n00b question: how do I run the Python tests in my local clone of this repository? Wanted to experiment.

First compile Python, as explained in "Quick reference" at https://devguide.python.org/. To run just the typing tests, the easiest way is just ./python.exe Lib/test/test_typing.py or similar depending on your OS and CWD.

n00b question: how do I run the Python tests in my local clone of this repository? Wanted to experiment.

First compile Python, as explained in "Quick reference" at https://devguide.python.org/. To run just the typing tests, the easiest way is just ./python.exe Lib/test/test_typing.py or similar depending on your OS and CWD.

Thanks, Jelle. Should have RTFM'ed but was being lazy :|

@mrahtz Some tests fail locally (and on CI, apparently).

@gvanrossum @JelleZijlstra @pradeep90 There's a high-level design point I've just realised we should discuss.

In the PEP itself, we stated that Unpack would primarily be for backwards compatibility. But in the current iteration of this PR, we're actually using it as the wrapper for unpacked types in general. For example, TypeVarTuple.__iter__ returns Unpack[self].

(@gvanrossum, in case you haven't followed the chain of comments on the PR so far, there were two main reasons to try doing it this way: 1. It means we can avoid the need for a middleman classes UnpackedTypeVarTuple and StarredTuple like we had before; and 2. It means we can also easily support unpacking of other types at runtime, which might be useful for people wanting to experiment with typing features - e.g. Jelle mentioned he'd like to experiment with Unpack[TypedDict type].)

This raises two questions:

1. How much does it matter if the way that Unpack is used in typing.py doesn't match the spirit in which the PEP introduced it? (It feels a bit ugly to me, but I wonder whether I'm being too perfectionist. An alternative might be to have two separate things, Unpack as the backwards-compatible version of *, then a separate wrapper Unpacked for eliminating the middleman classes.)
2. What should we do about unpacking a native tuple? In bpo-43224: Implement PEP 646 changes to genericaliasobject.c #31019, this essentially returns a special native 'starred tuple' type. I was about to change it so it instead returned typing.Unpack[self] - but I realise this is against the precedent set by the new union operator, where Foo | Bar also returns a special native type types.UnionType rather than typing.Union[Foo, Bar]. @gvanrossum in particular, do you know whether this was a deliberate decision? Is this a precedent we should respect?

Thanks for bringing this up! Here are my takes:

1. I agree there's some tension here, but I'd rather keep the number of classes lower. Having a separate Unpacked and Unpack doing virtually the same thing would be confusing.
2. types.UnionType was created specifically because we didn't want built-in syntax to create an instance of a Python-defined class. We should do the same for the PEP 646 syntax.

Exactly what Jelle says.

@JelleZijlstra @gvanrossum Thanks for confirming re Unpack design. In that case, we'll stick with the plan where unpacking things from typing.py (that is, Tuple and TypeVarTuple) results in an Unpack[], and unpacking a native tuple results in a special version of the native tuple type that's starred.

(Fixed a merge conflict)

On the subject of what to do about type substitution, Pradeep had the excellent idea of moving discussion to a future PR, given how long even this PR has taken. For now I've excised all the extra logic required, instead raising NotImplementedError in the cases it means we currently don't support (re-purposing the corresponding tests so it's clear exactly which things we currently can't deal with).

For what it's worth, though, given that both Pradeep and Jelle lean towards thinking simple is better, for the follow-up PR, I'll try drafting a prototype of Jelle's suggestion: having the subscription operator just return a new GenericAlias.

P.S. bedevere-bot: I have made the requested changes; please review again.

Thanks for making the requested changes!

@JelleZijlstra: please review the changes made to this pull request.

Congratulations!

Thanks @mrahtz for the code, and thanks @JelleZijlstra for the review. Thanks @pradeep90 for your help. This is a monumental change!

Fantastic! Thank you @JelleZijlstra and @pradeep90 for the review, and @gvanrossum for your continuing support :)