Within a realtime session, created with the turn_detection.interrupt_response set, the guardrail triggering fails and emit the following error event:

RealtimeErrorEvent(error=RealtimeError(message='Conversation already has an active response in progress: .... Wait until the response is finished before creating a new one.', type='invalid_request_error', code='conversation_already_has_active_response', event_id=None, param=None), event_id='event...', type='error')

My guess is that no response.cancel server event is sent during the interruption triggered by the guardrail, while it should always be.

In OpenAIRealtimeWebSocketModel._send_interrupt, response.cancel is triggered if turn_detection.interrupt_response is not set, and the guardrail process flow doesn't trigger a response.cancel event on its side.

Debug information

• Agents SDK version: 0.3.2
• Python version: Python 3.13

Repro steps

create an agent without setting a PlaybackTracker, and with turn_detection.interrupt_response set. Trigger a any guardrail.

Expected behavior

Guardrail interrupts and successfully send user input to the model with the tripped guardrail data.