I suspect the actual problem here is that the instruction set provided to the model were not explicit enough about what "destructive and irreversible" actual meant. I use Claude Code all the time. There's an option that has the agent stream out its though process instead of just giving a blinking "thinking..." prompt while it goes about its business. I like to open those and follow along, especially early on in any given iteration. It's kind of wild all of the different directions it can go, the gaps it finds in what I thought were clear instructions, and the assumptions it sometimes makes to try and figure out what I meant. Claude is a junior developer. It's a very fast junior developer, but it also makes all those same junior developer mistakes really fast. It should NEVER have permissions to do anything more than you would let any other junior dev do.