`issubclass()` assertions not narrowing types

[jwodder]

Consider the following (contrived) code:

from typing import Type

def untyped_function():
 return int

def foo() -> Type[int]:
 x = untyped_function()
 assert issubclass(x, int)
 return x

Running mypy on this with the --warn-return-any option produces the error "Returning Any from function declared to return "Type[int]"", even though the assert issubclass should convince mypy that x is of type Type[int] and not an Any.

Your Environment

• Mypy version used: 0.902
• Mypy command-line flags: --warn-return-any
• Mypy configuration options from mypy.ini (and other config files): none
• Python version used: Python 3.9.5
• Operating system and version: macOS 11.3.1

[kveretennicov]

We hit this issue in production code.

mypy Playground gist with repro and a workaround: https://mypy-play.net/?mypy=latest&python=3.10&gist=35deff0d7118af72d07a113a1f05c82b (mypy 0.961, Python 3.10)

[GertyP]

Just in case it's a useful test-case, I have a slight variation on this error, where it's clear from looking that we can deduce the narrowed return type in all conditional paths but where mypy complains that it thinks we could potentially be dealing with an 'Any' type -

def f(something: bool) -> dict:
 my_dict = None
 if something:
 exec('d = {}', globals(), locals())
 my_dict = locals()['d']
 assert isinstance(my_dict, dict)
 # ... No good if we use this assert.
 else:
 my_dict = {}
 # assert isinstance(my_dict, dict)
 # ... but all good if we use this assert
 return my_dict

[dave-v]

I think I'm seeing this error as well. Abbreviated example:

_registry: dict[str, type[Config] | Action]
...
for key, config in _registry.items():
 if config is NOOP: # NOOP is an Action object, and it's the only one that exists
 continue
 assert issubclass(config, Config)
 config.configure() # <- error: Item "Action" of "type[Config] | Action" has no attribute "configure" [union-attr]

(PyCharm's type checker is happy with this.)

If I add an additional assertion assert not isinstance(config, Action), then the error goes away, but I feel like that shouldn't be necessary: if config is indeed a subclass of Config, how can it also be an instance of Action? (Action inherits only from object.)

Edit: I found that assert not isinstance(config, Action) as the sole assertion works as well, but my question still stands - why isn't the issubclass check also enough to convince mypy?