I’m building a client around the OpenAI Responses API and sometimes the connection times out before I receive a response. The request may have already completed on the server, though, so sending it again could mean duplicate work—or repeat a tool action with side effects.

How are you handling a timed-out Responses API request in production? I’m interested in practical retry patterns, especially for requests that can trigger an external action, and what signals you use to decide whether it’s safe to try again.