Skip to content

Latest commit

 

History

History
210 lines (174 loc) · 15.5 KB

README.md

File metadata and controls

210 lines (174 loc) · 15.5 KB
Raw

browser-extension-framework

🛠 BrowserExtensionFramework is a zero-dependency RPC-centric framework for browser extensions (sometimes called web extensions).

BrowserExtensionFramework

NOTE: This project was developed on macOS. It is for my own personal use.

BrowserExtensionFramework is an RPC-centric web extension framework. I originally developed it while working on another project of mine: https://github.com/dgroomes/stackoverflow-look-back.

Here are some key points:

  • It supports Manifest V2 APIs only (Manifest V3 APIs are not supported)
  • It is useful for injecting JavaScript files into the web page
  • It is useful for two-way communication between components. E.g. web-page-to-background, popup-to-background, etc.
  • It depends on the RPC framework
  • If you do not need to inject JavaScript code into the web page, then you probably don't need this framework.

The API is complicated only because the architecture of a web extension can be complicated. Some extensions will use all JavaScript execution environments: background scripts, popup scripts, content scripts and web page scripts. It's challenging conceptually to even think about all these environments because we are used to programming in just one environment like the web page, or maybe a NodeJS app. Plus, writing a program for this environment requires a lot of message passing code, Promises code and logging (for debugging) code. That's where BrowserExtensionFramework comes in.

However, the framework cannot abstract away the JavaScript execution environments. The user of the API still needs to know how web extensions work and about each of the JavaScript execution environments. In that sense, this API does not offer a strong abstraction but rather a leaky abstraction. To make up for this, the framework offers block-level API documentation, design notes and inline code comments. The framework code is meant to be read. Please study it before using it!

The API is best introduced by way of example. See the README in the example/ directory which contains an example web extension.

RPC Framework

A significant portion of a non-trivial browser extension is often dedicated to Message Passing between the four components of an extension: (1) a background script (2) a popup script (3) a content script (4) the web page. Message passing is a fundamental and useful programming feature, but unfortunately in a browser extension environment the complexity of the code for message passing is exacerbated by the number of components (the aforementioned four) and the sometimes stark differences in APIs between browsers (Chromium vs Firefox). It's desirable to encapsulate the complexity of message passing behind an easy-to-use API that takes a message, does all of the behind the scenes work, and then returns a response. This description looks like a Remote Procedure Call system.

In this codebase, I've implemented a general-purpose Remote Procedure Call (RPC) API for web extensions.

It could be extracted into it's own project. And honestly, it's not a great implementation, but I came to it out of necessity.

The source code is laid out in a file structure that groups code by the execution context that the code runs in:

  • rpc/rpc.ts
    • The code in this file is foundational common code for the RPC framework. It is used in all contexts of a web extension: background scripts, popup scripts, content scripts, and the web page.
  • rpc/rpc-web-page.ts
    • The code in this file runs on the web page.
  • rpc/rpc-backend.ts/
    • The code in this file runs in the extension backend contexts: background workers, popups, and content scripts.
  • rpc/content-script-proxy.ts
    • The code in this file runs in a content script.

One thing I'm omitting with the RPC implementation is an "absolute unique identifier" to associate with each message. Without this uniqueness, it's potentially possible to "cross beams" and, for example, have an RPC client process a message that was not intended for it. I think this is virtually impossible though because we are in a browser environment where we exercise almost complete control of the environment. By contrast, an RPC system in a distributed system spanning different networks would need to handle these cases.

RPC Framework Usage Instructions

Browser extensions that use the RPC Framework must follow these steps to depend on and initialize the framework in the extension and web page contexts:

  1. Manifest changes
    • Unless you are bundling the RPC code directly into a final bundle.js-like file, then you must make these files accessible. The manifest.json file must allow access to the RPC JavaScript source code files as needed. Specifically, rpc.js, and rpc-backend.js must be added to the background scripts and rpc.js and rpc-web-page.js must be added to the web page.
  2. Initialize the content script proxy
    • The initializer function initializeProxy in content-script-middleware.ts must be executed.
  3. Initialize objects in the web page
    • The web page must initialize the RPC objects on the web page by calling initRpcWebPage(...)

Design

This project is made up of multiple npm workspaces:

  • framework/
    • This workspace represents the BrowserExtensionFramework. It pulls in the code from other workspaces and glues it together with the main framework code.
  • browser-types/chromium-types
    • TypeScript type declaration files for Chromium JavaScript APIs.
  • browser-types/firefox-types
    • TypeScript type declaration files for Firefox JavaScript APIs.
  • examples/detect-code-libraries
    • 'Detect Code Libraries' is an example browser extension that is built with BrowserExtensionFramework.

Wish List

General clean ups, TODOs and things I wish to implement for this project:

  • Consider publishing to npm. Publishing the compiled JavaScript and the TypeScript declaration files. A useful step to do before this would be to publish the distribution locally and consume it from the 'Detect Code Libraries' example project.
  • Runtime check the "externally_accessible" configuration list (I assume that's possible but I wouldn't be surprised if it wasn't) and warn if the current web page is not in the list. (This was the problem I had when I developed the example extension and I was confused).
  • Fix the double loading problem. This is all over the place. It double loads the source code when you click the extension browser action. It makes for an unusable experience except in the very narrow happy path.
  • The "init" functions initRpcBackground and initRpcWebPage should be encapsulated in the BackendWiring and PageWiring apis right?
  • Update the overall instructions. Especially since I started bundling the source code files into a "bundle" (because we're compiling TypeScript into 'entrypoint'-like files), the instructions have gotten stale. Also the whole 'RPC sub-framework' isn't really clear anymore.
  • Debugging in Firefox is broken for me on my mac. I can't get it to show logs or sources, even when I try an official extension example like 'beastify'. It does not work like the Extension workshop docs say it should. I need to try on my Windows computer.
  • Can I use interfaces for BackendWiring and PageWiring? It has been a journey to experiment with Deno, TypeScript, ts-loader and webpack and frequently face surprising restrictions in the way we can write TypeScript code. There are problems with something about final or non-extendable interfaces (seemed to be more of a Deno restriction) and then there are problems with cyclic dependencies (seemingly a ts-loader/webpack problem, I was not facing it with Deno). Anyway, BackendWiring is only usefully exposed to the user as an interface, not an abstract class. Can I make it an interface?
  • DONE I broke the project again. It turns our I still need the @types trick, or I need something else. The chrome global can't be found. I think I'm done with npm workspaces. What is more useful to me is TypeScript project references. Or frankly, just plain old "break up the .ts files in the same project".

Finished items

  • DONE Add an example web extension
  • DONE re-organize the directory layout. There's no need for a "browser-extension-framework/" directory. Too verbose.
  • DONE Consider abstracting away the required content script "thin bootstrap" files. For example, dcl-content-script.ts shouldn't have to exist. I thought it did earlier, but it's not needed. It can be replaced with a generic middleware content script.
  • DONE Support Firefox in the example. If the example supports both Chromium and Firefox, then I can build it, verify the behavior in both browsers, and have confidence that the framework still works.
  • DONE Are import maps going to save me from the awkward "installation-time" setter of the "browser descriptor? (e.g. 'chrome' of 'firefox')" and can it also be used for much more? Can I (should I) bundle only the Chromium-specific classes in the Chrome bundle? Also, this frees the architecture to not even use a class-oriented design. There's no need for dynamic dispatch if we can just get "vendor-specific dispatch" at build-time.
  • DONE Extract the api/ components into TypeScript interfaces so that the total amount of code presented to end-users is as small as feasible. By contrast, presenting the whole backend-wiring.ts file to end-users is too much. The public functions are BackendWiring.initialize and BackendWiring.injectInstrumentedPageScript. I don't want end-users to have to see the function bodies when browsing the API. Maybe move the implementation details into an impl/ or wiring/ directory (it doesn't really matter where).
  • DONE Stop using import maps for differentiating between Chromium/Firefox things. When it comes to publishing this library, I don't want to publish a Firefox artifact separately from a Chromium one. Node tooling is not equipped for consuming multi-flavor artifacts.
  • DONE Remove Deno for npm and Webpack. It was a rewarding experience and a quick start. But I need to understand a build a prototypical library and user/developer experience. There are so many quirks of browser-based JS modules that I can't afford to stray from mainstream.
  • DONE Remove Webpack and instead use Rollup and API Extractor. I figured out, through my work here that Rollup is a better fit for libraries. UPDATE: well this even drives the point home more strongly.. I found that Webpack continued to work well on the app side (DCL (detect-code-libraries)) because I couldn't get Rollup to work in DCL. Webpack is not dead! I think Rollup had trouble because of npm workspaces, but I couldn't prove it and didn't need to bother with it.
  • DONE Move Chrome and Firefox TypeScript type declaration files into a separate workspace. These type declarations are orthogonal to the implementation of BrowserExtensionFramework. They fit best in a separate workspace.
  • DONE Organize code into npm "workspaces". Workspaces were introduced with the release of npm 7 in 2020. I would consider workspaces a long-awaited feature because it's a tool that we have in most other programming languages and toolchains: Gradle (Java and JVM languages) has multi-project support, Rust has crates, Go has something (modules?) etc.
    • I used this command to init a workspace: npm init --workspace packages/framework and npm init --workspace packages/example-detect-code-libraries
  • DONE Use the index.ts (or .js, .mjs, whatever) convention instead of the api/ way. This is a common and well-understood convention.
  • DONE Resolve the warnings output from both Rollup and API Extractor. Resolving these errors will require making changes to the source code. I didn't want to make those changes when I introduced the new tools in order to minimize the diff.
  • DONE (UPDATE this did not work like I thought it did. I got false positives because node_modules/ had leftover goodies) (I got this resolved while doing the 'eject DCL' task. The trick was installing as 'devDependencies') TypeScript in combination with other tooling (npm, webpack) is so hard. I'm naming my Chromium and Firefox TypeScript type declaration packages with the name @types just to get things to work. The TypeScript compiler makes some assumptions about this package and things "just work" if the type declarations are in node_modules/@types. This is strange in my opinion because why should the owner of the @types package (DefinitelyTyped) have a monopoly on type declarations in the official TypeScript toolchain? People should be able to bring their own types. And you should be able to by using the typeRoots config in tsconfig.json, but I was having issues with it. (maybe it's an npm workspaces bug?). Below is the snippet I was trying. 
    "typeRoots": [
  "../node_modules/@types",
  "../node_modules/@dgroomes/chromium-types",
  "../node_modules/@dgroomes/firefox-types"
]
  • DONE Eject detect-code-libraries into its own standalone npm project. In other words, do not tread it is as workspace that's contained in the overall project. When detect-code-libraries is a standalone project, it acts closer to a real example project. One of the important tasks for a project that consumes BrowserExtensionFramework is to figure out how to import it.

Reference