I have found a different model should be used to do the review - like if Claude did the code, Codex should review. Models reviewing their own code is a recipe for disaster.