🐛 Bug Report: OpenAI instrumentation drops gen_ai.usage.reasoning_tokens on streamed chat completions

未关闭 适合新手
#4,660 1 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

维护者通常 1 天内回复

@jayfeng7 已经在做这个了。

开始于 2026年10月27日。

  • #4662 来自 @jayfeng7 —— 未关闭

评估

难度
2/5
预计耗时
1-3 小时
新手友好度
82/100
Issue 类型
缺陷
描述清晰度
描述清楚
活跃度
活跃
技术栈
python
领域
observability

调研方向

阅读 shared/chat_wrappers.py,特别是 _handle_response 和 _set_streaming_token_metrics,然后检查 tests/traces/test_chat.py 和 test_chat_reasoning。为包含 completion_tokens_details.reasoning_tokens 的 usage chunk 添加流式测试;当流式 span 除 token 指标外还记录 gen_ai.usage.reasoning_tokens 时,即表示完成。issue 作者表示愿意提交 PR,但没有报告负责人或关联的 pull request。

由索引模型根据 Issue 内容生成。

描述

Which component is this bug for?

OpenAI Instrumentation

📜 Description

A streamed chat completion never gets gen_ai.usage.reasoning_tokens, even when its usage chunk carries completion_tokens_details.reasoning_tokens. Without stream=True the span records it.

Only _handle_response sets the attribute (shared/chat_wrappers.py:334-354 at 8d9c93f), since #3336. Every stream path, async included, goes through _set_streaming_token_metrics (chat_wrappers.py:615), which sets input and output tokens and nothing else. The semantic-conventions page defines it as the reasoning part of completion_tokens, with no exception for streams. The only test, test_chat_reasoning (tests/traces/test_chat.py:1449), makes a non-stream call.

👟 Reproduction steps
  1. At 8d9c93f: pip install -e packages/opentelemetry-semantic-conventions-ai -e packages/opentelemetry-instrumentation-openai openai==3.35.1 opentelemetry-sdk.
  2. Run this (a mock transport, no network):
import json

import httpx
import openai
from opentelemetry.instrumentation.openai import OpenAIInstrumentor
from opentelemetry.sdk.trace import TracerProvider
from opentelemetry.sdk.trace.export import SimpleSpanProcessor
from opentelemetry.sdk.trace.export.in_memory_span_exporter import InMemorySpanExporter

exporter = InMemorySpanExporter()
provider = TracerProvider()
provider.add_span_processor(SimpleSpanProcessor(exporter))
OpenAIInstrumentor().instrument(tracer_provider=provider)

USAGE = {"prompt_tokens": 12, "completion_tokens": 300, "total_tokens": 312,
         "completion_tokens_details": {"reasoning_tokens": 256}}
BASE = {"id": "c1", "created": 0, "model": "o4-mini"}


def handler(request):
    if json.loads(request.content).get("stream"):
        chunks = [
            BASE | {"object": "chat.completion.chunk", "choices": [{"index": 0, "delta": {
                "role": "assistant", "content": "Paris."}, "finish_reason": "stop"}]},
            BASE | {"object": "chat.completion.chunk", "choices": [], "usage": USAGE},
        ]
        body = "".join(f"data: {json.dumps(c)}\n\n" for c in chunks) + "data: [DONE]\n\n"
        return httpx.Response(200, content=body.encode(),
                              headers={"content-type": "text/event-stream"})
    return httpx.Response(200, json=BASE | {"object": "chat.completion", "usage": USAGE,
        "choices": [{"index": 0, "message": {"role": "assistant", "content": "Paris."},
                     "finish_reason": "stop"}]})


client = openai.OpenAI(api_key="x", http_client=httpx.Client(transport=httpx.MockTransport(handler)))
for stream in (False, True):
    kw = {"stream_options": {"include_usage": True}} if stream else {}
    resp = client.chat.completions.create(model="o4-mini", stream=stream,
                                          messages=[{"role": "user", "content": "Capital of France?"}], **kw)
    if stream:
        for _ in resp:
            pass
    a = exporter.get_finished_spans()[-1].attributes
    print(f"stream={stream!s:<5} output_tokens={a.get('gen_ai.usage.output_tokens')} "
          f"reasoning_tokens={a.get('gen_ai.usage.reasoning_tokens')}")
    exporter.clear()
👍 Expected behavior

Both spans carry gen_ai.usage.reasoning_tokens=256.

👎 Actual Behavior with Screenshots
$ python repro_openai.py
stream=False output_tokens=300 reasoning_tokens=256
stream=True  output_tokens=300 reasoning_tokens=None
🤖 Python Version

3.12

📃 Provide any additional context for the Bug.

I ran the sync stream only, and read the async path from the code.

👀 Have you spent some time to check if this bug has been raised before?
  • I checked and didn't find similar issue
    • Searched open and closed issues and pull requests for reasoning_tokens and reasoning streaming. #4458, #4461 and #4462 cover Anthropic, #4488 covers Groq. No open pull request adds reasoning to chat_wrappers.py.
Are you willing to submit PR?

Yes I am willing to submit a PR!

@doronkopit5, I'd move the reasoning block into a helper that _handle_response and _set_streaming_token_metrics both call, so every stream path gets it, plus a streamed twin of test_chat_reasoning. Shall I open it?

主要语言
Python
星标
7.5k
派生
1.2k
平均合并
2 天 6 小时
30 天内合并 PR
2

环境准备

从这里开始

  1. 先读完整个 Issue,再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库,在一个分支上完成修改。
  4. 提交 Pull Request,并在描述里引用这个 Issue 编号。

traceloop/openllmetry 的其他 Issue

查看 traceloop/openllmetry 的全部 Issue

相似的 Issue

更多 Python Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。