Mypy *sometimes* doesn't see metaclass-provided properties for protocols

[posita]

This is just weird. Somehow, SupportsInt and other standard library Supports protocols (and subclasses) are "special" when used in composing protocols. I do not know if this is related to (or the same root cause as) python/mypy#7945.

# test_case.py
from abc import abstractmethod
from typing import Any, Protocol, SupportsInt, _ProtocolMeta, runtime_checkable

class MyProtocolMeta(_ProtocolMeta):
    def foo(cls) -> None:
        print("foo")

@runtime_checkable
class MySupportsInt(
    SupportsInt,
    Protocol,
    metaclass=MyProtocolMeta,
):
    pass

@runtime_checkable
class SupportsFoo(
    Protocol,
    metaclass=MyProtocolMeta,
):
    @abstractmethod
    def foo(self) -> Any:
        pass

@runtime_checkable
class SupportsBar(
    Protocol,
    metaclass=MyProtocolMeta,
):
    @abstractmethod
    def bar(self) -> Any:
        pass

@runtime_checkable
class SupportsComposedThing(
    MySupportsInt,  # <-- if you comment this out ...
    # SupportsFoo,  # <-- ... or uncomment this ...
    SupportsBar,
    Protocol,
    # metaclass=MyProtocolMeta,  # <-- ... or uncomment this ...
):
    pass

# ... then this error goes away
SupportsComposedThing.foo  # "Type[SupportsComposedThing]" has no attribute "foo"

assert SupportsComposedThing.__class__ is MyProtocolMeta  # <-- succeeds
MySupportsInt.foo  # <-- this is fine
SupportsFoo.foo  # <-- so is this
SupportsBar.foo  # <-- and this
# reveal_type(SupportsComposedThing)  # Revealed type is "def () -> test_case.SupportsComposedThing"

% mypy --version
mypy 0.910
% python --version
Python 3.9.7
% mypy --config=/dev/null test_case.py
/dev/null: No [mypy] section in config file
test_case.py:3: error: Module "typing" has no attribute "_ProtocolMeta"; maybe "Protocol"?
test_case.py:46: error: "Type[SupportsComposedThing]" has no attribute "foo"
Found 2 errors in 1 file (checked 1 source file)

[posita]

Wait, is this because the standard library's protocols themselves are lies? Why do (redundant?) definitions exist at all inside typing.pyi? If so, it may be a typeshed issue. But still, doesn't having to externally type your protocols sort of defeat the purpose of protocol types? What dark magic is this? Do we even need Supports definitions in typing.pyi anymore? (Apologies in advance for my ignorance. Or redundancy.)

---

Taking note of some history here to partly explain how we got here….

• Supports protocols are introduced into typing_extensions in early 2015, but protocols are still a private concept.
• typing.py makes it into CPython a couple months later with the initial set of Supports interfaces, but the notion of a protocol remains private.
• typing.pyi and the various Supports interface (re?)definitions are introduced several months later as ABCs (not protocols) in 2015.
• The ABCs in typing.pyi are converted to protocols in late 2017 alongside Mypy support for protocols as well as the public Protocol interface in typing.py via typing_extensions.
• Protocols officially make it into CPython in 2019.
• typing.pyi was later moved as part of a reorg earlier this year (2021).

[hauntsaninja]

mypy and other static type checkers don't look at CPython source code at all.

typeshed doesn't even have _ProtocolMeta (speaking generally, typeshed doesn't type typing.py internals too closely, since they tend to change a lot from version to version, type checkers special case them anyway, and those details tend to confuse users who aren't familiar with the inner workings of type checkers).

This means that _ProtocolMeta gets resolved to Any, so it's unsurprising to me that you get weirdness that only shows up later.

If you use the following I think you get the results you expect:

if TYPE_CHECKING:
    _ProtocolMeta = ABCMeta
else:
    from typing import _ProtocolMeta

[posita]

… [I]t's unsurprising to me that you get weirdness that only shows up later.

That's because you're a 17th level wizard specializing in typedivination. For the rest of us mere mortals, this is downright counterintuitive to say nothing of ergonomic. Why the inconsistent behavior among user-defined protocols and standard library ones? 😕 Or is that your point about "weirdness"?

Thanks for the work-around, though! ❤️