Since a Facebook cookie was added on prod, every reel came down titled
"Video": measured on dev, a signed-in read gives yt-dlp the placeholder
"Video" (an 18+ reel read as Chrome gives "Facebook") while the post's text
is still in the description. A new SourceAdapter.title hook takes the whole
first line of the description in that case, and both title writes (the early
fill and the finished download) go through it, so the row, the file name and
the share preview carry the post's text again. The line is not cut with an
ellipsis: normalize_title's NFKC turns "…" into "..." and collapses it to
".", which read as a finished sentence (seen in dev UAT).
The "Audio (MP3)" profile's extracted .mp3 was remuxed into an .mp4 by the
browser-safe step (the audio spec is clamped to container mp4 too) and served
as video/mp4, which iPad's WebKit will not play (prod job 268). Audio mode now
skips that step, so an .mp3 is served as audio/mpeg and an .m4a as audio/mp4.
Audio downloads version their cache identity on their own (v5), so a cached
remuxed audio re-derives without re-downloading every cached video.
Pause now re-reads the job under the same asset-then-job row lock as resume,
cancel, delete and redownload. Its unlocked write waited for a concurrent
cancel's lock and then set the released job back to paused, a holding status,
so a later delete released the same hold again and could free a file another
user's job still held (measured: the overlapping-request test ends 'paused'
instead of 'canceled' with the old read). The lock test now calls the route
handlers, so it also pins which helper each route uses.
Comments follow the code: resolve_url's contract names both verdicts and that a
failed page read is raised as itself; the inherited-failure diagnosis asks the
cookie gate as it stands when the job is claimed, so its advice fits the cookie
that job's Retry would use; the undecryptable-cookie warning no longer claims a
download goes ahead, since the worker parent reaches it only to diagnose.
A refused Facebook cookie on a group post is no longer "no video in this post":
measured on dev, Facebook answers with its login page (no post id at all), so the
adapter now raises its own POST_NOT_SHOWN verdict and the catalog explains it as
the refused cookie. The adapter owns both post verdicts and the "cannot parse
data" wall wording, so the Chrome retry and the catalog read one constant. The
group-post "no formats" entry is cookie-scoped (with a cookie yt-dlp only gets the
post's video link) and uses the adapter's own group-post test instead of a
"/groups/" substring.
Job transitions that move an asset's ref count (resume, cancel, delete,
redownload) re-read the job under a row lock, asset first, so two overlapping
requests can no longer release the same hold. get_or_create_asset loses an insert
race inside a SAVEPOINT instead of rolling back the caller's pending writes (Retry
lost its cleared error). Every worker-side job failure goes through
finalize_if_running and writes error_detail, so a cancel that landed first is kept
and a failure never shows an earlier one's copy. The worker asks the download
child's own cookie gate (credentials.usable_cookie) when a job inherits a failed
asset.
UI: the Facebook copy no longer points at a Retry button the failure window does
not have, says "paste" like Settings, and the HU wording is fixed
("Újraletöltés" on the re-download confirm). A category this client has no copy
for falls back to the generic sentence instead of a raw key.
Tests: route serialization of error_detail, the probed duration through a real
_download, the page-read cap and a failed read, a Retry off a pending asset, the
confirm-button labels, the lock and savepoint races.
Given a group post's link, yt-dlp either found no formats or took another
video from the page's feed data, so a "successful" retry of such a post was a
wrong download. With the user's cookie the Facebook adapter now reads the post
page, takes the single video node in the script blocks that carry the post id,
and downloads that video's watch link (the Chrome retry does the rest). No
single video in the post fails with its own message and the group-post copy,
reworded to say the post has no video; the post link is never handed to yt-dlp
as a fallback. Without a cookie nothing changes.
A Facebook or Instagram reel extracts with no duration, so the asset kept a
NULL duration_s although the file plays for minutes, and the Library row had
no trim action. The finalize probe now returns the duration with the codecs
(the file wins, info fills a gap), and runs before the sidecars so the .nfo
runtime gets it too.
Measured on dev: with the user's cookie some reels still answer "Cannot parse
data" and only extract when the request impersonates Chrome. Impersonation is
a second attempt behind that one error, only when a cookie is attached — links
that already download with the cookie keep their plain path, and without a
cookie the same error is the login wall.
curl_cffi is pinned inside yt-dlp's supported range (0.10-0.15; outside it
every target silently reports "unavailable"); the lock only gains the new
packages. A test fails the gate if the locked pair can no longer impersonate.
The login-wall copy tells the user to add their cookies and press Retry, but
Retry kept the job on the shared asset the worker never attaches a cookie to,
so it failed on the same wall. Resume now moves a pending/errored job onto the
asset enqueue would pick with the cookie (never off a file or a download in
flight).
Re-download salted its fresh asset with the job's previous asset id, which the
cookie gate cannot recompute, so even a cookie download rebuilt without the
cookie. The rebuild salt is now the new asset's own id (plus the user's cookie
salt when one applies), and is_authed_asset recognises it.
A failed Facebook download read "Download failed": yt-dlp words the login wall
(18+ reel, closed group) as "Cannot parse data", which no category matched.
Measured on dev with the owner's cookie before writing a rule — the same text is
the login wall without a cookie and a refused cookie with one.
- errors.diagnose(msg, url, cookie_attached) tries a host-scoped catalog first
(all-of markers, cookie state, link fragment) and falls back to the category.
First entries: facebook.login_wall, facebook.cookie_rejected,
facebook.group_post_video. The module doc is the recipe for the next entry.
- download_jobs.error_detail (0084) + backfill of the stored Facebook parse
failures that were unknown.
- The pipeline's failure branch passes the link and whether the cookie went
with it; resume and redownload clear code and detail (redownload never
cleared error_code).
- A job meeting an already-failed asset got the text but no code, so it read
"Download failed" where the first job said why. It is diagnosed again, with
the cookie state from credentials.is_authed_asset (now shared with the
cookie gate); the redacted network text classifies back to network.
- FE: failCopyKeys gives the row and the failure window one lookup — catalog
copy, else the category. Copy in EN and HU; a contract test pins every
backend entry to both locales.
Round five. 47 -> 34 -> 15 -> 7 -> 6 raw findings, and this round's sharpest was the
loop's own blind spot: **the branch did not pass the gate.** `prettier --check` is one of
its eight lanes and I had been running the other seven, so an unformatted `onError` from
round one sat on the branch for four rounds — `siftlode publish` would have refused it,
with no --skip flag to get past.
**The restart hint could still strand the restored data.** Round four said the scripts no
longer print a command they cannot resolve; they did. The COMPOSE_FILE fallback only fires
when `docker-compose.yml` is ABSENT, so in a source checkout it never fires and the hint
expanded to a bare, pasteable `docker compose up -d` — which reads the other file, whose
volumes are `pgdata`/`downloads` rather than `siftlode_pgdata`/`siftlode_downloads`, and
recreates the database on an empty one. When the file is unresolved the scripts now print
`-f <the compose file you started with>`: not runnable, which is the point.
**A comment naming the wrong owner.** `formats.py` still said a zero budget is "a refusal
for enqueue to make" — the premise round one disproved, three files away from the two
docstrings that now say the opposite. Anyone acting on it would delete the worker's refusal
and re-open the original inversion, so it now says where the refusal actually lives.
Also: the reset-flag rule was rewritten in three consecutive rounds and pinned by nothing —
two tests now cover a fragment losing its token, both by a pasted verify link and by
navigating away entirely; `backup.ps1` deletes a partially-copied dump before throwing, the
guard round four gave the docs but not the script it mirrors; and one comment that read as
if the asset were untouched by a line four below it.
Gate green, all eight lanes: ruff, ruff format, 3x tsc, eslint, prettier, knip,
vitest 687, pytest 950.
Round three, on round two's diff. The trend across the loop is 47 -> 34 -> 15 raw
findings; the ops half kept producing the sharpest ones because the gate cannot run it.
**`docker compose start` ignores a newly pinned IMAGE_TAG.** Round two removed the
automatic restart because it re-ran `alembic upgrade head` — and then told the operator to
run that exact command by hand, in both scripts and the docs. `start` relaunches the
container that already exists, which still carries the tag being rolled back FROM, so
pinning 0.79.0 and following the printed instruction brings 0.80.0 back up and re-applies
its migrations. Verified against a real daemon: after changing the pinned tag, `start` kept
the old image and only `up -d` picked up the new one. Both scripts and the doc now say
`up -d`, with the pin-first caveat where the operator reads it.
**The reset flag outlived its token.** Reverting round one's "each link clears the other"
went one step too far: a `#verify=` link pasted over a standing reset notice replaces the
whole fragment, so the reset token is gone — but `resetPending` stayed true, and being
checked first it hid the confirmation while promising a recovery that signing out could no
longer complete. The flag is now cleared when the fragment it describes is gone, which is
narrower than what was reverted and does not blank the signed-out card.
**A test that could not fail.** The shared-asset regression test seeded the asset as
`pending`, and `_reset_asset_pending` only acts on `downloading` — so the release half was a
no-op and deleting it kept the suite green. Seeded as the worker actually leaves it now.
Also: the malformed-URL guard answers with urlguard's structured detail rather than a bare
English string, so both halves of one typo class get the same localized copy, and the third
site in that module handing user input to `urlparse` was swept with it; the asset is
released after the job is terminal, not before; the Stats 403 toast gets the wizard handle
its own comment claimed it had; the pasted Windows recipe sets `$ErrorActionPreference`
inside the block (a scriptblock inherits the caller's, so a `SilentlyContinue` profile
swallowed every throw) and cleans up the container-side dump on the failure path too; the
restore warning moved to the warning stream so a `> log` redirect cannot file it away; and
two comments narrowed to what the code does.
Gate green: vitest 685, pytest 950.
Round two of the review loop, on round one's own diff. Two of its findings were
regressions the fixes themselves introduced, which is exactly what the loop is for:
**A per-user refusal was writing onto a shared row.** MediaAssets are deduped across
users, and the generic failure handler puts the message on the asset — right for an
upstream failure, wrong for "your storage is full". Two users queuing the same video share
one asset, so the first user hitting their cap failed the second user's job with the first
one's private reason, telling someone with an empty library to free up space. The refusal
now raises `_OwnerRefused`, caught separately: the job fails, the asset goes back to
`pending` for whoever else holds it. Pinned by a test that drives `_download` directly —
the refusal happens before extraction, so nothing needs stubbing.
**The restore scripts restarted the app, undoing the restore.** `docker compose start`
re-runs the container command, and that runs `alembic upgrade head` — so on the rollback
path this documents as recommended (restore the dump, THEN pin the older tag) the newer
image immediately re-applied the migrations the restore had just undone, and on a FAILED
restore it put the app back on a half-restored schema. Both are the opposite of why the
stop was added. Nothing is restarted automatically now; the scripts say what to run and
when, and the docs' snippets say the same.
**Reverted rather than patched:** the OnboardingWizard toast (the wizard already renders
that failure inline, next to a Retry button — three recoveries for one failure), and
`useVerifyEntry`'s "each new link clears the other". The latter blanked the signed-out
card's only banner, where nothing renders the reset flag to replace it. The two states are
independent again; what changes is that `dismiss` peels ONE layer, which is what the
original finding was actually about.
Also: `Stats` uses the YouTube helper rather than the bare one, so the route's English 403
detail no longer beats the translated fallback under a Hungarian UI; `urlparse` guarded
alongside `.port` (an unclosed bracket raised just as late), the reason renamed `malformed`
since "bad port" misdescribes it; three docstrings narrowed to what the code actually does
(the worker's byte check is per-job, not an aggregate ledger, and not on the edit path;
the redirect residual leaks a reachability oracle, which the note had denied; the
invalidation helper does not cover suspend); the pasted Windows recipe wrapped so `throw`
can stop it; the bash one-liner's failure branch no longer exits 0; and the tests that were
missing for the round-one fixes — the account panel's error paths, the invalidation set,
and the shared-asset case above.
Gate green: vitest 685, pytest 950.
The deep review of the branch raised 47 findings, 18 of which survived adversarial
verification. Grouped by what they actually were:
**The size ceiling inverted at its boundary.** `remaining_bytes` clamped to 0 for a user
already at their storage cap, and `build_ydl_opts` reads 0 as "no limit" — so the one
user over quota was the only one downloading with no ceiling at all. `check_enqueue`
could not catch it: it ran while the job was queued and the user was still under the cap.
The arithmetic moves into `quota.remaining_bytes` (where every other max_bytes decision
lives, and where it can be tested), and the worker now fails the job outright rather than
proceeding unbounded — as a non-retryable `quota_exceeded`, since retrying cannot help
until space is freed.
**Silent failures in the new account panel.** All three actions went through `quiet: true`
endpoints, which suppress the global error dialog on the promise that the caller reports
inline — with no `onError`, that promise was broken and a refusal rendered as nothing. On
the broken-SMTP instance this panel exists for, an admin would read silence as "the mail
went out". The co-invalidation set and the email-off predicate move to `lib/adminUsers`,
shared with the pending-access queue, so the two copies of these mutations cannot drift
apart again — which is how half of this was lost in the first place.
**A rollback recipe that destroyed data without working.** `alembic downgrade -1` steps
back one migration, but a release routinely ships several (0.80.0 shipped four): it drops
the newest one's columns and leaves the schema at a revision the older image also cannot
resolve, i.e. the crash loop it was meant to escape, minus the data. The dump restore is
now the recommended path, and the alembic route asks the target version for its own head
revision instead of counting backwards.
Also: redirects named in the urlguard docstring beside the DNS-rebinding residual they
share a fix with (SEC-17); a malformed port refused as `bad_port` instead of escaping as
a 500; the pull cooldown extended to the sibling route spending the same QuotaAction;
restore stops the app before dropping its tables; the PowerShell cleanup can no longer
decide the script's verdict on 5.1, nor report the wrong exit code; the documented
one-liners hardened to match the scripts; the registry's retention window stated where
pinning is advertised; the two banners made mutually exclusive so neither can shadow the
other; `hover:bg-hover` (a utility this project never generated) replaced by the
`selectable` recipe; the clipboard helper reused; and the sync cooldown's copy swept to
the two callers that discarded it.
Verified: gate green (vitest 678, pytest 949), and the panel, the row hover and the reset
banner UAT'd on localdev.
SEC-12: the download box accepts an arbitrary http(s) URL and yt-dlp fetches
it from the WORKER, so any approved user could aim the server at the compose
network, the LAN or a cloud metadata endpoint — and read the result off
job.error, where "connection refused" vs "HTTP Error 401" vs "timed out" is a
port scanner. downloads/urlguard.py refuses a URL whose host resolves to a
loopback/private/link-local/reserved address, at enqueue AND again in the
worker (DNS can move in between, and the worker is where the socket opens).
A network-class failure on a `url` job now reports one fixed sentence instead
of the raw transport text; the error CODE is still classified from the
original, so the localized explanation is unchanged. DNS rebinding is
explicitly NOT closed — that needs the connection gated inside yt-dlp — and
the module says so rather than implying more than it does.
SEC-13: `noplaylist` only demotes a video-inside-a-playlist URL to that video,
so a URL that is purely a playlist or channel still expanded into every
entry — one enqueue, unbounded downloads. `playlist_items` pins it to one.
Size was unbounded too: `check_enqueue` compares the CURRENT footprint against
the cap, and a download's size is unknown until it lands, so one job could
overshoot by any amount; the owner's remaining budget is now handed to yt-dlp
as `max_filesize`, applied AFTER the adapter merge so no source adapter can
raise a limit the quota layer set. A join's cut-list is capped at 100: an
accurate join emits two filters per segment into one ffmpeg filter_complex.
SEC-14: the YouTube budget is instance-wide, so one user holding down a
manual sync spent the day for everyone. Per-user (not per-IP — the cost
belongs to the account) 3-per-10-minutes limiters on the two bulk pulls,
subscriptions and playlists, with separate budgets so one does not close the
other. The review also listed /rss, /backfill and /enrich; E3 deleted them,
and search stays unlimited on purpose — it is interactive, not a bulk pull.
Gate: pytest 896 -> 924 (none skipped), vitest 671.
A downloaded video was silently demoted to a download-only "file" (no
in-app play/edit) whenever yt-dlp's info dict omitted vcodec/acodec — which
it does for Facebook and other sources. Finalize derived media_kind from
those nulls (`derive(None, None)` -> file), so a genuine H.264/AAC MP4
became non-playable in the library.
Fix: probe the finished file on disk (the ground truth) at finalize and
derive from that, falling back to the info dict only for a field the probe
couldn't read. This is the same ffprobe the upload ingest already used, now
extracted to a shared downloads/probe.py so the download worker and the
upload path share one classifier (and can't drift). Also backfills the
stored vcodec/acodec/width/height from the real file.
Tests: TestResolveMediaFacts pins probe-wins-over-info (incl. the Facebook
no-codec case and yt-dlp's literal "none"); upload's probe tests repoint to
the shared module.
Add the non-yt-dlp ingest path so a user-uploaded file joins the same download
library and inherits its function set. Any file type is accepted: media
(video/audio) gets the full set (metadata, share, play/edit), a generic file
(pdf/zip/docx/…) is download-only.
- MediaAsset.media_kind ("video"|"audio"|"file") — the stored discriminator that
gates library actions. Set at every finalize (download probe, editor carries the
source's kind, upload probe); consolidates the ad-hoc audio-detect that lived only
in links.py into one invariant (downloads.mediakind.derive). Migration 0076 adds
the column + backfills existing rows.
- downloads/upload.py: ingest_upload probes the file (ffprobe → codecs/dims/duration),
places it in a per-user .uploads tree, and creates an already-ready asset + done job
(the same cache-hit shortcut enqueue uses). Uploads never expire (irreplaceable).
- POST /api/downloads/upload: streaming multipart with the per-file cap
(download_upload_max_bytes, default 500 MiB) enforced on actual bytes read; quota
charges the known size up front (quota.check_upload).
- Security: a generic file is always served force-download (attachment +
application/octet-stream) on both the in-app and public endpoints, so uploaded
HTML/SVG can't execute inline. Edit + redownload reject non-media / upload sources.
- python-multipart added (first multipart endpoint in the app).
Each user stores their own login cookies (Netscape cookies.txt) for an
auth-walled download host (Facebook, Instagram, X, TikTok, Reddit) in
Settings; a download of an auth-walled link on that host is retried with
the acting user's cookie.
Two load-bearing security controls, both in downloads.credentials as the
single source of truth so enqueue and the worker can't drift:
- cache privacy: a cookie-authed download salts its MediaAsset format_sig
with the user id, so the authed asset is private and never cache-served
to another user. The worker attaches a cookie ONLY when the asset it is
downloading into IS that private, salted identity.
- host scope (SEC-12): a cookie is resolved only for a URL whose netloc
the credential owns (netloc-scoped _host_is); a look-alike/unrelated
host gets none.
cookies_enc is Fernet-encrypted at rest, only decrypted in-worker into a
0600 file under staging/.auth/ (survives the retry wipe + _produced_file
scan) and unlinked in finally. The blob is never returned by a route,
logged, audited, or written to a job's error/snapshot. cookiefile is set
after the adapter's extra merge so an adapter can't clobber it. Settings
UI is write-only (host + set/updated only, never echoed). New
download_credentials table + migration 0074; per-user routes; HU+EN i18n.
Tests: host-scope, cache-privacy (no cross-user serve), encrypt-at-rest,
write-only API, worker attach-gate; FE dom test for the write-only field.
The download/edit child (yt-dlp + ffmpeg) shares one LAN box with the API,
Postgres, and Plex playback. A CPU re-encode pegged every core for ~30s and
starved web requests + playback until it finished. Full-scope S4 mitigations:
- Child self-de-prioritizes at startup (inherited by every ffmpeg/yt-dlp
descendant): CPU nice increment + best-effort idle ionice class + an
OMP_NUM_THREADS budget, all config-gated and independently disable-able.
- Cap the browser-safe re-encode's encoder threads (`-threads`) so a transcode
reserves a core for foreground work instead of claiming all of them.
- Throttle the retention GC's whole-tree orphan sweep (yield IO after each
batch of files) so its rglob can't monopolize disk IO in the API process.
- Idle-back-off the Download Center (2s→15s when nothing queued/running) and
UsageBar (2s→15s when running_jobs is 0) polls; both snap back to 2s on the
next fetch after an enqueue (invalidateDownloads refetches immediately).
Tests: encode_thread_cap (pin/reserve-core/edge), -threads flag presence,
_deprioritize (nice/ionice/OMP branches), GC sweep throttle batching.
Round-2 review of the R1 fixes found that the two-independent-limiter progress
writer issued a redundant heartbeat every interval during an ACTIVE download: a
percent change writes progress (which already refreshes updated_at, the stall
watchdog's liveness source), then the sub-integer calls that follow fire the
separate beat budget — ~2 DB writes/interval instead of the throttle's intended 1.
Replace the second limiter with a shared `last_touch` clock (last DB write OR
beat). Progress writes keep their own `allow_write` budget (so a beat still can't
delay a real write — the R1 concern), but the heartbeat now fires only when
nothing touched the row for a full interval. Active download → ~1 write/interval;
a genuinely stuck re-encode → 1 heartbeat/interval. New regression test:
an unchanged-percent call in the same interval as a write does NOT heartbeat.
ruff clean, full DB-lane 365 pass.
Round-1 review fixes on the E3 branch:
- prune partial-fetch (correctness): `_paginate` now reports whether it reached
the server's reported `total`; `_prune_stale` prunes only on a COMPLETE
non-empty listing. A truncated fetch (a mid-sequence empty/short page before
total — a transient Plex hiccup) no longer deletes the rows it merely failed
to list. New test: a listing that claims 5 but returns 1 prunes nothing.
- `_prune_stale` (robustness): `stats.get("pruned", 0) + 1` — no KeyError when a
caller's stats dict omits the key.
- service transitions (conventions): pause/resume/redownload use the JobStatus
StrEnum instead of loose "queued"/"paused"/"error" literals, matching cancel.
- ffmpeg progress writer (correctness): a stuck-percent heartbeat now runs on its
OWN 1/interval budget, so it can't consume the progress-write slot and delay a
real percent change by up to an interval.
- backend/app/README.md (conventions): reword the transaction-ownership rule to
quote and extend CLAUDE.md's canonical ("handler/job-wrapper commits; helpers
flush; only quota owns its session") instead of stating a contradicting
"helpers MAY commit" carve-out.
ruff clean, full DB-lane 364 pass.
worker.py (~1090 L) mixed the PARENT supervisor (claim + supervise + the process
loop) with the CHILD pipeline (the yt-dlp/ffmpeg download+edit that runs in the
spawned subprocess). Extract them:
worker.py 1090 -> 475 parent: claim/supervise/loop
downloads/pipeline.py (new 592) the child pipeline (_download/_process_edit +
hooks/progress/ffmpeg helpers + _Aborted)
downloads/jobdb.py (new 73) the short-lived-session helpers both sides
share (_set_job/_job_status/_touch_job/
_requeue_if_running/_reset_asset_pending +
the staging-dir paths)
The graph is one-directional (runner -> pipeline -> jobdb; worker -> jobdb; no
worker<->pipeline edge) — the parent never calls the child directly, it spawns
runner.py as a subprocess. Consumers rewired: runner import, and two tests follow
the moved symbols (_media_info/_produced_file -> pipeline; the ffmpeg-callback
patches target pipeline where the names are looked up).
Pure code motion, verbatim bodies. ruff clean, full lane 363 pass (unchanged).
peter