I’m building a client around the OpenAI Responses API and want to handle streamed output cleanly. The event stream can include text as well as tool-call activity, so I’m unsure which event my interface should treat as the definitive end of a response.

For anyone who has implemented Responses API streaming events: how do you tell the difference between a completed response, an incomplete one, and a response that is handing off to a tool? I’d especially like to hear how you avoid showing a “done” state too early.