The NAS is the only host this container runs on. Building linux/arm64 as well
meant emulating it under QEMU on every release for a consumer that does not
exist, so both the second platform and the QEMU setup step are dropped.
Nothing in the image is architecture-specific; restoring arm64 is a one-line
change if it ever gains a home on the Pi.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Debian's ffmpeg package depends on libavdevice, which can capture from and
render to X11, Wayland, SDL and OpenGL. That pulls in Mesa, and Mesa pulls in
LLVM for its software rasteriser -- 127 MB of it -- plus Z3 at 27 MB and a
speech synthesiser at 28 MB, in an image whose only job is to encode MP3s on a
headless NAS.
Alpine's ffmpeg brings none of that. The runtime image goes from 576 MB to
189 MB. Small X11 and Wayland client libraries remain, but they are kilobytes
rather than megabytes.
The test suite passes unchanged inside the new image, which is the point of
running it there: the encoder under test is now a musl build from a different
distribution, and the tests cover exactly the tag, cover art and container
behaviour that could have differed. Verified separately that a CBR 256 encode
comes out at the expected bitrate with non-ASCII tags intact.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The deployment target is a container on TrueNAS Scale, so the flake sat on no
path between the source and the NAS. It was carried over from a sibling project
where the flake is the deployment mechanism; here it only added a second build
path and a second dependency pin.
Worse, it tested the wrong thing: `nix flake check` ran the suite against
nixpkgs' ffmpeg while the shipped artefact contains Debian's, and encoder and
muxer behaviour is exactly what these tests cover.
The Dockerfile gains a `test` stage that installs pytest and runs the suite
against the image's own ffmpeg; a failing test fails the build. CI runs
`docker build --target test` in place of the host-based Python setup, and the
push build states `target: runtime` so the published image is the lean stage
rather than the last one in the file. The runtime image carries neither the
tests nor pytest.
The release step now calls python3 rather than python, since setup-python is
no longer in the job to provide the alias.
Removes package.nix, flake.nix and flake.lock. A dev shell is a `nix shell`
away for anyone who wants one, and the README says so.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Three defects in how the mirror tracked its source, all of which show up on a
library that something else reorganises.
Pruning probed the source tree for a mirror file's original name, trying each
known extension in turn. A source saved as .FLAC was never found, so its mirror
file was deleted as an orphan and re-encoded on the next pass, for ever.
Pruning now works from the set of paths the pass actually accounted for, which
cannot disagree with the walk over letter case or extension coverage.
Two sources could also claim one mirror path -- 01 Song.flac beside a leftover
01 Song.mp3, which is what an interrupted upgrade leaves behind. Both encoded
to the same destination, whichever finished last won the race, and every later
pass found the other one stale. The best-quality source now wins, ties break on
path, and the loser is logged.
The source mtime was read after encoding rather than before. A file still being
written when the pass reached it would be stamped with its final mtime while
holding truncated audio, and would never be revisited.
Adds regression tests for all three, plus the format-upgrade, album-rename and
whole-library-deletion cases, each verified to fail before the change.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Adopts the release scheme from legacy-email-proxy so both repositories behave
the same way: the version is derived from conventional commits since the last
v* tag, the image is pushed under the full version, the truncated major.minor
and major forms, and latest, and non-release builds are published as
sha-<short>. Multi-arch (amd64 for the NAS, arm64 so the same image runs on a
Pi). Authentication uses the PACKAGES_SECRET repository secret.
The release step also writes the computed version into pyproject.toml and
commits it as chore(release) before tagging, so the packaging metadata cannot
drift behind the release. It skips the commit when the file already carries
that version, which would otherwise fail the job after the image had been
pushed.
compose.yaml and the README now reference the published image instead of
instructing the NAS to build one locally.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Apple's Music app cannot read FLAC, so getting a lossless library onto an iPod
requires a converted copy somewhere. This keeps that copy beside the library on
a NAS rather than on a laptop, and keeps it current unattended.
Walks a source tree and reproduces it path for path as MP3: tags and cover art
carried across, already-MP3 sources copied rather than re-encoded, and mirror
files whose source has gone deleted along with any directories they emptied.
The source library is never written to.
Freshness is tracked by modification time -- an encoded file is stamped with
its source's mtime, so a file is stale exactly when the two differ. That keeps
runs idempotent without a database that could fall out of step with whatever
owns the library, which here is Lidarr.
Encodes go to a temporary file and are renamed into place, so an interrupted
run cannot leave a truncated MP3 that the next run mistakes for finished work.
A lock file in the mirror root prevents overlapping passes.
Ships as a Python package with a console script, a container image with a
TrueNAS Scale compose file, and a Nix flake providing the package, an overlay
and a dev shell. The test suite runs real ffmpeg encodes rather than mocks --
the failures worth catching are in what ffmpeg does with tags, cover art and
container formats.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
