[srittau]

• Don't use deprecated pattern Incomplete | None = None.
• Use Iterable instead of Sequence as invariant alternative to list. (Keep Mapping as alternative for dict.)
• Remove "Don't use int | float. See Replace float with float | int typeshed#16059)

[jorenham]

Don't use deprecated pattern Incomplete | None = None.

Could you elaborate?

[srittau]

x: Incomplete | None = None is functionally equivalent to x=None, which is why we use the latter nowadays in typeshed. This prevents the argument to be flagged as "annotated" when it really isn't and is more readable. See also https://github.com/PyCQA/flake8-pyi/blob/main/ERRORCODES.md#Y067 and python/typeshed#14029.

[srittau]

Actually not functionally equivalent, but worse, since Incomplete = Any, so x: Incomplete | None = x: Any | None, while x=None = x: Unknown | None.

[jorenham]

since Incomplete = Any, so x: Incomplete | None = x: Any | None, while x=None = x: Unknown | None.

This is what Pyright does, but I'm pretty sure this inference behavior isn't defined the typing spec.

[srittau]

Sure, other type checkers might handle it differently, but avoiding Incomplete (i.e. an explicit Any) and an unnecessary annotation is still worthwhile, which is why we've been doing it in typeshed for 1.5 years now and added a check to flake8-pyi.

[jorenham]

My point is that we shouldn't advice using def incomplete2(z=None) -> None: ... because that relies on undefined behavior.

[srittau]

I disagree. This is the style section of a stub writing guide that mostly serves to reflect typeshed's best practices. Using x: Incomplete | None = None is not a best practice and actively discouraged for the reasons given. It uses the same undefined behavior as other items that are left unannotated, and we explicitly recommend to leave items unannotated.

[jorenham]

I guess I then disagree with typeshed's best practices then. Because in my view, it's always better to explicitly annotate something and follow the typing spec, than to rely on the particular design choice of certain type-checkers that's undefined in the spec.

[jorenham]

I don't agree with this, see #2387 (comment). So can we leave this out? The other bits I'm fine with, so if you want we can discuss this further in a separate PR for this part.

[AlexWaygood]

I feel like it's basically a de-facto standard at this point to treat an unannotated parameter as semantically distinct from a parameter annotated with Any (or an alias to Any), even if they are ultimately understood by the type checker as signifying that any value can be passed as an argument to the parameter without the call erroring. Mypy/zuban both emit no-untyped-def for the following snippet, pyright emits reportUnknownParameterType, pyrefly emits implicit-any-parameter, and ty will (after astral-sh/ruff#29917 is merged) complain with missing-parameter-annotation. Ruff also has rules that complain about missing parameter annotations. (Note that in all cases, these rules are disabled by default, and must be explicitly selected in configuration or similar.)

In all these cases, the type checker or linter is trying to alert you to the fact that your type hints are incomplete. But, perversely, explicitly annotating something with Incomplete silences all of these type checkers/linters: they no longer detect that your type hints are incomplete, giving you a misleading impression about the state of the type hints in your project.

[jorenham]

Standard or not, with the typing spec living next door, I still think we should not advice users to rely on undefined behavior, especially not if there are viable alternatives that are supported by the spec.

[jorenham]

And to be clear: I have no problem with typeshed using this (or anyone else for that matter). I actually agree that it's more honest to not annotate something instead of using Incomplete or something. And it's no secret that I go off-piste in the stubs I maintain myself. It's just that I don't think we should officially recommend doing this (or at least not until it's specified in the spec); that's all.

[srittau]

x=None is as much "undefined behavior" as every other unannotated item. Ultimately that means that we should remove the recommendation to leave items unannotated, which does not sound right at all.

[jorenham]

Ultimately that means that we should remove the recommendation to leave items unannotated, which does not sound right at all.

I actually think we should indeed not recommend that :p.