Realtime: close() keeps the previous connection's item and audio state, so a reconnect on the same model truncates and retrieves stale item ids

Author: kosesenaCreated Sep 17, 2026Updated Sep 17, 2026

Please read this first

Have you read the docs? Yes.
Have you searched for related issues? Yes. #4189 covers cleanup after a failed connection attempt and #4461 covers ending iteration after a clean server close; neither covers state that survives a successful close into the next connection.

Describe the bug

OpenAIRealtimeWebSocketModel.close() clears the response-scoped audio indexes but leaves three per-connection values in place: _current_item_id, the ModelAudioTracker (its last audio item and per-item states), and _created_session. connect() accepts a new connection on the same instance after close(), and RealtimeRunner reuses one model instance across runner.run() calls, so the second session starts with the first session's item ids and turn-detection settings.

On that second connection, before any assistant audio has been produced:

the first input_audio_buffer.speech_started emits RealtimeModelAudioInterruptedEvent(item_id="<old item>") and sends conversation.item.truncate for an item id the new server session has never seen;
the first conversation.item.input_audio_transcription.completed sends conversation.item.retrieve for that same old item id;
RealtimeModelSendInterrupt does the same as the first bullet;
until the new session.created arrives, _send_interrupt decides whether to cancel the response from the previous session's turn_detection.interrupt_response.

The server answers the foreign-id messages with error events, which reach the application as RealtimeError, and the application also receives an audio_interrupted for an item it cannot map to anything in the current session.

Debug information

Agents SDK version: main at 9f6f6b10 (also present in v0.22.2)
Python version: 3.12.13
Operating system: macOS 15
Model and model provider: OpenAIRealtimeWebSocketModel (OpenAI Realtime API over WebSocket)
Does the issue reproduce with the latest Agents SDK release? Yes
Does the issue occur consistently or intermittently? Consistently

Repro steps

Self-contained, no network: the WebSocket factory is replaced with a recorder.

python

import asyncio, base64, json
from agents.realtime.model_events import RealtimeModelAudioInterruptedEvent
from agents.realtime.openai_realtime import OpenAIRealtimeWebSocketModel


class RecordingWebSocket:
    def __init__(self):
        self._closed = asyncio.Event()
        self.sent = []

    def __aiter__(self):
        return self

    async def __anext__(self):
        await self._closed.wait()
        raise StopAsyncIteration

    async def send(self, payload):
        self.sent.append(json.loads(payload))

    async def close(self):
        self._closed.set()


SESSION_CREATED = {
    "type": "session.created", "event_id": "ev", "session": {
        "type": "realtime", "model": "gpt-realtime",
        "audio": {"input": {"turn_detection": {"type": "semantic_vad", "interrupt_response": True}},
                  "output": {"format": {"type": "audio/pcm", "rate": 24000}}},
    },
}


async def main():
    model = OpenAIRealtimeWebSocketModel()
    events = []

    class Listener:
        async def on_event(self, event):
            events.append(event)

    model.add_listener(Listener())
    sockets = []

    async def fake_connect(*args, **kwargs):
        sockets.append(RecordingWebSocket())
        return sockets[-1]

    model._create_websocket_connection = fake_connect

    # First session: one assistant audio item, then close.
    await model.connect({"api_key": "test", "initial_model_settings": {}})
    await model._handle_ws_event(SESSION_CREATED)
    await model._handle_ws_event({
        "type": "response.output_audio.delta", "event_id": "e1", "response_id": "resp_old",
        "item_id": "item_old", "output_index": 0, "content_index": 0,
        "delta": base64.b64encode(b"\x00\x01" * 2400).decode(),
    })
    await model._handle_ws_event({"type": "response.done", "event_id": "e2", "response": {"id": "resp_old"}})
    await model.close()
    events.clear()

    # Second session on the same instance: user starts speaking before any assistant audio.
    await model.connect({"api_key": "test", "initial_model_settings": {}})
    await model._handle_ws_event(SESSION_CREATED)
    await model._handle_ws_event({"type": "input_audio_buffer.speech_started", "event_id": "e3",
                                  "audio_start_ms": 0, "item_id": "item_user_new"})
    await model.close()

    print([e for e in events if isinstance(e, RealtimeModelAudioInterruptedEvent)])
    print([m for m in sockets[1].sent if m["type"] == "conversation.item.truncate"])


asyncio.run(main())

Output on main:

[RealtimeModelAudioInterruptedEvent(item_id='item_old', content_index=0, type='audio_interrupted')]
[{'audio_end_ms': 0, 'content_index': 0, 'item_id': 'item_old', 'type': 'conversation.item.truncate'}]

Expected behavior

Both lists are empty. A connection opened after close() should start with no item, audio, or session state from the previous connection, the same way close() already clears the response audio indexes. I have a fix with regression tests ready and will open a PR referencing this issue.

Source: openai/openai-agents-python

View original on GitHub View discussion on GitHub