yt-dlp asks for a bug report on an extraction failure it did not expect: the site answered in a way its extractor does not read (measured on a TikTok link: fails on every attempt under yt-dlp 2026.07.04, downloads under 2026.08.19). That failure read as "Something went wrong" with no retry; it is now its own category, site_changed, with a Retry in the failure window. A contract test holds every failure category to its copy in both languages.
yt-dlp refuses a file announced over `max_filesize` by logging one line and returning with no file
(measured, and pinned in test_formats against a local server): the worker found no output and
failed the job as `postprocess`. `formats.watching_size_cap` now sees that line through our yt-dlp
logger, and the download (and, should it ever happen, a YouTube clip's range fetch) fails as the
owner's refusal with DOWNLOAD_TOO_BIG_MESSAGE, classified `quota_exceeded`. The download's quota
copy says it does not fit in what is left — true at a full cap too — instead of "Storage full".
A pre-existing type error the change touched is fixed (baseline down by one).
- A converted upload over the space left (not zero) fails with CONVERTED_TOO_BIG_MESSAGE, not
"storage is full"; it classifies as quota_exceeded by the phrase it shares with the clip's.
- The public watch page says "preparing" only for a job a worker will take: a suspended owner's
queued job waits for as long as the suspension lasts, so the link says the video is gone, as
for a paused rebuild.
- The worker's claim docstring names the restart case: a running job put back in the queue waits.
A suspension now blocks the user's downloads as quota.resolve already advised: the worker leaves
their queued jobs unclaimed until it is lifted (the confirm says so). A cut clip larger than the
remaining storage is refused with its own message (not "storage is full"), and a clip's quota
refusal reads as a clip's in both languages. The cancel poll compares to JobStatus.RUNNING.
The code cited its plan instead of saying why: `E28 review J4`, `(R6 S3b)`, `S1 justified…`,
`deep-review round 5`. That plan lives in the private ops repo, so for every reader of this one the
citation was noise, and a sentence whose subject was the plan step ("round 2 showed…") hid its
lesson behind an id. Each one is gone and the "why" kept: a bare citation is deleted, a sentence
carrying a lesson is reworded around it ("a later attempt showed…").
Measured (`scripts/audit_metrics.py`, product code): 652 lines in 270 files carried an epic code by
the definition the next commit holds (507 / 241 by the old, two-digit-only one) -> 0. Across all
code, tests, scripts and configs: ~1100 epic-code lines plus ~320 sprint/round ids.
Behaviour unchanged, checked mechanically: every changed .py file has the same AST once docstrings
are blanked; css/sh/toml/yml/json files are identical once comments are stripped; the only
non-comment changes in TS/JS are test titles, one description string in a contract table, and five
ESLint messages, whose text is all the gates snapshot diff holds. Not touched: Markdown (docs and
the changelog keep their history), Alembic migrations, TV episode notation (`S2 · E5`), DESIGN.md
ids (`C1`, `B4`, `S3h`), tracker ids (`SEC-16`, `DX-17`).
One stale claim fixed on the way: test_hls_reaper's docstring said the player checkpoints "every
10 s whether or not it is playing", which plex/stream.py calls wrong; it now says what the code does.
Measured on the isolated test stack the same night, on the very same video ~90 minutes after it was
answering with the bot message: yt-dlp got the player response and picked its formats, and the CDN
then refused the stream itself — "unable to download video data: HTTP Error 403: Forbidden".
Signed out it did that on every attempt; with the owner's cookie attached the video downloaded in
full. So the wall moves between the player response and the stream, and the text it wears there was
costing the download twice.
retriable: the format URLs belong to the client that minted them, so the next client set is a real
second chance — but this text classifies as `network` and nothing in the YouTube marker list matched
it, so the ladder ended the job on attempt 1 and never tried the fallback.
catalog: `youtube.media_forbidden` for the cookie-less case, so the UI stops calling it a network
error when nothing was wrong with the network, and names what actually gets past it. Scoped exactly
as narrowly as the measurement: another host keeps the generic `network` category, and so does a
download that already carried a cookie — telling a signed-in user to sign in is the failure the
catalog exists to avoid. EN + HU copy.
A YouTube link can fail with "Sign in to confirm you're not a bot" while every
other YouTube download works — YouTube gates INDIVIDUAL videos behind a
signed-in session. Measured on prod job 273: all twelve of yt-dlp's player
clients fail on that one video while its neighbours download fine, the bgutil
PO-token sidecar is live and answering, and the newest yt-dlp fails identically.
So the download is not broken by our stack or a stale yt-dlp, and no retry,
client switch or upgrade gets past it. A cookie is the only way in.
The failure copy already sends the user to Settings -> Account -> Download
authentication, but YouTube was not in the cookie host registry, so that
dropdown could not help them: the app asked for something it did not offer.
credentials: YouTube joins SUPPORTED_HOSTS (youtube.com, youtu.be,
youtube-nocookie.com). The Settings dropdown, the SEC-12 host scope and the
private-asset cache salt all derive from the row, so that is the whole wiring.
adapters: a signed-in YouTube download runs only cookie-capable player clients.
Measured against yt-dlp's own INNERTUBE_CLIENTS table: with account cookies
attached yt-dlp REMOVES every client whose SUPPORTS_COOKIES is False, and a set
it empties that way raises "No player clients have been requested" — which is
not a per-client flake it retries past. Our primary set is android_vr, one of
those, so without this the cookie would have failed the download on attempt 1
and never reached the POT-backed web fallback it exists for. A test pins the
client list against the installed yt-dlp so a version bump turns the gate red
instead of silently un-authing every YouTube download, and the empty-set message
joins the retriable markers as the second line of defence.
errors: catalog entries for the wall with and without a cookie, so a user who
has already added cookies is told they were rejected (YouTube rotates exported
ones) rather than to add some. The marker is quote-agnostic: YouTube writes the
apostrophe as U+2019, so "you're not a bot" with a straight quote never matches
the real text. EN + HU copy.
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 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.
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.
5a — the public watch page rendered every asset in a <video>, but WebKit (all
iOS browsers) refuses to play an audio-only mp3 in <video>, so shared audio
links were broken on iPad while working on desktop Chromium. public_meta now
flags audio-only assets (codec present, no video stream) and WatchPage renders
a native <audio> in a padded surface card (no artwork exists for these).
5c — a login-walled Facebook post failed as a bare 'unknown', which offered a
misleading Retry. Classify closed-group / cookie-suggestion auth walls as
login_required and genuinely unsupported sites as a new unsupported_host class
(both non-retryable), and drop unknown from the modal's retryable set — an
unclassified failure is most often an unparsed auth wall. The queue row keeps
its own force-retry for real transients.
Three prod-observed download bugs, root-caused from a frozen prod queue with the
worker pinned at 100% CPU:
- ReDoS in normalize_title: _TRAILING_HASHTAGS had two overlapping unbounded
quantifiers around the required letter, so a Facebook title with a long hashtag
cluster that is NOT at the end ("... #a #b #c | Page") backtracked catastrophically.
It runs inside the yt-dlp progress hook holding the GIL, so one such download froze
the whole worker (and could stall feed enrichment too - sync/videos.py calls it).
Rewrite with a zero-width "has a letter" lookahead + a possessive tag body: one
unambiguous split, nothing to backtrack. 40-tag stress case: ~0.1ms.
- Reddit download killed by a cosmetic thumbnail convert: yt-dlp's
FFmpegThumbnailsConvertor runs "-bsf:v mjpeg2jpeg" on a JPEG-served-under-a-.png-name
(Reddit) and fails; being a postprocessor, it failed the WHOLE download even though
the media was fine. Drop it and produce the <base>.jpg poster/og:image ourselves via
editmod.thumbnail_to_jpg (plain ffmpeg decode->mjpeg, best-effort -> ensure_poster
cuts a frame on failure). The FB Messenger / MS Teams share preview keeps working
(posters verified as real JPEGs).
- classify: map "authentication is required" (Reddit's auth wall) to login_required
instead of unknown, so the UI explains it rather than offering a pointless retry.
Tests: ReDoS-timing + numeric-marker regression, the Reddit classify case, and
thumbnail_to_jpg (real JPEG out / None on bad input, never raises).
Failed downloads stored the raw yt-dlp/ffmpeg text and dumped it as a truncated red
one-liner in the queue. Classify the failure into a category (unavailable / login_required /
geo_blocked / format_unavailable / drm / postprocess / source_missing / network / unknown)
and surface it properly:
- backend: app/downloads/errors.classify maps the raw message to a category code; the worker
stores it in the new download_jobs.error_code (migration 0061) at both failure sites; resume
clears it; the serializer exposes it. Raw text kept as technical detail.
- frontend: the queue row shows a clean localized reason + a "details" link opening a modal
with the explanation, the raw text behind a disclosure, and a Retry only for transient
categories (network/postprocess/unknown).
- tests: classify() unit coverage over verbatim yt-dlp/ffmpeg message shapes.
peter