Validating /etc/omarchy.conf walks its parent chain to /, and that walk leaves
its own subject behind in the global failure_context. The comparison that
follows -- the one deciding whether the authorization actually names this
checkout -- sets no context of its own, so an unauthorized development tree
refused with:
refusing to publish: directory / (must be root-owned and not group- or
world-writable) failed validation
naming a directory that had just passed, and pointing the reader at a
filesystem problem that is not there. Seen on a worker VM running an
unauthorized checkout, where / is 0555 root-owned.
The hint line underneath was already right, which is what kept this from being
worse. Restore the context before the comparison so the first line agrees with
it, and assert both halves: that the refusal names the authorization, and that
it does not blame the root directory.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0115LngksSpXLD9NSXBEP3ki
Checking a queue for jobs and then deleting it leaves a window in between, and the sudo that does the deleting can sit at a password prompt for as long as someone takes to type. A job submitted in that window was cancelled by a deletion that had already decided the queue was empty. cupsreject closes the queue to new work first, which also stops more jobs piling onto one that is being left behind and can no longer route them. It comes from the cups package, which stays.
What a queue with jobs on it means is now said rather than implied: whatever is already at the printer finishes, whatever is still waiting cannot be routed with the daemon gone, and the person who owns them has to cancel what is left.
A queue that another administrator removed while this was running is the outcome wanted, not a failure to keep the package installed for, so a deletion that fails is checked against whether the destination is still there.
Co-Authored-By: Codex XHigh <noreply@openai.com>
Mutating each root-side control one at a time showed three that no test could
see: a destination directory root does not own, a single user-owned asset
inside an otherwise root-owned packaged directory, and an asset left group- or
world-writable by its own mode. Deleting any of the three left the suite green,
because the existing cases mark a whole tree untrusted and are caught by the
directory check before the per-file one is reached.
The harness already had the hook for the ownership pair: TEST_UNTRUSTED_SOURCE
makes the stat shim report a chosen prefix as uid 1000, so those two only need
it pointed at a destination directory and at a single file rather than at a
whole tree. A mode has to be real, so that case stages a copy of the packaged
tree the shim reports as root-owned and loosens one asset in it.
The empty logo is refused by the destination size bound rather than the
caller-side one, so that case pins the behaviour without isolating the check;
the two bounds are exactly redundant.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0115LngksSpXLD9NSXBEP3ki
cups-browsed is the daemon that watches the network and creates print queues by itself. Hardening it took a root daemon with a predictable cache down to a confined service account, but a daemon that turns anything advertising itself on the network into a print queue is a lot of exposure for a convenience, so it comes out of the default install while that is reworked. Only the discovery half: CUPS itself stays and printing keeps working, with each printer added by hand in Print Settings.
The migration disables the unit before removing the package because that is the only order that works: pacman deletes the unit file but not the enable symlink, and once the unit is gone systemd can no longer resolve it by name to clean that up.
It then removes the queues discovery generated. cups-browsed keeps those when it stops, since KeepGeneratedQueuesOnShutdown defaults to Yes, and they route through its own implicitclass backend, which goes with the package, so they cannot print again. Idle ones go. A queue with jobs on it is left alone and named: implicitclass only needs cups-browsed to choose a destination, so a job already past that point finishes on its own, and deleting the queue would abort it. One printer's job does not hold up the removal. A printer added by hand has an ipp:// or usb:// device and is left where it is.
A queue whose jobs cannot be asked about is left alone rather than assumed idle, including one named so that lpstat would misread it -- "all" is its word for every destination, and a leading dash or a comma reads as another option or a list.
Where CUPS does not answer at all, or a queue will not delete, discovery is still stopped but the package stays and no marker is written. omarchy-migrate records a migration for the user as soon as it exits zero, so that is where the machine stays until someone removes the package by hand, and the message says so rather than implying a retry.
The queue list is read under LC_ALL=C because lpstat translates "device for", and captured rather than piped, so a cupsd it cannot reach is reported instead of reading like a machine with nothing to clean up.
It removes with plain pacman -R rather than omarchy-pkg-drop, which passes -n and would discard /etc/cups/cups-browsed.conf instead of keeping it as a .pacsave. A removal meant to be temporary should not delete the machine's copy of its own configuration. Without -s either, so it only ever removes the package it names: sweeping newly unneeded dependencies is nothing today, but it is not a promise a rolling dependency graph can keep.
Queue names come off the network, since cups-browsed names its queues after what the printer advertised. CUPS allows every printable character but space, tab, / and #, and lpstat and lpadmin take a destination as an option value, so a name with a leading dash or a comma is reported rather than passed to them and guessed at.
Migration state is per user, so a machine-wide marker records the one removal. Without it, an account whose first migration run came after someone deliberately reinstalled discovery would quietly take it back out again.
The install-time override for cups-browsed.conf now waits for cups-browsed rather than for CUPS. Guarding it on a file CUPS still ships would write a configuration file for a package nothing installed, and pacman would later land the package's own copy beside it as a .pacnew.
The hardened configuration stays in the tree. omarchy-settings still ships the cups-browsed.conf override, the sysusers account and the service drop-in, so they are what discovery returns onto.
Co-Authored-By: Codex XHigh <noreply@openai.com>
Deleting the whole ancestor walk from validate_trusted_directory and checking
only the immediate parent left the suite green, so the invariant the design
rests on had no coverage: a writable ancestor lets an attacker swap a validated
directory out from under the leaf. Reject a run whose destination grandparent
is world-writable while the destination itself is pristine.
The harness also rewrites the root script's fixed /usr/share paths by string
substitution. A drifted string silently no-ops that rewrite and would point the
simulation at the real system tree, so assert each substitution landed. Assert
the scratch directory exists too: a failed mktemp left it empty and the suite
carried on building paths from it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-Authored-By: Codex XHigh <noreply@openai.com>
Every check inside the root shell is a bare [[ ]] or (( )) assertion that
aborts under set -e, so a refusal exited with status 1 and no output at all.
The floating-terminal wrapper then printed its green "Done!" for any status
but 130, so a failed boot-theme change read as a success.
The refusal a working machine actually hits is omarchy dev link, which points
OMARCHY_PATH at a checkout the desktop user owns. Name that case outright and
point at omarchy dev unlink; report every other rejection through an ERR trap
that names what failed validation.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Keep SDDM color substitution collision-free for White themes, based on the fix proposed in #8469.
Co-authored-by: itz4blitz <itz4blitz@users.noreply.github.com>
Neither half of the validation was covered. Dropping `&& -c $cached` from the wrapper left the whole file green: all three poison values fail on the pathname prefix, so none of them ever reached the character-device test. A path that matches the hiddev glob but is not a device now covers it, and it is the real case rather than a synthetic one -- the display replugs, the interface renumbers, and the cached node is gone. It is added only when the host has no such node, so a machine with the display attached cannot fail there spuriously.
The no-XDG_RUNTIME_DIR assertion had the same problem for the opposite reason: its decoy held a path the validation rejects on its own, so restoring the `${XDG_RUNTIME_DIR:-/tmp}` fallback left it passing. It asserts on the open now instead of on the contents -- a FIFO with no writer blocks whoever opens it, so a wrapper that consults the path hangs and one that ignores it exits. mkfifo is atomic and fails outright if the path is taken, so it still neither overwrites a file nor follows a symlink at the fixed path.
Clearing created_tmp_cache as soon as the decoy is removed keeps this run's EXIT trap from deleting a concurrent run's decoy at the same fixed path, which would have let that run pass against the old code.
🤖 Generated by Opus 5 in Claude Code.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-Authored-By: Codex XHigh <noreply@openai.com>
The scheme check ran after the interactive branch, which had already handed the URL to curl: a refused `file://` or `ftp://` URL was dereferenced first, the whole string went out to Google's favicon endpoint in a `domain=` query, and a fetch that happened to succeed left an orphan icon and a rebuilt icon cache behind. Validating immediately after normalization puts the refusal ahead of every dereference in both branches.
A leading space also kept the URL out of the scheme test entirely. `normalize_webapp_url` saw no scheme, prefixed it, and ` file:///etc/passwd` became `https:// file:///etc/passwd`, which passes `^https?://`. The desktop `Exec` field is unquoted, so it splits back into two arguments, and `omarchy-launch-webapp` forwards everything after the first to the browser, where a bare argument is another URL to open. Refusing whitespace closes that without touching desktop-entry escaping.
Scheme comparison is case-insensitive because schemes are: `HTTPS://example.com` installed before this check existed and has no reason to stop.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-Authored-By: Codex XHigh <noreply@openai.com>
The username prompt already refuses the service accounts a desktop user must not claim, cups and lp among them. A user who took cups-browsed would get a primary group of that name, and the CUPS authorization written here puts that group in SystemGroup, handing that desktop user the passwordless administration the rest of this change removes.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-Authored-By: Codex XHigh <noreply@openai.com>
cupsd compares directive names with _cups_strcasecmp, so a hand-edited "systemgroup sys root wheel" is live configuration, but matching $1 against the canonical spelling skipped it and appended a second directive at the end of the file. parse_groups accumulates the groups of every SystemGroup directive it reads rather than replacing them, so both lines took effect and wheel kept the passwordless administration this is meant to remove, with the migration reporting success.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-Authored-By: Codex XHigh <noreply@openai.com>
A bracket range is collated, not ASCII, so `[a-z]` admits `é` under
en_US.UTF-8 and refuses it under C: the same theme URL installed on one
desktop and was refused on the next, and the manual's stated set was only
true in the C locale. Pinning the comparison makes the written set the
shipped one everywhere.
git reads a URL as scp-style when a colon appears before any slash, so the
path after it need not contain one. Requiring a slash left the prefix on
`git@host:omarchy-blue-theme.git` and derived the name
`git@host:omarchy-blue`, which the allowlist then refuses -- a repo that
clones fine could no longer be installed at all.
Co-Authored-By: Codex XHigh <codex@openai.com>
The migration recomputed whether cups-browsed was active at the start of every run, after an earlier run may already have stopped it. A run interrupted between that stop and the completion marker left the service down, and the retry that followed read it as inactive, skipped the restart, and wrote the marker anyway: printer discovery stayed off until the next reboot with no migration left to bring it back. Keying the restart to whether the unit is enabled survives the interruption, and reports not-enabled for a unit the user masked or disabled, which restarting would fail on and abort the migration short of its marker.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-Authored-By: Codex XHigh <noreply@openai.com>
GNU sed's N auto-prints the pattern space and exits at end of input, so the
`:a;N;$!ba` slurp skipped every following s/// for a value with no newline in
it. That is every value except the injection attempt the escaping exists to
stop, so the Exec quoting's own backslashes were written to the file unescaped
and GLib refused to load the entry.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-Authored-By: Codex GPT-5.6-sol (xhigh) <noreply@openai.com>
The slash guard was the only thing keeping a name out of the directory structure, and nothing tested it: deleting it left the suite green, because creating the launcher directly in the applications directory already makes the redirect fail on its own, with a raw bash error instead of the message. The assertion is on the message now, alongside the traversal case the guard actually closes -- on quattro a name of `../../../../escaped` writes its launcher clean outside the applications directory.
The interactive prompt read the name, fetched the favicon, wrote it and updated the icon cache before the name was ever checked, so a URL typed into the Name field left an icon behind on every attempt. Validating as soon as the name is read covers both paths from one place.
Removing by name also scanned unconditionally, so a machine with no applications directory printed a find error where omarchy-remove-gaming-xbox-cloud does not hide stderr.
🤖 Generated by Opus 5 in Claude Code.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Run cups-browsed as a locked service account with a dedicated cache and a focused systemd sandbox. Restrict automatic queues to driverless IPP printers, remove wheel from passwordless CUPS administration, replace cups-pdf with Polkit-backed setup, and migrate existing systems safely.
Reported-By: Erik Hunstad (Bad Sector Labs)
Co-Authored-By: Daybreak Blue <noreply@openai.com>
Each of these is a Text rendering external data with no textFormat, written in
a form that passed silently. None exists in this tree, so they were holes in
the guard rather than live exposures — but a guard is only worth what it
catches, and every one of them is a single line someone could plausibly write.
Text /* why */ { strip_noise knew // and not /* */, so a block
comment between the type name and its brace hid
the element from every rule at once
QQ.Text { ... } a namespaced import made the name compare unequal
to `Text`, and the element was skipped outright
visible: textFormatEnabled textFormat was matched as a substring, so a
lookalike property exempted the whole block
component Info: a component root with its Text on the next line;
Text { the one-line form was covered and this was not
an unreadable subdirectory rglob() swallows a directory it cannot enter, so
a locked subtree scanned as though it were empty
The scan moves out of the heredoc into qml-text-format-scan.py, taking its root
as an argument, because nothing could run it over anything but the real tree —
and a scanner whose only input always passes cannot be shown to fail. The test
now runs it over nineteen fixtures, one per form above and one per form the
scan already handled, so a later edit that loosens it fails here instead of
going unnoticed until something renders a remote image.
Two limits stay open and are written down in the module docstring rather than
papered over: text assigned from elsewhere (a Binding element, PropertyChanges,
an onCompleted assignment, a property alias onto a child) is invisible to a
scanner that reads each element's own declaration, and a regex literal holding
a brace throws off the brace depth. Neither shape exists in this tree and both
need a QML parser, not another regex.
Co-Authored-By: Codex XHigh <noreply@openai.com>
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The card binds the body Text to styledBody, which rewrites newlines to <br/>
*after* sanitizeBody has run. That rewrite inserts tag syntax into text the
stripper deliberately kept: a kept tag may hold a `<` of its own, and `<x`,
newline, `<img src="http://host/x.png">` is one tag named `x` to both the
stripper and Qt, so it survives whole — until the rewrite splits it into
`<x<br/>` and a live image tag the input never contained.
Measured against Qt 6.11.2 with an offscreen StyledText and a local HTTP
server: that body issues the GET after this branch's sanitizer and issues
nothing before it, because the one-pass /<img[^>]*>/gi it replaces deleted the
inner substring outright. The whole-tag bound is still the right trade — it is
what stops the stripper manufacturing tags — but it only holds if nothing edits
the string afterwards.
So move the rewrite into NotificationLogic, next to the reasoning it depends
on, and strip again after it. What Qt parses is then what was checked last. The
tests assert on styledBody for the same reason, since sanitizeBody's output is
no longer the string that reaches the renderer, and a regex assertion pins the
card's binding because no JavaScript assertion can see a QML property.
The mute was reachable only as `omarchy-toggle crash-ignore/<program>`, which asks whoever runs it to know the flag layout, to reduce a binary's path to the name the watcher keys on, and to have read the rule that a name climbing out of that directory writes an unrelated toggle. All of that was carried in the skill's prose, which is the wrong place for a rule that has to hold: prose is advice, and the thing being advised about is a name the crashed program chose.
So it is a command now. `omarchy crash mute hyprland` silences that program, `off` lifts it, `toggle` flips it, and no argument lists what is muted. It takes the binary's path as readily as the name and reduces it the way the watcher does, so the `Executable:` line from `coredumpctl` can be handed straight to it; it refuses what is not one component of a name, so it cannot be talked into writing outside its own directory whatever it is given; and it re-reads the flag afterwards and reports what is now true rather than what was asked for. The listing counts only regular files, because that is all the watcher honours -- anything else in there would read as muted while the crashes kept arriving. A leading `--` is consumed so a program named `-h`, which the router would otherwise answer with its own help, can still be muted.
The watcher gained an unrelated fix that this uncovered. Its fields are read with `IFS=$'\t'`, and tab is IFS whitespace, so an empty field collapsed into the next delimiter and shifted every field after it along one: a crash whose comm was empty had a path read as its pid and was discarded as somebody else's. A process can set its comm to nothing, so that was reachable. Empty fields now arrive as a dash like missing ones, and a dash joins the empty and dot cases that fall back to `unknown`.
Co-Authored-By: Codex XHigh <noreply@openai.com>
A crash that is understood is not a crash that stops: an upstream bug waiting on a release, a program that dumps core every time it exits. The diagnosis explains it once and the toast keeps arriving, and the only answer Omarchy had was Crash Capture, which turns off every program's notifications in order to silence one.
The watcher already resolves a name to dedupe on and announces that same name in the toast, so the mute is keyed on it: a flag file under toggles/crash-ignore/, written by the existing omarchy-toggle and read by the existing omarchy-toggle-enabled. One flag per name rather than one list, so `on` mutes, `off` un-mutes, and `ls -A` shows what is muted, with no new file format and nothing to parse. It is the executable's basename wherever one was recorded, falling back to the process name, which the kernel truncates to fifteen characters -- muting the truncated form would match nothing, forever, while looking like it worked.
The name is not always a name, though, and the mute turns it into a path. A program picks its own comm and prctl takes anything, including slashes, and the watcher falls back to comm whenever a crash carries no absolute executable. So it is stripped to its last component first: without that, `a/../bar-off` is a legal comm aimed at an unrelated Omarchy flag, letting a crashing program suppress its own notification and letting a user who accepted the offered mute hide their bar instead. Stripping does not always leave a component either -- `/` leaves an empty string, which is no kind of array subscript and no kind of toast, and `.` or `..` names a directory that omarchy-toggle would touch and report success on, leaving a mute that never matches. Both fall back to `unknown`, the word omarchy-agent-crash already uses for a name it does not have, and which mutes like any other.
The skill offers this at the end of a diagnosis and never runs it unprompted, which makes it the single change a diagnosis may make to a system it otherwise only reads. It tells the agent to use the name it was handed rather than re-derive one, since the watcher resolved that name already and the two agree for ordinary names and not for strange ones; a diagnosis started by hand from `omarchy agent crash <pid>` is given no name and gets the derivation instead. It also says to treat the name as hostile text rather than as a word -- it is whatever the crashed program's author called a file, so a single quote inside one closes the quotes around it and the rest runs as the shell -- and to check the flag arrived rather than assume it.
Co-Authored-By: Codex XHigh <noreply@openai.com>
The name a theme installs under is derived from its repo URL, and holding it to
an allowlist made that allowlist a naming convention nobody had written down. It
was also tighter than the harm it exists to stop: `+` is not shell syntax and a
leading `_` is neither the `..` climb nor the dash basename reads as an option,
so `omarchy-c++-theme` was refused for nothing.
Widen the set to those two and say what it is where a theme author is already
picking a name. The leading character stays out of `.` and `-`, which is the
part that does the work.
Reported-by: Luis Alvarez (lalvarezt)
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Fd3RCHxwjEbMXoSYB9Aiso
The root rule matched only a file-level root Text, of which this tree has
exactly one. QML inline components are roots for the same reason — the
`text` of `component InfoValue: Text {` comes from every caller, so the
file it lives in never binds it — but they sit inside another element, so
the depth-1 test never saw them. Six went uncovered while the test
reported green, among them the network panel's InfoValue, which callers
bind to the IP address and gateway.
Six more ways to write a Text were read as clean rather than as unreadable:
an opening brace that is not last on its line, a brace on the line after
`Text`, a one-line block containing nested braces, a wrapped binding split
by a comment or a blank line before its `+` (which exempted a dynamic
binding as a literal), and a root Text indented from column zero. Require
the forms a line scanner can read instead of parsing QML; the tree already
writes every Text that way.
Last, a run that read no files reported success. A checkout with no shell/
QML now fails instead, since an all-clear from a scan that opened nothing
is the one answer this test must never give.
Each case is covered by a fixture that fails without its fix.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-Authored-By: OpenAI Codex (gpt-5, xhigh) <noreply@openai.com>
A theme installed from a git repo is named after the repo URL, and that name
becomes its directory name under ~/.config/omarchy/themes. Style > Unlock built
a command line out of the name the picker returned and handed it to
omarchy-launch-floating-terminal-with-presentation, which runs its argument as a
shell string -- so a theme directory called `a';id;'b` ran `id`. Themes are
already held to contributing colour and nothing that executes, which is why
omarchy-theme-set stages no .lua, terminal config, or vscode.json from one.
Hold the derived name to the characters a theme name needs, which stops it from
being dangerous at every place it lands rather than at the one found, and quote
it with printf %q on the way into the action for the names already on disk.
omarchy-theme-remove keeps its existing path-climb guard: its name reaches only
a quoted rm, and the same charset would strand a theme installed before this.
Reported-by: Luis Alvarez (lalvarezt)
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011WFcUm5HWFyxaVYdwAeWPP
QQuickStyledText skips the characters between `<` and the tag name with
QChar::isSpace(), which counts U+0085 NEL. JavaScript's `\s` does not, so
isImageTag() read no name at all from a tag written as `<`, U+0085, `img`,
kept it, and Qt then read `img` and issued the GET the stripper exists to
prevent. Measured against Qt 6.11.2 with an offscreen StyledText and a
local HTTP server.
Read the name by skipping everything that is not part of it rather than by
matching the separator, so the two definitions cannot drift apart again.
Over-skipping is the safe direction: it can only classify more runs as
images, and dropping a run never manufactures a tag.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
