Installed machines and the downstream template now resolve nomarchy from
git+https://git.bemagri.xyz/bernardo/nomarchy.git?ref=v1 (the lock's
original node carries the ref too), so `nix flake update` follows the
release branch instead of the forge's default branch. README quickstart
points at the real forge.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Promotes the throwaway harness that verified the installer into tools/:
- tools/vm/qmp.py QMP keystroke injection / typing / quit
- tools/vm/vncshot.py GL-safe screenshots via VNC readback (QMP
screendump shows "no surface" with virtio-vga-gl)
- tools/vm/gap-analysis.py drv-graph diff that converged the offline
pin set; run it when offline installs build
from source
- tools/test-install.sh the full offline regression: build ISO, boot
offline, unattended LUKS+swap install via a
config disk (typed long commands drop keys),
wait for poweroff, boot the installed disk,
screenshot the first boot for visual verdict
docs/TESTING.md §4 now points at the script.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The pre-activation log (now persisted) showed activate dying at its
profile operation: as the target user with no daemon in the chroot,
nix can't open the store db ("big-lock: Permission denied"). Run a
temporary nix-daemon inside the chroot and point activate at it via
NIX_REMOTE=daemon. HOME_MANAGER_BACKUP_EXT=bak makes file collisions
(e.g. a stray autogenerated hyprland.conf) move aside instead of
aborting. Verified in the chroot on a real installed disk: HM symlinks
land, collision backed up.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The HM pre-activation never ran: the script was written to /mnt/tmp and
nixos-enter mounts a fresh tmpfs over /tmp inside the chroot, so `bash
/tmp/...` found nothing and the best-effort guard swallowed it. Verified
interactively that every inner step (path: build as root, runuser leg)
works in the chroot — only the staging path was broken. Now staged in
/root, with set -x logging persisted to /var/log/nomarchy-hm-preactivate.log
on the target and the tail echoed into the install transcript on failure.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The chown fix made the flake user-owned, and root's libgit2 then refuses
git+file fetches of it ("not owned by current user") — which silently
killed the seeding step under set -e and would equally break
nixos-install --flake. Use path: refs for the live-side archive and the
nixos-install invocation (path flakes read flake.lock, skip libgit2),
and make seeding failure non-fatal as intended.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
End-to-end VM verification found the installed system perfect (LUKS,
auto-login, swapfile+resume, snapper) except: ~/.nomarchy was root-owned,
so libgit2 refused every `home-manager switch` ("repository not owned by
current user"). chown 1000:users after generation (the account doesn't
exist in the target yet — numeric ids).
The chroot pre-activation now builds via path: instead of git+file —
no libgit2, no ownership sensitivity — and the fallback message suggests
-b bak (a default-config Hyprland session leaves an autogenerated
hyprland.conf in the way otherwise; observed live).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Substituter plumbing through nixos-install proved unreliable (env config
is overridden by its hardcoded flags; the forwarded --substituters flag
still left the plan building gcc from source). Make every ISO path valid
in the target store with one nix copy --all before nixos-install — the
in-target build then resolves everything locally and only the per-machine
config drvs are built.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
nixos-install hardcodes --extra-substituters "auto?trusted=1" as a CLI
flag, and flags override NIX_CONFIG for the same setting — so the env
fix never reached the in-target build. nixos-install forwards a
--substituters flag verbatim; use that.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
nixos-install builds with --store /mnt and passes its own substituter as
"auto?trusted=1", which in that context resolves to the target store
itself — so nothing flowed from the ISO and nix fell back to a
full-source bootstrap (gcc/python/bash tarball fetches → offline death).
Host-store simulation (fresh local?root, substituters cleared) proves
"daemon?trusted=1" makes the whole install substitute + build offline:
1125 paths copied, only the per-machine config drvs built locally.
Also drop --offline from the chroot HM build (that flag disables ALL
substituters; plain empty substituters does what we want without it).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
`nix flake archive --to` rejects locally-evaluated source paths ("lacks
a signature by a trusted key"); enumerate the archive tree as JSON and
copy with --no-check-sigs instead.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Round 7 in-VM: autodetection emitted common-cpu-intel, whose nixos-hardware
module pulls the whole Intel graphics stack (intel-media-driver,
compute-runtime, libva, level-zero…) — none of it on the ISO, so the
offline install built graphics drivers from source. The representative
pin now imports the common-* profiles (intel+amd cpu/gpu, laptop, ssd,
pc), covering what the detector emits on ordinary machines.
Model-specific profiles may still need network.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Walking the install-shaped toplevel's drv graph against the ISO's actual
path set (tools/gap analysis) found why offline installs kept building
from source:
- system-path is rebuilt on EVERY machine: buildEnv's package-list order
follows module-graph order, which any hardware profile reshuffles →
pin its builders (texinfo, getconf)
- greetd.toml embeds the auto-login user and is generated by remarshal;
building remarshal offline drags its test closure (matplotlib → ffmpeg
→ x265…, ~1900 drvs) → pin pkgs.remarshal
- second-level tools for the always-rebuilt drvs: kmod.dev +
nukeReferences (modules-shrunk), xorg.lndir (unit dirs), libxslt.bin
(dbus config), python3Minimal (etc)
- the representative pin now has a vfat /boot like a real install
(mtools/dosfstools enter systemPackages via the vfat fs-type gate)
Verified on the host: the remaining offline build set for a foreign
username/hostname/UUID config is 66 trivial config drvs + eval-time file
copies from the pinned nixpkgs tree — zero network. Same check for the
per-username HM activation: 15 drvs, offline-clean.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The full system build closure adds ~134 GB to the store and would have
produced an unusable multi-ten-GB ISO. Back to targeted inputDerivation
pins; the still-missing second-level build tools get pinned explicitly
next, from a clean committed-tree gap analysis.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Round 6 still hit a from-source cascade: every per-machine drv (units,
initrd pieces, dbus config, modules-shrunk) needs its own build tools
(lndir, builder.pl, kmod-dev, nuke-refs, libxslt-bin, …), one level
deeper than targeted inputDerivation pins reach. Use the documented
mechanism instead: system.includeBuildDependencies = true on the
representative install pin. Bigger ISO, networkless installs.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Offline nixos-install still built from source: the drvs that embed the
user's hostname/username/UUIDs (etc, initrd, toplevel, HM activation)
must be REBUILT, and their builders need tools outside every runtime
closure (etc → python3-minimal, initrd → packing tools). Pin each drv's
inputDerivation into the ISO — the standard nixos-test offline trick.
Also: with no network every substituter query is a DNS retry storm that
ends in a nix goal.cc assertion crash (core dump observed); the installer
now probes connectivity and disables substituters when offline, both on
the live system and inside the nixos-enter HM pre-activation.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The offline nixos-install failed building from source: the generated
config enables snapper/auto-login/LUKS-initrd/swapfile-resume on a BTRFS
root, none of which the bare template (ext4 placeholder!) pulls in — so
snapper, nixos-rebuild-snap, btrfs-progs-lib and cryptsetup-bin were
absent from the ISO. The pinned system now mirrors the installer's output
(closure-diff verified).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Lock follows paths are absolute from the root node; nomarchy's nodes move
one level down in the composed downstream lock, so ["nixpkgs"] must
become ["nomarchy","nixpkgs"] — otherwise every follows dangles
("input 'nomarchy/home-manager/nixpkgs' follows a non-existent input").
Verified: the composed lock now resolves the full input tree with
`nix flake metadata --offline`.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
disko builds its format/mount script at install time; offline, the
script's tool closure (file, which, wrapper hooks) wasn't on the ISO and
nix tried to build it from source — which needs the network. Pre-building
one representative script (luks + swap variant) pins every input; the
user's actual invocation then differs only in the final text-generation
derivation, which builds offline in seconds.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Found by running the unattended install in an offline QEMU VM:
- compgen is missing from nixpkgs' non-interactive bash (the package
shebang) — replaced with plain glob tests in hardware-db and prewipe
- disko's eval resolves <nixpkgs> via NIX_PATH (a dead channel on the
ISO) and tried channels.nixos.org — export NIX_PATH to the pinned
nixpkgs source and point the flake registry at an empty baked file
- lock nomarchy by path (the source the ISO carries) instead of
git+https: git inputs clone even when narHash is known, which kills
offline installs; `original` keeps the forge URL so a later
`nix flake update` re-resolves normally. Works from dirty-tree ISOs too.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- LUKS2 is the installer default; in exchange the generated config sets
the new nomarchy.system.greeter.autoLogin (greetd initial_session) —
the disk passphrase already gates the machine.
- @snapshots subvolume + nomarchy.system.snapper.enable: hourly/daily
timeline snapshots of / and the nixos-rebuild-snap helper, ported from
the previous iteration (3bdfc35), guarded to no-op on non-BTRFS roots.
- @swap subvolume with a swapfile sized to RAM by default (disko
mkswapfile handles NOCOW); the installer computes the resume offset
and wires boot.resumeDevice + resume_offset for hibernation.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Ported from the previous iteration's installer (3bdfc35), adapted to the
mkFlake world. gum TUI: disk pick, optional LUKS2, user/hostname/timezone,
DMI → nixos-hardware autodetection (hardware-db.sh). disko partitions
GPT + 1 GiB ESP + BTRFS subvolumes; the generated machine flake lands at
~user/.nomarchy (one mkFlake call, /etc/nixos symlinks to it) and
nixos-install makes it bootable (UEFI/systemd-boot, v1 single disk).
Offline by construction: the target flake.lock is composed from the rev
the ISO was built from (compose-lock.py), `nix flake archive` seeds the
target store, and the ISO pre-builds the template system + desktop.
First boot lands themed: the HM generation is pre-activated in chroot.
mkFlake grows two things for this: hardwareProfile now also takes a list
(autodetection emits several common-* modules), and installed systems get
the standalone home-manager CLI from the pinned input. Unattended mode
(NOMARCHY_UNATTENDED=1 + env) documented in docs/TESTING.md §4.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Downstream users now own only system.nix and home.nix; their flake.nix is
generated once from the template and never hand-edited. mkFlake (lib.nix)
wires both layers and maps hardwareProfile = "<name>" to nixos-hardware
modules (new input, pinned here, nixpkgs follows ours; unknown names fail
with did-you-mean suggestions). Flake checks evaluate the template through
mkFlake — including a real profile — so drift fails `nix flake check`.
Also from the live-VM testing session: install the home-manager CLI in the
live ISO and pin flake inputs transitively (offline theme switching needs
stylix's own inputs too), plus swww→awww doc updates.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Found by booting the live ISO (docs/TESTING.md checklist):
- pin HM configType = "hyprlang" (26.05 stateVersion defaults to the new
Lua config and renders these settings as broken hl.$mod(...) calls,
booting Hyprland into emergency mode)
- drop dwindle:pseudotile (removed in Hyprland 0.55; aborts config parse)
- disable splash rendering alongside the logo
- launch awww-daemon and call awww with swww fallback (nixpkgs' swww is
the renamed awww fork; no `swww` binary exists anymore)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Full replacement of the previous iteration, rebuilt around three ideas:
- Pure evaluation: theme-state.json lives inside the flake and is read
via the nomarchy.stateFile option — no --impure, ever.
- All-Home-Manager theming: `nomarchy-theme-sync apply` writes the JSON
and runs `home-manager switch`; every theme change is one atomic,
rollbackable generation. Wallpaper (swww) is the sole runtime piece.
- Flat, downstream-first layout: modules/{nixos,home} with one
options.nix each, exported as nixosModules/homeModules + overlay +
flake template; system (nixos-rebuild) and desktop (home-manager
switch) rebuild paths are fully split.
Ships 21 themes imported from the previous iteration (palettes,
wallpapers, btop themes, six whole-swap Waybar identities), Stylix for
the GTK/Qt/cursor long tail, a live ISO target with offline theme
switching, and docs/TESTING.md with the QEMU verification workflow.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
