* Trilium: keep the installed version when an update download fails
update_script deleted /opt/trilium before fetch_and_deploy_gh_release, so a
failed download (GitHub answered 502 for the v0.107.0 archive) left the
container without Trilium, and every later update stopped at "No Trilium
Installation Found".
CLEAN_INSTALL=1 lets core clear the target only after the archive has
downloaded and extracted. If the deploy still fails, the service is started
again on the installed version and the pending marker makes the next update
retry. The install check now looks for trilium.service, so containers broken
by the old order recover with a plain update.
The "database" backup copied /opt/trilium/assets/db, which only holds files
shipped in the release; notes live in Trilium's data directory, which the
update never touches. Removed.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TjCs8G7xWJ12fzC1g4v81E
* Trilium: back up the notes database before updating
With the service stopped and before the new release is deployed, the update
copies document.db (and a -wal file left by an unclean stop) to
<data dir>/backup/backup-before-update.db. The copy is kept after the update,
overwritten by the next one, and listed on Trilium's Backup page. If it cannot
be written, the update does not start and Trilium is restarted.
An install whose TRILIUM_DATA_DIR is inside /opt/trilium is refused before
anything is stopped, since CLEAN_INSTALL replaces that directory.
The install check accepts /opt/trilium again as well as trilium.service, so no
install that passed the old check fails the new one.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TjCs8G7xWJ12fzC1g4v81E
---------
Co-authored-by: s31teg <23174852+s31teg@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* OpenCloud: bump to v9.0.0, configure the service account through env
- Pin v9.0.0 in the update and install scripts.
- Copy the service account from opencloud.yaml into opencloud.env as
OC_SERVICE_ACCOUNT_ID/OC_SERVICE_ACCOUNT_SECRET, which every service
falls back to (covers 9.0's new auth_guest). Replaces the sed block
that inserted sharing.service_account into the YAML.
- Rename GRAPH_SPACES_DEFAULT_QUOTA to OC_SPACES_DEFAULT_QUOTA.
- Remove the unused thumbnails/files cache.
- Add a commented OC_ENABLE_GUEST_LINKS block to opencloud.env.
* OpenCloud: warn on install if the service account cannot be read
Same check as the update path, in case the opencloud init output changes.
* OpenCloud: name the FQDN variables literally for unattended installs
The prompt loop builds var_opencloud_<server>_fqdn at runtime, so the PR
script test found no variables to pre-answer and the install exited 64.
---------
Co-authored-by: Sim Kai Long <2094270+SimKaiLong@users.noreply.github.com>
Debian's nginx security update now passes HTTP_HOST as $host in fastcgi_params, which drops the port. Wallabag, Investbrain, Invoice Ninja, Poznote and WatchState then redirected to port 80 and the browser got connection refused. Pass $http_host again after the include.
openresty.org stalls or refuses for part of the users and is the top cause of NPM failures; Alpine's distfiles serve the same tarball (same sha256 for 1.29.2.5).
upstream stopped shipping backend/node_modules in the release tarball
for any arch as of 3.3.1, so amd64 installs crash on boot with
ERR_MODULE_NOT_FOUND. run npm ci unconditionally in install and update.
CloudPanel (DevScripts PR 2614) installs through upstream's install.sh,
which core runs only after asking; CI cannot answer, so set
var_external_installer=yes for the build and the update.
A failed install on Incus also showed as passed: with dev_mode=keep the
engine still prints "Completed Successfully", and the line that gives the
failure away names the container ci-9000 there, not 9000, so the check
for it never matched.
They found the cause: marimo-team has an IP allow list, so GitHub
refuses the token from the lab with a 403 that core read as a rate
limit. Fixed in core. The quota line stays.
On the host the job token reads another repo's release fine (200), yet
installs and updates in the guest hit the anonymous limit. Probe there,
just before the update. Also drops the /tmp file named after $$.
marimo (ProxmoxVE PR 18672) hit the rate limit in its update although the
job token is passed as GITHUB_TOKEN there. Log what both the token and the
lab address have left at the start and when an update hits the limit.