Why do we need this? In my opinion, def __subclasshook__(cls, other): is enough.

For example, your own tests do not use this attribute.

Without this attribute, the abstract base class for all dataclasses would not itself be considered a dataclass by the dataclasses internal machinery, which would be quite bizarre in my opinion:

>>> from dataclasses import *
>>> is_dataclass(DataclassLike)
True
>>> del DataclassLike.__dataclass_fields__
>>> is_dataclass(DataclassLike)
False

But you are correct that this currently isn't covered. I will add a test for this :)

But, is DataclassLike a dataclass? :)

You can have two opinions here.

The contract here is that "all classes for which is_dataclass evaluates to True shall be considered subclasses of DataclassLike. But DataclassLike is considered a subclass of DataclassLike (because all classes are considered subclasses of themselves in Python). So it would be very strange, in my opinion, for there to be one class (DataclassLike itself) where is_dataclass(x) evaluates to False but it's nonetheless considered a subclass of DataclassLike.

Alternatively, I could make DataclassLike an actual dataclass, I suppose? Rather than "faking it". What do you think?

Do you like it better after adad189? :)