Python: FoundryChatClient crashes with AsyncStreamWrapper has no attribute 'headers' when azure-ai-projects experimental tracing is enabled

[EmilienMottet]

Describe the bug

After upgrading from agent-framework-openai==1.5.0 to 1.6.0, every streaming LLM call through FoundryChatClient crashes with:

agent_framework.exceptions.ChatClientException:
  ("<class 'agent_framework_foundry._chat_client.FoundryChatClient'>
    service failed to complete the prompt: 'AsyncStreamWrapper' object has no attribute 'headers'",
   AttributeError("'AsyncStreamWrapper' object has no attribute 'headers'"))

This breaks every workflow and every chat reply for any application running agent-framework>=1.6.0 together with azure-ai-projects experimental GenAI tracing enabled (AZURE_EXPERIMENTAL_ENABLE_GENAI_TRACING=true), which is a documented opt-in for end-to-end OTel spans.

Root cause

PR #5910 (commit 1b6f7d80) introduced four served_model = self._extract_served_model(raw_*_response.headers) reads inside _stream()/_get_response() to surface the x-ms-served-model header:

• python/packages/openai/agent_framework_openai/_chat_client.py:639 – raw_stream_response.headers
• python/packages/openai/agent_framework_openai/_chat_client.py:680 – raw_create_response.headers
• python/packages/openai/agent_framework_openai/_chat_client.py:709 – raw_response.headers (background poll)
• python/packages/openai/agent_framework_openai/_chat_client.py:731 – raw_response.headers (non-streaming)

When azure-ai-projects>=2.1.0 experimental tracing is active, azure/ai/projects/telemetry/_responses_instrumentor._async_wrapped_responses_stream wraps the raw streaming response into an AsyncStreamWrapper (defined inline at _responses_instrumentor.py:2929). That wrapper exposes .response / .stream_async_iter but not .headers, so raw_stream_response.headers raises AttributeError.

_extract_served_model (line 739) already tolerates headers=None (it returns None), so the cleanest fix is to read headers defensively with getattr(..., "headers", None).

To reproduce

1. pip install agent-framework>=1.6.0 azure-ai-projects>=2.1.0
2. Set AZURE_EXPERIMENTAL_ENABLE_GENAI_TRACING=true (Microsoft-documented opt-in)
3. Configure Azure Monitor OpenTelemetry so the experimental instrumentor activates
4. Instantiate a FoundryChatClient against an Azure OpenAI Responses deployment (any GPT model — reproduced with gpt-5.1-dzs and gpt-5.5-dzs)
5. Call get_streaming_response("hello")

Expected: streamed response.
Actual: immediate ChatClientException wrapping the AttributeError.

Environment

• agent-framework-core==1.6.0, agent-framework-foundry==1.6.0, agent-framework-openai==1.6.0
• azure-ai-projects==2.1.0
• Python 3.12 on Linux
• Reproduced both via FoundryChatClient.get_streaming_response() and via the agent loop (workflow.run(..., stream=True) / agent.run(..., stream=True))

Proposed fix

Use getattr(raw_*_response, "headers", None) for all four call sites in agent_framework_openai/_chat_client.py. _extract_served_model already handles None, so the served-model feature degrades gracefully (no header surfaced) instead of crashing when the response is wrapped by an instrumentor that doesn't proxy .headers. PR follows.

Workaround for users

Set AZURE_EXPERIMENTAL_ENABLE_GENAI_TRACING=false until the fix lands.

[github-actions]

🤖 Automated triage reproduction notes (agent-authored — trust but verify)

Agent analysis

Repro: OpenAIChatClient._stream() and _get_response() in python/packages/openai/agent_framework_openai/_chat_client.py (lines 639, 680, 709, 731) access raw_response.headers directly; when the response is wrapped by an instrumentor lacking .headers (like azure-ai-projects AsyncStreamWrapper), AttributeError is raised and caught as ChatClientException. Minimal repro: mock client.responses.create to return an object with .parse() and async iteration but no .headers attribute, then call _inner_get_response with stream=True. Fix: replace all four obj.headers with getattr(obj, "headers", None) — _extract_served_model already handles None.

• Failing test: python/packages/openai/tests/openai/test_issue_6028_async_stream_wrapper_no_headers.py
• Files examined: python/packages/openai/agent_framework_openai/_chat_client.py, python/packages/openai/tests/openai/test_openai_chat_client.py, python/packages/openai/pyproject.toml
• Tests run: python/packages/openai/tests/openai/test_issue_6028_async_stream_wrapper_no_headers.py
• Reported version: 1.6.0
• Current version: 1.6.0