Curious, which kind of model is used ? do y'all prefer using a lightweight model or a more capable one for code review ?
I would assume though advanced reasoning models would be best, especially since it's out of band. If I were building it maybe I'd run two adversarial models.