buffered_summary_chat：Testing does not work as expected after connecting to LLM

The actual performance is to continue to generate answers based on historical conversations rather than summarizing them