Python: [Bug]: FoundryAgent against a Prompt Agent fails with "Missing required parameter: 'model'" in 1.2.0–1.2.2 (works in 1.1.1)

[hrvoje83]

Description

Calling a Prompt Agent (created in the Foundry portal) via FoundryAgent fails on agent-framework 1.2.0, 1.2.1, and 1.2.2 with a 400 from the Responses API:

Missing required parameter: 'model'.

The same code works on 1.1.1. The agent is a Prompt Agent (kind: prompt), version 2, created in the Foundry portal. agent_version="2" is passed explicitly.

The Responses API should derive the model from the agent_reference server-side, so model should not be required. It looks like the request being built no longer attaches agent_reference to extra_body (or attaches it incorrectly) for the Prompt Agent path. Suspect PR #5447 ("Update FoundryAgent for hosted agent sessions") introduced the regression — the stack trace goes through agent_framework_openai._chat_client._get_response, which doesn't know about agent_reference on its own.

Versions

• agent-framework: 1.2.2 (also reproduces on 1.2.1 and 1.2.0)
• agent-framework-core: 1.2.2
• agent-framework-foundry: 1.2.2
• Python: [e.g. 3.13.x]
• OS: Windows [version]

Last known good: 1.1.1.

Reproduction

import os, sys, asyncio
from dotenv import load_dotenv, find_dotenv
from agent_framework.foundry import FoundryAgent
from azure.identity import ClientSecretCredential

load_dotenv(find_dotenv(), override=True)
endpoint = os.getenv("FOUNDRY_PROJECT_ENDPOINT")

credential = ClientSecretCredential(
    tenant_id=os.getenv("AZURE_TENANT_ID"),
    client_id=os.getenv("AZURE_CLIENT_ID"),
    client_secret=os.getenv("AZURE_CLIENT_SECRET"),
)

async def main():
    agent = FoundryAgent(
        project_endpoint=endpoint,
        agent_name="TestAgent",
        agent_version="2",
        credential=credential,
    )
    result = await agent.run("What is the capital of France?")
    print(result)

asyncio.run(main())