I’m building a voice feature with the OpenAI Realtime API, and I’m unsure what the client should do if the connection drops mid-conversation. Is there a supported way to resume the same session, or should I start a new one and send back the recent context?

I’m especially interested in avoiding duplicate audio or tool actions when reconnecting. If you’ve handled this in a real app, what state did you keep on the client, and what did you let the server own?