I’m working on a recursive approach to migrating large scale legacy systems with CLI AI tools. For months, I’ve been stuck with only being able to trust the output of simpler functions, but Codex and agent building is really opening the door on the possible routes forward. I’m curious if anyone might be up to share notes and discuss strategies. Before, I wouldn’t have trusted more than 400 lines (code) at a time - yes these still need to be reviewed. But, CLI AIs changed this - especially for well-structured code and goals. Suddenly 4k lines is much more achievable. Today I finished a round of experiments for a legacy 10k line long function. I even accidentally missed some base layer data i/o methods, but AI got almost entirely there. So, while I realize I am likely still limited from fully automating a system conversion, I think the designs that are possible today for simpler functions will work well once models improve - 6-12 months from now.