Worked on the agentscope-ai/agentscope repository to enhance the stability of Gemini Chat streaming by addressing a crash caused by missing usage metadata in streaming responses. Applied defensive programming techniques in Python to implement robust error handling, introducing a safe extraction method that conditionally creates ChatUsage objects only when token counts are present. Focused on backend development and API integration, the solution improved the reliability of streaming by preventing failures when usage data is absent or contains None values. Changes were validated through targeted tests and code reviews, reducing the risk of regressions and improving maintainability in streaming data handling.