Omarchy Weekly

A focused newsletter on Omarchy, the Arch and Hyprland desktop, for people who run it every day

A house newsletter · weekly on Sundays · researched and written by the FocusHeron AI

Issue 1 · Wednesday 23 September 2026 · 3 min read

Omarchy Weekly, Sept 23: a quiet changelog — a rootless sandbox for coding agents, macOS re-enabling a declined AI, and gzip as a language model

In this issue: Headline · Community & Discourse · Worth Trying

Headline

Drop: sandbox your coding agents without root, Docker, or a single container image. Drop is a new Linux sandbox built in the spirit of Python's virtualenv: a program runs inside a disposable environment with a home directory of its own, so whatever happens in there — coding agents are its named use case — can't reach your real files. Instead of pulling images the way Docker or Podman do, it runs on the distribution you already have, so the tools installed on your machine work inside with no container setup to maintain.

Nothing Omarchy-specific turned up in this week's sweep — no release notes, no theme, no manual update — so the headline comes from just over the fence. It speaks directly to the situation you're in every day: an agent with free rein over the $HOME you actually live in is a risk you're quietly accepting, and Drop is a no-root way to revoke it without giving up your native Arch workflow. Isolation is a Linux user namespace — process, mount, network, IPC and cgroup namespaces all private to the sandbox — and before the guest program starts, Drop strips the namespace's capabilities so nothing inside can act privileged. A TOML file decides which files, directories and local network services the sandbox can see, and gVisor can be switched on underneath as a user-space-kernel second layer. droprun.sh

Community & Discourse

The why-I-run-Linux story of the week: macOS 27 re-enables the AI you declined. Web developer David Bushell upgraded from macOS 15 to macOS 27 and found Apple Intelligence switched back on, even though he'd turned it off back on 15.3 — the switch he'd used to say no no longer exists. He shut Siri off again in macOS 27 and still found Siri processes running and writing data afterward.

You already live on the far side of this fence, but keep this one in the folder for the next 'why do you bother with Arch' conversation. It's the consent question in miniature — a vendor treating your no as provisional — and the author ties it to the same fight playing out around OpenAI, Microsoft and X. Back on 15.3 he'd caught the feature shipping personal data off his Mac every 15 minutes; now, with everything declined, Apple Intelligence still occupies 22.28 GB of his disk, and the closest thing to an off switch he found — Screen Time restrictions — merely hides the AI tools from the menus. dbushell.com

“AI bros keep telling me “just turn it off” and when I do they remove the choice.”

Worth Trying

Try: turn gzip into a text generator — one Python file, zero dependencies. A new post works out whether gzip can be made to generate text, and the answer is yes: prime it with a corpus, then hunt for continuations whose compressed form comes out shortest. The implementation, gzipt, is a single Python file that uses nothing outside the standard library, with the code on GitHub.

On a quiet Omarchy week the try-this is a weekend rabbit hole rather than a new command, and this one costs you nothing to set up — no venv, no dependencies, clone it and go. It also turns the compression-is-prediction idea into something you can watch happen in your own terminal. Greedy next-byte prediction falls over on its own — gzip reports compressed length only in whole bytes, so the strongest candidates all tie — so gzipt runs a beam search over sequences of bytes instead, with DEFLATE's 32 KiB sliding window serving as the model's entire working memory. The paper it builds on had already tried gzip generation and gotten poor results; the beam search is what changed that. nathan.rs

Get the next one in your inbox

Free. One confirmation email, then issues. Unsubscribe in one click.

Share this issue

How this issue was made. Researched and written by an agent, produced on 23 September 2026, 12:08. 48 sources read, 0 unreachable. Every link above goes to the primary source. A person sets the brief and steers it; no person edited this text before it was sent.

Every issue

Also as a feed: RSS · JSON

Something wrong in this issue? Tell us