Recursive sequence type of Sequence[NotAString] should not accept str

[jorenham]

With at least mypy 1.13.0, the false negative can be reproduced with:

from collections.abc import Sequence
from typing import TypeAlias

class Scalar: ...

Vector: TypeAlias = Sequence[Scalar]
Tensor: TypeAlias = Vector | Sequence["Tensor"]

accepted_vector: Vector = [Scalar()]     # true negative
rejected_vector: Vector = "duck"         # true positive: ducks aren't vectors

accepted_tensor: Tensor = [[[Scalar()]]] # true negative
rejected_tensor1: Tensor = object()      # true positive: objects aren't tensors
rejected_tensor2: Tensor = "duck"        # false negative: ducks are tensors...?

• mypy-play
• pyright-play (works as expected)

---

originally posted in #731 (comment)

[ilevkivskyi]

This works as expected (even if surprising) because str is a subclass of Sequence[str]. As an exercise, try writing a code where an actual runtime error happens because of this behavior (the point is that you can't).

[jorenham]

This works as expected (even if surprising) because str is a subclass of Sequence[str].

but str is not assignable to Scalar

edit: your argument assumes an infinite horizon, but within the context of typing, I think that a finite horizon would be more appropriate, analogous to how there's also a limit on the recursion depth at runtime

edit edit: if you'd parametrize str on its length then that will allow you to define an inductive basecase as str[1], and Sequence[str[1]] otherwise; so there's also no need for an infinite horizon

[jorenham]

As an exercise, try writing a code where an actual runtime error happens because of this behavior (the point is that you can't).

>>> import numpy as np
>>> np.array("not a tensor", np.float64)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
ValueError: could not convert string to float: 'not a tensor'

[jorenham]

FYI, this works as expected in pyright:

https://pyright-play.net/?strict=true&code=GYJw9gtgBAxmA28CmMAuBLMA7AzgOgEMAjGKdCABzBFSgGUkBHAVySxiQChRIpUBPCuiwBzMpWq0AKoKQBBeOgI5OnGPGU56MAhpAAuKHmOqAailTVDMivMXKoAXnpNW7JAG06OvQF1OUmw4VlA2dkpazuZo1FAAPi4sbBweAESBuNSp-pwEMBwUqEgAJgD6AG4WIdGWIE5QXj4EIAAUAJS%2BUF1dAMR8IKxQWEgiBBiVnCBIAFYWJRVVBlA1sc6pxcwwANap3Xu9-YNUOOjjSIYb21rNbADktJUxICq5%2BUiF80WZSxnBdc4eQHeXTNdq%2BcFQPqoAZIIYjMboCZTWZoT5BagARms6P%2BUDARBRqHa%2B0hh1hx1OiPOeIJFmuUyw9z4OJUyLmZS%2BfwATNjvvV1psdiSDsBdDhYcNRmcLoL6bDOdR8MYAPycIA

[davidhalter]

@ilevkivskyi I also feel like this is a Mypy issue. I am passing AFAIK all of Mypy's relevant recursion tests (except the ones with --disable-recursive-aliases) and this case seems to behave correct in my implementation.

[jorenham]

@KotlinIsland any thoughts on this?

[KotlinIsland]

I think its consistent with the concept/spec, but impractical... maybe. I think pyright is wrong/inconsistent here...

let's change how it works in basedmypy 🚀🚀🚀