I’m building a voice app with the OpenAI Realtime API and am trying to work out what a good recovery flow looks like when the connection drops mid-conversation.

If the client reconnects, should it resend any audio the server may not have received, or could that make the user’s speech appear twice? I’m also unsure how to handle an assistant response that was streaming when the connection failed. How are others tracking conversation state after reconnect and deciding what, if anything, to replay?