After a restart interrupts an open question, the thread keeps a pending-input count that nothing can clear, and the mobile app keeps failing to answer it #14190

Two of my Claude threads lost their open questions when T3 restarted their sessions. Afterward I could neither answer the questions nor stop them, my phone kept sending notifications for the threads, and I could not archive or delete them from the desktop app. The desktop app showed "Stale pending user-input request... Restart the turn to continue" when I tried to answer, and I found no way to stop the question.

What happened (two threads, same pattern)

1. Both threads were running in one project. T3 moved both sessions to starting at 03:19:13 UTC and again at 03:37:43 UTC on Sept 28, within a few milliseconds of each other each time. I don't know what triggered either restart.
2. Each restart interrupted a question call that was open. The agent's own notes say the call "got interrupted before I could confirm it registered", so it asked again. Each thread ended up with three requests for the same question, from Sept 27 evening, 03:19, and 03:37.
3. At 13:52 UTC I tried answering one of them from the desktop app. It failed with Stale pending user-input request ... Provider callback state does not survive app restarts or recovered sessions. Restart the turn to continue. I then interrupted the turn. The interrupt resolved only the newest request. The session went to stopped, and the older requests never got a user-input.resolved activity. The thread row kept a pending-input count of 1.
4. The mobile app (1.3.0) then sent 11 answers for one of the older requests between 23:17 and 23:25 UTC. Seven of them landed within 1.4 seconds. Each failed with No active provider session is bound to this thread. and appended a provider.user-input.respond.failed activity.

Why nothing clears it

• thread.user-input.dismiss rejects native questions with "This question needs an answer. Answer it or stop the turn." (apps/server/src/orchestration/decider.ts:1772). The session is already stopped, so there is no turn to stop.
• reconcileOrphanedNativeUserInputs (apps/server/src/serverRuntimeStartup.ts:846) exists to dismiss these on startup. The Nightly bundle contains that code. The database holds zero command receipts with the startup:user-input-resolved: prefix, so it never dismissed either thread. I have not found out why.

Expected

A question with no live session behind it gets resolved, so the phone stops offering it and stops notifying. A restart that interrupts a question call does not leave duplicate requests behind. Archive and delete from the desktop app work regardless of pending input.

Workaround that worked

t3 auth session issue --ttl 5m, then POST /api/orchestration/dispatch with a thread.archive command. The server returned 200 and both threads stopped notifying.

Not verified

• What restarted the sessions at 03:19 and 03:37 UTC. The trace logs I have do not cover those times.
• Why archive and delete from the desktop app did nothing. The server has no receipt for either command, so the request likely failed before it was sent.
• The full path from the pending question to the phone notification. I saw the failure activities and the repeated mobile answers, not the notification sender.

Environment

Desktop: T3 Code Nightly 0.0.43-nightly.20260927.2365 on macOS. Mobile app: 1.3.0. Provider: claudeAgent. Source checked: 1de563c149.

I searched open and closed issues for "user input stale pending" and "archive thread pending question" and found nothing.