* pocketbase-bot: accept the var_ names and two missing fields
cpu, ram, hdd, os and version were already reachable, but only under
the PocketBase names. People type what the ct scripts call them, so
"/pocketbase <slug> var_ram=4096" was rejected as an unknown field
while "ram=4096" worked. The bot already carried the mapping as
RESOURCE_TO_CT_VAR, for display only.
Normalise the keys in parseKVPairs, so both the field=value path and
the method path accept them, along with disk and memory as the other
two names people reach for. Matching is case-insensitive.
pin_reason and last_update_commit exist on the record and are worth
editing, but were not in ALLOWED_FIELDS. slug, script_created and
script_updated stay out: the first is the key the command looks the
record up by, the other two belong to the timestamp workflow. notes and
install_methods keep their own subcommands.
* pocketbase-bot: write the sync PR against the PR template
The sync PR body had its own Summary and Source headings, so the
autolabeler found none of the template checkboxes it looks for and the
PR came out with nothing but "needs triage". It also never referenced
the issue the command came from.
Write the body the way the template expects, with the Website update
box ticked, which is what a PocketBase sync is. Tested thoroughly stays
unticked and Tested on says not tested, because nothing here was run;
close-invalid-pr-template skips bot authors, so that costs nothing.
Reference the triggering number as Fixes when the command came from an
issue and as a plain mention when it came from a PR comment, where
Fixes would point the PR at itself. issue_comment carries both under
github.event.issue, so the new IS_PR_COMMENT tells them apart.
* ci: stop the .app header PR being closed as a new script
allowedBots carried "community-scripts-pr-app" but not the
"[bot]"-suffixed name GitHub actually reports, and the check is an exact
match, so the exemption never applied to it. push-app-to-main is listed
both ways; this one was not. PR #17304 was closed as an untested new
script submission because of it.
generate-app-headers.sh empties ct/headers, tools/headers and
vm/headers and writes them again, so every run reports those files as
added. That is what the autolabeler's new-script rule looks for, and
the vm rule matched vm/headers too. Exclude the header directories from
both, and skip them in the close workflow's own added-file fallback, so
the label cannot come back by another route.
Checked against minimatch with the shipped config: header files get
neither label, ct/*.sh, install/*.sh and vm/*.sh still get theirs.
* pocketbase-bot: label the sync PR as a bugfix
The sync corrects CT defaults that no longer match the PocketBase
record, so bugfix describes it better than website update, which is
meant for metadata changes on the site itself.
* frigate: move to 0.18.0
Bump the pinned release and follow the four build changes that matter
outside of Docker.
ffmpeg: 0.18 ships 8.0 as the default and keeps 7.0 and 5.0 alongside
it, so /etc/frigate.env has to name all three or the s6 run script
resolves a version that install_deps.sh never downloaded.
go2rtc: pin to v1.9.14, the version the Dockerfile fetches. "latest"
happened to work but shipped whatever AlexxIT had tagged that day
rather than the build Frigate was tested against.
OpenVINO: requirements-ov.txt dropped tensorflow and openvino-dev, and
build_ov_model.py no longer uses the Model Optimizer. omz_tools is
therefore gone, which left the first two branches of the labelmap
lookup dead; download coco_91cl_bkgr.txt the way the Dockerfile does.
Intel media driver: 0.18 builds intel-media-va-driver-non-free from
source for Battlemage, because the prebuilt noble/trixie packages need
a glibc that bookworm does not have. Run it before install_deps.sh, as
the deps-rootfs stage does, and drop the jammy repo the build adds so
the later trixie pull for libva2 is not resolved against it.
The remaining build scripts and requirement files changed too, but
those come from the checked-out tree and follow the version bump on
their own.
* frigate: stop the detector config from replacing the base config
The install wrote a config with mqtt, cameras, auth and detect, then the
detector branch wrote the file again instead of adding to it. Both
branches used a single redirect, so everything above them was discarded
and the result had neither mqtt nor cameras. frigate/config/config.py
declares both as Field() without a default, in 0.17.2 as well as in
0.18.0, so Frigate rejected the file and came up in safe mode:
mqtt - Field required
cameras - Field required
Move hwaccel_args into the base block, since both branches set it, and
append the detector and model sections instead of overwriting.
Checked by generating both branches and parsing the result: the OpenVINO
path yields auth, cameras, detect, detectors, ffmpeg, model, mqtt, the
CPU path the same without detectors, and the test camera survives in
both.
* frigate: write the config in one place
Only the detector and model section depends on the CPU check, but the
file was assembled in three heredocs writing to the same path. That
shape is what produced the safe-mode config: both branches used a plain
redirect and discarded everything above them.
Pick the variable part into DETECTOR_CONFIG first and write the file
once, so the layout is visible in one block and no branch can replace
what came before it.
Output is unchanged, verified by generating both branches from the
previous commit and from this one and diffing: byte-identical.
* Update default RAM and disk values in frigate.sh
Navidrome 0.61.x created cache, artwork and plugins under the data
folder whenever any navidrome command ran, including the ones the deb
postinstall runs as root. 0.62 stopped creating them that way but never
corrected the owner, so a container that passed through 0.61.x carries a
root-owned artwork directory. 0.64.0 is the first release to write into
it and fails with
writing image store: mkdir /var/lib/navidrome/artwork/hashed:
permission denied
which breaks newly resolved album covers and every playlist cover, since
those are composites.
Repair the three directories the same way navidrome/navidrome#6143 does:
the entries themselves only, since they were created empty, real
directories owned by root only, and chown -h, because the navidrome user
owns the data folder and could otherwise plant a symlink.
The trailing || true is not in the upstream copy and is needed here.
Their postinstall runs without set -e, while update_script runs under
catch_errors, which sets -Ee with an ERR trap. find exits 1 as soon as
one starting point is missing, so on a container without a plugins
directory the update would abort.
Fixes the report in #17247, confirmed there by the reporter's stat and
dpkg log: artwork is root-owned and dated to their 0.60.3 -> 0.61.1
upgrade.
* HomeAssistant-OS (VM): Refactor for improved readability
Refactor script to improve readability and maintainability. Updated function calls and variable assignments for better clarity.
* Modify author line in haos-vm.sh
Updated author information in the script header.
* Refactor Home Assistant OS version selection and caching
Refactor advanced settings dialog for Home Assistant OS version selection and simplify image caching logic.
* tor-snowflake: read the Go version from the module root
Snowflake is a single Go module with go.mod at the repository root, not
under proxy/. The install read /opt/tor-snowflake/proxy/go.mod, which
has never existed, so grep failed and the version came out empty. Go
was then fetched from
https://go.dev/dl/go.linux-amd64.tar.gz
with no version in the name, and the install aborted at that download.
Read the directive with a single awk instead of grep piped into awk.
The pipeline tripped catch_errors, which sets -Ee with pipefail and an
ERR trap, the moment the file was missing; that is the "exit code 2
while executing command awk" line that preceded the download failure.
Fall back to latest when the file cannot be read, so a future upstream
move of go.mod costs a Go version that is newer than the one pinned
rather than a failed build. setup_go already resolves latest, and also
resolves a bare major.minor to its newest patch, so a go directive
without a patch level stays fine.
Verified against v2.14.1: go.mod sits at the root and declares go
1.24.0, and the awk yields 1.24.0 from it and latest from a missing
file without tripping the ERR trap.
* Refactor GO_VERSION extraction in tor-snowflake.sh
Updated the method of extracting GO_VERSION from go.mod and set a default value if not found.
* Refactor: TrueNAS VM (core / improve functions / performance)
* Boot TrueNAS from sata0, the disk it actually installs to
The boot order named scsi0, which never exists: the system disk is
sata0, and imported passthrough disks start at scsi1 because SCSI_NR is
pre-incremented. So the order always fell through to ide2 and the
installer came back up after every reboot.
* Fix URL in truenas-vm.sh script
Docmost pins its package manager, so running pnpm goes through corepack,
which asks before fetching the pinned version:
Corepack is about to download .../pnpm-11.25.0.tgz
? Do you want to continue? [Y/n]
The pnpm calls run under $STD, so with verbose off the question is
never shown and the update sits on "Configuring Docmost" waiting for an
answer nobody can see. With verbose on the same update completes,
because the prompt is visible and gets answered, which is exactly what
the reporter observed.
Set COREPACK_ENABLE_DOWNLOAD_PROMPT=0 before the pnpm calls in both the
install and the update, the way fifteen other ct scripts and ten
install scripts already do.
One checklist now lists containers and VMs side by side, each tagged with its
type, with separate ALL CT / ALL VM entries. VMs stop via qm stop and are
removed with qm destroy --purge --destroy-unreferenced-disks.
Two fixes fall out of the rework: the destroy exit status is read from the
background job rather than from the spinner, so failures actually surface, and
a guest is only stopped once its deletion is confirmed.
bullseye left security.debian.org when its LTS ended, so the hardcoded
libssl1.1_1.1.1w-0+deb11u8 filename now 404s. Scan the security, the
security-archive and the archive pools and take the newest build on offer
for the host architecture.
Closes#17104
A container keeps the slug it was built with, so a renamed ct/ script leaves
the updater looking for a name that no longer exists.
Reported for pbs, renamed to proxmox-backup-server in 0e5f663df. The Alpine
merge on 2026-08-18 retired 29 more names the same way, so every container
installed from an alpine-* script before that date hits this too.
Candidates are only accepted when the target script really exists, so an
unknown slug still errors rather than running some other app's updater.
Fixes#16989
* FileFlows Node: Stop Spinner before read -rp
* fileflows: update install path for Node->Agent rename, detect service unit dynamically
* fileflows: tolerate no pre-existing fileflows units when checking for the new Agent unit
Replaces the misc/build.func bootstrap with the two-root core loader.
No other change: the engine contract (var_*, update_script, start,
build_container, description) is identical on both engines.
versitygw.sh additionally gains a trailing newline; it lacked one.
With this batch every ct/ script is on the core engine.
Replaces the misc/build.func bootstrap with the two-root core loader.
No other change: the engine contract (var_*, update_script, start,
build_container, description) is identical on both engines.
Replaces the misc/build.func bootstrap with the two-root core loader.
No other change: the engine contract (var_*, update_script, start,
build_container, description) is identical on both engines.
The 25 most-installed ct scripts still on misc/build.func, by install
count: homarr, influxdb, sonarr, radarr, networkoptimizer, authentik,
stirling-pdf, seerr, prowlarr, searxng, wordpress, zabbix, omada,
homeassistant, pegaprox, crafty-controller, iventoy, homelable,
flaresolverr, metube, netbird, casaos, obsidian-livesync, npmplus,
wazuh. ProxmoxVE goes from 115 migrated to 140.
Same three checks as #16749, since "migrate" has meant more than a line
swap before:
- None of the 25 has an alpine-* variant, so there is no merge to do.
- None references misc/ outside its bootstrap line.
- Two functions exist only in misc/ (_gl_asset_urls,
_send_abort_telemetry) and none of the 25 calls either.
So it is one line per script. Every head is byte-identical to the ones
migrated earlier, each diff is exactly 3+/1-, and all 25 parse.
Worth watching: scanopy is not in this set but sits at 32.9% success,
and casaos at 49.4% and networkoptimizer at 48.6% are in it. If those
two move, the engine is one changed variable among others.
vars.APP_ID named two different apps: 1065612 (community-scripts-pr-app)
here, 1108144 (app-header-generator) in ProxmoxVED. Folding both into
one GHAPP_HEADERS_ID would have pointed this repo at the wrong app, so
the PR app gets its own name.
* docker-vm, pve-privilege-converter: source from GitHub raw instead of gitea mirror
* docker-vm, pve-privilege-converter, vm-core.func: source from GitHub raw instead of gitea mirror
* Post the command that tests a ct/ or install/ change
Reviewing a script change meant working out the URL yourself, and the
obvious guess is wrong: curling the branch URL alone gives you the ct/
script from the PR and the install/ script from main, because each script
pins _CS_DEFAULT_URL to main and that pin is what fills
COMMUNITY_SCRIPTS_URL when it is unset. Frequently the install script is
the only thing that changed.
So the comment spells out both lines, per changed app.
Only for scripts already on the core bootstrap. The older one-liner
resolves everything from ProxmoxVE/main and ignores the variable, so a
command built for it would install main and look like it passed -- worse
than no comment. Those are named instead, with what to do about them.
pull_request_target for fork PRs, and nothing from the PR is checked out
or executed: the file list and the bootstrap line come from the API, and
a branch name that is not [A-Za-z0-9._/-]+ stops the run rather than
reaching a fenced code block.
* Update .github/workflows/pr-test-command.yml
Co-authored-by: Sam Heinz <sam@samheinz.com>
---------
Co-authored-by: Sam Heinz <sam@samheinz.com>
* calibre-web: point cps at explicit db path, src-layout package has no default anymore
* calibre-web: fix empty-library bootstrap, migrate pre-existing library on update
* Teach the PocketBase bot every field
The bot covered most of script_scripts but not categories, has_arm,
execute_in or app_vars, so those had to be edited by hand in the admin UI.
categories is a relation, so names are resolved against script_categories
and ids are accepted too - a copy out of the PocketBase UI works either way.
app_vars is a JSON column and rides the code-block 'set' path with its
content parsed rather than stored verbatim, so readers get an object.
cpu/ram/hdd/os/version live inside install_methods, and '/pocketbase immich
hdd=25' is how people ask for them. They now route to the default (non-Alpine)
method instead of being rejected as unknown fields, the reply names which
method was touched, and the value syncs into ct/<slug>.sh like the 'method'
path already does.
* Add screenshots and type to the bot
Two gaps were left. type is a relation to z_ref_script_types, so it needs the
same name-to-id resolution as categories - people write "ct", not a
fifteen-character id. Screenshots had no command at all.
The screenshot subcommand hands the URLs to the frontend's /api/screenshots
rather than fetching images inside a workflow. That endpoint already checks
the content type and size and attaches the file to PocketBase; doing it a
second time here would be a second set of bugs. It needs
SCREENSHOT_IMPORT_SECRET, and says so plainly when it is missing instead of
failing halfway.
slug stays deliberately out of reach. It is the URL, the JSON filename and the
ct/<slug>.sh path at once, so renaming it is a migration rather than an edit,
and the help text now says that instead of leaving people to wonder.
* Move the top 25 scripts onto the core engine
The engine work of the last few days reaches 30 of 561 ct scripts, about 5% of
ProxmoxVE traffic: retry on engine downloads, exit 227 instead of a misfiled
dpkg error, the umask fix that stops a hardened host producing containers apt
cannot resolve in, the TMPDIR guard, the toolchain restore. All of it has been
sitting where almost nobody runs it.
All eighteen at once rather than in waves. A slow rollout does not exercise the
paths only some scripts take, and broad exposure is what surfaces bugs -- a
deliberate call about release risk.
Checked before touching anything, because "migrate" meant far more than a line
swap last time:
- None of the eighteen has an alpine-* variant, so there is no merge to do.
- No script references misc/ outside its bootstrap line.
- Of the 61 functions that exist only in misc/, none is called by any of them.
So it is one line per script, and every head is now byte-identical to the ones
migrated earlier. With these, ProxmoxVE goes from 30 scripts on the core engine
to 48 -- and from roughly 5% of traffic to the majority, since these are the
ones people actually install.
Two to watch: immich sits at 44.7% success and vaultwarden at 42.1% before
this. If their numbers move, the engine is one of two changed variables rather
than the only one.
* Move update-apps onto the core engine
Entry 11 of the list and the only one that is not a ct script, so it was left
out of the previous commit. It is a host tool: it never used build.func at all,
it sources misc/core.func and misc/api.func directly.
The swap is therefore two lines rather than one, and worth checking rather than
assuming. It uses exactly five engine functions -- header_info,
init_tool_telemetry, msg_info, msg_ok, msg_error -- all present in the core, and
both files load standalone, which they had not had to do before: everywhere else
they arrive through build.func.
That completes the list. All 25 now run on the core engine.
Fixing this one matters beyond the migration: update-apps is what drives
unattended updates across every container on a host, and it is the path where
PHS_SILENT was being ignored (#16593). It now gets the engine that honours it.
* netbox: serve on plain HTTP too, port 80 forced HTTPS redirect broke reverse proxies
* Update netbox-install.sh for Apache configuration
Modify Apache configuration to serve NetBox on port 80.
* Set default ProxmoxVE raw URL in UPS script
Define `_CS_DEFAULT_URL` at the top of `ct/pve-ups.sh` and remove the outdated bootstrap comments, aligning the script with the newer URL-default pattern used for script sourcing.
* Branch var_cpu by OS where the two variants actually differ
The merged scripts branch var_ram, var_disk and var_version on var_os but
left var_cpu above the branch, so both variants inherited the Debian value.
On Vaultwarden that meant an Alpine container was told it wanted four cores
for an update that runs apk and restarts a service.
Only seven scripts are affected. Comparing each merged script against the
alpine-* script it replaced, 22 of the 29 already had the same CPU count on
both sides, so moving the line there would be churn with no behaviour change.
These seven did not:
docker, forgejo, ironclaw, syncthing, transmission, zigbee2mqtt 2 -> 1
vaultwarden 4 -> 1
The Alpine values are the ones those scripts carried before the merge, read
back out of the deleted alpine-* files rather than picked. The Debian arm
keeps what it has now.
Needs the matching core change: until build.func derives var_os inside a
container, an update never reaches the Alpine arm at all.
* fix: source url
* Update script to source build functions from URLs
---------
Co-authored-by: Tobias <96661824+CrazyWolf13@users.noreply.github.com>
DocuSeal switched to its own PDFium fork (upstream commit "adjust pdfium",
2026-08-15) and now attaches functions that only exist in that build, e.g.
FPDFPage_GetAnnotCountRaw from the added fpdf_annots_raw.h. The generic
bblanchon/pdfium-binaries library the script installed does not export them,
so lib/pdfium.rb raises FFI::NotFoundError while Rails eager-loads and both
docuseal.service and docuseal-sidekiq.service fail to start:
Unable to load application: FFI::NotFoundError: Function
'FPDFPage_GetAnnotCountRaw' not found in [libpdfium.so]
Install the library from docusealco/pdfium-binaries instead, matching the
upstream Dockerfile. It is only published as a musl build (DocuSeal's image is
Alpine based), so the musl runtime is installed and its library directory is
added to the loader search path - the shared object needs "libc.so" (musl) at
dlopen time. Verified on glibc: the library loads, resolves the raw annotation
functions and renders pages correctly.
The update path now refreshes PDFium too, so existing containers are repaired
by running "update" even when DocuSeal itself is already up to date.
CanbiZ (MickLesk)