Test failure:

 ======================================================================
FAIL: test_no_types (test.test_dataclasses.TestMakeDataclass.test_no_types)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "D:\a\cpython\cpython\Lib\test\test_dataclasses\__init__.py", line 4111, in test_no_types
    self.assertEqual(C.__annotations__, {'x': 'typing.Any',
    ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                                         'y': 'typing.Any',
                                         ^^^^^^^^^^^^^^^^^^
                                         'z': 'typing.Any'})
                                         ^^^^^^^^^^^^^^^^^^^
AssertionError: {'x': typing.Any, 'y': typing.Any, 'z': typing.Any} != {'x': 'typing.Any', 'y': 'typing.Any', 'z': 'typing.Any'}
- {'x': typing.Any, 'y': typing.Any, 'z': typing.Any}
+ {'x': 'typing.Any', 'y': 'typing.Any', 'z': 'typing.Any'}
?       +          +       +          +       +          +

is expected, I just don't want to touch tests at this point.

I mentioned it on the issue but with this change typing is imported on class creation because the dataclass call will inspect the annotations and hence trigger the import. It also imports typing even if there are no untyped fields.

Python 3.14.0a0 (heads/main:5592399313, Jul 24 2024, 19:11:33) [MSC v.1938 64 bit (AMD64)] on win32
Type "help", "copyright", "credits" or "license" for more information.
>>> import sys
>>> from dataclasses import make_dataclass
>>> print(sys.modules.get("typing"))
None
>>> A = make_dataclass('A', [('a_var', str)])
>>> print(sys.modules.get("typing"))
<module 'typing' from '...\\cpython\\Lib\\typing.py'>

If you want to avoid the typing import this way you need to put some placeholder for dataclass to use and replace the attribute afterwards. Possibly also modifying Field.type to be a property that does the same replacement.

This doesn't happen on the 3.13 version proposed because dataclasses doesn't evaluate the strings so the import doesn't get called by dataclasses when it looks at the annotations.

You might be able to get around @DavidCEllis's point by adding some internal-only flag to tell @dataclass to skip introspecting the annotations.

Perhaps you can take advantage of dataclasses calling annotate with format.FORWARDREF and return something that looks like a ForwardRef there (or a real ForwardRef if there is a way to make this evaluate correctly if typing has been imported).

If it's called with format.VALUE you could then do the import and return the typing.Any object.

Perhaps you can take advantage of dataclasses calling annotate with format.FORWARDREF and return something that looks like a ForwardRef there (or a real ForwardRef if there is a way to make this evaluate correctly if typing has been imported).

Yes, this is my go-to idea right now :)

This PR still does not account for the problem mentioned by @DavidCEllis
It still imports typing when dataclass is created, because dataclasses accesses .__annotations__ inside prepare_class function.

I will try to work around this problem.

But, since dataclasses.prepare_class sets __annotations__ in the created class, __annotate__ won't ever be called again, so I am not sure what to do.

cc @JelleZijlstra about my question in #122285

Now make_dataclass does not import typing on dataclass creation 🎉
And everything else seem to work just fine!

Thanks everyone for your help and ideas! 🤝

I realized there's now new way to convert annotations to source. Fixing!

Blocked by #125507