Grok turn stalls mid-reasoning: no tokens for 70+ minutes, UI stays Working, no turn_ended #7210

Summary

On a Grok thread, T3 Code stayed on Working for ~74 minutes after the provider stream died mid-reasoning. The user saw continuous activity. Local Grok session logs show the turn never ended.

This is not an app-restart ghost session, and it is not “the provider already finished.” The Grok turn itself froze after streaming_reasoning started, and T3 never timed it out or settled it.

Environment

• T3 Code desktop: 0.0.33 (T3 Code (Alpha).app, com.t3tools.t3code)
• Provider: Grok Build (grok-4.6-build, high reasoning)
• macOS, Apple Silicon
• Session dir: ~/.grok/sessions/

What happened

1. User asked for a small follow-up (same pattern as a previous ~10 minute Grok task).
2. Grok actually edited files for about 2 minutes.
3. Last tool completed, then first_token + phase_changed: streaming_reasoning.
4. Then 72 minutes of silence: no tools, no text, no turn_ended.
5. UI kept showing Working the whole time.
6. Session later emitted session_start / MCP reconnect, and only then did the next user message land.

Evidence from ~/.grok/sessions/<session>/events.jsonl

• 10:51:25Z turn_started (no turn_ended before the user interrupted)
• 10:51–10:53Z real tool calls (grep / read_file / search_replace)
• 10:53:04Z last tool_completed
• 10:53:07Z first_token, then many streaming_reasoning phase changes
• next event 12:05:40Z mcp_config_resolved / session reconnect
• 12:06:14Z user turn asking why it had been 74 minutes

updates.jsonl also has no entries in that 72-minute gap.

Expected

If the Grok stream stalls after streaming_reasoning with no further tokens, T3 should time out, mark the turn failed/interrupted, and stop showing Working.

Actual

The thread stayed Working for over an hour with no progress and no way to tell the stream was dead.

Related but not the same

• [Bug]: App restart during active turns leaves threads stuck on Working forever, with no startup reconciliation of dead provider sessions #4584: Working forever after app restart; session file frozen mid-turn. Same ghost-Working shape, different trigger (we did not restart at the stall).
• Chat shows 'working...' indefinitely after opencode CLI already finished responding #2644: Working after the provider already finished. Here the Grok turn itself never finished.
• Provider notification consumers are interrupted when startSession returns (Cursor, Grok, Codex) #5781 / fix(server): keep provider notification consumers alive past startSession #6538 (closed/merged 2026-08-15): GrokAdapter dropped the notification consumer on startSession. Different mechanism. That fix is in Nightly after v0.0.33; this report is from 0.0.33, so we will also try a current Nightly.

Offer

Happy to add Nightly version + a second trace if the stall still happens after #6538.