Loaders and `exports` field

[arcanis]

As a practical example, let's say that we have:

/node_modules/foo/package.json

{
  "name": "foo",
  "exports": {
    "./utils": "./sources/utils"
  }
}

/node_modules/foo/sources/utils.ts

// Content doesn't matter

/index.js

import 'foo/utils';

And let's say that we have a composed pipeline:

• The PnP resolver

  • will get foo/utils as input
  • will want to turn the foo segment into /node_modules/foo (and leave the rest untouched)

• The TS resolver

  • will want to add the.ts extension to whatever foo/utils resolves into

What should return the PnP resolver?

a. "/node_modules/foo/utils"
b. "/node_modules/foo/sources/utils"
c. "/node_modules/foo/sources/utils.js"
d. ["/node_modules/foo", "./utils"]

Option A is incorrect; the exports field doesn't apply when a file is loaded through an absolute path:

It is not a strong encapsulation since a direct require of any absolute subpath of the package such as require('/path/to/node_modules/pkg/subpath.js') will still load subpath.js.

Option B is good, but requires the PnP loader to not only resolve foo into node_modules/foo, but also ./util into ./sources/utils. To do that requires either to reimplement the exports resolution (with the risks in terms of correctness that come with it), or Node to provide a import.meta.resolveExports('foo/utils', '/node_modules/foo') utility that would do the job for us. This is what we currently do, by embedding a copy of resolve.exports into our loader.

Option C essentially prevents the TS loader from working, since the extension is already resolved.

Option D (which isn't supported by Node at the moment) lets the PnP loader resolves the bare identifier part which Node would presumably process through the same import.meta.resolveExports function described earlier, but keeping it internal.

[JakobJingleheimer]

Thanks for this!

I think this is addressed by somehow exposing resolve utils and/or some flavour of createResolve.

Looking at this, this seems like it would be covered by the createResolve({ ...resolveUtilsOverrides }) I mentioned in PR#33. But you've already said that's too limiting. More of the discussion seems to be there, so let's resolve there though.