botmux 3.4.2 → 3.5.0

package/dist/core/worker-pool.js

@@ -11,7 +11,7 @@ import { fileURLToPath } from 'node:url';
 import { ensureSkills, ensureAskSkill, ensurePluginSkills, ensureWhiteboardSkill, removeGlobalBotmuxSkills } from '../skills/installer.js';
 import { shouldInstallGlobalSkills } from '../skills/injection-mode.js';
 import { whiteboardEnabled } from '../services/whiteboard-store.js';
-import { installHook } from '../adapters/hook-installer.js';
+import { cleanupTraexAskHooks, installHook } from '../adapters/hook-installer.js';
 import { hookCommandFor } from '../adapters/hook-command.js';
 import { randomBytes, randomUUID } from 'node:crypto';
 import { config } from '../config.js';
@@ -24,6 +24,7 @@ import { loadFrozenCards, saveFrozenCards } from '../services/frozen-card-store.
 import { hashUrlForLog, cancelRiffTaskById } from '../adapters/backend/riff-backend.js';
 import { logger } from '../utils/logger.js';
 import { createCliAdapterSync } from '../adapters/cli/registry.js';
+import { traeHome } from '../services/traex-paths.js';
 import { botLocale, localeForBot, t as tr } from '../i18n/index.js';
 import { claudeJsonlPathForSession } from '../adapters/cli/claude-code.js';
 import { findUniqueClaudeSessionByCwd } from './session-discovery.js';
@@ -34,7 +35,7 @@ import { listDocSubscriptionsForSession, removeDocSubscription } from '../servic
 import { TmuxBackend } from '../adapters/backend/tmux-backend.js';
 import { HerdrBackend } from '../adapters/backend/herdr-backend.js';
 import { sandboxEnabled } from '../adapters/backend/sandbox.js';
-import { isSuspendableBackendType, getSessionPersistentBackendType, persistentSessionName, killPersistentSession, resolvePairedSpawnBackendType } from './persistent-backend.js';
+import { isSuspendableBackendType, getSessionPersistentBackendType, persistentBackendTargetForSession, killPersistentBackendTarget, managedTargetsForCliChange, resolvePairedSpawnBackendType, resolvePersistentBackendTarget } from './persistent-backend.js';
 import { getBot, getAllBots, resolveBrandLabel } from '../bot-registry.js';
 /** A random id minted once per daemon process (this lifetime). Stamped onto
  *  isolated persistent panes so a suspend→resume reattach (same id) is
@@ -79,7 +80,9 @@ import { isLocalCliOpenEnabled, isLocalCliOpenReady } from '../services/local-cl
 import { isSilentScheduledTurn } from './silent-schedule-turns.js';
 import { writeDeferredTopicBinding } from './deferred-topic-binding.js';
 import { deferWorkerSpawnDuringDeviceIsolation } from './device-isolation-activation.js';
+import { buildBotmuxLarkNativeSessionTitle, extractBotmuxLarkNativeSessionTitlePrompt, } from './session-title.js';
 import { acknowledgeSessionReady } from './session-ready-handshake.js';
+import { recordDispatchInputCommit } from './dispatch.js';
 const __filename = fileURLToPath(import.meta.url);
 const __dirname = dirname(__filename);
 const WORKER_SIGTERM_BACKSTOP_MS = 2_000;
@@ -652,13 +655,13 @@ export async function postPrivateSnapshotCard(ds, audience) {
  * Deliver the write-enabled session card (the "🔑 获取操作链接" card, which carries
  * a write-token terminal URL + manage buttons) privately to a single operator.
  *
- * Prefers an in-chat "visible-to-you" ephemeral card so the operator never has
- * to leave the conversation. Feishu's ephemeral API only works in plain `group`
- * chats — topic / thread groups reject with {@link LARK_CODE_EPHEMERAL_NOT_GROUP}
- * (18053) and p2p chats are unsupported — and chatType can't distinguish a topic
- * group from a regular one (both record 'group'), so we attempt ephemeral for any
- * non-p2p chat and fall back to a private DM on ANY failure. p2p chats skip the
- * doomed ephemeral attempt and DM directly (the DM lands in that same 1:1 chat).
+ * Prefers an in-chat "visible-to-you" ephemeral card in a flat group so the
+ * operator never has to leave the conversation. Thread-scope sessions and
+ * chat-scope sessions currently folded into a thread go straight to DM:
+ * Feishu may accept their ephemeral message without rendering it in the topic
+ * panel. p2p chats also DM directly (the DM lands in that same 1:1 chat).
+ *
+ * Other group chats attempt ephemeral first and fall back to DM on ANY failure.
  *
  * Both channels are private, so the DM fallback never leaks the write token —
  * unlike the private /card snapshot (which fails closed), here we fail OVER.
@@ -667,7 +670,10 @@ export async function postPrivateSnapshotCard(ds, audience) {
  */
 export async function deliverWriteLinkCard(ds, operatorOpenId, cardJson) {
     const who = operatorOpenId.substring(0, 8);
-    if (ds.chatType !== 'p2p') {
+    const replyTarget = ds.currentReplyTarget ?? ds.session.currentReplyTarget;
+    const isThreaded = ds.scope === 'thread'
+        || (!!replyTarget?.rootMessageId && replyTarget.quoteOnly !== true);
+    if (ds.chatType !== 'p2p' && !isThreaded) {
         try {
             await sendEphemeralCard(ds.larkAppId, ds.chatId, operatorOpenId, cardJson);
             logger.info(`[${tag(ds)}] write link delivered via ephemeral card to ${who}…`);
@@ -909,6 +915,12 @@ const skillsInstalledCliIds = new Set();
  */
 export function ensureCliSkills(cliId, cliPathOverride) {
     const adapter = createCliAdapterSync(cliId, cliPathOverride);
+    if (cliId === 'traex') {
+        cleanupTraexAskHooks([
+            join(traeHome(), 'hooks.json'),
+            join(traeHome(), 'cli', 'hooks.json'),
+        ]);
+    }
     // Claude-family CLIs deliver skills per-session via `--plugin-dir` (no global
     // leak), so they always materialise their plugin dir — the builtin-injection
     // mode does not apply to them. Everything below the branch is the global-dir
@@ -1084,6 +1096,10 @@ export function killWorker(ds) {
     // daemon-side copy synchronously; the worker may never get a chance to send
     // its ordered revoke IPC on close/crash paths.
     ds.managedTurnOrigin = undefined;
+    // The worker/CLI generation is ending — any outstanding stuck-warning card
+    // must be invalidated so a late click cannot inject keys into a replacement
+    // worker (or into nothing, if no replacement comes).
+    invalidateStuckWarning(ds, 'killWorker');
     if (!ds.worker || ds.worker.killed) {
         // No live worker to receive {type:'close'}, so its destroySession() — which
         // tears down the persistent backing session (tmux/herdr/zellij) — never
@@ -1149,7 +1165,7 @@ function destroyOrphanedBackingSession(ds) {
     if (!backendType)
         return;
     try {
-        killPersistentSession(backendType, persistentSessionName(backendType, ds.session.sessionId));
+        killPersistentBackendTarget(persistentBackendTargetForSession(ds));
         logger.info(`[${tag(ds)}] killWorker: no live worker — destroyed orphaned ${backendType} backing session`);
     }
     catch (err) {
@@ -1201,6 +1217,10 @@ export function suspendWorker(ds, reason = 'suspended_idle') {
     ds.workerToken = null;
     ds.workerViewToken = null;
     ds.managedTurnOrigin = undefined;
+    // The worker is being suspended — the CLI will be destroyed and cold-resumed
+    // later. Invalidate any stuck-warning card so a late click cannot inject keys
+    // after the CLI is gone (or into a different CLI on resume).
+    invalidateStuckWarning(ds, 'suspendWorker');
     // Screen state describes the process we just stopped. Keeping it would make
     // the dashboard hydrate this process-less logical session as idle/working.
     ds.lastScreenStatus = undefined;
@@ -1617,11 +1637,32 @@ export function sendWorkerInput(ds, payload, turnId, opts = {}) {
         return false;
     const normalized = typeof payload === 'string' ? { content: payload } : payload;
     const codexAppInput = codexAppInputForSession(ds, normalized.codexAppInput, turnId);
+    let nativeSessionTitlePrompt;
+    let nativeSessionTitle;
+    if (ds.session.nativeSessionTitleAwaitingContent && !ds.session.nativeSessionTitleUserDefined && !ds.adoptedFrom) {
+        const bot = getBot(ds.larkAppId);
+        const effectiveCliId = ds.session.cliId ?? bot.config.cliId;
+        if (effectiveCliId === 'codex') {
+            nativeSessionTitlePrompt = extractBotmuxLarkNativeSessionTitlePrompt(normalized.codexAppInput?.text ?? normalized.content, bot.botName ? [{ name: bot.botName }] : undefined);
+            if (nativeSessionTitlePrompt) {
+                nativeSessionTitle = buildBotmuxLarkNativeSessionTitle(nativeSessionTitlePrompt);
+                ds.session.nativeSessionTitle = nativeSessionTitle;
+                ds.session.nativeSessionTitleAwaitingContent = undefined;
+                if (ds.initConfig) {
+                    ds.initConfig.nativeSessionTitle = nativeSessionTitle;
+                    ds.initConfig.nativeSessionTitlePrompt = nativeSessionTitlePrompt;
+                }
+                sessionStore.updateSession(ds.session);
+            }
+        }
+    }
     const vcMeetingImTurnOrigin = resolveVcMeetingImTurnOrigin(ds.session, turnId);
     ds.worker.send({
         type: 'message',
         content: normalized.content,
         ...(codexAppInput ? { codexAppInput } : {}),
+        ...(nativeSessionTitle ? { nativeSessionTitle } : {}),
+        ...(nativeSessionTitlePrompt ? { nativeSessionTitlePrompt } : {}),
         ...(turnId ? { turnId } : {}),
         ...(opts.dispatchAttempt !== undefined ? { dispatchAttempt: opts.dispatchAttempt } : {}),
         ...(vcMeetingImTurnOrigin
@@ -1737,6 +1778,10 @@ export function forkWorker(ds, promptInput, resumeOrTurnId = false) {
         }
         sessionStore.updateSession(ds.session);
     }
+    // Reserve and durably publish the replacement lifetime before killing an
+    // existing worker. A failed reservation leaves the old worker untouched;
+    // a successful reservation immediately invalidates any late old-worker ACK.
+    const workerGeneration = reserveWorkerGeneration(ds);
     // Guard against double-fork: if a worker is already running, kill it first
     if (ds.worker && !ds.worker.killed) {
         logger.warn(`[${t}] Worker already running (pid: ${ds.worker.pid}), killing before re-fork`);
@@ -1769,6 +1814,38 @@ export function forkWorker(ds, promptInput, resumeOrTurnId = false) {
         reclaimParkedCrashDiagnostic(ds);
     const agentCfg = sessionAgentConfig(ds, botCfg);
     ensureCliEnv(agentCfg.cliId, agentCfg.cliPathOverride);
+    let nativeSessionTitle;
+    let nativeSessionTitlePrompt;
+    if (agentCfg.cliId === 'codex' && !ds.adoptedFrom) {
+        const isFreshNativeSession = !resume && !ds.session.cliSessionId;
+        const titlePrompt = extractBotmuxLarkNativeSessionTitlePrompt(promptPayload.codexAppInput?.text ?? prompt, bot.botName ? [{ name: bot.botName }] : undefined);
+        if (isFreshNativeSession && !ds.session.nativeSessionTitleUserDefined) {
+            ds.session.nativeSessionTitle = buildBotmuxLarkNativeSessionTitle(titlePrompt ? ds.session.title : undefined, bot.botName ? [{ name: bot.botName }] : undefined, ds.chatType === 'group' ? ds.session.chatDisplayName : undefined);
+            ds.session.nativeSessionTitleAwaitingContent = titlePrompt ? undefined : true;
+            nativeSessionTitlePrompt = titlePrompt;
+            sessionStore.updateSession(ds.session);
+        }
+        else if (ds.session.nativeSessionTitleAwaitingContent
+            && !ds.session.nativeSessionTitleUserDefined
+            && titlePrompt) {
+            ds.session.nativeSessionTitle = buildBotmuxLarkNativeSessionTitle(titlePrompt);
+            ds.session.nativeSessionTitleAwaitingContent = undefined;
+            nativeSessionTitlePrompt = titlePrompt;
+            sessionStore.updateSession(ds.session);
+        }
+        else if (isFreshNativeSession && !ds.session.nativeSessionTitle) {
+            ds.session.nativeSessionTitle = ds.session.title;
+            sessionStore.updateSession(ds.session);
+        }
+        if (isFreshNativeSession
+            || (resume && !!ds.session.cliSessionId && !!ds.session.nativeSessionTitle)
+            || (resume
+                && !!ds.session.nativeSessionTitleAwaitingContent
+                && !!ds.session.nativeSessionTitle)
+            || (!!nativeSessionTitlePrompt && !!ds.session.nativeSessionTitle)) {
+            nativeSessionTitle = ds.session.nativeSessionTitle;
+        }
+    }
     // Claude Code blocks on the interactive folder-trust dialog the first time
     // it runs in an untrusted workingDir; pre-accept it so the spawn doesn't hang.
     // Seed CLI (Claude Code fork) has the same dialog — drive both off the
@@ -1893,10 +1970,15 @@ export function forkWorker(ds, promptInput, resumeOrTurnId = false) {
         // getSessionPersistentBackendType). A brand-new session (no stamp) resolves
         // from live config, so a dashboard backend switch only affects NEW sessions.
         backendType: resolvePairedSpawnBackendType(agentCfg.cliId, ds.session.backendType, botCfg.backendType, config.daemon.backendType),
+        // Shared Herdr is not derivable from sessionId: preserve the exact host +
+        // managed-agent affinity across daemon/worker replacement.
+        persistentBackendTarget: ds.session.persistentBackendTarget,
         backendConfig: botCfg.riff,
         riffParentTaskId: ds.session.riffParentTaskId,
         riffRepoDirs: ds.session.riffRepoDirs,
         deferredScheduleRun: ds.session.deferredScheduleRun,
+        ...(nativeSessionTitle ? { nativeSessionTitle } : {}),
+        ...(nativeSessionTitlePrompt ? { nativeSessionTitlePrompt } : {}),
         prompt,
         ...(promptCodexAppInput ? { promptCodexAppInput } : {}),
         resume,
@@ -1937,7 +2019,7 @@ export function forkWorker(ds, promptInput, resumeOrTurnId = false) {
         sessionStore.updateSession(ds.session);
     }
     // Use shared handler for IPC messages and exit
-    setupWorkerHandlers(ds, worker, startupState);
+    setupWorkerHandlers(ds, worker, startupState, workerGeneration);
     ds.worker = worker;
     ds.spawnedAt = Date.now();
     ds.cliVersion = currentCliVersion;
@@ -1967,11 +2049,57 @@ export function forkWorker(ds, promptInput, resumeOrTurnId = false) {
     recordOwnershipForDaemonSession(ds);
 }
 // ─── Shared worker IPC handler ──────────────────────────────────────────────
-function setupWorkerHandlers(ds, worker, startupState = { ready: false, failureNotified: false }) {
+/**
+ * Clear the stuck-warning authority markers WITHOUT patching the card. Used by
+ * ACK paths (tui_keys_delivered / stuck_warning_expired) where the caller has
+ * already resolved the card to its final state — we only need to drop the
+ * nonce/cardId/turnId so a late duplicate click cannot re-inject keys.
+ */
+export function clearStuckWarningAuthority(ds) {
+    ds.stuckWarningCardId = undefined;
+    ds.stuckWarningTurnId = undefined;
+    ds.stuckWarningNonce = undefined;
+    ds.stuckWarningPageType = undefined;
+    ds.stuckWarningProcessing = false;
+    ds.stuckWarningCliLifetime = undefined;
+}
+/**
+ * Resolve any active stuck-warning card (PATCH it to "done") AND clear its
+ * markers. Called from every path where the CLI can recover or be replaced:
+ * prompt_ready, claude_exit, worker exit/kill/suspend/refork, and the worker's
+ * own stale-card notification. Centralising here avoids missing an exit path
+ * and leaving a clickable card that can inject keys into a different CLI.
+ *
+ * ACK paths (tui_keys_delivered / stuck_warning_expired) do NOT use this —
+ * they patch the card themselves with a context-specific message and then call
+ * clearStuckWarningAuthority() to drop the markers.
+ */
+export function invalidateStuckWarning(ds, reason) {
+    if (!ds.stuckWarningCardId && !ds.stuckWarningTurnId && ds.stuckWarningNonce === undefined)
+        return;
+    const t = tag(ds);
+    if (ds.stuckWarningCardId) {
+        const locDs = localeForBot(ds.larkAppId);
+        const resolvedCard = buildTuiPromptResolvedCard(tr('card.action.tui_done', undefined, locDs), locDs);
+        updateMessage(ds.larkAppId, ds.stuckWarningCardId, resolvedCard).catch(err => logger.debug(`[${t}] Failed to resolve stuck-warning card (${reason}): ${err}`));
+    }
+    logger.debug(`[${t}] invalidateStuckWarning (${reason}): turn=${ds.stuckWarningTurnId ?? 'none'} nonce=${ds.stuckWarningNonce ?? 'none'}`);
+    clearStuckWarningAuthority(ds);
+}
+function setupWorkerHandlers(ds, worker, startupState = { ready: false, failureNotified: false }, reservedWorkerGeneration) {
     const cb = requireCallbacks();
     const t = tag(ds);
-    const workerGeneration = (ds.workerGeneration ?? 0) + 1;
-    ds.workerGeneration = workerGeneration;
+    const workerGeneration = reservedWorkerGeneration
+        ?? reserveWorkerGeneration(ds);
+    if (ds.workerGeneration !== workerGeneration
+        || ds.session.workerGeneration !== workerGeneration) {
+        throw new Error('worker generation reservation changed before IPC setup');
+    }
+    // A new worker generation is starting. As a backstop, invalidate any
+    // stuck-warning card posted by the previous generation — explicit kill/suspend/
+    // exit paths should already have done this, but fork/refork/takeover paths
+    // can leave a stale card that would otherwise inject keys into the new CLI.
+    invalidateStuckWarning(ds, 'new_worker_generation');
     // Managed turn authority is issued by one concrete worker lifetime. A
     // replacement must advertise a fresh capability before daemon-mediated
     // exits may use it; carrying the old value across a restore/refork would
@@ -2063,6 +2191,29 @@ function setupWorkerHandlers(ds, worker, startupState = { ready: false, failureN
     worker.on('message', async (msg) => {
         const effectiveCliId = sessionCliId(ds, botCfg);
         switch (msg.type) {
+            case 'persistent_backend_target': {
+                ds.session.persistentBackendTarget = msg.target;
+                sessionStore.updateSession(ds.session);
+                break;
+            }
+            case 'turn_input_committed': {
+                // Bind the receipt to the exact live worker generation. A late ACK
+                // from a replaced worker cannot make the replacement appear to have
+                // accepted this dispatch turn.
+                if (ds.worker !== worker
+                    || ds.workerGeneration !== workerGeneration
+                    || ds.session.workerGeneration !== workerGeneration) {
+                    logger.warn(`[${t}] Ignored turn_input_committed from stale worker generation`);
+                    break;
+                }
+                if (recordDispatchInputCommit(ds.session, msg.turnId, workerGeneration)) {
+                    sessionStore.updateSession(ds.session);
+                }
+                else {
+                    logger.warn(`[${t}] Ignored unbound input commit turn=${msg.turnId.slice(0, 16)}`);
+                }
+                break;
+            }
             case 'session_ready_ack': {
                 if (ds.worker !== worker) {
                     logger.warn(`[${t}] Ignored session_ready_ack from stale worker generation`);
@@ -2096,6 +2247,10 @@ function setupWorkerHandlers(ds, worker, startupState = { ready: false, failureN
                 ds.workerViewToken = msg.viewToken ?? null;
                 // Persist port so it can be reused after daemon restart
                 ds.session.webPort = msg.port;
+                // Dashboard「复现命令」：worker 上报本次冷启的近似复现命令。只存内存字段、
+                // 绝不落盘（含凭证）；warm reattach 时 worker 不重算，故为空——这是有意的
+                // （reattach 不代表本次新算命令真的执行了）。worker 每次 ready 会重报。
+                ds.spawnCommand = msg.spawnCommand;
                 sessionStore.updateSession(ds.session);
                 const readOnlyUrl = buildTerminalUrl(ds);
                 const writeUrl = buildTerminalUrl(ds, { write: true });
@@ -2123,6 +2278,18 @@ function setupWorkerHandlers(ds, worker, startupState = { ready: false, failureN
                     logger.info(`[${t}] Managed/silent turn — suppressing ready/streaming card output`);
                     break;
                 }
+                // Pi can finish a very fast startup turn through `botmux send` before
+                // Herdr reports idle and the worker emits ready. Posting the initial
+                // card now would put a stale "Starting" card *after* the final reply,
+                // with no later transcript output to withdraw it (the explicit send
+                // intentionally suppresses bridge fallback). Keep the terminal/runtime
+                // state above, but suppress this already-completed turn's card.
+                if (msg.replyAlreadySent) {
+                    ds.streamCardPending = false;
+                    persistStreamCardState(ds);
+                    logger.info(`[${t}] Explicit reply landed before worker ready — skipping stale starting card`);
+                    break;
+                }
                 // Bot opted out of the streaming card: the terminal is up and the
                 // final answer will still arrive via `botmux send`; just don't post the
                 // live status card. (workerPort/token above are still set so the web
@@ -2277,6 +2444,8 @@ function setupWorkerHandlers(ds, worker, startupState = { ready: false, failureN
                 break;
             }
             case 'prompt_ready': {
+                if (ds.worker !== worker)
+                    break;
                 logger.info(`[${t}] ${getCliDisplayName(effectiveCliId)} is ready for input`);
                 // A live prompt means a (re)spawn reached a working CLI — clear the lazy
                 // cold-resume marker set when we parked a crash diagnostic shell. The
@@ -2319,6 +2488,8 @@ function setupWorkerHandlers(ds, worker, startupState = { ready: false, failureN
                             ...(followUpCodexAppInput ? { codexAppInput: followUpCodexAppInput } : {}),
                         }, { codexAppInputAccepted: !!followUpCodexAppInput });
                 }
+                // CLI reached its prompt — any previously posted stuck warning is stale.
+                invalidateStuckWarning(ds, 'prompt_ready');
                 break;
             }
             case 'cli_session_id': {
@@ -2340,6 +2511,21 @@ function setupWorkerHandlers(ds, worker, startupState = { ready: false, failureN
                 }
                 break;
             }
+            case 'native_session_title_generated': {
+                if (ds.worker !== worker || ds.session.nativeSessionTitleUserDefined)
+                    break;
+                const title = msg.title.trim();
+                if (!title)
+                    break;
+                ds.session.nativeSessionTitle = title;
+                ds.session.nativeSessionTitleAwaitingContent = undefined;
+                if (ds.initConfig) {
+                    ds.initConfig.nativeSessionTitle = title;
+                    ds.initConfig.nativeSessionTitlePrompt = undefined;
+                }
+                sessionStore.updateSession(ds.session);
+                break;
+            }
             case 'screen_update': {
                 // Wait for `ready` (workerPort) before any card work — the read link
                 // is the LOCAL log terminal for every backend including riff
@@ -2551,6 +2737,137 @@ function setupWorkerHandlers(ds, worker, startupState = { ready: false, failureN
                 }
                 break;
             }
+            case 'stuck_warning': {
+                // Ignore stuck_warning from a stale worker generation — the CLI may have
+                // been replaced and a new worker owns the session now.
+                if (ds.worker !== worker)
+                    break;
+                // AI-free StuckDetector fired: a written input hasn't produced a
+                // completed turn within the timeout AND the PTY has been quiet, AND the
+                // snapshot matches a known Codex hook-review screen. The detector only
+                // emits when it matches level 1 or level 2 — there is no "unknown
+                // stall" path, so we always build an interactive card with the exact
+                // keys the screen documents.
+                // Dedup: skip if we already posted a warning for THIS turn, OR if a
+                // stuck-warning card is already active (covers the no-turnId case where
+                // a second stall fires before the first card is resolved).
+                const isDuplicateTurn = msg.turnId !== undefined && ds.stuckWarningTurnId === msg.turnId;
+                const hasActiveCard = !!ds.stuckWarningCardId;
+                if (isDuplicateTurn || hasActiveCard) {
+                    logger.debug(`[${t}] Stuck warning dedup skipped (turn=${msg.turnId ?? 'none'}, activeCard=${hasActiveCard})`);
+                    break;
+                }
+                // Allocate a daemon-side nonce for this warning (NOT the worker's
+                // generation — the daemon is the authority on which warning is active).
+                // If the CLI recovers (prompt_ready) or the worker is replaced while the
+                // POST is in flight, invalidateStuckWarning clears the nonce; when the
+                // POST returns we check it is still current and, if not, resolve the card
+                // immediately instead of registering it as active.
+                // Allocate a daemon-side nonce for this warning from a monotonic
+                // counter that is NEVER cleared (even when the active authority is
+                // dropped). This prevents the nonce-reuse race: warning nonce=1 POST
+                // awaits → prompt_ready clears active nonce → warning nonce=1 again
+                // would let the old POST register against the new authority. With a
+                // monotonic counter the second warning gets nonce=2, and the old
+                // POST's nonce=1 check fails.
+                const nonce = (ds.stuckWarningNonceCounter ?? 0) + 1;
+                ds.stuckWarningNonceCounter = nonce;
+                ds.stuckWarningNonce = nonce;
+                ds.stuckWarningTurnId = msg.turnId;
+                ds.stuckWarningCliLifetime = msg.cliLifetime;
+                const pageType = msg.matchedPattern;
+                ds.stuckWarningPageType = pageType;
+                const secs = Math.round(msg.elapsedMs / 1000);
+                logger.info(`[${t}] Stuck warning: turn unresolved for ${secs}s (${pageType}) nonce=${nonce}`);
+                emitSessionLifecycleHook(ds, 'session.requires_attention', {
+                    reason: 'stuck_warning',
+                    elapsedMs: msg.elapsedMs,
+                    matchedPattern: pageType,
+                });
+                if (managedAuxUiSuppressed(msg.turnId, msg.dispatchAttempt))
+                    break;
+                let description;
+                let options;
+                if (pageType === 'hook review level 2') {
+                    // Level 2 — per-hook review: t=trust, Esc=go back. No Enter here.
+                    description = `⚠️ Codex 卡在单项 hook 审核界面——消息已写入但 ${secs}s 未完成处理。\n\n选择操作以继续，或打开终端手动处理：`;
+                    options = [
+                        { label: 't', text: '信任 (trust)', selected: false, type: 'confirm', keys: ['t'] },
+                        { label: 'Esc', text: '返回 (go back)', selected: false, type: 'select', keys: ['Escape'] },
+                    ];
+                }
+                else {
+                    // Level 1 — hooks browser: t=trust all, Enter=review hooks, Esc=close.
+                    description = `⚠️ Codex 卡在 hook 审核界面——消息已写入但 ${secs}s 未完成处理。\n\n选择操作以继续，或打开终端手动处理：`;
+                    options = [
+                        { label: 't', text: '信任全部 (trust all)', selected: false, type: 'confirm', keys: ['t'] },
+                        { label: 'Enter', text: '逐项审核 (review hooks)', selected: false, type: 'select', keys: ['Enter'] },
+                        { label: 'Esc', text: '关闭 (close)', selected: false, type: 'select', keys: ['Escape'] },
+                    ];
+                }
+                try {
+                    const cardJson = buildTuiPromptCard(sessionAnchorId(ds), ds.session.sessionId, description, options, false, undefined, loc);
+                    const cardMsgId = await scopedReply(cardJson, 'interactive', msg.turnId);
+                    // Authority check: if the warning was invalidated (prompt_ready,
+                    // CLI exit, worker replace) OR the worker was replaced while the POST
+                    // was in flight, resolve the card immediately and do NOT register it
+                    // as active — a late click would otherwise inject keys into a
+                    // recovered/replaced CLI. Verify the full tuple: same worker process,
+                    // same worker generation, same nonce.
+                    if (ds.worker !== worker || ds.workerGeneration !== workerGeneration || ds.stuckWarningNonce !== nonce) {
+                        logger.debug(`[${t}] Stuck warning card POSTed but authority stale (nonce=${nonce}, current=${ds.stuckWarningNonce ?? 'none'}, workerGen=${workerGeneration}) — resolving`);
+                        const locDs = localeForBot(ds.larkAppId);
+                        const resolvedCard = buildTuiPromptResolvedCard(tr('card.action.tui_done', undefined, locDs), locDs);
+                        updateMessage(ds.larkAppId, cardMsgId, resolvedCard).catch(err => logger.debug(`[${t}] Failed to resolve stale stuck-warning card: ${err}`));
+                        break;
+                    }
+                    ds.stuckWarningCardId = cardMsgId;
+                    publishAttentionPatch(ds);
+                }
+                catch (err) {
+                    logger.warn(`[${t}] Failed to post stuck warning card: ${err}`);
+                    // Card send failed — clear markers ONLY if this nonce is still
+                    // current AND we still own the worker. A newer warning (nonce+1) may
+                    // have started while this POST was in flight; we must not wipe the
+                    // newer state.
+                    if (ds.worker === worker && ds.workerGeneration === workerGeneration && ds.stuckWarningNonce === nonce) {
+                        clearStuckWarningAuthority(ds);
+                    }
+                }
+                break;
+            }
+            case 'stuck_warning_expired': {
+                // Ignore from a stale worker generation.
+                if (ds.worker !== worker || ds.workerGeneration !== workerGeneration)
+                    break;
+                // Worker refused to inject keys from a stuck-warning card because the
+                // current screen no longer matches the page type the card was built for.
+                // Resolve the card with a "page changed" message so the user knows the
+                // action was NOT performed, then clear the authority.
+                if (ds.stuckWarningNonce === msg.nonce && ds.stuckWarningCardId) {
+                    const locDs = localeForBot(ds.larkAppId);
+                    const resolvedCard = buildTuiPromptResolvedCard('页面已变化，未发送按键', locDs);
+                    updateMessage(ds.larkAppId, ds.stuckWarningCardId, resolvedCard).catch(err => logger.debug(`[${t}] Failed to update stuck-warning card on expired: ${err}`));
+                    clearStuckWarningAuthority(ds);
+                }
+                break;
+            }
+            case 'tui_keys_delivered': {
+                // Ignore from a stale worker generation.
+                if (ds.worker !== worker || ds.workerGeneration !== workerGeneration)
+                    break;
+                // Worker confirmed the keys were written to the PTY. Clear the card
+                // authority and render success. We only do this AFTER the worker ACK
+                // (not on click), so a rejected click (stuck_warning_expired) does not
+                // falsely report success.
+                if (ds.stuckWarningNonce === msg.nonce && ds.stuckWarningCardId) {
+                    const locDs = localeForBot(ds.larkAppId);
+                    const resolvedCard = buildTuiPromptResolvedCard(tr('card.action.tui_done', undefined, locDs), locDs);
+                    updateMessage(ds.larkAppId, ds.stuckWarningCardId, resolvedCard).catch(err => logger.debug(`[${t}] Failed to resolve stuck-warning card on delivered: ${err}`));
+                    clearStuckWarningAuthority(ds);
+                }
+                break;
+            }
             case 'claude_exit': {
                 // CLI-generation authority must not outlive the concrete worker/CLI
                 // pair that issued it. A delayed message from a replaced Node worker
@@ -2560,6 +2877,10 @@ function setupWorkerHandlers(ds, worker, startupState = { ready: false, failureN
                     break;
                 }
                 ds.managedTurnOrigin = undefined;
+                // The worker/CLI generation ended. Disable an outstanding stuck-warning
+                // card before any replacement worker can be attached; otherwise a late
+                // click could inject its keys into the replacement CLI.
+                invalidateStuckWarning(ds, 'claude_exit');
                 logger.info(`[${t}] ${getCliDisplayName(effectiveCliId)} exited (code: ${msg.code}, signal: ${msg.signal})`);
                 ds.hasHistory = true;
                 try {
@@ -2979,6 +3300,19 @@ function setupWorkerHandlers(ds, worker, startupState = { ready: false, failureN
             ds.worker = null;
             ds.workerPort = null;
             ds.managedTurnOrigin = undefined;
+            // This worker generation is gone. Invalidate any stuck-warning card it
+            // posted so a late click cannot inject keys into a replacement worker.
+            invalidateStuckWarning(ds, 'worker_exit');
+            // Fence this lifetime before a polling dispatcher can observe its last
+            // ACK. Keeping the old receipt is useful audit evidence, but the
+            // persisted current generation advances immediately so it cannot count
+            // as acceptance after the worker has died. A stale takeover worker never
+            // enters this branch and therefore cannot fence the replacement.
+            const fencedGeneration = Math.max(workerGeneration, ds.workerGeneration ?? 0, ds.session.workerGeneration ?? 0) + 1;
+            ds.workerGeneration = fencedGeneration;
+            ds.session.workerGeneration = fencedGeneration;
+            ds.session.pid = undefined;
+            sessionStore.updateSession(ds.session);
         }
         try {
             const notified = cb.onWorkerExit?.(ds, {
@@ -3345,6 +3679,28 @@ export const __testOnly_setupWorkerHandlers = setupWorkerHandlers;
 export const __testOnly_finishTurnReactions = finishTurnReactions;
 export const __testOnly_finalOutputDedupeKey = finalOutputDedupeKey;
 // ─── Fork adopt worker ──────────────────────────────────────────────────────
+function reserveWorkerGeneration(ds) {
+    const previousDaemonGeneration = ds.workerGeneration;
+    const previousSessionGeneration = ds.session.workerGeneration;
+    const workerGeneration = Math.max(previousDaemonGeneration ?? 0, previousSessionGeneration ?? 0) + 1;
+    ds.workerGeneration = workerGeneration;
+    ds.session.workerGeneration = workerGeneration;
+    try {
+        sessionStore.updateSession(ds.session);
+    }
+    catch (error) {
+        if (previousDaemonGeneration === undefined)
+            delete ds.workerGeneration;
+        else
+            ds.workerGeneration = previousDaemonGeneration;
+        if (previousSessionGeneration === undefined)
+            delete ds.session.workerGeneration;
+        else
+            ds.session.workerGeneration = previousSessionGeneration;
+        throw error;
+    }
+    return workerGeneration;
+}
 export function forkAdoptWorker(ds, opts) {
     const cb = requireCallbacks();
     const workerPath = join(__dirname, '..', 'worker.js');
@@ -3362,6 +3718,9 @@ export function forkAdoptWorker(ds, opts) {
         logger.warn(`[${t}] read-isolation bot: refusing to adopt existing CLI (would run unisolated); will cold-start isolated on next message`);
         return;
     }
+    // Reserve before replacing an existing bridge worker for the same reason as
+    // forkWorker: persistence failure must leave the old lifetime untouched.
+    const workerGeneration = reserveWorkerGeneration(ds);
     // Guard against double-fork
     if (ds.worker && !ds.worker.killed) {
         logger.warn(`[${t}] Worker already running, killing before adopt-fork`);
@@ -3541,7 +3900,7 @@ export function forkAdoptWorker(ds, opts) {
         sessionStore.updateSession(ds.session);
     }
     // Use shared handler
-    setupWorkerHandlers(ds, worker, startupState);
+    setupWorkerHandlers(ds, worker, startupState, workerGeneration);
     ds.worker = worker;
     ds.spawnedAt = Date.now();
     ds.cliVersion = '';
@@ -3705,9 +4064,17 @@ function cleanupPersistentBackendSessions(backendType, activeSessions_) {
     const currentCliId = config.daemon.cliId;
     if (!multiBot && lastCliId && lastCliId !== currentCliId) {
         logger.info(`CLI_ID changed (${lastCliId} → ${currentCliId}), killing all ${backendType} sessions`);
+        // Legacy per-topic hosts are still enumerable by bmx-* name.
         for (const name of backend.listBotmuxSessions()) {
             backend.killSession(name);
         }
+        // Machine-wide Herdr agents are not separate bmx-* sessions. Tear down
+        // each persisted managed target precisely so a CLI switch cannot reattach
+        // the old executable, while never stopping the shared `botmux` host or an
+        // explicitly adopted user pane.
+        for (const target of managedTargetsForCliChange(backendType, activeSessions_)) {
+            killPersistentBackendTarget(target);
+        }
     }
     else {
         const activeNames = new Set(activeSessions_.map(s => backend.sessionName(s.sessionId)));
@@ -3730,9 +4097,12 @@ function cleanupPersistentBackendSessions(backendType, activeSessions_) {
                 continue;
             }
             if (botCliId && sessionCliId !== botCliId) {
-                const name = backend.sessionName(session.sessionId);
-                logger.info(`CLI mismatch for ${session.sessionId.substring(0, 8)} (session=${sessionCliId}, bot=${botCliId}), killing ${backendType} ${name}`);
-                backend.killSession(name);
+                const target = resolvePersistentBackendTarget(backendType, session.sessionId, session.persistentBackendTarget);
+                const label = target.backendType === 'herdr' && target.agentName
+                    ? `${target.sessionName}/${target.agentName}`
+                    : target.sessionName;
+                logger.info(`CLI mismatch for ${session.sessionId.substring(0, 8)} (session=${sessionCliId}, bot=${botCliId}), killing ${backendType} ${label}`);
+                killPersistentBackendTarget(target);
             }
         }
     }