I was reading through the specification of dataclass_transform and it doesn't really specify what field_specifiers looks like. It mostly just says:

field_specifiers (tuple[Callable[..., Any], ...]) – Specifies a static list of supported classes or functions that describe fields, similar to dataclasses.field(). Defaults to ().

But dataclasses.field takes keyword-only arguments, which leaves it a bit ambiguous about whether positional arguments are ok. For example:

from typing import dataclass_transform, Any

def custom_field(default: object, *, init: bool = True) -> Any: ...

@dataclass_transform(field_specifiers=(custom_field,))
def build(x): ...

@build
class A:
    x: int = custom_field(default=0)

@build
class B:
    x: int = custom_field(0)

A()
B()

pyrefly, pyright, and mypy all flag B() as an error while ty accepts it without complaint.

My (uninformed) take is that this should be allowed -- custom_field(default=0) and custom_field(0) have the same runtime effect, so it feels like they should have the same type-checking effect too. This also matches how pydantic and attrs works (e.g. you can do pydantic.Field(1) and attr.ib(0)) -- today I suspect the type-checkers are special-casing this to make pydantic and attrs (since I've definitely used both without running into static type-checking errors about missing default values before).