🐛 fix(gateway): guard against stale operation after token refresh

Re-check the topic's running operationId after the async token-refresh await. A newer executeGatewayAgent call may have taken over for the same topic during that wait, which would cause two concurrent streams. Bail early if the operationId no longer matches. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
🐛 fix(agent-builder): explicitly sync editing agent ID to chatStore
2026-06-18 21:36:12 +00:00 · 2026-06-01 10:38:30 +08:00 · 2026-05-31 20:12:52 +08:00 · 2026-05-31 17:22:34 +08:00 · 2026-05-31 17:02:32 +08:00 · 2026-05-31 15:11:55 +08:00
2197 changed files with 125981 additions and 27512 deletions
@@ -14,7 +14,7 @@ In `NODE_ENV=development`, `AgentRuntimeService.executeStep()` automatically rec

 **Data flow**: executeStep loop -> build `StepPresentationData` -> write partial snapshot to disk -> on completion, finalize to `.agent-tracing/{timestamp}_{traceId}.json`

-**Context engine capture**: In `RuntimeExecutors.ts`, the `call_llm` executor emits a `context_engine_result` event after `serverMessagesEngine()` processes messages. This event carries the full `contextEngineInput` (DB messages, systemRole, model, knowledge, tools, userMemory, etc.) and the processed `output` messages (the final LLM payload).
+**Context engine capture**: In `RuntimeExecutors.ts`, the `call_llm` executor calls `ctx.tracingContextEngine(input, output)` after `serverMessagesEngine()` processes messages. `AgentRuntimeService.executeStep` buffers the call per step and forwards it to `OperationTraceRecorder.appendStep` as the typed `contextEngine` field. CE flows through this side channel rather than the `events` array so its heavy payload (agentDocuments, systemRole, …) never enters the Redis state pipeline (LOBE-9110).

 ## Package Location

@@ -199,9 +199,10 @@ interface StepSnapshot {
  messages?: any[]; // DB messages before step
  context?: { phase: string; payload?: unknown; stepContext?: unknown };
  events?: Array<{ type: string; [key: string]: unknown }>;
-  // context_engine_result event contains:
-  //   input: full contextEngineInput (messages, systemRole, model, knowledge, tools, userMemory, ...)
-  //   output: processed messages array (final LLM payload)
+  contextEngine?: {
+    input?: unknown; // contextEngineInput minus messages + toolsConfig (reconstructible from baseline)
+    output?: unknown; // processed messages array (final LLM payload)
+  };
 }
 ```

@@ -216,5 +217,5 @@ When using `--messages`, the output shows three sections (if context engine data
 ## Integration Points

 - **Recording**: `src/server/services/agentRuntime/AgentRuntimeService.ts` — in the `executeStep()` method, after building `stepPresentationData`, writes partial snapshot in dev mode
- **Context engine event**: `src/server/modules/AgentRuntime/RuntimeExecutors.ts` — in `call_llm` executor, after `serverMessagesEngine()` returns, emits `context_engine_result` event
+- **Context engine capture**: `src/server/modules/AgentRuntime/RuntimeExecutors.ts` — in `call_llm` executor, after `serverMessagesEngine()` returns, calls `ctx.tracingContextEngine(input, output)`. `AgentRuntimeService.executeStep` buffers it per step and passes it to `traceRecorder.appendStep` as the typed `contextEngine` field (kept off the `events` array to stay out of Redis state).
 - **Store**: `FileSnapshotStore` reads/writes to `.agent-tracing/` relative to `process.cwd()`
@@ -23,7 +23,7 @@ A builtin tool is a package the agent runtime can call. It ships **five faces**:
 | ------------------------------------------------------------------------------------ | --------------------------------------------- |
 | Where do files live? What does each face do? Wiring?                                 | [architecture.md](references/architecture.md) |
 | How do I name the tool, design APIs, write the manifest, executor, ExecutionRuntime? | [tool-design.md](references/tool-design.md)   |
-| How do I build Inspector / Render / Placeholder / Streaming / Intervention / Portal? | [ui.md](references/ui.md)                     |
+| How do I build Inspector / Render / Placeholder / Streaming / Intervention / Portal? | [ui/](references/ui/README.md)                |

 ---

@@ -2,7 +2,7 @@

 This doc covers everything that **isn't UI**: the tool's identifier, API surface, manifest, types, system prompt, ExecutionRuntime, and the executor that wires it into the frontend.

-For UI surfaces (Inspector / Render / Placeholder / Streaming / Intervention / Portal), see [ui.md](ui.md).
+For UI surfaces (Inspector / Render / Placeholder / Streaming / Intervention / Portal), see [ui/](ui/README.md).
 For where files live and how registries work, see [architecture.md](architecture.md).

 ---
@@ -156,7 +156,7 @@ export const TaskManifest: BuiltinToolManifest = {
  executors: ['client', 'server'],

  /* Default human intervention policy for all APIs that don't specify one.
-     Pair with an Intervention component (see ui.md). */
+     Pair with an Intervention component (see ui/intervention.md). */
  humanIntervention: 'never' | 'always' | { /* extended config */ },
 }
 ```
@@ -1,742 +0,0 @@
-# Tool UI Surfaces
-
-A builtin tool can ship up to **six client-side surfaces**, each with a different role in the chat UI. Only `Inspector` is required; the other five are added on demand and registered in their own central files.
-
-| Surface      | Required? | When the chat shows it                                                | Registered in                                 |
-| ------------ | --------- | --------------------------------------------------------------------- | --------------------------------------------- |
-| Inspector    | ✅ Always | Header strip of every tool call (one-line chip)                       | `packages/builtin-tools/src/inspectors.ts`    |
-| Render       | Optional  | Rich result card below the header, after the call returns             | `packages/builtin-tools/src/renders.ts`       |
-| Placeholder  | Optional  | Skeleton between "args streaming complete" and "result arrives"       | `packages/builtin-tools/src/placeholders.ts`  |
-| Streaming    | Optional  | Live output during execution (e.g. command stdout)                    | `packages/builtin-tools/src/streamings.ts`    |
-| Intervention | Optional  | Approval / edit-before-run dialog (when `humanIntervention` triggers) | `packages/builtin-tools/src/interventions.ts` |
-| Portal       | Optional  | Full-screen detail view (right-side or modal)                         | `packages/builtin-tools/src/portals.ts`       |
-
-The two reference tools to read end-to-end:
-
- **`builtin-tool-web-browsing/src/client/`** — Inspector + Render + Placeholder + Portal (no Intervention/Streaming).
- **`builtin-tool-local-system/src/client/`** — all six surfaces, including `components/` for shared building blocks.
-
---
-
-## Tool Render 设计原则（中文草案）
-
-这些原则用于判断一个 builtin tool 的 Inspector / Render / Placeholder / Streaming / Intervention / Portal 应该做什么，以及做到什么程度。
-
-1. **先保证折叠态可读。** 每个 API 都必须有 Inspector；用户不展开也应该能看懂 “正在做什么 / 对什么做 / 当前结果是什么”。Inspector 不应该只展示函数名和原始参数。
-2. **Inspector 是一句话，不是详情页。** 优先表达动作、关键对象、数量、状态，例如 “分析图片 3 张”“搜索 12 个结果”“读取 config.json”。长文本、列表和结构化结果放到 Render 或 Portal。
-3. **Inspector 要覆盖执行生命周期。** `args` 还在 streaming、工具执行中、执行完成、执行失败时都应该有稳定展示；必要时同时读取 `args`、`partialArgs` 和 `pluginState`，避免出现空白、跳变或只显示半截参数。
-4. **只有结构化结果才需要 Render。** 如果工具结果只是自然语言总结，通常不需要 Render；如果结果包含列表、媒体、文件、表格、代码、diff、地图、时间线、权限请求等结构，就应该提供 Render。
-5. **Render 要帮助用户检查结果，而不是复述参数。** Render 的主体应该围绕工具产物组织：可预览、可比较、可筛选、可定位。参数只作为上下文辅助出现，不要把 Render 做成一块更大的 args dump。
-6. **参数和结果要一起参与渲染。** 好的 Tool UI 通常同时用 `args` 解释意图，用 `pluginState` 展示真实执行结果；但 `pluginState` 只放结果域数据，不要反向塞入可以从 `args` 推导出的内容。
-7. **慢操作要有 Placeholder。** 如果工具通常需要等待网络、文件系统、模型或外部进程，Placeholder 应该先占住最终 Render 的版式，让用户知道即将看到什么，而不是只显示一个泛化 loading。
-8. **Streaming 只用于连续产物。** 搜索列表、日志、长文本、文件分析、分阶段计划适合 Streaming；一次性小结果不需要强行做 Streaming。Streaming UI 要能渐进追加，并且完成后自然过渡到最终 Render。
-9. **有风险的动作必须 Intervention。** 写文件、删除、发送、安装、执行命令、外部可见操作、权限敏感操作，都应该在执行前给出可理解的确认界面；确认文案要说明影响范围，而不是只问 “是否继续”。
-10. **错误、空态和截断都是正式状态。** Render 不能在失败、无结果、超长结果时退化成空白。错误要说明发生在哪一步；空态要告诉用户没有产物；超长内容要明确 “展示前 N 项 / 还有 N 项”。
-11. **信息密度要克制。** 默认展示最有判断价值的部分：标题、来源、状态、摘要、少量关键字段。大对象、长列表、原文、调试数据放进可展开区域或 Portal，避免把聊天流撑成后台管理页。
-12. **视觉上融入聊天流。** Tool UI 应该使用 `@lobehub/ui` / base-ui、`Flexbox`、`createStaticStyles` 和 `cssVar.*`，遵循现有间距、圆角、颜色、字号；不要为单个工具发明一套独立视觉语言。
-13. **Devtools fixture 是验收入口。** 新增或修改 Tool UI 时，应在 `/devtools` 里准备覆盖典型态、loading/streaming、空态、错误态、长内容态的 fixture；一个 API 如果在真实聊天里会出现，就不应该在 devtools 中缺席。
-14. **先做用户会看的 UI，再做调试 UI。** Raw JSON、trace、schema、内部 id 可以存在，但应默认收起或放到调试区；主界面先回答用户最关心的问题：工具做了什么，结果值不值得信任，下一步能做什么。
-
---
-
-## 0. Shared Style Rules
-
-These apply across every surface.
-
-### 0.1 Use `'use client'` at the top of every component file
-
-Tool surfaces are leaves in the chat tree and must not block server rendering.
-
-### 0.2 Prefer `createStaticStyles + cssVar.*`
-
-Zero-runtime CSS-in-JS — the styles compile once and read CSS variables at runtime.
-
-```tsx
-import { createStaticStyles, cssVar } from 'antd-style';
-
-const styles = createStaticStyles(({ css, cssVar }) => ({
-  chip: css`
-    padding-block: 2px;
-    padding-inline: 8px;
-    border-radius: 999px;
-    color: ${cssVar.colorText};
-    background: ${cssVar.colorFillTertiary};
-  `,
-}));
-```
-
-Fall back to `createStyles + token` only when you need runtime token computation (rare). Inline `style={{ color: cssVar.colorTextSecondary }}` is fine for one-off dynamic values.
-
-### 0.3 Use `@lobehub/ui`, not raw `antd`
-
-`Block`, `Text`, `Flexbox`, `Highlighter`, `Alert`, `Tooltip`, `Skeleton` all come from `@lobehub/ui`. Modals come from `@lobehub/ui/base-ui` (`createModal`, `useModalContext`, `confirmModal`) — see the **modal** skill.
-
-Memory note: `@lobehub/ui`'s `<Text type='secondary'>` is a lighter shade than `colorTextSecondary`. If you need that exact token color, write `<Text style={{ color: cssVar.colorTextSecondary }}>`.
-
-### 0.4 Always `memo` and set `displayName`
-
-```tsx
-export const SearchInspector = memo<BuiltinInspectorProps<SearchQuery, UniformSearchResponse>>(
-  ({ args /* … */ }) => {
-    /* … */
-  },
-);
-SearchInspector.displayName = 'SearchInspector';
-export default SearchInspector;
-```
-
-### 0.5 Always type with `BuiltinXProps<Args, State>` generics
-
-Don't widen to `any`. The Args generic is the JSON Schema params, the State generic is the executor's `state` field. The two should match `<Name>Params` and `<Name>State` from `types.ts`.
-
-### 0.6 Pull strings from `t('plugin')`
-
-```tsx
-const { t } = useTranslation('plugin');
-t('builtins.<identifier>.apiName.<api>');
-```
-
-Every Inspector should default to `t('builtins.<identifier>.apiName.<api>')` so it shows something while args stream in.
-
-### 0.7 Read store state from `@/store/chat`, not props
-
-Tool surfaces sometimes need cross-cutting state (loading, streaming buffer). Read it inside the component via Zustand selectors, not from props — props only carry args/state/messageId.
-
---
-
-## 1. Inspector — Header Chip (required)
-
-**Lifecycle:** Inspector renders for **every phase** of a tool call: while args are streaming in, while the executor is running, and after results come back. It's the only surface that's always visible.
-
-**Goal:** keep it to a single line. Show what's happening with as much context as is currently available.
-
-### Props (`BuiltinInspectorProps<Args, State>`)
-
-```ts
-interface BuiltinInspectorProps<Arguments = any, State = any> {
-  apiName: string;
-  args: Arguments; // final args (only after the assistant stops streaming)
-  identifier: string;
-  isArgumentsStreaming?: boolean; // args still arriving
-  isLoading?: boolean; // args complete, executor running
-  partialArgs?: Arguments; // partial JSON during streaming
-  pluginState?: State; // executor's `state` after success
-  result?: { content: string | null; error?: any };
-}
-```
-
-### State machine
-
-| Phase                               | What's available                                           | What to show                                               |
-| ----------------------------------- | ---------------------------------------------------------- | ---------------------------------------------------------- |
-| Args streaming, no useful field yet | `isArgumentsStreaming === true`, `partialArgs.X` undefined | Just the API title with `shinyTextStyles.shinyText`        |
-| Args streaming, key field arrived   | `partialArgs.X` populated                                  | Title + key field chip, still pulse-animated               |
-| Args complete, executor running     | `args` populated, `isLoading === true`                     | Same as above, still pulse-animated                        |
-| Result arrived                      | `pluginState` populated, `isLoading === false`             | Title + chips + result summary (count, identifier, status) |
-
-### Canonical example — Search
-
-`packages/builtin-tool-web-browsing/src/client/Inspector/Search/index.tsx`:
-
-```tsx
-'use client';
-
-import type { BuiltinInspectorProps, SearchQuery, UniformSearchResponse } from '@lobechat/types';
-import { Text } from '@lobehub/ui';
-import { cssVar, cx } from 'antd-style';
-import { memo } from 'react';
-import { useTranslation } from 'react-i18next';
-
-import { highlightTextStyles, inspectorTextStyles, shinyTextStyles } from '@/styles';
-
-export const SearchInspector = memo<BuiltinInspectorProps<SearchQuery, UniformSearchResponse>>(
-  ({ args, partialArgs, isArgumentsStreaming, isLoading, pluginState }) => {
-    const { t } = useTranslation('plugin');
-
-    const query = args?.query || partialArgs?.query || '';
-    const resultCount = pluginState?.results?.length ?? 0;
-    const hasResults = resultCount > 0;
-
-    if (isArgumentsStreaming && !query) {
-      return (
-        <div className={cx(inspectorTextStyles.root, shinyTextStyles.shinyText)}>
-          <span>{t('builtins.lobe-web-browsing.apiName.search')}</span>
-        </div>
-      );
-    }
-
-    return (
-      <div
-        className={cx(
-          inspectorTextStyles.root,
-          (isArgumentsStreaming || isLoading) && shinyTextStyles.shinyText,
-        )}
-      >
-        <span>{t('builtins.lobe-web-browsing.apiName.search')}:&nbsp;</span>
-        {query && <span className={highlightTextStyles.primary}>{query}</span>}
-        {!isLoading &&
-          !isArgumentsStreaming &&
-          pluginState?.results &&
-          (hasResults ? (
-            <span style={{ marginInlineStart: 4 }}>({resultCount})</span>
-          ) : (
-            <Text as="span" color={cssVar.colorTextDescription} fontSize={12}>
-              ({t('builtins.lobe-web-browsing.inspector.noResults')})
-            </Text>
-          ))}
-      </div>
-    );
-  },
-);
-SearchInspector.displayName = 'SearchInspector';
-export default SearchInspector;
-```
-
-### Inspector rules
-
- Wrap the whole row with `inspectorTextStyles.root` (provides correct flex / line-height baseline).
- Pulse with `shinyTextStyles.shinyText` whenever `isArgumentsStreaming || isLoading`.
- Show the i18n title first so the row is non-empty during the earliest streaming phase.
- Read both `args?.X` and `partialArgs?.X` together — `args` is final, `partialArgs` is in-stream.
- Use chips/tags for distinct facets (identifier, name, parent, status, count). Each chip should clip with `text-overflow: ellipsis` and have a `max-width` so long values don't blow out the chat bubble.
- Append `pluginState`-derived suffixes only **after** loading finishes — count or "(no results)" should not appear while still searching.
-
-### Inspector registry — `client/Inspector/index.ts`
-
-```ts
-import type { BuiltinInspector } from '@lobechat/types';
-
-import { TaskApiName } from '../../types';
-import { CreateTaskInspector } from './CreateTask';
-import { ListTasksInspector } from './ListTasks';
-/* … */
-
-export const TaskInspectors: Record<string, BuiltinInspector> = {
-  [TaskApiName.createTask]: CreateTaskInspector as BuiltinInspector,
-  [TaskApiName.listTasks]: ListTasksInspector as BuiltinInspector,
-  /* one entry per ApiName */
-};
-
-export { CreateTaskInspector } from './CreateTask';
-export { ListTasksInspector } from './ListTasks';
-/* re-export each */
-```
-
---
-
-## 2. Render — Rich Result Card (optional)
-
-**Lifecycle:** rendered **once the result arrives** (after Placeholder/Streaming hand off). Sits below the Inspector header.
-
-**Skip if** the API is read-only or the result is just text — the framework already shows the executor's `content` string. Add a Render only when there's a structured artifact worth seeing: a card, a chart, a diff, a list of files.
-
-### Props (`BuiltinRenderProps<Args, State, Content>`)
-
-```ts
-interface BuiltinRenderProps<Arguments = any, State = any, Content = any> {
-  apiName?: string;
-  args: Arguments; // final params from the LLM
-  content: Content; // executor's content string (or parsed)
-  identifier?: string;
-  messageId: string; // for store lookups
-  pluginError?: any; // from BuiltinToolResult.error
-  pluginState?: State; // executor's state
-  toolCallId?: string;
-}
-```
-
-### Two patterns
-
-**Pattern A — Single-file Render** (web-browsing CrawlSinglePage):
-
-```tsx
-// client/Render/CrawlSinglePage.tsx
-import type { BuiltinRenderProps, CrawlPluginState, CrawlSinglePageQuery } from '@lobechat/types';
-import { memo } from 'react';
-
-import PageContent from './PageContent';
-
-const CrawlSinglePage = memo<BuiltinRenderProps<CrawlSinglePageQuery, CrawlPluginState>>(
-  ({ messageId, pluginState, args }) => (
-    <PageContent messageId={messageId} results={pluginState?.results} urls={[args?.url]} />
-  ),
-);
-export default CrawlSinglePage;
-```
-
-**Pattern B — Folder with subcomponents** (web-browsing Search):
-
-```
-client/Render/Search/
-├── index.tsx           # composes the subcomponents, handles error states
-├── ConfigForm.tsx      # appears when pluginError.type === 'PluginSettingsInvalid'
-├── SearchQuery.tsx     # editable query header
-└── SearchResult.tsx    # result list
-```
-
-Use Pattern B when the Render has internal state (editing mode, expanded items), error variants, or is large enough to benefit from splitting.
-
-### Error handling in Render
-
-Renders are the canonical place to surface `pluginError` because the chat doesn't auto-render typed errors:
-
-```tsx
-if (pluginError) {
-  if (pluginError?.type === 'PluginSettingsInvalid') {
-    return <ConfigForm id={messageId} provider={pluginError.body?.provider} />;
-  }
-  return (
-    <Alert
-      title={pluginError?.message}
-      type="error"
-      extra={<Highlighter language="json">{JSON.stringify(pluginError.body, null, 2)}</Highlighter>}
-    />
-  );
-}
-```
-
-### Render rules
-
- **Return `null`** if there's nothing useful to draw yet (avoids empty cards during stream).
- Use `pluginState` for server-truth (ids, counts, server-assigned status) and `args` for what the LLM asked. **Combine — neither alone is enough.**
- For lists, summarize with a header line and show top N items with a "+N more" tail rather than rendering everything.
- For modals from a Render, use `@lobehub/ui/base-ui` (`createModal`, `useModalContext`, `confirmModal`) — see the **modal** skill.
-
-### Render registry — `client/Render/index.ts`
-
-```ts
-import type { BuiltinRender } from '@lobechat/types';
-
-import { TaskApiName } from '../../types';
-import CreateTaskRender from './CreateTask';
-import RunTasksRender from './RunTasks';
-
-export const TaskRenders: Record<string, BuiltinRender> = {
-  [TaskApiName.createTask]: CreateTaskRender as BuiltinRender,
-  [TaskApiName.runTasks]: RunTasksRender as BuiltinRender,
-  /* only the APIs with rich result UI — others fall back to text content */
-};
-
-export { default as CreateTaskRender } from './CreateTask';
-export { default as RunTasksRender } from './RunTasks';
-```
-
-### Render display control (rare)
-
-If the Render should hide for certain results (e.g. ClaudeCode's TodoWrite hides when the agent is mid-stream), add a `RenderDisplayControl` to `packages/builtin-tools/src/displayControls.ts`. See `ClaudeCodeRenderDisplayControls` for the pattern.
-
---
-
-## 3. Placeholder — Skeleton Between Args and Result (optional)
-
-**Lifecycle:** rendered when the args have finished streaming but the executor hasn't returned yet. Disappears when `pluginState` arrives. Bridges the moment of perceived lag.
-
-**Add for** APIs with noticeable execution time: web search, network crawl, file list, large grep. **Skip for** instant ops (status flips, calculator).
-
-### Props (`BuiltinPlaceholderProps<Args>`)
-
-```ts
-interface BuiltinPlaceholderProps<T extends Record<string, any> = any> {
-  apiName: string;
-  args?: T;
-  identifier: string;
-}
-```
-
-No `pluginState` — Placeholder lives entirely in the "executing" gap.
-
-### Canonical example — Search Placeholder
-
-`packages/builtin-tool-web-browsing/src/client/Placeholder/Search.tsx`:
-
-```tsx
-import type { BuiltinPlaceholderProps, SearchQuery } from '@lobechat/types';
-import { Flexbox, Icon, Skeleton } from '@lobehub/ui';
-import { createStaticStyles, cx } from 'antd-style';
-import { SearchIcon } from 'lucide-react';
-import { memo } from 'react';
-
-import { useIsMobile } from '@/hooks/useIsMobile';
-import { shinyTextStyles } from '@/styles';
-
-const styles = createStaticStyles(({ css, cssVar }) => ({
-  query: cx(
-    css`
-      padding: 4px 8px;
-      border-radius: 8px;
-      font-size: 12px;
-      color: ${cssVar.colorTextSecondary};
-      &:hover {
-        background: ${cssVar.colorFillTertiary};
-      }
-    `,
-    shinyTextStyles.shinyText,
-  ),
-}));
-
-export const Search = memo<BuiltinPlaceholderProps<SearchQuery>>(({ args }) => {
-  const { query } = args || {};
-  const isMobile = useIsMobile();
-
-  return (
-    <Flexbox gap={8}>
-      <Flexbox horizontal={!isMobile} gap={isMobile ? 8 : 40}>
-        <Flexbox horizontal align="center" className={styles.query} gap={8}>
-          <Icon icon={SearchIcon} />
-          {query ? query : <Skeleton.Block active style={{ height: 20, width: 40 }} />}
-        </Flexbox>
-        <Skeleton.Block active style={{ height: 20, width: 40 }} />
-      </Flexbox>
-      <Flexbox horizontal gap={12}>
-        {[1, 2, 3, 4, 5].map((id) => (
-          <Skeleton.Button active key={id} style={{ borderRadius: 8, height: 80, width: 160 }} />
-        ))}
-      </Flexbox>
-    </Flexbox>
-  );
-});
-```
-
-### Placeholder rules
-
- **Mirror the eventual Render's layout.** When the result arrives the Placeholder unmounts and the Render mounts; if they share dimensions, the chat doesn't jump.
- Use `Skeleton.Block` / `Skeleton.Button` from `@lobehub/ui` for placeholder shapes.
- Embed any args you have (e.g. the query text) — context helps the user know what's loading.
- Pulse with `shinyTextStyles.shinyText` if the Placeholder includes literal text.
-
-### Placeholder registry — `client/Placeholder/index.ts`
-
-```ts
-import { WebBrowsingApiName } from '../../types';
-import CrawlMultiPages from './CrawlMultiPages';
-import CrawlSinglePage from './CrawlSinglePage';
-import { Search } from './Search';
-
-export const WebBrowsingPlaceholders = {
-  [WebBrowsingApiName.crawlMultiPages]: CrawlMultiPages,
-  [WebBrowsingApiName.crawlSinglePage]: CrawlSinglePage,
-  [WebBrowsingApiName.search]: Search,
-};
-
-export { CrawlMultiPages, CrawlSinglePage, Search };
-```
-
---
-
-## 4. Streaming — Live Output During Execution (optional)
-
-**Lifecycle:** rendered **while the executor is still running** for APIs that emit incremental output. The component is responsible for fetching the in-flight stream from the chat store and rendering it.
-
-**Add for** long-running ops with continuous output: shell command execution (stdout/stderr), file write progress, code interpreter cells.
-
-### Props (`BuiltinStreamingProps<Args>`)
-
-```ts
-interface BuiltinStreamingProps<Arguments = any> {
-  apiName: string;
-  args: Arguments;
-  identifier: string;
-  messageId: string; // use to fetch the streaming buffer from store
-  toolCallId: string;
-}
-```
-
-Note there's **no `state` or `result` prop** — the Streaming component is for the in-flight phase. It pulls the live buffer from the store itself (typically via `chatToolSelectors.streamingContent(messageId)` or similar).
-
-### Canonical example — RunCommandStreaming
-
-`packages/builtin-tool-local-system/src/client/Streaming/RunCommand/index.tsx`:
-
-```tsx
-'use client';
-
-import type { BuiltinStreamingProps } from '@lobechat/types';
-import { Highlighter } from '@lobehub/ui';
-import { memo } from 'react';
-
-interface RunCommandParams {
-  command?: string;
-  description?: string;
-  timeout?: number;
-}
-
-export const RunCommandStreaming = memo<BuiltinStreamingProps<RunCommandParams>>(({ args }) => {
-  const { command } = args || {};
-  if (!command) return null;
-
-  return (
-    <Highlighter
-      animated
-      wrap
-      language="sh"
-      showLanguage={false}
-      style={{ padding: '4px 8px' }}
-      variant="outlined"
-    >
-      {command}
-    </Highlighter>
-  );
-});
-RunCommandStreaming.displayName = 'RunCommandStreaming';
-```
-
-For real-time output beyond just the command (stderr/stdout streaming), pull from the chat store:
-
-```tsx
-const buffer = useChatStore((state) =>
-  chatToolSelectors.streamingBuffer(messageId, toolCallId)(state),
-);
-```
-
-### Streaming rules
-
- Render `null` until you have something to display (avoids flash).
- For terminal-style output, use `Highlighter` with `animated` to show typing-like effect.
- The Streaming component must **unmount cleanly** when execution ends — typically the framework swaps it out for the Render automatically.
-
-### Streaming registry — `client/Streaming/index.ts`
-
-```ts
-import { LocalSystemApiName } from '../..';
-import { RunCommandStreaming } from './RunCommand';
-import { WriteFileStreaming } from './WriteFile';
-
-export const LocalSystemStreamings = {
-  [LocalSystemApiName.runCommand]: RunCommandStreaming,
-  [LocalSystemApiName.writeLocalFile]: WriteFileStreaming,
-};
-```
-
---
-
-## 5. Intervention — Approval / Edit-Before-Run (optional)
-
-**Lifecycle:** rendered **before the executor runs** for APIs whose manifest sets `humanIntervention`. The user sees a preview of the args, can edit them, then approves or skips/cancels.
-
-**Add for** destructive or sensitive ops: shell commands, file writes, file moves, payments, message broadcasts.
-
-### Props (`BuiltinInterventionProps<Args>`)
-
-```ts
-interface BuiltinInterventionProps<Arguments = any> {
-  apiName?: string;
-  args: Arguments;
-  identifier?: string;
-  interactionMode?: 'approval' | 'custom';
-  messageId: string;
-
-  /** Called when the user edits the args; the approve action awaits this. */
-  onArgsChange?: (args: Arguments) => void | Promise<void>;
-
-  /** Called on approve / skip / cancel. */
-  onInteractionAction?: (
-    action:
-      | { type: 'submit'; payload: Record<string, unknown> }
-      | { type: 'skip'; payload?: Record<string, unknown>; reason?: string }
-      | { type: 'cancel'; payload?: Record<string, unknown> },
-  ) => Promise<void>;
-
-  /** Register a callback to flush pending saves before approval. Returns cleanup. */
-  registerBeforeApprove?: (id: string, callback: () => void | Promise<void>) => () => void;
-}
-```
-
-### Canonical example — RunCommand Intervention
-
-`packages/builtin-tool-local-system/src/client/Intervention/RunCommand/index.tsx`:
-
-```tsx
-import type { RunCommandParams } from '@lobechat/electron-client-ipc';
-import type { BuiltinInterventionProps } from '@lobechat/types';
-import { Flexbox, Highlighter, Text } from '@lobehub/ui';
-import { memo } from 'react';
-
-const RunCommand = memo<BuiltinInterventionProps<RunCommandParams>>(({ args }) => {
-  const { description, command, timeout } = args;
-  return (
-    <Flexbox gap={8}>
-      <Flexbox horizontal justify="space-between">
-        {description && <Text>{description}</Text>}
-        {timeout && (
-          <Text style={{ fontSize: 12 }} type="secondary">
-            timeout: {formatTimeout(timeout)}
-          </Text>
-        )}
-      </Flexbox>
-      {command && (
-        <Highlighter wrap language="sh" showLanguage={false} variant="outlined">
-          {command}
-        </Highlighter>
-      )}
-    </Flexbox>
-  );
-});
-export default RunCommand;
-```
-
-### Intervention rules
-
- **Show a preview, not a form by default.** Editing UI is opt-in via `onArgsChange` and is usually inline (click to edit a code block, etc.).
- For args with debounced edit state (text fields), use `registerBeforeApprove(id, flushFn)` so the approve action waits for the debounce to flush. Always return the cleanup function.
- Call `onInteractionAction({ type: 'submit', payload })` when the user approves; `'skip'` if they skip with a reason; `'cancel'` if they cancel the whole turn.
- Add a corresponding `interventionAudit.ts` in the package root if the tool needs scope/path validation before approval (see `local-system/src/interventionAudit.ts`).
-
-### Intervention registry — `client/Intervention/index.ts`
-
-```ts
-import { LocalSystemApiName } from '../..';
-import EditLocalFile from './EditLocalFile';
-import RunCommand from './RunCommand';
-import WriteFile from './WriteFile';
-/* … */
-
-export const LocalSystemInterventions = {
-  [LocalSystemApiName.editLocalFile]: EditLocalFile,
-  [LocalSystemApiName.runCommand]: RunCommand,
-  [LocalSystemApiName.writeLocalFile]: WriteFile,
-  /* one entry per API that needs approval */
-};
-```
-
---
-
-## 6. Portal — Full-Screen Detail View (optional)
-
-**Lifecycle:** rendered when the user opens the tool message in a side panel or full-screen modal. One Portal per **tool**, not per API — the Portal switches on `apiName` internally.
-
-**Add for** tools whose results deserve a deep-dive view: search results with editable filters, page content with reader mode, code interpreter sessions.
-
-### Props (`BuiltinPortalProps<Args, State>`)
-
-```ts
-interface BuiltinPortalProps<Arguments = Record<string, any>, State = any> {
-  apiName?: string;
-  arguments: Arguments;
-  identifier: string;
-  messageId: string;
-  state: State;
-}
-```
-
-### Canonical example — Web-Browsing Portal
-
-`packages/builtin-tool-web-browsing/src/client/Portal/index.tsx`:
-
-```tsx
-import type { BuiltinPortalProps, CrawlPluginState, SearchQuery } from '@lobechat/types';
-import { memo } from 'react';
-
-import { WebBrowsingApiName } from '../../types';
-import PageContent from './PageContent';
-import PageContents from './PageContents';
-import Search from './Search';
-
-const Portal = memo<BuiltinPortalProps>(({ arguments: args, messageId, state, apiName }) => {
-  switch (apiName) {
-    case WebBrowsingApiName.search:
-      return <Search messageId={messageId} query={args as SearchQuery} response={state} />;
-
-    case WebBrowsingApiName.crawlSinglePage: {
-      const result = (state as CrawlPluginState).results.find((r) => r.originalUrl === args.url);
-      return <PageContent messageId={messageId} result={result} />;
-    }
-
-    case WebBrowsingApiName.crawlMultiPages:
-      return (
-        <PageContents
-          messageId={messageId}
-          results={(state as CrawlPluginState).results}
-          urls={args.urls}
-        />
-      );
-  }
-  return null;
-});
-export default Portal;
-```
-
-### Portal rules
-
- One Portal per tool — the file is the routing layer, subcomponents implement each API's view.
- Portals can read the chat store directly to detect "still streaming" and render a Skeleton internally (see `Search/index.tsx:20-46`).
- Layout assumes more space than the Render — use `Flexbox` with `height={'100%'}` and structure for a side panel viewport.
-
-### Portal registry — `packages/builtin-tools/src/portals.ts`
-
-```ts
-import { WebBrowsingManifest, WebBrowsingPortal } from '@lobechat/builtin-tool-web-browsing/client';
-import { type BuiltinPortal } from '@lobechat/types';
-
-export const BuiltinToolsPortals: Record<string, BuiltinPortal> = {
-  [WebBrowsingManifest.identifier]: WebBrowsingPortal as BuiltinPortal,
-};
-```
-
---
-
-## 7. `client/components/` — Shared Subcomponents
-
-Cross-cutting building blocks used by multiple surfaces live here, not duplicated in each surface folder.
-
-Examples from `web-browsing/src/client/components/`:
-
- `CategoryAvatar.tsx` — search category icon
- `EngineAvatar.tsx` — search engine logo (used in Inspector chip + Render list + Portal header)
- `SearchBar.tsx` — editable query bar (used in Render and Portal)
-
-Examples from `local-system/src/client/components/`:
-
- `FileItem.tsx` — single file row (used in ListFiles Render, SearchFiles Render, MoveLocalFiles Render)
- `FilePathDisplay.tsx` — path with truncation (used everywhere)
-
-### Rules
-
- Live under `client/components/`, exported via `client/components/index.ts`.
- Re-export from `client/index.ts` only if other packages need them; otherwise keep internal.
- Keep them dumb — props in, JSX out, no store reads. The store reads belong in the surface that composes them.
-
---
-
-## 8. `client/index.ts` — Package Public API
-
-Re-exports everything the registries need plus useful types/manifest:
-
-```ts
-// Inspector — required
-export { TaskInspectors } from './Inspector';
-
-// Render — only if any API has one
-export { TaskRenders, CreateTaskRender, RunTasksRender } from './Render';
-
-// Placeholder / Streaming / Intervention — only if used
-export { LocalSystemListFilesPlaceholder, LocalSystemSearchFilesPlaceholder } from './Placeholder';
-export { LocalSystemStreamings } from './Streaming';
-export { LocalSystemInterventions } from './Intervention';
-
-// Portal — single export per tool
-export { default as WebBrowsingPortal } from './Portal';
-
-// Reusable components if other packages need them
-export { CategoryAvatar, EngineAvatar, SearchBar } from './components';
-
-// Re-export manifest, identifier, types for convenience
-export { TaskManifest, TaskIdentifier } from '../manifest';
-export * from '../types';
-```
-
---
-
-## 9. Diagnostic Quick-Lookup
-
-| Symptom                                         | Surface to check                                                                                                  |     |                           |
-| ----------------------------------------------- | ----------------------------------------------------------------------------------------------------------------- | --- | ------------------------- |
-| No header at all on the tool call               | Inspector missing from `client/Inspector/index.ts` registry                                                       |     |                           |
-| Header shows the API name but no chips          | Inspector missing \`args?.X                                                                                       |     | partialArgs?.X\` fallback |
-| Header doesn't pulse during loading             | Missing `shinyTextStyles.shinyText` on `isArgumentsStreaming \|\| isLoading`                                      |     |                           |
-| Empty result card under header                  | Render returned `<div />` instead of `null` when no data                                                          |     |                           |
-| Layout jump when result arrives                 | Placeholder dimensions don't match Render dimensions                                                              |     |                           |
-| Approval dialog never appears                   | Manifest missing `humanIntervention`, or Intervention not in registry                                             |     |                           |
-| Approval click doesn't wait for inline edit     | Missing `registerBeforeApprove(id, flushFn)`                                                                      |     |                           |
-| Portal opens but blank                          | Switch in `Portal/index.tsx` doesn't cover the apiName                                                            |     |                           |
-| Strings show as `builtins.lobe-foo.apiName.bar` | Missing i18n key in `src/locales/default/plugin.ts` (or not seeded in dev locale files)                           |     |                           |
-| Wrong color shade on `<Text type="secondary">`  | `type='secondary'` is lighter than `colorTextSecondary` — pass via `style={{ color: cssVar.colorTextSecondary }}` |     |                           |
@@ -0,0 +1,36 @@
+# Tool UI Surfaces
+
+A builtin tool can ship up to **six client-side surfaces**, each with a different role in the chat UI. Only `Inspector` is required; the other five are added on demand and registered in their own central files.
+
+| Surface      | Required? | When the chat shows it                                                | Registered in                                 |
+| ------------ | --------- | --------------------------------------------------------------------- | --------------------------------------------- |
+| Inspector    | ✅ Always | Header strip of every tool call (one-line chip)                       | `packages/builtin-tools/src/inspectors.ts`    |
+| Render       | Optional  | Rich result card below the header, after the call returns             | `packages/builtin-tools/src/renders.ts`       |
+| Placeholder  | Optional  | Skeleton between "args streaming complete" and "result arrives"       | `packages/builtin-tools/src/placeholders.ts`  |
+| Streaming    | Optional  | Live output during execution (e.g. command stdout)                    | `packages/builtin-tools/src/streamings.ts`    |
+| Intervention | Optional  | Approval / edit-before-run dialog (when `humanIntervention` triggers) | `packages/builtin-tools/src/interventions.ts` |
+| Portal       | Optional  | Full-screen detail view (right-side or modal)                         | `packages/builtin-tools/src/portals.ts`       |
+
+The two reference tools to read end-to-end:
+
+- **`builtin-tool-web-browsing/src/client/`** — Inspector + Render + Placeholder + Portal (no Intervention/Streaming).
+- **`builtin-tool-local-system/src/client/`** — all six surfaces, including `components/` for shared building blocks.
+
+---
+
+## Files in this folder
+
+Read **principles** and **shared-rules** first — they apply to every surface. Then jump to the surface you're building.
+
+| File                               | What it covers                                                          |
+| ---------------------------------- | ----------------------------------------------------------------------- |
+| [principles.md](principles.md)     | Design principles — when each surface exists and how far to take it     |
+| [shared-rules.md](shared-rules.md) | Cross-surface rules: component skeleton, styling, single-layer surfaces |
+| [inspector.md](inspector.md)       | Inspector — header chip (required)                                      |
+| [render.md](render.md)             | Render — rich result card                                               |
+| [placeholder.md](placeholder.md)   | Placeholder — skeleton between args and result                          |
+| [streaming.md](streaming.md)       | Streaming — live output during execution                                |
+| [intervention.md](intervention.md) | Intervention — approval / edit-before-run                               |
+| [portal.md](portal.md)             | Portal — full-screen detail view                                        |
+| [composition.md](composition.md)   | Shared subcomponents (`client/components/`) + package public API        |
+| [diagnostics.md](diagnostics.md)   | Symptom → surface quick-lookup                                          |
@@ -0,0 +1,51 @@
+# Composition — Shared Components & Package API
+
+## `client/components/` — Shared Subcomponents
+
+Cross-cutting building blocks used by multiple surfaces live here, not duplicated in each surface folder.
+
+Examples from `web-browsing/src/client/components/`:
+
+- `CategoryAvatar.tsx` — search category icon
+- `EngineAvatar.tsx` — search engine logo (used in Inspector chip + Render list + Portal header)
+- `SearchBar.tsx` — editable query bar (used in Render and Portal)
+
+Examples from `local-system/src/client/components/`:
+
+- `FileItem.tsx` — single file row (used in ListFiles Render, SearchFiles Render, MoveLocalFiles Render)
+- `FilePathDisplay.tsx` — path with truncation (used everywhere)
+
+### Rules
+
+- Live under `client/components/`, exported via `client/components/index.ts`.
+- Re-export from `client/index.ts` only if other packages need them; otherwise keep internal.
+- Keep them dumb — props in, JSX out, no store reads. The store reads belong in the surface that composes them.
+
+---
+
+## `client/index.ts` — Package Public API
+
+Re-exports everything the registries need plus useful types/manifest:
+
+```ts
+// Inspector — required
+export { TaskInspectors } from './Inspector';
+
+// Render — only if any API has one
+export { TaskRenders, CreateTaskRender, RunTasksRender } from './Render';
+
+// Placeholder / Streaming / Intervention — only if used
+export { LocalSystemListFilesPlaceholder, LocalSystemSearchFilesPlaceholder } from './Placeholder';
+export { LocalSystemStreamings } from './Streaming';
+export { LocalSystemInterventions } from './Intervention';
+
+// Portal — single export per tool
+export { default as WebBrowsingPortal } from './Portal';
+
+// Reusable components if other packages need them
+export { CategoryAvatar, EngineAvatar, SearchBar } from './components';
+
+// Re-export manifest, identifier, types for convenience
+export { TaskManifest, TaskIdentifier } from '../manifest';
+export * from '../types';
+```
@@ -0,0 +1,15 @@
+# Diagnostic Quick-Lookup
+
+| Symptom                                         | Surface to check                                                                                                                      |
+| ----------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------- |
+| No header at all on the tool call               | Inspector missing from `client/Inspector/index.ts` registry                                                                           |
+| Header shows the API name but no chips          | Inspector missing `args?.X \|\| partialArgs?.X` fallback                                                                              |
+| Header doesn't pulse during loading             | Missing `shinyTextStyles.shinyText` on `isArgumentsStreaming \|\| isLoading`                                                          |
+| Empty result card under header                  | Render returned `<div />` instead of `null` when no data                                                                              |
+| Render looks "complex" / card-in-card           | Filled container (`colorFillQuaternary`) wrapping more filled boxes — flatten to single-layer, see [shared-rules.md](shared-rules.md) |
+| Layout jump when result arrives                 | Placeholder dimensions don't match Render dimensions                                                                                  |
+| Approval dialog never appears                   | Manifest missing `humanIntervention`, or Intervention not in registry                                                                 |
+| Approval click doesn't wait for inline edit     | Missing `registerBeforeApprove(id, flushFn)`                                                                                          |
+| Portal opens but blank                          | Switch in `Portal/index.tsx` doesn't cover the apiName                                                                                |
+| Strings show as `builtins.lobe-foo.apiName.bar` | Missing i18n key in `src/locales/default/plugin.ts` (or not seeded in dev locale files)                                               |
+| Wrong color shade on `<Text type="secondary">`  | `type='secondary'` is lighter than `colorTextSecondary` — pass via `style={{ color: cssVar.colorTextSecondary }}`                     |
@@ -0,0 +1,118 @@
+# Inspector — Header Chip (required)
+
+**Lifecycle:** Inspector renders for **every phase** of a tool call: while args are streaming in, while the executor is running, and after results come back. It's the only surface that's always visible.
+
+**Goal:** keep it to a single line. Show what's happening with as much context as is currently available.
+
+## Props (`BuiltinInspectorProps<Args, State>`)
+
+```ts
+interface BuiltinInspectorProps<Arguments = any, State = any> {
+  apiName: string;
+  args: Arguments; // final args (only after the assistant stops streaming)
+  identifier: string;
+  isArgumentsStreaming?: boolean; // args still arriving
+  isLoading?: boolean; // args complete, executor running
+  partialArgs?: Arguments; // partial JSON during streaming
+  pluginState?: State; // executor's `state` after success
+  result?: { content: string | null; error?: any };
+}
+```
+
+## State machine
+
+| Phase                               | What's available                                           | What to show                                               |
+| ----------------------------------- | ---------------------------------------------------------- | ---------------------------------------------------------- |
+| Args streaming, no useful field yet | `isArgumentsStreaming === true`, `partialArgs.X` undefined | Just the API title with `shinyTextStyles.shinyText`        |
+| Args streaming, key field arrived   | `partialArgs.X` populated                                  | Title + key field chip, still pulse-animated               |
+| Args complete, executor running     | `args` populated, `isLoading === true`                     | Same as above, still pulse-animated                        |
+| Result arrived                      | `pluginState` populated, `isLoading === false`             | Title + chips + result summary (count, identifier, status) |
+
+## Canonical example — Search
+
+`packages/builtin-tool-web-browsing/src/client/Inspector/Search/index.tsx`:
+
+```tsx
+'use client';
+
+import type { BuiltinInspectorProps, SearchQuery, UniformSearchResponse } from '@lobechat/types';
+import { Text } from '@lobehub/ui';
+import { cssVar, cx } from 'antd-style';
+import { memo } from 'react';
+import { useTranslation } from 'react-i18next';
+
+import { highlightTextStyles, inspectorTextStyles, shinyTextStyles } from '@/styles';
+
+export const SearchInspector = memo<BuiltinInspectorProps<SearchQuery, UniformSearchResponse>>(
+  ({ args, partialArgs, isArgumentsStreaming, isLoading, pluginState }) => {
+    const { t } = useTranslation('plugin');
+
+    const query = args?.query || partialArgs?.query || '';
+    const resultCount = pluginState?.results?.length ?? 0;
+    const hasResults = resultCount > 0;
+
+    if (isArgumentsStreaming && !query) {
+      return (
+        <div className={cx(inspectorTextStyles.root, shinyTextStyles.shinyText)}>
+          <span>{t('builtins.lobe-web-browsing.apiName.search')}</span>
+        </div>
+      );
+    }
+
+    return (
+      <div
+        className={cx(
+          inspectorTextStyles.root,
+          (isArgumentsStreaming || isLoading) && shinyTextStyles.shinyText,
+        )}
+      >
+        <span>{t('builtins.lobe-web-browsing.apiName.search')}:&nbsp;</span>
+        {query && <span className={highlightTextStyles.primary}>{query}</span>}
+        {!isLoading &&
+          !isArgumentsStreaming &&
+          pluginState?.results &&
+          (hasResults ? (
+            <span style={{ marginInlineStart: 4 }}>({resultCount})</span>
+          ) : (
+            <Text as="span" color={cssVar.colorTextDescription} fontSize={12}>
+              ({t('builtins.lobe-web-browsing.inspector.noResults')})
+            </Text>
+          ))}
+      </div>
+    );
+  },
+);
+SearchInspector.displayName = 'SearchInspector';
+export default SearchInspector;
+```
+
+## Inspector rules
+
+- Wrap the whole row with `inspectorTextStyles.root` (provides correct flex / line-height baseline).
+- Pulse with `shinyTextStyles.shinyText` whenever `isArgumentsStreaming || isLoading`.
+- Show the i18n title first so the row is non-empty during the earliest streaming phase.
+- Read both `args?.X` and `partialArgs?.X` together — `args` is final, `partialArgs` is in-stream.
+- Use chips/tags for distinct facets (identifier, name, parent, status, count). Each chip should clip with `text-overflow: ellipsis` and have a `max-width` so long values don't blow out the chat bubble.
+- Append `pluginState`-derived suffixes only **after** loading finishes — count or "(no results)" should not appear while still searching.
+- **Switch copy by phase.** If the verb implies an ongoing action ("Creating", "Searching", "Listing"), define `<api>.loading` and `<api>.completed` keys and select via `isArgumentsStreaming || isLoading ? loadingKey : completedKey`. Inspector chips persist in chat history — leaving "Creating task" frozen on a finished call reads as if the tool is still running. Read-only labels that are already noun-form ("View task") can keep a single key. See `CallSubAgentInspector` for the canonical two-key pattern.
+
+## Inspector registry — `client/Inspector/index.ts`
+
+```ts
+import type { BuiltinInspector } from '@lobechat/types';
+
+import { TaskApiName } from '../../types';
+import { CreateTaskInspector } from './CreateTask';
+import { ListTasksInspector } from './ListTasks';
+/* … */
+
+export const TaskInspectors: Record<string, BuiltinInspector> = {
+  [TaskApiName.createTask]: CreateTaskInspector as BuiltinInspector,
+  [TaskApiName.listTasks]: ListTasksInspector as BuiltinInspector,
+  /* one entry per ApiName */
+};
+
+export { CreateTaskInspector } from './CreateTask';
+export { ListTasksInspector } from './ListTasks';
+/* re-export each */
+```
@@ -0,0 +1,88 @@
+# Intervention — Approval / Edit-Before-Run (optional)
+
+**Lifecycle:** rendered **before the executor runs** for APIs whose manifest sets `humanIntervention`. The user sees a preview of the args, can edit them, then approves or skips/cancels.
+
+**Add for** destructive or sensitive ops: shell commands, file writes, file moves, payments, message broadcasts.
+
+## Props (`BuiltinInterventionProps<Args>`)
+
+```ts
+interface BuiltinInterventionProps<Arguments = any> {
+  apiName?: string;
+  args: Arguments;
+  identifier?: string;
+  interactionMode?: 'approval' | 'custom';
+  messageId: string;
+
+  /** Called when the user edits the args; the approve action awaits this. */
+  onArgsChange?: (args: Arguments) => void | Promise<void>;
+
+  /** Called on approve / skip / cancel. */
+  onInteractionAction?: (
+    action:
+      | { type: 'submit'; payload: Record<string, unknown> }
+      | { type: 'skip'; payload?: Record<string, unknown>; reason?: string }
+      | { type: 'cancel'; payload?: Record<string, unknown> },
+  ) => Promise<void>;
+
+  /** Register a callback to flush pending saves before approval. Returns cleanup. */
+  registerBeforeApprove?: (id: string, callback: () => void | Promise<void>) => () => void;
+}
+```
+
+## Canonical example — RunCommand Intervention
+
+`packages/builtin-tool-local-system/src/client/Intervention/RunCommand/index.tsx`:
+
+```tsx
+import type { RunCommandParams } from '@lobechat/electron-client-ipc';
+import type { BuiltinInterventionProps } from '@lobechat/types';
+import { Flexbox, Highlighter, Text } from '@lobehub/ui';
+import { memo } from 'react';
+
+const RunCommand = memo<BuiltinInterventionProps<RunCommandParams>>(({ args }) => {
+  const { description, command, timeout } = args;
+  return (
+    <Flexbox gap={8}>
+      <Flexbox horizontal justify="space-between">
+        {description && <Text>{description}</Text>}
+        {timeout && (
+          <Text style={{ fontSize: 12 }} type="secondary">
+            timeout: {formatTimeout(timeout)}
+          </Text>
+        )}
+      </Flexbox>
+      {command && (
+        <Highlighter wrap language="sh" showLanguage={false} variant="outlined">
+          {command}
+        </Highlighter>
+      )}
+    </Flexbox>
+  );
+});
+export default RunCommand;
+```
+
+## Intervention rules
+
+- **Show a preview, not a form by default.** Editing UI is opt-in via `onArgsChange` and is usually inline (click to edit a code block, etc.).
+- For args with debounced edit state (text fields), use `registerBeforeApprove(id, flushFn)` so the approve action waits for the debounce to flush. Always return the cleanup function.
+- Call `onInteractionAction({ type: 'submit', payload })` when the user approves; `'skip'` if they skip with a reason; `'cancel'` if they cancel the whole turn.
+- Add a corresponding `interventionAudit.ts` in the package root if the tool needs scope/path validation before approval (see `local-system/src/interventionAudit.ts`).
+
+## Intervention registry — `client/Intervention/index.ts`
+
+```ts
+import { LocalSystemApiName } from '../..';
+import EditLocalFile from './EditLocalFile';
+import RunCommand from './RunCommand';
+import WriteFile from './WriteFile';
+/* … */
+
+export const LocalSystemInterventions = {
+  [LocalSystemApiName.editLocalFile]: EditLocalFile,
+  [LocalSystemApiName.runCommand]: RunCommand,
+  [LocalSystemApiName.writeLocalFile]: WriteFile,
+  /* one entry per API that needs approval */
+};
+```
@@ -0,0 +1,93 @@
+# Placeholder — Skeleton Between Args and Result (optional)
+
+**Lifecycle:** rendered when the args have finished streaming but the executor hasn't returned yet. Disappears when `pluginState` arrives. Bridges the moment of perceived lag.
+
+**Add for** APIs with noticeable execution time: web search, network crawl, file list, large grep. **Skip for** instant ops (status flips, calculator).
+
+## Props (`BuiltinPlaceholderProps<Args>`)
+
+```ts
+interface BuiltinPlaceholderProps<T extends Record<string, any> = any> {
+  apiName: string;
+  args?: T;
+  identifier: string;
+}
+```
+
+No `pluginState` — Placeholder lives entirely in the "executing" gap.
+
+## Canonical example — Search Placeholder
+
+`packages/builtin-tool-web-browsing/src/client/Placeholder/Search.tsx`:
+
+```tsx
+import type { BuiltinPlaceholderProps, SearchQuery } from '@lobechat/types';
+import { Flexbox, Icon, Skeleton } from '@lobehub/ui';
+import { createStaticStyles, cx } from 'antd-style';
+import { SearchIcon } from 'lucide-react';
+import { memo } from 'react';
+
+import { useIsMobile } from '@/hooks/useIsMobile';
+import { shinyTextStyles } from '@/styles';
+
+const styles = createStaticStyles(({ css, cssVar }) => ({
+  query: cx(
+    css`
+      padding: 4px 8px;
+      border-radius: 8px;
+      font-size: 12px;
+      color: ${cssVar.colorTextSecondary};
+      &:hover {
+        background: ${cssVar.colorFillTertiary};
+      }
+    `,
+    shinyTextStyles.shinyText,
+  ),
+}));
+
+export const Search = memo<BuiltinPlaceholderProps<SearchQuery>>(({ args }) => {
+  const { query } = args || {};
+  const isMobile = useIsMobile();
+
+  return (
+    <Flexbox gap={8}>
+      <Flexbox horizontal={!isMobile} gap={isMobile ? 8 : 40}>
+        <Flexbox horizontal align="center" className={styles.query} gap={8}>
+          <Icon icon={SearchIcon} />
+          {query ? query : <Skeleton.Block active style={{ height: 20, width: 40 }} />}
+        </Flexbox>
+        <Skeleton.Block active style={{ height: 20, width: 40 }} />
+      </Flexbox>
+      <Flexbox horizontal gap={12}>
+        {[1, 2, 3, 4, 5].map((id) => (
+          <Skeleton.Button active key={id} style={{ borderRadius: 8, height: 80, width: 160 }} />
+        ))}
+      </Flexbox>
+    </Flexbox>
+  );
+});
+```
+
+## Placeholder rules
+
+- **Mirror the eventual Render's layout.** When the result arrives the Placeholder unmounts and the Render mounts; if they share dimensions, the chat doesn't jump.
+- Use `Skeleton.Block` / `Skeleton.Button` from `@lobehub/ui` for placeholder shapes.
+- Embed any args you have (e.g. the query text) — context helps the user know what's loading.
+- Pulse with `shinyTextStyles.shinyText` if the Placeholder includes literal text.
+
+## Placeholder registry — `client/Placeholder/index.ts`
+
+```ts
+import { WebBrowsingApiName } from '../../types';
+import CrawlMultiPages from './CrawlMultiPages';
+import CrawlSinglePage from './CrawlSinglePage';
+import { Search } from './Search';
+
+export const WebBrowsingPlaceholders = {
+  [WebBrowsingApiName.crawlMultiPages]: CrawlMultiPages,
+  [WebBrowsingApiName.crawlSinglePage]: CrawlSinglePage,
+  [WebBrowsingApiName.search]: Search,
+};
+
+export { CrawlMultiPages, CrawlSinglePage, Search };
+```
@@ -0,0 +1,71 @@
+# Portal — Full-Screen Detail View (optional)
+
+**Lifecycle:** rendered when the user opens the tool message in a side panel or full-screen modal. One Portal per **tool**, not per API — the Portal switches on `apiName` internally.
+
+**Add for** tools whose results deserve a deep-dive view: search results with editable filters, page content with reader mode, code interpreter sessions.
+
+## Props (`BuiltinPortalProps<Args, State>`)
+
+```ts
+interface BuiltinPortalProps<Arguments = Record<string, any>, State = any> {
+  apiName?: string;
+  arguments: Arguments;
+  identifier: string;
+  messageId: string;
+  state: State;
+}
+```
+
+## Canonical example — Web-Browsing Portal
+
+`packages/builtin-tool-web-browsing/src/client/Portal/index.tsx`:
+
+```tsx
+import type { BuiltinPortalProps, CrawlPluginState, SearchQuery } from '@lobechat/types';
+import { memo } from 'react';
+
+import { WebBrowsingApiName } from '../../types';
+import PageContent from './PageContent';
+import PageContents from './PageContents';
+import Search from './Search';
+
+const Portal = memo<BuiltinPortalProps>(({ arguments: args, messageId, state, apiName }) => {
+  switch (apiName) {
+    case WebBrowsingApiName.search:
+      return <Search messageId={messageId} query={args as SearchQuery} response={state} />;
+
+    case WebBrowsingApiName.crawlSinglePage: {
+      const result = (state as CrawlPluginState).results.find((r) => r.originalUrl === args.url);
+      return <PageContent messageId={messageId} result={result} />;
+    }
+
+    case WebBrowsingApiName.crawlMultiPages:
+      return (
+        <PageContents
+          messageId={messageId}
+          results={(state as CrawlPluginState).results}
+          urls={args.urls}
+        />
+      );
+  }
+  return null;
+});
+export default Portal;
+```
+
+## Portal rules
+
+- One Portal per tool — the file is the routing layer, subcomponents implement each API's view.
+- Portals can read the chat store directly to detect "still streaming" and render a Skeleton internally (see `Search/index.tsx:20-46`).
+- Layout assumes more space than the Render — use `Flexbox` with `height={'100%'}` and structure for a side panel viewport.
+
+## Portal registry — `packages/builtin-tools/src/portals.ts`
+
+```ts
+import { WebBrowsingManifest, WebBrowsingPortal } from '@lobechat/builtin-tool-web-browsing/client';
+import { type BuiltinPortal } from '@lobechat/types';
+
+export const BuiltinToolsPortals: Record<string, BuiltinPortal> = {
+  [WebBrowsingManifest.identifier]: WebBrowsingPortal as BuiltinPortal,
+};
+```
@@ -0,0 +1,19 @@
+# Tool Render 设计原则（中文草案）
+
+这些原则用于判断一个 builtin tool 的 Inspector / Render / Placeholder / Streaming / Intervention / Portal 应该做什么，以及做到什么程度。
+
+1. **先保证折叠态可读。** 每个 API 都必须有 Inspector；用户不展开也应该能看懂 “正在做什么 / 对什么做 / 当前结果是什么”。Inspector 不应该只展示函数名和原始参数。
+2. **Inspector 是一句话，不是详情页。** 优先表达动作、关键对象、数量、状态，例如 “分析图片 3 张”“搜索 12 个结果”“读取 config.json”。长文本、列表和结构化结果放到 Render 或 Portal。
+3. **Inspector 要覆盖执行生命周期。** `args` 还在 streaming、工具执行中、执行完成、执行失败时都应该有稳定展示；必要时同时读取 `args`、`partialArgs` 和 `pluginState`，避免出现空白、跳变或只显示半截参数。
+4. **文案要随状态切换时态。** 同一个动作在 loading 与 completed 两个阶段必须用不同的措辞：执行中用现在进行时（“正在创建任务 / Creating task / 正在搜索”），执行完成后切到完成态（“已创建任务 / Task created / 已找到 N 条”）。Inspector chip 会一直留在聊天记录里 —— 如果一直挂着 “正在 xxx”，几小时后回看历史时会读起来像还在跑。约定的 i18n 形式是 `<api>.loading` / `<api>.completed` 一对键（见 `lobe-agent.apiName.callSubAgent.{loading,completed}` 与 `lobe-claude-code.task.{create,list,update,get}.{loading,completed}`），渲染时按 `isArgumentsStreaming || isLoading` 决定取哪一个。只读 / 查询类（“查看任务” 这种本来就是名词性的）可以共用一个键。
+5. **只有结构化结果才需要 Render。** 如果工具结果只是自然语言总结，通常不需要 Render；如果结果包含列表、媒体、文件、表格、代码、diff、地图、时间线、权限请求等结构，就应该提供 Render。
+6. **Render 要帮助用户检查结果，而不是复述参数。** Render 的主体应该围绕工具产物组织：可预览、可比较、可筛选、可定位。参数只作为上下文辅助出现，不要把 Render 做成一块更大的 args dump。
+7. **参数和结果要一起参与渲染。** 好的 Tool UI 通常同时用 `args` 解释意图，用 `pluginState` 展示真实执行结果；但 `pluginState` 只放结果域数据，不要反向塞入可以从 `args` 推导出的内容。
+8. **慢操作要有 Placeholder。** 如果工具通常需要等待网络、文件系统、模型或外部进程，Placeholder 应该先占住最终 Render 的版式，让用户知道即将看到什么，而不是只显示一个泛化 loading。
+9. **Streaming 只用于连续产物。** 搜索列表、日志、长文本、文件分析、分阶段计划适合 Streaming；一次性小结果不需要强行做 Streaming。Streaming UI 要能渐进追加，并且完成后自然过渡到最终 Render。
+10. **有风险的动作必须 Intervention。** 写文件、删除、发送、安装、执行命令、外部可见操作、权限敏感操作，都应该在执行前给出可理解的确认界面；确认文案要说明影响范围，而不是只问 “是否继续”。
+11. **错误、空态和截断都是正式状态。** Render 不能在失败、无结果、超长结果时退化成空白。错误要说明发生在哪一步；空态要告诉用户没有产物；超长内容要明确 “展示前 N 项 / 还有 N 项”。
+12. **信息密度要克制。** 默认展示最有判断价值的部分：标题、来源、状态、摘要、少量关键字段。大对象、长列表、原文、调试数据放进可展开区域或 Portal，避免把聊天流撑成后台管理页。
+13. **视觉上融入聊天流。** Tool UI 应该使用 `@lobehub/ui` / base-ui、`Flexbox`、`createStaticStyles` 和 `cssVar.*`，遵循现有间距、圆角、颜色、字号；不要为单个工具发明一套独立视觉语言。具体的样式约定见 [shared-rules.md](shared-rules.md)。
+14. **Devtools fixture 是验收入口。** 新增或修改 Tool UI 时，应在 `/devtools` 里准备覆盖典型态、loading/streaming、空态、错误态、长内容态的 fixture；一个 API 如果在真实聊天里会出现，就不应该在 devtools 中缺席。
+15. **先做用户会看的 UI，再做调试 UI。** Raw JSON、trace、schema、内部 id 可以存在，但应默认收起或放到调试区；主界面先回答用户最关心的问题：工具做了什么，结果值不值得信任，下一步能做什么。
@@ -0,0 +1,101 @@
+# Render — Rich Result Card (optional)
+
+**Lifecycle:** rendered **once the result arrives** (after Placeholder/Streaming hand off). Sits below the Inspector header.
+
+**Skip if** the API is read-only or the result is just text — the framework already shows the executor's `content` string. Add a Render only when there's a structured artifact worth seeing: a card, a chart, a diff, a list of files.
+
+## Props (`BuiltinRenderProps<Args, State, Content>`)
+
+```ts
+interface BuiltinRenderProps<Arguments = any, State = any, Content = any> {
+  apiName?: string;
+  args: Arguments; // final params from the LLM
+  content: Content; // executor's content string (or parsed)
+  identifier?: string;
+  messageId: string; // for store lookups
+  pluginError?: any; // from BuiltinToolResult.error
+  pluginState?: State; // executor's state
+  toolCallId?: string;
+}
+```
+
+## Two patterns
+
+**Pattern A — Single-file Render** (web-browsing CrawlSinglePage):
+
+```tsx
+// client/Render/CrawlSinglePage.tsx
+import type { BuiltinRenderProps, CrawlPluginState, CrawlSinglePageQuery } from '@lobechat/types';
+import { memo } from 'react';
+
+import PageContent from './PageContent';
+
+const CrawlSinglePage = memo<BuiltinRenderProps<CrawlSinglePageQuery, CrawlPluginState>>(
+  ({ messageId, pluginState, args }) => (
+    <PageContent messageId={messageId} results={pluginState?.results} urls={[args?.url]} />
+  ),
+);
+export default CrawlSinglePage;
+```
+
+**Pattern B — Folder with subcomponents** (web-browsing Search):
+
+```
+client/Render/Search/
+├── index.tsx           # composes the subcomponents, handles error states
+├── ConfigForm.tsx      # appears when pluginError.type === 'PluginSettingsInvalid'
+├── SearchQuery.tsx     # editable query header
+└── SearchResult.tsx    # result list
+```
+
+Use Pattern B when the Render has internal state (editing mode, expanded items), error variants, or is large enough to benefit from splitting.
+
+## Error handling in Render
+
+Renders are the canonical place to surface `pluginError` because the chat doesn't auto-render typed errors:
+
+```tsx
+if (pluginError) {
+  if (pluginError?.type === 'PluginSettingsInvalid') {
+    return <ConfigForm id={messageId} provider={pluginError.body?.provider} />;
+  }
+  return (
+    <Alert
+      title={pluginError?.message}
+      type="error"
+      extra={<Highlighter language="json">{JSON.stringify(pluginError.body, null, 2)}</Highlighter>}
+    />
+  );
+}
+```
+
+## Render rules
+
+- **Return `null`** if there's nothing useful to draw yet (avoids empty cards during stream).
+- Use `pluginState` for server-truth (ids, counts, server-assigned status) and `args` for what the LLM asked. **Combine — neither alone is enough.**
+- For lists, summarize with a header line and show top N items with a "+N more" tail rather than rendering everything.
+- **Keep the Render single-layer** — the tool card is already your surface, so don't open with your own filled container and then nest more filled boxes inside it. See [shared-rules.md](shared-rules.md) → "Stay single-layer".
+- For modals from a Render, use `@lobehub/ui/base-ui` (`createModal`, `useModalContext`, `confirmModal`) — see the **modal** skill.
+
+## Render registry — `client/Render/index.ts`
+
+```ts
+import type { BuiltinRender } from '@lobechat/types';
+
+import { TaskApiName } from '../../types';
+import CreateTaskRender from './CreateTask';
+import RunTasksRender from './RunTasks';
+
+export const TaskRenders: Record<string, BuiltinRender> = {
+  [TaskApiName.createTask]: CreateTaskRender as BuiltinRender,
+  [TaskApiName.runTasks]: RunTasksRender as BuiltinRender,
+  /* only the APIs with rich result UI — others fall back to text content */
+};
+
+export { default as CreateTaskRender } from './CreateTask';
+export { default as RunTasksRender } from './RunTasks';
+```
+
+## Render display control (rare)
+
+If the Render should hide for certain results (e.g. ClaudeCode's TodoWrite hides when the agent is mid-stream), add a `RenderDisplayControl` to `packages/builtin-tools/src/displayControls.ts`. See `ClaudeCodeRenderDisplayControls` for the pattern.
@@ -0,0 +1,89 @@
+# Shared Style Rules
+
+These apply across every surface.
+
+## The component skeleton
+
+Every surface file is the same shape, so internalize it once instead of re-deriving it per rule. The skeleton below bakes in five mechanical conventions — copy it and fill the body:
+
+```tsx
+'use client'; // (a) leaves of the chat tree must not block server rendering
+
+import type { BuiltinInspectorProps, SearchQuery, UniformSearchResponse } from '@lobechat/types';
+import { memo } from 'react';
+import { useTranslation } from 'react-i18next';
+
+// (b) type with BuiltinXProps<Args, State> — never widen to `any`.
+//     Args = the JSON Schema params, State = the executor's `state` field;
+//     they should match <Name>Params / <Name>State from types.ts.
+export const SearchInspector = memo<BuiltinInspectorProps<SearchQuery, UniformSearchResponse>>(
+  ({ args, pluginState }) => {
+    const { t } = useTranslation('plugin'); // (c) all strings from the `plugin` namespace
+
+    // (d) cross-cutting state (loading, streaming buffer) comes from the store,
+    //     not props — props only carry args/state/messageId.
+    // const buffer = useChatStore((s) => chatToolSelectors.streamingBuffer(messageId)(s));
+
+    return <span>{t('builtins.<identifier>.apiName.search')}</span>;
+  },
+);
+SearchInspector.displayName = 'SearchInspector'; // (e) always memo + displayName
+export default SearchInspector;
+```
+
+- **(c)** Default an Inspector to `t('builtins.<identifier>.apiName.<api>')` so the row is non-empty while args stream in.
+- **(d)** Read the store via Zustand selectors inside the component; see [streaming.md](streaming.md) for the buffer selector.
+
+## Styling: `createStaticStyles + cssVar.*`, `@lobehub/ui` over `antd`
+
+Zero-runtime CSS-in-JS — styles compile once and read CSS variables at runtime:
+
+```tsx
+import { createStaticStyles, cssVar } from 'antd-style';
+
+const styles = createStaticStyles(({ css, cssVar }) => ({
+  chip: css`
+    padding-block: 2px;
+    padding-inline: 8px;
+    border-radius: 999px;
+    color: ${cssVar.colorText};
+    background: ${cssVar.colorFillTertiary};
+  `,
+}));
+```
+
+- Fall back to `createStyles + token` only when you need runtime token computation (rare). Inline `style={{ color: cssVar.colorTextSecondary }}` is fine for one-off dynamic values.
+- Components come from `@lobehub/ui` (`Block`, `Text`, `Flexbox`, `Highlighter`, `Alert`, `Tooltip`, `Skeleton`), not raw `antd`. Modals come from `@lobehub/ui/base-ui` (`createModal`, `useModalContext`, `confirmModal`) — see the **modal** skill.
+- Note: `<Text type='secondary'>` is a lighter shade than `colorTextSecondary`. For that exact token color, write `<Text style={{ color: cssVar.colorTextSecondary }}>`.
+
+## Stay single-layer — don't nest filled cards
+
+The framework already wraps every Render / Intervention in a tool card, so that card **is** your surface. A Render that opens with its own `background: ${cssVar.colorFillQuaternary}` container is already one card deep; put another filled box inside it (`colorBgContainer` / `colorFillTertiary`) and you get the card-in-card look that reads as "complex" — two or three stacked fills for what is really a flat list of fields.
+
+- **The outermost wrapper carries no fill.** Use a flat container with only `padding-block: 4px` for breathing room; let the tool card provide the card. (See `Agent/index.tsx`'s `container`.)
+- **At most one filled box, and only to delineate real content** — a Markdown preview, a diff, a code/result block. Labels, key–value fields, question/answer text, chips: render flat on the surface, separated by spacing or a hairline divider (`height: 1px; background: ${cssVar.colorFillSecondary}`), not by wrapping each in its own box.
+- **A box on a flat surface needs a visible fill.** Once the outer fill is gone, an inner `colorBgContainer` box can vanish against the tool card (same color). Use `colorFillTertiary` for the one content box so it still reads as delineated.
+- Don't wrap a single value in a box just to give it padding — that's the redundant-nesting smell (a `detailCard` around a `value` box around one string).
+
+```tsx
+// ❌ card-in-card: filled container wrapping a filled preview box
+container: css`
+  padding: 12px;
+  background: ${cssVar.colorFillQuaternary};
+`,
+previewBox: css`
+  background: ${cssVar.colorBgContainer};
+`,
+
+// ✅ single-layer: flat container, one visible content box
+container: css`
+  padding-block: 4px;
+`,
+previewBox: css`
+  background: ${cssVar.colorFillTertiary};
+`,
+```
+
+For the common "icon + file/title header, then one content box" shape, reuse `ToolResultCard` from `@lobechat/shared-tool-ui/components` instead of rebuilding it — it's already single-layer (flat wrapper, one `colorFillTertiary` content box) and is what CC `Read` / `Grep` / `Glob` / `Write` / `WebSearch` / `WebFetch` render through.
+
+The exception is a deliberate **panel** pattern — an `<Block variant="outlined">` with a header bar + list rows (CC `TodoWrite` / `Task`). There the single outlined block is the panel and the header fill is a header bar, not a nested card. One structured panel is fine; stacked decorative fills are not.
@@ -0,0 +1,83 @@
+# Streaming — Live Output During Execution (optional)
+
+**Lifecycle:** rendered **while the executor is still running** for APIs that emit incremental output. The component is responsible for fetching the in-flight stream from the chat store and rendering it.
+
+**Add for** long-running ops with continuous output: shell command execution (stdout/stderr), file write progress, code interpreter cells.
+
+## Props (`BuiltinStreamingProps<Args>`)
+
+```ts
+interface BuiltinStreamingProps<Arguments = any> {
+  apiName: string;
+  args: Arguments;
+  identifier: string;
+  messageId: string; // use to fetch the streaming buffer from store
+  toolCallId: string;
+}
+```
+
+Note there's **no `state` or `result` prop** — the Streaming component is for the in-flight phase. It pulls the live buffer from the store itself (typically via `chatToolSelectors.streamingContent(messageId)` or similar).
+
+## Canonical example — RunCommandStreaming
+
+`packages/builtin-tool-local-system/src/client/Streaming/RunCommand/index.tsx`:
+
+```tsx
+'use client';
+
+import type { BuiltinStreamingProps } from '@lobechat/types';
+import { Highlighter } from '@lobehub/ui';
+import { memo } from 'react';
+
+interface RunCommandParams {
+  command?: string;
+  description?: string;
+  timeout?: number;
+}
+
+export const RunCommandStreaming = memo<BuiltinStreamingProps<RunCommandParams>>(({ args }) => {
+  const { command } = args || {};
+  if (!command) return null;
+
+  return (
+    <Highlighter
+      animated
+      wrap
+      language="sh"
+      showLanguage={false}
+      style={{ padding: '4px 8px' }}
+      variant="outlined"
+    >
+      {command}
+    </Highlighter>
+  );
+});
+RunCommandStreaming.displayName = 'RunCommandStreaming';
+```
+
+For real-time output beyond just the command (stderr/stdout streaming), pull from the chat store:
+
+```tsx
+const buffer = useChatStore((state) =>
+  chatToolSelectors.streamingBuffer(messageId, toolCallId)(state),
+);
+```
+
+## Streaming rules
+
+- Render `null` until you have something to display (avoids flash).
+- For terminal-style output, use `Highlighter` with `animated` to show typing-like effect.
+- The Streaming component must **unmount cleanly** when execution ends — typically the framework swaps it out for the Render automatically.
+
+## Streaming registry — `client/Streaming/index.ts`
+
+```ts
+import { LocalSystemApiName } from '../..';
+import { RunCommandStreaming } from './RunCommand';
+import { WriteFileStreaming } from './WriteFile';
+
+export const LocalSystemStreamings = {
+  [LocalSystemApiName.runCommand]: RunCommandStreaming,
+  [LocalSystemApiName.writeLocalFile]: WriteFileStreaming,
+};
+```
@@ -1,13 +1,6 @@
 ---
 name: chat-sdk
-description: >
-  Build multi-platform chat bots with Chat SDK (`chat` npm package). Use when developers want to
-  (1) Build a Slack, Teams, Google Chat, Discord, GitHub, or Linear bot,
-  (2) Use the Chat SDK to handle mentions, messages, reactions, slash commands, cards, modals, or streaming,
-  (3) Set up webhook handlers for chat platforms,
-  (4) Send interactive cards or stream AI responses to chat platforms.
-  Triggers on "chat sdk", "chat bot", "slack bot", "teams bot", "discord bot", "@chat-adapter",
-  building bots that work across multiple chat platforms.
+description: "Build multi-platform chat bots with the Chat SDK (`chat` npm package) — Slack, Teams, Google Chat, Discord, GitHub, Linear. Use when building a chat bot, handling mentions / messages / reactions / slash commands / cards / modals / streaming, setting up a webhook handler, or sending interactive cards / streaming AI responses to a chat platform. Triggers on `@chat-adapter`, 'chat sdk', 'chat bot', 'slack bot', 'teams bot', 'discord bot', 'webhook handler', 'cross-platform bot'."
 user-invocable: false
 ---

@@ -1,6 +1,6 @@
 ---
-name: data-fetching
-description: Data fetching architecture guide using Service layer + Zustand Store + SWR. Use when implementing data fetching, creating services, working with store hooks, or migrating from useEffect. Triggers on data loading, API calls, service creation, or store data fetching tasks.
+name: data-fetching-architecture
+description: Standardized data-fetching pipeline guide — Service layer + Zustand Store + SWR. Use when implementing a data-fetching feature, creating a `xxxService`, adding a `useFetchXxx` hook, wiring `useClientDataSWR`, or migrating ad-hoc `useEffect + fetch` to the standard pipeline. Triggers on `lambdaClient`, `useClientDataSWR`, `xxxService`, `useFetchXxx`, 'data fetching', 'fetch architecture', 'service layer', 'SWR hook', 'migrate useEffect'.
 user-invocable: false
 ---

@@ -1,6 +1,6 @@
 ---
 name: docs-changelog
-description: 'Writing guide for website changelog pages under docs/changelog/*.mdx. Use when creating or editing product update posts in EN/ZH. Not for GitHub Release notes.'
+description: "Writing guide for website changelog pages under `docs/changelog/*.mdx` (NOT GitHub Release notes — those live in the `version-release` skill). Use when creating or editing a product update post in EN/ZH. Triggers on `docs/changelog/*.mdx`, 'changelog post', 'product update post', 'add a changelog', '更新日志', 'changelog 文案'."
 ---

 # Docs Changelog Writing Guide
@@ -397,35 +397,60 @@ The pattern is the same for every platform:

 Pick the file for your target platform — each contains activation, navigation, send-message, and verification snippets specific to that app:

-| Platform      | Reference                                          | Quick switcher |
-| ------------- | -------------------------------------------------- | -------------- |
-| Discord       | [references/discord.md](./references/discord.md)   | `Cmd+K`        |
-| Slack         | [references/slack.md](./references/slack.md)       | `Cmd+K`        |
-| Telegram      | [references/telegram.md](./references/telegram.md) | `Cmd+F`        |
-| WeChat / 微信 | [references/wechat.md](./references/wechat.md)     | `Cmd+F`        |
-| Lark / 飞书   | [references/lark.md](./references/lark.md)         | `Cmd+K`        |
-| QQ            | [references/qq.md](./references/qq.md)             | `Cmd+F`        |
+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:

-For **shared osascript patterns** (activate, type, paste, screenshot, read accessibility, common workflow template, gotchas), see [references/osascript-common.md](./references/osascript-common.md). Read this first if you're new to osascript automation.
+| 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

-Ready-to-use scripts in `.agents/skills/local-testing/scripts/`:
+**App / recording scripts** in `.agents/skills/local-testing/scripts/`:

 | Script                    | Usage                                               |
 | ------------------------- | --------------------------------------------------- |
 | `electron-dev.sh`         | Manage Electron dev env (start/stop/status/restart) |
-| `capture-app-window.sh`   | Capture screenshot of a specific app window         |
 | `record-electron-demo.sh` | Record Electron app demo with ffmpeg                |
 | `record-app-screen.sh`    | Record app screen (video + screenshots, start/stop) |
-| `test-discord-bot.sh`     | Send message to Discord bot via osascript           |
-| `test-slack-bot.sh`       | Send message to Slack bot via osascript             |
-| `test-telegram-bot.sh`    | Send message to Telegram bot via osascript          |
-| `test-wechat-bot.sh`      | Send message to WeChat bot via osascript            |
-| `test-lark-bot.sh`        | Send message to Lark / 飞书 bot via osascript       |
-| `test-qq-bot.sh`          | Send message to QQ bot via osascript                |
+
+**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

@@ -433,9 +458,9 @@ Ready-to-use scripts in `.agents/skills/local-testing/scripts/`:

 ```bash
 # Standalone usage
-./.agents/skills/local-testing/scripts/capture-app-window.sh "Discord" /tmp/discord.png
-./.agents/skills/local-testing/scripts/capture-app-window.sh "Slack" /tmp/slack.png
-./.agents/skills/local-testing/scripts/capture-app-window.sh "WeChat" /tmp/wechat.png
+./.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.
@@ -452,32 +477,48 @@ Examples:

 ```bash
 # Discord — test a bot in #bot-testing channel
-./.agents/skills/local-testing/scripts/test-discord-bot.sh "bot-testing" "!ping"
-./.agents/skills/local-testing/scripts/test-discord-bot.sh "bot-testing" "/ask Tell me a joke" 30
+./.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/scripts/test-slack-bot.sh "bot-testing" "@mybot hello"
-./.agents/skills/local-testing/scripts/test-slack-bot.sh "bot-testing" "/ask What is 2+2?" 20
+./.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/scripts/test-telegram-bot.sh "MyTestBot" "/start"
-./.agents/skills/local-testing/scripts/test-telegram-bot.sh "GPTBot" "Hello" 60
+./.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/scripts/test-wechat-bot.sh "文件传输助手" "test message" 5
-./.agents/skills/local-testing/scripts/test-wechat-bot.sh "MyBot" "Tell me a joke" 30
+./.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/scripts/test-lark-bot.sh "bot-testing" "@MyBot hello"
-./.agents/skills/local-testing/scripts/test-lark-bot.sh "bot-testing" "Help me with this" 30
+./.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/scripts/test-qq-bot.sh "bot-testing" "Hello bot" 15
-./.agents/skills/local-testing/scripts/test-qq-bot.sh "MyBot" "/help" 10
+./.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
@@ -517,4 +558,4 @@ Outputs to `.records/` directory (gitignored): `<name>.mp4` (video) + `<name>/`

 ### osascript

-See [references/osascript-common.md](./references/osascript-common.md#gotchas) for the full osascript gotchas list (accessibility permissions, `keystroke` non-ASCII issues, locale-specific app names, rate limiting, etc.).
+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.).
@@ -2,7 +2,7 @@

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

-See [osascript-common.md](./osascript-common.md) for shared patterns.
+See [osascript-common.md](../osascript-common.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/scripts/test-discord-bot.sh "bot-testing" "!ping"
-./.agents/skills/local-testing/scripts/test-discord-bot.sh "bot-testing" "/ask Tell me a joke" 30
+./.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
 ```
@@ -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"
+"$SCRIPT_DIR/../capture-app-window.sh" "$APP" "$SCREENSHOT"
 echo "[$APP] Done! Screenshot saved to $SCREENSHOT"
@@ -0,0 +1,232 @@
+# iMessage Desktop bridge regression test
+
+The iMessage channel is different from the other bot platforms: there is **no
+native app to drive with osascript**. Instead the Desktop app runs a local
+**BlueBubbles bridge** — a small HTTP server in the Electron main process that
+registers a webhook on a local [BlueBubbles](https://bluebubbles.app/) server,
+receives iMessage events, and forwards them to LobeHub Cloud.
+
+So the test surface is three layers:
+
+1. **Electron main IPC** — `imessageBridge.*` handlers (`getStatus`,
+   `testConfig`, `upsertConfig`, `removeConfig`, `start`, `stop`)
+2. **Local bridge HTTP server** — `http://127.0.0.1:<port>/webhooks/bluebubbles/<appId>?secret=<secret>`
+3. **BlueBubbles REST API** — `http://127.0.0.1:1234/api/v1/*` (webhook + server/info)
+
+## Prerequisites
+
+- A running **BlueBubbles server** (macOS, default `http://127.0.0.1:1234`) with
+  a known password. Sanity check:
+  ```bash
+  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`
+- 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/`
+  after switching branches — the new workspace package must be linked or the
+  main build fails to resolve `@lobechat/chat-adapter-imessage`.
+
+## Fast path: automated script
+
+```bash
+./.agents/skills/local-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
+its bridge config + BlueBubbles webhook on exit). Exit 0 = all green. It covers:
+
+- BlueBubbles reachable + password valid; Electron CDP reachable; IPC available
+- `testConfig` happy path → success
+- `testConfig` wrong password → rejected; unreachable URL → rejected
+- `upsertConfig` **first-time save → success** (Bug #1 regression guard, below)
+- `getStatus` → `running:true`, config persisted, password redacted (`blueBubblesPasswordSet`)
+- BlueBubbles webhook actually registered for the appId
+- Local bridge HTTP server: wrong secret → 401; valid secret → past auth
+
+The password is passed as argv (visible in `ps`) — local dev only, don't use a
+real secret on a shared machine.
+
+## Layer 1 — IPC probes (no UI)
+
+The renderer exposes the main-process handlers via `window.electronAPI.invoke`.
+This is the quickest way to exercise the bridge without clicking:
+
+```bash
+# baseline
+agent-browser --cdp 9222 eval \
+  "(async()=>JSON.stringify(await window.electronAPI.invoke('imessageBridge.getStatus',{})))()"
+
+# test a connection (note: password as a JS string)
+agent-browser --cdp 9222 eval --stdin << 'EVALEOF'
+(async function () {
+  try {
+    var r = await window.electronAPI.invoke('imessageBridge.testConfig', {
+      applicationId: 'probe',
+      blueBubblesServerUrl: 'http://127.0.0.1:1234',
+      blueBubblesPassword: 'PASTE_PW',
+      enabled: true,
+      webhookSecret: 'probe-secret',
+    });
+    return JSON.stringify(r);            // { success: true }
+  } catch (e) { return 'ERR: ' + (e.message || e); }
+})()
+EVALEOF
+```
+
+`upsertConfig` persists to the Electron store, starts the local HTTP server, and
+registers the BlueBubbles webhook. `removeConfig` + `stop` reverse it.
+
+## Layer 2 — full UI flow (agent-browser)
+
+The bridge settings only render in Desktop (`isDesktop` guard) under the agent's
+**Channel → iMessage** screen. The platform tile only appears as a real (non
+"Coming Soon") entry once the server registers `imessage` **and** the frontend
+drops it from `COMING_SOON_PLATFORMS` (`src/routes/(main)/agent/channel/const.ts`).
+
+```bash
+agent-browser --cdp 9222 open "http://localhost:5173/agent/<aid>/channel"
+agent-browser --cdp 9222 wait --load networkidle && agent-browser --cdp 9222 wait 1500
+
+# confirm the remote backend lists imessage (it must be registered + deployed)
+agent-browser --cdp 9222 eval --stdin << 'EVALEOF'
+(async function(){
+  var url='lobe-backend://lobe/trpc/lambda/agentBotProvider.listPlatforms?input='+encodeURIComponent('{"json":null,"meta":{"values":["undefined"],"v":1}}');
+  var d=await (await fetch(url,{credentials:'include'})).json();
+  var p=d.result?.data?.json||d;
+  return JSON.stringify(p.map(function(x){return x.id;}));
+})()
+EVALEOF
+
+# click the iMessage tile, then fill the form by ref
+agent-browser --cdp 9222 eval "(()=>{var b=[...document.querySelectorAll('aside button')].find(x=>/imessage/i.test(x.textContent));b&&b.click();})()"
+agent-browser --cdp 9222 wait 1500
+agent-browser --cdp 9222 snapshot -i | grep -iE "127.0.0.1:1234|Application ID|Webhook Secret|Test BlueBubbles|Save Bridge"
+```
+
+Field refs (from the snapshot): Application ID, Webhook Secret, BlueBubbles
+Server URL (`placeholder="http://127.0.0.1:1234"`), and a **nested** textbox right
+under the URL one is the BlueBubbles Password. Fill with `fill` (real input
+events — `eval`-setting React inputs won't fire onChange), click **Test
+BlueBubbles**, then **Save Bridge**. Read the antd toast immediately (it
+auto-dismisses):
+
+```bash
+agent-browser --cdp 9222 eval \
+  "JSON.stringify([...new Set([...document.querySelectorAll('.ant-message-custom-content')].map(n=>n.textContent.trim()))])"
+# Test  → "BlueBubbles connection passed"
+# Save  → "iMessage Desktop bridge saved"
+```
+
+Verify the end state via BlueBubbles + IPC:
+
+```bash
+curl -sS "http://127.0.0.1:1234/api/v1/webhook?password=<PW>" # webhook for the appId present
+agent-browser --cdp 9222 eval "(async()=>JSON.stringify(await window.electronAPI.invoke('imessageBridge.getStatus',{})))()"
+# running:true, serverUrl: http://127.0.0.1:33270, configs[].blueBubblesPasswordSet:true
+```
+
+Cleanup: `removeConfig` + `stop` via IPC, then `DELETE /api/v1/webhook/<id>` on
+BlueBubbles.
+
+## Outbound send test (desktop → BlueBubbles → iMessage)
+
+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
+```
+
+**Gotcha that bites everyone:** with `method=apple-script` and a _new_
+conversation, the HTTP POST often **times out** even though the message is
+sent. Never judge success by the HTTP response. Instead poll
+`POST /api/v1/message/query` and read the matching `isFromMe:true` row's
+`error` field:
+
+- `error: 0` (or null) → sent OK
+- non-zero `error` → real send failure
+
+The script does exactly this: fires the send, ignores the timeout, then matches
+its marker text in the message store and asserts `error == 0`.
+
+Two more notes:
+
+- Use a full E.164 handle (`iMessage;-;+<countrycode><number>`) or an Apple ID
+  email. Looking the chat up by guid afterwards may 404 if BB filed the message
+  under a differently-formatted guid — that's a lookup quirk, not a send failure.
+- Sending to _your own_ number round-trips: BB records both the outgoing
+  (`fromMe:true`) and an incoming copy (`fromMe:false`).
+
+## Inbound e2e test (iMessage → cloud agent → reply)
+
+Full inbound chain: a message arrives → BlueBubbles fires its `new-message`
+webhook → local bridge (`:33270`) → `forwardWebhook` POSTs to
+`<remote>/api/agent/webhooks/imessage/<appId>?secret=…` → cloud agent → reply
+flows back via Device Gateway → BB `sendText`.
+
+Prerequisites:
+
+- A cloud bot provider for the same `applicationId` exists and is **connected**
+  (Save Configuration + the device gateway connected — a _disconnected_ gateway
+  yields `DEVICE_NOT_FOUND` on connect and blocks the reply leg).
+- The `imessage` Labs toggle is on (otherwise the channel is gated to "Coming
+  Soon"), and `webhookSecret` matches on both ends (auto-generated on save).
+
+Two ways to drive it:
+
+1. **Second device / Apple ID (recommended).** Have _another_ Apple ID message
+   the BB-hosted number (e.g. "please reply pong"). The bot replies; you see it
+   on the other device. **No loop risk** — the reply goes to the other party,
+   not back to itself.
+2. **Send to your own number (quick, loop-aware).** `sendText` to the hosted
+   number; the loopback _incoming_ copy (`isFromMe:false`) triggers the bot.
+   Watch the reply land in `message/query` as a `fromMe:true` row.
+
+**Loop guard — why a self-send doesn't spin forever:** the Chat SDK adapter
+drops any `isFromMe` message before dispatch
+(`packages/chat-adapter-imessage/src/adapter.ts`: `if (message.isFromMe) return`).
+The bot's own reply (`isFromMe:true`) is never re-processed, so in the normal
+case (someone else → bot → reply to them) there is no loop. The self-send case
+is a **test-only edge**: the bot's reply also round-trips to your number, and
+only the adapter's `isFromMe` check stops a second pass. Keep the prompt
+conversational (so the bot doesn't keep finding something to answer), and
+**turn the `imessage` lab off / remove the config when done** — never leave a
+self-send bot running unattended.
+
+Watch the chain live:
+
+```bash
+tail -f /tmp/electron-dev.log | grep -iE "imessage|bridge|forward|Message API"
+# the agent reply shows up as a fromMe:true row with the bot's text:
+curl -sS -X POST "http://127.0.0.1:1234/api/v1/message/query?password=<PW>" \
+  -H 'Content-Type: application/json' -d '{"limit":5,"sort":"DESC"}'
+```
+
+`startTyping` will log a Private-API error unless BlueBubbles has the Private
+API helper set up (needs a jailbroken / SIP-disabled Mac) — it's logged and
+ignored; text replies still work.
+
+## Known bugs / gotchas
+
+- **Bug #1 — first-time save (fixed; guarded by the script).** BlueBubbles'
+  `GET /api/v1/webhook?url=<unregistered>` returns **HTTP 500**
+  (`Cannot read properties of null (reading 'events')`). The bridge must list
+  **all** webhooks and match client-side, never pass the `?url=` filter. If you
+  see `upsertConfig` fail with "An unhandled error has occurred!" originating in
+  `listWebhooks`, this regressed.
+- **Save leaves a half-state on webhook failure.** `upsertConfig` writes the
+  config + starts the HTTP server _before_ registering the webhook, so a webhook
+  failure still reports `running:true` with the config persisted but no
+  BlueBubbles webhook. Always assert the BlueBubbles webhook list, not just IPC
+  status.
+- **Unknown appId / forward failure → 500.** Posting to the local bridge for an
+  unknown appId, or when no cloud bot is bound, returns 500 (BlueBubbles retries
+  on 5xx). Auth (wrong secret → 401) is enforced before that.
+- **Backend deploy lag.** Desktop dev proxies tRPC through `lobe-backend://` to
+  the _remote_ server. iMessage only appears in `listPlatforms` once the server
+  registration is deployed there, regardless of local branch.
+- **Restart to load main-process fixes.** Editing `imessageBridgeSrv.ts` /
+  `@lobechat/chat-adapter-imessage` needs `electron-dev.sh restart` — main isn't
+  hot-replaced. On restart, enabled configs auto-register their webhook again.
@@ -0,0 +1,81 @@
+#!/usr/bin/env bash
+#
+# send-imessage-test.sh — Verify the outbound leg: desktop → BlueBubbles → iMessage
+#
+# Sends one real iMessage via the same REST call the Desktop bridge uses
+# (`POST /api/v1/message/text`, which BlueBubblesApiClient.sendText wraps) and
+# confirms it actually went out.
+#
+# KEY GOTCHA: with method=apple-script and a NEW conversation, the HTTP request
+# often TIMES OUT even though the message is sent. Do NOT treat the timeout as a
+# failure — instead poll `POST /api/v1/message/query` and check the message's
+# `error` field (0 = sent OK). This script does that for you.
+#
+# This sends a REAL message, so it has side effects. Target your own number.
+#
+# Usage:
+#   ./send-imessage-test.sh <bb_password> <target_e164> [message] [bb_url]
+#
+# Example (send to your own phone, E.164 with country code):
+#   ./send-imessage-test.sh 'my-bb-pass' '+15551234567'
+#
+set -euo pipefail
+
+BB_PASS="${1:?Usage: $0 <bb_password> <target_e164(+countrycode)> [message] [bb_url]}"
+TARGET="${2:?Need a target handle in E.164, e.g. +15551234567 (or an Apple ID email)}"
+MARKER="lobe-imsg-test-$(date +%s)"
+MESSAGE="${3:-[${MARKER}] desktop bridge → BlueBubbles → iMessage outbound check}"
+BB_URL="${4:-http://127.0.0.1:1234}"
+
+CHAT_GUID="iMessage;-;${TARGET}"
+
+echo "[send-test] target=${TARGET}  marker=${MARKER}"
+
+# 1) Fire the send. apple-script on a new chat may hang the HTTP response, so we
+#    cap it short and ignore a timeout — step 2 is the source of truth.
+python3 - "$BB_PASS" "$BB_URL" "$CHAT_GUID" "$MESSAGE" <<'PY' || true
+import json,sys,urllib.request,urllib.parse,uuid
+pw,base,guid,msg=sys.argv[1:5]
+url=base+"/api/v1/message/text?password="+urllib.parse.quote(pw)
+body={"chatGuid":guid,"message":msg,"method":"apple-script","tempGuid":str(uuid.uuid4())}
+req=urllib.request.Request(url,data=json.dumps(body).encode("utf-8"),
+    headers={"Content-Type":"application/json"},method="POST")
+try:
+    r=urllib.request.urlopen(req,timeout=8)
+    print("[send-test] HTTP",r.status,"(immediate response)")
+except urllib.error.HTTPError as e:
+    print("[send-test] HTTP",e.code,e.read().decode()[:200])
+except Exception as e:
+    print("[send-test] HTTP request returned no body (likely apple-script delay):",type(e).__name__)
+PY
+
+# 2) Source of truth: find our marker in the message store and read its error.
+echo "[send-test] verifying via message/query (the HTTP timeout above is expected)…"
+sleep 3
+python3 - "$BB_PASS" "$BB_URL" "$MARKER" <<'PY'
+import json,sys,time,urllib.request,urllib.parse
+pw,base,marker=sys.argv[1:4]
+url=base+"/api/v1/message/query?password="+urllib.parse.quote(pw)
+def query():
+    body={"limit":15,"offset":0,"with":["chats"],"sort":"DESC"}
+    req=urllib.request.Request(url,data=json.dumps(body).encode(),
+        headers={"Content-Type":"application/json"},method="POST")
+    return json.load(urllib.request.urlopen(req,timeout=12)).get("data") or []
+hit=None
+for _ in range(5):
+    for m in query():
+        if marker in (m.get("text") or "") and m.get("isFromMe"):
+            hit=m; break
+    if hit: break
+    time.sleep(2)
+if not hit:
+    print("[send-test] ✗ outbound message not found in BB store — send likely failed")
+    sys.exit(1)
+err=hit.get("error")
+if err in (0,None):
+    print("[send-test] ✓ outbound message sent (fromMe=True, error=%s)"%err)
+    print("[send-test]   → confirm it arrived in the Messages app on the target device")
+else:
+    print("[send-test] ✗ BlueBubbles reported send error=%s"%err)
+    sys.exit(1)
+PY
@@ -0,0 +1,187 @@
+#!/usr/bin/env bash
+#
+# test-imessage-bridge.sh — Regression test for the iMessage Desktop bridge
+#
+# Drives the Electron main-process `imessageBridge.*` IPC handlers plus the
+# local bridge HTTP server and the BlueBubbles server, asserting the full
+# connect/configure flow. Use it to regression-test PR work on the iMessage
+# channel (BlueBubbles bridge) without clicking through the UI every time.
+#
+# Prerequisites:
+#   1. BlueBubbles server running and reachable (default http://127.0.0.1:1234)
+#   2. Electron dev running with CDP — `electron-dev.sh start`
+#   3. `agent-browser` on PATH, connected to the same CDP port
+#
+# Usage:
+#   ./test-imessage-bridge.sh <bluebubbles_password> [bb_url] [cdp_port]
+#
+# Example:
+#   ./test-imessage-bridge.sh 'my-bb-password'
+#   ./test-imessage-bridge.sh 'my-bb-password' http://127.0.0.1:1234 9222
+#
+# Notes:
+#   - The password is passed as an argv, so it is visible in `ps`. This is a
+#     local dev tool; do not run it on shared machines with a real secret.
+#   - It uses a unique applicationId per run (imsg-regression-$$) and cleans up
+#     its own bridge config + BlueBubbles webhook on exit, so it is safe to
+#     re-run and does not disturb real configs.
+set -euo pipefail
+
+BB_PASS="${1:?Usage: $0 <bluebubbles_password> [bb_url] [cdp_port]}"
+BB_URL="${2:-http://127.0.0.1:1234}"
+CDP_PORT="${3:-9222}"
+
+APP_ID="imsg-regression-$$"
+SECRET="regression-secret-$$"
+
+PASS=0
+FAIL=0
+
+# ── Output helpers ───────────────────────────────────────────────────
+ok()   { echo "  ✓ $1"; PASS=$((PASS + 1)); }
+bad()  { echo "  ✗ $1 — $2"; FAIL=$((FAIL + 1)); }
+note() { echo "[imsg-test] $1"; }
+
+# ── BlueBubbles REST helpers ─────────────────────────────────────────
+bb_get_webhooks() {
+  curl -sS -m 8 "${BB_URL}/api/v1/webhook?password=${BB_PASS}"
+}
+
+# Delete every webhook whose URL mentions our APP_ID (cleanup is idempotent).
+bb_cleanup_webhooks() {
+  local ids
+  ids=$(bb_get_webhooks | python3 -c '
+import json,sys
+try: d=json.load(sys.stdin)
+except Exception: sys.exit(0)
+for w in (d.get("data") or []):
+    if "'"$APP_ID"'" in (w.get("url") or ""): print(w["id"])
+' 2>/dev/null || true)
+  for id in $ids; do
+    curl -sS -m 8 -X DELETE "${BB_URL}/api/v1/webhook/${id}?password=${BB_PASS}" >/dev/null 2>&1 || true
+  done
+}
+
+# ── IPC helper (drives the Electron renderer's electronAPI bridge) ───
+# Runs a JS snippet that returns a string token; prints the raw token.
+# The BlueBubbles password is base64-injected (atob) so special chars in the
+# secret never need shell/JS quoting.
+ipc_eval() {
+  local js="$1"
+  agent-browser --cdp "$CDP_PORT" eval -b "$(printf '%s' "$js" | base64)" 2>/dev/null
+}
+
+PASS_B64=$(printf '%s' "$BB_PASS" | base64)
+
+# Emit an inline JS object literal for the bridge config. $1 overrides the
+# password expression (defaults to atob of the real password); pass a JS string
+# literal like "'wrong'" to test the rejection path.
+ipc_config_js() {
+  local pwexpr="${1:-atob('${PASS_B64}')}"
+  printf "{applicationId:'%s',blueBubblesServerUrl:'%s',blueBubblesPassword:%s,enabled:true,webhookSecret:'%s'}" \
+    "$APP_ID" "$BB_URL" "$pwexpr" "$SECRET"
+}
+
+# ── Preflight ────────────────────────────────────────────────────────
+note "BlueBubbles: ${BB_URL}   CDP: ${CDP_PORT}   appId: ${APP_ID}"
+
+code=$(curl -sS -m 6 -o /dev/null -w '%{http_code}' \
+  "${BB_URL}/api/v1/server/info?password=${BB_PASS}" || echo 000)
+if [ "$code" = "200" ]; then ok "BlueBubbles reachable + password valid"; else
+  bad "BlueBubbles preflight" "HTTP $code (is BlueBubbles running on ${BB_URL}?)"
+  echo "Aborting — fix BlueBubbles first."; exit 1
+fi
+
+if ! curl -sf --max-time 3 "http://localhost:${CDP_PORT}/json/version" >/dev/null 2>&1; then
+  bad "Electron CDP preflight" "CDP ${CDP_PORT} unreachable — run electron-dev.sh start"
+  echo "Aborting."; exit 1
+fi
+ok "Electron CDP reachable"
+
+# Bridge must expose the IPC group (built from this branch's code).
+probe=$(ipc_eval "(async()=>{try{var s=await window.electronAPI.invoke('imessageBridge.getStatus',{});return 'OK:'+JSON.stringify(s);}catch(e){return 'ERR:'+(e.message||e);}})()")
+case "$probe" in
+  *OK:*) ok "imessageBridge IPC available" ;;
+  *) bad "imessageBridge IPC" "got: $probe (is the iMessage Desktop branch checked out?)"; echo "Aborting."; exit 1 ;;
+esac
+
+# Start clean: remove any leftover config for this appId + BB webhooks.
+ipc_eval "(async()=>{try{await window.electronAPI.invoke('imessageBridge.removeConfig',{applicationId:'${APP_ID}'});}catch(e){}return 'done';})()" >/dev/null
+bb_cleanup_webhooks
+
+# ── testConfig: happy path ───────────────────────────────────────────
+r=$(ipc_eval "(async()=>{try{var c=$(ipc_config_js);var x=await window.electronAPI.invoke('imessageBridge.testConfig',c);return 'OK:'+JSON.stringify(x);}catch(e){return 'ERR:'+(e.message||e);}})()")
+case "$r" in
+  *OK:*success*true*) ok "testConfig with valid password → success" ;;
+  *) bad "testConfig (valid)" "got: $r" ;;
+esac
+
+# ── testConfig: wrong password rejects ───────────────────────────────
+r=$(ipc_eval "(async()=>{try{var c=$(ipc_config_js "'definitely-wrong-password'");var x=await window.electronAPI.invoke('imessageBridge.testConfig',c);return 'OK:'+JSON.stringify(x);}catch(e){return 'ERR:'+(e.message||e);}})()")
+case "$r" in
+  *ERR:*) ok "testConfig with wrong password → rejected" ;;
+  *) bad "testConfig (wrong password)" "expected rejection, got: $r" ;;
+esac
+
+# ── testConfig: unreachable URL rejects ──────────────────────────────
+r=$(ipc_eval "(async()=>{try{var x=await window.electronAPI.invoke('imessageBridge.testConfig',{applicationId:'${APP_ID}',blueBubblesServerUrl:'http://127.0.0.1:65530',blueBubblesPassword:atob('${PASS_B64}'),enabled:true,webhookSecret:'${SECRET}'});return 'OK:'+JSON.stringify(x);}catch(e){return 'ERR:'+(e.message||e);}})()")
+case "$r" in
+  *ERR:*) ok "testConfig with unreachable URL → rejected" ;;
+  *) bad "testConfig (unreachable)" "expected rejection, got: $r" ;;
+esac
+
+# ── upsertConfig: FIRST-TIME registration (Bug #1 regression guard) ──
+# BlueBubbles' GET /webhook?url=<unregistered> returns HTTP 500. The bridge
+# must list ALL webhooks and match client-side, otherwise this first save
+# fails. This assertion guards that fix.
+r=$(ipc_eval "(async()=>{try{var c=$(ipc_config_js);var x=await window.electronAPI.invoke('imessageBridge.upsertConfig',c);return 'OK:'+JSON.stringify(x);}catch(e){return 'ERR:'+(e.message||e);}})()")
+case "$r" in
+  *OK:*success*true*) ok "upsertConfig first-time save → success (Bug #1 guard)" ;;
+  *) bad "upsertConfig (first-time)" "got: $r" ;;
+esac
+
+# ── getStatus: bridge running + config persisted ─────────────────────
+# Return a quote-free token so grep isn't tripped up by agent-browser's
+# JSON-string escaping of the eval result.
+r=$(ipc_eval "(async()=>{var s=await window.electronAPI.invoke('imessageBridge.getStatus',{});var c=(s.configs||[]).find(function(x){return x.applicationId==='${APP_ID}';});return 'RUN='+(s.running?'Y':'N')+' CFG='+(c?'Y':'N')+' PW='+((c&&c.blueBubblesPasswordSet)?'Y':'N');})()")
+echo "$r" | grep -q 'RUN=Y' && ok "bridge running" || bad "bridge running" "got: $r"
+echo "$r" | grep -q 'CFG=Y' && ok "config persisted" || bad "config persisted" "got: $r"
+echo "$r" | grep -q 'PW=Y'  && ok "password stored (redacted in status)" || bad "password stored" "got: $r"
+
+# ── BlueBubbles webhook actually registered ──────────────────────────
+if bb_get_webhooks | grep -q "${APP_ID}"; then
+  ok "BlueBubbles webhook registered for appId"
+else
+  bad "BlueBubbles webhook" "no webhook URL containing ${APP_ID}"
+fi
+
+# ── Local bridge HTTP server: secret enforcement ─────────────────────
+BRIDGE_URL=$(ipc_eval "(async()=>{var s=await window.electronAPI.invoke('imessageBridge.getStatus',{});return s.serverUrl||'';})()" | tr -d '"')
+if [ -n "$BRIDGE_URL" ]; then
+  # wrong secret → 401
+  code=$(curl -sS -m 6 -o /dev/null -w '%{http_code}' -X POST \
+    -H 'Content-Type: application/json' \
+    "${BRIDGE_URL}/webhooks/bluebubbles/${APP_ID}?secret=WRONG" \
+    -d '{"type":"new-message","data":{"guid":"x"}}' || echo 000)
+  [ "$code" = "401" ] && ok "local bridge rejects wrong secret (401)" || bad "local bridge wrong secret" "expected 401, got $code"
+
+  # right secret → passes auth (reaches forward; without a bound cloud bot it
+  # returns 5xx — that's fine, we're only asserting auth + routing here)
+  code=$(curl -sS -m 6 -o /dev/null -w '%{http_code}' -X POST \
+    -H 'Content-Type: application/json' \
+    "${BRIDGE_URL}/webhooks/bluebubbles/${APP_ID}?secret=${SECRET}" \
+    -d '{"type":"new-message","data":{"guid":"x","text":"hi"}}' || echo 000)
+  [ "$code" != "401" ] && ok "local bridge accepts valid secret (HTTP $code, past auth)" || bad "local bridge valid secret" "got 401 with correct secret"
+else
+  bad "local bridge URL" "getStatus returned no serverUrl"
+fi
+
+# ── Cleanup ──────────────────────────────────────────────────────────
+ipc_eval "(async()=>{try{await window.electronAPI.invoke('imessageBridge.removeConfig',{applicationId:'${APP_ID}'});await window.electronAPI.invoke('imessageBridge.stop',{});}catch(e){}return 'cleaned';})()" >/dev/null
+bb_cleanup_webhooks
+note "cleaned up config + BlueBubbles webhook for ${APP_ID}"
+
+# ── Summary ──────────────────────────────────────────────────────────
+echo ""
+echo "[imsg-test] PASS=${PASS}  FAIL=${FAIL}"
+[ "$FAIL" -eq 0 ] || exit 1
@@ -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 [osascript-common.md](../osascript-common.md) for shared patterns.

 ## Activate & Navigate

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

 ```bash
-./.agents/skills/local-testing/scripts/test-lark-bot.sh "bot-testing" "@MyBot hello"
-./.agents/skills/local-testing/scripts/test-lark-bot.sh "bot-testing" "Help me with this" 30
+./.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
 ```
@@ -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"
+"$SCRIPT_DIR/../capture-app-window.sh" "$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 [osascript-common.md](../osascript-common.md) for shared patterns.

 ## Activate & Navigate

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

 ```bash
-./.agents/skills/local-testing/scripts/test-qq-bot.sh "bot-testing" "Hello bot" 15
-./.agents/skills/local-testing/scripts/test-qq-bot.sh "MyBot" "/help" 10
+./.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
 ```
@@ -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"
+"$SCRIPT_DIR/../capture-app-window.sh" "$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 [osascript-common.md](../osascript-common.md) for shared patterns.

 ## Activate & Navigate

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

 ```bash
-./.agents/skills/local-testing/scripts/test-slack-bot.sh "bot-testing" "@mybot hello"
-./.agents/skills/local-testing/scripts/test-slack-bot.sh "bot-testing" "/ask What is 2+2?" 20
+./.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
 ```
@@ -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"
+"$SCRIPT_DIR/../capture-app-window.sh" "$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 [osascript-common.md](../osascript-common.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/scripts/test-telegram-bot.sh "MyTestBot" "/start"
-./.agents/skills/local-testing/scripts/test-telegram-bot.sh "GPTBot" "Hello" 60
+./.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
 ```
@@ -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"
+"$SCRIPT_DIR/../capture-app-window.sh" "$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 [osascript-common.md](../osascript-common.md) for shared patterns.

 ## Activate & Navigate

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

 ```bash
-./.agents/skills/local-testing/scripts/test-wechat-bot.sh "文件传输助手" "test message" 5
-./.agents/skills/local-testing/scripts/test-wechat-bot.sh "MyBot" "Tell me a joke" 30
+./.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
 ```
@@ -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"
+"$SCRIPT_DIR/../capture-app-window.sh" "$APP" "$SCREENSHOT"
 echo "[$APP] Done! Screenshot saved to $SCREENSHOT"
@@ -0,0 +1,93 @@
+# LobeHub gateway streaming + tab-switch test harness
+
+Captures store + DOM state at 200ms intervals so we can prove or disprove
+claims like "切回 tab 后消息回到了很早以前". Built for gateway-mode chat but
+works for any LobeHub streaming session.
+
+## Files
+
+`scripts/agent-gateway/`
+
+| File            | Role                                                             |
+| --------------- | ---------------------------------------------------------------- |
+| `probe.js`      | Injects a 200ms sampler + `__PROBE_EVENT` marker + `__switchTab` |
+| `probe-dump.js` | Stops the sampler and returns `{events, samples}` as JSON string |
+| `tab-switch.js` | Runs N round-trip switches between two tabs, marks each step     |
+| `analyze.mjs`   | Node post-processor: timeline + regression detection             |
+
+## Standard workflow
+
+```bash
+# 1. Start Electron with CDP
+./.agents/skills/local-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
+
+# 4. Send a tool-call message — manually or via type+press
+agent-browser --cdp 9222 eval "window.__PROBE_EVENT('SENT')"
+
+# 5. Run the multi-switch driver (auto-picks active tab as BACK and the
+#    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
+
+# 6. Wait for streaming to finish, then dump
+agent-browser --cdp 9222 eval --stdin \
+  < .agents/skills/local-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
+```
+
+The analyzer prints three sections: EVENTS, TIMELINE, REGRESSIONS. If
+REGRESSIONS is non-empty it means content/reasoning/childN dropped on the
+same topic — the symptom users describe.
+
+## What the probe tracks (and why)
+
+`chat.messagesMap` only stores the top-level `assistantGroup` shell. The
+actual streamed content, reasoning, and tool calls live in
+`assistantGroup.children: AssistantContentBlock[]`. Any probe that only
+reads `m.content` / `m.reasoning` will see zeros throughout streaming and
+miss everything that matters. probe.js walks both levels and sums:
+
+- `cT` total content length
+- `rT` total reasoning length
+- `toolT` total tool-call count
+- `childN` number of content blocks
+
+Plus DOM-side signals (`domLen`, search/crawl indicator counts) so you can
+tell store-side regressions apart from render-side regressions.
+
+## Gotchas
+
+- **Optimistic new-topic state.** Before the first chunk lands, messages
+  live under the `<scope>_new` key with `tmp_*` ids and no `topicId` field.
+  probe.js falls back to those when `activeTopicId` is null.
+- **Reasoning resets to 0 are not bugs.** When the assistant finishes
+  thinking and starts tool-use or text, the streaming reasoning buffer
+  empties and the finalised reasoning gets sealed into a completed block.
+  Filter these out manually if needed.
+- **DOM length jitters by a handful of chars** because counters like "(10)"
+  in tool-call labels change as results arrive. analyze.mjs only flags
+  `domLen` drops greater than 100 chars to ignore that noise.
+- **Never identify tabs by innerText.** The active tab's text embeds a
+  ` · <agent name>` suffix, so a search like `'LobeHub Growth'` matches the
+  active tab when the active agent happens to be LobeHub Growth — and you
+  end up clicking the tab you're already on. probe.js uses the stable
+  `data-contextmenu-trigger` attribute (a React `useId()` value that's set
+  per-tab and survives focus changes) plus `data-active="true"` to mark
+  the active one. Helpers exposed:
+  `__listTabs()` / `__clickTabByKey(key)` / `__clickTabByIndex(i)` /
+  `__activeTabKey()`.
+- **`tab-switch.js` fires-and-forgets.** The IIFE kicks off an async loop
+  and returns immediately so the agent-browser CLI eval doesn't blow past
+  its default 25 s timeout. Wait on the `SWITCH_LOOP_DONE` event marker
+  before dumping. Re-running while a loop is in flight is refused — the
+  chaotic data from overlapping runs is not worth debugging.
@@ -0,0 +1,243 @@
+// Analyzer for probe-events dumps. Reads a JSON file produced by `run.ts dump`
+// and prints a layered breakdown:
+//
+//   1. STREAM EVENTS — every non-chunk WS/SSE event in receipt order
+//   2. CHUNKS SUMMARY — collapsed per-step chunk counts (otherwise floods)
+//   3. ACTION CALLS — replaceMessages / refreshMessages / MARK:* with stack
+//   4. CORRELATION — calls ↔ nearest stream event within ±300ms
+//   5. PER-KEY ASSISTANT GROWTH — for each messagesMap key, when the leading
+//      assistant message's cLen / rLen actually moves (this is what reveals
+//      "chunks arrived but the message never grew" regressions)
+//   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>
+
+import { readFileSync } from 'node:fs';
+
+import type {
+  ProbeActionCall,
+  ProbeDump,
+  ProbeMessageSummary,
+  ProbeStreamEvent,
+  ProbeTimelineSample,
+} from './types';
+
+const file = process.argv[2];
+if (!file) {
+  console.error('usage: bun run analyze-events.ts <dump.json>');
+  process.exit(1);
+}
+
+const raw = readFileSync(file, 'utf8');
+// agent-browser eval --stdin wraps return values in quotes when the value is
+// a string — so the JSON file may be double-encoded depending on how it was
+// captured. Handle both.
+const parsedOnce = JSON.parse(raw) as ProbeDump | string;
+const dump: ProbeDump = typeof parsedOnce === 'string' ? JSON.parse(parsedOnce) : parsedOnce;
+
+const { streamEvents = [], actionCalls = [], timeline = [] } = dump;
+
+const pad = (v: unknown, n: number) => String(v).padStart(n);
+
+// ── META ───────────────────────────────────────────────────────────
+console.log('=== META ===');
+console.log(`  events:    ${streamEvents.length}`);
+console.log(`  calls:     ${actionCalls.length}`);
+console.log(`  timeline:  ${timeline.length}`);
+
+// ── 1. STREAM EVENTS (non-chunk) ───────────────────────────────────
+const nonChunkEvents = streamEvents.filter((e) => e.type !== 'stream_chunk');
+const chunkEvents = streamEvents.filter((e) => e.type === 'stream_chunk');
+
+console.log(
+  `\n=== STREAM EVENTS (${nonChunkEvents.length} non-chunk + ${chunkEvents.length} chunks elided) ===`,
+);
+for (const e of nonChunkEvents) {
+  const dataStr = e.dataKeys?.length ? ` [${e.dataKeys.join(',')}]` : '';
+  const data = e.data as Record<string, unknown> | undefined;
+  const uiHint = data?.uiMessagesPreview
+    ? ` uiPreview=${JSON.stringify(data.uiMessagesPreview)}`
+    : data?.uiMessagesTotal
+      ? ` uiTotal=${data.uiMessagesTotal}`
+      : '';
+  const phaseHint = data?.phase ? ` phase=${data.phase}` : '';
+  const extra = e.serverType ? ` serverType=${e.serverType}` : '';
+  console.log(
+    `  t=${pad(e.t, 7)}  [${(e.transport ?? '?').padEnd(3)}]  step=${pad(e.stepIndex ?? '-', 2)}  ` +
+      `type=${(e.type ?? '').padEnd(22)}  op=${e.opIdTail ?? '-'}${phaseHint}${uiHint}${extra}${dataStr}`,
+  );
+}
+
+// ── 2. CHUNK SUMMARY ───────────────────────────────────────────────
+console.log('\n=== CHUNKS SUMMARY (per step / chunkType) ===');
+const chunkBuckets = new Map<string, { count: number; firstT: number; lastT: number }>();
+for (const c of chunkEvents) {
+  const data = c.data as Record<string, unknown> | undefined;
+  const ct = (data?.chunkType as string | undefined) ?? '?';
+  const key = `step=${c.stepIndex ?? '-'}  chunkType=${ct.padEnd(8)}  op=${c.opIdTail}`;
+  const slot = chunkBuckets.get(key);
+  if (slot) {
+    slot.count += 1;
+    slot.lastT = c.t;
+  } else {
+    chunkBuckets.set(key, { count: 1, firstT: c.t, lastT: c.t });
+  }
+}
+for (const [k, v] of chunkBuckets) {
+  console.log(`  ${k}  count=${pad(v.count, 4)}  t=${pad(v.firstT, 7)}..${pad(v.lastT, 7)}`);
+}
+
+// ── 3. ACTION CALLS ───────────────────────────────────────────────
+console.log('\n=== ACTION CALLS (replace/refresh/MARK) ===');
+for (const c of actionCalls) {
+  if (c.name?.startsWith('MARK:')) {
+    console.log(`  t=${pad(c.t, 7)}  ${c.name}`);
+    continue;
+  }
+  const snapshot = (c.args as any)?.snapshot as
+    | Array<{ id: string; role: string; cLen: number; rLen: number }>
+    | undefined;
+  const snapStr = snapshot?.length
+    ? '  snapshot=' + snapshot.map((m) => `${m.id}:${m.role}/c${m.cLen}/r${m.rLen}`).join(' | ')
+    : '';
+  const summary =
+    c.name === 'replaceMessages'
+      ? `count=${c.args?.count} action=${(c.args?.params as any)?.action ?? '-'}${snapStr}`
+      : c.name === 'refreshMessages'
+        ? `ctx=${JSON.stringify(c.args?.context)}`
+        : c.error
+          ? `error=${c.error}`
+          : '';
+  console.log(`  t=${pad(c.t, 7)}  ${c.name.padEnd(20)} ${summary}`);
+  if (c.stack) {
+    const frames = c.stack
+      .split(' ← ')
+      .filter((f) => !!f && !f.includes('Object.<anonymous>'))
+      .slice(0, 3);
+    for (const f of frames) console.log(`             ↳ ${f}`);
+  }
+}
+
+// ── 4. CORRELATION ────────────────────────────────────────────────
+function nearestEventForCall(
+  call: ProbeActionCall,
+  windowMs = 300,
+): { event: ProbeStreamEvent; delta: number } | null {
+  let best: ProbeStreamEvent | null = null;
+  let bestDelta = Infinity;
+  for (const e of streamEvents) {
+    const d = Math.abs(e.t - call.t);
+    if (d < bestDelta && d <= windowMs) {
+      bestDelta = d;
+      best = e;
+    }
+  }
+  return best ? { event: best, delta: bestDelta } : null;
+}
+
+console.log('\n=== CORRELATION (replace/refresh ↔ nearest event within ±300ms) ===');
+for (const c of actionCalls) {
+  if (c.name !== 'refreshMessages' && c.name !== 'replaceMessages') continue;
+  const hit = nearestEventForCall(c);
+  if (hit) {
+    const phase = (hit.event.data as Record<string, unknown> | undefined)?.phase;
+    console.log(
+      `  t=${pad(c.t, 7)}  ${c.name.padEnd(16)} ← Δ${pad(hit.delta, 4)}ms ${hit.event.type}` +
+        (phase ? ` phase=${phase}` : ''),
+    );
+  } else {
+    console.log(`  t=${pad(c.t, 7)}  ${c.name.padEnd(16)} ← (no event nearby — external trigger)`);
+  }
+}
+
+// ── 5. PER-KEY ASSISTANT GROWTH ───────────────────────────────────
+// For each messagesMap key, find the trailing assistant message and report
+// the points in time where its cLen / rLen actually changed. If the timeline
+// shows chunks arriving but the assistant cLen never moves, that's the
+// signature of "dispatch queue blocked / messageId mismatch".
+console.log('\n=== PER-KEY ASSISTANT GROWTH ===');
+const keysEverSeen = new Set<string>();
+for (const s of timeline) for (const k of Object.keys(s.byKey ?? {})) keysEverSeen.add(k);
+
+for (const key of keysEverSeen) {
+  console.log(`\n  key=${key}`);
+  let lastSig: string | null = null;
+  for (const s of timeline) {
+    const slot = s.byKey?.[key];
+    if (!slot) continue;
+    const last = slot.msgs.at(-1) as ProbeMessageSummary | undefined;
+    if (!last) continue;
+    const sig = `${last.id}|c${last.cLen}|r${last.rLen}|n${slot.n}`;
+    if (sig === lastSig) continue;
+    lastSig = sig;
+    console.log(
+      `    t=${pad(s.t, 7)}  msgN=${pad(slot.n, 3)}  ` +
+        `lastAssistant=${last.id}  cLen=${pad(last.cLen, 5)}  rLen=${pad(last.rLen, 5)}` +
+        `  runOps=${s.runOps}`,
+    );
+  }
+}
+
+// ── 6. ROLLBACKS (active-topic msgN / childN / role drops) ─────────
+console.log('\n=== ROLLBACKS (active-topic msgN / childN / role drops) ===');
+let prev: ProbeTimelineSample | null = null;
+const rollbacks: Array<{ t: number; topic: string | null; drops: string[] }> = [];
+
+const flatten = (s: ProbeTimelineSample) => {
+  if (!s.activeTopic) return [];
+  return Object.entries(s.byKey ?? {})
+    .filter(([k]) => k.includes(s.activeTopic!))
+    .flatMap(([, v]) => v.msgs);
+};
+
+for (const s of timeline) {
+  if (s.err) {
+    prev = null;
+    continue;
+  }
+  if (!prev || prev.activeTopic !== s.activeTopic) {
+    prev = s;
+    continue;
+  }
+  const prevMsgs = flatten(prev);
+  const curMsgs = flatten(s);
+  const drops: string[] = [];
+
+  if (curMsgs.length < prevMsgs.length) drops.push(`msgN ${prevMsgs.length}→${curMsgs.length}`);
+
+  let prevChild = 0;
+  let curChild = 0;
+  for (const m of prevMsgs) prevChild += m.chN ?? 0;
+  for (const m of curMsgs) curChild += m.chN ?? 0;
+  if (curChild < prevChild) drops.push(`childN ${prevChild}→${curChild}`);
+
+  const prevById = new Map(prevMsgs.map((m) => [m.id, m]));
+  for (const m of curMsgs) {
+    const pr = prevById.get(m.id);
+    if (!pr) continue;
+    if (m.cLen < pr.cLen) drops.push(`cLen[${m.id}] ${pr.cLen}→${m.cLen}`);
+    if (m.rLen < pr.rLen) drops.push(`rLen[${m.id}] ${pr.rLen}→${m.rLen}`);
+  }
+
+  if (drops.length) rollbacks.push({ t: s.t, topic: s.activeTopic, drops });
+  prev = s;
+}
+
+if (rollbacks.length === 0) {
+  console.log('  (none)');
+} else {
+  for (const r of rollbacks) {
+    const nearEvent = streamEvents
+      .filter((e) => Math.abs(e.t - r.t) <= 300)
+      .map((e) => `${e.type}${(e.data as any)?.phase ? ':' + (e.data as any).phase : ''}`);
+    const nearCall = actionCalls
+      .filter((c) => Math.abs(c.t - r.t) <= 300 && !c.name?.startsWith('MARK:'))
+      .map((c) => c.name);
+    console.log(
+      `  t=${pad(r.t, 7)}  topic=${r.topic}  ${r.drops.join(' | ')}` +
+        (nearEvent.length ? `  near-event:[${nearEvent.join(',')}]` : '') +
+        (nearCall.length ? `  near-call:[${nearCall.join(',')}]` : ''),
+    );
+  }
+}
@@ -0,0 +1,119 @@
+#!/usr/bin/env node
+// Analyze a probe dump captured by probe.js + probe-dump.js.
+//
+//   node analyze.mjs /tmp/probe.json
+//
+// Prints:
+//   1. EVENTS — user-action markers with their relative timestamps
+//   2. TIMELINE — periodic samples (~1 per second + event-adjacent samples)
+//      showing every interesting field; columns:
+//        t(ms) | runOps | msgN | childN | content | reasoning | tools | domLen | search | crawl | topic | event
+//   3. REGRESSIONS — every place a tracked counter *dropped* on the same
+//      topic between adjacent samples. A "true" UI rollback shows up as a
+//      drop in content/reasoning/tools/childN/domLen without a topic change.
+//
+// Whitelisted transitions (not flagged):
+//   - topic change → all drops expected (focus moved away)
+//   - reasoning length 0 after content starts → reasoning gets sealed into a
+//     completed sub-block; the parent's running reasoning resets to ''.
+//   - msgN drop when topic transitions from `_new` placeholder to a real id.
+
+import fs from 'node:fs';
+
+const file = process.argv[2];
+if (!file) {
+  console.error('usage: node analyze.mjs <probe.json>');
+  process.exit(1);
+}
+
+const raw = JSON.parse(fs.readFileSync(file, 'utf8'));
+// probe-dump.js wraps the payload in JSON.stringify so agent-browser returns
+// it as a single quoted string. Unwrap.
+const data = typeof raw === 'string' ? JSON.parse(raw) : raw;
+const { events, samples } = data;
+
+const fmt = {
+  pad(v, n) {
+    return String(v).padStart(n);
+  },
+};
+
+console.log('=== EVENTS ===');
+for (const e of events) console.log(`  t=${fmt.pad(e.t, 7)}  ${e.name}`);
+
+console.log(
+  '\n=== TIMELINE (~1s cadence, plus event-adjacent samples) ===\n' +
+    '  t(ms)   runOps  msgN childN  content reasoning tools  domLen  search crawl  topic     event',
+);
+
+let lastSampledAt = -1e9;
+const eventBuckets = events.map((e) => e.t);
+for (let i = 0; i < samples.length; i++) {
+  const s = samples[i];
+  const nearEvent = eventBuckets.some((et) => Math.abs(et - s.t) < 110);
+  if (!nearEvent && s.t - lastSampledAt < 1000) continue;
+  lastSampledAt = s.t;
+
+  const ev = events.find((e) => Math.abs(e.t - s.t) < 110);
+  const evMarker = ev ? `  ◀ ${ev.name}` : '';
+  const topicSuffix = s.topicId ? s.topicId.slice(-6) : '(none)';
+  const search = s.ind?.search ?? 0;
+  const crawl = s.ind?.crawl ?? 0;
+  console.log(
+    `  ${fmt.pad(s.t, 6)} ` +
+      `${fmt.pad(s.runOps, 6)}  ` +
+      `${fmt.pad(s.msgN, 4)}  ` +
+      `${fmt.pad(s.childN ?? 0, 5)} ` +
+      `${fmt.pad(s.cT ?? 0, 8)} ` +
+      `${fmt.pad(s.rT ?? 0, 9)} ` +
+      `${fmt.pad(s.toolT ?? 0, 5)} ` +
+      `${fmt.pad(s.domLen ?? 0, 7)} ` +
+      `${fmt.pad(search, 6)} ` +
+      `${fmt.pad(crawl, 5)}  ` +
+      `${topicSuffix.padEnd(8)}${evMarker}`,
+  );
+}
+
+console.log('\n=== REGRESSIONS (same topic, value dropped) ===');
+const regressions = [];
+for (let i = 1; i < samples.length; i++) {
+  const prev = samples[i - 1];
+  const cur = samples[i];
+  if (!cur.topicId || prev.topicId !== cur.topicId) continue;
+
+  const drops = [];
+  if (cur.msgN < prev.msgN) drops.push(`msgN: ${prev.msgN}→${cur.msgN}`);
+  if ((cur.childN ?? 0) < (prev.childN ?? 0)) drops.push(`childN: ${prev.childN}→${cur.childN}`);
+  if ((cur.cT ?? 0) < (prev.cT ?? 0)) drops.push(`content: ${prev.cT}→${cur.cT}`);
+  if ((cur.rT ?? 0) < (prev.rT ?? 0)) drops.push(`reasoning: ${prev.rT}→${cur.rT}`);
+  if ((cur.toolT ?? 0) < (prev.toolT ?? 0)) drops.push(`tools: ${prev.toolT}→${cur.toolT}`);
+  // domLen jitters by a few chars from counter labels — only flag big drops.
+  if ((cur.domLen ?? 0) < (prev.domLen ?? 0) - 100) {
+    drops.push(`domLen: ${prev.domLen}→${cur.domLen}`);
+  }
+  if (drops.length === 0) continue;
+
+  const nearbyEv = events.filter((e) => Math.abs(e.t - cur.t) < 600).map((e) => e.name);
+  regressions.push({ t: cur.t, topic: cur.topicId.slice(-6), drops, nearbyEv });
+}
+
+if (regressions.length === 0) {
+  console.log('  (none)');
+} else {
+  for (const r of regressions) {
+    const evStr = r.nearbyEv.length ? `  near:[${r.nearbyEv.join(',')}]` : '';
+    console.log(`  t=${fmt.pad(r.t, 7)}  topic=${r.topic}  ${r.drops.join(' | ')}${evStr}`);
+  }
+}
+
+console.log(`\n=== SUMMARY ===`);
+console.log(`  samples: ${samples.length}`);
+console.log(`  events:  ${events.length}`);
+console.log(`  regressions: ${regressions.length}`);
+if (samples.length) {
+  const last = samples.at(-1);
+  console.log(
+    `  final: msgN=${last.msgN} childN=${last.childN ?? 0} content=${last.cT ?? 0} ` +
+      `reasoning=${last.rT ?? 0} tools=${last.toolT ?? 0} runOps=${last.runOps}`,
+  );
+}
@@ -0,0 +1,17 @@
+// Stop the probe and serialize collected data.
+//
+//   agent-browser --cdp 9222 eval --stdin < probe-dump.js > /tmp/probe.json
+//
+// The whole thing is wrapped in a JSON.stringify so agent-browser returns it
+// as a single quoted string — the analyzer double-parses to handle that.
+
+(function () {
+  if (window.__PROBE_TIMER) {
+    clearInterval(window.__PROBE_TIMER);
+    window.__PROBE_TIMER = null;
+  }
+  return JSON.stringify({
+    events: window.__PROBE_EVENTS || [],
+    samples: window.__PROBE_SAMPLES || [],
+  });
+})();
@@ -0,0 +1,37 @@
+// Stops the events-probe timeline timer and stashes the full capture as a
+// JSON string on `window.__PROBE_LAST_DUMP_JSON`. `run.ts` wraps the bundle
+// in an IIFE that returns that global, which `agent-browser eval` prints to
+// stdout — the runner then persists it under `.agent-gateway/`.
+
+import type { ProbeDump } from './types';
+
+declare global {
+  interface Window {
+    __PROBE_LAST_DUMP_JSON?: string;
+  }
+}
+
+const w = window;
+
+if (w.__PROBE_TIMELINE_TIMER) {
+  clearInterval(w.__PROBE_TIMELINE_TIMER);
+  w.__PROBE_TIMELINE_TIMER = null;
+}
+
+const mutations = w.__PROBE_MUTATIONS ?? [];
+
+const dump: ProbeDump & { mutations: typeof mutations } = {
+  meta: {
+    t0: w.__PROBE_T0 ?? 0,
+    collectedAt: Date.now(),
+    sampleCount: (w.__PROBE_MSG_TIMELINE ?? []).length,
+    eventCount: (w.__PROBE_STREAM_EVENTS ?? []).length,
+    callCount: (w.__PROBE_ACTION_CALLS ?? []).length,
+  },
+  streamEvents: w.__PROBE_STREAM_EVENTS ?? [],
+  actionCalls: w.__PROBE_ACTION_CALLS ?? [],
+  timeline: w.__PROBE_MSG_TIMELINE ?? [],
+  mutations,
+};
+
+w.__PROBE_LAST_DUMP_JSON = JSON.stringify(dump);
@@ -0,0 +1,637 @@
+// LobeHub gateway raw-event-stream probe.
+//
+// Gateway-mode chats subscribe via WebSocket — NOT via the `/api/agent/stream`
+// SSE endpoint (that one belongs to the direct/client durable-agent runtime).
+// `AgentStreamClient` (`packages/agent-gateway-client/src/client.ts`) opens
+// `new WebSocket('wss://.../ws?operationId=...')`, then parses JSON frames in
+// its `onmessage` handler and re-emits `agent_event.event` objects to the
+// chat store.
+//
+// To capture the RAW gateway events before the store touches them, we wrap
+// `window.WebSocket` so that for any socket whose URL contains `operationId=`
+// we intercept the `onmessage` handler / `addEventListener('message')` and
+// log every `agent_event` frame.
+//
+// We *also* keep the `window.fetch` hook for `/api/agent/stream` so this
+// probe still works for direct-mode runs — but gateway-mode events come
+// through the WebSocket path.
+//
+// Buffers (read via `dump`):
+//   __PROBE_STREAM_EVENTS  — raw events parsed off the wire
+//   __PROBE_ACTION_CALLS   — replaceMessages / refreshMessages calls (best-effort)
+//   __PROBE_MSG_TIMELINE   — 200ms snapshots of every messagesMap key
+
+import type {
+  ProbeActionCall,
+  ProbeMessageSummary,
+  ProbeStreamEvent,
+  ProbeTimelineSample,
+} from './types';
+
+// Bundled by esbuild as an IIFE. Top-level code runs once on injection.
+
+const w = window;
+
+// ── Buffers ─────────────────────────────────────────────────────────
+
+declare global {
+  interface Window {
+    __PROBE_MUTATIONS?: Array<{
+      t: number;
+      key: string;
+      n: number;
+      last?: { id: string; role: string; cLen: number; rLen: number; updatedAt?: unknown };
+      prevLast?: { id: string; role: string; cLen: number; rLen: number };
+      delta?: string;
+    }>;
+    __PROBE_STORE_UNSUB?: () => void;
+  }
+}
+
+const events: ProbeStreamEvent[] = (w.__PROBE_STREAM_EVENTS ??= []);
+const calls: ProbeActionCall[] = (w.__PROBE_ACTION_CALLS ??= []);
+const timeline: ProbeTimelineSample[] = (w.__PROBE_MSG_TIMELINE ??= []);
+const mutations = (w.__PROBE_MUTATIONS ??= []);
+events.length = 0;
+calls.length = 0;
+timeline.length = 0;
+mutations.length = 0;
+
+const t0 = Date.now();
+w.__PROBE_T0 = t0;
+const now = (): number => Date.now() - t0;
+
+// ── Helpers ─────────────────────────────────────────────────────────
+
+function summarizeData(data: unknown): Record<string, unknown> | unknown {
+  if (!data || typeof data !== 'object') return data;
+  const src = data as Record<string, unknown>;
+  const out: Record<string, unknown> = {};
+  for (const k of Object.keys(src)) {
+    const v = src[k];
+    if (v == null) {
+      out[k] = v;
+    } else if (Array.isArray(v)) {
+      out[k] = `Array(${v.length})`;
+      if (k === 'uiMessages') {
+        out.uiMessagesPreview = v.slice(0, 5).map((m: any) => ({
+          id: (m.id ?? '').slice(-8),
+          role: m.role,
+          cLen: (m.content ?? '').length,
+          children: (m.children ?? []).length,
+          tools: (m.tools ?? []).length,
+          reasoning: (m.reasoning?.content ?? '').length,
+        }));
+        out.uiMessagesTotal = v.length;
+      }
+    } else if (typeof v === 'object') {
+      const obj = v as Record<string, unknown>;
+      out[k] =
+        'Object{' +
+        Object.keys(obj)
+          .slice(0, 6)
+          .map((kk) => kk + (typeof obj[kk] === 'string' ? `=${(obj[kk] as string).length}ch` : ''))
+          .join(',') +
+        '}';
+    } else if (typeof v === 'string') {
+      out[k] = v.length > 100 ? v.slice(0, 100) + `…(${v.length})` : v;
+    } else {
+      out[k] = v;
+    }
+  }
+  return out;
+}
+
+function summarizeMessages(msgs: any[]): ProbeMessageSummary[] {
+  return (msgs ?? []).slice(0, 80).map((m) => ({
+    id: (m.id ?? '').slice(-8),
+    role: m.role,
+    cLen: (m.content ?? '').length,
+    rLen: (m.reasoning?.content ?? '').length,
+    tools: (m.tools ?? []).length,
+    chN: (m.children ?? []).length,
+  }));
+}
+
+function shortStack(): string {
+  const raw = new Error('probe-stack').stack ?? '';
+  return raw
+    .split('\n')
+    .slice(3)
+    .filter((l) => !l.includes('probe-events') && !l.includes('node_modules'))
+    .map((l) => l.trim().replace(/^at\s+/, ''))
+    .slice(0, 6)
+    .join(' ← ');
+}
+
+function recordAgentEvent(args: {
+  transport: 'ws' | 'sse';
+  opId: string | null;
+  agentEvent: any;
+  eventId?: string | null;
+  rawLen?: number;
+}): void {
+  const { transport, opId, agentEvent, eventId, rawLen } = args;
+  if (!agentEvent || typeof agentEvent !== 'object') return;
+  events.push({
+    t: now(),
+    transport,
+    opIdTail: (opId ?? '').slice(-10),
+    eventId: eventId ?? null,
+    type: agentEvent.type,
+    stepIndex: agentEvent.stepIndex,
+    dataKeys: agentEvent.data ? Object.keys(agentEvent.data) : [],
+    data: summarizeData(agentEvent.data) as Record<string, unknown>,
+    rawLen,
+  });
+}
+
+// ── 1. Patch window.WebSocket for gateway WS events ────────────────
+
+if (!w.__PROBE_ORIG_WEBSOCKET) w.__PROBE_ORIG_WEBSOCKET = w.WebSocket;
+const OrigWS = w.__PROBE_ORIG_WEBSOCKET;
+
+function extractOpIdFromWsUrl(url: string | URL): string | null {
+  const m = String(url ?? '').match(/operationId=([^&]+)/);
+  return m ? decodeURIComponent(m[1]) : null;
+}
+
+function isGatewayWs(url: string | URL): boolean {
+  return String(url ?? '').includes('operationId=');
+}
+
+function handleWsFrame(rawData: unknown, opId: string | null): void {
+  const rawLen = typeof rawData === 'string' ? rawData.length : -1;
+  let parsed: any;
+  try {
+    parsed = typeof rawData === 'string' ? JSON.parse(rawData) : null;
+  } catch {
+    events.push({
+      t: now(),
+      transport: 'ws',
+      opIdTail: (opId ?? '').slice(-10),
+      type: '_PARSE_ERROR_',
+      raw: typeof rawData === 'string' && rawData.length < 400 ? rawData : '(non-string or large)',
+    });
+    return;
+  }
+  if (!parsed) return;
+
+  if (parsed.type === 'agent_event') {
+    recordAgentEvent({
+      transport: 'ws',
+      opId,
+      agentEvent: parsed.event,
+      eventId: parsed.id,
+      rawLen,
+    });
+  } else {
+    events.push({
+      t: now(),
+      transport: 'ws',
+      opIdTail: (opId ?? '').slice(-10),
+      type: '_SERVER_MSG_',
+      serverType: parsed.type,
+      rawLen,
+    });
+  }
+}
+
+// Wrap the constructor. Instance `constructor` will still reflect OrigWS
+// (we share prototypes), so use the `_WS_OPEN_` sentinel events to confirm
+// the patch is firing.
+function PatchedWebSocket(this: WebSocket, url: string | URL, protocols?: string | string[]) {
+  const ws: WebSocket = protocols == null ? new OrigWS(url) : new OrigWS(url, protocols);
+  const opId = extractOpIdFromWsUrl(url);
+  if (!isGatewayWs(url)) return ws;
+
+  events.push({
+    t: now(),
+    transport: 'ws',
+    opIdTail: (opId ?? '').slice(-10),
+    type: '_WS_OPEN_',
+    url: String(url),
+  });
+
+  // One observer listener that always fires, regardless of how the consumer
+  // (AgentStreamClient uses `ws.onmessage = …`) subscribes.
+  ws.addEventListener('message', (e) => {
+    try {
+      handleWsFrame((e as MessageEvent).data, opId);
+    } catch {
+      /* swallow */
+    }
+  });
+
+  ws.addEventListener('close', () => {
+    events.push({
+      t: now(),
+      transport: 'ws',
+      opIdTail: (opId ?? '').slice(-10),
+      type: '_WS_CLOSE_',
+    });
+  });
+
+  return ws;
+}
+
+// Preserve prototype + static fields so `instanceof WebSocket` and
+// `WebSocket.OPEN` constants still work.
+(PatchedWebSocket as unknown as { prototype: WebSocket }).prototype = OrigWS.prototype;
+for (const k of Object.keys(OrigWS) as Array<keyof typeof OrigWS>) {
+  try {
+    (PatchedWebSocket as any)[k] = (OrigWS as any)[k];
+  } catch {
+    /* readonly */
+  }
+}
+(['CONNECTING', 'OPEN', 'CLOSING', 'CLOSED'] as const).forEach((k) => {
+  (PatchedWebSocket as any)[k] = (OrigWS as any)[k];
+});
+w.WebSocket = PatchedWebSocket as unknown as typeof WebSocket;
+
+// ── 2. Patch window.fetch for `/api/agent/stream` (direct-mode SSE) ─
+
+if (!w.__PROBE_ORIG_FETCH) w.__PROBE_ORIG_FETCH = w.fetch.bind(w);
+const origFetch = w.__PROBE_ORIG_FETCH;
+
+function isAgentStreamUrl(input: RequestInfo | URL): boolean {
+  let url = '';
+  if (typeof input === 'string') url = input;
+  else if (input instanceof URL) url = input.toString();
+  else if (input && typeof (input as Request).url === 'string') url = (input as Request).url;
+  return url.includes('/api/agent/stream');
+}
+
+function extractOpIdFromHttpUrl(input: RequestInfo | URL): string | null {
+  const url = typeof input === 'string' ? input : (input as Request | URL).toString();
+  const m = url.match(/operationId=([^&]+)/);
+  return m ? decodeURIComponent(m[1]) : null;
+}
+
+function pushFromSSEFrame(rawFrame: string, opId: string | null): void {
+  const lines = rawFrame.split('\n');
+  let dataJson = '';
+  let evtName = 'message';
+  for (const line of lines) {
+    if (line.startsWith('event:')) evtName = line.slice(6).trim();
+    else if (line.startsWith('data:')) dataJson += line.slice(5).trim();
+  }
+  if (!dataJson) return;
+  let parsed: any;
+  try {
+    parsed = JSON.parse(dataJson);
+  } catch {
+    events.push({
+      t: now(),
+      transport: 'sse',
+      opIdTail: (opId ?? '').slice(-10),
+      type: '_PARSE_ERROR_',
+      sseEvent: evtName,
+      raw: dataJson.length > 400 ? dataJson.slice(0, 400) + '…' : dataJson,
+    });
+    return;
+  }
+  recordAgentEvent({
+    transport: 'sse',
+    opId,
+    agentEvent: parsed,
+    eventId: null,
+    rawLen: dataJson.length,
+  });
+}
+
+async function teeAndDrain(response: Response, opId: string | null): Promise<Response> {
+  if (!response.body) return response;
+  const [a, b] = response.body.tee();
+
+  void (async () => {
+    const reader = b.getReader();
+    const decoder = new TextDecoder();
+    let buf = '';
+    try {
+      while (true) {
+        const { value, done } = await reader.read();
+        if (done) break;
+        buf += decoder.decode(value, { stream: true });
+        let idx: number;
+
+        while ((idx = buf.indexOf('\n\n')) !== -1) {
+          const frame = buf.slice(0, idx);
+          buf = buf.slice(idx + 2);
+          if (frame.trim()) pushFromSSEFrame(frame, opId);
+        }
+      }
+      if (buf.trim()) pushFromSSEFrame(buf, opId);
+    } catch (e: any) {
+      events.push({
+        t: now(),
+        transport: 'sse',
+        opIdTail: (opId ?? '').slice(-10),
+        type: '_TEE_ERROR_',
+        message: String(e?.message ?? e),
+      });
+    }
+  })();
+
+  return new Response(a, {
+    headers: response.headers,
+    status: response.status,
+    statusText: response.statusText,
+  });
+}
+
+w.fetch = async function patchedFetch(input: RequestInfo | URL, init?: RequestInit) {
+  const response = await origFetch(input as any, init);
+  if (!isAgentStreamUrl(input)) return response;
+  const opId = extractOpIdFromHttpUrl(input);
+  const url =
+    typeof input === 'string'
+      ? input.split('?')[0]
+      : (input as Request | URL).toString().split('?')[0];
+  events.push({
+    t: now(),
+    transport: 'sse',
+    opIdTail: (opId ?? '').slice(-10),
+    type: '_CONNECTED_',
+    url,
+    status: response.status,
+  });
+  return teeAndDrain(response, opId);
+} as typeof fetch;
+
+// ── 3. Wrap store actions (best-effort for "who called replace") ────
+
+// Side-global stash for the original chat-store actions. Re-installs ALWAYS
+// rewrap from the originals so updates to the probe body take effect
+// without a page reload — using only a `__probeWrapped` flag on the chat
+// state object would freeze the first-installed wrapper across re-installs.
+declare global {
+  interface Window {
+    __PROBE_ORIG_REFRESH_MESSAGES?: any;
+    __PROBE_ORIG_REPLACE_MESSAGES?: any;
+  }
+}
+
+try {
+  const chat = w.__LOBE_STORES?.chat?.();
+  if (chat) {
+    // First-time install: cache the originals. Re-install: restore from
+    // the cached originals before wrapping again.
+    if (!w.__PROBE_ORIG_REFRESH_MESSAGES) w.__PROBE_ORIG_REFRESH_MESSAGES = chat.refreshMessages;
+    if (!w.__PROBE_ORIG_REPLACE_MESSAGES) w.__PROBE_ORIG_REPLACE_MESSAGES = chat.replaceMessages;
+    const origRefresh = w.__PROBE_ORIG_REFRESH_MESSAGES;
+    const origReplace = w.__PROBE_ORIG_REPLACE_MESSAGES;
+    chat.refreshMessages = origRefresh;
+    chat.replaceMessages = origReplace;
+
+    chat.refreshMessages = async function probeRefresh(this: unknown, ...args: any[]) {
+      calls.push({
+        t: now(),
+        name: 'refreshMessages',
+        args: { context: args[0] ?? null },
+        stack: shortStack(),
+      });
+      return origRefresh.apply(this, args);
+    };
+    chat.replaceMessages = function probeReplace(this: unknown, ...args: any[]) {
+      const msgs = (args[0] as any[]) ?? [];
+      const snapshot = msgs.slice(-2).map((m) => ({
+        id: (m.id ?? '').slice(-8),
+        role: m.role,
+        cLen: (m.content ?? '').length,
+        rLen: (m.reasoning?.content ?? '').length,
+        updatedAt: m.updatedAt,
+      }));
+      calls.push({
+        t: now(),
+        name: 'replaceMessages',
+        args: { count: msgs.length, params: args[1] ?? null, snapshot } as any,
+        stack: shortStack(),
+      });
+
+      // Pair the call with a mutation row so the analyzer can build a
+      // single ordered timeline across replaceMessages + dispatchMessage.
+      const stackTop = shortStack().split(' ← ')[0]?.slice(0, 80);
+      const last = msgs.at(-1);
+      const lastSum = last
+        ? {
+            id: (last.id ?? '').slice(-8),
+            role: last.role,
+            cLen: (last.content ?? '').length,
+            rLen: (last.reasoning?.content ?? '').length,
+            updatedAt: last.updatedAt,
+          }
+        : undefined;
+      const params: any = args[1] ?? {};
+      const ctxKey = params.context
+        ? `main_${params.context.agentId ?? '?'}_${
+            params.context.topicId ? 'tpc_' + params.context.topicId : 'new'
+          }`.replace('main_tpc_', 'main_') // crude key inference
+        : '(no-ctx)';
+      mutations.push({
+        t: now(),
+        key: ctxKey,
+        n: msgs.length,
+        last: lastSum,
+        delta: `replaceMessages(action=${params.action ?? '-'})  src=${stackTop ?? '-'}`,
+      });
+
+      return origReplace.apply(this, args);
+    };
+  }
+} catch (e: any) {
+  calls.push({ t: now(), name: '_WRAP_ERROR_', error: String(e?.message ?? e) });
+}
+
+// ── 3.5. Mutation log — wrap the TWO ChatStore writers (replaceMessages,
+// internal_dispatchMessage) to record EVERY dbMessagesMap[key] reference
+// change with a one-line "before/after last assistant message" delta. This
+// reveals dispatchMessage-driven collapses that the replaceMessages wrap
+// alone cannot see.
+
+declare global {
+  interface Window {
+    __PROBE_ORIG_DISPATCH_MESSAGE?: any;
+  }
+}
+
+try {
+  const chat = w.__LOBE_STORES?.chat?.();
+  if (chat?.internal_dispatchMessage) {
+    if (!w.__PROBE_ORIG_DISPATCH_MESSAGE)
+      w.__PROBE_ORIG_DISPATCH_MESSAGE = chat.internal_dispatchMessage;
+    const origDispatch = w.__PROBE_ORIG_DISPATCH_MESSAGE;
+    chat.internal_dispatchMessage = origDispatch;
+
+    chat.internal_dispatchMessage = function probeDispatch(this: unknown, payload: any, ctx?: any) {
+      // Snapshot BEFORE — read the would-be target key + last message.
+      const before = (() => {
+        try {
+          const state = w.__LOBE_STORES?.chat?.();
+          if (!state) return null;
+          // Replicate state.internal_getConversationContext logic enough to
+          // resolve a key — but most callers pass operationId on ctx, and
+          // operationId-keyed lookup needs store internals. Easiest: snapshot
+          // ALL keys' last-assistant cLen and compare BEFORE vs AFTER below.
+          const map = state.dbMessagesMap ?? {};
+          const out: Record<string, any> = {};
+          for (const k of Object.keys(map)) {
+            const last = (map[k] ?? []).at(-1);
+            out[k] = last
+              ? {
+                  id: (last.id ?? '').slice(-8),
+                  cLen: (last.content ?? '').length,
+                  rLen: (last.reasoning?.content ?? '').length,
+                  n: map[k].length,
+                }
+              : { n: 0 };
+          }
+          return out;
+        } catch {
+          return null;
+        }
+      })();
+
+      const result = origDispatch.apply(this, [payload, ctx]);
+
+      // Snapshot AFTER — find which key(s) actually changed.
+      try {
+        const state = w.__LOBE_STORES?.chat?.();
+        if (state && before) {
+          const map = state.dbMessagesMap ?? {};
+          for (const k of Object.keys(map)) {
+            const last = (map[k] ?? []).at(-1);
+            const beforeSnap = before[k];
+            const afterSnap = last
+              ? {
+                  id: (last.id ?? '').slice(-8),
+                  cLen: (last.content ?? '').length,
+                  rLen: (last.reasoning?.content ?? '').length,
+                  n: map[k].length,
+                }
+              : { n: 0 };
+            const changed =
+              !beforeSnap ||
+              beforeSnap.n !== afterSnap.n ||
+              beforeSnap.id !== (afterSnap as any).id ||
+              beforeSnap.cLen !== (afterSnap as any).cLen ||
+              beforeSnap.rLen !== (afterSnap as any).rLen;
+            if (!changed) continue;
+            let delta = '';
+            if (beforeSnap?.id !== undefined && beforeSnap.id !== (afterSnap as any).id)
+              delta += `id:${beforeSnap.id}→${(afterSnap as any).id};`;
+            if (
+              beforeSnap?.cLen !== undefined &&
+              (afterSnap as any).cLen !== undefined &&
+              (afterSnap as any).cLen < beforeSnap.cLen
+            )
+              delta += `cLen↓${beforeSnap.cLen}→${(afterSnap as any).cLen};`;
+            if (
+              beforeSnap?.rLen !== undefined &&
+              (afterSnap as any).rLen !== undefined &&
+              (afterSnap as any).rLen < beforeSnap.rLen
+            )
+              delta += `rLen↓${beforeSnap.rLen}→${(afterSnap as any).rLen};`;
+            if (beforeSnap?.n !== undefined && afterSnap.n < beforeSnap.n)
+              delta += `n↓${beforeSnap.n}→${afterSnap.n};`;
+            mutations.push({
+              t: now(),
+              key: k,
+              n: afterSnap.n,
+              last: (afterSnap as any).id ? (afterSnap as any) : undefined,
+              prevLast: beforeSnap?.id ? beforeSnap : undefined,
+              delta: delta || `dispatch:${payload?.type}`,
+            });
+          }
+        }
+      } catch (e: any) {
+        mutations.push({
+          t: now(),
+          key: '_DISPATCH_PROBE_ERROR_',
+          n: -1,
+          delta: String(e?.message ?? e),
+        });
+      }
+      return result;
+    };
+  }
+} catch (e: any) {
+  calls.push({ t: now(), name: '_DISPATCH_WRAP_ERROR_', error: String(e?.message ?? e) });
+}
+
+// ── 4. Periodic per-key timeline snapshots ─────────────────────────
+
+function captureTimeline(): void {
+  try {
+    const c = w.__LOBE_STORES?.chat?.();
+    if (!c) return;
+    const msgsMap = (c.messagesMap ?? {}) as Record<string, any[]>;
+    const dbMap = (c.dbMessagesMap ?? {}) as Record<string, any[]>;
+    const byKey: ProbeTimelineSample['byKey'] = {};
+    for (const k of Object.keys(msgsMap)) {
+      const display = msgsMap[k] ?? [];
+      const db = dbMap[k] ?? [];
+      if (display.length === 0 && db.length === 0) continue;
+      byKey[k] = {
+        n: display.length,
+        dbN: db.length,
+        msgs: summarizeMessages(display),
+      };
+    }
+    const ops = Object.values((c.operations ?? {}) as Record<string, any>);
+    timeline.push({
+      t: now(),
+      activeTopic: ((c.activeTopicId as string | null) ?? '').slice(-10) || null,
+      keys: Object.keys(byKey),
+      byKey,
+      runOps: ops.filter((o: any) => o.status === 'running').length,
+    });
+  } catch (e: any) {
+    timeline.push({
+      t: now(),
+      activeTopic: null,
+      keys: [],
+      byKey: {},
+      runOps: 0,
+      err: e?.message ?? String(e),
+    });
+  }
+}
+captureTimeline();
+if (w.__PROBE_TIMELINE_TIMER) clearInterval(w.__PROBE_TIMELINE_TIMER);
+w.__PROBE_TIMELINE_TIMER = setInterval(captureTimeline, 200);
+
+// ── 5. Tab-switch helpers ──────────────────────────────────────────
+
+function listTopBarTabs(): HTMLElement[] {
+  return Array.from(
+    document.querySelectorAll<HTMLElement>(
+      '[data-insp-path*="TabItem.tsx"][data-contextmenu-trigger]',
+    ),
+  ).filter((t) => t.getBoundingClientRect().top < 30);
+}
+
+w.__listTabs = () =>
+  listTopBarTabs().map((t, i) => ({
+    i,
+    key: t.getAttribute('data-contextmenu-trigger'),
+    active: t.getAttribute('data-active') === 'true',
+    title: (t.innerText ?? '').slice(0, 60),
+  }));
+
+w.__clickTabByKey = (key: string) => {
+  const tab = listTopBarTabs().find((t) => t.getAttribute('data-contextmenu-trigger') === key);
+  if (!tab) return 'not found: ' + key;
+  if (tab.getAttribute('data-active') === 'true') return 'already active: ' + key;
+  tab.click();
+  return 'clicked key=' + key;
+};
+
+w.__PROBE_EVENT = (name: string) => {
+  calls.push({ t: now(), name: 'MARK:' + name });
+};
+
+// `run.ts` wraps the bundle in an IIFE and appends a `return <confirmation>`
+// after the bundle body — agent-browser then prints the confirmation back to
+// the operator. Nothing to do here at the end of the module body.
@@ -0,0 +1,204 @@
+// LobeHub chat streaming time-series probe.
+//
+// Inject into the renderer (via agent-browser eval) to record store + DOM
+// snapshots every 200ms during a streaming session. Designed to surface
+// "UI rolled back to an earlier state" symptoms — especially around
+// gateway-mode tab switches that happen while the assistant is still writing.
+//
+// Usage:
+//   agent-browser --cdp 9222 eval --stdin < probe.js
+//   # ...do test interactions, call window.__PROBE_EVENT('LABEL') to mark moments...
+//   agent-browser --cdp 9222 eval --stdin < probe-dump.js > /tmp/probe.json
+//   node analyze.mjs /tmp/probe.json
+//
+// What it captures per sample:
+//   - activeTopicId
+//   - msgN: top-level messages in chat.messagesMap for this topic
+//   - childN: total assistantGroup.children blocks across all msgs (THIS is
+//     where streaming content actually lives — top-level assistantGroup stays empty)
+//   - cT / rT / toolT: totals across messages AND their children
+//                       (content, reasoning, tool-call count)
+//   - perMsg: per-message breakdown so regressions can be located precisely
+//   - runOps: number of running operations (execServerAgentRuntime etc.)
+//   - domLen: total innerText length of the rendered chat list area
+//   - ind: visible UI indicators (Search pages, Crawled pages, Deeply Thought, Sending)
+//
+// Event markers: window.__PROBE_EVENT('NAME') records {t, name} into
+// __PROBE_EVENTS, used by the analyzer to align state changes with
+// user-driven actions (SENT, AWAY_1, BACK_1, ...).
+
+(function () {
+  if (window.__PROBE_TIMER) clearInterval(window.__PROBE_TIMER);
+  window.__PROBE_SAMPLES = [];
+  window.__PROBE_EVENTS = [];
+  const t0 = Date.now();
+
+  function snapshot() {
+    try {
+      const chat = window.__LOBE_STORES.chat();
+      const topicId = chat.activeTopicId;
+      const idTail = topicId ? topicId.replace('tpc_', '') : null;
+      const keys = Object.keys(chat.messagesMap || {});
+
+      // Collect messages for the active topic. Before a topic is committed,
+      // optimistic messages live under the `<agentScope>_new` key — fall
+      // back to those when no topic is active yet.
+      let msgs = [];
+      if (idTail) {
+        keys.forEach((k) => {
+          if (k.includes(idTail)) msgs = msgs.concat(chat.messagesMap[k] || []);
+        });
+      } else {
+        keys
+          .filter((k) => k.endsWith('_new'))
+          .forEach((k) => {
+            msgs = msgs.concat(chat.messagesMap[k] || []);
+          });
+      }
+
+      // Walk top-level + assistantGroup.children. children carry the actual
+      // streamed content / reasoning / tool calls; the parent assistantGroup
+      // remains a placeholder (cLen=0, rLen=0) for its whole lifetime.
+      let totalContent = 0;
+      let totalReason = 0;
+      let totalTools = 0;
+      let childCount = 0;
+      const perMsg = msgs.map((m) => {
+        const cLen = (m.content || '').length;
+        const rLen = ((m.reasoning && m.reasoning.content) || '').length;
+        const tools = (m.tools || []).length;
+        totalContent += cLen;
+        totalReason += rLen;
+        totalTools += tools;
+
+        const children = m.children || [];
+        let chC = 0;
+        let chR = 0;
+        let chT = 0;
+        children.forEach((c) => {
+          chC += (c.content || '').length;
+          chR += ((c.reasoning && c.reasoning.content) || '').length;
+          chT += (c.tools || []).length;
+        });
+        totalContent += chC;
+        totalReason += chR;
+        totalTools += chT;
+        childCount += children.length;
+
+        return {
+          id: (m.id || '').slice(-8),
+          role: m.role,
+          cLen,
+          rLen,
+          tools,
+          chCount: children.length,
+          chC,
+          chR,
+          chT,
+        };
+      });
+
+      const ops = Object.values(chat.operations || {});
+      const runningOps = ops.filter((o) => o.status === 'running');
+
+      // DOM probe: total rendered text in the chat scroll area (proxy for
+      // "how much is actually visible to the user").
+      const convScroll =
+        document.querySelector(
+          '[data-chat-list], [class*="ChatList"], [class*="ConversationList"]',
+        ) ||
+        document.querySelector('main [class*="scroll"]') ||
+        document.querySelector('main');
+      const domTxt = convScroll ? convScroll.innerText || '' : '';
+
+      const bodyTxt = document.body.innerText || '';
+      const searchMatches = (bodyTxt.match(/Search pages?:|Searched the web/g) || []).length;
+      const crawlMatches = (bodyTxt.match(/Crawl(ed|ing) pages?/g) || []).length;
+
+      window.__PROBE_SAMPLES.push({
+        t: Date.now() - t0,
+        topicId,
+        msgN: msgs.length,
+        childN: childCount,
+        cT: totalContent,
+        rT: totalReason,
+        toolT: totalTools,
+        perMsg,
+        runOps: runningOps.length,
+        runOpTypes: runningOps.map((o) => o.type),
+        domLen: domTxt.length,
+        ind: {
+          search: searchMatches,
+          crawl: crawlMatches,
+          sending: bodyTxt.includes('Sending message'),
+          deeplyThinking: bodyTxt.includes('Deeply Thinking'),
+          deeplyThought: bodyTxt.includes('Deeply Thought'),
+        },
+      });
+    } catch (e) {
+      window.__PROBE_SAMPLES.push({ t: Date.now() - t0, err: e.message });
+    }
+  }
+
+  snapshot();
+  window.__PROBE_TIMER = setInterval(snapshot, 200);
+  window.__PROBE_EVENT = function (name) {
+    window.__PROBE_EVENTS.push({ t: Date.now() - t0, name });
+  };
+
+  // Tab-switch helpers installed alongside the probe.
+  //
+  // The Electron tab bar mounts each tab as a div with data-insp-path
+  // ending in `TabItem.tsx:...`. The active tab is marked with
+  // data-active="true". DO NOT search by innerText — the active tab's text
+  // includes a ` · <agent name>` suffix that produces false matches when
+  // your search string happens to overlap with the agent name.
+  function listTabs() {
+    return Array.from(
+      document.querySelectorAll('[data-insp-path*="TabItem.tsx"][data-contextmenu-trigger]'),
+    ).filter((t) => t.getBoundingClientRect().top < 30);
+  }
+  function tabKey(el) {
+    // Stable for the tab's lifetime; survives focus changes.
+    return el.getAttribute('data-contextmenu-trigger');
+  }
+  function findActiveTab() {
+    return listTabs().find((t) => t.getAttribute('data-active') === 'true') || null;
+  }
+
+  // Click by stable key captured earlier (preferred for round-trips).
+  window.__clickTabByKey = function (key) {
+    const tab = listTabs().find((t) => tabKey(t) === key);
+    if (!tab) return 'not found: key=' + key;
+    if (tab.getAttribute('data-active') === 'true') return 'already active: ' + key;
+    tab.click();
+    return 'clicked key=' + key;
+  };
+
+  // Click by index in the tab strip (0-based, left-to-right).
+  window.__clickTabByIndex = function (i) {
+    const tabs = listTabs();
+    if (i < 0 || i >= tabs.length) return 'index out of range: ' + i + '/' + tabs.length;
+    const t = tabs[i];
+    if (t.getAttribute('data-active') === 'true') return 'already active: i=' + i;
+    t.click();
+    return 'clicked i=' + i + ' key=' + tabKey(t);
+  };
+
+  // Snapshot all tabs in order: [{key, active, title (first 60 chars of innerText)}]
+  window.__listTabs = function () {
+    return listTabs().map((t, i) => ({
+      i,
+      key: tabKey(t),
+      active: t.getAttribute('data-active') === 'true',
+      title: (t.innerText || '').slice(0, 60),
+    }));
+  };
+
+  window.__activeTabKey = function () {
+    const a = findActiveTab();
+    return a ? tabKey(a) : null;
+  };
+
+  return 'probe installed';
+})();
@@ -0,0 +1,211 @@
+// CLI for the agent-gateway probe.
+//
+// Bundles the TS probes with esbuild, pipes them into `agent-browser eval`,
+// and persists dumps under `.agent-gateway/` (gitignored) for later use as
+// streaming-replay test fixtures.
+//
+// Commands:
+//   bun run .agents/skills/local-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]
+//       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]
+//       Run analyze-events.ts on the dump. `path` defaults to the most
+//       recently modified file in `.agent-gateway/`.
+//
+// Optional flags:
+//   --cdp <port>     CDP port (default 9222)
+//   --browser <bin>  agent-browser binary (default 'agent-browser')
+
+import { spawn } from 'node:child_process';
+import { mkdirSync, readdirSync, statSync, writeFileSync } from 'node:fs';
+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
+const PROJECT_ROOT = path.resolve(SCRIPT_DIR, '../../../../..');
+const DUMP_DIR = path.join(PROJECT_ROOT, '.agent-gateway');
+
+interface Flags {
+  browser: string;
+  cdp: string;
+  positional: string[];
+}
+
+function parseFlags(argv: string[]): Flags {
+  const out: Flags = { cdp: '9222', browser: 'agent-browser', positional: [] };
+  for (let i = 0; i < argv.length; i++) {
+    const a = argv[i];
+    if (a === '--cdp') out.cdp = argv[++i] ?? out.cdp;
+    else if (a === '--browser') out.browser = argv[++i] ?? out.browser;
+    else out.positional.push(a);
+  }
+  return out;
+}
+
+async function bundle(entry: string): Promise<string> {
+  // Bun.build is built into the Bun runtime — no external dep needed.
+  const r = await Bun.build({
+    entrypoints: [path.join(SCRIPT_DIR, entry)],
+    target: 'browser',
+    format: 'esm',
+    minify: false,
+  });
+  if (!r.success) {
+    const msgs = r.logs.map((l) => `${l.level}: ${l.message}`).join('\n');
+    throw new Error(`bundle failed for ${entry}:\n${msgs}`);
+  }
+  return await r.outputs[0].text();
+}
+
+function wrapIife(body: string, returnExpr: string): string {
+  // Wrap as an IIFE that swallows the bundled top-level (top-level `const`
+  // declarations get scoped to the IIFE, so re-injection doesn't conflict)
+  // and returns the configured expression — which `agent-browser eval`
+  // captures and prints to stdout.
+  return `(() => {\n${body}\n;return ${returnExpr};\n})()`;
+}
+
+function runAgentBrowserEval(flags: Flags, script: string): Promise<string> {
+  return new Promise((resolveP, rejectP) => {
+    const child = spawn(flags.browser, ['--cdp', flags.cdp, 'eval', '--stdin'], {
+      stdio: ['pipe', 'pipe', 'inherit'],
+    });
+    let stdout = '';
+    child.stdout.on('data', (chunk: Buffer) => {
+      stdout += chunk.toString('utf8');
+    });
+    child.on('error', rejectP);
+    child.on('close', (code) => {
+      if (code === 0) resolveP(stdout);
+      else rejectP(new Error(`agent-browser exited ${code}`));
+    });
+    child.stdin.write(script);
+    child.stdin.end();
+  });
+}
+
+// agent-browser prints eval results as JSON (string values are quoted).
+function unquoteAgentBrowserResult(raw: string): string {
+  const trimmed = raw.trim();
+  if (trimmed.startsWith('"') && trimmed.endsWith('"')) {
+    try {
+      return JSON.parse(trimmed) as string;
+    } catch {
+      /* fall through */
+    }
+  }
+  return trimmed;
+}
+
+function isoStamp(): string {
+  const d = new Date();
+  const yyyy = d.getFullYear();
+  const mm = String(d.getMonth() + 1).padStart(2, '0');
+  const dd = String(d.getDate()).padStart(2, '0');
+  const hh = String(d.getHours()).padStart(2, '0');
+  const mi = String(d.getMinutes()).padStart(2, '0');
+  const ss = String(d.getSeconds()).padStart(2, '0');
+  return `${yyyy}${mm}${dd}-${hh}${mi}${ss}`;
+}
+
+function ensureDumpDir(): void {
+  mkdirSync(DUMP_DIR, { recursive: true });
+}
+
+function latestDump(): string | null {
+  ensureDumpDir();
+  const entries = readdirSync(DUMP_DIR)
+    .filter((f) => f.endsWith('.json'))
+    .map((f) => ({ f, mtime: statSync(path.join(DUMP_DIR, f)).mtimeMs }))
+    .sort((a, b) => b.mtime - a.mtime);
+  return entries[0] ? path.join(DUMP_DIR, entries[0].f) : null;
+}
+
+// ── Commands ────────────────────────────────────────────────────────
+
+async function cmdInstall(flags: Flags): Promise<void> {
+  const body = await bundle('probe-events.ts');
+  const installMsg = JSON.stringify(
+    'events probe installed: WebSocket+fetch interception. ' +
+      'WS captures operationId= sockets (gateway), fetch captures /api/agent/stream (direct).',
+  );
+  const script = wrapIife(body, installMsg);
+  const out = await runAgentBrowserEval(flags, script);
+  console.log(unquoteAgentBrowserResult(out));
+}
+
+async function cmdDump(flags: Flags): Promise<void> {
+  const name = flags.positional[1] ?? 'dump';
+  const body = await bundle('probe-dump.ts');
+  const script = wrapIife(body, 'window.__PROBE_LAST_DUMP_JSON');
+  const raw = await runAgentBrowserEval(flags, script);
+  const json = unquoteAgentBrowserResult(raw);
+  ensureDumpDir();
+  const filename = `${name}-${isoStamp()}.json`;
+  const dumpPath = path.join(DUMP_DIR, filename);
+  writeFileSync(dumpPath, json, 'utf8');
+  // Validate by parsing the meta header so we error early on bad capture
+  try {
+    const parsed = JSON.parse(json) as {
+      meta?: { eventCount?: number; callCount?: number; sampleCount?: number };
+    };
+    const meta = parsed.meta ?? {};
+    console.log(
+      `wrote ${dumpPath}  (${json.length} bytes  events=${meta.eventCount ?? '?'}  ` +
+        `calls=${meta.callCount ?? '?'}  samples=${meta.sampleCount ?? '?'})`,
+    );
+  } catch {
+    console.log(`wrote ${dumpPath}  (${json.length} bytes — JSON.parse failed; see file)`);
+  }
+}
+
+async function cmdAnalyze(flags: Flags): Promise<void> {
+  const target = flags.positional[1] ?? latestDump();
+  if (!target) {
+    console.error('no dump file found. run `dump` first or pass a path.');
+    process.exit(1);
+  }
+  const child = spawn('bun', ['run', path.join(SCRIPT_DIR, 'analyze-events.ts'), target], {
+    stdio: 'inherit',
+  });
+  await new Promise<void>((resolveP, rejectP) => {
+    child.on('error', rejectP);
+    child.on('close', (code) => (code === 0 ? resolveP() : rejectP(new Error(`exit ${code}`))));
+  });
+}
+
+// ── Entry point ─────────────────────────────────────────────────────
+
+const flags = parseFlags(process.argv.slice(2));
+const cmd = flags.positional[0];
+
+const usage = `usage:
+  bun run run.ts install [--cdp 9222]
+  bun run run.ts dump [name] [--cdp 9222]
+  bun run run.ts analyze [path]
+`;
+
+if (!cmd) {
+  console.error(usage);
+  process.exit(1);
+}
+
+try {
+  if (cmd === 'install') await cmdInstall(flags);
+  else if (cmd === 'dump') await cmdDump(flags);
+  else if (cmd === 'analyze') await cmdAnalyze(flags);
+  else {
+    console.error(`unknown command: ${cmd}\n\n${usage}`);
+    process.exit(1);
+  }
+} catch (e: any) {
+  console.error(e?.stack ?? e);
+  process.exit(1);
+}
@@ -0,0 +1,72 @@
+// Run N round-trip tab switches with event markers timed against the probe.
+//
+//   agent-browser --cdp 9222 eval --stdin < tab-switch.js
+//
+// Captures the currently-active tab as the BACK target and the rightmost
+// inactive tab as the AWAY target. Both are addressed by their stable
+// data-contextmenu-trigger key (NOT by visible title — the active tab's
+// innerText embeds a ` · <agent name>` suffix that breaks text matching).
+//
+// Fires the loop in the background and returns immediately so the
+// agent-browser eval doesn't have to await the full ROUND_TRIPS × DWELL_MS
+// duration. Wait on the `SWITCH_LOOP_DONE` event before dumping.
+//
+// Refuses to launch if a previous loop is still in flight.
+//
+// Requires probe.js to have been installed first (provides
+// window.__PROBE_EVENT / __listTabs / __clickTabByKey / __activeTabKey).
+
+(function () {
+  const ROUND_TRIPS = 4;
+  const DWELL_MS = 10_000;
+
+  if (!window.__PROBE_EVENT || !window.__listTabs || !window.__clickTabByKey) {
+    return 'probe not installed — eval probe.js first';
+  }
+  if (window.__SWITCH_LOOP_RUNNING) {
+    return 'switch loop already running — wait for SWITCH_LOOP_DONE first';
+  }
+
+  const tabs = window.__listTabs();
+  const activeTab = tabs.find((t) => t.active);
+  if (!activeTab) return 'no active tab — abort';
+
+  // Pick the first inactive tab as AWAY target. With multiple inactive tabs
+  // you'll usually want the one that's stable across the test — feel free
+  // to swap to tabs[tabs.length-1] if you want the rightmost.
+  const inactives = tabs.filter((t) => !t.active);
+  if (inactives.length === 0) return 'no inactive tab to switch to — abort';
+  const awayTab = inactives.at(-1); // rightmost inactive
+
+  const BACK_KEY = activeTab.key;
+  const AWAY_KEY = awayTab.key;
+
+  window.__SWITCH_LOOP_RUNNING = true;
+  window.__PROBE_EVENT('SWITCH_LOOP_CONFIG:back=' + BACK_KEY + ',away=' + AWAY_KEY);
+
+  (async function () {
+    function sleep(ms) {
+      return new Promise((r) => setTimeout(r, ms));
+    }
+
+    try {
+      window.__PROBE_EVENT('SWITCH_LOOP_START');
+      for (let i = 1; i <= ROUND_TRIPS; i++) {
+        window.__PROBE_EVENT('AWAY_' + i);
+        const awayResult = window.__clickTabByKey(AWAY_KEY);
+        window.__PROBE_EVENT('AWAY_' + i + '_RES:' + awayResult.slice(0, 50));
+        await sleep(DWELL_MS);
+
+        window.__PROBE_EVENT('BACK_' + i);
+        const backResult = window.__clickTabByKey(BACK_KEY);
+        window.__PROBE_EVENT('BACK_' + i + '_RES:' + backResult.slice(0, 50));
+        await sleep(DWELL_MS);
+      }
+      window.__PROBE_EVENT('SWITCH_LOOP_DONE');
+    } finally {
+      window.__SWITCH_LOOP_RUNNING = false;
+    }
+  })();
+
+  return 'switch loop kicked off (BACK=' + BACK_KEY + ', AWAY=' + AWAY_KEY + ')';
+})();
@@ -0,0 +1,113 @@
+// Shared types between the in-browser probe and the Node-side analyzer.
+// Kept tiny on purpose — anything the analyzer can re-derive is left off.
+
+export interface ProbeStreamEvent {
+  /** Summarized payload — long strings truncated, arrays printed as Array(N) */
+  data?: Record<string, unknown>;
+  /** Keys present on the event's `data` payload — useful at a glance */
+  dataKeys?: string[];
+  /** ServerMessage.id — gateway WS frames carry an event-id we may resume from */
+  eventId?: string | null;
+  message?: string;
+  /** Last 10 chars of the operationId (full id is excessively long) */
+  opIdTail: string;
+  raw?: string;
+  /** Raw frame byte length, when applicable */
+  rawLen?: number;
+  /** For non-agent_event server frames (auth_success, heartbeat_ack, …) */
+  serverType?: string;
+  sseEvent?: string;
+  status?: number;
+  stepIndex?: number;
+  /** Milliseconds since the probe's t0 (install time). */
+  t: number;
+  /** 'ws' for gateway WebSocket frames, 'sse' for direct /api/agent/stream */
+  transport: 'ws' | 'sse';
+  /** Either the AgentStreamEvent.type, or a probe sentinel like `_WS_OPEN_` */
+  type: string;
+  url?: string;
+}
+
+export interface ProbeActionCall {
+  args?: {
+    count?: number;
+    context?: unknown;
+    params?: unknown;
+  };
+  error?: string;
+  /** `replaceMessages` / `refreshMessages` / `MARK:<label>` / `_WRAP_ERROR_` */
+  name: string;
+  stack?: string;
+  t: number;
+}
+
+export interface ProbeMessageSummary {
+  /** children.length */
+  chN: number;
+  /** content.length */
+  cLen: number;
+  /** Last 8 chars of the message id */
+  id: string;
+  /** reasoning.content.length */
+  rLen: number;
+  role: string;
+  /** tools.length */
+  tools: number;
+}
+
+export interface ProbeTimelineSample {
+  /** Last 10 chars of activeTopicId, or null */
+  activeTopic: string | null;
+  /** Per-key breakdown: display count, db count, message summaries */
+  byKey: Record<
+    string,
+    {
+      n: number;
+      dbN: number;
+      msgs: ProbeMessageSummary[];
+    }
+  >;
+  err?: string;
+  /** All messagesMap keys that have content at this moment */
+  keys: string[];
+  /** Number of operations in 'running' status */
+  runOps: number;
+  t: number;
+}
+
+export interface ProbeDumpMeta {
+  callCount: number;
+  /** Date.now() at dump call */
+  collectedAt: number;
+  eventCount: number;
+  sampleCount: number;
+  /** Date.now() at probe install */
+  t0: number;
+}
+
+export interface ProbeDump {
+  actionCalls: ProbeActionCall[];
+  meta: ProbeDumpMeta;
+  streamEvents: ProbeStreamEvent[];
+  timeline: ProbeTimelineSample[];
+}
+
+/**
+ * Globals the probe attaches to `window`. Keeps `as any` casts at the boundary
+ * instead of sprinkling them through the probe body.
+ */
+declare global {
+  interface Window {
+    __clickTabByKey?: (key: string) => string;
+    __listTabs?: () => Array<{ i: number; key: string | null; active: boolean; title: string }>;
+    __LOBE_STORES?: Record<string, () => any>;
+    __PROBE_ACTION_CALLS?: ProbeActionCall[];
+    __PROBE_EVENT?: (label: string) => void;
+    __PROBE_MSG_TIMELINE?: ProbeTimelineSample[];
+    __PROBE_ORIG_FETCH?: typeof fetch;
+    __PROBE_ORIG_WEBSOCKET?: typeof WebSocket;
+    __PROBE_STREAM_EVENTS?: ProbeStreamEvent[];
+    __PROBE_T0?: number;
+    __PROBE_TIMELINE_TIMER?: ReturnType<typeof setInterval> | null;
+  }
+}
@@ -1,6 +1,6 @@
 ---
 name: pr
-description: "Create a PR for the current branch. Use when the user asks to create a pull request, submit PR, or says 'pr'."
+description: "Create a PR for the current branch (targets `canary` by default), including splitting one cross-layer branch into ordered stacked PRs so a lower layer (db / shared package / server TRPC) merges before its callers (desktop / CLI / UI). Use when the user asks to create / submit a PR, or to split a branch because clients call a server contract that isn't on the trunk yet. Triggers on 'pr', 'create pr', 'submit pr', 'open a PR', 'pull request', 'split this PR', 'stacked PR', 'backend should merge first', '提 PR', '提个 PR', '新建 PR', '拆 PR', '后端先合', '分层合并'."
 user-invocable: true
 ---

@@ -71,3 +71,82 @@ Use `.github/PULL_REQUEST_TEMPLATE.md` as the body structure. Key sections:

 - **Language**: All PR content must be in English
 - If a PR already exists for the branch, inform the user instead of creating a duplicate
+
+---
+
+# Stacked PRs (cross-layer feature)
+
+The steps above create **one** PR for the current branch. When a single branch lands across layers — `packages/database` schema/model → a shared `packages/*` lib → `src/server` TRPC → `apps/desktop` + `apps/cli` callers → `src/features` UI — shipping it as one PR can't merge safely: the clients call an endpoint that doesn't exist on the trunk until the same PR merges, so any partial/rollback or independent review breaks. Split it into **ordered PRs**, lower layer first.
+
+## The ordering rule
+
+A PR may only merge **after** every layer it calls is already on the trunk.
+
+- The **server contract** (new TRPC procedure, changed return shape, new table/model) merges first.
+- The **callers** (desktop, CLI, UI) merge after — they invoke that contract.
+- Tie-break with one question: _"if this merged alone to `canary` right now, would it build and behave?"_ If no, it belongs in a later PR.
+
+## Which file goes in which PR
+
+The non-obvious calls:
+
+- **Frontend that adapts to a contract change goes WITH the server PR.** If you widen a TRPC return shape (e.g. `listDevices` now returns `platform: string | null`), the component consuming it must change in the _same_ PR — otherwise the server PR breaks the build on its own. Contract + its in-repo consumers ship together.
+- **A new shared package goes with its consumer**, not the server, unless the server imports it too. A `@lobechat/*` package imported only by desktop/CLI ships in the client PR. Don't carry an unused package in the lower PR.
+- **Workspace dep declarations** (`package.json` `workspace:*`, `pnpm-workspace.yaml`) travel with the code that imports the package.
+
+## The git recipe — split an existing full branch
+
+Starting point: one branch (`feat/x`) with a single commit `<FULL>` containing everything, already pushed (so it's also safe on the remote).
+
+```bash
+# 1. Safety nets — make the full work unloseable before rewriting anything
+git branch backup/x-full <FULL>          # local ref to the full commit
+git branch feat/x-clients <FULL>         # the higher-layer branch starts here
+
+# 2. Rewrite the lower-layer branch to lower-layer files only
+git checkout feat/x                      # this becomes the SERVER PR
+git reset --hard origin/canary
+git checkout <FULL> -- <server/db files…>   # stages just those paths
+git commit -m "✨ feat(...): <server half>"
+git push --force-with-lease origin feat/x   # never --force; never push to canary
+
+# 3. Build the higher-layer branch STACKED on the lower branch
+git checkout feat/x-clients
+git reset --hard feat/x                  # base = the just-rewritten server HEAD
+git checkout backup/x-full -- <client/ui files…>   # only the remaining paths
+git commit -m "✨ feat(...): <client half>"
+git push -u origin feat/x-clients
+```
+
+Then open the higher PR **based on the lower branch**, not the trunk:
+
+```bash
+gh pr create --base feat/x --head feat/x-clients --title "…" --body "…"
+```
+
+`--base feat/x` keeps the diff client-only (no server files leak in) and makes it physically impossible to merge the clients before the server. **After the server PR merges to `canary`, retarget the client PR's base to `canary`** (GitHub usually auto-retargets when the base branch merges; note it in the PR body so a human confirms).
+
+## Verify the dependency actually holds
+
+The whole point is the higher layer needs the lower one. Prove it: on the stacked higher branch, type-check the caller and confirm the symbol the lower layer introduced resolves.
+
+```bash
+cd apps/cli && bun run type-check 2>&1 | grep -iE "connect\.ts|device\.register"
+# empty (re: your change) = the stacked base supplies device.register ✓
+```
+
+Filter to your touched files — this repo's standalone type-check emits pre-existing env noise (`__ELECTRON__`, `@/types/llm`, unbuilt `@lobechat/types`) that isn't yours.
+
+## PR + Linear bookkeeping
+
+- **Each PR closes only its own layer's issues.** Server PR: `Closes LOBE-<server>`. Client PR: `Closes LOBE-<pkg> / <desktop> / <cli>`. Don't let one PR's body claim another layer's issue.
+- Both PRs are `Part of LOBE-<parent>`.
+- On PR creation, move each closed sub-issue to **In Review** (not Done) and add a completion comment — see the `linear` skill.
+
+## Gotchas
+
+- **Never push to `canary`.** A split branch cut with `git checkout -b feat/x origin/canary` _tracks_ `origin/canary`, so a bare `git push` targets canary. Always `git push origin feat/x` with the explicit branch name.
+- **`--force-with-lease`, not `--force`** when rewriting the lower branch — it aborts if the remote moved under you.
+- **Back up before `reset --hard`.** Step 1's `backup/x-full` + the pushed remote branch mean the full commit is referenced by ≥3 refs before you rewrite anything. Verify with `git branch --contains <FULL>`.
+- **Lockfiles:** this monorepo commits no root `pnpm-lock.yaml`, so a new `workspace:*` dep needs no lockfile churn. In a repo that _does_ commit one, regenerate it on each branch after the split.
+- **Don't over-split.** Two PRs (contract / callers) is usually enough. A UI page that only reads an existing endpoint can be its own later PR, but don't fragment a single layer across PRs for its own sake.
@@ -1,6 +1,6 @@
 ---
 name: project-overview
-description: Complete project architecture and structure guide. Use when exploring the codebase, understanding project organization, finding files, or needing comprehensive architectural context. Triggers on architecture questions, directory navigation, or project overview needs.
+description: "LobeHub open-source monorepo architecture map — flat `apps/` + `packages/@lobechat/*` + `src/` layout, per-layer location table, and `src/business/` stubs that the cloud repo overrides. Use when exploring an unfamiliar part of the codebase, locating where a layer lives (store / service / router / schema / etc.), or onboarding to the monorepo. Triggers on 'where does X live', 'project structure', 'monorepo layout', `src/business/` stub, 'architecture overview', '项目结构', '架构总览'."
 user-invocable: false
 ---

@@ -13,11 +13,12 @@ user-invocable: false
 ## Project Description

 Open-source, modern-design AI Agent Workspace: **LobeHub** (previously LobeChat).
+This repo is the **open-source root** (`github.com/lobehub/lobehub`, package `@lobehub/lobehub`).

 **Supported platforms:**

 - Web desktop/mobile
- Desktop (Electron)
+- Desktop (Electron) — `apps/desktop`
 - Mobile app (React Native) — **separate repo, already launched** (not in this monorepo)

 **Logo emoji:** 🤯
@@ -47,30 +48,28 @@ Open-source, modern-design AI Agent Workspace: **LobeHub** (previously LobeChat)

 ## Monorepo Layout

-This is a monorepo extending the open-source `lobehub` submodule. Two repos:
-
- **cloud repo root** — `src/` and `packages/business/` (`config`, `const`, `model-runtime`) hold cloud-only SaaS code that overrides/extends the submodule. See `AGENTS.md` for the override mechanism.
- **`lobehub/` submodule** — the open-source product core.
-
-### `lobehub/` submodule — key directories
+Flat layout — `apps/`, `packages/`, and `src/` all sit at the repo root. No
+git submodules.

 ```
-lobehub/
+(repo root)
 ├── apps/
-│   ├── cli/                 # LobeHub CLI
-│   ├── desktop/             # Electron desktop app
-│   └── device-gateway/      # Device gateway service
-├── docs/                    # changelog, development, self-hosting, usage
-├── locales/                 # en-US, zh-CN, ...
-├── packages/                # ~80 @lobechat/* workspace packages — `ls` for the full set. Key ones:
-│   ├── agent-runtime/        # Agent runtime
+│   ├── cli/                  # LobeHub CLI
+│   ├── desktop/              # Electron desktop app
+│   └── device-gateway/       # Device gateway service
+├── docs/                     # changelog, development, self-hosting, usage
+├── locales/                  # en-US, zh-CN, ...
+├── packages/                 # ~80 @lobechat/* workspace packages — `ls` for the full set. Key ones:
+│   ├── agent-runtime/        # Agent runtime core
 │   ├── agent-signal/         # Agent Signal pipeline
-│   ├── builtin-tool-*/       # Builtin tool packages
-│   ├── builtin-tools/        # Builtin tool registries
+│   ├── agent-tracing/        # Tracing / snapshots
+│   ├── builtin-tool-*/       # Per-tool packages (calculator, web-browsing, claude-code, ...)
+│   ├── builtin-tools/        # Central registries that compose builtin-tool-*
 │   ├── context-engine/
 │   ├── database/             # src/{models,schemas,repositories}
 │   ├── model-bank/           # Model definitions & provider cards
 │   ├── model-runtime/        # src/{core,providers}
+│   ├── business/             # Open-source stubs (config, const, model-bank, model-runtime) — overridden by cloud
 │   ├── types/
 │   └── utils/
 └── src/
@@ -83,55 +82,54 @@ lobehub/
    ├── spa/                  # SPA entries + router config
    │   ├── entry.{web,mobile,desktop,popup}.tsx
    │   └── router/
-    ├── business/             # Open-source stubs (~50) overridden by cloud src/business/
+    ├── business/             # Open-source stubs (client/server) — cloud repo provides real impls
    ├── features/             # Domain business components
-    ├── store/                # ~28 zustand stores — `ls` for the full set
-    ├── server/               # featureFlags, globalConfig, modules, routers, services
+    ├── store/                # ~30 zustand stores — `ls` for the full set
+    ├── server/               # featureFlags, globalConfig, modules, routers, services, workflows, agent-hono
    └── ...                   # components, hooks, layout, libs, locales, services, types, utils
 ```

-### cloud repo — key directories
-
-```
-(cloud root)
-├── packages/business/        # Cloud overrides: config, const, model-runtime
-├── src/
-│   ├── business/             # Cloud impls of submodule stubs (client/server/locales)
-│   ├── routes/               # Cloud-only route groups: (cloud)/, embed/
-│   ├── store/                # Cloud-only stores (e.g. subscription/)
-│   ├── server/               # Cloud routers & services (billing, budget, risk control...)
-│   └── app/(backend)/cron/   # Vercel cron routes (schedules declared in root vercel.ts)
-└── vercel.ts                 # Cron schedule declarations
-```
-
-> File search rule: a path like `@/store/x` resolves cloud `src/store/x` first, then
-> `lobehub/packages/store/src/x`, then `lobehub/src/store/x`. Cloud override wins.
-
 ## Architecture Map

-| Layer            | Location                                             |
-| ---------------- | ---------------------------------------------------- |
-| UI Components    | `src/components`, `src/features`                     |
-| SPA Pages        | `src/routes/`                                        |
-| React Router     | `src/spa/router/`                                    |
-| Global Providers | `src/layout`                                         |
-| Zustand Stores   | `src/store`                                          |
-| Client Services  | `src/services/`                                      |
-| REST API         | `src/app/(backend)/webapi`                           |
-| tRPC Routers     | `src/server/routers/{async\|lambda\|mobile\|tools}`  |
-| Server Services  | `src/server/services` (can access DB)                |
-| Server Modules   | `src/server/modules` (no DB access)                  |
-| Feature Flags    | `src/server/featureFlags`                            |
-| Global Config    | `src/server/globalConfig`                            |
-| DB Schema        | `packages/database/src/schemas`                      |
-| DB Model         | `packages/database/src/models`                       |
-| DB Repository    | `packages/database/src/repositories`                 |
-| Third-party      | `src/libs` (analytics, oidc, etc.)                   |
-| Builtin Tools    | `src/tools`, `packages/builtin-tool-*`               |
-| Cloud-only       | `src/business/*`, `packages/business/*` (cloud repo) |
+| Layer            | Location                                            |
+| ---------------- | --------------------------------------------------- |
+| UI Components    | `src/components`, `src/features`                    |
+| SPA Pages        | `src/routes/`                                       |
+| React Router     | `src/spa/router/`                                   |
+| Global Providers | `src/layout`                                        |
+| Zustand Stores   | `src/store`                                         |
+| Client Services  | `src/services/`                                     |
+| REST API         | `src/app/(backend)/webapi`                          |
+| tRPC Routers     | `src/server/routers/{async\|lambda\|mobile\|tools}` |
+| Server Services  | `src/server/services` (can access DB)               |
+| Server Modules   | `src/server/modules` (no DB access)                 |
+| Feature Flags    | `src/server/featureFlags`                           |
+| Global Config    | `src/server/globalConfig`                           |
+| DB Schema        | `packages/database/src/schemas`                     |
+| DB Model         | `packages/database/src/models`                      |
+| DB Repository    | `packages/database/src/repositories`                |
+| Third-party      | `src/libs` (analytics, oidc, etc.)                  |
+| Builtin Tools    | `packages/builtin-tool-*`, `packages/builtin-tools` |
+| Open-source stub | `src/business/*`, `packages/business/*` (this repo) |

 ## Data Flow

 ```
 React UI → Store Actions → Client Service → TRPC Lambda → Server Services → DB Model → PostgreSQL
 ```
+
+## Note: Relationship to the Cloud Repo
+
+This open-source repo is consumed by a **separate, private cloud (SaaS) repo**
+as a git submodule mounted at `lobehub/`. The cloud repo provides:
+
+- **`src/business/{client,server}`** and **`packages/business/*`** implementations
+  that override the stubs shipped here.
+- Cloud-only routes (e.g. `(cloud)/`, `embed/`), cloud-only stores (e.g.
+  `subscription/`), cloud-only TRPC routers (billing, budget, risk control, …),
+  and Vercel cron routes under `src/app/(backend)/cron/`.
+- File-resolution order in cloud: `@/store/x` → cloud `src/store/x` first, then
+  `lobehub/packages/store/src/x`, then `lobehub/src/store/x`. **Cloud override wins.**
+
+When working in this repo alone, ignore the cloud layer — the stubs in
+`src/business/` and `packages/business/` are the source of truth here.
@@ -1,6 +1,6 @@
 ---
 name: react
-description: 'Use when writing or editing any `.tsx` under `src/**`. Triggers: createStaticStyles, createStyles, cssVar, antd-style, Flexbox, Center, Select, Modal, Drawer, Button, Tooltip, DropdownMenu, Popover, Switch, ScrollArea, Link, useNavigate, react-router-dom, next/link, desktopRouter, componentMap.desktop, .desktop.tsx, new component, new page, edit layout, add styles, zustand selector, @lobehub/ui, antd import.'
+description: "LobeHub React component conventions — base-ui (`@lobehub/ui/base-ui`) first for headless primitives (Select, Modal, DropdownMenu, ContextMenu, Popover, ScrollArea, Switch, Toast, FloatingSheet), then `@lobehub/ui` root, antd as last resort; styling via `antd-style` `createStaticStyles` + `cssVar.*` (zero-runtime preferred over `createStyles` + `token`); routing via `react-router-dom` (not `next/link`). Use when writing or editing any `.tsx` under `src/**`. Triggers on `createStaticStyles`, `createStyles`, `cssVar`, `antd-style`, `Flexbox`, `Center`, `Select`, `Modal`, `Drawer`, `Button`, `Tooltip`, `DropdownMenu`, `ContextMenu`, `Popover`, `Switch`, `ScrollArea`, `Toast`, `FloatingSheet`, `Link`, `useNavigate`, `react-router-dom`, `next/link`, `desktopRouter`, `componentMap.desktop`, `.desktop.tsx`, `base-ui`, `@lobehub/ui/base-ui`, 'new component', 'new page', 'edit layout', 'add styles', 'zustand selector', '@lobehub/ui', 'antd import'."
 user-invocable: false
 ---

@@ -17,22 +17,41 @@ user-invocable: false
 ## Component Priority

 1. **`src/components`** — project-specific reusable components
-2. **`@lobehub/ui/base-ui`** — headless primitives (Select, Modal, DropdownMenu, Popover, Switch, ScrollArea…)
-3. **`@lobehub/ui`** — higher-level components (ActionIcon, Markdown, DragPage…)
-4. **Custom implementation** — last resort; never reach for antd directly
+2. **`@lobehub/ui/base-ui`** — headless primitives. **If the component lives here, use it. Do NOT import the same-named root export.**
+3. **`@lobehub/ui`** — higher-level / antd-wrapping components (only when no base-ui equivalent)
+4. **antd** — only when neither base-ui nor `@lobehub/ui` root provides it
+5. **Custom implementation** — true last resort

-If unsure about available components, search existing code or check `node_modules/@lobehub/ui/es/index.mjs`.
+If unsure about available components, search existing code or check `node_modules/@lobehub/ui/es/index.mjs` and `node_modules/@lobehub/ui/es/base-ui/`.

-### Common @lobehub/ui Components
+### `@lobehub/ui/base-ui` — always prefer for these

-| Category     | Components                                                                      |
-| ------------ | ------------------------------------------------------------------------------- |
-| General      | ActionIcon, ActionIconGroup, Block, Button, Icon                                |
-| Data Display | Avatar, Collapse, Empty, Highlighter, Markdown, Tag, Tooltip                    |
-| Data Entry   | CodeEditor, CopyButton, EditableText, Form, FormModal, Input, SearchBar, Select |
-| Feedback     | Alert, Drawer, Modal                                                            |
-| Layout       | Center, DraggablePanel, Flexbox, Grid, Header, MaskShadow                       |
-| Navigation   | Burger, Dropdown, Menu, SideNav, Tabs                                           |
+| Component                                  | Import                                                                                                  |
+| ------------------------------------------ | ------------------------------------------------------------------------------------------------------- |
+| `Select` (+ `SelectProps`, `SelectOption`) | `import { Select } from '@lobehub/ui/base-ui';`                                                         |
+| `Modal` (imperative API)                   | `import { createModal, confirmModal, useModalContext, type ModalInstance } from '@lobehub/ui/base-ui';` |
+| `DropdownMenu`                             | `import { DropdownMenu } from '@lobehub/ui/base-ui';`                                                   |
+| `ContextMenu`                              | `import { ContextMenu } from '@lobehub/ui/base-ui';`                                                    |
+| `Popover`                                  | `import { Popover } from '@lobehub/ui/base-ui';`                                                        |
+| `ScrollArea`                               | `import { ScrollArea } from '@lobehub/ui/base-ui';`                                                     |
+| `Switch`                                   | `import { Switch } from '@lobehub/ui/base-ui';`                                                         |
+| `Toast`                                    | `import { Toast } from '@lobehub/ui/base-ui';`                                                          |
+| `FloatingSheet`                            | `import { FloatingSheet } from '@lobehub/ui/base-ui';`                                                  |
+
+For Modal specifically, see the dedicated **modal** skill — use the imperative `createModal({ content: … })` pattern over the legacy `<Modal open … />` declarative pattern. base-ui has its own `ModalHost` already mounted in `SPAGlobalProvider`.
+
+> Common slip: `import { Select } from '@lobehub/ui'` looks fine but it's the antd-backed Select. Use base-ui Select. Same for `Modal`, `DropdownMenu`, etc.
+
+### `@lobehub/ui` root — use when base-ui has no equivalent
+
+| Category     | Components                                                                            |
+| ------------ | ------------------------------------------------------------------------------------- |
+| General      | ActionIcon, ActionIconGroup, Block, Button, Icon                                      |
+| Data Display | Avatar, Collapse, Empty, Highlighter, Markdown, Tag, Tooltip                          |
+| Data Entry   | CodeEditor, CopyButton, EditableText, Form, Input, InputPassword, SearchBar, TextArea |
+| Feedback     | Alert, Drawer                                                                         |
+| Layout       | Center, DraggablePanel, Flexbox, Grid, Header, MaskShadow                             |
+| Navigation   | Burger, Menu, SideNav, Tabs                                                           |

 ## Layout

@@ -85,12 +104,15 @@ errorElement: <ErrorBoundary />;

 ## Common Mistakes

-| Mistake                                                           | Fix                                                               |
-| ----------------------------------------------------------------- | ----------------------------------------------------------------- |
-| Using `next/link` in SPA                                          | Use `react-router-dom` `Link`                                     |
-| Using antd directly                                               | Use `@lobehub/ui/base-ui` first, then `@lobehub/ui`               |
-| `createStyles` for static styles                                  | Use `createStaticStyles` + `cssVar`                               |
-| Editing only `desktopRouter.config.tsx`                           | Must edit both `.tsx` and `.desktop.tsx`                          |
-| Using `margin` for flex spacing                                   | Use `gap` prop on Flexbox                                         |
-| Accessing zustand store without selector                          | Use selectors to access store data (see zustand skill)            |
-| Text or icon-text actions built with `Flexbox`/`Text` + `onClick` | Use `Button type={'text'} size={'small'}` with `icon` when needed |
+| Mistake                                                            | Fix                                                                         |
+| ------------------------------------------------------------------ | --------------------------------------------------------------------------- |
+| Using `next/link` in SPA                                           | Use `react-router-dom` `Link`                                               |
+| Using antd directly                                                | Use `@lobehub/ui/base-ui` first, then `@lobehub/ui`                         |
+| `import { Select } from '@lobehub/ui'`                             | `import { Select } from '@lobehub/ui/base-ui'`                              |
+| `import { Modal } from '@lobehub/ui'` + `<Modal open>` declarative | `createModal` / `confirmModal` from `@lobehub/ui/base-ui` (see modal skill) |
+| `import { DropdownMenu/Popover/Switch } from '@lobehub/ui'`        | Import same name from `@lobehub/ui/base-ui` instead                         |
+| `createStyles` for static styles                                   | Use `createStaticStyles` + `cssVar`                                         |
+| Editing only `desktopRouter.config.tsx`                            | Must edit both `.tsx` and `.desktop.tsx`                                    |
+| Using `margin` for flex spacing                                    | Use `gap` prop on Flexbox                                                   |
+| Accessing zustand store without selector                           | Use selectors to access store data (see zustand skill)                      |
+| Text or icon-text actions built with `Flexbox`/`Text` + `onClick`  | Use `Button type={'text'} size={'small'}` with `icon` when needed           |
@@ -1,6 +1,6 @@
 ---
 name: review-checklist
-description: 'Common recurring mistakes in LobeHub code review — console leftovers, missing return await, hardcoded secrets, hardcoded i18n strings, desktop router pair drift, antd vs @lobehub/ui, non-idempotent migrations, cloud impact red flags. Use as a quick checklist when reviewing PRs, diffs, or branch changes.'
+description: "Common recurring mistakes in LobeHub code review — `console.*` leftovers, missing `return await`, hardcoded secrets, hardcoded i18n strings, desktop router pair drift, antd vs `@lobehub/ui`, non-idempotent migrations, cloud impact red flags. Use as a quick checklist when reviewing a PR, diff, or branch change. Triggers on 'code review', 'review the diff', 'review this PR', 'review changes', 'PR review checklist', '审一下', '审 PR'."
 user-invocable: false
 ---

@@ -0,0 +1,142 @@
+---
+name: skills-audit
+description: Weekly audit of `.agents/skills/*/SKILL.md` — surfaces duplicate / overlapping / stale skills, inconsistent descriptions, broken cross-references, and merge/delete candidates. Run as a recurring health-check, not during normal feature work.
+disable-model-invocation: true
+argument-hint: '[--verbose | --apply]'
+---
+
+# Skills Audit
+
+Periodic review of the project-local skill set under `.agents/skills/`. The goal is to catch drift before the catalog becomes confusing — too many skills, overlapping triggers, descriptions that no longer match the body, references to skills that were renamed/deleted.
+
+**Recommended cadence:** weekly, or after any week where >1 skill was added/renamed.
+
+## Procedure
+
+### 1 — Inventory
+
+Build a fresh census of all SKILL.md files. Do NOT trust any prior cached list.
+
+```bash
+find .agents/skills -name SKILL.md | wc -l                      # total count
+find .agents/skills -name SKILL.md -exec wc -l {} \; | sort -rn # by body length
+```
+
+Group by domain in a mental table (DB / state / UI / agent / testing / workflow / docs / etc.). Note new arrivals since last audit (`git log --since="1 week ago" -- .agents/skills/`).
+
+### 2 — Pull frontmatter for all skills
+
+```bash
+# Extract name + description for each SKILL.md
+for f in .agents/skills/*/SKILL.md; do
+  echo "=== $(basename $(dirname $f)) ==="
+  awk '/^---$/{c++; next} c==1' "$f" | head -20
+done
+```
+
+Read the description block of every skill. The body can stay unread unless step 4 flags it.
+
+### 3 — Detect overlap / redundancy
+
+For each pair within the same domain, ask:
+
+- **Same description**? → likely duplicate (one is probably a stale rename leftover, or a global-vs-local collision).
+- **Trigger keywords substantially overlap**? → either merge, OR tighten one description so the model can choose unambiguously.
+- **One skill's body says "see also: foo"**? → confirm `foo` still exists, AND confirm the cross-reference is still meaningful (the referenced skill may have absorbed the referrer's concerns).
+- **Skill duplicates content from `AGENTS.md`**? → fold into AGENTS.md or slim the skill to just the delta.
+
+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.
+
+### 4 — Description format consistency
+
+Apply the **standard template**:
+
+```
+{Topic + key conventions or scope}. Use when {scenarios — verbs + nouns}. Triggers on {`code-symbols`, 'natural phrases', '中文'}.
+```
+
+Skills with `disable-model-invocation: true` (user-invoked only, slash commands) don't need `Triggers on` — they're never auto-routed.
+
+Flag descriptions that:
+
+- ❌ Have NO `Use when` clause (model can't decide when to load it).
+- ❌ Have NO `Triggers on` clause (and aren't `disable-model-invocation`).
+- ❌ Use weird formats (numbered lists `(1)(2)(3)`, `Triggers:` colon instead of `Triggers on`, `MUST use when ...` as opening word).
+- ❌ Are dramatically terse for a 200+ line body, or dramatically verbose for a 60-line body.
+- ❌ Reference deleted/renamed skills.
+
+### 5 — Stale-skill check
+
+For narrow domain skills (e.g. `response-compliance`, one-off CLI workflows):
+
+```bash
+# Confirm the referenced code surface still exists
+rg -l "response-compliance|openresponses" packages/ src/              # adjust per skill
+git log --since="3 months ago" -- .agents/skills/ < skill > /SKILL.md # is it being maintained?
+```
+
+If the underlying surface is gone and the skill hasn't been edited in 3+ months → flag for archival.
+
+### 6 — Cross-reference integrity
+
+Any skill body mentioning another skill by name:
+
+```bash
+# Scan all skill bodies for skill-name references
+rg -o '`[a-z][a-z0-9-]+`' .agents/skills/*/SKILL.md | grep -v ':\s*$' | sort -u
+```
+
+For each name extracted, confirm `.agents/skills/<name>/SKILL.md` exists. Broken references happen after renames — fix them in the same audit pass.
+
+### 7 — Output report
+
+Produce a markdown summary back to the user with the same structure as the original audit (this skill was created during one):
+
+```markdown
+## 📊 Inventory
+
+{count, domain breakdown}
+
+## 🎯 Recommendations
+
+### 🔴 High confidence
+
+- {action} — {reason}
+
+### 🟡 Medium confidence
+
+- {action} — {reason needs verification}
+
+### 🟢 Low confidence / no-op
+
+- {item considered but skipping because ...}
+
+## 📋 Suggested order
+
+{table of actions with risk + LOC estimate}
+```
+
+End by asking the user which actions to apply — do NOT auto-apply unless the user passed `--apply` and even then confirm destructive deletes individually.
+
+## Output rules
+
+- Be specific. "Skill X overlaps with Y" is useless without naming the overlapping triggers.
+- Cite line numbers when flagging description / body issues.
+- Don't recommend merges unless the call sites would actually load the merged skill in the same context.
+- Don't recommend deletes for skills that haven't been touched recently — "unused" can mean "stable", not "dead".
+
+## What NOT to do
+
+- ❌ Don't rename skill directories without checking for cross-references AND user memory entries that name the old slug.
+- ❌ Don't normalize a description by removing trigger keywords just to fit the template — the keywords are the routing signal.
+- ❌ Don't fold a heavy 200+ line skill into another just because they share a domain — large skills get loaded selectively and merging makes everything load.
+- ❌ Don't propose `.agents/skills/INDEX.md` or `<domain>-<skill>` prefix renames unless the user explicitly asks — costs > benefits for cosmetic reorgs.
+
+## Related history
+
+- First audit: `chore/skills-audit` branch (2026-05-25) — deleted `source-command-dedupe`, renamed `data-fetching` → `data-fetching-architecture`, normalized 9 descriptions, created this skill.
@@ -1,44 +0,0 @@
---
-name: 'source-command-dedupe'
-description: 'Find duplicate GitHub issues'
---
-
-# source-command-dedupe
-
-Use this skill when the user asks to run the migrated source command `dedupe`.
-
-## Command Template
-
-Find up to 3 likely duplicate issues for a given GitHub issue.
-
-To do this, follow these steps precisely:
-
-1. Use an agent to check if the Github issue (a) is closed, (b) does not need to be deduped (eg. because it is broad product feedback without a specific solution, or positive feedback), or (c) already has a duplicates comment that you made earlier. If so, do not proceed.
-2. Use an agent to view a Github issue, and ask the agent to return a summary of the issue
-3. Then, launch 5 parallel agents to search Github for duplicates of this issue, using diverse keywords and search approaches, using the summary from #1
-4. Next, feed the results from #1 and #2 into another agent, so that it can filter out false positives, that are likely not actually duplicates of the original issue. If there are no duplicates remaining, do not proceed.
-5. Finally, comment back on the issue with a list of up to three duplicate issues (or zero, if there are no likely duplicates)
-
-Notes (be sure to tell this to your agents, too):
-
- Use `gh` to interact with Github, rather than web fetch
- Do not use other tools, beyond `gh` (eg. don't use other MCP servers, file edit, etc.)
- Make a todo list first
- For your comment, follow the following format precisely (assuming for this example that you found 3 suspected duplicates):
-
---
-
-Found 3 possible duplicate issues:
-
-1. <link to issue>
-2. <link to issue>
-3. <link to issue>
-
-This issue will be automatically closed as a duplicate in 3 days.
-
- If your issue is a duplicate, please close it and 👍 the existing issue instead
- To prevent auto-closure, add a comment or 👎 this comment
-
-> 🤖 Generated with Codex
-
---
@@ -1,6 +1,6 @@
 ---
 name: spa-routes
-description: MUST use when editing src/routes/ segments, src/spa/router/desktopRouter.config.tsx or desktopRouter.config.desktop.tsx (always change both together), mobileRouter.config.tsx, or when moving UI/logic between routes and src/features/.
+description: "SPA roots-vs-features split for LobeHub — thin route segments under `src/routes/` delegate to domain components under `src/features/`. Use when editing `src/routes/` segments, `src/spa/router/desktopRouter.config.tsx` or `desktopRouter.config.desktop.tsx` (MUST update both together — `desktopRouter.sync.test.tsx` enforces this), `mobileRouter.config.tsx`, `popupRouter.config.tsx`, any colocated `<name>.desktop.{ts,tsx}` variant (e.g. settings `componentMap.ts` × `componentMap.desktop.ts`, page-level `index.tsx` × `index.desktop.tsx`), or moving UI/logic between `routes/` and `features/`. Triggers on `desktopRouter.config`, `mobileRouter.config`, `popupRouter.config`, `componentMap.desktop`, `index.desktop.tsx`, `.desktop.tsx` variant, `src/routes/**`, `src/features/**`, 'add a route', 'new page', 'route segment', '路由'."
 user-invocable: false
 ---

@@ -94,6 +94,27 @@ Anything that changes the tree (new segment, renamed `path`, moved layout, new c

 ---

+## 3b. Other `.desktop.{ts,tsx}` variants inside `src/routes/`
+
+The router pair is **not** the only `.desktop` variant pattern in this repo. Some route trees colocate a `<name>.desktop.{ts,tsx}` next to its base `<name>.{ts,tsx}` — Vite's resolver swaps in the `.desktop` file for Electron builds. Same drift risk as the router pair: editing only one side can break Electron silently.
+
+Known variants today:
+
+| Base file (web)                                       | Desktop file (Electron)                                       | Purpose                                                                                                                                    |
+| ----------------------------------------------------- | ------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------ |
+| `src/routes/(main)/settings/features/componentMap.ts` | `src/routes/(main)/settings/features/componentMap.desktop.ts` | Settings tab → component map. Web uses dynamic `import()`; desktop uses sync imports. `componentMap.sync.test.ts` enforces identical keys. |
+| `src/routes/(main)/agent/index.tsx`                   | `src/routes/(main)/agent/index.desktop.tsx`                   | Page entry. Desktop variant overrides the web page wholesale (e.g. extra popup guards).                                                    |
+| `src/routes/(main)/group/index.tsx`                   | `src/routes/(main)/group/index.desktop.tsx`                   | Same pattern as agent.                                                                                                                     |
+
+**Rules:**
+
+1. After editing **any** `.ts`/`.tsx` under `src/routes/`, glob the same directory for a `<filename>.desktop.{ts,tsx}` sibling. If one exists, apply the equivalent change there in the same commit.
+2. When adding a new SettingsTab, register it in **both** `componentMap.ts` (with `dynamic(...)`) and `componentMap.desktop.ts` (with a sync `import`). `componentMap.sync.test.ts` will fail the build otherwise.
+3. When adding a new desktop-only page wholesale-override, prefer a single base file with platform-aware code over introducing a new `.desktop.tsx` variant — only add a new variant when the two trees genuinely diverge (different store wiring, different popup guards, etc.).
+4. When deleting, remove **both** files together.
+
+---
+
 ## 4. How to Divide Files (route vs feature)

 | Question                                                 | Put in `src/routes/`                                     | Put in `src/features/`       |
@@ -1,6 +1,6 @@
 ---
 name: store-data-structures
-description: Zustand store data structure patterns for LobeHub. Covers List vs Detail data structures, Map + Reducer patterns, type definitions, and when to use each pattern. Use when designing store state, choosing data structures, or implementing list/detail pages.
+description: "Zustand store data-shape patterns for LobeHub — List vs Detail split, Map + Reducer, type definitions sourced from `@lobechat/types` (not `@lobechat/database`). Use when designing store state, choosing between Array (list) and `Record<string, Detail>` (detail map), or implementing a list/detail page pair. Triggers on `messagesMap`, `topicsMap`, `Record<string, Detail>`, 'list vs detail', 'store data shape', 'normalize state', 'state structure'."
 user-invocable: false
 ---

@@ -310,5 +310,5 @@ export interface BenchmarkListItem {

 ## Related Skills

- `data-fetching` — how to fetch and update this data
+- `data-fetching-architecture` — how to fetch and update this data
 - `zustand` — general Zustand patterns
@@ -48,6 +48,7 @@ user-invocable: false
 - Replace magic numbers/strings with well-named constants
 - Defer formatting to tooling
 - Prefer **named exports** over `export default` — keeps refactor renames and IDE auto-import in sync, and avoids the `default` re-naming drift you get with `import Foo from './foo'`. Reserve `export default` for files where the framework requires it (Next.js page/route/layout, React.lazy targets, config files like `vitest.config.ts`)
+- Before adding local helpers for common guards/parsing/normalization (record checks, string extraction, empty-string handling, timing helpers, JSON-safe utilities, etc.), search `packages/utils` first. If the helper already exists or clearly belongs there, import it from `@lobechat/utils` (or the relevant `@lobechat/utils/*` subpath) instead of duplicating tiny helpers across feature files.

 ## UI and Theming

@@ -1,6 +1,6 @@
 ---
 name: upstash-workflow
-description: 'Upstash Workflow implementation guide. Use when creating async workflows with QStash, implementing fan-out patterns, or building 3-layer workflow architecture (process → paginate → execute).'
+description: "Upstash Workflow + QStash implementation guide for LobeHub — 3-layer architecture (process → paginate → execute), fan-out patterns. Use when creating an async workflow, implementing fan-out (paginate → execute), or wiring `serve()` + `context.run` / `context.call` steps. Triggers on `serve()`, `context.run`, `context.call`, `context.sleep`, `qstash`, 'async workflow', 'fan-out workflow', 'QStash workflow'."
 user-invocable: false
 ---

@@ -1,6 +1,10 @@
 # Issue Triage Guide

-This guide is used for batch triaging GitHub issues - analyzing issues and applying appropriate labels.
+This guide is used for triaging GitHub issues — analyzing issues and applying only the most essential business-domain labels.
+
+## Core Principle
+
+**Each issue should have 1-3 labels that describe its core business domain.** Do NOT apply redundant labels that can be inferred from other labels. Less is more.

 ## Workflow

@@ -20,23 +24,76 @@ For each issue number, run:
 gh issue view [ISSUE_NUMBER] --json number,title,body,labels,comments
 ```

-### Step 3: Analyze and Select Labels
+### Step 3: Select Labels (1-3 per issue)

-Extract information from the issue template and content:
+Only apply labels from these THREE categories:

-#### Template Fields Mapping
+#### Category 1: Technology Carrier

- 📦 Platform field → `platform:web/desktop/mobile`
- 💻 Operating System → `os:windows/macos/linux/ios`
- 🌐 Browser → `device:pc/mobile`
- 📦 Deployment mode → `deployment:server/client/pglite`
- Platform (hosting) → `hosting:cloud/self-host/vercel/zeabur/railway`
+The runtime environment or technology wrapper where the issue occurs:

-#### Provider Detection
+| Label | When to apply |
+|-------|--------------|
+| `electron` | Desktop/Electron-specific issues. This REPLACES `platform:desktop`, `os:*`, `deployment:*`, `hosting:*` — do NOT add those. |
+| `pwa` | PWA/mobile-app-specific issues |
+| `docker` | Docker-specific deployment issues |

-**IMPORTANT**: Always check issue title and body for provider mentions!
+**Rule**: If `electron` is applied, do NOT add `platform:desktop`, `os:*`, `deployment:*`, or `hosting:*`. The `electron` label already implies all of these.

-**Official Providers** (check for these keywords in title/body):
+#### Category 2: Feature / Component
+
+The functional area affected. Select the 1-2 MOST relevant:
+
+Core Features:
+
+- `feature:agent` - Agent/Assistant functionality
+- `feature:topic` - Topic/Conversation management
+- `feature:marketplace` - Agent/plugin marketplace
+- `feature:settings` - Settings and configuration
+
+Content & Knowledge:
+
+- `feature:editor` - Lobe Editor / rich text / markdown rendering
+- `feature:markdown` - Markdown rendering (if separate from editor)
+- `feature:files` - File upload/management
+- `feature:knowledge-base` - Knowledge base and RAG
+- `feature:export` - Export functionality
+
+Model Capabilities:
+
+- `feature:tool` - Tool calling and function execution
+- `feature:streaming` - Streaming responses
+- `feature:vision` - Vision/multimodal capabilities
+- `feature:image` - AI image generation
+- `feature:tts` - Text-to-speech
+
+Technical:
+
+- `feature:api` - Backend API
+- `feature:auth` - Authentication/authorization
+- `feature:sync` - Cloud sync functionality
+- `feature:search` - Search functionality
+- `feature:mcp` - MCP integration
+- `feature:thread` - Thread/Subtopic functionality
+
+Collaboration:
+
+- `feature:group-chat` - Group chat functionality
+- `feature:memory` - Memory feature
+- `feature:team-workspace` - Team workspace
+- `feature:im-integration` - IM and bot integration
+
+Other:
+
+- `feature:schedule-task` - Scheduled task functionality
+
+**Rule**: Pick only the 1-2 most specific feature labels. Don't stack multiple features unless the issue genuinely spans multiple areas.
+
+#### Category 3: Model Provider
+
+Only when the issue is SPECIFICALLY about a provider's behavior:
+
+**Official Providers** (check title and body for these keywords):

 - `openai`, `gpt` → `provider:openai`
 - `gemini` → `provider:gemini`
@@ -57,197 +114,100 @@ Extract information from the issue template and content:
 **Third-party Aggregation Providers**:

 - `aihubmix`, `AIHubMix`, `AIHUBMIX` → `provider:aihubmix`
- Check environment variables like `AIHUBMIX_*` in issue body
+- `zenmux` → `provider:zenmux`

-**Multiple Providers**: If issue mentions multiple providers, add ALL applicable provider labels.
+**Rule**: Only add a provider label if the issue is specifically about that provider's behavior (e.g., "Gemini returns error X"). Do NOT add provider labels just because the issue template mentions a provider.

-### Label Categories
-
-#### a) Issue Type (select ONE if applicable)
-
- `💄 Design` - UI/UX design issues
- `📝 Documentation` - Documentation improvements
- `⚡️ Performance` - Performance optimization
-
-#### b) Priority (select ONE if applicable)
-
- `priority:high` - Critical issues, data loss, security, maintainer mentions "urgent"/"serious"/"critical"
- `priority:medium` - Important issues affecting multiple users, significant functionality impact
- `priority:low` - Nice to have, minor issues, edge cases
-
-**Priority Guidelines**:
-
- Set `priority:high` for: data loss, authentication failures, deployment blockers, critical bugs
- Set `priority:medium` for: feature bugs affecting multiple users, workflow issues
- Set `priority:low` for: cosmetic issues, feature requests, configuration questions
-
-#### c) Platform (select ALL applicable)
-
- `platform:web`
- `platform:desktop`
- `platform:mobile`
-
-#### d) Device (for platform:web, select ONE)
-
- `device:pc`
- `device:mobile`
-
-#### e) Operating System (select ALL applicable)
-
- `os:windows`
- `os:macos`
- `os:linux`
- `os:ios`
- `os:android`
-
-#### f) Hosting Platform (select ONE)
-
- `hosting:cloud` - Official LobeHub Cloud
- `hosting:self-host` - Self-hosted deployment
- `hosting:vercel` - Vercel deployment
- `hosting:zeabur` - Zeabur deployment
- `hosting:railway` - Railway deployment
-
-#### g) Deployment Mode (select ONE if mentioned)
-
- `deployment:server` - Server-side database mode
- `deployment:client` - Client-side database mode
- `deployment:pglite` - PGLite mode
-
-**Additional deployment tags**:
-
- `docker` - If using Docker deployment
- `electron` - If desktop/Electron specific
-
-#### h) Model Provider (select ALL applicable)
-
-See "Provider Detection" section above for complete list.
-
-**IMPORTANT**: Always scan issue title and body for provider keywords!
-
-#### i) Feature/Component (select ALL applicable)
-
-Core Features:
-
- `feature:settings` - Settings and configuration
- `feature:agent` - Agent/Assistant functionality
- `feature:topic` - Topic/Conversation management
- `feature:marketplace` - Agent marketplace
-
-File & Knowledge:
-
- `feature:files` - File upload/management
- `feature:knowledge-base` - Knowledge base and RAG
- `feature:export` - Export functionality
-
-Model Capabilities:
-
- `feature:streaming` - Streaming responses
- `feature:tool` - Tool calling
- `feature:vision` - Vision/multimodal capabilities
- `feature:image` - AI image generation
- `feature:dalle` - DALL-E specific
- `feature:tts` - Text-to-speech
-
-Technical:
-
- `feature:api` - Backend API
- `feature:auth` - Authentication/authorization
- `feature:sync` - Cloud sync functionality
- `feature:search` - Search functionality
- `feature:mcp` - MCP integration
- `feature:editor` - Lobe Editor
- `feature:markdown` - Markdown rendering
- `feature:thread` - Thread/Subtopic functionality
-
-Collaboration:
-
- `feature:group-chat` - Group chat functionality
- `feature:memory` - Memory feature
- `feature:team-workspace` - Team workspace
-
-#### j) Workflow/Status
+#### Special Labels (use sparingly)

+- `i18n` - Internationalization / translation issues
 - `Duplicate` - Only if duplicate of an OPEN issue (mention issue number)
- `needs-reproduction` - Cannot reproduce, needs more information
- `good-first-issue` - Good for first-time contributors
 - `🤔 Need Reproduce` - Needs reproduction steps
+- `good-first-issue` - Good for first-time contributors

 ### Step 4: Apply Labels

-Add labels (comma-separated, no spaces after commas):
-
-```bash
-gh issue edit [ISSUE_NUMBER] --add-label "label1,label2,label3"
-```
-
-Remove "unconfirm" label if adding other labels:
-
 ```bash
+gh issue edit [ISSUE_NUMBER] --add-label "label1,label2"
 gh issue edit [ISSUE_NUMBER] --remove-label "unconfirm"
 ```

-**Important**: Combine both commands when possible for efficiency.
-
 ### Step 5: Log Summary

-For each issue, provide reasoning (2-4 sentences):
+For each issue, provide a brief reasoning (1-2 sentences) explaining why each label was chosen.

- Labels applied and why
- Key factors from issue template/comments
- Provider detection reasoning (if applicable)
+## What NOT to Label
+
+These categories are INTENTIONALLY OMITTED — do NOT apply them:
+
+| Do NOT apply | Reason |
+|-------------|--------|
+| `platform:web`, `platform:desktop`, `platform:mobile` | Inferred from `electron`/`pwa` or issue context |
+| `os:windows`, `os:macos`, `os:linux`, `os:ios`, `os:android` | Low triage value; inferred from `electron` |
+| `device:pc`, `device:mobile` | Redundant with platform |
+| `hosting:cloud`, `hosting:self-host`, `hosting:vercel`, etc. | Low triage value unless deployment-specific |
+| `deployment:server`, `deployment:client`, `deployment:pglite` | Low triage value; inferred from `electron` |
+| `priority:high`, `priority:medium`, `priority:low` | Maintainers judge priority themselves |
+| `🐛 Bug`, `💄 Design`, `📝 Documentation`, `⚡️ Performance` | Issue type is already indicated by GitHub issue template |
+| `Inactive` | Handled separately; do NOT add during triage |
+
+## Examples
+
+### Example 1: Electron desktop bug
+
+**Issue**: "Connection failure when executing tasks on macOS desktop app"
+
+**Analysis**: Desktop Electron app issue with task scheduling.
+
+**Labels**: `electron,feature:schedule-task`
+
+**Why**: `electron` covers the desktop platform. `feature:schedule-task` identifies the affected feature. No need for `platform:desktop`, `os:macos`, `hosting:cloud`, `priority:*`, or `Bug`.
+
+### Example 2: Provider-specific issue
+
+**Issue**: "Gemini tool calling returns empty response on desktop"
+
+**Analysis**: Desktop app issue, but the core problem is Gemini provider behavior with tool calling.
+
+**Labels**: `electron,provider:gemini`
+
+**Why**: `electron` for the desktop context. `provider:gemini` because the issue is about Gemini's behavior. The tool calling aspect is secondary — the provider is the key domain.
+
+### Example 3: Feature-specific issue
+
+**Issue**: "Underscore auto-escaped in markdown editor"
+
+**Analysis**: Markdown rendering bug in the editor component.
+
+**Labels**: `feature:markdown`
+
+**Why**: Single label is sufficient — the issue is purely about markdown rendering. No need for platform, OS, or priority labels.
+
+### Example 4: Web-only feature request
+
+**Issue**: "Add search functionality to plugin marketplace"
+
+**Analysis**: Feature request for marketplace search. Web platform, no specific provider.
+
+**Labels**: `feature:marketplace,feature:search`
+
+**Why**: Two feature labels capture the core domain. No platform label needed — it's a web app by default.
+
+### Example 5: Ollama self-hosted issue
+
+**Issue**: "Ollama model not loading on self-hosted Docker deployment"
+
+**Analysis**: Provider-specific issue with Ollama on Docker.
+
+**Labels**: `docker,provider:ollama`
+
+**Why**: `docker` for the deployment context, `provider:ollama` for the model provider. No need for `hosting:self-host` or `platform:*`.

 ## Important Rules

-1. **Read Carefully**: Read issue template fields AND issue body/title for complete context
-2. **Provider Detection**: ALWAYS check title and body for provider keywords (including aihubmix, etc.)
-3. **Multiple Categories**: Use ALL applicable labels from different categories
-4. **Label Prefixes**: Always use proper prefixes (`feature:`, `provider:`, `os:`, `platform:`, etc.)
-5. **Maintainer Comments**: Check maintainer comments for priority/status hints
-6. **No Comments**: Only apply labels, DO NOT post comments to issues
-7. **Batch Efficiency**: Process issues in parallel when possible
-
-## Common Patterns
-
-### Provider in Environment Variables
-
-If issue body contains `AIHUBMIX_*`, add `provider:aihubmix`
-
-### Multiple Provider Issues
-
-If comparing providers (e.g., "works with OpenAI but not Gemini"), add both provider labels
-
-### Desktop Issues
-
-Desktop issues often need: `platform:desktop`, `electron`, specific `os:*`, and `deployment:client` or `deployment:server`
-
-### Knowledge Base Issues
-
-Usually need: `feature:knowledge-base`, often with `feature:files`, may need `provider:*` for embedding models
-
-### Tool Calling Issues
-
-Usually need: `feature:tool`, specific `provider:*`, may need `feature:mcp` if MCP-related
-
-### Streaming Issues
-
-Usually need: `feature:streaming`, specific `provider:*`, check for timeout/performance issues
-
-## Example Triage
-
-**Issue #8850**: "aihubmix 的优惠 app 没有生效"
-
-**Analysis**:
-
- Title contains "aihubmix" → `provider:aihubmix`
- Template shows: Windows, Chrome, Docker, Client mode
- About API discount codes not working
-
-**Labels Applied**:
-
-```bash
-gh issue edit 8850 --add-label "provider:aihubmix,platform:web,os:windows,deployment:client,hosting:self-host,docker"
-gh issue edit 8850 --remove-label "unconfirm"
-```
-
-**Reasoning**: AIHubMix provider discount feature not working. Client mode deployment on Windows with Docker. Provider detection from title keyword "aihubmix".
+1. **1-3 labels per issue** — Never exceed 3 labels. If you find yourself adding more, you're being too granular.
+2. **`electron` replaces all platform/OS/deployment labels** — Never combine `electron` with `platform:desktop`, `os:*`, `deployment:*`, or `hosting:*`.
+3. **Provider only when relevant** — Only add `provider:*` if the issue is specifically about that provider's behavior.
+4. **No priority, no type** — Do NOT add `priority:*`, `🐛 Bug`, `💄 Design`, etc. Maintainers handle these.
+5. **No comments** — Only apply labels. Do NOT post comments to issues.
+6. **Remove `unconfirm`** — Always remove the `unconfirm` label when applying triage labels.
@@ -77,6 +77,9 @@ OPENAI_API_KEY=sk-xxxxxxxxx
 # use a proxy to connect to the Anthropic API
 # ANTHROPIC_PROXY_URL=https://api.anthropic.com

+# Anthropic SDK client timeout in milliseconds
+# ANTHROPIC_CLIENT_TIMEOUT=295000
+
 # ## Google AI  ####

 # GOOGLE_API_KEY=xxxxxxxxxxxxxxxxxxxxxxxxxxxxx
@@ -1,5 +1,13 @@
+#!/usr/bin/env sh
+set -e
+
+[ "${HUSKY-}" = "0" ] && exit 0
+
+export PATH="node_modules/.bin:$PATH"
+
 BRANCH=$(git branch --show-current)
 if [ "$BRANCH" = "dev" ] || [ "$BRANCH" = "main" ]; then
  npm run type-check
 fi
-npx --no-install lint-staged
+
+lint-staged
@@ -75,7 +75,7 @@ runs:

        # 1. 上传安装包到版本目录
        echo "📦 Uploading release files to s3://$S3_BUCKET/$CHANNEL/$VERSION/"
-        for file in release/*.dmg release/*.zip release/*.exe release/*.AppImage release/*.deb release/*.rpm release/*.snap release/*.tar.gz; do
+        for file in release/*.dmg release/*.zip release/*.exe release/*.AppImage release/*.deb release/*.rpm release/*.snap release/*.tar.gz release/*.blockmap; do
          if [ -f "$file" ]; then
            filename=$(basename "$file")
            echo "   ↗️ $filename"
@@ -10,6 +10,10 @@ inputs:
    description: Pass-through to actions/setup-node package-manager-cache
    required: false
    default: 'false'
+  bun-version:
+    description: Bun version
+    required: false
+    default: '1.3.2'

 runs:
  using: composite
@@ -21,6 +25,8 @@ runs:

    - name: Install bun
      uses: oven-sh/setup-bun@v2
+      with:
+        bun-version: ${{ inputs.bun-version }}

    - name: Setup Node.js
      uses: actions/setup-node@v6
@@ -59,7 +59,14 @@ jobs:
    timeout-minutes: 30
    steps:
      - name: Checkout
-        uses: actions/checkout@v6
+        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
@@ -16,14 +16,14 @@ permissions:
 jobs:
  run:
    permissions:
-      issues: write # for actions-cool/issues-helper to update issues
-      pull-requests: write # for actions-cool/issues-helper to update PRs
+      issues: write
+      pull-requests: write
    runs-on: ubuntu-latest
    steps:
      - name: Auto Comment on Issues Closed
        uses: wow-actions/auto-comment@v1
        with:
-          GITHUB_TOKEN: ${{ secrets.GH_TOKEN}}
+          GITHUB_TOKEN: ${{ secrets.GH_TOKEN }}
          issuesClosed: |
            ✅ @{{ author }}

@@ -51,11 +51,4 @@ jobs:
            The growth of project is inseparable from user feedback and contribution, thanks for your contribution! If you are interesting with the lobehub developer community, please join our [discord](https://discord.com/invite/AYFPHvv2jT) and then dm @arvinxx or @canisminor1990. They will invite you to our private developer channel. We are talking about the lobe-chat development or sharing ai newsletter around the world.
          emoji: 'hooray'
          pr-emoji: '+1, heart'
-      - name: Remove inactive
-        if: github.event.issue.state == 'open' && github.actor == github.event.issue.user.login
-        uses: actions-cool/issues-helper@v3
-        with:
-          actions: 'remove-labels'
-          token: ${{ secrets.GH_TOKEN }}
-          issue-number: ${{ github.event.issue.number }}
-          labels: 'Inactive'
+
@@ -1,63 +0,0 @@
-name: Issue Close Require
-
-on:
-  schedule:
-    - cron: '0 0 * * *'
-
-permissions:
-  contents: read
-
-jobs:
-  issue-check-inactive:
-    permissions:
-      issues: write # for actions-cool/issues-helper to update issues
-      pull-requests: write # for actions-cool/issues-helper to update PRs
-    runs-on: ubuntu-latest
-    steps:
-      - name: check-inactive
-        uses: actions-cool/issues-helper@v3
-        with:
-          actions: 'check-inactive'
-          token: ${{ secrets.GH_TOKEN }}
-          inactive-label: 'Inactive'
-          inactive-day: 60
-
-  issue-close-require:
-    permissions:
-      issues: write # for actions-cool/issues-helper to update issues
-      pull-requests: write # for actions-cool/issues-helper to update PRs
-    runs-on: ubuntu-latest
-    steps:
-      - name: need reproduce
-        uses: actions-cool/issues-helper@v3
-        with:
-          actions: 'close-issues'
-          token: ${{ secrets.GH_TOKEN }}
-          labels: '✅ Fixed'
-          inactive-day: 3
-          body: |
-            👋 @{{ author }}
-            <br/>
-            Since the issue was labeled with `✅ Fixed`, but no response in 3 days. This issue will be closed. If you have any questions, you can comment and reply.
-      - name: need reproduce
-        uses: actions-cool/issues-helper@v3
-        with:
-          actions: 'close-issues'
-          token: ${{ secrets.GH_TOKEN }}
-          labels: '🤔 Need Reproduce'
-          inactive-day: 3
-          body: |
-            👋 @{{ author }}
-            <br/>
-            Since the issue was labeled with `🤔 Need Reproduce`, but no response in 3 days. This issue will be closed. If you have any questions, you can comment and reply.
-      - name: need reproduce
-        uses: actions-cool/issues-helper@v3
-        with:
-          actions: 'close-issues'
-          token: ${{ secrets.GH_TOKEN }}
-          labels: "🙅🏻‍♀️ WON'T DO"
-          inactive-day: 3
-          body: |
-            👋 @{{ github.event.issue.user.login }}
-            <br/>
-            Since the issue was labeled with `🙅🏻‍♀️ WON'T DO`, and no response in 3 days. This issue will be closed. If you have any questions, you can comment and reply.
@@ -20,7 +20,7 @@ jobs:
      - uses: actions/checkout@v6

      - name: Clean issue notice
-        uses: actions-cool/issues-helper@v3
+        uses: actions-cool/issues-helper@e361abf610221f09495ad510cb1e69328d839e1c # v3.7.6
        with:
          actions: 'close-issues'
          labels: '🚨 Sync Fail'
@@ -37,7 +37,7 @@ jobs:

      - name: Sync check
        if: failure()
-        uses: actions-cool/issues-helper@v3
+        uses: actions-cool/issues-helper@e361abf610221f09495ad510cb1e69328d839e1c # v3.7.6
        with:
          actions: 'create-issue'
          title: '🚨 同步失败 | Sync Fail'
@@ -35,7 +35,15 @@ jobs:
      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/builtin-tool-lobe-agent model-bank'

    steps:
-      - uses: actions/checkout@v6
+      - 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
@@ -101,7 +109,15 @@ jobs:
    name: Test App (shard ${{ matrix.shard }}/3)
    runs-on: ubuntu-latest
    steps:
-      - uses: actions/checkout@v6
+      - 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
@@ -128,7 +144,15 @@ jobs:
    name: Merge and Upload App Coverage
    runs-on: ubuntu-latest
    steps:
-      - uses: actions/checkout@v6
+      - 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
@@ -161,7 +185,15 @@ jobs:
    runs-on: ubuntu-latest

    steps:
-      - uses: actions/checkout@v6
+      - 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
@@ -207,7 +239,15 @@ jobs:
          - 5432:5432

    steps:
-      - uses: actions/checkout@v6
+      - 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
@@ -28,6 +28,9 @@ prd
 # Recordings
 .records/

+# Agent-gateway probe captures (local debugging dumps)
+.agent-gateway/
+
 # Temporary files
 .temp/
 temp/
@@ -96,7 +99,7 @@ sitemap*.xml
 robots.txt

 # Git hooks
-.husky/prepare-commit-msg
+.githooks/prepare-commit-msg

 # Documents and media
 *.pdf
@@ -106,6 +109,7 @@ vertex-ai-key.json

 # Agent tracing snapshots
 .agent-tracing/
+.llm-generation-tracing/

 # AI coding tools
 .local/
@@ -8,7 +8,7 @@
 .history
 .temp
 .env.local
-.husky
+.githooks
 .npmrc
 .gitkeep
 venv
@@ -59,4 +59,4 @@ Dockerfile*

 # misc
 # add other ignore file below
-.next
+.next
@@ -7,6 +7,7 @@ Guidelines for using AI coding agents in this LobeHub repository.
 - Next.js 16 + React 19 + TypeScript
 - SPA inside Next.js with `react-router-dom`
 - `@lobehub/ui`, antd for components; antd-style for CSS-in-JS — **prefer `createStaticStyles` with `cssVar.*`** (zero-runtime); only fall back to `createStyles` + `token` when styles genuinely need runtime computation. See `.cursor/docs/createStaticStyles_migration_guide.md`.
+- **Component priority**: `@lobehub/ui/base-ui` (headless primitives) **first**, then `@lobehub/ui` root, then antd as last resort. When the component exists in base-ui, use it — never reach for the root or antd counterpart. Base-ui covers `Select`, `Modal` / `createModal` / `confirmModal`, `DropdownMenu`, `ContextMenu`, `Popover`, `ScrollArea`, `Switch`, `Toast`, `FloatingSheet`. Prefer `@lobehub/ui/base-ui` for new code and migrate root-package call sites opportunistically.
 - react-i18next for i18n; zustand for state management
 - SWR for data fetching; TRPC for type-safe backend
 - Drizzle ORM with PostgreSQL; Vitest for testing
@@ -2,6 +2,39 @@

 # Changelog

+### [Version 2.2.0](https://github.com/lobehub/lobe-chat/compare/v2.1.59-canary.27...v2.2.0)
+
+<sup>Released on **2026-05-18**</sup>
+
+#### 💄 Styles
+
+- **pricing**: restore DeepSeek models to official pricing.
+
+#### 🐛 Bug Fixes
+
+- **conversation**: animate only the last markdown block + drop clearMessages hotkey.
+
+<br/>
+
+<details>
+<summary><kbd>Improvements and Fixes</kbd></summary>
+
+#### Styles
+
+- **pricing**: restore DeepSeek models to official pricing, closes [#14911](https://github.com/lobehub/lobe-chat/issues/14911) ([e566688](https://github.com/lobehub/lobe-chat/commit/e566688))
+
+#### What's fixed
+
+- **conversation**: animate only the last markdown block + drop clearMessages hotkey, closes [#14906](https://github.com/lobehub/lobe-chat/issues/14906) ([469a8e6](https://github.com/lobehub/lobe-chat/commit/469a8e6))
+
+</details>
+
+<div align="right">
+
+[![](https://img.shields.io/badge/-BACK_TO_TOP-151515?style=flat-square)](#readme-top)
+
+</div>
+
 ## [Version 2.1.58](https://github.com/lobehub/lobe-chat/compare/v2.1.57...v2.1.58)

 <sup>Released on **2026-05-13**</sup>
@@ -219,7 +219,7 @@ ENV \
    # AiHubMix
    AIHUBMIX_API_KEY="" AIHUBMIX_MODEL_LIST="" \
    # Anthropic
-    ANTHROPIC_API_KEY="" ANTHROPIC_MODEL_LIST="" ANTHROPIC_PROXY_URL="" \
+    ANTHROPIC_API_KEY="" ANTHROPIC_CLIENT_TIMEOUT="" ANTHROPIC_MODEL_LIST="" ANTHROPIC_PROXY_URL="" \
    # Amazon Bedrock
    ENABLED_AWS_BEDROCK="" AWS_ACCESS_KEY_ID="" AWS_SECRET_ACCESS_KEY="" AWS_REGION="" AWS_BEDROCK_MODEL_LIST="" \
    # Azure OpenAI
@@ -4,9 +4,11 @@

 # LobeHub

-LobeHub is the ultimate space for work and life: <br/>
-to find, build, and collaborate with agent teammates that grow with you.<br/>
-We’re building the world’s largest human–agent co-evolving network.
+LobeHub organizes your agents into 7×24 operation.
+
+It hires, schedules, reports on your entire AI team.
+
+You stay in charge — without staying online.

 **English** · [简体中文](./README.zh-CN.md) · [Official Site][official-site] · [Changelog][changelog] · [Documents][docs] · [Blog][blog] · [Feedback][github-issues-link]

@@ -25,7 +27,6 @@ We’re building the world’s largest human–agent co-evolving network.
 [![][github-stars-shield]][github-stars-link]
 [![][github-issues-shield]][github-issues-link]
 [![][github-license-shield]][github-license-link]<br>
-[![][sponsor-shield]][sponsor-link]

 **Share LobeHub Repository**

@@ -37,9 +38,9 @@ We’re building the world’s largest human–agent co-evolving network.
 [![][share-mastodon-shield]][share-mastodon-link]
 [![][share-linkedin-shield]][share-linkedin-link]

-<sup>Agent teammates that grow with you</sup>
+<sup>Your Chief Agent Operator</sup>

-[![][github-trending-shield]][github-trending-url]
+<a href="https://www.producthunt.com/products/lobehub?embed=true&amp;utm_source=badge-top-post-badge&amp;utm_medium=badge&amp;utm_campaign=badge-lobehub-2" target="_blank" rel="noopener noreferrer"><img alt="LobeHub - Your Chief Agent Operator for multi-agent work | Product Hunt" width="250" height="54" src="https://api.producthunt.com/widgets/embed-image/v1/top-post-badge.svg?post_id=1147569&amp;theme=light&amp;period=daily&amp;t=1779247564355"></a> <a href="https://trendshift.io/repositories/19224" target="_blank"><img src="https://trendshift.io/api/badge/repositories/19224" alt="lobehub%2Flobehub | Trendshift" style="width: 250px; height: 55px;" width="250" height="55"/></a>

 [![](https://vercel.com/oss/program-badge.svg)](https://vercel.com/oss)

@@ -52,30 +53,10 @@ We’re building the world’s largest human–agent co-evolving network.

 - [👋🏻 Getting Started & Join Our Community](#-getting-started--join-our-community)
 - [✨ Features](#-features)
+  - [Operator: Agents as the Unit of Work](#operator-agents-as-the-unit-of-work)
  - [Create: Agents as the Unit of Work](#create-agents-as-the-unit-of-work)
  - [Collaborate: Scale New Forms of Collaboration Networks](#collaborate-scale-new-forms-of-collaboration-networks)
  - [Evolve: Co-evolution of Humans and Agents](#evolve-co-evolution-of-humans-and-agents)
-  - [MCP Plugin One-Click Installation](#mcp-plugin-one-click-installation)
-  - [MCP Marketplace](#mcp-marketplace)
-  - [Desktop App](#desktop-app)
-  - [Smart Internet Search](#smart-internet-search)
-  - [Chain of Thought](#chain-of-thought)
-  - [Branching Conversations](#branching-conversations)
-  - [Artifacts Support](#artifacts-support)
-  - [File Upload /Knowledge Base](#file-upload-knowledge-base)
-  - [Multi-Model Service Provider Support](#multi-model-service-provider-support)
-  - [Local Large Language Model (LLM) Support](#local-large-language-model-llm-support)
-  - [Model Visual Recognition](#model-visual-recognition)
-  - [TTS & STT Voice Conversation](#tts--stt-voice-conversation)
-  - [Text to Image Generation](#text-to-image-generation)
-  - [Plugin System (Function Calling)](#plugin-system-function-calling)
-  - [Agent Market (GPTs)](#agent-market-gpts)
-  - [Support Local / Remote Database](#support-local--remote-database)
-  - [Support Multi-User Management](#support-multi-user-management)
-  - [Progressive Web App (PWA)](#progressive-web-app-pwa)
-  - [Mobile Device Adaptation](#mobile-device-adaptation)
-  - [Custom Themes](#custom-themes)
-  - [`*` What's more](#-whats-more)
 - [🛳 Self Hosting](#-self-hosting)
  - [`A` Deploying with Vercel, Zeabur , Sealos or Alibaba Cloud](#a-deploying-with-vercel-zeabur--sealos-or-alibaba-cloud)
  - [`B` Deploying with Docker](#b-deploying-with-docker)
@@ -95,7 +76,7 @@ We’re building the world’s largest human–agent co-evolving network.

 <br/>

-<https://github.com/user-attachments/assets/6710ad97-03d0-4175-bd75-adff9b55eca2>
+<https://github.com/user-attachments/assets/0a33365f-b786-48b5-9ed6-f8af7927bccb>

 ## 👋🏻 Getting Started & Join Our Community

@@ -104,9 +85,9 @@ By adopting the Bootstrapping approach, we aim to provide developers and users w

 Whether for users or professional developers, LobeHub will be your AI Agent playground. Please be aware that LobeHub is currently under active development, and feedback is welcome for any [issues][issues-link] encountered.

-| [![](https://api.producthunt.com/widgets/embed-image/v1/featured.svg?post_id=1065874&theme=light&t=1769347414733)](https://www.producthunt.com/products/lobehub?embed=true&utm_source=badge-featured&utm_medium=badge&utm_campaign=badge-lobehub) | We are live on Product Hunt! We are thrilled to bring LobeHub to the world. If you believe in a future where humans and agents co-evolve, please support our journey. |
-| :------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | :-------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
-| [![][discord-shield-badge]][discord-link]                                                                                                                                                                                                         | Join our Discord community! This is where you can connect with developers and other enthusiastic users of LobeHub.                                                    |
+| [![](https://api.producthunt.com/widgets/embed-image/v1/featured.svg?post_id=1065874&theme=light&t=1769347414733)](https://www.producthunt.com/products/lobehub?launch=lobehub-2&embed=true&utm_source=badge-featured&utm_medium=badge&utm_campaign=badge-lobehub) | We are live on Product Hunt! We are thrilled to bring LobeHub to the world. If you believe in a future where humans and agents co-evolve, please support our journey. |
+| :----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | :-------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| [![][discord-shield-badge]][discord-link]                                                                                                                                                                                                                          | Join our Discord community! This is where you can connect with developers and other enthusiastic users of LobeHub.                                                    |

 > \[!IMPORTANT]
 >
@@ -130,7 +111,26 @@ Today’s agents are one-off, task-driven tools. They lack context, live in isol

 LobeHub is a work-and-lifestyle space to find, build, and collaborate with agent teammates that grow with you. In LobeHub, we treat **Agents as the unit of work**, providing an infrastructure where humans and agents co-evolve.

-![](https://hub-apac-1.lobeobjects.space/blog/assets/2204cde2228fb3f583f3f2c090bc49fb.webp)
+![](https://github.com/user-attachments/assets/89d1c402-a62b-4794-82ea-17e5ee1a6165)
+
+### Operator: Agents as the Unit of Work
+
+Hires, schedules, and reports on your entire AI team.
+
+- **More productivity. Fewer tools**: Bring all your agents under one roof.
+- **IM Gateway**: Agents where you already chat.
+
+![](https://github.com/user-attachments/assets/7b08d6d9-9dff-4b06-a919-324630554509)
+
+[![][back-to-top]](#readme-top)
+
+<div align="right">
+
+[![][back-to-top]](#readme-top)
+
+</div>
+
+![](https://github.com/user-attachments/assets/81e89324-fc66-4024-99a3-aa8e16ec8184)

 ### Create: Agents as the Unit of Work

@@ -139,6 +139,8 @@ Building a personalized AI team starts with the **Agent Builder**. You can descr
 - **Unified Intelligence**: Seamlessly access any model and any modality—all under your control.
 - **10,000+ Skills**: Connect your agents to the skills you use every day with a library of over 10,000 tools and MCP-compatible plugins.

+![](https://github.com/user-attachments/assets/949b8166-486d-4750-ad7a-cfe7bfcb84e3)
+
 [![][back-to-top]](#readme-top)

 <div align="right">
@@ -158,6 +160,8 @@ LobeHub introduces **Agent Groups**, allowing you to work with agents like real
 - **Project**: Organize work by project to keep everything structured and easy to track.
 - **Workspace**: A shared space for teams to collaborate with agents, ensuring clear ownership and visibility across the organization.

+![](https://github.com/user-attachments/assets/e51526c6-e09c-4a5a-9cec-dcd3fd68a3a8)
+
 [![][back-to-top]](#readme-top)

 <div align="right">
@@ -175,113 +179,7 @@ The best AI is one that understands you deeply. LobeHub features **Personal Memo
 - **Continual Learning**: Your agents learn from how you work, adapting their behavior to act at the right moment.
 - **White-Box Memory**: We believe in transparency. Your agents use structured, editable memory, giving you full control over what they remember.

-<div align="right">
-
-[![][back-to-top]](#readme-top)
-
-</div>
-
-<details>
-<summary>More Features</summary>
-
-![][image-feat-mcp]
-
-### MCP Plugin One-Click Installation
-
-**Seamlessly Connect Your AI to the World**
-
-Unlock the full potential of your AI by enabling smooth, secure, and dynamic interactions with external tools, data sources, and services. LobeHub's MCP (Model Context Protocol) plugin system breaks down the barriers between your AI and the digital ecosystem, allowing for unprecedented connectivity and functionality.
-
-Transform your conversations into powerful workflows by connecting to databases, APIs, file systems, and more. Experience the freedom of AI that truly understands and interacts with your world.
-
-[![][back-to-top]](#readme-top)
-
-![][image-feat-mcp-market]
-
-### MCP Marketplace
-
-**Discover, Connect, Extend**
-
-Browse a growing library of MCP plugins to expand your AI's capabilities and streamline your workflows effortlessly. Visit [lobehub.com/mcp](https://lobehub.com/mcp) to explore the MCP Marketplace, which offers a curated collection of integrations that enhance your AI's ability to work with various tools and services.
-
-From productivity tools to development environments, discover new ways to extend your AI's reach and effectiveness. Connect with the community and find the perfect plugins for your specific needs.
-
-[![][back-to-top]](#readme-top)
-
-![][image-feat-desktop]
-
-### Desktop App
-
-**Peak Performance, Zero Distractions**
-
-Get the full LobeHub experience without browser limitations—comprehensive, focused, and always ready to go. Our desktop application provides a dedicated environment for your AI interactions, ensuring optimal performance and minimal distractions.
-
-Experience faster response times, better resource management, and a more stable connection to your AI assistant. The desktop app is designed for users who demand the best performance from their AI tools.
-
-[![][back-to-top]](#readme-top)
-
-![][image-feat-web-search]
-
-### Smart Internet Search
-
-**Online Knowledge On Demand**
-
-With real-time internet access, your AI keeps up with the world—news, data, trends, and more. Stay informed and get the most current information available, enabling your AI to provide accurate and up-to-date responses.
-
-Access live information, verify facts, and explore current events without leaving your conversation. Your AI becomes a gateway to the world's knowledge, always current and comprehensive.
-
-[![][back-to-top]](#readme-top)
-
-[![][image-feat-cot]][docs-feat-cot]
-
-### [Chain of Thought][docs-feat-cot]
-
-Experience AI reasoning like never before. Watch as complex problems unfold step by step through our innovative Chain of Thought (CoT) visualization. This breakthrough feature provides unprecedented transparency into AI's decision-making process, allowing you to observe how conclusions are reached in real-time.
-
-By breaking down complex reasoning into clear, logical steps, you can better understand and validate the AI's problem-solving approach. Whether you're debugging, learning, or simply curious about AI reasoning, CoT visualization transforms abstract thinking into an engaging, interactive experience.
-
-[![][back-to-top]](#readme-top)
-
-[![][image-feat-branch]][docs-feat-branch]
-
-### [Branching Conversations][docs-feat-branch]
-
-Introducing a more natural and flexible way to chat with AI. With Branch Conversations, your discussions can flow in multiple directions, just like human conversations do. Create new conversation branches from any message, giving you the freedom to explore different paths while preserving the original context.
-
-Choose between two powerful modes:
-
- **Continuation Mode:** Seamlessly extend your current discussion while maintaining valuable context
- **Standalone Mode:** Start fresh with a new topic based on any previous message
-
-This groundbreaking feature transforms linear conversations into dynamic, tree-like structures, enabling deeper exploration of ideas and more productive interactions.
-
-[![][back-to-top]](#readme-top)
-
-[![][image-feat-artifacts]][docs-feat-artifacts]
-
-### [Artifacts Support][docs-feat-artifacts]
-
-Experience the power of Claude Artifacts, now integrated into LobeHub. This revolutionary feature expands the boundaries of AI-human interaction, enabling real-time creation and visualization of diverse content formats.
-
-Create and visualize with unprecedented flexibility:
-
- Generate and display dynamic SVG graphics
- Build and render interactive HTML pages in real-time
- Produce professional documents in multiple formats
-
-[![][back-to-top]](#readme-top)
-
-[![][image-feat-knowledgebase]][docs-feat-knowledgebase]
-
-### [File Upload /Knowledge Base][docs-feat-knowledgebase]
-
-LobeHub supports file upload and knowledge base functionality. You can upload various types of files including documents, images, audio, and video, as well as create knowledge bases, making it convenient for users to manage and search for files. Additionally, you can utilize files and knowledge base features during conversations, enabling a richer dialogue experience.
-
-<https://github.com/user-attachments/assets/faa8cf67-e743-4590-8bf6-ebf6ccc34175>
-
-> \[!TIP]
->
-> Learn more on [📘 LobeHub Knowledge Base Launch — From Now On, Every Step Counts](https://lobehub.com/blog/knowledge-base)
+![](https://github.com/user-attachments/assets/5c6e16f0-7f47-4baf-9aeb-3a00deb8ff5b)

 <div align="right">

@@ -289,277 +187,6 @@ LobeHub supports file upload and knowledge base functionality. You can upload va

 </div>

-[![][image-feat-privoder]][docs-feat-provider]
-
-### [Multi-Model Service Provider Support][docs-feat-provider]
-
-In the continuous development of LobeHub, we deeply understand the importance of diversity in model service providers for meeting the needs of the community when providing AI conversation services. Therefore, we have expanded our support to multiple model service providers, rather than being limited to a single one, in order to offer users a more diverse and rich selection of conversations.
-
-In this way, LobeHub can more flexibly adapt to the needs of different users, while also providing developers with a wider range of choices.
-
-#### Supported Model Service Providers
-
-We have implemented support for the following model service providers:
-
-<!-- PROVIDER LIST -->
-
-<details><summary><kbd>See more providers (+-10)</kbd></summary>
-
-</details>
-
-> 📊 Total providers: [<kbd>**0**</kbd>](https://lobechat.com/discover/providers)
-
- <!-- PROVIDER LIST -->
-
-At the same time, we are also planning to support more model service providers. If you would like LobeHub to support your favorite service provider, feel free to join our [💬 community discussion](https://github.com/lobehub/lobehub/discussions/1284).
-
-<div align="right">
-
-[![][back-to-top]](#readme-top)
-
-</div>
-
-[![][image-feat-local]][docs-feat-local]
-
-### [Local Large Language Model (LLM) Support][docs-feat-local]
-
-To meet the specific needs of users, LobeHub also supports the use of local models based on [Ollama](https://ollama.ai), allowing users to flexibly use their own or third-party models.
-
-> \[!TIP]
->
-> Learn more about [📘 Using Ollama in LobeHub][docs-usage-ollama] by checking it out.
-
-<div align="right">
-
-[![][back-to-top]](#readme-top)
-
-</div>
-
-[![][image-feat-vision]][docs-feat-vision]
-
-### [Model Visual Recognition][docs-feat-vision]
-
-LobeHub now supports OpenAI's latest [`gpt-4-vision`](https://platform.openai.com/docs/guides/vision) model with visual recognition capabilities,
-a multimodal intelligence that can perceive visuals. Users can easily upload or drag and drop images into the dialogue box,
-and the agent will be able to recognize the content of the images and engage in intelligent conversation based on this,
-creating smarter and more diversified chat scenarios.
-
-This feature opens up new interactive methods, allowing communication to transcend text and include a wealth of visual elements.
-Whether it's sharing images in daily use or interpreting images within specific industries, the agent provides an outstanding conversational experience.
-
-<div align="right">
-
-[![][back-to-top]](#readme-top)
-
-</div>
-
-[![][image-feat-tts]][docs-feat-tts]
-
-### [TTS & STT Voice Conversation][docs-feat-tts]
-
-LobeHub supports Text-to-Speech (TTS) and Speech-to-Text (STT) technologies, enabling our application to convert text messages into clear voice outputs,
-allowing users to interact with our conversational agent as if they were talking to a real person. Users can choose from a variety of voices to pair with the agent.
-
-Moreover, TTS offers an excellent solution for those who prefer auditory learning or desire to receive information while busy.
-In LobeHub, we have meticulously selected a range of high-quality voice options (OpenAI Audio, Microsoft Edge Speech) to meet the needs of users from different regions and cultural backgrounds.
-Users can choose the voice that suits their personal preferences or specific scenarios, resulting in a personalized communication experience.
-
-<div align="right">
-
-[![][back-to-top]](#readme-top)
-
-</div>
-
-[![][image-feat-t2i]][docs-feat-t2i]
-
-### [Text to Image Generation][docs-feat-t2i]
-
-With support for the latest text-to-image generation technology, LobeHub now allows users to invoke image creation tools directly within conversations with the agent. By leveraging the capabilities of AI tools such as [`DALL-E 3`](https://openai.com/dall-e-3), [`MidJourney`](https://www.midjourney.com/), and [`Pollinations`](https://pollinations.ai/), the agents are now equipped to transform your ideas into images.
-
-This enables a more private and immersive creative process, allowing for the seamless integration of visual storytelling into your personal dialogue with the agent.
-
-<div align="right">
-
-[![][back-to-top]](#readme-top)
-
-</div>
-
-[![][image-feat-plugin]][docs-feat-plugin]
-
-### [Plugin System (Function Calling)][docs-feat-plugin]
-
-The plugin ecosystem of LobeHub is an important extension of its core functionality, greatly enhancing the practicality and flexibility of the LobeHub assistant.
-
-<video controls src="https://github.com/lobehub/lobehub/assets/28616219/f29475a3-f346-4196-a435-41a6373ab9e2" muted="false"></video>
-
-By utilizing plugins, LobeHub assistants can obtain and process real-time information, such as searching for web information and providing users with instant and relevant news.
-
-In addition, these plugins are not limited to news aggregation, but can also extend to other practical functions, such as quickly searching documents, generating images, obtaining data from various platforms like Bilibili, Steam, and interacting with various third-party services.
-
-> \[!TIP]
->
-> Learn more about [📘 Plugin Usage][docs-usage-plugin] by checking it out.
-
-<!-- PLUGIN LIST -->
-
-| Recent Submits                                                                                                             | Description                                                                                                                                     |
-| -------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------- |
-| [Shopping tools](https://lobechat.com/discover/plugin/ShoppingTools)<br/><sup>By **shoppingtools** on **2026-01-12**</sup> | Search for products on eBay & AliExpress, find eBay events & coupons. Get prompt examples.<br/>`shopping` `e-bay` `ali-express` `coupons`       |
-| [SEO Assistant](https://lobechat.com/discover/plugin/seo_assistant)<br/><sup>By **webfx** on **2026-01-12**</sup>          | The SEO Assistant can generate search engine keyword information in order to aid the creation of content.<br/>`seo` `keyword`                   |
-| [Video Captions](https://lobechat.com/discover/plugin/VideoCaptions)<br/><sup>By **maila** on **2025-12-13**</sup>         | Convert Youtube links into transcribed text, enable asking questions, create chapters, and summarize its content.<br/>`video-to-text` `youtube` |
-| [WeatherGPT](https://lobechat.com/discover/plugin/WeatherGPT)<br/><sup>By **steven-tey** on **2025-12-13**</sup>           | Get current weather information for a specific location.<br/>`weather`                                                                          |
-
-> 📊 Total plugins: [<kbd>**40**</kbd>](https://lobechat.com/discover/plugins)
-
- <!-- PLUGIN LIST -->
-
-<div align="right">
-
-[![][back-to-top]](#readme-top)
-
-</div>
-
-[![][image-feat-agent]][docs-feat-agent]
-
-### [Agent Market (GPTs)][docs-feat-agent]
-
-In LobeHub Agent Marketplace, creators can discover a vibrant and innovative community that brings together a multitude of well-designed agents,
-which not only play an important role in work scenarios but also offer great convenience in learning processes.
-Our marketplace is not just a showcase platform but also a collaborative space. Here, everyone can contribute their wisdom and share the agents they have developed.
-
-> \[!TIP]
->
-> By [🤖/🏪 Submit Agents][submit-agents-link], you can easily submit your agent creations to our platform.
-> Importantly, LobeHub has established a sophisticated automated internationalization (i18n) workflow,
-> capable of seamlessly translating your agent into multiple language versions.
-> This means that no matter what language your users speak, they can experience your agent without barriers.
-
-> \[!IMPORTANT]
->
-> We welcome all users to join this growing ecosystem and participate in the iteration and optimization of agents.
-> Together, we can create more interesting, practical, and innovative agents, further enriching the diversity and practicality of the agent offerings.
-
-<!-- AGENT LIST -->
-
-| Recent Submits                                                                                                                                                                 | Description                                                                                                                                                                                                              |
-| ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
-| [Turtle Soup Host](https://lobechat.com/discover/assistant/lateral-thinking-puzzle)<br/><sup>By **[CSY2022](https://github.com/CSY2022)** on **2025-06-19**</sup>              | A turtle soup host needs to provide the scenario, the complete story (truth of the event), and the key point (the condition for guessing correctly).<br/>`turtle-soup` `reasoning` `interaction` `puzzle` `role-playing` |
-| [Academic Writing Assistant](https://lobechat.com/discover/assistant/academic-writing-assistant)<br/><sup>By **[swarfte](https://github.com/swarfte)** on **2025-06-17**</sup> | Expert in academic research paper writing and formal documentation<br/>`academic-writing` `research` `formal-style`                                                                                                      |
-| [Gourmet Reviewer🍟](https://lobechat.com/discover/assistant/food-reviewer)<br/><sup>By **[renhai-lab](https://github.com/renhai-lab)** on **2025-06-17**</sup>                | Food critique expert<br/>`gourmet` `review` `writing`                                                                                                                                                                    |
-| [Minecraft Senior Developer](https://lobechat.com/discover/assistant/java-development)<br/><sup>By **[iamyuuk](https://github.com/iamyuuk)** on **2025-06-17**</sup>           | Expert in advanced Java development and Minecraft mod and server plugin development<br/>`development` `programming` `minecraft` `java`                                                                                   |
-
-> 📊 Total agents: [<kbd>**505**</kbd> ](https://lobechat.com/discover/assistants)
-
- <!-- AGENT LIST -->
-
-<div align="right">
-
-[![][back-to-top]](#readme-top)
-
-</div>
-
-[![][image-feat-database]][docs-feat-database]
-
-### [Support Local / Remote Database][docs-feat-database]
-
-LobeHub supports the use of both server-side and local databases. Depending on your needs, you can choose the appropriate deployment solution:
-
- **Local database**: suitable for users who want more control over their data and privacy protection. LobeHub uses CRDT (Conflict-Free Replicated Data Type) technology to achieve multi-device synchronization. This is an experimental feature aimed at providing a seamless data synchronization experience.
- **Server-side database**: suitable for users who want a more convenient user experience. LobeHub supports PostgreSQL as a server-side database. For detailed documentation on how to configure the server-side database, please visit [Configure Server-side Database](https://lobehub.com/docs/self-hosting/advanced/server-database).
-
-Regardless of which database you choose, LobeHub can provide you with an excellent user experience.
-
-<div align="right">
-
-[![][back-to-top]](#readme-top)
-
-</div>
-
-[![][image-feat-auth]][docs-feat-auth]
-
-### [Support Multi-User Management][docs-feat-auth]
-
-LobeHub supports multi-user management and provides flexible user authentication solutions:
-
- **Better Auth**: LobeHub integrates `Better Auth`, a modern and flexible authentication library that supports multiple authentication methods, including OAuth, email login, credential login, magic links, and more. With `Better Auth`, you can easily implement user registration, login, session management, social login, multi-factor authentication (MFA), and other functions to ensure the security and privacy of user data.
-
-<div align="right">
-
-[![][back-to-top]](#readme-top)
-
-</div>
-
-[![][image-feat-pwa]][docs-feat-pwa]
-
-### [Progressive Web App (PWA)][docs-feat-pwa]
-
-We deeply understand the importance of providing a seamless experience for users in today's multi-device environment.
-Therefore, we have adopted Progressive Web Application ([PWA](https://support.google.com/chrome/answer/9658361)) technology,
-a modern web technology that elevates web applications to an experience close to that of native apps.
-
-Through PWA, LobeHub can offer a highly optimized user experience on both desktop and mobile devices while maintaining high-performance characteristics.
-Visually and in terms of feel, we have also meticulously designed the interface to ensure it is indistinguishable from native apps,
-providing smooth animations, responsive layouts, and adapting to different device screen resolutions.
-
-> \[!NOTE]
->
-> If you are unfamiliar with the installation process of PWA, you can add LobeHub as your desktop application (also applicable to mobile devices) by following these steps:
->
-> - Launch the Chrome or Edge browser on your computer.
-> - Visit the LobeHub webpage.
-> - In the upper right corner of the address bar, click on the <kbd>Install</kbd> icon.
-> - Follow the instructions on the screen to complete the PWA Installation.
-
-<div align="right">
-
-[![][back-to-top]](#readme-top)
-
-</div>
-
-[![][image-feat-mobile]][docs-feat-mobile]
-
-### [Mobile Device Adaptation][docs-feat-mobile]
-
-We have carried out a series of optimization designs for mobile devices to enhance the user's mobile experience. Currently, we are iterating on the mobile user experience to achieve smoother and more intuitive interactions. If you have any suggestions or ideas, we welcome you to provide feedback through GitHub Issues or Pull Requests.
-
-<div align="right">
-
-[![][back-to-top]](#readme-top)
-
-</div>
-
-[![][image-feat-theme]][docs-feat-theme]
-
-### [Custom Themes][docs-feat-theme]
-
-As a design-engineering-oriented application, LobeHub places great emphasis on users' personalized experiences,
-hence introducing flexible and diverse theme modes, including a light mode for daytime and a dark mode for nighttime.
-Beyond switching theme modes, a range of color customization options allow users to adjust the application's theme colors according to their preferences.
-Whether it's a desire for a sober dark blue, a lively peach pink, or a professional gray-white, users can find their style of color choices in LobeHub.
-
-> \[!TIP]
->
-> The default configuration can intelligently recognize the user's system color mode and automatically switch themes to ensure a consistent visual experience with the operating system.
-> For users who like to manually control details, LobeHub also offers intuitive setting options and a choice between chat bubble mode and document mode for conversation scenarios.
-
-<div align="right">
-
-[![][back-to-top]](#readme-top)
-
-</div>
-
-### `*` What's more
-
-Beside these features, LobeHub also have much better basic technique underground:
-
- [x] 💨 **Quick Deployment**: Using the Vercel platform or docker image, you can deploy with just one click and complete the process within 1 minute without any complex configuration.
- [x] 🌐 **Custom Domain**: If users have their own domain, they can bind it to the platform for quick access to the dialogue agent from anywhere.
- [x] 🔒 **Privacy Protection**: All data is stored locally in the user's browser, ensuring user privacy.
- [x] 💎 **Exquisite UI Design**: With a carefully designed interface, it offers an elegant appearance and smooth interaction. It supports light and dark themes and is mobile-friendly. PWA support provides a more native-like experience.
- [x] 🗣️ **Smooth Conversation Experience**: Fluid responses ensure a smooth conversation experience. It fully supports Markdown rendering, including code highlighting, LaTex formulas, Mermaid flowcharts, and more.
-
-</details>
-
 > ✨ more features will be added when LobeHub evolve.

 <div align="right">
@@ -855,28 +482,10 @@ This project is [LobeHub Community License](./LICENSE) licensed.
 [docs-dev-guide]: https://lobehub.com/docs/development/start
 [docs-docker]: https://lobehub.com/docs/self-hosting/server-database/docker-compose
 [docs-env-var]: https://lobehub.com/docs/self-hosting/environment-variables
-[docs-feat-agent]: https://lobehub.com/docs/usage/features/agent-market
-[docs-feat-artifacts]: https://lobehub.com/docs/usage/features/artifacts
-[docs-feat-auth]: https://lobehub.com/docs/usage/features/auth
-[docs-feat-branch]: https://lobehub.com/docs/usage/features/branching-conversations
-[docs-feat-cot]: https://lobehub.com/docs/usage/features/cot
-[docs-feat-database]: https://lobehub.com/docs/usage/features/database
-[docs-feat-knowledgebase]: https://lobehub.com/blog/knowledge-base
-[docs-feat-local]: https://lobehub.com/docs/usage/features/local-llm
-[docs-feat-mobile]: https://lobehub.com/docs/usage/features/mobile
-[docs-feat-plugin]: https://lobehub.com/docs/usage/features/plugin-system
-[docs-feat-provider]: https://lobehub.com/docs/usage/features/multi-ai-providers
-[docs-feat-pwa]: https://lobehub.com/docs/usage/features/pwa
-[docs-feat-t2i]: https://lobehub.com/docs/usage/features/text-to-image
-[docs-feat-theme]: https://lobehub.com/docs/usage/features/theme
-[docs-feat-tts]: https://lobehub.com/docs/usage/features/tts
-[docs-feat-vision]: https://lobehub.com/docs/usage/features/vision
 [docs-function-call]: https://lobehub.com/blog/openai-function-call
 [docs-plugin-dev]: https://lobehub.com/docs/usage/plugins/development
 [docs-self-hosting]: https://lobehub.com/docs/self-hosting/start
 [docs-upstream-sync]: https://lobehub.com/docs/self-hosting/advanced/upstream-sync
-[docs-usage-ollama]: https://lobehub.com/docs/usage/providers/ollama
-[docs-usage-plugin]: https://lobehub.com/docs/usage/plugins/basic
 [fossa-license-link]: https://app.fossa.com/projects/git%2Bgithub.com%2Flobehub%2Flobehub
 [fossa-license-shield]: https://app.fossa.com/api/projects/git%2Bgithub.com%2Flobehub%2Flobehub.svg?type=large
 [github-action-release-link]: https://github.com/actions/workflows/lobehub/lobehub/release.yml
@@ -898,29 +507,7 @@ This project is [LobeHub Community License](./LICENSE) licensed.
 [github-releasedate-shield]: https://img.shields.io/github/release-date/lobehub/lobehub?labelColor=black&style=flat-square
 [github-stars-link]: https://github.com/lobehub/lobehub/stargazers
 [github-stars-shield]: https://img.shields.io/github/stars/lobehub/lobehub?color=ffcb47&labelColor=black&style=flat-square
-[github-trending-shield]: https://trendshift.io/api/badge/repositories/2256
-[github-trending-url]: https://trendshift.io/repositories/2256
-[image-banner]: https://github.com/user-attachments/assets/0fe626a3-0ddc-4f67-b595-3c5b3f1701e0
-[image-feat-agent]: https://github.com/user-attachments/assets/b3ab6e35-4fbc-468d-af10-e3e0c687350f
-[image-feat-artifacts]: https://github.com/user-attachments/assets/7f95fad6-b210-4e6e-84a0-7f39e96f3a00
-[image-feat-auth]: https://github.com/user-attachments/assets/80bb232e-19d1-4f97-98d6-e291f3585e6d
-[image-feat-branch]: https://github.com/user-attachments/assets/92f72082-02bd-4835-9c54-b089aad7fd41
-[image-feat-cot]: https://github.com/user-attachments/assets/f74f1139-d115-4e9c-8c43-040a53797a5e
-[image-feat-database]: https://github.com/user-attachments/assets/f1697c8b-d1fb-4dac-ba05-153c6295d91d
-[image-feat-desktop]: https://github.com/user-attachments/assets/a7bac8d3-ea96-4000-bb39-fadc9b610f96
-[image-feat-knowledgebase]: https://github.com/user-attachments/assets/7da7a3b2-92fd-4630-9f4e-8560c74955ae
-[image-feat-local]: https://github.com/user-attachments/assets/1239da50-d832-4632-a7ef-bd754c0f3850
-[image-feat-mcp]: https://github.com/user-attachments/assets/1be85d36-3975-4413-931f-27e05e440995
-[image-feat-mcp-market]: https://github.com/user-attachments/assets/bb114f9f-24c5-4000-a984-c10d187da5a0
-[image-feat-mobile]: https://github.com/user-attachments/assets/32cf43c4-96bd-4a4c-bfb6-59acde6fe380
-[image-feat-plugin]: https://github.com/user-attachments/assets/66a891ac-01b6-4e3f-b978-2eb07b489b1b
-[image-feat-privoder]: https://github.com/user-attachments/assets/e553e407-42de-4919-977d-7dbfcf44a821
-[image-feat-pwa]: https://github.com/user-attachments/assets/9647f70f-b71b-43b6-9564-7cdd12d1c24d
-[image-feat-t2i]: https://github.com/user-attachments/assets/708274a7-2458-494b-a6ec-b73dfa1fa7c2
-[image-feat-theme]: https://github.com/user-attachments/assets/b47c39f1-806f-492b-8fcb-b0fa973937c1
-[image-feat-tts]: https://github.com/user-attachments/assets/50189597-2cc3-4002-b4c8-756a52ad5c0a
-[image-feat-vision]: https://github.com/user-attachments/assets/18574a1f-46c2-4cbc-af2c-35a86e128a07
-[image-feat-web-search]: https://github.com/user-attachments/assets/cfdc48ac-b5f8-4a00-acee-db8f2eba09ad
+[image-banner]: https://github.com/user-attachments/assets/5f78ae58-ed4f-4d38-8037-96109fbba58c
 [image-star]: https://github.com/user-attachments/assets/3216e25b-186f-4a54-9cb4-2f124aec0471
 [issues-link]: https://img.shields.io/github/issues/lobehub/lobehub.svg?style=flat
 [lobe-chat-plugins]: https://github.com/lobehub/lobe-chat-plugins
@@ -958,8 +545,6 @@ This project is [LobeHub Community License](./LICENSE) licensed.
 [share-whatsapp-shield]: https://img.shields.io/badge/-share%20on%20whatsapp-black?labelColor=black&logo=whatsapp&logoColor=white&style=flat-square
 [share-x-link]: https://x.com/intent/tweet?hashtags=chatbot%2CchatGPT%2CopenAI&text=Check%20this%20GitHub%20repository%20out%20%F0%9F%A4%AF%20LobeHub%20-%20An%20open-source%2C%20extensible%20%28Function%20Calling%29%2C%20high-performance%20chatbot%20framework.%20It%20supports%20one-click%20free%20deployment%20of%20your%20private%20ChatGPT%2FLLM%20web%20application.&url=https%3A%2F%2Fgithub.com%2Flobehub%2Flobehub
 [share-x-shield]: https://img.shields.io/badge/-share%20on%20x-black?labelColor=black&logo=x&logoColor=white&style=flat-square
-[sponsor-link]: https://opencollective.com/lobehub 'Become ❤️ LobeHub Sponsor'
-[sponsor-shield]: https://img.shields.io/badge/-Sponsor%20LobeHub-f04f88?logo=opencollective&logoColor=white&style=flat-square
 [submit-agents-link]: https://github.com/lobehub/lobe-chat-agents
 [submit-agents-shield]: https://img.shields.io/badge/🤖/🏪_submit_agent-%E2%86%92-c4f042?labelColor=black&style=for-the-badge
 [submit-plugin-link]: https://github.com/lobehub/lobe-chat-plugins
@@ -4,8 +4,11 @@

 # LobeHub

-LobeHub 是一个工作与生活空间，用于发现、构建并与会随着您一起成长的 Agent 队友协作。<br/>
-在 LobeHub 中，我们将 **Agent 视为工作单元**，提供一个让人类与 Agent 共同进化的基础设施。
+LobeHub 帮你把专属 Agent 组织成 7×24 不打烊的高效队伍：
+
+自动为你招募适配的 AI 队友、调度任务排班、汇总生成工作报告，
+
+你始终掌控全局，从此不用再时刻在线盯守，真正解放自己的时间。

 [English](./README.md) · **简体中文** · [官网][official-site] · [更新日志][changelog] · [文档][docs] · [博客][blog] · [反馈问题][github-issues-link]

@@ -24,7 +27,6 @@ LobeHub 是一个工作与生活空间，用于发现、构建并与会随着您
 [![][github-stars-shield]][github-stars-link]
 [![][github-issues-shield]][github-issues-link]
 [![][github-license-shield]][github-license-link]<br>
-[![][sponsor-shield]][sponsor-link]

 **分享 LobeHub 给你的好友**

@@ -35,9 +37,9 @@ LobeHub 是一个工作与生活空间，用于发现、构建并与会随着您
 [![][share-weibo-shield]][share-weibo-link]
 [![][share-mastodon-shield]][share-mastodon-link]

-<sup>Agent teammates that grow with you</sup>
+<sup>你的首席 Agent 运营官</sup>

-[![][github-trending-shield]][github-trending-url]
+<a href="https://www.producthunt.com/products/lobehub?embed=true&amp;utm_source=badge-top-post-badge&amp;utm_medium=badge&amp;utm_campaign=badge-lobehub-2" target="_blank" rel="noopener noreferrer"><img alt="LobeHub - Your Chief Agent Operator for multi-agent work | Product Hunt" width="250" height="54" src="https://api.producthunt.com/widgets/embed-image/v1/top-post-badge.svg?post_id=1147569&amp;theme=light&amp;period=daily&amp;t=1779247564355"></a> <a href="https://trendshift.io/repositories/19224" target="_blank"><img src="https://trendshift.io/api/badge/repositories/19224" alt="lobehub%2Flobehub | Trendshift" style="width: 250px; height: 55px;" width="250" height="55"/></a>
 [![][github-hello-shield]][github-hello-url]

 </div>
@@ -49,30 +51,10 @@ LobeHub 是一个工作与生活空间，用于发现、构建并与会随着您

 - [👋🏻 开始使用 & 交流](#-开始使用--交流)
 - [✨ 特性一览](#-特性一览)
+  - [运营：你制定策略，我们负责运行 Agent。](#运营你制定策略我们负责运行-agent)
  - [创建：以 Agent 为工作单元](#创建以-agent-为工作单元)
  - [协作：扩展新型协作网络](#协作扩展新型协作网络)
  - [进化：人类与 Agent 的共生进化](#进化人类与-agent-的共生进化)
-  - [MCP](#mcp)
-  - [发现、连接、扩展](#发现连接扩展)
-  - [巅峰性能，零干扰](#巅峰性能零干扰)
-  - [在线知识，按需获取](#在线知识按需获取)
-  - [思维链 (CoT)](#思维链-cot)
-  - [分支对话](#分支对话)
-  - [支持白板 (Artifacts)](#支持白板-artifacts)
-  - [文件上传 / 知识库](#文件上传--知识库)
-  - [多模型服务商支持](#多模型服务商支持)
-  - [支持本地大语言模型 (LLM)](#支持本地大语言模型-llm)
-  - [模型视觉识别 (Model Visual)](#模型视觉识别-model-visual)
-  - [TTS & STT 语音会话](#tts--stt-语音会话)
-  - [Text to Image 文生图](#text-to-image-文生图)
-  - [插件系统 (Tools Calling)](#插件系统-tools-calling)
-  - [助手市场 (GPTs)](#助手市场-gpts)
-  - [支持本地 / 远程数据库](#支持本地--远程数据库)
-  - [支持多用户管理](#支持多用户管理)
-  - [渐进式 Web 应用 (PWA)](#渐进式-web-应用-pwa)
-  - [移动设备适配](#移动设备适配)
-  - [自定义主题](#自定义主题)
-  - [`*` 更多特性](#-更多特性)
 - [🛳 开箱即用](#-开箱即用)
  - [`A` 使用 Vercel、Zeabur 、Sealos 或 阿里云计算巢 部署](#a-使用-vercelzeabur-sealos-或-阿里云计算巢-部署)
  - [`B` 使用 Docker 部署](#b-使用-docker-部署)
@@ -93,7 +75,7 @@ LobeHub 是一个工作与生活空间，用于发现、构建并与会随着您

 <br/>

-<https://github.com/user-attachments/assets/6710ad97-03d0-4175-bd75-adff9b55eca2>
+<https://github.com/user-attachments/assets/0a33365f-b786-48b5-9ed6-f8af7927bccb>

 ## 👋🏻 开始使用 & 交流

@@ -102,9 +84,9 @@ LobeHub 是一个工作与生活空间，用于发现、构建并与会随着您

 不论普通用户与专业开发者，LobeHub 旨在成为所有人的 AI Agent 实验场。LobeHub 目前正在积极开发中，有任何需求或者问题，欢迎提交 [issues][issues-link]

-| [![](https://api.producthunt.com/widgets/embed-image/v1/featured.svg?post_id=1065874&theme=light&t=1769347414733)](https://www.producthunt.com/products/lobehub?embed=true&utm_source=badge-featured&utm_medium=badge&utm_campaign=badge-lobehub) | 我们已在 Product Hunt 上线！我们很高兴将 LobeHub 推向世界。如果您相信人类与 Agent 共同进化的未来，请支持我们的旅程。 |
-| :------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | :------------------------------------------------------------------------------------------------------------------- |
-| [![][discord-shield-badge]][discord-link]                                                                                                                                                                                                         | 加入我们的 Discord 社区！这是你可以与开发者和其他 LobeHub 热衷用户交流的地方                                         |
+| [![](https://api.producthunt.com/widgets/embed-image/v1/featured.svg?post_id=1065874&theme=light&t=1769347414733)](https://www.producthunt.com/products/lobehub?launch=lobehub-2&embed=true&utm_source=badge-featured&utm_medium=badge&utm_campaign=badge-lobehub) | 我们已在 Product Hunt 上线！我们很高兴将 LobeHub 推向世界。如果您相信人类与 Agent 共同进化的未来，请支持我们的旅程。 |
+| :----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | :------------------------------------------------------------------------------------------------------------------- |
+| [![][discord-shield-badge]][discord-link]                                                                                                                                                                                                                          | 加入我们的 Discord 社区！这是你可以与开发者和其他 LobeHub 热衷用户交流的地方                                         |

 > \[!IMPORTANT]
 >
@@ -127,7 +109,26 @@ LobeHub 是一个工作与生活空间，用于发现、构建并与会随着您

 LobeHub 是一个工作与生活空间，用于发现、构建并与会随着您一起成长的 Agent 队友协作。在 LobeHub 中，我们将 **Agent 视为工作单元**，提供一个让人类与 Agent 共同进化的基础设施。

-![](https://hub-apac-1.lobeobjects.space/blog/assets/2204cde2228fb3f583f3f2c090bc49fb.webp)
+![](https://github.com/user-attachments/assets/89d1c402-a62b-4794-82ea-17e5ee1a6165)
+
+### 运营：你制定策略，我们负责运行 Agent。
+
+雇用、排程并汇报你整个 AI 团队的工作
+
+- **更高生产力，更少工具**：将你所有的 Agent 集中在一个平台。
+- **IM 网关**： Agent 连接到您每天使用的技能。
+
+![](https://github.com/user-attachments/assets/7b08d6d9-9dff-4b06-a919-324630554509)
+
+[![][back-to-top]](#readme-top)
+
+<div align="right">
+
+[![][back-to-top]](#readme-top)
+
+</div>
+
+![](https://github.com/user-attachments/assets/81e89324-fc66-4024-99a3-aa8e16ec8184)

 ### 创建：以 Agent 为工作单元

@@ -136,6 +137,8 @@ LobeHub 是一个工作与生活空间，用于发现、构建并与会随着您
 - **统一智能**：无缝访问任何模型与任何模态 —— 全部由您掌控。
 - **1 万 + 技能**：通过超过 10,000 个工具和与 MCP 兼容的插件，将 Agent 连接到您每天使用的技能。

+![](https://github.com/user-attachments/assets/949b8166-486d-4750-ad7a-cfe7bfcb84e3)
+
 [![][back-to-top]](#readme-top)

 <div align="right">
@@ -155,6 +158,8 @@ LobeHub 引入了 **Agent Groups**，让您可以像对待真实队友一样与
 - **项目（Project）**：按项目组织工作，保持一切结构化且易于跟踪。
 - **工作区（Workspace）**：供团队与 Agent 协作的共享空间，确保明确的所有权和组织内的可见性。

+![](https://github.com/user-attachments/assets/e51526c6-e09c-4a5a-9cec-dcd3fd68a3a8)
+
 [![][back-to-top]](#readme-top)

 <div align="right">
@@ -172,105 +177,7 @@ LobeHub 引入了 **Agent Groups**，让您可以像对待真实队友一样与
 - **持续学习**：您的 Agent 会从您的工作方式中学习，调整其行为以在恰当时刻采取行动。
 - **白盒记忆**：我们相信透明性。您的 Agent 使用结构化、可编辑的记忆，让您完全掌控它们记住的内容。

-<div align="right">
-
-[![][back-to-top]](#readme-top)
-
-</div>
-
-<details>
-<summary>更多特性</summary>
-
-[![](https://github.com/user-attachments/assets/1be85d36-3975-4413-931f-27e05e440995)](https://lobehub.com/mcp)
-
-### MCP
-
-通过启用与外部工具、数据源和服务的平滑、安全和动态交互，释放你的 AI 的全部潜力。基于 MCP（模型上下文协议）的插件系统打破了 AI 与数字生态系统之间的壁垒，实现了前所未有的连接性和功能性。
-
-将对话转化为强大的工作流程，连接数据库、API、文件系统等。体验真正理解并与你的世界互动的 AI Agent。
-
-[![][back-to-top]](#readme-top)
-
-![][image-feat-mcp-market]
-
-### 发现、连接、扩展
-
-浏览不断增长的 MCP 插件库，轻松扩展你的 AI 能力并简化工作流程。访问 [lobehub.com/mcp](https://lobehub.com/mcp) 探索 MCP 市场，提供精选的集成集合，增强你的 AI 与各种工具和服务协作的能力。
-
-从生产力工具到开发环境，发现扩展 AI 覆盖范围和效率的新方式。与社区连接，找到满足特定需求的完美插件。
-
-[![][back-to-top]](#readme-top)
-
-![][image-feat-desktop]
-
-### 巅峰性能，零干扰
-
-获得完整的 LobeHub 体验，摆脱浏览器限制 —— 轻量级、专注且随时就绪。我们的桌面应用程序为你的 AI 交互提供专用环境，确保最佳性能和最小干扰。
-
-体验更快的响应时间、更好的资源管理和与 AI 助手的更稳定连接。桌面应用专为要求 AI 工具最佳性能的用户设计。
-
-[![][back-to-top]](#readme-top)
-
-![][image-feat-web-search]
-
-### 在线知识，按需获取
-
-通过实时联网访问，你的 AI 与世界保持同步 —— 新闻、数据、趋势等。保持信息更新，获取最新可用信息，使你的 AI 能够提供准确和最新的回复。
-
-访问实时信息，验证事实，探索当前事件，无需离开对话。你的 AI 成为通向世界知识的门户，始终保持最新和全面。
-
-[![][back-to-top]](#readme-top)
-
-[![][image-feat-cot]][docs-feat-cot]
-
-### [思维链 (CoT)][docs-feat-cot]
-
-体验前所未有的 AI 推理过程。通过创新的思维链（CoT）可视化功能，您可以实时观察复杂问题是如何一步步被解析的。这项突破性的功能为 AI 的决策过程提供了前所未有的透明度，让您能够清晰地了解结论是如何得出的。
-
-通过将复杂的推理过程分解为清晰的逻辑步骤，您可以更好地理解和验证 AI 的解题思路。无论您是在调试问题、学习知识，还是单纯对 AI 推理感兴趣，思维链可视化都能将抽象思维转化为一种引人入胜的互动体验。
-
-[![][back-to-top]](#readme-top)
-
-[![][image-feat-branch]][docs-feat-branch]
-
-### [分支对话][docs-feat-branch]
-
-为您带来更自然、更灵活的 AI 对话方式。通过分支对话功能，您的讨论可以像人类对话一样自然延伸。在任意消息处创建新的对话分支，让您在保留原有上下文的同时，自由探索不同的对话方向。
-
-两种强大模式任您选择：
-
- **延续模式**：无缝延展当前讨论，保持宝贵的对话上下文
- **独立模式**：基于任意历史消息，开启全新话题探讨
-
-这项突破性功能将线性对话转变为动态的树状结构，让您能够更深入地探索想法，实现更高效的互动体验。
-
-[![][back-to-top]](#readme-top)
-
-[![][image-feat-artifacts]][docs-feat-artifacts]
-
-### [支持白板 (Artifacts)][docs-feat-artifacts]
-
-体验集成于 LobeHub 的 Claude Artifacts 能力。这项革命性功能突破了 AI 人机交互的边界，让您能够实时创建和可视化各种格式的内容。
-
-以前所未有的灵活度进行创作与可视化：
-
- 生成并展示动态 SVG 图形
- 实时构建与渲染交互式 HTML 页面
- 输出多种格式的专业文档
-
-[![][back-to-top]](#readme-top)
-
-[![][image-feat-knowledgebase]][docs-feat-knowledgebase]
-
-### [文件上传 / 知识库][docs-feat-knowledgebase]
-
-LobeHub 支持文件上传与知识库功能，你可以上传文件、图片、音频、视频等多种类型的文件，以及创建知识库，方便用户管理和查找文件。同时在对话中使用文件和知识库功能，实现更加丰富的对话体验。
-
-<https://github.com/user-attachments/assets/faa8cf67-e743-4590-8bf6-ebf6ccc34175>
-
-> \[!TIP]
->
-> 查阅 [📘 LobeHub 知识库上线 —— 此刻起，跬步千里](https://lobehub.com/zh/blog/knowledge-base) 了解详情。
+![](https://github.com/user-attachments/assets/5c6e16f0-7f47-4baf-9aeb-3a00deb8ff5b)

 <div align="right">

@@ -278,262 +185,6 @@ LobeHub 支持文件上传与知识库功能，你可以上传文件、图片、

 </div>

-[![][image-feat-privoder]][docs-feat-provider]
-
-### [多模型服务商支持][docs-feat-provider]
-
-在 LobeHub 的不断发展过程中，我们深刻理解到在提供 AI 会话服务时模型服务商的多样性对于满足社区需求的重要性。因此，我们不再局限于单一的模型服务商，而是拓展了对多种模型服务商的支持，以便为用户提供更为丰富和多样化的会话选择。
-
-通过这种方式，LobeHub 能够更灵活地适应不同用户的需求，同时也为开发者提供了更为广泛的选择空间。
-
-#### 已支持的模型服务商
-
-我们已经实现了对以下模型服务商的支持：
-
-<!-- PROVIDER LIST -->
-
-<details><summary><kbd>See more providers (+-10)</kbd></summary>
-
-</details>
-
-> 📊 Total providers: [<kbd>**0**</kbd>](https://lobechat.com/discover/providers)
-
- <!-- PROVIDER LIST -->
-
-同时，我们也在计划支持更多的模型服务商，以进一步丰富我们的服务商库。如果你希望让 LobeHub 支持你喜爱的服务商，欢迎加入我们的 [💬 社区讨论](https://github.com/lobehub/lobehub/discussions/6157)。
-
-<div align="right">
-
-[![][back-to-top]](#readme-top)
-
-</div>
-
-[![][image-feat-local]][docs-feat-local]
-
-### [支持本地大语言模型 (LLM)][docs-feat-local]
-
-为了满足特定用户的需求，LobeHub 还基于 [Ollama](https://ollama.ai) 支持了本地模型的使用，让用户能够更灵活地使用自己的或第三方的模型。
-
-> \[!TIP]
->
-> 查阅 [📘 在 LobeHub 中使用 Ollama][docs-usage-ollama] 获得更多信息
-
-<div align="right">
-
-[![][back-to-top]](#readme-top)
-
-</div>
-
-[![][image-feat-vision]][docs-feat-vision]
-
-### [模型视觉识别 (Model Visual)][docs-feat-vision]
-
-LobeHub 已经支持 OpenAI 最新的 [`gpt-4-vision`](https://platform.openai.com/docs/guides/vision) 支持视觉识别的模型，这是一个具备视觉识别能力的多模态应用。
-用户可以轻松上传图片或者拖拽图片到对话框中，助手将能够识别图片内容，并在此基础上进行智能对话，构建更智能、更多元化的聊天场景。
-
-这一特性打开了新的互动方式，使得交流不再局限于文字，而是可以涵盖丰富的视觉元素。无论是日常使用中的图片分享，还是在特定行业内的图像解读，助手都能提供出色的对话体验。
-
-<div align="right">
-
-[![][back-to-top]](#readme-top)
-
-</div>
-
-[![][image-feat-tts]][docs-feat-tts]
-
-### [TTS & STT 语音会话][docs-feat-tts]
-
-LobeHub 支持文字转语音（Text-to-Speech，TTS）和语音转文字（Speech-to-Text，STT）技术，这使得我们的应用能够将文本信息转化为清晰的语音输出，用户可以像与真人交谈一样与我们的对话助手进行交流。
-用户可以从多种声音中选择，给助手搭配合适的音源。 同时，对于那些倾向于听觉学习或者想要在忙碌中获取信息的用户来说，TTS 提供了一个极佳的解决方案。
-
-在 LobeHub 中，我们精心挑选了一系列高品质的声音选项 (OpenAI Audio, Microsoft Edge Speech)，以满足不同地域和文化背景用户的需求。用户可以根据个人喜好或者特定场景来选择合适的语音，从而获得个性化的交流体验。
-
-<div align="right">
-
-[![][back-to-top]](#readme-top)
-
-</div>
-
-[![][image-feat-t2i]][docs-feat-t2i]
-
-### [Text to Image 文生图][docs-feat-t2i]
-
-支持最新的文本到图片生成技术，LobeHub 现在能够让用户在与助手对话中直接调用文生图工具进行创作。
-通过利用 [`DALL-E 3`](https://openai.com/dall-e-3)、[`MidJourney`](https://www.midjourney.com/) 和 [`Pollinations`](https://pollinations.ai/) 等 AI 工具的能力， 助手们现在可以将你的想法转化为图像。
-同时可以更私密和沉浸式地完成你的创作过程。
-
-<div align="right">
-
-[![][back-to-top]](#readme-top)
-
-</div>
-
-[![][image-feat-plugin]][docs-feat-plugin]
-
-### [插件系统 (Tools Calling)][docs-feat-plugin]
-
-LobeHub 的插件生态系统是其核心功能的重要扩展，它极大地增强了 ChatGPT 的实用性和灵活性。
-
-<video controls src="https://github.com/lobehub/lobehub/assets/28616219/f29475a3-f346-4196-a435-41a6373ab9e2" muted="false"></video>
-
-通过利用插件，ChatGPT 能够实现实时信息的获取和处理，例如自动获取最新新闻头条，为用户提供即时且相关的资讯。
-
-此外，这些插件不仅局限于新闻聚合，还可以扩展到其他实用的功能，如快速检索文档、生成图象、获取电商平台数据，以及其他各式各样的第三方服务。
-
-> 通过文档了解更多 [📘 插件使用][docs-usage-plugin]
-
-<!-- PLUGIN LIST -->
-
-| 最近新增                                                                                                             | 描述                                                                                                               |
-| -------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------ |
-| [购物工具](https://lobechat.com/discover/plugin/ShoppingTools)<br/><sup>By **shoppingtools** on **2026-01-12**</sup> | 在 eBay 和 AliExpress 上搜索产品，查找 eBay 活动和优惠券。获取快速示例。<br/>`购物` `e-bay` `ali-express` `优惠券` |
-| [SEO 助手](https://lobechat.com/discover/plugin/seo_assistant)<br/><sup>By **webfx** on **2026-01-12**</sup>         | SEO 助手可以生成搜索引擎关键词信息，以帮助创建内容。<br/>`seo` `关键词`                                            |
-| [视频字幕](https://lobechat.com/discover/plugin/VideoCaptions)<br/><sup>By **maila** on **2025-12-13**</sup>         | 将 Youtube 链接转换为转录文本，使其能够提问，创建章节，并总结其内容。<br/>`视频转文字` `you-tube`                  |
-| [天气 GPT](https://lobechat.com/discover/plugin/WeatherGPT)<br/><sup>By **steven-tey** on **2025-12-13**</sup>       | 获取特定位置的当前天气信息。<br/>`天气`                                                                            |
-
-> 📊 Total plugins: [<kbd>**40**</kbd>](https://lobechat.com/discover/plugins)
-
- <!-- PLUGIN LIST -->
-
-<div align="right">
-
-[![][back-to-top]](#readme-top)
-
-</div>
-
-[![][image-feat-agent]][docs-feat-agent]
-
-### [助手市场 (GPTs)][docs-feat-agent]
-
-在 LobeHub 的助手市场中，创作者们可以发现一个充满活力和创新的社区，它汇聚了众多精心设计的助手，这些助手不仅在工作场景中发挥着重要作用，也在学习过程中提供了极大的便利。
-我们的市场不仅是一个展示平台，更是一个协作的空间。在这里，每个人都可以贡献自己的智慧，分享个人开发的助手。
-
-> \[!TIP]
->
-> 通过 [🤖/🏪 提交助手][submit-agents-link] ，你可以轻松地将你的助手作品提交到我们的平台。我们特别强调的是，LobeHub 建立了一套精密的自动化国际化（i18n）工作流程， 它的强大之处在于能够无缝地将你的助手转化为多种语言版本。
-> 这意味着，不论你的用户使用何种语言，他们都能无障碍地体验到你的助手。
-
-> \[!IMPORTANT]
->
-> 我欢迎所有用户加入这个不断成长的生态系统，共同参与到助手的迭代与优化中来。共同创造出更多有趣、实用且具有创新性的助手，进一步丰富助手的多样性和实用性。
-
-<!-- AGENT LIST -->
-
-| 最近新增                                                                                                                                                         | 描述                                                                                                              |
-| ---------------------------------------------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------- |
-| [海龟汤主持人](https://lobechat.com/discover/assistant/lateral-thinking-puzzle)<br/><sup>By **[CSY2022](https://github.com/CSY2022)** on **2025-06-19**</sup>    | 一个海龟汤主持人，需要自己提供汤面，汤底与关键点（猜中的判定条件）。<br/>`海龟汤` `推理` `互动` `谜题` `角色扮演` |
-| [学术写作助手](https://lobechat.com/discover/assistant/academic-writing-assistant)<br/><sup>By **[swarfte](https://github.com/swarfte)** on **2025-06-17**</sup> | 专业的学术研究论文写作和正式文档编写专家<br/>`学术写作` `研究` `正式风格`                                         |
-| [美食评论员🍟](https://lobechat.com/discover/assistant/food-reviewer)<br/><sup>By **[renhai-lab](https://github.com/renhai-lab)** on **2025-06-17**</sup>        | 美食评价专家<br/>`美食` `评价` `写作`                                                                             |
-| [Minecraft 资深开发者](https://lobechat.com/discover/assistant/java-development)<br/><sup>By **[iamyuuk](https://github.com/iamyuuk)** on **2025-06-17**</sup>   | 擅长高级 Java 开发及 Minecraft 开发<br/>`开发` `编程` `minecraft` `java`                                          |
-
-> 📊 Total agents: [<kbd>**505**</kbd> ](https://lobechat.com/discover/assistants)
-
- <!-- AGENT LIST -->
-
-<div align="right">
-
-[![][back-to-top]](#readme-top)
-
-</div>
-
-[![][image-feat-database]][docs-feat-database]
-
-### [支持本地 / 远程数据库][docs-feat-database]
-
-LobeHub 支持同时使用服务端数据库和本地数据库。根据您的需求，您可以选择合适的部署方案：
-
- 本地数据库：适合希望对数据有更多掌控感和隐私保护的用户。LobeHub 采用了 CRDT (Conflict-Free Replicated Data Type) 技术，实现了多端同步功能。这是一项实验性功能，旨在提供无缝的数据同步体验。
- 服务端数据库：适合希望更便捷使用体验的用户。LobeHub 支持 PostgreSQL 作为服务端数据库。关于如何配置服务端数据库的详细文档，请前往 [配置服务端数据库](https://lobehub.com/zh/docs/self-hosting/advanced/server-database)。
-
-无论您选择哪种数据库，LobeHub 都能为您提供卓越的用户体验。
-
-<div align="right">
-
-[![][back-to-top]](#readme-top)
-
-</div>
-
-[![][image-feat-auth]][docs-feat-auth]
-
-### [支持多用户管理][docs-feat-auth]
-
-LobeHub 支持多用户管理，提供了灵活的用户认证方案：
-
- **Better Auth**：LobeHub 集成了 `Better Auth`，一个现代化且灵活的身份验证库，支持多种身份验证方式，包括 OAuth、邮件登录、凭证登录、魔法链接等。通过 `Better Auth`，您可以轻松实现用户的注册、登录、会话管理、社交登录、多因素认证 (MFA) 等功能，确保用户数据的安全性和隐私性。
-
-<div align="right">
-
-[![][back-to-top]](#readme-top)
-
-</div>
-
-[![][image-feat-pwa]][docs-feat-pwa]
-
-### [渐进式 Web 应用 (PWA)][docs-feat-pwa]
-
-我们深知在当今多设备环境下为用户提供无缝体验的重要性。为此，我们采用了渐进式 Web 应用 [PWA](https://support.google.com/chrome/answer/9658361) 技术，
-这是一种能够将网页应用提升至接近原生应用体验的现代 Web 技术。通过 PWA，LobeHub 能够在桌面和移动设备上提供高度优化的用户体验，同时保持轻量级和高性能的特点。
-在视觉和感觉上，我们也经过精心设计，以确保它的界面与原生应用无差别，提供流畅的动画、响应式布局和适配不同设备的屏幕分辨率。
-
-> \[!NOTE]
->
-> 若您未熟悉 PWA 的安装过程，您可以按照以下步骤将 LobeHub 添加为您的桌面应用（也适用于移动设备）：
->
-> - 在电脑上运行 Chrome 或 Edge 浏览器 .
-> - 访问 LobeHub 网页 .
-> - 在地址栏的右上角，单击 <kbd>安装</kbd> 图标 .
-
-<div align="right">
-
-[![][back-to-top]](#readme-top)
-
-</div>
-
-[![][image-feat-mobile]][docs-feat-mobile]
-
-### [移动设备适配][docs-feat-mobile]
-
-针对移动设备进行了一系列的优化设计，以提升用户的移动体验。目前，我们正在对移动端的用户体验进行版本迭代，以实现更加流畅和直观的交互。如果您有任何建议或想法，我们非常欢迎您通过 GitHub Issues 或者 Pull Requests 提供反馈。
-
-<div align="right">
-
-[![][back-to-top]](#readme-top)
-
-</div>
-
-[![][image-feat-theme]][docs-feat-theme]
-
-### [自定义主题][docs-feat-theme]
-
-作为设计工程师出身，LobeHub 在界面设计上充分考虑用户的个性化体验，因此引入了灵活多变的主题模式，其中包括日间的亮色模式和夜间的深色模式。
-除了主题模式的切换，还提供了一系列的颜色定制选项，允许用户根据自己的喜好来调整应用的主题色彩。无论是想要沉稳的深蓝，还是希望活泼的桃粉，或者是专业的灰白，用户都能够在 LobeHub 中找到匹配自己风格的颜色选择。
-
-> \[!TIP]
->
-> 默认配置能够智能地识别用户系统的颜色模式，自动进行主题切换，以确保应用界面与操作系统保持一致的视觉体验。对于喜欢手动调控细节的用户，LobeHub 同样提供了直观的设置选项，针对聊天场景也提供了对话气泡模式和文档模式的选择。
-
-<div align="right">
-
-<div align="right">
-
-[![][back-to-top]](#readme-top)
-
-</div>
-
-</div>
-
-### `*` 更多特性
-
-除了上述功能特性以外，LobeHub 所具有的设计和技术能力将为你带来更多使用保障：
-
- [x] 💎 **精致 UI 设计**：经过精心设计的界面，具有优雅的外观和流畅的交互效果，支持亮暗色主题，适配移动端。支持 PWA，提供更加接近原生应用的体验。
- [x] 🗣️ **流畅的对话体验**：流式响应带来流畅的对话体验，并且支持完整的 Markdown 渲染，包括代码高亮、LaTex 公式、Mermaid 流程图等。
- [x] 💨 **快速部署**：使用 Vercel 平台或者我们的 Docker 镜像，只需点击一键部署按钮，即可在 1 分钟内完成部署，无需复杂的配置过程。
- [x] 🔒 **隐私安全**：所有数据保存在用户浏览器本地，保证用户的隐私安全。
- [x] 🌐 **自定义域名**：如果用户拥有自己的域名，可以将其绑定到平台上，方便在任何地方快速访问对话助手。
-
-</details>
-
 > ✨ 随着产品迭代持续更新，我们将会带来更多更多令人激动的功能！

 <div align="right">
@@ -867,28 +518,10 @@ This project is [LobeHub Community License](./LICENSE) licensed.
 [docs-dev-guide]: https://lobehub.com/docs/development/start
 [docs-docker]: https://lobehub.com/zh/docs/self-hosting/server-database/docker-compose
 [docs-env-var]: https://lobehub.com/docs/self-hosting/environment-variables
-[docs-feat-agent]: https://lobehub.com/docs/usage/features/agent-market
-[docs-feat-artifacts]: https://lobehub.com/docs/usage/features/artifacts
-[docs-feat-auth]: https://lobehub.com/docs/usage/features/auth
-[docs-feat-branch]: https://lobehub.com/docs/usage/features/branching-conversations
-[docs-feat-cot]: https://lobehub.com/docs/usage/features/cot
-[docs-feat-database]: https://lobehub.com/docs/usage/features/database
-[docs-feat-knowledgebase]: https://lobehub.com/blog/knowledge-base
-[docs-feat-local]: https://lobehub.com/docs/usage/features/local-llm
-[docs-feat-mobile]: https://lobehub.com/docs/usage/features/mobile
-[docs-feat-plugin]: https://lobehub.com/docs/usage/features/plugin-system
-[docs-feat-provider]: https://lobehub.com/docs/usage/features/multi-ai-providers
-[docs-feat-pwa]: https://lobehub.com/docs/usage/features/pwa
-[docs-feat-t2i]: https://lobehub.com/docs/usage/features/text-to-image
-[docs-feat-theme]: https://lobehub.com/docs/usage/features/theme
-[docs-feat-tts]: https://lobehub.com/docs/usage/features/tts
-[docs-feat-vision]: https://lobehub.com/docs/usage/features/vision
 [docs-function-call]: https://lobehub.com/zh/blog/openai-function-call
 [docs-plugin-dev]: https://lobehub.com/docs/usage/plugins/development
 [docs-self-hosting]: https://lobehub.com/docs/self-hosting/start
 [docs-upstream-sync]: https://lobehub.com/docs/self-hosting/advanced/upstream-sync
-[docs-usage-ollama]: https://lobehub.com/docs/usage/providers/ollama
-[docs-usage-plugin]: https://lobehub.com/docs/usage/plugins/basic
 [fossa-license-link]: https://app.fossa.com/projects/git%2Bgithub.com%2Flobehub%2Flobehub
 [fossa-license-shield]: https://app.fossa.com/api/projects/git%2Bgithub.com%2Flobehub%2Flobehub.svg?type=large
 [github-action-release-link]: https://github.com/lobehub/lobehub/actions/workflows/release.yml
@@ -911,29 +544,8 @@ This project is [LobeHub Community License](./LICENSE) licensed.
 [github-releasedate-link]: https://github.com/lobehub/lobehub/releases
 [github-releasedate-shield]: https://img.shields.io/github/release-date/lobehub/lobehub?labelColor=black&style=flat-square
 [github-stars-link]: https://github.com/lobehub/lobehub/stargazers
-[github-stars-shield]: https://github.com/user-attachments/assets/3216e25b-186f-4a54-9cb4-2f124aec0471
-[github-trending-shield]: https://trendshift.io/api/badge/repositories/2256
-[github-trending-url]: https://trendshift.io/repositories/2256
-[image-banner]: https://github.com/user-attachments/assets/0fe626a3-0ddc-4f67-b595-3c5b3f1701e0
-[image-feat-agent]: https://github.com/user-attachments/assets/b3ab6e35-4fbc-468d-af10-e3e0c687350f
-[image-feat-artifacts]: https://github.com/user-attachments/assets/7f95fad6-b210-4e6e-84a0-7f39e96f3a00
-[image-feat-auth]: https://github.com/user-attachments/assets/80bb232e-19d1-4f97-98d6-e291f3585e6d
-[image-feat-branch]: https://github.com/user-attachments/assets/92f72082-02bd-4835-9c54-b089aad7fd41
-[image-feat-cot]: https://github.com/user-attachments/assets/f74f1139-d115-4e9c-8c43-040a53797a5e
-[image-feat-database]: https://github.com/user-attachments/assets/f1697c8b-d1fb-4dac-ba05-153c6295d91d
-[image-feat-desktop]: https://github.com/user-attachments/assets/a7bac8d3-ea96-4000-bb39-fadc9b610f96
-[image-feat-knowledgebase]: https://github.com/user-attachments/assets/7da7a3b2-92fd-4630-9f4e-8560c74955ae
-[image-feat-local]: https://github.com/user-attachments/assets/1239da50-d832-4632-a7ef-bd754c0f3850
-[image-feat-mcp-market]: https://github.com/user-attachments/assets/bb114f9f-24c5-4000-a984-c10d187da5a0
-[image-feat-mobile]: https://github.com/user-attachments/assets/32cf43c4-96bd-4a4c-bfb6-59acde6fe380
-[image-feat-plugin]: https://github.com/user-attachments/assets/66a891ac-01b6-4e3f-b978-2eb07b489b1b
-[image-feat-privoder]: https://github.com/user-attachments/assets/e553e407-42de-4919-977d-7dbfcf44a821
-[image-feat-pwa]: https://github.com/user-attachments/assets/9647f70f-b71b-43b6-9564-7cdd12d1c24d
-[image-feat-t2i]: https://github.com/user-attachments/assets/708274a7-2458-494b-a6ec-b73dfa1fa7c2
-[image-feat-theme]: https://github.com/user-attachments/assets/b47c39f1-806f-492b-8fcb-b0fa973937c1
-[image-feat-tts]: https://github.com/user-attachments/assets/50189597-2cc3-4002-b4c8-756a52ad5c0a
-[image-feat-vision]: https://github.com/user-attachments/assets/18574a1f-46c2-4cbc-af2c-35a86e128a07
-[image-feat-web-search]: https://github.com/user-attachments/assets/cfdc48ac-b5f8-4a00-acee-db8f2eba09ad
+[github-stars-shield]: https://img.shields.io/github/stars/lobehub/lobehub?color=ffcb47&labelColor=black&style=flat-square
+[image-banner]: https://github.com/user-attachments/assets/5f78ae58-ed4f-4d38-8037-96109fbba58c
 [image-star]: https://github.com/user-attachments/assets/c3b482e7-cef5-4e94-bef9-226900ecfaab
 [issues-link]: https://img.shields.io/github/issues/lobehub/lobehub.svg?style=flat
 [lobe-chat-plugins]: https://github.com/lobehub/lobe-chat-plugins
@@ -969,8 +581,6 @@ This project is [LobeHub Community License](./LICENSE) licensed.
 [share-whatsapp-shield]: https://img.shields.io/badge/-share%20on%20whatsapp-black?labelColor=black&logo=whatsapp&logoColor=white&style=flat-square
 [share-x-link]: https://x.com/intent/tweet?hashtags=chatbot%2CchatGPT%2CopenAI&text=%E6%8E%A8%E8%8D%90%E4%B8%80%E4%B8%AA%20GitHub%20%E5%BC%80%E6%BA%90%E9%A1%B9%E7%9B%AE%20%F0%9F%A4%AF%20LobeHub%20-%20%E5%BC%80%E6%BA%90%E7%9A%84%E3%80%81%E5%8F%AF%E6%89%A9%E5%B1%95%E7%9A%84%EF%BC%88Function%20Calling%EF%BC%89%E9%AB%98%E6%80%A7%E8%83%BD%E8%81%8A%E5%A4%A9%E6%9C%BA%E5%99%A8%E4%BA%BA%E6%A1%86%E6%9E%B6%E3%80%82%0A%E5%AE%83%E6%94%AF%E6%8C%81%E4%B8%80%E9%94%AE%E5%85%8D%E8%B4%B9%E9%83%A8%E7%BD%B2%E7%A7%81%E4%BA%BA%20ChatGPT%2FLLM%20%E7%BD%91%E9%A1%B5%E5%BA%94%E7%94%A8%E7%A8%8B%E5%BA%8F&url=https%3A%2F%2Fgithub.com%2Flobehub%2Flobehub
 [share-x-shield]: https://img.shields.io/badge/-share%20on%20x-black?labelColor=black&logo=x&logoColor=white&style=flat-square
-[sponsor-link]: https://opencollective.com/lobehub 'Become ❤ LobeHub Sponsor'
-[sponsor-shield]: https://img.shields.io/badge/-Sponsor%20LobeHub-f04f88?logo=opencollective&logoColor=white&style=flat-square
 [submit-agents-link]: https://github.com/lobehub/lobe-chat-agents
 [submit-agents-shield]: https://img.shields.io/badge/🤖/🏪_submit_agent-%E2%86%92-c4f042?labelColor=black&style=for-the-badge
 [submit-plugin-link]: https://github.com/lobehub/lobe-chat-plugins
@@ -1,6 +1,6 @@
 .\" Code generated by `npm run man:generate`; DO NOT EDIT.
 .\" Manual command details come from the Commander command tree.
-.TH LH 1 "" "@lobehub/cli 0.0.15" "User Commands"
+.TH LH 1 "" "@lobehub/cli 0.0.22" "User Commands"
 .SH NAME
 lh \- LobeHub CLI \- manage and connect to LobeHub services
 .SH SYNOPSIS
@@ -1,6 +1,6 @@
 {
  "name": "@lobehub/cli",
-  "version": "0.0.15",
+  "version": "0.0.22",
  "type": "module",
  "bin": {
    "lh": "./dist/index.js",
@@ -30,6 +30,7 @@
  "devDependencies": {
    "@lobechat/agent-gateway-client": "workspace:*",
    "@lobechat/device-gateway-client": "workspace:*",
+    "@lobechat/device-identity": "workspace:*",
    "@lobechat/heterogeneous-agents": "workspace:*",
    "@lobechat/local-file-shell": "workspace:*",
    "@trpc/client": "^11.8.1",
@@ -44,7 +45,7 @@
    "picocolors": "^1.1.1",
    "superjson": "^2.2.6",
    "tsdown": "^0.21.4",
-    "typescript": "^5.9.3",
+    "typescript": "^6.0.3",
    "ws": "^8.18.1"
  },
  "publishConfig": {
@@ -1,6 +1,7 @@
 packages:
  - '../../packages/agent-gateway-client'
  - '../../packages/device-gateway-client'
+  - '../../packages/device-identity'
  - '../../packages/heterogeneous-agents'
  - '../../packages/local-file-shell'
  - '../../packages/types'
@@ -70,6 +70,26 @@ export async function getTrpcClient(): Promise<TrpcClient> {
  return _client;
 }

+/**
+ * Build a Lambda tRPC client from an already-resolved auth context, without
+ * re-running credential discovery. Use this when the caller already holds a
+ * token (e.g. `lh connect --token <jwt>`) — `getTrpcClient` would re-resolve
+ * via env/stored creds and `process.exit(1)` when none exist, which would
+ * abort an otherwise-valid explicit-token session.
+ */
+export function createLambdaClient(auth: {
+  serverUrl: string;
+  token: string;
+  tokenType: 'apiKey' | 'jwt' | 'serviceToken';
+}): TrpcClient {
+  const headers =
+    auth.tokenType === 'apiKey' ? { 'X-API-Key': auth.token } : { 'Oidc-Auth': auth.token };
+
+  return createTRPCClient<LambdaRouter>({
+    links: [httpLink({ headers, transformer: superjson, url: `${auth.serverUrl}/trpc/lambda` })],
+  });
+}
+
 export async function getToolsTrpcClient(): Promise<ToolsTrpcClient> {
  if (_toolsClient) return _toolsClient;

@@ -6,6 +6,7 @@ import { getTrpcClient } from '../api/client';
 import { confirm, outputJson, printBoxTable, printTable, timeAgo } from '../utils/format';
 import { log } from '../utils/logger';
 import { registerBotMessageCommands } from './botMessage';
+import { registerBotMessengersCommands } from './botMessengers';

 // ── Access policy helpers ──────────────────────────────

@@ -475,6 +476,9 @@ export function registerBotCommand(program: Command) {
  // Register message subcommand group
  registerBotMessageCommands(bot);

+  // Register messengers subcommand group (System Bot installations + account links)
+  registerBotMessengersCommands(bot);
+
  // ── platforms ───────────────────────────────────────────

  bot
@@ -806,7 +810,7 @@ export function registerBotCommand(program: Command) {
    name: 'group-allowlist',
  });

-  // ── watch-keywords (LOBE-8891) ────────────────────────
+  // ── watch-keywords () ────────────────────────

  registerWatchKeywordsCommand(bot);

@@ -0,0 +1,417 @@
+import { mkdtemp, writeFile } from 'node:fs/promises';
+import { tmpdir } from 'node:os';
+import path from 'node:path';
+
+import { Command } from 'commander';
+import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest';
+
+import { registerBotMessageCommands } from './botMessage';
+
+const { mockTrpcClient } = vi.hoisted(() => ({
+  mockTrpcClient: {
+    botMessage: {
+      replyToThread: { mutate: vi.fn() },
+      sendDirectMessage: { mutate: vi.fn() },
+      sendMessage: { mutate: vi.fn() },
+    },
+  },
+}));
+
+const { getTrpcClient: mockGetTrpcClient } = vi.hoisted(() => ({
+  getTrpcClient: vi.fn(),
+}));
+
+vi.mock('../api/client', () => ({ getTrpcClient: mockGetTrpcClient }));
+vi.mock('../utils/logger', () => ({
+  log: { debug: vi.fn(), error: vi.fn(), info: vi.fn(), warn: vi.fn() },
+  setVerbose: vi.fn(),
+}));
+
+describe('bot message send --attachment', () => {
+  let exitSpy: ReturnType<typeof vi.spyOn>;
+  let consoleSpy: ReturnType<typeof vi.spyOn>;
+
+  beforeEach(() => {
+    exitSpy = vi.spyOn(process, 'exit').mockImplementation((() => {}) as any);
+    consoleSpy = vi.spyOn(console, 'log').mockImplementation(() => {});
+    mockGetTrpcClient.mockResolvedValue(mockTrpcClient);
+    mockTrpcClient.botMessage.sendMessage.mutate.mockReset();
+    mockTrpcClient.botMessage.sendMessage.mutate.mockResolvedValue({ messageId: 'm-1' });
+    mockTrpcClient.botMessage.sendDirectMessage.mutate.mockReset();
+    mockTrpcClient.botMessage.sendDirectMessage.mutate.mockResolvedValue({
+      channelId: 'dm-1',
+      messageId: 'm-dm-1',
+    });
+    mockTrpcClient.botMessage.replyToThread.mutate.mockReset();
+    mockTrpcClient.botMessage.replyToThread.mutate.mockResolvedValue({ messageId: 'm-tr-1' });
+  });
+
+  afterEach(() => {
+    exitSpy.mockRestore();
+    consoleSpy.mockRestore();
+  });
+
+  function createProgram() {
+    const program = new Command();
+    program.exitOverride();
+    const bot = program.command('bot');
+    registerBotMessageCommands(bot);
+    return program;
+  }
+
+  it('passes a remote URL through as fetchUrl', async () => {
+    const program = createProgram();
+    await program.parseAsync([
+      'node',
+      'test',
+      'bot',
+      'message',
+      'send',
+      'bot-1',
+      '--target',
+      'ch-1',
+      '--message',
+      'hi',
+      '--attachment',
+      'https://cdn.example.com/foo.png',
+    ]);
+
+    expect(mockTrpcClient.botMessage.sendMessage.mutate).toHaveBeenCalledWith(
+      expect.objectContaining({
+        attachments: [
+          expect.objectContaining({
+            fetchUrl: 'https://cdn.example.com/foo.png',
+            mimeType: 'image/png',
+            name: 'foo.png',
+            type: 'image',
+          }),
+        ],
+        botId: 'bot-1',
+        channelId: 'ch-1',
+        content: 'hi',
+      }),
+    );
+  });
+
+  it('base64-encodes a local file path', async () => {
+    const dir = await mkdtemp(path.join(tmpdir(), 'lh-cli-attach-'));
+    const filePath = path.join(dir, 'tiny.txt');
+    await writeFile(filePath, 'hello');
+
+    const program = createProgram();
+    await program.parseAsync([
+      'node',
+      'test',
+      'bot',
+      'message',
+      'send',
+      'bot-1',
+      '--target',
+      'ch-1',
+      '--message',
+      'm',
+      '--attachment',
+      filePath,
+    ]);
+
+    const call = mockTrpcClient.botMessage.sendMessage.mutate.mock.calls[0][0];
+    expect(call.attachments).toHaveLength(1);
+    expect(call.attachments[0]).toMatchObject({
+      mimeType: 'text/plain',
+      name: 'tiny.txt',
+      type: 'file',
+    });
+    expect(call.attachments[0].data).toBe(Buffer.from('hello').toString('base64'));
+    expect(call.attachments[0].fetchUrl).toBeUndefined();
+  });
+
+  it('accepts multiple --attachment flags', async () => {
+    const program = createProgram();
+    await program.parseAsync([
+      'node',
+      'test',
+      'bot',
+      'message',
+      'send',
+      'bot-1',
+      '--target',
+      'ch-1',
+      '--message',
+      'm',
+      '--attachment',
+      'https://cdn.example.com/a.png',
+      '--attachment',
+      'https://cdn.example.com/b.pdf',
+    ]);
+
+    const call = mockTrpcClient.botMessage.sendMessage.mutate.mock.calls[0][0];
+    expect(call.attachments).toHaveLength(2);
+    expect(call.attachments[0]).toMatchObject({ type: 'image', name: 'a.png' });
+    expect(call.attachments[1]).toMatchObject({ type: 'file', name: 'b.pdf' });
+  });
+
+  it('omits attachments field when no flag is given', async () => {
+    const program = createProgram();
+    await program.parseAsync([
+      'node',
+      'test',
+      'bot',
+      'message',
+      'send',
+      'bot-1',
+      '--target',
+      'ch-1',
+      '--message',
+      'm',
+    ]);
+
+    const call = mockTrpcClient.botMessage.sendMessage.mutate.mock.calls[0][0];
+    expect(call.attachments).toBeUndefined();
+  });
+});
+
+describe('bot message dm --attachment', () => {
+  let exitSpy: ReturnType<typeof vi.spyOn>;
+  let consoleSpy: ReturnType<typeof vi.spyOn>;
+
+  beforeEach(() => {
+    exitSpy = vi.spyOn(process, 'exit').mockImplementation((() => {}) as any);
+    consoleSpy = vi.spyOn(console, 'log').mockImplementation(() => {});
+    mockGetTrpcClient.mockResolvedValue(mockTrpcClient);
+    mockTrpcClient.botMessage.sendDirectMessage.mutate.mockReset();
+    mockTrpcClient.botMessage.sendDirectMessage.mutate.mockResolvedValue({
+      channelId: 'dm-1',
+      messageId: 'm-dm-1',
+    });
+  });
+
+  afterEach(() => {
+    exitSpy.mockRestore();
+    consoleSpy.mockRestore();
+  });
+
+  function createProgram() {
+    const program = new Command();
+    program.exitOverride();
+    const bot = program.command('bot');
+    registerBotMessageCommands(bot);
+    return program;
+  }
+
+  it('sends a DM with a remote-URL attachment', async () => {
+    const program = createProgram();
+    await program.parseAsync([
+      'node',
+      'test',
+      'bot',
+      'message',
+      'dm',
+      'bot-1',
+      '--user-id',
+      'u-1',
+      '--message',
+      'hi',
+      '--attachment',
+      'https://cdn.example.com/foo.png',
+    ]);
+
+    expect(mockTrpcClient.botMessage.sendDirectMessage.mutate).toHaveBeenCalledWith(
+      expect.objectContaining({
+        attachments: [
+          expect.objectContaining({
+            fetchUrl: 'https://cdn.example.com/foo.png',
+            type: 'image',
+          }),
+        ],
+        botId: 'bot-1',
+        content: 'hi',
+        userId: 'u-1',
+      }),
+    );
+  });
+
+  it('omits attachments when no flag is given', async () => {
+    const program = createProgram();
+    await program.parseAsync([
+      'node',
+      'test',
+      'bot',
+      'message',
+      'dm',
+      'bot-1',
+      '--user-id',
+      'u-1',
+      '--message',
+      'plain',
+    ]);
+    const call = mockTrpcClient.botMessage.sendDirectMessage.mutate.mock.calls[0][0];
+    expect(call.attachments).toBeUndefined();
+  });
+});
+
+describe('bot message thread reply --attachment', () => {
+  let exitSpy: ReturnType<typeof vi.spyOn>;
+  let consoleSpy: ReturnType<typeof vi.spyOn>;
+
+  beforeEach(() => {
+    exitSpy = vi.spyOn(process, 'exit').mockImplementation((() => {}) as any);
+    consoleSpy = vi.spyOn(console, 'log').mockImplementation(() => {});
+    mockGetTrpcClient.mockResolvedValue(mockTrpcClient);
+    mockTrpcClient.botMessage.replyToThread.mutate.mockReset();
+    mockTrpcClient.botMessage.replyToThread.mutate.mockResolvedValue({ messageId: 'm-tr-1' });
+  });
+
+  afterEach(() => {
+    exitSpy.mockRestore();
+    consoleSpy.mockRestore();
+  });
+
+  function createProgram() {
+    const program = new Command();
+    program.exitOverride();
+    const bot = program.command('bot');
+    registerBotMessageCommands(bot);
+    return program;
+  }
+
+  it('replies to a thread with attachments', async () => {
+    const program = createProgram();
+    await program.parseAsync([
+      'node',
+      'test',
+      'bot',
+      'message',
+      'thread',
+      'reply',
+      'bot-1',
+      '--thread-id',
+      'th-1',
+      '--message',
+      'reply',
+      '--attachment',
+      'https://cdn.example.com/a.png',
+    ]);
+
+    expect(mockTrpcClient.botMessage.replyToThread.mutate).toHaveBeenCalledWith(
+      expect.objectContaining({
+        attachments: [
+          expect.objectContaining({
+            fetchUrl: 'https://cdn.example.com/a.png',
+            type: 'image',
+          }),
+        ],
+        botId: 'bot-1',
+        content: 'reply',
+        threadId: 'th-1',
+      }),
+    );
+  });
+});
+
+describe('bot message send via System Bot messenger install (@id)', () => {
+  let exitSpy: ReturnType<typeof vi.spyOn>;
+  let consoleSpy: ReturnType<typeof vi.spyOn>;
+
+  beforeEach(() => {
+    exitSpy = vi.spyOn(process, 'exit').mockImplementation((() => {}) as any);
+    consoleSpy = vi.spyOn(console, 'log').mockImplementation(() => {});
+    mockGetTrpcClient.mockResolvedValue(mockTrpcClient);
+    mockTrpcClient.botMessage.sendMessage.mutate.mockReset();
+    mockTrpcClient.botMessage.sendMessage.mutate.mockResolvedValue({ messageId: 'm-mi-1' });
+    mockTrpcClient.botMessage.sendDirectMessage.mutate.mockReset();
+    mockTrpcClient.botMessage.sendDirectMessage.mutate.mockResolvedValue({ messageId: 'm-mi-2' });
+    mockTrpcClient.botMessage.replyToThread.mutate.mockReset();
+    mockTrpcClient.botMessage.replyToThread.mutate.mockResolvedValue({ messageId: 'm-mi-3' });
+  });
+
+  afterEach(() => {
+    exitSpy.mockRestore();
+    consoleSpy.mockRestore();
+  });
+
+  function createProgram() {
+    const program = new Command();
+    program.exitOverride();
+    const bot = program.command('bot');
+    registerBotMessageCommands(bot);
+    return program;
+  }
+
+  it('@-prefixed positional arg routes to messengerInstallationId on send', async () => {
+    const program = createProgram();
+    await program.parseAsync([
+      'node',
+      'test',
+      'bot',
+      'message',
+      'send',
+      '@inst_abc',
+      '--target',
+      'C1',
+      '--message',
+      'hi',
+    ]);
+
+    const call = mockTrpcClient.botMessage.sendMessage.mutate.mock.calls[0][0];
+    expect(call.messengerInstallationId).toBe('inst_abc');
+    expect(call.botId).toBeUndefined();
+    expect(call.channelId).toBe('C1');
+  });
+
+  it('@-prefixed routes on dm', async () => {
+    const program = createProgram();
+    await program.parseAsync([
+      'node',
+      'test',
+      'bot',
+      'message',
+      'dm',
+      '@inst_xyz',
+      '--user-id',
+      'U1',
+      '--message',
+      'hi',
+    ]);
+    const call = mockTrpcClient.botMessage.sendDirectMessage.mutate.mock.calls[0][0];
+    expect(call.messengerInstallationId).toBe('inst_xyz');
+    expect(call.botId).toBeUndefined();
+  });
+
+  it('@-prefixed routes on thread reply', async () => {
+    const program = createProgram();
+    await program.parseAsync([
+      'node',
+      'test',
+      'bot',
+      'message',
+      'thread',
+      'reply',
+      '@inst_thr',
+      '--thread-id',
+      'T1',
+      '--message',
+      'r',
+    ]);
+    const call = mockTrpcClient.botMessage.replyToThread.mutate.mock.calls[0][0];
+    expect(call.messengerInstallationId).toBe('inst_thr');
+  });
+
+  it('plain (non-@) positional stays as botId', async () => {
+    const program = createProgram();
+    await program.parseAsync([
+      'node',
+      'test',
+      'bot',
+      'message',
+      'send',
+      'uuid-bot-id',
+      '--target',
+      'C1',
+      '--message',
+      'hi',
+    ]);
+    const call = mockTrpcClient.botMessage.sendMessage.mutate.mock.calls[0][0];
+    expect(call.botId).toBe('uuid-bot-id');
+    expect(call.messengerInstallationId).toBeUndefined();
+  });
+});
@@ -1,3 +1,6 @@
+import { readFile } from 'node:fs/promises';
+import { basename, extname } from 'node:path';
+
 import { DEFAULT_BOT_HISTORY_LIMIT } from '@lobechat/const';
 import type { Command } from 'commander';
 import pc from 'picocolors';
@@ -6,6 +9,111 @@ import { getTrpcClient } from '../api/client';
 import { confirm, outputJson, printTable, truncate } from '../utils/format';
 import { log } from '../utils/logger';

+type AttachmentInput = {
+  data?: string;
+  fetchUrl?: string;
+  mimeType?: string;
+  name?: string;
+  type: 'image' | 'file' | 'video' | 'audio';
+};
+
+const MIME_EXT_MAP: Record<string, string> = {
+  '.bmp': 'image/bmp',
+  '.gif': 'image/gif',
+  '.jpeg': 'image/jpeg',
+  '.jpg': 'image/jpeg',
+  '.m4a': 'audio/mp4',
+  '.mp3': 'audio/mpeg',
+  '.mp4': 'video/mp4',
+  '.ogg': 'audio/ogg',
+  '.pdf': 'application/pdf',
+  '.png': 'image/png',
+  '.svg': 'image/svg+xml',
+  '.txt': 'text/plain',
+  '.wav': 'audio/wav',
+  '.webm': 'video/webm',
+  '.webp': 'image/webp',
+};
+
+const inferMime = (path: string): string | undefined => MIME_EXT_MAP[extname(path).toLowerCase()];
+
+const inferAttachmentType = (mimeType?: string): AttachmentInput['type'] => {
+  if (!mimeType) return 'file';
+  if (mimeType.startsWith('image/')) return 'image';
+  if (mimeType.startsWith('video/')) return 'video';
+  if (mimeType.startsWith('audio/')) return 'audio';
+  return 'file';
+};
+
+/**
+ * Resolve a list of `--attachment` flag values into `AttachmentInput[]`. Each
+ * entry is either a URL or a local file path. Returns `undefined` when no
+ * flags were passed so callers can omit the field on the wire entirely (the
+ * TRPC schema treats absent vs empty differently). Bails the process on
+ * load failures — a silently-dropped attachment would be worse than a
+ * loud error here.
+ */
+const resolveAttachmentFlags = async (flags: string[]): Promise<AttachmentInput[] | undefined> => {
+  if (flags.length === 0) return undefined;
+  const out: AttachmentInput[] = [];
+  for (const raw of flags) {
+    try {
+      out.push(await parseAttachmentArg(raw));
+    } catch (error) {
+      log.error(`Failed to load attachment "${raw}": ${(error as Error).message}`);
+      process.exit(1);
+    }
+  }
+  return out;
+};
+
+/**
+ * Parse a single `--attachment <value>` argument. Accepted forms:
+ *   - `https://…` / `http://…`           → fetchUrl, type inferred from extension
+ *   - any other string                    → treated as a local file path;
+ *                                           bytes are read + base64-encoded
+ */
+const parseAttachmentArg = async (raw: string): Promise<AttachmentInput> => {
+  if (/^https?:\/\//.test(raw)) {
+    const pathname = new URL(raw).pathname;
+    const mimeType = inferMime(pathname);
+    return {
+      fetchUrl: raw,
+      mimeType,
+      name: basename(pathname) || undefined,
+      type: inferAttachmentType(mimeType),
+    };
+  }
+  const bytes = await readFile(raw);
+  const mimeType = inferMime(raw);
+  return {
+    data: bytes.toString('base64'),
+    mimeType,
+    name: basename(raw),
+    type: inferAttachmentType(mimeType),
+  };
+};
+
+/**
+ * Resolve the `<botIdOrAtKey>` positional argument into a `{ botId? |
+ * messengerInstallationId? }` shape that matches the TRPC send procedures'
+ * `exactly-one-of` constraint.
+ *
+ * Convention: a value prefixed with `@` is treated as a System Bot
+ * messenger installation id (e.g. `@inst_abc123`); anything else is a
+ * per-agent bot id. The `@` was chosen because `agent_bot_providers`.id is
+ * always a UUID — no UUID starts with `@`, so the prefix unambiguously
+ * disambiguates without breaking the existing UUID-only call sites.
+ */
+const resolveSendTargetArg = (
+  value: string,
+): { botId?: string; messengerInstallationId?: string } => {
+  if (value.startsWith('@')) {
+    return { messengerInstallationId: value.slice(1) };
+  }
+  return { botId: value };
+};
+
 export function registerBotMessageCommands(bot: Command) {
  const message = bot
    .command('message')
@@ -14,20 +122,40 @@ export function registerBotMessageCommands(bot: Command) {
  // ── send ────────────────────────────────────────────────

  message
-    .command('send <botId>')
-    .description('Send a message to a channel')
+    .command('send <botIdOrAtKey>')
+    .description(
+      'Send a message to a channel. Pass a per-agent bot id, or "@<messenger-install-id>" ' +
+        'to send through a System Bot messenger installation (see `lh bot messengers list`).',
+    )
    .requiredOption('--target <channelId>', 'Target channel / conversation ID')
    .requiredOption('--message <text>', 'Message content')
+    .option(
+      '--attachment <pathOrUrl>',
+      'Attach a file by local path or remote URL (repeatable). ' +
+        'Local paths are base64-encoded; http(s) URLs are passed as fetchUrl.',
+      collectOptions,
+      [],
+    )
    .option('--reply-to <messageId>', 'Reply to a specific message')
    .option('--json', 'Output JSON')
    .action(
      async (
-        botId: string,
-        options: { json?: boolean; message: string; replyTo?: string; target: string },
+        botIdOrAtKey: string,
+        options: {
+          attachment: string[];
+          json?: boolean;
+          message: string;
+          replyTo?: string;
+          target: string;
+        },
      ) => {
+        const attachments = await resolveAttachmentFlags(options.attachment);
+        const target = resolveSendTargetArg(botIdOrAtKey);
+
        const client = await getTrpcClient();
        const result = await client.botMessage.sendMessage.mutate({
-          botId,
+          ...target,
+          attachments,
          channelId: options.target,
          content: options.message,
          replyTo: options.replyTo,
@@ -39,8 +167,56 @@ export function registerBotMessageCommands(bot: Command) {
        }

        const r = result as any;
+        const suffix = attachments?.length ? ` with ${attachments.length} attachment(s)` : '';
        console.log(
-          `${pc.green('✓')} Message sent${r.messageId ? ` (${pc.dim(r.messageId)})` : ''}`,
+          `${pc.green('✓')} Message sent${r.messageId ? ` (${pc.dim(r.messageId)})` : ''}${suffix}`,
+        );
+      },
+    );
+
+  // ── dm (direct message) ─────────────────────────────────
+
+  message
+    .command('dm <botIdOrAtKey>')
+    .description(
+      'Send a direct message to a platform user. Pass a per-agent bot id, or ' +
+        '"@<messenger-install-id>" for a System Bot install.',
+    )
+    .requiredOption('--user-id <id>', 'Target user ID on the platform')
+    .requiredOption('--message <text>', 'Message content')
+    .option(
+      '--attachment <pathOrUrl>',
+      'Attach a file by local path or remote URL (repeatable). ' +
+        'Local paths are base64-encoded; http(s) URLs are passed as fetchUrl.',
+      collectOptions,
+      [],
+    )
+    .option('--json', 'Output JSON')
+    .action(
+      async (
+        botIdOrAtKey: string,
+        options: { attachment: string[]; json?: boolean; message: string; userId: string },
+      ) => {
+        const attachments = await resolveAttachmentFlags(options.attachment);
+        const target = resolveSendTargetArg(botIdOrAtKey);
+
+        const client = await getTrpcClient();
+        const result = await client.botMessage.sendDirectMessage.mutate({
+          ...target,
+          attachments,
+          content: options.message,
+          userId: options.userId,
+        });
+
+        if (options.json) {
+          outputJson(result);
+          return;
+        }
+
+        const r = result as any;
+        const suffix = attachments?.length ? ` with ${attachments.length} attachment(s)` : '';
+        console.log(
+          `${pc.green('✓')} DM sent${r.messageId ? ` (${pc.dim(r.messageId)})` : ''}${suffix}`,
        );
      },
    );
@@ -450,21 +626,43 @@ export function registerBotMessageCommands(bot: Command) {
    });

  thread
-    .command('reply <botId>')
-    .description('Reply to a thread')
+    .command('reply <botIdOrAtKey>')
+    .description(
+      'Reply to a thread. Pass a per-agent bot id, or "@<messenger-install-id>" ' +
+        'for a System Bot install.',
+    )
    .requiredOption('--thread-id <id>', 'Thread ID')
    .requiredOption('--message <text>', 'Reply content')
-    .action(async (botId: string, options: { message: string; threadId: string }) => {
-      const client = await getTrpcClient();
-      const result = await client.botMessage.replyToThread.mutate({
-        botId,
-        content: options.message,
-        threadId: options.threadId,
-      });
+    .option(
+      '--attachment <pathOrUrl>',
+      'Attach a file by local path or remote URL (repeatable). ' +
+        'Local paths are base64-encoded; http(s) URLs are passed as fetchUrl.',
+      collectOptions,
+      [],
+    )
+    .action(
+      async (
+        botIdOrAtKey: string,
+        options: { attachment: string[]; message: string; threadId: string },
+      ) => {
+        const attachments = await resolveAttachmentFlags(options.attachment);
+        const target = resolveSendTargetArg(botIdOrAtKey);

-      const r = result as any;
-      console.log(`${pc.green('✓')} Reply sent${r.messageId ? ` (${pc.dim(r.messageId)})` : ''}`);
-    });
+        const client = await getTrpcClient();
+        const result = await client.botMessage.replyToThread.mutate({
+          ...target,
+          attachments,
+          content: options.message,
+          threadId: options.threadId,
+        });
+
+        const r = result as any;
+        const suffix = attachments?.length ? ` with ${attachments.length} attachment(s)` : '';
+        console.log(
+          `${pc.green('✓')} Reply sent${r.messageId ? ` (${pc.dim(r.messageId)})` : ''}${suffix}`,
+        );
+      },
+    );

  // ── channel (subcommand group) ──────────────────────────

@@ -0,0 +1,365 @@
+import { Command } from 'commander';
+import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest';
+
+import type * as FormatModule from '../utils/format';
+import { registerBotMessengersCommands } from './botMessengers';
+
+const { mockTrpcClient } = vi.hoisted(() => ({
+  mockTrpcClient: {
+    messenger: {
+      availablePlatforms: { query: vi.fn() },
+      getMyLink: { query: vi.fn() },
+      listMyInstallations: { query: vi.fn() },
+      listMyLinks: { query: vi.fn() },
+      setActiveAgent: { mutate: vi.fn() },
+      unlink: { mutate: vi.fn() },
+      uninstallInstallation: { mutate: vi.fn() },
+    },
+  },
+}));
+
+const { getTrpcClient: mockGetTrpcClient } = vi.hoisted(() => ({
+  getTrpcClient: vi.fn(),
+}));
+
+vi.mock('../api/client', () => ({ getTrpcClient: mockGetTrpcClient }));
+vi.mock('../utils/logger', () => ({
+  log: { debug: vi.fn(), error: vi.fn(), info: vi.fn(), warn: vi.fn() },
+  setVerbose: vi.fn(),
+}));
+
+// `confirm` always answers yes — we test uninstall/unlink under the explicit
+// `--yes` flag too, but for the prompt path we want a deterministic answer.
+vi.mock('../utils/format', async () => {
+  const actual = await vi.importActual<typeof FormatModule>('../utils/format');
+  return { ...actual, confirm: vi.fn(async () => true) };
+});
+
+describe('bot messengers', () => {
+  let exitSpy: ReturnType<typeof vi.spyOn>;
+  let consoleSpy: ReturnType<typeof vi.spyOn>;
+  let errorSpy: ReturnType<typeof vi.spyOn>;
+
+  beforeEach(() => {
+    exitSpy = vi.spyOn(process, 'exit').mockImplementation((() => {}) as any);
+    consoleSpy = vi.spyOn(console, 'log').mockImplementation(() => {});
+    errorSpy = vi.spyOn(console, 'error').mockImplementation(() => {});
+    mockGetTrpcClient.mockResolvedValue(mockTrpcClient);
+    for (const fn of [
+      mockTrpcClient.messenger.availablePlatforms.query,
+      mockTrpcClient.messenger.getMyLink.query,
+      mockTrpcClient.messenger.listMyInstallations.query,
+      mockTrpcClient.messenger.listMyLinks.query,
+      mockTrpcClient.messenger.setActiveAgent.mutate,
+      mockTrpcClient.messenger.unlink.mutate,
+      mockTrpcClient.messenger.uninstallInstallation.mutate,
+    ]) {
+      fn.mockReset();
+    }
+  });
+
+  afterEach(() => {
+    exitSpy.mockRestore();
+    consoleSpy.mockRestore();
+    errorSpy.mockRestore();
+  });
+
+  function createProgram() {
+    const program = new Command();
+    program.exitOverride();
+    const bot = program.command('bot');
+    registerBotMessengersCommands(bot);
+    return program;
+  }
+
+  function renderedOutput(): string {
+    return consoleSpy.mock.calls.map((c) => String(c[0])).join('\n');
+  }
+
+  // ── installations ──────────────────────────────────────
+
+  describe('list', () => {
+    it('renders the installation table with SEND ARG hint', async () => {
+      mockTrpcClient.messenger.listMyInstallations.query.mockResolvedValueOnce([
+        {
+          applicationId: 'A1',
+          id: 'inst_abc',
+          installedAt: '2026-01-15T00:00:00Z',
+          platform: 'slack',
+          tenantId: 'T1',
+          tenantName: 'Acme Corp',
+        },
+      ]);
+      await createProgram().parseAsync(['node', 'test', 'bot', 'messengers', 'list']);
+      const out = renderedOutput();
+      expect(mockTrpcClient.messenger.listMyInstallations.query).toHaveBeenCalled();
+      expect(out).toContain('inst_abc');
+      expect(out).toContain('Acme Corp');
+      // The hint should explain how to use the id with the send commands
+      expect(out).toContain('@<INSTALLATION ID>');
+    });
+
+    it('reports empty state with install guidance', async () => {
+      mockTrpcClient.messenger.listMyInstallations.query.mockResolvedValueOnce([]);
+      await createProgram().parseAsync(['node', 'test', 'bot', 'messengers', 'list']);
+      const out = renderedOutput();
+      expect(out).toContain('No System Bot installations connected.');
+      expect(out).toContain('Settings → Messenger');
+    });
+
+    it('--json passes through the payload', async () => {
+      const payload = [{ id: 'inst_only', platform: 'discord', tenantId: 'g1' }];
+      mockTrpcClient.messenger.listMyInstallations.query.mockResolvedValueOnce(payload);
+      await createProgram().parseAsync(['node', 'test', 'bot', 'messengers', 'list', '--json']);
+      expect(renderedOutput()).toContain('"id": "inst_only"');
+    });
+  });
+
+  describe('view', () => {
+    it('prints details for a matching install', async () => {
+      mockTrpcClient.messenger.listMyInstallations.query.mockResolvedValueOnce([
+        {
+          applicationId: 'A1',
+          id: 'inst_match',
+          installedAt: '2026-01-15T00:00:00Z',
+          platform: 'slack',
+          scope: 'chat:write,users:read',
+          tenantId: 'T1',
+          tenantName: 'Acme Corp',
+        },
+      ]);
+      await createProgram().parseAsync(['node', 'test', 'bot', 'messengers', 'view', 'inst_match']);
+      const out = renderedOutput();
+      expect(out).toContain('inst_match');
+      expect(out).toContain('slack');
+      expect(out).toContain('Acme Corp');
+      expect(out).toContain('chat:write,users:read');
+    });
+
+    it('exits non-zero when install missing', async () => {
+      mockTrpcClient.messenger.listMyInstallations.query.mockResolvedValueOnce([]);
+      await createProgram().parseAsync([
+        'node',
+        'test',
+        'bot',
+        'messengers',
+        'view',
+        'inst_missing',
+      ]);
+      expect(errorSpy).toHaveBeenCalled();
+      expect(exitSpy).toHaveBeenCalledWith(1);
+    });
+
+    it('--json missing install emits JSON null + exit 1 (scriptable)', async () => {
+      mockTrpcClient.messenger.listMyInstallations.query.mockResolvedValueOnce([]);
+      await createProgram().parseAsync([
+        'node',
+        'test',
+        'bot',
+        'messengers',
+        'view',
+        'inst_missing',
+        '--json',
+      ]);
+      // No human-readable error log; the JSON-pipe consumer gets `null`.
+      expect(errorSpy).not.toHaveBeenCalled();
+      expect(consoleSpy.mock.calls.map((c) => String(c[0])).join('\n')).toContain('null');
+      expect(exitSpy).toHaveBeenCalledWith(1);
+    });
+  });
+
+  describe('uninstall', () => {
+    it('--yes skips confirm and calls the mutation', async () => {
+      mockTrpcClient.messenger.uninstallInstallation.mutate.mockResolvedValueOnce({
+        success: true,
+      });
+      await createProgram().parseAsync([
+        'node',
+        'test',
+        'bot',
+        'messengers',
+        'uninstall',
+        'inst_abc',
+        '--yes',
+      ]);
+      expect(mockTrpcClient.messenger.uninstallInstallation.mutate).toHaveBeenCalledWith({
+        installationId: 'inst_abc',
+      });
+      expect(renderedOutput()).toContain('revoked');
+    });
+
+    it('confirms before calling when --yes is omitted', async () => {
+      mockTrpcClient.messenger.listMyInstallations.query.mockResolvedValueOnce([
+        { id: 'inst_abc', platform: 'slack', tenantId: 'T1', tenantName: 'Acme' },
+      ]);
+      mockTrpcClient.messenger.uninstallInstallation.mutate.mockResolvedValueOnce({
+        success: true,
+      });
+      await createProgram().parseAsync([
+        'node',
+        'test',
+        'bot',
+        'messengers',
+        'uninstall',
+        'inst_abc',
+      ]);
+      // Mocked confirm returns true → mutation still fires
+      expect(mockTrpcClient.messenger.uninstallInstallation.mutate).toHaveBeenCalled();
+    });
+  });
+
+  describe('platforms', () => {
+    it('renders the platforms table', async () => {
+      mockTrpcClient.messenger.availablePlatforms.query.mockResolvedValueOnce([
+        { appId: 'A123', id: 'slack', name: 'Slack' },
+        { botUsername: 'lobehub_bot', id: 'telegram', name: 'Telegram' },
+      ]);
+      await createProgram().parseAsync(['node', 'test', 'bot', 'messengers', 'platforms']);
+      const out = renderedOutput();
+      expect(out).toContain('slack');
+      expect(out).toContain('A123');
+      expect(out).toContain('lobehub_bot');
+    });
+
+    it('handles empty platform list gracefully', async () => {
+      mockTrpcClient.messenger.availablePlatforms.query.mockResolvedValueOnce([]);
+      await createProgram().parseAsync(['node', 'test', 'bot', 'messengers', 'platforms']);
+      expect(renderedOutput()).toContain('No System Bot platforms');
+    });
+  });
+
+  // ── links ──────────────────────────────────────────────
+
+  describe('links list', () => {
+    it('renders the links table', async () => {
+      mockTrpcClient.messenger.listMyLinks.query.mockResolvedValueOnce([
+        {
+          activeAgentId: 'agent_1',
+          platform: 'slack',
+          platformUserId: 'U1',
+          platformUsername: 'alice',
+          tenantId: 'T1',
+        },
+      ]);
+      await createProgram().parseAsync(['node', 'test', 'bot', 'messengers', 'links', 'list']);
+      const out = renderedOutput();
+      expect(out).toContain('agent_1');
+      expect(out).toContain('alice');
+    });
+
+    it('reports empty state', async () => {
+      mockTrpcClient.messenger.listMyLinks.query.mockResolvedValueOnce([]);
+      await createProgram().parseAsync(['node', 'test', 'bot', 'messengers', 'links', 'list']);
+      expect(renderedOutput()).toContain('No account links yet');
+    });
+  });
+
+  describe('links view', () => {
+    it('shows the link detail', async () => {
+      mockTrpcClient.messenger.getMyLink.query.mockResolvedValueOnce({
+        activeAgentId: 'agent_2',
+        platform: 'slack',
+        platformUserId: 'U2',
+        platformUsername: 'bob',
+        tenantId: 'T2',
+      });
+      await createProgram().parseAsync([
+        'node',
+        'test',
+        'bot',
+        'messengers',
+        'links',
+        'view',
+        'slack',
+        '--tenant',
+        'T2',
+      ]);
+      expect(mockTrpcClient.messenger.getMyLink.query).toHaveBeenCalledWith({
+        platform: 'slack',
+        tenantId: 'T2',
+      });
+      expect(renderedOutput()).toContain('agent_2');
+    });
+
+    it('exits non-zero on missing link', async () => {
+      mockTrpcClient.messenger.getMyLink.query.mockResolvedValueOnce(null);
+      await createProgram().parseAsync([
+        'node',
+        'test',
+        'bot',
+        'messengers',
+        'links',
+        'view',
+        'discord',
+      ]);
+      expect(errorSpy).toHaveBeenCalled();
+      expect(exitSpy).toHaveBeenCalledWith(1);
+    });
+  });
+
+  describe('links set-agent', () => {
+    it('passes agentId through to setActiveAgent', async () => {
+      mockTrpcClient.messenger.setActiveAgent.mutate.mockResolvedValueOnce({ success: true });
+      await createProgram().parseAsync([
+        'node',
+        'test',
+        'bot',
+        'messengers',
+        'links',
+        'set-agent',
+        'slack',
+        '--agent',
+        'agent_xyz',
+        '--tenant',
+        'T1',
+      ]);
+      expect(mockTrpcClient.messenger.setActiveAgent.mutate).toHaveBeenCalledWith({
+        agentId: 'agent_xyz',
+        platform: 'slack',
+        tenantId: 'T1',
+      });
+    });
+
+    it('clears the agent when --agent none is passed', async () => {
+      mockTrpcClient.messenger.setActiveAgent.mutate.mockResolvedValueOnce({ success: true });
+      await createProgram().parseAsync([
+        'node',
+        'test',
+        'bot',
+        'messengers',
+        'links',
+        'set-agent',
+        'telegram',
+        '--agent',
+        'none',
+      ]);
+      expect(mockTrpcClient.messenger.setActiveAgent.mutate).toHaveBeenCalledWith({
+        agentId: null,
+        platform: 'telegram',
+        tenantId: undefined,
+      });
+    });
+  });
+
+  describe('links unlink', () => {
+    it('passes platform + tenant to the unlink mutation with --yes', async () => {
+      mockTrpcClient.messenger.unlink.mutate.mockResolvedValueOnce({ success: true });
+      await createProgram().parseAsync([
+        'node',
+        'test',
+        'bot',
+        'messengers',
+        'links',
+        'unlink',
+        'slack',
+        '--tenant',
+        'T1',
+        '--yes',
+      ]);
+      expect(mockTrpcClient.messenger.unlink.mutate).toHaveBeenCalledWith({
+        platform: 'slack',
+        tenantId: 'T1',
+      });
+    });
+  });
+});
@@ -0,0 +1,305 @@
+/**
+ * `lh bot messengers ...` — manages the user's System Bot installations
+ * (Slack workspaces, Discord guilds, Telegram), distinct from per-agent bots.
+ *
+ * Mirrors `bot ...` (per-agent CRUD) and `bot message ...` (send/read), but
+ * operates on `messenger_installations` (workspace-scoped) and
+ * `messenger_account_links` (per-user routing). Subcommands talk directly to
+ * `lambdaClient.messenger.*` — there's no shared CLI-side service layer for
+ * this domain yet.
+ *
+ * **uninstall vs unlink** (recurring confusion — surface in command help):
+ * - `uninstall <installationId>` revokes the install for the **whole
+ *   workspace**. Other users in that workspace can no longer use the bot.
+ * - `links unlink <platform>` only removes the **current user's** account
+ *   binding. Workspace stays installed; colleagues are unaffected.
+ */
+import type { Command } from 'commander';
+import pc from 'picocolors';
+
+import { getTrpcClient } from '../api/client';
+import { confirm, outputJson, printTable } from '../utils/format';
+
+const PLATFORMS = ['telegram', 'slack', 'discord'] as const;
+type MessengerPlatform = (typeof PLATFORMS)[number];
+
+const validatePlatform = (value: string): MessengerPlatform => {
+  if (!(PLATFORMS as readonly string[]).includes(value)) {
+    throw new Error(`Unknown messenger platform: ${value}. Valid values: ${PLATFORMS.join(', ')}.`);
+  }
+  return value as MessengerPlatform;
+};
+
+export function registerBotMessengersCommands(bot: Command) {
+  const messengers = bot
+    .command('messengers')
+    .description(
+      'Manage System Bot messenger installations (Slack workspaces, Discord guilds, Telegram) ' +
+        'and per-user account links',
+    );
+
+  // ── installations ──────────────────────────────────────
+
+  messengers
+    .command('list')
+    .description('List all System Bot installations the current user has connected.')
+    .option('--json', 'Output JSON')
+    .action(async (options: { json?: boolean }) => {
+      const client = await getTrpcClient();
+      const installations = await client.messenger.listMyInstallations.query();
+
+      if (options.json) {
+        outputJson(installations);
+        return;
+      }
+
+      if (installations.length === 0) {
+        console.log('No System Bot installations connected.');
+        console.log(
+          `\nRun ${pc.dim('lh bot messengers platforms')} to see what's available, then install via ` +
+            `${pc.dim('Settings → Messenger')} (OAuth requires a browser).`,
+        );
+        return;
+      }
+
+      const rows = installations.map((i: any) => [
+        i.id || '',
+        i.platform || '',
+        i.tenantName || i.tenantId || '(global)',
+        i.applicationId || '',
+        i.installedAt ? new Date(i.installedAt).toISOString().slice(0, 10) : '',
+      ]);
+      printTable(rows, ['INSTALLATION ID', 'PLATFORM', 'TENANT', 'APP ID', 'INSTALLED']);
+      console.log(
+        `\nUse ${pc.dim('@<INSTALLATION ID>')} as the positional argument on ` +
+          `${pc.dim('lh bot message send/dm/thread reply')} to route through a System Bot install.`,
+      );
+    });
+
+  messengers
+    .command('view <installationId>')
+    .description('Show detail for one installation.')
+    .option('--json', 'Output JSON')
+    .action(async (installationId: string, options: { json?: boolean }) => {
+      const client = await getTrpcClient();
+      const installations = (await client.messenger.listMyInstallations.query()) as any[];
+      const install = installations.find((i) => i.id === installationId);
+
+      if (!install) {
+        // Under `--json`, scripts need a parseable output even on miss —
+        // emit `null` to stdout (rather than a human-readable error), then
+        // exit non-zero so error handling still works in pipelines.
+        if (options.json) {
+          outputJson(null);
+        } else {
+          console.error(pc.red(`Installation not found: ${installationId}`));
+        }
+        process.exit(1);
+        return;
+      }
+
+      if (options.json) {
+        outputJson(install);
+        return;
+      }
+
+      console.log(`${pc.bold('Installation')} ${pc.dim(install.id)}`);
+      console.log(`  Platform:       ${install.platform}`);
+      console.log(`  Tenant:         ${install.tenantName || install.tenantId || '(global)'}`);
+      if (install.tenantId && install.tenantName) {
+        console.log(`  Tenant ID:      ${install.tenantId}`);
+      }
+      console.log(`  Application ID: ${install.applicationId}`);
+      if (install.scope) console.log(`  OAuth Scope:    ${install.scope}`);
+      if (install.installedAt) {
+        console.log(`  Installed:      ${new Date(install.installedAt).toISOString()}`);
+      }
+      if (install.enterpriseId) {
+        console.log(`  Enterprise ID:  ${install.enterpriseId}`);
+      }
+      if (install.isEnterpriseInstall) {
+        console.log(`  Enterprise:     yes`);
+      }
+    });
+
+  messengers
+    .command('uninstall <installationId>')
+    .description(
+      'Revoke a workspace install. AFFECTS EVERY USER IN THAT WORKSPACE — for Slack this freezes ' +
+        'the bot; for Discord it removes the audit entry (a guild admin must remove the bot ' +
+        'separately). To disconnect only your own account, use `bot messengers links unlink`.',
+    )
+    .option('--yes', 'Skip confirmation prompt')
+    .action(async (installationId: string, options: { yes?: boolean }) => {
+      const client = await getTrpcClient();
+
+      if (!options.yes) {
+        const installations = (await client.messenger.listMyInstallations.query()) as any[];
+        const install = installations.find((i) => i.id === installationId);
+        const label = install
+          ? `${install.platform} (${install.tenantName || install.tenantId || 'global'})`
+          : installationId;
+        const ok = await confirm(
+          `${pc.yellow('⚠')}  Uninstall ${pc.bold(label)} — this revokes the install for the whole workspace. Continue?`,
+        );
+        if (!ok) {
+          console.log('Aborted.');
+          return;
+        }
+      }
+
+      await client.messenger.uninstallInstallation.mutate({ installationId });
+      console.log(`${pc.green('✓')} Installation ${pc.dim(installationId)} revoked.`);
+    });
+
+  messengers
+    .command('platforms')
+    .description('List the platforms available for System Bot OAuth install.')
+    .option('--json', 'Output JSON')
+    .action(async (options: { json?: boolean }) => {
+      const client = await getTrpcClient();
+      const platforms = await client.messenger.availablePlatforms.query();
+
+      if (options.json) {
+        outputJson(platforms);
+        return;
+      }
+
+      if (platforms.length === 0) {
+        console.log('No System Bot platforms are configured on this deployment.');
+        return;
+      }
+
+      const rows = platforms.map((p: any) => [
+        p.id || '',
+        p.name || '',
+        p.appId || '',
+        p.botUsername || '',
+      ]);
+      printTable(rows, ['ID', 'NAME', 'APP ID', 'BOT USERNAME']);
+      console.log(
+        `\nInstalls are initiated via ${pc.dim('Settings → Messenger')} in the web UI ` +
+          '(OAuth needs a browser).',
+      );
+    });
+
+  // ── account links ──────────────────────────────────────
+
+  const links = messengers
+    .command('links')
+    .description('Manage per-user account links — routing of inbound IM to your agents');
+
+  links
+    .command('list')
+    .description('List all your account links across platforms and tenants.')
+    .option('--json', 'Output JSON')
+    .action(async (options: { json?: boolean }) => {
+      const client = await getTrpcClient();
+      const linkRows = await client.messenger.listMyLinks.query();
+
+      if (options.json) {
+        outputJson(linkRows);
+        return;
+      }
+
+      if (linkRows.length === 0) {
+        console.log('No account links yet. Complete verify-im on a platform first.');
+        return;
+      }
+
+      const rows = linkRows.map((l: any) => [
+        l.platform || '',
+        l.tenantId || '(global)',
+        l.activeAgentId || pc.dim('(unset)'),
+        l.platformUsername || l.platformUserId || '',
+      ]);
+      printTable(rows, ['PLATFORM', 'TENANT', 'ACTIVE AGENT', 'PLATFORM USER']);
+    });
+
+  links
+    .command('view <platform>')
+    .description('Show one account link.')
+    .option('--tenant <id>', 'Tenant scope (Slack workspace id). Omit for global-bot platforms.')
+    .option('--json', 'Output JSON')
+    .action(async (platform: string, options: { json?: boolean; tenant?: string }) => {
+      const client = await getTrpcClient();
+      const platformValidated = validatePlatform(platform);
+      const link = await client.messenger.getMyLink.query({
+        platform: platformValidated,
+        tenantId: options.tenant,
+      });
+
+      if (!link) {
+        console.error(
+          pc.red(
+            `No link found for ${platform}${options.tenant ? ` (tenant ${options.tenant})` : ''}`,
+          ),
+        );
+        process.exit(1);
+        return;
+      }
+
+      if (options.json) {
+        outputJson(link);
+        return;
+      }
+
+      console.log(`${pc.bold('Link')} ${pc.dim(link.platform)}`);
+      if (link.tenantId) console.log(`  Tenant ID:        ${link.tenantId}`);
+      console.log(`  Platform User ID: ${link.platformUserId}`);
+      if (link.platformUsername) {
+        console.log(`  Platform User:    ${link.platformUsername}`);
+      }
+      console.log(`  Active Agent:     ${link.activeAgentId ?? pc.dim('(unset)')}`);
+    });
+
+  links
+    .command('set-agent <platform>')
+    .description('Change which agent receives inbound IM on a platform link.')
+    .requiredOption('--agent <id>', 'Agent id to route to, or "none" to clear the active agent.')
+    .option('--tenant <id>', 'Tenant scope (Slack workspace id). Omit for global-bot platforms.')
+    .action(async (platform: string, options: { agent: string; tenant?: string }) => {
+      const client = await getTrpcClient();
+      const platformValidated = validatePlatform(platform);
+      const agentId = options.agent === 'none' ? null : options.agent;
+
+      await client.messenger.setActiveAgent.mutate({
+        agentId,
+        platform: platformValidated,
+        tenantId: options.tenant,
+      });
+
+      const scope = options.tenant ? ` (tenant ${options.tenant})` : '';
+      const target = agentId === null ? 'cleared' : `set to agent ${pc.dim(agentId)}`;
+      console.log(`${pc.green('✓')} Active agent for ${platform}${scope} ${target}.`);
+    });
+
+  links
+    .command('unlink <platform>')
+    .description(
+      'Remove your account link for a platform. Workspace install is unaffected — colleagues ' +
+        'can still use the bot.',
+    )
+    .option('--tenant <id>', 'Tenant scope (Slack workspace id). Omit for global-bot platforms.')
+    .option('--yes', 'Skip confirmation prompt')
+    .action(async (platform: string, options: { tenant?: string; yes?: boolean }) => {
+      const client = await getTrpcClient();
+      const platformValidated = validatePlatform(platform);
+
+      if (!options.yes) {
+        const ok = await confirm(
+          `Unlink your account from ${pc.bold(platform)}${options.tenant ? ` (tenant ${options.tenant})` : ''}?`,
+        );
+        if (!ok) {
+          console.log('Aborted.');
+          return;
+        }
+      }
+
+      await client.messenger.unlink.mutate({
+        platform: platformValidated,
+        tenantId: options.tenant,
+      });
+      console.log(`${pc.green('✓')} Unlinked.`);
+    });
+}
@@ -15,6 +15,7 @@ vi.mock('../auth/resolveToken', () => ({
  }),
 }));
 vi.mock('../settings', () => ({
+  loadOrCreateConnectionId: vi.fn().mockReturnValue('test-connection-id'),
  loadSettings: vi.fn().mockReturnValue(null),
  normalizeUrl: vi.fn((url?: string) => (url ? url.replace(/\/$/, '') : undefined)),
  saveSettings: vi.fn(),
@@ -8,8 +8,11 @@ import type {
  ToolCallRequestMessage,
 } from '@lobechat/device-gateway-client';
 import { GatewayClient } from '@lobechat/device-gateway-client';
+import type { IdentitySource } from '@lobechat/device-identity';
+import { deriveDeviceId } from '@lobechat/device-identity';
 import type { Command } from 'commander';

+import { createLambdaClient } from '../api/client';
 import { getValidToken } from '../auth/refresh';
 import { resolveToken } from '../auth/resolveToken';
 import { CLI_API_KEY_ENV } from '../constants/auth';
@@ -25,7 +28,7 @@ import {
  stopDaemon,
  writeStatus,
 } from '../daemon/manager';
-import { loadSettings, normalizeUrl, saveSettings } from '../settings';
+import { loadOrCreateConnectionId, loadSettings, normalizeUrl, saveSettings } from '../settings';
 import { executeToolCall } from '../tools';
 import { cleanupAllProcesses } from '../tools/shell';
 import { log, setVerbose } from '../utils/logger';
@@ -192,8 +195,24 @@ async function runConnect(options: ConnectOptions, isDaemonChild: boolean) {

  const resolvedGatewayUrl = gatewayUrl || OFFICIAL_GATEWAY_URL;

+  // Resolve a stable device identity. An explicit `--device-id` wins (lets a
+  // user pin a VM to a fixed identity); otherwise derive from the machine id so
+  // the same machine + user maps to one device across reconnects.
+  const identity: { deviceId: string; identitySource: IdentitySource } | undefined =
+    options.deviceId
+      ? { deviceId: options.deviceId, identitySource: 'fallback' }
+      : auth.userId
+        ? deriveDeviceId(auth.userId)
+        : undefined;
+
+  // Freeform channel label (`cli` by default); `LOBEHUB_CLI_CHANNEL` lets a
+  // dev build tag itself `cli-dev` so the gateway can prioritise / display it.
+  const channel = process.env.LOBEHUB_CLI_CHANNEL || 'cli';
+
  const client = new GatewayClient({
-    deviceId: options.deviceId,
+    channel,
+    connectionId: loadOrCreateConnectionId(),
+    deviceId: identity?.deviceId ?? options.deviceId,
    gatewayUrl: resolvedGatewayUrl,
    logger: isDaemonChild ? createDaemonLogger() : log,
    serverUrl: auth.serverUrl,
@@ -248,14 +267,14 @@ async function runConnect(options: ConnectOptions, isDaemonChild: boolean) {

  // Handle tool call requests
  client.on('tool_call_request', async (request: ToolCallRequestMessage) => {
-    const { requestId, toolCall } = request;
+    const { requestId, timeout, toolCall } = request;
    if (isDaemonChild) {
      appendLog(`[TOOL] ${toolCall.apiName} (${requestId})`);
    } else {
      log.toolCall(toolCall.apiName, requestId, toolCall.arguments);
    }

-    const result = await executeToolCall(toolCall.apiName, toolCall.arguments);
+    const result = await executeToolCall(toolCall.apiName, toolCall.arguments, timeout);

    if (isDaemonChild) {
      appendLog(`[RESULT] ${result.success ? 'OK' : 'FAIL'} (${requestId})`);
@@ -386,6 +405,25 @@ async function runConnect(options: ConnectOptions, isDaemonChild: boolean) {
    process.exit(0);
  });

+  // Register this device in the server registry before opening the WS, so the
+  // row exists by the time the gateway reports it online. Best-effort: a
+  // failure must not block the connection.
+  if (identity) {
+    try {
+      // Reuse the already-resolved auth (respects `--token` mode) instead of
+      // getTrpcClient(), which re-discovers creds and exits when none are found.
+      const trpc = createLambdaClient(auth);
+      await trpc.device.register.mutate({
+        deviceId: identity.deviceId,
+        hostname: os.hostname(),
+        identitySource: identity.identitySource,
+        platform: process.platform,
+      });
+    } catch (err) {
+      error(`Device registration failed (non-fatal): ${(err as Error).message}`);
+    }
+  }
+
  // Connect
  await client.connect();
 }
@@ -8,11 +8,20 @@ import { registerHeteroCommand } from './hetero';
 const { mockSpawnAgent } = vi.hoisted(() => ({
  mockSpawnAgent: vi.fn(),
 }));
+const { mockGetTrpcClient, mockHeteroFinishMutate, mockHeteroIngestMutate } = vi.hoisted(() => ({
+  mockGetTrpcClient: vi.fn(),
+  mockHeteroFinishMutate: vi.fn(),
+  mockHeteroIngestMutate: vi.fn(),
+}));

 vi.mock('@lobechat/heterogeneous-agents/spawn', () => ({
  spawnAgent: mockSpawnAgent,
 }));

+vi.mock('../api/client', () => ({
+  getTrpcClient: mockGetTrpcClient,
+}));
+
 vi.mock('../utils/logger', () => ({
  log: { debug: vi.fn(), error: vi.fn(), info: vi.fn(), warn: vi.fn() },
  setVerbose: vi.fn(),
@@ -77,6 +86,17 @@ describe('hetero exec command', () => {
    }) as any);
    stdoutSpy = vi.spyOn(process.stdout, 'write').mockImplementation(() => true);
    mockSpawnAgent.mockReset();
+    mockHeteroIngestMutate.mockReset();
+    mockHeteroFinishMutate.mockReset();
+    mockGetTrpcClient.mockReset();
+    mockHeteroIngestMutate.mockResolvedValue({ ack: true });
+    mockHeteroFinishMutate.mockResolvedValue({ ack: true });
+    mockGetTrpcClient.mockResolvedValue({
+      aiAgent: {
+        heteroFinish: { mutate: mockHeteroFinishMutate },
+        heteroIngest: { mutate: mockHeteroIngestMutate },
+      },
+    });
  });

  afterEach(() => {
@@ -341,4 +361,288 @@ describe('hetero exec command', () => {
    expect(exitSpy).toHaveBeenCalledWith(2);
    expect(mockSpawnAgent).not.toHaveBeenCalled();
  });
+
+  describe('--resume auto-retry on session-not-found', () => {
+    it('retries without --resume when the error stream event indicates the session is gone', async () => {
+      // First spawn: exits non-zero, emits a resume-not-found error event
+      const resumeNotFoundEvent = {
+        data: {
+          error: 'No conversation found with session ID cc-stale',
+          message: 'No conversation found with session ID cc-stale',
+        },
+        operationId: 'op-r1',
+        stepIndex: 0,
+        timestamp: 1,
+        type: 'error',
+      };
+      mockSpawnAgent
+        .mockReturnValueOnce(createFakeHandle({ events: [resumeNotFoundEvent], exitCode: 1 }))
+        // Second spawn: succeeds
+        .mockReturnValueOnce(createFakeHandle({ exitCode: 0 }));
+
+      await runCmd([
+        'hetero',
+        'exec',
+        '--type',
+        'claude-code',
+        '--prompt',
+        'do the thing',
+        '--resume',
+        'cc-stale',
+        '--operation-id',
+        'op-r1',
+      ]);
+
+      // Two spawns: first with --resume, retry without
+      expect(mockSpawnAgent).toHaveBeenCalledTimes(2);
+      expect(mockSpawnAgent.mock.calls[0][0]).toMatchObject({ resumeSessionId: 'cc-stale' });
+      expect(mockSpawnAgent.mock.calls[1][0]).not.toHaveProperty('resumeSessionId');
+      expect(mockSpawnAgent.mock.calls[1][0].resumeSessionId).toBeUndefined();
+
+      // Final exit code comes from the retry (0 → success)
+      expect(exitSpy).toHaveBeenCalledWith(0);
+    });
+
+    it('retries without --resume when stderr contains a session-not-found message', async () => {
+      // First spawn: exits non-zero with no events, but stderr has the pattern
+      mockSpawnAgent
+        .mockReturnValueOnce(
+          createFakeHandle({
+            exitCode: 1,
+            stderrChunks: ['Error: No conversation found with session ID xyz\n'],
+          }),
+        )
+        .mockReturnValueOnce(createFakeHandle({ exitCode: 0 }));
+
+      await runCmd([
+        'hetero',
+        'exec',
+        '--type',
+        'claude-code',
+        '--prompt',
+        'continue',
+        '--resume',
+        'xyz',
+        '--operation-id',
+        'op-r2',
+      ]);
+
+      expect(mockSpawnAgent).toHaveBeenCalledTimes(2);
+      expect(mockSpawnAgent.mock.calls[1][0].resumeSessionId).toBeUndefined();
+      expect(exitSpy).toHaveBeenCalledWith(0);
+    });
+
+    it('retries without --resume when the error indicates context overflow', async () => {
+      const contextOverflowEvent = {
+        data: {
+          error: 'prompt is too long: 215168 tokens > 200000 maximum',
+          message: 'prompt is too long: 215168 tokens > 200000 maximum',
+        },
+        operationId: 'op-ctx',
+        stepIndex: 0,
+        timestamp: 1,
+        type: 'error',
+      };
+      mockSpawnAgent
+        .mockReturnValueOnce(createFakeHandle({ events: [contextOverflowEvent], exitCode: 1 }))
+        .mockReturnValueOnce(createFakeHandle({ exitCode: 0 }));
+
+      await runCmd([
+        'hetero',
+        'exec',
+        '--type',
+        'claude-code',
+        '--prompt',
+        'next question',
+        '--resume',
+        'cc-longctx',
+        '--operation-id',
+        'op-ctx',
+      ]);
+
+      expect(mockSpawnAgent).toHaveBeenCalledTimes(2);
+      expect(mockSpawnAgent.mock.calls[0][0]).toMatchObject({ resumeSessionId: 'cc-longctx' });
+      expect(mockSpawnAgent.mock.calls[1][0].resumeSessionId).toBeUndefined();
+      expect(exitSpy).toHaveBeenCalledWith(0);
+    });
+
+    it('does NOT retry on a non-resume error exit', async () => {
+      // Exit code 1 but no resume-related error message
+      mockSpawnAgent.mockReturnValueOnce(
+        createFakeHandle({ exitCode: 1, stderrChunks: ['rate limit exceeded\n'] }),
+      );
+
+      await runCmd([
+        'hetero',
+        'exec',
+        '--type',
+        'claude-code',
+        '--prompt',
+        'hi',
+        '--resume',
+        'cc-valid',
+      ]);
+
+      expect(mockSpawnAgent).toHaveBeenCalledTimes(1);
+      expect(exitSpy).toHaveBeenCalledWith(1);
+    });
+
+    it('does NOT retry when --resume is not provided', async () => {
+      const errorEvent = {
+        data: { error: 'No conversation found', message: 'No conversation found' },
+        operationId: 'op-nr',
+        stepIndex: 0,
+        timestamp: 1,
+        type: 'error',
+      };
+      mockSpawnAgent.mockReturnValueOnce(createFakeHandle({ events: [errorEvent], exitCode: 1 }));
+
+      await runCmd([
+        'hetero',
+        'exec',
+        '--type',
+        'claude-code',
+        '--prompt',
+        'fresh run',
+        '--operation-id',
+        'op-nr',
+      ]);
+
+      // No --resume → no interception → no retry
+      expect(mockSpawnAgent).toHaveBeenCalledTimes(1);
+      expect(exitSpy).toHaveBeenCalledWith(1);
+    });
+
+    it('does NOT suppress the resume-error event from JSONL output', async () => {
+      const resumeNotFoundEvent = {
+        data: {
+          error: 'No conversation found with session ID old',
+          message: 'No conversation found with session ID old',
+        },
+        operationId: 'op-jsonl',
+        stepIndex: 0,
+        timestamp: 1,
+        type: 'error',
+      };
+      mockSpawnAgent
+        .mockReturnValueOnce(createFakeHandle({ events: [resumeNotFoundEvent], exitCode: 1 }))
+        .mockReturnValueOnce(createFakeHandle({ exitCode: 0 }));
+
+      await runCmd([
+        'hetero',
+        'exec',
+        '--type',
+        'claude-code',
+        '--prompt',
+        'do thing',
+        '--resume',
+        'old',
+        '--render',
+        'jsonl',
+      ]);
+
+      // The error event is still emitted to JSONL (for observability) even
+      // though it was withheld from the ingester.
+      const lines = stdoutSpy.mock.calls
+        .map((c) => c[0])
+        .filter((s): s is string => typeof s === 'string');
+      const errorLine = lines.find((l) => {
+        try {
+          return JSON.parse(l).type === 'error';
+        } catch {
+          return false;
+        }
+      });
+      expect(errorLine).toBeDefined();
+    });
+  });
+
+  it('sends full text snapshots before tools and waits for finish until all server ingests ack', async () => {
+    const callOrder: string[] = [];
+    mockHeteroIngestMutate.mockImplementation(async ({ events }: any) => {
+      const first = events[0];
+      callOrder.push(`ingest:${first.type}:${first.data?.chunkType ?? 'terminal'}`);
+      return { ack: true };
+    });
+    mockHeteroFinishMutate.mockImplementation(async () => {
+      callOrder.push('finish');
+      return { ack: true };
+    });
+
+    mockSpawnAgent.mockReturnValue(
+      createFakeHandle({
+        events: [
+          {
+            data: { chunkType: 'text', content: 'hello ' },
+            operationId: 'op-server',
+            stepIndex: 0,
+            timestamp: 1,
+            type: 'stream_chunk',
+          },
+          {
+            data: { chunkType: 'text', content: 'world' },
+            operationId: 'op-server',
+            stepIndex: 0,
+            timestamp: 2,
+            type: 'stream_chunk',
+          },
+          {
+            data: {
+              chunkType: 'tools_calling',
+              toolsCalling: [
+                {
+                  apiName: 'Bash',
+                  arguments: '{"cmd":"ls"}',
+                  id: 'tc-1',
+                  identifier: 'bash',
+                  type: 'default',
+                },
+              ],
+            },
+            operationId: 'op-server',
+            stepIndex: 1,
+            timestamp: 3,
+            type: 'stream_chunk',
+          },
+          {
+            data: { reason: 'success' },
+            operationId: 'op-server',
+            stepIndex: 1,
+            timestamp: 4,
+            type: 'agent_runtime_end',
+          },
+        ],
+        exitCode: 0,
+      }),
+    );
+
+    await runCmd([
+      'hetero',
+      'exec',
+      '--type',
+      'claude-code',
+      '--prompt',
+      'hi',
+      '--topic',
+      'topic-1',
+      '--operation-id',
+      'op-server',
+      '--render',
+      'none',
+    ]);
+
+    expect(mockHeteroIngestMutate).toHaveBeenCalledTimes(3);
+    expect(mockHeteroIngestMutate.mock.calls[0][0].events[0].data).toMatchObject({
+      chunkType: 'text',
+      content: 'hello world',
+      snapshotMode: 'replace',
+      snapshotSeq: 1,
+    });
+    expect(callOrder).toEqual([
+      'ingest:stream_chunk:text',
+      'ingest:stream_chunk:tools_calling',
+      'ingest:agent_runtime_end:terminal',
+      'finish',
+    ]);
+  });
 });
@@ -1,4 +1,5 @@
 import { randomUUID } from 'node:crypto';
+import { once } from 'node:events';
 import { readFile } from 'node:fs/promises';
 import path from 'node:path';

@@ -6,17 +7,51 @@ import type {
  AgentContentBlock,
  AgentImageSource,
  AgentPromptInput,
+  AgentStreamEvent,
 } from '@lobechat/heterogeneous-agents/spawn';
 import { spawnAgent } from '@lobechat/heterogeneous-agents/spawn';
 import type { Command } from 'commander';

 import { getTrpcClient } from '../api/client';
-import { BatchIngester, NoopIngestSink } from '../utils/BatchIngester';
 import { log } from '../utils/logger';
 import { TrpcIngestSink } from '../utils/TrpcIngestSink';

 const SUPPORTED_AGENT_TYPES = new Set(['claude-code', 'codex']);

+/**
+ * Patterns that indicate a `--resume <sessionId>` run should be retried
+ * without `--resume`.  Two classes of failure:
+ *
+ *   1. Session file missing (sandbox recycled): the container is ephemeral
+ *      (~1 h idle TTL), so a new sandbox has an empty `~/.claude/projects/`
+ *      and the stored session id is stale.
+ *
+ *   2. Context overflow (long conversation): the resumed session carries all
+ *      accumulated history; when the combined token count exceeds the model's
+ *      context window the API rejects the request immediately after CC
+ *      initialises.  Starting fresh (no `--resume`) drops the old history and
+ *      lets CC respond to the new prompt alone.
+ *
+ * Checked against:
+ *   - `error` stream events emitted by the CC adapter from CC's result event
+ *   - Accumulated stderr output (fallback when CC exits without a result event)
+ */
+const RESUME_RETRY_PATTERNS = [
+  // Session file missing — sandbox was recycled
+  /no conversation found/i,
+  /session.*not found/i,
+  /conversation.*not found/i,
+  /resume.*not found/i,
+  // Context overflow — API rejected the resumed session's accumulated history
+  /prompt.*too long/i,
+  /context.*too long/i,
+  /context window.*exceed/i,
+  /maximum.*context.*length/i,
+] as const;
+
+const looksLikeNeedsRetryWithoutResume = (text: string): boolean =>
+  RESUME_RETRY_PATTERNS.some((p) => p.test(text));
+
 interface ExecOptions {
  command?: string;
  cwd?: string;
@@ -166,6 +201,85 @@ const resolvePrompt = async (options: ExecOptions): Promise<ResolvedPrompt> => {
  return buildPromptFromText(raw, images);
 };

+class SerialServerIngester {
+  private accumulatedText = '';
+  private fatalError: Error | null = null;
+  private inflight: Promise<void> = Promise.resolve();
+  private nextSnapshotSeq = 0;
+  private pendingTextEvent: AgentStreamEvent | undefined;
+  private timer: ReturnType<typeof setTimeout> | null = null;
+
+  constructor(
+    private readonly sink: TrpcIngestSink,
+    private readonly snapshotFlushMs = 200,
+  ) {}
+
+  push(event: AgentStreamEvent): void {
+    if (this.fatalError) return;
+
+    if (
+      event.type === 'stream_chunk' &&
+      event.data?.chunkType === 'text' &&
+      typeof event.data?.content === 'string'
+    ) {
+      this.accumulatedText += event.data.content;
+      this.pendingTextEvent = event;
+      if (this.timer) clearTimeout(this.timer);
+      this.timer = setTimeout(() => {
+        this.timer = null;
+        this.queuePendingTextSnapshot();
+      }, this.snapshotFlushMs);
+      return;
+    }
+
+    this.queuePendingTextSnapshot();
+    this.enqueue(async () => {
+      await this.sink.ingest([event]);
+    });
+  }
+
+  async drain(): Promise<void> {
+    this.queuePendingTextSnapshot();
+    try {
+      await this.inflight;
+    } catch {
+      // `fatalError` is re-thrown below.
+    }
+    if (this.fatalError) throw this.fatalError;
+  }
+
+  private enqueue(task: () => Promise<void>) {
+    this.inflight = this.inflight.then(task).catch((err) => {
+      this.fatalError = err instanceof Error ? err : new Error(String(err));
+      throw this.fatalError;
+    });
+  }
+
+  private queuePendingTextSnapshot() {
+    if (!this.pendingTextEvent || this.fatalError) return;
+    if (this.timer) {
+      clearTimeout(this.timer);
+      this.timer = null;
+    }
+
+    const baseEvent = this.pendingTextEvent;
+    this.pendingTextEvent = undefined;
+    const snapshotEvent: AgentStreamEvent = {
+      ...baseEvent,
+      data: {
+        ...baseEvent.data,
+        content: this.accumulatedText,
+        snapshotMode: 'replace',
+        snapshotSeq: ++this.nextSnapshotSeq,
+      },
+    };
+
+    this.enqueue(async () => {
+      await this.sink.ingest([snapshotEvent]);
+    });
+  }
+}
+
 const exec = async (options: ExecOptions): Promise<void> => {
  if (!SUPPORTED_AGENT_TYPES.has(options.type)) {
    log.error(
@@ -209,128 +323,256 @@ const exec = async (options: ExecOptions): Promise<void> => {
  // server-ingest mode.  The tRPC client reads LOBEHUB_JWT (operation-scoped
  // JWT injected by the server) for authentication.
  const agentType = options.type as 'claude-code' | 'codex';
-  let sink: InstanceType<typeof TrpcIngestSink> | InstanceType<typeof NoopIngestSink>;
+  let sink: TrpcIngestSink | undefined;
+  let serverIngester: SerialServerIngester | undefined;
  if (serverIngest) {
    const client = await getTrpcClient();
-    sink = new TrpcIngestSink(client, agentType, operationId, options.topic!);
-  } else {
-    sink = new NoopIngestSink();
+    sink = new TrpcIngestSink(
+      client,
+      agentType,
+      operationId,
+      options.topic!,
+      process.env.LOBEHUB_ASSISTANT_MESSAGE_ID,
+    );
+    serverIngester = new SerialServerIngester(sink);
  }
-  const ingester = new BatchIngester(sink);

-  // `spawnAgent` is async and can reject DURING image normalization — fetch
-  // failures, missing local --image paths, decode errors. Surface those as a
-  // clean error + exit code instead of an unhandled promise rejection / stack
-  // trace, mirroring the validation try/catch above.
-  let handle: Awaited<ReturnType<typeof spawnAgent>>;
-  try {
-    handle = await spawnAgent({
+  /**
+   * Spawn one agent process and stream all its events into the server ingester.
+   *
+   * When `interceptResumeErrors` is true, any `error`-type event whose
+   * message matches `RESUME_RETRY_PATTERNS` is withheld from the
+   * ingester and signals a retry instead.  This keeps the server's
+   * operation state clean: no terminal error event is pushed, so the
+   * retry's events land on the same operationId without confusing the
+   * renderer.
+   *
+   * Returns:
+   *   code / signal — child exit info
+   *   sessionId     — CC session id from `system.init` (undefined on resume failure)
+   *   ingestError   — true when a batch could not be flushed after retries
+   *   resumeNotFound — true when a resume-not-found error was intercepted
+   *   stderrContent  — accumulated stderr (only when interceptResumeErrors=true)
+   */
+  const runOneAgent = async (
+    spawnOpts: Parameters<typeof spawnAgent>[0],
+    interceptResumeErrors: boolean,
+  ): Promise<{
+    code: number | null;
+    ingestError: boolean;
+    resumeNotFound: boolean;
+    sessionId: string | undefined;
+    signal: NodeJS.Signals | null;
+    stderrContent: string;
+  }> => {
+    // `spawnAgent` is async and can reject DURING image normalization — fetch
+    // failures, missing local --image paths, decode errors.
+    let handle: Awaited<ReturnType<typeof spawnAgent>>;
+    try {
+      handle = await spawnAgent(spawnOpts);
+    } catch (err) {
+      log.error('Failed to start agent:', err instanceof Error ? err.message : String(err));
+      process.exit(1);
+    }
+
+    // Always collect stderr — used for resume-error detection AND for
+    // surfacing a meaningful error message to the server when CC fails
+    // without emitting a structured error event.  Cap at 8 KB so the
+    // collector doesn't grow unboundedly on a chatty run.
+    // Always pipe to process.stderr too so users see auth prompts / warnings.
+    const STDERR_CAP = 8 * 1024;
+    let stderrContent = '';
+    const stderrEnded = once(handle.stderr, 'end').then(() => undefined);
+    handle.stderr.on('data', (chunk: Buffer) => {
+      if (stderrContent.length < STDERR_CAP) {
+        stderrContent += chunk.toString();
+      }
+    });
+    handle.stderr.pipe(process.stderr);
+
+    // Ctrl-C → SIGINT to the child's process group.
+    // Repeated Ctrl-C escalates to SIGKILL.
+    let interrupted = false;
+    const onSigint = async () => {
+      if (interrupted) {
+        handle.kill('SIGKILL');
+        return;
+      }
+      interrupted = true;
+      handle.kill('SIGINT');
+      if (serverIngester && sink) {
+        try {
+          await serverIngester.drain();
+          await sink.finish({ result: 'cancelled' });
+        } catch {
+          // best-effort; process is exiting anyway
+        }
+      }
+    };
+    const onSigterm = async () => {
+      handle.kill('SIGTERM');
+      if (serverIngester && sink) {
+        try {
+          await serverIngester.drain();
+          await sink.finish({ result: 'cancelled' });
+        } catch {
+          // best-effort
+        }
+      }
+    };
+    process.on('SIGINT', onSigint);
+    process.on('SIGTERM', onSigterm);
+
+    // Stream events. Each event is optionally written as JSONL and pushed
+    // into the ingester.  When intercepting resume errors, a matching
+    // `error` event is withheld from the ingester and flags a retry instead.
+    let resumeNotFound = false;
+    const ingestError = false;
+    try {
+      for await (const event of handle.events) {
+        if (interceptResumeErrors && event.type === 'error') {
+          const data = event.data as Record<string, unknown> | undefined;
+          const msg = String(data?.message ?? data?.error ?? '');
+          if (looksLikeNeedsRetryWithoutResume(msg)) {
+            resumeNotFound = true;
+            // Emit to JSONL for observability but do NOT push to ingester —
+            // we are about to retry; the server must not see a terminal error.
+            if (emitJsonl) process.stdout.write(`${JSON.stringify(event)}\n`);
+            continue;
+          }
+        }
+        if (emitJsonl) process.stdout.write(`${JSON.stringify(event)}\n`);
+        serverIngester?.push(event);
+      }
+    } catch (err) {
+      log.error(
+        'Stream error from agent process:',
+        err instanceof Error ? err.message : String(err),
+      );
+      if (serverIngester && sink) {
+        try {
+          await serverIngester.drain();
+          await sink.finish({
+            error: { message: String(err), type: 'stream_error' },
+            result: 'error',
+          });
+        } catch {
+          // best-effort
+        }
+      }
+      process.exit(1);
+    } finally {
+      process.off('SIGINT', onSigint);
+      process.off('SIGTERM', onSigterm);
+    }
+
+    const { code, signal } = await handle.exit;
+    await stderrEnded;
+
+    // Fallback stderr detection: CC may exit non-zero without emitting a
+    // result event (e.g. it writes to stderr and quits immediately).
+    if (
+      interceptResumeErrors &&
+      !resumeNotFound &&
+      code !== 0 &&
+      looksLikeNeedsRetryWithoutResume(stderrContent)
+    ) {
+      resumeNotFound = true;
+    }
+
+    return {
+      code,
+      ingestError,
+      resumeNotFound,
+      sessionId: handle.sessionId,
+      signal,
+      stderrContent,
+    };
+  };
+
+  // ─── First run (with --resume if provided) ───────────────────────────────
+
+  const interceptResume = !!options.resume;
+  const first = await runOneAgent(
+    {
      agentType: options.type,
      command: options.command,
      cwd: options.cwd || process.cwd(),
      operationId,
      prompt: resolved.prompt,
      resumeSessionId: options.resume,
-    });
-  } catch (err) {
-    log.error('Failed to start agent:', err instanceof Error ? err.message : String(err));
-    process.exit(1);
+    },
+    interceptResume,
+  );
+
+  // ─── Auto-retry without --resume when the session cannot be used ─────────
+  //
+  // Two classes of failure detected via `RESUME_RETRY_PATTERNS`:
+  //   A. Sandbox recycled: container is ephemeral (~1 h idle TTL); new sandbox
+  //      has no CC session files so `--resume <staleId>` is rejected with a
+  //      "no conversation found" error.
+  //   B. Context overflow: the resumed session carries accumulated history that
+  //      pushes the combined token count past the model limit; the API rejects
+  //      the call with a "prompt is too long" error.
+  //
+  // In both cases we transparently restart CC without `--resume` so it starts a
+  // fresh session.  The server's `heteroSessionId` is updated with the new id,
+  // breaking the stale-session loop.
+  let result = first;
+  if (first.resumeNotFound) {
+    log.info('Resume failed (session not found or context overflow) — retrying without --resume');
+    result = await runOneAgent(
+      {
+        agentType: options.type,
+        command: options.command,
+        cwd: options.cwd || process.cwd(),
+        operationId,
+        prompt: resolved.prompt,
+        // No resumeSessionId — start fresh
+      },
+      false, // no need to intercept resume errors on a fresh run
+    );
  }

-  // Forward the child's stderr to ours so users see CLI errors / warnings
-  // (auth prompts, missing-binary errors, etc.) in the terminal.
-  handle.stderr.pipe(process.stderr);
+  // ─── Drain + finish ───────────────────────────────────────────────────────

-  // Ctrl-C → SIGINT to the child's process group so the spawned CLI gets a
-  // chance to clean up. Repeated Ctrl-C escalates to SIGKILL via the
-  // standard "double-tap" pattern most CLIs implement themselves.
-  // In server-ingest mode, drain the ingester and call heteroFinish before
-  // exiting so the server knows the operation was cancelled.
-  let interrupted = false;
-  const onSigint = async () => {
-    if (interrupted) {
-      handle.kill('SIGKILL');
-      return;
-    }
-    interrupted = true;
-    handle.kill('SIGINT');
-    if (serverIngest) {
-      try {
-        await ingester.drain();
-        await sink.finish({ result: 'cancelled' });
-      } catch {
-        // best-effort; process is exiting anyway
-      }
-    }
-  };
-  process.on('SIGINT', onSigint);
-  process.on('SIGTERM', async () => {
-    handle.kill('SIGTERM');
-    if (serverIngest) {
-      try {
-        await ingester.drain();
-        await sink.finish({ result: 'cancelled' });
-      } catch {
-        // best-effort
-      }
-    }
-  });
+  const { code, signal, sessionId } = result;

-  // Stream events. Each event is optionally written as JSONL and always
-  // pushed into the ingester (which batches and sends to the server).
-  let ingestError = false;
-  try {
-    for await (const event of handle.events) {
-      if (emitJsonl) {
-        process.stdout.write(`${JSON.stringify(event)}\n`);
-      }
-      ingester.push(event);
-    }
-  } catch (err) {
-    log.error('Stream error from agent process:', err instanceof Error ? err.message : String(err));
-    if (serverIngest) {
-      try {
-        await ingester.drain();
-        await sink.finish({
-          result: 'error',
-          error: { message: String(err), type: 'stream_error' },
-        });
-      } catch {
-        // best-effort
-      }
-    }
-    process.exit(1);
-  } finally {
-    process.off('SIGINT', onSigint);
-  }
-
-  // Pass the child's exit code through. In server-ingest mode, drain the
-  // ingester and call heteroFinish before exiting.
-  const { code, signal } = await handle.exit;
-
-  if (serverIngest) {
+  if (serverIngester && sink) {
    try {
-      await ingester.drain();
+      await serverIngester.drain();
    } catch (err) {
      log.error(
        'Failed to flush events to server:',
        err instanceof Error ? err.message : String(err),
      );
-      ingestError = true;
+      result = { ...result, ingestError: true };
    }

-    const exitedClean = !ingestError && (code === 0 || signal === 'SIGTERM');
+    const exitedClean = !result.ingestError && (code === 0 || signal === 'SIGTERM');
+
+    // When the run failed, pass stderr as the error detail so the server can
+    // surface a useful message instead of the generic "Agent execution failed"
+    // fallback.  Trim to the last 1 KB — the tail is most informative and
+    // keeps the tRPC payload small.
+    const stderrTail = result.stderrContent.trim();
+    const finishError =
+      !exitedClean && stderrTail
+        ? { message: stderrTail.slice(-1024), type: 'AgentRuntimeError' }
+        : undefined;
+
    try {
      await sink.finish({
+        error: finishError,
        result: exitedClean ? 'success' : 'error',
-        sessionId: handle.sessionId,
+        sessionId,
      });
    } catch (err) {
      log.error('Failed to send heteroFinish:', err instanceof Error ? err.message : String(err));
    }
  }

-  if (code !== null) process.exit(ingestError ? 1 : code);
+  if (code !== null) process.exit(result.ingestError ? 1 : code);
  if (signal === 'SIGINT') process.exit(130);
  if (signal === 'SIGTERM') process.exit(143);
  if (signal === 'SIGKILL') process.exit(137);
@@ -12,16 +12,38 @@ export function registerNotifyCommand(program: Command) {
    .requiredOption('-c, --content <content>', 'Message content')
    .option('--agent-id <agentId>', 'Agent ID (overrides topic default)')
    .option('--thread-id <threadId>', 'Thread ID for threaded conversations')
+    .option(
+      '--role <role>',
+      'Message role: user (default, triggers agent reply) | assistant (writes directly as agent message)',
+      'user',
+    )
+    .option(
+      '--message-id <messageId>',
+      'When --role assistant: update an existing message instead of creating a new one (keeps a single bubble)',
+    )
+    .option(
+      '--continue',
+      'When --role assistant: trigger a follow-up agent turn after writing the message',
+    )
    .option('--json', 'Output JSON')
    .action(
      async (options: {
        agentId?: string;
        content: string;
+        continue?: boolean;
        json?: boolean;
+        messageId?: string;
+        role?: 'assistant' | 'user';
        threadId?: string;
        topic: string;
      }) => {
-        log.debug('notify: topic=%s, agentId=%s', options.topic, options.agentId);
+        log.debug(
+          'notify: topic=%s, agentId=%s, role=%s, messageId=%s',
+          options.topic,
+          options.agentId,
+          options.role,
+          options.messageId,
+        );

        const client = await getTrpcClient();

@@ -29,6 +51,9 @@ export function registerNotifyCommand(program: Command) {
          const result = await client.agentNotify.notify.mutate({
            agentId: options.agentId,
            content: options.content,
+            continue: options.continue,
+            messageId: options.messageId,
+            role: options.role,
            threadId: options.threadId,
            topicId: options.topic,
          });
@@ -6,9 +6,13 @@ import { registerTopicCommand } from './topic';

 const { mockTrpcClient } = vi.hoisted(() => ({
  mockTrpcClient: {
+    message: {
+      getMessages: { query: vi.fn() },
+    },
    topic: {
      batchDelete: { mutate: vi.fn() },
      createTopic: { mutate: vi.fn() },
+      getTopicDetail: { query: vi.fn() },
      getTopics: { query: vi.fn() },
      recentTopics: { query: vi.fn() },
      removeTopic: { mutate: vi.fn() },
@@ -41,6 +45,18 @@ describe('topic command', () => {
        (fn as ReturnType<typeof vi.fn>).mockReset();
      }
    }
+    for (const method of Object.values(mockTrpcClient.message)) {
+      for (const fn of Object.values(method)) {
+        (fn as ReturnType<typeof vi.fn>).mockReset();
+      }
+    }
+    // Default stub for getTopicDetail
+    mockTrpcClient.topic.getTopicDetail.query.mockResolvedValue({
+      favorite: false,
+      id: 't1',
+      title: 'Test Topic',
+      updatedAt: new Date().toISOString(),
+    });
  });

  afterEach(() => {
@@ -203,4 +219,130 @@ describe('topic command', () => {
      expect(mockTrpcClient.topic.recentTopics.query).toHaveBeenCalledWith({ limit: 10 });
    });
  });
+
+  describe('view', () => {
+    it('should display topic metadata and messages', async () => {
+      mockTrpcClient.message.getMessages.query.mockResolvedValue([
+        { content: 'Hello world', id: 'm1', role: 'user' },
+        { content: 'Hi there', id: 'm2', role: 'assistant' },
+      ]);
+
+      const program = createProgram();
+      await program.parseAsync(['node', 'test', 'topic', 'view', 't1']);
+
+      expect(mockTrpcClient.topic.getTopicDetail.query).toHaveBeenCalledWith(
+        expect.objectContaining({ id: 't1' }),
+      );
+      expect(mockTrpcClient.message.getMessages.query).toHaveBeenCalledWith(
+        expect.objectContaining({ topicId: 't1' }),
+      );
+      expect(consoleSpy).toHaveBeenCalled();
+    });
+
+    it('should skip message query entirely when --no-messages flag is set', async () => {
+      const program = createProgram();
+      await program.parseAsync(['node', 'test', 'topic', 'view', 't1', '--no-messages']);
+
+      // getTopicDetail is still called (for metadata)
+      expect(mockTrpcClient.topic.getTopicDetail.query).toHaveBeenCalled();
+      // but getMessages must NOT be called
+      expect(mockTrpcClient.message.getMessages.query).not.toHaveBeenCalled();
+    });
+
+    it('should output json when --json flag is set', async () => {
+      mockTrpcClient.message.getMessages.query.mockResolvedValue([
+        { content: 'Hello', id: 'm1', role: 'user' },
+      ]);
+
+      const program = createProgram();
+      await program.parseAsync(['node', 'test', 'topic', 'view', 't1', '--json']);
+
+      const calls = consoleSpy.mock.calls.flat().join('');
+      const parsed = JSON.parse(calls);
+      expect(parsed.topic.id).toBe('t1');
+      expect(parsed.messages).toHaveLength(1);
+      expect(parsed.messages[0]).toHaveProperty('role', 'user');
+      expect(parsed.messages[0]).toHaveProperty('content', 'Hello');
+    });
+
+    it('should output json with empty messages for --no-messages --json', async () => {
+      const program = createProgram();
+      await program.parseAsync(['node', 'test', 'topic', 'view', 't1', '--no-messages', '--json']);
+
+      expect(mockTrpcClient.message.getMessages.query).not.toHaveBeenCalled();
+      const calls = consoleSpy.mock.calls.flat().join('');
+      const parsed = JSON.parse(calls);
+      expect(parsed.topic.id).toBe('t1');
+      expect(parsed.messages).toHaveLength(0);
+    });
+
+    it('should respect -L for message page size', async () => {
+      mockTrpcClient.message.getMessages.query.mockResolvedValue([]);
+
+      const program = createProgram();
+      await program.parseAsync(['node', 'test', 'topic', 'view', 't1', '-L', '10']);
+
+      expect(mockTrpcClient.message.getMessages.query).toHaveBeenCalledWith(
+        expect.objectContaining({ pageSize: 10, topicId: 't1' }),
+      );
+    });
+
+    it('should slice messages with --from and --to', async () => {
+      mockTrpcClient.message.getMessages.query.mockResolvedValue([
+        { content: 'msg1', id: 'm1', role: 'user' },
+        { content: 'msg2', id: 'm2', role: 'assistant' },
+        { content: 'msg3', id: 'm3', role: 'user' },
+      ]);
+
+      const program = createProgram();
+      await program.parseAsync(['node', 'test', 'topic', 'view', 't1', '--from', '2', '--to', '3']);
+
+      // Should print only m2 and m3 (index 1 and 2)
+      const output = consoleSpy.mock.calls.flat().join('\n');
+      expect(output).toContain('msg2');
+      expect(output).toContain('msg3');
+      expect(output).not.toContain('msg1');
+    });
+
+    it('should render tool calls inline', async () => {
+      mockTrpcClient.message.getMessages.query.mockResolvedValue([
+        {
+          content: "I'll search for that.",
+          id: 'm1',
+          role: 'assistant',
+          tools: [
+            {
+              function: { arguments: '{"query":"lobehub"}', name: 'web_search' },
+              id: 'call_1',
+              type: 'function',
+            },
+          ],
+        },
+        { content: 'search results...', id: 'm2', role: 'tool' },
+      ]);
+
+      const program = createProgram();
+      await program.parseAsync(['node', 'test', 'topic', 'view', 't1']);
+
+      const output = consoleSpy.mock.calls.flat().join('\n');
+      expect(output).toContain('web_search');
+      expect(output).toContain('lobehub');
+    });
+
+    it('should render threaded messages with indentation', async () => {
+      mockTrpcClient.message.getMessages.query.mockResolvedValue([
+        { content: 'Parent message', id: 'm1', parentId: null, role: 'user' },
+        { content: 'Thread reply', id: 'm2', parentId: 'm1', role: 'assistant' },
+      ]);
+
+      const program = createProgram();
+      await program.parseAsync(['node', 'test', 'topic', 'view', 't1']);
+
+      const output = consoleSpy.mock.calls.flat().join('\n');
+      expect(output).toContain('Parent message');
+      expect(output).toContain('Thread reply');
+      // thread reply should appear after parent (basic ordering check)
+      expect(output.indexOf('Thread reply')).toBeGreaterThan(output.indexOf('Parent message'));
+    });
+  });
 });
@@ -332,4 +332,170 @@ export function registerTopicCommand(program: Command) {

      printTable(rows, ['ID', 'TITLE', 'UPDATED']);
    });
+
+  // ── view ──────────────────────────────────────────────
+
+  topic
+    .command('view <id>')
+    .description('View topic details and its messages')
+    .option('-L, --limit <n>', 'Max messages to fetch per page', '50')
+    .option('--from <n>', 'Show messages starting from this index (1-based)', '1')
+    .option('--to <n>', 'Show messages up to this index (inclusive)')
+    .option('--no-messages', 'Skip messages, show topic metadata only')
+    .option('--json', 'Output JSON')
+    .action(
+      async (
+        id: string,
+        options: {
+          from?: string;
+          json?: boolean;
+          limit?: string;
+          messages?: boolean;
+          to?: string;
+        },
+      ) => {
+        const client = await getTrpcClient();
+
+        // ── 1. Fetch topic detail (single query by id) ──
+        const topicDetail = await client.topic.getTopicDetail.query({ id } as any);
+
+        // ── 2. Fetch messages only when needed ──
+        if (options.messages === false) {
+          // --no-messages: skip message query entirely
+          if (options.json) {
+            console.log(JSON.stringify({ messages: [], topic: topicDetail ?? { id } }, null, 2));
+            return;
+          }
+          console.log('');
+          console.log(
+            `${pc.bold('Topic:')}   ${pc.cyan((topicDetail as any)?.title ?? id)}  ${pc.dim(`(${id})`)}`,
+          );
+          console.log('');
+          return;
+        }
+
+        const msgLimit = Number.parseInt(options.limit || '50', 10);
+        const msgResult = await client.message.getMessages.query({
+          pageSize: msgLimit,
+          topicId: id,
+        } as any);
+        const allMessages: any[] = Array.isArray(msgResult)
+          ? msgResult
+          : ((msgResult as any).items ?? []);
+
+        // Apply --from / --to slicing (1-based)
+        const fromIdx = Math.max(1, Number.parseInt(options.from || '1', 10)) - 1;
+        const toIdx = options.to ? Number.parseInt(options.to, 10) : allMessages.length;
+        const messages = allMessages.slice(fromIdx, toIdx);
+
+        if (options.json) {
+          console.log(
+            JSON.stringify(
+              {
+                messages: messages.map((m: any) => ({
+                  content: m.content ?? null,
+                  createdAt: m.createdAt ?? null,
+                  id: m.id,
+                  parentId: m.parentId ?? null,
+                  role: m.role,
+                  threadId: m.threadId ?? null,
+                  tools: m.tools ?? null,
+                })),
+                topic: { id },
+              },
+              null,
+              2,
+            ),
+          );
+          return;
+        }
+
+        // ── Header ──
+        const t = topicDetail as any;
+        console.log('');
+        console.log(`${pc.bold('Topic:')}   ${pc.cyan(t?.title ?? id)}  ${pc.dim(`(${id})`)}`);
+        if (t?.favorite) console.log(`${pc.bold('Favorite:')} ★`);
+        if (t?.updatedAt) console.log(`${pc.bold('Updated:')}  ${timeAgo(t.updatedAt)}`);
+        if (t?.status) console.log(`${pc.bold('Status:')}   ${t.status}`);
+        if (t?.model) console.log(`${pc.bold('Model:')}    ${t.model}${t.provider ? ` (${t.provider})` : ''}`);
+        console.log('');
+
+        // ── Messages ──
+        if (messages.length === 0) {
+          console.log(pc.dim('  (no messages)'));
+          return;
+        }
+
+        // Build parentId → children map for thread display
+        const childrenOf = new Map<string | null, any[]>();
+        for (const m of messages) {
+          const key = m.parentId ?? null;
+          if (!childrenOf.has(key)) childrenOf.set(key, []);
+          childrenOf.get(key)!.push(m);
+        }
+
+        const printMessage = (m: any, depth: number) => {
+          const indent = '  '.repeat(depth + 1);
+          const roleLabel =
+            m.role === 'user'
+              ? pc.green('user     ')
+              : m.role === 'tool'
+                ? pc.yellow('tool     ')
+                : pc.blue('assistant');
+          const threadMark = depth > 0 ? pc.dim('↳ ') : '';
+
+          // Full content (no truncation)
+          const content = (m.content || '').trim();
+          if (content) {
+            console.log(`${indent}${threadMark}${roleLabel}  ${content}`);
+          }
+
+          // Tool calls (assistant requesting tools)
+          if (m.tools && Array.isArray(m.tools) && m.tools.length > 0) {
+            for (const tool of m.tools) {
+              const toolName = tool.function?.name ?? tool.id ?? 'unknown';
+              const toolArgs = tool.function?.arguments
+                ? (() => {
+                    try {
+                      return JSON.stringify(JSON.parse(tool.function.arguments), null, 2)
+                        .split('\n')
+                        .map((l: string) => `${indent}    ${l}`)
+                        .join('\n');
+                    } catch {
+                      return `${indent}    ${tool.function.arguments}`;
+                    }
+                  })()
+                : '';
+              console.log(`${indent}  ${pc.yellow('⚙')} ${pc.bold(toolName)}`);
+              if (toolArgs) console.log(toolArgs);
+            }
+          }
+
+          // Render thread children recursively
+          const children = childrenOf.get(m.id) ?? [];
+          for (const child of children) {
+            printMessage(child, depth + 1);
+          }
+        };
+
+        // Print only top-level messages (parentId === null/undefined, or parentId not in current page)
+        const msgIds = new Set(messages.map((m: any) => m.id));
+        const topLevel = messages.filter(
+          (m: any) => !m.parentId || !msgIds.has(m.parentId),
+        );
+
+        for (const m of topLevel) {
+          printMessage(m, 0);
+        }
+
+        if (allMessages.length > msgLimit) {
+          console.log('');
+          console.log(
+            pc.dim(
+              `  … total ${allMessages.length} messages, showing ${fromIdx + 1}–${Math.min(toIdx, allMessages.length)}. Use -L / --from / --to to paginate.`,
+            ),
+          );
+        }
+      },
+    );
 }
@@ -0,0 +1,51 @@
+import fs from 'node:fs';
+import os from 'node:os';
+import path from 'node:path';
+
+export interface TaskEntry {
+  agentId?: string;
+  agentType: 'hermes' | 'openclaw';
+  operationId: string;
+  pid: number;
+  startedAt: string;
+  taskId: string;
+  topicId: string;
+}
+
+function getRegistryPath(): string {
+  return path.join(os.homedir(), '.lobehub', 'task-registry.json');
+}
+
+function readRegistry(): Record<string, TaskEntry> {
+  try {
+    return JSON.parse(fs.readFileSync(getRegistryPath(), 'utf8')) as Record<string, TaskEntry>;
+  } catch {
+    return {};
+  }
+}
+
+function writeRegistry(entries: Record<string, TaskEntry>): void {
+  const dir = path.dirname(getRegistryPath());
+  fs.mkdirSync(dir, { mode: 0o700, recursive: true });
+  fs.writeFileSync(getRegistryPath(), JSON.stringify(entries, null, 2), { mode: 0o600 });
+}
+
+export function saveTask(entry: TaskEntry): void {
+  const registry = readRegistry();
+  registry[entry.taskId] = entry;
+  writeRegistry(registry);
+}
+
+export function getTask(taskId: string): TaskEntry | undefined {
+  return readRegistry()[taskId];
+}
+
+export function removeTask(taskId: string): void {
+  const registry = readRegistry();
+  delete registry[taskId];
+  writeRegistry(registry);
+}
+
+export function listTasks(): TaskEntry[] {
+  return Object.values(readRegistry());
+}
@@ -1,3 +1,10 @@
 import { createProgram } from './program';
+import { log } from './utils/logger';

-createProgram().parse(process.argv, { from: 'node' });
+void createProgram()
+  .parseAsync(process.argv, { from: 'node' })
+  .catch((error: unknown) => {
+    const message = error instanceof Error ? error.message : String(error);
+    log.error(message);
+    process.exit(1);
+  });
@@ -5,7 +5,13 @@ import path from 'node:path';
 import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest';

 import { log } from '../utils/logger';
-import { loadSettings, normalizeUrl, resolveServerUrl, saveSettings } from './index';
+import {
+  loadOrCreateConnectionId,
+  loadSettings,
+  normalizeUrl,
+  resolveServerUrl,
+  saveSettings,
+} from './index';

 const tmpDir = path.join(os.tmpdir(), 'lobehub-cli-test-settings');
 const settingsDir = path.join(tmpDir, '.lobehub');
@@ -91,4 +97,22 @@ describe('settings', () => {

    expect(resolveServerUrl()).toBe('https://app.lobehub.com');
  });
+
+  it('should create a connectionId once and reuse it across calls', () => {
+    const first = loadOrCreateConnectionId();
+    expect(first).toMatch(/[\da-f-]{36}/);
+
+    // Persisted in its own file, independent of settings.json.
+    expect(fs.existsSync(path.join(settingsDir, 'connection-id'))).toBe(true);
+    expect(loadOrCreateConnectionId()).toBe(first);
+  });
+
+  it('should keep the connectionId even when settings.json is cleared', () => {
+    const id = loadOrCreateConnectionId();
+    // Clearing official-server settings unlinks settings.json — connectionId must survive.
+    saveSettings({ serverUrl: 'https://app.lobehub.com/' });
+
+    expect(fs.existsSync(settingsFile)).toBe(false);
+    expect(loadOrCreateConnectionId()).toBe(id);
+  });
 });
@@ -1,3 +1,4 @@
+import { randomUUID } from 'node:crypto';
 import fs from 'node:fs';
 import os from 'node:os';
 import path from 'node:path';
@@ -14,6 +15,9 @@ export interface StoredSettings {
 const LOBEHUB_DIR_NAME = process.env.LOBEHUB_CLI_HOME || '.lobehub';
 const SETTINGS_DIR = path.join(os.homedir(), LOBEHUB_DIR_NAME);
 const SETTINGS_FILE = path.join(SETTINGS_DIR, 'settings.json');
+// Kept in its own file rather than settings.json, which is unlinked whenever
+// all server/gateway URLs are default — the connectionId must persist regardless.
+const CONNECTION_ID_FILE = path.join(SETTINGS_DIR, 'connection-id');

 export function normalizeUrl(url: string | undefined): string | undefined {
  return url ? url.replace(/\/$/, '') : undefined;
@@ -54,6 +58,31 @@ export function saveSettings(settings: StoredSettings): void {
  fs.writeFileSync(SETTINGS_FILE, JSON.stringify(normalized, null, 2), { mode: 0o600 });
 }

+/**
+ * Stable per-install connection routing key for `lh connect`. Decoupled from
+ * the (machine-derived, shared-across-clients) deviceId so the gateway only
+ * replaces this install's own stale socket — a co-running desktop app on the
+ * same machine keeps its connection. Persisted under the CLI home dir, so a
+ * separate `LOBEHUB_CLI_HOME` (e.g. a dev build) naturally gets its own id.
+ */
+export function loadOrCreateConnectionId(): string {
+  try {
+    const existing = fs.readFileSync(CONNECTION_ID_FILE, 'utf8').trim();
+    if (existing) return existing;
+  } catch {
+    // not yet created
+  }
+
+  const id = randomUUID();
+  try {
+    fs.mkdirSync(SETTINGS_DIR, { mode: 0o700, recursive: true });
+    fs.writeFileSync(CONNECTION_ID_FILE, id, { mode: 0o600 });
+  } catch {
+    // best-effort: an unwritable home dir just means a fresh id per run
+  }
+  return id;
+}
+
 export function loadSettings(): StoredSettings | null {
  if (!fs.existsSync(SETTINGS_FILE)) return null;

@@ -0,0 +1,251 @@
+import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest';
+
+import { removeTask, saveTask } from '../../daemon/taskRegistry';
+import { runHeteroTask } from '../heteroTask';
+
+// ─── Mocks ───
+
+const spawnMock = vi.hoisted(() => vi.fn());
+const execFileSyncMock = vi.hoisted(() => vi.fn());
+
+vi.mock('node:child_process', () => ({
+  execFileSync: execFileSyncMock,
+  spawn: spawnMock,
+}));
+
+// task registry — use real implementation backed by a temporary in-memory map
+const taskStore: Record<string, any> = {};
+vi.mock('../../daemon/taskRegistry', () => ({
+  getTask: vi.fn((id: string) => taskStore[id]),
+  listTasks: vi.fn(() => Object.values(taskStore)),
+  removeTask: vi.fn((id: string) => {
+    delete taskStore[id];
+  }),
+  saveTask: vi.fn((entry: any) => {
+    taskStore[entry.taskId] = entry;
+  }),
+}));
+
+vi.mock('../../api/client', () => ({
+  getTrpcClient: vi.fn().mockResolvedValue({
+    agentNotify: {
+      notify: { mutate: vi.fn().mockResolvedValue(undefined) },
+    },
+  }),
+}));
+
+vi.mock('../../utils/logger', () => ({
+  log: { error: vi.fn(), info: vi.fn(), warn: vi.fn() },
+}));
+
+// ─── Helpers ───
+
+function makeMockChild(pid = 9999) {
+  const listeners: Record<string, Array<(...a: any[]) => void>> = {};
+  return {
+    on: vi.fn((event: string, cb: (...a: any[]) => void) => {
+      (listeners[event] ??= []).push(cb);
+    }),
+    pid,
+    unref: vi.fn(),
+    _emit: (event: string, ...args: any[]) => listeners[event]?.forEach((cb) => cb(...args)),
+  };
+}
+
+// ─── Tests ───
+
+describe('runHeteroTask (openclaw)', () => {
+  beforeEach(() => {
+    vi.clearAllMocks();
+    // Clear task store
+    for (const key of Object.keys(taskStore)) delete taskStore[key];
+    execFileSyncMock.mockReturnValue('/usr/local/bin/lh\n');
+  });
+
+  afterEach(() => {
+    vi.restoreAllMocks();
+  });
+
+  it('always injects buildNotifyProtocol into the prompt regardless of session history', async () => {
+    const child = makeMockChild();
+    spawnMock.mockReturnValue(child);
+
+    await runHeteroTask({
+      agentType: 'openclaw',
+      operationId: 'op-1',
+      prompt: 'what time is it',
+      taskId: 'task-1',
+      topicId: 'topic-1',
+    });
+
+    expect(spawnMock).toHaveBeenCalledTimes(1);
+    const [, spawnArgs] = spawnMock.mock.calls[0] as [string, string[]];
+    const msgIdx = spawnArgs.indexOf('--message');
+    const messageArg = spawnArgs[msgIdx + 1];
+
+    expect(messageArg).toContain('what time is it');
+    expect(messageArg).toContain('lh notify');
+    expect(messageArg).toContain('MSG_ID');
+  });
+
+  it('always injects protocol even on the second turn of the same session', async () => {
+    const child1 = makeMockChild(1111);
+    const child2 = makeMockChild(2222);
+    spawnMock.mockReturnValueOnce(child1).mockReturnValueOnce(child2);
+
+    // First turn
+    await runHeteroTask({
+      agentType: 'openclaw',
+      operationId: 'op-1',
+      prompt: 'hello',
+      taskId: 'task-1',
+      topicId: 'topic-1',
+    });
+    // Simulate process exit so task is removed
+    child1._emit('close', 0, null);
+
+    // Second turn (same topicId)
+    await runHeteroTask({
+      agentType: 'openclaw',
+      operationId: 'op-2',
+      prompt: 'follow up',
+      taskId: 'task-2',
+      topicId: 'topic-1',
+    });
+
+    expect(spawnMock).toHaveBeenCalledTimes(2);
+    for (const call of spawnMock.mock.calls) {
+      const args = call[1] as string[];
+      const msg = args[args.indexOf('--message') + 1];
+      expect(msg).toContain('lh notify');
+    }
+  });
+
+  it('kills an existing concurrent process for the same topicId before spawning', async () => {
+    const killSpy = vi.spyOn(process, 'kill').mockImplementation(() => true);
+
+    const child1 = makeMockChild(1111);
+    spawnMock.mockReturnValueOnce(child1);
+    await runHeteroTask({
+      agentType: 'openclaw',
+      operationId: 'op-1',
+      prompt: 'msg1',
+      taskId: 'task-1',
+      topicId: 'topic-same',
+    });
+    // task-1 is still "running" (close not fired)
+
+    const child2 = makeMockChild(2222);
+    spawnMock.mockReturnValueOnce(child2);
+    await runHeteroTask({
+      agentType: 'openclaw',
+      operationId: 'op-2',
+      prompt: 'msg2',
+      taskId: 'task-2',
+      topicId: 'topic-same',
+    });
+
+    expect(killSpy).toHaveBeenCalledWith(1111, 'SIGTERM');
+    expect(spawnMock).toHaveBeenCalledTimes(2);
+  });
+
+  it('does not kill processes for a different topicId', async () => {
+    const killSpy = vi.spyOn(process, 'kill').mockImplementation(() => true);
+
+    const child1 = makeMockChild(3333);
+    spawnMock.mockReturnValueOnce(child1);
+    await runHeteroTask({
+      agentType: 'openclaw',
+      operationId: 'op-1',
+      prompt: 'a',
+      taskId: 'task-a',
+      topicId: 'topic-A',
+    });
+
+    const child2 = makeMockChild(4444);
+    spawnMock.mockReturnValueOnce(child2);
+    await runHeteroTask({
+      agentType: 'openclaw',
+      operationId: 'op-2',
+      prompt: 'b',
+      taskId: 'task-b',
+      topicId: 'topic-B',
+    });
+
+    expect(killSpy).not.toHaveBeenCalled();
+  });
+
+  it('saves task entry with correct fields after spawn', async () => {
+    const child = makeMockChild(5555);
+    spawnMock.mockReturnValue(child);
+
+    await runHeteroTask({
+      agentId: 'agent-1',
+      agentType: 'openclaw',
+      operationId: 'op-x',
+      prompt: 'test',
+      taskId: 'task-x',
+      topicId: 'topic-x',
+    });
+
+    expect(saveTask).toHaveBeenCalledWith(
+      expect.objectContaining({
+        agentType: 'openclaw',
+        pid: 5555,
+        taskId: 'task-x',
+        topicId: 'topic-x',
+      }),
+    );
+  });
+
+  it('passes --session-id and --agent args to openclaw', async () => {
+    const child = makeMockChild();
+    spawnMock.mockReturnValue(child);
+
+    await runHeteroTask({
+      agentType: 'openclaw',
+      operationId: 'op-1',
+      prompt: 'hello',
+      taskId: 'task-1',
+      topicId: 'my-topic-id',
+    });
+
+    const [, spawnArgs] = spawnMock.mock.calls[0] as [string, string[]];
+    expect(spawnArgs).toContain('--session-id');
+    expect(spawnArgs[spawnArgs.indexOf('--session-id') + 1]).toBe('my-topic-id');
+    expect(spawnArgs).toContain('--agent');
+    expect(spawnArgs).toContain('--local');
+  });
+
+  it('removes task and ignores already-exited process when killing concurrent task', async () => {
+    const killSpy = vi.spyOn(process, 'kill').mockImplementation(() => {
+      throw new Error('No such process');
+    });
+
+    const child1 = makeMockChild(7777);
+    spawnMock.mockReturnValueOnce(child1);
+    await runHeteroTask({
+      agentType: 'openclaw',
+      operationId: 'op-1',
+      prompt: 'msg1',
+      taskId: 'task-1',
+      topicId: 'topic-gone',
+    });
+
+    const child2 = makeMockChild(8888);
+    spawnMock.mockReturnValueOnce(child2);
+    // Should not throw even though kill fails
+    await expect(
+      runHeteroTask({
+        agentType: 'openclaw',
+        operationId: 'op-2',
+        prompt: 'msg2',
+        taskId: 'task-2',
+        topicId: 'topic-gone',
+      }),
+    ).resolves.not.toThrow();
+
+    expect(removeTask).toHaveBeenCalledWith('task-1');
+    killSpy.mockRestore();
+  });
+});
@@ -0,0 +1,61 @@
+import { execFileSync } from 'node:child_process';
+
+export interface CheckPlatformCapabilityParams {
+  platform: 'hermes' | 'openclaw';
+}
+
+export interface CheckPlatformCapabilityResult {
+  available: boolean;
+  reason?: string;
+  version?: string;
+}
+
+/**
+ * Probe whether a specific agent platform is available on this device.
+ * Dispatched by the server via `device.checkCapability` tRPC procedure.
+ *
+ * - openclaw: runs `openclaw --version` and parses the output
+ * - hermes:   hits the gateway health endpoint on the configured port
+ */
+export async function checkPlatformCapability(
+  params: CheckPlatformCapabilityParams,
+): Promise<CheckPlatformCapabilityResult> {
+  const { platform } = params;
+
+  if (platform === 'openclaw') {
+    try {
+      const output = execFileSync('openclaw', ['--version'], {
+        encoding: 'utf8',
+        timeout: 5000,
+      }).trim();
+      // output is typically "openclaw x.y.z"
+      const version = output.split(/\s+/).at(-1);
+      return { available: true, version };
+    } catch (err) {
+      return {
+        available: false,
+        reason: err instanceof Error ? err.message : 'openclaw not found or failed to run',
+      };
+    }
+  }
+
+  if (platform === 'hermes') {
+    try {
+      const output = execFileSync('hermes', ['--version'], {
+        encoding: 'utf8',
+        timeout: 5000,
+      }).trim();
+      // output is typically "Hermes Agent vX.Y.Z (...)"
+      const versionMatch = output.match(/v(\d+\.\d+\.\d+)/);
+      const version = versionMatch ? versionMatch[1] : output.split(/\s+/).at(-1);
+      return { available: true, version };
+    } catch (err) {
+      return {
+        available: false,
+        reason: err instanceof Error ? err.message : 'hermes not found or failed to run',
+      };
+    }
+  }
+
+  return { available: false, reason: `Unknown platform: ${platform as string}` };
+}
@@ -0,0 +1,183 @@
+import { execFileSync } from 'node:child_process';
+import fs from 'node:fs';
+import os from 'node:os';
+import path from 'node:path';
+
+import type { RemoteHeterogeneousAgentType } from '@lobechat/heterogeneous-agents';
+
+export interface GetAgentProfileParams {
+  /** Agent ID to query (openclaw only). Defaults to the default agent. */
+  agentId?: string;
+  platform: RemoteHeterogeneousAgentType;
+}
+
+export interface AgentProfileResult {
+  avatar?: string;
+  description?: string;
+  title?: string;
+}
+
+// Files to look for a description (tried in order)
+const IDENTITY_FILES = ['IDENTITY.md', 'SOUL.md'];
+
+/**
+ * Try to extract a description from the workspace identity file.
+ * Looks for Creature / Vibe / Description fields in IDENTITY.md or SOUL.md.
+ */
+function readDescriptionFromWorkspace(workspacePath: string): string | undefined {
+  for (const filename of IDENTITY_FILES) {
+    const filePath = path.join(workspacePath, filename);
+    if (!fs.existsSync(filePath)) continue;
+
+    const content = fs.readFileSync(filePath, 'utf8');
+    const match = content.match(/\*{0,2}(?:Creature|Vibe|Description):?\*{0,2}\s*(.+)/i);
+    if (!match) continue;
+
+    const value = match[1].trim();
+    // Skip unfilled template placeholders like _(pick something)_ or (TBD)
+    if (/^[_*(（].*[）)*_]$|^(?:tbd|todo|n\/?a|none|待定|未定)$/i.test(value)) continue;
+    return value;
+  }
+}
+
+interface OpenClawAgentEntry {
+  id: string;
+  identityEmoji?: string;
+  identityName?: string;
+  isDefault?: boolean;
+  workspace?: string;
+}
+
+function getOpenClawProfile(agentId?: string): AgentProfileResult {
+  let output: string;
+  try {
+    output = execFileSync('openclaw', ['agents', 'list', '--json'], {
+      encoding: 'utf8',
+      timeout: 5000,
+    });
+  } catch {
+    return {};
+  }
+
+  let agents: OpenClawAgentEntry[];
+  try {
+    agents = JSON.parse(output) as OpenClawAgentEntry[];
+  } catch {
+    return {};
+  }
+
+  const agent = agentId
+    ? agents.find((a) => a.id === agentId)
+    : (agents.find((a) => a.isDefault) ?? agents[0]);
+
+  if (!agent) return {};
+
+  const title = agent.identityName || undefined;
+  const avatar = agent.identityEmoji || '🦞'; // OpenClaw brand mascot as default
+
+  // Description is not exposed by the CLI — read from the workspace IDENTITY.md
+  const description = agent.workspace ? readDescriptionFromWorkspace(agent.workspace) : undefined;
+
+  return { avatar, description, title };
+}
+
+/**
+ * Read the active Hermes profile name from `hermes profile list` output.
+ * The active profile is marked with ◆ in the first column.
+ */
+function getActiveHermesProfileName(): string | undefined {
+  try {
+    const output = execFileSync('hermes', ['profile', 'list'], {
+      encoding: 'utf8',
+      timeout: 5000,
+    });
+    const match = output.match(/◆(\S+)/);
+    return match?.[1];
+  } catch {
+    return undefined;
+  }
+}
+
+/**
+ * Read the filesystem path of a Hermes profile from `hermes profile show <name>`.
+ */
+function getHermesProfilePath(profileName: string): string | undefined {
+  try {
+    const output = execFileSync('hermes', ['profile', 'show', profileName], {
+      encoding: 'utf8',
+      timeout: 5000,
+    });
+    const match = output.match(/^Path:\s+(.+)/m);
+    const raw = match?.[1]?.trim();
+    // Expand leading `~` — Node does not auto-expand home-dir shorthands.
+    return raw?.replace(/^~(?=\/|$)/, os.homedir());
+  } catch {
+    return undefined;
+  }
+}
+
+/**
+ * Extract a one-line description from a Hermes SOUL.md file.
+ * Strips HTML comments and Markdown headings, then returns the first
+ * non-empty line of actual content.
+ */
+function readHermesSoulDescription(soulPath: string): string | undefined {
+  try {
+    const content = fs.readFileSync(soulPath, 'utf8');
+    // Loop until stable to handle any malformed/nested comment sequences.
+    let stripped = content;
+    let previous: string;
+    do {
+      previous = stripped;
+      stripped = stripped
+        .replaceAll(/<!--[\s\S]*?-->/g, '') // strip complete HTML comments
+        .replaceAll(/[<>]/g, '') // strip any remaining HTML delimiter chars
+        .replaceAll(/^#+\s.*$/gm, ''); // strip Markdown headings
+    } while (stripped !== previous);
+    const line = stripped
+      .split('\n')
+      .map((l) => l.trim())
+      .find((l) => l.length > 0);
+    return line || undefined;
+  } catch {
+    return undefined;
+  }
+}
+
+function getHermesProfile(): AgentProfileResult {
+  const profileName = getActiveHermesProfileName();
+  if (!profileName) return {};
+
+  const profilePath = getHermesProfilePath(profileName);
+  const description = profilePath
+    ? readHermesSoulDescription(path.join(profilePath, 'SOUL.md'))
+    : undefined;
+
+  return {
+    avatar: '⚡',
+    description,
+    title: profileName,
+  };
+}
+
+/**
+ * Fetch the agent profile (title, avatar, description) from the platform
+ * installed on this device. Dispatched by the server via `device.getAgentProfile`.
+ *
+ * - openclaw: `openclaw agents list --json` for name + emoji, workspace
+ *             IDENTITY.md for description fallback
+ * - hermes:   active profile name + SOUL.md description
+ */
+export async function getAgentProfile(params: GetAgentProfileParams): Promise<AgentProfileResult> {
+  const { platform, agentId } = params;
+
+  if (platform === 'openclaw') {
+    return getOpenClawProfile(agentId);
+  }
+
+  if (platform === 'hermes') {
+    return getHermesProfile();
+  }
+
+  return {};
+}
@@ -0,0 +1,341 @@
+import { execFileSync, spawn } from 'node:child_process';
+import fs from 'node:fs';
+import os from 'node:os';
+import path from 'node:path';
+
+import type { RemoteHeterogeneousAgentType } from '@lobechat/heterogeneous-agents';
+
+import { getTrpcClient } from '../api/client';
+import { getTask, listTasks, removeTask, saveTask } from '../daemon/taskRegistry';
+import { log } from '../utils/logger';
+
+// ─── Hermes session persistence ───
+// Maps topicId → hermes session_id so multi-turn conversations can resume
+// the same session across separate `runHeteroTask` invocations.
+
+const LOBEHUB_DIR_NAME = process.env.LOBEHUB_CLI_HOME || '.lobehub';
+const HERMES_SESSIONS_FILE = path.join(os.homedir(), LOBEHUB_DIR_NAME, 'hermes-sessions.json');
+
+function getHermesSessionId(topicId: string): string | undefined {
+  try {
+    const data = JSON.parse(fs.readFileSync(HERMES_SESSIONS_FILE, 'utf8')) as Record<
+      string,
+      string
+    >;
+    return data[topicId];
+  } catch {
+    return undefined;
+  }
+}
+
+function saveHermesSessionId(topicId: string, sessionId: string): void {
+  let data: Record<string, string> = {};
+  try {
+    data = JSON.parse(fs.readFileSync(HERMES_SESSIONS_FILE, 'utf8')) as Record<string, string>;
+  } catch {
+    // File doesn't exist yet — start fresh.
+  }
+  data[topicId] = sessionId;
+  fs.mkdirSync(path.dirname(HERMES_SESSIONS_FILE), { recursive: true });
+  fs.writeFileSync(HERMES_SESSIONS_FILE, JSON.stringify(data), 'utf8');
+}
+
+/** Resolve the absolute path to the `lh` binary to avoid PATH issues in child processes. */
+function resolveLhPath(): string {
+  try {
+    return execFileSync('which', ['lh'], { encoding: 'utf8' }).trim();
+  } catch {
+    return 'lh';
+  }
+}
+
+export interface RunHeteroTaskParams {
+  agentId?: string;
+  agentType: RemoteHeterogeneousAgentType;
+  cwd?: string;
+  operationId: string;
+  prompt: string;
+  taskId: string;
+  topicId: string;
+}
+
+export interface CancelHeteroTaskParams {
+  signal?: 'SIGINT' | 'SIGKILL' | 'SIGTERM';
+  taskId: string;
+}
+
+async function sendAutoNotify(
+  topicId: string,
+  taskId: string,
+  text: string,
+  agentId?: string,
+): Promise<void> {
+  try {
+    const client = await getTrpcClient();
+    await client.agentNotify.notify.mutate({
+      agentId,
+      content: text,
+      role: 'assistant',
+      topicId,
+    });
+  } catch (err) {
+    log.error('Failed to send auto-notify:', err instanceof Error ? err.message : String(err));
+  }
+}
+
+/**
+ * Signal remote hetero task completion to the server so it can publish
+ * `agent_runtime_end` to the gateway WS and close the frontend subscription.
+ * Called on clean process exit (code=0, no signal) — error exits go through
+ * `sendAutoNotify` which writes an error message AND triggers completion via
+ * the `done` flag.
+ */
+async function sendDoneSignal(topicId: string, agentId?: string): Promise<void> {
+  try {
+    const client = await getTrpcClient();
+    await client.agentNotify.notify.mutate({
+      agentId,
+      content: '',
+      done: true,
+      role: 'assistant',
+      topicId,
+    });
+  } catch (err) {
+    log.error('Failed to send done signal:', err instanceof Error ? err.message : String(err));
+  }
+}
+
+/**
+ * Build the notify protocol injected into the first message of a new hetero-agent session.
+ * Tells the agent how to push updates back to the LobeHub user via `lh notify`.
+ */
+function buildNotifyProtocol(lhPath: string, topicId: string): string {
+  return (
+    `## Context: This task was dispatched by LobeHub\n\n` +
+    `This conversation / task was sent to you by the **LobeHub platform** on behalf of a user. You are running as a background agent; the user is waiting for your response inside the LobeHub chat interface.\n\n` +
+    `**When to call notify**: any time you have something meaningful to tell the user — a key finding, a decision you made, a result, a question, or your final answer. Think of it as speaking directly to the user in the chat window.\n\n` +
+    `**What to hide**: internal work details such as tool call sequences, file reads, intermediate command output, retries, or low-level reasoning steps. The user cares about outcomes and insights, not your step-by-step mechanics.\n\n` +
+    `## Sending messages back to the user\n\n` +
+    `Use the \`${lhPath} notify\` command. All your updates appear as a **single message bubble** in the UI — create it once and update it in place.\n\n` +
+    `**Step 1 — Open the bubble on your first meaningful update** (captures the messageId):\n` +
+    `\`\`\`\n` +
+    `MSG_ID=$(${lhPath} notify --topic ${topicId} --role assistant --content "Starting..." --json | grep -o '"messageId":"[^"]*"' | cut -d'"' -f4)\n` +
+    `\`\`\`\n\n` +
+    `**Step 2 — Update the same bubble as you make progress**:\n` +
+    `\`\`\`\n` +
+    `${lhPath} notify --topic ${topicId} --role assistant --message-id "$MSG_ID" --content "Still working..."\n` +
+    `\`\`\`\n\n` +
+    `**Step 3 — Replace with your complete, final response when done**:\n` +
+    `\`\`\`\n` +
+    `${lhPath} notify --topic ${topicId} --role assistant --message-id "$MSG_ID" --content "<your full response here>"\n` +
+    `\`\`\`\n\n` +
+    `Rules:\n` +
+    `- Always use \`--json\` on the first call and capture \`messageId\` from the output.\n` +
+    `- Always pass \`--message-id\` on every subsequent call so updates overwrite the same bubble.\n` +
+    `- Write what matters to the user — not implementation steps or internal tool calls.\n` +
+    `- Call notify at least once when the task is done, even if there were no intermediate updates.`
+  );
+}
+
+export async function runHeteroTask(params: RunHeteroTaskParams): Promise<string> {
+  const { agentId, agentType, cwd, operationId, prompt, taskId, topicId } = params;
+  const workDir = cwd || process.cwd();
+  const lhPath = resolveLhPath();
+
+  if (agentType === 'openclaw') {
+    // openclaw agent --local is one-shot: each invocation processes one message and exits.
+    // The --session-id links turns into the same conversation history on disk.
+    // Requires the `openclaw` binary to be on PATH with Node >=22.19.
+    const openclawAgent = process.env.OPENCLAW_AGENT_ID ?? 'main';
+
+    // Always inject the notify protocol so openclaw knows how to report results
+    // back to the LobeHub UI — even if the previous turn failed and the session
+    // history was not cleanly committed.
+    const enrichedPrompt = `${prompt}\n\n${buildNotifyProtocol(lhPath, topicId)}`;
+
+    // Kill any existing openclaw process for this topicId before spawning a new one.
+    // openclaw serialises session writes; a concurrent process holding the session
+    // lock will cause the new one to exit with code 1.
+    for (const existing of listTasks()) {
+      if (existing.topicId === topicId && existing.agentType === 'openclaw') {
+        try {
+          process.kill(existing.pid, 'SIGTERM');
+        } catch {
+          // Already exited — nothing to do.
+        }
+        removeTask(existing.taskId);
+      }
+    }
+
+    const child = spawn(
+      'openclaw',
+      [
+        'agent',
+        '--agent',
+        openclawAgent,
+        '--session-id',
+        topicId,
+        '--message',
+        enrichedPrompt,
+        '--local',
+      ],
+      {
+        cwd: workDir,
+        detached: true,
+        env: { ...process.env },
+        stdio: 'ignore',
+      },
+    );
+
+    const pid = child.pid;
+    if (pid === undefined) {
+      throw new Error('Failed to get PID for openclaw process');
+    }
+    child.unref();
+
+    saveTask({
+      agentId,
+      agentType,
+      operationId,
+      pid,
+      startedAt: new Date().toISOString(),
+      taskId,
+      topicId,
+    });
+    log.info(`OpenClaw task started: taskId=${taskId} pid=${pid} agent=${openclawAgent}`);
+
+    // On exit: notify the server so it can close the frontend gateway WS subscription.
+    // - Abnormal exit (signal or non-zero code): write an error message bubble.
+    // - Clean exit (code=0, no signal): openclaw already sent its final message via
+    //   `lh notify`; just send a done signal to publish `agent_runtime_end`.
+    child.on('close', (code, signal) => {
+      removeTask(taskId);
+      if (code !== 0 || signal !== null) {
+        const text = signal
+          ? `Task cancelled (signal: ${signal})`
+          : `Task failed (exit code: ${code})`;
+        // Send error message first, THEN signal done (sequential).
+        // Fire-and-forget both, but ensure done is always sent even if notify fails.
+        void sendAutoNotify(topicId, taskId, text, agentId).finally(() =>
+          sendDoneSignal(topicId, agentId),
+        );
+      } else {
+        // Clean exit — openclaw already sent its final message; just signal done.
+        void sendDoneSignal(topicId, agentId);
+      }
+    });
+
+    return JSON.stringify({ pid, taskId });
+  }
+
+  if (agentType === 'hermes') {
+    // Kill any existing hermes process for this topicId before spawning a new one.
+    for (const existing of listTasks()) {
+      if (existing.topicId === topicId && existing.agentType === 'hermes') {
+        try {
+          process.kill(existing.pid, 'SIGTERM');
+        } catch {
+          // Already exited — nothing to do.
+        }
+        removeTask(existing.taskId);
+      }
+    }
+
+    // Resume the previous session for this topic if one exists.
+    const existingSessionId = getHermesSessionId(topicId);
+    const hermesArgs: string[] = ['chat', '--query', prompt, '--quiet', '--accept-hooks'];
+    if (existingSessionId) {
+      hermesArgs.push('--resume', existingSessionId);
+    }
+
+    // Hermes prints "session_id: <id>\n<response>" to stdout in --quiet mode.
+    // We capture stdout, parse both fields on exit, and relay the response via notify.
+    const child = spawn('hermes', hermesArgs, {
+      cwd: workDir,
+      detached: true,
+      env: { ...process.env },
+      stdio: ['ignore', 'pipe', 'ignore'],
+    });
+
+    const pid = child.pid;
+    if (pid === undefined) throw new Error('Failed to get PID for hermes process');
+    child.unref();
+
+    saveTask({
+      agentId,
+      agentType,
+      operationId,
+      pid,
+      startedAt: new Date().toISOString(),
+      taskId,
+      topicId,
+    });
+    log.info(`Hermes task started: taskId=${taskId} pid=${pid}`);
+
+    let stdout = '';
+    child.stdout.on('data', (chunk: Buffer) => {
+      stdout += chunk.toString();
+    });
+
+    child.on('close', (code, signal) => {
+      removeTask(taskId);
+
+      if (code !== 0 || signal !== null) {
+        const text = signal
+          ? `Task cancelled (signal: ${signal})`
+          : `Task failed (exit code: ${code})`;
+        void sendAutoNotify(topicId, taskId, text, agentId).finally(() =>
+          sendDoneSignal(topicId, agentId),
+        );
+        return;
+      }
+
+      // Parse "session_id: <id>" from the first line, response from the rest.
+      const sessionIdMatch = stdout.match(/^session_id:\s*(\S+)/m);
+      const sessionId = sessionIdMatch?.[1];
+      const response = stdout.replace(/^session_id:[^\n]*\n?/, '').trim();
+
+      if (sessionId) saveHermesSessionId(topicId, sessionId);
+
+      if (response) {
+        void sendAutoNotify(topicId, taskId, response, agentId).finally(() =>
+          sendDoneSignal(topicId, agentId),
+        );
+      } else {
+        void sendDoneSignal(topicId, agentId);
+      }
+    });
+
+    return JSON.stringify({ pid, taskId });
+  }
+
+  throw new Error(`Unsupported agentType: ${agentType as string}`);
+}
+
+export async function cancelHeteroTask(params: CancelHeteroTaskParams): Promise<string> {
+  const { signal = 'SIGINT', taskId } = params;
+  const entry = getTask(taskId);
+
+  if (!entry) {
+    return JSON.stringify({ message: `No task found with taskId: ${taskId}`, success: false });
+  }
+
+  // Both openclaw and hermes: kill by PID and let the child's close handler send the notify.
+  try {
+    process.kill(entry.pid, signal);
+  } catch (err) {
+    // Process already exited — exit handler won't fire; clean up manually.
+    log.warn(
+      `Failed to send ${signal} to pid ${entry.pid}: ${err instanceof Error ? err.message : String(err)}`,
+    );
+    removeTask(taskId);
+    await sendAutoNotify(
+      entry.topicId,
+      taskId,
+      'Task already completed or cancelled',
+      entry.agentId,
+    );
+  }
+
+  return JSON.stringify({ pid: entry.pid, signal, taskId });
+}
@@ -1,4 +1,5 @@
 import { log } from '../utils/logger';
+import { checkPlatformCapability } from './checkPlatformCapability';
 import {
  editLocalFile,
  globLocalFiles,
@@ -8,9 +9,14 @@ import {
  searchLocalFiles,
  writeLocalFile,
 } from './file';
+import { getAgentProfile } from './getAgentProfile';
+import { cancelHeteroTask, runHeteroTask } from './heteroTask';
 import { getCommandOutput, killCommand, runCommand } from './shell';

 const methodMap: Record<string, (args: any) => Promise<unknown>> = {
+  cancelHeteroTask,
+  checkPlatformCapability,
+  getAgentProfile,
  editFile: editLocalFile,
  getCommandOutput,
  globFiles: globLocalFiles,
@@ -19,6 +25,7 @@ const methodMap: Record<string, (args: any) => Promise<unknown>> = {
  listFiles: listLocalFiles,
  readFile: readLocalFile,
  runCommand,
+  runHeteroTask,
  searchFiles: searchLocalFiles,
  writeFile: writeLocalFile,

@@ -34,6 +41,7 @@ const methodMap: Record<string, (args: any) => Promise<unknown>> = {
 export async function executeToolCall(
  apiName: string,
  argsStr: string,
+  timeout?: number,
 ): Promise<{
  content: string;
  error?: string;
@@ -46,8 +54,12 @@ export async function executeToolCall(

  try {
    const args = JSON.parse(argsStr);
+    const finalArgs =
+      typeof timeout === 'number' && Number.isFinite(timeout) && !('timeout' in args)
+        ? { ...args, timeout }
+        : args;

-    const result = await handler(args);
+    const result = await handler(finalArgs);
    const content = typeof result === 'string' ? result : JSON.stringify(result);

    return { content, success: true };
@@ -16,6 +16,7 @@ export class TrpcIngestSink implements IngestSink {
    private readonly agentType: 'claude-code' | 'codex',
    private readonly operationId: string,
    private readonly topicId: string,
+    private readonly assistantMessageId?: string,
  ) {}

  async finish(params: Parameters<IngestSink['finish']>[0]): Promise<void> {
@@ -30,6 +31,7 @@ export class TrpcIngestSink implements IngestSink {
  async ingest(events: AgentStreamEvent[]): Promise<void> {
    await this.client.aiAgent.heteroIngest.mutate({
      agentType: this.agentType,
+      assistantMessageId: this.assistantMessageId,
      events: events as any,
      operationId: this.operationId,
      topicId: this.topicId,
--- a/Show More
+++ b/Show More