Built a Linux OS where you tell it what you want and it actually does
it, not another chatbot sidebar

Been frustrated with every "AI PC" announcement for the past two years. They're all the same thing. A chat window. That can't do anything. You ask it to open your files and it tells you how to open your files yourself. So I started building an OS that actually does what you say. You type "prepare my coding workspace" and it opens your editor at the last file you were working on, restores your terminal to the right directory, starts the language server, and has your last session context ready. Not by magic, by actually controlling the OS. It's Linux underneath. Full Linux.

Every app you use today still works. apt, flatpak, your terminal, your dotfiles, everything. The AI layer sits on top and augments it, it doesn't replace it. You can ignore it completely and use it like a normal Linux desktop if you want. The part I'm most proud of is something I'm calling HAL, Human Approval Layer. The AI cannot just do things. Every action it wants to take gets risk-scored first. Opening a file? Silent, just happens. Moving files around?

You get a notification with an undo button. Deleting something? Confirmation dialog. Touching anything system-level? Full explanation of what it wants to do and why, explicit approve or deny. The AI serves you not the other way around. No "are you sure you want to do that?" for every little thing, but also no AI quietly reorganizing your home directory while you sleep. Privacy was the other thing I couldn't compromise on after watching the Recall disaster. Your semantic memory, the index of your files and work context, lives in \~/.cognos/memory/. Normal directory. You can open it in a file manager. The AI processes literally cannot make network connections, enforced by kernel-level nftables rules not a settings toggle. Runs completely air-gapped if you want. I've been building this with a PhD student who works on safety-critical neural implant systems at UCSD.

His background in safety-critical architectures has been really useful for the HAL formal model , turns out "how do you gate autonomous actions in a life-critical system" is basically the same problem whether you're building a brain implant controller or an OS. ore infrastructure is working , HAL skeleton, CI pipeline with HAL authorship enforcement, UNIPKG unified package manager, AI terminal assistant, kernel config. Building out the intent engine and memory system next.

Repo: https://github.com/iCrewZero/COGNOS Would love feedback from people who actually use Linux daily. EIt is rn in the development phase not production phase.

https://redd.it/1tovxmj
@r_linux