Keep a Codex Review Finding Reproducible
A review comment becomes useful when another person can reproduce the problem. A confident severity label alone does not establish that the code is wrong.
For each consequential Codex finding, ask for the triggering input, the affected path and the observed behavior. Check that the finding applies to the revision under review, then turn a confirmed defect into a regression test where practical.
Keep unsupported findings separate from verified ones. A reviewer may infer a caller contract incorrectly or overlook an existing validation step. Disagreeing with a comment is reasonable when the code supplies contrary evidence. Record that evidence in the review so the discussion teaches the next reviewer something more useful than “the agent was mistaken.”
