neuron/packages/opencode/src/plugin/openai

OpenAI Responses WebSocket

Enabled by default on local, dev, and beta. On latest and prod, set OPENCODE_EXPERIMENTAL_WEBSOCKETS=true.

Flow

1. A streamed POST /responses request arrives.
2. If it has no session-id or x-session-affinity header, use HTTP.
3. Title requests use HTTP.
4. If that session's socket is busy or already in fallback mode, use HTTP.
5. Otherwise, reuse its open socket or open a new one.
6. Send response.create and return WebSocket events as SSE.

Lifetime

• Connect timeout: 15 seconds.
• Idle timeout: 5 minutes.
• After a completed response, keep the socket for reuse.
• Reuse a socket for up to 55 minutes, then replace it on the next request.

Retries

• If WebSocket setup fails or it fails before its first event, replay over HTTP and keep that session on HTTP until idle-pruned.
• If the server returns websocket_connection_limit_reached before output, reconnect up to 5 times, then follow the same HTTP fallback.
• If a WebSocket fails after its first event, fail the stream. Do not replay partial output.
• Abort or cancel closes the socket.

Next Steps

• previous_response_id continuation.
• Optional second WebSocket for concurrent requests in one session. Currently these use HTTP.