Interesting idea but I found out that AI is pretty inconsistent on how it structures or breakdowns commits, highlighy dependand on the model and/or user prompting. So there is a chance you might only get to replay single move or commits that are large enough to make it harder to know whats happening
Fair point — it depends heavily on how the agent commits. I have this in my CLAUDE.md:
10. *Commit frequently* — Many small commits over few large ones
Without it, yeah, you can get one giant commit and the replay adds nothing.