Skip to content

Sync upstream quattro through d4985887 - #103

Merged
mark-groves merged 33 commits into
quattrofrom
cursor/upstream-changes-pr-d498
Oct 29, 2026
Merged

mark-groves merged 33 commits into
quattrofrom
cursor/upstream-changes-pr-d498

Conversation

@cursor

@cursor cursor Bot commented Oct 28, 2026

Copy link
Copy Markdown

Upstream quattro moved from a9cedf93 (27 Oct, already on this fork via #99) to d4985887 (28 Oct, oligarchy tests omacom#16926). This replays all 33 first-parent commits, including merges omacom#14390, omacom#16800, and omacom#16926. Four files also differ on the fork (AGENTS.md, default/i18n/en/cli.json, install/omarchy-other.packages, shell/shell.qml); the hunks do not overlap, so they merged on their own. Face, Grok Bot, and Cursor stay.

What changed

Backgrounds

Screensaver

Audio

  • rtkit is in the base package list so PipeWire can get realtime priority. Without it, audio threads stay at normal scheduling and crackle under load, and a stale realtime portal can hand clients a hard realtime budget of 0 that the kernel then enforces by killing the process (Give PipeWire realtime with rtkit and stop the silent app kills omacom/omarchy#16824, PipeWire runs without realtime priority on a stock install (rtkit is not included) omacom/omarchy#9390).
  • omarchy-realtime-repair raises those zero budgets before rtkit is installed, restarts a portal that still reports zero, then restarts the caller's audio units if the PipeWire server is not realtime yet. Migration 1789747667 runs that repair and stays pending when something it should have fixed or inspected is left behind. The Quattro upgrade runs the same repair after installing Omarchy and before the base package list.
  • PipeWire and JACK clients ask rtkit themselves (rtportal.enabled = false), matching the servers. WirePlumber is not covered: it does not read a drop-in for this.
  • The extra aarch64 rtkit line is dropped now that the base list carries it.

Displays and input

Capture

Surface Pro 11

  • Board support for the Snapdragon X Elite OLED model (microsoft,denali-oled only; the LCD and X Plus models share the DMI name and are left alone). It drops a leftover Intel keyboard module file, pins that model's device tree into the UKI, and installs iptsd, surface-pro-11-sensors, and surface-pro-11-support (Add Microsoft Surface Pro 11 board support omacom/omarchy#14390).

Tests and checks

Catalogs are current through omacom#16951.

Fork overlay

Face, Grok Bot, and Cursor stay. FaceChromeCanvas stays immediately before FilterInput. Cursor stays the official tarball. linux and linux-headers stay in install/omarchy-other.packages. The power-profile menu provider stays out and the write stays busctl. English catalogs keep the face, Grok Bot, and Cursor rows. Fork delta vs d4985887 is 221 paths, the same set as quattro vs a9cedf93.

#101, #102, and topic draft #94 (cursor/factory-reset-gum-argv-421a) were left alone. No other open sync.

Open in Web View Automation 

dhh and others added 30 commits October 28, 2026 23:03
Cut opens the downloaded video in Omacut, like the screen recording toast.

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
* Skip the Kitty parser checks when only try-omarchy's wrapper answers

try-omarchy VMs put a wrapper named kitty first on PATH that exits 127
until Kitty is installed, so the test's command check found it, ran
`kitty +runpy`, and failed with its install hint instead of skipping.
Ask Kitty itself for its version before using it.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

* Use the scratch shell's names in the runtime smoke test's shell.json checks

The unreadable-shell.json checks added in 463a9a1 were written against
another file's fixture: they read $test_home, $TMPDIR and $QS_PID and
call shell_ipc, none of which this file defines. It runs on the scratch
shell from base-test.sh, whose names are SCRATCH_HOME, SCRATCH_DIR,
SCRATCH_PID and scratch_ipc. Under set -u the first of them stopped the
file at every run on a desktop, so the checks never ran; CI never reaches
them, since the file needs a compositor. With the names right, they pass.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

---------

Co-authored-by: darshil dhameliya <dhameliyadarshil@gmail.com>
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
…omacom#16859)

The transparent bar check from omacom#8307 sets the bar's background with
omarchy-bar, which edits shell.json from outside the shell. Elsewhen
saves its first-run cities once its timezone probe returns, and that
save writes the shell's whole in-memory config back to shell.json. When
the probe returns after the edit, the file goes back to an opaque bar,
the bar never turns transparent and nothing samples the wallpaper.

On a desktop the probe returns before the test gets there. In a VM it
returned later in 7 of 8 runs: the file read "background": "opaque"
with Elsewhen's cities saved, 30 s after omarchy-bar had reported
"Bar background set to none".

The unreadable shell.json checks further down already waited for that
save. The wait is now one helper, called before both edits, and allows
30 s, since it returns as soon as the cities are saved.
…m#16715)

* Let Omasnap's scroll capture see the screen behind its overlay

* no-mistakes(review): Drop source-grep assertion for removed no_screen_share rule
…macom#15382)

* Keep the idle spawn test's exec logger off the vfork child's stack

Qt 6.12 starts a QProcess child with clone(CLONE_VM | CLONE_VFORK) on a stack
carved out of the parent's own, about 8KB above the parent's frame in clone().
The logger's execve wrapper built its line in an 8KB stack buffer, ran over
the return address the parent resumes on, and the test shell segfaulted
before it became ready. Build the line in pages mapped for the call instead.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Size the vfork child stack in the logger's comment by SIGSTKSZ

Qt 6.12's forkfd gives the child `char childStack[SIGSTKSZ]`, and under _GNU_SOURCE glibc computes SIGSTKSZ at run time: 4 * the kernel's minimum signal stack, never below 8192. On a CPU with AVX-512 it is 13504, so the old 8KB buffer only overran the child stack on some machines; saying "about 8KB" hides why the test passes on one machine and crashes on another.

Co-Authored-By: Codex Medium <noreply@openai.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Omarchybot <317366263+omarchybot@users.noreply.github.com>
Co-authored-by: Codex Medium <noreply@openai.com>
…om#16862)

The check ran on every pull request and decided inside the job whether
anything i18n-relevant had changed, so a PR touching none of it still
took a runner for about 20 s, checked out the whole history and showed a
check that tested nothing. The same file patterns are now the workflow's
paths filter: an unrelated PR starts no job, and one that matches runs
both steps directly.

Without the in-job diff against the base the checkout no longer needs
the full history; both steps pass from a depth-1 clone.

It ran on every PR so that it could become a required status check.
Nothing on quattro requires it, and a path-filtered check must not be
required, so the header says to use a job-level condition if it ever is.
…m#16872)

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
* Play a background's intro when changing backgrounds

Choosing a background, cycling to the next one, or reselecting the current
theme now plays the selected still's matching intro through the same cover
handoff a theme switch uses.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Order theme switches around background intros

A theme switch now waits for a background intro's opening fade before
capturing its outgoing frame, and a theme intro's deferred Hyprland
reload waits for a background intro that interrupted it to finish.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Keep desktop double-clicks working while a video background plays

OWE's video layer takes no input and the shell hides its own background
layer under it, so double-clicks on the desktop went nowhere during
intros and video wallpapers. A clear layer above the video now takes
them while the background is suspended.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Let a newer background intro take over a playing one

Background intros share the reload lock, so a second selection no longer
waits behind the first clip; only a theme's deferred reload excludes them.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Play background intros straight through OWE

A background change keeps the theme and its palette, so it needs none of
the theme switch's shell cover: the selector plays the clip from its first
frame and hands the shell the new still afterwards. This drops the cover
handshake, its tokens and the lock that ordered it against theme switches.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
* Fade background intros in like theme intros

Cutting straight to a background's intro left the desktop black until OWE
presented the clip's first frame. A background change now uses the theme
switch's cover: the outgoing frame stays up until the clip is moving, then
crossfades out. Each change carries its own token, so a quick second
selection cannot reveal the first one's cover early.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Hand the shell the new still before a background intro plays

OWE gives the desktop back to the shell as the intro ends, and the shell
still held the old background until the new one arrived, so it flashed
for a few frames. It now takes the new still as soon as the cover hides
it. The cover captures the current frame before the intro is prepared
into the same renderer, and the desktop double-click no longer refreshes
the background after the picker, which started an ordinary transition
under the cover.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Fade from a playing intro instead of cutting

A theme or background change to a still without an intro of its own ran
the shell's ordinary transition while a playing intro hid the shell's
background, so OWE handed the desktop back after it had finished and
the change landed as a cut. Such a change now holds the playing frame
under the intro cover, and the cover waits for OWE to hand back a
decoded still before it fades.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Open the cover with the background cut for stills without an intro

A change to a still without an intro of its own, made while an intro
plays, held the playing frame under the intro cover and faded it like an
intro. It now opens that cover with the same slanted cut as an ordinary
background transition, so only intros fade.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Present the background cover before repointing the link

Repointing the background link cancels a playing intro and hands the
desktop back to the shell, so a cover still decoding then appeared after
the handoff and flashed the old frame over the new still. The cover is
now on screen before the link changes, as on a theme switch.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* Sequence background intros in the shell

A background change and a theme switch's intro were sequenced from bash,
which could only poll the shell and OWE for each step and kept tokens in
files and a lock to stay ordered. The shell's new BackgroundSwitch
service now owns the sequence: omarchy-theme-bg-set hands a change over
with one call, and omarchy-theme-set hands its intro over once its cover
is up. Each step waits on the event that allows it, from the cover
reaching the screen to OWE handing the desktop back, and Hyprland's
reload after a theme's intro waits until no intro plays.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Settle background switches after the reveal starts

A reveal clears its cover token just before its animation starts, so a
queued selection could start in between; settling now waits for both. A
theme's reload is decided once the theme switch lock is free, since a
newer switch may start an intro while holding it, and a handover without
the background plugin still brings its cover down.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Order background switches against theme switches and each other

Background work still being set up made a theme switch wait, and work a
theme switch overtook anyway is dropped, so it can no longer replace the
theme's cover or repoint the link after it. Link updates run one at a
time and end on the latest selection.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Start the cover's reveal after the palette's restyle

The new palette restyles the whole shell in the frame the reveal began,
and the fade started in that frame was already some 60 ms, about 40%,
along when it was first drawn, so a theme's intro stuttered into view.
The fade or cut now starts two frames later and eases in from the
outgoing frame.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Cycle backgrounds from the shell's pending selection

The shell repoints the background link only once a change is under way,
so quick presses of omarchy-theme-bg-next read an old link and picked the
same background twice. The shell now answers which background the latest
selection names, and cycling moves on from that.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Stage the login intro marker under a fixed name

The marker is written under the intro lock, so only one writer stages it
at a time; a fixed staging name replaces mktemp and its cleanup trap.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Simplify the background switch handover

omarchy-theme-bg-set finds the intro itself and hands it over, so the
shell decides synchronously and needs no resolver process, generation
counter or resolve queue. A background change no longer prepares its
clip separately: it starts once the cover is up. omarchy-theme-set drops
its fallback for a shell from before the handover, since updates
restart the shell, and the shell no longer relinks what it already
linked.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Hold the theme lock through Hyprland's reload

The reload after a theme's intro released the theme selection lock
before reloading, so a theme switch could start its own intro in
between; it now reloads while holding the lock. The shell reports a
background change as selected from the moment it begins, so cycling
during a snapshot moves on from it, and omarchy-theme-set falls back to
the ordinary transition again when a shell an update could not restart
lacks the handover.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Order background changes with theme switches

omarchy-theme-bg-set takes the theme selection lock again, so a
background chosen during a theme switch waits for it instead of moving
the link back afterwards. A selection queued behind a reveal goes back
through omarchy-theme-bg-set, which finds its intro under the settings
as they are by then.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Replace the terminal screensaver with the native Wayland one
)

Entrances within 250ms, exits about two thirds of that, one strong
ease-out everywhere, windows popping from 95% rather than 87%, and no
linear or ease-in curves. The Quake console keeps its slide but
retracts on the same ease-out.

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* Add default touchpad gestures

Three fingers swipe between workspaces (empty ones included), pull down
the AI console, open the Omarchy menu, and move the active window with
Super held. Four fingers carry the active window to the neighbouring
workspace or toggle fullscreen.

Hyprland rejects a second gesture on the same fingers and direction, so
omarchy_default_gestures = false turns them off, and a migration sets it
for anyone who already has gestures of their own.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Put the AI console on three fingers up and the menu on down

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Drop the Super + three-finger window move gesture

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Find gestures anywhere in the hypr config and keep symlinks in the migration

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
The live probe edits files the shell watches: the catalogs, the
language selection and the plugin directory. It wrote them in place,
which truncates first, and Quickshell's FileView reloads on the
change, so a reload between the truncation and the write reads an
empty file. With an empty catalog or plugin directory,
t("test.hello") returns its key and .arg("Ada") logs "QString::arg:
Argument missing", which fails the final log check. CI hit this on a
loaded runner.

Write each file beside its target and rename it over, so a reload
reads either the old file or the new one.

In a VM with every vCPU kept busy, the test failed 10 of 40 runs on
that warning before this change and 0 of 40 after. Renaming only the
catalogs was not enough (13 of 40 failed): the language and plugin
directory files are watched too.
…acom#16887)

The test starts sleep with FCITX_CONFIG_HOME set and at once reads
the profile path from that process's /proc environ. Popen returns
when the child's close-on-exec pipe closes, early in the kernel's
exec, but the new environment is readable only once create_elf_tables
sets env_end (fs/binfmt_elf.c). Until then the read is empty
(fs/proc/base.c), refused, or the parent's, so the lookup falls back
to the caller's own $HOME and the assertion fails on
/home/<user>/.config/fcitx5/profile, as it did in CI.

On a 7.2 kernel, 842 of 2000 reads made right after Popen returned
lacked the child's environment, which appeared p50 26 us and p99 44 us
later. With the test's VM pinned to one host core, as vCPUs share a
core on a busy runner host, the test failed that way in 12 of 150
runs before this change and 0 of 150 after.

A real Fcitx owns its bus name long after its exec, so the shell's
lookup is unaffected; only the test needs to wait.
* Persist monitor scaling per monitor instead of on the catch-all

Scaling the focused monitor rewrote omarchy_monitor_scale, so Hyprland's
reload of monitors.lua applied the new scale to every screen. Save it as an
hl.monitor rule for that monitor instead, updating an existing one in place.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Leave monitor rules alone that scaling cannot safely edit

A multi-line catch-all or a scale given as an expression would have been
rewritten into broken Lua, so persisting now skips those configs instead.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Skip scaling persistence beside an unrecognized monitor rule and write atomically

A multi-line rule naming the focused monitor would have gained a competing
single-line rule, and rewriting monitors.lua in place let Hyprland's reload
catch it half written. Leave such configs alone, and build the new file
beside the real one before renaming it into place through any symlink.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
…om#16824)

* Install rtkit so PipeWire gets realtime priority

PipeWire requests realtime scheduling from RealtimeKit at startup, but
rtkit is not a dependency of Arch's pipewire packages and no rlimits
fallback is configured either. On a stock install every data-loop thread
runs at SCHED_OTHER nice 0, so audio produces xruns and crackling under
sustained CPU/GPU load (gaming, GPU inference, compilation).

Ship rtkit in the base package list, and repair existing installs with a
migration that installs it and restarts the running audio units so the
priority request succeeds without a reboot.

Fixes omacom#9390

* Install rtkit on fresh installs and restart the portal before PipeWire

omarchy-other.packages only seeds the ISO's offline mirror; the installer pacstraps omarchy-base.packages, so rtkit listed only in the other list never reaches a fresh install, and fresh installs mark this migration done.

xdg-desktop-portal reads rtkit's RTTimeUSecMax once, at startup. Restarting the audio units while it still reports 0 gave WirePlumber, and the shell as it reconnected, a hard RLIMIT_RTTIME of 0 with SCHED_RR data loops. Restarting the portal first gives them 200000 us like the daemon.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Drop the aarch64 rtkit entry now that the base list carries it

omarchy-pkg-defaults composes the base list first and de-duplicates, so the entry is redundant, and its comment saying nothing pulls rtkit in on aarch64 would mislead the next reader.

Co-Authored-By: Sam Diaz <265920095+szaidi-code@users.noreply.github.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Ask rtkit for realtime directly from PipeWire and JACK clients

PipeWire's client.conf and jack.conf load module-rt with the realtime
portal enabled. A portal started before rtkit was installed, or one
whose single read of rtkit at startup failed
(flatpak/xdg-desktop-portal#2110), reports a realtime budget of zero,
and module-rt applies it as a hard RLIMIT_RTTIME of 0 to the process,
which rtkit then accepts: the process's thread runs realtime and the
kernel kills it at its first RT watchdog check. On a 4.0.6 VM where
rtkit was installed and the portal could not be restarted, 4 of 17 new
clients died that way.

Set rtportal.enabled = false through module.rt.args, which module-rt
reads from the client's own config, as the stock pipewire.conf and
pipewire-pulse.conf do for the servers and Omarchy's speaker tuning
host does since omacom#12958. Clients then ask rtkit itself and take its real
budget. Without rtkit they get no zero limit. On a VM with the portal
made stale, a new client got 200000/200000.

WirePlumber is not covered: it creates its PipeWire context with no
config of its own and passes PipeWire only its context.modules list,
so a top-level module.rt.args in wireplumber.conf.d never reaches
module-rt. A drop-in there was tried and left WirePlumber at a hard 0
with realtime threads. Reaching it would mean replacing WirePlumber's
whole module list. Flatpak apps read their runtime's own PipeWire
config and are not affected either.

omarchy-settings copies etc/ wholesale, so these ship with no
packaging change.

* Add a command that installs rtkit and repairs zero realtime budgets

Without rtkit, xdg-desktop-portal reports a realtime budget of 0 and
PipeWire's client module-rt applies it as a hard RLIMIT_RTTIME of 0. The
shell is such a client, so everything it launches inherits the zero.
Once rtkit is installed it accepts any hard limit at or below its own
budget, 0 included, so a thread of such a process becomes realtime and
the kernel kills the process at its first RT watchdog check. A 4.0.6 VM
showed 44 such kills in the minute after omacom#12428's migration, from a
terminal opened before the update.

Raising the zeros has to come before the install. dbus-broker watches
its service directories and reloads when they change, so rtkit can be
activated as soon as pacman extracts it, before any install script or
ALPM hook runs. That rules out a dependency or a scriptlet.

omarchy-realtime-repair runs in two parts:

- Once per machine, as root, while rtkit is missing or an earlier run
  stopped after installing it: raise every logged-in regular user's
  hard zero (UID 1000 and up, so never root's or a system account's
  services) to rtkit's default 200 ms, install rtkit, raise the zeros
  to its live budget, restart each logged-in user's portal that still
  reports zero, and raise the zeros that appeared meanwhile. A
  root-owned marker under /var/lib/omarchy, written before the install
  and removed at the end, makes an interrupted run repeat this. A
  machine that got rtkit any other way, a fresh install included, has
  no marker.
- Every run, for the caller only: raise the caller's own zeros, restart
  the caller's portal if it reports zero, raise again, and restart the
  caller's audio units if the PipeWire server is not realtime yet,
  since PipeWire asks rtkit only at startup. The server is
  pipewire.service's main process: the speaker tuning host is also
  named pipewire and can be realtime while the server is not. Without
  zeros this needs no sudo, so a second user, or a non-admin one,
  finishes without it.

Each pass repeats until it finds no zero, so children forked meanwhile
are covered. Each target is checked again just before prlimit, by start
time, owner and limit, so a reused PID or a limit someone else has just
set is left alone. The machine-wide scan runs as root: with /proc
mounted hidepid a user cannot read other users' processes, so their
zeros would be missed (hidepid=2) or every run would fail on them
(hidepid=1). A user's own scan skips processes it does not own.

It never prompts. Under omarchy update sudo is already cached; without
it the command stops before changing anything and says the next update
finishes it. Anything it cannot inspect, the user list, a portal's
state, a live process's start time, owner or limits, PipeWire's main
process, is a failure rather than nothing to do. The start time is
read from /proc/<pid>/stat after the last ") ", so a command name with
spaces, parentheses or a newline cannot shift it.

The test runs the command against a fake proc tree with only the
system boundary stubbed.

* Install rtkit on existing machines through omarchy-realtime-repair

The migration installed rtkit while processes started without it still
carried a hard RLIMIT_RTTIME of 0, and it marked itself done when the
portal restart failed. Run omarchy-realtime-repair instead, which
raises the zeros before rtkit can be activated and fails, keeping the
migration pending, while anything is left that it could not repair or
inspect. A second run changes nothing.

* Raise new zero budgets even when a portal restart fails

A client can read a portal's old zero budget just before the portal
restarts. The repair raised such zeros in a final pass, but a later
portal that failed to restart exited the command before that pass, and
the client kept a zero with rtkit already installed. The same held for
the caller's own portal. Restart each portal in a subshell, always run
the final pass, and only then fail, so the migration still stays
pending.

Run as root, the command now does the machine-wide part alone: root has
no desktop session of its own, and omarchy-upgrade-to-quattro runs it
as root.

The test stubs id, so the caller is a regular user whoever runs the
suite; as root, the caller and root's own services were the same
account and a case failed. It passes as a regular user and as root, and
each new case fails with its fix removed.

* Repair realtime budgets before the 4.0 upgrade installs rtkit

omarchy-upgrade-to-quattro installs the base package list in the
running 3.x session, and the list now carries rtkit. rtkit can make an
app that took a zero realtime budget without it realtime, and the
kernel kills it, before the upgrade reaches its migrations. Those
migrations would also find rtkit installed and skip the machine-wide
repair.

Run omarchy-realtime-repair as root once the new omarchy package is in
and before any other package install: it raises the zero budgets,
installs rtkit itself, and repairs every logged-in user's portal. The
base list then finds rtkit installed, and the migration needs no sudo,
so an upgrade run as root does not stop on it. A failed repair stops
the upgrade before the base list, like a failed migration does.

* Keep rtkit masked through its own install

dbus-broker reloads its service directories as soon as pacman extracts
rtkit, so a client's realtime request can activate rtkit-daemon in the
middle of the transaction, before the sysusers hook has created the
rtkit user. It then fails "Failed to find user 'rtkit'" into systemd's
start limit, the repair's budget read failed and it exited before
restarting the portals, and rtkit came back about 10 s later with the
portals still handing out zeros. In VMs under load this hit 4 of 18
upgrade runs; in one, 79 portal clients holding a zero were made
realtime and killed.

Mask rtkit-daemon at runtime before the install, so no client can
activate it until the transaction is done, raise the zeros once more,
unmask it, reset a failed start, and wait for rtkit to answer before
restarting the portals. A trap unmasks it if anything fails, and a
runtime mask does not survive a reboot. A client that asked while it
was masked stays at normal priority rather than taking a grant.

The repair also:
- skips an empty logged-in user list instead of repairing user "";
- installs rtkit with sudo -n pacman rather than omarchy-pkg-add, which
  runs plain sudo and could prompt;
- always does the machine-wide part when run as root, so a user's own
  run after a root-run 4.0 upgrade needs no sudo even where rtkit was
  installed by hand;
- finds the caller's session bus under /run/user when its environment
  lacks it, as omarchy-upgrade-to-quattro's migrations run, and repairs
  no portal or audio units for a caller that has no session;
- skips /proc entries a user's own sweep cannot read, which hidepid=1
  shows but hides, instead of staying pending on every run.

* Follow the repository's Bash style in the realtime repair

AGENTS.md asks for unquoted variables and quoted string literals inside
[[ ]], and for full if blocks rather than an exit in one branch that
leaves the following statements unreachable. Apply both to
omarchy-realtime-repair and its test: the comparisons quote their
literals and not their variables, and the caller's own part moves into
repair_own_session, which the command runs in an if block unless it
runs as root, and which returns early when the caller has no session
bus.

The compared values are start times, UIDs and user names, which hold no
glob characters, so no comparison changes meaning. The test passes 40 of
40 as a regular user and as root.

* Say when rtkit stays masked after the repair stops

The repair masks rtkit-daemon for its install and an EXIT trap unmasks
it. When cached sudo expires while pacman runs, as it can for
omarchy-migrate run by hand, the trap's sudo -n fails too and rtkit
silently stays masked. That state grants nothing and kills nothing, and
the pending marker makes the next run with sudo unmask it first, as a
reboot does; but the user was not told. The trap now says so.

The test lets sudo expire after the install, checks that rtkit stays
masked, the migration stays pending and the message is printed, and
that the next run with sudo releases rtkit and finishes.

* Let one realtime repair run at a time

Each user's omarchy update holds only its own lock, under that user's
runtime directory, and omarchy-migrate takes none, so two users can run
the rtkit migration at once. Both mask rtkit and start the install; the
second one's pacman fails on db.lck, and its exit trap unmasks rtkit
while the first one's transaction is still extracting it, which is the
window the mask exists to close.

The machine-wide part now takes a flock on a root-owned file in /run,
which every user can open read-only, and a second run waits for the
first, then finds rtkit installed and finishes. sudo applies the
caller's umask, so under umask 077 the file would be root's alone and
no user could open it again; the mode is set explicitly, keeping the
inode the lock sits on.

* Say how to finish the realtime repair when zeros keep coming

An app that keeps starting PipeWire clients against a stale portal
gets a fresh zero every time, and after ten passes the repair stops
before installing rtkit. The error listed PIDs only; it now says to
close the apps that keep starting them and run the update again.

* Indent the rtkit PipeWire drop-ins with two spaces

The repository indents with two spaces, as its WirePlumber drop-ins do. PipeWire's config parser ignores the whitespace, so nothing changes at runtime.

---------

Co-authored-by: Zack Siri <me@zacksiri.dev>
Co-authored-by: Omarchybot <317366263+omarchybot@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Sam Diaz <265920095+szaidi-code@users.noreply.github.com>
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
…com#16896)

* Time the lock's first appearance once the scratch shell is quiet

The lock service answers while the shell is still drawing its first
frames of the wallpaper, bar, dock and notifications. A GPU draws them
in milliseconds. Mesa's software renderer, which a machine without a
GPU uses, takes a second or two, mostly in the buffer swap, and a lock
preview timed meanwhile waits for those frames: in a GPU-less VM on
GitHub's runners the fastest of three first appearances measured 587
to 1440 ms against budgets of about 1200 ms, and failed 10 of 18 runs.

Wait until the shell uses at most one clock tick of CPU in 200 ms, for
at most 10 s, before timing. The lock's own work starts only with the
preview request, after the wait. In the same VMs the fastest first
appearance is then 42 to 67 ms, on EPYC 7763, EPYC 9V74 and Xeon 8573C
hosts, and 0 of 18 runs fail; the laptop the budget came from measured
74 ms.

* Fail the lock test by name when the scratch shell dies while settling

A shell that exited during the quiet wait read as quiet: its stat
could not be read, every sample came back 0, and the test went on to
time a preview that could not appear, failing one step later under
the wrong name. A sample that cannot be read now prints the scratch
log and fails as the shell having stopped.
GDK_SCALE is one value for the session, and taking it from whichever monitor
was scaled last oversized XWayland apps on every other screen, such as 4x apps
on a 2x laptop panel after scaling an external display up. Follow the lowest
scale instead, so they come out slightly small rather than too big.

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
emirb and others added 3 commits October 28, 2026 23:04
The lock service answers once the shell has built its plugin
services, QML work on the shell's main thread. Its budget scaled with
process_speed_percent, the cost of 100 fork/execs, which does not
track that work: in KVM guests on GitHub's runners, Xeon 8370C and
8573C hosts start processes 25 to 40% faster than EPYC hosts but
build the shell no faster, so their budget came out at about 2100ms
against a build of 2100 to 2300ms, and the check failed about half
the time there. A bash loop and quickshell --version tracked the
build no better.

The time until the shell first answers, once Quickshell has loaded
its config and before the plugin services, is the same kind of work
on the same thread. The view build took 2.6 to 3.2 times it on EPYC
7763, 9V45 and 9V74 and Xeon 8370C, 8573C and 6973P-C hosts, so the
budget is now four times the fastest of those answers. Across 47
runs on 12 runners the build used 66 to 78% of that budget on every
CPU, and a lock service that spends an extra second building failed
it 48 times out of 48. A slower start before the first answer moves
both and is left to bar-startup-budget-test.sh.
The leak checks counted /proc/$$/fd with find inside a command
substitution. Bash makes the substitution's pipe, forks, and closes
the pipe's write end in the parent only afterwards (subst.c:7208,
7251 and 7425 in bash 5.3), so a find that runs before that close
counts the write end as well. In full-suite runs the test then failed
one of its "leaked FD" checks now and then while the same five
descriptors were open before and after: 2 of 9 runs in VMs, and one
of three in CI.

Count with a glob of /proc/$$/fd in the test's own shell instead. In
a VM, 20000 counts the old way read 6 instead of 5 eleven times, and
two more under load; the glob read 5 in all 40000. The test passed 20
of 20 runs and still failed on a descriptor deliberately left open.
@cursor
cursor Bot marked this pull request as ready for review October 28, 2026 23:07
@mark-groves
mark-groves merged commit 17a9e31 into quattro Oct 29, 2026
1 check passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

8 participants