[Bug] AttributeError in streaming due to missing None check on chunk.model (lines 919, 949)

[abhijaysingh]

What happened?

---

Bug Description

The BaseOpenAIChatCompletionClient.create_stream() method crashes with AttributeError: 'NoneType' object has no attribute 'model' when the OpenAI stream yields None chunks. This is especially prevalent in compiled binaries and causes premature stream termination.

Affected File: autogen_ext/models/openai/_openai_client.py
Affected Lines: 919, 949
Version: autogen-ext 0.7.5

Root Cause

# Line 908: Iterate over stream chunks
async for chunk in chunks:
    if first_chunk:
        ...

    # Line 919:  Accesses chunk.model without None check
    maybe_model = chunk.model  # AttributeError if chunk is None!

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

The code accesses chunk.model before checking if chunk is None, despite the type annotation allowing None at line 893:

chunk: ChatCompletionChunk | None = None

Steps to Reproduce

import asyncio
import os
from autogen_core.models import CreateResult
from autogen_ext.models.openai import OpenAIChatCompletionClient
from autogen_core.models import UserMessage


async def test():
    client = OpenAIChatCompletionClient(
        model=os.getenv("OPENAI_MODEL"),
        base_url=os.getenv("OPENAI_BASE_URL"),
        api_key=os.getenv("OPENAI_API_KEY"),
    )

    try:
        while True:
            prompt = input("Enter a prompt: ")
            messages = [UserMessage(content=prompt, source="user")]
            stream = client.create_stream(
                messages=messages,
                tools=[],
                extra_create_args={},
            )
            async for chunk in stream:
                if isinstance(chunk, CreateResult):
                    print(chunk.content, end="", flush=True)
                if isinstance(chunk, str):
                    print(chunk, end="", flush=True)

            print("\n\n")
    except KeyboardInterrupt:
        print("\nExiting...")
    except Exception as e:
        print(f"Error: {e}")
    await client.close()


if __name__ == "__main__":
    asyncio.run(test())

Trigger conditions:

• Run in compiled binary (PyInstaller/Nuitka)
• Use long-running streams
• High latency or heavy API load

Expected Behavior

Stream should skip None chunks and complete with final CreateResult.

Actual Behavior

AttributeError: 'NoneType' object has no attribute 'model'

Stream terminates prematurely without yielding final result.

Proposed Fix

Add None check before accessing attributes:

async for chunk in chunks:
    # Add this check immediately
    if chunk is None:
        continue

    if first_chunk:
        ...

    # Now safe to access
    maybe_model = chunk.model
    ...

Same fix needed at line 949.

Impact

Severity: High

• Breaks streaming in compiled binaries
• Unpredictable failures depending on network/API behavior
• Affects OpenAIChatCompletionClient and AzureOpenAIChatCompletionClient

Workaround

Monkeypatch the method to filter None chunks:

from autogen_ext.models.openai import BaseOpenAIChatCompletionClient

original = BaseOpenAIChatCompletionClient._create_stream_chunks

async def patched(self, *args, **kwargs):
    async for chunk in original(self, *args, **kwargs):
        if chunk is not None:
            yield chunk

BaseOpenAIChatCompletionClient._create_stream_chunks = patched

Environment

autogen-core==0.7.5
autogen-ext==0.7.5
openai==1.65.5
Python==3.12
Deployment: Compiled binary (Nuitka)

Additional Notes

• Lines 1082-1108 (_create_stream_chunks) and 1110-1140 (_create_stream_chunks_beta_client) both need the same fix
• Type annotation at line 893 correctly indicates chunk can be None
• OpenAI stream can yield None for keepalives, heartbeats, or under load
• Issue is more frequent in compiled binaries due to different async timing

---

Fix complexity: Low - Single line addition: if chunk is None: continue