Realtime API response.done events missing detailed usage fields

[ginzlabs]

response.done events from the Realtime API do not include the full token usage breakdown shown in the Playground and documentation. Only total input and output counts are returned.

Currently, the event contains only:

"usage": {
  "input_tokens": 132,
  "output_tokens": 121
}

But the Playground/docs show a full structure:

"usage": {
  "total_tokens": 253,
  "input_tokens": 132,
  "output_tokens": 121,
  "input_token_details": {
    "text_tokens": 119,
    "audio_tokens": 13,
    "image_tokens": 0,
    "cached_tokens": 64,
    "cached_tokens_details": {
      "text_tokens": 64,
      "audio_tokens": 0,
      "image_tokens": 0
    }
  },
  "output_token_details": {
    "text_tokens": 30,
    "audio_tokens": 91
  }
}

Expected behavior: response.done should return the complete usage object as documented.

[ginzlabs]

I tracked down the root cause — the problem is in packages/agents-realtime/openaiRealtimeEvents.ts with the Zod schema validation.

The schema is currently expecting input_tokens_details / output_tokens_details (plural tokens), but the Realtime WebSocket API actually returns input_token_details / output_token_details (singular token).

Because of this mismatch, the validation fails and strips out the detailed usage fields, which is why only the minimal input_tokens / output_tokens values are coming through in response.done.

[leecjohnny]

I also face this issue, as a workaround, I'm reading the events from the webrtc datachannel directly.