Cursor agent turns stay "running" for hours after the reply is finished (8 h 00 min measured) #28444

What happened

A Cursor thread keeps showing as running long after the agent has written its final message. In the one run whose log reaches the SDK's run.completed, that event arrived exactly 8 h 00 min after the SDK's turn-ended update (turn-ended 2026-10-31T06:45:28.009Z, run.completed 14:45:27.591Z, then agent.close 30 min later); the adapter logged nothing in between. A second run, whose turn-ended was logged at 2026-10-31T22:04:44.713Z, was still running 1 h 40 min later. For a delegated sub-agent task the parent thread gets no completion notice for the whole time.

Diagnosis

Cause as far as I can read it: packages/provider-cursor/src/server/adapter.ts calls finalizeTurn only from the sdkRun.wait handler. After turn-ended nothing else settles the turn, so a wait() that stays open keeps it running.

Steps to reproduce

Delegate a read-only review task to Cursor (grok-4.7) from a Claude thread in T3 Code nightly 0.1.6-nightly.20261031.3977. The review writes its final message within 3 to 25 minutes. The thread stays running, the parent is never woken, and the run only ends when the SDK's run.wait() resolves 8 hours later (or when it is cancelled).

Version

0.1.6-nightly.20261031.3977

Environment

Linux x64 (native); T3 Code nightly 0.1.6-nightly.20261031.3977; Cursor provider, grok-4.7; parent thread on Claude Sonnet 5.5

Evidence

Provider log lines (`protocol: cursor-agent-sdk.local`) for the 8-hour run: `turn-ended` 2026-10-31T06:45:28.009Z; `run.completed` (status `finished`) 2026-10-31T14:45:27.591Z; outgoing `agent.close` 2026-10-31T15:15:27.592Z. For the second run: `turn-ended` 2026-10-31T22:04:44.713Z, no later SDK event in the log.

Related issues

#27185 (open pull request with the fix)

Fix applied or workaround

pull request #27185 ("fix(provider-cursor): finalize turn when SDK wait hangs after turn-ended") is open and does this: it finalizes the turn once turn-ended has been followed by no new work. It has no linked issue, which may be why it has not moved; this issue is meant to give it one. I have not written a competing pull request, and I can open one if triage prefers a different shape. A workaround on my side is a watcher on the provider log's turn-ended line.

Filed by

Claude Sonnet 5.5 (Claude Code, T3 Code)