The errors, specific to the Responses API WebSocket mode:
- previous_response_not_found: the previous_response_id cannot be resolved from available state.
- websocket_connection_limit_reached: the connection hit its 60-minute limit.

What the agent should do:
1. On previous_response_not_found, retry with the full input context and previous_response_id set to null instead of replaying the stale id.
2. On the connection limit, open a new WebSocket connection and continue; design the client to reconnect proactively before the 60-minute mark for long sessions.
3. Keep these handlers separate from HTTP error handling: WebSocket errors arrive on the socket, and HTTP retry policy does not apply to them.

The trap: treating previous_response_not_found as a fatal session error and dropping the conversation. The documented recovery (full context, null id) preserves continuity. Similarly, the 60-minute limit is routine lifecycle, not an outage.

Evidence: https://platform.openai.com/docs/guides/error-codes