Swarm + goal mode: quota 403 kills all subagents silently — session idles for hours with no surfaced error, failed subagents never resumed

Bug description

In /goal + /swarm mode (YOLO permission), when the Kimi subscription quota is exhausted mid-run, the API returns 403 permission_error to subagent and main-agent turns. Every affected turn fails terminally, and afterwards the entire session goes completely silent — zero log events for ~5.5 hours. No retries, no subagent respawn, no error aggregation, no surfaced status. The swarm looks alive but is dead. Manual "Resume the active goal" steers fail instantly (~1 s) with the same 403 and nothing is queued. After the user switched auth to an API key the next day, work resumed — but 8 of the 13 quota-killed subagents were never re-run (their wire files end exactly at the 403); the main agent spawned replacements instead, with no record of the original failures being tracked.

Environment

• kimi-code 0.29.2 (npm @moonshot-ai/kimi-code, global install; verified via ~/.kimi-code/updates/rollout.log)
• Node v22.23.1, Ubuntu 26.04 LTS x86_64 VPS
• Mode: /goal + yolo + /swarm (manual; per wire events permission.set_mode, swarm_mode.enter, goal.create)
• Model k3 (kimi-for-coding/k3); auth at incident start: Kimi subscription; later switched to API key
• Swarm: 17 subagents in the first wave

Timeline (from logs/kimi-code.log and per-agent wire.jsonl, all UTC, 2026-07-27)

• 16:38–17:47 — session created, goal set, 17 subagents running normally.

• 17:48:15–17:49:26 — quota wall: 13 subagents + main agent all fail within ~70 s. Representative log lines:

  WARN  llm request failed  turnStep=0.39 attempt=1/10 model=k3 errorName=APIStatusError errorMessage="403 {\"error\":{\"type\":\"permission_error\",\"message\":\"You've reached your usage limit for this billing cycle. ...\"}}" statusCode=403 agentId=agent-4
  ERROR turn failed  turnId=0 agentId=agent-4
  

  Note attempt=1/10 immediately followed by turn failed — fine for a 403 — but nothing else happens either.

• 17:49:26 → 23:07:28: total silence, zero log events for ~5.5 h.

• 23:07:28 user steers "Resume the active goal." → main turn fails ~1 s later with the identical 403. Same again at 2026-07-28T00:47:33. Another ~11.8 h of silence.

• 12:37:44 (after API-key switch): resume succeeds. 5 of 13 killed subagents show later activity; 8 never ran again — wire files end exactly at the 403 moment.

Expected vs actual

• Expected: on a non-retryable auth/quota error, the swarm/goal scheduler should surface a clear blocking status (pause the swarm, mark subagents failed, tell the user what happened and how to resume), and resume/retry failed subagents once auth is restored.
• Actual: every turn failed independently and silently; the session idled for hours with no log activity; manual resumes failed in ~1 s with no queued retry; after auth was fixed, most failed subagents were never re-run and their failures are untracked.

Suspected cause (speculation)

Quota 403 is treated as a per-turn terminal error with no escalation path to the swarm/goal scheduler; subagent task state is not marked failed-and-resumable, so the scheduler has nothing left to do and goes idle while the UI (per user recollection, not in logs) still shows the swarm as active.

Repro notes

Not deterministic without a near-exhausted subscription quota, but conceptually: start a /goal + /swarm session with 10+ subagents on a subscription account, exhaust the billing-cycle quota mid-run, observe all turns fail with 403 and the session go idle with no surfaced status; then "Resume the active goal" and observe instant 403 with no queued retry.

Attachments (available on request)

Full session artifacts preserved: logs/kimi-code.log (quota errors at lines 803–925; silence gaps 17:49Z→23:07Z and 00:47Z→12:37Z), agents/main/wire.jsonl, per-subagent agents/agent-0…41/wire.jsonl (the 8 never-resumed subagents' wires end exactly at their 403 timestamps), state.json, and rollout.log proving version 0.29.2 at incident time. Happy to attach any of these.

Start with logs/kimi-code.log, the main and per-agent wire.jsonl files, state.json, and rollout.log to correlate the quota 403 failures with the later silent periods. Trace how failed turns and active goal/swarm state are recorded, then verify that a non-retryable quota failure is surfaced and that affected subagents become resumable after authentication is restored.