I have a similar workflow but I disagree with Codex/GPT-5.4 reviews being very useful. For example, in a lot of cases they suggest over-engineering by handling edge cases that won't realistically happen.