Drop compiling libjxl, jpegli, libheif, libraw, ImageMagick and libvips from
source. The install pulls them from the base image the release's Dockerfile
pins and stops with a clear error when it can't (skopeo/umoci need Proxmox VE
9.1+). The update fetches and unpacks the new image before it stops or changes
anything, so a failure leaves the running install as it is, and installs it
only when the Immich release changes. ~/.immich_base_image records the image;
the old compile leftovers and ~/.immich_library_revisions are removed.
The website now posts each requested change as its own comment with the
command and two checkboxes. Ticking one edits the comment: Accept runs
the command as if a maintainer had typed it, Decline records the
refusal, and the issue closes once every change is decided. Only users
with write access can decide. app_vars is checked field by field before
it is saved.
* fix(whisparr-eros): set the first login and save it
Eros 3.6 turns on forms login and creates the user whisparr with a password that only appears in its log, so a fresh install showed a login nobody knew the password for (#18616). The install now sets a random password through WHISPARR__AUTH__INITIALPASSWORD for the first start only, removes it again, and saves the login in ~/whisparr-eros.creds; the password stays out of the logs.
* whisparr-eros: set the first password in the unit, write creds with cat
The tar copy out of the unpacked base image ran outside the if, so a failed or
partial unpack hit the ERR trap and aborted the install instead of compiling
from source.
* Paperless-AI: install and update to 4.0
4.0 is a rewrite as a single Node.js process. requirements.txt is gone,
so every install since its release on 2026-11-10 failed at pip, and an
update would have failed the same way.
Install and update now build it the way upstream's image does. An
update from 3.x removes the Python RAG service and rewrites the unit
without data/.env as its environment: 4.0 migrates that file on first
start and treats any setting passed as a variable as locked.
* Paperless-AI: migrate pre-RAG 3.x units too, drop comments
* fix(truenas-vm): improve ISO fetching logic and exclusions
* fix(truenas-vm): enhance ISO selection and error handling
* fix(truenas-vm): improve ISO selection and error handling
* fix(truenas-vm): enhance ISO lookup and selection process to filter by stable releases only and improve user dialog
The ARM runner's resolver cannot resolve ports.ubuntu.com since this
afternoon; apt then hangs. The name is now looked up over DNS-over-HTTPS
when needed, and a failing apt prints the runner's network state.
Setting is_dev to false left platforms as DevScripts had them, usually empty or pve only, although a container in ProxmoxVE is synced to Incus from then on. 36 production records had no platforms because of it. The migration now sets pve and incus for containers, pve where nothing was set, and amd64 for an empty architectures.
- ARM runs on GitHub's ubuntu-24.04-arm runner (free for public repos), set
up by arm-runner-setup.sh: Incus 6.0 LTS from Zabbly (key pinned, APT
pin), screenshots through Playwright. ARM rows are shown but never count:
not in the headline, the check, the review or the approval, and the job
runs with continue-on-error.
- var_arm64=no is not built on ARM; unset is built with var_arm64=yes.
- arm-script-test.yml builds one script on ARM by hand, posting nothing.
- Update test: only the versions update_script checks are set back.
- The screenshot browser gets a writable HOME.
* fix(pocketbase): install and update from the new host
PocketBase moved from GitHub to ganigeorgiev.com in October 2026; the GitHub repo is archived at v0.40.6 and gets no new releases. Install and update now read the latest stable tag from the new release API and download the zip from its fixed download path. The update no longer runs pocketbase update, which older binaries point at GitHub.
* fix(pocketbase): query the release API with curl_with_retry
Containers without /etc/apt/sources.list.d/debian.sources (Incus images) got
testing.sources from the install, but the update only looked in debian.sources
and ran sed on the missing file. Check every *.sources file, and write
testing.sources like the install does when debian.sources can't take it.
- vm/ scripts run unattended (VM_UNATTENDED=1) on the Proxmox hosts, IDs
9100-9199, one VM per host at a time. Checks: success line, the VM keeps
running, a DHCP lease, and a ${VM_IP} URL answering 200. Console
screenshot from QEMU when there is no web UI. ARM64-only scripts are
listed as not built.
- Every ct/ script now builds on pve and incus (new incus-test runners,
incus-ci-host-setup.sh); the comment gets a Platform column.
- Spinner frames are dropped from logs and live output.
- The sweeper ignores and removes locks of killed jobs (older than 3h).
3.4.0 adds FUTO Backups, which run restic from PATH and use restic-proxy when it is there. Both are built from the refs the server Dockerfile pins, and an update only rebuilds them when those change. pigz is installed, which database backups now use for compression when present. mise skips its own restic, and the extism-js fallback follows the js-pdk version in mise.toml (v1.7.0) instead of a fixed v1.6.0.
Move the build-only toolchain (autoconf, automake, libtool, cmake, meson,
ninja-build, gcc-13) into the fallback branch and drop the unused ccache,
so the prebuilt-library path no longer installs them. The runtime library
-dev packages stay in the common path because they pull the shared
libraries the prebuilt libvips loads at runtime.
Also restore the base-images clone and source-dir creation in the fallback,
which the previous commit dropped.
The installer compiles libjxl, jpegli, libheif, libraw, ImageMagick and
libvips from source, which takes 15min-2h. These are the same artifacts
Immich already ships in its ghcr.io/immich-app/base-server-dev base image.
Pull /usr/local out of that image (pinned by digest, parsed from the
release's server/Dockerfile) with daemonless skopeo + umoci and reuse it.
Verify vips is usable before committing to this path; on any failure, or
with IMMICH_LIBS_FROM_SOURCE=1, fall back to the unchanged source compile.
A version-matched ~/.immich_library_revisions is written so the update
script does not needlessly recompile the libraries.
Update test: after a passing install, every version file the install
wrote (~/.<app>) is set back to 0.0.0-ci and the container's own
`update` runs unattended, pulling the PR's ct/ script, so
update_script's real path is exercised. It must report "Updated
successfully" without errors, and services and URL are checked again.
Apps without a version file get "nothing to update", a GitHub rate
limit counts as not tested. Separate Install and Update columns in the
report; a failed update requests changes like a failed install.
The job's read-only token is passed to the build so release lookups are
not rate limited (all test containers share one public IP).
/test <app> in an issue builds that app from main, install and update,
and answers in the issue; no check run, review or approval.
A second test host (pve-ci-2) now serves both repositories. The setup
script registered runners as pve-ci-<repo> with --replace, which would
have made it take over the first host's registration; they are now
named <host>-<repo>.
Anyone with write access can push to a push-app-to-main branch, and the
bot's approval counts toward the main ruleset's required review. The
approval now only covers what a migration contains: if the PR changes
anything besides ct/*.sh and install/*.sh (workflows, misc/, ...), it is
tested as usual but not approved, and the result comment says why.
PRs that the push-app-to-main app opens on a branch of this repository
(scripts moving over from DevScripts, reviewed and tested there,
writable only by writers) are now built on pull_request_target for
every push, pinned to the pushed commit; an event for an older push is
ignored. When every script passes, the bot approves the tested commit,
which counts toward the main ruleset's required review; a new push
drops it as stale, and a later run that does not fully pass withdraws
it. Untestable (CPU) and setup errors are never approved. All other
PRs still need /test and are never approved.
A /test on a PR that only changes vm/ (#18578, TrueNAS) found nothing
to build and stopped without a word. The gate now always answers:
"Nothing to build" with the reason for each file (vm/ not covered,
tools/ not built, install/ script without a ct/ script of the same
name, more than 10 scripts), and when some scripts are built the rest
are listed as "Not built" in the progress and the result comment.
File names from the PR are put in code spans with backticks and line
breaks removed.