Default it to "Nemu FM" and derive the tab title, logo, and dynamic
document title from one STATION_NAME constant for easy renaming.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
If the stream drops (server restart, network loss), the player now retries
goLive() on error/ended/stalled with a delay that doubles each failure, capped
at 30s, and resets once audio flows again.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
yt-dlp's flat extraction mis-attributes bandcamp uploads: on a label
account the artist becomes the label, and many files ship untagged. After
a download, guess (artist, title, album) from the richer download metadata
and the "Artist - Title" filename shape, confirm against MusicBrainz, and
write ID3 tags into the file. Existing tags are respected; MusicBrainz is
primary, filename parsing only a fallback.
The file's tags are the single source of truth: the provider reads them
back on the next pick, so the corrected identity (incl. MBID) is in place
before the scheduler keys and anti-repeat-checks the track, and cache
reuse inherits it too.
Fetchers now return (path, Track) so the corrected metadata flows back.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
/download now accepts ?file=<name> to fetch any cached file, not just the
current track. History entries carry that filename token (via /history), so
the web UI renders each aired track as a download link. A shared
serve_attachment helper validates the request (basename-only, real audio file,
no hidden/.part files) before streaming it; LRU-evicted tracks return 404.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
On the pinned savonet/liquidsoap:v2.4.5 image, `source.on_track` no longer
exists as a module function — on_track is a method on the source value. The
old `music = source.on_track(music, …)` form fails type-checking, so the whole
radio.liq is rejected and the stream container never starts. Register the
handler in place via the method form instead.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Add a read-only field with the bare /radio.mp3 URL and a copy button
(clipboard API, with a select+execCommand fallback), so the stream can
easily be opened in VLC or another external player.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Firefox kept resuming the <audio> element from a stale buffer after a
pause, drifting behind the live point. Load the stream with an anti-cache
query parameter, and on resume-from-pause reconnect to live rather than
replaying the buffered audio.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The fallback played the whole /cache directory, which at cold start holds
only the 2-3 tracks being pre-fetched — so it looped them until the
request.dynamic buffer filled. Restrict the fallback to tracks already
aired: the ingest daemon exposes them at GET /fallback.m3u (played_at set,
still on disk), and the stream fetches that into a local /tmp/fallback.m3u
that playlist watches. Cold start is now silent (assumed) instead of a tight
loop, and a mid-stream drain degrades across the whole listening history.
A local file (not a remote playlist URL) is used to avoid Liquidsoap's http
resolver mis-sniffing the response as text/html; mime_type is forced so an
empty header-only m3u still parses.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Add Drone pipelines building both container images per architecture and
merging them into multi-arch manifests under registry.nemunai.re/radieo.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Give the outgoing HTTP clients (Navidrome, MusicBrainz, ListenBrainz) a
transport-level retry budget (RADIEO_HTTP_RETRIES, default 2) so a brief
connection blip on a remote endpoint no longer drops a fetch or a lookup.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Add a Docker healthcheck probing ingest's /healthz (via python, since the slim
image has neither curl nor wget) and make the stream start only once ingest is
service_healthy, so it never briefly falls back on an empty cache at boot.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Handle SIGTERM (and SIGINT) to stop the HTTP server cleanly instead of being
killed after Docker's grace period, so `docker compose down` returns in ~1s and
the queue, HTTP clients and database are closed properly. shutdown() runs on a
helper thread since it must not run on the serving thread.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Serve a small web page at http://<host>:8000/ (stream/index.html) from the
Liquidsoap harbor, alongside the /radio.mp3 stream. It shows the track
currently on air and refreshes it from a /nowplaying JSON endpoint, fed by the
broadcast source's live metadata — accurate even though the ingest daemon runs
a track ahead (prefetch).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The fallback playlist probed every file in /cache, including .gitkeep and
in-progress .part downloads, logging ffmpeg "Invalid data" warnings at startup.
Filter candidates with check_next to keep only real audio files and skip hidden
ones, removing the noise.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Add a third playback source: a ListenBrainz recommendations Atom feed. Each
suggestion already carries a MusicBrainz recording MBID, title and artist, so
it is keyed directly by MBID (source-agnostic identity, no extra lookup) and
resolved to a concrete file — Navidrome search3 first, then a yt-dlp
ytsearch1: fallback.
- providers/listenbrainz.py: parse the Atom/HTML feed, anti-repeat on the MBID
key, resolve Navidrome-then-yt-dlp. Feed may be an http(s) URL or a local
path (for testing).
- subsonic.py: add search_songs (search3) for resolution.
- canonicalizer.py: short-circuit when a Track already has an MBID, so
feed-provided MBIDs are trusted and MusicBrainz is not hit.
- __main__.py: wire the provider in; register the yt-dlp fetcher as a
resolution backend even when the yt-dlp source is off; close providers on
shutdown.
- config/compose/.env.example: RADIEO_LISTENBRAINZ_URL + weight.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Give tracks a source-agnostic identity so the same song from different
sources no longer replays in a loop.
- Canonicalizer resolves (artist, title) to a MusicBrainz recording MBID
(no API key; ~1 req/s, descriptive User-Agent, best-effort). Hits and
confirmed misses are cached in SQLite; transient errors are not.
- Track.key becomes mbid:<id> when resolved, else a normalized
name:<artist>|<title> fallback — still source-agnostic.
- Scheduler now owns the authoritative anti-repeat on the canonical key,
canonicalizing the drawn track with a bounded retry; providers keep a
cheap recent-locator filter to limit retries.
- db: canonical_cache table, history.locator column with migration for
existing databases, recent_locators().
- Canonicalization can be turned off via RADIEO_CANONICAL_ENABLED=0.
Verified: MBID hit/cache/miss, cross-source key collapse, scheduler
dodging a recent play, schema migration, and full stack (Navidrome +
yt-dlp) with zero Python tracebacks and a valid 192 kbps MP3 stream.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Insert a 3s crossfade after the fallback so it applies to both
daemon-driven transitions and cache fallbacks, before mksafe.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Add a second playback source and a weighted scheduler mixing it with
Navidrome:
- Scheduler picks a provider by SOURCE_WEIGHTS, falling through to the
others when one has nothing ready, so no source can stall playback.
- YtdlpProvider reads a hand-maintained config/urls.txt; container URLs
(playlist/album/label/artist) are flat-extracted and one entry is
drawn at random, honouring the anti-repeat window. Adds Track.source_url.
- YtdlpFetcher downloads bestaudio via the yt-dlp library, reusing the
atomic hidden-temp-then-rename pattern; Liquidsoap decodes the result.
- Queue now dispatches to a fetcher registry keyed by backend.
- Sweep orphaned download temp files on daemon startup (leftovers from a
killed container otherwise pile up and trip the stream fallback).
Verified end-to-end: yt-dlp opus decoded and served as 192 kbps MP3, and
the 3:1 default mix observed in play history.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Replace the directory-scan queue with a real ingestion pipeline:
provider -> fetcher -> cache -> ready queue, driven by a background
prefetch thread.
- subsonic.py: minimal OpenSubsonic client (salted-token auth,
getPlaylists/getPlaylist, raw streaming download).
- providers/navidrome.py: pick tracks from a playlist (by name or id),
with anti-repeat and periodic playlist reload.
- fetchers/subsonic.py: atomic download into the shared cache.
- db.py: SQLite state — append-only play history (anti-repeat + stats)
and cache_files LRU retention (keep the N most recently played).
- queue.py: prefetch buffer + retention on play; graceful degradation
to the stream's local-cache fallback when no source is configured.
- api.py: GET /next now carries real title/artist metadata.
- Config via .env (Navidrome credentials), persistent state/ volume,
httpx dependency.
Verified end-to-end against a live Navidrome: playlist resolved,
tracks downloaded and broadcast, retention and history correct.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Add the Python `ingest` container exposing `GET /next`, which returns the
next track as an annotated Liquidsoap URI (or an empty body when nothing is
ready). Liquidsoap switches from a static playlist to a `request.dynamic`
source pulling from the daemon, with the local cache as fallback and mksafe
for guaranteed continuous output.
For now the daemon just cycles through the files already in the cache; the
download providers (Navidrome, yt-dlp, ListenBrainz) come in later milestones.
Also commit the implementation plan (PLAN.md).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Liquidsoap (v2.4.5) container that plays the /cache directory in random
order and broadcasts it over HTTP at :8000/radio.mp3 (MP3 192 kbps).
mksafe guarantees a continuous stream (silence when the cache is empty).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
