pacman refuses to install over a file it doesn't own, so any path an
Omarchy package starts shipping that a script had already written by
hand aborts the whole upgrade:
omarchy-settings-dev: /usr/lib/systemd/user/omarchy-fcitx5.service exists in filesystem
Errors occurred, no packages were upgraded.
The mitigation was a hand-maintained --overwrite allowlist, and it only
worked with foresight: an entry had to ship a release before the package
took the path, because pacman checks conflicts during transaction
prepare, so the script running the upgrade is the one already on disk. A
missed entry left people hard-stuck, since the upgrade that would
deliver the entry is the one refusing to run.
So react instead. omarchy-update-system-pkgs now does the ordinary thing
and hands a failed transaction to omarchy-update-system-pkgs-when-
conflicted, which moves the offending files out of the way and runs the
upgrade again. Nothing has to be predicted, and the allowlist is gone.
Moving rather than overwriting is what makes it small: no --overwrite
argument to build, no glob escaping, no separate backup step, and a
leftover directory is cleared too, which --overwrite cannot do at all.
Files go to /var/lib/omarchy/replaced/<original path>, not next to the
original. A sibling copy is not inert -- SDDM reads every file in
sddm.conf.d whatever its extension, and systemd-sleep runs every
executable in system-sleep -- and the directory a future path lives in
is unknowable, which is the whole point of a mechanism for paths nobody
predicted.
What it will not do:
- Take a file another package owns. pacman reports those with a
"(owned by x)" suffix, so the end anchor excludes them, and pacman -Qo
re-checks the live database before anything moves.
- Move a subset. If any reported conflict isn't recoverable the retry is
doomed anyway, and moving leaves that config inactive for nothing --
worse than the stuck-but-intact state.
- Leave anything inactive that was live. A failed retry, a failed move
partway through the loop, or an interrupt all put back whatever the
upgrade didn't install.
- Act on a report handed to it by hand. It is internal to the update,
and an old report would clear live files for an upgrade that isn't
happening.
Also adds a test that fails when a script writes a path under /usr that
no PKGBUILD installs, since not creating these is cheaper than
recovering from them. It reads the destination off the command, so a
path assembled from variables still slips through; the two known cases
are recorded with their reasons.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
omarchy-migrate-notify handed its notification to `systemd-run --scope`,
which is synchronous: the calling process becomes the payload, so
systemd-run does not return until the toast has been answered -- and if
the toast is clicked, not until the migration terminal it opens has been
closed.
omarchy-migrate-notify.service is Type=oneshot, which defaults to
TimeoutStartUSec=infinity, so the unit sat in activating for exactly that
long. On a machine that had not yet picked up c7e327b0, which orders the
notifier after graphical-session.target, that held the target open --
and wayland-wm-app-daemon.service is ordered after the same target, so it
never started. Every keybinding goes through uwsm-app, which waits ten
seconds for that daemon's pipes and then reports "App failure -- Timed
out waiting for pipes!" instead of launching anything. Keybindings were
dead for two minutes and seventeen seconds, until a pacman hook ran
`systemctl reload user@*.service`, whose re-exec of the user manager
broke the scope's bus connection and let the oneshot finish. That also
made systemd-run exit non-zero, so the notifier fell through to the
terminal fallback meant for having no user manager at all, and printed
the migration list into the journal.
Hand the notification to a transient service instead. systemd-run
returns once the unit has started rather than once the payload is done,
so the oneshot completes in milliseconds and nothing ordered after the
target waits on a toast. Put it in background-graphical.slice, which
systemd ships with PartOf=graphical-session.target, so an unanswered
toast ends at logout rather than outliving the session.
The terminal a clicked toast opens is unaffected: uwsm-app re-registers
it into app-graphical.slice, outside this unit's cgroup.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
fcitx5 is what turns the CapsLock compose sequences in ~/.XCompose into
text for Wayland clients -- CapsLock m s for an emoji, CapsLock space n
for your name. It was launched fire-and-forget from Hyprland's autostart
via uwsm-app, so nothing supervised it, nothing restarted it, and it
logged nowhere. When it went away, every compose sequence stopped
working for the rest of the session with no visible symptom beyond
"emoji input is busted", and no record of why.
That is not hypothetical: it was found dead on a running machine with no
coredump, no OOM kill, and nothing in the journal to explain it.
Move it to a systemd user service:
- Restart=always, not on-failure. fcitx5 exits 0 when it finds another
instance already owning org.fcitx.Fcitx5, and a clean exit still
leaves the user with no input method.
- After=/PartOf=graphical-session.target. It needs WAYLAND_DISPLAY and
DISPLAY, which uwsm imports into the user manager before reaching the
target, and its wayland connection dies with the compositor.
- ConditionEnvironment=WAYLAND_DISPLAY. After= is ordering only and does
not stop the unit from being started while the target is inactive. An
update over SSH has a live user manager (pam_systemd) and no graphical
session, and a fcitx5 started there comes up blind -- then stays
active, so the later target activation won't pull in a working one,
because Wants= does not restart what is already running. Skipping the
start leaves the unit enabled and healthy for the next graphical login.
The migration hands over inside a live session only: it enables without
--now, and only when graphical-session.target is active does it drop the
autostart-launched process and start the unit. Because that kills a
fcitx5 that was working a moment ago, a failed start is reported instead
of leaving the session mute with the migration marked complete.
omarchy-restart-xcompose now drives the unit. It still clears any fcitx5
running outside it first: that process owns the bus name, so the unit's
instance would exit on arrival and the stale one would keep serving the
old table -- a restart that reports success and changes nothing.
Side benefit: fcitx5 now logs to the journal under its own unit, so the
next disappearance leaves a record.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The retired omarchy-update-user-notify.path stays loaded in sessions that
started before the update removing it, and pacman writes the migrations
directory mid-transaction, so it fired a critical toast for migrations that
omarchy-migrate was about to apply a step later. Migration 1785095882 stops
that watcher, but migrations run after pacman, so it lands 11 seconds too
late to prevent the toast it exists to retire.
Check the lock omarchy-update holds for its whole pipeline instead of
trusting that no trigger exists. That covers the stale watcher and anything
added later: during an update every pending migration is by definition
already being applied. The check repeats after waiting for the notification
server, which is long enough for an update to start underneath it.
Only this user's runtime directory is read, never the /tmp path the updater
falls back to without XDG_RUNTIME_DIR. A shared lock file belongs to whoever
created it first, so honouring it would let one user silence another user's
notification; a redundant toast is the better failure.
The sleep inhibitor now starts with the lock descriptor closed. It outlives
the step that starts it, so an update killed before restore_update_inhibitors
left it holding the flock indefinitely. That already blocked later updates,
and now that the notifier reads the same lock it would have silenced
migration notices at every login.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
omarchy-migrate-notify.service is a Type=oneshot wanted by
graphical-session.target, and systemd complements a target's Wants= with an
implicit After=, so the target waited for the notifier to exit. The notifier
does not exit quickly: it sends the notification through systemd-run --scope,
which is synchronous, and omarchy-notification-send -a blocks until the user
clicks. The target stayed in activating for as long as the toast was up.
wayland-wm-app-daemon.service is After=graphical-session.target and nothing
wants it, so uwsm-app starts it on demand. Clicking the notification runs
omarchy-launch-floating-terminal-with-presentation, which execs uwsm-app,
which blocks on a systemctl --user restart of that daemon -- a job queued
behind the very target the clicked notifier was holding open. The terminal
never opened; uwsm-app gave up on its own pipe timeout instead.
Declaring After= on the wanted unit suppresses the implicit dependency rather
than forming a cycle, so the target is reached without waiting and the
notifier runs behind it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Both sides tightened the same center-layout assertion. Taking quattro's:
it asserts the weather/update adjacency the test name is about instead of
pinning the whole row, which is what kept breaking it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
A cloned bar widget is copied out on its own, so a relative JS import
lands in a directory without the file and the clone fails to load. Point
those imports back at the bundled module, the same way Qt.resolvedUrl()
references were already rewritten.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* Send files to a tailnet machine with Taildrop
The panel gets a send button next to the copy one on every machine that
Tailscale grades as a Taildrop target, and `s` does the same from the
keyboard. Picking runs through the XDG portal chooser, so it looks like
the file dialog every other app opens.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TJQJfHXXApUk6En8EZisHg
* Save incoming Taildrop files and say so
Linux keeps Taildrop files in the daemon's inbox until someone asks for
them, so nothing arrived until you ran `tailscale file get` by hand. A
user service now stages each delivery next to the downloads directory,
hands it over under a free name, and announces it — with a preview when
it's an image, and a click to open it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TJQJfHXXApUk6En8EZisHg
* Float every portal dialog, not just the titled ones
The portal only ever shows dialogs, and the title regex missed any
chooser an app names something else — ours included.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TJQJfHXXApUk6En8EZisHg
* Re-run the Taildrop enable now that the unit ships
The unit was never installed to /usr/lib/systemd/user/, so the enable had
nothing to act on and machines that already ran the migration carry a marker
for a no-op. Rename it so they get a working pass, and report what systemctl
says instead of a bare failure line.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* Wait for the file chooser on the connection that asked for it
The portal answers a request with a Response signal directed at the connection
that made it, and dbus-daemon delivers directed signals only to that
connection. gdbus monitor registers with AddMatch rather than BecomeMonitor, so
it never saw the reply: every pick left omarchy-file-select blocked on a read
that could not arrive, taking omarchy-tailscale-send down with it before it
reached either its notification or the transfer.
Make the call and wait for the signal on one connection, and give up after ten
minutes so an unanswered dialog cannot strand the caller.
Drop the "Sending to" notification while here, so a send reports once.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* Mark Taildrop notifications with the panel's send glyph
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* Stop a hung tailscale poll from freezing the panel
Each poll is skipped while its own process is still running, so one that never
exits leaves the panel showing whatever it last read, for good: the peer list
keeps a woken machine missing, and opening the panel cannot help because open
runs the same refresh that hits the same guard.
Reap anything still running fifteen seconds after a refresh, well inside the
thirty second interval, so the next tick starts clean.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* Place a moved widget where an added one lands
'move omarchy.media left' named a section, not a slot, but the section went
through as an explicit target, which resolves a missing index by appending. The
widget landed on the far end of the row instead of after the section anchor
where 'add' puts it.
Its test has never run: the assertion covering this went in four hours after an
unrelated layout change had already stopped the file, and the runner stops the
whole suite at the first failure.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* Keep the config test from failing on things it is not about
The center layout assertion pinned the whole row, so parking the indicators
left of the clock broke a test named for update sitting next to weather. Assert
that adjacency instead.
The package-defaults check reads PKGBUILDs from the omarchy-pkgs repo and blew
up with a traceback wherever that is not a sibling checkout. Skip it when the
checkout is absent, honour OMARCHY_PKGS_ROOT when it is somewhere else, and
keep failing when it is present and wrong.
Between them these stopped the suite eighty files early.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* Make the file chooser a Python command rather than a bash host for one
The portal work was a heredoc wedged inside a bash script that existed only to
parse two flags. Drop the host: argparse covers the flags, and the file says at
the top why it is the one command here not written in bash.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* Tell a chooser that never opened apart from one that was dismissed
Three fixes from review:
The poll watchdog rearmed on every refresh, so a refresh interval shorter than
its timeout — the setting goes down to five seconds — pushed the deadline ahead
of a hung process forever. Arm it on the launch that needs watching and leave
it alone.
omarchy-file-select exited 1 both for nothing picked and for a chooser that
could not run, and omarchy-tailscale-send read it through a process
substitution, which drops the status anyway. A session bus that was not there
looked exactly like someone changing their mind. Separate the two exits and
read them with a command substitution.
Delivery picked a free name and then renamed, which overwrites anything that
takes the name in between. Link to the name instead: link(2) refuses one that
is taken, so the check and the claim are the same step.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
* Anchor the webcam overlay to the recorded region, not the monitor
Recording a single window on an ultrawide put the camera outside the frame.
The overlay is placed twice, and neither placement knows what is being
captured: the window rules in default/hypr/apps/webcam-overlay.lua position it
from monitor_w/monitor_h, and omarchy-capture-webcam-resize recomputes the same
corner from hyprctl monitors. Both resolve to the monitor's bottom-right, so a
window anywhere but the far right of the display records without the camera in
it, and dragging it into frame by hand is undone by the next resize.
The geometry is already known. select_capture_target resolves the picked window
to a region rectangle before start_webcam_overlay runs, and a bare click in
smart mode snaps that rectangle to the window's own bounds. gpu-screen-recorder
takes the region in the compositor's logical coordinate space, which is the
same space window moves take, so the value needs no conversion to be reused as
an anchor.
Publish it for the resize helper, which becomes the single place that positions
the overlay. It anchors to the region when one is recorded and falls back to
the monitor otherwise, so full-monitor captures, the portal backend, and manual
resizes outside a recording all keep their current behaviour. A malformed or
empty file falls back the same way.
Presets scale from the anchor's height rather than the monitor's, so the camera
keeps its proportion of the frame instead of covering a small capture outright.
A tall, narrow region cannot fit a preset derived from its height, so widths are
capped to the space available and heights follow at the same 8:9 aspect.
Placement happens inside start_webcam_overlay rather than after it returns.
Correcting the position once that function had returned left the move adjacent
to gpu-screen-recorder starting, and the camera was recorded sliding the last
stretch into its corner over the opening frames. The overlay is waited for
explicitly instead, positioned, and only then does capture start.
Waiting for the map is what the blind second was partly guessing at, so the
remainder is trimmed to hold the delay before capture where it was. Starting
later is not free: it eats the opening words of whatever is being narrated.
* Keep the webcam size ladder usable in a narrow region
Capping each preset's width to the region separately collapsed small,
medium and large onto the same width, so Super + Alt + [ and ] had
nothing to step between, and integer division left small taller than
medium. Cap the height the presets scale from instead, which shrinks
the ladder as a whole and keeps the three sizes ordered and distinct.
Cover the anchoring in the test suite: a region the camera follows, the
fallback for an unreadable one, and the narrow-region ladder. Point
XDG_RUNTIME_DIR at the test's own directory while doing so — the resize
helper now reads a region file from there, and the existing geometry
assertions would pick up a real one from a live recording.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: David Heinemeier Hansson <david@hey.com>
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
omarchy-update-user-notify.path watched /usr/share/omarchy/migrations, but
pacman writes that directory during every update, including the blessed
omarchy update, which runs omarchy-migrate a step later. The watcher fired a
critical notification for the migrations the update was already applying in
the visible terminal. A watcher cannot tell that apart from a bypassed
pacman -Syu, so the only trigger that never collides with a running update is
a once-per-login check.
The service that already ran at graphical-session.target is now the whole
mechanism, renamed after the command it runs. That is also all the second-user
case needs: markers are per-user, so anyone who did not run the update finds
them missing at their next login.
Login timing means the toast can be sent before the shell has claimed
org.freedesktop.Notifications, so the notifier waits for a live server first.
The wait is omarchy-first-run's, lifted into omarchy-notification-wait rather
than duplicated.
The package keeps omarchy-update-user-notify.service as a symlink onto the new
unit. Existing users hold an absolute wants symlink to the old path, and the
migration that repoints it only runs for users who run an update, which is the
opposite of who the notifier is for.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
libfprint-git only existed to carry the focaltech_moc driver and the
FocalTech FT9349 (2808:a97a) ID ahead of a release. libfprint 1.94.100
has both and is in the Arch repos, so fingerprint setup installs stock
libfprint again and a migration swaps existing installs back.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
pacman checks file conflicts during transaction prepare, so the --overwrite
entry added alongside /etc/systemd/zram-generator.conf never got to run: the
copy of omarchy-update-system-pkgs driving the upgrade is the one already on
disk. Every machine carrying archinstall's copy aborted the transaction.
zram-generator.conf(5) reserves /etc for the local admin and has vendors ship
snippets under /usr/lib/systemd/zram-generator.conf.d/, where drop-ins outrank
the main config file. Nothing collides, and the tuning wins whether or not
archinstall's file is still around.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LjeyQZsNBxqyYy9z8KaKm7
A Dell XPS 14 with no fingerprint reader was invited to set one up on first
run. Its Synaptics webcam bridge, 06cb:0701 "SVP7500" bound to usbio-bridge,
matched omarchy-hw-fingerprint on the bare 06cb vendor ID.
The vendor list already excludes Elan and STMicro for being multi-purpose, but
Synaptics is just as multi-purpose and was left in. Dropping 06cb is not an
option: Synaptics and Validity are among the most common real reader vendors,
and many of those readers report no product string at all, so the vendor ID is
the only signal that finds them.
The vendor guess is now qualified instead. libfprint drives every reader it
supports from userspace over libusb, so a real reader sits with no kernel
driver bound to any of its interfaces, while the bridges, touchpads and cameras
that share these vendor IDs all bind one. The guard applies only to the fuzzy
vendor-ID path; a device whose product descriptor names it a fingerprint reader
is still trusted outright, since that signal is precise enough on its own and
should not be lost to a driver that happened to claim the device.
The menu entry was a second surface for the same bug. Setup > Security >
Fingerprint carried no when clause, unlike the Remove entry beside it, so it
appeared on every machine and dead-ended in the setup script's own hardware
check. It now gates on the same detector.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017q97Rsk8KwgaPUVq5ArLgM
Closing the lid could suspend the machine with the session still exposed.
omarchy-sleep-lock held a delay inhibitor and waited for Quickshell to report
the session secure, but a delay inhibitor is a timer rather than a promise:
logind suspends once InhibitDelayMaxSec expires, locked or not. The default is
five seconds, and closing the lid also reconfigures displays -- exactly when
the lock service is waiting for the screen set to settle before it can secure.
The race is now off the critical path. switch:on:Lid Switch runs
omarchy-system-lid-close, which requests the lock the moment the lid closes,
before logind has decided to suspend at all, so the inhibitor window usually
finds the session already secure. A docked lid close does not suspend
(HandleLidSwitchDocked defaults to ignore) and must not lock either, since that
is clamshell mode still in use on the external display, so the handler gates on
the same closed-and-undocked pair logind itself keys on.
Suspends that never touch the lid -- idle timeout, the system menu, low battery
-- still arrive through the inhibitor, so that path gets room to work too. The
shipped drop-in raises InhibitDelayMaxSec to 15s, and the helper derives its
budget from logind's live InhibitDelayMaxUSec rather than assuming the drop-in
landed: a machine that has not reloaded logind yet, or that carries its own
override, gets a budget that fits what logind will actually allow. It leaves
logind a fifth of its own window and caps at 12s, so a hand-raised window
cannot strand a closed laptop awake in a bag.
The wait itself had three defects. Its deadline arithmetic read EPOCHREALTIME
assuming a period, so under any comma-decimal locale the subtraction parsed as
bash's comma operator and silently voided the deadline, leaving only the
attempt counter to stop it. The lock request shared the status polls' timeout
and exited on first failure, so a shell 300ms slow meant suspending unlocked;
it now has its own budget and is simply retried, since asking again is
idempotent. And a refusal the shell reports on stdout with a zero exit --
missing-pam -- read as success, burning the whole window on a lock that could
never happen.
Every call is bounded by what is left of the budget rather than by an estimate
of what the step should cost, so the deadline holds on hardware slower than
anything the constants were fitted to.
Failure is still possible and it used to be silent. It now writes to the
journal and raises a critical notification, which lands on the screen the user
unlocks into.
Incidentally, monitor-recovery-test asserted lid probing against
omarchy-hw-clamshell after that logic moved to omarchy-hw-laptop-closed. It
aborted the file under set -e, skipping the nine assertions behind it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The user override was written with mode="assign", but it loads after the
package default has already assigned JetBrainsMono, so it never displaced
it. prepend_first puts the chosen family at the head of the list.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
It was the user override file for the first Quickshell bar, back when the
bar layered it over bar-defaults.json. Unifying everything into
shell.json left an empty file that installs and the quattro upgrade kept
faithfully copying around.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
tmux only raises an activity flag for windows that are not currently
selected, so a long-running command in the window you left selected --
the common case, since you switch away by moving your Hyprland focus
elsewhere rather than by selecting another tmux window -- finished
silently and the indicator never lit.
Windows now also count as waiting when they are selected in an attached
client, every client showing them is unfocused, and their window_activity
is newer than an @omarchy_unfocused_activity watermark. The watermark is
stamped by a new `track` subcommand wired to client-focus-in/out and the
existing select-window hooks, so it records where attention last was.
The hooks pass #{window_id} and #{window_activity} as arguments, which
tmux expands when the hook fires; run-shell -b would otherwise let output
arriving during the handoff be swallowed by the new watermark. The focus
hooks take index 100 to leave a user's own bindings alone.
That state has no hook of its own, so it needs polling to be noticed.
The probe therefore moves out of the indicator and into a service plugin,
alongside nightlight and battery. A bar surface exists per monitor and
each one instantiates every indicator twice, once per block, so a timer
on the indicator meant a shell-out per instance per tick -- four probe
processes every three seconds on a two-monitor machine, forever. One
service polls for the whole shell instead.
Sharing the state also fixes what per-instance polling would have papered
over: each indicator used to own its own count, so a timer-driven update
only refreshed the bar it ran on and left the other monitors stale.
Refreshes still arrive over the existing indicator broadcast, which now
coalesces into a single run no matter how many bars relay it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018CC9kSQv8ZEogaxDoeKBzL
The tmux alert hooks fire omarchy-shell through run-shell, and run-shell
builds its own environment rather than passing the server's along, so the
hook ran without WAYLAND_DISPLAY. qs matches instances by display, so every
refresh from a bell died at "omarchy-shell is not running", which -q then
swallowed by design.
That left the indicator with no way to turn itself on. Its five second poll
only runs while the indicator is already active, so an alert raised after
startup was invisible until the shell happened to restart with the window
still flagged. The icon appeared roughly never.
Recovering the display from the socket in XDG_RUNTIME_DIR fixes it in the
one place that needs it, and covers any other caller reaching the shell from
a stripped environment rather than just the tmux hooks.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Omarchy installed zram-generator but never shipped a config for it, so the
only zram tuning any machine had was whatever the ISO wrote once, and the VM
sysctls were untouched kernel defaults aimed at spinning disks.
The generator's own default caps the device at 4G, which on a large machine
sends reclaim to the hibernation swapfile far earlier than it needs to go.
Ship min(ram / 2, 8192) instead, matching what Fedora settles on, and state
the priority explicitly since it has to sit above the pri=0 that
omarchy-hibernation-setup gives the disk swapfile.
The four sysctls follow from swap being RAM rather than a disk. swappiness
above 100 says evicting an anonymous page beats dropping a page-cache page
that would have to be read back, which is true once swap is compressed
memory. page-cluster drops to one page per fault because the default of 8
buys readahead for a seek zram doesn't have and pays a decompression for
each page. Zeroing watermark_boost_factor stops fragmentation from producing
reclaim bursts while memory is still free, and raising watermark_scale_factor
gives kswapd room to reclaim in the background instead of letting allocations
stall in direct reclaim. The last two are what actually addresses the stutter
people notice and misread as swap being used too eagerly: proactive swapping
to zram is the cure, and synchronous direct reclaim is the disease.
Every ISO-installed machine already has an unowned zram-generator.conf, which
pacman would refuse to overwrite, so both upgrade paths need it listed.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Reported to suit the XPS 16 as well, so the tuning now matches both machines and
the directory is named for the pair rather than the one model.
Gating moves from the DMI product name to the product SKU, which is what Dell keys
the Cirrus speaker firmware on -- 10280db9 for the XPS 14, 10280dba for the XPS 16
-- so it identifies the speaker hardware itself instead of a marketing name. SKUs
are compared as whole values, not substrings, so this cannot widen to the rest of
the XPS line the way a short product-name match would: product_family on these
machines is "Dell Laptops".
match_sku and match_dmi are now lists, letting one tuning name every model it
covers. A plain string still works as a single entry.
The measurements in tuning.conf were taken on the XPS 14 (0DB9). The XPS 16 is
covered on report rather than measurement, and tuning.conf and the graph both say
so, because the profile's bass lift and high-pass were fitted against one
machine's excursion and distortion limits.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Laptop speakers ship voiced by the vendor's Windows DSP layer, which Linux does
not get. A tuning restores that as a PipeWire filter-chain in front of the
internal speaker sink, matched to the machine by DMI string and expected sink.
Adding a laptop is a directory under default/audio/tunings with two files and no
new code: matching is data. The XPS 14 DA14260 tuning included here was derived by
measuring the xps-audio-linux EasyEffects profile (MIT) and fitting a biquad chain
to it, so no impulse response or other upstream asset is redistributed. It measures
1.24 dB RMS against that reference, and matches its dynamic range within 0.1 LU --
the reference's multiband compressor turned out to contribute nothing, so a linear
chain replaces it. Bass Q is capped deliberately: a closer magnitude fit swung
group delay 31 ms across 63-80 Hz, which smears bass transients.
The graph runs as its own PipeWire client under its own config name rather than
loading into the audio daemon. The daemon only reads its config at startup, so a
daemon-loaded tuning could only be switched by restarting PipeWire -- which drops
every PulseAudio client's connection, and applications that do not reconnect
(Spotify) then have to be restarted by hand. Hosting it separately also contains
failure, since a malformed tuning breaks only that service.
Three things about the surrounding audio graph needed fixing for this to behave:
- Volume must live downstream of the tuning. omarchy-audio-output-sink is now the
single definition of which sink an output's volume really uses, shared by the
volume keys, the output switcher's OSD and the audio panel, so they cannot
disagree. It resolves the current default output, which keeps it correct when
headphones are selected while a tuning exists.
- The tuning's own output is a movable sink input, so rerouting "all streams" to a
newly selected output would drag the processing onto headphones, or into the
tuning's own sink, which is a cycle. It is pinned, and stream moves are limited
to streams carrying an application.name.
- The physical sink a tuning fronts is not independently selectable, since picking
it would only bypass the tuning, so it is kept out of the output list.
Applying happens at first-run, not finalize-user, because finalize-user also runs
in the ISO chroot where there is no audio server and nothing would retry.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The on-battery toggle bought a fraction of a watt while the radio
idled -- twenty-some minutes across a full discharge in the best case
-- and paid for it with 20-300ms latency spikes on idle links and
outright link drops on Intel BE200/BE211 firmware. NetworkManager can
keep power save off natively with one conf.d line, which replaces the
udev rule, the setter script, and any per-card deny-listing.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
An unsupported terminal notified once per monitor and each monitor
paid a wasted wait; check before the loop instead. The per-read
timeout also let a chatty event stream extend the wait indefinitely,
so spend the 5 seconds against one deadline.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
