Registry marks live idle agents "done" within minutes of spawn — silently disables submit verification and hard-fails sends

[EtanHey]

Found by the PR#404 reviewer while measuring, and flagged by it as the top follow-up. Not caused by #404 — it gates #404.

Behavior

The registry transitions live, idle agents to done within minutes of spawn. The agents are alive and responsive; only the record is wrong.

Two measured consequences (N=29 probe run, branch build)

1. send_to(mode=agent) hard-fails against a perfectly healthy agent, because the state gate sees done. Both mode=agent errors in the run were this, and nothing else.
2. Submit verification is silently skipped on surface sends — shouldVerifySubmit is false when the record is not ready/idle, so 7 of 10 surface sends in the run never verified. That is precisely the path where a false submitted receipt then slipped through (the fix: make agent delivery receipts trustworthy #404 round-3 blocker).

So a stale record does not just mislead a reader — it disables the delivery machinery that everything else now depends on, and caps how much of #404's fix is reachable in real use.

Related

markAgentWorking only transitions from idle, so an agent parked in done never climbs back to working either — which is why dispatch_to_agent's busy-queue branch never fired live during the same run.

Expected

An agent that is alive and whose pane parses as ready/idle/working must not be recorded done. done should mean the work ended, not "we stopped hearing from it." The live-reality derivation from #398 is the obvious lever — list_agents already reconciles against the screen; the state gates that guard delivery should consult the same reconciled truth rather than the raw record.

— cmuxlayerClaude (lead) · claude-code/claude-fable-5

[EtanHey]

Additional evidence from the PR #404 round-3 review (now merged as v0.4.34), which sharpens why this issue caps the delivery work:

The surface path has no allow_busy escape. The agent path does, and the reviewer used allow_busy:true to get past the stale-done gate on its one mode=agent send this round. Surface sends have no equivalent, so with this bug in play every surface send in the round read as unverified — shouldVerifySubmit requires a ready/idle agent record (server.ts:7775), and a stale done record silently turns verification off.

Consequence worth stating plainly: #404 made receipts honest, but a stale done record means the honest answer is usually "unverified." The fix's real-world reach is gated by this issue, not by its own design.

Also recorded from the same review: the round's headline "0% error rate" was achieved partly because the reviewer adapted around this gate. Without allow_busy:true it expected the same two errors as the previous round. The reviewer put that caveat in its own report rather than letting the number read better than it was.

— cmuxlayerClaude (lead) · claude-code/claude-fable-5

[EtanHey]

Root cause found — and it is the same defect as PR #411's FN-2, in a second location

maybeMarkTaskDone (src/agent-engine.ts:3100) transitions an agent to state:"done" when hasGroundTruthDone() returns true. That predicate (agent-engine.ts:1833) is:

if (!this.transcriptHasSettledDone(agent)) return false;
// ...then only checks the screen does not CONTRADICT it
return !this.screenContradictsTranscriptDone(agent.cli, screen.text);

transcriptHasSettledDone → claudeDone() (src/harness-session.ts:188-220) → true when the last assistant message has stop_reason:"end_turn". That is the model finished its turn, not the task is complete.

So for every Claude agent: finish any turn → transcript says end_turn → an idle screen does not contradict it → state:"done". That is why live agents go done "within minutes of spawn" — it is really after their first turn, which for a freshly-spawned worker is the same thing.

Codex is not affected the same way: its equivalent is genuinely task-scoped (task_complete sets, task_started clears).

Why this matters twice over

PR #411 hit this exact predicate as FN-2 and fixed it inside the halt classifier only — gating halt detection on it meant idle_without_done could never fire for a Claude agent. The fix did not touch maybeMarkTaskDone, which is the code that writes the wrong state in the first place. So the halt lane worked around the symptom in its own path while the source kept producing it.

This is the second time the same misread signal has caused a distinct bug (ledger #27), which argues for fixing it at the source rather than per-consumer:

• either claudeDone() stops being used as a task-completion signal anywhere and is renamed to what it is (claudeTurnEnded), with task-completion for Claude requiring an actual DONE artifact;
• or every consumer is audited. There are at least three (maybeMarkTaskDone, the halt classifier, hasRecordedOutputDoneEvidence's feeder).

The first is the real fix. A predicate whose name says "done" and whose meaning is "turn ended" will keep being misread — it already has been, by two separate lanes.

Downstream damage already measured

• send_to(mode=agent) hard-fails against healthy agents whose record says done (both mode=agent errors in the fix: make agent delivery receipts trustworthy #404 measurement round were this, and nothing else).
• Submit verification is silently skipped when the record is not ready/idle — 7 of 10 surface sends in that round never verified, which is the path that let a false submitted receipt through.
• markAgentWorking only transitions from idle, so an agent parked in done never climbs back to working — which is why dispatch_to_agent's busy-queue branch never fired live.

— cmuxlayerClaude (lead) · claude-code/claude-fable-5

[EtanHey]

Additional evidence (2026-08-19): the P12 evaluator's resume-probe — which preserved its public agent_id across resume when tested live — degraded overnight to auto-claude-3768979C-... with session_id null while its pane sat unattended. So identity churn also hits RESUMED agents left idle, not only discovery-time races.

— cmuxlayerClaude (lead) · claude-code/claude-fable-5