Generated from docs/internals/ in the main repository. Edits made here are overwritten on the next push to main.

The reasoning behind the repository's flake, moved out of the file so it reads as code. Every entry is reachable from a # Why: pointer at the line it explains.

Deliberately not in the root AGENTS.md: that file is symlinked as CLAUDE.md and loads into every session, so a thousand lines of build reasoning there would cost context on every task that has nothing to do with it. This page is read when somebody follows a pointer.

• inputs — nixpkgs is the package set; hyprland, zen-browser, microvm and the rest are pinned separately because they publish on their own schedule. Some pins are deliberate and must not be bumped by an automated job; the nightly review reports them and flake-update.yml names nixpkgs explicitly for that reason.
• overlay — how a package reaches a configuration. A module cannot take packages from inputs.self.packages: those are built from nixarchy's nixpkgs, and mixing instances makes buildEnv refuse a profile holding two builds of the same version.
• packages — the ISO, the installer, the doctor, #try, #vm.
• checks — 37 of them; tests/AGENTS.md covers what each proves.
• nixosConfigurations — reference is the machine the installer writes, vm is the smoke-test guest, iso/iso-net are the images.

• A check here and in no workflow is not a check. build.yml asserts every checks.* is run by some workflow, because one shipped that way and its pull request went green without it ever executing.
• self.rev is absent on a dirty tree, and installer/mkFlake.nix throws rather than pinning a user's machine to a store path that may be garbage collected. A build from an uncommitted tree fails on purpose.

nixConfig = {

For the person who has never seen this repository and types nix run github:olafkfreund/nixarchy#try (or #doctor, or #vm): without this, nothing tells their nix that this flake's packages and closures are already built and cached, and Hyprland alone becomes a compile. nixConfig is honoured only from the flake you invoke -- the same rule the sops-nix comment below leans on from the other side -- so the module's programs.nixarchy.binaryCaches cannot reach a bare nix run; only this can. Nix asks before trusting either entry (or accepts them silently for users already trusting the caches), and a "no" costs nothing but the download it was offered.

What this does NOT cover, measured rather than assumed: the ISO images themselves are not pushed to cachix -- they ship as release assets -- so #try detects that case with a dry-run and falls back to the release download instead of leaning on these substituters.

home-manager = {

Deliberately UNPINNED, unlike hyprland, sops-nix and microvm below, and not for lack of tags (home-manager has none; its releases are release-XX.XX branches that pair with STABLE nixpkgs, which is not what this flake tracks). master is co-developed against nixpkgs-unstable, and nixpkgs above is itself a branch ref: one nix flake update moves both to their tips together, which is the only pairing home-manager tests. Pinning one side of that pair would CREATE skew -- a frozen home-manager evaluating renamed and removed nixpkgs attrs a few updates from now -- which is the exact failure pinning is supposed to prevent. Pin this the day nixpkgs is pinned, and not before.

hyprland.url = "github:hyprwm/Hyprland";

The compositor comes from upstream's own flake, and their binary cache serves it, so no machine compiles it. That is only true because inputs.nixpkgs.follows is deliberately NOT set on it: hyprwm ask consumers not to override their nixpkgs, and overriding it forfeits the cache and rebuilds the compositor from source.

A branch, not a revision, and that is the whole lesson here. This input used to read

hyprland.url = "github:hyprwm/Hyprland/0bd11c7a04a63d2785abd53363f09d552175d67d";

with the revision in the URL. nix flake update cannot move that, so the nightly bump skipped it silently, and nix run .#review watched six pins with hyprland among none of them. It sat at 0.56.0 while nixpkgs reached 0.56.2 and Arch shipped 0.56.2-2 — the pin meant to keep us ahead of nixpkgs had put us behind it, and two weeks of drift looked exactly like none. It would have looked the same at two years.

A ref is visible to the tooling that exists to catch this: nix flake update moves it, and pkgs/flake-pins.py reports it as a ref rather than a frozen rev.

What this buys, and what it costs, measured rather than assumed:

Until #1248 the key for hyprwm's cache was wrong, so Hyprland came from nixarchy.cachix.org or was built; checks.cache-keys now verifies the key against a real signature.

That last row is the real cost. Evaluating this input pulls hyprwm's whole build graph — aquamarine, hyprutils, hyprlang, hyprtoolkit, xdph and a nixpkgs and systems for each — from GitHub, every update. On 2026-09-07 a machine's update failed on exactly that: GitHub returned 504 for the aquamarine archive, reproducible from outside the VM, and nothing could proceed. nixpkgs' hyprland fetches none of it.

The trade was made deliberately: the compositor is what this desktop is, and getting its fixes on hyprwm's schedule is worth an occasional upstream outage during an update — which is retryable, where waiting on a nixpkgs bump is not.

Omarchy itself pins nothing: install/omarchy-base.packages names hyprland with no version, so upstream gets whatever Arch shipped that day. Tracking hyprwm's branch is the closest a flake gets to that, and usually ahead of it.

Its packages, not its NixOS module (#1083). The module set programs.hyprland.package at mkDefault on every machine that imported nixarchy, including one with nixarchy off. nixarchy sets package and portalPackage itself under programs.nixarchy.enable, and links /share/hypr as the module did. A configuration that uses the module's own options (plugins, settings, extraConfig, topPrefixes, bottomPrefixes) imports inputs.hyprland.nixosModules.default itself.

The >= 0.55 assertion in modules/nixos.nix stays regardless. Omarchy 4.x configures Hyprland through the Lua API that landed in 0.55, and that is a requirement rather than a preference.

Hyprland loads its GPU drivers into its own process, not into a separate one it talks to over a socket -- so the driver it opens has to run against the glibc Hyprland itself was built with. #1154 bumped our own nixpkgs and left Hyprland's untouched, and the result was GLIBC_2.43 not found (required by .../libEGL_mesa.so.0): our nixpkgs' Mesa had moved ahead of the compositor that dlopens it, and the session never started.

So hardware.graphics.package and package32 are set from Hyprland's own nixpkgs pin (inputs.hyprland.inputs.nixpkgs), the same one programs.hyprland.package already comes from, rather than from this flake's own. checks.graphics-glibc holds the pairing: it fails whenever the drivers' glibc is newer than the compositor's.

hyprPkgs = inputs.hyprland.inputs.nixpkgs.legacyPackages.${pkgs.stdenv.hostPlatform.system};
hardware.graphics.package = lib.mkOverride 900 hyprPkgs.mesa;

Only where the system's glibc is at least as new (Issue #1211). That driver is system-wide, so every OpenGL program on the machine loads it, not only Hyprland. On a nixos-26.05 machine Hyprland's pin is still unstable (the installer template never overrides it), so its Mesa needed a newer glibc (2.44 on 2026-10-07) than everything built from 26.05 (2.42). The SDDM greeter could not load it, and the machine had no login screen. glibc is backward compatible, not forward, so the override applies only when lib.versionAtLeast pkgs.glibc.version hyprPkgs.glibc.version. Where it does not, the system keeps nixpkgs' own Mesa, which the greeter, the user's programs and Hyprland's newer glibc can all load. The same rule covers a Hyprland bump that moves its glibc ahead of this flake's nixpkgs: the system Mesa is kept there too, until the two meet again.

Not lib.mkDefault. nixos/modules/hardware/graphics.nix sets both options at mkDefault itself, unconditionally once hardware.graphics.enable is on -- which Hyprland turns on -- so a second mkDefault here ties with nixpkgs' own and evaluation refuses with "defined multiple times". mkOverride 900 sits between that default (1000) and a plain assignment (100): nixpkgs' default loses, and a machine's own plain hardware.graphics.package = ...; still wins. Its own lib.mkDefault does not -- keeping the two nixpkgs pins in step is this module's job, not an adopter's, unless they say otherwise with something stronger than a default.

The cost is paid in store space rather than eval time: the two pins' Mesa builds are only the same derivation while nothing in Mesa's own dependency chain has moved between them, which is not guaranteed to stay true. Measured 2026-10-02, on main's pins: identical store path (mesa-26.2.3, a closure of 60 paths and about 1.1 GB by nix path-info -S), so there is no second Mesa on disk today -- but the day the two pins diverge there, a machine carries both. The extra cost is then the paths the two closures do not share, well under that 1.1 GB.

An overlay on mesa no longer reaches the drivers. They come from Hyprland's nixpkgs, which sees none of the machine's nixpkgs.overlays. A patched or replaced Mesa now needs a plain hardware.graphics.package = pkgs.mesa; (and package32), and then keeping its glibc in step with Hyprland's is that machine's job: checks.graphics-glibc only covers the reference machine.

A second Mesa in extraPackages collides, so #1163 refuses it. hardware.graphics.package is only one entry in the buildEnv NixOS builds from package plus extraPackages; a different Mesa listed there (its own pkgs.mesa, say) ships the same share/glvnd/egl_vendor.d/50_mesa.json, and buildEnv collides. An assertion in modules/nixos.nix catches this at evaluation -- while this module's Mesa is still the one in effect, and only against a Mesa with a different outPath -- and names both store paths. The way out is either: drop mesa from extraPackages (this module's Mesa already provides it), or set hardware.graphics.package (and package32) yourself, which then makes its glibc your own job to keep in step with Hyprland's, same as the overlay case above. checks.graphics-mesa-clash holds the pairing.

nix-flatpak.url = "github:gmodena/nix-flatpak/v0.8.0";

Declarative Flatpaks, for the software nixpkgs genuinely does not carry.

nixpkgs' own services.flatpak is infrastructure only -- the daemon, the portals, polkit, FUSE -- and manages no applications. It does not even add the flathub remote. So the honest answer this repo has been giving is data/arch-extras.nix's note = "then: flatpak install flathub ...", which is advice to do the one thing this project exists to prevent: install something imperatively that will not survive a rebuild.

This adds services.flatpak.packages, overloading the same option namespace nixpkgs uses -- worth knowing, because a reader will look for packages in nixpkgs' module and not find it.

Declared, not reproducible, and the distinction is real: the app IDS are in your configuration and travel to a new machine, but the BITS are whatever Flathub serves that day unless every entry pins a commit. Rollback restores the list, not the version. Upstream nixpkgs has declined to bless the equivalent (PR #347605, open since October 2024) on exactly that objection. We take it anyway, because the alternative on offer is a shell command in a comment.

No follows: v0.7.0 has no inputs at all, which makes this the cheapest kind of dependency -- nothing to override, nothing to drift.

zen-browser = {

Zen is not in nixpkgs and upstream maintains its own flake, which tracks Zen's releases far more closely than a derivation here ever would.

Pinned to a COMMIT because upstream's only tags are automated twilight-* build tags, not releases of the flake -- main is the release channel, the same position sops-nix and microvm already hold in this file. What makes the pin worth its bump burden here is that this flake's default.nix couples to nixpkgs' wrapFirefox SIGNATURE: at this rev it overrides both the ffmpeg_7 and ffmpeg_8 formals, so it needs a nixpkgs that declares both -- ours does -- and a branch ref here hands whatever main has become to every downstream lock, against whatever nixpkgs THEY have, where an eval throw is a break our CI structurally cannot see (#320 was this, in one direction; stable consumers hit it in the other). A commit makes the pair deliberate.

What the pin costs: Zen security releases stop arriving on their own. Bump this routinely alongside nix flake update, and expect the bump to become MANDATORY the day nixpkgs drops its ffmpeg_7 formal, when this rev's override starts throwing. Never track a branch here.

hypr-rdp = {

The RDP server behind the remote-desktop feature (#159). Not in nixpkgs -- nothing under that name as of 2026-09 -- and the two RDP servers that are there cannot serve Hyprland: krdp needs the org.freedesktop.portal.RemoteDesktop interface, which xdg-desktop-portal-hyprland 1.4.1 does not advertise, and xrdp only reaches Wayland through a wayvnc chain built from an unreleased commit. So nixarchy carries the package until nixpkgs does.

The cost, stated rather than hidden: a flake input lands in EVERY user's lock, including the Mode A machine that will never enable RDP. It is one more ref nix flake update can move and one more repository that has to keep existing. That is paid because the thing it buys -- reaching this desktop from a Windows machine with nothing installed on it -- has no other implementation at all, at any price. See the alternatives survey on #159; every other route is closed, not merely worse.

What this project is taking on: pinned v0.1.6, one primary author. The protocol stack is not theirs -- it is IronRDP, Devolutions' maintained Rust implementation -- so what this author owns is the Hyprland glue: capture, input, audio, clipboard. That is a real dependency on a small project, and a bad rebuild here breaks the machine you are remote to, from where you cannot fix it. Hence a tag. Never a branch, and bump it deliberately.

The input has a named exit: when hypr-rdp lands in nixpkgs, delete this and point the module at pkgs.hypr-rdp.

That exit is now IN FLIGHT rather than hypothetical -- NixOS/nixpkgs#560204 ("hypr-rdp: init at 0.1.5") is open. When it merges and reaches the nixpkgs revision this flake pins, the deletion is:

1. remove this input and its follows
2. modules/services/hypr-rdp.nix: inputs.hypr-rdp.packages.${system}... becomes pkgs.hypr-rdp
3. check nothing else reads inputs.hypr-rdp (grep; the module is the only consumer today)

Do NOT delete it merely because the nixpkgs PR merged: a merge into nixpkgs master is not the same as being present in the revision we pin, and removing the input before then leaves the module reaching for an attribute that does not exist yet. The condition is pkgs.hypr-rdp evaluating against OUR lock, not a green tick on a pull request.

follows is right here and wrong for hyprland above -- upstream's pkg/nix/package.nix is a plain rustPlatform.buildRustPackage whose cargoHash does not depend on which nixpkgs supplies ffmpeg, and they publish no binary cache to forfeit by overriding it.

gliff = {
  url = "github:omacom/gliff/v0.4.0";
  flake = false;
};

gliff (omacom/gliff, MIT, Rust) is Omarchy's remote desktop over SSH: a gliff client, a gliff-server that ssh spawns on the far machine, and a gliff-probe for diagnosing both. Upstream's install/omarchy-base.packages names it (omacom/omarchy#14712), so a fresh Omarchy has it; nixpkgs does not carry it, so nixarchy builds it (#1226).

Why flake = false. Upstream ships no flake.nix, so there is nothing to take packages from. The input is a source tree and pkgs/gliff/default.nix builds it. Its Cargo.lock has no git sources, so cargoLock.lockFile needs no output hashes: nothing here to update by hand except the tag.

Why a tag. The same argument as hypr-rdp above: a bad rebuild breaks the machine you are remote to. Bump it deliberately, never track a branch. The exit is nixpkgs carrying gliff: delete the input and the overlay entry, and point modules/nixos.nix at it.

The RUNPATH trap. gliff loads Vulkan with ash::Entry::load(), which dlopens libvulkan.so.1 by name. Nothing links it, so nothing puts the loader on the RUNPATH; without it gliff does not fail, it falls back to the CPU tier and every machine looks fine. postFixup adds vulkan-loader to all three binaries (before wrapGApp, so it lands on the ELF that becomes .gliff-wrapped) and checks.gliff-runpath reads it back. The gliff client also runs ssh by name, so the wrapper puts openssh on its PATH. Measured on an RX 7900 XT: gliff-probe gpu reports PASS for VA-API H.264 encode and decode.

Why a preinstall and not an omarchy runtime dependency. The upstream remote-session plugin only pgreps for gliff-server; it does not need the package to build. As a preinstall, programs.nixarchy.preinstallsExclude = [ "gliff" ] can drop it, which a runtime dependency would not allow.

No cache allowlist entry. As a preinstall gliff rides inside vm-toplevel and reference-toplevel, which the cache already holds; what it adds is what cache-budget.sh reports on the PR.

sops-nix = {

Declarative secrets, adopted for one concrete reason rather than as a general capability. #121 deliberately did not take a secrets mechanism ("a separate decision with a key-management story attached"), and #122 got away without one because hashedPasswordFile is consumed by NixOS itself, so the cleartext could live outside git and be pointed at.

hypr-rdp removed that dodge. When #155 / PR #184 adopted sops-nix, v0.1.5 took a password only inline in config.toml or via -p; a command-line secret was readable in /proc/*/cmdline. That historical need to render TOML selected sops-nix over agenix's raw secret files.

The pinned v0.1.6 supports password_file from config or CLI and refuses a missing or empty named file. It still warns and serves without authentication when no credentials are set, or gives only a HalfCredentials warning for a username with an empty inline password. The current module keeps its guard: sops-nix writes a private password staging file, and at service start ExecStartPre checks it and writes private TOML. Migrating to upstream password_file is separate work.

Pinned to a COMMIT because sops-nix publishes no release tags at all -- its only tag, assets, is from 2021 and is not a release. master is the release channel. This is the same call flake.nix already makes for hyprland above: a commit is exactly as reproducible as a tag. Bump it deliberately; never track a branch.

What this costs a user's lock: ONE node. sops-nix declares exactly one input, nixpkgs, which follows ours, so nothing transitive arrives. (Its dev inputs live in a private flake under dev/ that consumers never see.) Its nixConfig asks for cache.thalheim.io, which does NOT apply to us -- nixConfig is honoured only from the flake you invoke, not from inputs -- so no substituter and no trust is granted by taking this.

Inert on import, which is what makes it safe for Mode A: both halves of the upstream module are gated, lib.mkIf (cfg.secrets != { }) and lib.mkIf (config.sops.templates != { }). A machine that defines neither gets no unit, no activation script and no package. nixarchy sets no sops scalar of its own -- sops.age.sshKeyPaths already defaults to the ed25519 keys from services.openssh.hostKeys upstream -- so the mkDefault rule in modules/services/default.nix never even arises here. tests/options.nix asserts the inertness rather than trusting this note.

nixarchy-pkg = {

nixarchy-pkg is the first of nixarchy's own shell plugins in programs.nixarchy.defaultPlugins. It is installed and turned on, once, wherever nixarchy is enabled (modules/AGENTS.md has the reversal of "plugins present, never on" and its reasoning). It is an input for nixi's reasons: same maintainer, its own release cadence and checks, and a copy here would be a second place to maintain it. It uses follows because it is QML plus a bash adapter, so there is nothing to build and no binary cache to forfeit, and one nixpkgs in the lock beats two.

Measured at dd937f2 (2026-09-19): the plugin output is 120 KiB. The input's source tree, which lib.inputSources carries onto the offline ISO and the installed host, is 6.0 MiB, mostly its docs images.

It is not in checks. modules/home.nix's validatedPlugins runs omarchy-plugin-validate on it in every home that installs it, the reference machine included, so a pin that moves to a broken manifest fails there. A separate checks.nixarchy-pkg would need a workflow edit (AGENTS.md section 4) and would check nothing more.

Bumping the pin. The repo has no tags, so there is nothing to track:

1. Pick the commit on main, and read what changed since the current pin (gh api repos/olafkfreund/nixarchy-pkg/compare/<old>...<new>).
2. Edit the rev in the URL, then nix flake lock. The lock diff must touch nixarchy-pkg alone; a second nixpkgs means follows broke.
3. Check that the manifest id is still nixarchy.pkg. The build fails if it is not, because the default set is keyed by it.
4. Build checks.menu-verbs, checks.options and checks.plugin. The row, the default and the enable-once hook all name the plugin.

nixarchy-podman = {

nixarchy-podman is the second of nixarchy's own plugins in programs.nixarchy.defaultPlugins, and the first with a gate: it is installed where virtualisation.podman.enable is true -- the Podman services row or boxes -- and nowhere else, because a podman panel on a machine without podman is a panel that fails. It is an input, with follows, for the reasons nixarchy-pkg's entry above gives: pure QML, nothing to build, one nixpkgs.

Measured at 03d9f02 (2026-09-19): the plugin output is 106 KiB, and ships its MIT LICENSE (its flake lists it among the copied files). The input's source tree, which lib.inputSources carries onto the offline ISO and the installed host whether or not podman is on, is 5.2 MiB, mostly docs images.

Not in checks, like nixarchy-pkg: validatedPlugins validates it in every home that installs it, and checks.menu-verbs builds a machine with boxes, so podman is on there and the Apps ▸ Podman row's id is checked against its manifest.

Bumping the pin. As nixarchy-pkg's, with two differences: the branch is master, not main, and the id to confirm in step 3 is nixarchy.podman.

nixarchy-gliff-manager = {

Gliff-Manager is Gavin Nugent's bar panel for gliff machines (MIT; his copyright is kept in the plugin's LICENSE). Upstream assumes Arch, so nixarchy pins our fork, olafkfreund/nixarchy-gliff-manager, which reads gliff's version from gliff itself, takes the channel from /etc/os-release, finds the NixOS-wrapped process names (.gliff-wrapped, .Hyprland-wrapp), runs no package manager and does not offer the Arch session fix on NixOS. It is a default plugin (nixarchy.gliff-manager), gated on preinstalls and not preinstallsExclude containing gliff: wherever nixarchy installs gliff.

The input follows nixpkgs and home-manager (the fork's flake uses home-manager only in its own check), so nixarchy pulls in nothing extra.

Bumping the pin. nixarchy pins a commit (the release tag's sha). Read the diff against both the fork and upstream (28allday/Gliff-Manager, whose last-synced commit is in the fork's UPSTREAM file), confirm the id is still nixarchy.gliff-manager, and build checks.options and checks.plugin.

nixarchy-devenv = {

nixarchy-devenv is the panel for per-project environments, and the second default with a gate: it is installed where programs.nixarchy.services.devenv.enable is true and nowhere else, for the same reason the Podman panel follows podman -- with no devenv behind it there is nothing to list and nothing it could create. An input with follows, like the others: QML plus one shell script, nothing to build, one nixpkgs.

It also carries the catalogue. The eight presets this repo used to hold in data/devenv-presets.nix moved there unchanged, with eight more beside them (java, java-maven, kotlin, dotnet, php, ruby, flutter, and a cloud generator that runs cloud-projects-templates). Keeping a second catalogue here would be drift waiting to happen, so pkgs/dev-init.nix went with it and nixarchy dev ... dispatches to the plugin's nixarchy-devenv command.

Measured at e003f00 (2026-09-19): the plugin output is 148 KiB (a 121.7 KiB closure) and ships its MIT LICENSE, which checks.options asserts. The CLI is 28 KiB of script, but its closure is 57.2 MiB: it is a writeShellApplication over coreutils, findutils, gnugrep, gnused and jq. It goes only to machines that turned devenv on -- which are already carrying devenv itself, and devenv bundles its own Nix. devenv, git and nix are deliberately not runtime inputs of it: it calls them by name from PATH, so a machine that never enabled devenv gets none of that closure through this.

Bumping the pin. As nixarchy-pkg's; the branch is main and the id to confirm is nixarchy.devenv. Run nix run .#devenv-presets after the bump: that check runs the pinned plugin's own templates check, so a bump that breaks a template fails there rather than in somebody's project.

nixarchy-gltui = {

Another of nixarchy's own default plugins, for the same reasons as nixarchy-pkg above: same maintainer, its own cadence and checks, and follows because it is QML plus python with nothing to build. modules/home.nix installs nixarchy's copy rather than upstream's output as-is: a runCommand that adds menu.managed (nixarchy declares the menu rows) and takes the MIT LICENSE from the source tree, because upstream's package leaves it out.

Measured at 0b827c6 (2026-09-19): the plugin output is 58 KiB, and the source tree lib.inputSources carries onto the ISO is 208 KiB. Its runtime tools cost more: glab is 49.9 MiB (its 88.6 MiB closure is otherwise glibc and friends every system already has), and python3 and xdg-utils are already in the reference closure.

Bumping the pin, as for nixarchy-pkg:

1. Pick the commit on main and read the diff (gh api repos/olafkfreund/nixarchy-gltui/compare/<old>...<new>).
2. Edit the rev, then nix flake lock. The lock diff touches nixarchy-gltui alone.
3. The manifest id must still be olafkfreund.gitlab-pipelines; the build fails if not.
4. If upstream's menu.py changed how it decides to register, re-read it: menu.managed only works if it still checks for that file.
5. Build checks.options, checks.menu-verbs and checks.plugin.

nixarchy-ghtui = {

gltui's origin, installed the same way and for the same reasons (#770 above): a runCommand in modules/home.nix adds menu.managed and takes the MIT LICENSE from the source tree when upstream's package omits it.

Measured at dfba799 (2026-09-19): the plugin output is 84 KiB, and the source tree on the ISO is 304 KiB. gh is 40.1 MiB (its 80.6 MiB closure is otherwise glibc and friends every system already has); python3 and xdg-utils are already in the reference closure.

Bumping the pin: as for nixarchy-gltui above, with olafkfreund/nixarchy-ghtui, the id olafkfreund.github-actions, and the same menu.py register check.

nixarchy-flatsnap = {

nixarchy-flatsnap is two things, and nixarchy takes both from this one input:

• The plugin (packages.default), a default in modules/home.nix like nixarchy-pkg. It is on wherever nixarchy is, and defaultPlugins.flatsnap = false opts out.
• The NixOS module (nixosModules.default), imported unconditionally in modules/nixos.nix next to nix-flatpak. It reads ~/.config/nixarchy/flatsnap.nix (which nixarchy-apply copies, #904) and turns it into services.flatpak.packages and, for Snaps, nix-snapd plus a reconciler unit.

nixarchy is now the one place nix-snapd is imported. default is the plugin's module and nix-snapd. A downstream flake that also imports nix-snapd.nixosModules.default declares services.snap twice, and evaluation fails before any assertion can run, so no check here can catch it for them. The manual's page says what to remove, and quotes the error.

Inert until used. With nothing declared there is no snapd, no setuid snap-confine, and no nixarchy-flatsnap-snaps unit. tests/options.nix (flatsnapSnapd) holds that both ways.

Measured at 12d33d5 (2026-09-23): the plugin output is 42 KiB. snapd's closure is 1.0 GiB, and reaches a machine only once a Snap is declared. The input follows nixarchy's nixpkgs and nix-flatpak, and adds one nix-snapd node (plus its flake-parts and flake-compat) to the lock.

Bumping the pin: a commit on the plugin's main, then nix flake lock --update-input nixarchy-flatsnap. Check that the lock still has one nix-snapd and no second nixpkgs, and that options (flatsnapIsADefault, flatsnapSnapd, the menu row) and menu-verbs pass. The plugin's own nix flake check (a VM test and a gating eval) is its release gate.

nixarchy-menu = {

nixarchy-menu is a Raycast-style replacement for the Omarchy menu (id nixarchy.menu, clonedFrom: omarchy.menu). It hands off to every default agent, and it has Nixi and skill-backed help rows. nixarchy takes its packages.<sys>.plugin.

Installed but off since #1283; turned on at first login only with programs.nixarchy.menu.default = "nixarchy". The enable-once hook reads enableOnceIds, which leaves the menu out unless asked for, while install, the clone restore and the placement map still read every declared default. Setup ▸ Defaults ▸ Menu switches at runtime and writes the same enabled-once marker the hook would, so the restore below still works after a click. It was on by default from #1052 to #1283; defaultPlugins.menu = false still declines it entirely. #946 shipped it off, through a per-entry enableByDefault = false: a menu = false in the defaultPlugins default attrset would not have survived a host setting { podman = false; }, which replaces that attrset wholesale. With the menu on, nothing used the field and it was deleted; menu is deliberately absent from that attrset, because absent means on.

It takes the stock menu's place. The entry has placement = "", so omarchy-plugin-enable gets no section, and the shell puts the clone in omarchy.menu's own bar slot. Before enabling it, the enable-once hook turns off any other enabled plugin with the same clonedFrom (an old evindor.keystroke, say). It has to be done in that order: disabling a clone after another is on restores the stock menu beside it.

Turning it off brings the stock menu back. Setting menu back to false removes the plugin link, but removing files never runs the shell's restore, so omarchy.menu would stay disabled with nothing in its place. The hook's enabled-once marker therefore records what a clone replaces. At the next login, for a marker whose clone is no longer declared and whose files are gone, the hook re-enables that source beside the clone's orphaned bar entry, disables the orphan, and removes the marker. The limit: the hook exists only while at least one default resolves. With every default off, run omarchy plugin enable omarchy.menu once by hand.

A hand install wins until removed. A real directory at ~/.config/omarchy/plugins/nixarchy.menu, for example from nixarchy-menu's own bin/nixarchy-menu install, is left alone, as for every plugin. The hook logs the fix to nixarchy-default-plugins:

rm -rf ~/.config/omarchy/plugins/nixarchy.menu   # then log in again

The inputs follow. Its only inputs are nixpkgs and omarchy (flake = false), so both follow nixarchy's and the lock gains one node. Since #1070 omarchy is also a build input of the plugin output, not checks alone: one file (MenuModel.js, about 18 KB) is copied into the store and imported by absolute path. The 128 MiB tree itself is not referenced.

nixi does the same since nixi-nixarchy#84, which is why inputs.nixi.inputs.omarchy follows ours.

Measured at 8775661 (2026-09-24): the closure is 84.7 MiB, of which the Smart Match engine and both models are most. Since #1052 that reaches every installed machine, and the offline ISO's reference closure.

Bumping the pin: move url to a newer commit on nixarchy-menu main -- now a release commit, pinned at v1.0.1 -- then nix flake lock --update-input nixarchy-menu. checks.options builds it through the plugin validator, which fails if the manifest id stops being nixarchy.menu.

nixarchy-hyprsettngs = {

Hyprforge is Aziz's Hyprland studio for Omarchy (MIT, id aziz.hyprforge). nixarchy pins it through olafkfreund's fork, nixarchy-hyprsettngs, which carried no commits of its own when it was pinned at 2392ad6, identical to upstream's HEAD.

flake = false, packaged here, as herdr is. Upstream has no flake. modules/home.nix's hyprforgePlugin is a runCommand that copies the tree without test/ and preview.png, and fails unless LICENSE still reads Copyright (c) 2026 Aziz. MIT requires the notice to travel with the code, and checks.options asserts it too (hyprforgePackaged).

The session has to load what it saves. Hyprforge writes ~/.config/hypr/hyprforge.lua and expects ~/.config/hypr/hyprland.lua to require it. nixarchy's session runs the store hyprland.lua instead, so pkgs/omarchy adds Hyprforge's own optional-require line there (pkgs/AGENTS.md#hyprforge-s-hook-lives-in-the-store-hyprland-lua). checks.session writes the file and reads the option back through hyprctl.

Its tools are lua5_4 (what Arch ships as lua, and what baseline.lua was written against; nixpkgs' plain lua is 5.2), diffutils for cmp, libnotify and wl-clipboard, all lowPrio. hyprctl, coreutils and omarchy-shell are already on the session's PATH.

No menu row and no keybinding. Its service writes a .desktop entry, and the Omarchy launcher, which nixarchy-menu uses too, lists desktop entries.

Measured at 2392ad6 (2026-09-29): the plugin is 307.6 KiB. lua5_4's closure is 41.6 MiB, nearly all of it glibc and friends that every machine already carries.

Bumping the pin. This is third-party code that runs unsandboxed in every user's shell, so read the diff before moving it:

1. gh api repos/olafkfreund/nixarchy-hyprsettngs/compare/<pinned>...<new>. If the fork is behind upstream, compare upstream too (AbdulazizAlwabel/omarchy-hyprforge).
2. Check that the hook line is still what Engine.js's hasHook looks for (hypr.hyprforge), and that LICENSE still names its holder.
3. Move url, then nix flake lock --update-input nixarchy-hyprsettngs. The lock diff touches this node alone.

nixarchy-herdr = {

The third default plugin, with one difference: upstream has no flake, so the input is flake = false and there is nothing to follows. nixarchy is the packager. modules/home.nix's herdrSessions is a runCommand that copies the tree without its design documents (intent/, spec/, plan/, tests/, preview.png) and runs patchShebangs on bin/. Upstream's two scripts say #!/bin/bash and are run by path, which works today only through envfs. The build fails unless LICENSE still names both holders, Jankees van Woezik and olafkfreund.

herdr itself comes from nixpkgs (0.9.0 at the pin). herdr update downloads a new binary into the path it runs from, so it cannot update a store copy. Update herdr by bumping nixpkgs, or install your own build ahead of it on PATH, as the maintainer's machine does.

Measured at 6bb0a4c (2026-09-19): the plugin output is 681 KiB, and the source tree on the ISO is 983 KiB. herdr adds 25.0 MiB; the rest of its 71.0 MiB closure, jq and iproute2 are already in the reference system.

Bumping the pin:

1. Pick the commit on master and read the diff (gh api repos/olafkfreund/nixarchy-herdr/compare/<old>...<new>).
2. Edit the rev, then nix flake lock. The lock diff touches nixarchy-herdr alone.
3. The manifest id must still be nixarchy.herdr; the build fails if not.
4. Build checks.menu-verbs: every herdr <level> <sub> the new herdr-sessions sends must be one the pinned herdr lists. Re-run it on a nixpkgs bump too, since herdr can rename a subcommand from that side.
5. Build checks.options and checks.plugin.

nixarchy-distrobox = {

The fifth default plugin, gated like distrobox itself on programs.nixarchy.services.boxes.enable. Its templates come from nixarchy: boxes.nix writes data/box-templates.nix to /etc/nixarchy/box-templates.ini as a distrobox assemble file, and modules/home.nix's distroboxPanel points the manifest's default templatesFile at it with jq. A path the user sets in Setup → Plugins still wins, and nothing writes shell.json. checks.options reads the pinned plugin's own accepted-key lists (ASSEMBLE_BOOLS, ASSEMBLE_SINGLE, ASSEMBLE_CUMULATIVE in Model.js) and fails if a template sets a key the panel refuses. The package ships its LICENSE. The panel can promote since dd9e89c: p on a row copies a programs.nixarchy.services.boxes.machines.<name> snippet, which is what the retired nixarchy box promote printed. That was the last thing #801 wanted from this side, and #801 has since retired the command.

Measured at dd9e89c (2026-09-20): the plugin output is 131 KiB. The source tree on the ISO is 2.9 MiB, mostly its docs site's screenshots.

Bumping the pin:

1. Pick the commit on main and read the diff (gh api repos/olafkfreund/nixarchy-distrobox/compare/<old>...<new>).
2. Edit the rev, then nix flake lock. The lock diff touches nixarchy-distrobox alone.
3. The manifest id must still be nixarchy.distrobox, and .barWidget.defaults.templatesFile must still be the setting's key. checks.options fails if either moved.
4. Build checks.options (the accepted-key read) and checks.menu-verbs.

nixarchy-microvm = {

The fourth default plugin, and the Sandbox group now: its five child rows are gone (each called a verb with no name, #781). The panel lists both kinds of VM in one place: disposable ones from nixarchy vm, and permanent ones from programs.nixarchy.services.microvm.machines. It drives nixarchy vm through argv arrays in Model.js, and reads what the CLI can do from nixarchy vm help (the --detach, console and set-template lines, #762). So checks.menu-verbs checks every verb the pinned Model.js runs against the CLI's own dispatch. Its permanent-VM features call nixarchy.pkg's script by path (Model.js:988), and nixarchy.pkg is a default too.

Upstream's Home Manager module is not imported: its microvm-binds.lua would duplicate the seeded Super+Alt+V. The package output ships its LICENSE.

Measured at 481e6c5 (2026-09-19): the plugin output is 145 KiB, and the source tree on the ISO is 921 KiB. It needs no packages the reference system lacks.

Bumping the pin:

1. Pick the commit on main and read the diff (gh api repos/olafkfreund/nixarchy-microvm/compare/<old>...<new>).
2. Edit the rev, then nix flake lock. The lock diff touches nixarchy-microvm alone, and follows keeps it on our nixpkgs.
3. The manifest id must still be nixarchy.microvm.
4. Build checks.menu-verbs: every nixarchy-vm <verb> the new Model.js runs must be one the CLI accepts. Re-check nixarchy vm help against the three capability regexes in Model.js if either side reworded them.
5. Build checks.options and checks.plugin.

nixarchy-plugin-browser = {

Setup ▸ Plugins ▸ Add Plugin opens this panel instead of upstream's Git-URL prompt. It searches the marketplace catalog (plugins.omarchy.org) and audits a plugin before it is installed. The audit clones it at a pinned commit and scans it inside bubblewrap. It gives two verdicts: a security one, and a NixOS-compatibility one (FHS paths, pacman/yay, writes to /etc, bundled ELFs). An install lands disabled at the audited commit. It is a default because Add Plugin is on every machine, and without it that row installs whatever a URL points at, unchecked (#361).

packages carries the plugin's CLI (omarchy-plugin-audit, omarchy-plugin-browser, nixarchy-plugin-fix) and bubblewrap. The audit fails closed without bwrap, and it runs with a fixed PATH that includes /etc/profiles/per-user/$USER/bin. So the per-user profile is the right place for bubblewrap, and turning the plugin off removes it again. Like every default's tools, both are lowPrio (#809).

Upstream's stock row stays one row down as setup.plugin.add-url ("Add Plugin from URL"), unaudited, for a URL you already have. It is also what is left once the panel is turned off, because setup.plugin.add has a when. The flake's own homeManagerModules is not used: nixarchy seeds Super+Alt+U itself.

A changed menu row needs a re-login. The session keeps the OMARCHY_PATH tree it logged in with, and the menu reads its defaults from there. So after a switch that brings this in, Add Plugin still opens the Git-URL prompt until you log out and back in. omarchy-restart-shell alone is not enough, and neither are the keybindings, which run with Hyprland's login-time environment. Found on razer, 2026-09-23.

Measured at cd3a560 (2026-09-23): the plugin output is 204 KiB, the CLI wrappers 20 KiB, and the source tree on the ISO 544 KiB. bubblewrap 0.12.0 is 112 KiB; the rest of its closure is glibc and libraries the reference system already has. The catalog (about 8 MB, cached for an hour) and one preview per opened plugin are fetched at run time, only when the panel is used.

Bumping the pin:

1. Pick the commit on master and read the diff (gh api repos/olafkfreund/nixarchy-plugin-browser/compare/<old>...<new>).
2. Edit the rev, then nix flake lock. The lock diff touches nixarchy-plugin-browser alone, and follows keeps it on our nixpkgs.
3. The manifest id must still be io.github.olafkfreund.nixarchy-plugin-browser: the menu row, the seeded bind and tests/options.nix name it.
4. Build checks.options and checks.plugin. The plugin's own checks.default runs its scanner tests and the pacman/yay grep; ours runs the grep again on the installed copy.

microvm = {

#221/#222: the foundation of the sandboxes epic. A guest whose /nix/store is a 9p share of the HOST's store -- microvm.storeOnDisk = false, computed by upstream itself the moment microvm.shares contains a source = "/nix/store" entry -- so no image is ever built for a template and the closure a user runs is store paths the host already had.

follows = "nixpkgs", and deliberately not the reflex hyprland.url above declines: microvm.cachix.org (this flake's own nixConfig, inert unless a consumer opts in -- which nothing here does) carries the non-qemu hypervisors, and every guest this repo builds is qemu. What a user actually downloads on first launch is the template CLOSURE, built from THIS lock and served by nixarchy.cachix.org -- following buys nothing to forfeit there. What it buys instead is disko's argument, not hyprland's: a guest and host that share glibc, systemd and openssh store paths rather than each carrying a second copy, because the same nixpkgs built both.

Pinned to a COMMIT. The newest tag, v0.5.0, is from April 2024 and predates options modules/microvm/guest.nix depends on (microvm.shares gained fields since); upstream publishes no newer tags and treats main as its release channel -- the same position sops-nix already takes in this file, for the same reason. Bump it deliberately; never track a branch.

Lock cost, stated rather than hidden: TWO nodes for this one input, not one. microvm.nix declares nixpkgs (which follows ours, so nothing transitive) and spectrum, a flake = false tree it reads only to patch cloud-hypervisor's graphics support -- a hypervisor nothing in this repo will ever build, on a patch nothing here will ever apply. spectrum still lands in every user's flake.lock, including the Mode A machine that never asked for a VM at all. Paid because the alternative -- vendoring a runner ourselves -- is the copy-upstream-and-keep-it-green trade #221 already surveyed and rejected.

nixarchyRev = self.shortRev or self.dirtyShortRev or "unknown";

Which nixarchy built this machine (#208), for nixarchy-version.

self is the only place the answer exists, and it is only reachable here -- pkgs/omarchy/default.nix cannot see the flake it belongs to -- so the two values are computed once and threaded into the package.

Not a tag: a flake cannot know its own, and the ordinary consumer flake.nix tracks main and has none. Not a hand-kept literal beside omarchyVersion either; that is the class of stale string #201 had to fix. A rev cannot go stale.

dirtyShortRev is what a working copy with uncommitted changes gets instead of shortRev -- it reads "800af40-dirty", which is the honest thing for a machine built from someone's checkout to say. Neither exists when the source is not a git tree at all, hence the last fallback.

libva-utils

vainfo, for the Graphics section. Declared because writeShellApplication prepends runtimeInputs to the caller's PATH. Without a declared vainfo, a machine without it on PATH reports "no VAAPI driver answered" rather than "vainfo is missing". It ran while this was being written because the developer's PATH supplied it, which is exactly how that goes unnoticed.

No pciutils: the GPUs come from /sys/bus/pci, the same way the Bluetooth check reads /sys/class/bluetooth, and sysfs hands over the PCI address already in the form the bus IDs need.

path = final.lib.splitString "." (app.attr or name);

nixarchy-apps first, then the top level. Apps this repo packages live under nixarchy-apps, so probing only the top level returned null for every one of them and the fallback answered with the attribute name. That is right by luck when the attribute matches the binary -- once, omacut, ttfx -- and wrong when it does not: zen for a package installing bin/zen-beta, hey-cli for one installing bin/hey. Both were reported as absent on machines that had them.

podman

podman info / podman inspect, for the Boxes section. Safe to pin here -- unlike distrobox, podman does not resolve anything relative to argv[0], so a /nix/store path to it carries none of pkgs/box.nix's hazard. distrobox itself is deliberately NOT here: it is looked up by bare name (command -v distrobox), the same wrapper-only rule that command's own header explains. hyprctl is deliberately not here either -- the client has to match the RUNNING compositor, which the session's PATH carries and this list cannot; the Graphics section's comment in verify.sh makes the argument.

zen-browser = zen-browser.packages.${final.stdenv.hostPlatform.system}.default;

Not built here -- upstream's own flake, re-exported into the overlay so nixarchy-doctor can find it.

The doctor asks pkgs for an app's meta.mainProgram to learn which command it puts on PATH, and falls back to the attribute name when the lookup fails. zen-browser lived only in packages., so the lookup returned null and the fallback answered "zen" -- a command that does not exist, because this package installs bin/zen-beta. The doctor could never report Zen as present.

Exactly the vscode/code trap the doctor was written to avoid, reached by a different road: not a wrong mainProgram, but a package the probe could not see. The Install menu's copy of the probe had the same bug for longer -- it reads cfg.apps..package instead of this overlay; see appBinary in modules/apps.nix.

retroarch = final.retroarch.withCores (

nixpkgs' retroarch is retroarch-with-cores built with an EMPTY core list, so installing it gives an emulator that can run nothing. retroarch-full is the obvious fix and the wrong one: it pulls unfree cores, and an unfree package in the app list aborts the whole rebuild rather than failing on its own.

So: every core Omarchy's own picker offers that nixpkgs ships under a free licence, minus the two that dwarf the rest. snes9x and genesis-plus-gx are unfree, hence bsnes and blastem for those systems.

hypr-rdp = inputs.hypr-rdp.packages.${final.stdenv.hostPlatform.system}.hypr-rdp;

The RDP daemon, re-exported from its own flake so the module that will run it (#156) can say pkgs.hypr-rdp and never mention an input. Built by upstream's pkg/nix/package.nix against OUR nixpkgs, through the follows on the input.

This indirection is what makes the input's exit cheap: when nixpkgs carries hypr-rdp, delete the input and this attribute, and every pkgs.hypr-rdp in the tree keeps resolving -- to nixpkgs' own.

Lazy, so a machine that never enables RDP never builds it. Being in the overlay is not being on the system.

It carries one patch, and that is the wrong way round (#1031). Hyprland 0.56 dropped the legacy keyword IPC request and answers unknown request. hypr-rdp already falls back to eval hl.monitor({...}) for exactly this, but its guard, is_non_legacy_parser_error(), matched only the older refusal (non-legacy parsers), so the daemon created its headless output, failed to set its mode, and exited. pkgs/hypr-rdp/lua-fallback-on-unknown-request.patch widens that match and adds a test that pins the new message.

The override also sets doCheck = true with checkFlags = [ "hyprland::tests" ]. Upstream's package sets doCheck = false, so without this the patch's test would compile into nothing and a bump that dropped the fix would stay green. The filter keeps the run to the module the patch touches (15 tests, about a second), and leaves unrun the rest of the suite, which upstream chose not to run.

Retire it when a hypr-rdp tag carries the wider matcher. Nothing has to be remembered: on that bump the patch stops applying and the build fails. Delete the file and the overrideAttrs together. When the input itself goes (see above), the patch goes with it.

quickshell = final.quickshell.overrideAttrs (_: rec {

nixpkgs is on 0.3.0, whose session lock aborts on DPMS.

WlSessionLock::updateSurfaces reached qFatal -- "Tried to show lockscreen surfaces without active lock" -- when screens slept and woke while locked. quickshell dies, and because the Wayland session-lock protocol deliberately keeps the compositor locked when its lock client disappears, the machine is left blank with nothing to type a password into. Reboot is the only way out. Seen three times in one night, ~65-70 MB of coredump each.

v0.3.1 fixes it, and says so in as many words: "Fixed session lock crashes on sleep, wake, DPMS, and unlocking." Three commits touch wayland/lock; 897fcdaa is this one -- it null-checks the wayland output and skips placeholder screens instead of asserting.

overrideAttrs rather than a fork: nixpkgs keeps the build, the Qt wrapper and the dependency set, and only the tag moves. Passed here rather than set on the overlay, so a machine using quickshell for something of its own keeps nixpkgs'.

DELETE THIS once nixpkgs ships >= 0.3.1.

nixarchy-plymouth = final.runCommand "nixarchy-plymouth-theme" { } ''

The boot splash, and nothing else on the disk with it.

The omarchy package already carries share/plymouth/themes/omarchy, and on a desktop that is the right place for it: the machine has the package anyway. The live image does not. It deliberately does not import nixosModules.nixarchy (see installer/cd.nix), so naming the desktop package in boot.plymouth.themePackages would be the only reason it were there -- and nixpkgs' plymouth module puts themePackages into environment.etc, so the named package joins the system closure whole: 411.9 MiB on an image budgeted at 2.00 GiB, which is GitHub's limit on a release asset rather than a preference. These assets are 180 KiB.

Copied out of the same package rather than assembled again, so the theme is built in exactly one place and the image's splash cannot differ from the installed one. Nothing is rewritten here: omarchy.plymouth already points at /etc/plymouth/themes/omarchy, so this output has no store references and costs only its own size.

install = pkgsFor.${system}.writeShellApplication {

nix run github:olafkfreund/nixarchy#verify, from inside a running Omarchy session. Everything in checks/ runs in a machine with no GPU, no Bluetooth radio, no network and no sound; this asks the questions that leaves unanswered. nix run github:olafkfreund/nixarchy#install, as root, from a NixOS live ISO. Formats ONE disk with installer/disk-config.nix, writes the generated flake to /mnt/etc/nixos and runs nixos-install against it.

Every answer ends up as text in that flake: the machine it produces belongs to the flake, not to this script.

release-notes = pkgsFor.${system}.writeShellApplication {

nix run .#release-notes <from-tag> <to-ref> -- what changed for someone running nixarchy, between two releases. release.yml puts its output above the download instructions on the release page.

An app for the same reason review is one: it evaluates the option set at two revisions and asks GitHub what upstream shipped, and a check has neither a network nor a second checkout. The half that needs neither is checks.release-notes, which drives this whole script against a fixture repository.

It carries omarchy-package-delta.sh rather than reimplementing it: the question "what did this Omarchy release add and drop" already has an answer here, and two of them would disagree eventually.

devenv-presets = pkgsFor.${system}.writeShellApplication {

nix run .#devenv-presets scaffolds the eight templates this repo used to carry -- go, jupyter, ml, node, python, react, rust, typescript -- with a real devenv, and asks devenv to evaluate what it wrote. Since #802 the catalogue and the scaffolder live in nixarchy-devenv, so this runs the plugin's own templates-check app over those ids rather than a copy of the proof.

The name is fixed. devenv-presets is a required status check on main, and build.yml's job of that name runs this attribute. Renaming either would be a workflow change and a branch-protection change, which are a human's (AGENTS.md 4 and 11) -- so the attribute keeps the name and changes what it runs. That is also why the check's scope stayed at the eight: the plugin's other templates are its own CI's business, and its cloud generator needs the network at create time.

This is still the whole safety net under a catalogue of strings. lines is a string, so a template naming an option devenv renamed is a valid Nix file and a broken project, and nothing in nix flake check would ever say so.

NOT in checks, and that is not an oversight. #150 proposed it as one on the reasoning that a full devenv shell needs the network but evaluation does not. That is wrong, twice over: evaluating a devenv project fetches the inputs devenv.yaml names (github:cachix/devenv-nixpkgs/rolling), and devenv's own flake reaches them through import-from-derivation -- nix flake show github:cachix/devenv fails with allow-import-from-derivation is disabled before it prints anything. A sandboxed derivation has neither, so checks.devenv-presets could not run at all. A workflow job that has a network does, and build.yml has one.

What the old runner got wrong. It moved HOME into a temporary directory but inherited XDG_DATA_HOME, and devenv keeps its trust database at $XDG_DATA_HOME/devenv/allowed. Every run therefore wrote its throwaway scaffolds into the trust list of whoever ran it: 80 dead entries on p620, all vmtest/tmp/tmp.*/<preset>. The plugin's check puts HOME, every XDG_* path and DEVENV_HOME inside one temporary root, and compares the caller's allow list checksum before and after, failing if it moved.

inherit

Screencasts of a real session, scene by scene -- see tests/demo/default.nix. Not checks: each boots a desktop, drives it and encodes a GIF, which is minutes of work nobody wants on every push.

nix build .#demo the full tour (mp4 + gif) nix run .#demo-record -- re-record ONE scene's GIF and drop it in docs/img/features/ nix run .#demo-verify -- f.gif audit any GIF: sampled frames must change, and must OCR to what the scene claims to show

update =

nix run .#update -- rewrites the pinned versions and hashes in place. CI runs it nightly and opens a PR (update.yml).

Derived, not enumerated: every nixarchy-apps package carrying a passthru.updateScript is run, so a new hand-pinned package is covered the day it lands, without an edit here. This line used to name once alone -- beside a comment claiming once was the only hand pin -- while hey-cli sat pinned with an updateScript nothing reached, and needed a human (#195, 2026-09-05).

nix eval .#update.pinned lists the covered attributes; update.yml builds exactly those after a rewrite, off the same set. A package pinned by hand WITHOUT an updateScript is still invisible to this. The one there was, grok-bot, was removed in #1201.

installer-vm =

A VM that installs onto a blank disk, WITH a network -- which checks.install cannot have, because it is a sandboxed derivation and its whole point is that a rebuild afterwards cannot fetch anything. See installer/vm.nix.

Wrapped rather than exported raw: the generated run script creates two sparse qcow2 images and only discovers they do not fit hours later, from inside the guest. installer/vm-preflight.sh checks first and says so. The sizes come from the configuration itself so the check cannot drift away from what the VM actually asks for.

try =

The front door: nix run github:olafkfreund/nixarchy#try boots the installer ISO in a local UEFI VM, wizard and all, for someone who knows nothing about this repository. -- --net takes the network image instead.

Deliberately NOT depending on the ISO derivation: referencing it here would make nix run .#try download -- or worse, silently build -- 5.6 GB before the script's first line ran. The script resolves the image itself at runtime, dry-run first, so "this is a download of X" or "STOP, this would be a source build" is said BEFORE either happens. See installer/try.sh for the rest of the refusals; checks.try-preflight exercises each of them.

Also not a wrapper around .#installer-vm: that VM answers the wizard itself from a baked answers file -- a test harness, and exactly the thing a person trying nixarchy must not get.

// lib.concatMapAttrs (

Two runners per data/microvm-templates.nix entry, both built by lib.mkMicrovm above from the same template and the same modules/microvm/guest.nix -- only microvm.cpu differs between them.

microvm-<t> passes nothing extra: cpu = null is upstream's own default, which on x86_64-linux means -enable-kvm -cpu host,+x2apic,-sgx (lib/runners/qemu.nix in the pinned commit -- -enable-kvm is added exactly when cpu == null). That is what nearly every user runs, and #221 says outright it is not provable in CI without nested virtualisation.

microvm-<t>-tcg sets microvm.cpu = "max". The same cpu != null branch that changes the -cpu argument ALSO removes -enable-kvm from the same conditional -- one option, both effects -- which is what makes this the runner a host with no /dev/kvm (including nixarchy running inside a VM) can boot, and the one CI builds. No separate "disable kvm" option exists to set; there is only this one.

machineOpts on the -tcg variant, and every entry is load-bearing:

• pit = "on", pic = "on". Upstream's default microvm machine sets both off, which is correct under KVM -- kvmclock is the guest's clock -- and fatal under TCG, which has no kvmclock: with no PIT, no PIC and no HPET on this machine type, early timer init has nothing to calibrate from, and the kernel triple-faults or wedges right after "Poking KASLR", outcome depending on where KASLR happened to land. Measured on a real A/B: the shipped runner under forced TCG produced zero output for five minutes; the same runner with pit/pic on booted to a login prompt in 2m13s. -cpu qemu64 wedged identically, so the CPU model was ruled out.

• accel = "tcg", PINNED -- not upstream's kvm:tcg fallback chain. The fallback is how the wedge above shipped unnoticed: qemu silently takes KVM wherever /dev/kvm exists (including inside a test VM on a host with nested virtualisation), so every local run of the "-tcg" artifact was measuring a KVM boot, and the one environment that exercised real TCG was CI -- at the moment something first tried to boot it. This variant's entire reason to exist is hosts with no /dev/kvm; a KVM host should run the normal runner. Do not "optimise" the fallback back in: an artifact that never runs as what its name promises is how this bug lived long enough to gate a PR.

• the rest reproduce upstream's x86 defaults (machineOpts REPLACES the whole set, so omitting one is turning it off). pcie = "on" because the 9p shares sit on virtio-pci.

vm-big = nixpkgs.lib.nixosSystem {

The same VM with room to run a model.

.#vm is a smoke test and is sized like one: 8GB, and an ephemeral root, which is right for catching a broken desktop and wrong for anything to do with programs.nixarchy.localAi. Two reasons, and the second is the one that is not obvious:

• 8GB does not hold a useful model. qwen3:8b is 5.2GB of weights before any cache.

• diskImage = null puts the root on a tmpfs, so the weights live in RAM. They are then paid for twice, once to store and once to load, out of the same pool -- and the service is OOM-killed while every visible number says it had room. Observed, not theorised.

So this one has a real disk, which also means a pulled model survives a restart rather than being re-downloaded every time.

nix run .#vm-big ssh -p 2224 omarchy@localhost (password: omarchy)

It writes nixarchy-vm-big.qcow2 into the working directory and REUSES it, which is the point here and is exactly what .#vm avoids. Delete that file to start clean.

This is the VM you keep (#817). Not only the model VM: it is the one to develop and test features in, because it is the only one whose state survives. So it also turns on podman and boxes, which the smoke test does not -- nixarchy.podman and nixarchy.distrobox are gated on those services, and with them off the two panels most worth trying by hand are invisible. It now resolves seven default plugins where .#vm resolves five.

Keep the qcow2 under /mnt/data/vmtest/, never /tmp: that is a 32GB tmpfs and a VM disk there competes with the machine's RAM and loses quietly.

Rebuild it monthly:

rm nixarchy-vm-big.qcow2 && nix run .#vm-big

Monthly rather than "when it misbehaves", because the failure a stale disk causes is one you do not notice. That is not hypothetical, and it is why the other VM has no disk at all: Omarchy persists every notification under ~/.local/state/omarchy/notifications/history/ and its shell replays that directory on start, so failures already fixed in the package kept reappearing on screen from an old disk. .#vm's diskImage = null is load-bearing (vm/configuration.nix), and the obligation it hands over is stated there too -- set a path and you own clearing it. This is that obligation, written down and given a date.

Boxes pulls images, so a vm-big run offline has a Distrobox panel that lists templates and cannot fetch one. Honest behaviour, not a fault.

{

One module that imports the machine's own two, rather than two entries in this list. It has to be shaped this way, because the generated flake's hosts//default.nix is shaped this way: imports inside a module and entries in modules are merged in different orders, so the flat form gives a different environment.systemPackages ORDER -- same packages, same closure, different list -- and system-path hashes that order into chosenOutputs. A different system-path is a different toplevel, and on an ISO with no network the installed machine can no longer copy the one baked here: it has to build it, which means stdenv, which means the source bootstrap from hex0-seed.

Nothing about the packages differs. Only the order does, and only the order has to.

lib.mkMicrovm =

One closure per template, shared by every VM a user ever names from it. #221's "the name is never a Nix argument" is exactly what makes this possible: template and modules are the only things this function is allowed to vary on, and neither is a per-VM value.

template is a path from data/microvm-templates.nix -- plain NixOS plus microvm.nix's own options, per that catalogue's bar. modules/microvm/guest.nix is what every template gets underneath it, imported here rather than by each template file so a template cannot forget it. modules exists for exactly one caller below: the -tcg packages, which add microvm.cpu = "max" without a second template.

Returns declaredRunner, upstream's own name for config.microvm.runner.${config.microvm.hypervisor} -- a package, which is what packages.<system>.microvm-<template> has to be.

iso-budget =

Both images, against the budgets recorded in installer/cd.nix.

A check rather than a comment because a number nobody enforces is a number that drifts. The nightly runs this; it costs nothing beyond the ISO builds it already does.

The iso-net budget is the load-bearing one, and it is not a taste question: GitHub refuses a release asset over 2 GiB, so an iso-net that crosses it stops being publishable as a single file and release.yml would have to start splitting it. Failing here says so while there is still something to do about it, rather than at the next tag.