This worked in Python 3.6, but in Python 3.7 and later creating a mock with a spec specifying a subscripted generic class does not mock any of the attributes of the class, because those attributes are not returned by dir().

For example:

# cat test.py
from typing import Generic, TypeVar
from unittest import mock

T = TypeVar('T')

class Foo(Generic[T]):
    def bar(self) -> None:
        pass

m = mock.MagicMock(spec=Foo[int])
m.bar()


# python3.11 test.py
Traceback (most recent call last):
  File "/root/test.py", line 11, in <module>
    m.bar()
    ^^^^^^^
  File "/usr/lib/python3.11/unittest/mock.py", line 635, in __getattr__
    raise AttributeError("Mock object has no attribute %r" % name)
    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
AttributeError: Mock object has no attribute 'bar'

This seems to be an issue with typing than mock since mock just uses the output from dir() . I am not able to bisect the relevant change but below is the output of dir(Foo[int]) in Python 3.6 and master.

Python 3.6.9

['__abstractmethods__', '__args__', '__class__', '__delattr__', '__dict__', '__dir__', '__doc__', '__eq__', '__extra__', '__format__', '__ge__', '__getattribute__', '__gt__', '__hash__', '__init__', '__init_subclass__', '__le__', '__lt__', '__module__', '__ne__', '__new__', '__next_in_mro__', '__orig_bases__', '__origin__', '__parameters__', '__reduce__', '__reduce_ex__', '__repr__', '__setattr__', '__sizeof__', '__slots__', '__str__', '__subclasshook__', '__tree_hash__', '__weakref__', '_abc_cache', '_abc_generic_negative_cache', '_abc_generic_negative_cache_version', '_abc_registry', '_gorg', 'bar']


master branch : 

['__args__', '__call__', '__class__', '__delattr__', '__dict__', '__dir__', '__doc__', '__eq__', '__format__', '__ge__', '__getattr__', '__getattribute__', '__getitem__', '__gt__', '__hash__', '__init__', '__init_subclass__', '__instancecheck__', '__le__', '__lt__', '__module__', '__mro_entries__', '__ne__', '__new__', '__or__', '__origin__', '__parameters__', '__reduce__', '__reduce_ex__', '__repr__', '__ror__', '__setattr__', '__sizeof__', '__slots__', '__str__', '__subclasscheck__', '__subclasshook__', '__weakref__', '_inst', '_name', '_paramspec_tvars', '_typevar_types', 'copy_with']

Isn’t the solution to use the unspecialized class?

Not exactly sure if this is a bug, but the reason is that Foo[int] used to be a class, now it's a plain object. It's a change brought in 3.7 by PEP 560.

3.6:
>>> isinstance(Foo[int], type)
True
>>> Foo[int].__dir__
<method '__dir__' of 'object' objects
>>> type(Foo[int].__dir__)
<class 'method_descriptor'>

main:
>>> isinstance(Foo[int], type)
False
>>> Foo[int].__dir__
<built-in method __dir__ of _GenericAlias object at 0x000001B44DF0A850>
>>> type(Foo[int].__dir__)
<class 'method_descriptor'>

The fix is just a 2-line change in either _GenericAlias or _BaseGenericAlias:
+    def __dir__(self):
+        return dir(self.__origin__)

Should we go ahead with this?

New changeset d6e13747161d7b634b47d2d3d212ed3be4a21fab by Ken Jin in branch 'main':
bpo-45755: [typing] Reveal class attributes of generic in generic aliases in `dir()` (GH-29962)
https://github.com/python/cpython/commit/d6e13747161d7b634b47d2d3d212ed3be4a21fab

Hi, I won't be backporting this to 3.9 since I don't like having jarring behavior changes so late into the bugfix cycle of a Python version.

I'm tempted to backport to 3.10. For now, I'll be conservative and just merge it in main. Please do tell me if any of you feel that I should backport to 3.10 too.

3.6-3.8 are in security-fix only mode so they won't get any bugfixes.

Thanks Kevin for the interesting bug report, and Karthikeyan for saving me a ton of time debugging this!

I think a 3.10 backport would be appreciated.

New changeset 87539cc716fab47cd4f501f2441c4ab8e80bce6f by Miss Islington (bot) in branch '3.10':
bpo-45755: [typing] Reveal class attributes of generic in generic aliases in `dir()` (GH-29962)
https://github.com/python/cpython/commit/87539cc716fab47cd4f501f2441c4ab8e80bce6f