The rabbit hole is real once the loop goes from prompt to tool call to file diff to replan and suddenly your shell feels obsolete. The part that bites later is not model quality, it's control surfaces like checkpoints, review gates, and a clean event log so you can see why it touched a file. Curious if your main loop is planner plus executor or just one model doing the whole thing.