Guides / Codex rate limit recovery
Codex hit a rate limit. Your coding session does not have to die.
If Codex stops mid-task, first identify which limit you hit. A plan allowance, a credit balance, and an OpenAI API rate limit are different conditions and do not all clear by retrying. The working tree, changed files, recent commits, and local notes remain your evidence for the unfinished task.
OpenAI usage guidance checked September 23, 2026. Codex usage and limits ยท API rate-limit guidance
Tell a plan limit from an API rate limit
Codex plan or credit limit: OpenAI says to check Settings or the usage dashboard for the exhausted allowance, remaining credits, and any reset time. In an active Codex CLI session, /status shows usage information. Do not treat every plan-limit message as a transient API throttle.
OpenAI API 429: if a separate API-backed workflow returns a rate-limit response, follow the Retry-After header when provided, reduce request rate or volume, and use bounded exponential backoff with jitter when no delay is provided. An API 429 and a Codex plan allowance are not interchangeable.
Preserve the task before resuming
- Save open files and confirm the project directory and current branch.
- Check tracked and untracked changes; do not assume the interrupted agent committed its work.
- Review recent commits, the diff, local notes, and actual test output.
- Write down the next small action and distinguish verified facts from reconstructed intent.
For plans with shared allowances, OpenAI notes that usage can depend on model, task complexity, context, reasoning, speed, and tools. Check the current usage surface for the account rather than relying on a copied static quota.
Create a continuation packet
- Run ShardStitch from the dashboard, VS Code/Cursor extension, or CLI.
- Review the generated packet; treat disk-derived facts as evidence and inferred intent as something to confirm.
- Select Codex again after usage becomes available, or choose another supported target.
- Copy the packet into the new session or use the target's supported handoff file.
- Ask the receiving AI to inspect current repository state before it edits.
What ShardStitch can reconstruct
ShardStitch does not need the interrupted Codex session to summarize itself. It reads what survived locally: the working tree, Git state, recent commits, project notes, dependency shape, and available verification artifacts. That can ground a continuation packet, but it cannot recover a decision that existed only in the unavailable conversation.
What the next AI gets
- What changed in the repository.
- Which files are risky to touch.
- What the task was trying to accomplish.
- What should happen next.