ChatCompletion streaming with AzureOpenAI provider

[ciresimo]

Please read this first

• Have you read the docs?Agents SDK docs Yes
• Have you searched for related issues? Others may have had similar requests Yes

Question

I need to integrate an agent in a third-party front-end that accepts only ChatCompletions messages (the platform was firstly built to use a "plain" OpenAI call)
I'm using a AsyncAzureOpenAI client (but I've tried using the standard AsyncOpenAI following the indications reported here https://learn.microsoft.com/en-us/azure/ai-foundry/openai/api-version-lifecycle?view=foundry-classic&tabs=python) which by default streams using Response API
I've tried changing the behavior by using set_default_openai_api("chat_completions") as suggested in the documentation. I've also tried with OpenAIProvider by setting use_responses=True and adding it to the runner
Runner.run_streamed(starting_agent=agent, input=prompt, run_config=RunConfig(model_provider=PROVIDER)). In none of the cases anything changes and I still get ResponseStreamEvents like this

RawResponsesStreamEvent(data=ResponseTextDeltaEvent(content_index=0, delta=' per', item_id='fake_id', logprobs=[], output_index=0, sequence_number=270, type='response.output_text.delta'), type='raw_response_event')

I'm wondering if I'm miss some settings, there's something wrong with my setup or if this is a known limitations

set_default_openai_api seems to have an effect only for OpenAIProvider, is this correct? And how should it be used to achieve the correct result?

Not sure if I gave all relevant information. Let me know if more details from my side are required. I leve here a snippet of the code I've being using

import asyncio
from openai import AsyncAzureOpenAI
from openai import AsyncOpenAI

from agents import Agent, OpenAIChatCompletionsModel, set_default_openai_api, OpenAIProvider, Runner, RunConfig

async def test_api():
    
    set_default_openai_api("chat_completions")
    async_client = AsyncOpenAI(
        api_key=AZURE_OPENAI_API_KEY,
        base_url="https://<company>.openai.azure.com/openai/v1/"
    ) # Same with AsyncAzureOpenAI
    MODEL = OpenAIChatCompletionsModel(
    model="gpt-4.1",
    openai_client=async_client
    )
    # agent definition
    agent = Agent(
        name="Pippo",
        model=MODEL,
        instructions="You are my personal assistant"
    )

    PROVIDER = OpenAIProvider(openai_client=async_client, use_responses=True)

    result =  Runner.run_streamed(
        starting_agent=agent, input="Hello",  run_config=RunConfig(model_provider=PROVIDER)  )


    async for event in result.stream_events():
        print(event)


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

[seratch]

Hi, thanks for asking the question.

Here are what's happening:

• The SDK always normalizes streaming to Responses-style events. Even when the Chat Completions API is called, Runner emits RawResponsesStreamEvent/ResponseTextDeltaEvent, not raw ChatCompletionChunk.
• In your snippet you passed use_responses=True into OpenAIProvider, which forces Responses mode and overrides set_default_openai_api("chat_completions").

To properly use Chat Completions instead, you can remove the explicit use_responses=True (or set it to False). That lets set_default_openai_api("chat_completions") take effect and the provider will call chat.completions under the hood.

Example

set_default_openai_api("chat_completions")
client = AsyncAzureOpenAI( ... )
provider = OpenAIProvider(openai_client=client)  # omit use_responses or set to False
agent = Agent(
    name="Pippo",
    model=OpenAIChatCompletionsModel("gpt-4.1", client),
    instructions="You are my personal assistant",
)
result = Runner.run_streamed(starting_agent=agent, input="Hello", run_config=RunConfig(model_provider=provider))

That said, the events returned from Runner is still Responses-style format. So your frontend is unable to consume the events as Chat Completions chunks. There is no built-in solution for your use case so far, and we don't have immediate plans to unlock Chat Completions style responses / events directly from Runner. If the code you're changing is relatively simple, directly using client.chat.completions.create may be quicker.

[ciresimo]

Thank you so much for your explanation
I had alredy tried with use_responses=False together with set_default_openai_api("chat_completions") and playing with keeping just one or the other, without seeing any difference in the output format. So if I understood correctly, these options have an effect only to the request sent to Azure, not to the output I'm getting from the Runner.
I need to use the Runner, so I think I will modify the chunks as they come out before sending them
Thanks for your help