Fix(Bug): Automatic_function_calling config persists across requests, causing UNEXPECTED_TOOL_CALL when tools=None

[Muhammad-Rebaal]

Fixes #1818

Hi @janasangeetha !

I've identified and fixed this issue
The bug was in the should_disable_afc() function in google/genai/_extra_utils.py. The function defaulted to enabling AFC even when tools=None or tools=[], because it checked the automatic_function_calling config before checking whether any tools were actually provided.

Thank you!

[janasangeetha]

Hey @Muhammad-Rebaal
Thanks for your contributions!
The branch is out-of-date. Please update the same.

[BenjaminKazemi]

Please test tests/afc/test_get_max_remote_calls_for_afc.py as well. Thanks for your contribution.

[BenjaminKazemi]

Please change the test cases as much as possible instead of changing the assertion. By changing only the assertion, the tests are not validating the actual test case. E.g. in this test case, pass in a tool and keep the assertion.

[Muhammad-Rebaal]

Thanks for the feedback! I've updated the test cases as suggested.

Instead of changing the assertions, I've added tools parameter to the three failing tests:

test_afc_enable_unset_max_1 , test_afc_enable_unset_max_1_0 , test_afc_enable_true_max_unset

[BenjaminKazemi]

Thanks! In addition to that would you mind adding an example code that reproduce the bug? AFC logic may seem simple in the first glance but there are corner cases. Would be idea if you could add an example to regenerate the issue or refer me to an existing one.

[Muhammad-Rebaal]

My pleasure , Here's the minimal that can reproduce the error mentioned by @npkanaka :

from google import genai
from google.genai import types

client = genai.Client(api_key="#####")
MODEL = "gemini-2.5-pro"

tools = [
    types.Tool(
        function_declarations=[
            types.FunctionDeclaration(
                name="noop", description="noop", parameters={"type": "object", "properties": {}}
            )
        ]
    )
]

# Call 1: with tools, force function_call
resp1 = client.models.generate_content(
    model=MODEL,
    contents="Call the noop tool. Do not answer in text. Only return a function call to noop with empty args.",
    config=types.GenerateContentConfig(tools=tools),
)
print("Call 1:", resp1)

# Call 2: tools=None, fresh
resp2 = client.models.generate_content(
    model=MODEL,
    contents="hello from call without tools",
    config=types.GenerateContentConfig(tools=None),
)
print("Call 2:", resp2)

# Call 3: history includes prior function_call, tools=None
history = [
    types.Content(
        role="user",
        parts=[types.Part(text="Call the noop tool. Do not answer in text. Only return a function call to noop with empty args.")],
    ),
    resp1.candidates[0].content,  # contains function_call
    types.Content(
        role="user",
        parts=[types.Part(text="Now answer normally (no tools). Just say hi in plain text.")],
    ),
]
resp3 = client.models.generate_content(
    model=MODEL,
    contents=history,
    config=types.GenerateContentConfig(tools=None),
)
print("Call 3:", resp3)

Although, as you said there can be corner cases, so kindly let me know what is up to you. I'd be happy to implement that under your mentorship.