Consider the following code on 3.11 main:

from typing import Annotated, ClassVar, get_type_hints

class DC:
    a: ClassVar[int] = 3
    b: ClassVar["int"] = 3
    c: "ClassVar[int]" = 3
    d: Annotated[ClassVar[int], (2, 5)] = 3
    e: Annotated[ClassVar["int"], (2, 5)] = 3
    f: "Annotated[ClassVar[int], (2, 5)]" = 3

class DC_Special_ForwardRef:
    g: Annotated["ClassVar[int]", (] = 3

# OK
assert get_type_hints(DC, globals(), locals()) == {
    "a": ClassVar[int],
    "b": ClassVar[int],
    "c": ClassVar[int],
    "d": ClassVar[int],
    "e": ClassVar[int],
    "f": ClassVar[int],
}

# TypeError: typing.ClassVar[int] is not valid as type argument
get_type_hints(DC_Special_ForwardRef, globals(), locals())

Currently, the Annotated["ClassVar[int]", (2, 5)] annotation raises at runtime when get_type_hints is called, but all the other forward reference annotations are okay.

My understanding is this is because when typing._type_check runs on a type where special forms are allowed it's possible for the typing._type_convert it calls it itself run a typing._type_check on contained forward references. However, if that forward reference was itself a special form then it's possible to get an error because typing._type_check doesn't pass on that special forms are allowed.

I have drafted a patch to pass on this information. This will become important in the future as more special forms are allowed to wrap each other, such as allowing Final and ClassVar to nest each other in dataclasses, or when Required and NotRequired land. In the future we may also want to reconsider runtime restrictions on special forms in typing.py entirely and instead choose to leave this to type checkers.

Is my analysis/patch approach okay? Forward references can be tricky. Should we be discussing runtime restrictions in typing.py more generally in the future?

typo: the line is g: Annotated["ClassVar[int]", (2, 5)] = 3 in the code sample. Somehow I left it at only ( for the annotation instead of (2,5)

I wonder if this is at all similar to bpo-41370.

I did try the proposed patch in bpo-41370 and verified it didn't resolve the issue so I'm not certain they strictly overlap, but I also haven't had time to fully digest the underlying issues in bpo-41370 yet.

I think it does have relevance for changes we want to make for dataclasses re: Annotated, though: https://bugs.python.org/issue46511

Agree it's not the same issue, but there's similarity -- both are due to putting a stringized annotation (presumably a forward ref) somewhere inside another construct. whether it's list["N"] or Annotated["ClassVar[int]"].

New changeset ced5005 by Gregory Beauregard in branch 'main':
bpo-46539: Pass status of special typeforms to forward references (GH-30926)
ced5005

New changeset 3757703 by Miss Islington (bot) in branch '3.9':
bpo-46539: Pass status of special typeforms to forward references (GH-30926)
3757703

New changeset bfcb414 by Miss Islington (bot) in branch '3.10':
bpo-46539: Pass status of special typeforms to forward references (GH-30926)
bfcb414