I built a little terminal tool because Codex kept getting away from me

I've been building a little terminal tool called Savepoint because I kept losing the plot once Codex got beyond small changes. I'd ask for one thing, Codex would touch a bunch more files, run the tests and tell me everything looked good. And I'd be sitting there thinking... probably? I can tell whether the feature works. I'm much less confident deciding whether every extra change was sensible, necessary, or whether I've let a very confident robot renovate the kitchen because I asked it to fix a tap. So I've started keeping jobs smaller and putting a stop between chunks of work. Sometimes I use a fresh session to check what got built rather than asking the thing that wrote it whether it did a good job. Savepoint is basically me turning that habit into a small terminal workflow. Still not sure whether this is useful discipline or an elaborate coping mechanism because AI let me build software faster than I learned how software works. How are people here handling this with bigger Codex changes? /review ? Fresh session? Another model? Just inspect the diff and hope for the best? Repo if anyone wants to poke at it: github.com/anipatke/savepoint preview.redd.it/jfzcuwefqith1.png