OpenAIChatCompletionClient.create_stream returns finish_reason "unknown" when the finish chunk also carries usage (seen with GPT-6 Sol/Luna/Astra on Amazon Bedrock)

[kimnamu]

What happened?

Thanks for the usage-chunk handling in create_stream (#3972) and the normalized stop reasons (#5027). GPT-6 Sol, Luna and Astra on Amazon Bedrock already stream through OpenAIChatCompletionClient with just base_url, api_key and model_info set. One result field still differs from create().

Describe the bug

create_stream reads finish_reason only from chunks where usage is None (_openai_client.py L948). The Bedrock runtime endpoint /openai/v1/chat/completions puts usage on the GPT-6 finish chunk, with or without include_usage. So the finish reason is never read and the streamed result says "unknown", including for length truncation and tool calls. #4875 (DeepSeek, "ValueError: No stop reason found" always raised when llm usage returned) looks like the same code path; #5027 turned that error into "unknown".

live, main@027ecf0a	create()	create_stream() today	with the fix below
runtime us-east-1 us.openai.gpt-6-sol and global.openai.gpt-6-sol: text / truncated / tool call¹	stop / length / function_calls	❌ unknown ×3	✅ stop / length / function_calls
runtime us-east-1 us.openai.gpt-6-luna and global.openai.gpt-6-luna: same 3	same	❌ unknown ×3	✅ same as create()
runtime us.openai.gpt-6-astra (us-east-1, us-west-2) and global.openai.gpt-6-astra (us-east-1): text / truncated	stop / length	❌ unknown ×2	✅ stop / length
runtime openai.gpt-oss-120b-1:0 (no usage on finish chunk)	stop / length / function_calls	✅ same as create()	✅ unchanged
Mantle openai.gpt-6-sol (us-east-1, 3 cases) and openai.gpt-6-astra (us-west-2, text / truncated), no usage on finish chunk	stop / length (/ function_calls)	✅ same as create()	✅ unchanged
streamed usage, GPT-6 Sol text	—	(8, 5)	✅ (8, 5) unchanged

¹ with reasoning_effort="none", which Bedrock chat completions needs for Sol/Luna function tools.

To Reproduce

import asyncio
import os

from autogen_core.models import ModelInfo, UserMessage
from autogen_ext.models.openai import OpenAIChatCompletionClient


async def main() -> None:
    for model in ["us.openai.gpt-6-sol", "us.openai.gpt-6-luna", "us.openai.gpt-6-astra"]:
        client = OpenAIChatCompletionClient(
            model=model,
            base_url="https://bedrock-runtime.us-east-1.amazonaws.com/openai/v1",
            api_key=os.environ["AWS_BEARER_TOKEN_BEDROCK"],
            model_info=ModelInfo(vision=True, function_calling=True, json_output=True, family="unknown", structured_output=True),
        )
        messages = [UserMessage(content="Write a long poem about the sea.", source="user")]
        args = {"max_completion_tokens": 5}
        result = await client.create(messages, extra_create_args=args)
        async for streamed in client.create_stream(messages, extra_create_args=args):
            pass
        print(f"{model}: create={result.finish_reason} create_stream={streamed.finish_reason}")


asyncio.run(main())

us.openai.gpt-6-sol: create=length create_stream=unknown
us.openai.gpt-6-luna: create=length create_stream=unknown
us.openai.gpt-6-astra: create=length create_stream=unknown

Expected behavior

create_stream reports the same finish_reason as create().

Additional context

Raw stream tail, proposed one-line fix, test RED/GREEN

Last two SSE events from us.openai.gpt-6-sol (trimmed; us.openai.gpt-6-astra has the same shape):

data: {"choices":[{"delta":{},"finish_reason":"stop","index":0}],...,"usage":{"completion_tokens":5,...}}
data: {"choices":[],...,"usage":{"completion_tokens":5,...}}

Fix: drop the chunk.usage is None condition. The stop_reason is None check already keeps an earlier finish reason when a LiteLLM-style usage chunk (with finish_reason=None) follows, and the existing usage-chunk tests still pass. The Azure AI client already reads finish_reason whenever it is set.

-            stop_reason = choice.finish_reason if chunk.usage is None and stop_reason is None else stop_reason
+            stop_reason = choice.finish_reason if stop_reason is None else stop_reason

A mocked test (content chunk → finish chunk with usage → choices=[] usage chunk), parametrized for stop and length:

# main
E       AssertionError: assert 'unknown' == 'stop'
E       AssertionError: assert 'unknown' == 'length'
2 failed, 86 deselected in 2.68s
# with fix
2 passed, 86 deselected in 2.41s
tests/models/test_openai_model_client.py: 55 passed, 33 skipped

Astra has no tool-call row because Bedrock's chat completions endpoint returns 400 for function tools with Astra ("To use function tools, use /v1/responses"), on main and with the fix alike.

Released 0.7.5 has the same line. Happy to open a PR with the fix and the test.

Which packages was the bug in?

Python Extensions (autogen-ext)

AutoGen library version.

Python dev (main branch)

Other library version.

0.7.5 has the same code

Model used

us.openai.gpt-6-sol, us.openai.gpt-6-luna, us.openai.gpt-6-astra (and their global. profiles)

Model provider

AWS Bedrock

Other model provider

No response

Python version

3.11

.NET version

No response

Operating system

MacOS

[michaelkillgta]

Hi! I'd like to take this. Plan is to make OpenAIChatCompletionClient.create_stream report the real finish reason instead of "unknown" when the stream ends cleanly, add a regression test, and open a PR.