core.fileMode=false on a Windows checkout, so git never picked up the executable
bit from disk. Both files are 100755 in the index now; a fresh Linux clone gets
them runnable instead of needing chmod.
Adds T16-T20 to the Debian harness (103 passed, 0 failed, 2 skipped):
T16 the polkit rule is installed, mode 644, covers the two update actions,
does not over-grant package-remove/reinstall/untrusted, and is idempotent.
T17 the real decision, in both directions: a sudo-group user is authorised
(pkcheck rc=0) with the rule, and still prompted (rc=2) without it. The
control is the point - without it the test cannot fail.
T18 vdi-offline-update: bash -n, unit verification, the throttle stamp, that
a second run is skipped, that a stale stamp lets it run, that pkcon exit 5
is reported as 'no updates available' rather than a download failure, and
that /system-update is never armed without a prepared payload.
T19 the preseed still generates a script that parses, delegates to
vdi-offline-update instead of an inline pkcon sequence, and enables cron
and the timer.
T20 make-iso.sh: refuses to overwrite its input or an existing output, then
checks the built image - preseed byte-identical, all 40 kernel lines
patched, gtk.cfg patched, no doubled parameters, El Torito present with
both BIOS and UEFI images, no source file missing, and size within 16MB of
the original (the bloat regression).
Skips stay honest: T12 needs gpg, T13's ks.cfg assertion is vacuous because
the file was never in the repo.
Ports the update behaviour of the Fedora clients to Debian, and adds the
Debian equivalent of putting ks.cfg on the installation medium.
Updates
-------
Debian's shipped polkit policy only lets a sudo-group user trigger
upgrade-system and trigger-offline-update without a password. The actions
pkcon refresh and pkcon update actually use (system-sources-refresh and
system-update) are auth_admin, so a desktop user was prompted - a
behavioural difference from Fedora. update_local_repo.py now installs
/etc/polkit-1/rules.d/60-vdi-no-password-updates.rules granting exactly
those two plus package-install to group sudo. package-remove,
package-reinstall and package-install-untrusted stay auth_admin.
The offline update itself ports unchanged: Debian 13 PackageKit 1.3.1
supports offline-get-prepared and offline-trigger, and
packagekit-offline-update.service is wired into system-update.target.
Scheduling moves from an @reboot cron line to vdi-offline-update.timer
(6h, Persistent), with /root/auto_update.sh delegating to the same script
so there is one implementation instead of two.
Both are installed by update_local_repo.py, not only by the preseed, so
already-provisioned machines pick them up on their next 10-minute cron run.
install ISO
-----------
debian/make-iso.sh bakes preseed.cfg into a Debian netinst ISO and adds
auto=true priority=critical preseed/file=/cdrom/preseed.cfg to every boot
menu entry, then verifies its own output.
Measured, not assumed
---------------------
* subject.local == true never matches in this environment: pkcheck returns
2 with the clause and 0 without it, same sudo-group user. It was in the
first version of the rule and is removed; the reasoning is in the source.
* pkcon exits 5 for "no packages to update", which is the normal answer on
a patched box and was being logged as a download failure. Fixed.
* offline-get-prepared exits 4 when nothing is prepared, so the Fedora
"if ... offline-trigger" guard is load-bearing - triggering anyway can
leave a dangling /system-update symlink that breaks the next boot.
* Extracting a netinst and rebuilding with "xorriso -as mkisofs" grows it
792MB -> 988MB because Debian shares identical file extents; replaying
the original session with -map keeps dedup and the boot record.
* A hardcoded menu list misses isolinux/gtk.cfg, where the BIOS *default*
entry lives, so the machine still prompted. Menus are discovered now.
Tests: 90 passed, 0 failed, 2 skipped. T17 asserts the polkit decision in
both directions (authorised with the rule, prompted without it), because a
check that cannot fail proves nothing.
Two runs of the harness share /var/lib/apt/lists and /etc/apt/sources.list.d.
With two in flight, one run's teardown deleted the other's temp repo while apt
was reading it, producing 'Could not open file .../vdi-debian-test.<other>' -
a failure that looks like a port bug but is a collision. Take a flock for the
duration of the run, and clear indexes left by earlier runs whose temp
directory no longer exists.
The provisioning path now runs on both package families. Fedora behaviour is
unchanged: the Fedora publishing code issues the same ten operations in the
same order as before, verified by recording the call sequence from both
implementations against a stubbed environment.
update_local_repo.py
Detects the family from /etc/os-release and dispatches to sync_fedora() or
sync_debian(). Release assets are filtered by extension, so neither family
sees the other's packages. The Debian path builds a flat apt-ftparchive
repository at the same /var/local/citrix-repo, publishes
local-citrix-repo.list, and runs apt-get update.
Fixes found while testing:
- tqdm is now an optional dependency. A stock Debian 13 image has no pip and
no tqdm, and dying at import would take down provisioning over a progress
bar.
- sync_debian prunes stale *.deb files before repopulating. The Fedora path
gets this for free by writing fixed filenames; Debian filenames carry the
version, so without pruning every old version stayed indexed and apt always
picked the highest, making rollback impossible.
- The apt source file is resolved next to the script rather than from the
working directory, with a hardcoded canonical fallback, so a foreign cwd or
a missing file cannot break the 10-minute cron job.
local-citrix-repo.list
The apt counterpart of local-citrix-repo.repo, which stays as-is for dnf.
Uses trusted=yes rather than a signature: the repository is generated on the
machine from packages this script downloaded itself, so a signature would only
prove the machine signed its own work, and it would add a gpg dependency that
Debian 13 does not ship by default. Index integrity is still enforced through
the hashes in Release.
debian/preseed.cfg
The debian-installer equivalent of ks.cfg. Same locale, keyboard, timezone,
passwords, inventory-derived hostname/user/password, disk auto-detection,
Firefox managed bookmarks, Citrix and TeamViewer installs, auto-update script,
cron jobs and disable-sleep unit. Every directive was checked against the
templates in the real trixie udebs. Documented gaps, all inherent to preseed
vs kickstart: uppercase usernames are rejected by user-setup and so are
created in late_command; d-i has no lsblk, openssl, chvt or openvt, so disk
detection reads /sys/block and the inventory prompt goes to /dev/console with
a kernel-cmdline override and a timeout; pip install tqdm becomes
python3-tqdm from apt because of PEP 668.
debian/test/
Harness covering the Debian path on a real Debian 13 machine: repo build,
sources.list validity, apt visibility after purging the list cache, install
resolution, idempotency, stale-package pruning, corrupt-index rejection, and
Fedora non-regression. 52 pass, 3 skip (each skip states why).
Known limitations are listed in README.md under "What the tests do NOT prove":
no real d-i install has been run, and the generated post-install script is
verified as well-formed rather than executed against a target root.