Imports proposal - first draft

[guybedford]

This provides details on the corresponding package "imports" proposal as the second field as part of this proposal, extending the similar types of mappings we have for "exports" into cases that can work for internal aliasing as well.

Notes have been provided where details are still to be determined, in particular whether or what custom symbol will be used for disambiguation in mappings.

Help fleshing out more edge cases in the examples here would be great.

[guybedford]

//cc @robpalme @devongovett

[ljharb]

🎉

[ljharb]

cc @GeoffreyBooth i think this exact use case is a reason it’s useful, in that other thread, to have exports override whatever main points to, instead of providing main itself.

[jkrems]

Can you elaborate? This suggests having: {"main": "./foo.mjs", "exports": false} which would break if exports would completely overwrite main..? It does work if it desugars to merge semantics (Object.assign({ ".": mainValue || [] }, exports)).

[ljharb]

that’s a good point, but I’m more thinking conceptually.

[GeoffreyBooth]

i think this exact use case is a reason it’s useful, in that other thread, to have exports override whatever main points to, instead of providing main itself.

I’m not sure I follow; you mean the reason that exports: false is useful?

A package needs to export something, right? exports: false and no main would be the package equivalent of unreachable code?

I think I agree with @jkrems that merge semantics make the most sense for this, since overriding main with exports: false would have the effect of creating an unusable package in all versions of Node that support exports. We’re obviously not dropping support for main, so if exports allows for setting the main as well it would just take precedence.

[ljharb]

i don’t mean that. I mean that exports has the ability to override whatever main points to. exports false wouldn’t be using that ability.

[GeoffreyBooth]

It’s a question of semantics. The docs could simply define exports: false as “no exports other than whatever is defined in "main".” or something like that. That’s probably what they should say.

Because if you think about it, the lack of "." in exports wouldn’t mean that there’s no main export defined for the package. I.e. we’d want to support:

"main": "./main.js",
"exports": { "./foo": "./foo/index.js" }

Where require('pkg') and require('pkg/foo') both work. This would be an example of the merging/Object.assign approach, rather than exports overriding main in the sense that this package would not have any main export defined because there’s no "." in exports.

[jkrems]

It sounds like you both support the same semantics and the only disagreement is how it should be worded? My vote is for calling it "merge" because as @GeoffreyBooth points out, it's only the "." key that actually overrides anything. But I think everyone agrees that we should bring back dot-main in some form (the ability to specify main in exports in a way to wins over the top-level main field).

[GeoffreyBooth]

Perhaps "." should have its own small section in the docs. Something like:

"." can specify the root entry point for a package, i.e. the file loaded for require('pkg') or import 'pkg'. Historically this has been defined by the package.json "main" field, and "main" will continue to be supported. If "main" is defined and "." is not, the value in "main" will determine the package root entry point; if both "main" and "." are defined, the latter takes precedence. If "main" is defined and exports is set to false, the value in "main" is the only export available for the package.

Since "main" doesn’t define an object it’s a little weird to discuss merging at all. We’re really only talking about the precedence of "main" versus ".".

[guybedford]

Hopefully we can discuss this main problem at the coming meeting.

I would ideally like to get the same spec here merged so that we can also start to discuss it further though, and wouldn't want these main considerations to block that.

Is there something we can do here to resolve this further? Any note or clarifications?

[GeoffreyBooth]

I don't think there's any disagreement? It seems like we're just discussing the best way to explain the behavior to users, which is something we can worry about when we get to writing the docs.

[ljharb]

we don’t have to bikeshed here, but i prefer ~

[guybedford]

Definitely open to that.

[ljharb]

what if package A has an imports map of “x” to “B/y”, and package B has an imports map of “A/x”?

[guybedford]

Do you mean something like:

A/main.js import '#x'
B/y.js: import 'A/main.js'

where #x -> B/y and B/y -> B/y.js is an exports map in B?

That would construct a cycle between the modules, yes, but the cycle point was about the resolver itself getting caught in a resolution cycle.

For example, if we supported package exports mapping into external packages then you could set up a cycle where B/x -> A/x -> B/x sort of thing, where the resolver cannot complete the resolve operation.

[jkrems]

Happy to land this as-is (maybe with the minor typo fix) and go from there!

[bmeck]

I'm curious how this interacts with https://github.com/nodejs/node/blob/master/doc/api/policy.md#dependency-redirection which are controlling at the app level and not package level.

[jkrems]

My gut feeling would be "policy applies to the raw specifier, preventing any resolution including imports". But it may be worth clarifying it in the proposal.

[guybedford]

Seems like we may be good to merge here then?

Further feedback welcome.

[devongovett]

I like ~ for this instead of #. Parcel and a few other tools already support ~/foo to mean foo within the nearest folder with package.json. Supporting ~foo (without the slash) to mean the foo named import within the package.json kinda makes sense too. ~ always refers to the folder with package.json, and you can either refer to a file or a named import from there. Not sure if Node is interested in the ~/ specifier as well, but it would leave the option open for the future.

[jkrems]

I think node may be interested in ~/ (or something like it) to mean "this package as exported". Right now there's no good way to unit test the public interface when using exports for example. If not ~, we'd need to find another character for this.

[devongovett]

I guess that's what I meant. ~/ gets you to the package root, and it's normal resolution after that. So if there is a foo export, then ~/foo would refer to that as it would normally. And ~foo would refer to an import. ~ or ~/ by itself could refer to main.

[jkrems]

Is it worth opening a dedicated issue to resolve the sigil question? My current thoughts:

We have three kinds of specifiers that people have asked for (implied: that we may or may not want to support):

1. Getting the public interface of the package. This will allow people using exports to actually unit test their packages (since exports do not apply to relative specifiers). Examples: ✩ (the main/default export), ✩/subpath
2. Adding custom aliases that are only valid inside of the package boundary. Examples: ✩data/emoji.json, ✩fetch.
3. Accessing arbitrary paths relative to the package boundary. Examples: ✩/src/model.mjs.

Of these, only (1) and (3) actually conflict. (2) could share a symbol with either one of them. A concern raised by @guybedford was that if (1) and (2) share a symbol, it may be confusing.

So to me the options are:

1. One sigil, no (3):
   • Use ~ and ~/ to mean "this package as if it was imported by name".
   • Allow ~<name> to be used for custom aliases within the package.
2. Two sigils, optional support for (3):
   • Use ~ and ~/` to mean "this package as if it was imported by name".
   • Use #<name> for "private names", aliases only visible inside of the package.
   • (optional) Use #/ for "paths relative to the package boundary".

For packages that don't use exports, ~/ and #/ are effectively the same but that would change once they choose to remap subpaths and/or lock themselves down.

[jkrems]

My preference right now would be ~ + #<name> without support for importing non-public paths relative to the project root. There would still be design space for adding #/ in the future if it becomes truly necessary.

[devongovett]

Isn't # problematic since it's meaningful to URL parsing? ESM specifiers are URLs, so wouldn't that be considered a hash?

[jkrems]

While ESM uses URLs as cache keys in the browser, it has more restrictions for specifiers. The only kinds of specifiers it allows are:

1. Relative specifiers starting with ./, ../, or /.
2. An absolute URL, including protocol.

See: https://html.spec.whatwg.org/#resolve-a-module-specifier

So neither ? nor # may start a specifier unless something like an import map is involved. Even though both are valid relative URLs in other contexts like certain HTML attributes.

[guybedford]

I thought we would promise not to bikeshed yet :P

But if we must then my preference would be to use ~/ for the internal root and #/ or something else for the public interface.

Under that logic, perhaps we should use ~name?

But yeah opening a new issue to hash / tilde this out seems to make sense!

[jkrems]

Forked this to an issue: #41

[jkrems]

(Resolved conflicts using UI after merging the validations PR. I think I picked the right lines.)

[guybedford]

Looks good, shall we merge and move the symbol discussions to their own threads then?

[guybedford]

(I can also get behind changing to ~ if people want now too)

[jkrems]

Let's merge and discuss the sigil question separately (#41).