Implementing shadow.remote API

Since version 0.8.0 of Chlorine, there’s a new way to evaluate ClojureScript code: that’s the Shadow-CLJS Remote API. It is basically a new REPL (not nREPL, no Socket REPL) over WebSockets to try to solve problems when translating other REPLs to ClojureScript. So, to understand why these problems exist, I’ll first introduce the difference between ClojureScript and Clojure.

On Clojure, you’re always inside a JVM. This means that compilation happens on the same JVM that your REPL, and your code is running. If you practice REPL-Driven Development, even your tests are running on the same JVM. In practical terms, it means that when you fire up your REPL, you already have everything ready to run code, compile code, and evaluate forms.

On ClojureScript, the compiler is written in Clojure – that means it’s running on the JVM. So, to produce Javascript code you don’t need a Javascript environment – and that’s when things become confusing, because when exactly will you run the REPL? Let’s try from another angle: if you start the REPL on compilation time, you can’t evaluate code (because there’s no Javascript generated, nor any Javascript engine running). If you start the REPL when you run the compiled code, this REPL can become unusable if you stop the Javascript environment, and also you have to coordinate lots of state and translations between formats.

Implementing a nREPL Client

Some people asked me on how did I implement the nREPL client on Chlorine. So, I’m writing this post!

nREPL is a simple protocol. It uses “sessions” that are used to isolate evaluation contexts and other things (for example, on Chlorine every connection connect to two REPLs: one “primary” and one “auxiliar” that is used to run commands like autocomplete / goto var definition, and so on). On nREPL, this isn’t necessary: you just connect to a single REPL and use two different sessions. Just for the record, because the way Chlorine works, I didn’t implement it like this (because I would have to rewrite lots of code – maybe in the future).

Now, to explain the protocol, let’s separate things in parts: the first thing I do (and I was already doing in the past) is to connect to a socket. Then, I looked at the documentation for nREPL to understand how to send and receive commands to the REPL. Now, there are some details on the way that every operation is implemented that I simply ignored because it was not necessary to understand then from the perspective of my application…

nREPL on… Chlorine?

When I started the Chlorine project, I just thought it would be great if I could target all Clojure-like REPLs that already exist but didn’t have tooling support. At the time, this would include Lumo and Plank, mostly. Also, Shadow-CLJS and Figwheel have some “clunky run-some-code-and-transform-in-cljs” way of working that simply didn’t click with me.

Now, almost a year later, Chlorine supports Clojure, ClojureScript (Shadow-CLJS, Lumo, Plank, or even over clj), ClojureCLR, Arcadia, Babashka, Clojerl (Clojure on Erlang) and Joker (Clojure on Go, also a linter). But the reality is that working with a pure Socket REPL is really hard – a socket REPL works exactly like a regular one, printing namespaces after each code, and so on. Also, there are some strange decisions on some REPLs, mostly likely ClojureScript (that is the second most used Clojure flavor), so things are not always easy. To put things in perspective, currently Chlorine uses 3 ways to evaluate code: It uses unrepl, that only works on Clojure, or uses internal APIs of shadow-cljs (that obviously only works for shadow-cljs), and for other implementations it uses a kind of a hack – it evaluates the code, inside a trycatch, and it returns a vector where the first element is a symbol in a specific format that Chlorine will understand and then link that with the response. This “hacky way” is currently being used for every other implementation except Clojure and Shadow-CLJS. Things work (autocomplete works too), but it is not pretty and sometimes have strange results.

As a matter of fact, I was already thinking about removing UNREPL (it’s really hard to implement new features on it, and some good ideas only work in theory – for example, the ability to evaluate long strings / collections and render only a part at a time aren’t that good with lots of edge-cases) and, to do it, I though about a better, non-hacky way to evaluate things on some Socket-REPLs (that, again, would only work on some REPLs – ClojureScript REPLs will probably never support “upgradable REPLs” because of the way they work) – the only thing that I had to understand is how to implement this “upgraded REPL”…

Then, recently, Babashka added an initial support for nREPL, with an insane low amount of lines. So, I’ve tried to implement a way to evaluate code over nREPL… and it was really simple to do it, using a npm library that already did it. But implementing like this meant that the user would need to know if the host/port to connect is a Socket REPL, or a nREPL (and the user does not know – lots of tools like lein and shadow-cljs show an nREPL port to be connected).