Permit inferring Self for unannotated self

[hauntsaninja]

https://discuss.python.org/t/permit-inferring-self-for-unannotated-self/65190
https://typing--1860.org.readthedocs.build/en/1860/spec/annotations.html

[stroxler]

Overall looks good to me but I was confused by some of the wording

[stroxler]

I'm not 100% sure what this last sentence means.

I think type[Self] is a fairly precise type - it's a type wrapping a type variable bound above by type[<current-class>] but can be narrowed on callsites where we know more.

That seems as precise as I can imagine being? And I don't think it's any less precise than the type of self, which behaves similarly.

[hauntsaninja]

Yeah, that sentence is from PEP 484, it confused me as well. I thought it was maybe talking about the usual thing where args to cls(...) are hopeless, but I don't really know. Maybe we just remove it?

[JelleZijlstra]

Yes, that sentence makes no sense to me either, I think we should remove it.

[erictraut]

Except for the last sentence, I think it looks reasonable.

I do think the spec should eventually be updated to mandate that self has the implied type Self and cls has the implied type type[Self]. Many important use cases involving the Self type break if this isn't the case. In retrospect, PEP 673 (which introduced the Self type) should have specified this IMO. But if you'd prefer to hold off on that change, then your proposed wording seems fine.

[carljm]

Thanks for clarifying this!

[carljm]

I don't think it is accurate to make an explicit equivalence here between the type Self and "the type of the containing class". This is not actually what the specification for Self says: Self is not "the type of the containing class," it is "a typevar with a bound of the containing class". This distinction is important for how the Self type is handled when used in other arguments, and in inheritance cases.

I suggest we simply don't attempt a mini-definition of the meaning of Self here, and allow the Self spec to handle this in depth.

Suggested change

	assumed to have the type of the containing class or :ref:`Self
	<self>`, and for class methods the type object type corresponding to
	the containing class object or ``type[Self]``.
	assumed to have the type :ref:`Self <self>`, and for class methods ``type[Self]``.

[JelleZijlstra]

I interpreted this as not an attempt to define Self, but a way to permit type checkers to use an alternative to inferring Self. A type checker may either infer self to be of the type of the containing class (as mypy does: https://mypy-play.net/?mypy=latest&python=3.12&gist=ab4b370dd76d1d446bde35bca7de7bf6 ), or of type Self (as pyright does: https://pyright-play.net/?strict=true&code=MYGwhgzhAEAaBcAoaLoBMCmAzaBbDALgBYAUEGIWAlNALQB80AcgPYB2GSq30AThgDcMYEAH0CATwAOGMhWpA ).

[carljm]

Oh! I see how that could be the interpretation of the "or" here. If that's the intention, then I suggest we make it a bit more explicit. (I'll make a separate suggested edit, since it would involve one more line than I included in this one.)

[carljm]

Personally I think inferring Self is the better option here, and we could also consider just specifying that, but I see that may be a bigger change than was intended in this PR.

[carljm]

Suggested change

	If the argument is not annotated, then for instance methods it is
	assumed to have the type of the containing class or :ref:`Self
	<self>`, and for class methods the type object type corresponding to
	the containing class object or ``type[Self]``.
	If the argument is not annotated, then for instance methods it may be
	inferred to have either the type of the containing class, or the type :ref:`Self
	<self>`. For class methods it may be inferred to have either the type object
	type corresponding to the containing class object, or ``type[Self]``.