dataclass needs to examine the annotations of class objects it decorates. Due to __annotations__ being badly designed, it currently pulls the __annotations__ out of the class dict (cls.__dict__.get('__annotations__')). While this works today, this wouldn't work properly for classes defiend in modules using from __future__ import co_annotations if PEP 649 is accepted. It's also not recommended best practices for working with annotations. (Which, admittedly, I wrote.)

Best practices call for using inspect.get_annotations to get the annotations on any object. This does exactly what dataclass wants; it doesn't inherit base class annotations if no child class defines annotations, and it always returns a dict. (Empty, if appropriate.)

dataclass has in the past resisted using inspect.get_annotations because it didn't want to incur the runtime cost of importing inspect. However, as of this time in CPython trunk, inspect is always imported during startup, and dataclass is already importing it anyway. It's time dataclass changed to best practices here.