How to Enable Tracing For Realtime Runner?

[Ali-Olliek]

Please read this first

• Have you read the docs?Agents SDK docs
• Yes ✔️
• Have you searched for related issues? Others may have had similar requests
• Yes ✔️

Question

Describe your question. Provide details if available.

I was checking OpenAI's Dashboard to check the traces, and it shows that my Realtime Traces are not appearing on the dashboard.
I used to be able to see them with the proper traces up until 25th of September, I was using version 0.3.0 of the Library.

I am providing the following object to the Runner method

{
    "model_settings": {
    "voice": "ash",
    "turn_detection": "server_vad",
    "model_name": "gpt-realtime",
    "tracing": {
            "workflow_name": "Realtime Workflow",
            "group_id": "123",
            "metadata": {
                "site": "test.com
            }
        }
    }
}

I even set the default values of the tracing object in the DEFAULT_MODEL_SETTINGS at agents/realtime/openai_realtime.py, but I still could not see the traces in the dashboard.

I removed any set_tracing_export_api_key() and associated code from my codebase assuming it would force using the DEFAULT_MODEL_SETTINGS.

I would like to know how to force tracing for my Realtime flow and whether there are some prerequisites to tracing that I am not following.

Thank you

[seratch]

Hi, thanks for writing in.

I used to be able to see them with the proper traces up until 25th of September, I was using version 0.3.0 of the Library.

Does this mean you see tracing behavior difference with a newer version (0.3.1 - 0.3.3) and if we switch back to 0.3.0, it works without any issues?

[Ali-Olliek]

Hello @seratch, thanks for your reply

Does this mean you see tracing behavior difference with a newer version (0.3.1 - 0.3.3) and if we switch back to 0.3.0, it works without any issues?

No even on 0.3.0 It is not working anymore, I tried 0.3.0, 0.3.2 and 0.2.11 and I wasn't able to see the traces on the dashboard.

I tried to mimic my production setup and configuration using this small script:

"""
OpenAI Realtime Agent POC with Audio Playback

This demonstrates a working realtime agent that:
1. Sends a text message to the agent
2. Captures audio responses
3. Plays the audio through your speakers
"""

import asyncio
import os
import uuid
from agents.realtime import RealtimeAgent, RealtimeRunner
import pyaudio
import queue


class AudioPlayer:
    """Simple audio player for PCM16 audio"""

    def __init__(self, sample_rate=24000, channels=1):
        self.sample_rate = sample_rate
        self.channels = channels
        self.audio_queue = queue.Queue()
        self.p = pyaudio.PyAudio()
        self.stream = None
        self.playing = False

    def start(self):
        """Start the audio playback stream"""
        self.stream = self.p.open(
            format=pyaudio.paInt16,
            channels=self.channels,
            rate=self.sample_rate,
            output=True,
            frames_per_buffer=1024,
            stream_callback=self._callback,
        )
        self.playing = True
        self.stream.start_stream()
        print("🔊 Audio player started")

    def _callback(self, in_data, frame_count, time_info, status):
        """Callback for audio playback"""
        try:
            data = self.audio_queue.get_nowait()
            return (data, pyaudio.paContinue)
        except queue.Empty:
            return (bytes(frame_count * 2), pyaudio.paContinue)

    def add_audio(self, audio_bytes):
        """Add audio data to playback queue"""
        if self.playing:
            self.audio_queue.put(audio_bytes)

    def stop(self):
        """Stop the audio player"""
        self.playing = False
        if self.stream:
            self.stream.stop_stream()
            self.stream.close()
        self.p.terminate()
        print("🔇 Audio player stopped")


async def main():
    api_key = os.getenv("OPENAI_API_KEY")
    if not api_key:
        print("❌ Error: OPENAI_API_KEY environment variable not set")
        return

    # Create audio player
    audio_player = AudioPlayer()
    audio_player.start()

    # Create the agent
    agent = RealtimeAgent(
        name="Assistant",
        instructions="You are a helpful voice assistant. Keep responses brief and conversational.",
    )

    # Set up the runner with configuration
    runner = RealtimeRunner(
        starting_agent=agent,
        config={
            "model_settings": {
                "model_name": "gpt-realtime",
                "voice": "ash",
                "modalities": ["audio"],
                "input_audio_format": "pcm16",
                "output_audio_format": "pcm16",
                "input_audio_transcription": {"model": "gpt-4o-mini-transcribe"},
                "turn_detection": {"type": "semantic_vad", "interrupt_response": True},
                "tracing": {
                    "workflow_name": "Realtime Agent",
                    "group_id": str(uuid.uuid4()),
                    "metadata": {
                        "hello": "test"
                    }
                }
            }
        },
    )

    # Start the session
    session = await runner.run(model_config={"api_key": api_key})

    async with session:
        print("=" * 60)
        print("🎙️  Realtime Agent Session Started")
        print("=" * 60)
        print()

        # Send a text message to trigger response
        print("📤 Sending message: 'Hello, how are you?'")
        await session.send_message("Hello, how are you?")

        audio_chunks_received = 0

        # Process events
        async for event in session:
            try:
                if event.type == "agent_start":
                    print(f"🤖 Agent started: {event.agent.name}")

                elif event.type == "agent_end":
                    print(f"✅ Agent ended: {event.agent.name}")

                elif event.type == "audio":
                    # Play the audio chunk
                    audio_chunks_received += 1
                    if audio_chunks_received == 1:
                        print("🎵 Receiving audio...")
                    audio_player.add_audio(event.audio.data)

                elif event.type == "audio_end":
                    print(f"🎵 Audio ended (received {audio_chunks_received} chunks)")
                    # Wait a bit for audio to finish playing
                    await asyncio.sleep(2)
                    break

                elif event.type == "audio_interrupted":
                    print("⚠️  Audio interrupted")

                elif event.type == "error":
                    print(f"❌ Error: {event.error}")

                elif event.type == "response_text_delta":
                    # Print text as it comes in
                    print(f"💬 {event.delta}", end="", flush=True)

                elif event.type == "response_text_done":
                    print()  # New line after text is done

                elif event.type in ["history_updated", "history_added"]:
                    pass  # Skip these frequent events

                elif event.type == "raw_model_event":
                    pass

            except Exception as e:
                print(f"❌ Error processing event: {e}")

    # Cleanup
    audio_player.stop()
    print()
    print("=" * 60)
    print("Session ended successfully!")
    print("=" * 60)

if __name__ == "__main__":
    try:
        asyncio.run(main())
    except KeyboardInterrupt:
        print("\n\n⚠️  Session interrupted by user")
    except Exception as e:
        print(f"\n\n❌ Fatal error: {e}")

I used the correct API key for the project, and I expected to see the traces which did not appear either.

[alexanderfilipzik]

+1

[hauvinenai-rgb]

+1

[kucharzyk-sebastian]

Same here. It feels like the issue started happening when the library moved to gpt-realtime as the default model.

[Ali-Olliek]

@kucharzyk-sebastian For me the last traces appeared on the 25th of September.
The gpt-realtime update I remember was on the 12th of September, plus I was able to see gpt-realtime for the final traces as well, so I am not sure the issue is the gpt-realtime being the default model.

If we can all check when the last traces had occurred (on September 25th - like me - for example) maybe we can help the team identify the issue faster.

[kucharzyk-sebastian]

@Ali-Olliek - one of my agents stopped reporting traces on the 25th of September and the other on 7th of October. Both of these were upgraded from openai-agents 0.2.9 to 0.3.0 and openai 1.101.0 to 1.107.0 around the time the traces stopped appearing. I hope that helps the team.

[kucharzyk-sebastian]

Hi @seratch . This issue affects production environments and effectively prevents from debugging and benchmarking. Could you share information on when we can expect a fix? I'd really appreciate that since we need either have that fixed or switch to a different tracing provider.

[yb90200]

+1