diff --git a/.planning/conversation-stale-running-fix/findings.md b/.planning/conversation-stale-running-fix/findings.md
new file mode 100644
index 0000000..79dabd2
--- /dev/null
+++ b/.planning/conversation-stale-running-fix/findings.md
@@ -0,0 +1,79 @@
+# 调研记录
+
+## 用户现象
+
+- 助手正文和两个 Shell Command 均已完整结束，工具退出码为 0。
+- 页面底部仍显示“运行中”，停止按钮仍存在。
+- 随后发送“还在么”时插入失败，服务端提示：`Codex 任务失败：no active turn to steer`。
+
+## 初始假设
+
+- cc-web 的本地/持久化会话仍认为 turn 正在运行，而 Codex app-server 已释放 active turn。
+- 发送路径依据陈旧运行态选择 `steer`，缺少对 app-server 明确“无活动 turn”响应的状态修复与普通 `turn/start` 降级。
+
+## 工作区约束
+
+- 检查开始时 Git 工作树干净。
+- `.planning/.active_plan` 属于并发中的 `subagent-card-metadata`，本任务使用独立 scoped 目录且不覆盖 active 指针。
+- Trellis 当前任务也属于其他工作，本任务不切换共享指针。
+
+## 项目上下文
+
+- 独立计划审查已通过，无阻塞问题；建议在进度记录中保留实际测试命令与 60 秒超时证据。
+- `codebase-memory-mcp` 项目 `home-cc-web` 索引状态为 `ready`，当前包含 3246 个节点、7821 条边，无需重新全量索引。
+
+## 初步代码命中
+
+- 后端 `handleCodexAppSteerMessage` 直接读取 `activeCodexAppTurns` 中的 `threadId/turnId`，请求 `turn/steer`；失败分支目前只发送 `failed` 状态与 `codexapp_steer_failed` 错误，未见状态自愈或普通 turn 降级。
+- 前端 `handleServerMessage` 只把 `codex_app_steer_status` 映射成 pending/inserted/failed 展示；是否运行的权威来源仍需沿 `done`、`session_info` 和 `session_list` 消息继续确认。
+- `handleLoadSession` 返回的 `isRunning` 来自服务端 `isSessionRunning(sessionId)`，说明截图底部“运行中”不是单纯 DOM 残留，核心漂移大概率位于后端活动 turn Map 或其终态通知处理。
+- mock 已有 `turn/steer` 和 `turn/completed` 基础能力，`scripts/regression.js` 也已有运行中 steer 正常路径，可在其附近追加竞态回归。
+
+## 状态机证据
+
+- `isSessionRunning` 直接以 `activeCodexAppTurns.has(sessionId)` 判定 Codex App 运行态；只要 Map 条目没有删除，所有页面和再次发送都会继续认为 turn 活跃。
+- `processCodexAppNotification` 只有收到 `turn/completed` 或 `error` 才返回 `done: true`；`handleCodexAppNotification` 随后调用 `handleCodexAppTurnComplete` 删除活动条目并广播 `done`。
+- `handleMessage` 在持久化新普通消息之前先检查 `activeCodexAppTurns`，存在条目便无条件转入 `handleCodexAppSteerMessage`。
+- steer 路径会先把用户消息持久化，再异步请求 `turn/steer`。失败分支目前仅显示错误，不删除陈旧活动条目，也不为已经持久化的用户消息启动新 turn，因此会同时造成“持续运行中”和“消息插入失败”。
+- 最小安全修复应只识别 app-server 的权威终态/turn 不匹配错误：先收敛仍指向同一旧 entry 的本地状态，再用已持久化的用户消息启动新 turn；不能再次经过 `handleMessage`，否则会重复写入用户历史。
+
+## 前端恢复态证据
+
+- 前端用 `isGenerating` 控制停止按钮和“插入”发送分支，用 `currentSessionRunning` 控制“运行中”徽标，两者不是同一状态源。
+- 正常 `done` 会调用 `finishGenerating` 同时清理两者；但断线/切后台后收到权威 `resume_session_result(isRunning=false)` 时，仅清理徽标并重新读取仍为 true 的 `isGenerating`，因此停止按钮和 steer 视觉状态可继续残留。
+- 前端最小修复是在当前会话的 idle 恢复结果中调用 `finishGenerating`；该函数已有空 streaming bubble 与历史计数保护，无需单独操作停止按钮。
+- 本次先不扩大到 resume requestId/epoch 的乱序重构；现有服务端会以同一 requestId 先发 `resume_generating` 再发最终 `resume_session_result`，直接在前一个事件清 pending 会让简单匹配方案误拒绝最终响应。
+
+## 修复范围决策
+
+- 后端：只对 `no active turn to steer` 这类明确“无活动 turn”的权威错误自动收敛并新开 turn；不把一般 expectedTurnId 不匹配、超时或网络错误视为安全降级条件。
+- 前端：用权威 idle 恢复结果收敛本地生成态，解决漏收 `done` 后的停止按钮残留。
+- 暂不扩展 `thread/status/changed`；其非 systemError 状态可能出现在 turn 开始前，笼统作为终态会误结束真实任务。
+
+## 测试入口
+
+- 项目完整回归入口为 `timeout 60s npm run regression`。
+- 现有 `scripts/regression.js --target` 只有 composer、未路由通知和子代理卡片三个定向目标；本次应新增独立 stale-turn 目标，避免每次红绿循环都启动完整回归。
+- 前端已有 `assertFrontendGenerationControlsContract` 与 `assertSessionSwitchResilienceContract` 源码合同，可在前者或本次新目标中断言 idle resume 必须调用完整生成态收敛函数。
+
+## 实现结果
+
+- mock 新增确定性场景：app-server 内部 turn 已结束但故意不发 `turn/completed`，下一次 steer 返回真实 `no active turn to steer`。
+- 后端新增窄匹配分类器，仅在旧 entry 身份仍一致时调用现有完成流程保存输出并清理状态，再直接以已持久化消息启动新 turn；一般 expectedTurnId 不匹配继续报错。
+- 前端 idle `resume_session_result` 改为调用 `finishGenerating`，同步清理 `isGenerating`、停止按钮、streaming 状态与运行徽标。
+- 定向回归覆盖消息/输出不重复、自动转新 turn、最终 idle、无失败 toast，以及 expectedTurnId 不匹配负例。
+
+## 独立审查修正
+
+- 首轮实现会在 follow-up 用户消息之后追加旧 turn 助手输出，刷新后历史顺序错误；现通过 timestamp+content 定位，把旧助手消息插入 follow-up 之前，并补严格索引顺序断言。
+- 完成旧 turn 时同步 flush 跨会话回复可能先创建新 entry；现仅在 stale replacement 期间延迟 flush，并在替代 turn 启动前再次确认 Map 为空。
+- `done` 会短暂清掉前端 `isGenerating`；现替代 turn 创建成功后、首个 delta 前发送空 `resume_generating`，恢复停止/插入控制状态。
+- 替代 turn 优先沿用旧 entry 的 `mcpContext`，避免线程级 MCP 上下文退化。
+- 独立质量审查最终通过，无阻塞问题。
+
+## 部署验证
+
+- 重启前通过 ccweb 会话列表确认只有当前对话处于 `running`，不存在会被中断的其他会话。
+- `pm2 restart ccweb --update-env` 导致调用连接中断，但 PM2 进程创建时间、restart 计数与启动日志确认重启已经执行。
+- 重启后 `ccweb` 状态为 `online`，unstable restart 为 0，事件循环 P95 约 1.61 ms。
+- 本地 `http://127.0.0.1:8002/` 返回 HTTP 200，探针耗时约 3.9 ms。
diff --git a/.planning/conversation-stale-running-fix/progress.md b/.planning/conversation-stale-running-fix/progress.md
new file mode 100644
index 0000000..b62666c
--- /dev/null
+++ b/.planning/conversation-stale-running-fix/progress.md
@@ -0,0 +1,20 @@
+# 进度日志
+
+- 2026-07-17：读取项目指令、Trellis 工作流及 `planning-with-files`、`todo-list-csv` 技能。
+- 2026-07-17：确认工作树干净，已有共享规划任务正在使用 active 指针，本任务采用独立 scoped 计划。
+- 2026-07-17：根据截图固化状态漂移与 `steer` 失败的初始因果假设，建立实施与验收计划。
+- 2026-07-17：独立计划审查通过；代码索引就绪，开始并行定位后端终态与前端发送判定链路。
+- 2026-07-17：初步定位 `handleCodexAppSteerMessage`、`isSessionRunning` 与现有 steer 回归入口；确认失败分支缺少明显的状态自愈。
+- 2026-07-17：完成核心状态机梳理；根因收敛为“终态通知缺失/乱序后 Map 陈旧，steer 权威拒绝未触发本地收敛与新 turn 降级”。
+- 2026-07-17：前后端并行排查完成；另确认前端 idle 恢复只清徽标、不清 `isGenerating`，会保留停止按钮与插入视觉状态。
+- 2026-07-17：定位阶段完成，进入失败回归编写；计划仅窄匹配明确无 active turn，保留其他 steer 错误原行为。
+- 2026-07-17：确认完整回归入口和现有定向 target 结构；实施代理将新增本问题专用 target 以保留红绿测试证据。
+- 2026-07-17：实施代理完成测试先行修复；旧实现定向回归按预期失败于 steer 错误和 done 超时，修复后连续三次通过。
+- 2026-07-17：主线程复跑 `timeout 60s node scripts/regression.js --target codexapp-stale-running` 通过；四个修改文件语法检查和 `git diff --check` 通过。
+- 2026-07-17：进入完整回归与独立边界审查阶段。
+- 2026-07-17：首轮独立审查发现历史顺序与跨会话 flush 重入问题，已补实现及红绿回归。
+- 2026-07-17：最终复审进一步指出 done→running 的短暂 UI 窗口；已发送 `resume_generating` 收敛并保留旧 MCP 上下文。
+- 2026-07-17：主线程再次运行定向回归、四文件语法检查、`git diff --check` 和 `timeout 60s npm run regression`，全部通过；独立审查最终放行。
+- 2026-07-17：进入最终差异、部署条件与临时跟踪文件清理。
+- 2026-07-17：重启前确认只有当前对话 running；执行 cc-web 重启后 PM2 online、HTTP 200、事件循环与进程状态正常。
+- 2026-07-17：最终行号、差异统计和 `git diff --check` 核验完成，任务闭环。
diff --git a/.planning/conversation-stale-running-fix/task_plan.md b/.planning/conversation-stale-running-fix/task_plan.md
new file mode 100644
index 0000000..045e125
--- /dev/null
+++ b/.planning/conversation-stale-running-fix/task_plan.md
@@ -0,0 +1,35 @@
+# 对话结束后仍显示运行中修复计划
+
+## 目标
+
+修复 Codex App 对话实际已结束但 cc-web 仍显示“运行中”的偶发状态漂移，确保后续消息不会误走 `steer` 并触发 `no active turn to steer`。
+
+## 阶段
+
+1. [complete] 定位运行状态结束与再次发送的完整调用链
+2. [complete] 补充卡住运行态与 steer 失败的回归用例
+3. [complete] 实现终态事件驱动的运行状态收敛
+4. [complete] 实现无活动 turn 时的安全发送降级
+5. [complete] 运行定向回归与语法静态检查
+6. [complete] 执行完整回归并独立审查边界
+7. [complete] 核对差异风险并清理临时跟踪文件
+
+## 验收标准
+
+- turn 正常完成、失败或中断后，会话最终稳定为非运行态。
+- 状态漂移时再次发送不会把用户消息永久失败在 `steer`。
+- 不引入重复 turn、重复消息或吞掉真实 steer 错误。
+- 回归测试覆盖终态缺失/乱序和 `no active turn to steer` 竞态。
+- 定向测试、语法检查与完整回归均通过，单次测试最长 60 秒。
+
+## 约束
+
+- 代码理解优先使用 `codebase-memory-mcp`，`rg` / `nl` 仅做校验。
+- 不覆盖现有 `.planning` 任务，不切换共享 Trellis 当前任务指针。
+- 未确认其他会话均非 `running` 前不重启 cc-web。
+- 代码注释与新增文档使用简体中文。
+
+## 错误记录
+
+| 错误 | 尝试 | 处理 |
+|---|---:|---|
diff --git a/dist-exe/cc-web-bun-linux-x64-baseline.tar.gz b/dist-exe/cc-web-bun-linux-x64-baseline.tar.gz
index f026b16..a44a52f 100644
Binary files a/dist-exe/cc-web-bun-linux-x64-baseline.tar.gz and b/dist-exe/cc-web-bun-linux-x64-baseline.tar.gz differ
diff --git a/public/app.js b/public/app.js
index 1622647..1d41176 100644
--- a/public/app.js
+++ b/public/app.js
@@ -5574,7 +5574,7 @@
         clearPendingSessionResumeRequest(msg.sessionId || currentSessionId, msg.requestId);
         setCurrentSessionRunningState(!!msg.isRunning);
         if (!msg.isRunning && currentSessionId && msg.sessionId === currentSessionId) {
-          updateGenerationControls();
+          finishGenerating(msg.sessionId || currentSessionId);
         }
         break;
 
diff --git a/scripts/mock-codex-app-server.js b/scripts/mock-codex-app-server.js
index 7dab4a2..ca11af4 100755
--- a/scripts/mock-codex-app-server.js
+++ b/scripts/mock-codex-app-server.js
@@ -364,6 +364,47 @@ function completeTurn(thread, turnId, text, status = 'completed') {
   thread.steers = [];
 }
 
+function completeTurnWithoutTerminalNotification(thread, turnId, text) {
+  if (thread.activeTurnId !== turnId) return;
+  const responseText = `Codex App stale turn output: ${text}`;
+  send({
+    method: 'item/agentMessage/delta',
+    params: {
+      threadId: thread.id,
+      turnId,
+      itemId: 'agent-msg',
+      delta: responseText,
+    },
+  });
+  send({
+    method: 'item/completed',
+    params: {
+      threadId: thread.id,
+      turnId,
+      completedAtMs: Date.now(),
+      item: {
+        id: 'agent-msg',
+        type: 'agentMessage',
+        content: [{ type: 'text', text: responseText }],
+        status: 'completed',
+      },
+    },
+  });
+  send({
+    method: 'thread/tokenUsage/updated',
+    params: {
+      threadId: thread.id,
+      turnId,
+      tokenUsage: tokenUsage(text),
+    },
+  });
+
+  // 模拟 app-server 已结束内部 turn，但终态通知在传输中丢失。
+  thread.activeTurnId = null;
+  thread.timer = null;
+  thread.steers = [];
+}
+
 function completeGoalBackgroundTurn(thread, objective) {
   const turnId = `goal-turn-${crypto.randomUUID()}`;
   const text = `Goal background output: ${objective}`;
@@ -685,6 +726,26 @@ function startTurn(params) {
     },
   });
 
+  if (/^codexapp stale running first$/i.test(text)) {
+    completeTurnWithoutTerminalNotification(thread, turnId, text);
+    return { turn: { id: turnId, status: 'running', items: [] } };
+  }
+
+  if (/^codexapp expected turn mismatch first$/i.test(text)) {
+    send({
+      method: 'item/agentMessage/delta',
+      params: {
+        threadId: thread.id,
+        turnId,
+        itemId: 'agent-msg',
+        delta: `Codex App expected turn mismatch fixture: ${text}`,
+      },
+    });
+    // 保留一个不同的活动 turn，确保 steer 返回“不匹配”而不是“无活动 turn”。
+    thread.activeTurnId = `app-turn-mismatch-${crypto.randomUUID()}`;
+    return { turn: { id: turnId, status: 'running', items: [] } };
+  }
+
   if (/runtime warning/i.test(text)) {
     const message = 'Heads up: Long threads and multiple compactions can cause the model to be less accurate. Start a new thread when possible to keep threads small and targeted.';
     for (let i = 0; i < 2; i += 1) {
@@ -819,7 +880,15 @@ function interruptTurn(params) {
 
 function steerTurn(params) {
   const thread = ensureThread(params.threadId, params);
-  if (!thread.activeTurnId || thread.activeTurnId !== params.expectedTurnId) {
+  if (!thread.activeTurnId) {
+    return {
+      error: {
+        code: -32001,
+        message: 'no active turn to steer',
+      },
+    };
+  }
+  if (thread.activeTurnId !== params.expectedTurnId) {
     return {
       error: {
         code: -32001,
diff --git a/scripts/regression.js b/scripts/regression.js
index bd24239..c210f2a 100644
--- a/scripts/regression.js
+++ b/scripts/regression.js
@@ -132,10 +132,11 @@ async function startServer(env) {
   };
 }
 
-function connectWs(port, password) {
+function connectWs(port, password, options = {}) {
   return new Promise((resolve, reject) => {
     const ws = new WebSocket(`ws://127.0.0.1:${port}/ws`);
     const messages = [];
+    const receivedMessages = options.trackReceived ? [] : null;
     let settled = false;
 
     ws.on('open', () => {
@@ -144,9 +145,10 @@ function connectWs(port, password) {
     ws.on('message', (buf) => {
       const msg = JSON.parse(String(buf));
       messages.push(msg);
+      if (receivedMessages) receivedMessages.push(msg);
       if (msg.type === 'auth_result' && msg.success) {
         settled = true;
-        resolve({ ws, messages, token: msg.token });
+        resolve({ ws, messages, receivedMessages, token: msg.token });
       }
       if (msg.type === 'auth_result' && !msg.success) {
         settled = true;
@@ -1423,6 +1425,10 @@ function assertSessionSwitchResilienceContract() {
   assert(frontendSource.includes('function requestSessionResume'), 'Frontend should request running-session resume without full history reload');
   assert(frontendSource.includes("type: 'resume_session'"), 'Frontend should use resume_session for reconnecting running conversations');
   assert(frontendSource.includes("case 'resume_session_result':"), 'Frontend should handle lightweight resume results');
+  assert(
+    /case 'resume_session_result':[\s\S]*?if \(!msg\.isRunning && currentSessionId && msg\.sessionId === currentSessionId\) \{[\s\S]*?finishGenerating\(msg\.sessionId \|\| currentSessionId\);[\s\S]*?\}[\s\S]*?break;/.test(frontendSource),
+    'Frontend idle resume result should finish generation state for the current session'
+  );
   assert(frontendSource.includes('recoverCurrent: true'), 'Frontend fallback load_session should preserve the current running view');
   const visibilityStart = frontendSource.indexOf("document.addEventListener('visibilitychange'");
   const visibilityEnd = visibilityStart >= 0 ? frontendSource.indexOf("if (!authToken)", visibilityStart) : -1;
@@ -1515,6 +1521,175 @@ function assertSessionSwitchResilienceContract() {
   );
 }
 
+function assertCodexAppStaleRunningRecoveryContract() {
+  const serverSource = fs.readFileSync(SERVER_PATH, 'utf8');
+  const completeBlock = extractFunctionSource(serverSource, 'handleCodexAppTurnComplete');
+  const steerBlock = extractFunctionSource(serverSource, 'handleCodexAppSteerMessage');
+  assert(
+    completeBlock.includes('!options.deferPendingCrossConversationFlush') && completeBlock.includes('flushPendingCrossConversationReplies(sessionId)'),
+    'Stale recovery should be able to defer pending cross-conversation reply flushes'
+  );
+  assert(
+    steerBlock.includes('deferPendingCrossConversationFlush: true'),
+    'Stale steer recovery should defer pending cross-conversation reply flushes while replacing the turn'
+  );
+  assert(
+    steerBlock.includes('!activeCodexAppTurns.has(sessionId)') && steerBlock.includes('handleCodexAppMessage(ws, refreshedSession'),
+    'Stale steer recovery should only start the replacement turn while the active turn map is empty'
+  );
+  assert(
+    steerBlock.includes('mcpContext: entry.mcpContext || options.mcpContext || {}'),
+    'Stale steer recovery should prefer the original active entry MCP context'
+  );
+}
+
+async function runCodexAppStaleRunningRegression() {
+  const tempRoot = fs.mkdtempSync(path.join(os.tmpdir(), 'cc-web-stale-running-regression-'));
+  const configDir = path.join(tempRoot, 'config');
+  const sessionsDir = path.join(tempRoot, 'sessions');
+  const logsDir = path.join(tempRoot, 'logs');
+  const homeDir = path.join(tempRoot, 'home');
+  mkdirp(configDir);
+  mkdirp(sessionsDir);
+  mkdirp(logsDir);
+  mkdirp(homeDir);
+
+  const port = await getFreePort();
+  const password = 'StaleRunning!234';
+  await withServer({
+    PORT: String(port),
+    CC_WEB_PASSWORD: password,
+    CC_WEB_INTERNAL_MCP_TOKEN: 'StaleRunningMcp!234',
+    CC_WEB_CONFIG_DIR: configDir,
+    CC_WEB_SESSIONS_DIR: sessionsDir,
+    CC_WEB_LOGS_DIR: logsDir,
+    HOME: homeDir,
+    CLAUDE_PATH: MOCK_CLAUDE,
+    CODEX_PATH: MOCK_CODEX_APP_SERVER,
+  }, async () => {
+    const { ws, messages, receivedMessages } = await connectWs(port, password, { trackReceived: true });
+    await nextMessage(messages, ws, (msg) => msg.type === 'session_list');
+
+    ws.send(JSON.stringify({ type: 'new_session', agent: 'codexapp', cwd: homeDir, mode: 'yolo' }));
+    const sessionInfo = await nextMessage(messages, ws, (msg) => (
+      msg.type === 'session_info' && msg.agent === 'codexapp' && msg.cwd === homeDir
+    ));
+    const sessionId = sessionInfo.sessionId;
+
+    ws.send(JSON.stringify({
+      type: 'message',
+      text: 'codexapp stale running first',
+      sessionId,
+      mode: 'yolo',
+      agent: 'codexapp',
+    }));
+    await nextMessage(messages, ws, (msg) => (
+      msg.type === 'session_list' && msg.sessions.some((session) => session.id === sessionId && session.isRunning)
+    ));
+    const staleOutput = await nextMessage(messages, ws, (msg) => (
+      msg.type === 'text_delta' && msg.sessionId === sessionId && /Codex App stale turn output/.test(msg.text || '')
+    ));
+    assert(/codexapp stale running first/.test(staleOutput.text || ''), 'Stale running fixture should emit the first turn output');
+    await sleep(150);
+
+    ws.send(JSON.stringify({
+      type: 'message',
+      text: 'codexapp stale running follow-up',
+      sessionId,
+      mode: 'yolo',
+      agent: 'codexapp',
+      clientMessageId: 'regression-stale-running-follow-up',
+    }));
+    await nextMessage(messages, ws, (msg) => (
+      msg.type === 'codex_app_steer_status' &&
+      msg.sessionId === sessionId &&
+      msg.clientMessageId === 'regression-stale-running-follow-up' &&
+      msg.status === 'pending'
+    ), 5000);
+    await nextMessage(messages, ws, (msg) => msg.type === 'done' && msg.sessionId === sessionId, 5000);
+    await nextMessage(messages, ws, (msg) => (
+      msg.type === 'session_list' && msg.sessions.some((session) => session.id === sessionId && !session.isRunning)
+    ), 5000);
+    await nextMessage(messages, ws, (msg) => (
+      msg.type === 'session_list' && msg.sessions.some((session) => session.id === sessionId && session.isRunning)
+    ), 5000);
+    const replacementResume = await nextMessage(messages, ws, (msg) => (
+      msg.type === 'resume_generating' && msg.sessionId === sessionId
+    ), 5000);
+    assert(replacementResume.text === '' && Array.isArray(replacementResume.toolCalls) && replacementResume.toolCalls.length === 0, 'Stale replacement should resume generation with an empty streaming payload');
+    const recoveredStatus = await nextMessage(messages, ws, (msg) => (
+      msg.type === 'codex_app_steer_status' &&
+      msg.sessionId === sessionId &&
+      msg.clientMessageId === 'regression-stale-running-follow-up' &&
+      msg.status === 'inserted'
+    ), 5000);
+    assert(!/失败/.test(recoveredStatus.message || ''), 'Recovered stale steer should update the UI as a non-failure');
+    const recoveredHint = await nextMessage(messages, ws, (msg) => (
+      msg.type === 'system_message' && msg.sessionId === sessionId && /已自动开始新一轮对话/.test(msg.message || '')
+    ), 5000);
+    assert(recoveredHint.transient === true, 'Recovered stale steer hint should be transient');
+    await nextMessage(messages, ws, (msg) => (
+      msg.type === 'text_delta' && msg.sessionId === sessionId && /codexapp stale running follow-up/.test(msg.text || '')
+    ), 5000);
+    const replacementResumeIndex = receivedMessages.findIndex((msg) => msg.type === 'resume_generating' && msg.sessionId === sessionId);
+    const replacementDeltaIndex = receivedMessages.findIndex((msg) => msg.type === 'text_delta' && msg.sessionId === sessionId && /codexapp stale running follow-up/.test(msg.text || ''));
+    assert(replacementResumeIndex >= 0 && replacementResumeIndex < replacementDeltaIndex, 'Stale replacement should restore generating UI before its first text delta');
+    await nextMessage(messages, ws, (msg) => msg.type === 'done' && msg.sessionId === sessionId, 5000);
+    const finalList = await nextMessage(messages, ws, (msg) => (
+      msg.type === 'session_list' && msg.sessions.some((session) => session.id === sessionId && !session.isRunning)
+    ), 5000);
+    assert(finalList.sessions.some((session) => session.id === sessionId && !session.isRunning), 'Recovered follow-up should finish idle');
+    assert(!messages.some((msg) => msg.code === 'codexapp_steer_failed'), 'Recovered stale steer should not emit codexapp_steer_failed');
+
+    const stored = JSON.parse(fs.readFileSync(path.join(sessionsDir, `${sessionId}.json`), 'utf8'));
+    assert(stored.messages.filter((message) => message.role === 'user' && message.content === 'codexapp stale running follow-up').length === 1, 'Recovered follow-up user message should persist exactly once');
+    assert(stored.messages.filter((message) => message.role === 'assistant' && /Codex App stale turn output/.test(String(message.content || ''))).length === 1, 'Stale first turn output should persist exactly once');
+    assert(stored.messages.filter((message) => message.role === 'assistant' && /Codex App mock handled: codexapp stale running follow-up/.test(String(message.content || ''))).length === 1, 'Recovered follow-up output should persist exactly once');
+    const staleAssistantIndex = stored.messages.findIndex((message) => message.role === 'assistant' && /Codex App stale turn output/.test(String(message.content || '')));
+    const followUpUserIndex = stored.messages.findIndex((message) => message.role === 'user' && message.content === 'codexapp stale running follow-up');
+    const recoveredAssistantIndex = stored.messages.findIndex((message) => message.role === 'assistant' && /Codex App mock handled: codexapp stale running follow-up/.test(String(message.content || '')));
+    assert(staleAssistantIndex < followUpUserIndex && followUpUserIndex < recoveredAssistantIndex, 'Recovered history should keep stale assistant before follow-up user before recovered assistant');
+    assert(!fs.existsSync(path.join(sessionsDir, `${sessionId}-run`)), 'Recovered follow-up should clean the Codex App run directory after completion');
+
+    ws.send(JSON.stringify({ type: 'new_session', agent: 'codexapp', cwd: homeDir, mode: 'yolo' }));
+    const mismatchSession = await nextMessage(messages, ws, (msg) => (
+      msg.type === 'session_info' && msg.agent === 'codexapp' && msg.cwd === homeDir && msg.sessionId !== sessionId
+    ));
+    ws.send(JSON.stringify({
+      type: 'message',
+      text: 'codexapp expected turn mismatch first',
+      sessionId: mismatchSession.sessionId,
+      mode: 'yolo',
+      agent: 'codexapp',
+    }));
+    await nextMessage(messages, ws, (msg) => (
+      msg.type === 'text_delta' && msg.sessionId === mismatchSession.sessionId && /expected turn mismatch fixture/.test(msg.text || '')
+    ), 5000);
+    await sleep(150);
+    ws.send(JSON.stringify({
+      type: 'message',
+      text: 'codexapp mismatch follow-up',
+      sessionId: mismatchSession.sessionId,
+      mode: 'yolo',
+      agent: 'codexapp',
+      clientMessageId: 'regression-expected-turn-mismatch',
+    }));
+    const mismatchFailed = await nextMessage(messages, ws, (msg) => (
+      msg.type === 'codex_app_steer_status' &&
+      msg.sessionId === mismatchSession.sessionId &&
+      msg.clientMessageId === 'regression-expected-turn-mismatch' &&
+      msg.status === 'failed'
+    ), 5000);
+    assert(/失败/.test(mismatchFailed.message || ''), 'Expected turn mismatch should keep the original failed steer status');
+    const mismatchError = await nextMessage(messages, ws, (msg) => (
+      msg.type === 'error' && msg.sessionId === mismatchSession.sessionId && msg.code === 'codexapp_steer_failed'
+    ), 5000);
+    assert(/expectedTurnId does not match active turn/.test(mismatchError.message || ''), 'Expected turn mismatch should not trigger stale-turn recovery');
+    assert(!messages.some((msg) => msg.type === 'system_message' && msg.sessionId === mismatchSession.sessionId && /已自动开始新一轮对话/.test(msg.message || '')), 'Expected turn mismatch should not start a replacement turn');
+    ws.close();
+  });
+}
+
 function assertUnlimitedImageAttachmentsContract() {
   const frontendSource = fs.readFileSync(PUBLIC_APP_PATH, 'utf8');
   const serverSource = fs.readFileSync(SERVER_PATH, 'utf8');
@@ -1722,6 +1897,13 @@ async function main() {
       console.log('Subagent card metadata regression checks passed.');
       return;
     }
+    if (regressionTarget === 'codexapp-stale-running') {
+      await runCodexAppStaleRunningRegression();
+      assertSessionSwitchResilienceContract();
+      assertCodexAppStaleRunningRecoveryContract();
+      console.log('Codex App stale running regression checks passed.');
+      return;
+    }
     throw new Error(`Unknown regression target: ${regressionTarget}`);
   }
 
diff --git a/server.js b/server.js
index 1bc5da1..66028fa 100644
--- a/server.js
+++ b/server.js
@@ -10075,7 +10075,7 @@ function handleCodexAppTurnComplete(sessionId, options = {}) {
   }
 
   if (session && (assistantContent.trim() || assistantToolCalls.length > 0) && !hasCodexAppTurnMessage(session, turnKey)) {
-    session.messages.push({
+    const assistantMessage = {
       role: 'assistant',
       content: assistantContent,
       toolCalls: assistantToolCalls,
@@ -10084,7 +10084,21 @@ function handleCodexAppTurnComplete(sessionId, options = {}) {
       codexAppThreadId: entry.threadId || null,
       codexAppTurnId: entry.turnId || null,
       interrupted: !!options.interrupted,
-    });
+    };
+    const beforeUserMessage = options.beforeUserMessage;
+    const beforeUserMessageIndex = beforeUserMessage
+      ? session.messages.findIndex((message) => (
+          message?.role === 'user'
+          && message.timestamp === beforeUserMessage.timestamp
+          && message.content === beforeUserMessage.content
+        ))
+      : -1;
+    if (beforeUserMessageIndex >= 0) {
+      // stale steer 已先持久化用户消息，把旧 turn 输出恢复到该消息之前。
+      session.messages.splice(beforeUserMessageIndex, 0, assistantMessage);
+    } else {
+      session.messages.push(assistantMessage);
+    }
     session.updated = new Date().toISOString();
     if (!entry.ws) session.hasUnread = true;
     saveSession(session);
@@ -10095,7 +10109,9 @@ function handleCodexAppTurnComplete(sessionId, options = {}) {
   if (entry.crossConversationReplyRequestId) {
     completeCrossConversationReply(entry.crossConversationReplyRequestId, entry, session);
   }
-  flushPendingCrossConversationReplies(sessionId);
+  if (!options.deferPendingCrossConversationFlush) {
+    flushPendingCrossConversationReplies(sessionId);
+  }
   plog(completionError ? 'WARN' : 'INFO', 'codex_app_turn_complete', {
     sessionId: sessionId.slice(0, 8),
     threadId: entry.threadId || null,
@@ -10140,6 +10156,10 @@ function handleCodexAppTurnFailure(sessionId, err) {
   handleCodexAppTurnComplete(sessionId, { error: entry.lastError });
 }
 
+function isCodexAppNoActiveTurnError(err) {
+  return /\bno active turn to steer\b/i.test(String(err?.message || err || ''));
+}
+
 function handleCodexAppSteerMessage(ws, msg, options = {}) {
   const sessionId = sanitizeId(msg?.sessionId || '');
   const entry = activeCodexAppTurns.get(sessionId);
@@ -10231,6 +10251,42 @@ function handleCodexAppSteerMessage(ws, msg, options = {}) {
       message: `已引导对话: ${previewInlineText(textValue)}`,
     });
   }).catch((err) => {
+    if (isCodexAppNoActiveTurnError(err) && activeCodexAppTurns.get(sessionId) === entry) {
+      // app-server 已确认旧 turn 不存在：先收敛旧输出，再复用已持久化的消息直接启动新 turn。
+      handleCodexAppTurnComplete(sessionId, {
+        beforeUserMessage: {
+          timestamp: persistedUserMessage.timestamp,
+          content: persistedUserMessage.content,
+        },
+        deferPendingCrossConversationFlush: true,
+      });
+      const refreshedSession = loadSession(sessionId);
+      const restarted = refreshedSession
+        && isCodexAppSession(refreshedSession)
+        && !activeCodexAppTurns.has(sessionId)
+        ? handleCodexAppMessage(ws, refreshedSession, runtimeTextValue, [], {
+            mcpContext: entry.mcpContext || options.mcpContext || {},
+          })
+        : null;
+      if (restarted?.ok) {
+        wsSend(entry.ws || ws, {
+          type: 'resume_generating',
+          sessionId,
+          text: '',
+          toolCalls: [],
+        });
+        sendSteerStatus('inserted', '已转为新对话');
+        wsSend(entry.ws || ws, {
+          type: 'system_message',
+          sessionId,
+          tone: 'info',
+          transient: true,
+          autoDismissMs: 5000,
+          message: '检测到上一轮已结束，已自动开始新一轮对话。',
+        });
+        return;
+      }
+    }
     sendSteerStatus('failed', '插入失败');
     wsSend(entry.ws || ws, {
       type: 'error',