1. enrich-buy-url.py re-queried every no-match track on every daily run (external,
rate-limited lookups over the whole backlog), which blew past its 30-min
timeout. Cache "tried, no match" with a BUY_URL_TRIED marker tag and skip
those for 30 days (--force still re-checks). This is what made "Look up buy
links" fail with TIMEOUT after 1800s.
2. Move maintenance:enrich_buy_url to run last (09:45) in the 9am block, after
the fingerprint index (09:25) and status report (09:30). enrich starting at
09:10 and holding the shared lock is what left "Rebuild fingerprint index"
skipped_lock.
3. pipeline-status.sh: guard the qobuz/app_id read with a file-exists check.
`< missing 2>/dev/null` still leaks the shell's redirection error (opened
before 2>/dev/null applies), so the daily report logged
"qobuz/app_id: No such file or directory" every run.
4. dedup confirm_and_apply: a candidate whose delete target no longer exists
(already removed by an earlier dedup/upgrade/hand) was filtered out and the
apply silently no-op'd, leaving it stuck in the queue with no way to delete
or clear it (the 100 gecs case). Now mark such candidates applied so a Delete
click clears them. Covered by tests/test_dedup_review.py.
Tests: 46 green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- R11: pin requirements.txt to the versions running in the image, so a rebuild
can't pull a breaking upstream release. Documented how to upgrade.
- R12: log rotation now also prunes job_runs rows older than 90 days (nulling
the manual_imports FK reference first), so the table doesn't grow without
bound. dedup/genre candidates are review state and left alone.
- U5: credential form shows non-secret fields (URLs, usernames, prefs) as plain
text so they can be verified while typing; only real secrets stay masked.
Dashboard IP card reworded from gluetun-specific to generic "Outbound IP".
- S9: enforce ALLOWED_EMAIL at the OIDC callback, before any user row or session
is created, instead of only on later requests.
Verified: pinned build resolves; prune deletes only old rows and keeps the
manual_imports record; credentials page renders text+password inputs; a
disallowed email gets 403 with no user row, an allowed one succeeds.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Root cause of a missed 9am notification: the read-only pipeline_status_report
runs at 09:30, but enrich_buy_url (09:10, --apply) held the single global
pipeline lock long past 09:30, so the status report recorded skipped_lock and
never sent its Telegram digest. build_fingerprint_index (09:25) was starved the
same way.
- run_job/run_lib_script gain use_lock; the status report runs with
use_lock=False since it only reads logs, pings Navidrome, and sends Telegram.
It can no longer be blocked by a long write job, and running concurrently is
safe.
- _lib was silently dropping its timeout argument, so maintenance jobs had no
timeout and a hung one held the lock until the next container restart. Thread
timeout through _lib -> run_lib_script -> run_job, cap enrich_buy_url at 30m
and build_fingerprint_index at 60m, and give the status report a 5m cap.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Security (P0):
- Remove committed session-secret default; auto-generate and persist a
random secret to the config volume when SESSION_SECRET is unset
(prevents forgeable session cookies / auth bypass).
- Validate playlist names to a safe charset and render sldl configs via
literal Python substitution instead of sed (closes a command-injection
and path-traversal path through playlist names).
- shlex-quote credential values written to shell-sourced env files, and
strip newlines from values patched into .conf files.
- Render playlist .conf files 0600; warn at startup if the master key is
co-located with the config volume; document keeping it separate.
Portability:
- Configurable timezone via TZ (default UTC) instead of hardcoded Edmonton.
- Remove personal defaults (navidrome user "andrew", ephemeral.club URLs).
- Ship generic example seeds; move the cross-album dedup keep-list and the
legacy playlist import to editable config files; drop the personal
_upgrade.csv.
- Generic VPN reference in docker-compose.snippet.yml.
First-run experience:
- Redirect to /setup instead of 500 when OIDC is unconfigured; surface a
missing master key inline; entrypoint exits with an actionable message
when the config folder is not writable.
- Add unauthenticated /health (JSON) and /setup (checklist) diagnostics.
Docs:
- Write docs/ARCHITECTURE.md and docs/MIGRATION.md (previously referenced
but missing); expand README with ownership, backups, advanced settings,
and migration guidance.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Settings is now one top-nav entry with Credentials and Jobs as sidebar
sub-pages (jobs moved from /jobs to /settings/jobs). Manual import page
replaces the bare file input and free-text filename field with a drag-drop
dropzone and a proper file-picker table. Genres page drops the "force"
checkbox for two explicit buttons (Preview / Fix genres now). Dedup's
"Scan now" runs both the file-naming and acoustic passes together, and the
table labels which pass caught each candidate instead of showing the raw
pass name. .btn-ghost gets a visible border so it reads as a real button
next to Delete/Danger actions instead of looking unaligned. Job names
throughout the UI are now human-readable instead of raw job_key strings.
Also includes the fpcalc exit-code fix from earlier (fingerprint index
was discarding valid fingerprints on files with a benign decode warning).
job descriptions
Dashboard: format breakdown (FLAC/MP3/etc. counts), approximate library
size (bitrate*length/8, same approximation `beet stats` itself uses --
verified to match its "100.5 GiB" output exactly against the real
library), and a credential auth-state summary per scope (configured y/n
plus, for Bandcamp, cookie expiry parsed locally from the stored cookie
jar -- deliberately not a live network probe like pipeline-status.sh's
Qobuz check, since this renders on every dashboard load).
Library: was one unfiltered page dumping all 4072 tracks. Added search
(artist/title/album), playlist and format filter dropdowns, and real
SQL-level pagination (LIMIT/OFFSET, not fetch-everything-then-slice).
beets_service gained count_items()/distinct_formats() to support this.
Playlists: cron_expr is still the stored/scheduled representation, but the
UI now shows and edits a plain daily time picker instead of raw cron
syntax -- every playlist schedule today is a simple daily HH:MM anyway.
playlist_service.cron_to_time()/time_to_cron() convert at the router
boundary; verified round-trip against all real playlist cron values.
Jobs: each maintenance job now shows a short one-line description of what
it actually does (MAINTENANCE_JOB_DESCRIPTIONS), and "next run" is
formatted consistently with playlists' time style (HH:MM, with a day
qualifier for non-today runs -- maintenance jobs can be weekly/monthly,
unlike playlists' plain daily schedule).
Verified end-to-end against the real production data (4072 tracks, 100.5
GiB, real credentials): stats/format-breakdown/search/pagination all
correct, all 7 credential scopes report configured with Bandcamp's real
cookie expiry (325 days), and a full authenticated page-render sweep
(dashboard, library with every filter combination, playlist detail,
jobs) all returned 200 with no template errors.
dedup-library.sh: additive --json (one NDJSON line per candidate deletion
to stdout, alongside the unchanged human log) and --only-paths FILE (in
--apply mode, only actually delete entries whose path is in FILE; without
it, --apply deletes everything as before -- existing direct callers are
unaffected). Without --only-paths the safety re-verification is free:
--apply --only-paths re-runs all 4 passes from scratch on every invocation,
so if a group's ranking changed since a scan (e.g. the old keep_path is
gone), the fresh pass assigns the previously-"delete" path the KEEP role
instead and the only-paths allowlist naming it is simply never consulted --
no duplicate ranking logic needed in the review service.
spotify-genre.py: additive --json emitting one JSON line per genre change
(dry-run or --apply) for genre_review_service to persist.
pipeline_runner.run_job_capture(): like run_job() but captures stdout as
text (still under the same shared lock, still writes a job_runs row) for
callers that need to parse structured output rather than just log it.
dedup_review_service.scan() persists dry-run candidates into
dedup_runs/dedup_candidates. confirm_and_apply() re-checks confirmed
candidates still exist before invoking --apply --only-paths, so nothing is
ever deleted without an explicit confirm -- matches the false-negative-
biased dedup preference. scheduler_service's maintenance:dedup job now
goes through this (still dry-run only, every day).
genre_review_service.run() wraps spotify-genre.py for both dry-run preview
and the real scheduled --apply run, persisting every run's diff into
genre_runs/genre_candidates either way -- genre writes keep their current
auto-apply behavior (low-risk, reversible, GENRE_LOCK-protected) but are
now reviewable after the fact. lock_artist_genre() gives a one-click revert
path when a run gets something wrong.
Added minimal routers+templates for /dedup (scan, review, confirm-and-
delete) and /genres (preview, review, lock-old-genre).
Verified end-to-end against REAL duplicate files (not mocked): built an
actual FLAC+MP3 duplicate pair in a real beets library, ran dedup-library.sh
--json and confirmed correct JSON output, verified --apply --only-paths
with an empty confirm list deletes nothing and with the real confirmed path
deletes exactly that file (DB + disk) while preserving the FLAC, and ran
the full dedup_review_service scan->confirm->apply flow through the same
fixture. genre_review_service and spotify-genre.py --json verified against
mocked/direct output (spotify-genre.py's own artist-genre lookup needs a
live Spotify API call, out of reach in this sandbox). Confirmed the full
app boots with all five routers registered.
pipeline_runner: subprocess execution wrapper replacing flock -n with a
single asyncio.Lock (deliberately one global lock, not per-resource --
matches the old flock's all-jobs-share-one-lock behavior rather than
over-engineering it). Records one job_runs row per invocation
(started/finished/status/exit_code/summary/log_path/triggered_by).
If the lock is already held, records status='skipped_lock' immediately
instead of silently dropping the run.
Found and fixed a real concurrency bug via testing: the obvious
asyncio.wait_for(lock.acquire(), timeout=0) idiom for a non-blocking
try-acquire is broken in asyncio -- the wrapping Task's first iteration and
the timeout-0 callback race with no guaranteed ordering, so it timed out on
literally every call, including the very first uncontended one. Fixed using
lock.locked() + acquire(), relying on acquire()'s fast path never
suspending when uncontended.
scheduler_service: AsyncIOScheduler with the default in-memory job store
(NOT SQLAlchemyJobStore -- it pickles job functions to persist them, and
the _lib()/_beet() factory closures here aren't picklable; MemoryJobStore
avoids this since register_all_jobs() rebuilds the whole schedule from code
+ DB on every startup anyway). All 19 maintenance jobs ported 1:1 from
/etc/cron.d/sldl-maintenance with their exact cron expressions; dedup
deliberately has no --apply (dry-run only on schedule, per the
false-negative-biased dedup preference). Playlist jobs are synced live from
the playlists table (sync_playlist_jobs), so add/remove-in-the-UI takes
effect with no redeploy -- wired into playlist_service's create/update/
delete. Maintenance job enable/disable persists to the scheduled_jobs table
so a paused job stays paused across a restart despite the in-memory job
store. trigger_now() supports "Run now" by invoking a job's function
immediately with triggered_by='manual', bypassing its schedule.
Wired into main.py's lifespan (start on boot, shutdown on exit).
Verified via integration tests: 36 jobs register correctly (17 seeded
playlists + 19 maintenance) with correct cron triggers; live playlist
create/delete correctly adds/removes its scheduler job; maintenance
enable/disable persists to the DB and takes effect live; trigger_now
correctly bypasses the schedule; pipeline_runner correctly handles
success/failure/timeout/concurrent-lock-contention with real subprocesses;
and the full app boots with the scheduler running and shuts down cleanly.
andrew