FormPage.__pydantic_init_subclass__ doesn't behave as expected

[Viicos]

I came across this implementation while trying to evaluate potential regressions in the next Pydantic release (2.11):

pydantic-forms/pydantic_forms/core/shared.py

Lines 45 to 53 in 0dbbd50

	@classmethod
	def __pydantic_init_subclass__(cls, /, **kwargs: Any) -> None:
	# The default and requiredness of a field is not a property of a field
	# In the case of DisplayOnlyFieldTypes, we do kind of want that.
	# Using this method we set the right properties after the form is created
	for field in cls.model_fields.values():
	if field.frozen:
	field.validate_default = False

This is not working as you would expect, because Pydantic constructs a core schema used during validation based on model fields; and this happens before this __pydantic_init_subclass__ hook is called.

This core schema is used to understand how validation should be applied, and it contains the logic related to validate_default:

from pydantic import BaseModel, Field

class FormPage(BaseModel):
    @classmethod
    def __pydantic_init_subclass__(cls, /, **kwargs):
        for field in cls.model_fields.values():
            if field.frozen:
                field.validate_default = False


class Sub(FormPage):
    a: int = Field(default="wrong", frozen=True, validate_default=True)

Sub.model_fields
{'a': FieldInfo(annotation=int, required=False, default='test', frozen=True, validate_default=False)}

Sub.__pydantic_core_schema__
{
│   'type': 'model',
│   'cls': <class '__main__.Sub'>,
│   'schema': {
│   │   'type': 'model-fields',
│   │   'fields': {
│   │   │   'a': {
│   │   │   │   'type': 'model-field',
│   │   │   │   'schema': {'type': 'default', 'schema': {'type': 'int'}, 'validate_default': True, 'default': 'test'},  # Still True!
│   │   │   │   'frozen': True,
│   │   │   │   'metadata': {}
│   │   │   }
│   │   },
│   │   'model_name': 'Sub',
│   │   'computed_fields': []
│   },
│   'config': {'title': 'Sub'},
│   'ref': '__main__.Sub:95507792084592',
│   'metadata': {'<stripped>'}
}

Model()
# ValidationError, default value is still validated

To make sure the core schema is reconstructed, you can call model_rebuild(force=True):

class FormPage(BaseModel):
    @classmethod
    def __pydantic_init_subclass__(cls, /, **kwargs):
        needs_rebuild = False
        for field in cls.model_fields.values():
            if field.frozen:
                field.validate_default = False
                needs_rebuild = True

        if needs_rebuild:
            cls.model_rebuild(force=True)

Note that Pydantic does not really explicitly support this pattern of rebuilding a model after altering the model fields.

On a related note, relying on model_fields just after class creation might lead to incorrect results. The reason for this is that subclasses of FormPage might use unresolvable forward references. This PR gives more context: pydantic/pydantic#11388.

In 2.11, we might raise a warning when incomplete model fields are being accessed (cross ref this issue), but be assured that if we do so, we will provide another kind of hook similar to __pydantic_init_subclass__, that is only called when fields are guaranteed to be complete.