Repository navigation
Description
What happened
An existing Claude thread stopped producing output after a gateway usage limit. Continue, steering, and a later Stop followed by "start again" did not restore progress. The resumed run remained Working for about 16 minutes with no assistant output. A failed child notification then queued behind the active run. A fresh takeover thread failed promptly with the same gateway 429.
Diagnosis
Confirmed from provider protocol logs and read-only SQLite projections on the exact release tag b5554ac. The resumed query accepted the prompt: command_lifecycle queued and started arrived. No further native frames arrived before triage stopped it. Run and attempt remained running, completedAt null, with no pending user request. provider-turn.start had succeeded. The child task correctly settled failed; its completion message was queued behind the still-running parent, the queue alone does not prove deadlock.
ClaudeAdapterV2.ts:11962 handles the prompt lifecycle; a matching started marks promptStarted. The available evidence does not establish why the CLI stopped emitting events, or whether the root cause is in T3, Claude Code, or CLIProxyAPI.
Separately, this build misses the gateway reset when (last error: ...) precedes (reset after ...), in ClaudeAdapterV2.ts:4767. PR #37692, merged after this nightly, fixes that parsing issue. It does not demonstrate a fix for the silent resumed query.
Steps to reproduce
Observed sequence; no deterministic isolated repro yet:
- Use Claude Agent through a CLIProxyAPI executable, model claude-fable-5-2, effort high, full access, resuming a long existing session.
- Receive a direct gateway 429 reporting all credentials cooling down.
- Continue and steer; the CLI acknowledges steering as queued but emits no assistant output.
- Stop, then send "start again". A resumed query opens, acknowledges queued/started, then produces no further frames.
- A child result arriving meanwhile queues behind the parent.
Expected: a visible indication/recovery path for a silent accepted native query; preserve useful gateway cooldown information. Actual: Working continues until manually stopped.
Version
0.1.10-nightly.20261116.4787 (desktop/server; also triage CLI)
Environment
macOS darwin arm64 27.2.0; Node v24.22.1; T3 Code Nightly desktop hosting local server. Custom claude_proxy instance using CLIProxyAPI claude-proxy executable, Claude Agent SDK query protocol. CLI version not established. Original client surface not independently established.
Evidence
2026-11-16 UTC, redacted identifiers:
16:56:04 prior run failed: API Error: Request rejected (429) · All credentials for model claude-fable-5-2 are cooling down (last error: rate_limit_error: This request would exceed your account's rate limit. Please try again later.) (reset after 44h4m3s)
17:11:40.525 outgoing query.open, resume=<existing-session>
17:11:40.532 outgoing prompt.offer, uuid=<prompt>
17:11:40.534 incoming command_lifecycle, same uuid, state=queued
17:11:40.537 incoming command_lifecycle, same uuid, state=started
[No further native frames until Stop]
17:22:00.564 child task failed: same gateway 429 for claude-sonnet-5-5, reset after 43h38m7s
17:22:00.608 completion notice queued as next run; parent still running
17:27:22.890 outgoing query.interrupt
17:27:22.933 outgoing query.close
17:27:24.957 parent interrupted successfully
Before recovery: pendingRequestCount=0, parent completedAt=null. Start effect succeeded; no failed outbox effect among recent operations.Related issues
#33001: related silent resumed query, but that report had no incoming frames and Stop hung; here queued/started arrive and Stop works. #37226: child finalization failure, whereas this child settled and delivered its result to the queue. #37692 fixes the separate cooldown parsing defect.
Fix applied or workaround
Triage used t3_thread_interrupt(stop=true) on the affected top-level thread. Parent run became interrupted, the queued child-only notification became cancelled, activeRunId became null, and checkpoint capture succeeded. No server restart, direct database edits, or source patch. The separate gateway credential cooldown remains; work needs available provider capacity.
Filed by
Codex, gpt-6.1-sol (high), via t3 triage
Activity
fix(server): a Claude prompt that starts and then goes silent fails a…