useNautilus = false assigns FileChooser to gtk and drops the nautilus D-Bus
registration; the GTK3 dialog then follows the same settings.ini as thunar.
The deployed niri-portals.conf is gone: /etc/xdg-desktop-portal/ ranks below
/etc/xdg/, where programs.niri writes the same file. Its one extra key,
Inhibit=none, moves to desktop.nix.
Bitwarden Inc.'s own build, addressed by GUID because the store carries
third-party forks under the same name. normal_installed like the others, so it
stays removable.
languagePacks installs the pack and the policy reports it as active, but the
interface stayed English: installing is not switching, the same split the theme
already had. Firefox does not fall back to LANG either, although de_DE.UTF-8
reaches the process. RequestedLocales is what switches to it.
Three settings from the same first run: address autofill, payment autofill, and
the three Firefox Suggest switches. The autofill pair are preferences. Suggest
is a policy instead, because its preferences already ship as false and the
profile sets no user value while the switches were on — something enables them
at runtime, which a preference set as a default does not survive. ImproveSuggest
is named rather than left to follow from the other two, so the data collection
cannot come back if Suggest is ever turned on again.
Host-local for the same reason as the node toolchain. python3 belongs next to
uv: its default python-preference reaches for a system interpreter before it
downloads one, so the common case needs no download at all. A project pinned
to an older version still gets one fetched, and those FHS binaries find their
loader through the nix-ld that modules/zed.nix turns on.
Two changes that cannot be separated in the result: the Firefox
preconfiguration is what made desktop.nix untenable, and the split is what it
now lives in.
Firefox replaces LibreWolf, with Betterfox as its preference base and the
settings that Betterfox does not cover expressed as policies: uBlock Origin,
the Catppuccin theme, no Pocket, no Sync, autoscroll, DuckDuckGo, no
translation offer, no password prompts, no bookmarks toolbar, about:blank as
home and new tab. Both add-ons are normal_installed and the preferences keep
status "default", so none of it is locked away — undoing settings by hand was
the reason to leave LibreWolf.
desktop.nix carried thirteen topics over 637 lines, and the existing split
followed the NixOS option namespace rather than the topic: services.pipewire
here, its configPackages there; services.gvfs and services.udisks2 apart from
their packages; Bluetooth likewise. Nine modules now each hold one topic with
its options, packages, configs and services together. desktop.nix keeps the
compositor, the /etc deployment, portals and what belongs to no topic.
The refactor changes nothing about the built system, and that was checked
after each of the eight steps rather than assumed. The plain store path does
not serve as the criterion: moving a package to another module reorders
environment.systemPackages, which changes the system-path hash and propagates
into every file naming it. What was compared instead is the package set
through diff-closures and every differing file with store hashes normalised
away. Two steps reordered the user tmpfiles list, which is neutral only
because no two rules name the same path — checked, not presumed.
Four host-local things that the first install turned up.
The LUKS prompt and the boot menu were both unreadable on the 3840x2400
panel, for different reasons: nixos-hardware puts amdgpu into the initrd, so
the panel is already at its native mode when cryptsetup asks, and systemd-boot
draws through UEFI text output and inherits whatever the firmware left behind.
Keeping the driver out of the initrd fixes the first, consoleMode = "0" the
second.
The Hybrid USB-C Dock drives its monitors over DisplayLink, not DisplayPort alt
mode, so the panel's own connectors never report a screen. Naming the driver
pulls in evdi, the service and the udev rules.
nodejs and pnpm are host-local because no other machine here builds JavaScript.
The camera scripts walk USB descriptors with standard tools and reach each
other by name, but pkexec replaces the environment with a fixed FHS PATH and
the boot unit brings its own, so neither the tools nor the sibling scripts
were found through an inherited PATH. /run/wrappers/bin leads the list because
the system environment carries polkit's own unprivileged pkexec as well, which
aborts unless it is the setuid wrapper.
fprintd put pam_fprintd ahead of pam_unix in every PAM stack, gtklock and
sudo included, for a reader nobody enrolled. The Arch install never had
fprintd either.
The host still carried snapper, which the desktop dropped in d3cc03c:
disk.nix creates no .snapshots subvolumes, so the timer would fail every
hour without ever taking a snapshot. The account sat in the docker group
while docker itself was off, and stateVersion named 26.05 although the
machine is first installed on 26.11.
Also drops the stale snapper mentions elsewhere and adds git-firefly to
Zed's auto-installed extensions.
nixpkgs patches Kvantum to look for themes along XDG_DATA_DIRS
(tsujan/Kvantum 01989083), so a theme package on the system path is
found without anything being copied into the home directory. Linking
/share/Kvantum is what makes it reachable there; the system path carries
only the /share subdirectories it is told to.
The rule this replaces never did what it promised. 'C' recreates the
destination with the mode of its source, and a store directory is
r-xr-xr-x, so copying into it fails for anyone but root and leaves an
empty directory behind — silently, with exit 0 (systemd 261). Only root
was unaffected, which is why a user unit was the one place it broke.
Qt5 clients keep the default style: the patch takes in the Qt6 build
only (NixOS/nixpkgs#355277). VLC is the only Qt5 client here.
waypaper reads only the copy of its config in the home directory, and
that copy is seeded with tmpfiles 'C', which never overwrites. The store
path substituted into it therefore froze at the moonarch revision it was
generated against, and the next garbage collect took that source away —
leaving the session with a black background and no error anywhere.
/usr/share/backgrounds/moonarch.jpg is relinked on every activation, so
the seeded config keeps naming a file that exists, whatever moonarch
revision is current.
gtklock and greetd keep their store paths: their configs are regenerated
on every activation and are held by the system profile, so neither can
go stale the way the seeded copy did.
Without configurationLimit every generation keeps its entry on the ESP.
The desktop had grown to 58 of them, 232M of a 1G partition, and nothing
ever removed one. Ten keeps a useful rollback range and lets the builder
clear the rest on every switch.
moonarch replaced its direct gtklock calls with the moonarch-lock
wrapper. stasis.rune and PowerPopout.qml come from moonarchSrc, so the
new command arrived here at once while the script did not -- the
moonarch-scripts whitelist did not carry it.
Wrapped rather than left bare like the other seven: it has two callers
and must not depend on either one's PATH. A missing gtklock means no
lock screen at all, a missing niri or jq only the wrong monitor.
Picks up the gtklock switch and the GTK theme fix, so the tree is
self-contained again: `nixos-rebuild switch` no longer needs
`--override-input moonarch`, which 0cbd829 had left it needing.
Also brings in app2unit from 20ee46f, which landed on moonarch's main while
this work was local. NixOS already carries app2unit in the system
environment, so nothing here changes because of it.
Verified: both hosts build to exactly the store paths the --override-input
build produced, so the pin is the revision that was tested.
moonarch 6ea02b5 moves to the same Catppuccin Mocha Lavender theme, so the
rewrite no longer bridges Colloid to Catppuccin. It now bridges two names
for the identical theme: Arch installs the upstream 1.0.3 release zips,
whose directories carry the full build name with the +default tweak marker,
while nixpkgs builds the same source with no tweaks and drops the suffix.
Neither side can take the other's name without going out of its way, so
this stays.
Comment rewritten to say that, and to keep the reason it matters: GTK3
reads settings.ini directly — the dconf entry only reaches applications
that ask GSettings, which on Wayland the GTK3 ones do not — so a name that
does not resolve puts every GTK3 application on Adwaita with no error
anywhere. That is exactly the state moonarch was in until 6ea02b5, where
the named theme was installed by no package at all.
Verified with --override-input: both hosts build, and the deployed
settings.ini carries the nixpkgs name while moonarch's source carries the
Arch one.
moonarch 4f36035 names gtklock itself now, in stasis.rune and in
PowerPopout.qml, so the two substitute derivations that were rewriting them
at deploy time have nothing left to do. stasisConfig, quickshellConfig,
lockCommand and lockArgv go; both consumers read moonarch's tree directly
again.
gtklockConfig stays generated rather than copied. moonarch ships
defaults/xdg/gtklock/config.ini, but it names /usr/share/moonarch for the
background and resolves the modules by bare name against /usr/lib/gtklock.
Both are store paths here, so the file is built rather than deployed — the
same reason PipeWire's config takes its own route.
NOT SELF-CONTAINED: flake.lock still pins moonarch from before 4f36035, and
that revision still says swaylock. Until the lock is bumped, a plain
`nixos-rebuild switch` deploys configs naming a locker that is no longer
installed, and stasis fails silently on a missing binary. Build with
`--override-input moonarch ~/Projects/moonarch` until then, or bump first.
Committed at Dominik's request ahead of the push that makes the bump
possible.
Verified with --override-input: both hosts build, the seeded stasis.rune
and the deployed PowerPopout.qml come straight from moonarch and name
gtklock, and the generated config.ini carries the blurred wallpaper and
both module store paths.
Looking at whether gtklock was worth it turned up two things.
The GTK layer was never on Mocha. colloid-gtk-theme with the catppuccin
tweak produces Frappé with a grey accent — base #292c3c, selection #ccd0da
— and #1e1e2e, #b4befe and #cba6f7 appear nowhere in its gtk.css. Kvantum,
the swaylock ring and Zed were all Mocha, so GTK was the one layer out of
step, unnoticed because the two dark bases sit close together.
catppuccin-gtk carries Colloid as a git submodule and only recolors it, so
the switch is a recolor with identical widget shapes.
The old objection to gtklock is gone: 4.0.0 links libgtk-session-lock.so.0,
so it speaks ext-session-lock-v1 like swaylock and the compositor still
holds the lock if it crashes. Only the ≤2.x layer-shell versions unlocked
on a crash. What it adds over swaylock is the modules — avatar, clock and
power buttons — which swaylock cannot do at all.
The blur stays baked into the wallpaper. gtklock cannot blur either: GTK3
has no filter CSS property at all, and it takes no screenshot of the
session, so there is nothing to blur at lock time.
Two rewrites of moonarch config were needed and both take the command from
one lockCommand binding: stasis names the locker in four places, and the
Quickshell power popout spawns it itself. Routing the popout through
`stasis trigger lock-screen` would have named it once, but the Lock button
would then fail silently whenever stasis is down. The reason to want it
turned out not to exist — the plan continues after a manual lock anyway,
because stasis takes the state from logind's LockedHint rather than from
having started the locker. Verified by locking by hand and watching DPMS
fire on schedule.
gtkSettings rewrites the Colloid name out of moonarch's settings.ini files.
Without it every GTK3 application would have fallen back to Adwaita: GTK3
reads those files directly and the dconf entry does not reach it on
Wayland.
Verified after rebuild: both hosts build, /etc/pam.d/gtklock carries the
same `auth sufficient pam_unix.so` swaylock had, the seeded stasis.rune
names gtklock at all four call sites, and the popout spawns it.
A rebuild against an unchanged lock installs nothing new, so a machine
built daily can still run months-old packages. A user timer reads the
locked lastModified of the nixpkgs input and notifies once it passes
seven days, offering a terminal that runs the update.
A plain fstab entry is invisible to Thunar and friends — they list
removable media and gvfs entries, so the disk could only be reached by
typing its path. x-gvfs-show puts it in the sidebar, and because the entry
is noauto it appears as a drive that is not mounted yet, which is what it
usually is.
Note when changing this: the gvfs volume monitor reads fstab at startup,
so an already-running one has to be restarted before the entry shows up.
The Crucial P1 was mounted through two hand-written blocks in the host
because moonarch.disk knows only roles that stage 1 waits for, and a disk
that dies every other day must not be one of those. disko has a switch for
exactly this — initrdUnlock = false keeps a container out of the initrd —
so the role fits into the layout after all.
moonarch.disk.storage partitions and formats the disk like the other two,
but its container is opened after stage 1 by a key file on the encrypted
root disk, the filesystem mounts on first access, systemd stops waiting
after five seconds, and an idle timeout unmounts it again. A missing or
dead disk is then a missing directory, not a stalled boot.
The crypttab entry addresses the partition by partlabel, which disko
derives from the names in the layout, so no UUID has to be pasted in and a
replaced disk needs no edit here.
Verified after reboot: container opens as `storage` without a passphrase
prompt, /storage mounts on access, and the initrd holds only `crypted`.
The module wrote the configs and the timers, but nothing creates the
.snapshots subvolumes they need, so snapper-timeline failed every hour
with errno:2 and there was never a single snapshot to roll back to.
Removing it rather than fixing it: rollbacks of / are already covered by
NixOS generations, and snapshots of /home would have shared a disk with
/home. That answers "I deleted the wrong file" but not "the disk is
gone", and the second is the failure that actually occurred here.
The P1 left the layout when /home moved to the XPG. It comes back at
/storage, carrying only data whose loss costs nothing — the Steam library
to begin with.
Every option is chosen so the disk can no longer hold up the boot the way
it used to. The container is opened after stage 1 through a keyfile on the
encrypted system disk instead of in the initrd, the filesystem mounts on
first access only, and device-timeout together with nofail lets systemd
give up after five seconds when the drive is gone. idle-timeout unmounts
it again once nothing uses it, which keeps it cold, and cold is when this
drive works.
Verified on the machine: storage.automount pulls the cryptsetup unit in on
first access through the device dependency, so no explicit ordering is
needed.
The 1T Crucial P1 that carried /home fell off the PCIe bus four times in
four days — controller is down, CSTS=0xffffffff, reset failing with
ENODEV — each time forcing /home read-only mid-session.
Measured on the machine: the controller sits 17 K above the rest of the
drive at idle and reaches 90 C under pure read load while the other
sensors stay at 45-48. Throughput starts at 1.5 GB/s cold and collapses
in a sawtooth as it heats, down to 768 KiB/s over a 15 minute run, with
single reads stalling up to 1.465 s. The XPG under the same fio run
holds ~1 GB/s at 42 C. SMART is otherwise clean (no media errors, no
error log entries, 5% used) and P3CR013 is the final firmware, so there
is nothing left to repair.
/home therefore moves to the system disk, which modules/disk.nix already
does when home.enable is off. The P1 stays in its slot with no
partition, container or mount referencing it.
Steam gets XCURSOR_THEME from the session variables and still drew the
core pointer: parts of it resolve the theme literally named 'default'
instead of reading the variable. Xcursor looks that name up along
XCURSOR_PATH, where ~/.icons comes first, so a link there covers those
lookups as well.
Quickshell has no scope support of its own, so apps started from its
launcher inherited the cgroup of quickshell.service and systemd killed
them along with the unit on every restart — a rebuild that rewrites the
unit was enough, and Librewolf counted each one as a crash until it
refused to restore the session. app2unit is the start prefix the QML now
uses; see moonarch's LauncherPopout.qml.
The unit also gained a start limit. After a rebuild sent niri SIGTERM,
graphical-session.target stayed up while the Wayland socket was gone, so
Restart=on-failure spun 32 times in two minutes against a dead display.
Three tries per 30s ends that; the window slides, so a single crash after
a normal run still gets its restart.
moonarch 7a24cdb puts a suspend step back into the desktop plan; it was lost
when that plan stopped inheriting the ac: blocks. systemctl has to be on the
unit PATH for it, or the step fails silently like swaylock did.
services.fwupd.enable already puts the package in systemPackages, so the
entry only mattered where the daemon is off: on the desktop it left a
working fwupdmgr with no unit to talk to.
Steam is started by the user manager and only inherits XCURSOR_PATH, so
Xcursor falls back to the missing ~/.icons/default and draws the core
theme. The name matches modules/greetd.nix.
The port swapped moonarch's wdisplays-persistent for the nixpkgs wdisplays,
which cannot save -- so the kanshi config stayed empty and its unit was
skipped on every boot. nwg-displays supports niri and writes the layout to
~/.config/niri/monitor.kdl, which niri applies itself.
niri prefers a home config and does not merge, so config.kdl is seeded with
an include of /etc/niri/config.kdl, like moonarch does for stasis.rune.
wlsunset, both cliphist units and kanshi each only need the compositor
socket, so the set from the root oneshots applies. 9.4 UNSAFE to 0.9 SAFE,
verified by starting each one under the full set with `systemd-run --user`.
ProtectHome is out because it hides /run/user too, where the socket lives.
stasis gets five directives at most: swaylock authenticates through the
setuid-root unix_chkpwd, and every directive needing a mount namespace makes
`systemd --user` mount it nosuid, so /etc/shadow becomes unreadable.
Both units read one state file and write a single sysfs attribute, yet ran
as unrestricted root with network access and the full capability set.
systemd-analyze security scored them 7.5 EXPOSED, now 0.9 SAFE.
ProtectKernelTunables and PrivateUsers stay off on purpose; the reasons sit
in the comment, since without them they read as forgotten.
stasis does not run the configured actions through a login shell, so
/etc/profile never applies and nothing resolves via
/run/current-system/sw/bin. The old comment claimed the opposite and listed
only the binaries stasis spawns directly, which left swaylock unreachable:
`niri msg` worked, so the chain looked healthy while the lock step never ran.
The failure is silent. The spawn fails, stasis logs the action as started,
sees the immediate exit, treats the step as finished and stops there -- the
daemon then sits in "waiting for idle" with no lock and no dpms.
Verified: with an absolute path in the config, `stasis trigger lock_screen`
locks; with a bare `swaylock` and the old PATH it exits instantly.
brightnessctl comes along for the laptop's brightness step, which has the
same problem.
niri sets both in its environment block, which only reaches what niri spawns.
Since quickshell became a user service, everything started from its launcher
inherits the user manager's environment instead and comes up unstyled. As
session variables they travel through /etc/profile into niri-session's
import-environment and reach the user manager too.
Brings the logout fix: the power menu quits niri instead of terminating the
login session, which left the compositor running and blocked every later login.
niri-greeter.kdl spawns it by name. The regreet module only ever references the
binary through the default_session command it sets with mkDefault, and that is
overridden here with the niri greeter -- so nothing installed it and the greeter
came up with no UI, quit itself, and started over.
The option defaults to false as soon as a host sets initial_session, on the
grounds that a restart would trigger the autologin again. greetd(5) says it
does not: the initial session runs only on the first launch since boot, marked
by a runfile under /run. Left at the default, greetd exits with the session and
leaves a dead VT that only a power cycle gets out of.
Qt applications are wrapped, so each looks for plugins in its own closure and
never finds a style installed next to it -- VLC ships no Kvantum plugin and so
came up in the stock Qt style however the theme was configured. qt.enable adds
the profile's Qt5 and Qt6 plugin directories, which the wrappers append to.
platformTheme and style stay null: they would only pull in packages already
installed, and niri's environment block sets both variables.
kvantum.kvconfig selects catppuccin-mocha-lavender, but nothing installed it,
so Kvantum fell back to its built-in default and every Qt application went
unthemed. Of the four places it searches (style/Kvantum.cpp) three are under
$HOME and the fourth is a DATADIR compiled into the style plugin, which under
Nix is the plugin's own store path -- so a theme package in systemPackages can
never be found. Seeded like the other home-only configs.
As a spawn-at-startup it inherited niri's environment; the unit got systemd's
minimal one, so every widget that shells out failed to start its process --
niri, nmcli, powerprofilesctl, the moonarch scripts and a dozen more.
nevaforget
nevaforget