fix(agents): release session write-lock immediately when holder PID is dead

yplongapp · yplongapp · commit f00c6556b6b2 · 2026-07-06T19:58:54.000+08:00
When an agent run fails abnormally during tool execution, the session write-lock remains held in memory. The watchdog's force-release threshold inherited the full maxHoldMs (up to 17 minutes for a 15-minute agent timeout), causing all subsequent requests to the same session to fail with SessionWriteLockTimeoutError until the watchdog fires. Fix: read the lock file payload during each watchdog check and release the lock immediately if the holder PID is no longer alive, regardless of maxHoldMs. Fixes #100872
diff --git a/src/agents/session-write-lock.ts b/src/agents/session-write-lock.ts
@@ -286,6 +286,26 @@ function releaseAllLocksSync(): void {
 async function runLockWatchdogCheck(nowMs = Date.now()): Promise<number> {
   let released = 0;
   for (const held of SESSION_LOCKS.heldEntries()) {
+    // Release locks whose holder process is no longer alive, regardless of
+    // maxHoldMs. This prevents sessions from being locked for the full agent
+    // timeout duration (up to 17 minutes) when the owner process terminates
+    // abnormally during tool execution. (#100872)
+    const lockPayload = await readLockPayload(held.lockPath);
+    if (lockPayload) {
+      const pid =
+        isValidLockNumber(lockPayload.pid) && lockPayload.pid > 0 ? lockPayload.pid : null;
+      if (pid !== null && !isPidAlive(pid)) {
+        process.stderr.write(
+          `[session-write-lock] releasing lock held by dead pid ${pid}: ${held.lockPath}\n`,
+        );
+        const didRelease = await held.forceRelease();
+        if (didRelease) {
+          released += 1;
+        }
+        continue;
+      }
+    }
+
     const maxHoldMs =
       typeof held.metadata.maxHoldMs === "number"
         ? held.metadata.maxHoldMs
