[WIP] AI Chat (Gemini)

[belabh]

2025-09-18_13-06-54.mp4

I just made an AI chatting module 😄

[belabh]

update.mp4

UPDATE!!! 😃

Made new tools
Now the AI can read and edit the config file
Model selector

[El-Cid-02]

How do I install this on my computer?

[belabh]

It's still work in progress. I need to fix a few issues before I can open a PR

[El-Cid-02]

Alright, we are looking forward to it

[Arch-Martin]

Bro I need this!!!!

[Cleboost]

We need an API key to use it?

[belabh]

Yes, you need an API key

[Arch-Martin]

Bro, how can I implement this on my system? I NEED IT, PLEASE! HELP ME!!!

[Cleboost]

It's still work in progress. I need to fix a few issues before I can open a PR
#642 (reply in thread)

[belabh]

update.mp4

UPDATE (again) 😄

Added:

• new tools: run command (runs commands on your machine), search tool
• tool confirmation
• markdown code blocks with syntax highlighting
• tool tips
• personalities (The AI could be funny, nerdy, or even rude)
• UI improvements

[leithXD]

Can i ask how did you achieve that transparency?

[belabh]

Can i ask how did you achieve that transparency?

~/.config/caelestia/shell.json:

{
  "appearance": {
    "transparency": {
      "enabled": true,
      "base": 0.6,
      "layers": 0.5
    }
}

[leithXD]

Thats so cool Hope it comes soon i wanna try xD

[madeupandprobablydoesnotexist]

[I wrote up a rant on different providers then realized this more sane option]

I think any actual ability to config this AI chat would be awesome as it should be convenient to use. Gemini is nice, but often the API will just not respond, which leads to quickshell freezing. Rather than juggling a bunch of providers, a lot of services offer the same openai sdk compatible endpoint. Most other awesome options such as groq, openrouter, pollinations(?), etc. all offer this. It's honestly a surprise that the normal ai chat does not have the option of simply supplying a different url and api key for it. I think end-4's dotfiles added more options of providers, so these all could be added manually, but that's ultimately not necessary.

[SonnyBoy-7]

I appreciate the UI, clean and functional. The feature of an assistant aware of the shell context, while still able to leverage typical LLM capabilities, looks very useful.
It would be nice if a configurable system prompt (e.g. via an .md file referenced in ~/.config/caelestia/shell.json) were allowed, as well as support for multiple providers with OpenAI API compatibility and local models. Forcing “personalities” on the assistant feels redundant (IMO); I would personally prefer a neutral but configurable system prompt, which is more flexible. Nonetheless, from what you showed, it looks like a valuable feature and I’m sure I’m not the only one looking forward to it.

[belabh]

For the system prompt part, it’s already customizable
You can either use the default systemPrompts/personalities or set it to custom/xxxx, and it will load from ~/.config/caelestia/ai/prompts/xxxx.md
There’s also a userPrompt option for user preferences or notes

As for more providers, that’s not something I can add right now. It would require refactoring the service to properly support Gemini and others, and I don’t currently have the time for that

[SonnyBoy-7]

Thanks for the clarification on the custom system prompt setup, that's exactly what I was looking for.

I'm really looking forward to the PR so I can take a closer look at the codebase. I'm not a developer myself, but I'd love to help if possible. Once the code is available, perhaps we (as this community) could explore adapting a more flexible approach for provider selection, without compromising the slick UI and features you're already building.
I imagine you're already aware of possible approaches, so it might just be a matter of finding the right balance.

Either way, this WIP looks like a really valuable addition to the shell. Thanks for putting in the work and sharing your progress.

[nuggetpilot]

Maybe they can look at how end-4 added other models in their dotfiles, because just relying on Gemini alone is very frustrating in quickshell as it often has moments where the API won't respond and quickshell freezes. Thankfully the openai sdk is used by almost all providers and hopefully won't take long for them to let you put a custom url & API key, as this sounds like a busted copilot

[Lazy-BOT19]

Hello, I love it I was just about to do something similar, is it done? Have you published it yet? lmk

[Arch-Martin]

Bro, please I need this!!!! 😲😲😲😲

[Arch-Martin]

How much time???????????????!?!??!?!?!?!?!?!?!?!?!?!?!??!?!?!?!?!?!?!?!?!??!?!?!?!?!??!?!?

[MicheleLandini]

is there still much left?