🐛 Bug Report: OpenAI Agents: with two concurrent runs of the same agent, a handoff span is recorded in the other run's trace

未关闭
#4,635 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

@rupak-eng 已经在做这个了。

开始于 2026年10月19日。

  • #4639 来自 @rupak-eng —— 未关闭

评估

难度
4/5
预计耗时
3-5 天
新手友好度
30/100
Issue 类型
缺陷
描述清晰度
描述清楚
活跃度
停滞
技术栈
python

调研方向

Start in the OpenAI Agents SDK instrumentation, especially _start_handoff_span() and _find_agent_span(from_agent), then reproduce the issue with the supplied concurrent run script. Check how each run's trace context is tracked and how the handoff target is named; done when the handoff span appears under run B's assistant.agent span in trace 2. PR #4639 is already open against this issue.

由索引模型根据 Issue 内容生成。

描述

Which component is this bug for?

OpenAI Instrumentation

📜 Description

Package: opentelemetry-instrumentation-openai-agents (OpenAI Agents SDK instrumentation; it is not in the component list).

When an app runs the same agent twice at the same time (for example a server handling two requests), the handoff span of one run is recorded in the trace of the other run, and the run that handed off has no handoff span in its own trace.

_start_handoff_span() takes the parent from _find_agent_span(from_agent), which searches the open agent spans of all runs by agent name and returns the first match, not the agent span of the handoff's own trace. The handoff span is also named assistant → unknown.handoff, since to_agent is not known yet when the span starts.

👟 Reproduction steps
# OpenAI Agents SDK: two concurrent runs of the same agent (two requests to one app). The handoff span of run B is
# recorded in run A's trace, and run B's trace has no handoff span.
import asyncio, json, sys
from importlib.metadata import version
import openai
try:
    import httpx2 as httpx                         # openai >= 3
except ImportError:
    import httpx
from opentelemetry.sdk.trace import SpanProcessor, TracerProvider
from agents import Agent, Runner, function_tool, set_default_openai_client, set_trace_processors
from opentelemetry.instrumentation.openai_agents import OpenAIAgentsInstrumentor

class Spans(SpanProcessor):
    def __init__(self): self.ended = []
    def on_end(self, span): self.ended.append(span)

spans = Spans(); provider = TracerProvider(); provider.add_span_processor(spans)

SCRIPTS = {"A": ["tool", "text"], "B": ["tool", "handoff", "text"]}    # only run B hands off
TURN = {}
def respond(request):                              # a local stand-in for the Responses API, one script per run
    body = json.loads(request.content)
    first_user = next(i for i in body["input"] if i.get("role") == "user")
    run = first_user["content"].split()[-1]                           # the user asks "question A" or "question B"
    n = TURN[run] = TURN.get(run, 0) + 1
    step = SCRIPTS[run][n - 1] if n <= len(SCRIPTS[run]) else "text"
    item = {"tool": {"type": "function_call", "id": f"fc_{run}{n}", "call_id": f"call_{run}{n}", "name": "lookup",
                     "arguments": json.dumps({"q": "weather"}), "status": "completed"},
            "handoff": {"type": "function_call", "id": f"fc_{run}{n}", "call_id": f"call_{run}{n}",
                        "name": "transfer_to_specialist", "arguments": "{}", "status": "completed"},
            "text": {"type": "message", "id": f"msg_{run}{n}", "status": "completed", "role": "assistant",
                     "content": [{"type": "output_text", "annotations": [], "text": f"answer for {run}"}]}}[step]
    resp = {"id": f"resp_{run}{n}", "object": "response", "created_at": 0, "status": "completed", "model": "m",
            "parallel_tool_calls": True, "tool_choice": "auto", "tools": [], "output": [item],
            "usage": {"input_tokens": 3, "output_tokens": 2, "total_tokens": 5,
                      "input_tokens_details": {"cached_tokens": 0}, "output_tokens_details": {"reasoning_tokens": 0}}}
    events = [{"type": "response.created", "response": dict(resp, status="in_progress", output=[])},
              {"type": "response.output_item.added", "output_index": 0,
               "item": dict(item, status="in_progress", **({"content": []} if item["type"] == "message" else {"arguments": ""}))},
              {"type": "response.output_item.done", "output_index": 0, "item": item},
              {"type": "response.completed", "response": resp}]
    text = "".join(f"event: {e['type']}\ndata: {json.dumps(dict(e, sequence_number=i))}\n\n" for i, e in enumerate(events))
    return httpx.Response(200, headers={"content-type": "text/event-stream"}, content=text.encode())

set_trace_processors([])                           # no export to OpenAI
set_default_openai_client(openai.AsyncOpenAI(api_key="x", base_url="http://mock/v1", max_retries=0,
                          http_client=httpx.AsyncClient(transport=httpx.MockTransport(respond))), use_for_tracing=False)

@function_tool
def lookup(q: str) -> str:
    """Look something up."""
    return "sunny"

specialist = Agent(name="specialist", instructions="Answer in detail.", model="m")
assistant = Agent(name="assistant", instructions="Answer briefly.", model="m", tools=[lookup], handoffs=[specialist])

async def main():
    print(f"opentelemetry-instrumentation-openai-agents {version('opentelemetry-instrumentation-openai-agents')} | "
          f"openai-agents {version('openai-agents')} | python {sys.version.split()[0]}")
    OpenAIAgentsInstrumentor().instrument(tracer_provider=provider)
    run_a = Runner.run_streamed(assistant, "question A")      # run A: a tool call, then an answer
    run_b = Runner.run_streamed(assistant, "question B")      # run B: a tool call, a handoff to the specialist, an answer
    async for _ in run_a.stream_events():         # both runs are in flight; the app reads A, then B
        pass
    async for _ in run_b.stream_events():
        pass
    traces = {}
    for s in sorted(spans.ended, key=lambda s: s.start_time):
        traces.setdefault(s.context.trace_id, []).append(s.name)
    for i, names in enumerate(traces.values(), 1):
        print(f"trace {i}: {names}")

asyncio.run(main())
👍 Expected behavior

Each run's spans are in its own trace: the handoff span is in trace 2, under run B's assistant.agent span.

👎 Actual Behavior with Screenshots
opentelemetry-instrumentation-openai-agents 0.62.4 | openai-agents 0.24.0 | python 3.13.7
trace 1: ['Agent Workflow', 'assistant.agent', 'openai.response', 'lookup.tool', 'openai.response', 'assistant → unknown.handoff']
trace 2: ['Agent Workflow', 'assistant.agent', 'openai.response', 'lookup.tool', 'openai.response', 'specialist.agent', 'openai.response']
🤖 Python Version

3.13 (also seen with 3.12)

📃 Provide any additional context for the Bug.

Trace 1 is run A, which calls the tool and answers; it never hands off. Trace 2 is run B, which calls the tool, hands
off to specialist and answers. opentelemetry-instrumentation-openai-agents 0.62.4 (the latest release; main has no changes to this package since), openai-agents 0.24.0 (also 0.23.1). Related: PR #4097.

👀 Have you spent some time to check if this bug has been raised before?
  • I checked and didn't find similar issue
Are you willing to submit PR?

None

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。