Code-GUI bidirectional editing via LSP | Hacker News Reader

Code-GUI bidirectional editing via LSP

jamesbvaughan.com

241 pointsjamesbvaughan68 comments

Hey nice - this stuff is so much fun to me. I've worked a number of experiments like this too, especially related to live coding. Love seeing it in the wild.

I built a small project where you can live-code Love2D. The running program updates in real time (no saving needed) and see all values update in real-time, via LSP.

https://github.com/jasonjmcghee/livelove

And also added the same kind of interactivity like a number slider and color picker that replace text inline, like yours (though via vs code extension: https://gist.github.com/jasonjmcghee/17a404bbf15918fda29cf69...)

Here's another experiment where I made it so you could "drag and drop" to choose a position for something by manipulating the editor / replacing a computed position with a static one, on keypress.

https://clj.social/@jason/113550406525463981

There's so much cool stuff you can do here.

I love your project. I am making something with LSP and your code was a great example of what is possible.

Glad it was useful! LSP is definitely under-utilized. It's great for the "malleable software" / small tools for yourself kind of stuff especially.

That's cool! I suspected that someone must have done something like this before.

Isn't autocad lisp powered?

I’ve been wanting to get into game programming and was looking at Love2D. I don’t know much Lua outside of what I’ve picked up with NeoVim configs.

This looks like it could be a great way to get my feet wet - I don’t do well with math and physics programming, but I used to make things in Flash back in the day that were similar to the particles demo and being able to quickly change things and see the updates makes it a lot easier for me to grok. Thanks for sharing!

Do it! It's a lot of fun.

Highly recommend adding code definitions https://github.com/LuaCATS/love2d

And getting the lua LSP.

It reminds me a lot of processing / p5js. So easy to get something fun up and running quickly.

This is close to the functionality that Borland Delphi had back in the 1990s. The pascal language and the design of their GUI toolkit were a really good impedance match, so you could freely switch between GUI design and editing the text version of it.

Doing so for languages like C++, was a sea of boilerplate that you couldn't touch, which is why I never moved away from Pascal. Similar fragility was evident in WxPython and it's builder.

I'm glad to see that LLMs can provide a match for less well suited Language/GUI pairs. We all deserve to get that kind of productivity.

Delphi's form files weren't pascal. The gui-sync was in my memory just editing the class definition when adding components / event handlers.

Borland C++ Builder did the same thing as Delphi, used the Delphi VCL class library, worked great.

For Slint [https://slint.dev], a (native) GUI Toolkit, I've also developed a LSP server that do live preview and editing. You can try it online at https://slintpad.com : if you click on the toolbar button to enable the right panel, you can edit the properties from the UI, and this is all done through the LSP and can be integrated in any editor that supports it.

cool! how does this work? e.g. how do you know which UI element matches which text element, do you track it while rendering? How do you propagate changes in the UI? do you update the text and then re-render the whole UI?

I use the the code lenses of code action feature of the LSP so that the user can start a preview. The LSP server will then open a native window on desktop.

Every time the users do some change in the code, the editor sends the document and we re-render the preview.

I use the textDocument/documentHighlight request to know when elements are being selected from the code so I can highlight them in the preview.

When selecting an element in the preview UI, my LSP server sends a window/showDocument to position the cursor at the right location. And if the user changes property or do change in the file, we do a workspace/applyEdit command with the changes.

Btw, the code is there: https://github.com/slint-ui/slint/tree/master/tools/lsp

This is a great idea: I'd never think of using LSP for this!

As a software developer, I always get frustrated when I am doing some graphical work and struggle to neatly parametrize whatever I am drawing (wooden cabinets and furniture, room layouts, installation plans...) and switch between coding where that makes most sense and GUI where it doesn't.

The best I've gotten was FreeCAD with Python bindings (I've got a couple of small libraries to build out components for me), but while you can use your own editor, the experience is not very seamless.

And then I start imagining tools like the one here, but obviously doing it just right for me (balancing the level of coding or GUI work).

I think we’re a lot of developers who are frustrated by constantly needing to choose between code or GUI… both have their use case, but I feel like there must be a system that combines both.

How about a visual programming language? Plenty of 3D and CAD software uses a VPL for procedural design, which helps a ton to bring out the benefits of both

Any examples?

Grasshopper for Rhino is a big one, or VisualCAD/CAM. Blender, Houdini, or Unreal Blueprints are others. It gives you the exploration benefits of a UI plus the procedural benefits of code. Inputs tend to be auto-bound which makes exploration much, much faster.

I'm working on one for web map data visualization: https://youtu.be/GJ0ftFB8r4I?t=871

Let's lobby for adding UI semantics to MCP. Generating ad-hoc GUIs seems like a major and common enough feature to warrant it.

It's concerning to me that the LSP idea is .. a thing. Casey Muratori observed years ago that it's just a way worse way of doing libraries. Like, you're introducing HTTP where there could just be a function call into a DLL/SO. What's the benefit there? Just make vim/emacs/$editor speak some native protocol and be done with it. Then you GUI is just welded directly into the running editor process.. right??

There's no security risk there that wasn't present before as far as I can tell because you were already planning on running the LSP on your local machine..

You're going to love MCP.

What is up with mcp... feels like CORBA

It's like if CORBA and COBOL had a baby :P

Seriously though, it does combine the "RPC IDL" aspect of the former, with the "Use English" aspect of the latter.

I’ve definitely wondered about this. Like why does the language-understanding standard have to be a server specifically? Just cause it’s a good architecture to build around?

I think I’ve heard that vscode has benefitted hugely from it starting out with a client-server architecture from the start, since it started as a browser based editor. Things like editing code directly on servers via ssh or in containers is easy for vscode cause its client-server all the way down.

Vscode and LSP are both Microsoft products, maybe Microsoft has been pushing the client server thing?

> I’ve definitely wondered about this. Like why does the language-understanding standard have to be a server specifically?

I don't think it does. I think it's a bad architectural decision that web bros thought sounded cute.

> Things like editing code directly on servers via ssh or in containers

I mean, vim and emacs have supported editing over ssh for like .. longer than I've been alive probably.

> I think I’ve heard that vscode has benefitted hugely from [clinet-server architecture]

IMO VSCode is a giant steaming pile; I'm not sure what the huge benefits could have been. It's intolerably slow, uses an insane amount of system resources, and the debugger barely works most of the time.

there are many practical benefits:

1. naturally async

2. each server and the editor itself can be written in its own language and runtime easily

3. servers can just crash or be killed because of oom errors and your editor won't be affected

4. in a lot of languages it is easier to write a server then to call/export c abi

5. the editor can run in a browser and connect to a remote server

6. you can have a remote central server

all of those things are done in practice

IMO, many of these are marginal or nonce gains for the huge cost of introducing a network call in the middle of your program.

1. There's nothing stopping you from shoving the library code onto a thread. For something like this request-response style usage pattern, that sounds extremely straight-forward, especially since the previous paradigm was probably an async server anyways. The calling code (from the editor) obviously already had to be support async calls, so no real change there.

2. If $LANGUAGE can be used to write a server, it should be able to build a DLL. I realize this is not practically true, but I also don't support the notion that we should be writing even light systems stuff like this in JS or python.

3. lol. You're telling me you're worried about a process eating 32GB of ram parsing text files ..? If some core part of my editing workflow is crashing or running out of memory, it's going to be a disruptive enough event that my editor might as well just crash. The program I'm working on uses a lot of memory, my editor better not.

4. I guess ..? Barely seems like that's worth mentioning because .. counterpoint, debugging networked, multi-process programs is massively harder than debugging a singular process.

5. Why would I want (other than extremely niche applications, shadertoy comes to mind) an editor to run in a browser? If I have a browser, I can run an editor that connects to a remote machine. Furthermore, the `library-over-http` approach of LSPs doesn't really buy you anything in this scenario that using a single process wouldn't.. you can just send all the symbol information to the browser.. it's just not that big.

6. Wut?

1. True, but in practice it's not always the case, for example older versions of resharper sometimes slowed down typing speed. It's much harder to fuck it up when there is a network request in the middle.

2. What about languages like Java and Go?

3. a. The experience of lsp server crashing is much better than an editor crashing, the editor usually automatically restart it. I had lsp servers crash without me noticing at all.

b. Memory problems in both lsp servers and traditional IDE analysis are extremely common in my experience. It seem to me that the problem is that there are a lot of pathological cases where the analysis enters a loop and keeps allocating memory.

4. When mixing runtimes I actually find it easier to have multiple processes because I can attach a specialized debugger to each process but this is definitely an important point.

5. good counter argument, I retract my point

6. what I meant is that for very large code bases it can be beneficial to run a central lsp server that many people can connect to because most of the index is shared between all of them and the parsing+indexing itself is very costly. I heard Google were doing something like that but I don't have more information.

1. Choosing a shitty architecture to absorb the downside of low quality tools seems like a bad decision, but I guess that's just where the industry's at right now.

2. AFAIK Go can both compile and load DLLs now. A language that doesn't have native compilation facilities (Java, JS, python, etc) would have to have extra tooling that loads the runtime and program from a DLL. A quick google search tells me Java has this tooling already. I'm sure other languages would too.

3 a,b. See point 1.

4. Yeah, that's fair, although I'd probably still rather have a single process.

6. Also a good point. Might be a useful paradigm for large codebases. Although at that point the usage pattern is so different that the calculus for choosing the architecture is different, and I stand by my choice. Google can afford to build fancy shit for themselves.

first of all I want to say that I have a lot fun talking to you! Honestly I would also prefer dll based approach, if you really want to see the limits of the architecture of Lsp I recommend you to read the discussion about adding syntax highlighting to the protocol, but I see why the people behind LSP did what they did.

One final point!

I think there is a difference between absorbing low quality tools and absorbing low quality code. I think that it makes a lot of sense for a plugin system to design for low quality plugins. I have two examples from my job, the first is a micro services based parsing infrastructure, where each team is responsible with parsing it's own formats, which are oftentimes layered apon formats parsed by different teams. I believe this takes about 10x-100x more resources than it should, but it has the benefit that a rouge parser can't takes everything down with it. The second example is internal scripting capability of a different system where a lot of work was done to make sure the script can't do anything stupid.

In both of those cases the system is designed for low quality code because it is more cost effective to make sure code can be low quality without affecting the overall system, than to make sure low quality code doesn't exists.

Isn’t the python based build123d the current best CAD in code solution? The problem with OpenSCAD is that it cannot export solid geometry, just a final mesh.

More broadly, I was genuinely shocked to realize, when I was playing with it, that there is no cross CAD file format that captures even simple design concepts like “this hole is aligned to the center of this plate” or even “this is a 2mm fillet”. STEP (the file format) mostly just captures final geometry.

I think CAD people just … redesign the part again if they need to move from say Fusion 360 to FreeCAD or whatever. How do they live like that?!

I don't think they change programs very often, and only very rarely in the middle of a project.

I think this maybe why!

Of course it’s great for vendor lock-in.

But given how many companies need to work with diverse suppliers, there must be a whole bunch of re-creating models happening. There is no chance that everybody is using the same CAD tool

The big example of this is Airbus on the A380 [1], design offices were using different versions of CATIA, we added some features to STEP to allow wiring harnesses to be modelled explicitly.

[1] https://en.wikipedia.org/wiki/Airbus_A380#Production_and_del...

STEP is capable of capturing what you describe, it is down to the "user group" of customers of the CAD vendors to ask for it to be implemented by each CAD system in terms of what they import and export.

We put in for some funding for the next edition of STEP AP242 for me to be able to work more closely with the user group to improve this area.

That's awesome! Yeah it seems strange to me that we don't have a standard way of tracking how a model is built up and parameterized (basically Fusion 360's history mode). I'd naively assumed that this was a solved problem in the CAD world - given the whole point of parametric CAD is to be able to easily tweak the distance between two points and have the whole model update.

It's really hard to explain if you don't know how CAD kernels produce final BREP shapes via their process trees, but, try give an example- something like "this is a 2mm fillet" requires a 'fillet solver' that is deeply ingrained in the kernel. It isn't something that would ever be portable, say, from Fusion360 to CATIA natively because those are completely different kernels with different ways of 'solving' the model (and not just fillets, but, everything).

That is why STEP containing the final BREP manifold solid is the standard interchange that it is - it is a final representation of the solved output that IS portable, and anything else is... difficult.

This is helpful but surely we can save a consistent description of the process step and the input parameters alongside the final geometry in the same file format?

Yes, we could, but there are very few open source fully parametric couplings of kernels and the file format out there. FreeCAD (using OpenCASCADE as a kernel) is one of the few. And that is nothing close to the vastness of modern CAD+PLM systems in level of functionality.

There is a reason that systems like Dassault's CATIA are both ubiquitous in aerospace and automotive and closed source- it is literally the culmination of probably 3000+ person-years of programming, stemming from the 1970s. The same for many others... making them interoperable would mean making the entire chain of systems open source and there isn't a reason for them to do so.

For small projects though, building DSLs and graphical tools on top of OpenCASCADE (which is GNU LGPL version 2.1 licensed) is about the closest you could get right now, particularly, in my opinion, building something like a pseudo-gui+textural tool with the Python package Build123D.

The interoperability problem stems from CAD kernels using proprietary B-rep representations and constraint solvers, with STEP's AP242 standard attempting to address this by including Product Manufacturing Information (PMI) and semantic annotations, though adoption remains fragmented.

I’d have assumed that somebody had a format for just the basic operations … but it looks like even that is just being considered

While it doesn’t use an LSP (which is a great idea!), this project of mine does bidirectional mapping between a GUI and a shell pipeline:

Screenshots and GIFs for the explanation!

https://github.com/williamcotton/guish

It is referenced down in the wonderful tree of knowledge, but I thought I would lift this info up to the top level. A system like this called a "bimodal editor" from the work of Ravi Chugh [1] see the project Sketch-n-Sketch [2] these kinds of systems got a big kick in the pants with Bret Victor's talk Inventing On Principle [3]. That in turn kicked off a wonderful dynamic execution Clojure IDE called lighttable [4]

There are tons of other great research projects and researchers in this fascinating space. Comment is not meant to exclude.

[1] https://people.cs.uchicago.edu/%7Erchugh/

[2] https://ravichugh.github.io/sketch-n-sketch/

[3] https://vimeo.com/906418692

[4] https://web.archive.org/web/20250616031034/http://lighttable... (demo video plays)

Brian Hempel https://people.cs.uchicago.edu/%7Ebrianhempel/ also from uchicago has some great bimodal projects.

See https://new.rt.ht for bidi i⇄o sync demos!

The PDF ones are especially fun!

The whole playground is built for bidi sync!

Wait, what is this?! It's really cool

See more at https://rtcode.io

I'm about to set off on this journey too with a code-first PCB editing suite for myself to work on. Seems like this is a good piece of prior art to reference as I need to build similar tooling. Thanks for posting!

The idea of a language server also being some other kind of server is almost fiendishly clever to me and I'm shocked I hadn't seen other people play with this before. Definitely can think of applications to game programming -- a sufficiently dedicated engine team could fork their language's LS and allow you to e.g. integrate the map editor's feedback into the source code. I know hot-reloading is considered super valuable in games, but this is almost the dual of that idea. So clever!

We do quite a lot of this using webviews in vscode over at Darwinium. Similar process that involves postMessages between the UI container and the host containing the LSP. Biggest challenge I had with this was figuring out how to map a diagnostic from the language server (which is inherently range-based) to a UI component that would display it - it involves a fairly intricate pubsub system and mapping UI components explicitly to parts of a tree. Would love to share it eventually.

Reminds me of the excitement I got on seeing https://www.youtube.com/watch?v=4tyTgyzUJqM. I still get excited about trying something like this again pretty much every time I see it. Have yet to really get any traction on anything. :(

A lot of folks had fun watching Minecraft built using a live code session, if I recall.

Hot reload in Java is still a crucial part of my workflow for modding Minecraft, but it doesn't have the other direction where an action in the game could alter the code

Ah, fair. Seems that is probably more applicable to image based development than it is typical batch processed source files? Doesn't seem too strange to have your code editing a data file, for example. No reason you couldn't treat your source as a data object, in some sense. Though, I'll note Common Lisp explicitly warns against editing literal objects. Such that I'm assuming this was considered and largely agreed to be a bad idea. :D

I've suggested this sort of interactivity for (Open)PythonSCAD:

https://lists.openscad.org/empathy/thread/GAX4QYYRUC3CEH572I...

The devil is in the details though, and I worry about the UI becoming cluttered and unmanageable.

I know it's not quite the same, and maybe it's just the demo video doing it, but this reminds me of the Light Table IDE http://lighttable.com/

Here's a similar example using lean widgets (in the live programming part of the talk): https://www.youtube.com/watch?v=c5LOYzZx-0c

I know basically nothing about CAD, but I know thet fornjot exists and wondered if it might be useful to your purposes, in case you don't know about that.

But well, the project is very cool and I love the idea of using LSP for something more!

> Write in the browser with your text editor.

This is really interesting being able to do bi-directional editing. This is desperately needed for accessibility software to get at underlying text. I would use it for semantic editing of code by voice. However, getting at GUI elements would be amazing.

This is awesome! If someone created a Interface Builder on top of that, I feel like I might even go back to making iOS apps. Xcode makes the whole experience so terrible that it sucks the joy out of it.

i wonder: could this kind of bidirectional flow be used to turn any LSP-capable editor (without it having to have any collaborative editing function integrated) into a collaborative editor that real-time-syncs any modification by any connected user with all other users who have the modified file open?

Very cool concept. Would love to see something like this for web development. Where making changes in the browser could be persisted seamlessly to code

This stuff is a dream of mine, editors in different modes that can collaborate seamlessly. Keep up the good work!

Whoa … you can use various unicode characters in titles on HN?

I was surprised by this too! I actually did a search for that character in order to check before posting this: https://hn.algolia.com/?q=%E2%87%84

I was a bit concerned that HN's algorithm might down-weight posts with non-ASCII characters in titles to discourage people from trying to attract attention with them, but it seems like it's fine?

So if we replace LSP with MCP, the UI might control the code in interesting ways!

See also: zoo.dev, started by an acquaintance of mine.

I guess it is about time for the usual Lisp, Smalltalk, Cedar and Oberon remarks.

Why do cool ideas take so much time to be embraced in mainstream?

Retoric question, naturally they weren't VC friendly with exponential growth capitalising user acquisition. /s

Reply on news.ycombinator.com