harness: a belt-spawned turn waits in line for its serving seat instead of exiting

[gunbai-bot]

A belt-spawned harness turn (worker, reviewer, auditor, supervisor) that finds no admissible serving seat now waits in line instead of exiting.

Why

harness_seat already classifies "every admissible group busy" as HarnessSeatPending, and says the remedy is to ask again in a moment. But harness_cli turned it into ExitFailure { code: 1 } before any provider call. Three things followed:

• The dashboard read Agent failed.
• The belt recorded a yield without a captured submission.
• A continuation turn was spent on a turn that never reached a model.

On srv2 (2026-10-02/03), this used up all 3 turns of the shell-dag-live-deploy-restart-tailscale lineage. Group-b was saturated by traffic outside the seat ledger, and group-a was down. None of those turns ran a model round, and the supervisor's verdict said so itself.

What

• gunbc.harness.harness_cli harness_bind_seat_waiting. On Pending, the harness announces turn.placement_waiting (wait n of cap, plus the reason) on the attempt's own event stream, sleeps 60s, and re-binds. It's bounded at 1440 waits; at the cap it exits pending as before, loudly.
  • Refused, an unwritable announcement, or a failed sleep each end the wait at once.
  • The wait only defers asking. It never places a turn beside occupants the ledger can't explain.
• gunbc.harness.harness_turn harness_event_placement_waiting. The event's producer.
• gunbc.roadmap_provider_events:
  • New event kind HarnessPlacementWaiting, which maps to ProviderAwaitingSeat { detail }, a non-terminal state.
  • The jq projector keeps wait, wait_cap and reason. Without that, an unknown type projects to {type} alone and loses the reason.
• Every ProviderExecutionState consumer gains the arm. The agent segment shows as active — waiting for a serving seat — …reason; there's no commit, the handoff awaits the worker, and the attempt isn't terminal, so the lineage holds and no second attempt is spawned.

Evidence

• Witnesses:
  • roadmap_provider_events_witness_test: the harness's own wait line goes through the real reader, the reason and counter reach the detail, the label is awaiting_seat, a placed turn leaves the wait behind, and the projector keeps the reason.
  • roadmap_workflow_progress_witness_test: a waiting worker is an active agent with its reason, and a wait ending in a nonzero exit is failed.
  • roadmap_belt_actuate_witness_test: a waiting worker is not terminal until it exits.
• Run on srv2 against the deployed binary (deploy branch srv2/placement-wait, the same change on the old deployed revision cc5bbdcc): provider_events 25 pass, workflow_progress 29 pass, belt_actuate 140 pass and 1 fail.
  • The one failure, the_two_production_lineages_escalate_and_convene_naming_their_derived_supervisor, also fails at cc5bbdcc without this change (baseline run).
• Production route: deployed to srv2 (revision e1cc8709).
  • Attempt 9e12b415dd3899e4 emitted turn.placement_waiting and waited about 100 times. The issue page showed Agent · active — waiting for a serving seat — wait 9 of 1440: group-a … unavailable; group-b: the engine reports 7 occupant(s) while the seat ledger admitted 0.
  • Later, the supervisor turn and the renewed worker turn (01feb0b65dee1377) were seated through the ledger once group-b drained.
• Not yet run on this exact head: CI is the first run of the witnesses at main + this commit.

Known follow-ups (not in this PR)

• The page header still reads "Agent working" while waiting; the segment line underneath is correct.
• A launch with a missing worktree is admitted and shown as "Workspace · complete". Launch should check that the workspace exists.
• Exited worker and supervisor units block re-dispatch (stale_session) until a belt tick's teardown pass runs, so with the timer off they have to be cleared by hand.

🤖 Generated with Claude Code