fix: guard against None chunks in OpenAI streaming client

[jahanzaib-iqbal-dev]

Summary

Fixes AttributeError: 'NoneType' object has no attribute 'model' that occurs when the OpenAI streaming response yields a None chunk.

• Adds a None guard at the top of the streaming loop (before any attribute access) so None chunks are silently skipped.
• Removes a redundant duplicate maybe_model = chunk.model assignment that appeared later in the same loop body (the assignment already happens earlier in the loop).

Root Cause

In _openai_client.py, chunk.model was accessed before checking if chunk is None. When the API yields a None sentinel chunk (observed in compiled binaries and under heavy endpoint load), this caused an unhandled AttributeError that crashed the streaming session:

# Line 908: Iterate over stream chunks
async for chunk in chunks:
    if first_chunk:
        ...
    # Line 919: chunk.model accessed with NO None guard — crashes if chunk is None
    maybe_model = chunk.model

    # Line 923: Empty chunk check comes AFTER — too late
    if len(chunk.choices) == 0:
        ...

The type annotation on line 893 (chunk: ChatCompletionChunk | None = None) already documents that chunk can be None, but the guard was missing.

Fix

async for chunk in chunks:
    # Guard against None chunks which can be yielded by some endpoints.
    # https://github.com/microsoft/autogen/issues/7130
    if chunk is None:
        continue
    ...
    # Set the model from the latest chunk (only runs after None guard)
    maybe_model = chunk.model

Testing

• Added test_openai_chat_completion_client_create_stream_none_chunk in test_openai_model_client.py which injects None chunks at both the start and in the middle of the stream and verifies the stream completes successfully without raising AttributeError.
• All existing streaming tests continue to pass.

Fixes #7130

[jahanzaib-iqbal-dev]

@microsoft-github-policy-service agree