PydanticAI Github Bot Found 1 issues similar to this one:

1. "model_serializer in "wrap" mode called twice when model_dump called with serialize_as_any=True #11139" (90% similar)

Defining a custom __get_pydantic_core_schema__() can lead to unexpected results, and we decided not to explicitly support it.

There's a first issue happening: for Pydantic models, the core schema is cached in the __pydantic_core_schema__ property. For MyModel, you end up with MyModel.__pydantic_core_schema__ = {'type': 'union', 'choices': [...]} (constructed from your __get_pydantic_core_schema__() implementation). Then, when MyParentModel encounters MyModel for field nested, it will call the __get_pydantic_core_schema__() method again. When calling handler(cls), Pydantic will use the cached __pydantic_core_schema__ property (which already contains the 'union' schema). You then end up with a nested schema ({'type': 'union', 'choices': [{'type': 'union', 'choices': [...]}, ...]}).

What you need to do is manually reusing the cached attribute:

    @classmethod
    def __get_pydantic_core_schema__(cls, source, handler):
        schema = cls.__dict__.get('__pydantic_core_schema__')
        if schema is not None and isinstance(schema, dict):
            return cls.__pydantic_core_schema__
        return pydantic_core.core_schema.union_schema(
            [
                handler(cls),
                # anything goes here, using duplicate schema for the example
                handler(cls),
            ]
        )

The main issue still persists though, and I think it is indeed related to #11505. We knew the fix wasn't robust enough, so a different approach needs to be taken.

thank you for the workaround 🙏.

Defining a custom get_pydantic_core_schema() can lead to unexpected results, and we decided not to explicitly support it.

does this mean this will get dropped completely? i'm heavily relying on it, especially for custom types, hence the ask.

Sorry I meant Defining a custom __get_pydantic_core_schema__() on Pydantic models, it is supported and will stay this way for custom types as documented.