Hi, I explored a bit the usage of AI for omarchy and imo this is already partially usefull with real potential.

# Approach

<img width="200" height="300" alt="Image" src="https://github.com/user-attachments/assets/74f4f72a-c32c-4150-893e-7db2b9ffe544" />

Using this agent model, a possible workflow is:

1. Goal: User asks for something. Agent interviews if unclear until task/goal is defined.
2. Reasoning: Agent figures out what needs to change.
    - Needs: context, tools
    - Tools: Provided by LLM CLI (Read, Write, Execute, WebFetch)
    - Context: The critical part (more below)
3. Feedback loop: Shows diff and plan
4. Feedback loop: User approves or rejects or corrects
5. Execute & Rollback if it fails

# omarchy-config-agent
So this agent does basically the workflow sketched above
'inquire → plan → diff → approve/reject → execute → rollback if needed'

Code here https://github.com/parnoldx/omarchy-config-agent

Tested with claude code but should also work with gemini and others if the cmds are added

## How It Works
Omarchy being opinionated helps, the environment is defined, we don't need to handle e.g. KDE, and the choosen components use text config files.
I gave the agent context for some example components (base omarchy/utils,arch,hyprland,waybar):
In case e.g. hyprland it knows:
- Config file locations
- Introspection/validation (hyprctl commands)
- Default configs and examples
- Official wiki links (local wiki could be better?, for arch wiki I point to local clone if present)
So it has enough knowledge to make a good config change plan for hyprland.

With good docs/context, the agent was successful.
So e.g it changed in my tests sucessfully border radius, gaps, clock format, open a terminal calendar when clicking the clock, changed keybinding from spotify to yt music, ... 

When docs were missing (remove the locale icon in the top bar), it guessed wrong but I caught it in the diff.

So it will:
- Shows exactly what will change
- User decides. Learn if he wants, or just approve if yolo
- Rollback if something breaks

It can also just give "educated" answers when it has context. Like giving exact how-tos and similar stuff. For the answers it will also do introspection and so on.

The rollback part I did not dive deep yet what's the most reliable approach. btrfs snapshots? undo bash scripts? I am also no linux expert so I am sure there are people who have better ideas here.

Permissions are in yolo mode right now. In production possibly more usage of Allow or Disallow Tools

## What needs to be done
If the agents has all **context** it needs I think it can get really capable.

Imo what it comes down to is giving it the proper documentation. Some docs like how to introspect and validate are agent specific but after they are in place the agent will be probably pretty capable already. Then there will be edge cases for specific stuff pretty much like the discord help topics or github discussions are right now about how to achive something. So these atomic how-tos just have to be accessible as context for the agent in whatever way. They can be consumed by the agent and also by people copying cmds and editing files manual since it can just be markdown.
This probably also creates better documentation overall - clear, structured docs accessible to both agent and humans.

## Questions
- Is this a direction that makes sense and is a agent in whatever form even wanted? Any concerns? 
- Should the agent handle everything a user could do? Or are some things too risky?
    - Window manager config: clearly yes
    - Package installation: probably?
    - Boot configuration, Kernel parameters: scary? But the normal user also does not know shit about it and just paste some googled cmds
    Or is "show diff + require approval + rollback" enough safety for everything?
- How should this integrate?
    - I just use claude code TUI as interface for this experiment because it has everything needed already built in but if they have the functionality this can also be achieved by some generic LLM cli with all the cloud models or local models. In theory this could also be something custom...