A developer enables thinking on a tool-using agent. After Claude returns a response containing a thinking block and a tool_use block, the developer's code runs the tool and builds the next request. To save tokens, the code keeps only the tool_use block in the assistant message. Reasoning quality then drops, and in some configurations the API returns an error. What is the correct way to send the tool result back?