The version came from scraping technitium.com, which failed for some hosts ("Command substitution exited with status 1", 6 runs on 4 hosts in three days) and has no rate-limit fallback. The GitHub release carries the same version and goes through the engine's release lookup.
* fix(caddy): skip the xCaddy test build below 2 GB RAM
The container defaults to 512 MB (256 MB on Alpine), and Go's linker was
killed in xcaddy build: 14 runs on 6 hosts in a week. The build only
produced an unused binary, so below 2 GB xCaddy is installed without it
and a warning says why; install and update alike.
* Change directory to /opt before building xCaddy
The install replaced the file, which dropped JAVA_OPTS with install.data_dir, so creating a tenant failed. The DB settings are now appended, and an update restores the defaults on existing containers. The package upgrade keeps the edited file. Fixes#18483.
The server unpacks its frontend into /tmp/nexterm-sea-*. systemd-tmpfiles removes it after 10 days while the process keeps running, and every page returns 404. Fixes#18488.
npm 12 rejects --prod, and --production still starts PLANKA through nodemon. Starting node app.js directly needs HOME: Sails 1.5.18 only looks for the project's .sailsrc when HOME is set, and systemd sets none without User=, so the app lifted with the wrong paths and returned 503. The unit now sets NODE_ENV and HOME, and an update repairs existing units even without a new release. Follow-up to #18477.
* MeTube: install Deno without its shell setup
With -y, deno's install.sh also fetches jsr:@deno/installer-shell-setup to edit the shell profile, and when jsr.io did not answer the install stopped ("JSR package manifest ... failed to load"). Deno goes to /usr/local/bin, which is on PATH, so the shell setup is not needed; CI=1 keeps the installer from asking for it on a terminal.
* Modify Deno installation command
Update Deno installation command to include CI environment variable.
* OPNsense VM: size the temp space check to the unpacked image
The script refused to run unless TEMP_DIR had 20 GB free, a figure left
over from when the download went there too. The download now goes to the
image cache and only the unpacked FreeBSD image (2.5 GB) lands in
TEMP_DIR, so hosts with 8 to 16 GB free in /tmp stopped with exit 214.
Check the size xz reports for the image instead.
* OPNsense VM: say why the EFI disk could not be allocated
pvesm's error went to stderr only, so the two exit 220 runs in telemetry
carry no reason. As msg_error it reaches the log and the trace, e.g.
"disk image '.../vm-969-disk-0.qcow2' already exists" after an earlier
run left its disk on a storage that qm destroy does not clean.
* Clean up comments in opnsense-vm.sh
Removed comments about image cache and size calculation.
Advanced mode prefills the LAN VLAN with 999. On a bridge with no
physical port that is not VLAN-aware (an internal vmbr1, the usual LAN
choice) pve-bridge cannot tag the tap, and qm start failed with "no
physical interface on bridge 'vmbr1'". The tag is dropped there with a
warning, for LAN and WAN alike.
The scripts base still pointed at DevScripts, where this script was
removed on 10-09, so every run was reported as a DevScripts run.
Portainer and the agent are written as Quadlet .container units, which
need Podman 4.4. Debian 12 ships 4.3.1, so systemctl start portainer
failed with "Unit portainer.service not found". Without Quadlet both
prompts are skipped with a warning.
* fix(truenas-vm): let a failed ISO lookup reach its own error
When download.truenas.com did not answer, the lookup handed an empty
path to an associative array (bad array subscript) and the pipeline
failed under pipefail, so the run ended with exit 1 at the SELECTED_ISO
line instead of the "Could not find a stable ISO" message. The index
fetch now fails on HTTP errors and retries, empty lines are skipped, and
the assignment no longer aborts on an empty result.
* fix(truenas-vm): use the engine's retry flags for the ISO index
* fix(proxmox-mail-gateway): take the repo suite from the container
The install hardcoded trixie, so a Debian 12 container got the PMG 9
repository and pmg-api could not resolve its trixie dependencies (held
broken packages). Suite and key now follow the container's codename;
bookworm still has PMG 8 with proxmox-mailgateway-container.
* fix(proxmox-mail-gateway): read the codename inline
* fix(immich): take the watchfiles wheel and retry GeoNames longer
The v3.3.5 lock pins watchfiles 0.21.0, which has no wheel for the
Python 3.13 machine-learning venv, so every ML install downloaded a Rust
toolchain and compiled it. Where static.rust-lang.org reset the
connection, uv sync failed three times. --upgrade-package watchfiles
takes 1.3.0, an abi3 wheel; only uvicorn's reload mode uses it.
download.geonames.org dropped connections on 11-03/04 and the first file
failed after three 10 s connects; those four calls now retry six times
with a 30 s connect timeout.
* Immich: bump release version to v3.3.6
v3.3.6 disables onnxruntime's telemetry in immich_ml, which runs from
source here, so nothing else changes. The lock still pins watchfiles
0.21.0, so the wheel upgrade stays.
The compose file from main now uses ARCANE_ENCRYPTION_KEY and
ARCANE_PROJECTS_DIRECTORY, which the latest image (2.17.3) does not read,
so it panicked on a missing ENCRYPTION_KEY. Arcane 3 in turn refuses to
start when only the old names are set. Install and update now keep both
names in compose.yaml, which also carries existing installs over to 3.
The .env was never referenced by the compose file, but compose parses it,
and the unanchored ENCRYPTION_KEY sed turned the renamed line into
ARCANE_#ENCRYPTION_KEY and broke every install. New installs no longer
download it; update repairs the broken line.
The FreeBSD image boots in under a minute when both NICs get DHCP, but
dhclient waits 60 seconds on each NIC without a DHCP server, which is the
normal case for a firewall's LAN before setup. On vx-pve an isolated LAN
took 110 seconds; the reporter of #18464 measured 210. The keys then went
into the boot output. The image prints its login prompt on the serial
console as well, so the script now reads serial0 and continues when it
appears, up to 15 minutes, or waits the old 3 minutes if the console
cannot be read.
Tududi 2.0.0 dropped SQLite and refuses to start next to a populated 1.x
database. New installs set up PostgreSQL and let start.sh create the
schema; db:init would sync the tables without SequelizeMeta and make
db-prepare treat the database as an old 1.x one. Installs without
DATABASE_URL stay on 1.15.0, the final SQLite release, which also brings
back installs that already updated to 2.0.0.
Uploads lived in backend/uploads, which CLEAN_INSTALL wiped on every
update; they move to /opt/tududi-uploads.
n8n-nodes-base 2.46.0 moved @napi-rs/canvas-linux-arm64-musl and
-linux-x64-musl from optional into hard dependencies. npm enforces
os/cpu/libc on those, and no combination satisfies both arm64 and x64,
so every install fails with EBADPLATFORM. n8n 2.46.4 reached the latest
tag on 2026-11-03 and the failures started four minutes later.
--force skips the platform check. The matching gnu binary is still
installed through @napi-rs/canvas's own optional dependencies, so canvas
keeps working.
* Wiki.js: install 3.x
Wiki.js 3 lives in requarks/wiki, needs Node.js 26 and creates its own ltree and pg_trgm extensions in the wiki schema, and starts with node --no-experimental-webstorage backend. dataPath moves out of the app folder, which every update replaces; the admin password is generated and saved to ~/wikijs.creds. Updates follow upstream's guide and put Puppeteer back when it was installed. 2.x containers keep getting 2.x updates from requarks/wiki-v2, since 3.x cannot upgrade them in place, and are told how to export and import.
* Refactor credential output to use heredoc syntax
close-invalid-pr-template.yml now comments and labels (missing pr template,
stale) instead of closing immediately, clears itself once fixed. stale_pr_close.yml
gains the engagement-gated/keep-open/maintainer-reset behavior already used in
DevScripts, so a template-stale PR only closes via its own re-check.
request-script.yml discussion template gets an optional Justification field;
a new workflow verifies stars/age against GitHub for the submitted repo and
either does nothing (eligible), labels investigation (ineligible + justified),
or closes the discussion (ineligible, no justification).
Moved the AI Assistance section to the end of pull_request_template.md.
vzdump and pct restore ran with their output thrown away, so a failed
backup only said "Backup failed for container N" (exit 235). That is the
second most common update-apps failure and none of them can be told
apart. Keep the output in the run log and put its error line in the
message, e.g. "CT is locked (backup)".
var_backup_storage was matched with grep -w, so "truenas" passed when
only "truenas-backup" exists and vzdump then failed on a storage that
does not exist. Match the whole name, and list the valid ones in the
error instead of a spinner line after it.
Name the app in "Update failed for container N", so a failed update
says which app's update failed. Declining the start dialog now exits 0
like the other tools instead of failing with whiptail's 1.
* OpenWrt VM: read the stable release from downloads.openwrt.org
The version came from a line of text on the openwrt.org wiki front page.
One host got that page without the line three times in a row while
others resolved fine at the same moment, and the run ended with exit 115
before anything was downloaded. .versions.json on the download server
is what the OpenWrt firmware selector reads and names the stable release
explicitly.
* Update vm/openwrt-vm.sh
Co-authored-by: Tobias <96661824+CrazyWolf13@users.noreply.github.com>
---------
Co-authored-by: Tobias <96661824+CrazyWolf13@users.noreply.github.com>
clear exits 1 for a TERM the host has no terminfo entry for, such as
xterm-ghostty or xterm-kitty over SSH, and when TERM is unset under cron.
The tools that run under errexit died on it in header_info within a
second, before showing anything, which telemetry counts as
"exit_code=1 | General error" with no step for update-lxcs, update-repo,
fstrim, clean-lxcs, execute and storage-share-helper.
Fall back to the escape sequence the engine's _cs_clear uses.
Both whiptail prompts ran bare under errexit, so No on the first question
or Cancel in the template list ended in the error handler and was reported
as a failed run (3 of 5 failures in three days).
Monitor-All is now stopped right before the download instead of at the
start, so leaving at a prompt no longer leaves it stopped.
The user was the first passwd entry with a UID from 1000 up. On a Proxmox
host with Ceph that is ceph (UID 64045, shell nologin), and su - ceph fails
with "This account is currently not available". Accounts without a login
shell are skipped now.
apt update on a host with the enterprise repository and no subscription exits
100 on the 401. apt_update_safe retries, says why and carries on, as the other
scripts already do since #17462.
The script ran its own apt-get update before installing
libguestfs-tools, so one broken or stale repository on the host (a 404
for a hand-added proxmox.com or debian.org entry) ended the run with
exit 100. vm_prepare_cloud_image installs the tools itself and carries
on when an update fails, so the install can still succeed from the
repositories that did update.
Cinephage depends on camoufox instead of camoufox-js, whose bin is called camoufox, so the install stopped at "node_modules/.bin/camoufox-js: No such file or directory" (6 runs in three days).
- VectorChord moved to supervc-stack/VectorChord. With the GitHub API
rate limited, core reads the release from github.com, finds no asset
under the old name and stops with "No suitable .deb asset found".
- mise installed every tool in Immich's mise.toml. Java (for the Dart
OpenAPI generator), terragrunt and opentofu are not needed for the
plugin build, and the Java index timed out on slow links.
- GeoNames failed after the 1-3 h library build; it now downloads first.
- Stop on x86-64 CPUs without SSE4.2: the web build ends there with
"Illegal instruction" after the libraries are built.
The script replaced the engine's error_handler with one that sent the
failed command as the exit code, so every failure reached telemetry as
exit 199 "exit code not reported", without the failed command or
category. Its signal traps sent words instead of codes the same way.
It also ran its own unguarded apt-get update for libguestfs-tools, so a
host with the enterprise repository and no subscription (401) stopped
there. vm_prepare_cloud_image installs the tools itself, tolerating a
repository that fails to update.
The container menu lists stopped containers too, and pct exec on one exits
255: 12 of 41 failures in three days. Start it the way add-netbird-lxc does
and wait for its default route.
The install runs under bash, which an Alpine container from the template does
not have; lxc-attach then exits 127 before the Alpine branch is reached.
Install bash there first, and stop with a clear message on any OS other than
Debian, Ubuntu or Alpine instead of failing on apt-get.
WGDashboard builds its frontend with Vite during the install, and at 512 MB the kernel killed node: "Failed to compile frontend", then wg-dashboard would not start (11 runs in three days). The build finishes at 1 GB.
* Zabbix: install on Debian 12 again
The install and the update fetched Zabbix's release package for Debian 13 and php8.4-pgsql, so a Debian 12 container stopped with "Unable to locate package php8.4-pgsql" (12 runs in three days). Zabbix publishes 7.0 and 7.4 for Debian 12 too: take the container's Debian version and the distribution's php-pgsql.
* Zabbix: leave the update as it was
The update refuses Debian 12 on purpose, so its release URL stays as it was.
Nothing checked that the host is a Backup Server. On a PVE host the
first step stopped only because tar found no /etc/proxmox-backup (exit
2); answering "no" to that backup would have gone on to delete
/etc/apt/sources.list.d/*.list, add the PBS repositories and
dist-upgrade. Stop with the same message post-pbs-install uses.
Discourse enabled libvips image processing by default and its worker refuses anything older than 8.18; Debian 13 ships 8.16, so the service failed to boot after an update and on fresh installs. Build the latest libvips release from source like discourse_docker does, on install and on update when a new release is out.
The add-on runs on the Proxmox host, where the enterprise repository without a
subscription answers 401 and apt update exits 100 (2 of 3 failures in three
days). apt_update_safe retries, points at the enterprise repository and
carries on; the netdata repository was fetched fine.
Cancel in the first menu was reported as a failed run.
wazuh-install.sh writes everything to ~/wazuh-install.output, so a failed run left only "bash exited with status 1" in the log and the telemetry (8 runs in three days, three of them with the default resources). The end of that output now goes into the install log.
pct fstrim exits 255 when the container is locked, for example while a
backup runs ("CT is locked (backup)"). Under errexit the command
substitution ended the whole run there, before the output check below it
and before the remaining containers, and telemetry filed it as
"DPKG: Fatal internal error". Print the result and go on instead.
The skip message printed $name, which is only set once a container has
run the command. When a non-Debian container (Alpine, for example) came
before any Debian one, set -u ended the run there with "name: unbound
variable" and exit 1; later it printed the previous container's name.
The target list came from storages with "images" content, but pct
restore puts the container's rootfs there and refuses a storage without
"rootdir" ("storage 'X' does not support container directories"). A
storage that holds VM disks only was offered, and picking it ended the
run after the backup with "Conversion failed" (exit 235).
On a PVE 9 host without deb822 .sources files the script asks whether to
disable the legacy .list sources, then read the answer from $? on the
next line. Under set -e a "No" ended the whole run with exit 1 right
there, so the "Kept legacy sources as-is" branch never ran and nothing
after it was offered. Test whiptail in the if instead.
Since the release of 2026-10-28, Server/ holds only FileFlows' own assemblies and the rest comes from ../Libraries. Copying the archive over an older install left the old third-party DLLs in Server/, which the runtime loads first, and the server fails with 'Could not load file or assembly Microsoft.Data.Sqlite, Version=10.0.13.0'. Upstream's own upgrade removes Server, FlowRunner, Libraries and Agent the same way; Data stays.
The MAC-matched network snippet is a workaround for Ubuntu 26.04's NIC rename. Without a storage for snippets the run failed after the VM was built; Debian now keeps the static ipconfig0 Proxmox already set, as it did before September.
Ubuntu 26.04 needs the MAC-matched network snippet, and without a storage for snippets the run failed only after the VM was created. It now stops before downloading the image.
* Norish: fetch Obscura from the repo Norish pins
Norish now builds Obscura from norish-recipes/obscura (pin.env OBSCURA_REPO); the v0.2.5-norish.1 tag does not exist in h4ckf0r0day/obscura, so installs and updates stopped with a GitHub 404.
* Norish: run the v0.26 server on Node.js 26
v0.26.0 dropped dist-server and starts with node --import tsx server/index.ts from apps/web, and requires Node 26. Updates rewrite the unit.
PVE 9.2 reports the import as 'Successfully imported disk as unused0:...', which the local parser missed, and its pvesm fallback matched the VMID column, so imports on local-zfs ended in 'Unable to determine imported disk reference'.
* Aurral: build the pnpm workspace and run backend/server.ts
v2.12.0 moved from npm workspaces to pnpm and the backend entry to server.ts, so npm ci --workspace frontend found no workspace. Updates point existing units at server.ts.
* Aurral: install only the frontend and backend packages
The workspace root's prepare script runs vp config, which needs git; the upstream image filters the install to the two packages as well.
* Aurral: install git for the workspace prepare script
The root prepare script runs vp config, which aborts when git is missing and only notes the absent .git directory otherwise.
* PS5-MQTT: build the yarn workspace and run the packaged server
v1.8.0 turned the repo into a yarn 4 workspace with Plug'n'Play: ps5-mqtt/package.json is gone and npm install there fails. The server now runs from the packaged tree with its own node_modules, as upstream's image does; updates move the unit's working directory.
* PS5-MQTT: give the build 2 GB of RAM
Building the client and server bundles with rspack is killed by the OOM killer at 512 MB.
* Drop the legacy engine copies from misc/
Nothing has loaded misc/*.func since ct/, install/, vm/ and tools/ moved to
community-scripts/core. misc/api.func stays until vm/owncloud-vm.sh gets its
DevScripts rewrite.
misc/images/ moves to .github/assets/. The tombstone stub for removed scripts
sourced misc/build.func and called variables(), which core's PVE engine does
not define; it now sources core.func.
* Drop misc/api.func
vm/owncloud-vm.sh was its last caller; DevScripts holds that script's rewrite
onto core/pve/vm-core.func. Merge this after the promotion.
* fix(discourse): track the Ruby version Discourse requires
Discourse main now pins `ruby "~> 4.0.0"` in its Gemfile, so `bundle install`
aborts with exit 18 on containers built against the hardcoded Ruby 3.4.4. The
update path never touched Ruby at all, so every existing install breaks.
Discourse ships no .ruby-version, so read the pinned version from
Gemfile.lock's RUBY VERSION stanza and hand it to setup_ruby — on install
(repo is now cloned before the runtime setup) and on update (after the git
pull, so a bumped requirement is picked up). Also silence bundler's root
warning, split the update into separate progress steps, and skip the
unused global Rails gem.
* move ruby into msg block
* move git
Removed comment about Ruby version in install script.
Since 2.7.1 the deb's postinst appends a live service: block with the secret, so uncommenting the stock service: line left two service keys and Vikunja fell back to its defaults. Updates merge the duplicate on existing installs.
When a merged PR came from a DevScripts migration, the telemetry service
re-attributes that script's DevScripts records. Needs the
TELEMETRY_ADMIN_TOKEN secret; without it the step is skipped.
Stalwart v1.0.0 (26 Oct 2026) ships a new data store layout, so v0.16.x
deployments cannot be upgraded in place. Data has to be copied into a new,
empty v1 store with stalwartlabs/stalwart-migrate, which requires v0.16.27
as its source version and needs an interactive prepare step plus an offline
final cutover.
The update path pulled "latest" and would have swapped the binary for v1 on
existing containers, leaving the mail server unable to start on its old store.
Pin install and update to v0.16.27 instead and surface the reason through
check_for_gh_release. Containers that already pulled v1.0.0 are put back on
v0.16.27 by the next update run.
update_script ran apt upgrade and reported success while MeshCentral stayed on the version it was installed with, so 1.1.x installs never got 1.3.1, which fixes an actively exploited flaw. Compare the installed package with npm and install the newer one. Fixes#18205
mise 2026.10.19 defaults lockfile_mode to generate, which locks every platform during mise install. Immich's mise.toml enables the lockfile, and Java 21.0.2 has no linux-arm64-musl build, so the update failed on any host. Fixes#18236
MikroTik RouterOS asked for machine type, CPU model and disk cache and used none of them, nor its discard/ssd flags. HAOS and TrueNAS dropped the chosen disk cache, UniFi OS Server the discard/ssd flags. Fixes#18211
hermes update refuses a checkout its caller does not own, so the
root-run update stopped working for every install. The installer and
hermes update now run as hermes, and the dashboard becomes a user unit
next to the gateways, so Hermes Desktop and the dashboard can update
and restart it themselves. libatomic1 is installed up front, the one
package Hermes would otherwise ask sudo for.
update moves existing containers over once: the system unit becomes a
user unit, leftover root processes are stopped and /home/hermes is
handed back to hermes.
* post-pmg-install: hide the subscription nag in the PMG 9 mobile quarantine
PMG 9 replaced pmgmanagerlib-mobile.js with the Yew/WASM app in
pmg-mobile-quarantine-ui, so the mobile hook no longer found its file.
The status check is compiled into the WASM, so the helper now injects a
fetch wrapper into the mobile index template that reports the
subscription as active to the app. The API response itself is unchanged.
All patches move into /usr/local/bin/pmg-remove-nag.sh behind a single
apt hook. "no" now also removes the mobile hook and the helper, and the
reinstall covers pmg-mobile-quarantine-ui so its template is restored.
Fixes#16217
* post-pmg-install: reinstall the mobile quarantine UI only where it exists
pmg-mobile-quarantine-ui is not in the PMG 8 (bookworm) repository, so
naming it in the reinstall made apt abort there and skip the widget
toolkit reinstall as well.
tsc on the 5.0 server needs about 3 GB of heap; the 2 GB default left it 2.3 GB and the update died in the build. npm ci runs with --ignore-scripts like upstream, so better-sqlite3 keeps its prebuilt binary instead of an addon built against the running Node's headers. kitinerary-extractor sits in libexec on Debian, so link it into PATH where TREK looks for it.
The mobile UI patch removed the dialog and card only when their text
contained "subscription", so translated UIs (de: "Subskription", fr:
"Abonnement", ...) kept the nag. It also deleted DOM nodes Yew still
owns and polled every 300 ms for good.
The helper now injects a fetch wrapper into the mobile index template
that fills the empty subscription level of each node in the /nodes
response, so the app never shows the alert or card. The API response
itself is unchanged; the node list shows "-" instead of "No
subscription". The mobile UI package is reinstalled alongside the widget
toolkit so old blocks are dropped, and "no" also removes the helper.
* install: answer every prompt from the environment, survive a closed stdin
Every interactive prompt takes a var_<app>_<name> override and treats a
closed stdin like Enter (`|| true`), so an install keeps going after the
console that started it is gone. Values the script cannot do without end
with exit 64 and name the variable to set. Validation loops stop asking
when nobody can answer instead of spinning on an empty read.
* mongodb, paperless-gpt: accept yes as well as y
Upstream added PROXCENTER_SKIP_TYPECHECK for builds from a released tag;
the type check alone needs about 3 GB. A 4.6 GB heap on 6 GB of RAM got
the build killed.
An empty answer, or no TTY, ended the install with 64. The prompt now
defaults to 17 and takes var_postgresql_version; the Alpine list is
16/17/18, which is what 3.24 ships.
The lookup ran under pipefail with no timeout: a page without the link
ended the install at `head -1`, a stalled connection hung it for half an
hour, and the fallback version never ran. The installer takes
--accept-eula, --no-swap-check and --no-check-disk; piped answers
shifted with the prompts, and a "y" to the swap question tried swapon
inside the container.
Every failed install this week ran on Debian 12, where the installer
rejects the Podman that ships there. The default is 13; an override now
fails before the 785 MB download.
The Ollama release binaries carry no SYCL build, so the 2.8 GB download
was never used. Its handled failure was also what later systemctl
failures were reported as.
CanbiZ (MickLesk)
