The builder image is now multi-arch (linux/amd64 + linux/arm64, pushed
as srfplay-builder:latest after building all three plugin zips in each
variant), so jobs no longer need an amd64 host. runs-on moves from
linux/amd64 to ubuntu-latest, which draco-x86, freebox and oracle-a1
all carry, so a job no longer waits behind long jobs on draco while
the ARM runners sit idle.
.gitea/runner/cloud-init.yaml sets up such a runner on an Oracle Cloud
Always Free Ampere VM (Ubuntu 24.04): Docker, act_runner registered
instance-wide with the instance's label set, and a weekly image prune.
The registration token and the optional console password hash are
placeholders here and must be filled in locally before use.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The upload loop used IFS=$'\t', but the job steps run under /bin/sh,
which is dash in the builder image. dash does not expand $'\t', so IFS
became the characters $, \ and t and every path was split on "t":
"artifacts/srfplay_1.2.0.9.zip" reached curl as "s/srfplay_...", which
failed with exit 26 after the release had already been created. This
broke both the v1.2.0 release and the nightly.
cut -f3 is plain POSIX. Checked by running the upload and manifest
steps under the builder image's dash with curl stubbed.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The single net8.0 build compiled against Jellyfin 10.9.11 still loads on
10.11 and 12, so channels and browsing keep working, but 10.11 changed
APIs in ways that only fail at runtime. A binary audit of the DLL against
the real server assemblies found 9 broken references on both 10.11 and
12 (none on 10.9/10.10):
- TaskTriggerInfo.Type became an enum. GetDefaultTriggers throws, so
Jellyfin silently gives every SRF task a 24h fallback trigger. The
recording scheduler therefore ran about once a day and scheduled
recordings never started.
- ILibraryManager.GetItemList changed its return type: the expiration
check crashed on every run.
- Jellyfin.Data.Entities.User moved (IUserManager.Users, GetUserData,
SaveUserData, PlaybackProgressEventArgs.Users): resume cleanup returned
500 and the playback-stop guard threw on every stop.
Jellyfin 12 also rejects the legacy X-Emby-Token header with an empty
401, which the recordings UI surfaced as "Unexpected end of JSON input".
The pages now send Authorization: MediaBrowser Token="...", verified on
10.9, 10.11 and 12.
The plugin now multi-targets net8.0/net9.0/net10.0 against Jellyfin
10.9.11/10.11.0/12.0.0 (the oldest package of each line, so each DLL
loads on every patch release of it), with JELLYFIN_10_11_OR_GREATER and
JELLYFIN_12_OR_GREATER for the differences. Trigger construction moves
into Utilities/TaskTriggers so the version switch lives in one place.
.gitea/scripts/build-plugins.sh builds one zip per generation. jprm has
no targetAbi flag, so it rewrites build.yaml per flavour and restores
it. Jellyfin installs the highest version whose targetAbi it satisfies,
so the flavour code goes into the last version segment:
release v1.2.0 -> 1.2.0.9 / 1.2.0.11 / 1.2.0.12
nightly -> 1.0.<date>.<run>09 / 11 / 12
The PR, nightly and release workflows use the script and publish one
manifest entry per zip. Verified end to end against a three-entry
manifest: 10.11 installs the .11 build, 12.1 installs the .12 build.
The existing .NET 10 builder image builds all three unchanged.
Supporting changes: the test project multi-targets too, because jprm
publishes the whole solution for one framework at a time, and it drops
its Microsoft.Extensions.Logging 8.0.1 pin, which is a NU1605 downgrade
under the newer Jellyfin packages. CA1873 (new in the .NET 10 analyzers)
is silenced beside CA1848. The release manifest step checks out
origin/master like the nightly does since 7e1973a.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
`git checkout master` after `git fetch origin master` switches to the
LOCAL master -- the one actions/checkout left at this run's own SHA --
not the ref just fetched. A nightly build takes long enough that master
has usually moved by the time it finishes, so the manifest gets its new
entry prepended to a stale copy and the push is rejected:
! [rejected] master -> master (non-fast-forward)
Run 1045 failed this way, and the log is unambiguous about it: "Your
branch is behind 'origin/master' by 1 commit, and can be fast-forwarded"
appears one line before the commit that could not be pushed.
`checkout -B master origin/master` is the fix, and it is right rather
than merely convenient. This step appends one version entry to the
manifest as it stands on master, so the base it edits has to be master.
Editing the copy this build happened to start from and then forcing it
through would silently drop whatever landed in between.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
dtourolleandClaude Opus 5.5