When Codex reaches a usage limit: correcting an older GPT-5.1-Codex-Max guide

A Codex usage limit does not establish code quality or explain why the allowance was consumed. Record the actual limit message and reset time, then preserve the diff and unfinished work.

Separate a limit from an execution failure

The API catalog and a ChatGPT subscription allowance are different. The OpenAI model catalog does not show your Codex balance. Do not infer fixed request counts or a guaranteed next-day reset.

Write a resumption note

Save the branch, changed files, successful checks, failed checks, and one next action.

Goal: explain CSV import errors
Changed: parser and error display
Verified: empty-file test
Unfinished: invalid dates
Next: add invalid-date input and expected error

This is an illustrative note. Do not record tests you did not run or carry secret values into the next session.

Fix the CSV expectations before resuming

For the note above, prepare separate inputs for an empty row, a missing header, and a row with an invalid date. Define outcomes such as skipping empty rows, stopping an import with no header, and reporting the line number of an invalid date. These are illustrative acceptance criteria, not tests reported as executed.

Reuse the inputs after switching models to check that resuming did not lose an earlier requirement.

Check again after resuming

Read the saved diff before assigning the unfinished action. When switching models or settings, compare the same input against the same expected result. Measure progress by verified behavior rather than by exhausting an allowance.

Correction record

The previous version linked exhausting GPT-5.1-Codex-Max limits to high quality and asserted development-time savings and reliable stock-analysis logic. Those claims are withdrawn because no reproducible comparison or dataset is supplied. The historical URL remains, but this is not a promise that the old model is available today.