[Attrs] Invalid typing of init arguments in subclasses of generics

[galcik]

from typing import Generic, List, TypeVar
import attr

T_BASE_ITEM = TypeVar("T_BASE_ITEM", bound="BaseItem")


class BaseItem(object):
 pass


class ConcreteItem(BaseItem):
 pass


@attr.s
class Container(Generic[T_BASE_ITEM]):
 items = attr.ib(type=List[T_BASE_ITEM], factory=list)


@attr.s
class ConcreteContainer(Container[ConcreteItem]):
 pass


container1 = ConcreteContainer()
container1.items.append(ConcreteItem())
container1.items.extend([ConcreteItem()])
container1.items = [ConcreteItem()]

container2 = ConcreteContainer(items=[ConcreteItem()])

fails ONLY on the last line:

List item 0 has incompatible type "ConcreteItem"; expected "T_BASE_ITEM"

It seems that typing for the attribute items works as expected. However, if items is used as a keyword argument of the __init__ generated by attrs, typing fails.

Note also that if the following class is added:

@attr.s
class ItemWrapper(Generic[T_BASE_ITEM]):
 item = attr.ib(type=T_BASE_ITEM)

mypy fails with error:

'T_BASE_ITEM' is a type variable and only valid in type context

[ilevkivskyi]

Yes, this is clearly a bug. The cause is that __init__ is generated for every subclass that is an attrs class. When __init__ (or any other attribute) is accessed on a superclass of a normal class, mypy takes care to map_instance_to_supertype/expand_type_by_instance, while the plugin doesn't do this.

@euresti will you have time to work on this?

[ilevkivskyi]

Btw, most likely dataclasses have the same problem.

[natemcmaster]

Thanks for the tip about map_instance_to_supertype. I have proposed a fix in #9383 which uses that API to resolve typevars.