Codex Timeout Leaves External Actions Uncertain
A lost connection can leave a task's most consequential step in limbo.
The task may stop before an action starts. It may also lose contact while the action is running, even as the destination accepts the request. Retrying without checking could repeat work that already took effect.
On October 13, OpenAI reported degraded performance for GPT-6.1 Sol in ChatGPT Work and Codex. The disruption affected ChatGPT and Codex Web. The API and CLI were affected too, as was the VS Code extension. OpenAI's incident notice recorded mitigation and later recovery. UptimeRobot's chronology puts the initial notice at 08:10 UTC and the mitigation update at 09:30 UTC. The incident was marked resolved at 16:36 UTC.
A user report in the Codex GitHub repository logged 25 HTTP 503 responses across October 12-13. It also recorded 59 model-capacity messages. The report lists 17 pre-compaction errors. These are one user's observations, not platform-wide totals. A 503 or capacity rejection confirms a server-side error, but says nothing about whether an external operation started or the destination accepted it. The public incident record does not describe what happened to in-flight tasks. It reports no duplicate deployments. Nor does it mention duplicate pull requests, messages or writes.
Check the destination before starting again
Inspect local files, then compare local status with the diffs. Search Git for the intended branch or commit; check the pull-request title separately. Review deployment history and the active version. Check tickets, then inspect messages. Look for scheduled jobs and relevant application records. For APIs, search by the client-generated operation key or business identifier.
A matching remote object created during the incident window is stronger evidence than a missing final chat response. Preserve timestamps and IDs from the agent task and the destination.
A separate report covered a scheduled-job issue involving ChatGPT Work errors. Here, the question is whether a stateful action reached its destination.
Replay only what remains undone
When an API supports idempotency, reuse the original key for an uncertain mutation. Keep the request body unchanged if the API contract requires an exact match. For versioned records, read the current revision and honor If-Match or another compare-and-set precondition. In Git, continue from the existing branch and commit instead of creating a second branch with equivalent changes.
If the destination has no idempotency control, search for an existing operation with a stable task or operation identifier before making another change. A human-readable title may not be unique enough.
OpenAI's ChatGPT troubleshooting guide recommends regenerating a failed response after transient errors. That advice fits text-only replies, not every agent workflow. A new Codex Cloud task gets a separate workspace and cannot recover another task's uncommitted changes. Read the destination state first, then reconcile what happened.
OpenAI resolved the provider incident at 16:36 UTC.
