Usage data unavailable when streaming run is aborted via AbortSignal

[sonh1230]

Describe the bug

When a streaming run is aborted via AbortSignal, result.state.usage returns all zeros (inputTokens: 0, outputTokens: 0, totalTokens: 0), even though the model has already generated partial content and consumed tokens.

This happens because:

1. The SDK accumulates usage only when a response_done event arrives (added in PR feat: track token usage while streaming responses for openai models #750)
2. When signal.abort() is called, the SDK catches the AbortError internally in #runStreamLoop and returns silently — the response_done event never arrives
3. The consumer's for await loop ends normally with no error, and result.state.usage remains at zero

This makes it impossible to track token consumption for cancelled runs, which is important for billing and quota management in production applications.

Debug information

• Agents SDK version: v0.4.6+
• Runtime environment: Node.js 22

Repro steps

import { Agent, run } from '@openai/agents';

async function main() {
  const controller = new AbortController();
  const agent = new Agent({ name: 'Test', instructions: 'You are a helpful assistant.' });

  const result = await run(agent, 'Write a long essay about AI', {
    stream: true,
    signal: controller.signal,
  });

  let content = '';
  for await (const event of result) {
    if (event.type === 'raw_model_stream_event') {
      const data = event.data as any;
      if (data.type === 'output_text_delta' && data.delta) {
        content += data.delta;
        if (content.length > 100) {
          controller.abort();
        }
      }
    }
  }

  console.log(`Content length: ${content.length}`); // > 0
  console.log(`Usage: ${JSON.stringify(result.state?.usage)}`); // All zeros!
}

[seratch]

Thanks for taking the time to report this issue and you're right. However, the usage data comes at the end of streaming, so there is no way to attach the data to an aborted streaming result.

[sonh1230]

@seratch
Thanks for the quick response!

I understand that usage data arrives at the end of the stream. However, for production applications requiring per-turn billing and quota tracking, this is a significant gap—tokens are actually consumed and billed by the API, but the SDK reports zero, making accurate cost attribution impossible for cancelled runs.

Are there any plans to address this in a future release? Or, is there any alternative way to track the actual usage for cancelled requests at this moment?

[seratch]

I understand this is important, but we don't have immediate plans to improve it as it won't be done only by this SDK side changes. As a workaround, your app can estimate consumed tokens using the tiktoken library. The estimate for input tokens can be reasonable, while the one for output tokens is especially inaccurate. This is still only an estimate, so we don't plan to implement something like this as part of the SDK.

I am aware that this is not an ideal answer for your, but I hope this clarifies things at this moment.