With a switch to Only One Webpack™ we need a way to help developers generate the component index without a concurrent watch task. The best way to do this is to have developers import their components, but how do they do that when we support skins? The answer in this commit is to change skinning.
Skinning now expects to receive your list of overrides instead of the react-sdk+branded components. For Riot this means we send over *only* the Vector components and not Vector+react-sdk.
Components can then be annotated with the `replaceComponent` decorator to have them be skinnable. The decorator must take a string with the dot path of the component because we can't reliably calculate it ourselves, sadly.
The decorator does a call to `getComponent` which is where the important part of the branded components not including the react-sdk is important: if the branded app includes the react-sdk then the decorator gets executed before the skin has finished loading, leading to all kinds of fun errors. This is also why the skinner lazily loads the react-sdk components to avoid importing them too early, breaking the app.
The decorator will end up receiving null for a component because of the getComponent loop mentioned: the require() call is still in progress when the decorator is called, therefore we can't error out. All usages of getComponent() within the app are safe to not need such an error (the return won't be null, and developers shouldn't use getComponent() after this commit anyways).
The AuthPage, being a prominent component, has been converted to demonstrate this working. Changes to riot-web are required to have this work.
The reskindex script has also been altered to reflect these skinning changes - it no longer should set the react-sdk as a parent. The eventual end goal is to get rid of `getComponent()` entirely as it'll be easily replaced by imports.
If we have account data in local storage but nothing in the crypto store, it
generally means the browser has evicted IndexedDB out from under us. This adds a
modal to explain the situation and offer to completely clear storage to get
things back to normal.
Fixes https://github.com/vector-im/riot-web/issues/9109
Hopefully makes the syntax a bit nicer. Also uses ES6 async import
rather than require.ensure which is now deprecated. Also also
displays an error if the component fails to load rather than falling
over in a heap, which is nice.
It turns out that the assertion made in
https://github.com/matrix-org/matrix-react-sdk/pull/1213 about `async`
functions returning bluebird promises was only correct when babel used an
inline version of the `asyncToGenerator` helper; in react-sdk we are using
`babel-transform-runtime` which means that we use a separate
`babel-runtime/helpers/asyncToGenerator`, which returns a native (or core-js)
Promise.
This meant that we were still in the situation where some methods returned
native Promises, and some bluebird ones, which is exactly the situation I
wanted to resolve by switching to bluebird in the first place: in short,
unless/until we get rid of all code which assumes Promises have a `done` method
etc, we need to make sure that everything returns a bluebird promise.
(Aside: there was debate over whether in the long term we should be trying to
wean ourselves off bluebird promises by assuming all promises are native. The
conclusion was that the complexity hit involved in doing so outweighed any
benefit of a potential future migration away from bluebird).
It does something, but things I don't like:
* it churns for 15 seconds webpacking everything. Do we really need to get
webpack involved here?
* I don't think there's any way to control which tests get run and which don't.
Other things I'd want to fix up include:
* Make it run on jsdom or phantomjs instead of Chrome
* figure out how to configure babel without a .babelrc
Travis Ralston