Responses API streaming crashes with 'AsyncStream' object has no attribute 'id' when using .with_raw_response

Which component is this bug for?

OpenAI Instrumentation

Description

client.responses.with_raw_response.create(stream=True, ...) crashes async_responses_get_or_create_wrapper with:

AttributeError: 'AsyncStream' object has no attribute 'id'

This is a variant of #4058 / PR #4078 that the fix didn't cover. #4078 taught parse_response/async_parse_response to unwrap APIResponse/AsyncAPIResponse (used by .with_streaming_response), but .with_raw_response.create() on Responses returns a LegacyAPIResponse (via openai._legacy_response.async_to_raw_response_wrapper), which was already unwrapped before #4078. The problem isn't the unwrapping itself — it's that LegacyAPIResponse.parse(), when the underlying request was stream=True, returns a Stream/AsyncStream instance, not a parsed Response.

The only isinstance(response, (Stream, AsyncStream)) check in async_responses_get_or_create_wrapper runs before parsing, on the raw, unparsed return of wrapped():

https://github.com/traceloop/openllmetry/blob/main/packages/opentelemetry-instrumentation-openai/opentelemetry/instrumentation/openai/v1/responses_wrappers.py#L666

When the call goes through .with_raw_response, that raw return is a LegacyAPIResponse wrapper, not a Stream/AsyncStream, so the check misses it. Execution falls through, async_parse_response() calls .parse() on it, which (per the self._stream branch in openai._legacy_response.LegacyAPIResponse._parse) returns an AsyncStream. There's no post-parse recheck for Stream/AsyncStream, so it crashes reading .id:

https://github.com/traceloop/openllmetry/blob/main/packages/opentelemetry-instrumentation-openai/opentelemetry/instrumentation/openai/v1/responses_wrappers.py#L746

This combination (.with_raw_response + stream=True) is exactly what agent_framework_openai (Microsoft Agent Framework) uses as of >=1.6.0, to read the x-ms-served-model response header — so every agentic streaming turn instrumented with this package crashes.

Reproduction steps

Self-contained (mocks the HTTP transport, no API key needed):

import asyncio
import httpx
from opentelemetry.instrumentation.openai import OpenAIInstrumentor
from opentelemetry.sdk.trace import TracerProvider
from opentelemetry import trace

trace.set_tracer_provider(TracerProvider())
OpenAIInstrumentor().instrument()

from openai import AsyncOpenAI  # noqa: E402

SSE_BODY = (
    b'event: response.created\n'
    b'data: {"type":"response.created","response":{"id":"resp_123","object":"response",'
    b'"created_at":0,"status":"in_progress","model":"gpt-4o-mini","output":[]}}\n\n'
    b'event: response.completed\n'
    b'data: {"type":"response.completed","response":{"id":"resp_123","object":"response",'
    b'"created_at":0,"status":"completed","model":"gpt-4o-mini","output":[],'
    b'"usage":{"input_tokens":1,"output_tokens":1,"total_tokens":2}}}\n\n'
)

def handler(request: httpx.Request) -> httpx.Response:
    return httpx.Response(
        200,
        headers={"content-type": "text/event-stream", "x-ms-served-model": "gpt-4o-mini"},
        content=SSE_BODY,
    )

async def main():
    client = AsyncOpenAI(api_key="sk-test", http_client=httpx.AsyncClient(transport=httpx.MockTransport(handler)))
    raw = await client.responses.with_raw_response.create(
        model="gpt-4o-mini",
        input="hello",
        stream=True,
    )
    stream = raw.parse()
    async for event in stream:
        pass
    print("no crash")

asyncio.run(main())

Expected behavior

No crash — either trace the stream normally (matching the non-with_raw_response path) or skip tracing gracefully.

Actual Behavior

Reproduced on both:

• opentelemetry-instrumentation-openai==0.62.3 + openai==2.50.0
• opentelemetry-instrumentation-openai==0.62.3 + openai==3.14.1 (latest)

File ".../opentelemetry/instrumentation/openai/v1/responses_wrappers.py", line 746, in async_responses_get_or_create_wrapper
    existing_data = responses.get(parsed_response.id)
                                  ^^^^^^^^^^^^^^^^^^
AttributeError: 'AsyncStream' object has no attribute 'id'

Python Version

3.12

Additional context

We're currently working around this in our own codebase by unwrapping just the Responses API span instrumentation after OpenAIInstrumentor().instrument() runs.

Are you willing to submit PR?

Yes I am willing to submit a PR if someone can confirm my assessment of the issue.