[1.16 regression] narrowing on is_dataclass

[henryiii]

Bug Report

If you use isinstance(item, type), this used to be used in conjunction with dataclasses.is_dataclass to identify types and instances. It's the recommendation in the CPython documentation, in fact: https://docs.python.org/3/library/dataclasses.html#dataclasses.is_dataclass. However, the latests MyPy no longer narrows this correctly (from what I can tell, unless I'm doing something wrong), and still thinks I might have an instance, which then breaks later when creating an instance, etc.

Note that I can flip the order, in which case I just get builtins.type for both versions. This isn't enough to fix it for some of the more complex usages in scikit-build-core, though.

To Reproduce

import dataclasses

raw_target: object

if isinstance(raw_target, type) and dataclasses.is_dataclass(raw_target):
    reveal_type(raw_target)

Expected Behavior

This is from 1.15:

tmp.py:6: note: Revealed type is "Type[_typeshed.DataclassInstance]"

Actual Behavior

tmp.py:6: note: Revealed type is "Union[_typeshed.DataclassInstance, Type[_typeshed.DataclassInstance]]"

Your Environment

• Mypy version used: uvx --from git+https://github.com/python/mypy mypy tmp.py was used, so latest commit, also latest release 1.15 version for comparison.
• Python version used: 3.13

Noticed in hauntsaninja/mypy_primer#174.

[sterliakov]

This is a regression, and a rather bad one. Note also that this is somehow dependent on checks order. MRE without stdlib dependency:

from typing import TypeIs

class A: ...

def is_a(o: object) -> TypeIs[type[A] | A]:
    return isinstance(o, int)

raw_target: object

if is_a(raw_target) and isinstance(raw_target, type):
    reveal_type(raw_target)  # N: Revealed type is "builtins.type"
if isinstance(raw_target, type) and is_a(raw_target):
    reveal_type(raw_target)  # N: Revealed type is "Union[type[__main__.A], __main__.A]"

https://mypy-play.net/?mypy=master&python=3.13&flags=strict&gist=6e2a57a00f5aa275ebbb2abbefa972e6

My MRE exhibits the same behaviour on 1.15.0, so it isn't the same. And this is correct - A and type can have a common LSP-compatible subclass.

This issue bisects to #17678; is_dataclass is an overloaded function:

# HACK: `obj: Never` typing matches if object argument is using `Any` type.
@overload
def is_dataclass(obj: Never) -> TypeIs[DataclassInstance | type[DataclassInstance]]: ...  # type: ignore[narrowed-type-not-subtype]  # pyright: ignore[reportGeneralTypeIssues]
@overload
def is_dataclass(obj: type) -> TypeIs[type[DataclassInstance]]: ...
@overload
def is_dataclass(obj: object) -> TypeIs[DataclassInstance | type[DataclassInstance]]: ...

And this looks like a bug - after first isinstance we should have raw_target already narrowed to type, so the second overload should match.

[sterliakov]

@hauntsaninja could you please create a label for PEP 742 / TypeIs?

[sterliakov]

I get it now. When find_isinstance_check_helper is applied, we check all arguments at once, which means wrong @overload is selected - mypy only stores narrowed types when all typemaps are known. cc @sobolevn as original author of #17678 - this sounds like a painful limitation, any overload resolution here won't take preceding narrowing statements into account.

Splitting the ladder works:

import dataclasses

raw_target: object

if isinstance(raw_target, type): 
    if dataclasses.is_dataclass(raw_target):
        reveal_type(raw_target)  # N: Revealed type is "type[_typeshed.DataclassInstance]"