Add Steel as an optional plugin system

[mattwparas]

Notes:

• I still need to rebase up with the latest master changes, however doing so causes some headache with the lock file, so I'll do it after some initial feedback. Also, this depends on the event system in Add an event system #8021.
• The large diff size is a combination of lock file changes + the dependency on the event system PR. The diff has ended up quite large with all of the other stuff
• I'm currently pointing to the master branch of steel as a dependency. This will point to a stable release on crates once I cut a release.

---

Opening this just to track progress on the effort and gather some feedback. There is still work to be done but I would like to gather some opinions on the direction before I continue more.

You can see my currently functioning helix config here and there are instructions listed in the STEEL.md file. The main repo for steel lives here, however much documentation is in works and will be added soon.

The bulk of the implementation lies in the engine.rs and scheme.rs files.

Design

Given prior conversation about developing a custom language implementation, I attempted to make the integration with Steel as agnostic of the engine as possible to keep that door open.

The interface I ended up with (which is subject to change and would love feedback on) is the following:

pub trait PluginSystem {
    /// If any initialization needs to happen prior to the initialization script being run,
    /// this is done here. This is run before the context is available.
    fn initialize(&self) {}

    fn engine_name(&self) -> PluginSystemKind;

    /// Post initialization, once the context is available. This means you should be able to
    /// run anything here that could modify the context before the main editor is available.
    fn run_initialization_script(&self, _cx: &mut Context) {}

    /// Allow the engine to directly handle a keymap event. This is some of the tightest integration
    /// with the engine, directly intercepting any keymap events. By default, this just delegates to the
    /// editors default keybindings.
    #[inline(always)]
    fn handle_keymap_event(
        &self,
        _editor: &mut ui::EditorView,
        _mode: Mode,
        _cxt: &mut Context,
        _event: KeyEvent,
    ) -> Option<KeymapResult> {
        None
    }

    /// This attempts to call a function in the engine with the name `name` using the args `args`. The context
    /// is available here. Returns a bool indicating whether the function exists or not.
    #[inline(always)]
    fn call_function_if_global_exists(
        &self,
        _cx: &mut Context,
        _name: &str,
        _args: &[Cow<str>],
    ) -> bool {
        false
    }

    /// This is explicitly for calling a function via the typed command interface, e.g. `:vsplit`. The context here
    /// that is available is more limited than the context available in `call_function_if_global_exists`. This also
    /// gives the ability to handle in progress commands with `PromptEvent`.
    #[inline(always)]
    fn call_typed_command_if_global_exists<'a>(
        &self,
        _cx: &mut compositor::Context,
        _input: &'a str,
        _parts: &'a [&'a str],
        _event: PromptEvent,
    ) -> bool {
        false
    }

    /// Given an identifier, extract the documentation from the engine.
    #[inline(always)]
    fn get_doc_for_identifier(&self, _ident: &str) -> Option<String> {
        None
    }

    /// Fuzzy match the input against the fuzzy matcher, used for handling completions on typed commands
    #[inline(always)]
    fn available_commands<'a>(&self) -> Vec<Cow<'a, str>> {
        Vec::new()
    }

    /// Retrieve a theme for a given name
    #[inline(always)]
    fn load_theme(&self, _name: &str) -> Option<Theme> {
        None
    }

    /// Retrieve the list of themes that exist within the runtime
    #[inline(always)]
    fn themes(&self) -> Option<Vec<String>> {
        None
    }

    /// Fetch the language configuration as monitored by the plugin system.
    ///
    /// For now - this maintains backwards compatibility with the existing toml configuration,
    /// and as such the toml error is exposed here.
    #[inline(always)]
    fn load_language_configuration(&self) -> Option<Result<Configuration, toml::de::Error>> {
        None
    }
}

If you can implement this, the engine should be able to be embedded within Helix. On top of that, I believe what I have allows the coexistence of multiple scripting engines, with a built in priority for resolving commands / configurations / etc.

As a result, Steel here is entirely optional and also remains completely backwards compatible with the existing toml configuration. Steel is just another layer on the existing configuration chain, and as such will be applied last. This applies to both the config.toml and the languages.toml. Keybindings can be defined via Steel as well, and these can be buffer specific, language specific, or global. Themes can also be defined from Steel code and enabled, although this is not as rigorously tested and is a relatively recent addition. Otherwise, I have been using this as my daily driver to develop for the last few months.

I opted for a two tiered approach, centered around a handful of design ideas that I'd like feedback on:

The first, there is a init.scm and a helix.scm file - the helix.scm module is where you define any commands that you would like to use at all. Any function exposed via that module is eligible to be used as a typed command or via a keybinding. For example:

;; helix.scm

(provide shell)

;;@doc
;; Specialized shell - also be able to override the existing definition, if possible.
(define (shell cx . args)
  ;; Replace the % with the current file
  (define expanded (map (lambda (x) (if (equal? x "%") (current-path cx) x)) args))
  (helix.run-shell-command cx expanded helix.PromptEvent::Validate))

This would then make the command :shell available, and it will just replace the % with the current file. The documentation listed in the @doc doc comment will also pop up explaining what the command does:

Once the helix.scm module is require'd - then the init.scm file is run. One thing to note is that the helix.scm module does not have direct access to a running helix context. It must act entirely stateless of anything related to the helix context object. Running init.scm gives access to a helix object, currently defined as *helix.cx*. This is something I'm not sure I particularly love, as it makes async function calls a bit odd - I think it might make more sense to make the helix context just a global inside of a module. This would also save the hassle that every function exposed has to accept a cx parameter - this ends up with a great deal of boilerplate that I don't love. Consider the following:

;;@doc
;; Create a file under wherever we are
(define (create-file cx)
  (when (currently-in-labelled-buffer? cx FILE-TREE)
    (define currently-selected (list-ref *file-tree* (helix.static.get-current-line-number cx)))
    (define prompt
      (if (is-dir? currently-selected)
          (string-append "New file: " currently-selected "/")
          (string-append "New file: "
                         (trim-end-matches currently-selected (file-name currently-selected)))))

    (helix-prompt!
     cx
     prompt
     (lambda (cx result)
       (define file-name (string-append (trim-start-matches prompt "New file: ") result))
       (temporarily-switch-focus cx
                                 (lambda (cx)
                                   (helix.vsplit-new cx '() helix.PromptEvent::Validate)
                                   (helix.open cx (list file-name) helix.PromptEvent::Validate)
                                   (helix.write cx (list file-name) helix.PromptEvent::Validate)
                                   (helix.quit cx '() helix.PromptEvent::Validate)))

       (enqueue-thread-local-callback cx refresh-file-tree)))))

Every function call to helix built ins requires passing in the cx object - I think just having them be able to reference the global behind the scenes would make this a bit ergonomic. The integration with the helix runtime would make sure whether that variable actually points to a legal context, since we pass this in via reference, so it is only alive for the duration of the call to the engine.

Async functions

Steel has support for async functions, and has successfully been integrated with the tokio runtime used within helix, however it requires constructing manually the callback function yourself, rather than elegantly being able to use something like await. More to come on this, since the eventual design will depend on the decision to use a local context variable vs a global one.

Built in functions

The basic built in functions are first all of the function that are typed and static - i.e. everything here:

• Typed
• Static

However, these functions don't return values so aren't particularly useful for anything but their side effects to the editor state. As a result, I've taken the liberty of defining functions as I've needed/wanted them. Some care will need to be decided what those functions actually exposed are.

Examples

Here are some examples of plugins that I have developed using Steel:

File tree

Source can be found here

filetree.webm

Recent file picker

Source can be found here

recent-files.webm

This persists your recent files between sessions.

Scheme indent

Since steel is a scheme, there is a relatively okay scheme indent mode that only applied on .scm files, which can be found here. The implementation requires a little love, but worked enough for me to use helix to write scheme code 😄

Terminal emulator

I did manage to whip up a terminal emulator, however paused the development of it while focusing on other things. When I get it back into working shape, I will post a video of it here. I am not sure what the status is with respect to a built in terminal emulator, but the one I got working did not attempt to do complete emulation, but rather just maintained a shell to interact with non-interactively (e.g. don't try to launch helix in it, you'll have a bad time 😄 )

Steel as a choice for a language

I understand that there is skepticism around something like Steel, however I have been working diligently on improving it. My current projects include shoring up the documentation, and working on an LSP for it to make development easier - but I will do that in parallel with maintaining this PR. If Steel is not chosen and a different language is picked, in theory the API I've exposed should do the trick at least with matching the implementation behavior that I've outlined here.

Pure rust plugins

As part of this, I spent some time trying to expose a C ABI from helix to do rust to rust plugins directly in helix without a scripting engine, with little success. Steel supports loading dylibs over a stable abi (will link to documentation once I've written it). I used this to develop the proof of concept terminal emulator. So, you might not be a huge fan of scheme code, but in theory you can write mostly Rust and use Steel as glue if you'd like - you would just be limited to the abi compatible types.

System compatibility

I develop off of Linux and Mac - but have not tested on windows. I have access to a windows system, and will get around to testing on that when the time comes.

[zetashift]

I've been using this Steel + Helix integration for a bit. My config is here: https://github.com/zetashift/helix-config

The main downside here imho is that Helix isn't as great for Lisps(or Scheme in this case?) as Emacs for example.

But besides that this worked nicely. I have no real complaints and I actually did not need to script so many things as with my VSCode or NeoVim configs because a lot just works ™️ .

[pascalkuthe]

The main downside here imho is that Helix isn't as great for Lisps(or Scheme in this case?) as Emacs for example.

I think we can definitely improve, this is just the first prototype after all. You can't do everything at once. For example, I want to fully revamp the config system in the future (and remove the toml based config) so that is works even better with scheme

[zetashift]

I think we can definitely improve, this is just the first prototype after all. You can't do everything at once. For example, I want to fully revamp the config system in the future (and remove the toml based config) so that is works even better with scheme

Oh! I'm very sorry if I came over as demanding. I meant more like, in my experience this was what I stumbled upon, and it's not even a permanent thing, because a LSP for Steel might be in the works as well.

I completely agree that you can't do everything at once, nor should somebody be obliged to do it.

And maybe some lower hanging fruit can be picked up by people like me. For example if the tree-sitter grammar for scheme can improve things for Steel than that's another avenue.

[pascalkuthe]

No worries I didn't read it as negative. I ws more trying to highlight that this is just an initial prototype that can & will Improve (although Mathew already did a great job!)

[mattwparas]

And maybe some lower hanging fruit can be picked up by people like me. For example if the tree-sitter grammar for scheme can improve things for Steel than that's another avenue.

This is something that people with more familiarity with tree sitter and indents could be able to answer - are indent queries capable of matching a proper lisp indent mode? I did spend some time trying to mimic it but without much success

[Omnikar]

Github says the ending newline is missing in this file.

[Triton171]

This is something that people with more familiarity with tree sitter and indents could be able to answer - are indent queries capable of matching a proper lisp indent mode? I did spend some time trying to mimic it but without much success

I don't know Scheme / Lisp but it looks like the indentation should not be too difficult to model with tree-sitter indent queries (just using @indent and @align). I wrote a very simple query that correctly predicts the indentation of the first 100 lines of your scheme indent code:

(list) @indent
; Align lists to the second element (except if the list starts with `define`)
(list . (symbol) @first . (_) @anchor
  (#not-eq? @first "define")
  (#set! "scope" "tail")) @align

In order to improve this, I'd need to understand in which cases lists are aligned to the first, in which to the second element and when they are not aligned at all. Is this documented somewhere? If I find the time, I'll also try to understand the scheme indent implementation you wrote. Now that the plugin system is making progress, it might finally be time for me to learn the language 😄

[mattwparas]

I don't know Scheme / Lisp but it looks like the indentation should not be too difficult to model with tree-sitter indent queries (just using @indent and @align). I wrote a very simple query that correctly predicts the indentation of the first 100 lines of your scheme indent code:

(list) @indent
; Align lists to the second element (except if the list starts with `define`)
(list . (symbol) @first . (_) @anchor
  (#not-eq? @first "define")
  (#set! "scope" "tail")) @align

In order to improve this, I'd need to understand in which cases lists are aligned to the first, in which to the second element and when they are not aligned at all. Is this documented somewhere? If I find the time, I'll also try to understand the scheme indent implementation you wrote. Now that the plugin system is making progress, it might finally be time for me to learn the language 😄

This is great, I really need to learn more about tree sitter idents! I used this as a reference to get started: https://github.com/ds26gte/scmindent#how-subforms-are-indented

Note, my implementation is not complete and was mostly an exercise in 1. seeing if it could be done and 2. getting something far enough to make editing scheme code more pleasant, I would not verbatim use it as a canonical reference.

[Triton171]

Thanks a lot for the reference @mattwparas. I created a PR (#8720) with indent queries that should cover everything from it except for some simplifications regarding keywords. Feel free to try it out and report any issues it has (you can just copy the indents.scm file into the corresponding folder in your runtime directory).

[pascalkuthe]

already left some notes but haven't got trough everything.

I think we should be mindful of not exposing too many implementation details and not tying the API too closely to how helix is currently implemented (as that would make evolution harder and may feel clunky).

In the longterm I would like a lot more focus on configuration but to get a good steel API for that we probably need to revamp our config system internally first so that its not tied to closely to toml/serde (And that there is a separate config of a mutable buffer config) so it may be better to do that in future PRs.

Ideally I would like something like the following to work eventually:

(config
  (theme "everforest")
  (soft-wrap
    enable #true
    wrap-at-text-width #true)
  (text-width 85)

  (keymap "normal"
    ("=" write)
    ("j" my_custom_function)))

[pascalkuthe]

nit: I think the name engine is a bit ambiguous. Could we call it something like script_engine to make it clearer what the file actually does

[pascalkuthe]

this is too specical cased IMO.

While nice for a proof of concept the editor should have a generic concept of per buffer keymaps and then plugins should just interface with that

[mattwparas]

I would agree, it requires a little hacking from the steel side to make it work as you can tell. Do you think it is worth investing in that machinery to expose that from the editor side now?

[pascalkuthe]

I am not sure how suitable threadlocals really are.

I think tokio can and does move our main loop to different threads (and I also don't want to constrain interaction with plugins to a single thread).

In the worst case these could be globals but for almost all of these it feels like its really adding more extensability to the editor here instead of in the editor itself.

I think we should instead take the approach of making the core editor more extensible and then exposing those APIs to plugins.

These APIs would ofcourse require access to somekind of context. This would be similar to the context problem you mentinoed in your description. I think the best way to handle that one would be if we stored the context as a global inside the engine (using run_with_reference).

In any callbacks that want access to the context we check for that global (even user code could do it themselves right?) and throw an error if a function that needs a context is called without context access.

[mattwparas]

I'm currently refactoring the implementation to instead just from the steel side of things, have one variable that refers to the context. Whenever we call steel code, we'll do like you said - just refresh the context variable with the current context for the duration of the steel code scope, otherwise it'll be some poisoned value that will error.

With respect to thread locals - I'm really not sure how else to implement the interaction with the steel engine if not through a thread local, and any callbacks that I set up are done on a thread local queue that tokio is processing.

Can you elaborate on:

I also don't want to constrain interaction with plugins to a single thread

Does this mean you want the engine to be utilizing multiple threads? Or you want the main loop to be able to be moved to different threads (or both)?

As it stands, the engine implementation is not thread safe, which would make this rather tricky to implement. I don't know if tokio is actively moving the main loop to another thread or not, as I haven't run into any strange behavior with the interaction between the editor and the engine up to this point.

[mattwparas]

To put it another way - if we didn't want to have the main thread interact with the scripting engine, then it would have to take the context by value for this to work (at least, that is what I'm currently thinking)

[pascalkuthe]

If tokio lets you hold the engine struct across await points, I think the main loop only ever runs on a single thread. I would like to be able to eventually make Engine Send so we can send it across thread. I could definitely imagine that we would like to run some plugin code in the background (or in response to some async event).

Requiring engine to be Sync won't work since you would need to pub mutexes everywhere which would be excessively slow. But Send could be reasonable so we can have Mutex<Engine> and share that across threads. That is usually a reasonably easy requirement. The main thing you can't do is expose something like Rc to the user. I know you use refcounting internally (so Send is currently not implemented) but it could be safe to unsafe impl Send for Engine assuming that you never hand any refcounted types out to rust (or if you do make sure to use Arc or a non-clonable reference).

I think thread_locals are quite alright for the Context since the context would be explicitly provided everytime we call into the Engine.

[mattwparas]

So that the current moment, I've added some code to have the whole editor panic if the engine gets used on another thread to see if it is happening. I have been using this as a daily driver for a few weeks like this and haven't yet run into it. So while I agree thread local isn't the best, it also seems to be acting as intended, which I think is just by coincidence

[pascalkuthe]

this is an instance of exposing too many implementation details.

As I said in the other comment, context should be somekind of hidden global (which can have different capabilities depending on callsite).

The user should only know that he can do X/Y at a certain callsite. the existance of context and editor struct is an implementation detail that plugins should not be aware of

[mattwparas]

So just to clarify - don't expose the editor struct at all, any details go directly through the context?

I'm fine with that, just confirming the intent here.

[pascalkuthe]

yeah don't expose theses structs at all. These should not be editor-cursor and cx->cursor but instead just active-cursor and the internal structs are handled in the background.

[pascalkuthe]

what are the naming conventions in steel? These look like rust paths to me

[mattwparas]

There aren't any explicit naming conventions, although in general I prefer kebab case. At some points I tend to adopt Rust naming conventions in order to make interaction with the system more obvious. This is certainly open for discussion and I don't have any strong opinions.

[pascalkuthe]

hmm I am not sure how I fell about using rust naming conventions. I think syntactically it feels nice to stay a bit with conventions typically used in scheme/lisp. For example racket conventions could work well: https://docs.racket-lang.org/style/Textual_Matters.html

I think in this case the name new is not needed at all and we could just call it (position row col) following the naming conventions for constructors in racket (https://docs.racket-lang.org/reference/define-struct.html)

[mattwparas]

yeah that works for me!

[pascalkuthe]

this PR is not intended to discuss the choice of plugin language. We already made our choice after lengthy discussion. This PR is only intended for reviewing and discussing this particular implementation. I will mark such comments offtopic

[kirawi]

As an aside question, I recall that we were talking about sandboxing on Matrix a while back. Is this still planned/implemented here?

[pascalkuthe]

As an aside question, I recall that we were talking about sandboxing on Matrix a while back. Is this still planned/implemented here?

Not fully, right now there are no restrictions what a VM van do but I talked about this a while ago with Mathew and it should be possible to add capability based security to plygins in the future.

This would entail a one VM per plugin approach (and how to handle interop in that case) but those should be solvable problems (the same problems would have occurred with any sandboxing including wasm)

[i1Fury]

Very excited, great stuff Paras ❤️

[Ciflire]

Hi, just wanted to share a small plugin that I was personnaly missing helix-discord-rpc which just allows you to show that you are Playing Helix in discord. I had a lot of help with setting things up on the discord but finally it works, the fork clearly provides solid ground for building plugins and hope it gets merged soon !

Thank you mattwparas for your work and support

[gerblesh]

if we're adding to the plugin list, I made modeline.hx to support basic emacs and vim modelines. Currently requires a custom branch that allows regex matching on ropes but presumably it could be retrofitted to use a dylib for regex instead. Thanks again mattwparas for your hard work! Super exciting to see this PR coming to a close

[quantonganh]

@mattwparas

I'm trying to make "Goto definition" decompile a java class (like IntelliJ or nvim-jdtls does):

• Add support for jdtls' class-file-contents feature #11559
• Go to Definition' (gd) isn't working for Java's Standard Library using jdtls #14142

It took me a few hours to find out that the client needs to set classFileContentsSupport = true during initialization:

[language-server.jdtls.config]
extendedClientCapabilities.classFileContentsSupport = true

This allows the server to respond with something like:

{
  "jsonrpc": "2.0",
  "id": 2,
  "result": [
    {
      "uri": "jdt://contents/java.base/java.util/ArrayList.class?=karls-languages/%5C/opt%5C/homebrew%5C/Cellar%5C/openjdk%5C@21%5C/21.0.7%5C/libexec%5C/openjdk.jdk%5C/Contents%5C/Home%5C/lib%5C/jrt-fs.jar%60java.base=/javadoc_location=/https:%5C/%5C/docs.oracle.com%5C/en%5C/java%5C/javase%5C/21%5C/docs%5C/api%5C/=/%3Cjava.util(ArrayList.class",
      "range": {
        "start": {
          "line": 167,
          "character": 11
        },
        "end": {
          "line": 167,
          "character": 20
        }
      }
    }
  ]
}

As you can see, the URL scheme is jdt://, which Helix currently does not support:

2025-08-08T14:52:01.392 helix_term::commands::lsp [WARN] discarding invalid or unsupported URI: unsupported scheme 'jdt' in URL jdt://contents/java.base/java.util/ArrayList.class...

So my question is: is it possible to intercept the server's response using a plugin? If so, we could parse the jdt:// scheme, send additional LSP commands, and navigate to the decompiled file.

Thanks,

[jdrst]

To add to the plugin list, i started working on a c# plugin. This is still very barebones though.

[piotrkwarcinski]

Sharing hx-tmux-navigator might be helpful for anyone using tmux+helix. It allows you to use the same keybindings to navigate within and between helix splits and tmux panes.

[jerabaul29]

Curious: where should we discussed "wishlist" etc for plugins? For example, I would love to have a plugin that lets me integrate ClaudeCode and helix (like what exists between eg VSC and helix, i.e. make CC aware of which file I am in, where I am in the file, what my selection is etc). Should there be an issue / discussion for this, or should we just discuss it here (but this may clutter?)? :)

[noor-tg]

Curious: where should we discussed "wishlist" etc for plugins? For example, I would love to have a plugin that lets me integrate ClaudeCode and helix (like what exists between eg VSC and helix, i.e. make CC aware of which file I am in, where I am in the file, what my selection is etc). Should there be an issue / discussion for this, or should we just discuss it here (but this may clutter?)? :)

I think there is steel discord somewhere 🤔.
But it will be better if you are willing to develop it yourself.

[gyreas]

So my question is: is it possible to intercept the server's response using a plugin? If so, we could parse the jdt:// scheme, send additional LSP commands, and navigate to the decompiled file.

you can also write an LSP server thats a wrapper around jdtls that forwards messages between it and the client, except for GotoDefinition which it intercepts and updates before forwarding a response

[godalming123]

I thought that I would share this after working on it a bit more, but I haven't found any time to work on it for a while, so I'll share it now. It's a plugin called select-project.hx, and it allows you to set helix's working directory. @mattwparas, I was wondering how you would close the current picker and open the file picker when the user selects a project.