Still on Arch, btw

Six months is long enough for an experiment to become a habit, and a habit to become a preference, and a preference to stop announcing itself entirely. That is roughly where things stand now. The return to Linux has stopped feeling like a return and started feeling like where I live. And where I live, specifically, is CachyOS — an Arch-based distribution I migrated to on both my desktop and laptop after spending enough time with stock Arch to confirm the original intuition, then spending enough more to notice what could be sharper still. CachyOS is not a reinvention. It is Arch distilled: the same philosophy, a kernel and scheduler tuned for performance, and an installer that doesn't treat the first hour as a hazing ritual. Boot times and shutdown times remain dramatically superior to Windows. That hasn't faded with familiarity; if anything, the contrast has sharpened, because you stop waiting for something to be impressive and start waiting for the day it isn't.

The KDE desktop customisation deserves its own mention, not because it is flashy — though it can be — but because it is genuinely complete in a way that matters. The ability to truly make the desktop your own is not a marketing phrase here. Every visual element, every panel, every spacing decision bends to preference rather than prescription. The result is a machine that feels inhabited rather than rented.

A few things still don't work. My Elgato 4K Pro capture card is, as far as Linux is concerned, a very expensive piece of nothing. My DVB-S2 card sits in a PCIe slot doing close to the same. I had known about these limitations before the move; I had expected them to sting more than they do. But there's something clarifying about returning to Linux and discovering that the hardware you haven't reached for in months wasn't pulling its weight anyway. Both of those devices were investments in use cases that quietly collapsed long before the OS change. The kernel didn't take anything from me. It just made explicit what I'd been pretending otherwise.

Gaming was the other shoe I'd been waiting to drop. It hasn't. What Steam and Valve have done with Proton is genuinely astonishing — the right word, not a casual one. Out of the box. That phrase sounds like marketing hyperbole until you actually try it and find yourself three hours into a game you expected to spend two hours configuring. Not everything works, but the proportion that doesn't has become small enough that it no longer defines the experience. For years Windows held gaming as the unchallengeable trump card. That card has been played down to a bluff.

I should say clearly that I am not a certain kind of Linux user. I don't use vim. I don't use Emacs. I have no particular intention of learning either, and I have made my peace with that. My terminal text editor is micro, running a Nord theme, which does everything I need a terminal text editor to do without requiring me to memorise a modal system on top of learning everything else. I use ncspot for music, btop for system monitoring. I am not performing austerity. I have simply found the tools that fit, and they happen to be FOSS, and they happen to be good.

Which brings me to something I didn't anticipate writing about when I sat down to draft this. One of the animating complaints in the original post was the tendency to bolt features onto surfaces where they don't belong — AI woven into context menus, search bars, sidebars, notifications, places that didn't ask for it and don't benefit from it. Microslop was the word I used in my head, if not on the page. I stand by it. But the last six months introduced a complication: I started using Claude Code, and on Linux, it doesn't feel anything like that.

Part of what I've done with it has nothing to do with code. I had it reorganise years of accumulated Red Alert 2: Yuri's Revenge maps into a sensible folder structure — one of those tasks that is simple in concept and tedious in execution, the kind you defer indefinitely until something external forces the issue. I also used it to set up multiple instances of the game with different mod configurations: Mental Omega, the standard CnCNet setup that ships with the Steam copies, and Apocalypse Mod, each with quick launch entries in Lutris. It worked. More than worked — it was the least friction I have ever had with that particular category of setup task.

But the reason it felt like a feature rather than an imposition is structural, not incidental. AI agents trained on language and tooling are substantially trained on Linux CLI tools — sed, jq, find, ffmpeg, the whole kit. When you're working in that environment, the agent is operating in its native habitat. It doesn't reach for workarounds. It uses the same tools you'd use if you remembered every flag and every syntax. And then — this is the part that matters — you close it. It goes away. It is not residing in your taskbar. It is not reappearing in your notification shade. It is not asking whether you'd like help writing your next sentence every time you open a new window. You summoned it, it worked, you closed it, the desktop is quiet again.

There is an idea here that I think is worth building toward: a Linux distribution with an agent harness designed specifically for that distribution. Natural language as a configuration layer — not replacing the terminal, but sitting comfortably adjacent to it. Describe the thing you want done; the agent translates it into the commands that would do it; you review, confirm, proceed. And when the task is finished, there is a button. You press it. The AI goes away. That is the entire pitch. Not ambient intelligence. Not always-on assistance. Invokable utility with a close button. The operating system gets out of the way, and so does the agent that serves it.

The original post ended with the observation that good systems don't shout. Six months later, I'd add a corollary: neither should the tools that run on them. The measure of whether something belongs in your workflow is not whether it is powerful — most things are powerful — it is whether it knows when it's done.