Claude background subagents still stay Running after their completion arrives in a wake continuation (0.1.2, follow-up to #17809) #24337

Area

apps/server

Summary

#17809 was closed on Oct 22 as not reproducible on main. It still happens on the v0.1.2 release (macOS desktop app, Claude CLI 2.1.310, claude-opus-5-5, Claude Agent provider).

Timeline (one thread, event log + provider log)

Four background Agent calls (run_in_background: true, spawn_depth: 1) were all launched inside run 1 (07:56:52Z to 08:00:50Z). For one of them, native id a2eda89e0e1e87a18:

• task_started at 07:59:22.379Z, in run 1. The first subagent.updated running was written at 07:59:22.415Z.
• Run 1 ends with its result at 08:00:50Z.
• Run 13 is a wake continuation, started at 08:54:25Z for another task's task_notification.
• During run 13, at 09:05:24.220Z, the SDK sends task_notification {status: "completed"} for a2eda89e0e1e87a18, with its final summary. No task_started for this task appears after run 1.
• The row never goes terminal. Another subagent.updated with status: running was written at 22:21:29Z, 13 hours after the completion. Server shutdown recovery at 01:21:55Z the next day finally cancelled it.

All four behaved the same way. Five more background subagents in three other threads on the same machine and day did too. They had all finished: each one's own transcript ends in a final end_turn answer or a user interrupt.

Background agents whose final notification was preceded by a fresh task_started (interim completions, SendMessage resumes) settled correctly.

Workarounds tried

• subagent.stop on a stuck row is a no-op. The row stays running.
• thread.stop on the idle parent thread ends the stuck rows (as cancelled) and leaves the thread usable. We run that from a watcher as a stopgap.

Related

• Claude wake latch: a wake queued behind a held run swallows later wakes, so idle-launched subagents stay Running #22033 / PR fix(server): settle late-discovered Claude subagents #22049: subagents launched while the root is idle. That is a different path; these were launched inside an active run.
• [Bug]: Claude background subagent completion is overwritten by a late 'running' update, and Stop/cancel can't settle it #22279: a late running overwrite, possibly what wrote the 22:21Z update.
• [Bug]: Nested Claude subagent stays Running after the SDK reports it completed during a user turn #16355: nested subagents. We hit that too, separately.

Provider-log lifecycle frames for the thread are available on request.