Skip to content

Backport 44 bug fixes from quattro to v4-5-0 - #17386

Merged
ryanrhughes merged 45 commits into
v4-5-0from
backport/v4-5-0-quattro-fixes
Nov 4, 2026
Merged

ryanrhughes merged 45 commits into
v4-5-0from
backport/v4-5-0-quattro-fixes

Conversation

@omarchybot

@omarchybot omarchybot commented Nov 2, 2026 •

Copy link
Copy Markdown
Collaborator

Backports 44 bug fixes that landed on quattro, plus a fix to the dropdown test they bring, after it split from v4-5-0. Each one is cherry-picked with -x from its quattro commit, in the order it landed there.

All 44 applied without hand edits. Three needed a three-way merge (#13916, #14183, #17284), and their resulting diffs match the originals line for line. #17284 builds on #17280, which is also included.

Left out on purpose: new features, such as the native Wayland screensaver and the Neovim theme transparency options (#6451), and fixes to things that exist only on quattro: the rtkit follow-ups #17103 and #17096 (the rtkit change #16824 isn't on v4-5-0), the background intro fixes #17082 and #17179, and the display panel fix #17229.

Fixes

System and hardware

Desktop and shell

Accounts and integrations

Testing

./test/shell on this branch: 408 of 413 files pass. The five that fail behave the same on plain v4-5-0:

gb10-sleep-test.sh, which covers the GB10 idle fix, passes. The GB10 fix needs no new packages: it uses bash, udev and the existing omarchy-hw-aarch64-gb10. #15076 renames the Dropbox dependency to libappindicator, which the rc Arch mirror carries (12.10.1-2).

codex review --base origin/v4-5-0 found nothing in the fixes. It flagged two problems in the dropdown test brought in with #13916 and #6725: it doesn't clear QT_QPA_PLATFORMTHEME, and on Qt 6.12 an unqualified Color keeps the 2px and asymmetric border cases from changing the border. The last commit here fixes both, the same change as in #17388 for quattro. All 56 dropdown cases pass with the real borders.

The other test failures listed above are fixed in a separate backport of #17388.

🤖 Generated with Claude Code

https://claude.ai/code/session_01CY6YBoMt9VUxZ6fJqSXaj1

omarchybot and others added 30 commits November 2, 2026 21:25
pgrep prints the PID of every match, so omarchy-font-set wrote the Ghostty or Foot PID to stdout whenever one was running. Only its exit status is wanted here.

Co-authored-by: Omarchybot <317366263+omarchybot@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
(cherry picked from commit 11312b4)
…5910)

* Answer the webcam menu guard from udev instead of opening cameras

omarchy-hw-webcam ran v4l2-ctl over every video node, opening and powering up
each camera whenever the menu opened. udev already records each node's V4L
capabilities; read that, and fall back to asking the devices only for a node
udev has no record of.

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

* Read every udev record before asking the webcams

A node udev had no record for sent the guard straight to
omarchy-capture-webcam-list, which opens every camera, even when a
later node's record already answered. Check all recorded nodes first
and ask the devices only when none of them is a capture node and some
had no record.

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

* Trust udev only for a real camera's capture record

udev records a node's capabilities once, when it appears. A v4l2loopback camera with exclusive_caps=1 (OBS's virtual camera, the v4l2-relayd bridge) switches between Video Output and Video Capture as its producer starts and stops, so its record can say either: taken as the answer, an output record hid the webcam recording row while the camera streamed, and a capture record kept it after the stream stopped. Only a capture record from a node with a parent device answers now; a virtual node, and any node without a capture record, goes to the live query as on quattro.

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

* Answer from udev only for a camera this user can open

omarchy-capture-webcam-list enumerates through v4l2-ctl, which skips a node it cannot open read-write, so the picker finds nothing on a machine whose only camera is locked down (a udev rule setting MODE=0000, a node outside the session's uaccess). udev's record says nothing about access, so the guard answered yes there and the menu offered a row that could not work. A node this user cannot read and write is now left to the live query, which answers as quattro did.

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

---------

Co-authored-by: Ryan Doupe <ryan.doupe@gmail.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>
(cherry picked from commit 2afe858)
… wake (#15908)

* Keep the keyboard backlight level across a second blank before a wake

The lock blanks with `omarchy-brightness-keyboard off`, which saved the current level with `brightnessctl -s` before setting 0, and wakes with `restore`. Since the lock re-arms its blank timer after a resume (and when a screen arrives) without waking in between, a second blank saved the 0 the first one had set, and unlock restored 0.

`off` now records the level only when nothing is already saved, and `restore` consumes it, so the level from before the first blank survives any number of blanks. Blank and wake run as separate processes, so both take a lock; a level that cannot be read is left lit, and a restore the device rejects keeps the level for the next wake.

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

* Treat an empty keyboard backlight record as no record, and test the lock

A save that creates the record and then fails to write it, as a full runtime directory does, left an empty file. off checked only that the file existed, so it blanked anyway, and restore discarded the empty record without restoring: the keyboard stayed dark. off now removes a record it could not write and blanks only over a non-empty one, and an empty record left behind is replaced by the next off.

The tests now hold a blank paused mid-save while a wake starts, which fails when the flock is removed, and check that a rejected restore leaves the backlight where it was.

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

* Check the blank over an empty record, and that the paused blank started

Without the first check, a save guard that left an empty record in place passed: off skipped the blank and restore found the light still on. Without the second, a blank that never reached its save let the overlap test run restore first.

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

* Silence and reap the overlap test's background jobs

docs/testing.md asks background fixtures to send their output elsewhere and be cleaned up on exit. The exit trap releases the paused blank and waits, since killing the backgrounded subshell would leave the script under it running against the deleted directory.

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>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
(cherry picked from commit 0c29fa5)
* Keep the flat edges of bar icons level

Scaled bar glyphs were drawn with the curve renderer, whose antialiasing
overshoots where a corner curve meets a straight edge at bar sizes. The
active workspace's rounded square showed it as a top edge that stepped up
at its right end. Distance-field rendering keeps the edge flat while
holding the same exact fitted size and center.

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

* Center bar icons on their actual ink

Qt reports a glyph's tight bounds in whole pixels. At bar sizes that adds
up to a pixel of ink that isn't there, so icons were fit to a box taller
than they are and centered on its middle rather than theirs. The active
workspace's square came out 10.8px instead of 12 and sat almost a device
pixel above the icons beside it. Measuring a hundred times larger keeps
the bounds to the outline.

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

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
(cherry picked from commit 3de3c32)
The knob was kept inside the track while the ticks and fill spanned its full
width, so a slider at 100% still showed track past the knob.

Co-authored-by: Erik Johansson <172146456+ejuro@users.noreply.github.com>
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
Co-authored-by: Omarchybot <317366263+omarchybot@users.noreply.github.com>
(cherry picked from commit 0ba7646)
Button's implicitWidth counted the label only while it was visible.
Inside a hidden parent Qt reports every child as not visible, so a
hidden Button collapsed to its padding. Plugins that measure labels
with a hidden Button, like Pigeon's topic tabs, then cut every label
down to an ellipsis. Check for text and icon instead.

(cherry picked from commit 33d6472)
* Keep one pending-migrations reminder instead of one per login

omarchy-migrate-notify runs at every login and sends a critical toast
while migrations are pending. Since notifications are restored from
live.json (#15932), the last login's reminder comes back next to the
new one, so a machine whose migrations stay pending gains one identical
card per login: 15 after 15 logins on a 4.5.0-rc2 VM, pushing the
first-run cards off the screen.

The freedesktop spec's transient hint asks the server to by-pass its
persistence. The shell now honours it: a transient card is shown but
never written to live.json, so it is not restored and leaves no history;
an update that turns a card transient takes its stored row out.
omarchy-notification-send gains --transient, and the reminder uses it,
since it is sent again at every login while it still applies.

* Take -p and --transient only as bare words in omarchy-notification-send

The option parser split --flag=value before matching, so the two flags
matched their valued forms too: --transient=false switched transient on,
and --print-id=1 printed the id. Flags take no value; a valued form is
now text wherever it stands, like any unknown word: the headline before
it, the description after it.

(cherry picked from commit 64833fa)
…mpty (#17077)

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
(cherry picked from commit 717175d)
…15076)

* Install and remove Dropbox's libappindicator under its current name

Arch renamed libappindicator-gtk3 to libappindicator, which provides and replaces the old name. The installer asked for the old name and left omarchy-pkg-add's post-install check to depend on pacman -Q resolving the provide, which older pacman did not do. The remover never matched it at all, because omarchy-pkg-drop compares exact installed names, so removing Dropbox left libappindicator behind.

Fixes #11726

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

* Leave libappindicator installed when another package needs it

Naming it beside the Dropbox packages put it in one pacman -Rns transaction, so a retained package that requires it (network-manager-applet, remmina, rustdesk) made pacman refuse the whole removal and leave Dropbox installed. Drop it separately, and only when nothing directly requires it; optional dependents do not block removal, hence -tt.

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

---------

Co-authored-by: Omarchybot <317366263+omarchybot@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Codex Medium <noreply@openai.com>
(cherry picked from commit d494d35)
* Ignore vendor hotkey keyboard devices

* Ignore vendor hotkey layout events

---------

Co-authored-by: Omarchybot <317366263+omarchybot@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
(cherry picked from commit 0ca79f2)
Inkscape draws its own GTK file chooser rather than asking the portal, so neither the portal rule nor the per-app dialog list in system.lua matches it. Left to Hyprland, the chooser opens at whatever size GTK remembers, which on a fresh install is larger than a 1280x800 screen, with Cancel and Open off the edge, and GTK keeps that size for next time.

Fixes #11995

Co-authored-by: Omarchybot <317366263+omarchybot@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
(cherry picked from commit f77ed36)
omarchy-hibernation-available measured swap and looked for the resume hook, but never asked the kernel whether it would take a hibernation request. A machine where hibernation is disabled -- lockdown, nohibernate, secretmem, an active CXL device -- still got the Hibernate menu row, and selecting it ran `systemctl hibernate` detached with its output on /dev/null, so logind's refusal was invisible and the machine simply stayed awake.

The kernel lists "disk" in /sys/power/state exactly while it will accept the request; the same predicate is what makes /sys/power/disk read [disabled]. Raising a test machine's lockdown to integrity reproduced both readings and the false success.

An image size that exists but does not read was a second way to say yes: named bare in the arithmetic comparison it counts as zero, and any swap at all beats zero.

The sysfs, swaps and resume-config paths take an override the way omarchy-power-present does, so the test can construct a kernel that refuses.

🤖 Generated by Opus 5 in Claude Code. Reviewed by Codex XHigh.

Co-authored-by: Codex XHigh <codex@openai.com>
(cherry picked from commit c97617b)
Stopping a recorder this script did not start, with no saved pid and no recording file, still finalized and posted "Screen recording saved" for an empty name: a broken "-preview.png" thumbnail and a click that opens nothing. Now that the whole script runs under one lock, that is the last way the empty-name toast from #11508 is reached.

Co-authored-by: Omarchybot <317366263+omarchybot@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Codex Medium <noreply@openai.com>
(cherry picked from commit 0fc420c)
* Stop reloading a local plugin on its own Python bytecode writes

The local plugin watcher reloads a plugin on every write under its directory. A plugin whose Python helper imports modules from that directory makes CPython write __pycache__/*.pyc, each write reloads the plugin, the reload restarts the helper, and the helper's next lazy import writes the next .pyc. The 150ms debounce coalesces a burst but cannot break that cycle, since each write comes from the helper the previous reload started.

Ignore paths under __pycache__ (which covers the temp name CPython writes before renaming) and stray .pyc/.pyo files. Hidden files inside a plugin are still watched: entry points may live there, and the manual promises a reload for any saved file.

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

* Ignore only __pycache__ below a plugin, not names that look like bytecode

Matching .pyc/.pyo and __pycache__ against the whole plugin-relative path also matched plugin directories themselves: moving in a plugin whose id is acme.pyc (a valid id) was ignored, as was every change to a plugin dir named __pycache__. Python 3 writes bytecode only into __pycache__, so the suffix check was never needed for the reload loop; drop it and require __pycache__ to sit inside a plugin.

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

---------

Co-authored-by: Omarchybot <317366263+omarchybot@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Codex Medium <noreply@openai.com>
(cherry picked from commit c726e6a)
…16243)

* Keep bar tooltips hidden until their surface has the output's scale

Each time a tooltip opens, Qt maps a fresh popup surface at a whole-number scale (2 on a 1.6 output, then 1) before the compositor's fractional scale arrives, and the first frame is drawn at the wrong size. Hold the bubble at opacity 0 until the popup's devicePixelRatio matches the bar's, then reveal it on the next turn, the way OverlayWindow already does for overlays.

Fixes #14673

* Keep a revealed bar tooltip shown through Qt's transient scale

At a whole-number scale the popup maps at the bar's ratio, draws its first frame, then dips to 1 and back. Resetting revealed on that dip hid the bubble for two to five frames after it had appeared, a blink the base branch does not have. Reset it when the popup hides instead, on visible itself: tied to scaleReady, a hide during the dip would leave it set for the next show.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Codex Medium <noreply@openai.com>
Co-Authored-By: Muse Spark 1.3 <noreply@muse.dev>

---------

Co-authored-by: Omarchybot <317366263+omarchybot@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Codex Medium <noreply@openai.com>
Co-authored-by: Muse Spark 1.3 <noreply@muse.dev>
(cherry picked from commit 72cb55e)
A load bigger than the adapter drains the battery while it still reads 100%, and the AC-at-100% fallback then said "Fully charged" and hid the live draw and the time left. The fallback now holds only when the battery is not discharging, or its draw is idle.

Co-authored-by: Omarchybot <317366263+omarchybot@users.noreply.github.com>
Co-authored-by: Codex Medium <noreply@openai.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
(cherry picked from commit f53e0b2)
A sender named in several words never matched its window: "Telegram Desktop" became "telegramdesktop", which no class equals, and the omarchy-hyprland-focus-app fallback reads the raw name as a regex. That path is every click on a replayed or restored card, which has no live sender to run its default action, so clicking a replayed Telegram notification focused nothing. Such a name now also matches a class ending in its words, as org.telegram.desktop ends in telegram desktop. The helper keeps its regex argument, which omarchy-launch-browser relies on.

Co-authored-by: Omarchybot <317366263+omarchybot@users.noreply.github.com>
Co-authored-by: Codex Medium <noreply@openai.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
(cherry picked from commit 85241e9)
* Notice images overwritten in place in the image picker cache

The fast row-cache signature held only each directory's mtime, which an in-place save never changes, so a cache hit kept serving the old thumbnail. It now lists each file's size and mtime with one find per directory, so the check still costs a single process per directory.

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

* Say what keeping failed-video rows uncached actually buys

Rebuilding the rows does not retry a video whose .failed marker still matches its key; only a repaired file, whose size or mtime gives it a new key, is converted again. The comment said the rebuild itself retried it.

Co-Authored-By: Codex Medium <noreply@openai.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>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
(cherry picked from commit 1ee5fd6)
Off set 6500K, which hyprsunset still renders as a tint (green 0.9965, blue 0.9806), so the screen never got back to the untouched state the shipped identity profile starts it in. Off now also sends `identity`, from both the toggle and the shell service; hyprsunset keeps reporting 6500K, which --status and the bar already read as off.

The toggle also decided on and off by `== 6500` while --status uses `< 6000`, so with hyprsunset at its own 6000K default the first toggle set 6500K instead of turning night light on. It now uses the same threshold as --status.

Co-authored-by: Omarchybot <317366263+omarchybot@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Codex Medium <noreply@openai.com>
(cherry picked from commit 0857083)
The system panel parsed top's summary line in the user's locale. A decimal-comma locale prints idle as "84,3", which deleting commas turned into 843, so es_ES showed CPU at -743%; a locale that translates the labels (de_DE prints "%CPU(s)") gave no reading at all. Wherever the label matched, a fully idle machine read as 100%, because top writes "ni,100.0 id" with no space and deleting the comma glued the two fields together.

Running top under LC_ALL=C pins both the labels and the decimal point, and overrides an LC_ALL the user set, which LC_NUMERIC alone would not. With commas then only separators, they become spaces rather than nothing.

Fixes #12315

Co-authored-by: Omarchybot <317366263+omarchybot@users.noreply.github.com>
Co-authored-by: Codex Medium <noreply@openai.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
(cherry picked from commit c304026)
* Remember a DDC display whose brightness cannot be read

A display ddcutil detects but which does not answer getvcp 10 had its cache entry deleted on every failed read, so each call re-ran the full detect scan and the read again. Some monitors drop and re-train their link when probed, so with the monitor panel polling they blanked over and over. Cache the failure as unavailable for the same 60 seconds as a display detect does not find.

Fixes #6855

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

* Retry a known DDC bus after one failed read

Caching every failed read as unavailable also caught a monitor that normally answers: one flaky getvcp from a panel poll hid its brightness control and turned its brightness keys into no-ops for 60 seconds. A failed read on a bus whose cache entry carries the range, which only a successful read writes, now drops the cache, as before, so the next call detects again; any other failed read is left alone for the window. The bare bus find_bus writes before reading does not count, so a caller overlapping a first probe cannot erase the negative cache it leaves. The display in #6855 never has a range cached, so it is still probed at most once a minute.

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

---------

Co-authored-by: Omarchybot <317366263+omarchybot@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Codex Medium <noreply@openai.com>
(cherry picked from commit 37b1f3a)
)

* Stop the Dropbox status poll from voiding the pending login link

`dropbox-cli status` asks an unlinked daemon for a link URL every time, and the daemon answers each ask with a fresh nonce, so the widget's poll replaced the URL Log in had just opened while the user was still signing in. While no account is linked, read whether the daemon is running from its pid file instead, the same check dropbox-cli makes.

Fixes #12637

* Cover a stale info.json naming a missing Dropbox folder

info.json can outlive an unlink, so the poll's guard has to rest on the folder existing, not on info.json naming an account; a guard on the account path alone would bring the nonce rotation back while Log in is showing.

Co-Authored-By: Codex Medium <noreply@openai.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>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
(cherry picked from commit b628987)
…#15007)

* Open a new terminal in HOME when the focused window is not a terminal

omarchy-cmd-terminal-cwd took the newest child of the focused window and used its directory whenever its executable was a shell, without asking whether the window was a terminal at all. A browser keeping a helper shell in ~/Downloads therefore sent SUPER+RETURN, and Files, to ~/Downloads. A terminal's shell leads its own session on the terminal's pty, which every terminal Omarchy supports sets up before exec'ing the shell; a helper some other app spawned does neither, so the child must now be a session leader with a controlling tty before its directory is used.

The test's stand-in terminal now runs its shell under script(1), which gives it a pty and session as a real terminal does, and a stand-in browser keeps a helper shell in Downloads. Each fixture is waited for until its shell is actually running in its directory, so neither assertion can pass on a fixture that never started.

Fixes #10840

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

* Test an app sharing a terminal's pty, not only a detached one

A browser started from a terminal's shell inherits that shell's pty, and so does its helper shell, so a controlling tty alone does not tell a terminal's shell from a helper. The new fixture runs a stand-in browser under script(1) and asserts its helper still gets HOME; it fails if the check is reduced to tty_nr != 0, which the existing fixtures did not catch when the suite runs without a pty.

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

---------

Co-authored-by: Omarchybot <317366263+omarchybot@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Codex Medium <noreply@openai.com>
(cherry picked from commit 65f629e)
…4549)

Hyprland carries a floating window to another monitor at its offset from the old monitor's origin, so a window popped out with Super+O lands off-center there, or partly off a smaller or portrait screen. A dispatcher move fires window.move_to_workspace before the window lands, so the center is deferred a loop; a mouse drag has already landed on its monitor by then and keeps where it was dropped.

Co-authored-by: Omarchybot <317366263+omarchybot@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Codex Medium <noreply@openai.com>
(cherry picked from commit 60a912b)
…tream (#14585)

With nothing playing, the media service fell back to the first player holding a PipeWire stream, even one the bar refuses to show. An idle Chrome that once played media keeps a Stopped MPRIS session with only art and length, so pausing another player handed it the selection and the widget vanished with nothing left to click. The idle fallback now prefers a player with a title or artist over one without, unless the hidden one is the player the user last chose, so cycling sources onto it still works. It moves into MediaModel.js so the tests exercise the selection itself.

Co-authored-by: Omarchybot <317366263+omarchybot@users.noreply.github.com>
Co-authored-by: Codex Medium <noreply@openai.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
(cherry picked from commit 31020ac)
* Let the Bitwarden extension popout keep the size it asks for

The extension window was tagged floating-window, which forces every window carrying it to 875x600. Bitwarden opens its popout at the width chosen in its settings (380, 480 or 600) and a height of 630, so the forced size left it the wrong shape. Float and centre it directly instead, keeping no_screen_share.

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

* Match the Bitwarden extension popout in every Chromium browser

The rule matched only chrome-<id>-Default, so the popout in Brave (brave-<id>...), in a second Chromium profile (-Profile_1) or with the popup path appended (__popup_index.html) got none of it: no float and no screen-share protection, and Brave tiles it into a wide, half-drawn window. Matching on the extension ID alone covers them, as #8341 does. The test is adapted from #8341's, keeping its float and screen-share checks and asserting that no size or size cap is set.

Co-Authored-By: Ty Richards <me@tyrichards.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Drop the Bitwarden rule test

The test restated the two declarative rules in bitwarden.lua and counted them, so it guarded nothing a reader of the file cannot see and broke whenever a rule was added, which AGENTS.md asks tests not to do.

The broader `.*nngceckbapebfimnlniiiahkandclblb.*` match stays: the popout's class is `brave-...` in Brave and ends in `-Profile_1` in a second Chromium profile, and on quattro those popouts tile because the old exact class never matches them.

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>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Ty Richards <me@tyrichards.com>
(cherry picked from commit ab85004)
Hyprland 0.56's keybind:unbind() removes every bind on the handle's chord, not just the handle's own, so closing the region picker took any user bind on RETURN, CTRL + RETURN, TAB, CTRL + TAB or the arrows with it until the next reload. The picker's binds are now created once per config load and switched on and off with set_enabled(), which leaves other binds on the same chord alone. They carry no description, so the keybindings menu does not list them while they are off.

Any unbind of one of those chords expires the cached handle, and in 0.56 every handle method but tostring() dereferences an expired handle and aborts the compositor. So each handle is checked through tostring() before use and re-created when it has expired.

Fixes #12546

Co-authored-by: Omarchybot <317366263+omarchybot@users.noreply.github.com>
Co-authored-by: Codex Medium <noreply@openai.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
(cherry picked from commit 0a1dd20)
Claude-Session: https://claude.ai/code/session_01DMKCCA2jSkTTwEwdZepjQd

Co-authored-by: Claude <noreply@anthropic.com>
(cherry picked from commit 2f1a2c5)
manuaudio and others added 14 commits November 2, 2026 21:25
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
(cherry picked from commit a3c3a6a)
* Preserve terminal font size when changing font family

omarchy-font-set rewrote the Foot font line with a hardcoded size=9,
silently undoing the size chosen by omarchy display text size. Now the
existing :size=N is kept (falling back to 9 when absent), matching the
behavior of kitty, alacritty, and ghostty.

Fixes #6957

* Resolve the test font deterministically via fc-match

* Harden font size test for minimal environments

* Keep the primary Foot font's size when the line lists fallbacks

Foot's font= takes a comma-separated fallback list, and the greedy match took the last :size= on the line, so a fallback's size replaced the one omarchy-display-text-size had set on the primary font.

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

* Stub fc-list in the font-set size test

The test picked a family with fc-match under the real HOME and then had omarchy-font-set validate it with fc-list under the temporary one, so a font installed only in the user's home, or fontconfig with no fonts, failed the test for reasons unrelated to size. A fixed family answered by a stub makes it hermetic and removes the skip that reported as a pass.

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

---------

Co-authored-by: David Heinemeier Hansson <david@hey.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Codex Medium <noreply@openai.com>
(cherry picked from commit 09ed48f)
#14103)

* Leave a dangling shared symlink alone when setting up an agent account

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

* Only skip a dangling shared link, and test failed-setup cleanup

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

* Test failed account setup error

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Omarchybot <317366263+omarchybot@users.noreply.github.com>
(cherry picked from commit f51ecf1)
* Paint symbolic tray icons in the bar foreground through their alpha

MultiEffect's colorization multiplies the source's luminance by colorizationColor, so a symbolic icon with a dark fill (Adwaita's #2e3436, Cryptomator's #222222) stayed dark on a dark bar. Fill a rectangle with the foreground and mask it by the icon instead, so the icon's own colour is discarded. The mask ramps smoothly over the icon's alpha (threshold 0.5, spread 1.0) rather than the default hard cutoff, which turned every anti-aliased edge pixel opaque.

The mock tray item takes its icon from the environment so the new test can publish a dark symbolic icon through a real StatusNotifierItem. The test grabs the icon itself, checks every opaque pixel is the foreground and a corner stays transparent, and skips when the output is asleep, since a grab then never returns.

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

* Check that a symbolic tray icon keeps soft edges

The mask ramps over the icon's alpha so anti-aliased edges stay partly transparent, but the test only looked at opaque pixels and one corner, so a hard cutoff that fattens the icon still passed. Require at least one partly transparent pixel in the grab.

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

---------

Co-authored-by: Omarchybot <317366263+omarchybot@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Codex Medium <noreply@openai.com>
(cherry picked from commit b94447f)
cyclenext walks Hyprland's z-order vector and bringactivetotop rotates the
active window to the back of that same vector, so raising on the way back put
the window just left exactly where "previous" looks next and focus ping-ponged
between two windows. Lowering the window being left instead of raising the one
being landed on keeps the newly focused window on top and still steps one window
per press, so the chord stays the inverse of Alt + Tab.

(cherry picked from commit 1f7c8c5)
…igration (#13432)

* Undo Compose's $$ escaping when migrating a legacy Windows VM password

The legacy compose under ~/.config/windows holds the password as Compose reads it, so a $$ there reached the guest as $. Migration copied the raw text into the RDP credentials and re-escaped it for the new compose, leaving RDP and the new container environment with $$ while the existing guest, whose disk is preserved, still has $. Decoding $$ at the migration boundary keeps both on the password the guest actually has. This also makes #4882 safe to land on dev, which writes every $ in the password as $$.

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

* Undo the legacy password's YAML escapes too when migrating

Compose parses the legacy PASSWORD as a YAML double-quoted string before it interpolates $$, so a legacy "a\\b\"c$$d" reached the guest as a\b"c$d. Decoding only $$ left \\ and \" in the RDP credential and the rewritten compose. unescape() already reverses exactly these layers for every other read of the compose, so use it here.

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

---------

Co-authored-by: omarchybot <omarchybot@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Omarchybot <317366263+omarchybot@users.noreply.github.com>
(cherry picked from commit 51b348b)
#14684)

* Use the latin1 console keymaps for German and French in the setup form

The kbd `de`, `fr` and `fr_CH` keymaps are the old 7-bit national variants: `de` puts @ [ ] where ü ö ä belong, `fr` puts { } | where é è ù belong, and `fr_CH` leaves the accented keys unassigned. The installer loads the chosen keymap on the console, so a German or French user could not type their own letters in the name or password fields. The `-latin1` maps are the standard layouts, and systemd derives the same XKB layout from them, so the installed desktop is unchanged.

Fixes #8680

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

* Test that both French entries take their latin1 console maps

Only German was asserted, so reverting French or French (Switzerland) to `fr` or `fr_CH` passed the suite. Folds in the French assertion from #15003, extended to French (Switzerland), which only this pull request changes.

Co-Authored-By: Omarchybot <317366263+omarchybot@users.noreply.github.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>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
(cherry picked from commit 4157e42)
…6385)

* Count Codex sessions it has compressed to .jsonl.zst

Codex 0.163 compresses a rollout it considers finished to .jsonl.zst and deletes the plain file, so the scan's *.jsonl glob found nothing and every local stat in the Codex tab stayed at zero. Read both, through Python 3.14's compression.zstd.

Codex publishes one form before removing the other, in both directions, so for a moment a session exists twice; the plain file is the one Codex reads, and the only one counted.

Co-Authored-By: Dennis Gram <48596094+dennisgram@users.noreply.github.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Fail the compressed-session tests when their fixture is not in place

The script runs without set -e, so a zstd call that failed left the fixture in the wrong form and the assertion after it passed against quattro's collector, or without the both-forms filter. Each fixture now checks it has the files it means to test.

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

* Read compressed Codex rollouts through collectors.py's guarded zstd import

collectors.py already imports compression.zstd under a guard, leaving zstd None on a Python without it, as DeepSeek's reader expects. A second, unconditional import made that guard dead and would stop every collector loading on such an interpreter. Without zstd, a .jsonl.zst rollout is now one that couldn't be read, like a damaged one, rather than an error.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: muse-wonder High <noreply@meta.com>

---------

Co-authored-by: Omarchybot <317366263+omarchybot@users.noreply.github.com>
Co-authored-by: Dennis Gram <48596094+dennisgram@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Codex Medium <noreply@openai.com>
Co-authored-by: muse-wonder High <noreply@meta.com>
(cherry picked from commit ada0b51)
* Restore gvfs after resume from a transient unit instead of a background subshell

The post hook backgrounded its gvfs-daemon restart, but systemd kills everything left in systemd-suspend.service's cgroup when the unit finishes, so the restart never ran and /run/user/$UID/gvfs stayed unmounted after every resume. Where user sessions are frozen, the leftover child instead held up the thaw for about 90 seconds. Hand the restart to a systemd-run timer unit that lives outside that cgroup, and run it as the user with --uid rather than sudo, which goes through PAM and stalls on a fingerprint prompt nobody can answer.

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

* Fire the gvfs restart within a second of its delay

A transient timer inherits AccuracySec=1min, so the restart could land anywhere up to a minute after resume, leaving the FUSE gateway down that long. The test now also checks the scheduling arguments and fails if the hook itself fails.

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

* Say when the gvfs scheduling check had no user bus to check

Without a user bus under /run/user the loop asserts nothing, so report it as skipped rather than passed.

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

---------

Co-authored-by: Omarchybot <317366263+omarchybot@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Codex Medium <noreply@openai.com>
(cherry picked from commit 620fd0c)
Firmware that doesn't know the present rate leaves current_now in place
but fails every read of it, so the readable-file check passed, the read
came back empty, and the draw showed as 0W. Only take the sysfs reading
when it is a number, and otherwise keep UPower's rate.

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
(cherry picked from commit 4761d7b)
* Keep GB10 CPU cores out of deep idle states

The GB10 firmware's ACPI _LPI table understates how long its deeper idle
states take to leave, so cpuidle parks idle cores there and every
cross-core wakeup costs ~300 us instead of ~10 us. On a DGX Spark that
made the desktop sluggish and Speedometer 3.1 score 10. A udev rule keeps
each core in state0 (WFI) on GB10 machines, with a migration for
existing installs.

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

* Apply the GB10 idle setting to the running cores

The CPUs are already added when the hardware setup runs on an image's
first boot, so the udev rule alone left deep idle on until the next
reboot. Write the running cores' idle states directly as well, which
also lets the migration drop its udev trigger. Fold the test into the
GB10 hardware test and check the cores' states instead of the rule text.

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

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
(cherry picked from commit bbfdfdd)
* Give the lock's password field its focus back after a VT switch

The NVIDIA driver forces a VT switch around every suspend, which takes the
keyboard away; Qt drops the field's focus with it and nothing restores it, so
typing after a lid-open resume went nowhere until the field was clicked.

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

* Hold off blanking the lock for longer after a resume

The countdown restarts as the kernel thaws, seconds before the panel lights
on NVIDIA, so the lock went dark again just as it appeared.

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

* Hold off the blank after a resume that found the lock already dark

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

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
(cherry picked from commit 1843592)
…7280)

The driver rebind only covers i2c_hid_acpi and intel_quicki2c, so on other
touchpads it did nothing. Cycling the device off and on in Hyprland first
makes libinput release a held button on every platform, without sudo.

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
(cherry picked from commit 901af70)
* Reset only the touchpad in omarchy-restart-trackpad

Rebinding every i2c_hid_acpi device also reset the keyboard and
touchscreen. Pick the device udev marks ID_INPUT_TOUCHPAD instead.

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

* Test that restart-trackpad resets only the touchpad

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

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
(cherry picked from commit 3e7f8ed)
@greptile-apps

greptile-apps Bot commented Nov 2, 2026 •

Copy link
Copy Markdown
Contributor

RetriggerView in GreptileConfidence Score: 5/5

[High impact] The follow-up changes appear safe to merge; the three earlier lower-impact findings remain open.

Findings

  1. P2 Image previews stay stale ▶
  2. P2 Shell behavior goes untested ▶
  3. P2 Tests omit required failure settings ▶

Summary

Backports 44 fixes for hardware, desktop controls, accounts, and integrations.

  • GB10 cores stay out of deep idle states.
  • The lock screen keeps focus and restores keyboard lighting after wake.
  • Power and display controls use live readings and finish their reset work.
  • The webcam menu checks recorded camera capabilities before opening devices.
  • Trackpad recovery leaves nearby keyboards and touchscreens alone.
  • Window shortcuts and placement follow the window's actual desktop context.
  • Bitwarden popouts keep their requested size and Inkscape dialogs float.
  • Hidden buttons keep the same measured width as visible buttons.
  • Dropdowns toggle on repeated clicks and size their popups to the rows.
  • Slider controls reach both ends of their track accurately.
  • Bar icons, tooltips, and keyboard indicators render at the right scale and color.
  • The media widget keeps a visible paused track ahead of hidden idle players.
  • Transient reminders stay visible without being saved in notification history.
  • Dropbox status checks avoid replacing a pending login link.
  • Agent accounts and session totals keep working with existing links and formats.
  • Changing a terminal font keeps its existing text size.
  • The image picker notices files overwritten in place.
  • Plugin reloads ignore Python bytecode cache writes.
  • German and French setup choices use their latin1 console keymaps.

Reviews (2) · Last reviewed commit: "Apply the dropdown test's border cases t..." · Reviewed by Greptile

Comment thread bin/omarchy-menu-images
# place leaves the directory untouched. One find per directory keeps it cheap.
for dir in "${image_dirs[@]}"; do
[[ -d $dir ]] && rows_fast_signature+="$dir:$(stat -Lc '%Y' "$dir")"$'\n'
[[ -d $dir ]] && rows_fast_signature+="$dir"$'\n'"$(find -L "$dir" -maxdepth 1 -type f -printf '%f:%s:%T@\n' 2>/dev/null)"$'\n'

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P2 Image previews stay stale

The overwrite fix still misses an image replaced with another of the same size within one second. The new fast signature sees fractional modification times, but rows_signature and thumbnail_for() still use %s:%Y, which keeps only whole seconds. The picker reuses the old rows and thumbnail, then saves the new fast signature, so reopening it does not refresh the preview.

Use the same timestamp precision throughout the cache keys and cover a same-size overwrite.

Comment on lines +107 to +109
rg -F '(temp === dayTemperature ? " && hyprctl hyprsunset identity" : "")' "$ROOT/shell/plugins/services/nightlight/Service.qml" >/dev/null ||
fail "nightlight service resets to the identity matrix when turning off"
pass "nightlight service resets to the identity matrix when turning off"

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P2 Shell behavior goes untested

The new check searches Service.qml for an exact command string instead of running the shell’s off path. It can pass even if runApply() is never called or fails to run the command. It also breaks when equivalent code is written differently.

AGENTS.md forbids tests that grep source for the string just added. Replace this with a behavior test that records the commands run when the shell turns nightlight off. This repository requirement must be satisfied before merging.

File Used: AGENTS.md (source)

Comment on lines +1 to +2
#!/bin/bash
source "$(dirname "$0")/base-test.sh"

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P2 Tests omit required failure settings

This new test and test/shell.d/hyprland-pop-test.sh omit set -euo pipefail. Failed setup commands can therefore continue into later checks and be missed or reported as a different problem.

docs/testing.md requires these settings before sourcing base-test.sh. Add them to both files before merging.

File Used: docs/testing.md (source)

The dropdown test set an unqualified Color.shellValues, which Qt 6.12
resolves to QtQuick's Color type, so the 2px and asymmetric border cases
ran on the default border. Use Commons.Color as shell-dev.md says. Clear
QT_QPA_PLATFORMTHEME too, so the GTK theme plugin doesn't try to open a
display under the offscreen platform.

(cherry picked from the dropdown half of commit 0a54d6f)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CY6YBoMt9VUxZ6fJqSXaj1
@ryanrhughes
ryanrhughes merged commit 0658da2 into v4-5-0 Nov 4, 2026
2 checks passed
@emirb emirb added this to the 4.5 milestone Nov 5, 2026
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.