.NET: Python: [Bug]: Multi-turn tool calls fail with both Responses API and Chat Completions API when used via `agent-framework-ag-ui` (Foundry project endpoint)

[nomhiro]

Description

Summary

Using agent-framework-ag-ui from Python with a Microsoft Foundry project inference endpoint, the first turn of a tool-calling chat works correctly, but any second or later turn that follows a tool call fails with a 400 from the upstream OpenAI API.

The failure mode differs by client type but both surface a consistent root cause: when AG-UI replays conversation history on subsequent turns, the messages produced by the prior tool-calling turn are not paired correctly with their tool results before being sent back to the LLM.

This affects all four AG-UI features we tested (Agentic Chat, Human-in-the-Loop, Shared State, Agentic Generative UI) as soon as the agent has invoked any tool at least once.

Code Sample

## Reproduction

Minimal server (full repo: <will-be-linked>):


import os
from agent_framework import Agent, tool
from agent_framework.openai import OpenAIChatClient  # or OpenAIChatCompletionClient — both fail
from agent_framework_ag_ui import add_agent_framework_fastapi_endpoint
from fastapi import FastAPI
from fastapi.middleware.cors import CORSMiddleware
from dotenv import load_dotenv

load_dotenv()

@tool
def search_cars(car_class: str | None = None) -> list[dict]:
    return [{"car_id": "SUV-001", "class": "suv", "daily_jpy": 12800}]

client = OpenAIChatClient(  # also tried OpenAIChatCompletionClient — same outcome, different error
    model=os.environ["FOUNDRY_MODEL"],
    base_url=os.environ["FOUNDRY_BASE_URL"],  # ".../openai/v1"
    api_key=os.environ["FOUNDRY_API_KEY"],
)

agent = Agent(
    name="ShirokumaChat",
    instructions="You are a car rental booking assistant.",
    client=client,
    tools=[search_cars],
)

app = FastAPI()
app.add_middleware(CORSMiddleware, allow_origins=["*"], allow_methods=["*"], allow_headers=["*"])
add_agent_framework_fastapi_endpoint(app, agent, "/agentic_chat")


Steps:

1. Start the server on `localhost:8888`.
2. In AG-UI Dojo (running locally), open the Agentic Chat tab for `Microsoft Agent Framework (Python)`.
3. Send turn 1: "Show me your SUV cars" → tool is called, response streams correctly. ✅
4. Send turn 2 in the **same** chat session (anything, e.g. "What's the price?") → server returns 200 but the stream contains a `RUN_ERROR` and no text response. ❌

Error Messages / Stack Traces

## Expected behavior

Subsequent turns should continue the conversation, with the agent having access to the tool result from the previous turn.

## Actual behavior

### With `OpenAIChatClient` (Responses API)


openai.BadRequestError: Error code: 400 - {
  'error': {
    'message': 'No tool output found for function call call_cLJJ7tsAQL9j7ZK3J1g8ca30.',
    'type': 'invalid_request_error',
    'param': 'input',
    'code': None
  }
}


Full traceback (abridged):


File ".venv/Lib/site-packages/agent_framework_openai/_chat_client.py", line 647, in _stream
    async for chunk in await client.responses.create(stream=True, **run_options):
File ".venv/Lib/site-packages/openai/resources/responses/responses.py", line 2626, in create
    return await self._post(...)
agent_framework.exceptions.ChatClientException: OpenAIChatClient service failed to complete the prompt: 400 - No tool output found for function call call_cLJJ7tsAQL9j7ZK3J1g8ca30.


### With `OpenAIChatCompletionClient` (Chat Completions API)


openai.BadRequestError: Error code: 400 - {
  'error': {
    'message': "An assistant message with 'tool_calls' must be followed by tool messages responding to each 'tool_call_id'. The following tool_call_ids did not have response messages: call_7jOeA41CWTxLIyheCJnxqr6G, call_cWxiZjHImVV1KTq6ajOqTbnB",
    'type': 'invalid_request_error',
    'param': 'messages.[3].role',
    'code': None
  }
}


Full traceback (abridged):


File ".venv/Lib/site-packages/agent_framework_openai/_chat_completion_client.py", line 522, in _stream
    async for chunk in await client.chat.completions.create(stream=True, **options_dict):
agent_framework.exceptions.ChatClientException: OpenAIChatCompletionClient service failed to complete the prompt:
  An assistant message with 'tool_calls' must be followed by tool messages responding to each 'tool_call_id'.

Package Versions

agent-framework-core 1.4.0, agent-framework-ag-ui 1.0.0rc1, agent-framework-openai 1.4.0, ag-ui-protocol 0.1.18, openai 2.37.0, fastapi 0.133.0, pydantic 2.13.4

Python Version

3.12.11

Additional Context

Frontend: AG-UI Dojo (apps/dojo in ag-ui-protocol/ag-ui, latest main), integration = Microsoft Agent Framework (Python), pointed at our local server via AGENT_FRAMEWORK_PYTHON_URL=http://localhost:8888.

Backend endpoint: Microsoft Foundry project inference URL of the form
https://<resource>.services.ai.azure.com/api/projects/<project>/openai/v1
authenticated with an API key.

Why both APIs fail (analysis)

Both errors point at the same root cause from different angles:

• Chat Completions strictly requires (assistant{tool_calls}, tool{tool_call_id}) pairs to be interleaved in the messages array. The 400 says the tool role messages are missing from the replayed history.
• Responses API expects function call results to be presented as function_call_output items in input. The 400 says the output for a specific call_id is missing.

In both cases, AG-UI's HttpAgent (or whatever produces the replayed history sent to the Python backend) is dropping or mis-pairing the tool result messages from the previous turn. The server-side agent-framework-ag-ui pipeline appears to pass this incomplete history through to the chat client without reconstructing the missing pairs.

Related issues

• Python: [Bug]: Tool calling not working with Responses API #3304 — Python: Tool calling not working with Responses API (matches our Responses-API error exactly)
• .NET: AG-UI: Multi-turn tool calls replay produces invalid OpenAI tool_call history #2699 — .NET: AG-UI: Multi-turn tool calls replay produces invalid OpenAI tool_call history (.NET equivalent of our Chat Completions error)
• Python: AG-UI, Tool calls suppress shared state message injection #3010 — Python: AG-UI shared state breaks after first tool call
• Python: [Bug]: MCP Tool Calls + Reasoning Not Emitted as AG-UI Events (Foundry) #4213 — Python: MCP tool calls + reasoning not emitted as AG-UI events (Foundry)
• Python: [AG-UI] Error occurs when using with_plan_review in Magentic workflow #3331 — Python: with_plan_review in Magentic workflow + AG-UI
• AG-UI: tool result messages missing from conversation history on subsequent turns (causes LLM 400) CopilotKit/CopilotKit#3884 — AG-UI: tool result messages missing from conversation history on subsequent turns (client-side cause), proposes PR .NET: AF version bump 12-feb #3885

This issue intentionally reports the end-to-end symptom on Python + Foundry with both client types, because the existing issues are scattered (per-API, per-platform, per-feature) and the impression from a user's perspective is that AG-UI is currently unusable for any multi-turn tool-calling scenario in Python — which is not made visible in the AG-UI integration docs or the Microsoft Community Hub announcement post.

Requests

1. Confirm whether the root cause is server-side (agent-framework-ag-ui history reconstruction) or client-side (AG-UI Dojo / HttpAgent not sending tool results back) — or both.
2. Add a documentation note about the current limitation on the AG-UI integration page until fixed.
3. Coordinate with CopilotKit#3884 / PR .NET: AF version bump 12-feb #3885 if the fix needs to land on both sides.

Workaround tried

We are evaluating the workaround suggested in CopilotKit#3884: subclass AgentFrameworkAgent to scan input messages and inject synthetic tool (or function_call_output) entries for any tool_call_id that lacks a matching result before handing off to the chat client. Will follow up here with results.