Semantic Consistency Policy Optimization for Reinforcement Learning of LLM Agents