voxtype's build from source fails now and then at link time, with
Vulkan shaders missing: 0c0ddf5 built it in two runs and not the third,
with the same packages. The baseline isn't what's under test, and the
old installers skip what they already installed, so try once more.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
mise.run's installer added a blank line before its activate line, and
migrations/1793453501.sh removes only the activate line, so machines
installed before then keep a harmless blank line a fresh install lacks.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
013acac's update (in use 2026-08-14 to 2026-10-19) copies the new update
over itself with sudo cp while bash is still reading it, so bash reads on
from byte 1484 of the new file: today, the middle of the logo, and the
update dies with a syntax error. Put a line at exactly that byte that
execs the new update, in a here-document a normal run skips. Found by
30-upgrade-from-baseline at ca658dc; the test replays 013acac's update.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
30's fix-up commit now also adds clang and cmake, without which voxtype
no longer builds from source (later commits install them), and stops
ca658dc's bitwarden.sh failing on its unset LINE under the set -u that
install-kickstart-nvim.sh leaves on.
The fingerprint recorded only the line numbers of .bashrc's coredesktop
lines, so 0c0ddf5's showed a line moved without saying what was added.
Record the whole file, sorted.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
CoreDesktop installed it to build cline-desktop, which cline-desktop-bin
has since replaced, so machines installed before kept it where a fresh
install has none. Found by 30-upgrade-from-baseline at 0c0ddf5.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
makepkg builds pkg-debug alongside an AUR package, and removing the
package leaves it behind: the migration that dropped asdf-vm left
asdf-vm-debug. Found by 30-upgrade-from-baseline at 0c0ddf5, whose
machine kept it where a fresh install has none.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
ca658dc installs extensions with plain gext install, which waits for a
GNOME Shell dialog no one clicks; add --filesystem, as HEAD does. ca658dc
and d8d93a4 run flatpak install as the user, which polkit refuses over
SSH; allow it for the install only, with a rule removed afterwards.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
An AUR package built with makepkg's debug option comes with pkg-debug,
which nothing depends on, so pacman -Rns leaves it behind. voxtype-debug
then conflicts with voxtype-bin-debug, and update fails on machines
installed before voxtype-bin replaced voxtype. Found by
30-upgrade-from-baseline at 0c0ddf5.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
clang-libs and llvm-lit were split off clang and llvm after the archive
date, so the old clang and llvm conflicted with their files and the
downgrade failed, skipping the scenario. Remove packages from Arch's repos
that the archive doesn't have yet before downgrading.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Every baseline installs herdr from source, which pins zig 0.16, and Arch
now has 0.17, so their installs fail. Install them with herdr-bin (which
provides herdr, and is what HEAD installs) in a commit on top, then put
the clone back at the baseline itself, since update pulls --ff-only.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
install.sh at every baseline in smoketest/baselines clones from
https://gitea.federated.computer/david/coredesktop.git, which wasn't in
COREDESKTOP_URLS. So 30-upgrade-from-baseline installed today's real
upstream (via Gitea's redirect) instead of the baseline, and failed when
the old install.sh asked for files that have since gone.
Add a test that every URL install.sh has ever cloned from is in the list.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
selected-layouts, last-version-name-installed and overridden-settings are
Tiling Shell's own state from the machine the file came from: the layout
of each of its 8 workspaces, the version it last ran, and the GNOME
settings to restore if it's turned off. Loaded on other machines, they
reset users' layouts whenever the file changed and could restore the
wrong GNOME settings. Existing machines keep their current values.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
gnome-extensions.sh loaded gnome-settings.dconf unconditionally, so each
refresh (install.sh runs it) undid users' changes to those settings,
bypassing apply-gnome-settings.sh's check that the file has changed. That
script, which install.sh runs later, already loads the settings when
needed. The smoketest's 11-refresh caught it: Tiling Shell's
selected-layouts went back to the file's value.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Since a65f07c ran each hook in a subshell, a hook that does its work as
it's sourced (12-migrations, 15-services, 20-aur, 30-flatpak and others)
had its failure lost: the run_update check after source left the
subshell's status at 0, so update said "Done." and exited 0. The smoketest's
40-failing-migration caught it. update now exits the subshell with the
sourced hook's status.
Also, the smoketest scenarios that check whether a migration ran look for
"Running migration " so install.sh's "Running migrations..." heading
doesn't match (11-refresh failed on it).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
run_in_vm prints how long it has run and the log's latest line, so an
hour-long install.sh shows progress without tail -f.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
upstream_set pushes the commit under test from the working copy, so
coredesktop-dev's pre-push hook ran make check on every push to the VM.
Inside the smoketest those checks saw VM_GIT_URL and friends, the
smoketest tests pushed to the real VM and failed, and the hook stopped
the push, failing the installed image's build. The push now uses
--no-verify, and the tests clear the smoketest's variables.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
So a long command (install.sh takes an hour or more) can be followed with
tail -f; run_in_vm prints the log's path when it starts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
coredesktop-dev now gives its VMs QEMU's standard VGA, which qemu-base
has, so the extra package and doctor's check are gone.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
New machines get VSCodium alone. A migration uninstalls the VS Code
Flatpak once on existing machines; its settings stay in ~/.var/app. The
README says how to install VS Code for extensions only Microsoft's
marketplace has, such as Remote-SSH; update then leaves it alone.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
install.sh now stops early, with a clear message, if this isn't 64-bit
Arch, it's run as root, there's no internet, or sudo doesn't work.
system/preflight.sh then asks before going on with too little disk space
or / not on btrfs, and offers to switch another login screen to GDM and
PulseAudio to PipeWire. Keeping another login screen puts GDM in
services-off, so install and update stop failing to enable it; a
migration does the same on existing machines. It says nothing when all
is well, as refresh runs install.sh again.
Intel GPUs also get intel-media-driver for hardware video decoding, with
a migration for existing machines.
The mise migration links ~/.local/bin/mise to /usr/bin/mise instead of
deleting it, so shells opened before the update don't print an error at
every prompt.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Brings in the shellcheck lint and bats tests. Both sides fixed the same
shellcheck warnings; this keeps the remote's versions of 20-aur.sh,
80-gnome-extensions.sh and .shellcheckrc, which silence false positives
per line rather than for the whole repo.
The Gitea workflow now runs make lint, so it checks the same scripts as
the Makefile, and the README points contributors at make check. The test
helper's REPO, used only by the .bats files, is marked for shellcheck,
and the update test expects the pull to be fast-forward only.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
make lint runs shellcheck on every tracked bash script, make test runs
the bats suite in tests/, and make check runs both. The tests cover the
update orchestrator (hook isolation, failure reporting, pull and
free-space handling) and utils.sh (migrations, package install and
removal, services, initramfs build check), using stub commands so they
never touch the real system.
Fixes from the first lint run: 20-aur.sh stops if a cd fails rather
than building or deleting in the wrong directory, and
80-gnome-extensions.sh refuses to rm -rf with an empty path. The rest
are false positives, silenced in .shellcheckrc or per line.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
davidandClaude Opus 5.5