Update Tool dataclass to accept a custom function_schema argument

[jonchun]

Unsure if the tests should be importing _pydantic like that, but the idea behind the test is to add a function schema for a different function and checking that the generated json schema matches the function schema passed rather than the one that would have been generated without it.

Resolves #878.

[sydney-runkle]

LGTM overall, thanks for the contribution! I commented with one test change request :)

[sydney-runkle]

Maybe we could just do assert agent._function_tools['x_tool']._parameters_json_schema == snapshot(...)?

[jonchun]

17df4fc

@sydney-runkle updated

[dmontagu]

I like this, but if we expose a way to override the function_schema (which I see could be a useful convenience in some scenarios where a function is defined using *args and **kwargs), I think we should move pydantic_ai._pydantic to be a public module (or at least the things in it related to FunctionSchema), and rename it.

[jonchun]

@dmontagu - Any suggestions? There's quite a lot/almost all of _pydantic is related to FunctionSchema or building it. I had the same thought as you but figured it was internal intentionally because of this comment

This module has to use numerous internal Pydantic APIs and is therefore brittle to changes in Pydantic.

I think maybe we can pull out only the FunctionSchema TypedDict so that it is public?

[sydney-runkle]

I don't think we should make pydantic_ai._pydantic public - I think we should preserve its privacy so that we have the ability to make changes quickly as needed for now.

That being said, let's expose FunctionSchema. I think in tools.py would be appropriate for now.

[jonchun]

@sydney-runkle - how does the latest commit look? I had to change a few things around due to the circular dependency.