I’m not sure what one has to do with the other. I used LLMs to accelerate development of dialog-tool, and did the same for dialog-ide. I’ve looked over the generated code and it’s quite good, generally well structured and well documented. If LLMs disappeared, I could continue maintaining either, just not at the same pace. I’m sure there are more people out there who understand TypeScript and VSC than Clojure.
The website describes VS Code as “The open source AI code editor: Your home for multi-agent development” and vibe coding is pretty much the only thing the page talks about. Even when coding by hand, the website makes it clear that it will be running an LLM in the background to provide inline suggestions.
And that’s just not something I want out of an IDE.
(To be clear, my concerns are about VS Code, not the Dialog IDE components specifically. But unlike dgt, it looks like Dialog IDE can’t work without it.)
I used to use vscode, and while I don’t like it for other reasons (i use vim now), the AI is definitely possible to turn off. It’s more that microsoft has some strange obsession with AI and LLMs, and also marketing i guess. so i wouldn’t say VScode is just for vibecoding
edit: oop did this not actually save as a reply? i havent actually used the forums here in a bit
I really wouldn’t categorize this as a Microsoft / not Microsoft phenomena. Visual Studio proper, the original Visual Studio, has been stuffed with it as well. It is opt-in by default but fortunately you can go on an expedition into the IDE settings to turn it all off.
The part that probably matters most is the GitHub Copilot integration and the side effects of using it. This opens up your private repositories as training data for their AI campaigns. I’m probably over simplifying it because I’m tired but I had to not only disable all the AI unsolicited help from the IDE but I also had to opt out of it in my GitHub account. All users globally had some new AI based usage settings added to their online portal and they were turned on by default.
That’s good to hear! In that case I’m more willing to try it out. I do prefer command-line tools that I can script to work exactly how I want them to, but I’ve used Inform 7 for years via its IDE without anything terrible happening.
Okay, for anyone else who’s like me on this—you need to go into File > Preferences > Settings, search for chat.disableAIFeatures, and tick that checkbox. That seems to get rid of all of it, and I didn’t need to provide any of my logins or personal information.
And the syntax highlighting is working, which I’m going to take as a sign that the IDE is properly installed! Next question: is it possible to import an existing dgt project?
You will have to manually convert dialog.edn to dialog.json.
But the Skein files will port over correctly? Those are the ones I’m worried about; the EDN file is easy enough.
Yes, though you might see tiny mismatches for whitespace at the end of some knots. You’ll see this after replay all and bless transcript (and clicking the error count) is your friend. Some logic was slightly different and (for an installed base of maybe 3 people) it wasn’t worth chasing it down.
But every other aspect was carefully duplicated; I too have a huge investment in my skein for Sand-dancer and a few others.
Version 0.6.0 will have the ability to create or open a skein from the command line with dgbuild, so you should be able to have a quite reasonable Emacs workflow. You might want to dip into VSC to set up exports, perhaps.
Version 0.6.0
- Add
dgbuild bundle [export-name]- builds the web page (out/web/+ zip) for adialog.jsonexport configuration headlessly, so a GitHub Action can publish a release - Add
dgbuild new-skein [name]/dgbuild open-skein [name]- create or open a skein and drive its full interactive browser UI with no VS Code; auto-opens a browser (--no-opento skip), with an in-UI Quit button that prompts to save when there are unsaved changes - Fix the Trace panel indenting each nested level cumulatively more than the last, pushing deep traces far off the right edge
On my 2019 Macbook Pro (2.8 GHz 8-core i9, 32 Gb ram) it runs the 278 individual paths of the stress test in 1 minute, 10 seconds.
My monster Mac studio does the same in 15.5 seconds.
I’ll be looking forward to feedback on performance against further real projects. Perhaps I can pull off a magic trick to make it more performant, if that’s a problem.