Guess it depends on the project or model, I haven't used it in an ide for a while now, since agents have been a thing. I have it set in it's "soul" to always keep backups, make sure it can restore the backups before making changes, or if i'm in a good stable place, I have it work on a duplicate rather than the main thing so I can make sure it isn't breaking anything. Every once in a while it fucks up but I just tell it to reverse/restore and it works out fine. Never had any issues with github. In the case of agents, telling it to plan, always ask for confirmation, work readonly on every task until the user allows writing, etc. helps a lot.

You just have to be ultra specific with instructions which feels counter intuitive to using llms or ai in general. Also helps to only use it for meaningless personal projects and not real work I guess lol.