Thank you very much. Great ideas. LSP support was (and likely still is) planned via a plugin system, which is already a part of rhun. I'm trying to be really careful with feature creep and load time, so I postponed some features to stabilize the codebase and gather user feedback first.
Edit: I have built rhunpad[0], its slightly built for myself
Now, I was having some issues with LLM not able to really work with the native file manager capability [I would say that this probably happened because I wanted there to be two types of system, debug and prod, where debug asked for folder everytime in scratchpad so that I can test it behaviour and prod only asked once]
And I try to write the prompt and pass it through a middle layer like chatgpt to then pass it onto the final layer and it seems that it preferred the macOS file picker in that sense
but finally I have created the project, I think that it was a really pushing the limits vibe-coding project but quite impressed nonetheless that an open weights model was still able to do these edits, there are some other issues as well but I think that this was just created by me for myself to write thoughts in a scratchpad like way.
I will be waiting for a v1 release or a stable release/the plugin system that you are mentioning for a more proper release as well but just wanted to let you know that I built it on top of your project and I think that there are just so many more ideas that I can think of if someone really wants to test the limits of LLM's while getting something decent. Creating a monkeytype alternative comes in my mind as well
(I did use the word rhunpad, I hope that it can be okay and if not then I am more than happy to change the name as well!)
Anyways, I'd be curious as to what you think of it and as always have a nice day and take care!
Yup, I had left the agent (GLM 5.3 on maki) to work on creating rhunpad and I must say that its unwieldy for the model to handle the codebase although I guess it is still able to work though but definitely testing the limits of the model in terms of making it overthink substantially more.
I would really like a plugin system which can make the codebase more easier to wield for these custom use-cases. Atom created Electron for their editor which then made literally everything use a browser for a native app, I can imagine that perhaps a similar yet hopefully more minimalist primitives can be provided by a strong plugin system to help other people make easier (side-editors?), plus points for using that for LSP purposes as well, also good point on preventing future bloat, I think that the future looks good provided that the codebase becomes managable for AI to use as its definitely testing the limits of AI in some sense imo which is strangely why I am even more curious as to how this project pans out as well and in terms of how it maintains the software/source code.