@konstin thanks for this!

Could you also announce this PR on discuss.python.org/packaging. Since this could be considered a spec change, it is worth announcing broadly (I personally don't think this change requires a PEP, but better check).

Could you also announce this PR on discuss.python.org/packaging. Since this could be considered a spec change, it is worth announcing broadly (I personally don't think this change requires a PEP, but better check).

Sure: https://discuss.python.org/t/clarify-that-the-direct-url-json-url-field-must-be-a-spec-compliant-url/46518

I wonder if this is too restricted regarding to git urls. If I understand correctly, you can't just convert a scp-like url to a valid url, cf servo/rust-url#220 and linked issues for example. If you can't convert it what should an installer that supports scp-like urls write into direct_url.json? Or does this change mean that installers should not support scp-like urls?

Wouldn't prefixing them with ssh:// make them valid urls?

I assume for github they are equivalent, but if this statement is correct they are not always equivalent:

However, the user@host and ssh:// url forms are not equivalent--in the user@host form, the path is usually relative to the user's homedir, while in the ssh:// form, the path is relative to root. (This is the behaviour when pulling over a vanilla ssh+shell setup--obviously hosting tools can make their own choices.)

(from https://secure.phabricator.com/T11004)

I have no idea how common that is. Maybe, that's just an edge case that can be neglected?

FYI: We decided to do the normalization in Poetry for now even if it might not always be equivalent in theory.

@konstin Could you rebase?

Thank you!