Repository navigation
Description
Problem
A Codex chat in the ChatGPT macOS desktop app appeared stuck for roughly six minutes during a web search. Local read-only diagnostics show a successful 3.4-second code-mode operation, followed by an approximately 300-second outer tool dispatch and a Responses WebSocket idle timeout. The chat then recovered automatically after reconnecting.
Environment
- macOS 27.0.1 (26A434).
- Bundled executable inspected after the incident: codex-cli 0.160.0. This is installed-version evidence, not independent attribution of the running app-server binary.
- Model recorded in the affected turn: gpt-6.1-sol, medium reasoning.
- Date: October 5, 2026. Timeline below is UTC.
- Local desktop execution, signed in with ChatGPT.
Observed timeline
| UTC | Evidence |
|---|---|
| 06:44:59.604 | exec invokes web search through tools.web__run |
| 06:45:03 | codex.code_mode.host_timing: host operation completed, duration 3403844375 ns; span includes outcome="interrupted", aborted=false |
| 06:47:23 | A follow-up message is accepted by turn/steer, response duration 2 ms; no new assistant output yet |
| 06:49:59.860 | Rollout records successful custom_tool_call_output: "Script completed", "Wall time 3.4 seconds", with populated search results |
| 06:49:59 | codex.tool_call logs handler_duration_ms=300253 and total_duration_ms=300253 |
| 06:49:59 | responses_retry logs "stream disconnected before completion: idle timeout waiting for websocket", retry 1/5 |
| 06:50:00–01 | A new Responses WebSocket connection is established successfully |
| 06:50:16 | New assistant/tool activity appears |
| 06:50:37 | Turn completes successfully; native thread status becomes idle |
Expected
A completed tool operation should be delivered promptly and generation should continue. If a stream stalls, the UI and diagnostics should distinguish waiting/reconnection from productive execution and preserve follow-up input.
Investigation requested
Please investigate the gap between code-mode host completion and outer tool dispatch completion: was dispatch waiting on response-stream completion or another synchronization boundary? Why did it take the 300-second WebSocket idle timeout to release the result and resume the turn? Please clarify the meaning of outcome="interrupted" for successful code-mode operations; the same label also occurs on successful short calls in another chat.
The evidence establishes the timeout and automatic reconnect, but does not establish whether the initiating fault was in the client, code-mode synchronization, a network intermediary, or the service. It is not a confirmed deadlock, and no deterministic reproduction has been established. No app restart, configuration change, or project edit was needed for this recovery. Automatic recovery is not a permanent fix.
Related symptom: #50086; a shared root cause is not established. Raw transcripts, project paths, credentials and unrelated logs are intentionally excluded. Redacted local timing records remain available.
Activity
github-actions commentedon Oct 5, 2026
Potential duplicates detected. Please review them and close your issue if it is a duplicate.
Powered by Codex Action