How to recover Codex tasks after OpenAI’s GPT-6.1 Sol incident

OpenAI resolved the [GPT‑6.1 Sol incident](https://status.openai.com/incidents/jwna590w) at 16:36 UTC after listing Codex Web, API, CLI and VS Code as affected. If a cloud task stopped during that window, open the original task first. OpenAI's [Codex Cloud guide](https://help.openai.com/en/articles/20001545‑using‑codex‑cloud) says that starting a new task creates a separate workspace and does not recover another task's uncommitted changes. Continuing in the original task preserves the best chance of seeing its files, command history and last response together. If the original task is unavailable, record its link or ID and the last access time before contacting Support. Do not assume an empty new workspace means the old changes were erased.

Treat the last visible assistant message as incomplete evidence. Check the repository itself: git status --short --branch, the diff, generated files, test output and any commits created during the task. Preserve unrelated dirty work.

Then inspect external artifacts that the task may have touched, such as a pull request, deployment, issue, database migration or package release. A response can time out after an external service accepted an action. The recovery question is therefore what state exists now, not simply whether the chat showed a success message.

Write down the last confirmed step and the first uncertain step. That boundary gives the resumed task a precise place to continue without repeating completed work.

Begin with reads: file status, current branch, test reports, remote object state and provider dashboards. Once the actual state is known, retry only the missing operation with the same stable identifier where the integration supports idempotency. For a coding task, that may mean re‑running a failed test rather than regenerating the patch. For a pull request, read the remote PR list before creating another. For an API workflow, query the application record before replaying a POST. Avoid broad cleanup while recovering. Resetting a branch, deleting files or starting from a fresh workspace can destroy the evidence needed to distinguish partial completion from no completion.

The incident covered GPT‑6.1 Sol on the named surfaces, but OpenAI's status page notes that individual availability varies. If the original task still fails after resolution, test one small read‑only action on the same surface and then on one alternative supported surface. A failure limited to one client may point to a stale app session, local network control or extension. A failure tied to one repository may point to permissions or environment setup. A failure across multiple clients with the same account is stronger support evidence. OpenAI's [troubleshooting guide](https://help.openai.com/en/articles/7996703‑troubleshooting‑chatgpt‑error‑messages) recommends reloading the app, checking the status page and collecting timestamps and diagnostics when errors persist. Use those steps after preserving task state.

After resuming, verify the intended outcome directly. Review the diff, run the relevant test, inspect the created remote object and confirm that no duplicate action exists. Commit or otherwise preserve important work once it is sound. Record the task ID, affected surface, incident‑window timestamp, recovery action and final state. This produces a reusable incident note instead of relying on chat history alone. OpenAI's incident was shared infrastructure; the completion state remains task‑specific. Follow [OpenAI's OpenTools organization page](https://opentools.ai/organizations/openai) for future Codex and Work updates. *Figure: A Codex recovery path from the original task through local and external state checks to the smallest safe retry. Sources: [OpenAI Status](https://status.openai.com/incidents/jwna590w) and [Using Codex Cloud](https://help.openai.com/en/articles/20001545‑using‑codex‑cloud). Figure by OpenTools Team; no third‑party expressive material used.*