Seven gpu_tpu scripts now write through the journal for host changes only — driver and package installs, vfio binding, modprobe.d, /etc/modules, GRUB and udev — while VM and LXC configuration stays out; add_gpu_vm and install_coral_lxc journal only their host writes. amd_gpu_tools records its install, and pmx_journal.sh gains pmx_record_uninstall while changes_journal.py retires an installed package when ProxMenux removes it. What each function writes is byte-identical to before, verified by tests/journal/verify_journal_migration.py.
The change journal exists so a sysadmin sees what ProxMenux changed on the host — its own configuration, packages and services — not how it uses the host or configures a guest. Several scripts recorded operations that are neither: disk passthrough to a VM, container conversions, VM import/export, mounting a share into an LXC. Those are restored to their original, uninstrumented form. Scripts that operate on the host while also installing a package now record only the package: format-disk keeps its exFAT-tools install, the UUP ISO builder its build dependencies, and the share/host scripts their packages, services and /etc/fstab writes, while the mount and unmount operations they used to log are dropped.
The page now reads in three sections: what ProxMenux optimized after install (each function under its menu name), what its other host scripts changed (by script), and what it installed (packages and utilities, each referencing the script that installed it). A file reads as created or modified with its diff, a service shows its state transition, and the undo line appears only when a revert is possible.
ProxMenux modifies the host: it rewrites configuration files, installs packages, enables services. Until now nobody could say afterwards what had changed, and showing the script does not answer that question — a four-hundred-line function may alter two values, and the reader has no way to know which two. This adds the two halves of an answer.
The change journal records what ProxMenux does as it does it. Eleven bash primitives capture the previous state, apply the change and record it in the same step, writing to a spool that the Monitor reads back. One hundred and thirteen functions across twenty-five scripts are instrumented, covering post-install, shared storage, security tooling, container conversions, disk operations and the PVE 8 to 9 upgrade path. The page shows the difference — rotate 7 becoming rotate 14 — and never the script. Restore and backup scripts are deliberately left out: a restore puts the host back to a state some other script already recorded.
The Audit and reports page answers the other half: what state is this host in, regardless of who put it there. Forty-three checks across seven areas read the host and classify each result as critical, warning, observation, conformant, unverified or not applicable, with the evidence they read attached to each one. A declared policy lets the reader say what this particular host is expected to do — which guests must have a backup, which storages are essential — so the report judges the host against its own intent rather than a generic template. An inventory records the hardware, network and guest topology behind those readings, a comparison shows what moved between two runs, and six report profiles produce a printable document scoped to what the reader needs. Everything is available in the eight supported languages.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
An LXC running its workload in Docker could not answer two questions it
already had the data for: which application is in there, and whether a
newer version exists.
**Which version is available.** The Updates tab resolved that number only
for docker.io, and only when a version tag happened to share the digest of
the tag in use. On ghcr.io, lscr.io or quay.io the row said "New image
available" with no number at all. The image a pull would install carries
its own version label, so it is now read from the registry by digest —
over the same protocol and Bearer challenge the digest comparison already
uses, and through the same label lookup the installed version uses, now
shared as _docker_version_from_labels.
The question this answers is the one the tab asks: what do I get if I
re-pull this tag. Not "what is the newest upstream release", which is a
different number whenever a tag is pinned or the publisher tags releases
differently from images.
Docker Hub keeps priority on docker.io: its tag API is not a pull and does
not spend the anonymous pull-rate budget, and official images carry no
labels for the registry path to read. The digest still decides whether an
update exists; this only names it, and declines to name it when the answer
would be a guess — no build for this platform, no labels, an unreadable
manifest, a moving tag, a rebuild of the same version, or two sides whose
versions came from different label keys. Each refusal is recorded in
available_version_source.
Attestation manifests are skipped explicitly: they advertise
unknown/unknown and their config blob is a provenance document, not an
image. Every document is fetched by digest and verified against it, the
config read is bounded, and it is cached per digest, which never changes
content. The CDN redirect is followed by hand, dropping Authorization:
urllib re-sends it to the redirect target and signed-URL storage rejects a
second auth mechanism.
**Which application it is.** The probe already read "1.37.2" out of a
Vaultwarden container and get_suggestions discarded it, so the panel
answered "No new applications were detected" about an application whose
version it had just measured. Containerised applications are now offered
for registration like any other, with their name, logo, published ports
and installed version.
What they do not get is an update path of their own, because they do not
have one: updating Vaultwarden means pulling and recreating its image. A
new update_via=docker marker records that delegation, so one release stays
one badge, one notification and one button. The marker is validated rather
than inferred from installed_via, since docker_exec with an upstream is a
legitimate registration someone may already rely on; combining it with an
upstream is rejected instead of silently stripped, because registering an
app that checks GitHub behind a delegation promising it will not is worse
than an error message.
Three failure modes the delegation had to be defended against: detector
auto-healing would have migrated the app onto a leftover /root/.<app>
marker and quietly un-delegated it; saving replaces the whole record, so
the editor carries the marker explicitly rather than dropping it on the
first port edit; and the release-age hold gates on a publish date a
delegated app never has, which deferred the whole schedule forever.
Their version is resolved server-side through the container the detector
declares — not through the app's name or image, since Immich's compose
service and image are both immich-server while the application is immich.
The annotation happens on the way out of both endpoints rather than into
their caches: the App tab's cache is invalidated by events, not by time,
and the Docker inventory it reads is built asynchronously, so annotating
before storing froze a response taken before the first scan.
The rows carry that name too. display_name was already computed and
already used by the bulk-update section; the image row, the update
notification and the CT badge now use it as well. A delegated app's
pending update counts in the badge only while its image is not already
being counted, so registering just the application does not leave the
container looking up to date, and registering both does not count twice.
Catalog: four detectors verified on real containers, following the rules
in the file. vaultwarden and immich gain docker fallbacks for installs
where the native marker does not exist. netalertx is new — note its
repository is netalertx/NetAlertX; the Docker Hub namespace 404s.
technitiumdns is new and uses Technitium's own update endpoint rather than
GitHub releases: its marker reads 15.4 while the release tag is v15.4.0,
and _version_tuple compares (15,4,0) > (15,4) as an update that would
never clear.
Verified live against ghcr.io (Immich 3.1.0), docker.io (Vaultwarden
1.37.2) and lscr.io (Radarr 6.3.0.10514-ls314), plus postgres:16, which
correctly reports no version because official images carry no labels.
Exercised end to end on Proxmox VE 9.2.4 with NetAlertX reporting
26.6.3 -> 26.9.0. 31 new unit tests cover the resolution rules, every
refusal, the delegation contract and the container-to-image pairing.
VAIO73
MacRimiandGitHub
MacRimi
github-actions[bot]
ProxMenuxBot
byGarcia