Worked on the vercel/ai repository to enhance error handling in the MCP Client Tool, focusing on improving model visibility and reliability within tool workflows. Addressed a bug by implementing targeted error propagation, allowing raw error messages to reach the LLM directly when a tool returns an error, while maintaining strict output schema validation for successful responses. This approach reduced blind spots during tool failures and facilitated more effective debugging and recovery. The work involved updating local and CI validation processes, expanding test coverage with new cases, and improving release hygiene. Utilized TypeScript and full stack development skills with an emphasis on testing.