chatccc 0.2.212 → 0.2.214

package/.agents/skills/create-chatccc-feishu-app/SKILL.md

@@ -0,0 +1,85 @@
+---
+name: create-chatccc-feishu-app
+description: Create, clone, or configure a Feishu enterprise self-built bot application for ChatCCC. Use when asked to create a ChatCCC Feishu app, imitate an existing lowercase `chatccc` app, prepare a bot for another computer, or configure its bot capability, `im:`/`cardkit:` permissions, events, credentials, version, and internal release from this repository's README.
+---
+
+# Create a ChatCCC Feishu App
+
+Use the logged-in Feishu developer console to create and completely configure a ChatCCC bot app. Treat this repository's README as the minimum requirement and the existing app named exactly lowercase `chatccc` as the configuration template.
+
+## 1. Establish the target and authorization
+
+- Read the `README.md` section `#### 飞书（推荐）` before changing the console.
+- Determine the requested app name, target computer, description, icon, availability, and whether the user wants configuration only or a real release. Derive a concise description when one is not supplied; for example, `用飞书聊天控制<目标电脑>上的 Claude Code / Cursor / Codex。`.
+- Use the browser surface the user requested. If the user specifies Chrome DevTools/CDP, connect directly to `chromeDevtools.port` or the README default `15166`; do not substitute extension-based Chrome control.
+- Obtain action-time confirmation before creating the app, granting permissions, saving event/callback subscriptions, or publishing. One explicit instruction such as `完整配置并真正发布` may authorize the listed actions as a batch.
+
+## 2. Inventory the lowercase template
+
+1. Open `https://open.feishu.cn/app?from=devbotmenu` in the authenticated browser.
+2. Open the app named exactly lowercase `chatccc` in a separate tab.
+3. Record its non-secret configuration:
+   - app description and icon;
+   - application capabilities, especially `机器人`;
+   - every granted permission and its identity type;
+   - event and callback subscription modes;
+   - every event and callback identifier;
+   - version availability, external sharing choices, and release state.
+4. Never reveal, log, export to chat, or persist the template App Secret.
+
+Prefer current console data over a stale hard-coded list. At the time this workflow was verified, lowercase `chatccc` had 62 application-identity permissions, one user-identity permission, 12 application events, and one callback. Use those counts as a sanity check, not as a replacement for inspecting the template.
+
+## 3. Create the target app
+
+1. Choose `创建企业自建应用` and fill the requested name and description.
+2. Reuse the lowercase `chatccc` icon only when imitation was requested, and verify its crop preview.
+3. Click the final `创建` action only after confirmation.
+4. Click once and wait for SPA navigation. Verify the exact app name, App ID, URL, and status before retrying; the console can temporarily show stale text after a successful mutation.
+
+## 4. Clone the complete ChatCCC configuration
+
+Perform the configuration in this order:
+
+1. Add the `机器人` application capability and verify that the page offers `删除能力`, which means the capability is enabled.
+2. Clone permissions with `权限管理` > `批量处理` > `批量导入/导出权限`:
+   - export from lowercase `chatccc`;
+   - import into the target app;
+   - confirm all additions and verify both application-identity and user-identity totals;
+   - do not save the export JSON in the repository or chat.
+3. Under `事件与回调` > `事件配置`, set the subscription mode to `使用长连接接收事件`, then add every event present in the template. The verified baseline is:
+   - `im.chat.disbanded_v1`
+   - `im.chat.member.bot.added_v1`
+   - `im.chat.member.bot.deleted_v1`
+   - `im.chat.member.user.added_v1`
+   - `im.chat.member.user.deleted_v1`
+   - `im.chat.member.user.withdrawn_v1`
+   - `im.message.bot_muted_v1`
+   - `im.message.message_read_v1`
+   - `im.message.reaction.created_v1`
+   - `im.message.reaction.deleted_v1`
+   - `im.message.recalled_v1`
+   - `im.message.receive_v1`
+4. Under `回调配置`, independently set `使用长连接接收回调` and add `card.action.trigger`. Configuring the event mode does not configure the callback mode.
+5. Verify that every listed event shows a sufficient permission as `已开通`.
+
+The README minimum is the bot capability, relevant `im:` and `cardkit:` permissions, `im.message.receive_v1`, and `card.action.trigger`; a clone request requires the complete current template configuration, not only that minimum.
+
+## 5. Create and publish a version
+
+1. Open `版本管理与发布` and create a version such as `1.0.0`.
+2. Keep mobile and desktop default abilities set to `机器人` and write a meaningful update note.
+3. Match the requested or template availability. For a personal computer, default to the owner-only internal range and leave external group/private-chat sharing disabled unless explicitly requested.
+4. Save the version, then complete the separate `确认发布` action. A successful version creation alone is not a release.
+5. Refresh the version list and require all of the following evidence before reporting success:
+   - application status `已启用`;
+   - banner `当前修改均已发布`;
+   - version status `已发布` with a publication time.
+6. If administrator review is required, submit everything authorized and report the exact remaining approval state.
+
+## 6. Verify and hand off safely
+
+- Re-open the bot, permission, event/callback, and version pages after publication. Verify the bot is enabled, permission identity counts match, both subscription modes are long connection, event/callback identifiers match, and the version is published.
+- Treat network responses and a refreshed page as stronger evidence than an immediately stale SPA view. Never repeat a create, add, or publish action only because the first page read is stale.
+- Report the app name, App ID, capability, permission counts, event/callback counts, version, release state, and any remaining action.
+- Never print the App Secret. If the user explicitly asks to configure ChatCCC, transfer the App ID and App Secret directly into the local private configuration without echoing the secret in commands, logs, screenshots, files tracked by Git, or chat.
+- Do not modify local ChatCCC configuration or transmit credentials elsewhere unless explicitly authorized.

package/.agents/skills/create-chatccc-feishu-app/agents/openai.yaml

@@ -0,0 +1,4 @@
+interface:
+  display_name: "Create ChatCCC Feishu App"
+  short_description: "Create and configure a Feishu bot app for ChatCCC"
+  default_prompt: "Use $create-chatccc-feishu-app to create and configure a Feishu enterprise bot app for this ChatCCC project."

package/.claude/skills/create-chatccc-feishu-app/SKILL.md

@@ -0,0 +1,85 @@
+---
+name: create-chatccc-feishu-app
+description: Create, clone, or configure a Feishu enterprise self-built bot application for ChatCCC. Use when asked to create a ChatCCC Feishu app, imitate an existing lowercase `chatccc` app, prepare a bot for another computer, or configure its bot capability, `im:`/`cardkit:` permissions, events, credentials, version, and internal release from this repository's README.
+---
+
+# Create a ChatCCC Feishu App
+
+Use the logged-in Feishu developer console to create and completely configure a ChatCCC bot app. Treat this repository's README as the minimum requirement and the existing app named exactly lowercase `chatccc` as the configuration template.
+
+## 1. Establish the target and authorization
+
+- Read the `README.md` section `#### 飞书（推荐）` before changing the console.
+- Determine the requested app name, target computer, description, icon, availability, and whether the user wants configuration only or a real release. Derive a concise description when one is not supplied; for example, `用飞书聊天控制<目标电脑>上的 Claude Code / Cursor / Codex。`.
+- Use the browser surface the user requested. If the user specifies Chrome DevTools/CDP, connect directly to `chromeDevtools.port` or the README default `15166`; do not substitute extension-based Chrome control.
+- Obtain action-time confirmation before creating the app, granting permissions, saving event/callback subscriptions, or publishing. One explicit instruction such as `完整配置并真正发布` may authorize the listed actions as a batch.
+
+## 2. Inventory the lowercase template
+
+1. Open `https://open.feishu.cn/app?from=devbotmenu` in the authenticated browser.
+2. Open the app named exactly lowercase `chatccc` in a separate tab.
+3. Record its non-secret configuration:
+   - app description and icon;
+   - application capabilities, especially `机器人`;
+   - every granted permission and its identity type;
+   - event and callback subscription modes;
+   - every event and callback identifier;
+   - version availability, external sharing choices, and release state.
+4. Never reveal, log, export to chat, or persist the template App Secret.
+
+Prefer current console data over a stale hard-coded list. At the time this workflow was verified, lowercase `chatccc` had 62 application-identity permissions, one user-identity permission, 12 application events, and one callback. Use those counts as a sanity check, not as a replacement for inspecting the template.
+
+## 3. Create the target app
+
+1. Choose `创建企业自建应用` and fill the requested name and description.
+2. Reuse the lowercase `chatccc` icon only when imitation was requested, and verify its crop preview.
+3. Click the final `创建` action only after confirmation.
+4. Click once and wait for SPA navigation. Verify the exact app name, App ID, URL, and status before retrying; the console can temporarily show stale text after a successful mutation.
+
+## 4. Clone the complete ChatCCC configuration
+
+Perform the configuration in this order:
+
+1. Add the `机器人` application capability and verify that the page offers `删除能力`, which means the capability is enabled.
+2. Clone permissions with `权限管理` > `批量处理` > `批量导入/导出权限`:
+   - export from lowercase `chatccc`;
+   - import into the target app;
+   - confirm all additions and verify both application-identity and user-identity totals;
+   - do not save the export JSON in the repository or chat.
+3. Under `事件与回调` > `事件配置`, set the subscription mode to `使用长连接接收事件`, then add every event present in the template. The verified baseline is:
+   - `im.chat.disbanded_v1`
+   - `im.chat.member.bot.added_v1`
+   - `im.chat.member.bot.deleted_v1`
+   - `im.chat.member.user.added_v1`
+   - `im.chat.member.user.deleted_v1`
+   - `im.chat.member.user.withdrawn_v1`
+   - `im.message.bot_muted_v1`
+   - `im.message.message_read_v1`
+   - `im.message.reaction.created_v1`
+   - `im.message.reaction.deleted_v1`
+   - `im.message.recalled_v1`
+   - `im.message.receive_v1`
+4. Under `回调配置`, independently set `使用长连接接收回调` and add `card.action.trigger`. Configuring the event mode does not configure the callback mode.
+5. Verify that every listed event shows a sufficient permission as `已开通`.
+
+The README minimum is the bot capability, relevant `im:` and `cardkit:` permissions, `im.message.receive_v1`, and `card.action.trigger`; a clone request requires the complete current template configuration, not only that minimum.
+
+## 5. Create and publish a version
+
+1. Open `版本管理与发布` and create a version such as `1.0.0`.
+2. Keep mobile and desktop default abilities set to `机器人` and write a meaningful update note.
+3. Match the requested or template availability. For a personal computer, default to the owner-only internal range and leave external group/private-chat sharing disabled unless explicitly requested.
+4. Save the version, then complete the separate `确认发布` action. A successful version creation alone is not a release.
+5. Refresh the version list and require all of the following evidence before reporting success:
+   - application status `已启用`;
+   - banner `当前修改均已发布`;
+   - version status `已发布` with a publication time.
+6. If administrator review is required, submit everything authorized and report the exact remaining approval state.
+
+## 6. Verify and hand off safely
+
+- Re-open the bot, permission, event/callback, and version pages after publication. Verify the bot is enabled, permission identity counts match, both subscription modes are long connection, event/callback identifiers match, and the version is published.
+- Treat network responses and a refreshed page as stronger evidence than an immediately stale SPA view. Never repeat a create, add, or publish action only because the first page read is stale.
+- Report the app name, App ID, capability, permission counts, event/callback counts, version, release state, and any remaining action.
+- Never print the App Secret. If the user explicitly asks to configure ChatCCC, transfer the App ID and App Secret directly into the local private configuration without echoing the secret in commands, logs, screenshots, files tracked by Git, or chat.
+- Do not modify local ChatCCC configuration or transmit credentials elsewhere unless explicitly authorized.

package/.claude/skills/create-chatccc-feishu-app/agents/openai.yaml

@@ -0,0 +1,4 @@
+interface:
+  display_name: "Create ChatCCC Feishu App"
+  short_description: "Create and configure a Feishu bot app for ChatCCC"
+  default_prompt: "Use $create-chatccc-feishu-app to create and configure a Feishu enterprise bot app for this ChatCCC project."

package/.cursor/skills/create-chatccc-feishu-app/SKILL.md

@@ -0,0 +1,85 @@
+---
+name: create-chatccc-feishu-app
+description: Create, clone, or configure a Feishu enterprise self-built bot application for ChatCCC. Use when asked to create a ChatCCC Feishu app, imitate an existing lowercase `chatccc` app, prepare a bot for another computer, or configure its bot capability, `im:`/`cardkit:` permissions, events, credentials, version, and internal release from this repository's README.
+---
+
+# Create a ChatCCC Feishu App
+
+Use the logged-in Feishu developer console to create and completely configure a ChatCCC bot app. Treat this repository's README as the minimum requirement and the existing app named exactly lowercase `chatccc` as the configuration template.
+
+## 1. Establish the target and authorization
+
+- Read the `README.md` section `#### 飞书（推荐）` before changing the console.
+- Determine the requested app name, target computer, description, icon, availability, and whether the user wants configuration only or a real release. Derive a concise description when one is not supplied; for example, `用飞书聊天控制<目标电脑>上的 Claude Code / Cursor / Codex。`.
+- Use the browser surface the user requested. If the user specifies Chrome DevTools/CDP, connect directly to `chromeDevtools.port` or the README default `15166`; do not substitute extension-based Chrome control.
+- Obtain action-time confirmation before creating the app, granting permissions, saving event/callback subscriptions, or publishing. One explicit instruction such as `完整配置并真正发布` may authorize the listed actions as a batch.
+
+## 2. Inventory the lowercase template
+
+1. Open `https://open.feishu.cn/app?from=devbotmenu` in the authenticated browser.
+2. Open the app named exactly lowercase `chatccc` in a separate tab.
+3. Record its non-secret configuration:
+   - app description and icon;
+   - application capabilities, especially `机器人`;
+   - every granted permission and its identity type;
+   - event and callback subscription modes;
+   - every event and callback identifier;
+   - version availability, external sharing choices, and release state.
+4. Never reveal, log, export to chat, or persist the template App Secret.
+
+Prefer current console data over a stale hard-coded list. At the time this workflow was verified, lowercase `chatccc` had 62 application-identity permissions, one user-identity permission, 12 application events, and one callback. Use those counts as a sanity check, not as a replacement for inspecting the template.
+
+## 3. Create the target app
+
+1. Choose `创建企业自建应用` and fill the requested name and description.
+2. Reuse the lowercase `chatccc` icon only when imitation was requested, and verify its crop preview.
+3. Click the final `创建` action only after confirmation.
+4. Click once and wait for SPA navigation. Verify the exact app name, App ID, URL, and status before retrying; the console can temporarily show stale text after a successful mutation.
+
+## 4. Clone the complete ChatCCC configuration
+
+Perform the configuration in this order:
+
+1. Add the `机器人` application capability and verify that the page offers `删除能力`, which means the capability is enabled.
+2. Clone permissions with `权限管理` > `批量处理` > `批量导入/导出权限`:
+   - export from lowercase `chatccc`;
+   - import into the target app;
+   - confirm all additions and verify both application-identity and user-identity totals;
+   - do not save the export JSON in the repository or chat.
+3. Under `事件与回调` > `事件配置`, set the subscription mode to `使用长连接接收事件`, then add every event present in the template. The verified baseline is:
+   - `im.chat.disbanded_v1`
+   - `im.chat.member.bot.added_v1`
+   - `im.chat.member.bot.deleted_v1`
+   - `im.chat.member.user.added_v1`
+   - `im.chat.member.user.deleted_v1`
+   - `im.chat.member.user.withdrawn_v1`
+   - `im.message.bot_muted_v1`
+   - `im.message.message_read_v1`
+   - `im.message.reaction.created_v1`
+   - `im.message.reaction.deleted_v1`
+   - `im.message.recalled_v1`
+   - `im.message.receive_v1`
+4. Under `回调配置`, independently set `使用长连接接收回调` and add `card.action.trigger`. Configuring the event mode does not configure the callback mode.
+5. Verify that every listed event shows a sufficient permission as `已开通`.
+
+The README minimum is the bot capability, relevant `im:` and `cardkit:` permissions, `im.message.receive_v1`, and `card.action.trigger`; a clone request requires the complete current template configuration, not only that minimum.
+
+## 5. Create and publish a version
+
+1. Open `版本管理与发布` and create a version such as `1.0.0`.
+2. Keep mobile and desktop default abilities set to `机器人` and write a meaningful update note.
+3. Match the requested or template availability. For a personal computer, default to the owner-only internal range and leave external group/private-chat sharing disabled unless explicitly requested.
+4. Save the version, then complete the separate `确认发布` action. A successful version creation alone is not a release.
+5. Refresh the version list and require all of the following evidence before reporting success:
+   - application status `已启用`;
+   - banner `当前修改均已发布`;
+   - version status `已发布` with a publication time.
+6. If administrator review is required, submit everything authorized and report the exact remaining approval state.
+
+## 6. Verify and hand off safely
+
+- Re-open the bot, permission, event/callback, and version pages after publication. Verify the bot is enabled, permission identity counts match, both subscription modes are long connection, event/callback identifiers match, and the version is published.
+- Treat network responses and a refreshed page as stronger evidence than an immediately stale SPA view. Never repeat a create, add, or publish action only because the first page read is stale.
+- Report the app name, App ID, capability, permission counts, event/callback counts, version, release state, and any remaining action.
+- Never print the App Secret. If the user explicitly asks to configure ChatCCC, transfer the App ID and App Secret directly into the local private configuration without echoing the secret in commands, logs, screenshots, files tracked by Git, or chat.
+- Do not modify local ChatCCC configuration or transmit credentials elsewhere unless explicitly authorized.

package/.cursor/skills/create-chatccc-feishu-app/agents/openai.yaml

@@ -0,0 +1,4 @@
+interface:
+  display_name: "Create ChatCCC Feishu App"
+  short_description: "Create and configure a Feishu bot app for ChatCCC"
+  default_prompt: "Use $create-chatccc-feishu-app to create and configure a Feishu enterprise bot app for this ChatCCC project."

package/README.md

@@ -334,7 +334,7 @@ Codex 的默认模型和推理强度可继续由 `~/.codex/config.toml` 管理
 
 **微信：** 扫码登录后，在机器人私聊里发送 `/new` 或指定 Agent 的 `/new ...` 命令即可开始。功能与飞书基本一致，但展示为纯文本。
 
-**会话停滞保护：** 当 Agent 连续 3 分钟处于“正在生成回复”且回复字符数没有变化、同时尚未报告权威终态时，ChatCCC 会结束旧 CLI，并优先补发一次“完成了吗？如果没完成继续”；恢复轮再次发生相同停滞时不再递归续跑。Codex 只有 `turn.completed` 才算权威终态，阶段性的 `agent_message` 不算；任一 Agent 报告权威终态后若输出流仍超过 10 秒未关闭，ChatCCC 会强制清理该 CLI 并按正常完成收尾，不会重复询问 Agent。
+**会话停滞保护：** 当 Agent 连续 3 分钟停在“正在启动 Agent”且没有任何事件，或停在“正在生成回复”且回复字符数没有变化，同时尚未报告权威终态时，ChatCCC 会结束旧 CLI，并优先补发一次“完成了吗？如果没完成继续”；恢复轮再次发生相同停滞时不再递归续跑。`/new claude` 和 `/new cursor` 在等待底层 init 事件时也使用 3 分钟超时并主动清理 SDK/CLI。思考、搜索和工具调用阶段不按回复字符数误判，由进程资源监控负责识别真正僵死。Codex 只有 `turn.completed` 才算权威终态，阶段性的 `agent_message` 不算；任一 Agent 报告权威终态后若输出流仍超过 10 秒未关闭，ChatCCC 会强制清理该 CLI 并按正常完成收尾，不会重复询问 Agent。
 
 ## 可用指令
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "chatccc",
-  "version": "0.2.212",
+  "version": "0.2.214",
   "description": "Feishu bot bridge for Claude Code",
   "license": "Apache-2.0",
   "type": "module",
@@ -16,6 +16,9 @@
     "demo/ilink_echo_probe.ts",
     "agent-prompts/",
     "im-skills/",
+    ".agents/skills/create-chatccc-feishu-app/",
+    ".claude/skills/create-chatccc-feishu-app/",
+    ".cursor/skills/create-chatccc-feishu-app/",
     "images/img_readme_*.jpg",
     "images/img_readme_*.png",
     "images/avatars/status_*.png",

package/src/__tests__/claude-adapter.test.ts

@@ -379,6 +379,20 @@ describe("createClaudeAdapter", () => {
     await expect(adapter.closeSession("any-sid")).resolves.toBeUndefined();
   });
 
+  it("does not start SDK session creation when the startup signal is already aborted", async () => {
+    const adapter = createClaudeAdapter({
+      model: "claude-sonnet-4-6",
+      effort: "high",
+      isEmpty: () => false,
+    });
+    const controller = new AbortController();
+    controller.abort();
+
+    await expect(
+      adapter.createSession("F:\\repo", controller.signal),
+    ).rejects.toThrow("Claude session creation aborted");
+  });
+
   // -------------------------------------------------------------------------
   // getSessionInfo 行为契约
   //   - cwd 决定 /git 是否可用

package/src/__tests__/cursor-adapter.test.ts

@@ -347,19 +347,32 @@ describe("normalizeCursorMessage", () => {
 // createCursorAdapter — 工厂函数测试
 // ---------------------------------------------------------------------------
 
-describe("createCursorAdapter", () => {
+describe("createCursorAdapter", () => {
   it("returns adapter with correct displayName and sessionDescPrefix", () => {
     const adapter = createCursorAdapter();
     expect(adapter.displayName).toBe("Cursor");
     expect(adapter.sessionDescPrefix).toBe("Cursor Session:");
   });
 
-  it("closeSession does not throw", async () => {
-    const adapter = createCursorAdapter();
-    await expect(adapter.closeSession("any-id")).resolves.toBeUndefined();
-  });
-
-  // -------------------------------------------------------------------------
+  it("closeSession does not throw", async () => {
+    const adapter = createCursorAdapter();
+    await expect(adapter.closeSession("any-id")).resolves.toBeUndefined();
+  });
+
+  it("aborts a zero-output createSession stream instead of leaving Cursor Agent alive", async () => {
+    const store = createInMemoryMetaStore();
+    const spawnImpl = (() =>
+      createHangingMockCursorProcess({})) as CursorSpawnForTest;
+    const adapter = createCursorAdapter({ metaStore: store, spawn: spawnImpl });
+    const controller = new AbortController();
+
+    const pending = adapter.createSession("F:/repo", controller.signal);
+    controller.abort();
+
+    await expect(pending).rejects.toThrow("Cursor session creation aborted");
+  });
+
+  // -------------------------------------------------------------------------
   // getSessionInfo 行为契约
   //   - cwd 决定 /git 是否可用
   //   - model 决定 /state、/sessions 显示的是否是 Cursor 真实模型

package/src/__tests__/orchestrator.test.ts

@@ -427,7 +427,7 @@ describe("handleCommand WeChat processing ack", () => {
 
       await handleCommand(platform, "使用新的默认 Agent", "feishu-p2p", "ou-user", Date.now(), "p2p");
 
-      expect(createCursorSession).toHaveBeenCalledWith(homedir());
+      expect(createCursorSession).toHaveBeenCalledWith(homedir(), expect.any(AbortSignal));
       expect(oldPrompt).not.toHaveBeenCalled();
       expect(cursorPrompt).toHaveBeenCalledWith(
         "sid-new-cursor",
@@ -549,7 +549,7 @@ describe("handleCommand WeChat processing ack", () => {
 
     await handleCommand(platform, "帮我看一下日志", "feishu-p2p", "ou-user", Date.now(), "p2p");
 
-    expect(createSession).toHaveBeenCalledWith(homedir());
+    expect(createSession).toHaveBeenCalledWith(homedir(), expect.any(AbortSignal));
     expect(platform.createGroup).not.toHaveBeenCalled();
     expect(platform.updateChatInfo).not.toHaveBeenCalled();
     expect(prompt).toHaveBeenCalledWith(
@@ -596,7 +596,7 @@ describe("handleCommand WeChat processing ack", () => {
 
     await handleCommand(platform, "继续", "feishu-p2p", "ou-user", Date.now(), "p2p");
 
-    expect(createSession).toHaveBeenCalledWith(homedir());
+    expect(createSession).toHaveBeenCalledWith(homedir(), expect.any(AbortSignal));
     expect(prompt).toHaveBeenCalledWith(
       "sid-feishu-migrated",
       expect.stringContaining("继续"),
@@ -633,7 +633,7 @@ describe("handleCommand WeChat processing ack", () => {
 
     await handleCommand(platform, "/abd帮我看一下日志", "feishu-p2p", "ou-user", Date.now(), "p2p");
 
-    expect(createSession).toHaveBeenCalledWith(homedir());
+    expect(createSession).toHaveBeenCalledWith(homedir(), expect.any(AbortSignal));
     expect(platform.createGroup).not.toHaveBeenCalled();
     expect(platform.updateChatInfo).not.toHaveBeenCalled();
     const userText = prompt.mock.calls[0][1];
@@ -706,7 +706,7 @@ describe("handleCommand WeChat processing ack", () => {
 
     await handleCommand(platform, "/newh", "feishu-p2p", "ou-user", Date.now(), "p2p");
 
-    expect(createSession).toHaveBeenCalledWith(homedir());
+    expect(createSession).toHaveBeenCalledWith(homedir(), expect.any(AbortSignal));
     expect(platform.createGroup).not.toHaveBeenCalled();
     expect(platform.updateChatInfo).not.toHaveBeenCalled();
     const registry = await loadSessionRegistryForBinding();

package/src/__tests__/package-files.test.ts

@@ -0,0 +1,24 @@
+import { existsSync, readFileSync } from "node:fs";
+import { join } from "node:path";
+
+import { describe, expect, it } from "vitest";
+
+const CREATE_APP_SKILL_DIRS = [
+  ".agents/skills/create-chatccc-feishu-app/",
+  ".claude/skills/create-chatccc-feishu-app/",
+  ".cursor/skills/create-chatccc-feishu-app/",
+];
+
+describe("npm package files", () => {
+  it("ships the Feishu app creation skill for every supported agent", () => {
+    const root = process.cwd();
+    const packageJson = JSON.parse(readFileSync(join(root, "package.json"), "utf8")) as {
+      files?: string[];
+    };
+
+    for (const skillDir of CREATE_APP_SKILL_DIRS) {
+      expect(packageJson.files).toContain(skillDir);
+      expect(existsSync(join(root, skillDir, "SKILL.md"))).toBe(true);
+    }
+  });
+});

package/src/__tests__/session.test.ts

@@ -3,7 +3,7 @@ import { mkdtemp, rm, readFile } from "node:fs/promises";
 import { tmpdir } from "node:os";
 import { dirname, join } from "node:path";
 
-const killProcessTreeMock = vi.hoisted(() => vi.fn(async () => {}));
+const killProcessTreeMock = vi.hoisted(() => vi.fn(async (_pid?: number) => {}));
 vi.mock("../adapters/proc-tree-kill.ts", () => ({
   killProcessTree: killProcessTreeMock,
 }));
@@ -109,8 +109,9 @@ import {
   _clearAdapterCacheForTest,
   setSessionPlatform,
   recordChatPlatform,
-  _getPlatformForChatForTest,
-  runAgentSession,
+  _getPlatformForChatForTest,
+  initClaudeSession,
+  runAgentSession,
   stopSession,
   startUnifiedDisplayLoop,
   stopUnifiedDisplayLoop,
@@ -226,7 +227,7 @@ function mockPlatform(name: string): PlatformAdapter {
   };
 }
 
-describe("resetState", () => {
+describe("resetState", () => {
   it("clears all maps and sets", () => {
     chatSessionMap.set("chat1", {
       gen: 1, close: () => {}, cardId: null, stopped: false,
@@ -245,9 +246,58 @@ describe("resetState", () => {
     expect(sessionInfoMap.size).toBe(0);
     expect(processedMessages.size).toBe(0);
   });
-});
-
-describe("chat platform routing", () => {
+});
+
+describe("initClaudeSession startup watchdog", () => {
+  afterEach(() => {
+    _clearAdapterCacheForTest();
+    _resetResponseStallTimeoutForTest();
+    vi.useRealTimers();
+  });
+
+  it("aborts session creation when an Agent never emits its init event", async () => {
+    vi.useFakeTimers();
+    _setResponseStallTimeoutForTest(100);
+
+    let aborted = false;
+    const adapter: ToolAdapter = {
+      displayName: "Silent Cursor",
+      sessionDescPrefix: "Cursor Session:",
+      createSession: async (_cwd: string, signal?: AbortSignal) => {
+        await new Promise<void>((resolve) => {
+          if (signal?.aborted) {
+            aborted = true;
+            resolve();
+            return;
+          }
+          signal?.addEventListener("abort", () => {
+            aborted = true;
+            resolve();
+          }, { once: true });
+        });
+        throw new Error("adapter aborted");
+      },
+      prompt: async function* () {},
+      getSessionInfo: async () => undefined,
+      closeSession: async () => {},
+    };
+    _setAdapterForToolForTest("cursor", adapter);
+
+    let outcome: unknown;
+    void initClaudeSession("cursor", "F:\\repo").then(
+      (value) => { outcome = value; },
+      (err) => { outcome = err; },
+    );
+
+    await vi.advanceTimersByTimeAsync(101);
+
+    expect(aborted).toBe(true);
+    expect(outcome).toBeInstanceOf(Error);
+    expect((outcome as Error).message).toContain("3 minutes");
+  });
+});
+
+describe("chat platform routing", () => {
   beforeEach(() => {
     resetState();
   });
@@ -774,6 +824,88 @@ describe("runAgentSession response stall watchdog", () => {
     });
   });
 
+  it("auto-ends an Agent that stays in starting state without emitting any event", async () => {
+    vi.setSystemTime(0);
+    _setResponseStallTimeoutForTest(180_000);
+    _setResponseStallCheckIntervalForTest(1_000);
+    _setProcessAliveForTest(() => true);
+
+    const platform = mockPlatform("feishu");
+    setSessionPlatform(platform);
+    bindChatToSession("sid-starting-stall", "chat-starting-stall");
+    recordLastActiveChat("sid-starting-stall", "chat-starting-stall");
+
+    const closeSession = vi.fn();
+    const adapter: ToolAdapter = {
+      displayName: "Silent Agent",
+      sessionDescPrefix: "Agent Session:",
+      createSession: async () => ({ sessionId: "sid-starting-stall" }),
+      getSessionInfo: async (sid) => ({ sessionId: sid, cwd: "F:\\repo" }),
+      closeSession: async () => {},
+      prompt: async function* (
+        _sid: string,
+        _text: string,
+        _cwd: string,
+        signal?: AbortSignal,
+        options?: ToolPromptOptions,
+      ) {
+        options?.onSessionCreated?.(closeSession);
+        options?.onProcessStart?.({ pid: 4343 });
+        await new Promise<void>((resolve) => {
+          if (signal?.aborted) {
+            resolve();
+            return;
+          }
+          signal?.addEventListener("abort", () => resolve(), { once: true });
+        });
+        if (false) yield { type: "assistant", blocks: [] };
+      },
+    };
+    _setAdapterForToolForTest("codex", adapter);
+
+    const runPromise = runAgentSession(
+      "sid-starting-stall",
+      "prompt",
+      platform,
+      "chat-starting-stall",
+      Date.now(),
+      "codex",
+    );
+
+    await vi.waitFor(() => {
+      expect(activePrompts.get("sid-starting-stall")).toMatchObject({
+        processPid: 4343,
+        responseProgress: {
+          totalChars: 0,
+          unchangedSince: expect.any(Number),
+        },
+      });
+      expect(closeSession).toHaveBeenCalledTimes(0);
+    });
+
+    await vi.advanceTimersByTimeAsync(181_001);
+    const stateAfterDeadline = mockStreamStates.get("sid-starting-stall");
+    const closeCountAfterDeadline = closeSession.mock.calls.length;
+    const treeKillCountAfterDeadline = killProcessTreeMock.mock.calls
+      .filter(([pid]) => pid === 4343)
+      .length;
+
+    // 旧实现不会结束真正的零事件启动；先清理测试会话，避免失败用例悬挂。
+    if (stateAfterDeadline?.status !== "auto_ended") {
+      stopSession("sid-starting-stall");
+      await vi.advanceTimersByTimeAsync(1_000);
+    }
+    await runPromise;
+
+    expect(stateAfterDeadline).toMatchObject({
+      status: "auto_ended",
+      finalReply: "",
+      autoEndedAt: expect.any(Number),
+    });
+    expect(closeCountAfterDeadline).toBe(1);
+    expect(treeKillCountAfterDeadline).toBeGreaterThanOrEqual(1);
+  });
+
   it("self-heals a missing trigger-chat binding and gives automatic recovery the normal card lifecycle", async () => {
     vi.setSystemTime(0);
     _setResponseStallTimeoutForTest(100);
@@ -1142,7 +1274,7 @@ describe("runAgentSession response stall watchdog", () => {
     expect(consumeQueued).not.toHaveBeenCalled();
 
     await vi.advanceTimersByTimeAsync(200);
-    expect(consumeQueued).toHaveBeenCalledTimes(1);
+    await vi.waitFor(() => expect(consumeQueued).toHaveBeenCalledTimes(1));
     expect(consumeQueued).toHaveBeenCalledWith(
       platform,
       expect.objectContaining({ text: "queued user prompt" }),
@@ -1215,7 +1347,7 @@ describe("runAgentSession response stall watchdog", () => {
     expect(receivedPrompts[1]).toContain(recoveryPrompt);
     expect(platform.sendText).toHaveBeenCalledWith(
       "chat-recovery-limit",
-      "⚠️ 自动续跑仍连续 3 分钟没有新回复，本次不再自动继续。",
+      "⚠️ 自动续跑仍连续 3 分钟没有启动进展或新回复，本次不再自动继续。",
     );
   });
 
@@ -1474,7 +1606,7 @@ describe("unified display loop terminal card update", () => {
     expect(card.header.template).toBe("orange");
     expect(platform.sendText).toHaveBeenCalledWith(
       "chat-auto-ended",
-      "⚠️ 已自动结束：连续 3 分钟处于“正在生成回复”且回复字符总数没有变化。本轮没有可发送的回复内容。",
+      "⚠️ 已自动结束：连续 3 分钟没有启动进展或回复字符变化。本轮没有可发送的回复内容。",
     );
     expect(displayCards.has("chat-auto-ended")).toBe(false);
   });

package/src/__tests__/stop-session.test.ts

@@ -11,10 +11,15 @@ import { describe, it, expect, vi, beforeEach } from "vitest";
 import type { StreamState } from "../stream-state.ts";
 
 // mock stream-state，使用模块内可观测的 Map 记录读写
-const stateStore = new Map<string, StreamState>();
-const writeCalls: StreamState[] = [];
-
-vi.mock("../stream-state.ts", () => ({
+const stateStore = new Map<string, StreamState>();
+const writeCalls: StreamState[] = [];
+const killProcessTreeMock = vi.hoisted(() => vi.fn(async (_pid?: number) => {}));
+
+vi.mock("../adapters/proc-tree-kill.ts", () => ({
+  killProcessTree: killProcessTreeMock,
+}));
+
+vi.mock("../stream-state.ts", () => ({
   readStreamState: async (sid: string): Promise<StreamState | null> => {
     return stateStore.get(sid) ?? null;
   },
@@ -72,9 +77,10 @@ async function flush(): Promise<void> {
 }
 
 beforeEach(() => {
-  activePrompts.clear();
-  stateStore.clear();
-  writeCalls.length = 0;
+  activePrompts.clear();
+  stateStore.clear();
+  writeCalls.length = 0;
+  killProcessTreeMock.mockClear();
 });
 
 describe("stopSession 行为护栏", () => {
@@ -128,13 +134,29 @@ describe("stopSession 行为护栏", () => {
     expect(activePrompts.get("sid-C")?.stopped).toBe(true);
   });
 
-  it("调用 adapter 提供的 closeSession 以主动关闭底层 SDK session", () => {
+  it("调用 adapter 提供的 closeSession 以主动关闭底层 SDK session", () => {
     const closeSession = vi.fn();
     seedRunningSession("sid-D", "partial output", closeSession);
 
     const ok = stopSession("sid-D");
 
-    expect(ok).toBe(true);
-    expect(closeSession).toHaveBeenCalledTimes(1);
-  });
-});
+    expect(ok).toBe(true);
+    expect(closeSession).toHaveBeenCalledTimes(1);
+  });
+
+  it("按完整进程树停止 CLI，不先杀壳进程造成后代进程逃逸", async () => {
+    seedRunningSession("sid-tree");
+    activePrompts.get("sid-tree")!.processPid = 4242;
+    const processKillSpy = vi.spyOn(process, "kill").mockImplementation(() => true);
+
+    try {
+      expect(stopSession("sid-tree")).toBe(true);
+      await flush();
+
+      expect(killProcessTreeMock).toHaveBeenCalledWith(4242);
+      expect(processKillSpy).not.toHaveBeenCalled();
+    } finally {
+      processKillSpy.mockRestore();
+    }
+  });
+});

package/src/adapters/adapter-interface.ts

@@ -152,11 +152,12 @@ export interface ToolAdapter {
   /** 群描述中会话 ID 的前缀，如 "Claude Session:" */
   readonly sessionDescPrefix: string;
 
-  /**
-   * 创建新会话，返回会话 ID。
-   * 适配器内部需处理后台流消费（静默消费 stream 中除 init 外的所有事件）。
-   */
-  createSession(cwd: string): Promise<CreateSessionResult>;
+  /**
+   * 创建新会话，返回会话 ID。
+   * 适配器内部需处理后台流消费（静默消费 stream 中除 init 外的所有事件）。
+   * signal 用于上层在长期收不到 init 事件时终止底层 SDK/CLI。
+   */
+  createSession(cwd: string, signal?: AbortSignal): Promise<CreateSessionResult>;
 
   /**
    * 向已有会话发送提示文本，返回归一化消息的异步迭代器。

package/src/adapters/claude-adapter.ts

@@ -452,9 +452,14 @@ class ClaudeAdapter implements ToolAdapter {
     this.maxTurn = options.maxTurn ?? 0;
   }
 
-  async createSession(cwd: string): Promise<CreateSessionResult> {
+  async createSession(cwd: string, signal?: AbortSignal): Promise<CreateSessionResult> {
     logMcpConfig();
     const abortController = new AbortController();
+    const removeAbortListener = bridgeAbortSignal(signal, abortController);
+    if (abortController.signal.aborted) {
+      removeAbortListener?.();
+      throw new Error("Claude session creation aborted");
+    }
     let sessionId: string | undefined;
     const session = unstable_v2_createSession(
       toSdkSessionOptions(buildSdkOptions({
@@ -486,6 +491,7 @@ class ClaudeAdapter implements ToolAdapter {
         }
       }
     } finally {
+      removeAbortListener?.();
       closeSdkSession(session);
     }
 

package/src/adapters/cursor-adapter.ts

@@ -511,34 +511,40 @@ class CursorAdapter implements ToolAdapter {
     this.badJsonIdleTimeoutMs = badJsonIdleTimeoutMs;
   }
 
-  async createSession(cwd: string): Promise<CreateSessionResult> {
+  async createSession(cwd: string, signal?: AbortSignal): Promise<CreateSessionResult> {
+    if (signal?.aborted) throw new Error("Cursor session creation aborted");
     const handle = spawnAgent(["ok"], cwd, undefined, this.modelOverride, undefined, this.spawnImpl);
     const proc = handle.proc;
     const stats = createCursorStreamStats();
     this.activeProcs.add(proc);
+    const onAbort = () => { void killProcessTree(proc.pid); };
+    signal?.addEventListener("abort", onAbort, { once: true });
 
-    for await (const msg of readJsonLines(proc, undefined, "createSession", null, stats, this.badJsonIdleTimeoutMs)) {
-      if (msg.type === "system" && msg.subtype === "init" && msg.session_id) {
-        const sessionId = msg.session_id;
-        await this.metaStore
-          .set(sessionId, { cwd: msg.cwd ?? cwd, model: msg.model })
-          .catch(() => {});
-        this.activeProcs.delete(proc);
-        await killProcessTree(proc.pid);
-        return { sessionId };
+    try {
+      for await (const msg of readJsonLines(proc, signal, "createSession", null, stats, this.badJsonIdleTimeoutMs)) {
+        if (msg.type === "system" && msg.subtype === "init" && msg.session_id) {
+          const sessionId = msg.session_id;
+          await this.metaStore
+            .set(sessionId, { cwd: msg.cwd ?? cwd, model: msg.model })
+            .catch(() => {});
+          return { sessionId };
+        }
       }
-    }
 
-    await killProcessTree(proc.pid);
-    this.activeProcs.delete(proc);
-    const closeInfo = await handle.waitForClose();
-    const visibleMessage = formatCursorAgentEmptyOutputMessage({
-      exitCode: closeInfo.code,
-      stdoutLength: stats.stdoutLength,
-      stderr: closeInfo.stderr,
-    });
-    if (visibleMessage) throw new Error(visibleMessage);
-    throw new Error("No session ID in Cursor init event");
+      if (signal?.aborted) throw new Error("Cursor session creation aborted");
+      const closeInfo = await handle.waitForClose();
+      const visibleMessage = formatCursorAgentEmptyOutputMessage({
+        exitCode: closeInfo.code,
+        stdoutLength: stats.stdoutLength,
+        stderr: closeInfo.stderr,
+      });
+      if (visibleMessage) throw new Error(visibleMessage);
+      throw new Error("No session ID in Cursor init event");
+    } finally {
+      signal?.removeEventListener("abort", onAbort);
+      await killProcessTree(proc.pid);
+      this.activeProcs.delete(proc);
+    }
   }
 
   async *prompt(

package/src/response-stall.ts

@@ -1,20 +1,20 @@
-/** A snapshot of response output progress while the Agent is generating a reply. */
+/** A snapshot of visible output progress while the Agent is starting or replying. */
 export interface ResponseProgressObservation {
   totalChars: number;
   unchangedSince: number;
 }
 
 /**
- * Tracks how long the displayed response character count has remained unchanged.
- * Leaving the responding phase clears the window; returning starts a fresh one.
+ * Tracks how long the displayed output character count has remained unchanged.
+ * Leaving a monitored phase clears the window; returning starts a fresh one.
  */
 export function observeResponseProgress(
   previous: ResponseProgressObservation | undefined,
-  isResponding: boolean,
+  isMonitoredPhase: boolean,
   totalChars: number,
   now = Date.now(),
 ): ResponseProgressObservation | undefined {
-  if (!isResponding) return undefined;
+  if (!isMonitoredPhase) return undefined;
   if (previous?.totalChars === totalChars) return previous;
   return { totalChars, unchangedSince: now };
 }

package/src/session.ts

@@ -27,6 +27,7 @@ import {
   formatAgentActivityTitle,
   updateAgentActivity,
 } from "./agent-activity.ts";
+import type { AgentActivityKind } from "./agent-activity.ts";
 import { simplifyToolUse, simplifyToolResult } from "./simplify.ts";
 import { logTrace } from "./trace.ts";
 import type { UnifiedBlock } from "./adapters/adapter-interface.ts";
@@ -170,7 +171,7 @@ export const RESPONSE_STALL_RECOVERY_PROMPT = "完成了吗？如果没完成继
 export const RESPONSE_STALL_RECOVERY_NOTICE =
   `检测到会话停滞，正在自动确认并继续。\n\n${RESPONSE_STALL_RECOVERY_PROMPT}`;
 export const RESPONSE_STALL_RECOVERY_EXHAUSTED_NOTICE =
-  "⚠️ 自动续跑仍连续 3 分钟没有新回复，本次不再自动继续。";
+  "⚠️ 自动续跑仍连续 3 分钟没有启动进展或新回复，本次不再自动继续。";
 const RESPONSE_STALL_RECOVERY_DELAY_MS = 200;
 let processMonitorIntervalMs = DEFAULT_PROCESS_MONITOR_INTERVAL_MS;
 let responseStallTimeoutMs = DEFAULT_RESPONSE_STALL_TIMEOUT_MS;
@@ -314,12 +315,20 @@ function turnFinalStatus(status: "running" | "done" | "stopped" | "error" | "aut
 }
 
 function formatAutoEndedReply(finalReply: string): string {
-  const reason = "⚠️ 已自动结束：连续 3 分钟处于“正在生成回复”且回复字符总数没有变化。";
+  const reason = "⚠️ 已自动结束：连续 3 分钟没有启动进展或回复字符变化。";
   return finalReply
     ? `${reason}以下回复可能不完整。\n\n${finalReply}`
     : `${reason}本轮没有可发送的回复内容。`;
 }
 
+/**
+ * 只监控用户无法判断是否仍有进展的两个阶段。思考、工具调用和搜索可能合法地
+ * 长时间不产生回复字符，由资源监控负责识别真正僵死，不能在这里误杀。
+ */
+function monitorsOutputProgress(kind: AgentActivityKind): boolean {
+  return kind === "starting" || kind === "responding";
+}
+
 function formatTerminalReply(
   status: "running" | "done" | "stopped" | "error" | "auto_ended",
   finalReply: string,
@@ -1010,15 +1019,40 @@ function formatToolConfigForLog(tool: string, sessionModel?: string, sessionId?:
   return `model=${anthropicConfigDisplay(getModelForSession(sessionId))}, subagentModel=${anthropicConfigDisplay(CLAUDE_SUBAGENT_MODEL)}, effort=${anthropicConfigDisplay(getEffectiveEffortForTool(tool, sessionId))}`;
 }
 
-export async function initClaudeSession(tool: string, overrideCwd?: string, chatId?: string): Promise<{ sessionId: string; cwd: string }> {
-  const cwd = overrideCwd ?? (await getDefaultCwd(chatId));
-  const adapter = getAdapterForTool(tool);
+export async function initClaudeSession(tool: string, overrideCwd?: string, chatId?: string): Promise<{ sessionId: string; cwd: string }> {
+  const cwd = overrideCwd ?? (await getDefaultCwd(chatId));
+  const adapter = getAdapterForTool(tool);
   console.log(
-    `[${ts()}] [STEP 1/5] Creating ${adapter.displayName} session (${formatToolConfigForLog(tool)}, cwd=${cwd})`
-  );
-
-  const result = await adapter.createSession(cwd);
-  const sessionId = result.sessionId;
+    `[${ts()}] [STEP 1/5] Creating ${adapter.displayName} session (${formatToolConfigForLog(tool)}, cwd=${cwd})`
+  );
+
+  // Claude/Cursor 创建会话时需要先等待 SDK/CLI 的 init 事件。它们若在首个
+  // 事件前卡死，正式 turn 尚未建立，runAgentSession 的看门狗无法介入。
+  // 因此创建入口也使用相同的三分钟阈值，并通过 AbortSignal 释放底层资源。
+  const createController = new AbortController();
+  let createTimeout: ReturnType<typeof setTimeout> | undefined;
+  const timeoutError = new Error(
+    `${adapter.displayName} session creation timed out after 3 minutes without an init event`,
+  );
+  const timeoutPromise = new Promise<never>((_resolve, reject) => {
+    createTimeout = setTimeout(() => {
+      // 先固定对外错误，再 abort 适配器，避免适配器自己的 abort 错误赢得竞态。
+      reject(timeoutError);
+      createController.abort();
+    }, responseStallTimeoutMs);
+    createTimeout.unref?.();
+  });
+
+  let result: Awaited<ReturnType<ToolAdapter["createSession"]>>;
+  try {
+    result = await Promise.race([
+      adapter.createSession(cwd, createController.signal),
+      timeoutPromise,
+    ]);
+  } finally {
+    if (createTimeout) clearTimeout(createTimeout);
+  }
+  const sessionId = result.sessionId;
   console.log(`[${ts()}]   → sessionId: ${sessionId}`);
 
   await saveSessionTool(sessionId, tool);
@@ -1334,6 +1368,16 @@ export async function runAgentSession(
 
   const runningPrompt = activePrompts.get(sessionId);
   if (runningPrompt) {
+    // 必须在消费第一个事件前建立零字符基线。部分 CLI 卡死时只启动了进程，
+    // 甚至一个事件都不会 yield；若等循环体更新进度，这种会话会永久停在
+    // “正在启动 Agent”，也永远触发不了三分钟保护。
+    runningPrompt.responseProgress = observeResponseProgress(
+      undefined,
+      true,
+      0,
+      activityTracker.activity.startedAt,
+    );
+
     const checkResponseStall = async () => {
       const current = activePrompts.get(sessionId);
       if (!current || current !== runningPrompt) {
@@ -1346,7 +1390,7 @@ export async function runAgentSession(
         || current.resourceStuck
         || current.autoEnded
         || current.finalResponseObserved
-        || activityTracker.activity.kind !== "responding"
+        || !monitorsOutputProgress(activityTracker.activity.kind)
         || !hasResponseStalled(current.responseProgress, Date.now(), responseStallTimeoutMs)
       ) {
         return;
@@ -1402,7 +1446,7 @@ export async function runAgentSession(
       current.controller.abort();
       await killProcessTree(current.processPid);
       console.warn(
-        `[${ts()}] [RESPONSE-STALL] Session ${sessionId} auto-ended after 3 minutes without reply character changes`,
+        `[${ts()}] [RESPONSE-STALL] Session ${sessionId} auto-ended after 3 minutes without startup or reply progress`,
       );
     };
 
@@ -1466,8 +1510,10 @@ export async function runAgentSession(
       if (prompt && !prompt.autoEnded) {
         const totalChars = state.accumulatedContent.length + pickFinalReply(state).length;
         prompt.responseProgress = observeResponseProgress(
-          prompt.responseProgress,
-          activityTracker.activity.kind === "responding",
+          // starting → responding 本身是一次有效进展，即便首个文本块仍为空也应
+          // 重新计时；其它活动阶段会清空观察窗口。
+          activityChanged ? undefined : prompt.responseProgress,
+          monitorsOutputProgress(activityTracker.activity.kind),
           totalChars,
           Date.now(),
         );
@@ -1644,7 +1690,7 @@ export async function runAgentSession(
           autoRecoveryTarget = { chatId: activeAutoEnded, platform: pp };
         }
       }
-      console.warn(`[${ts()}] Session ${sessionId} auto-ended after stalled response output (content chunks: ${state.chunkCount})`);
+      console.warn(`[${ts()}] Session ${sessionId} auto-ended after stalled startup or response output (content chunks: ${state.chunkCount})`);
       if (tid) logTrace(tid, "SESSION_END", { sessionId, outcome: "response_stall", chunks: state.chunkCount });
     } else if (wasAbnormalExit) {
       for (const cid of finalizationChatIds) {
@@ -2149,21 +2195,20 @@ export function stopSession(sessionId: string): boolean {
   clearPromptResponseStallMonitor(sessionId);
   clearPromptProcessMonitor(sessionId);
   clearPromptFinalResponseCloseTimer(sessionId);
-  cancelQueuedMessage(sessionId);
-  try {
-    prompt.closeSession?.();
-  } catch (err) {
+  cancelQueuedMessage(sessionId);
+
+  // 先发起整棵进程树清理，再触发 close/abort。Windows 上 CLI 由
+  // cmd.exe → node → 实际二进制组成；若先 process.kill(cmd.exe)，taskkill
+  // 随后便无法从已消失的根 PID 找到后代，正是幽灵 Codex/Cursor 的来源。
+  // killProcessTree 在返回 Promise 前已启动 taskkill，因此这里无需阻塞。
+  void killProcessTree(prompt.processPid);
+  try {
+    prompt.closeSession?.();
+  } catch (err) {
     console.warn(`[${ts()}] [STOP] closeSession failed for ${sessionId}: ${(err as Error).message}`);
-  }
-  prompt.controller.abort();
-
-  // 强制杀死 CLI 子进程。controller.abort() 只在 for-await 收到下一条
-  // stream 消息时才被检测到——如果 agent 陷入无输出的计算循环，abort 信号
-  // 永远不会生效。直接 process.kill 让 stdout/stderr 管道关闭，stream 立即结束。
-  if (prompt.processPid !== undefined) {
-    try { process.kill(prompt.processPid); } catch { /* 进程可能已退出 */ }
-  }
-  console.log(`[${ts()}] [STOP] Session ${sessionId} aborted`);
+  }
+  prompt.controller.abort();
+  console.log(`[${ts()}] [STOP] Session ${sessionId} aborted`);
 
   // fire-and-forget：立刻把 stream-state.status 改成 stopped，
   // 让 display loop 下一次扫到立刻渲染"已停止"卡片，不必再等几秒。