Anyone ever seen these weird chat precaution halts? They keep repeating after compactions.

A previously hallucinated rule in an orchestration skill gated against spawning agents with any other model and reasoning effort than the one the parent uses. I instructed the model to change that skill and it patched it. After 3+ hours into the run, this starts repeating and locking up the process. It sends an involuntary message to the agent each time I accept. Anyone ever seen this before?