Steal uses `browser` field in package.json even when running on Node

[stucox]

Test case — using the facts that:

• Nunjucks has both browser and Node versions
• It declares the location of its browser build with a browser field in package.json (which I know Steal observes)
• nunjucks.FileSystemLoader is only available in the Node environment

— so when loading nunjucks on Node (with or without Steal), we’d expect nunjucks.FileSystemLoader to be available.

$ npm install -DE steal
$ npm install -DE nunjucks
$ node
> var nunjucks = require('nunjucks')
> var steal = require('steal')
> !!nunjucks.FileSystemLoader
true
> steal.import('nunjucks').then(nunjucks => console.log(!!nunjucks.FileSystemLoader));
[...]
false

I’ve checked with a bit of debugging that Steal is indeed loading nunjucks/browser/nunjucks.js, rather than nunjucks/index.js.

Removing the browser field from Nunjucks’ package.json and repeating the above throws an error (within Nunjucks) re. fs not being available, which at least implies it’s loading the version intended for Node (which hopefully I can solve that with steal-builtins or similar… anyway, that’s my problem to deal with).

[matthewp]

Running Steal code in Node is a general problem. Some times you want it to use the browser field, such as when you are doing a build, and then some times you want to run the code in a Node context.

I'm not sure how to fix this, to be honest. I plan on making some larger architecture changes and want to outline those in https://github.com/stealjs/rfcs. Leaving this open so I'm sure to consider this case.

[stucox]

Having thought about this a bit, I settled on the following (fairly non-revolutionary) mindset:

• For me, Steal is all about portable code — which shouldn’t assume either platform. Ideally a package wouldn’t have a “browser mode” / “node mode”… or would at least make the bulk of its functionality available from a generic entry point (which Steal could use), with browser / Node wrappers if necessary.
• That said, it’s not feasible to impose this on every library we might want to use (aintnobodygottime to raise that many PRs), so maybe simply making this configurable per package is enough? Then if necessary someone can provide different system configs for browser / server. A bit more manual, but more control to the user.

P.S. I like the rfcs repo idea!