Review requested:

• @nodejs/modules

I'm a bit concerned about two pieces here:

1. The "*" syntax was previously valid as an exports pattern afaik and also may end up appearing in import maps, potentially with different semantics[1]. So we're potentially conflicting with two things here.
2. If this or something very similar doesn't end up in import maps, this is the moment where just the package.json data is no longer enough to convert exports to an import map. Building an import map now requires downloading the package contents and scanning for matches. Ad-hoc resolution still works fully in-memory (no file access required) but import map-based resolution[2] doesn't.

[1] Relevant import maps issue: WICG/import-maps#7
[2] By "fully in-memory import map-based resolution" I mean getting the package.json metadata, transforming it into an import map, and then running "vanilla" import map resolution based on it - all without having access to the package contents.

All that said: I am very excited about this refactoring!

Note as well that if we no longer have the ability to generate an import map directly from package.json, then many of the arguments against automatic index and extension resolution no longer apply, so this will have an impact on that discussion.

The "*" syntax was previously valid as an exports pattern afaik and also may end up appearing in import maps, potentially with different semantics[1]. So we're potentially conflicting with two things here.

The semantics in [1] are actually incredibly similar. Thanks for pointing this out I had forgotten about that!

If this or something very similar doesn't end up in import maps

That will take many years at this rate. Another good reason to have the import maps discussion - Node.js does not have to be a "passive player" when it comes to import maps, it might have a role in actively driving forward concepts around them.

Note as well that if we no longer have the ability to generate an import map directly from package.json, then many of the arguments against automatic index and extension resolution no longer apply, so this will have an impact on that discussion.

@ljharb the pattern can be seen as a file system based expansion to build the full import map. That is import map = function (package.json + file system). Agreed the same can be said for extension searching. Bare specifiers, even with subpaths, are a very separate case to relative though. We do not have extension searching on the agenda though, do you want to re-add a discussion / proposal here?

@guybedford I think it can wait on the outcome of this. We've already debated extension searching extensively, and recently discussed how its proponents feel frustrated that "no searching" seems to have won via inertia. However, if the landscape changes by adding this proposal, I would be happy to add an agenda item to discuss enabling normal node resolution by default for ESM.

@ljharb ok, that would likely set this proposal back though. Did you follow my argument wrt bare specifier extension searching versus relative specifier extension searching?

I don't think they're actually different at all (in the absence of "exports", or in a LHS-exposed directory). In node, specifiers are specifiers, it's just that the filesystem context might determine what file is loaded.

@ljharb the difference is that with "exports" patterns there exists a processing step to enumerate all "exports" from the pattern by looking at the files in the package, because the space of all possibilities of specifiers within the package is full enumerated / encapsulated by exports. With relative or absolute specifiers, the space of all possibilities is the entire user file system - not something which is practical to enumerate.

@guybedford ah, i see what you're saying. I think that "relative paths below package.json" and "bare identifiers" should both be fine in that regard, and I'm still of the position that ESM should never allow filesystem-absolute paths anyways.

I've rebased this PR to a separate refactoring PR in #34744.

@guybedford I've updated the title based on the refactor being refactored :P

Any further feedback on this PR @nodejs/modules-active-members? I still think this is a necessary extension for the sorts of use cases we have to deal with for large numbers of exports, where directory exports may expose unwanted modules and force the usage of file extensions.

The patterning system is based on always ending in a wildcard, with arbitrary number of replacements of that pattern on the RHS. If anyone has use cases, feedback or suggestions or edits to this model please say so.

Any explicit concerns would be great to hear as well.

One important property of this that convinced me: It allows writing generic specifiers for dual packages. This may be worth mentioning as an example?

"exports": {
  "./features/*": {
    "require": "./features/*.cjs",
    "import": "./features/*.mjs"
  }
}

// Now both of these work:
require('pkg/features/some-feature');
import('pkg/features/some-feature');