Python: [Bug]: MCP sampling callback cannot return structured tool-use responses

[SensorLock]

Description

MCPTool.sampling_callback forwards MCP sampling tools and toolChoice to the chat client, but cannot convert the resulting Agent Framework function_call content back into an MCP structured sampling response.

When a model follows a supplied output-schema tool, it returns Content(type="function_call"). The callback passes this through _prepare_message_for_mcp(), which handles text, image, data, and URI content but not function calls. It consequently returns: Failed to get right content types from the response.

Expected behavior

Convert Agent Framework function calls into MCP tool-use results:
Content(type="function_call")
-> ToolUseContent
-> CreateMessageResultWithTools(stopReason="toolUse")

Actual behavior

Structured sampling fails even though the tool schema is correctly forwarded to the chat client.

Impact

FastMCP Context.sample(result_type=some_dataclass) and other MCP sampling clients using tools cannot receive structured results through Agent Framework. Plain text sampling continues to work.

Suggested fix

Update sampling_callback to return CreateMessageResultWithTools when the chat response contains function calls, including updating its return annotation. The conversion should inspect all response messages and preserve each function call’s ID, name, and arguments.

Code Sample

import asyncio
from typing import Any, cast
from unittest.mock import MagicMock

from agent_framework import ChatResponse, Content, MCPStdioTool, Message
from mcp import types
from mcp.client.session import ClientSession
from mcp.shared.context import RequestContext


class StructuredClient:
    async def get_response(self, messages, **kwargs):
        # Simulate a model following the sampling request's output-schema tool.
        return ChatResponse(
            model="test-model",
            messages=[
                Message(
                    role="assistant",
                    contents=[
                        Content.from_function_call(
                            call_id="call-1",
                            name="Answer",
                            arguments={"answer": "hello"},
                        )
                    ],
                )
            ],
        )


async def main():
    tool = MCPStdioTool(
        name="repro",
        command="unused",  # No subprocess is started.
        client=StructuredClient(),
        sampling_approval_callback=lambda params: True,
    )

    result = await tool.sampling_callback(
        cast(RequestContext[ClientSession, Any], MagicMock()),
        types.CreateMessageRequestParams(
            messages=[
                types.SamplingMessage(
                    role="user",
                    content=types.TextContent(
                        type="text",
                        text="Answer the question",
                    ),
                )
            ],
            maxTokens=100,
            tools=[
                types.Tool(
                    name="Answer",
                    description="Return a structured answer",
                    inputSchema={
                        "type": "object",
                        "properties": {
                            "answer": {"type": "string"},
                        },
                        "required": ["answer"],
                    },
                )
            ],
        ),
    )

    print(result)


asyncio.run(main())

Error Messages / Stack Traces

Failed to get right content types from the response.

Package Versions

agent-framework-core==1.11.0 agent-framework-openai==1.10.1

Python Version

3.12

Additional Context

No response

[github-actions]

🤖 Automated triage reproduction notes (agent-authored — trust but verify)

Agent analysis

Repro: MCPStdioTool.sampling_callback() in python/packages/core/agent_framework/_mcp.py fails when the chat client returns Content(type="function_call") because _prepare_content_for_mcp() (line 779) has no handler for function_call content and sampling_callback() (line 1549) only searches for TextContent|ImageContent. Minimal repro: mock a chat client that returns Content.from_function_call(call_id="call-1", name="Answer", arguments={"answer":"hello"}), pass tools in CreateMessageRequestParams, and call sampling_callback — it returns ErrorData instead of CreateMessageResultWithTools. The fix requires adding a function_call→ToolUseContent case in _prepare_content_for_mcp and updating sampling_callback to return CreateMessageResultWithTools when tool-use content is present.

• Failing test: python/packages/core/tests/core/test_mcp_sampling_function_call_bug.py
• Files examined: python/packages/core/agent_framework/_mcp.py, python/packages/core/tests/core/test_mcp.py, python/packages/core/pyproject.toml
• Tests run: python/packages/core/tests/core/test_mcp_sampling_function_call_bug.py::test_sampling_callback_with_function_call_content
• Reported version: 1.11.0
• Current version: 1.11.0

[jstar0]

Hi, I'd like to take this on if the scope matches the maintainers' direction.

I propose a focused bug fix that:

• adds regression coverage for one or multiple function_call contents across response messages;
• converts each call to MCP ToolUseContent, preserving the call ID, name, and arguments;
• returns CreateMessageResultWithTools with a toolUse stop reason when tool calls are present; and
• leaves the existing text/image sampling behavior and sampling approval/rate/token guardrails unchanged.

Could a maintainer confirm this direction and assign the issue before I start implementation?