Metaclasses with type annotated attributes do not add `__annotations__ = {}` to derived classes.

[matthewwardrop]

Bug report

Bug description:

This can cause odd behaviour for subclasses which then propagate their parent's annotations rather than their own.

class MyType(type):
    METACLASS_VAR: int = 1

class A(metaclass=MyType):
    value: str = "HI"

class B(A):
    pass

B.__annotations__  # {"value": str} instead of {}; but if you remove the type annotation in `MyType`, you get {}

This issue only seems to affect versions prior to 3.14

CPython versions tested on:

3.10, 3.13, 3.14

Operating systems tested on:

Linux

[JelleZijlstra]

This is discussed in https://peps.python.org/pep-0749/#pre-existing-bugs and fixed in 3.14 (unless you use from __future__ import annotations).

[matthewwardrop]

Oh... just noted that as I did a quick test on Python 3.14. I tried search for this in Google, but didn't find it (probably had the wrong keywords). Thanks for linking me on @JelleZijlstra . Feel free to close this issue if this is considered a non-issue going forward.

[DavidCEllis]

It is somewhat of an obscure place for this to be documented but it is also probably a fairly rare thing to come across? I think technically this was supposed to be fixed in 3.10 where it was fixed for classes without the metaclass.

Whether or not this is considered worth fixing for __future__ annotations or as a bugfix on 3.13, perhaps it may be worth putting something in the annotations best practices page about this? Currently it indicates that in 3.10 and newer you can use getattr if you can't use inspect.get_annotations which would lead you to this bug.

[JelleZijlstra]

I wouldn't mind documenting this in the best practices doc.

We could also fix this (in 3.13, and in 3.14+ when from __future__ import annotations is involved), but I suspect any fix will also break some other ways of using annotations and it's better left alone.