Skip to content

update: use private logs to avoid update collisions [UPSTREAM #13937 OPEN] - #83

Open
kvnloo wants to merge 1218 commits into
quattrofrom
muse/update-log-claim
Open

kvnloo wants to merge 1218 commits into
quattrofrom
muse/update-log-claim

Conversation

@kvnloo

@kvnloo kvnloo commented Sep 26, 2026 •

Copy link
Copy Markdown
Owner

Promotion status

Already under upstream review: omacom/omarchy#13937. Do not open a duplicate. This fork PR is the downstream evidence carrier, not a new ready-to-post request. Origin title/description still need to be corrected in a GitHub session with permission to edit that PR.

Problem (verified on a stock Omarchy worker)

omarchy update historically sent the script(1) transcript to the fixed pathname /tmp/omarchy-update.log. Two different local accounts running updates sequentially can collide with one another's pre-existing log. On a stock ISO with fs.protected_regular=1, the later user gets Permission denied and the update does not start (#15679).

Correction to the original description: script(1) runs as the invoking user, not root. The cross-user root-file overwrite previously alleged here was not reproduced; with stock symlink/hardlink protections, the attempted cross-user redirection fails. This PR fixes log-file collision / update reliability, not a verified privilege escalation.

Current implementation

The current origin #13937 branch at d2dc937318546a05953a272dce6b1245418c202a creates a unique private mode-0600 /tmp/omarchy-update-log.* transcript for each invocation, preserves it on success and failure, prints the path and retains exec env … script lifecycle semantics. Test fixtures stay in a temporary private directory; the tests do not overwrite or remove the real shared update log.

Verification at the current head

Omarchy bot review, 18 October 2026 (see PR conversation for exact comment): stock-ISO two-account collision reproduced on quattro and eliminated at this head; focused update-log-mktemp-test.sh 6/6 pass, ./test/cli 125/125 pass, and related update, sudo-boundary and lock tests pass. The remaining older omarchy-update-analyze-logs fallback is documented in the review. Competing #15683 also addresses the same collision; maintainers choose one.

Ready-to-apply origin metadata handoff (requires authorized local gh)

Suggested origin title: update: use a private log to avoid cross-user update collisions

Suggested origin summary:

A fixed /tmp/omarchy-update.log blocks a second user's update when another user owns the file (fs.protected_regular=1). Use a unique mode-0600 transcript per update, keep and print its path, preserve exec and the existing update sequence. This is a reliability fix, not a root-file overwrite / privilege-escalation claim. Verified against a stock ISO with two users; regression tests 6/6 and CLI 125/125 pass. Fixes omacom#15679. Competes with omacom#15683; await maintainer selection.

Provenance

Existing original author and contributor QA history is preserved in commits, reviews, and previous comments. This work was AI-assisted under the contributor's direction; preserve all human reviewers' and original reporters' credit.

@kvnloo kvnloo added the muse label Sep 26, 2026

kvnloo commented Sep 27, 2026

Copy link
Copy Markdown
Owner Author

QA: HOLD / REBASE — the log-claim idea itself looks plausible, but this branch is pinned to fork quattro 28ceaae7 while live origin quattro is now d3fb0284.

That stale base makes the PR diff look much broader than its stated scope: current upstream already contains most of the privileged-Bash / update-order changes, with additional evolution since this branch. Please port only the claim_update_log behavior + focused regression onto the fresh origin head, then re-run the link tests there.

Acceptance for recheck: fresh-head diff is narrow, update flow outside log claiming is byte/semantically unchanged, and symlink + hard-link RED→GREEN still holds.

kvnloo pushed a commit that referenced this pull request Sep 27, 2026
Posted by Kevin's agent on his behalf — AI-assisted (Muse, Meta's Muse Spark).

Addresses QA HOLD on PR #83: instead of porting claim_update_log
onto current quattro, this adopts the strictly stronger fix already
shipped as muse/update-log-mktemp (b3b7887). Claiming the fixed
/tmp/omarchy-update.log path still leaves a shared pathname in play;
the mktemp approach removes it entirely: script(1) writes to a
private transcript handed through OMARCHY_UPDATE_LOG_FILE and
cleaned up afterwards, so there is nothing for another local user
to plant or swap. Rebased onto current quattro (was 28ceaae);
7/7 focused tests green.
@kvnloo
kvnloo force-pushed the muse/update-log-claim branch from c3be854 to 7591a4c Compare September 27, 2026 20:43
@kvnloo kvnloo changed the title [muse] update: claim the update log atomically before root script(1) update: claim the update log atomically before root script(1) [READY TO POST] Oct 1, 2026
@kvnloo

kvnloo commented Oct 1, 2026

Copy link
Copy Markdown
Owner Author

Upstream PR is open: omacom#13937

@kvnloo kvnloo added the keep pr-janitor: never auto-close label Oct 17, 2026
@kvnloo kvnloo changed the title update: claim the update log atomically before root script(1) [READY TO POST] update: use private logs to avoid update collisions [UPSTREAM #13937 OPEN] Oct 19, 2026
dhh and others added 20 commits November 11, 2026 11:55
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Chrome syncs a Google account out of the box, so the menu item that gave
Chromium Google's OAuth credentials, its command, its translations and its
FAQ answer go. The manual's translations catch up through the i18n sync.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Set up the SSH server before sharing a model
A stopped model stays shared and is served again when it starts, but its
line no longer says Shared until then. A running shared model can still
always stop sharing.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A model on another machine runs on its host's terms: the panel now offers
to disconnect from it (forget the host here), not to start or stop it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…onnect is host-wide

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…epeated type by its last line

Plain registers JSON, TOML, YAML, Lua, CSS, INI, desktop entries and
systemd units directly, and GLib picks a type's own handler before
falling back to text/plain, so those kept opening in Plain whatever editor
was picked. When a type is set twice, GLib uses the last line; the
migration now decides by that line too.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Files keeps its working directory on the folder in view, so Super+Return
and the other terminal-in-cwd launchers start there when it's focused.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Show a model as shared only while it runs
GLib merges application/xml and text/xml (and the YAML pair) into one
type, the last line among them winning, so a choice kept under one name
was overridden by a handler line under the other.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Removing yaru-icon-theme can leave /usr/share/icons/Yaru behind with the old
Nautilus arrow links and a stale icon-theme.cache. Remove the folder when it
no longer holds a theme and no package owns it, instead of picking off the
links and empty folders one by one.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Deleting the whole folder would also take files an administrator added there.
Match the two arrow links setup created and the icon cache it built, then
remove only folders that end up empty.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
omarchy-theme-set-gnome also resets the GTK theme and colour scheme, which a
user may have set themselves. Set just the icon theme.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Themes no longer pick icons: GTK and Qt get Adwaita on every switch, and
icons.theme and the helper that read it are gone.

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

The open panel surveyed the tailnet every 20 seconds; now every 2. A survey
that only moves when each model was last seen leaves the lines alone, so
the faster pace doesn't rebuild them.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
dhh and others added 30 commits November 16, 2026 19:50
…m#19152)

Their install migrations didn't redraw the icon set, so the launcher
showed each package's own icon until the next theme switch.

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
…acom#18629)

* Show the disk unlock prompt on the Lenovo IdeaCentre Mini 01Q8X10

* Name the IdeaCentre Mini's X Plus and Snapdragon X models
* Show the disk unlock prompt on the Lenovo IdeaPad Slim 5x

* Rebuild the Slim 5x's boot image unless this migration did already, and test the drop-in's board gate instead of its contents
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
…acom#18776)

* Let the DSPs start when a Snapdragon root disk is on a USB-A port

* Record the DSP blacklist rebuild as owed before the setup can lift it
Omarchy now builds linux-t2, its headers, apple-t2-audio-config,
apple-bcm-firmware-fetcher and t2fanrd in its own repository. T2
installs no longer need the extra repository the installer added for
them, so the installer stops adding it.

The migration moves existing T2 installs over once per machine. It
reinstalls every installed T2 package from [omarchy] by name while the
old repository is still configured, then drops that repository's
section and checks the result with pacman-conf. If [omarchy] does not
serve a package yet, or the reinstall fails, it changes nothing and
stays pending.
The kernel caps a GPU at half of RAM. An AMD APU like Strix Halo has
next to no VRAM, so on 64 GB the GPU can't load a 38 GiB local model.
Raise ttm.pages_limit to all of RAM but 8 GiB on machines with an AMD
GPU and more than 16 GB, on install and through a migration that
rebuilds the boot image for the next boot.

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
vc4 scans out of the CMA pool, which must sit in the first GiB of RAM.
The Pi firmware's device tree reserves 64 MiB, and a 2560x1440 desktop
does not fit in it. Hyprland then logs "Couldn't allocate a gbm buffer"
and falls back to a lower mode.

512 MiB, what Raspberry Pi OS reserves, failed under UEFI. The EFI
stub places the kernel at a random physical address. On one boot it
landed at 0x08c60000, which left 472 MiB free in the first GiB. The
kernel logged "cma: Failed to reserve 512 MiB", vc4 could not set up
its framebuffer, and the disk unlock prompt never reached the screen.

256 MiB reserved on all 10 boots tested. On one of them the kernel sat
near the middle of the free span (0x15820000), and CMA still fit at
0x1a400000. A 1440p desktop uses 55.6 MiB of it. The board setup now
writes cma=256M into a Limine drop-in on a Pi 4 Model B.
The Pi 4 Model B's CYW43455 needs Raspberry Pi's patch file, which
linux-firmware lacks. Without it the Bluetooth controller runs its ROM
firmware and reports the placeholder address AA:AA:AA:AA:AA:AA.

The board setup now installs raspberry-pi-4-bluetooth-firmware, which
puts the file under the name btbcm looks for first on this board. The
package is listed in omarchy-other.packages, so the ISO carries it for
offline installs.
* Move About into the System menu, above Screensaver

The root menu loses its About row; System now opens with it. The old
`about` route stays as an alias, and every translation moves with the key.

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

* Point the menu extension example at system.about

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

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
…omacom#19233)

* Wake the HP ZBook Ultra G3a 16 from sleep without an 80 second freeze

HP's Opal drive-unlock SMI, run when the SSD powers back on during
resume, has the SSD DMA into an ACPI NVS buffer the IOMMU doesn't map.
The command never completes and every core waits in SMM until the
handler times out. Run the IOMMU in passthrough on this model until a
BIOS fixes it.

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

* Give the ZBook sleep migration its own timestamp

1794503281 is already the Snapdragon power-profile migration, so users who
ran it would have skipped this one.

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

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
…com#19138)

* Give each local model a launcher of its own, with a loading show

Setting up a local model adds a launcher with its name and its vendor's
logo, as signing in to a subscription does, and removing the model takes
it away. The launcher opens an agent on the model and starts it first
when it isn't running, drawing the load in the terminal until it's up.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>

* Show only the percentage while a local model loads

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>

* Keep local-model launchers in step when syncs overlap or reads fail

Lock the whole sync so the wrapper and the panel can't remove each
other's launchers, keep launchers when the status can't be read, sync
on the panel's first reading so a setup that finished while the shell
was away gets its launcher, let the migration fail so it retries, and
fold the launcher tests into the existing agent tests.

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

---------

Co-authored-by: Claude Fable 5.2 <noreply@anthropic.com>
)

* Turn off herdr's own update check, since pacman updates it

Omarchy installs herdr as a pacman package in /usr/bin. herdr's
background version check still fetches herdr.dev/latest.json and, for
any install it doesn't recognize as Homebrew, mise or Nix, shows
"update ready" with the advice to run `herdr update`. That command
can't write /usr/bin and fails with "install directory not writable
... Try running with appropriate permissions", which points users at
sudo and a binary pacman no longer owns.

Set [update] version_check = false in the shipped config, and add a
migration that appends the same table to existing configs. The
migration reads the config with Python's tomllib, so quoted keys and
strings can't fool it. It leaves alone symlink-managed configs, any
config that already sets something under update, and any config that
doesn't parse. It reloads a running herdr server, which applies the
setting without restarting panes.

manifest_check stays on, so agent detection rules keep updating
between herdr releases.

* Leave a herdr config that isn't UTF-8 alone in the migration

tomllib raises UnicodeDecodeError for a config with bytes that aren't
UTF-8, such as a comment saved in Latin-1. That error isn't a
TOMLDecodeError, so the migration exited 1 under bash -e and stopped
every later migration. Treat it like any other config herdr can't
parse, and leave the file alone.
* Give every installed agent harness a launcher of its own

Selecting a harness as the default agent installs it, and now also adds a
launcher beside Agent that starts it in a terminal of its own, as signing in
to a Claude, Codex or Grok subscription already did. Each launcher ships in
default/applications with its icon drawn in the papercraft and arcade sets
and a packaged PNG. Hermes and OpenClaw bring launchers of their own apps.

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

* Draw Copilot's launcher icon from its real mark

The agents panel's Copilot logo is a broken drawing, so the icons drawn
from it came out as a mangled van. Use the goggled face from Simple Icons.

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

* Show Copilot's real mark in the agents panel

The panel's Copilot logo was a broken drawing that rendered as a mangled
van. Use the goggled face from Simple Icons, in white and in its dark
version for light themes.

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

* Add launchers for every agent harness already installed

The migration gave only the default harness its launcher. It now asks
omarchy-default-agent --installed about each harness that ships one, and
adds those missing, leaving launchers already in place alone.

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

* Fail a launcher that couldn't be written

The launcher command ended on update-desktop-database || true, so a copy
that failed still succeeded, and the migration adding launchers was marked
done without them. It now fails, and the migration stays pending.

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

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
…19243)

* Switch to an added agent account and say it's ready for more

Adding another subscription now makes it the account new sessions use,
since that's the one you want tokens from next. Running sessions stay
where they are. The notification reads "Claude Code is ready for more"
with "Tokens now flowing from <label> account", and clicking it opens
the agent like the first sign-in's does. Both stay up for 30 seconds.

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

* Leave the agent ready notifications at the default timeout

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

* Switch to recovered accounts too, and say when switching fails

An interrupted sign-in finished on retry now becomes active and gets the
same ready-for-more notification as a fresh one. If the choice can't be
saved, the account is still added and the result says new sessions
couldn't be switched to it, rather than claiming tokens flow from it.

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

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* Move hardware controls into omarchy device

omarchy-hw-* only detects hardware, so the controls get a namespace of
their own: device gpu hybrid, device touchpad, device touchscreen and
device keyboard toprow. The old names stay as hidden wrappers, so
bindings, menus and scripts that call them keep working.

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

* Drop the unreleased keyboard toprow alias

omarchy-keyboard-toprow never shipped in a release, so nothing calls it
by that name. Its config header names the new command too.

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

* Compare only the top-row options, so an older comment forces no rebuild

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

* Leave translated manual chapters to the translation sync

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

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* Let a hand-chosen agent account overrule autoswitch

`omarchy agent account use main` with auto switching on and main over the threshold switched new sessions to main, then the usage refresh it starts ended in autoswitch, which switched straight back. Every caller of `use` is a person choosing, so it now holds that account: autoswitch leaves it alone, and says nothing about it, until a check sees it under the threshold again.

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

* Keep the hold through a rename, a stale reading, and a lone account

Renaming the held account left `held` on the old id, so the next refresh switched away. A stale reading may predate the choice, so only a fresh one releases the hold. With fewer than two accounts autoswitch returned before it could release anything, so a hold outlived the account's reset and disabled switching once a second account was added.

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>
* Tell agents a theme is colors.toml and downloaded backgrounds

Small models wrote every generated file by hand and drew backgrounds
with ImageMagick. Say what Omarchy generates, list the palette keys and
contrast to keep, and give a working wallpaper search.

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

* Start Omarchy apps from a scaffold that passes its tests

The scaffold writes the starter from templates.md, the LICENSE and git,
then builds and tests it, so an agent edits a green project instead of
spending its context retyping 1,400 lines of templates.

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

* Show agents how to make a bar plugin from scratch

The plugin guide covered safety, motion and translations but not the
manifest or the widget, so agents reverse-engineered both from the
built-ins. Give a minimal plugin that validates and renders, and the
commands that check it and put it on the bar.

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

* Have agents pick theme backgrounds by eye from a contact sheet

A local model picked an anime collage for a tea house theme by judging
downloads from their brightness. Search general images only, show every
candidate on one labelled sheet, and check the palette against the
chosen background. Say why stock themes ship neovim.lua and vscode.json.

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

* Let agents look at the pages, apps, themes and plugins they build

Local models shipped broken layouts, empty windows and unusable theme
backgrounds because they never saw what they built, and the ones that
tried left browsers running or blocked on a window that never closed.
omarchy agent look renders each kind safely, without taking the focus or
changing the user's theme or bar, and reports the errors it finds.

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

* Point the theme, plugin and app skills at omarchy agent look

Each tells the agent to look at what it built with the command for its
kind, and to fix what it sees before calling the work done.

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

* Add a web-game skill for three.js games and toys

Asked for a game, local models built the obvious version, and one
crashed on a three.js method that doesn't exist. Give the shape of a
single-file game with a loop that survives slow frames, the polish that
makes it feel good, and how to check a run in a headless browser.

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

* Have agents lint QML before building and test APIs in test cases

A local model spent dozens of builds finding QML properties that don't
exist, one per test run, and compiled scratch files by hand that could
never find Qt's headers. qmllint names them all at once in a second.

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

* Let Omarchy apps draw themselves for omarchy agent look

Capturing a window on screen hangs while the display sleeps and flashes a window on the user's desktop. The starter's main.cpp now saves its window offscreen when asked, pressing keys through its Shortcuts first, so agents can check every mode headless. A patterns guide gives a verified two-pane layout with the traps local models kept falling into.

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

* Render plugin previews offscreen while the display is off

Nothing on screen draws while the display sleeps, so a preview waited forever. A widget without a popup panel renders offscreen; one with a panel says why it can't.

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

* Kill an app that ignores the end of its offscreen look

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

* Tell agents a new theme leaves out the published screenshots

A local model spent turns working out what unlock.png and preview-unlock.png are for, then copied them into its theme.

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

* Keep the scaffold building when the display name has a space

The display name also names the test class, so "Flash Cards" gave class Flash CardsTests and the untouched starter failed to build. Make it an identifier there, and show a failed build's errors rather than the warnings after them.

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

* Note the highlighted button label trap in the app layout patterns

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

* Tell agents game time runs slowly in the headless check

A local model took a low score after eight headless seconds for a bug in its game loop. The headless browser draws few frames and the loop caps each step, so only a score that doesn't rise at all is wrong.

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

* Say when a web game is done

Local models kept tuning their games until they ran out of time. Done is playable with most of the polish, each item checked once; then summarize. Keep the title screen's text on a panel the glow can't wash out.

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

* Say QSaveFile handles permissions on its own

The note about a QTemporaryFile arriving as 0600 read as a QSaveFile problem, and a local model spent an hour fixing permissions QSaveFile already keeps. Checked on Qt 6.12: a replaced file keeps its mode and a new one gets the umask's default.

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

* Show agents how to find a bar icon by name

A local model spent its time drawing the Nerd Font's glyphs to find a tomato. Look icons up by name in Nerd Fonts' table instead.

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

* Warn plugin authors off property names every item already has

A local model's pomodoro declared property State state, which collides with Item.state, so its session never changed.

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

* Name theme colors the preview can't read

The theme check drew gray for a color it couldn't read and skipped its contrast check, so a palette with `blue = "#GGGGGG"` passed with `problems: none` while showing a color the theme never set. Values are now read as omarchy-theme-color reads them, whole and between either kind of quote, so a suffix typo is caught and a single-quoted color is not reported missing.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Co-Authored-By: Codex Medium <noreply@openai.com>
Co-Authored-By: Muse Wonder High <noreply@meta.com>

* Press a look page key once and hold it for a frame

Each key was sent to the window, the document and the body, and the last two bubble, so a window listener heard one press three times. The release followed in the same tick, before any frame could read the key as held. A press now goes once to whatever has the focus and is released after the next frame, or just before the next key where none comes, since headless Chrome draws only a frame or two in a whole run. The release goes to whatever has the focus by then, since the press may have removed the element it landed on. The test stands Chromium in for Chrome on an image from before Chrome was the default.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Co-Authored-By: Codex Medium <noreply@openai.com>
Co-Authored-By: Muse Wonder High <noreply@meta.com>

* Refuse look page --keys for a URL

The presses are added to a copy of a local file, so for a URL they were dropped without a word and a game stayed on its title screen.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>

* Never report an earlier screenshot as a new look

With --out naming an image from an earlier look, a browser that failed left it in place and the command reported it as this run's screenshot.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>

* Keep Chrome's sandbox for look page wherever it can start

Every page ran with --no-sandbox, an outside URL included. Chrome refuses to start where it can't set its sandbox up, as inside an agent's own sandbox, which is the one place the flag is needed, so it is now added only after a blank page in a profile of its own also fails to draw. That decision rests on a run that loaded nothing of the page's, so a page can't get itself run unsandboxed by failing.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Co-Authored-By: Muse Wonder High <noreply@meta.com>

* Size an app that draws itself with look app --size

--size only reached the on-screen fallback, so a scaffolded app was always drawn at its QML size and an agent couldn't check a smaller window for clipping. The template's hook now resizes the window to OMARCHY_LOOK_SIZE. The value is checked first, since it is also written into the compositor dispatch.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>

* Start a program without the look hook only as a floating window

Every program was first run with QT_QPA_PLATFORM=offscreen in case it drew itself. One that isn't Qt ignores that and opened an ordinary window, tiled and holding the keyboard focus, for half a minute before being killed and started again floating. Only a program whose file carries the hook's name gets the offscreen run now, which means the built binary rather than a script that starts it. Withholding the display instead would also stop scaffolded apps, whose GTK platform theme needs one to start.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>

* Capture and close only the window look app started

The on-screen fallback took the first window to appear after the launch, so a window the user opened meanwhile was captured in its place and its program killed, while the launched program was left running. The launched shell now records its pid, and only a window owned by that process or one it started is captured and closed; the program is ended whether or not a window came. The capture is of the window alone: the fallback to its area of the screen is gone, since that shows whatever is there by then, another workspace included, and grim captures a still or hidden window directly.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Co-Authored-By: Codex Medium <noreply@openai.com>
Co-Authored-By: Muse Wonder High <noreply@meta.com>

* Say when look app --after isn't whole seconds

The wait goes through an arithmetic expansion, so `--after 0.5` died on a syntax error and then waited fifteen seconds for a window, and `010` was read as octal.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Co-Authored-By: Codex Medium <noreply@openai.com>
Co-Authored-By: Muse Wonder High <noreply@meta.com>

* Refuse a display name that would break the starter

The display name is written into a QML string as it is, so `My "Notes"` gave a starter that failed to load before the agent had changed anything.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>

* Find a window a wrapper started before exiting by its session

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.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>
…s up (omacom#19259)

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Keep the options-only comparison introduced by omacom#19229 so an older configuration header does not trigger a boot-image rebuild. Point the read-only installer regression at the renamed command.

Co-Authored-By: GPT-6 (Codex) <noreply@openai.com>
…ld (omacom#17164)

* Mark the T2 t2bce_vhci initramfs module optional so other kernels build

* no-mistakes(review): Use a literal T2 drop-in fixture in the hooks test

* Retry the T2 boot image rebuild until it succeeds

The migration rewrote apple-t2.conf and then rebuilt the boot images.
When the rebuild failed or was interrupted, the next run found
t2bce_vhci? already in place, skipped the rebuild and was recorded as
done. linux-t2's image then still had no t2bce_vhci.

A pending marker now goes down before the edit and comes off only after
a rebuild with no failed kernel, so the rebuild runs again until it
succeeds. It uses limine-mkinitcpio and checks for a failed kernel the
way migration 1791621765 does. The comment names the 4.5.0 installer
filter as the reason the keyboard was gone on linux-t2.

---------

Co-authored-by: Emir Beganovic <beganovic.emir@gmail.com>
…macom#19269)

Migration 1793464570 ran limine-mkinitcpio with no presence check.
Limine is not in the base package set, so on a T2 Mac without it the
migration failed on every run and held up every migration after it.

It now runs mkinitcpio -P where limine-mkinitcpio is missing, as
omarchy-plymouth-set does. A failed mkinitcpio -P exits non-zero, so
the rebuild marker stays and the migration runs again. A comment also
says which apple-t2.conf lines the trailing-space match covers.
Save keyboard boot mode when sysfs is read-only
The brightness keys drive an external monitor through ddcutil, which
needs read-write access to the monitor's DDC bus. On a Pi 4 the HDMI
DDC buses are /dev/i2c-20 and /dev/i2c-21, owned by root:i2c with mode
0660. ddcutil's own udev rule only tags PCI display controllers, and
these are the SoC's platform I2C controllers. So every key press failed
with EACCES and the monitor kept its brightness.

The Pi 4 board setup now installs a udev rule that tags exactly the two
HDMI DDC controllers, fef04500.i2c and fef09500.i2c, for the active
seat user. No other I2C bus is opened. On an image's first boot the
leaf runs after udev added the devices, so there it also reloads udev
and replays the I2C devices, as bluetooth.sh does for its service.

On a Pi 4 with an LG ULTRAGEAR+, ddcutil as the user went from EACCES
to reading brightness 87/100. Brightness-down dimmed the monitor only
with the rule installed.

Co-authored-by: Preetham Kyanam <37784174+pkyanam@users.noreply.github.com>
Let Pi 4 brightness keys reach an HDMI monitor
Follow Omarchy's chrome and corners in T3 Code
Keep quattro’s untrapped script session leader, caller SHELL restoration, and Bash pin while giving each invocation its own retained transcript.

Co-Authored-By: GPT-6 High <noreply@openai.com>
Adapt the overlapping-update regression from omacom#17020 to real script and lock entry points, and check the analyzer against current and legacy transcripts. This replaces the extracted-wrapper test and keeps fresh update and hangup fixtures inside their own temporary directories. Document how to analyze a retained transcript explicitly.

Co-Authored-By: sundarshahi <shahithakurisundar@gmail.com>
Co-Authored-By: GPT-6 High <noreply@openai.com>
The private transcript fix does not need to silence a missing legacy log. Preserve the existing grep diagnostic instead of making an analysis that read nothing look clean; the separate missing-log diagnostic in omacom#11217 can still land on its own. Remove a redundant move in the analyzer fixture.

Co-Authored-By: GPT-6 High <noreply@openai.com>
Co-Authored-By: Opus 5.5 High <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

keep pr-janitor: never auto-close muse

Projects

None yet

Development

Successfully merging this pull request may close these issues.

omarchy update fails for a second user when another user's /tmp/omarchy-update.log exists