Looks good, thanks for sharing. I'm still too comfortable with LazyVim to go this route ... Maybe now with Pack, it'll be easier to jump into it.

I’m a big fan of distrobox although I don’t use it much anymore.

You can sandbox the neovim appimage with AM and then it will ask you what locations you want to give access to. https://github.com/ivan-hc/AM

am -i nvim

am --sandbox nvim

If you want to go the "packaging way", you could use nix's nixCats-nvim to make a fully hermetic nvim installation where you track the origin of all the dependencies (LSPs too) and plugins, all with receipts and hashes and all the good stuff of a reproducible build system. The security industry likes reproducible build systems because there's only one way you can go from source to the artifact.

Then, you package that in e.g. a docker container (which nix can build for you, too) and ship where you need it.

there's already official appimages for neovim, I use them (but not with firejail) and they work fine. but the lsp stuff is all installed outside, my guess is it'd be really annoying to mantain a custom compiled appimage

the best option is probably to just run neovim inside a docker container, you can then mount the directories every time you run a neovim container (~/.config/nvim, ~/.local/share/nvim, optionally .local/state/nvim and your undodir/undofile, there may be more I'm not sure).

assuming you want to isolate your home directory, what gets annoying is giving it access to only the code you're editing, I've yet to try this but my next plan is to give it read access to $HOME, then read/write access to the neovim dirs, and then take an argument to mount the project directory

some other options which I also haven't tried are

• distrobox, which just creates containers as well but by default wants to mount your entire homedir inside them
• devcontainers
• lspcontainers which isolate the LSP binaries themselves
• LXC/LXD/Incus containers which behave a bit different from docker containers, could be worth looking into

no

I've used gen.nvim and codecompanion.nvim, both with a local ollama server running.

I think gen.nvim was a little more straightforward to use, but in general I don't find LLM-aided coding to be that helpful to me.

I gave a try to ollama.nvim, but I'm not conviced (not by the plugin, but by using LLM directly in IDE). Because of security reasons, I cannot send code to public LLMs, so I have to either use my company's internal LLM (GPT4o), but which just have a front end, no API. Either I have to use local LLM through ollama for ex.

I tried several models, but they are to slow or too dumb. In the end, when I need help, I copy/past code into LLM portal front end.

I'm still a bit sceptical over LLMs but I've used copilot for line code completion for a few years now.

More recently I've started playing about with agents. I've been using avante.nvim for a couple of months with Claude and quite like it. The setup was a bit of a faff but it works quite well and has support for the more mainstream agents.

Over the last few days I've started trying out augment code which has a neovim plugin. Still on the fence about that one.

I'm no expert when it comes to any of this but this is what I've been messing about with. Hope it's useful!

I was also sceptical, but then I started to use codecompanion.nvim with Hugging Face recently. It works quite well and can be helpful for coding.