This is an experimental implementation to replace the current live development architecture with something more flexible that isn't tied solely to Chrome Developer Tools. 
It's based on the current Live Development code in Brackets and has been incubated at njx/brackets-livedev2.

### What's working

It can be enabled by setting ```livedev.multibrowser``` preference to true. If enabled, it will launch will launch the page in your default browser when clicking on the LiveDevelopment icon as it works today. You should then also be able to copy and paste the URL from that browser into any other browser, live edits will then update all connected browsers at once.

This is still an experimental implementation, the basic functionality for CSS/HTML editing is working but there are could be some scenarios that might be partially or entirely not covered yet. Page reload when saving changes on related Javascript files is also working. Documents that are loaded by the current HTML live doc are being tracked by ```DocumentObserver``` at the browser side which relies on DOM MutationObserver for monitoring added/removed stylesheets and Javascript files (just monitoring ```<link>``` and ```<script>``` added/removed nodes so far. 

### Basic architecture

The primary difference in this architecture is that communication with the browser is done via an injected script rather than CDT's native remote debugging interface, and the browser connects back to Brackets rather than Brackets connecting to the browser. This makes it so:

* launching a preview, injecting scripts into the HTML, and establishing the connection between the previewed page and Brackets are relatively simple and largely decoupled
* live preview can work in any browser, not just Chrome
* multiple browsers can connect to the same live preview session in Brackets
* browsers could theoretically connect from anywhere on the network that can see Brackets (though right now it's only implemented for localhost)
* opening dev tools in the browser doesn't break live preview

Communication between Brackets and the browser is factored into three layers:

1. a low-level "transport" layer, which is responsible for launching live preview in the browser and providing a simple textual message bus between the browser and Brackets.
2. the "protocol" layer, which sits on top of the transport layer and provides the actual semantic behavior (currently just "evaluate in browser")
3. the injected RemoteFunctions script, which is the same as in today's LiveDevelopment and provides Brackets-specific functionality (highlighting, DOM edit application) on top of the core protocol.

The reason for this factoring is so that the transport layer can be swapped out for different use cases, and so that anything higher-level we need that can be easily built in terms of eval doesn't have to be built into the protocol.

(We could arguably get rid of the distinction between (2) and (3), and basically roll all the Brackets functionality into the "protocol" layer by simply merging the RemoteFunctions script into the protocol remote script. The only reason to keep the protocol layer separate, IMO, is if we want to keep it compatible with CDT, a la RemoteDebug - so it only provides the functionality that CDT does.)

The transport layer currently implemented uses a WebSocket server in Node, coupled with an injected script in the browser that connects back to that server. However, this could easily be swapped out for a different transport layer that supports a preview iframe directly inside Brackets, where the communication is via `postMessage()`.

The protocol layer currently exposes a very simple API that currently just contains specific protocol functions: 
 * "evaluate"  which evals in the browser
 * "reload" which reload the page in the browser
 * "navigate" which allow the browser navigate to a given URL
 
The over-the-wire protocol is a JSON message that more or less looks like the CDT wire protocol, although it's not an exact match right now - again, we could decide to make it exactly mimic CDT if we wanted.

If we want to eventually reintroduce a CDT connection (or hook up to RemoteDebug), we have two choices: we could either just implement it as a separate transport, or we could implement it as a separate protocol impl entirely. Implementing it as a transport would be easier, and would be fine for talking to our own injected script; but it would only make sense for talking to CDT-specific functionality if we were very good about our wire protocol looking like the CDT wire protocol in general. Otherwise, we would probably want to consider swapping out the protocol entirely.

### Explanation of the flow

I've created a [really crappy block diagram](https://raw.githubusercontent.com/wiki/njx/brackets-livedev2/livedev2-block-diagram.png) of how the various bits talk to each other.

Here's a short summary of what happens when the user clicks on the Live Preview button on an HTML page.

1. LiveDevelopment creates a LiveHTMLDocument for the page, passing it the protocol handler (LiveDevProtocol). LiveHTMLDocument manages communication between the editor and the browser for HTML pages.
2. LiveDevelopment tells StaticServer that this path has a live document. StaticServer is in charge of actually serving the page and associated assets to the browser. (Note: eventually I think we should get rid of this step - StaticServer shouldn't know anything about live documents directly; it should just have a way of request instrumented text for HTML URLs.)
3. LiveDevelopment tells the launcher to open the page via the StaticServer URL, then it opens the page in the default browser.
4. The browser requests the page from StaticServer. StaticServer notes that there is a live document for this page, and requests an instrumented version of the page from LiveHTMLDocument. (The current "requestFilterPaths" mechanism for this could be simplified, I think.)
5. LiveHTMLDocument instruments the page for live editing using the existing HTMLInstrumentation mechanism, and additionally includes remote scripts provided by the protocol (LiveDevProtocolRemote), transport (LivePreviewTransportRemote), remote functions (RemoteFunctions) and document observation (DocumentObserver) which tracks related documents. The transport script includes the URL for the WebSocket server created when the transport is started after being set to the protocol layer.
6. The instrumented page is sent back to StaticServer, which responds to the browser with the instrumented version. Other files requested by the browser are simply returned directly by StaticServer.
7. As the browser loads the page, it encounters the injected scripts. The transport script connects back to the Service WorkerTransport's WebSocket server created in step 3 and sends it a "connect" message to tell it what URL has been loaded in the browser. The LivePreviewTransport assigns the socket a client id so it can keep track of which socket is associated with which page instance, then raises a "connect" event.
8. LivDevProtocol receives the "connect" event and makes a note of the associated client ID updating the current active connections. LiveDevelopment also receives the "connect" event, it checks that the URL that comes on the message matches the current live doc instance and updates session status to STATUS_ACTIVE in case this is the first client connected.
After the live document is loaded in the browser, DocumentObserver scan related stylesheets and Javascript files and send a message 'DocumentRelated' which contains their URLs. LiveDevelopment receives the message, extract stylesshes and creates a LiveCSSDocument instance per each of them. LiveHTMLDocument listen to the same message and further notifications (added/removed) to keep info about its related docs which is needed by LiveDocument (eg. when saving changes of a JS file).
9. As the user makes live edits or changes selection, LiveHTMLDocument calls the protocol handler's "evaluate" function to call functions from the injected RemoteFunctions.
10. The protocol's "evaluate" method packages up the request as a JSON message and sends it via the transport.
11. The remote transport handler unpacks the message and passes it to the remote protocol handler, which finally interprets it and evals its content.
12. If another browser loads the same page (from the StaticServer URL), steps 4-8 repeat, with LiveHTMLDocument just adding the new connection's client ID to its list. Future evals are then sent to all the associated client IDs for the page.


### Changes from existing LiveDevelopment code

* the existing code for talking to Chrome Developer Tools via the remote debugging interface is gone for now
* CSSDocument and HTMLDocument were renamed to LiveCSSDocument and LiveHTMLDocument, with a new LiveDocument base class
* the "agents" are all gone - a lot of them were dead code anyway; other functionality was rolled into LiveDocument
* communication is factored into transport and protocol layers (see above)
* HTMLInstrumentation and HTMLSimpleDOM were modified slightly (which is why they're copied into the extension), to make it possible to inject the remote scripts and to fix an issue with re-instrumenting the HTML when a second browser connects to Live Development. The former change is harmless; the latter change would need some review or possibly more work in order to merge into master. 


#### Unit tests

We would definitely need a good suite of unit tests for the new functionality. There are just a few integration tests that were 'migrated' from the current suite of LiveDevelopment but, it would need more granular unit tests and be able to run them on multiple browsers.