fix(codex): prevent false turn idle timeouts under event-loop pressure (#86948)

Username · claude · Username · commit 78ec4fa2cf9c · 2026-05-27T07:05:37.000+08:00
The completion-idle timer (60 s) fires based on when notifications are *processed* inside the chained notification queue, not when they are *received* from the app-server binary over stdio. Since ea16a5e added a `setImmediate` yield between each queued dispatch, every notification is delayed by at least one event-loop tick. Under heavy load (tool-call-intensive turns) the ticks accumulate, and the `turn/completed` notification can arrive from the binary but sit in the queue for longer than 60 s — at which point the idle timer fires and aborts the turn. Fix: 1. Touch `turnCompletionLastActivityAt` and `turnAttemptLastProgressAt` at *enqueue* time (when readline delivers the line) so the idle timer sees live binary I/O even when the queue is backed up. 2. Guard all three idle-timeout handlers (`fireTurnCompletionIdleTimeout`, `fireTurnAttemptIdleTimeout`, `fireTurnTerminalIdleTimeout`) with the existing `terminalTurnNotificationQueued` flag so they never fire when `turn/completed` has already been received but not yet processed. Closes #86948 Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
diff --git a/extensions/codex/src/app-server/run-attempt.ts b/extensions/codex/src/app-server/run-attempt.ts
@@ -1979,7 +1979,12 @@ export async function runCodexAppServerAttempt(
   };
 
   const fireTurnAttemptIdleTimeout = () => {
-    if (completed || runAbortController.signal.aborted || !turnAttemptIdleWatchArmed) {
+    if (
+      completed ||
+      terminalTurnNotificationQueued ||
+      runAbortController.signal.aborted ||
+      !turnAttemptIdleWatchArmed
+    ) {
       return;
     }
     const idleMs = Math.max(0, Date.now() - turnAttemptLastProgressAt);
@@ -2014,6 +2019,7 @@ export async function runCodexAppServerAttempt(
   const fireTurnCompletionIdleTimeout = () => {
     if (
       completed ||
+      terminalTurnNotificationQueued ||
       runAbortController.signal.aborted ||
       !turnCompletionIdleWatchArmed ||
       activeAppServerTurnRequests > 0
@@ -2053,6 +2059,7 @@ export async function runCodexAppServerAttempt(
   const fireTurnTerminalIdleTimeout = () => {
     if (
       completed ||
+      terminalTurnNotificationQueued ||
       runAbortController.signal.aborted ||
       !turnTerminalIdleWatchArmed ||
       activeAppServerTurnRequests > 0
@@ -2581,6 +2588,17 @@ export async function runCodexAppServerAttempt(
     if (isTerminalTurnNotificationForTurn(notification, turnId)) {
       terminalTurnNotificationQueued = true;
     }
+    // Touch idle-watch timestamps at receive time (not just at process time)
+    // so that the completion-idle timer sees live binary I/O even when the
+    // notification queue is backed up behind setImmediate yields under heavy
+    // event-loop load.  See openclaw/openclaw#86948.
+    if (correlation.matchesActiveTurn !== false) {
+      const now = Date.now();
+      turnCompletionLastActivityAt = now;
+      turnCompletionLastActivityReason = `notification:${notification.method}`;
+      turnAttemptLastProgressAt = now;
+      turnAttemptLastProgressReason = `notification:${notification.method}`;
+    }
     notificationQueue = notificationQueue.then(
       () => handleNotification(notification),
       () => handleNotification(notification),
