I find that when an LLM jumps into tasks it was not told to do (or even worse, doing things it was explicitly told not to), it is a good sign the context is too full, and you should do a controlled hand-off to a new instance.