Python: [Bug]: Non-streaming FoundryChatClient / FoundryAgent drop oauth_consent_request output items (empty answer, no consent link) #8975

Description

When a Foundry tool (toolbox / MCP connection) needs the user's OAuth consent, the service returns an oauth_consent_request output item with a consent_link and marks the response incomplete.

The streaming path surfaces it: try_parse_oauth_consent_event (added for #3950) turns response.output_item.added / response.oauth_consent_requested into Content.from_oauth_consent_request(...).

The non-streaming path does not. RawFoundryChatClient._parse_response_from_openai and RawFoundryAgentChatClient._parse_response_from_openai only post-process reasoning items / agent_session_id, and the base RawOpenAIChatClient._parse_response_from_openai hits case _: logger.debug("Unparsed output of type ...") for the consent item. agent.run() (non-streaming is the default) therefore returns an empty assistant message, finish_reason=None, no exception, and the consent link is lost. The caller cannot tell the user to sign in.

• Expected: the response contains an oauth_consent_request content with the consent_link, same as with stream=True.
• Actual: the response has no content at all.

Code Sample

Self-contained repro (no Azure resources; the HTTP layer is mocked, the real OpenAI SDK parses the payload):

import asyncio, importlib
from unittest.mock import MagicMock
from openai import AsyncOpenAI, DefaultAsyncHttpxClient
from agent_framework_foundry import FoundryAgent

# httpx (openai 2.x) or httpx2 (openai 3.x), whichever the installed SDK uses
httpx = importlib.import_module(DefaultAsyncHttpxClient.__mro__[1].__module__.partition(".")[0])

def handler(request):
    return httpx.Response(200, json={
        "id": "resp_1", "object": "response", "created_at": 0, "model": "gpt-5.4",
        "status": "incomplete",
        "output": [{
            "type": "oauth_consent_request", "id": "oauth-item-1",
            "consent_link": "https://consent-host.example.com/login?data=abc123",
            "server_label": "github",
        }],
    })

async def main():
    oai = AsyncOpenAI(api_key="x", http_client=DefaultAsyncHttpxClient(transport=httpx.MockTransport(handler)))
    project = MagicMock()
    project.get_openai_client.return_value = oai
    agent = FoundryAgent(project_client=project, agent_name="toolbox-agent")
    response = await agent.run("List my repos")
    print([c.type for m in response.messages for c in m.contents])  # [] -> consent link lost

asyncio.run(main())

Same result with FoundryChatClient(...).get_response(...). With stream=True the consent content is returned.

Root cause

• python/packages/foundry/agent_framework_foundry/_chat_client.py:338 (RawFoundryChatClient._parse_response_from_openai) and python/packages/foundry/agent_framework_foundry/_agent.py:423 (RawFoundryAgentChatClient._parse_response_from_openai) never look for oauth_consent_request items; only the streaming overrides (:367 / :434) call try_parse_oauth_consent_event.
• The item then falls through to python/packages/openai/agent_framework_openai/_chat_client.py:3543 (logger.debug("Unparsed output of type ...")).

Proposed fix

Add a small helper in _oauth_helpers.py that converts oauth_consent_request items of a completed response's output into Content.from_oauth_consent_request (same HTTPS validation as the streaming path), and append the result to the message in both non-streaming overrides. PR with regression tests for both clients is ready.

Code Sample

Error Messages / Stack Traces

Package Versions

agent-framework-foundry: 1.14.0, agent-framework-openai: 1.15.0, agent-framework-core: 1.20.0 (main @ e9857bd)

Python Version

Python 3.12

Additional Context

Related: #3950 (streaming events silently dropped). That fix covers the streaming path only.