Strategy-Augmented Planning for Large Language Models via Opponent Exploitation