+ * as stopped while the server is still waiting for sub-agents.
 */
 const STREAM_END_STATUSES = new Set<AgentState['status']>([
  'done',
  'error',
  'interrupted',
  'waiting_for_human',
-  'waiting_for_async_tool',
 ]);

 const hasEnteredStreamEndState = (
  previousStatus?: AgentState['status'],
  nextStatus?: AgentState['status'],
-): nextStatus is
-  | 'done'
-  | 'error'
-  | 'interrupted'
-  | 'waiting_for_human'
-  | 'waiting_for_async_tool' => {
+): nextStatus is 'done' | 'error' | 'interrupted' | 'waiting_for_human' => {
  const wasStreamEnd = previousStatus ? STREAM_END_STATUSES.has(previousStatus) : false;
  return Boolean(nextStatus && STREAM_END_STATUSES.has(nextStatus) && !wasStreamEnd);
 };
@@ -61,6 +61,7 @@ import { chainCompressContext } from '@lobechat/prompts';
 import {
  type ChatToolPayload,
  type ExecSubAgentParams,
+  type ExecVirtualSubAgentParams,
  type MessageToolCall,
  type UIChatMessage,
 } from '@lobechat/types';
@@ -73,6 +74,7 @@ import { TopicModel } from '@/database/models/topic';
 import { UserModel } from '@/database/models/user';
 import { type LobeChatDatabase } from '@/database/type';
 import { fileEnv } from '@/envs/file';
+import { type ExecutionPlan, isDeviceCapablePlan } from '@/helpers/executionTarget';
 import { serverMessagesEngine } from '@/server/modules/Mecha/ContextEngineering';
 import { type EvalContext } from '@/server/modules/Mecha/ContextEngineering/types';
 import { initModelRuntimeFromDB } from '@/server/modules/ModelRuntime';
@@ -202,6 +204,51 @@ const isEmptyModelCompletion = (params: {
  return true;
 };

+type ReasoningReplayNode = {
+  children?: ReasoningReplayNode[];
+  members?: ReasoningReplayNode[];
+  reasoning?: unknown;
+};
+
+const stripAssistantReasoningForReplay = (messages: UIChatMessage[]): UIChatMessage[] => {
+  const stripMessage = <T extends ReasoningReplayNode>(message: T): T => {
+    let changed = false;
+
+    const children = message.children?.map((child) => {
+      const strippedChild = stripMessage(child);
+      if (strippedChild !== child) changed = true;
+      return strippedChild;
+    });
+
+    const members = message.members?.map((member) => {
+      const strippedMember = stripMessage(member);
+      if (strippedMember !== member) changed = true;
+      return strippedMember;
+    });
+
+    if ('reasoning' in message) changed = true;
+    if (!changed) return message;
+
+    const { reasoning: _reasoning, ...messageWithoutReasoning } = message;
+
+    return {
+      ...messageWithoutReasoning,
+      ...(children ? { children } : {}),
+      ...(members ? { members } : {}),
+    } as T;
+  };
+
+  let changed = false;
+
+  const strippedMessages = messages.map((message) => {
+    const strippedMessage = stripMessage(message);
+    if (strippedMessage !== message) changed = true;
+    return strippedMessage;
+  });
+
+  return changed ? strippedMessages : messages;
+};
+
 const GEN_AI_FUNCTION_TOOL_TYPE: ToolType = 'function';

 type ToolFailureKind = 'replan' | 'retry' | 'stop';
@@ -277,7 +324,7 @@ const buildPostProcessUrl = (
 };

 /**
- * Build the per-tool-call server sub-agent runner injected into the tool
+ * Build the per-tool-call server virtual sub-agent runner injected into the tool
 * execution context. Closes over the current tool payload + parent message so
 * the `callSubAgent` server tool can fork a child op without re-deriving the
 * message anchor (which it cannot do correctly from its own context).
@@ -285,17 +332,18 @@ const buildPostProcessUrl = (
 * The runner creates the pending placeholder tool message that anchors the
 * isolation thread (so the UI shows a loading state and the completion bridge
 * has a message to backfill), then kicks off the child op asynchronously and
- * returns immediately. Returns `undefined` when sub-agent execution is not
- * available (no `execSubAgent` callback, or missing agent/topic context).
+ * returns immediately. Returns `undefined` when virtual sub-agent execution is
+ * not available (no `execVirtualSubAgent` callback, or missing agent/topic
+ * context).
 */
-const buildServerSubAgentRunner = (
+const buildServerVirtualSubAgentRunner = (
  ctx: RuntimeExecutorContext,
  state: AgentState,
  chatToolPayload: ChatToolPayload,
  parentMessageId: string,
 ): ServerSubAgentRunner | undefined => {
-  const execSubAgent = ctx.execSubAgent;
-  if (!execSubAgent) return undefined;
+  const execVirtualSubAgent = ctx.execVirtualSubAgent;
+  if (!execVirtualSubAgent) return undefined;

  const agentId = state.metadata?.agentId;
  const topicId = ctx.topicId ?? state.metadata?.topicId;
@@ -318,16 +366,15 @@ const buildServerSubAgentRunner = (
        topicId,
      });

-      // 2. Fork the child op anchored to the placeholder. `resumeParentOnComplete`
-      //    tells execSubAgent to register the completion bridge that
-      //    backfills this tool message and resumes the parent op.
-      const result = (await execSubAgent({
+      // 2. Fork the virtual child op anchored to the placeholder. The virtual
+      //    entry marks the child as `isSubAgent` and registers the completion
+      //    bridge that backfills this tool message and resumes the parent op.
+      const result = (await execVirtualSubAgent({
        agentId: targetAgentId ?? agentId,
        groupId: state.metadata?.groupId ?? undefined,
        instruction,
        parentMessageId: placeholder.id,
        parentOperationId: ctx.operationId,
-        resumeParentOnComplete: true,
        timeout,
        title: description,
        topicId,
@@ -341,7 +388,7 @@ const buildServerSubAgentRunner = (
          await ctx.messageModel.deleteMessage(placeholder.id);
        } catch (error) {
          log(
-            'buildServerSubAgentRunner: failed to clean up placeholder %s: %O',
+            'buildServerVirtualSubAgentRunner: failed to clean up placeholder %s: %O',
            placeholder.id,
            error,
          );
@@ -476,11 +523,17 @@ export interface RuntimeExecutorContext {
  discordContext?: any;
  evalContext?: EvalContext;
  /**
-   * Callback to spawn a sub-agent task server-side.
+   * Callback to run a legacy agent invocation server-side.
   * Injected by AiAgentService so exec_sub_agent / exec_sub_agents executors
-   * can dispatch callAgent-triggered tasks without a circular import.
+   * can dispatch callAgent-triggered runs without a circular import.
   */
  execSubAgent?: (params: ExecSubAgentParams) => Promise<unknown>;
+  /**
+   * Callback to fork a `lobe-agent.callSubAgent` virtual child run. Unlike
+   * execSubAgent, this path installs the async completion bridge and marks the
+   * child operation as a sub-agent.
+   */
+  execVirtualSubAgent?: (params: ExecVirtualSubAgentParams) => Promise<unknown>;
  hookDispatcher?: HookDispatcher;
  loadAgentState?: (operationId: string) => Promise<AgentState | null>;
  messageModel: MessageModel;
@@ -532,17 +585,23 @@ export const createRuntimeExecutors = (
    const provider = llmPayload.provider || state.modelRuntimeConfig?.provider;
    // Resolve tools via ToolResolver (unified tool injection).
    //
-    // Belt-and-suspenders: even if `aiAgent.execAgent` ever forgets to clear
-    // `state.metadata.activeDeviceId` for a non-trusted sender, swallowing
-    // it here keeps `buildStepToolDelta` from re-injecting `local-system` —
-    // the engine's enabledToolIds exclusion alone is not enough, since the
-    // delta builder treats activeDeviceId as an independent activation
-    // signal and only dedupes against already-enabled tools.
+    // Single-track device gate: `buildStepToolDelta` treats activeDeviceId as
+    // an independent activation signal (it only dedupes against already-
+    // enabled tools), so any id that reaches it WILL inject local-system. The
+    // execution plan is the only authority on whether this session may touch
+    // a device — swallow the id for non-device-capable plans (`none`,
+    // `sandbox`) and for denied senders, even if `state.metadata.activeDeviceId`
+    // was populated by a bug or a mid-run side effect. Plans absent on old /
+    // resumed operations fall back to the policy-only gate.
    const devicePolicy = state.metadata?.deviceAccessPolicy as
      | { canUseDevice: boolean; reason: DeviceAccessReason }
      | undefined;
+    const executionPlan = state.metadata?.executionPlan as ExecutionPlan | undefined;
+    const planAllowsDevice = !executionPlan || isDeviceCapablePlan(executionPlan);
    const activeDeviceId =
-      devicePolicy?.canUseDevice === false ? undefined : state.metadata?.activeDeviceId;
+      devicePolicy?.canUseDevice === false || !planAllowsDevice
+        ? undefined
+        : state.metadata?.activeDeviceId;
    const operationToolSet: OperationToolSet = state.operationToolSet ?? {
      enabledToolIds: [],
      executorMap: state.toolExecutorMap ?? {},
@@ -660,7 +719,7 @@ export const createRuntimeExecutors = (

    try {
      type ContentPart = { text: string; type: 'text' } | { image: string; type: 'image' };
-      let shouldPersistAssistantReasoning = false;
+      let shouldReplayAssistantReasoning = false;
      let preserveThinkingForPayload: boolean | undefined;

      // Process messages through serverMessagesEngine to inject system role, knowledge, etc.
@@ -699,19 +758,21 @@ export const createRuntimeExecutors = (
          modelSupportsPreserveThinkingFromCard ||
          (!modelCard && providerSupportsPreserveThinkingFallback);

-        shouldPersistAssistantReasoning =
-          preserveThinkingRequested && modelSupportsPreserveThinking;
+        shouldReplayAssistantReasoning = preserveThinkingRequested && modelSupportsPreserveThinking;
        preserveThinkingForPayload =
          modelSupportsPreserveThinking && typeof preserveThinkingConfigured === 'boolean'
            ? preserveThinkingConfigured
            : undefined;
+        const messagesForContext = shouldReplayAssistantReasoning
+          ? (llmPayload.messages as UIChatMessage[])
+          : stripAssistantReasoningForReplay(llmPayload.messages as UIChatMessage[]);

        // Extract <refer_topic> tags from messages and fetch summaries.
        // Skip if messages already contain injected topic_reference_context
        // (e.g., from client-side contextEngineering preprocessing) to avoid double injection.
        let topicReferences;
        const alreadyHasTopicRefs = (
-          llmPayload.messages as Array<{ content: string | unknown }>
+          messagesForContext as Array<{ content: string | unknown }>
        ).some(
          (m) => typeof m.content === 'string' && m.content.includes('topic_reference_context'),
        );
@@ -720,7 +781,7 @@ export const createRuntimeExecutors = (
          const topicModel = new TopicModel(ctx.serverDB, ctx.userId, ctx.workspaceId);
          const messageModel = new MessageModelClass(ctx.serverDB, ctx.userId, ctx.workspaceId);
          topicReferences = await resolveTopicReferences(
-            llmPayload.messages as Array<{ content: string | unknown }>,
+            messagesForContext as Array<{ content: string | unknown }>,
            async (topicId) => topicModel.findById(topicId),
            async (topicId) => {
              const topic = await topicModel.findById(topicId);
@@ -762,7 +823,7 @@ export const createRuntimeExecutors = (
          agentConfig?.slug === 'web-onboarding' ||
          resolved.enabledToolIds.includes('lobe-web-onboarding');
        const alreadyHasOnboardingContext = (
-          llmPayload.messages as Array<{ content: string | unknown }>
+          messagesForContext as Array<{ content: string | unknown }>
        ).some((message) => {
          if (typeof message.content !== 'string') return false;

@@ -1043,7 +1104,7 @@ export const createRuntimeExecutors = (
                name: kb.name ?? '',
              })),
          },
-          messages: llmPayload.messages as UIChatMessage[],
+          messages: messagesForContext,
          model,
          provider,
          systemRole: agentConfig.systemRole ?? undefined,
@@ -1071,14 +1132,14 @@ export const createRuntimeExecutors = (
          CONTEXT_ENGINEERING_SPAN_NAME,
          {
            attributes: buildContextEngineeringAttributes({
-              hasImages: (llmPayload.messages as Array<{ content?: unknown }>).some(
+              hasImages: (messagesForContext as Array<{ content?: unknown }>).some(
                (m) =>
                  Array.isArray(m.content) &&
                  (m.content as Array<{ type?: string }>).some((p) => p?.type === 'image_url'),
              ),
              historyCompressed:
-                Array.isArray(llmPayload.messages) &&
-                llmPayload.messages.some((m: { role?: string }) => m?.role === 'compressedGroup'),
+                Array.isArray(messagesForContext) &&
+                messagesForContext.some((m: { role?: string }) => m?.role === 'compressedGroup'),
              knowledgeCount:
                (contextEngineInput.knowledge?.knowledgeBases?.length ?? 0) +
                (contextEngineInput.knowledge?.fileContents?.length ?? 0),
@@ -1086,7 +1147,7 @@ export const createRuntimeExecutors = (
                (contextEngineInput.knowledge?.knowledgeBases?.length ?? 0) > 0 ||
                (contextEngineInput.knowledge?.fileContents?.length ?? 0) > 0,
              memoryInjected: Boolean(contextEngineInput.userMemory?.memories),
-              messageCount: llmPayload.messages.length,
+              messageCount: messagesForContext.length,
              operationId,
              stepIndex,
              systemRoleLength: contextEngineInput.systemRole?.length,
@@ -1639,9 +1700,10 @@ export const createRuntimeExecutors = (
                };
              }

-              const persistedReasoning = shouldPersistAssistantReasoning
-                ? finalReasoning
-                : undefined;
+              // preserveThinking only gates whether reasoning is replayed into the
+              // next LLM payload (state.messages); the DB copy powers UI display
+              // after refresh and must always be saved.
+              const replayedReasoning = shouldReplayAssistantReasoning ? finalReasoning : undefined;

              try {
                // Build metadata object
@@ -1675,7 +1737,7 @@ export const createRuntimeExecutors = (
                  content: finalContent,
                  imageList: imageList.length > 0 ? imageList : undefined,
                  metadata: Object.keys(metadata).length > 0 ? metadata : undefined,
-                  reasoning: persistedReasoning,
+                  reasoning: finalReasoning,
                  search: grounding,
                  tools: persistedTools,
                });
@@ -1708,7 +1770,7 @@ export const createRuntimeExecutors = (
              newState.messages.push({
                content,
                id: assistantMessageItem.id,
-                reasoning: persistedReasoning,
+                reasoning: replayedReasoning,
                role: 'assistant',
                tool_calls: stateToolCalls,
              });
@@ -2421,7 +2483,7 @@ export const createRuntimeExecutors = (
                scope: state.metadata?.scope,
                serverDB: ctx.serverDB,
                skipResultTruncation: true,
-                subAgent: buildServerSubAgentRunner(
+                subAgent: buildServerVirtualSubAgentRunner(
                  ctx,
                  state,
                  chatToolPayload,
@@ -2663,14 +2725,15 @@ export const createRuntimeExecutors = (

        log('[%s:%d] Tool execution completed', operationId, stepIndex);

-        // When the tool result carries an execSubAgent / execSubAgents state the
-        // GeneralChatAgent needs `stop: true` in the payload to detect it and
-        // emit the matching exec_sub_agent / exec_sub_agents instruction.  Without
-        // this flag the agent falls through to the normal LLM-call path and the
-        // sub-agent is never spawned.
-        const execTaskStateType = executionResult.state?.type as string | undefined;
-        const isExecTaskState =
-          execTaskStateType === 'execSubAgent' || execTaskStateType === 'execSubAgents';
+        // When a legacy callAgent task result carries execSubAgent / execSubAgents
+        // state, the GeneralChatAgent needs `stop: true` in the payload to detect
+        // it and emit the matching exec_sub_agent / exec_sub_agents instruction.
+        // Without this flag the agent falls through to the normal LLM-call path
+        // and the background agent run is never spawned.
+        const legacyAgentInvocationStateType = executionResult.state?.type as string | undefined;
+        const isLegacyAgentInvocationState =
+          legacyAgentInvocationStateType === 'execSubAgent' ||
+          legacyAgentInvocationStateType === 'execSubAgents';

        executeToolSpan.setAttributes(
          buildExecuteToolResultAttributes({ attempts: execution.attempts, success: isSuccess }),
@@ -2686,7 +2749,7 @@ export const createRuntimeExecutors = (
              isSuccess,
              // Pass tool message ID as parentMessageId for the next LLM call
              parentMessageId: toolMessageId,
-              ...(isExecTaskState && { stop: true }),
+              ...(isLegacyAgentInvocationState && { stop: true }),
              toolCall: chatToolPayload,
              toolCallId: chatToolPayload.id,
            },
@@ -2993,7 +3056,7 @@ export const createRuntimeExecutors = (
                    scope: state.metadata?.scope,
                    serverDB: ctx.serverDB,
                    skipResultTruncation: true,
-                    subAgent: buildServerSubAgentRunner(
+                    subAgent: buildServerVirtualSubAgentRunner(
                      ctx,
                      state,
                      chatToolPayload,
@@ -176,6 +176,19 @@ describe('AgentRuntimeCoordinator', () => {
      });
    });

+    it('should not publish end event when status changes to waiting_for_async_tool because the same stream will resume', async () => {
+      const operationId = 'test-operation-id';
+      const previousState = { status: 'running', stepCount: 3 };
+      const newState = { status: 'waiting_for_async_tool', stepCount: 4 };
+
+      mockStateManager.loadAgentState.mockResolvedValue(previousState);
+
+      await coordinator.saveAgentState(operationId, newState as any);
+
+      expect(mockStateManager.saveAgentState).toHaveBeenCalledWith(operationId, newState);
+      expect(mockStreamManager.publishAgentRuntimeEnd).not.toHaveBeenCalled();
+    });
+
    it('should not publish end event when status was already done', async () => {
      const operationId = 'test-operation-id';
      const previousState = { status: 'done', stepCount: 5 };
@@ -291,6 +304,22 @@ describe('AgentRuntimeCoordinator', () => {
      });
    });

+    it('should not publish end event when status becomes waiting_for_async_tool because deferred tools resume this operation', async () => {
+      const operationId = 'test-operation-id';
+      const stepResult = {
+        executionTime: 1000,
+        newState: { status: 'waiting_for_async_tool', stepCount: 4 },
+        stepIndex: 4,
+      };
+
+      mockStateManager.loadAgentState.mockResolvedValue({ status: 'running', stepCount: 3 });
+
+      await coordinator.saveStepResult(operationId, stepResult as any);
+
+      expect(mockStateManager.saveStepResult).toHaveBeenCalledWith(operationId, stepResult);
+      expect(mockStreamManager.publishAgentRuntimeEnd).not.toHaveBeenCalled();
+    });
+
    it('should publish end event when status becomes interrupted', async () => {
      const operationId = 'test-operation-id';
      const stepResult = {
@@ -408,8 +408,11 @@ describe('RuntimeExecutors', () => {
      );
    });

-    describe('reasoning persistence gate', () => {
-      it('should persist assistant reasoning with tool calls when preserveThinking is enabled on a supported model', async () => {
+    // preserveThinking gates whether reasoning is replayed into the next LLM
+    // payload (state.messages). The DB copy powers UI display after refresh and
+    // is always persisted regardless of the gate.
+    describe('reasoning replay gate', () => {
+      it('should replay assistant reasoning with tool calls when preserveThinking is enabled on a supported model', async () => {
        const toolCallPayload = [
          {
            function: { arguments: '{}', name: 'search' },
@@ -474,7 +477,7 @@ describe('RuntimeExecutors', () => {
        );
      });

-      it('should not persist assistant reasoning when preserveThinking is not enabled', async () => {
+      it('should persist reasoning to DB but not replay it when preserveThinking is not enabled', async () => {
        const mockChat = vi.fn().mockImplementation(async (_payload, options) => {
          await options?.callback?.onThinking?.('hidden reasoning');
          await options?.callback?.onText?.('answer');
@@ -498,9 +501,14 @@ describe('RuntimeExecutors', () => {
        const assistant = result.newState.messages.at(-1) as any;

        expect(assistant.reasoning).toBeUndefined();
+        // DB persistence must NOT be gated — UI shows reasoning after refresh
+        expect(mockMessageModel.update).toHaveBeenCalledWith(
+          'msg-123',
+          expect.objectContaining({ reasoning: { content: 'hidden reasoning' } }),
+        );
      });

-      it('should persist assistant reasoning when preserveThinking is enabled on a supported model', async () => {
+      it('should replay assistant reasoning when preserveThinking is enabled on a supported model', async () => {
        const mockChat = vi.fn().mockImplementation(async (_payload, options) => {
          await options?.callback?.onThinking?.('preserved reasoning');
          await options?.callback?.onText?.('answer');
@@ -546,7 +554,7 @@ describe('RuntimeExecutors', () => {
        );
      });

-      it('should persist reasoning for unknown custom deployments on supported providers', async () => {
+      it('should replay reasoning for unknown custom deployments on supported providers', async () => {
        const mockChat = vi.fn().mockImplementation(async (_payload, options) => {
          await options?.callback?.onThinking?.('custom deployment reasoning');
          await options?.callback?.onText?.('answer');
@@ -592,9 +600,9 @@ describe('RuntimeExecutors', () => {
        );
      });

-      it('should not persist reasoning when model does not declare preserveThinking capability', async () => {
+      it('should persist reasoning to DB but not replay it when model does not declare preserveThinking capability', async () => {
        const mockChat = vi.fn().mockImplementation(async (_payload, options) => {
-          await options?.callback?.onThinking?.('reasoning that should not be saved');
+          await options?.callback?.onThinking?.('reasoning on an unsupported model');
          await options?.callback?.onText?.('answer');
          return new Response('done');
        });
@@ -634,6 +642,13 @@ describe('RuntimeExecutors', () => {
          expect.not.objectContaining({ preserveThinking: expect.any(Boolean) }),
          expect.anything(),
        );
+        // DB persistence must NOT be gated — UI shows reasoning after refresh
+        expect(mockMessageModel.update).toHaveBeenCalledWith(
+          'msg-123',
+          expect.objectContaining({
+            reasoning: { content: 'reasoning on an unsupported model' },
+          }),
+        );
      });
    });

@@ -771,9 +786,9 @@ describe('RuntimeExecutors', () => {
        });
        vi.mocked(initModelRuntimeFromDB).mockResolvedValueOnce({ chat: mockChat } as any);

-        // Reasoning only lands in the finalized message when preserveThinking is
-        // enabled on a supported model; otherwise it is intentionally dropped.
-        // Enable it here so this still guards reasoning_part capture (not drop).
+        // Reasoning is only replayed into state.messages when preserveThinking is
+        // enabled on a supported model. Enable it here so this asserts
+        // reasoning_part capture via the state replay path.
        const ctxWithThinking: RuntimeExecutorContext = {
          ...ctx,
          agentConfig: { chatConfig: { preserveThinking: true }, plugins: [], systemRole: 'test' },
@@ -1596,6 +1611,168 @@ describe('RuntimeExecutors', () => {
        );
      });

+      it('should strip stored assistant reasoning before context processing when replay gate is off', async () => {
+        const ctxWithConfig: RuntimeExecutorContext = {
+          ...ctx,
+          agentConfig: {
+            plugins: [],
+            systemRole: 'test',
+          },
+        };
+        const executors = createRuntimeExecutors(ctxWithConfig);
+        const state = createMockState();
+        const messages = [
+          {
+            content: 'Previous answer',
+            reasoning: { content: 'stored reasoning should stay display-only' },
+            role: 'assistant',
+          },
+          { content: 'Continue', role: 'user' },
+        ];
+
+        await executors.call_llm!(
+          {
+            payload: {
+              messages,
+              model: 'gpt-4',
+              provider: 'openai',
+            },
+            type: 'call_llm' as const,
+          },
+          state,
+        );
+
+        const engineInput = engineSpy.mock.calls[0][0];
+        expect(engineInput.messages[0]).toEqual({
+          content: 'Previous answer',
+          role: 'assistant',
+        });
+        expect(messages[0]).toEqual(
+          expect.objectContaining({
+            reasoning: { content: 'stored reasoning should stay display-only' },
+          }),
+        );
+      });
+
+      it('should strip stored reasoning from grouped assistant messages before context processing when replay gate is off', async () => {
+        const ctxWithConfig: RuntimeExecutorContext = {
+          ...ctx,
+          agentConfig: {
+            plugins: [],
+            systemRole: 'test',
+          },
+        };
+        const executors = createRuntimeExecutors(ctxWithConfig);
+        const state = createMockState();
+        const groupedChild = {
+          content: 'Grouped answer',
+          id: 'group-child-1',
+          reasoning: { content: 'grouped child reasoning should stay display-only' },
+          role: 'assistant',
+        };
+        const councilMember = {
+          content: 'Council member answer',
+          id: 'member-1',
+          reasoning: { content: 'member reasoning should stay display-only' },
+          role: 'assistant',
+        };
+        const nestedCouncilChild = {
+          content: 'Nested council answer',
+          id: 'member-child-1',
+          reasoning: { content: 'nested member reasoning should stay display-only' },
+          role: 'assistant',
+        };
+        const messages = [
+          {
+            children: [groupedChild],
+            content: '',
+            id: 'group-1',
+            role: 'assistantGroup',
+          },
+          {
+            content: '',
+            id: 'council-1',
+            members: [
+              councilMember,
+              {
+                children: [nestedCouncilChild],
+                content: '',
+                id: 'member-group-1',
+                role: 'assistantGroup',
+              },
+            ],
+            role: 'agentCouncil',
+          },
+          { content: 'Continue', role: 'user' },
+        ];
+
+        await executors.call_llm!(
+          {
+            payload: {
+              messages,
+              model: 'gpt-4',
+              provider: 'openai',
+            },
+            type: 'call_llm' as const,
+          },
+          state,
+        );
+
+        const engineInput = engineSpy.mock.calls[0][0];
+        expect(engineInput.messages[0].children[0]).not.toHaveProperty('reasoning');
+        expect(engineInput.messages[1].members[0]).not.toHaveProperty('reasoning');
+        expect(engineInput.messages[1].members[1].children[0]).not.toHaveProperty('reasoning');
+        expect(groupedChild).toHaveProperty('reasoning');
+        expect(councilMember).toHaveProperty('reasoning');
+        expect(nestedCouncilChild).toHaveProperty('reasoning');
+      });
+
+      it('should keep stored assistant reasoning before context processing when replay gate is enabled', async () => {
+        const ctxWithConfig: RuntimeExecutorContext = {
+          ...ctx,
+          agentConfig: {
+            chatConfig: { preserveThinking: true },
+            plugins: [],
+            systemRole: 'test',
+          },
+        };
+        const executors = createRuntimeExecutors(ctxWithConfig);
+        const state = createMockState({
+          modelRuntimeConfig: {
+            model: 'qwen3.6-plus',
+            provider: 'qwen',
+          },
+        });
+
+        await executors.call_llm!(
+          {
+            payload: {
+              messages: [
+                {
+                  content: 'Previous answer',
+                  reasoning: { content: 'reasoning to replay' },
+                  role: 'assistant',
+                },
+                { content: 'Continue', role: 'user' },
+              ],
+              model: 'qwen3.6-plus',
+              provider: 'qwen',
+            },
+            type: 'call_llm' as const,
+          },
+          state,
+        );
+
+        const engineInput = engineSpy.mock.calls[0][0];
+        expect(engineInput.messages[0]).toEqual(
+          expect.objectContaining({
+            content: 'Previous answer',
+            reasoning: { content: 'reasoning to replay' },
+            role: 'assistant',
+          }),
+        );
+      });
+
      it('should not call serverMessagesEngine when agentConfig is not set', async () => {
        const executors = createRuntimeExecutors(ctx); // ctx without agentConfig
        const state = createMockState();
@@ -508,12 +508,34 @@ describe('createServerAgentToolsEngine', () => {
      expect(result.enabledToolIds).not.toContain(LocalSystemManifest.identifier);
    });

-    it('should disable LocalSystem when runtimeMode is explicitly set to cloud', () => {
+    it('should disable LocalSystem when executionTarget is sandbox', () => {
      const context = createMockContext();
      const engine = createServerAgentToolsEngine(context, {
        agentConfig: {
+          agencyConfig: { executionTarget: 'sandbox' },
+          plugins: [LocalSystemManifest.identifier],
+        },
+        canUseDevice: true,
+        deviceContext: { gatewayConfigured: true, deviceOnline: true, autoActivated: true },
+        model: 'gpt-4',
+        provider: 'openai',
+      });
+
+      const result = engine.generateToolsDetailed({
+        toolIds: [LocalSystemManifest.identifier],
+        model: 'gpt-4',
+        provider: 'openai',
+      });
+
+      expect(result.enabledToolIds).not.toContain(LocalSystemManifest.identifier);
+    });
+
+    it('should disable LocalSystem when executionTarget is none', () => {
+      const context = createMockContext();
+      const engine = createServerAgentToolsEngine(context, {
+        agentConfig: {
+          agencyConfig: { executionTarget: 'none' },
          plugins: [LocalSystemManifest.identifier],
-          chatConfig: { runtimeEnv: { runtimeMode: { desktop: 'cloud' } } },
        },
        canUseDevice: true,
        deviceContext: { gatewayConfigured: true, deviceOnline: true, autoActivated: true },
@@ -574,6 +596,50 @@ describe('createServerAgentToolsEngine', () => {
      expect(result.enabledToolIds).not.toContain(RemoteDeviceManifest.identifier);
    });

+    it('should disable RemoteDevice when executionTarget is none — 无设备 means NO device', () => {
+      const context = createMockContext();
+      const engine = createServerAgentToolsEngine(context, {
+        agentConfig: {
+          agencyConfig: { executionTarget: 'none' },
+          plugins: [RemoteDeviceManifest.identifier],
+        },
+        canUseDevice: true,
+        deviceContext: { gatewayConfigured: true },
+        model: 'gpt-4',
+        provider: 'openai',
+      });
+
+      const result = engine.generateToolsDetailed({
+        toolIds: [RemoteDeviceManifest.identifier],
+        model: 'gpt-4',
+        provider: 'openai',
+      });
+
+      expect(result.enabledToolIds).not.toContain(RemoteDeviceManifest.identifier);
+    });
+
+    it('should disable RemoteDevice when executionTarget is sandbox — sandbox and devices are mutually exclusive', () => {
+      const context = createMockContext();
+      const engine = createServerAgentToolsEngine(context, {
+        agentConfig: {
+          agencyConfig: { executionTarget: 'sandbox' },
+          plugins: [RemoteDeviceManifest.identifier],
+        },
+        canUseDevice: true,
+        deviceContext: { gatewayConfigured: true },
+        model: 'gpt-4',
+        provider: 'openai',
+      });
+
+      const result = engine.generateToolsDetailed({
+        toolIds: [RemoteDeviceManifest.identifier],
+        model: 'gpt-4',
+        provider: 'openai',
+      });
+
+      expect(result.enabledToolIds).not.toContain(RemoteDeviceManifest.identifier);
+    });
+
    it('should disable RemoteDevice when device is already auto-activated', () => {
      const context = createMockContext();
      const engine = createServerAgentToolsEngine(context, {
@@ -28,7 +28,7 @@ import { ToolsEngine } from '@lobechat/context-engine';
 import { type RuntimeEnvMode, type RuntimePlatform } from '@lobechat/types';
 import debug from 'debug';

-import { resolveRuntimeMode } from '@/helpers/executionTarget';
+import { executionTargetToRuntimeMode, resolveExecutionTarget } from '@/helpers/executionTarget';
 import {
  buildAllowedBuiltinTools,
  DEVICE_TOOL_IDENTIFIERS,
@@ -129,6 +129,7 @@ export const createServerAgentToolsEngine = (
    canUseDevice = false,
    deviceContext,
    disableLocalSystem = false,
+    executionPlan,
    globalMemoryEnabled = false,
    hasAgentDocuments = false,
    hasEnabledKnowledgeBases = false,
@@ -144,20 +145,26 @@ export const createServerAgentToolsEngine = (
  // back to the caller) is removed.
  const hasDeviceProxy = !!deviceContext?.gatewayConfigured;

-  // Platform key is used only to look up the user's per-platform
-  // `runtimeMode` preference. A server configured with a device-gateway is
-  // serving desktop-class users; otherwise the caller is treated as web.
+  // A server configured with a device-gateway is serving desktop-class users
+  // (the unset-target default resolves to `local`); otherwise the caller is
+  // treated as web.
  const platform: RuntimePlatform = hasDeviceProxy ? 'desktop' : 'web';

-  // Tool gate derived from the single `agencyConfig.executionTarget` param
-  // (sandbox → cloud tools, local → local-system tools, device → gateway), with
-  // a no-regression fallback to the legacy per-platform `runtimeMode` for agents
-  // that predate `executionTarget`.
-  const runtimeMode: RuntimeEnvMode = resolveRuntimeMode(
-    agentConfig.agencyConfig,
-    agentConfig.chatConfig?.runtimeEnv?.runtimeMode?.[platform],
-    platform === 'desktop',
-  );
+  // Tool gate derived from the run's resolved execution plan (sandbox → cloud
+  // tools, local → local-system tools, device → gateway). Callers that don't
+  // resolve a plan (focused sub-agent engines) fall back to deriving the
+  // effective target from agencyConfig.
+  const executionTarget =
+    executionPlan?.target ??
+    resolveExecutionTarget(agentConfig.agencyConfig, {
+      isDesktop: platform === 'desktop',
+    });
+  const runtimeMode: RuntimeEnvMode = executionTargetToRuntimeMode(executionTarget);
+  // Device tools (local-system, remote-device proxy) only exist for
+  // device-capable targets. `none` means NO device — the proxy that could
+  // activate one mid-run must not be offered either; `sandbox` and devices
+  // are mutually exclusive.
+  const deviceCapable = executionTarget === 'local' || executionTarget === 'device';

  const searchMode = agentConfig.chatConfig?.searchMode ?? 'auto';
  const isSearchEnabled = searchMode !== 'off';
@@ -207,13 +214,14 @@ export const createServerAgentToolsEngine = (
    // System-level rules (may override user selection for specific tools)
    [CloudSandboxManifest.identifier]: runtimeMode === 'cloud',
    [KnowledgeBaseManifest.identifier]: hasEnabledKnowledgeBases,
-    // Local-system: gated by `canUseDevice` (resolveDeviceAccessPolicy)
-    // first — keeps external bot senders out before runtime checks even
-    // run. Then user must have opted into local runtime on this platform
-    // (`runtimeMode === 'local'`) AND have an online, auto-activated
-    // device registered with the device-gateway.
+    // Local-system: the user must have opted into local runtime
+    // (`runtimeMode === 'local'`) AND have an online, auto-activated device
+    // registered with the device-gateway. Access policy (external bot
+    // senders) is enforced upstream: `resolveExecutionPlan` degrades denied
+    // targets to `none`, and `buildAllowedBuiltinTools` +
+    // `excludeIdentifiers` physically drop the manifest for
+    // `canUseDevice=false` turns.
    [LocalSystemManifest.identifier]:
-      canUseDevice &&
      !disableLocalSystem &&
      runtimeMode === 'local' &&
      hasDeviceProxy &&
@@ -222,16 +230,13 @@ export const createServerAgentToolsEngine = (
    [MemoryManifest.identifier]: globalMemoryEnabled,
    // Only auto-enable in bot conversations; otherwise let user's plugin selection take effect
    ...(isBotConversation && { [MessageManifest.identifier]: true }),
-    // Remote-device proxy: shown only when the server has a proxy but
-    // no specific device is auto-activated yet (user must pick).
-    //
-    // `canUseDevice` is the first short-circuit: external bot senders
-    // (and unconfigured bot owners) never reach the proxy, both because
-    // it would let them poke at the owner's machine AND because its
-    // systemRole would otherwise leak the device list into the LLM
-    // context — see the gated injection in `aiAgent.execAgent`.
+    // Remote-device proxy: shown only for device-capable targets when the
+    // server has a proxy but no specific device is auto-activated yet (user
+    // must pick). External bot senders never reach it: the plan degrades
+    // denied targets to `none` (→ not deviceCapable) and the physical
+    // manifest walls drop it for `canUseDevice=false` turns.
    [RemoteDeviceManifest.identifier]:
-      canUseDevice && hasDeviceProxy && !deviceContext?.autoActivated,
+      deviceCapable && hasDeviceProxy && !deviceContext?.autoActivated,
    [AgentDocumentsManifest.identifier]: hasAgentDocuments,
    [WebBrowsingManifest.identifier]: isSearchEnabled,
  };
@@ -1,10 +1,7 @@
 import { type LobeToolManifest, type PluginEnableChecker } from '@lobechat/context-engine';
-import {
-  type LobeAgentAgencyConfig,
-  type LobeBuiltinTool,
-  type LobeTool,
-  type RuntimeEnvConfig,
-} from '@lobechat/types';
+import { type LobeAgentAgencyConfig, type LobeBuiltinTool, type LobeTool } from '@lobechat/types';
+
+import type { ExecutionPlan } from '@/helpers/executionTarget';

 /**
 * Installed plugin with manifest
@@ -67,7 +64,6 @@ export interface ServerCreateAgentToolsEngineParams {
       * `plugins` and `alwaysOnToolIds`. Undefined / true → agent mode.
       */
      enableAgentMode?: boolean;
-      runtimeEnv?: RuntimeEnvConfig;
      searchMode?: 'off' | 'on' | 'auto';
      /**
       * Overrides the `enableAgentMode` derivation. `custom` = the toolset is
@@ -97,6 +93,12 @@ export interface ServerCreateAgentToolsEngineParams {
  };
  /** Whether to suppress the local-system builtin while preserving other tools. */
  disableLocalSystem?: boolean;
+  /**
+   * The run's resolved execution plan (see `resolveExecutionPlan`). When
+   * provided, its effective `target` drives the runtime tool gate; when
+   * omitted the engine derives the target from `agencyConfig` directly.
+   */
+  executionPlan?: ExecutionPlan;
  /** Whether the user's global memory setting is enabled */
  globalMemoryEnabled?: boolean;
  /** Whether agent has agent documents */
@@ -408,7 +408,7 @@ export const initModelRuntimeFromDB = async (
 ): Promise<ModelRuntime> => {
  // 1. Get user's provider configuration from database
  // NOTE: workspace-scoped ai_infra is deferred until the ai_infra surrogate-`_id`
-  // PK migration (LOBE-10056) lands; AiProviderModel stays personal-scoped for now.
+  // PK migration lands; AiProviderModel stays personal-scoped for now.
  const aiProviderModel = new AiProviderModel(db, userId);

  // Use getAiProviderById with KeyVaultsGateKeeper.getUserKeyVaults as decryptor
@@ -119,7 +119,7 @@ describe('aiChatRouter', () => {
    expect(mockCreateUserAndAssistantMessages).toHaveBeenCalledTimes(1);
    expect(mockCreateUserAndAssistantMessages).toHaveBeenCalledWith(
      expect.any(Object),
-      expect.objectContaining({ touchTopicUpdatedAt: false }),
+      expect.not.objectContaining({ touchTopicUpdatedAt: expect.anything() }),
    );

    expect(mockGet).toHaveBeenCalledWith(
@@ -161,7 +161,7 @@ describe('aiChatRouter', () => {
    expect(mockCreateMessage).toHaveBeenCalled();
    expect(mockCreateUserAndAssistantMessages).toHaveBeenCalledWith(
      expect.any(Object),
-      expect.objectContaining({ touchTopicUpdatedAt: true }),
+      expect.not.objectContaining({ touchTopicUpdatedAt: expect.anything() }),
    );
    expect(mockGet).toHaveBeenCalledWith(
      expect.objectContaining({
@@ -1032,6 +1032,93 @@ describe('aiChatRouter', () => {
      }
    });

+    it('marks input completion runtime 4xx errors to skip tRPC handler logging', async () => {
+      const { initModelRuntimeFromDB } = await import('@/server/modules/ModelRuntime');
+      const runtimeError = {
+        error: { message: 'rate limited' },
+        errorType: AgentRuntimeErrorType.RateLimitExceeded,
+      };
+
+      vi.mocked(initModelRuntimeFromDB).mockRejectedValueOnce(runtimeError);
+
+      const caller = aiChatRouter.createCaller({ ...mockCtx, serverDB: {} } as any);
+
+      try {
+        await caller.outputJSON({
+          messages: [{ content: 'test', role: 'user' }],
+          model: 'gpt-4o',
+          provider: 'openai',
+          tracing: { scenario: 'input_completion' },
+        });
+        throw new Error('Expected outputJSON to throw');
+      } catch (error) {
+        expect(error).toBeInstanceOf(TRPCError);
+        expect((runtimeError as any).__lobeSilentTRPCErrorLog).toBe(true);
+      }
+    });
+
+    it('does not mark non-input-completion runtime errors as silent', async () => {
+      const { initModelRuntimeFromDB } = await import('@/server/modules/ModelRuntime');
+      const runtimeError = {
+        error: { message: 'rate limited' },
+        errorType: AgentRuntimeErrorType.RateLimitExceeded,
+      };
+
+      vi.mocked(initModelRuntimeFromDB).mockRejectedValueOnce(runtimeError);
+
+      const caller = aiChatRouter.createCaller({ ...mockCtx, serverDB: {} } as any);
+
+      try {
+        await caller.outputJSON({
+          messages: [{ content: 'test', role: 'user' }],
+          model: 'gpt-4o',
+          provider: 'openai',
+          tracing: { scenario: 'topic_title' },
+        });
+        throw new Error('Expected outputJSON to throw');
+      } catch (error) {
+        expect(error).toBeInstanceOf(TRPCError);
+        expect((runtimeError as any).__lobeSilentTRPCErrorLog).toBeUndefined();
+      }
+    });
+
+    it('maps raw provider 4xx errors to BAD_REQUEST instead of internal errors', async () => {
+      const { initModelRuntimeFromDB } = await import('@/server/modules/ModelRuntime');
+
+      // Raw SDK APIError shape: carries an HTTP status but no errorType — the
+      // generateObject path rethrows upstream errors verbatim (e.g. a BYOK
+      // gateway rejecting response_format json_schema).
+      const providerError = Object.assign(
+        new Error(
+          '400 Error from provider (DeepSeek): This response_format type is unavailable now',
+        ),
+        { status: 400 },
+      );
+      const mockGenerateObject = vi.fn().mockRejectedValue(providerError);
+
+      vi.mocked(initModelRuntimeFromDB).mockResolvedValue({
+        generateObject: mockGenerateObject,
+      } as any);
+
+      const caller = aiChatRouter.createCaller({ ...mockCtx, serverDB: {} } as any);
+
+      try {
+        await caller.outputJSON({
+          messages: [{ content: 'test', role: 'user' }],
+          model: 'deepseek-v4-flash-free',
+          provider: 'opencodezen',
+        });
+        throw new Error('Expected outputJSON to throw');
+      } catch (error) {
+        expect(error).toBeInstanceOf(TRPCError);
+        expect(error).toMatchObject({
+          cause: providerError,
+          code: 'BAD_REQUEST',
+          message: providerError.message,
+        });
+      }
+    });
+
    it('should handle tools parameter when provided', async () => {
      const { initModelRuntimeFromDB } = await import('@/server/modules/ModelRuntime');

@@ -4,12 +4,15 @@ import { pushTokenRouter } from '@/server/routers/lambda/pushToken';

 const mockUpsert = vi.fn();
 const mockUnregister = vi.fn();
+const mockDeleteByExpoTokenAndDevice = vi.fn();

 vi.mock('@/database/models/pushToken', () => ({
  PushTokenModel: vi.fn(() => ({
    unregister: mockUnregister,
    upsert: mockUpsert,
  })),
+  deletePushTokenByExpoTokenAndDevice: (...args: unknown[]) =>
+    mockDeleteByExpoTokenAndDevice(...args),
 }));

 const createCaller = (ctxOverrides: Partial<any> = {}) => {
@@ -91,18 +94,90 @@ describe('pushTokenRouter', () => {
  });

  describe('unregister', () => {
-    it('should call model.unregister with deviceId', async () => {
+    it('should delete by (expoToken, deviceId) when expoToken is provided', async () => {
+      mockDeleteByExpoTokenAndDevice.mockResolvedValueOnce(undefined);
+      const caller = createCaller();
+
+      const result = await caller.unregister({
+        deviceId: 'device-1',
+        expoToken: 'ExponentPushToken[abc]',
+      });
+
+      expect(mockDeleteByExpoTokenAndDevice).toHaveBeenCalledWith(expect.anything(), {
+        deviceId: 'device-1',
+        expoToken: 'ExponentPushToken[abc]',
+      });
+      expect(result).toEqual({ success: true });
+      // Legacy (userId, deviceId) path must not fire when expoToken is present
+      expect(mockUnregister).not.toHaveBeenCalled();
+    });
+
+    it('should fall back to (userId, deviceId) for legacy clients with a session', async () => {
+      // Path B — v1.0.7 only sends deviceId; if the request still carries a
+      // valid session we MUST delete the row, otherwise PushChannel keeps
+      // notifying a signed-out device (Expo DeviceNotRegistered only fires on
+      // uninstall, not logout).
      mockUnregister.mockResolvedValueOnce(undefined);
      const caller = createCaller();

-      await caller.unregister({ deviceId: 'device-1' });
+      const result = await caller.unregister({ deviceId: 'device-1' });

      expect(mockUnregister).toHaveBeenCalledWith('device-1');
+      expect(mockDeleteByExpoTokenAndDevice).not.toHaveBeenCalled();
+      expect(result).toEqual({ success: true });
+    });
+
+    it('should silently succeed without expoToken AND without session', async () => {
+      // Path C — v1.0.7 + dead session: the only safe move is silent OK.
+      // Orphan row will be cleaned up by the process-push-receipts worker via
+      // Expo DeviceNotRegistered receipts. Returning 200 here stops the storm.
+      const caller = createCaller({ userId: undefined });
+
+      const result = await caller.unregister({ deviceId: 'device-1' });
+
+      expect(mockDeleteByExpoTokenAndDevice).not.toHaveBeenCalled();
+      expect(mockUnregister).not.toHaveBeenCalled();
+      expect(result).toEqual({ success: true });
+    });
+
+    it('should succeed for an unauthenticated caller carrying expoToken', async () => {
+      // New clients (>=1.0.8) hit Path A regardless of session.
+      const caller = createCaller({ userId: undefined });
+
+      const result = await caller.unregister({
+        deviceId: 'device-1',
+        expoToken: 'ExponentPushToken[abc]',
+      });
+
+      expect(result).toEqual({ success: true });
+      expect(mockDeleteByExpoTokenAndDevice).toHaveBeenCalled();
+      expect(mockUnregister).not.toHaveBeenCalled();
+    });
+
+    it('should prefer expoToken precision over the legacy userId fallback', async () => {
+      // If both are available, always take Path A — the (expoToken, deviceId)
+      // pair is more precise and doesn't risk deleting a wrong row.
+      const caller = createCaller();
+
+      await caller.unregister({
+        deviceId: 'device-1',
+        expoToken: 'ExponentPushToken[abc]',
+      });
+
+      expect(mockDeleteByExpoTokenAndDevice).toHaveBeenCalled();
+      expect(mockUnregister).not.toHaveBeenCalled();
    });

    it('should reject empty deviceId', async () => {
      const caller = createCaller();
      await expect(caller.unregister({ deviceId: '' })).rejects.toThrow();
    });
+
+    it('should reject empty expoToken when provided', async () => {
+      const caller = createCaller();
+      await expect(
+        caller.unregister({ deviceId: 'device-1', expoToken: '' }),
+      ).rejects.toThrow();
+    });
  });
 });
@@ -36,6 +36,7 @@ export const compareDocumentHistoryItemsInputSchema = z.object({
 });

 export const updateDocumentInputSchema = z.object({
+  breakAutosaveWindow: z.boolean().optional(),
  content: z.string().optional(),
  editorData: z.string().optional(),
  fileType: z.string().optional(),
@@ -58,6 +59,7 @@ export interface DocumentHistoryListItem {
  isCurrent: boolean;
  savedAt: string;
  saveSource: DocumentHistorySaveSource;
+  userId: string;
 }

 export interface ListHistoryOutput {
@@ -123,6 +125,7 @@ export interface CompareHistoryItemsInput {
 }

 export interface UpdateDocumentInput {
+  breakAutosaveWindow?: boolean;
  content?: string;
  editorData?: string;
  fileType?: string;
@@ -85,6 +85,7 @@ export const agentSignalRouter = router({
      return enqueueAgentSignalSourceEvent(sourceEvent, {
        agentId: input.agentId,
        userId: ctx.userId,
+        workspaceId: ctx.workspaceId ?? undefined,
      });
    }),
  listReceipts: agentSignalProcedure
--- a/Show More
+++ b/Show More