Use json.dumps with ensure_ascii=False to avoid converting Chinese descriptions into Unicode escape sequences. This makes logs easier to read and significantly improves LLM tool selection accuracy (from ~48% to ~67%) by preserving the original natural language context.