[Gastown] Refinery repeatedly gets stuck working on phantom/closed bead IDs after gt_done

[kilo-code-bot]

What happened?

The refinery agent repeatedly enters a working state hooked to bead IDs that don't exist in any active bead list (phantom IDs: 2636ff8f, c3c341c6). This happens immediately after a polecat calls gt_done on a completed bead. The refinery hooks to a nonexistent bead ID and stays stuck in working indefinitely, blocking the review queue. Requires manual gt_agent_reset each time. Has occurred at least 3 times in this rig within the same session. The escalation message is "Refinery stuck on closed bead after gt_done" / "Agent stuck in working state on closed bead after gt_done".

Area

Merge Queue / Refinery

Context

• Town ID: a1d1b54a-0bc6-4b39-a508-24ba66b213cd
• Agent: Mayor (e9b2bcb7-4965-41f9-97da-f11b68c2c520)
• Rig ID: cb8729a0-f334-45f0-8a1b-ec875043733f

Recent Errors

Refinery hooked to bead 2636ff8f-2cd1-4eb4-9ba6-398c7c8e3e4f (not found in bead list), then c3c341c6-94fb-47dc-89a6-6a8e36f8617d (not found in bead list). Both times status=working, required manual reset.

Filed automatically by the Mayor via gt_report_bug.

[jrf0110]

Investigation Results

Ran diagnostics against town a1d1b54a via Analytics Engine.

Corroborated: YES

• 439 dispatch failures over 2 days (80 refinery, 359 polecat)
• Refinery 1ae508ae was repeatedly dispatched to 20+ different bead IDs, with beads seeing 10-11 failures each
• Smoking gun: Bead 2636ff8f was closed at 13:46:26, but the refinery was dispatched to other beads seconds later that also failed — the reconciler kept dispatching even after the target bead was closed
• Zombie agent e9b2bcb7 stuck in working with no hook for the entire observation window — 12,004 invariant [7] violations (one per 5s tick)
• Reconciler in a tight loop emitting transition_agent actions every ~30s that never resolve

Root cause: Agent hook/status desynchronization after gt_done. The refinery calls gt_done, bead closes, but the reconciler dispatches the agent to the next pending bead before the unhook event is fully processed. The dispatch fails because the agent is in a stale state. This is the same class of bug as #1653 (no circuit breaker on dispatch failures).

Related issues: #1653 (dispatch circuit breaker), #1584 (role-based tool filtering)