Python: [Bug]: Auto function calling removes null arguments

[cecheta]

Description

When using auto function calling, any null arguments that the model generates are removed before the function is invoked. If the parameter is required, then the function will fail to be invoked.

Code Sample

import asyncio
from random import randint
from typing import Annotated, Literal

from agent_framework import Agent
from agent_framework.foundry import FoundryChatClient
from azure.identity import AzureCliCredential
from pydantic import Field


def get_weather(
    location: Annotated[str, Field(description="The location to get the weather for.")],
    unit: Annotated[
        Literal["C", "F"] | None,
        Field(
            description="The unit for temperature. Only set this if the user specified a unit in their query, "
            "otherwise set it to null."
        ),
    ],
) -> str:
    """Get the weather for a given location."""
    conditions = ["sunny", "cloudy", "rainy", "stormy"]
    return f"The weather in {location} is {conditions[randint(0, 3)]} with a high of {randint(10, 30)}°{unit or 'C'}."


async def main() -> None:
    client = FoundryChatClient(
        credential=AzureCliCredential(),
    )

    agent = Agent(
        client=client,
        name="WeatherAgent",
        instructions="You are a helpful weather agent. Use the get_weather tool to answer questions.",
        tools=[get_weather],
    )

    result = await agent.run("What's the weather like in Seattle?")
    print(f"Agent: {result}")


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

Error Messages / Stack Traces

The model generates the following function call:

{"location":"Seattle","unit":null}

If you then inspect the function call result, you get the following:

[{"call_id": "call_6WcSppkJ2mMxLYpHPptu0l40", "type": "function_call_output", "output": "Error: Argument parsing failed."}]

Package Versions

agent-framework-core: 1.4.0

Python Version

Python 3.13.6

Additional Context

No response

[github-actions]

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

Agent analysis

Repro: FunctionTool.invoke() in packages/core/agent_framework/_tools.py fails when a required-but-nullable parameter (e.g. unit: Literal["C","F"] | None) receives None — model_dump(exclude_none=True) at lines 638-639 (invoke path) and 1495 (_auto_invoke_function path) strips the key, then _validate_arguments_against_schema raises TypeError for missing required args. Minimal repro: create a @tool function with a required Optional param, call .invoke(arguments={"location": "Seattle", "unit": None}) and observe the TypeError.

• Failing test: python/packages/core/tests/core/test_null_args_bug.py::test_invoke_with_null_argument_preserved
• Files examined: python/packages/core/agent_framework/_tools.py, python/packages/core/tests/core/test_null_args_bug.py, python/packages/core/pyproject.toml
• Tests run: tests/core/test_null_args_bug.py::test_invoke_with_null_argument_preserved (FAILED), tests/core/test_null_args_bug.py::test_model_dump_exclude_none_strips_null_args (PASSED)
• Reported version: 1.4.0
• Current version: 1.4.0