peter/siftlode
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

release: 0.118.0 (improvement/e59-small-residuals)

Browse Source
This commit is contained in:
peter
2026-11-08 01:23:33 +01:00
parent 5a93a9a8e7 e08278fae2
commit 45b5e19fb8
77 changed files with 3742 additions and 574 deletions
Download Patch File Download Diff File Expand all files Collapse all files
CHANGELOG.md
+12
View File
@@ -7,6 +7,18 @@ edit the release notes and re-run the generator.
Upgrading, pinning a version and rolling back: [docs/self-hosting.md](docs/self-hosting.md#upgrading).
## 0.118.0 — 2026-11-08
### New
- Surprise me now leans on the channels and genres you watch, and one pick in three stays a blind draw, so it remains a surprise.
- On a connection too slow for a Plex video, playback waits and goes on where it was, and says how fast the video needs and how fast it arrives.
### Fixed
- A 24/7 stream's archive no longer adds years to a list's total: a total names the videos it could not count as "+N without length". The Feed sorted by Surprise me opens faster.
- Plex films the server's GPU cannot decode (VC-1, older DivX or XviD) start in a few seconds instead of about 20. Text over the Plex picture is readable on any frame.
## 0.117.0 — 2026-11-06
### New
CLAUDE.md
+1 -1
View File
@@ -65,7 +65,7 @@ how it gets built. Concretely:
`routes/*` (HTTP only: validation, auth-dependency, serialization) → domain packages `sync/`, `plex/`,
`downloads/`, `reader/`, `discover/`, `catalog/`, `youtube/`, `issues/` (shared by routes + scheduler + worker) → flat core (`auth`, `apitokens`, `quota`, `jobruns`,
`sysconfig`, `state`, `audit`, `notifications`, `userscope`, `ratelimit`, `bodylimit`, `progress`, `collation`,
`security`, `logs`, `logsink`, `diagnostics`, `debugsession`) → `models/db/config`. **Dependency direction is one-way: core never imports routes.**
`security`, `logs`, `logsink`, `diagnostics`, `debugsession`, `draw`) → `models/db/config`. **Dependency direction is one-way: core never imports routes.**
The domain is four **units**, each owning its tables: Capture (`downloads/`), Reader (`reader/`, `discover/`, the playlist
sync), Catalog (`catalog/`, the `sync/` ingests) and Plex (`plex/`). A unit's route does no database work: it reads through
the unit's `queries`, writes through its `service` (which flushes) and commits. A unit's code reads another unit's tables
VERSION
+1 -1
View File
@@ -1 +1 @@
0.117.0
0.118.0
backend/.basedpyright/baseline.json
-138
View File
@@ -185,118 +185,6 @@
"lineCount": 1
}
},
{
"code": "reportArgumentType",
"range": {
"startColumn": 50,
"endColumn": 56,
"lineCount": 1
}
},
{
"code": "reportArgumentType",
"range": {
"startColumn": 50,
"endColumn": 56,
"lineCount": 1
}
},
{
"code": "reportArgumentType",
"range": {
"startColumn": 50,
"endColumn": 56,
"lineCount": 1
}
},
{
"code": "reportArgumentType",
"range": {
"startColumn": 50,
"endColumn": 56,
"lineCount": 1
}
},
{
"code": "reportArgumentType",
"range": {
"startColumn": 50,
"endColumn": 56,
"lineCount": 1
}
},
{
"code": "reportArgumentType",
"range": {
"startColumn": 50,
"endColumn": 56,
"lineCount": 1
}
},
{
"code": "reportArgumentType",
"range": {
"startColumn": 50,
"endColumn": 56,
"lineCount": 1
}
},
{
"code": "reportArgumentType",
"range": {
"startColumn": 50,
"endColumn": 56,
"lineCount": 1
}
},
{
"code": "reportArgumentType",
"range": {
"startColumn": 50,
"endColumn": 56,
"lineCount": 1
}
},
{
"code": "reportArgumentType",
"range": {
"startColumn": 50,
"endColumn": 56,
"lineCount": 1
}
},
{
"code": "reportArgumentType",
"range": {
"startColumn": 50,
"endColumn": 56,
"lineCount": 1
}
},
{
"code": "reportArgumentType",
"range": {
"startColumn": 50,
"endColumn": 56,
"lineCount": 1
}
},
{
"code": "reportArgumentType",
"range": {
"startColumn": 50,
"endColumn": 56,
"lineCount": 1
}
},
{
"code": "reportArgumentType",
"range": {
"startColumn": 50,
"endColumn": 56,
"lineCount": 1
}
},
{
"code": "reportArgumentType",
"range": {
@@ -364,25 +252,7 @@
}
}
],
"./app/downloads/adapters.py": [
{
"code": "reportArgumentType",
"range": {
"startColumn": 30,
"endColumn": 38,
"lineCount": 1
}
}
],
"./app/downloads/backfill.py": [
{
"code": "reportArgumentType",
"range": {
"startColumn": 34,
"endColumn": 38,
"lineCount": 1
}
},
{
"code": "reportReturnType",
"range": {
@@ -413,14 +283,6 @@
}
],
"./app/downloads/pipeline.py": [
{
"code": "reportArgumentType",
"range": {
"startColumn": 34,
"endColumn": 38,
"lineCount": 1
}
},
{
"code": "reportAttributeAccessIssue",
"range": {
backend/alembic/versions/0117_videos_length_sort_indexes.py
+76
View File
@@ -0,0 +1,76 @@
"""videos: the duration sorts ride an index on the video's length, not on the raw column
Revision ID: 0117_videos_length_sort_indexes
Revises: 0116_followed_explored_channels
Create Date: 2026-11-07
A YouTube length past a day is not a length (`catalog.models.LENGTH_MAX_S`): a 24/7 stream's archive
reports the years it has run. The feed's duration sorts now order by `Video.length_seconds`, which
is NULL for such a video, so "longest first" no longer opens on a stream archive. Migration 0096's
two duration indexes are on the raw column and no longer match that ORDER BY; they are replaced by
the same indexes over the length expression. The expression is spelled out here as SQLAlchemy emits
it, with the limit as a literal (an index matches a query's constant, never a bound parameter);
tests/test_feed_sort_indexes.py fails if the two drift apart.
CONCURRENTLY (autocommit_block), re-runnable, as in 0096.
"""
from alembic import op
from sqlalchemy import text
revision: str = "0117_videos_length_sort_indexes"
down_revision: str | None = "0116_followed_explored_channels"
branch_labels: str | None = None
depends_on: str | None = None
# Frozen with this migration: `catalog.models.LENGTH_MAX_S` is 24 h.
LENGTH = "(CASE WHEN (duration_seconds > 86400) THEN NULL ELSE duration_seconds END)"
NEW = {
"ix_videos_length_desc": f"{LENGTH} DESC NULLS LAST, id",
"ix_videos_length_asc": f"{LENGTH} ASC NULLS LAST, id",
}
OLD = {
"ix_videos_duration_desc": "duration_seconds DESC NULLS LAST, id",
"ix_videos_duration_asc": "duration_seconds ASC NULLS LAST, id",
}
def _valid(name: str) -> bool | None:
"""The index's `indisvalid`, or None when there is no index of that name."""
row = (
op.get_bind()
.execute(
text(
"SELECT i.indisvalid FROM pg_index i JOIN pg_class c ON c.oid = i.indexrelid "
"WHERE c.relname = :name"
),
{"name": name},
)
.first()
)
return None if row is None else bool(row[0])
def _build(name: str, columns: str) -> None:
"""CREATE INDEX CONCURRENTLY, re-runnable: an INVALID leftover of a failed build is dropped first
(see 0096 `_build`)."""
if _valid(name) is False:
op.execute(f"DROP INDEX CONCURRENTLY IF EXISTS {name}")
op.execute(f"CREATE INDEX CONCURRENTLY IF NOT EXISTS {name} ON videos ({columns})")
def upgrade() -> None:
with op.get_context().autocommit_block():
for name, columns in NEW.items():
_build(name, columns)
for name in OLD:
op.execute(f"DROP INDEX CONCURRENTLY IF EXISTS {name}")
def downgrade() -> None:
with op.get_context().autocommit_block():
for name, columns in OLD.items():
_build(name, columns)
for name in NEW:
op.execute(f"DROP INDEX CONCURRENTLY IF EXISTS {name}")
backend/alembic/versions/0118_videos_shuffle_key.py
+40
View File
@@ -0,0 +1,40 @@
"""videos: a random key per video, so a seeded shuffle walks an index instead of sorting the scope
Revision ID: 0118_videos_shuffle_key
Revises: 0117_videos_length_sort_indexes
Create Date: 2026-11-07
The feed's Shuffle sort and the Home surprise ordered the whole feed scope by md5(id || seed) on every
call: ~350 ms on prod for a scope of ~213k videos, and a Shuffle page deeper in the list cost more
(767 ms on page 7 on localdev, against 4 ms for Newest). Each video now carries a random number drawn
once (`Video.shuffle_key`); a seed picks a point on that ring and the draw reads the index from there
(`feedquery.shuffle_walk`).
One ADD COLUMN with a volatile default rewrites the table and fills every row with its own value. On
localdev (278k videos) that took 28 s, against 84 s for an UPDATE of a nullable column, so the rewrite
it is: the table is locked for that time, and a deploy has the app down while it migrates anyway.
The index is (shuffle_key, id): the walk's keyset cursor compares the pair.
"""
from alembic import op
revision: str = "0118_videos_shuffle_key"
down_revision: str | None = "0117_videos_length_sort_indexes"
branch_labels: str | None = None
depends_on: str | None = None
def upgrade() -> None:
op.execute(
"ALTER TABLE videos ADD COLUMN IF NOT EXISTS shuffle_key double precision "
"NOT NULL DEFAULT random()"
)
with op.get_context().autocommit_block():
op.execute("DROP INDEX CONCURRENTLY IF EXISTS ix_videos_shuffle_key")
op.execute("CREATE INDEX CONCURRENTLY ix_videos_shuffle_key ON videos (shuffle_key, id)")
def downgrade() -> None:
with op.get_context().autocommit_block():
op.execute("DROP INDEX CONCURRENTLY IF EXISTS ix_videos_shuffle_key")
op.execute("ALTER TABLE videos DROP COLUMN IF EXISTS shuffle_key")
backend/alembic/versions/0119_plex_transcode_gpu.py
+65
View File
@@ -0,0 +1,65 @@
"""plex_items.transcode_hw: the facet says whether the GPU ENCODES, under the measured codec sets
Measured on prod (2026-11-07): the iGPU has no VC-1 Advanced decode, so a VC-1 film never ran on the
GPU although the facet said it did; VC-1, XviD and DivX now take the CPU-decode → GPU-encode path; and an
HDR film runs on the GPU through the OpenCL tonemap although the facet said software. The rule is
`plex.sync.transcode_hw_capable`; the two codec sets below are pinned to `plex.sync` by
tests/test_plex_playback_class.py.
Rewritten from the stored codec and bit depth. HDR is not stored, so a 10-bit HW-decodable item is set to
software here and its probe watermark cleared: the bit-depth backfill (`plex.sync._probe_media_batch`,
a bounded batch per Plex sync) reads its HDR flag again and moves an HDR one back to the GPU.
"""
from typing import Sequence, Union
from alembic import op
revision: str = "0119_plex_transcode_gpu"
down_revision: Union[str, None] = "0118_videos_shuffle_key"
branch_labels: Union[str, Sequence[str], None] = None
depends_on: Union[str, Sequence[str], None] = None
HW_DECODABLE_CODECS = ("h264", "hevc", "h265", "vp9", "mpeg2video")
CPU_DECODE_CODECS = ("vc1", "mpeg4", "msmpeg4v3")
def _in(codecs: tuple[str, ...]) -> str:
return ", ".join(f"'{c}'" for c in codecs)
def upgrade() -> None:
op.execute(
f"""
UPDATE plex_items
SET transcode_hw = CASE
WHEN lower(codec_video) IN ({_in(HW_DECODABLE_CODECS)})
THEN coalesce(video_bit_depth, 8) <= 8
WHEN lower(codec_video) IN ({_in(CPU_DECODE_CODECS)})
THEN coalesce(video_bit_depth, 8) <= 8
ELSE false
END
WHERE playable = 'transcode'
"""
)
op.execute(
f"""
UPDATE plex_items
SET media_probed_at = NULL
WHERE playable = 'transcode'
AND lower(codec_video) IN ({_in(HW_DECODABLE_CODECS)})
AND video_bit_depth > 8
"""
)
def downgrade() -> None:
# The old rule: a decodable codec (VC-1 counted) at 8 bits; HDR and every other codec software.
op.execute(
f"""
UPDATE plex_items
SET transcode_hw = lower(codec_video) IN ({_in(HW_DECODABLE_CODECS + ("vc1",))})
AND coalesce(video_bit_depth, 8) <= 8
WHERE playable = 'transcode'
"""
)
backend/app/README.md
+2 -2
View File
@@ -25,8 +25,8 @@ Three layers, by *who* runs the code — not by feature:
background run's durable row, its words in `runstatus`),
`notifications`, `userscope`, `ratelimit`, `progress`, `logs` (the one logging
configuration), `logsink` (its `log_events` table), `diagnostics` (a user's consented log
level), `debugsession` (a recording that tags its user's lines, E16) and `collation` (A→Z in the
reader's alphabet, E30), … are single-file primitives many layers import. They depend on nothing above them (a core
level), `debugsession` (a recording that tags its user's lines, E16), `collation` (A→Z in the
reader's alphabet, E30) and `draw` (the weighted, seeded draw "Surprise me" deals from), … are single-file primitives many layers import. They depend on nothing above them (a core
module must never import a `routes/…` module).
**Extraction trigger: the SECOND consumer.** Keep new logic inline in its first caller. The moment a
backend/app/catalog/models.py
+53 -2
View File
@@ -10,16 +10,22 @@ from sqlalchemy import (
Boolean,
Computed,
DateTime,
Float,
ForeignKey,
Integer,
JSON,
String,
Text,
UniqueConstraint,
case,
func,
literal_column,
null,
)
from sqlalchemy.dialects.postgresql import TSVECTOR
from sqlalchemy.ext.hybrid import hybrid_property
from sqlalchemy.orm import Mapped, mapped_column, relationship
from sqlalchemy.sql.elements import ColumnElement
from app.db import Base
from app.models import TimestampMixin
@@ -114,6 +120,20 @@ class Subscription(Base, TimestampMixin):
# hidden from feeds/search and revisited by the refresh job until they settle into a VOD.
LIVE_OR_UPCOMING = ("live", "upcoming")
# A YouTube length past a day is not a length. A 24/7 stream's archive (`was_live`) reports the years
# the stream has run (measured on prod, 2026-11-06: 122,962,914 s), and one such video made a Watch
# later total read 44,073 hours. YouTube caps an upload at 12 hours, so nothing real is lost. Such a
# length is unknown wherever a length is used: the API sends none (`Video.length_seconds`), sums leave
# it out (a playlist's counts it as `without_length`), and the duration filter, sort and Home budget never
# match it. The raw `duration_seconds` stays for what it does prove — that the video is not a Short —
# and for the jobs that ask whether YouTube has answered yet (test_video_length holds who reads it).
LENGTH_MAX_S = 24 * 3600
def known_length(seconds: int | None) -> int | None:
"""`seconds` as a video's length, or None when it is not one (see LENGTH_MAX_S)."""
return seconds if seconds is not None and seconds <= LENGTH_MAX_S else None
class Video(Base, TimestampMixin):
"""A YouTube video. Shared across users; per-user state lives elsewhere."""
@@ -141,8 +161,8 @@ class Video(Base, TimestampMixin):
)
thumbnail_url: Mapped[str | None] = mapped_column(String(1024))
# The feed's views / duration / title sorts each ride a migration-defined index in their exact
# ORDER BY, one per direction (and per ICU collation for title): migration 0096, held to
# `feedquery.order_by` by test_feed_sort_indexes.
# ORDER BY, one per direction (and per ICU collation for title): migration 0096 (the duration
# pair since 0117, on `length_seconds`), held to `feedquery.order_by` by test_feed_sort_indexes.
duration_seconds: Mapped[int | None] = mapped_column(Integer)
view_count: Mapped[int | None] = mapped_column(BigInteger)
like_count: Mapped[int | None] = mapped_column(BigInteger)
@@ -197,6 +217,37 @@ class Video(Base, TimestampMixin):
last_checked_at: Mapped[datetime | None] = mapped_column(DateTime(timezone=True))
unavailable_since: Mapped[datetime | None] = mapped_column(DateTime(timezone=True))
unavailable_reason: Mapped[str | None] = mapped_column(String(24))
# A random number in [0, 1), drawn once when the row is written: the ring a seeded shuffle walks
# (`feedquery.shuffle_walk`, index ix_videos_shuffle_key on (shuffle_key, id), migration 0118).
shuffle_key: Mapped[float] = mapped_column(Float, nullable=False, server_default=func.random())
@hybrid_property
def length_seconds(self) -> int | None:
"""The video's length, or None when it is unknown — not fetched yet, or not a length at all
(`LENGTH_MAX_S`). What every reader of a length reads."""
return known_length(self.duration_seconds)
@length_seconds.inplace.expression
@classmethod
def _length_seconds_expression(cls) -> ColumnElement[int | None]:
# The limit is a literal, not a bound parameter: the duration sort indexes (migration 0117)
# are built on this very expression, and Postgres matches an index to a query's constant only.
return case(
(cls.duration_seconds > literal_column(str(LENGTH_MAX_S)), null()),
else_=cls.duration_seconds,
)
@hybrid_property
def without_length(self) -> bool:
"""A video whose length is unknown — live, an archive past a day, or not fetched yet: what a
sum of lengths leaves out and names as "+N without length". Every such video, so the count
and the sum never disagree about what was left out."""
return self.length_seconds is None
@without_length.inplace.expression
@classmethod
def _without_length_expression(cls) -> ColumnElement[bool]:
return cls.length_seconds.is_(None)
# Weighted full-text search document (DB-generated, never written by the app): title (A) >
# creator keywords + search queries (B) > truncated description (C). ts_rank honours the
backend/app/catalog/queries.py
+3 -2
View File
@@ -44,7 +44,8 @@ def channel_titles(db: Session, ids: set[str]) -> dict[str, str]:
def channel_aggregates(channel_ids: list[str]) -> Select:
"""Per-channel stored count, last upload and total duration, in one grouped pass. Every column it
"""Per-channel stored count, last upload and total duration (the lengths it knows: a video with
none is left out, `Video.length_seconds`), in one grouped pass. Every column it
reads is in `ix_videos_channel_pub_cover` (migration 0095), so Postgres answers it from the index
alone; a test holds that (test_channel_list_index_only)."""
return (
@@ -52,7 +53,7 @@ def channel_aggregates(channel_ids: list[str]) -> Select:
Video.channel_id,
func.count(Video.id),
func.max(Video.published_at),
func.coalesce(func.sum(Video.duration_seconds), 0),
func.coalesce(func.sum(Video.length_seconds), 0),
)
.where(Video.channel_id.in_(channel_ids))
.group_by(Video.channel_id)
backend/app/discover/queries.py
+101 -29
View File
@@ -10,16 +10,25 @@ explore-visible and in-progress for free, and cannot drift from them. The Plex h
narrows EVERY rail by the Plex sidebar's `LibraryFilters`, which is right for a surface that HAS a
sidebar and wrong for a landing page that does not.
**One seed drives both halves.** The feed orders a shuffle by `md5(id || seed)` so a keyset cursor
stays valid across pages; the Plex surprise rail uses `ORDER BY random()`, which is fine for a single
un-paged draw but cannot be reproduced. A cross-source reroll needs both halves to move together —
otherwise "Mást kérek" re-draws the Plex picks while the YouTube ones stand still — so the Plex side
here takes the feed's seeded-md5 idiom instead of `random()`.
**One seed drives both halves.** The YouTube half starts on the ring of stored random keys at the
seed's point and shuffles that window by `md5(id || seed)`; the Plex surprise rail uses
`ORDER BY random()`, which is fine for a single un-paged draw but cannot be reproduced. A
cross-source reroll needs both halves to move together — otherwise "Mást kérek" re-draws the Plex
picks while the YouTube ones stand still — so the Plex side here orders by the same seeded md5
instead of `random()`.
"""
from sqlalchemy import and_, func, or_, select
from datetime import datetime, timezone
from sqlalchemy import (
and_,
func,
or_,
select,
)
from sqlalchemy.orm import Session, aliased
from app import draw
from app.discover.budget import band
from app.feedquery import (
feed_columns,
@@ -27,13 +36,14 @@ from app.feedquery import (
not_blocked_cond,
order_by,
serialize,
shuffle_walk,
sort_keys,
)
from app.models import User
from app.plex.models import PlexItem, PlexShow, PlexState
from app.catalog.models import Channel, Video
from app.reader.models import PlaylistItem, VideoState
from app.plex import rollup
from app.plex import affinity, rollup
from app.plex.cards import display_title, leaf_card, movie_card, show_card
from app.plex.query import apply_show_state_filter, content_libs
from app.userscope import owned_by
@@ -67,7 +77,8 @@ _FEED_BASE: dict = {
def _seeded(col, seed: int):
"""The feed's deterministic shuffle expression, reused for Plex so one seed moves both halves."""
"""The Home surprise's seeded order (md5 of the key and the seed), the same for Plex so one
seed moves both halves."""
return func.md5(func.concat(col, str(seed)))
@@ -112,17 +123,75 @@ def yt_surprise(db: Session, user: User, budget: str, seed: int, limit: int) ->
Shorts are opened up only for the shortest band: "I have ten minutes" is the one answer a Short is
actually a good response to, and the feed's own default keeps them out otherwise."""
lo, hi = band(budget)
params = dict(_FEED_BASE, min_duration=lo, max_duration=hi, include_shorts=budget == "short")
query, rank_expr = filtered_query(db, user, **params)
keys = sort_keys("shuffle", seed, rank_expr)
# Over-fetch, then drop anything already started. The feed's `unwatched` scope keeps watched and
# hidden out but deliberately keeps in-progress rows in (they belong in the feed). Here they must
# go: the resume tier has already had its chance at them, and it may have passed precisely because
# the REMAINING time did not fit the budget — offering the same item back as a surprise, whose
# runtime does fit, would label a half-finished thing "Meglepetés" and contradict the question.
rows = db.execute(query.order_by(*order_by(keys)).limit(limit * 4)).all()
out = [serialize(r) for r in rows]
return [r for r in out if not r["position_seconds"]][:limit]
params: dict = {
**_FEED_BASE,
"min_duration": lo,
"max_duration": hi,
"include_shorts": budget == "short",
}
query, _rank = filtered_query(db, user, **params)
# A window of the ring, not the whole scope: the seed's SURPRISE_WINDOW videos read along the
# stored random key (`shuffle_walk`, a millisecond), then shuffled by the seed. Dealing straight
# from the ring would favour for good every video that sits after a wide gap of it (the first
# video past a random point is the one whose gap the point fell into); shuffling the window makes
# each video's chance the window's share, even to within a few percent. A scope no larger than
# the window is the whole scope, in exactly the md5 order the draw always had.
#
# Anything already started is dropped. The feed's `unwatched` scope keeps watched and hidden out
# but deliberately keeps in-progress rows in (they belong in the feed). Here they must go: the
# resume tier has already had its chance at them, and it may have passed precisely because the
# REMAINING time did not fit the budget — offering the same item back as a surprise, whose runtime
# does fit, would label a half-finished thing "Meglepetés" and contradict the question.
window = [
r
for r in map(serialize, shuffle_walk(db, query, seed, SURPRISE_WINDOW))
if not r["position_seconds"]
]
blind = draw.shuffled(window, seed, key=_video_id)
channels = _channel_affinity(db, user)
if not channels:
return blind[:limit]
def weight(row: dict) -> float:
return 1.0 + CHANNEL_WEIGHT * channels.get(row["channel_id"], 0.0)
leaning = draw.leaning(window, seed, key=_video_id, weight=weight)
return draw.blend(leaning, blind, seed, limit, key=_video_id)
# How many videos of the ring a surprise draws from: wide enough that a channel the user watches has
# a few in it, small enough to read in a millisecond or two.
SURPRISE_WINDOW = 500
def _video_id(row: dict) -> str:
return row["id"]
# How much a channel the user watches most outweighs one they never watch (1 + this), and how far back
# the watched history counts: the latest WATCHED_RECENT watches, the recent ones weighing more.
CHANNEL_WEIGHT = 3.0
WATCHED_RECENT = 300
WATCHED_HALF_LIFE_DAYS = 90
def _channel_affinity(db: Session, user: User) -> dict[str, float]:
"""Each channel the user watched lately, scored 0..1 against the most watched: the sum of its
watches, each weighing half as much every WATCHED_HALF_LIFE_DAYS. Empty with no history."""
rows = db.execute(
select(Video.channel_id, VideoState.watched_at)
.join(Video, Video.id == VideoState.video_id)
.where(owned_by(VideoState.user_id, user), VideoState.status == "watched")
.order_by(VideoState.watched_at.desc().nullslast())
.limit(WATCHED_RECENT)
).all()
now = datetime.now(timezone.utc)
scores: dict[str, float] = {}
for channel_id, watched_at in rows:
r = draw.recency(watched_at, now, WATCHED_HALF_LIFE_DAYS)
scores[channel_id] = scores.get(channel_id, 0.0) + r
return draw.scaled(scores)
def yt_video_row(db: Session, user: User, video_id: str) -> dict | None:
@@ -289,21 +358,24 @@ def plex_surprise(
if not movie_libs:
return []
lo, hi = band(budget)
q = (
select(PlexItem)
.where(
PlexItem.kind == "movie",
PlexItem.library_id.in_(movie_libs),
_unseen_movie_conds(uid),
)
.order_by(_seeded(PlexItem.rating_key, seed))
.limit(limit)
q = select(PlexItem).where(
PlexItem.kind == "movie",
PlexItem.library_id.in_(movie_libs),
_unseen_movie_conds(uid),
)
if lo is not None:
q = q.where(PlexItem.duration_s >= lo)
if hi is not None:
q = q.where(PlexItem.duration_s <= hi)
items = db.execute(q).scalars().all()
shuffled = _seeded(PlexItem.rating_key, seed)
blind = db.execute(q.order_by(shuffled, PlexItem.id).limit(limit)).scalars().all()
leaning = affinity.for_user(db, uid)
if leaning:
key = draw.weighted_key(draw.seeded_u(PlexItem.rating_key, seed), leaning.weight(PlexItem))
weighted = db.execute(q.order_by(key, shuffled, PlexItem.id).limit(limit)).scalars().all()
items = draw.blend(list(weighted), list(blind), seed, limit, key=lambda it: it.id)
else:
items = list(blind)
states = rollup.item_state_map(db, uid, [it.id for it in items])
return [movie_card(it, states.get(it.id), lang=lang) for it in items]
backend/app/downloads/adapters.py
+8 -2
View File
@@ -14,10 +14,10 @@ Resolution works both before extraction (from source_kind / URL host) and after
import re
from urllib.parse import urlparse
import yt_dlp
from yt_dlp.networking.impersonate import ImpersonateTarget
from app.config import settings
from app.downloads import egress
from app.downloads.stated import StatedFailure
# Facebook's "Cannot parse data": the page yt-dlp got was a wall, not the content. Both the Chrome retry
@@ -33,6 +33,9 @@ class SourceAdapter:
"""Generic source: yt-dlp handles everything and the ref is already a fetchable URL."""
key = "generic"
# SEC-17: every fetch of this source's download goes through the egress gate (`egress`). A
# source's own host list is no guarantee — its pages, redirects and formats can point anywhere.
egress_guarded = True
# --- extraction (worker._extract_with_fallback / backfill) ---------------------------------
def resolve_url(self, url: str, ydl_opts: dict) -> str:
@@ -92,6 +95,9 @@ class SourceAdapter:
class YouTubeAdapter(SourceAdapter):
key = "youtube"
# Not through the gate: YouTube's fetches go to Google's fixed hosts, and the PO-token provider
# it needs is the sidecar on our internal network, which the gate exists to refuse.
egress_guarded = False
# Errors worth retrying with a different player-client set (YouTube per-client flakiness).
_RETRIABLE = (
"not a bot",
@@ -179,7 +185,7 @@ class FacebookAdapter(SourceAdapter):
post_id = facebook_group_post_id(url)
if post_id is None or not ydl_opts.get("cookiefile"):
return url
with yt_dlp.YoutubeDL(ydl_opts) as ydl:
with egress.open_ydl(ydl_opts) as ydl:
raw = ydl.urlopen(url).read(self._POST_PAGE_MAX_BYTES)
page = raw.decode("utf-8", "replace")
# Measured: a refused or expired cookie gets the login page (~450 KB, a login
backend/app/downloads/backfill.py
+7 -4
View File
@@ -16,7 +16,6 @@ Run once per environment:
from urllib.parse import urlsplit
import yt_dlp
from sqlalchemy import select, update
from sqlalchemy.orm import Session
@@ -25,7 +24,7 @@ from app.config import settings
from app.db import SessionLocal
from app.downloads import adapters
from app.downloads import edit as editmod
from app.downloads import formats, service
from app.downloads import egress, formats, service
from app.downloads.models import DownloadJob, MediaAsset
from app.models import DISPLAY_TEXT_MAX
from app.titles import normalize_title
@@ -43,11 +42,15 @@ def extract_meta(url: str, adapter: adapters.SourceAdapter, asset_id: int) -> di
def _extract_meta_scoped(url: str, adapter: adapters.SourceAdapter, asset_id: int) -> dict | None:
for clients in adapter.client_sets():
opts = formats.build_ydl_opts(
{}, "%(id)s.%(ext)s", lambda d: None, extra=adapter.ydl_opts_extra(clients)
{},
"%(id)s.%(ext)s",
lambda d: None,
extra=adapter.ydl_opts_extra(clients),
egress=adapter.egress_guarded,
)
opts["skip_download"] = True
try:
with yt_dlp.YoutubeDL(opts) as ydl:
with egress.open_ydl(opts) as ydl:
info = ydl.extract_info(url, download=False)
if info:
return info
backend/app/downloads/egress.py
+297
View File
@@ -0,0 +1,297 @@
"""The egress gate: every fetch of a user-supplied download goes through here, and this is where the
address is judged — at the moment of the connection, not when the URL was typed (SEC-17).
`urlguard` judges the URL the user typed, at enqueue and again in the worker. What that cannot see is
every request yt-dlp makes after it: a public host answering 3xx with an internal `Location`, a page
whose formats point inside, a name that resolves public for the check and private for the fetch (DNS
rebinding). So a guarded download runs with yt-dlp's `proxy` set to this gate, a small HTTP proxy on
the worker's loopback. For each request it resolves the host ONCE, refuses the connection if any
address the name has is one the server must not fetch from (`urlguard.is_forbidden`), and otherwise
connects to the very address it vetted — there is no second lookup for a rebinding answer to slip into.
Measured (yt-dlp 2026.8.19, 2026-11-06) on the worker image, with a logging proxy: the `proxy` option
reaches every path a download takes — the Requests handler, the CurlCFFI one an impersonated request
uses, the native HLS downloader and an ffmpeg HLS download (segments included). ffmpeg itself also
speaks rtmp, rtsp, ftp, tcp and more, which no HTTP proxy covers, so a guarded download keeps only the
formats it fetches over http(s) (`HttpOnlyFormats`, added by `open_ydl`) and fetches HLS natively.
A refused connection is answered `403` before anything is dialled, so a live internal port and a dead
one read the same; `errors.user_error_text` then stores the one redacted sentence for both.
Not used for YouTube: those downloads talk to Google's fixed hosts and to the PO-token sidecar, which
lives on the internal network on purpose (`adapters.YouTubeAdapter.egress_guarded`).
"""
import ipaddress
import select
import socket
import socketserver
import threading
from collections.abc import Callable
from typing import Any, cast
from urllib.parse import urlsplit
import yt_dlp
from yt_dlp.postprocessor.common import PostProcessor
from app import logs
from app.downloads import urlguard
log = logs.get_logger(__name__)
Resolver = Callable[..., list]
Forbidden = Callable[[ipaddress.IPv4Address | ipaddress.IPv6Address], bool]
_BLOCKED = b"HTTP/1.1 403 Forbidden\r\nContent-Length: 0\r\nConnection: close\r\n\r\n"
_BAD_GATEWAY = b"HTTP/1.1 502 Bad Gateway\r\nContent-Length: 0\r\nConnection: close\r\n\r\n"
_BAD_REQUEST = b"HTTP/1.1 400 Bad Request\r\nContent-Length: 0\r\nConnection: close\r\n\r\n"
_MAX_HEAD = 64 * 1024
# A tunnel with no traffic either way for this long is closed. Above yt-dlp's own socket_timeout
# (30 s, `formats.build_ydl_opts`), so the gate never ends a transfer yt-dlp would still wait for.
_IDLE_S = 120
# Hop-by-hop headers a proxy must not pass on; `Connection: close` replaces them, so one connection
# carries one request and a reused client connection can never be steered to a second host unvetted.
_HOP = (b"proxy-", b"connection:", b"keep-alive:")
def vetted_address(
host: str, port: int, *, resolve: Resolver = socket.getaddrinfo, forbidden: Forbidden
) -> tuple[str, int] | None:
"""The address to connect to for `host:port`, or None when it must not be fetched: the name has
no answer, or ANY of its addresses is forbidden (a round-robin answer is not a coin flip). IPv4
only, as every download is (`source_address` 0.0.0.0 in `formats.build_ydl_opts`)."""
try:
infos = resolve(host, port, socket.AF_INET, socket.SOCK_STREAM)
except (socket.gaierror, UnicodeError, OSError):
# silent-ok: "no answer" is a refusal, and the caller's refusal is the line that is logged.
return None
addrs = [info[4] for info in infos]
if not addrs:
return None
for addr in addrs:
try:
ip = ipaddress.ip_address(addr[0])
except ValueError:
# silent-ok: an answer that is not an address is refused like a forbidden one, and the
# caller logs the refusal.
return None
if forbidden(ip):
return None
return addrs[0][0], addrs[0][1]
def _read_head(sock: socket.socket) -> bytes | None:
buf = b""
while b"\r\n\r\n" not in buf:
if len(buf) > _MAX_HEAD:
return None
chunk = sock.recv(8192)
if not chunk:
return None
buf += chunk
return buf
def _pipe(a: socket.socket, b: socket.socket) -> None:
try:
_relay(a, b)
except OSError:
# silent-ok: one side hung up mid-transfer (yt-dlp closes a connection once it has read a
# file's size, or gives up on a slow host); whichever side cared reports it itself.
return
def _relay(a: socket.socket, b: socket.socket) -> None:
socks = [a, b]
while True:
ready, _, _ = select.select(socks, [], [], _IDLE_S)
if not ready:
return
for s in ready:
data = s.recv(65536)
if not data:
return
(b if s is a else a).sendall(data)
class _Handler(socketserver.BaseRequestHandler):
def handle(self) -> None:
client: socket.socket = self.request
head = _read_head(client)
if head is None:
return
head_part, _, body = head.partition(b"\r\n\r\n")
lines = head_part.split(b"\r\n")
try:
method, target, version = lines[0].decode("latin1").split(" ", 2)
except ValueError:
# silent-ok: the only client is yt-dlp on this loopback; a malformed request is answered
# 400 and reaches its own error, which the job records.
client.sendall(_BAD_REQUEST)
return
if method == "CONNECT":
host, _, port_s = target.rpartition(":")
host = host.strip("[]")
upstream = self._dial(host, port_s)
if upstream is None:
return
with upstream:
client.sendall(b"HTTP/1.1 200 Connection established\r\n\r\n")
if body:
upstream.sendall(body)
_pipe(client, upstream)
return
u = urlsplit(target)
if u.scheme != "http" or not u.hostname:
client.sendall(_BLOCKED)
return
try:
port = u.port or 80
except ValueError:
# silent-ok: as for a malformed request line above — answered 400, recorded by the job.
client.sendall(_BAD_REQUEST)
return
upstream = self._dial(u.hostname, str(port))
if upstream is None:
return
path = (u.path or "/") + (f"?{u.query}" if u.query else "")
headers = [h for h in lines[1:] if not h.lower().startswith(_HOP)]
request = b"\r\n".join(
[f"{method} {path} {version}".encode("latin1"), *headers, b"Connection: close"]
)
with upstream:
upstream.sendall(request + b"\r\n\r\n" + body)
_pipe(client, upstream)
def _dial(self, host: str, port_s: str) -> socket.socket | None:
"""Connect to the address `host` was vetted at, or answer the client and return None."""
client: socket.socket = self.request
try:
port = int(port_s)
except ValueError:
# silent-ok: as for a malformed request line above — answered 400, recorded by the job.
client.sendall(_BAD_REQUEST)
return None
gate = cast("_Gate", self.server)
addr = vetted_address(host, port, resolve=gate.resolve, forbidden=gate.forbidden)
if addr is None:
log.warning(
"egress refused a connection to %s:%s",
host, # log-pii-ok: the host a download tried to reach, never its path or query
port,
extra=logs.event("download.egress.refused", port=port),
)
client.sendall(_BLOCKED)
return None
try:
return socket.create_connection(addr, timeout=_IDLE_S)
except OSError:
# silent-ok: a public host that does not answer; yt-dlp reports it as the job's own
# network failure, which the job records.
client.sendall(_BAD_GATEWAY)
return None
class _Gate(socketserver.ThreadingTCPServer):
daemon_threads = True
allow_reuse_address = True
def __init__(self, resolve: Resolver, forbidden: Forbidden):
self.resolve = resolve
self.forbidden = forbidden
super().__init__(("127.0.0.1", 0), _Handler)
@property
def url(self) -> str:
return f"http://127.0.0.1:{self.server_address[1]}"
def handle_error(self, request, client_address) -> None:
# socketserver's default prints a bare traceback to stderr, outside the log; a connection the
# gate could not finish is a line with its code instead.
log.exception("egress gate dropped a connection", extra=logs.event("download.egress.error"))
def start_gate(
*, resolve: Resolver = socket.getaddrinfo, forbidden: Forbidden = urlguard.is_forbidden
) -> _Gate:
"""A gate serving on its own thread. The worker uses the one `proxy_url` starts; a test starts
its own, with a resolver and a judgement it controls."""
gate = _Gate(resolve, forbidden)
threading.Thread(target=gate.serve_forever, name="egress-gate", daemon=True).start()
return gate
# The options key `formats.build_ydl_opts` marks a guarded download with; `open_ydl` reads it. yt-dlp
# keeps unknown keys in its params and never acts on them.
GUARDED = "siftlode_egress_guarded"
def open_ydl(opts: dict) -> yt_dlp.YoutubeDL:
"""The one way this package opens yt-dlp (a contract test holds it): a guarded download also
gets `HttpOnlyFormats`, which yt-dlp takes only as an instance, not through the options dict."""
# yt-dlp's stub types its params as a TypedDict; these are the plain dict `formats.build_ydl_opts`
# assembles, the same every caller handed yt-dlp before.
ydl = yt_dlp.YoutubeDL(opts) # type: ignore[arg-type]
if opts.get(GUARDED):
ydl.add_post_processor(HttpOnlyFormats(), when="pre_process")
return ydl
_gate: _Gate | None = None
_gate_lock = threading.Lock()
def proxy_url() -> str:
"""This process's gate, started on first use (the worker's jobs share it)."""
global _gate
with _gate_lock:
if _gate is None:
_gate = start_gate()
return _gate.url
# The protocols fetched over http(s) through the proxy option — the only ones the gate sees: yt-dlp's
# own fetchers (http, f4m, ism, the DASH segments and their generator, native HLS — which a bare m3u8
# also gets, as a guarded download sets hls_prefer_native) and ffmpeg's http input (a live stream or
# a clip), which takes the proxy too. Everything else (rtmp, rtsp, mms, ws) ffmpeg or another tool
# would dial directly.
_HTTP_PROTOCOLS = frozenset(
{
"http",
"https",
"f4m",
"ism",
"m3u8",
"m3u8_native",
"http_dash_segments",
"http_dash_segments_generator",
}
)
class HttpOnlyFormats(PostProcessor):
"""Drops every format a guarded download would fetch past the gate. Runs as `pre_process`, which
yt-dlp re-reads its formats after ("The pre-processors may have modified the formats"), so the
selector never sees them. With nothing left, yt-dlp reports no formats — a refusal, not a leak."""
def set_downloader(self, downloader):
# Not the job's postprocessor hooks: they would report this filter as a "processing" step
# before a byte is fetched (yt-dlp attaches them here, once per call).
self._downloader = downloader
def run(self, information):
# yt-dlp's stub types the info dict as a TypedDict with optional keys; it is a dict here.
info = cast("dict[str, Any]", information)
if info.get("formats"):
info["formats"] = [f for f in info["formats"] if _over_http(f)]
elif info.get("url") and not _over_http(info):
info["url"] = None
return [], information
def _over_http(fmt: dict) -> bool:
url = fmt.get("manifest_url") or fmt.get("url") or ""
if urlsplit(url).scheme not in ("http", "https"):
return False
protocol = fmt.get("protocol") or urlsplit(fmt.get("url") or "").scheme
return all(p in _HTTP_PROTOCOLS for p in protocol.split("+"))
backend/app/downloads/formats.py
+15 -1
View File
@@ -28,6 +28,7 @@ import urllib.parse
from collections.abc import Iterator
from app import logs
from app.downloads import egress as egress_gate
log = logs.get_logger(__name__)
@@ -294,6 +295,8 @@ def build_ydl_opts(
extra: dict | None = None,
cookiefile: str | None = None,
max_bytes: int | None = None,
*,
egress: bool,
) -> dict:
"""Assemble yt-dlp options for a single download to `outtmpl` (a full path template).
@@ -309,7 +312,11 @@ def build_ydl_opts(
a file announced larger than it rather than filling the disk and only then being caught by the
quota accounting. Its ffmpeg downloader (a `download_ranges` fetch) does not read it, nor does
a fragmented download (YouTube's formats: measured, 718 MB fetched with 1 MB left) — so the
pipeline also holds every produced file to what is left by its real size."""
pipeline also holds every produced file to what is left by its real size.
`egress` (SEC-17) routes every fetch through the worker's egress gate, which refuses a connection
into our own network, and keeps only the formats fetched over http(s). Required, with no default:
a new caller has to say which kind of download it builds (`SourceAdapter.egress_guarded`)."""
s = normalize(spec)
h = s["max_height"]
opts: dict = {
@@ -364,6 +371,13 @@ def build_ydl_opts(
if cookiefile:
opts["cookiefile"] = cookiefile
# SEC-17. AFTER the adapter merge, like the cookie: an adapter must not be able to point a
# guarded download past the gate.
if egress:
opts["proxy"] = egress_gate.proxy_url()
opts["hls_prefer_native"] = True
opts[egress_gate.GUARDED] = True
if s["mode"] == "a":
opts["format"] = "bestaudio/best"
opts["postprocessors"].append(
backend/app/downloads/pipeline.py
+4 -2
View File
@@ -17,7 +17,7 @@ from app.downloads import clip as clipmod
from app.downloads import edit as editmod
from app.downloads import errors as dlerrors
from app.downloads import adapters, credentials, formats, mediakind, playable, probe, quota
from app.downloads import backfill
from app.downloads import backfill, egress
from app.downloads import service, storage
from app.downloads import urlguard
from app.downloads.jobdb import (
@@ -394,6 +394,7 @@ def _extract_with_fallback(
_make_progress_hook(job_id, asset_id),
cookiefile=cookiefile,
max_bytes=max_bytes,
egress=adapter.egress_guarded,
),
)
last_exc: Exception | None = None
@@ -405,6 +406,7 @@ def _extract_with_fallback(
extra=adapter.ydl_opts_extra(clients),
cookiefile=cookiefile,
max_bytes=max_bytes,
egress=adapter.egress_guarded,
)
opts["postprocessor_hooks"] = [_make_pp_hook(job_id)]
if section is not None:
@@ -415,7 +417,7 @@ def _extract_with_fallback(
# (Per-ATTEMPT only; a single extract_info that itself runs past the stall timeout is still
# killed — acceptable, as socket_timeout + extractor_retries bound one attempt well under it.)
try:
with yt_dlp.YoutubeDL(opts) as ydl:
with egress.open_ydl(opts) as ydl:
return ydl.extract_info(target, download=True)
except yt_dlp.utils.DownloadError as exc:
last_exc = exc
backend/app/downloads/service.py
+2 -2
View File
@@ -153,7 +153,7 @@ def populate_from_catalog(db: Session, asset: MediaAsset) -> bool:
return False
asset.title = video.title
asset.thumbnail_url = video.thumbnail_url
asset.duration_s = video.duration_seconds
asset.duration_s = video.length_seconds
if video.published_at:
asset.upload_date = video.published_at.date()
channel = db.get(Channel, video.channel_id)
@@ -337,7 +337,7 @@ def _clip_source_facts(db: Session, spec: dict) -> dict:
"title": video.title,
"uploader": channel.title if channel else None,
"thumbnail_url": video.thumbnail_url,
"length_s": video.duration_seconds,
"length_s": video.length_seconds,
}
item = plex_lookups.by_rating_key(db, PlexItem, spec["ref"])
if item is None:
backend/app/downloads/urlguard.py
+8 -18
View File
@@ -10,22 +10,12 @@ What this closes: a URL whose host resolves to a loopback / private / link-local
is refused, at enqueue AND again in the worker before extraction. Both, because DNS can answer
differently between the two moments, and the worker is where the socket is actually opened.
What it does NOT close, stated plainly rather than implied, because both need the same fix — gating
the connection itself with a custom socket factory inside yt-dlp, a much larger change than this:
* DNS rebinding: a host that passes the worker's check and then answers yt-dlp's own resolution
with an internal address.
* HTTP redirects: a public host that answers 3xx with an internal ``Location``. yt-dlp follows it,
and only the URL the user TYPED was ever judged here. An attacker-controlled host or any public
open redirect reaches it, so this is the cheaper of the two. It does NOT usually yield the
response body — a non-media body fails the mp4 remux and stores nothing — but it does hand back
the readout above: the redaction in ``errors.user_error_text`` fires only for the ``network``
class, so a dead internal port (refused → ``network`` → redacted) is distinguishable from a live
one (an HTTP answer that is not media → ``unsupported_host``/``format_unavailable``, raw text
kept). Tracked as SEC-17.
The two checks here stop the straightforward case (typing ``http://127.0.0.1:6379`` or an internal
hostname into the box), which is the reachable one.
What this alone cannot see is every request yt-dlp makes AFTER the typed URL: a public host
answering 3xx with an internal ``Location``, and DNS rebinding (a name that passes this check and
then resolves inside for yt-dlp's own lookup). Those are closed one layer down, by `egress`: a guarded
download's every connection goes through a gate that judges the address it actually dials, with
this module's `is_forbidden`. The checks here stay as the early, structured refusal — a user typing
``http://127.0.0.1:6379`` hears why at once instead of from a failed job.
Not used for YouTube jobs: those resolve to a fixed public host from an 11-char id, so there is no
user-controlled destination to guard.
@@ -62,7 +52,7 @@ class BlockedUrl(Exception):
self.reason: UrlBlockedReason = reason
def _is_forbidden(ip: ipaddress.IPv4Address | ipaddress.IPv6Address) -> bool:
def is_forbidden(ip: ipaddress.IPv4Address | ipaddress.IPv6Address) -> bool:
"""Every address family we must not open a connection to. Checked explicitly rather than
leaning on ``is_private`` alone, whose coverage differs between IPv4 and IPv6."""
# ::ffff:127.0.0.1 is loopback wearing an IPv6 hat — unwrap before judging it.
@@ -113,5 +103,5 @@ def check_url(raw: str) -> None:
ip = ipaddress.ip_address(addr)
except ValueError: # getaddrinfo returned something unparseable; treat as unusable
raise BlockedUrl(REASON_PRIVATE) from None
if _is_forbidden(ip):
if is_forbidden(ip):
raise BlockedUrl(REASON_PRIVATE)
backend/app/draw.py
+118
View File
@@ -0,0 +1,118 @@
"""A weighted draw that stays a surprise: what "Surprise me" deals from.
A title's chance to come first is its weight over the pool's (Efraimidis–Spirakis: order by
-ln(1-u)/w, u uniform per row). The weight is 1 + the title's affinity to what the user watches
(`plex.affinity`, `discover.queries`), so a title nobody leans towards still comes up — only less
often. Two things keep it a surprise rather than a ranking:
- **u is the seed's md5 of the row's key**, the Home surprise's seeded order. With every weight 1
the order is exactly that md5 order, so a user with no history sees the unweighted draw, and one
seed still moves both sources of the Home hero together. A draw that is not seeded (the Plex
Discover rail) takes u from random().
- **One slot of three is blind** (`blend`): dealt from the unweighted order, so the same few
favourites cannot fill the rail.
The YouTube surprise draws over a window of the feed scope rather than all of it (`yt_surprise`):
the window is read in SQL, and `shuffled` / `leaning` order it here with the same u.
Both affinities weigh a watch by `recency` and score against the strongest with `scaled`, so the two
sources lean by one rule.
"""
import hashlib
import math
from collections.abc import Callable, Hashable
from datetime import datetime
from typing import Any
from sqlalchemy import BigInteger, Float, cast, func, literal
from sqlalchemy.dialects.postgresql import BIT
from sqlalchemy.orm import InstrumentedAttribute
from sqlalchemy.sql.elements import ColumnElement
# 2**32: the first 8 hex digits of an md5 as a fraction of one.
_SPAN = 4294967296.0
# A watch with no date (an old import) counts as one this many half-lives ago.
UNDATED_HALF_LIVES = 2
def recency(watched_at: datetime | None, now: datetime, half_life_days: float) -> float:
"""A watch's weight in an affinity: 1 now, half as much every `half_life_days`."""
if watched_at is None:
return 0.5**UNDATED_HALF_LIVES
return 0.5 ** (max(0.0, (now - watched_at).total_seconds() / 86400) / half_life_days)
def scaled(scores: dict[str, float]) -> dict[str, float]:
"""Each score over the strongest, so the strongest is 1; empty when nothing scored."""
top = max(scores.values(), default=0.0)
return {k: v / top for k, v in scores.items()} if top else {}
def seeded_u(
key: ColumnElement[Any] | InstrumentedAttribute[Any], seed: int
) -> ColumnElement[float]:
"""A uniform u in [0, 1) per row, from md5(key || seed) — the order `md5(...)` ascending gives."""
digest = func.md5(func.concat(key, str(seed)))
head = cast(cast(literal("x") + func.substr(digest, 1, 8), BIT(32)), BigInteger)
return cast(head, Float) / _SPAN
def weighted_key(u: ColumnElement[float], weight: ColumnElement[float]) -> ColumnElement[float]:
"""The row's place in a weighted draw, ascending: -ln(1-u)/w. Monotonic in u when every w is 1,
so the unweighted order is unchanged."""
return -func.ln(1.0 - u) / weight
def random_u() -> ColumnElement[float]:
return func.random()
def u_of(key: str, seed: int) -> float:
"""`seeded_u` in Python: the same u for the same key and seed."""
return int(hashlib.md5(f"{key}{seed}".encode()).hexdigest()[:8], 16) / _SPAN
def shuffled(items: list, seed: int, key: Callable[[Any], str]) -> list:
"""`items` in the seed's md5 order — the order `ORDER BY md5(key || seed)` gives."""
return sorted(items, key=lambda it: hashlib.md5(f"{key(it)}{seed}".encode()).hexdigest())
def leaning(
items: list, seed: int, key: Callable[[Any], str], weight: Callable[[Any], float]
) -> list:
"""`items` in the weighted draw's order (`weighted_key` in Python), ties in the md5 order."""
order = {key(it): i for i, it in enumerate(shuffled(items, seed, key))}
return sorted(
items,
key=lambda it: (-math.log(1.0 - u_of(key(it), seed)) / weight(it), order[key(it)]),
)
def blend(
weighted: list, blind: list, offset: int, n: int, key: Callable[[Any], Hashable] = lambda x: x
) -> list:
"""`n` picks: slot i comes from `blind` when (i + offset) % 3 == 0, else from `weighted`, each
skipping what an earlier slot took (`key` tells two picks apart). A side that runs out leaves the
rest to the other."""
out: list = []
taken: set = set()
sources = (iter(weighted), iter(blind))
def nxt(source):
for item in source:
if key(item) not in taken:
return item
return None
for i in range(n):
first, second = (sources[1], sources[0]) if (i + offset) % 3 == 0 else sources
item = nxt(first)
if item is None:
item = nxt(second)
if item is None:
break
taken.add(key(item))
out.append(item)
return out
backend/app/feedquery.py
+71 -12
View File
@@ -13,7 +13,19 @@ import binascii
import json
from datetime import date, datetime, timedelta, timezone
from sqlalchemy import BigInteger, Select, and_, case, cast, false, func, or_, select
from sqlalchemy import (
BigInteger,
Select,
and_,
case,
cast,
false,
func,
literal,
or_,
select,
tuple_,
)
from sqlalchemy.orm import Session, aliased
from app.collation import collated
@@ -68,7 +80,7 @@ def feed_columns(user: User, state) -> list:
Channel.handle.label("channel_handle"),
Video.published_at,
Video.thumbnail_url,
Video.duration_seconds,
Video.length_seconds.label("length_seconds"),
Video.view_count,
Video.is_short,
Video.live_status,
@@ -88,7 +100,7 @@ def serialize(row) -> dict:
"channel_url": _channel_url(row.channel_id, row.channel_handle),
"published_at": iso(row.published_at),
"thumbnail_url": row.thumbnail_url,
"duration_seconds": row.duration_seconds,
"duration_seconds": row.length_seconds,
"view_count": row.view_count,
"is_short": row.is_short,
"live_status": row.live_status,
@@ -260,9 +272,9 @@ def filtered_query(
# OR across the picked channels; every other filter still applies on top.
query = query.where(Video.channel_id.in_(channel_ids))
if min_duration is not None:
query = query.where(Video.duration_seconds >= min_duration)
query = query.where(Video.length_seconds >= min_duration)
if max_duration is not None:
query = query.where(Video.duration_seconds <= max_duration)
query = query.where(Video.length_seconds <= max_duration)
if max_age_days is not None:
cutoff = datetime.now(timezone.utc).timestamp() - max_age_days * 86400
query = query.where(Video.published_at >= datetime.fromtimestamp(cutoff, tz=timezone.utc))
@@ -503,10 +515,10 @@ def sort_keys(sort: str, seed: int, rank_expr=None, lang: str | None = None) ->
"views": [{"expr": Video.view_count, "asc": False, "nullable": True, "kind": "num"}],
"views_asc": [{"expr": Video.view_count, "asc": True, "nullable": True, "kind": "num"}],
"duration_desc": [
{"expr": Video.duration_seconds, "asc": False, "nullable": True, "kind": "num"}
{"expr": Video.length_seconds, "asc": False, "nullable": True, "kind": "num"}
],
"duration_asc": [
{"expr": Video.duration_seconds, "asc": True, "nullable": True, "kind": "num"}
{"expr": Video.length_seconds, "asc": True, "nullable": True, "kind": "num"}
],
"title": [{"expr": title, "asc": True, "nullable": True, "kind": "str"}],
"title_desc": [{"expr": title, "asc": False, "nullable": True, "kind": "str"}],
@@ -536,15 +548,17 @@ def sort_keys(sort: str, seed: int, rank_expr=None, lang: str | None = None) ->
},
pub,
],
# Deterministic per-seed shuffle: the same seed yields the same total order, so
# the cursor stays valid across pages of one shuffled run.
# The ring of stored random keys from the seed's point (`shuffle_walk`): the keys at or after
# it, then the ones before. The same seed yields the same total order, so the cursor stays
# valid across pages; `feed_page` reads it as two index ranges, never as this sort.
"shuffle": [
{
"expr": func.md5(func.concat(Video.id, str(seed))),
"expr": case((Video.shuffle_key < shuffle_start(seed), 1), else_=0),
"asc": True,
"nullable": False,
"kind": "str",
}
"kind": "num",
},
{"expr": Video.shuffle_key, "asc": True, "nullable": False, "kind": "num"},
],
}
# Relevance (search-result ranking) only exists when there's a query to rank against;
@@ -700,6 +714,43 @@ class BadCursor(ValueError):
"""A `cursor` this module did not write (or one from another sort): the caller's bad input."""
# The golden ratio's fraction: consecutive seeds land far apart on the ring, never next to each other.
_GOLDEN = 0.6180339887498949
def shuffle_start(seed: int) -> float:
"""The point in [0, 1) where a seed's shuffle starts on the ring of `Video.shuffle_key`.
Folded to 32 bits first: an int past a float's range overflowed the product, and one past 2**53
left the product no fraction to land on. A seed the client draws (under a billion) is unchanged."""
return ((seed % 2**32) * _GOLDEN) % 1.0
def shuffle_walk(
db: Session, query: Select, seed: int, n: int, after: tuple[int, float, str] | None = None
) -> list:
"""Up to `n` rows of `query` in the seed's ring order: shuffle_key from the seed's point upwards,
then from 0 up to that point. `after` is a cursor's (phase, key, id): the walk resumes past it.
Each arc is an ascending range read along ix_videos_shuffle_key with the scope's filters checked
row by row, so a draw costs the rows it reads, not a sort of the scope — the md5(id || seed) order
it replaced read and sorted every row the filters let through (~350 ms on prod for ~213k)."""
start = shuffle_start(seed)
phase_from, key, last_id = after if after is not None else (0, None, None)
rows: list = []
for phase in (0, 1):
if phase < phase_from or len(rows) >= n:
continue
arc = Video.shuffle_key >= start if phase == 0 else Video.shuffle_key < start
q = query.where(arc)
if key is not None and phase == phase_from:
q = q.where(
tuple_(Video.shuffle_key, Video.id) > tuple_(literal(key), literal(last_id))
)
rows += db.execute(q.order_by(Video.shuffle_key, Video.id).limit(n - len(rows))).all()
return rows
def feed_page(
db: Session,
user: User,
@@ -734,6 +785,14 @@ def feed_page(
k["value"] = v
if ranked and window is not None:
rows = windowed_relevance_rows(db, query, keys, window, cursor_id, limit + 1)
elif sort == "shuffle":
after = None
if vals is not None and cursor_id is not None:
try:
after = (int(vals[0]), float(vals[1]), str(cursor_id))
except (TypeError, ValueError, OverflowError) as e: # Infinity, a 400-digit key
raise BadCursor("Invalid feed cursor") from e
rows = shuffle_walk(db, query, seed, limit + 1, after)
else:
query = query.order_by(*order_by(keys))
if cursor_id is not None:
backend/app/plex/affinity.py
+110
View File
@@ -0,0 +1,110 @@
"""What a user's Plex history leans towards, as a weight for a surprise draw (`app.draw`).
Two signals, both already in the database (no TMDB call):
- **genres** of the titles the user watched, the recent ones weighing more (a half-life of
HALF_LIFE_DAYS), each scored 0..1 against the strongest;
- the **TMDB recommendations** stored on those titles (`recommendations`), each scored by the most
recent title that recommends it.
A title counts once however many of its episodes were watched — a binge of one show would otherwise
outvote every film — and only the RECENT_TITLES latest titles count. A candidate's weight is
1 + GENRE_WEIGHT × (the mean score of its genres) + REC_WEIGHT × (its recommendation score), so a
title with no affinity keeps weight 1 and still comes up.
"""
import json
from dataclasses import dataclass, field
from datetime import datetime, timezone
from sqlalchemy import Float, cast, func, literal, literal_column, select
from sqlalchemy.dialects.postgresql import JSONB
from sqlalchemy.orm import Session
from sqlalchemy.sql.elements import ColumnElement
from app import draw
from app.plex.models import PlexItem, PlexShow, PlexState
from app.userscope import owned_by
HALF_LIFE_DAYS = 180
RECENT_TITLES = 200
GENRE_WEIGHT = 2.0
REC_WEIGHT = 3.0
@dataclass(frozen=True)
class Affinity:
genres: dict[str, float] = field(default_factory=dict)
movie_recs: dict[str, float] = field(default_factory=dict)
show_recs: dict[str, float] = field(default_factory=dict)
def __bool__(self) -> bool:
return bool(self.genres or self.movie_recs or self.show_recs)
def weight(self, model: type[PlexItem] | type[PlexShow]) -> ColumnElement[float]:
"""The draw weight of a `model` row (a movie or a show)."""
scores = cast(literal(json.dumps(self.genres)), JSONB)
recs = self.show_recs if model is PlexShow else self.movie_recs
genre = func.jsonb_array_elements_text(model.genres).table_valued("value").alias("g")
genre_aff = (
select(
func.coalesce(
func.avg(func.coalesce(cast(scores.op("->>")(genre.c.value), Float), 0.0)),
0.0,
)
)
.select_from(genre)
.scalar_subquery()
)
rec_aff = func.coalesce(
cast(cast(literal(json.dumps(recs)), JSONB).op("->>")(model.guid), Float), 0.0
)
return literal_column("1.0") + GENRE_WEIGHT * genre_aff + REC_WEIGHT * rec_aff
def for_user(db: Session, uid: int) -> Affinity:
"""The user's affinity from their latest watched titles; empty (`not aff`) with no history."""
# One key per title: a show for an episode, the item for a film. concat() of a NULL show_id is
# just "s", which nullif turns back into NULL so the film's key takes over.
title = func.coalesce(
func.nullif(func.concat("s", PlexItem.show_id), "s"), func.concat("m", PlexItem.id)
).label("title")
latest = (
select(title, PlexState.watched_at, PlexItem.id.label("item_id"), PlexItem.show_id)
.join(PlexItem, PlexItem.id == PlexState.item_id)
.where(owned_by(PlexState.user_id, uid), PlexState.status == "watched")
.distinct(title)
.order_by(title, PlexState.watched_at.desc().nullslast())
.subquery()
)
rows = db.execute(
select(
latest.c.watched_at,
PlexItem.genres,
PlexItem.recommendations,
PlexShow.genres,
PlexShow.recommendations,
)
.select_from(latest)
.join(PlexItem, PlexItem.id == latest.c.item_id)
.outerjoin(PlexShow, PlexShow.id == latest.c.show_id)
.order_by(latest.c.watched_at.desc().nullslast())
.limit(RECENT_TITLES)
).all()
if not rows:
return Affinity()
now = datetime.now(timezone.utc)
genres: dict[str, float] = {}
movie_recs: dict[str, float] = {}
show_recs: dict[str, float] = {}
for watched_at, item_genres, item_recs, show_genres, show_recs_json in rows:
r = draw.recency(watched_at, now, HALF_LIFE_DAYS)
for g in (show_genres if show_genres is not None else item_genres) or []:
genres[g] = genres.get(g, 0.0) + r
for rec in (show_recs_json if show_recs_json is not None else item_recs) or []:
if not rec.get("tmdb_id"):
continue
bucket = show_recs if rec.get("media_type") == "tv" else movie_recs
guid = f"tmdb:{rec['tmdb_id']}"
bucket[guid] = max(bucket.get(guid, 0.0), r)
return Affinity(draw.scaled(genres), draw.scaled(movie_recs), draw.scaled(show_recs))
backend/app/plex/discover.py
+21 -13
View File
@@ -22,7 +22,8 @@ from sqlalchemy.orm import Session
from app.collation import sorted_words
from app.discover.queries import RAIL_LIMIT
from app.models import User
from app.plex import rollup
from app import draw
from app.plex import affinity, rollup
from app.plex.cards import SHOW_NEW, display_title, leaf_card, movie_card, show_card
from app.plex.models import PlexItem, PlexShow, PlexState
from app.plex.query import (
@@ -399,22 +400,25 @@ def discover_page(db: Session, user: User, scope: str, lang: str, f: LibraryFilt
}
def _draw(db: Session, stmt, key_col, per: int, exclude: list[str]) -> list:
def _draw(db: Session, stmt, key_col, per: int, exclude: list[str], weight=None) -> list:
"""`per` random rows of `stmt`, the previous draw's (`exclude`) left out while the pool allows: a
reshuffle over a small filtered pool kept dealing the same few titles. What the fresh rows leave
short is topped up from the excluded ones, so a pool smaller than a draw still fills the rail."""
fresh = (
db.execute(stmt.where(key_col.not_in(exclude)).order_by(func.random()).limit(per))
.scalars()
.all()
if exclude
else []
)
short is topped up from the excluded ones, so a pool smaller than a draw still fills the rail.
With a `weight` (`plex.affinity`) the draw leans on it, every third slot left blind (`app.draw`)."""
def pick(q, n: int) -> list:
blind = db.execute(q.order_by(func.random()).limit(n)).scalars().all()
if weight is None:
return list(blind)
key = draw.weighted_key(draw.random_u(), weight)
leaning = db.execute(q.order_by(key).limit(n)).scalars().all()
return draw.blend(list(leaning), list(blind), 1, n, key=lambda r: r.rating_key)
fresh = pick(stmt.where(key_col.not_in(exclude)), per) if exclude else []
if len(fresh) == per:
return list(fresh)
return fresh
taken = [key_col.not_in([r.rating_key for r in fresh])] if fresh else []
rest = db.execute(stmt.where(*taken).order_by(func.random()).limit(per - len(fresh))).scalars()
return [*fresh, *rest.all()]
return [*fresh, *pick(stmt.where(*taken), per - len(fresh))]
def _kept(
@@ -474,6 +478,8 @@ def surprise(
if keep:
return {"scope": scope, "items": _kept(db, user.id, movie_ids, show_ids, keep, lang)}
per = min(limit, SURPRISE_PER_KIND)
# What the user's history leans towards; with none, every title weighs the same (a plain draw).
leaning = affinity.for_user(db, user.id)
movies = []
if movie_ids:
# Exclude movies this user has watched.
@@ -501,6 +507,7 @@ def surprise(
PlexItem.rating_key,
per,
exclude,
leaning.weight(PlexItem) if leaning else None,
)
shows = []
if show_ids:
@@ -523,6 +530,7 @@ def surprise(
PlexShow.rating_key,
per,
exclude,
leaning.weight(PlexShow) if leaning else None,
)
mcards = _movie_cards(db, user.id, movies, lang)
scards = _show_cards(db, user.id, shows, lang)
backend/app/plex/models.py
+9 -8
View File
@@ -227,16 +227,17 @@ class PlexItem(Base, TimestampMixin, UpdatedAtMixin):
playable: Mapped[str | None] = mapped_column(String(12), index=True)
probed_at: Mapped[datetime | None] = mapped_column(DateTime(timezone=True))
# Transcode HW-capability facet (the playback-class filter). NULL for non-transcode items; for a
# `playable="transcode"` item: True = the prod iGPU can hardware-transcode it (codec is HW-decodable
# AND the source is 8-bit SDR), False = software libx264 only (AV1/XviD, or a 10-bit/HDR source that
# breaks the VAAPI nv12 pipeline — see plex/stream.py). Derived by `plex.sync.transcode_hw_capable`:
# set codec-only at sync (bit depth unknown from the cheap listing), then refined by the bit-depth
# backfill (`_probe_media_batch`) which downgrades a 10-bit/HDR HEVC to False. The one filterable
# home of the invariant `_want_hw`'s source gate computes at play time — kept in sync via the shared
# classifier so the catalog facet and the runtime encoder choice never disagree by codec.
# `playable="transcode"` item: True = the prod iGPU ENCODES it (a HW-decodable 8-bit SDR source, a
# true-HDR one through the OpenCL tonemap, or an 8-bit source in a codec the CPU decodes for it —
# VC-1, XviD, DivX), False = software libx264 (AV1, or a 10-bit SDR source — see plex/stream.py).
# Derived by `plex.sync.transcode_hw_capable`: set codec-only at sync (bit depth unknown from the
# cheap listing), then refined by the bit-depth backfill (`_probe_media_batch`), which downgrades a
# 10-bit SDR HEVC to False. The filterable twin of the rule `_select_encode` applies at play time;
# the two share their codec sets, not their code, so a change to one tree is a change to both.
transcode_hw: Mapped[bool | None] = mapped_column(Boolean, index=True)
# Video bit depth (8/10/12…) from the Plex media Stream, filled by the bit-depth backfill. NULL =
# not yet probed. Surfaced on the info page (e.g. "HEVC · 10-bit") and feeds `transcode_hw`.
# not yet probed. Surfaced on the info page (e.g. "HEVC · 10-bit") and feeds `transcode_hw` (a
# 10-bit SDR source drops to software; an HDR one keeps the GPU through the tonemap).
video_bit_depth: Mapped[int | None] = mapped_column(Integer)
# Watermark for the bit-depth backfill: when this item's media Stream was last probed for
# bit-depth/HDR. NULL = never (the backfill's work-list). Only HW-decodable transcode items are
backend/app/plex/stream.py
+86 -41
View File
@@ -4,9 +4,10 @@ Playback is from the LOCAL physical file. Browser-compatible files (playable="di
raw with HTTP range requests. Everything else that is h264 video (playable="remux") is remuxed on
the fly to HLS with **video stream-copy** (cheap, I/O-bound — no video re-encode) and audio to AAC
when needed. Non-h264 video (playable="transcode": HEVC/VP9/AV1/MPEG-2/XviD…) needs a full re-encode
to H.264: on the prod host that runs on the Intel iGPU via VAAPI (`h264_vaapi`, near-zero CPU), with
software libx264 as the automatic fallback where the render node is absent or the source codec isn't
hardware-decodable. Same HLS session/seek machinery either way — only the video encoder differs.
to H.264: on the prod host that is encoded on the Intel iGPU via VAAPI (`h264_vaapi`) — decoded there
too when it can, else decoded on the CPU and uploaded (VC-1, XviD, DivX) — with software libx264 as
the fallback where the render node is absent, for a source neither GPU route takes (AV1, 10-bit SDR),
and when a GPU session produces nothing. Same HLS session/seek machinery either way.
Seek model (Jellyfin-style): one ffmpeg session per item, started at a given offset via `-ss`
(fast keyframe seek). A seek beyond the generated region restarts the session at the new offset,
@@ -35,7 +36,7 @@ from app.config import settings
from app.downloads.edit import encode_thread_cap
from app.plex.models import PlexItem
from app.plex import paths
from app.plex.sync import HDR_TRANSFERS, HW_DECODABLE_CODECS
from app.plex.sync import CPU_DECODE_CODECS, HDR_TRANSFERS, HW_DECODABLE_CODECS
log = logs.get_logger(__name__)
_HLS_ROOT = (
@@ -120,6 +121,7 @@ class HlsSession:
hw: bool = False,
tonemap: bool = False,
user_id: int = 0,
cpu_decode: bool = False,
):
# Unguessable per-session id, carried in the URLs the client is handed. A seek-restart
# replaces the viewer's registered session for the item — and the old player's next request
@@ -135,6 +137,7 @@ class HlsSession:
self.transcode = transcode # video re-encoded (non-h264 source), not stream-copied
self.hw = hw # transcode ran on the VAAPI hardware encoder (else libx264 software)
self.tonemap = tonemap # HW HDR/10-bit path: OpenCL tonemap between VAAPI decode and encode
self.cpu_decode = cpu_decode # HW encode of frames the CPU decoded (a codec the GPU can't)
self.start_s = start_s # the REQUESTED seek offset
# The REAL media start: with `-ss X -c:v copy` ffmpeg can only cut at the nearest earlier
# keyframe K<=X, and hls.js zero-bases the timeline to K (so video.currentTime=0 is K, not X).
@@ -188,8 +191,11 @@ _HLS_TAIL = [
# * VAAPI hardware (`h264_vaapi`) — the prod default once the iGPU is passed in. A full GPU pipeline
# (decode + scale + encode all on-device) offloads the CPU almost entirely, so several concurrent
# 1080p transcodes are fine and the API stays responsive.
# * CPU decode → VAAPI encode (`format=nv12,hwupload` + `h264_vaapi`) — a codec the GPU cannot decode
# but whose 8-bit picture it encodes (`CPU_DECODE_CODECS`: VC-1, XviD, DivX). The encode is the
# expensive half, so this still takes most of the load off the CPU.
# * libx264 software — the automatic fallback where `/dev/dri` is absent (dev box, or prod before the
# iGPU enablement) or the source codec isn't hardware-decodable (AV1/XviD on Gen 9.5). CPU-bound, so
# iGPU enablement) or the source fits neither GPU route (AV1, a 10-bit SDR source). CPU-bound, so
# it reserves a core via `-threads` (see encode_thread_cap).
# Everything else — the `-copyts`/`_HLS_TAIL`/segment/seek-restart model and the AAC audio path — is
# unchanged, so hls.js, subtitles, multi-audio and seeking keep working exactly as for remux.
@@ -198,10 +204,10 @@ _VAAPI_DEVICE = "/dev/dri/renderD128"
# here; its presence is the precondition for the HW HDR→SDR tonemap graph (see _opencl_present).
_OPENCL_ICD_DIR = Path("/etc/OpenCL/vendors")
# Codecs the prod iGPU (UHD 630 / Quick Sync Gen 9.5) can hardware-DECODE — the precondition for the
# full-GPU VAAPI pipeline. A source outside it (AV1, MPEG-4 ASP/XviD…) falls back to software even with
# the render node present, because the GPU can't decode it to feed its own encoder. THE single source is
# plex.sync, so the runtime encoder choice here and the catalog `transcode_hw` facet can't disagree by
# codec; aliased to the old name to keep the call sites below unchanged.
# full-GPU VAAPI pipeline. A source outside it is decoded on the CPU when its codec is in
# `CPU_DECODE_CODECS` (the GPU still encodes), else transcoded in software. THE single source of both
# sets is plex.sync, so the runtime encoder choice here and the catalog `transcode_hw` facet can't
# disagree by codec; aliased to the old name to keep the call sites below unchanged.
_HW_DECODABLE = HW_DECODABLE_CODECS
# Color-transfer values that mark a true-HDR source (PQ / HLG). A HW-decodable HDR source can't take the
# plain scale_vaapi=format=nv12 path (Gen 9.5 VPP refuses it) — it needs the OpenCL tonemap graph
@@ -302,16 +308,22 @@ def _opencl_present() -> bool:
return False
def _hwaccel_in_opts(hw: bool, tonemap: bool = False) -> list[str]:
def _hwaccel_in_opts(hw: bool, tonemap: bool = False, cpu_decode: bool = False) -> list[str]:
"""VAAPI decode options, placed before `-i` so the source is decoded (and kept) on the GPU.
`-hwaccel_output_format vaapi` keeps frames as GPU surfaces all the way to `h264_vaapi` — no
round-trip to system memory. Empty for the software path.
The tonemap variant additionally derives an OpenCL device FROM the VAAPI one (`opencl=ocl@va`) so
tonemap_opencl can map the decoded surfaces zero-copy (see _venc_opts); it needs named devices, so it
uses `-init_hw_device` + `-hwaccel_device` rather than the plain path's implicit `-vaapi_device`."""
uses `-init_hw_device` + `-hwaccel_device` rather than the plain path's implicit `-vaapi_device`.
The CPU-decode variant opens the VAAPI device for the FILTER graph only (`hwupload` needs one) and
asks for no `-hwaccel`: the codec is one the GPU refuses to decode. This exact form is the one
measured on prod with the HLS muxer and `-copyts` (and to the end of the file)."""
if not hw:
return []
if cpu_decode:
return ["-init_hw_device", f"vaapi=va:{_VAAPI_DEVICE}", "-filter_hw_device", "va"]
if tonemap:
return [
"-init_hw_device",
@@ -359,6 +371,7 @@ def _venc_opts(
preset: str = "veryfast",
crf: int = 20,
max_height: int = 0,
cpu_decode: bool = False,
) -> list[str]:
"""The video-encoder args that replace `-c:v copy` for a transcode session (returns the copy args
when not transcoding). A re-encode must MINT its own keyframes at the segment boundaries — the copy
@@ -385,6 +398,10 @@ def _venc_opts(
"h264_vaapi",
*force_kf,
]
if hw and cpu_decode:
# The CPU decoded the frames: convert them to the encoder's 8-bit nv12 and upload them to the
# device `-filter_hw_device` named (see _hwaccel_in_opts).
return ["-vf", "format=nv12,hwupload", "-c:v", "h264_vaapi", *force_kf]
if hw:
# scale_vaapi with format=nv12 downconverts a 10-bit (p010) HEVC source to 8-bit nv12 so
# h264_vaapi (8-bit High) can encode it; for an 8-bit source it's a cheap format pass-through.
@@ -422,13 +439,18 @@ def _render_node_present() -> bool:
return Path(_VAAPI_DEVICE).exists()
def _select_encode(db: DbSession, item: PlexItem, transcode: bool, src: Path) -> tuple[bool, bool]:
"""Pick the transcode video pipeline for this session — returns `(hw, tonemap)`:
* `(False, False)` → software libx264 (the fallback: no render node, non-decodable source, or a
10-bit/HDR source with no OpenCL runtime to tonemap on the GPU).
* `(True, False)` → full-GPU VAAPI (`scale_vaapi=format=nv12` + `h264_vaapi`) — an 8-bit SDR source.
* `(True, True)` → VAAPI decode → OpenCL tonemap → VAAPI encode — a true-HDR (PQ/HLG) source, when
the OpenCL runtime is present (Gen 9.5 VPP can't tonemap natively).
def _select_encode(
db: DbSession, item: PlexItem, transcode: bool, src: Path
) -> tuple[bool, bool, bool]:
"""Pick the transcode video pipeline for this session — returns `(hw, tonemap, cpu_decode)`:
* `(False, False, False)` → software libx264 (the fallback: no render node, a source neither GPU
route takes, or a 10-bit/HDR source with no OpenCL runtime to tonemap on the GPU).
* `(True, False, False)` → full-GPU VAAPI (`scale_vaapi=format=nv12` + `h264_vaapi`) — an 8-bit SDR
source the GPU decodes.
* `(True, True, False)` → VAAPI decode → OpenCL tonemap → VAAPI encode — a true-HDR (PQ/HLG)
source, when the OpenCL runtime is present (Gen 9.5 VPP can't tonemap natively).
* `(True, False, True)` → CPU decode → `hwupload` → VAAPI encode — an 8-bit SDR source in a codec
the GPU cannot decode (`CPU_DECODE_CODECS`).
A 10-bit *SDR* source is deliberately NOT a `(True, True)` case: it needs a plain 10→8-bit downconvert,
not the HDR tone curve (which would visibly darken/flatten an SDR picture), so it drops to software.
@@ -438,31 +460,40 @@ def _select_encode(db: DbSession, item: PlexItem, transcode: bool, src: Path) ->
backstops a session that produces nothing — but still routing a true-HDR source through the tonemap
graph so it doesn't hit the scale_vaapi wall); "auto" (default) → HW only when the render node is
present AND the iGPU can hardware-decode the source, with tonemap chosen only for a true-HDR (PQ/HLG)
source that has an OpenCL runtime, while a 10-bit SDR source drops to software."""
source that has an OpenCL runtime, while a 10-bit SDR source drops to software. In both "on" and
"auto" a codec the GPU is known not to decode goes the CPU-decode route: asking the GPU to decode it
only bought the fallback's wasted second before libx264."""
sw = (False, False, False)
if not transcode:
return (False, False)
return sw
mode = sysconfig.get_str(db, "plex_transcode_hw")
if mode == "off" or not _render_node_present():
return (False, False)
return sw
codec = (item.codec_video or "").lower()
if mode == "on":
if codec in CPU_DECODE_CODECS:
return (True, False, True)
# Trust the admin: HW wherever the render node exists, skipping the decodability gate (the runtime
# backstop catches a miss). Only a true-HDR (PQ/HLG) source needs the tonemap graph — a 10-bit SDR
# source takes the plain scale path here, NOT the HDR tone curve. Probe only when OpenCL exists
# (short-circuit), so "on" stays probe-free where tonemap can't apply anyway.
tonemap = _opencl_present() and _probe_video_kind(src) == _VK_HDR
return (True, tonemap)
return (True, tonemap, False)
# auto
if (item.codec_video or "").lower() not in _HW_DECODABLE:
return (False, False)
if codec in CPU_DECODE_CODECS:
# The upload converts to 8-bit nv12: only an 8-bit SDR picture comes through it unchanged.
return (True, False, True) if _probe_video_kind(src) == _VK_HWSAFE else sw
if codec not in _HW_DECODABLE:
return sw
kind = _probe_video_kind(src)
if kind == _VK_HWSAFE:
return (True, False) # 8-bit SDR → full-GPU scale path
return (True, False, False) # 8-bit SDR → full-GPU scale path
if kind == _VK_HDR:
# PQ/HLG: HW OpenCL tonemap when the runtime is present, else software libx264.
return (True, True) if _opencl_present() else (False, False)
return (True, True, False) if _opencl_present() else sw
# _VK_SDR10 — 10-bit SDR: a plain 10→8-bit downconvert, done exactly on the CPU. Never the HDR tone
# curve (it would alter an SDR picture), and off the unverified scale_vaapi 10-bit path — software.
return (False, False)
return sw
def _want_hw(db: DbSession, item: PlexItem, transcode: bool, src: Path) -> bool:
@@ -471,16 +502,17 @@ def _want_hw(db: DbSession, item: PlexItem, transcode: bool, src: Path) -> bool:
return _select_encode(db, item, transcode, src)[0]
# CPU-nice increment for the SOFTWARE transcode child. The Plex ffmpeg is a Popen in the long-lived API
# process (which must NOT itself be niced), so — unlike the download worker's _deprioritize (run once in
# its own main()) — we lower priority per-child via preexec_fn in the fork. The GPU path needs none of
# this. Reuses the download child's nice default so all co-located re-encoders share one handicap.
# CPU-nice increment for a transcode child that works the CPU: a software encode, or a CPU decode
# feeding the GPU encoder. The Plex ffmpeg is a Popen in the long-lived API process (which must NOT
# itself be niced), so — unlike the download worker's _deprioritize (run once in its own main()) — we
# lower priority per-child via preexec_fn in the fork. The full-GPU path needs none of this. Reuses
# the download child's nice default so all co-located re-encoders share one handicap.
_SW_TRANSCODE_NICE = settings.download_child_nice
def _sw_child_preexec() -> None:
"""Runs in the forked child (POSIX) before exec: renice this software-transcode ffmpeg down so it
can't starve the co-located API / Postgres / playback. Best-effort — a failed renice must not fail
"""Runs in the forked child (POSIX) before exec: renice this CPU-bound transcode ffmpeg down so
it can't starve the co-located API / Postgres / playback. Best-effort — a failed renice must not fail
playback."""
try:
os.nice(_SW_TRANSCODE_NICE)
@@ -540,13 +572,15 @@ def _ffmpeg_cmd(
crf: int = 20,
max_height: int = 0,
loudnorm: bool = False,
cpu_decode: bool = False,
) -> list[str]:
# For a transcode session the video is re-encoded (VAAPI or libx264) instead of stream-copied; the
# hwaccel decode options go before -i, the encoder args replace `-c:v copy` below. Everything else is
# identical to remux. `tonemap` is the HDR/10-bit HW variant (OpenCL tonemap between VAAPI decode and
# encode) — it needs matching decode-side (named devices) and encode-side (the OpenCL filter graph) opts.
venc = _venc_opts(transcode, hw, tonemap, preset, crf, max_height)
args = ["ffmpeg", "-nostdin", "-loglevel", "error", *_hwaccel_in_opts(hw, tonemap)]
# `cpu_decode` is the HW encode of CPU-decoded frames: a device for the filter graph, no -hwaccel.
venc = _venc_opts(transcode, hw, tonemap, preset, crf, max_height, cpu_decode)
args = ["ffmpeg", "-nostdin", "-loglevel", "error", *_hwaccel_in_opts(hw, tonemap, cpu_decode)]
# Keep ORIGINAL timestamps on the output so the first segment carries the true absolute PTS of the
# keyframe ffmpeg actually seeked to. hls.js still zero-bases playback to that PTS; we read it back
# (see _probe_first_pts) to learn the real content offset. Without -copyts the segment PTS is
@@ -800,7 +834,11 @@ def start_session(
return directory / segment_name(i, _VIDEO_VARIANT if variants else None)
def _spawn(
hw: bool, tonemap: bool = False, *, replace_only: "HlsSession | None" = None
hw: bool,
tonemap: bool = False,
cpu_decode: bool = False,
*,
replace_only: "HlsSession | None" = None,
) -> "HlsSession | None":
# `replace_only` makes the supersede identity-aware for the runtime fallback: the whole
# check-and-replace happens under ONE _lock, so a newer same-key session (a concurrent
@@ -815,10 +853,12 @@ def start_session(
preset = sysconfig.get_str(db, "plex_transcode_preset")
crf = sysconfig.get_int(db, "plex_transcode_crf")
max_height = sysconfig.get_int(db, "plex_transcode_max_height")
# Nice the SOFTWARE transcode child (not the HW path, and not remux) so a CPU-bound re-encode
# can't starve the co-located API/playback. preexec_fn runs in the forked child (POSIX); on a
# platform without fork (the API always runs in a Linux container in prod) it's simply omitted.
preexec = _sw_child_preexec if (transcode and not hw and os.name == "posix") else None
# Nice the child that works the CPU — a SOFTWARE transcode, or a CPU decode feeding the GPU
# encoder (not the full-GPU path, and not remux) — so it can't starve the co-located
# API/playback. preexec_fn runs in the forked child (POSIX); on a platform without fork (the API
# always runs in a Linux container in prod) it's simply omitted.
cpu_bound = transcode and (not hw or cpu_decode)
preexec = _sw_child_preexec if (cpu_bound and os.name == "posix") else None
doomed: list[tuple[HlsSession, Path | None]] = []
# Finished even when the new session fails to start (no space, no ffmpeg, a DB error): what
# was moved aside under the lock is otherwise left for a restart's sweep.
@@ -851,6 +891,7 @@ def start_session(
crf=crf,
max_height=max_height,
loudnorm=loudnorm,
cpu_decode=cpu_decode,
),
stdout=subprocess.DEVNULL,
stderr=subprocess.DEVNULL,
@@ -866,6 +907,7 @@ def start_session(
hw=hw,
tonemap=tonemap,
user_id=user_id,
cpu_decode=cpu_decode,
)
_sessions[reg] = s
# The one line that says which pipeline a session actually got. Everything downstream
@@ -879,6 +921,7 @@ def start_session(
transcode=transcode,
hw=hw,
tonemap=tonemap,
cpu_decode=cpu_decode,
variants=len(variants or ()),
start_s=round(start_s, 1),
),
@@ -937,7 +980,8 @@ def start_session(
if not wait_for(_seg(i), timeout=_WARMUP_SEG_TIMEOUT_S, proc=s.proc):
break
log.info(
"plex hls session start key=%s req=%.1f real=%.3f multi=%s audio=%s transcode=%s hw=%s tonemap=%s",
"plex hls session start key=%s req=%.1f real=%.3f multi=%s audio=%s transcode=%s hw=%s"
" tonemap=%s cpu_decode=%s",
key,
start_s,
s.media_start_s,
@@ -946,6 +990,7 @@ def start_session(
transcode,
s.hw,
s.tonemap,
s.cpu_decode,
)
return s
backend/app/plex/sync.py
+34 -21
View File
@@ -60,14 +60,20 @@ def classify_playable(vcodec: str | None, acodec: str | None, container: str | N
# Video codecs the prod iGPU (UHD 630 / Quick Sync Gen 9.5) can hardware-DECODE — the precondition for
# the full-GPU VAAPI transcode pipeline. THE single source of this set: plex.stream imports it here so
# the runtime encoder choice (`_want_hw`) and this catalog facet can never disagree by codec. A source
# outside it (AV1, MPEG-4 ASP/XviD…) is software-only even on the iGPU (it can't decode it to feed its
# own encoder).
HW_DECODABLE_CODECS = frozenset({"h264", "hevc", "h265", "vp9", "vc1", "mpeg2video"})
# the runtime encoder choice (`_select_encode`) and this catalog facet can never disagree by codec. VC-1
# is NOT in it: measured on prod, the iHD driver answers "No support for codec vc1 profile 3" (VC-1
# Advanced, what a Blu-ray remux carries), so every session that asked the GPU to decode one died and
# fell back to libx264.
HW_DECODABLE_CODECS = frozenset({"h264", "hevc", "h265", "vp9", "mpeg2video"})
# Codecs the iGPU cannot decode but whose picture it can still ENCODE: the CPU decodes, the frames go up
# to the GPU as 8-bit nv12, and h264_vaapi encodes (plex.stream's `cpu_decode` path). Measured on prod:
# three segments in ~5 s for a 1080p VC-1 remux (libx264: ~17 s) and ~1 s for an SD XviD or DivX
# episode. Only codecs that are always 8-bit SDR, and only ones measured; any other codec stays software.
CPU_DECODE_CODECS = frozenset({"vc1", "mpeg4", "msmpeg4v3"})
# Color-transfer values marking an HDR source (PQ / HLG). THE single source of this set: plex.stream
# (the tonemap choice) and downloads.clip (a clip's tonemap) import it here, so a transfer added for
# one is tonemapped by all. A 10-bit/HDR source breaks the VAAPI scale=nv12 path on Gen 9.5 (see
# plex.stream) → software transcode, so it counts as NOT HW-capable.
# one is tonemapped by all. An HDR source breaks the plain VAAPI scale=nv12 path on Gen 9.5, so it takes
# the OpenCL tonemap graph (see plex.stream) — still the GPU encoder.
HDR_TRANSFERS = frozenset({"smpte2084", "arib-std-b67"})
@@ -78,18 +84,24 @@ def transcode_hw_capable(
hdr: bool = False,
) -> bool | None:
"""The stored `PlexItem.transcode_hw` facet: for a `playable="transcode"` item, whether the prod
iGPU can HARDWARE-transcode it (else software libx264 only); None for a non-transcode item. HW needs
a HW-decodable codec AND an 8-bit SDR source — a 10-bit/HDR source (bit_depth>8 or an HDR transfer)
breaks the VAAPI nv12 pipeline and falls back to software. `bit_depth`/`hdr` are unknown at sync
time (the cheap listing carries no Stream), so they default to 8-bit SDR (codec-only) and the
bit-depth backfill refines a 10-bit/HDR HEVC to False later."""
iGPU ENCODES it ("Transcoded on the hardware encoder"), else software libx264; None for a
non-transcode item. The same answer plex.stream's `_select_encode` gives on prod (render node and
OpenCL runtime present, policy "auto"):
* a HW-decodable codec: 8-bit SDR → the full-GPU path; HDR → the OpenCL tonemap, also on the GPU;
10-bit SDR → software (an exact downconvert, never the HDR tone curve);
* a CPU-decode codec (`CPU_DECODE_CODECS`) at 8 bits → CPU decode, GPU encode;
* any other codec → software.
`bit_depth`/`hdr` are unknown at sync time (the cheap listing carries no Stream), so they default to
8-bit SDR and the bit-depth backfill refines a 10-bit SDR HEVC to False later."""
if playable != "transcode":
return None
if (vcodec or "").lower() not in HW_DECODABLE_CODECS:
return False
if bit_depth is not None and bit_depth > 8:
return False
return not hdr
codec = (vcodec or "").lower()
deep = bit_depth is not None and bit_depth > 8
if codec in HW_DECODABLE_CODECS:
return hdr or not deep
if codec in CPU_DECODE_CODECS:
return not deep and not hdr
return False
def _video_stream_props(meta: dict) -> tuple[int | None, bool]:
@@ -266,8 +278,9 @@ def sync(db: Session) -> dict:
# a Plex shape change, a DB error or our own bug, and `%s` told them apart badly.
log.exception("the Plex people-enrich batch failed")
# Probe a bounded batch of HW-decodable transcode items for bit-depth/HDR, so the
# playback-class filter's HW/SW split is precise (a 10-bit/HDR HEVC → software) over
# successive runs. Only HW-decodable-codec transcodes are probed — the rest are SW by codec.
# playback-class filter's HW/SW split is precise (a 10-bit SDR HEVC → software) over
# successive runs. Only HW-decodable-codec transcodes are probed — the rest are classed
# by codec alone.
try:
stats["media_probed"] = _probe_media_batch(db, plex)
except Exception: # best-effort; never fail the sync over a backfill
@@ -294,15 +307,15 @@ def sync(db: Session) -> dict:
# whole library is covered over successive runs (each fetched item is stamped people_enriched_at once).
_ENRICH_BATCH = 60
# Same bounded-per-run pattern for the bit-depth probe. Only HW-decodable-codec transcode items are
# probed (typically a small minority — the bulk of transcodes are old codecs that are SW by codec), so
# probed (typically a small minority — the bulk of transcodes are old codecs classed by codec), so
# a modest batch covers the work-list in a run or two; a stuck item retries next run (watermark unset).
_PROBE_BATCH = 60
def _probe_media_batch(db: Session, plex: PlexClient, limit: int = _PROBE_BATCH) -> int:
"""Rich-fetch up to `limit` not-yet-probed HW-decodable transcode items, read the video Stream's
bit-depth + HDR, and store `video_bit_depth` + the refined `transcode_hw` (a 10-bit/HDR HEVC →
False/software). Stamps `media_probed_at` on every attempt (success OR a Plex miss with no Stream),
bit-depth + HDR, and store `video_bit_depth` + the refined `transcode_hw` (a 10-bit SDR HEVC →
False/software; an HDR one stays on the GPU through the tonemap). Stamps `media_probed_at` on every attempt (success OR a Plex miss with no Stream),
so an item is probed once and not re-fetched every sync; a transient fetch FAILURE leaves the
watermark unset to retry. Oldest-id first. Returns how many items were probed this run."""
rows = (
backend/app/reader/queries.py
+21 -11
View File
@@ -146,12 +146,17 @@ def item_counts(db: Session, ids: list[int]) -> dict[int, int]:
)
def durations(db: Session, ids: list[int]) -> dict[int, int]:
"""Each playlist's running time: the sum of its videos' lengths."""
def durations(db: Session, ids: list[int]) -> dict[int, tuple[int, int]]:
"""Each playlist's running time — the sum of its videos' known lengths — and how many videos
it left out for having none (`Video.without_length`)."""
return {
pid: int(total or 0)
for pid, total in db.execute(
select(PlaylistItem.playlist_id, func.coalesce(func.sum(Video.duration_seconds), 0))
pid: (int(total or 0), int(left_out or 0))
for pid, total, left_out in db.execute(
select(
PlaylistItem.playlist_id,
func.coalesce(func.sum(Video.length_seconds), 0),
func.count().filter(Video.without_length),
)
.join(Video, Video.id == PlaylistItem.video_id)
.where(PlaylistItem.playlist_id.in_(ids))
.group_by(PlaylistItem.playlist_id)
@@ -191,11 +196,16 @@ def covers(db: Session, ids: list[int]) -> dict[int, str | None]:
)
def playlist_totals(db: Session, playlist_id: int) -> tuple[int, int, str | None]:
"""One playlist's item count, running time and cover. An outer join, so an empty list still
answers (0, 0, None)."""
count, total = db.execute(
select(func.count(PlaylistItem.id), func.coalesce(func.sum(Video.duration_seconds), 0))
def playlist_totals(db: Session, playlist_id: int) -> tuple[int, int, int, str | None]:
"""One playlist's item count, running time, the videos the running time left out for having no
length (`Video.without_length`) and cover. An outer join, so an empty list still answers
(0, 0, 0, None)."""
count, total, left_out = db.execute(
select(
func.count(PlaylistItem.id),
func.coalesce(func.sum(Video.length_seconds), 0),
func.count().filter(Video.without_length),
)
.select_from(PlaylistItem)
.outerjoin(Video, Video.id == PlaylistItem.video_id)
.where(PlaylistItem.playlist_id == playlist_id)
@@ -207,7 +217,7 @@ def playlist_totals(db: Session, playlist_id: int) -> tuple[int, int, str | None
.order_by(PlaylistItem.position)
.limit(1)
)
return count or 0, total or 0, cover
return count or 0, total or 0, left_out or 0, cover
def playlist_rows(db: Session, user: User, playlist_id: int) -> list:
backend/app/routes/_common.py
+6 -2
View File
@@ -6,9 +6,9 @@ and "fetch a row this user owns or 404" idioms have one definition, one status c
one existence-leak policy.
"""
from typing import Any, Protocol, TypeVar
from typing import Annotated, Any, Protocol, TypeVar
from fastapi import HTTPException, Request
from fastapi import HTTPException, Query, Request
from pydantic import BaseModel, ConfigDict, Field
from sqlalchemy.orm import Session
@@ -21,6 +21,10 @@ from app.youtube.client import YouTubeError, refusal_reason
M = TypeVar("M", bound=Base)
# A shuffle / surprise seed off the query string. The client draws one under a billion; bounded so a
# hand-typed 400-digit one is a 422, not an int no float or bigint can hold further down.
Seed = Annotated[int, Query(ge=-(2**53), le=2**53)]
# Resume-position thresholds of a player that only keeps a resume point (the YouTube player, the
# library's): positions below this are "didn't really start" and within this of the end are
# "basically finished" — neither is worth storing, so the position is cleared instead (no progress
backend/app/routes/discover.py
+3 -3
View File
@@ -23,7 +23,7 @@ from app.discover import pins, rails
from app.discover.budget import Budget, DEFAULT_BUDGET
from app.models import User
from app.plex.cards import norm_lang
from app.routes._common import not_found
from app.routes._common import Seed, not_found
router = APIRouter(prefix="/api/discover", tags=["discover"])
@@ -44,7 +44,7 @@ class ReorderIn(BaseModel):
@router.get("")
def get_discover(
budget: Budget = DEFAULT_BUDGET,
seed: int = 0,
seed: Seed = 0,
lang: str = "en",
user: User = Depends(current_user),
db: Session = Depends(get_db),
@@ -79,7 +79,7 @@ MAX_EXCLUDE = 32
@router.get("/pick")
def get_pick(
budget: Budget = DEFAULT_BUDGET,
seed: int = 0,
seed: Seed = 0,
exclude: str = "",
start: int = 0,
lang: str = "en",
backend/app/routes/feed.py
+6 -4
View File
@@ -10,6 +10,7 @@ from sqlalchemy.orm.exc import StaleDataError
from app import jobruns, quota, scheduler
from app.catalog import queries as catalog_queries, service as catalog_service
from app.catalog.models import known_length
from app.auth import current_user
from app.db import get_db, release_read_connection
from app.mailtext import lang_for_user
@@ -19,6 +20,7 @@ from app.models import User
from app.reader import queries, service, takeout
from app.routes._common import (
ProgressIn,
Seed,
capped_body,
drain_body,
kept_position,
@@ -88,7 +90,7 @@ def _feed_params(
def get_feed(
params: dict = Depends(_feed_params),
sort: str = "newest",
seed: int = 0,
seed: Seed = 0,
limit: int = Query(default=60, le=200),
cursor: str | None = None,
user: User = Depends(current_user),
@@ -287,7 +289,7 @@ def get_video_detail(
"channel_title": v.channel.title if v.channel else None,
"published_at": iso(v.published_at),
"view_count": v.view_count,
"duration_seconds": v.duration_seconds,
"duration_seconds": v.length_seconds,
}
# The DB path above is free; the off-catalog YouTube lookup below spends a quota unit.
@@ -315,8 +317,8 @@ def get_video_detail(
"channel_title": snippet.get("channelTitle"),
"published_at": snippet.get("publishedAt"),
"view_count": int(views) if views is not None else None,
"duration_seconds": parse_iso8601_duration(
items[0].get("contentDetails", {}).get("duration")
"duration_seconds": known_length(
parse_iso8601_duration(items[0].get("contentDetails", {}).get("duration"))
),
}
backend/app/routes/playlists.py
+29 -9
View File
@@ -41,11 +41,16 @@ router = APIRouter(prefix="/api/playlists", tags=["playlists"])
# It guards EVERY route that spends under QuotaAction.PLAYLISTS_PULL, not just the bulk one: a
# scripted loop over the single-playlist re-pull drains the same shared budget, and limiting one of
# two routes with the same quota action is the asymmetric fix this codebase keeps re-learning.
# Deliberately NOT on the push side (`push-plan` / `push`): those run on every push attempt,
# including the no-op "already up to date" path, so a 3-per-10-minute allowance would refuse
# legitimate work. Their exposure is tracked as SEC-18 instead.
_playlist_pull_limiter = RateLimiter(max_events=3, window_seconds=600)
# SEC-18, the push side: `push-plan` reads the live list (1 unit a page) and `push` reads and writes
# it, both against the same instance-wide budget. They get a budget of their own, counted in calls
# and far wider than the pull's: a press in the UI is one plan and, unless nothing changed, one push,
# so no person reaches it, while a scripted loop stops at it. A big push is not this limit's job —
# `push` already refuses one the day's remaining budget cannot cover.
PUSH_ALLOWANCE = 20
_playlist_push_limiter = RateLimiter(max_events=PUSH_ALLOWANCE, window_seconds=600)
def _check_pull_cooldown(user: User) -> None:
"""Shared gate so the two PLAYLISTS_PULL routes cannot drift apart again."""
@@ -58,6 +63,7 @@ def _check_pull_cooldown(user: User) -> None:
PlaylistRefusal = Literal[
"name_required",
"not_on_youtube",
"push_cooldown",
"quota_exceeded",
"sync_all_busy",
"sync_busy",
@@ -89,10 +95,13 @@ def _video_or_404(db: Session, video_id: str) -> None:
def _summary(
pl: "Playlist",
count: int,
total_duration: int = 0,
duration: tuple[int, int] = (0, 0),
has_video: bool | None = None,
cover: str | None = None,
) -> dict:
"""`duration` is `queries.durations`'s pair: the known running time, and the videos it left out
for having no length — the "+N without length" beside it."""
total_duration, left_out = duration
out = {
"id": pl.id,
"name": pl.name,
@@ -102,6 +111,7 @@ def _summary(
"dirty": pl.dirty,
"item_count": count,
"total_duration_seconds": total_duration,
"without_length": left_out,
"cover_thumbnail": cover,
# When this list last CHANGED (rename, add/remove/move/reorder) — the rail's
# "recently changed first" sort reads it. See `service.touch` for why it isn't automatic.
@@ -132,7 +142,7 @@ def list_playlists(
_summary(
p,
counts.get(p.id, 0),
durations.get(p.id, 0),
durations.get(p.id, (0, 0)),
(p.id in member) if contains else None,
covers.get(p.id),
)
@@ -244,9 +254,16 @@ def revert_youtube(
raise youtube_refusal(exc, playlist=True) from exc
def _check_push_cooldown(user: User) -> None:
"""Shared gate of the two push routes, so they spend one allowance and cannot drift apart."""
if not _playlist_push_limiter.allow(str(user.id)):
raise _refused(429, "push_cooldown")
def _pushable(db: Session, user: User, playlist_id: int) -> "Playlist":
"""A local playlist the user may push to YouTube (create or update). The built-in
Watch later and read-only YouTube mirrors are never pushable."""
"""A playlist the user may push to YouTube: a local one (created there on its first push) or a
mirror of their own YouTube list (its local edits merged back). Only the built-in Watch later is
never pushable."""
pl = _own_playlist(db, user, playlist_id)
if pl.kind == "watch_later":
raise _refused(400, "watch_later_not_pushable")
@@ -263,6 +280,7 @@ def push_plan(
the quota estimate, so the UI can show a confirm. For a linked playlist this reads the live
YouTube state (cheap, 1 unit/page). Write scope is required — syncing is a YouTube write."""
pl = _pushable(db, user, playlist_id)
_check_push_cooldown(user) # after the refusal: a list that cannot be pushed spends nothing
try:
with quota.attribute(user.id, quota.QuotaAction.PLAYLISTS_PUSH):
plan = plan_merge(db, user, pl)
@@ -288,6 +306,8 @@ def push(
with sync_claim(db, pl.id) as free:
if not free:
raise _refused(409, "sync_busy")
# Spent only by a push that runs, as Revert's pull: a busy list or Watch later costs nothing.
_check_push_cooldown(user)
try:
with quota.attribute(user.id, quota.QuotaAction.PLAYLISTS_PUSH):
# Read the live playlist ONCE (linked playlists only) and share it with both the
@@ -349,8 +369,8 @@ def rename_playlist(
raise _refused(400, "name_required")
service.rename_playlist(db, pl, name)
db.commit()
count, total, cover = queries.playlist_totals(db, pl.id)
return _summary(pl, count, total, cover=cover)
count, total, left_out, cover = queries.playlist_totals(db, pl.id)
return _summary(pl, count, (total, left_out), cover=cover)
@router.delete("/{playlist_id}")
backend/tests/test_abuse_limits.py
+106
View File
@@ -125,3 +125,109 @@ def test_the_single_playlist_repull_shares_the_pull_cooldown(db, as_user, make_u
r = c.post("/api/playlists/sync-youtube")
assert r.status_code == 429
assert r.json()["detail"] == {"code": "sync_cooldown"}
def _pushable_list(db, make_user, monkeypatch, plroutes):
"""A local (never exported) list and a user with the write scope; the merge itself is stubbed —
the limiter is under test, not the YouTube round trip."""
from app import auth
from app.reader.models import Playlist
user = make_user()
pl = Playlist(user_id=user.id, name="Local", source="local")
db.add(pl)
db.commit()
monkeypatch.setattr(auth, "has_write_scope", lambda u: True)
monkeypatch.setattr(plroutes, "plan_merge", lambda d, u, p, live=None: {"units_estimate": 0})
monkeypatch.setattr(plroutes, "merge_playlist", lambda d, u, p, live=None: {"added": 0})
return user, pl
def test_the_push_side_goes_on_cooldown_after_its_allowance(db, as_user, make_user, monkeypatch):
# SEC-18: push-plan reads the live list and push writes it, both against the instance-wide
# YouTube budget. A press in the UI is one plan + one push, so the allowance is generous — but a
# scripted loop stops at it, with a reason the UI can say.
from app.routes import playlists as plroutes
user, pl = _pushable_list(db, make_user, monkeypatch, plroutes)
c = as_user(user)
allowance = plroutes.PUSH_ALLOWANCE
codes = [c.get(f"/api/playlists/{pl.id}/push-plan").status_code for _ in range(allowance)]
assert codes == [200] * allowance
r = c.get(f"/api/playlists/{pl.id}/push-plan")
assert r.status_code == 429
assert r.json()["detail"] == {"code": "playlist", "reason": "push_cooldown"}
def test_plan_and_push_draw_on_one_allowance(db, as_user, make_user, monkeypatch):
# The asymmetric fix again, avoided up front: a limit on one of the two routes leaves the same
# drain open next door.
from app.routes import playlists as plroutes
user, pl = _pushable_list(db, make_user, monkeypatch, plroutes)
c = as_user(user)
for i in range(plroutes.PUSH_ALLOWANCE):
if i % 2:
r = c.get(f"/api/playlists/{pl.id}/push-plan")
else:
r = c.post(f"/api/playlists/{pl.id}/push")
assert r.status_code == 200, (i, r.json())
r = c.post(f"/api/playlists/{pl.id}/push")
assert r.status_code == 429
assert r.json()["detail"] == {"code": "playlist", "reason": "push_cooldown"}
def test_a_refused_push_spends_no_allowance(db, as_user, make_user, monkeypatch):
# The allowance counts pushes that run: Watch later (never pushable) and a list another sync
# holds are refused before any YouTube call, so they must not leave the next real press at 429.
from contextlib import contextmanager
from app.reader.models import Playlist
from app.routes import playlists as plroutes
user, pl = _pushable_list(db, make_user, monkeypatch, plroutes)
wl = Playlist(user_id=user.id, name="Watch later", kind="watch_later")
db.add(wl)
db.commit()
c = as_user(user)
over = plroutes.PUSH_ALLOWANCE + 1
assert {c.get(f"/api/playlists/{wl.id}/push-plan").status_code for _ in range(over)} == {400}
assert {c.post(f"/api/playlists/{wl.id}/push").status_code for _ in range(over)} == {400}
@contextmanager
def busy(_db, _id):
yield False
with monkeypatch.context() as m:
m.setattr(plroutes, "sync_claim", busy)
assert {c.post(f"/api/playlists/{pl.id}/push").status_code for _ in range(over)} == {409}
assert c.post(f"/api/playlists/{pl.id}/push").status_code == 200
def test_a_mirrored_youtube_list_is_pushable(db, as_user, make_user, monkeypatch):
# A mirror of the user's own YouTube list takes local edits (`dirty`) and syncs them back; only
# Watch later is refused.
from app.reader.models import Playlist
from app.routes import playlists as plroutes
user, _ = _pushable_list(db, make_user, monkeypatch, plroutes)
mirror = Playlist(
user_id=user.id, name="Mine", source="youtube", yt_playlist_id="PLx", dirty=True
)
db.add(mirror)
db.commit()
assert as_user(user).get(f"/api/playlists/{mirror.id}/push-plan").status_code == 200
def test_the_push_cooldown_is_per_user(db, as_user, make_user, monkeypatch):
from app.routes import playlists as plroutes
a, pa = _pushable_list(db, make_user, monkeypatch, plroutes)
b, pb = _pushable_list(db, make_user, monkeypatch, plroutes)
ca = as_user(a)
codes = [
ca.get(f"/api/playlists/{pa.id}/push-plan").status_code
for _ in range(plroutes.PUSH_ALLOWANCE + 1)
]
assert codes == [200] * plroutes.PUSH_ALLOWANCE + [429]
assert as_user(b).get(f"/api/playlists/{pb.id}/push-plan").status_code == 200
backend/tests/test_clip.py
+3
View File
@@ -531,6 +531,9 @@ def test_a_youtube_clip_hands_the_range_to_yt_dlp(monkeypatch, tmp_path):
def __enter__(self):
return self
def add_post_processor(self, pp, when="post_process"):
pass # the egress format filter; nothing here to filter
def __exit__(self, *_a):
return False
backend/tests/test_egress.py
+412
View File
@@ -0,0 +1,412 @@
"""SEC-17: the egress gate judges the address a guarded download actually dials.
Two loopback addresses stand in for the two sides: 127.0.0.2 plays a public host, 127.0.0.1 our own
network. The gate under test refuses exactly 127.0.0.1, so every assertion is about the judgement and
the wiring, with real sockets, a real yt-dlp and a real ffmpeg — no mocked transport.
"""
import contextlib
import http.server
import ipaddress
import socket
import socketserver
import struct
import threading
import time
from pathlib import Path
import pytest
import yt_dlp
from yt_dlp.utils import download_range_func
from app.downloads import adapters, egress, errors, formats
PUBLIC = "127.0.0.2"
INTERNAL = "127.0.0.1"
def _forbid_internal(ip: ipaddress.IPv4Address | ipaddress.IPv6Address) -> bool:
return str(ip) == INTERNAL
@pytest.fixture
def gate():
g = egress.start_gate(forbidden=_forbid_internal)
yield g
g.shutdown()
g.server_close()
class _Listener:
"""A TCP listener that only counts who connected — the internal service under attack."""
def __init__(self, host: str = INTERNAL, port: int = 0):
self.hits = 0
self.sock = socket.socket()
self.sock.setsockopt(socket.SOL_SOCKET, socket.SO_REUSEADDR, 1)
self.sock.bind((host, port))
self.sock.listen()
self.port = self.sock.getsockname()[1]
threading.Thread(target=self._serve, daemon=True).start()
def _serve(self):
while True:
try:
conn, _ = self.sock.accept()
except OSError:
return
self.hits += 1
with contextlib.suppress(OSError):
conn.sendall(b"HTTP/1.1 200 OK\r\nContent-Length: 2\r\n\r\nok")
conn.close()
def close(self):
self.sock.close()
def _public_server(routes: dict, port: int = 0, directory: Path | None = None):
"""An HTTP server on the 'public' address: `routes` maps a path to (status, headers); anything
else is served from `directory`."""
class H(http.server.SimpleHTTPRequestHandler):
def __init__(self, *a, **k):
super().__init__(*a, directory=str(directory or "."), **k)
def log_message(self, *a):
pass
def do_GET(self):
if self.path in routes:
status, headers, body = routes[self.path]
self.send_response(status)
for k, v in headers.items():
self.send_header(k, v)
self.send_header("Content-Length", str(len(body)))
self.end_headers()
self.wfile.write(body)
return
super().do_GET()
srv = socketserver.ThreadingTCPServer((PUBLIC, port), H)
srv.daemon_threads = True
threading.Thread(target=srv.serve_forever, daemon=True).start()
return srv
def _guarded_opts(tmp_path: Path, gate_url: str, **extra) -> dict:
opts = formats.build_ydl_opts({}, str(tmp_path / "%(id)s.%(ext)s"), lambda d: None, egress=True)
opts.update(retries=0, fragment_retries=0, extractor_retries=0, socket_timeout=5)
opts["proxy"] = gate_url # the test's gate (it lets 127.0.0.2 through), not the worker's
opts.update(extra)
return opts
def _download_error(opts: dict, url: str) -> str:
with pytest.raises(yt_dlp.utils.DownloadError) as exc, egress.open_ydl(opts) as ydl:
ydl.extract_info(url, download=True)
return str(exc.value)
# --- the judgement ---------------------------------------------------------------------------------
def test_a_name_with_any_forbidden_answer_is_refused():
def resolve(answers):
return lambda host, port, *a: [(2, 1, 6, "", (ip, port)) for ip in answers]
assert egress.vetted_address(
"x", 80, resolve=resolve([PUBLIC]), forbidden=_forbid_internal
) == (
PUBLIC,
80,
)
assert (
egress.vetted_address(
"x", 80, resolve=resolve([PUBLIC, INTERNAL]), forbidden=_forbid_internal
)
is None
)
def no_answer(*a):
raise socket.gaierror("no such name")
assert egress.vetted_address("x", 80, resolve=no_answer, forbidden=_forbid_internal) is None
def test_the_real_judgement_refuses_our_own_networks():
# The worker's gate uses urlguard's judgement unchanged — the same one the enqueue check uses.
for addr in ("127.0.0.1", "10.1.2.3", "172.18.0.5", "192.168.1.1", "169.254.169.254"):
assert egress.vetted_address(addr, 80, forbidden=egress.urlguard.is_forbidden) is None, addr
assert egress.vetted_address("93.184.215.14", 80, forbidden=egress.urlguard.is_forbidden)
def test_the_gate_dials_the_address_it_vetted_not_a_second_answer(gate):
# DNS rebinding: a name that answers public for the check and private for the fetch. The gate
# resolves once per connection and dials that answer.
public = _public_server({"/": (200, {}, b"public")})
internal = _Listener(INTERNAL, public.server_address[1])
calls = []
def rebinding(host, port, *a):
calls.append(host)
ip = PUBLIC if len(calls) == 1 else INTERNAL
return [(2, 1, 6, "", (ip, port))]
gate.resolve = rebinding
try:
with socket.create_connection(("127.0.0.1", gate.server_address[1])) as s:
s.sendall(
f"GET http://rebind.test:{public.server_address[1]}/ HTTP/1.1\r\n"
f"Host: rebind.test\r\n\r\n".encode()
)
answer = b"".join(iter(lambda: s.recv(4096), b""))
assert answer.endswith(b"public")
assert calls == ["rebind.test"]
assert internal.hits == 0
finally:
public.shutdown()
internal.close()
def test_a_tunnel_into_the_network_is_refused_before_it_is_dialled(gate):
internal = _Listener()
try:
with socket.create_connection(("127.0.0.1", gate.server_address[1])) as s:
s.sendall(f"CONNECT {INTERNAL}:{internal.port} HTTP/1.1\r\n\r\n".encode())
answer = s.recv(4096)
assert answer.startswith(b"HTTP/1.1 403")
assert internal.hits == 0
finally:
internal.close()
def test_a_client_that_hangs_up_mid_transfer_is_not_an_error(gate, monkeypatch):
# yt-dlp closes a connection once it has read what it needed; seen live as a traceback the gate
# printed for every such download.
errors_seen = []
monkeypatch.setattr(gate, "handle_error", lambda *a: errors_seen.append(a))
public = _public_server({"/big": (200, {}, b"x" * 4_000_000)})
try:
s = socket.create_connection(("127.0.0.1", gate.server_address[1]))
s.sendall(f"GET http://{PUBLIC}:{public.server_address[1]}/big HTTP/1.1\r\n\r\n".encode())
s.recv(1024)
s.setsockopt(socket.SOL_SOCKET, socket.SO_LINGER, struct.pack("ii", 1, 0))
s.close() # a reset, not a polite close
time.sleep(0.5)
assert errors_seen == []
finally:
public.shutdown()
# --- a real yt-dlp through it ----------------------------------------------------------------------
def test_a_redirect_into_the_network_reaches_it_unguarded_and_not_through_the_gate(tmp_path, gate):
internal = _Listener()
public = _public_server(
{"/v.mp4": (302, {"Location": f"http://{INTERNAL}:{internal.port}/v.mp4"}, b"")}
)
url = f"http://{PUBLIC}:{public.server_address[1]}/v.mp4"
try:
# The hole, as it was: yt-dlp follows the 3xx into our network.
bare = {"outtmpl": str(tmp_path / "bare" / "%(id)s.%(ext)s"), "quiet": True, "retries": 0}
with contextlib.suppress(yt_dlp.utils.DownloadError), yt_dlp.YoutubeDL(bare) as ydl:
ydl.extract_info(url, download=True)
assert internal.hits >= 1
before = internal.hits
_download_error(_guarded_opts(tmp_path, gate.url), url)
assert internal.hits == before
finally:
public.shutdown()
internal.close()
def test_a_live_and_a_dead_internal_port_leave_the_same_text(tmp_path, gate):
live = _Listener()
dead = _Listener()
dead_port = dead.port
dead.close()
public = _public_server(
{
"/live.mp4": (302, {"Location": f"http://{INTERNAL}:{live.port}/x"}, b""),
"/dead.mp4": (302, {"Location": f"http://{INTERNAL}:{dead_port}/x"}, b""),
}
)
base = f"http://{PUBLIC}:{public.server_address[1]}"
try:
stored = []
for name in ("live", "dead"):
msg = _download_error(_guarded_opts(tmp_path / name, gate.url), f"{base}/{name}.mp4")
stored.append(errors.user_error_text("url", msg, errors.classify(msg)))
assert stored[0] == stored[1]
assert live.hits == 0
finally:
public.shutdown()
live.close()
def test_a_format_fetched_past_the_gate_is_never_chosen(tmp_path):
# ffmpeg speaks rtmp, rtsp, tcp… directly, where no HTTP proxy is asked. A guarded download keeps
# only the formats fetched over http(s).
opts = _guarded_opts(tmp_path, "http://127.0.0.1:9")
info = {
"id": "x",
"title": "x",
"extractor": "generic",
"extractor_key": "Generic",
"webpage_url": f"http://{PUBLIC}/page",
"formats": [
{"format_id": "rtmp", "url": f"rtmp://{INTERNAL}/live", "ext": "flv", "height": 1080},
{"format_id": "web", "url": f"http://{PUBLIC}/a.mp4", "ext": "mp4", "height": 360},
],
}
with egress.open_ydl(opts) as ydl:
chosen = ydl.process_ie_result(dict(info), download=False)
assert chosen["format_id"] == "web"
only_rtmp = dict(info, formats=[info["formats"][0]])
with (
pytest.raises(yt_dlp.utils.YoutubeDLError, match="Requested format is not available"),
egress.open_ydl(opts) as ydl,
):
ydl.process_ie_result(only_rtmp, download=False)
def test_a_format_yt_dlp_fetches_itself_is_kept():
# f4m, ism and the DASH generator fetch through yt-dlp's own urlopen; a bare m3u8 goes to the
# native HLS fetcher under hls_prefer_native (set for every guarded download). All ask the gate.
kept = [
{"format_id": p, "url": f"https://{PUBLIC}/{p}", "protocol": p}
for p in ("f4m", "ism", "m3u8", "m3u8_native", "http_dash_segments_generator", "https")
]
raw = {"format_id": "rtmp", "url": f"rtmp://{PUBLIC}/live", "protocol": "rtmp"}
_, info = egress.HttpOnlyFormats().run({"formats": [*kept, raw]})
assert [f["format_id"] for f in info["formats"]] == [f["format_id"] for f in kept]
def test_a_single_url_result_past_the_gate_loses_its_url():
_, info = egress.HttpOnlyFormats().run({"url": f"rtmp://{INTERNAL}/live"})
assert info["url"] is None
_, info = egress.HttpOnlyFormats().run({"url": f"https://{PUBLIC}/a.mp4"})
assert info["url"] == f"https://{PUBLIC}/a.mp4"
def test_the_format_filter_says_no_phase_of_its_own(tmp_path):
# It is a post-processor only to run before the format choice; the job's postprocessor hook
# must not see it (it painted every guarded job "processing" before a byte was fetched).
calls: list[dict] = []
opts = _guarded_opts(tmp_path, "http://127.0.0.1:9", postprocessor_hooks=[calls.append])
with egress.open_ydl(opts) as ydl:
ydl.run_all_pps("pre_process", {"id": "x", "formats": []})
assert calls == []
def test_an_allowed_tunnel_carries_bytes_both_ways(gate):
# The CONNECT path a site served over https takes: a 200, then the bytes the client sent after
# the head (in the same packet here) reach the far end, and its answer comes back.
far = socket.socket()
far.bind((PUBLIC, 0))
far.listen()
def echo():
conn, _ = far.accept()
with conn:
conn.sendall(b"PONG:" + conn.recv(64))
threading.Thread(target=echo, daemon=True).start()
try:
with socket.create_connection(("127.0.0.1", gate.server_address[1]), timeout=5) as s:
s.sendall(f"CONNECT {PUBLIC}:{far.getsockname()[1]} HTTP/1.1\r\n\r\nPING".encode())
got = b""
while b"PONG:PING" not in got:
chunk = s.recv(4096)
if not chunk:
break
got += chunk
assert got.startswith(b"HTTP/1.1 200")
assert got.endswith(b"PONG:PING")
finally:
far.close()
def test_the_workers_own_gate_refuses_our_network():
# Every other gate here is built with a test rule; this is the one `proxy_url()` starts.
internal = _Listener()
try:
port = int(egress.proxy_url().rsplit(":", 1)[1])
with socket.create_connection(("127.0.0.1", port), timeout=5) as s:
s.sendall(f"CONNECT {INTERNAL}:{internal.port} HTTP/1.1\r\n\r\n".encode())
assert s.recv(4096).startswith(b"HTTP/1.1 403")
assert internal.hits == 0
finally:
internal.close()
def _clip_hits(tmp_path: Path, segment: str, proxy: str) -> int:
"""How many connections the internal listener takes while a clip (`download_ranges`, which
yt-dlp hands to ffmpeg even for an HLS format) is cut from a playlist whose one segment is
`segment` — formatted with the listener's address."""
internal = _Listener()
playlist = (
"#EXTM3U\n#EXT-X-VERSION:3\n#EXT-X-TARGETDURATION:2\n#EXTINF:2.0,\n"
+ segment.format(host=INTERNAL, port=internal.port)
+ "\n#EXT-X-ENDLIST\n"
).encode()
public = _public_server(
{"/p.m3u8": (200, {"Content-Type": "application/vnd.apple.mpegurl"}, playlist)}
)
try:
opts = _guarded_opts(tmp_path, proxy, download_ranges=download_range_func([], [(0, 1)]))
with contextlib.suppress(yt_dlp.utils.DownloadError), egress.open_ydl(opts) as ydl:
ydl.extract_info(f"http://{PUBLIC}:{public.server_address[1]}/p.m3u8", download=True)
return internal.hits
finally:
public.shutdown()
internal.close()
def test_a_clip_cut_by_ffmpeg_reaches_a_segment_only_through_the_gate(tmp_path, gate):
# Without the gate, ffmpeg fetches a hostile playlist's segment from inside our network — the
# control that shows this path is real. Through the gate it is never dialled.
segment = "http://{host}:{port}/s.ts"
assert _clip_hits(tmp_path / "bare", segment, proxy="") >= 1
assert _clip_hits(tmp_path / "gated", segment, proxy=gate.url) == 0
def test_ffmpeg_itself_refuses_a_raw_tcp_segment(tmp_path):
# A boundary, not our code: the HTTP proxy cannot see a tcp:// segment, and measured, ffmpeg's
# HLS reader does not open one from an http playlist. This fails if an ffmpeg upgrade starts to.
assert _clip_hits(tmp_path, "tcp://{host}:{port}/", proxy="") == 0
# --- the wiring ------------------------------------------------------------------------------------
def test_only_youtube_downloads_skip_the_gate():
assert adapters.SourceAdapter.egress_guarded is True
assert adapters.resolve_adapter("url", "https://www.facebook.com/x").egress_guarded is True
assert adapters.resolve_adapter("url", "https://vimeo.com/1").egress_guarded is True
assert adapters.resolve_adapter("youtube", "dQw4w9WgXcQ").egress_guarded is False
def test_a_guarded_download_goes_through_the_gate_whatever_its_adapter_adds():
opts = formats.build_ydl_opts(
{}, "/tmp/x", lambda d: None, extra={"proxy": "", "hls_prefer_native": False}, egress=True
)
assert opts["proxy"] == egress.proxy_url()
assert opts["hls_prefer_native"] is True
assert opts[egress.GUARDED] is True
assert "proxy" not in formats.build_ydl_opts({}, "/tmp/x", lambda d: None, egress=False)
def test_every_yt_dlp_this_app_opens_goes_through_open_ydl():
app = Path(egress.__file__).resolve().parents[1]
offenders = [
f"{p.relative_to(app)}:{n}"
for p in sorted(app.rglob("*.py"))
if p.name != "egress.py"
for n, line in enumerate(p.read_text(encoding="utf-8").splitlines(), 1)
if "YoutubeDL(" in line
]
assert offenders == []
backend/tests/test_facebook_impersonation.py
+14 -1
View File
@@ -18,7 +18,7 @@ import yt_dlp
from yt_dlp.networking.exceptions import TransportError
from yt_dlp.networking.impersonate import ImpersonateTarget
from app.downloads import adapters, pipeline
from app.downloads import adapters, egress, pipeline
from app.downloads import errors as dlerrors
_URL = "https://www.facebook.com/reel/806379462330494/?fs=e&fs=e"
@@ -73,6 +73,9 @@ def ydl(monkeypatch):
def __enter__(self):
return self
def add_post_processor(self, pp, when="post_process"):
pass # the egress format filter; nothing here to filter
def __exit__(self, *exc):
return False
@@ -157,6 +160,16 @@ def test_a_cookie_group_post_downloads_the_posts_own_video(ydl, tmp_path):
# Never the post link: that is where the feed video comes from.
assert web["extracted"] == [watch, watch]
assert [o.get("impersonate") for o in seen[1:]] == [None, _CHROME]
# The post page read and both attempts all go through the egress gate.
assert all(o.get(egress.GUARDED) and o["proxy"] == egress.proxy_url() for o in seen), seen
def test_a_youtube_download_skips_the_gate(ydl, tmp_path):
seen, outcomes, _ = ydl
outcomes += [{"id": "dQw4w9WgXcQ"}]
yt = adapters.resolve_adapter("youtube", "dQw4w9WgXcQ")
pipeline._extract_with_fallback(1, 1, "https://youtu.be/dQw4w9WgXcQ", {}, tmp_path, yt)
assert seen and not any(o.get(egress.GUARDED) or o.get("proxy") for o in seen)
def test_a_group_post_without_a_single_video_fails_saying_so(ydl, tmp_path):
backend/tests/test_facebook_titles.py
+6 -2
View File
@@ -10,7 +10,7 @@ from sqlalchemy import select, update
from app.config import settings
from app.db import SessionLocal
from app.downloads import adapters, backfill, pipeline
from app.downloads import adapters, backfill, egress, pipeline
from app.downloads.models import DownloadJob, MediaAsset
_URL = "https://www.facebook.com/reel/1860682468422402/"
@@ -36,6 +36,9 @@ def _fake_ydl(monkeypatch, outcome):
def __enter__(self):
return self
def add_post_processor(self, pp, when="post_process"):
pass # the egress format filter; nothing here to filter
def __exit__(self, *exc):
return False
@@ -43,7 +46,7 @@ def _fake_ydl(monkeypatch, outcome):
seen.append({**self.opts, "url": url})
return outcome
monkeypatch.setattr(backfill.yt_dlp, "YoutubeDL", FakeYDL)
monkeypatch.setattr("yt_dlp.YoutubeDL", FakeYDL)
return seen
@@ -58,6 +61,7 @@ def test_a_placeholder_with_nothing_behind_it_takes_the_signed_out_pages_text(mo
assert len(seen) == 1
assert seen[0]["skip_download"] is True
assert not seen[0].get("cookiefile")
assert seen[0][egress.GUARDED] is True and seen[0]["proxy"] == egress.proxy_url()
def test_a_real_title_reads_nothing_more(monkeypatch):
backend/tests/test_failure_lines.py
+12 -2
View File
@@ -359,6 +359,9 @@ def test_the_metadata_backfill_names_the_asset_and_not_the_download(sink, monkey
def __enter__(self):
return self
def add_post_processor(self, pp, when="post_process"):
pass # the egress format filter; nothing here to filter
def __exit__(self, *exc):
return False
@@ -367,7 +370,7 @@ def test_the_metadata_backfill_names_the_asset_and_not_the_download(sink, monkey
f"ERROR: [youtube] dQw4w9WgXcQ: Video unavailable ({target})"
)
monkeypatch.setattr(backfill.yt_dlp, "YoutubeDL", _Refusing)
monkeypatch.setattr("yt_dlp.YoutubeDL", _Refusing)
assert backfill.extract_meta(url, adapters.YouTubeAdapter(), 11) is None
lines = [o for o in sink.objects() if "meta extract failed" in o["msg"]]
@@ -386,13 +389,16 @@ def test_the_metadata_backfill_takes_out_a_generic_name_whatever_it_holds(sink,
def __enter__(self):
return self
def add_post_processor(self, pp, when="post_process"):
pass # the egress format filter; nothing here to filter
def __exit__(self, *exc):
return False
def extract_info(self, target, download):
raise yt_dlp.utils.DownloadError("ERROR: [generic] Our Trip: Day 1: Unable to download")
monkeypatch.setattr(backfill.yt_dlp, "YoutubeDL", _Refusing)
monkeypatch.setattr("yt_dlp.YoutubeDL", _Refusing)
assert backfill.extract_meta(url, adapters.SourceAdapter(), 11) is None
lines = [o for o in sink.objects() if "meta extract failed" in o["msg"]]
@@ -449,6 +455,9 @@ def test_a_retried_attempt_names_the_host_and_not_the_video(sink, monkeypatch, t
def __enter__(self):
return self
def add_post_processor(self, pp, when="post_process"):
pass # the egress format filter; nothing here to filter
def __exit__(self, *exc):
return False
@@ -463,6 +472,7 @@ def test_a_retried_attempt_names_the_host_and_not_the_video(sink, monkeypatch, t
resolve_url=lambda target, opts: target,
ydl_opts_extra=lambda clients: {},
retriable=lambda text: True,
egress_guarded=False,
)
monkeypatch.setattr(pipeline.yt_dlp, "YoutubeDL", _YDL)
monkeypatch.setattr(pipeline, "_touch_job", lambda job_id: None)
backend/tests/test_formats.py
+16 -7
View File
@@ -15,16 +15,20 @@ from app.downloads import formats
def _opts(spec):
return formats.build_ydl_opts(spec, "/tmp/%(id)s.%(ext)s", lambda d: None)
return formats.build_ydl_opts(spec, "/tmp/%(id)s.%(ext)s", lambda d: None, egress=False)
def test_adapter_extra_opts_are_merged_in():
# The SourceAdapter contributes source-specific opts (e.g. YouTube's extractor_args) via `extra`.
extra = {"extractor_args": {"youtube": {"player_client": ["web"]}}}
opts = formats.build_ydl_opts({}, "/tmp/%(id)s.%(ext)s", lambda d: None, extra=extra)
opts = formats.build_ydl_opts(
{}, "/tmp/%(id)s.%(ext)s", lambda d: None, extra=extra, egress=False
)
assert opts["extractor_args"]["youtube"]["player_client"] == ["web"]
# No extra → no extractor_args (a generic source contributes nothing).
assert "extractor_args" not in formats.build_ydl_opts({}, "/tmp/x", lambda d: None)
assert "extractor_args" not in formats.build_ydl_opts(
{}, "/tmp/x", lambda d: None, egress=False
)
def test_mp4_default_selects_avc1_with_progressive_fallback():
@@ -113,7 +117,7 @@ def test_a_playlist_url_can_only_yield_one_item():
def test_the_remaining_quota_becomes_a_hard_size_ceiling():
opts = formats.build_ydl_opts({}, "/tmp/x", lambda d: None, max_bytes=5_000_000)
opts = formats.build_ydl_opts({}, "/tmp/x", lambda d: None, max_bytes=5_000_000, egress=False)
assert opts["max_filesize"] == 5_000_000
@@ -122,7 +126,7 @@ def test_no_budget_means_no_size_filter():
# enqueue, not an impossible yt-dlp filter to express here.
for budget in (None, 0, -1):
assert "max_filesize" not in formats.build_ydl_opts(
{}, "/tmp/x", lambda d: None, max_bytes=budget
{}, "/tmp/x", lambda d: None, max_bytes=budget, egress=False
)
@@ -130,7 +134,12 @@ def test_a_source_adapter_cannot_raise_the_size_ceiling():
# The adapter merge happens after max_filesize is set, so this pins that no adapter option can
# widen a storage cap — the mirror of the cookiefile rule, in the opposite direction.
opts = formats.build_ydl_opts(
{}, "/tmp/x", lambda d: None, extra={"max_filesize": 999_999_999}, max_bytes=5_000_000
{},
"/tmp/x",
lambda d: None,
extra={"max_filesize": 999_999_999},
max_bytes=5_000_000,
egress=False,
)
assert opts["max_filesize"] == 5_000_000
@@ -197,7 +206,7 @@ def test_yt_dlp_refuses_a_file_over_the_cap_without_raising_and_the_watch_sees_i
for cap, refused in ((1_000, True), (10_000_000, False)):
out = tmp_path / f"out-{cap}"
opts = formats.build_ydl_opts(
{}, str(out / "%(id)s.%(ext)s"), lambda d: None, max_bytes=cap
{}, str(out / "%(id)s.%(ext)s"), lambda d: None, max_bytes=cap, egress=False
)
with formats.watching_size_cap() as over, yt_dlp.YoutubeDL(opts) as ydl:
ydl.extract_info(url, download=True)
backend/tests/test_migration_0110.py
+3
View File
@@ -45,6 +45,9 @@ def test_an_interval_override_moves_into_sysconfig_and_back(db):
}
assert conn.execute(text("SELECT to_regclass('scheduler_settings')")).scalar() is None
assert scheduler.load_intervals(db)["rss_poll"] == 25
# The way down rebuilds an index CONCURRENTLY (0117), which waits for every open transaction —
# the read above left this session's open.
db.rollback()
command.downgrade(cfg, "0109_job_runs")
try:
backend/tests/test_migration_0119.py
+70
View File
@@ -0,0 +1,70 @@
"""Migration 0119's data step, DB-lane: the stored `transcode_hw` facet follows the rule that it says
whether the GPU does the encode, and a 10-bit decodable item is probed again so its HDR flag decides.
The schema half is checked by test_migrations.py; a facet value left stale is invisible there.
"""
from alembic import command
from alembic.config import Config
from sqlalchemy import text
from app.db import engine
def test_the_stored_facet_is_rewritten_by_the_gpu_encode_rule(db):
cfg = Config("alembic.ini")
command.downgrade(cfg, "0118_videos_shuffle_key")
try:
with engine.begin() as conn:
lib = conn.execute(
text(
"INSERT INTO plex_libraries (plex_key, title, kind, enabled)"
" VALUES ('m1', 'Movies', 'movie', true) RETURNING id"
)
).scalar_one()
def item(key, playable, codec, hw, depth=None, probed=False):
conn.execute(
text(
"INSERT INTO plex_items (rating_key, library_id, kind, title, playable,"
" codec_video, transcode_hw, video_bit_depth, media_probed_at)"
" VALUES (:k, :l, 'movie', :k, :p, :c, :hw, :d,"
" CASE WHEN :pr THEN now() END)"
),
{
"k": key,
"l": lib,
"p": playable,
"c": codec,
"hw": hw,
"d": depth,
"pr": probed,
},
)
item("vc1", "transcode", "VC1", True, 8, probed=True) # the old list said decodable
item("xvid", "transcode", "mpeg4", False)
item("divx", "transcode", "msmpeg4v3", False)
item("av1", "transcode", "av1", False)
item("h8", "transcode", "hevc", True, 8, probed=True)
item("h10", "transcode", "hevc", False, 10, probed=True) # SDR or HDR: unknown here
item("remux", "remux", "h264", None)
finally:
command.upgrade(cfg, "head")
with engine.connect() as conn:
rows = {
k: (hw, probed)
for k, hw, probed in conn.execute(
text("SELECT rating_key, transcode_hw, media_probed_at IS NOT NULL FROM plex_items")
).tuples()
}
assert rows == {
"vc1": (True, True),
"xvid": (True, False),
"divx": (True, False),
"av1": (False, False),
"h8": (True, True),
"h10": (False, False), # re-queued: the probe reads its HDR flag again
"remux": (None, False),
}
backend/tests/test_plex_playback_class.py
+56 -20
View File
@@ -32,15 +32,25 @@ class TestTranscodeHwCapable:
assert sync.transcode_hw_capable("transcode", "hevc") is True
assert sync.transcode_hw_capable("transcode", "vp9", bit_depth=8) is True
def test_non_decodable_codec_is_sw(self):
# AV1 / MPEG-4 ASP (XviD) can't be HW-decoded on Gen 9.5 → software regardless of bit depth.
def test_a_codec_neither_path_knows_is_sw(self):
assert sync.transcode_hw_capable("transcode", "av1") is False
assert sync.transcode_hw_capable("transcode", "mpeg4", bit_depth=8) is False
assert sync.transcode_hw_capable("transcode", "av1", bit_depth=8) is False
def test_10bit_or_hdr_hevc_is_sw(self):
# A HW-decodable codec that is 10-bit or HDR still breaks the VAAPI nv12 pipeline → software.
def test_a_cpu_decoded_8bit_codec_is_encoded_on_the_gpu(self):
# VC-1 (the iGPU has no VC-1 Advanced decode), XviD, DivX: the CPU decodes, the GPU encodes.
for codec in ("vc1", "mpeg4", "msmpeg4v3"):
assert sync.transcode_hw_capable("transcode", codec) is True
assert sync.transcode_hw_capable("transcode", "mpeg4", bit_depth=10) is False
def test_10bit_sdr_hevc_is_sw(self):
# 10-bit SDR: an exact downconvert on the CPU (libx264), never the HDR tone curve.
assert sync.transcode_hw_capable("transcode", "hevc", bit_depth=10) is False
assert sync.transcode_hw_capable("transcode", "hevc", bit_depth=8, hdr=True) is False
def test_hdr_hevc_is_encoded_on_the_gpu(self):
# HDR goes through the OpenCL tonemap between VAAPI decode and encode (the prod image ships the
# runtime); the facet said software for it while every such session ran on the GPU.
assert sync.transcode_hw_capable("transcode", "hevc", bit_depth=10, hdr=True) is True
assert sync.transcode_hw_capable("transcode", "hevc", bit_depth=8, hdr=True) is True
class TestVideoStreamProps:
@@ -87,6 +97,9 @@ class TestSyncDerivation:
row2 = PlexItem(rating_key="y", kind="movie")
sync._apply_item(row2, lib_id=1, kind="movie", meta=self._meta("av1"))
assert row2.transcode_hw is False # av1 not HW-decodable
row3 = PlexItem(rating_key="z", kind="movie")
sync._apply_item(row3, lib_id=1, kind="movie", meta=self._meta("vc1"))
assert row3.transcode_hw is True # CPU decode, GPU encode
def test_probed_value_survives_resync_when_codec_unchanged(self):
row = PlexItem(rating_key="x", kind="movie")
@@ -124,21 +137,24 @@ class TestSyncDerivation:
assert row.video_bit_depth is None and row.media_probed_at is None # re-probe scheduled
def test_migration_codec_set_matches_hw_decodable_codecs():
# Migration 0077 hardcodes the HW-decodable codec set in raw SQL and its comment mandates it stay
# identical to plex.sync.HW_DECODABLE_CODECS. A comment isn't enforcement (house rule) — pin the
# equality so adding/removing a codec from the frozenset without editing the migration fails here
# (else already-migrated prod rows keep a wrong transcode_hw the runtime derivation wouldn't catch).
def test_migration_codec_sets_match_the_sync_sets():
# The newest migration that rewrites `transcode_hw` hardcodes both codec sets in raw SQL. Pin them
# to plex.sync, so a codec added to or removed from a frozenset without a migration that rewrites
# the stored rows fails here (else prod rows keep a facet the runtime choice disagrees with).
mig = (
Path(__file__).resolve().parent.parent
/ "alembic"
/ "versions"
/ "0077_plex_playback_class.py"
/ "0119_plex_transcode_gpu.py"
)
m = re.search(r"lower\(codec_video\)\s+IN\s+\(([^)]*)\)", mig.read_text())
assert m, "could not find the codec IN(...) list in migration 0077"
codecs = {c.strip().strip("'\"").lower() for c in m.group(1).split(",")}
assert codecs == {c.lower() for c in sync.HW_DECODABLE_CODECS}
src = mig.read_text()
sets = {}
for name in ("HW_DECODABLE_CODECS", "CPU_DECODE_CODECS"):
m = re.search(name + r"\s*=\s*\(([^)]*)\)", src)
assert m, f"could not find {name} in migration 0119"
sets[name] = {c.strip().strip("'\"").lower() for c in m.group(1).split(",") if c.strip()}
assert sets["HW_DECODABLE_CODECS"] == {c.lower() for c in sync.HW_DECODABLE_CODECS}
assert sets["CPU_DECODE_CODECS"] == {c.lower() for c in sync.CPU_DECODE_CODECS}
class _FakePlex:
@@ -182,22 +198,42 @@ def test_probe_media_batch_refines_hw_to_sw_for_10bit(db):
codec_video="mpeg4",
transcode_hw=False,
)
db.add_all([hevc8, hevc10, xvid])
hdr10 = PlexItem(
rating_key="hdr",
library_id=lib.id,
kind="movie",
title="hdr",
playable="transcode",
codec_video="hevc",
transcode_hw=True,
)
db.add_all([hevc8, hevc10, xvid, hdr10])
db.commit()
plex = _FakePlex(
{
"h8": {"Media": [{"Part": [{"Stream": [{"streamType": 1, "bitDepth": 8}]}]}]},
"h10": {"Media": [{"Part": [{"Stream": [{"streamType": 1, "bitDepth": 10}]}]}]},
"hdr": {
"Media": [
{
"Part": [
{"Stream": [{"streamType": 1, "bitDepth": 10, "colorTrc": "smpte2084"}]}
]
}
]
},
}
)
probed = sync._probe_media_batch(db, plex)
db.refresh(hevc8)
db.refresh(hevc10)
db.refresh(xvid)
# Only the two HW-decodable HEVC items are probed; the XviD is SW-by-codec and never fetched.
assert probed == 2
db.refresh(hdr10)
# Only the HW-decodable HEVC items are probed; the XviD's class follows from its codec alone.
assert probed == 3
assert hevc8.video_bit_depth == 8 and hevc8.transcode_hw is True and hevc8.media_probed_at
assert hevc10.video_bit_depth == 10 and hevc10.transcode_hw is False # downgraded to software
assert hevc10.video_bit_depth == 10 and hevc10.transcode_hw is False # 10-bit SDR: software
assert hdr10.video_bit_depth == 10 and hdr10.transcode_hw is True # HDR: the GPU tonemap
assert xvid.media_probed_at is None # not in the work-list
# A second run has nothing left to do (watermarks set).
assert sync._probe_media_batch(db, plex) == 0
backend/tests/test_plex_stream.py
+85 -14
View File
@@ -284,6 +284,24 @@ class TestTranscodeCommand:
)
assert cmd[cmd.index("-vf") + 1] == "scale_vaapi=format=nv12" # not a software scale
def test_cpu_decode_uploads_the_frames_to_the_gpu_encoder(self):
# A source the GPU cannot decode (VC-1 Advanced on Gen 9.5): the CPU decodes, the frames go to
# nv12 and up to the GPU, and h264_vaapi encodes. No -hwaccel: asking the GPU to decode is what
# failed.
cmd = stream._ffmpeg_cmd(
self.SRC, 0.0, self.OUT, None, transcode=True, hw=True, cpu_decode=True
)
assert "-hwaccel" not in cmd
assert cmd[cmd.index("-init_hw_device") + 1] == f"vaapi=va:{stream._VAAPI_DEVICE}"
assert cmd[cmd.index("-filter_hw_device") + 1] == "va"
assert cmd.index("-init_hw_device") < cmd.index("-i")
assert cmd[cmd.index("-vf") + 1] == "format=nv12,hwupload"
assert _venc(cmd) == "h264_vaapi"
assert "libx264" not in cmd
assert (
cmd[cmd.index("-force_key_frames") + 1] == f"expr:gte(t,n_forced*{stream._SEG_SECONDS})"
)
def test_hdr_tonemap_uses_the_opencl_graph(self):
# The HDR/10-bit HW path: VAAPI decode → OpenCL tonemap → VAAPI encode. Gen 9.5 can't tonemap in
# the VAAPI VPP, so this is the only HW route for a 10-bit/HDR source.
@@ -365,19 +383,48 @@ class TestWantHw:
is True
)
def test_auto_falls_back_to_software_for_a_non_decodable_source(self, monkeypatch):
# AV1 / MPEG-4 ASP (XviD) aren't decodable on Gen 9.5 → the GPU can't feed its own encoder.
def test_auto_falls_back_to_software_for_a_codec_neither_path_knows(self, monkeypatch):
# AV1 is neither GPU-decodable on Gen 9.5 nor on the CPU-decode list (it can be 10-bit HDR).
self._policy(monkeypatch, "auto")
self._hwsafe(monkeypatch, True) # decodability gate must fire BEFORE the probe
self._hwsafe(monkeypatch, True) # the codec gate must fire BEFORE the probe
monkeypatch.setattr(stream, "_render_node_present", lambda: True)
assert (
stream._want_hw(None, _FakeItem(codec_video="av1"), transcode=True, src=self.SRC)
is False
)
assert (
stream._want_hw(None, _FakeItem(codec_video="mpeg4"), transcode=True, src=self.SRC)
is False
)
@pytest.mark.parametrize("codec", ["vc1", "mpeg4", "msmpeg4v3"])
def test_auto_cpu_decodes_what_the_gpu_cannot_and_still_encodes_on_it(self, monkeypatch, codec):
# Measured on prod (UHD 630, iHD 25.2.3): "No support for codec vc1 profile 3", so every VC-1
# HW session died and fell back to libx264 (~19 s to start). Decoding on the CPU and encoding on
# the GPU produced three segments in ~5 s (VC-1 1080p) and ~1 s (XviD, DivX SD).
self._policy(monkeypatch, "auto")
self._kind(monkeypatch, stream._VK_HWSAFE)
self._opencl(monkeypatch, True)
monkeypatch.setattr(stream, "_render_node_present", lambda: True)
assert stream._select_encode(
None, _FakeItem(codec_video=codec), transcode=True, src=self.SRC
) == (True, False, True)
def test_auto_a_cpu_decoded_codec_that_is_not_8bit_sdr_stays_software(self, monkeypatch):
# The upload path converts to nv12 for the encoder; only an 8-bit SDR picture is known to come
# through it unchanged.
self._policy(monkeypatch, "auto")
self._kind(monkeypatch, stream._VK_SDR10)
monkeypatch.setattr(stream, "_render_node_present", lambda: True)
assert stream._select_encode(
None, _FakeItem(codec_video="mpeg4"), transcode=True, src=self.SRC
) == (False, False, False)
def test_on_cpu_decodes_a_codec_the_gpu_cannot(self, monkeypatch):
# "on" trusts the GPU, but asking it to DECODE a codec it is known not to decode only buys the
# fallback's wasted second; the upload path is the GPU route for such a source.
self._policy(monkeypatch, "on")
self._opencl(monkeypatch, False)
monkeypatch.setattr(stream, "_render_node_present", lambda: True)
assert stream._select_encode(
None, _FakeItem(codec_video="vc1"), transcode=True, src=self.SRC
) == (True, False, True)
def test_auto_hdr_source_falls_back_to_software_without_opencl(self, monkeypatch):
# A decodable HEVC that is true-HDR (PQ/HLG) can't use scale_vaapi=format=nv12 (Gen 9.5), and with
@@ -388,7 +435,7 @@ class TestWantHw:
monkeypatch.setattr(stream, "_render_node_present", lambda: True)
assert stream._select_encode(
None, _FakeItem(codec_video="hevc"), transcode=True, src=self.SRC
) == (False, False)
) == (False, False, False)
def test_auto_hdr_source_uses_hw_tonemap_with_opencl(self, monkeypatch):
# With the OpenCL runtime present, a true-HDR HEVC tone-maps on the GPU (VAAPI decode → OpenCL
@@ -399,7 +446,7 @@ class TestWantHw:
monkeypatch.setattr(stream, "_render_node_present", lambda: True)
assert stream._select_encode(
None, _FakeItem(codec_video="hevc"), transcode=True, src=self.SRC
) == (True, True)
) == (True, True, False)
def test_auto_10bit_SDR_source_uses_software_not_tonemap(self, monkeypatch):
# F1 regression: a 10-bit *SDR* (bt709) source is NOT HDR, so it must NOT get the hable HDR tone
@@ -411,7 +458,7 @@ class TestWantHw:
monkeypatch.setattr(stream, "_render_node_present", lambda: True)
assert stream._select_encode(
None, _FakeItem(codec_video="hevc"), transcode=True, src=self.SRC
) == (False, False)
) == (False, False, False)
def test_auto_8bit_source_uses_plain_hw_not_tonemap(self, monkeypatch):
# An 8-bit SDR source takes the plain full-GPU scale path — never the tonemap graph — even where
@@ -422,7 +469,7 @@ class TestWantHw:
monkeypatch.setattr(stream, "_render_node_present", lambda: True)
assert stream._select_encode(
None, _FakeItem(codec_video="hevc"), transcode=True, src=self.SRC
) == (True, False)
) == (True, False, False)
def test_off_forces_software_even_for_a_decodable_source(self, monkeypatch):
self._policy(monkeypatch, "off")
@@ -460,7 +507,7 @@ class TestWantHw:
monkeypatch.setattr(stream, "_render_node_present", lambda: True)
assert stream._select_encode(
None, _FakeItem(codec_video="hevc"), transcode=True, src=self.SRC
) == (True, True)
) == (True, True, False)
def test_on_10bit_SDR_uses_plain_hw_not_tonemap(self, monkeypatch):
# F1 regression under "on": a 10-bit SDR source is NOT HDR, so even with OpenCL present "on" must
@@ -471,7 +518,7 @@ class TestWantHw:
monkeypatch.setattr(stream, "_render_node_present", lambda: True)
assert stream._select_encode(
None, _FakeItem(codec_video="hevc"), transcode=True, src=self.SRC
) == (True, False)
) == (True, False, False)
def test_on_without_opencl_is_plain_hw(self, monkeypatch):
# "on" with no OpenCL runtime → plain HW, tonemap=False, regardless of source class (the runtime
@@ -483,7 +530,7 @@ class TestWantHw:
monkeypatch.setattr(stream, "_render_node_present", lambda: True)
assert stream._select_encode(
None, _FakeItem(codec_video="hevc"), transcode=True, src=self.SRC
) == (True, False)
) == (True, False, False)
class TestColdStartWarmup:
@@ -620,6 +667,30 @@ class TestTranscodeStartSession:
assert cmd[cmd.index("-vf") + 1] == "scale_vaapi=format=nv12" # no SW cap
assert kwargs.get("preexec_fn") is None # GPU encode is cheap — never niced
def test_a_cpu_decode_session_encodes_on_the_gpu_and_falls_back_to_software(
self, capturing_env, monkeypatch
):
monkeypatch.setattr(
stream, "_select_encode", lambda db, item, transcode, src: (True, False, True)
)
seen = {"n": 0}
def flaky_wait(path, timeout=20.0, proc=None):
seen["n"] += 1
return seen["n"] > 1 # the upload session yields nothing; the software respawn does
monkeypatch.setattr(stream, "wait_for", flaky_wait)
s = stream.start_session(
db=None, user_id=7, item=_FakeItem(codec_video="vc1"), start_s=0.0, transcode=True
)
(up_cmd, *_), up_kwargs = _CapturingProc.calls[0]
assert "format=nv12,hwupload" in up_cmd and _venc(up_cmd) == "h264_vaapi"
if os.name == "posix": # the decode is CPU work: niced like a software transcode
assert up_kwargs.get("preexec_fn") is stream._sw_child_preexec
(sw_cmd, *_), _ = _CapturingProc.calls[1]
assert _venc(sw_cmd) == "libx264" and "format=nv12,hwupload" not in sw_cmd
assert s is not None and s.hw is False and s.cpu_decode is False
def test_hdr_tonemap_transcode_threads_the_opencl_graph_and_flags_the_session(
self, capturing_env, monkeypatch
):
backend/tests/test_reader_service.py
+2 -2
View File
@@ -252,7 +252,7 @@ def test_an_add_is_idempotent_and_lands_where_asked(db, make_user):
db.commit()
assert _ids(db, pl) == ["c", "b"]
assert queries.item_counts(db, [pl.id]) == {pl.id: 2}
assert queries.durations(db, [pl.id]) == {pl.id: 120}
assert queries.durations(db, [pl.id]) == {pl.id: (120, 0)}
assert queries.containing(db, [pl.id], "b") == {pl.id}
@@ -294,7 +294,7 @@ def test_a_rename_and_a_delete(db, make_user):
service.add_item(db, pl, "a", mark_dirty=False)
service.rename_playlist(db, pl, "New")
db.commit()
assert queries.playlist_totals(db, pl.id) == (1, 60, None)
assert queries.playlist_totals(db, pl.id) == (1, 60, 0, None)
assert pl.name == "New"
service.delete_playlist(db, pl)
db.commit()
backend/tests/test_shuffle_walk.py
+223
View File
@@ -0,0 +1,223 @@
"""A seeded shuffle walks a stored random key instead of sorting the whole scope. DB lane.
Every video carries a random number drawn once (`Video.shuffle_key`). A seed picks a point on that
ring; the feed's Shuffle sort pages along the ring from there and wraps, and the Home surprise reads
a window of it and shuffles that window by the seed (`discover.queries.yt_surprise`). Ordering the
scope by md5(id || seed) cost ~350 ms per call on prod and grew with every Shuffle page.
"""
import base64
import random
from contextlib import contextmanager
from datetime import datetime, timedelta, timezone
from sqlalchemy import event, insert, text
from app import feedquery
from app.catalog.models import Channel, Subscription, Video
from app.discover import queries
from app.discover.queries import _FEED_BASE
NOW = datetime.now(timezone.utc)
KEYS = {"k05": 0.05, "k20": 0.20, "k40": 0.40, "k60": 0.60, "k80": 0.80, "k95": 0.95}
def _seed(db, make_user, keys=KEYS):
user = make_user()
db.add(Channel(id="UCring", title="Ring"))
db.flush()
db.add(Subscription(user_id=user.id, channel_id="UCring"))
db.add_all(
[
Video(
id=vid,
channel_id="UCring",
title=vid,
duration_seconds=20 * 60,
published_at=NOW - timedelta(days=i + 1),
shuffle_key=key,
)
for i, (vid, key) in enumerate(keys.items())
]
)
db.commit()
return user
def _ring(seed, keys=KEYS):
"""The ring order from the seed's point: the keys at or after it ascending, then the rest."""
r = feedquery.shuffle_start(seed)
return sorted(keys, key=lambda v: (keys[v] < r, keys[v]))
def _seed_landing_between(lo, hi):
return next(s for s in range(1000) if lo < feedquery.shuffle_start(s) < hi)
def test_the_walk_starts_at_the_seeds_point_and_wraps(db, make_user):
user = _seed(db, make_user)
query, _ = feedquery.filtered_query(db, user, **_FEED_BASE)
seed = _seed_landing_between(0.45, 0.55) # starts at k60, wraps to k05 after k95
got = [r.id for r in feedquery.shuffle_walk(db, query, seed, 6)]
assert got == ["k60", "k80", "k95", "k05", "k20", "k40"]
assert [r.id for r in feedquery.shuffle_walk(db, query, seed, 4)] == got[:4]
def test_the_feed_shuffle_pages_along_the_ring(db, make_user, as_user):
user = _seed(db, make_user)
client = as_user(user)
for seed in (_seed_landing_between(0.1, 0.15), _seed_landing_between(0.85, 0.9)):
walked, cursor = [], None
for _ in range(5):
url = f"/api/feed?sort=shuffle&seed={seed}&limit=2" + (
f"&cursor={cursor}" if cursor else ""
)
body = client.get(url).json()
walked += [it["id"] for it in body["items"]]
cursor = body["next_cursor"]
if not body["has_more"]:
break
assert walked == _ring(seed), seed
def test_a_new_seed_opens_the_shuffle_elsewhere(db, make_user, as_user):
user = _seed(db, make_user)
client = as_user(user)
first = {
seed: client.get(f"/api/feed?sort=shuffle&seed={seed}&limit=2").json()["items"][0]["id"]
for seed in (_seed_landing_between(0.1, 0.15), _seed_landing_between(0.65, 0.7))
}
assert sorted(first.values()) == ["k20", "k80"]
@contextmanager
def _statements(db):
seen: list[str] = []
def grab(conn, cursor, statement, *args):
seen.append(statement)
engine = db.get_bind()
event.listen(engine, "before_cursor_execute", grab)
try:
yield seen
finally:
event.remove(engine, "before_cursor_execute", grab)
def _draws_of_videos(seen):
return [s for s in seen if "FROM videos" in s and "LIMIT" in s]
def test_neither_draw_orders_the_scope_by_a_hash(db, make_user):
# Each draw reads the videos as an ascending range of the key (what ix_videos_shuffle_key serves),
# never ordered by a hash or by the ring's CASE (which no index can serve).
user = _seed(db, make_user)
with _statements(db) as surprise:
queries.yt_surprise(db, user, "any", 7, 3)
with _statements(db) as page:
feedquery.feed_page(
db, user, dict(_FEED_BASE), sort="shuffle", seed=7, limit=2, cursor=None, lang="en"
)
for seen in (surprise, page):
draws = _draws_of_videos(seen)
assert draws
for s in draws:
assert "md5(" not in s.lower()
assert "ORDER BY videos.shuffle_key, videos.id" in s.replace("\n", " "), s
def test_the_walk_reads_the_ring_index_and_sorts_nothing(db, make_user):
# The text of the query is not the plan: each arc must be served by ix_videos_shuffle_key in key
# order (an index on another shape, or none, leaves a Sort of the whole scope). A scope of a few
# thousand videos, analysed, so the planner weighs it as it does a real one.
user = _seed(db, make_user)
rng = random.Random(7)
db.execute(
insert(Video),
[
{
"id": f"bulk{i:05}",
"channel_id": "UCring",
"title": "b",
"duration_seconds": 600,
"published_at": NOW - timedelta(minutes=i),
"shuffle_key": rng.random(),
}
for i in range(4000)
],
)
db.commit()
db.execute(text("ANALYZE videos"))
query, _ = feedquery.filtered_query(db, user, **_FEED_BASE)
captured: list[tuple[str, object]] = []
def grab(conn, cursor, statement, params, *args):
if "FROM videos" in statement and "LIMIT" in statement:
captured.append((statement, params))
engine = db.get_bind()
event.listen(engine, "before_cursor_execute", grab)
try:
seed = _seed_landing_between(0.45, 0.55)
feedquery.shuffle_walk(db, query, seed, 6) # the first arc, from the seed's point
feedquery.shuffle_walk(db, query, seed, 6, (1, 0.3, "bulk00001")) # the wrap, resumed
finally:
event.remove(engine, "before_cursor_execute", grab)
assert len(captured) == 2
conn = db.connection()
for statement, params in captured:
plan = "\n".join(r[0] for r in conn.exec_driver_sql("EXPLAIN " + statement, params))
assert "ix_videos_shuffle_key" in plan, plan
assert "Sort Key: videos.shuffle_key" not in plan, plan
def test_the_surprise_reads_a_window_of_the_ring(db, make_user, monkeypatch):
user = _seed(db, make_user)
monkeypatch.setattr(queries, "SURPRISE_WINDOW", 3)
for seed in range(20):
drawn = {r["id"] for r in queries.yt_surprise(db, user, "any", seed, 3)}
assert drawn == set(_ring(seed)[:3]), seed
def test_a_wide_gap_on_the_ring_does_not_make_a_favourite(db, make_user):
# One video after an empty half of the ring, twenty packed into the other half. Taking the first
# video after the seed's point would deal the lonely one about half the time; the window is
# shuffled by the seed, so it leads about as often as any other.
keys = {"lonely": 0.5} | {f"p{i:02}": 0.51 + i * 0.02 for i in range(20)}
user = _seed(db, make_user, keys)
leads = [queries.yt_surprise(db, user, "any", s, 1)[0]["id"] for s in range(90)]
assert leads.count("lonely") < 15, leads.count("lonely")
def test_any_seed_lands_on_the_ring():
# An int too large for a float overflowed the multiplication: a 500 for a hand-typed seed.
for seed in (10**400, -(10**400), 2**53 + 1, -7):
assert 0 <= feedquery.shuffle_start(seed) < 1, seed
def test_a_seed_out_of_range_is_refused_not_a_500(db, make_user, as_user):
client = as_user(_seed(db, make_user))
huge = "9" * 400
for url in (
f"/api/feed?sort=shuffle&seed={huge}",
f"/api/discover?seed={huge}",
f"/api/discover/pick?seed={huge}",
):
assert client.get(url).status_code == 422, url
def test_a_shuffle_cursor_past_a_float_is_a_bad_cursor(db, make_user, as_user):
client = as_user(_seed(db, make_user))
for k in ("[Infinity, 0.5]", f"[0, {'9' * 400}]"):
raw = '{"k": %s, "id": "k20"}' % k
cursor = base64.urlsafe_b64encode(raw.encode()).decode()
assert client.get(f"/api/feed?sort=shuffle&seed=1&cursor={cursor}").status_code == 400, k
def test_a_new_video_gets_its_own_key(db, make_user):
_seed(db, make_user, {})
db.add_all([Video(id=f"n{i}", channel_id="UCring", title="n") for i in range(3)])
db.commit()
keys = [db.get(Video, f"n{i}").shuffle_key for i in range(3)]
assert len(set(keys)) == 3 and all(0 <= k < 1 for k in keys)
backend/tests/test_surprise_affinity.py
+349
View File
@@ -0,0 +1,349 @@
"""Surprise me leans on what the user watches, and stays a surprise.
The Home hero's surprise tier and the Plex Discover rail were uniform draws over the unseen. They are
now weighted draws (`app.draw`): a title's weight is 1 + its affinity — for Plex the genres of the
films and shows the user watched (recent ones weigh more) and the TMDB recommendations stored on them;
for YouTube the channels the user watches. One slot of three stays a blind draw, and with no watch
history the order is exactly the old one. DB lane.
"""
import hashlib
from datetime import datetime, timedelta, timezone
from sqlalchemy import select, text
from app import draw
from app.catalog.models import Channel, Subscription, Video
from app.discover import queries
from app.plex import affinity, discover as plex_discover
from app.plex.models import PlexItem, PlexLibrary, PlexShow, PlexState
from app.plex.query import LibraryFilters
from app.reader.models import VideoState
NOW = datetime.now(timezone.utc)
SEEDS = range(90)
# ── the draw itself ──────────────────────────────────────────────────────────────────────────────
def test_every_third_slot_is_blind_and_nothing_repeats():
weighted = ["a", "b", "c", "d", "e", "f"]
blind = ["f", "e", "d", "c", "b", "a"]
# offset 0: slots 0 and 3 are blind
assert draw.blend(weighted, blind, 0, 6) == ["f", "a", "b", "e", "c", "d"]
# offset 1: slot 2 and 5
assert draw.blend(weighted, blind, 1, 6) == ["a", "b", "f", "c", "d", "e"]
assert draw.blend(weighted, blind, 2, 3) == ["a", "f", "b"]
def test_a_short_side_is_topped_up_from_the_other():
assert draw.blend(["a"], ["a", "b", "c"], 0, 3) == ["a", "b", "c"]
assert draw.blend(["a", "b", "c"], [], 0, 3) == ["a", "b", "c"]
def test_a_watch_weighs_by_its_age_an_undated_one_two_half_lives_and_a_future_one_now():
# Both affinities weigh every watch through this one rule; an old import has no date, and a
# clock ahead of the server's must not weigh more than a watch made now.
assert draw.recency(NOW, NOW, 30) == 1.0
assert draw.recency(NOW - timedelta(days=30), NOW, 30) == 0.5
assert draw.recency(None, NOW, 30) == 0.25
assert draw.recency(NOW + timedelta(days=5), NOW, 30) == 1.0
def test_scores_are_scaled_to_the_strongest_and_nothing_scored_is_empty():
assert draw.scaled({"a": 4.0, "b": 1.0}) == {"a": 1.0, "b": 0.25}
assert draw.scaled({}) == {}
# ── YouTube: the channels the user watches ───────────────────────────────────────────────────────
def _md5_order(ids, seed):
return sorted(ids, key=lambda i: hashlib.md5(f"{i}{seed}".encode()).hexdigest())
def _seed_yt(db, make_user, *, watched_in_fav: int):
user = make_user()
fav, other = Channel(id="UCfav", title="Fav"), Channel(id="UCother", title="Other")
db.add_all([fav, other])
db.flush()
db.add_all(
[
Subscription(user_id=user.id, channel_id="UCfav"),
Subscription(user_id=user.id, channel_id="UCother"),
]
)
def video(vid, ch, days):
return Video(
id=vid,
channel_id=ch,
title=vid,
duration_seconds=20 * 60,
published_at=NOW - timedelta(days=days),
)
# One unseen video per channel, plus the user's watched history on the favourite.
db.add_all([video("fav_new", "UCfav", 1), video("other_new", "UCother", 1)])
for i in range(watched_in_fav):
db.add(video(f"fav_seen{i}", "UCfav", 10 + i))
db.flush()
for i in range(watched_in_fav):
db.add(
VideoState(
user_id=user.id,
video_id=f"fav_seen{i}",
status="watched",
watched_at=NOW - timedelta(days=i),
)
)
db.commit()
return user
def _first(rows):
return rows[0]["id"] if rows else None
def test_a_watched_channel_wins_most_draws_but_not_all(db, make_user):
user = _seed_yt(db, make_user, watched_in_fav=12)
leads = [_first(queries.yt_surprise(db, user, "any", s, 2)) for s in SEEDS]
fav = leads.count("fav_new") / len(leads)
# weight 1 + 3 against 1: 0.8 in a weighted slot, 0.5 in the blind one (a third of the seeds)
assert 0.6 < fav < 0.9, fav
def test_the_blind_slot_deals_the_plain_shuffle_however_strong_the_lean(db, make_user):
user = _seed_yt(db, make_user, watched_in_fav=12)
for s in range(0, 30, 3): # (0 + seed) % 3 == 0: the first slot is the blind one
got = _first(queries.yt_surprise(db, user, "any", s, 2))
assert got == _md5_order(["fav_new", "other_new"], s)[0]
def test_with_no_history_the_youtube_order_is_the_old_shuffle(db, make_user):
user = _seed_yt(db, make_user, watched_in_fav=0)
for s in range(12):
got = [r["id"] for r in queries.yt_surprise(db, user, "any", s, 2)]
assert got == _md5_order(["fav_new", "other_new"], s)
def test_a_watched_video_is_never_drawn(db, make_user):
user = _seed_yt(db, make_user, watched_in_fav=12)
drawn = {r["id"] for s in SEEDS for r in queries.yt_surprise(db, user, "any", s, 5)}
assert drawn == {"fav_new", "other_new"}
# ── Plex: genres and recommendations of what the user watched ────────────────────────────────────
def _seed_plex(db, make_user, *, history: bool):
user = make_user()
lib = PlexLibrary(plex_key="77", title="Movies", kind="movie", enabled=True)
db.add(lib)
db.flush()
tmdb_ids = iter(range(5000, 6000))
def movie(key, genres, recs=None):
m = PlexItem(
rating_key=key,
library_id=lib.id,
kind="movie",
title=key,
duration_s=100 * 60,
genres=genres,
guid=f"tmdb:{next(tmdb_ids)}",
recommendations=recs,
)
db.add(m)
return m
horror = movie("horror_new", ["Horror"])
comedy = movie("comedy_new", ["Comedy"])
tipped = movie("tipped_new", ["Western"])
db.flush()
tip = [{"tmdb_id": int(tipped.guid.split(":")[1]), "media_type": "movie", "title": "t"}]
seen = [movie(f"horror_seen{i}", ["Horror"], tip if i == 0 else None) for i in range(3)]
db.flush()
if history:
for i, m in enumerate(seen):
db.add(
PlexState(
user_id=user.id,
item_id=m.id,
status="watched",
watched_at=NOW - timedelta(days=i),
)
)
db.commit()
from app import sysconfig
sysconfig.set_value(db, "plex_enabled", True)
db.commit()
return user, horror, comedy, tipped
def test_the_genre_and_the_recommendation_of_what_was_watched_win_most_draws(db, make_user):
user, *_ = _seed_plex(db, make_user, history=True)
leads = [_first(queries.plex_surprise(db, user.id, "any", s, 3, lang="en")) for s in SEEDS]
# Weights: Horror 1 + 2 (its genre), the recommended Western 1 + 3, the Comedy nobody chose 1. In
# a weighted slot the Comedy leads 1 time in 8, in the blind third 1 in 3: about 0.19 against
# 0.81 for the other two together (a uniform draw: 1 in 3 each). It still leads now and then —
# a surprise, not a ranking.
comedy = leads.count("comedy_new")
assert comedy > 0
assert leads.count("horror_new") > 1.5 * comedy and leads.count("tipped_new") > 1.5 * comedy
assert leads.count("horror_new") + leads.count("tipped_new") > 3 * comedy
def test_with_no_history_the_plex_order_is_the_old_shuffle(db, make_user):
user, *_ = _seed_plex(db, make_user, history=False)
keys = ["horror_new", "comedy_new", "tipped_new"] + [f"horror_seen{i}" for i in range(3)]
for s in range(12):
got = [c["id"] for c in queries.plex_surprise(db, user.id, "any", s, 6, lang="en")]
assert got == _md5_order(keys, s)
def test_a_watched_film_is_never_drawn(db, make_user):
user, *_ = _seed_plex(db, make_user, history=True)
drawn = {
c["id"] for s in SEEDS for c in queries.plex_surprise(db, user.id, "any", s, 6, lang="en")
}
assert drawn == {"horror_new", "comedy_new", "tipped_new"}
def test_the_plex_blind_slot_deals_the_plain_shuffle(db, make_user):
user, *_ = _seed_plex(db, make_user, history=True)
unseen = ["horror_new", "comedy_new", "tipped_new"]
for s in range(0, 30, 3):
got = _first(queries.plex_surprise(db, user.id, "any", s, 3, lang="en"))
assert got == _md5_order(unseen, s)[0]
def test_a_binged_show_counts_once_and_recent_watches_weigh_more(db, make_user):
user = make_user()
lib = PlexLibrary(plex_key="78", title="Shows", kind="show", enabled=True)
mlib = PlexLibrary(plex_key="79", title="Movies", kind="movie", enabled=True)
db.add_all([lib, mlib])
db.flush()
show = PlexShow(rating_key="sitcom", library_id=lib.id, title="Sitcom", genres=["Comedy"])
db.add(show)
db.flush()
eps = [
PlexItem(
rating_key=f"ep{i}",
library_id=lib.id,
kind="episode",
title=f"e{i}",
show_id=show.id,
season_number=1,
episode_number=i,
)
for i in range(10)
]
films = [
PlexItem(
rating_key=f"h{i}", library_id=mlib.id, kind="movie", title=f"h{i}", genres=["Horror"]
)
for i in range(2)
]
old = PlexItem(
rating_key="w0", library_id=mlib.id, kind="movie", title="w0", genres=["Western"]
)
db.add_all([*eps, *films, old])
db.flush()
for it in [*eps, *films]:
db.add(PlexState(user_id=user.id, item_id=it.id, status="watched", watched_at=NOW))
db.add(
PlexState(
user_id=user.id, item_id=old.id, status="watched", watched_at=NOW - timedelta(days=720)
)
)
db.commit()
genres = affinity.for_user(db, user.id).genres
# Ten episodes of one show are one title: two horror films outweigh it.
assert genres["Horror"] == 1.0
assert genres["Comedy"] == 0.5
# Four half-lives ago weighs a sixteenth of today.
assert abs(genres["Western"] - 0.5 / 16) < 0.01
def test_the_discover_rail_leans_the_same_way(db, make_user):
# The rail draws with random(), not a seed; setseed makes this session's draws repeatable.
user, *_ = _seed_plex(db, make_user, history=True)
db.execute(text("SELECT setseed(0.42)"))
leads = [
plex_discover.surprise(db, user, "movie", "en", 3, [], [], LibraryFilters())["items"][0][
"id"
]
for _ in SEEDS
]
assert leads.count("comedy_new") > 0
assert leads.count("horror_new") + leads.count("tipped_new") > 3 * leads.count("comedy_new")
# ── whose history: only the user's own ───────────────────────────────────────────────────────────
def test_another_users_youtube_history_does_not_lean_my_draw(db, make_user):
watcher = _seed_yt(db, make_user, watched_in_fav=12)
me = make_user()
db.add_all(
[
Subscription(user_id=me.id, channel_id="UCfav"),
Subscription(user_id=me.id, channel_id="UCother"),
]
)
db.commit()
assert queries._channel_affinity(db, watcher)
assert queries._channel_affinity(db, me) == {}
def test_another_users_plex_history_does_not_lean_my_draw(db, make_user):
_seed_plex(db, make_user, history=True)
me = make_user()
assert not affinity.for_user(db, me.id)
keys = ["horror_new", "comedy_new", "tipped_new"] + [f"horror_seen{i}" for i in range(3)]
for s in range(12):
got = [c["id"] for c in queries.plex_surprise(db, me.id, "any", s, 6, lang="en")]
assert got == _md5_order(keys, s)
# ── the rail: its blind slot, and a show's own recommendations ───────────────────────────────────
def test_the_rails_third_slot_is_the_blind_draw(db, make_user, monkeypatch):
# The rail draws with random(), so the blend is watched rather than its odds: every leaning draw
# goes through it, with the blind order beside the weighted one and the third slot blind.
user, *_ = _seed_plex(db, make_user, history=True)
calls = []
real = draw.blend
def spy(weighted, blind, offset, n, key):
calls.append((offset, n, sorted(map(key, weighted)), sorted(map(key, blind))))
return real(weighted, blind, offset, n, key=key)
monkeypatch.setattr(draw, "blend", spy)
got = plex_discover.surprise(db, user, "movie", "en", 3, [], [], LibraryFilters())["items"]
assert calls and all(offset == 1 and n == 3 for offset, n, *_ in calls)
pool = sorted(["horror_new", "comedy_new", "tipped_new"])
assert calls[0][2] == pool and calls[0][3] == pool
assert len(got) == 3
def test_a_show_leans_on_the_shows_recommended_not_the_films(db, make_user):
lib = PlexLibrary(plex_key="80", title="Shows", kind="show", enabled=True)
db.add(lib)
db.flush()
tipped = PlexShow(rating_key="tv1", library_id=lib.id, title="t", guid="tmdb:1", genres=[])
db.add(tipped)
db.commit()
shows = affinity.Affinity(show_recs={"tmdb:1": 1.0})
films = affinity.Affinity(movie_recs={"tmdb:1": 1.0})
def weight(aff):
return db.execute(select(aff.weight(PlexShow)).where(PlexShow.id == tipped.id)).scalar_one()
assert weight(shows) == 1 + affinity.REC_WEIGHT
assert weight(films) == 1
backend/tests/test_video_length.py
+218
View File
@@ -0,0 +1,218 @@
"""A YouTube length past a day is unknown, everywhere a length is used (`catalog.models.LENGTH_MAX_S`).
Measured on prod (2026-11-06): two 24/7 streams' archives report 122,962,914 s and 34,762,459 s, and the
owner's Watch later total read "44,073 hr 55 min". Such a video now has no length on any API answer, the
totals leave it out (a playlist's counts it as `without_length`), and the duration filter and sort pass it over
like a video whose length is not fetched yet. An AST check holds who may still read the raw column.
DB lane (the AST check needs none).
"""
import ast
from datetime import datetime, timezone
from pathlib import Path
from sqlalchemy.dialects import postgresql
from app.catalog.models import LENGTH_MAX_S, Channel, Subscription, Video, known_length
from app.reader.models import Playlist, PlaylistItem
YEARS = 122_962_914 # the measured archive
HOUR = 3600
def _seed(db, user):
db.add(Channel(id="UClen", title="Len"))
db.flush()
db.add(Subscription(user_id=user.id, channel_id="UClen", priority=0, hidden=False))
rows = [
("short", 600, "none"),
("long", 2 * HOUR, "none"),
("day", LENGTH_MAX_S, "was_live"), # exactly a day still is a length
("archive", YEARS, "was_live"),
("onair", None, "live"),
("unfetched", None, "none"), # not live, but left out of the sum: counted all the same
]
for i, (vid, seconds, live) in enumerate(rows):
db.add(
Video(
id=vid,
channel_id="UClen",
published_at=datetime(2024, 1, i + 1, tzinfo=timezone.utc),
title=vid,
duration_seconds=seconds,
live_status=live,
is_short=False,
via_search=False,
via_explore=False,
)
)
pl = Playlist(user_id=user.id, name="Later")
db.add(pl)
db.flush()
for i, (vid, _s, _l) in enumerate(rows):
db.add(PlaylistItem(playlist_id=pl.id, video_id=vid, position=i))
db.commit()
return pl
KNOWN_SUM = 600 + 2 * HOUR + LENGTH_MAX_S
def test_the_rule():
assert known_length(None) is None
assert known_length(LENGTH_MAX_S) == LENGTH_MAX_S
assert known_length(LENGTH_MAX_S + 1) is None
assert known_length(YEARS) is None
def test_the_sql_length_carries_its_limit_as_a_literal():
# The duration sort indexes (migration 0117) are built on this expression with 86400 in it. A
# bound parameter matches no index once the driver reuses a prepared, generic plan, and the
# sort-index test cannot see that: it compiles with literal binds.
compiled = Video.length_seconds.expression.compile(dialect=postgresql.dialect())
assert compiled.params == {}
assert f"> {LENGTH_MAX_S}" in str(compiled)
def test_a_years_long_archive_has_no_length_in_the_feed_or_a_playlist(db, make_user, as_user):
user = make_user()
pl = _seed(db, user)
client = as_user(user)
feed = client.get("/api/feed?show=all&include_live=true&limit=50")
assert feed.status_code == 200
by_id = {v["id"]: v["duration_seconds"] for v in feed.json()["items"]}
assert by_id["archive"] is None
assert by_id["day"] == LENGTH_MAX_S
detail = client.get(f"/api/playlists/{pl.id}")
assert detail.status_code == 200
items = {v["id"]: v["duration_seconds"] for v in detail.json()["items"]}
assert items["archive"] is None
one = client.get("/api/videos/archive")
assert one.status_code == 200
assert one.json()["duration_seconds"] is None
def test_an_off_catalog_archive_has_no_length_either(db, make_user, as_user, monkeypatch):
# A video not in the catalog is looked up on YouTube (videos.list); its length is judged the same.
from app.routes import feed as feed_routes
class Yt:
def __init__(self, *_):
pass
def __enter__(self):
return self
def __exit__(self, *exc):
return False
def get_videos(self, ids):
hours = {"offyears": YEARS // HOUR, "offshort": 1}[ids[0]]
return [
{"snippet": {}, "statistics": {}, "contentDetails": {"duration": f"PT{hours}H"}}
]
monkeypatch.setattr(feed_routes, "YouTubeClient", Yt)
client = as_user(make_user())
assert client.get("/api/videos/offyears").json()["duration_seconds"] is None
assert client.get("/api/videos/offshort").json()["duration_seconds"] == HOUR
def test_the_totals_leave_it_out_and_count_what_they_left_out(db, make_user, as_user):
user = make_user()
pl = _seed(db, user)
client = as_user(user)
listed = client.get("/api/playlists")
assert listed.status_code == 200
(summary,) = [p for p in listed.json() if p["id"] == pl.id]
assert summary["total_duration_seconds"] == KNOWN_SUM
assert summary["without_length"] == 3 # the archive, the one on air, the unfetched one
renamed = client.patch(f"/api/playlists/{pl.id}", json={"name": "Later 2"})
assert renamed.status_code == 200
assert renamed.json()["total_duration_seconds"] == KNOWN_SUM
assert renamed.json()["without_length"] == 3
channels = client.get("/api/channels")
assert channels.status_code == 200
(ch,) = channels.json()
assert ch["total_duration_seconds"] == KNOWN_SUM
# The channel table shows the sum alone (the count did not fit its column), so it is not sent.
assert "without_length" not in ch and "live_without_length" not in ch
def test_the_duration_filter_and_sort_pass_it_over(db, make_user, as_user):
user = make_user()
_seed(db, user)
client = as_user(user)
base = "/api/feed?show=all&include_live=true&limit=50"
longer = client.get(f"{base}&min_duration={HOUR}")
assert longer.status_code == 200
assert {v["id"] for v in longer.json()["items"]} == {"long", "day"}
longest = client.get(f"{base}&sort=duration_desc")
assert longest.status_code == 200
order = [v["id"] for v in longest.json()["items"]]
assert order[:3] == ["day", "long", "short"]
assert set(order[3:]) == {"archive", "onair", "unfetched"}
# Who may read `duration_seconds` itself, by (file, function): what each one asks is not "how long".
RAW_READERS = {
# The column is defined and turned into a length here.
("catalog/models.py", "length_seconds"),
("catalog/models.py", "_length_seconds_expression"),
# Ingest writes it; the live refresh asks whether YouTube has answered for an archive yet; the
# Shorts classification needs only "longer than a Short", which a years-long value proves.
("sync/videos.py", "apply_video_details"),
("sync/videos.py", "refresh_live"),
("sync/videos.py", "run_shorts_classification"),
("catalog/search.py", "classify_shorts"),
("catalog/queries.py", "work_queue"),
# Not a video's column: the player's own report, and the watch history's looked-up lengths, which
# count a watched video as more than a glance (a watched archive is that too).
("routes/_common.py", "kept_position"),
("routes/plex/playback.py", "item_progress"),
("reader/queries.py", "_counted_channels"),
("reader/queries.py", "watched_channels"),
("reader/queries.py", "watched_channel_videos"),
}
def _raw_reads() -> set[tuple[str, str]]:
root = Path(__file__).resolve().parents[1] / "app"
found: set[tuple[str, str]] = set()
for path in root.rglob("*.py"):
if "alembic" in path.parts:
continue
rel = path.relative_to(root).as_posix()
tree = ast.parse(path.read_text(encoding="utf-8"))
in_a_function: set[int] = set()
for fn in ast.walk(tree):
if not isinstance(fn, (ast.FunctionDef, ast.AsyncFunctionDef)):
continue
for node in ast.walk(fn):
if _reads_raw(node):
found.add((rel, fn.name))
in_a_function.add(id(node))
# A read outside every def (module or class level, a lambda there) is named `<module>`.
if any(_reads_raw(n) and id(n) not in in_a_function for n in ast.walk(tree)):
found.add((rel, "<module>"))
return found
def _reads_raw(node: ast.AST) -> bool:
return isinstance(node, ast.Attribute) and node.attr == "duration_seconds"
def test_only_the_listed_readers_take_the_raw_column():
found = _raw_reads()
stray = sorted(found - RAW_READERS)
assert not stray, f"read `length_seconds` (a length) instead of the raw column: {stray}"
# A listed reader that stopped reading it leaves the list.
assert not RAW_READERS - found, sorted(RAW_READERS - found)
backend/tests/test_ytdlp_logging.py
+2 -2
View File
@@ -216,7 +216,7 @@ def test_a_line_with_no_url_is_left_alone(sink):
def test_the_download_options_carry_the_logger():
opts = formats.build_ydl_opts({}, "%(id)s.%(ext)s", lambda d: None)
opts = formats.build_ydl_opts({}, "%(id)s.%(ext)s", lambda d: None, egress=False)
assert opts["logger"] is formats._YDL_LOG
@@ -227,7 +227,7 @@ def test_a_source_adapter_cannot_silence_it():
adapter must not be able to take away the one account of what went wrong that a failed job
leaves behind."""
opts = formats.build_ydl_opts(
{}, "%(id)s.%(ext)s", lambda d: None, extra={"logger": None, "quiet": False}
{}, "%(id)s.%(ext)s", lambda d: None, extra={"logger": None, "quiet": False}, egress=False
)
assert opts["logger"] is formats._YDL_LOG
docs/DESIGN.md
+160 -152
View File
@@ -568,6 +568,9 @@ of a full-width panel and nothing on the right.
the first tap however long the video played).
- **The bars stay up while the pointer rests on them** (F22, owner); leaving starts a fresh 3 s. YouTube and
Plex Web do the same: before, the row faded out from under a mouse parked on Next.
- **Text over the picture sits on a dark panel** (2026-11-07): loading, stalls, errors and the notices are
readable on any frame. A stall that lasts 5 s on a link measured slower than the stream says "Slow
connection" with both rates.
- **Volume is the native `Slider`** (A6); the hand-drawn bar had no hover, focus, keyboard or screen-reader
state. A pointer drag gives the focus back, so Space and the arrows keep driving the player.
- **The seek bar:** the track stays thin inside a 2.75rem (44px) hit area with `touch-action: none`, so a finger
@@ -718,155 +721,160 @@ of a full-width panel and nothing on the right.
## Decision log
| Date | Decision | Chosen | Over | Why |
| ---------- | ---------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| 2026-09-24 | Floating button | Back to top: a round 48 glass FAB; "round" = floats over something (media or the page) | A Material 3 rounded square; a labelled pill; the old 64 circle | The common web convention for back to top; round tells it apart from every toolbar button; 64 was oversized |
| 2026-09-24 | Device contract | Tablet portrait + landscape and up; phones out | A 1024px floor with a "too small" message | Portrait tablet use is real, and the app already renders at 820 portrait |
| 2026-09-24 | Card actions on touch | Fewer visible, the rest in a "…" menu | 44px buttons that no longer fit the card | Six 44px targets need more width than a card has |
| 2026-09-24 | Action placement | One menu with several triggers, never gesture-only, a desktop background menu that mirrors the toolbar | Visible buttons for everything | Fewer non-critical buttons on screen, without hiding any function |
| 2026-09-24 | Text selection | Text stays selectable; a click that ends a selection does nothing | Click-to-play on titles | The user must be able to copy any text the app shows |
| 2026-09-24 | A1 Button shape | r8 + a rung per context; glass is a Button variant | r12; a pill | Nested corners (r8 in r12 in r16); the pill already means "a filter or a choice" |
| 2026-09-24 | A2 Icon buttons | Two sizes, r8; a round button only on media | One size; all round | A lone toolbar icon at the row size looks lost; round rows read as tags |
| 2026-09-24 | A3 Touch size | 44 for every control primitive | 36 + gaps; Button only | One rule (HIG 44) instead of two sizes nobody remembers |
| 2026-09-24 | A4 Input | Bordered, a 55% fg-into-bg edge (≥ 3:1) | Filled; today's look as a primitive | The only option that clears WCAG 1.4.11 without a heavy fill on glass |
| 2026-09-24 | A5 Dropdown trigger | Pill in a toolbar, r8 in a form | Pill everywhere; r8 everywhere | Matches the neighbours in each context |
| 2026-09-24 | A6 Checkbox, switch, slider | Native, accent-tinted, the row is the target | Drawn primitives | Free keyboard / screen-reader / forced-colours support; only 3 sites |
| 2026-09-24 | B1 Density | Compact 32 / 28 / 24 by explicit height | Standard 38 / 32 / 26 | The owner's read of the lab; touch is covered separately (A3), and one toolbar row lines up at 28 |
| 2026-09-24 | B2 Scheme set | Midnight, Forest→cyan, Slate→fuchsia, Matrix→lime | Today's six | The accent collided with danger (YouTube), warn (Slate) and success (Forest, Matrix); Midnight and Starship were one scheme twice; 4 of 6 light accents failed 4.5:1 |
| 2026-09-24 | B3 State layers | Text colour at 10% / 16% over anything | Material's 8 / 12%; today's opacity + bg-card | Visible in every scheme and mode (≥ 1.2:1); today's row hover was 1.00:1 on a card |
| 2026-09-24 | B4 Page chrome | One sticky band + fade; filters fold when short | A floating glass bar; everything scrolls | Tabs and filters always in reach, in one place; a keyboard-up tablet keeps its content |
| 2026-09-24 | B5 Surfaces | One class + radius per role | Today's per-site mix | Nested corners and no per-site alpha drift |
| 2026-09-24 | B6 /watch | The app's surfaces, native video controls | Flat and minimal | The page strangers see should look like the product; native controls work on any device |
| 2026-09-24 | B7 Nav rail | Soft tint + edge bar, optional labels (a user setting), slim | Solid accent (recommended); a Material rail | The owner's combination: calmer than a solid fill, labels for who wants them, no width cost |
| 2026-09-24 | B8 Loading | Skeleton by shape, spinner in a control | A spinner everywhere | Nothing jumps when data lands; a long wait doesn't look stuck |
| 2026-09-24 | B9 Fade | Toolbars / headers, panel tabs, dialog edges, chip rows | Thumbnails; buttons | The north-star depth where content scrolls under chrome; not on content or small controls |
| 2026-09-24 | C1 Action classification | The table (lab C1): state + primary at rest, everyday toggles on hover, "…" on touch, destructive into the menu | Visible buttons per site, as today | One action list per object, nothing gesture-only, fewer buttons without losing a function |
| 2026-09-24 | C2 Side panels | A tap opens an overlay drawer; pin is a separate button ≥ 1024 | A tap pins and the content reflows | Below 768 the filters were unreachable; on touch a tap pinned and pushed the grid |
| 2026-09-24 | C3 Players on touch | Tap = show the controls; 44px seek hit area; 44 buttons; the YT description behind ⓘ | Today (tap also pauses; 7.5px seek bar) | Measured live: you couldn't reach the seek bar without pausing, and the hover-only description never opened on a tap |
| 2026-09-24 | C4 Dialogs | Focus by purpose; a pinned footer | Focus on Close; the footer scrolls | A form starts in its first field, and Save stays on screen with the keyboard up |
| 2026-09-24 | C5 Plex player controls | The owner's lean set: no Stop / Restart / ±10 s buttons, Back by position, Next only with a next item, Resume + From the start on the card, touch double-tap ±10 s | Every control in one row (it ran off-screen below 1024); Back by a press counter; ±10 s keyboard-only | The row fits a portrait tablet; the decision to restart is made on the card; a tablet keeps the 10 s skip |
| 2026-09-24 | Glass tuner | Retired; the schemes are the only colour choice | Keep it with ranges clamped to the rules | Free tuning is the drift the constitution exists to stop (owner) |
| 2026-09-24 | Matrix toned down | Moss #86a05a at Midnight's lightness, near-neutral greys; every white-on-accent part → `--accent-fg` | The B2 lime #a3e635 | The owner found the lime neon and tiring; it was the loudest accent (L 0.85, C 0.21), and fixed white parts sat on it at 1.5:1 |
| 2026-09-24 | Field primitives | `Input` / `Textarea` / `InnerInput` + `fieldClass`, `SelectTrigger`, the native controls; the lint rule "no raw `<input>`" | `inputCls` and ~40 hand-styled boxes | Six edges, four radii and two focus looks for one control; the ≥ 3:1 edge held nowhere |
| 2026-09-25 | Surface roles | `SURFACE` roles in `ui/surfaces.ts`; the lint rule "no bare glass class"; menus p-1 with r8 rows | Three tiers paired ad hoc with five radii and three menu paddings | ~140 sites drifted; nested corners only hold if the radius comes with the role |
| 2026-09-25 | Static labels | Flat tonal `Badge`, no edge; a filtering tag is an edged chip | A glass pill with an edge, like the chips | Edged = pressable at a glance; a static tag that looks like a chip gets clicked and does nothing |
| 2026-09-25 | Header search | A 40 glass pill with the field edge (`lg`); what sits inside it is round | 44, r16, a glass box with a faint edge | The one rounded box among pills, and its edge under 3:1; the hero search is the YouTube/Gmail pill |
| 2026-09-25 | Switch row (F12) | The whole settings row toggles (`SwitchRow`; a hand-laid row points a label at the switch) | Only the 36×24 switch toggles | The iOS / Material settings row; the text was a dead click beside the control it names |
| 2026-09-25 | Edges and hovers (F14) | The ≥ 3:1 edge on the field family (fields, the form trigger); labelled pills keep the faint edge and the accent-edge hover | The ≥ 3:1 edge and a state-layer hover on every control | 1.4.11 asks an edge only where nothing else identifies the control; a heavy edge on every chip and tab weighs the whole UI down |
| 2026-09-25 | Rating and live (F16) | Own tokens: `--rating` (theme-scoped gold), `--live`/`--live-fg`, `--rating-media`, `--media-chrome`; palette classes lint-refused | Map them to `--warn` / `--danger` | A gold star and a red LIVE are what the eye expects; as warn/danger they would read as a status. The old amber score was 2:1 on light |
| 2026-09-25 | Fjord trial (B2 exception) | A fifth scheme, temporary: Midnight's neutrals, Playnite's steel accent (#337599 lifted to #5aa1cb) and steel rules; exit = fold into Midnight or remove | Retune Midnight straight away; borrow only the qualities | The owner wants to live with the calmer palette app-wide before it replaces the default; it breaks "own hue family" on purpose, and only until that call |
| 2026-09-25 | Whole-card row (owner) | A card row's B3 layer is the card's `::before` under its content; the open button's hit layer spans the card, the actions sit above it | The name button's own hover, ending at the first badge or button | A row that half-answers the pointer reads as broken; a badge click did nothing |
| 2026-09-25 | Badge tones (F29) | `Badge tone` neutral / accent / warn at 6%; light `--warn` #b45309 → #a14a07; edged label spans lint-refused | Keep edged status pills (a private Badge in Users, Settings "Granted", the suggestions count) | Edged = pressable at a glance; amber-700 was 4.58:1 as text and fell under 4.5 on its own tint |
| 2026-09-26 | Page band built (B4) | Controls only in the band (intros scroll); the fold shows the same row in place; one `h1` per page (`PageHeading` steps the hidden one aside); `SectionLabel` headings; Skip to content | A popover of the folded filters; the page title kept visible | A folded row is still the same controls, nothing duplicated; the rail and the tab title already name the page |
| 2026-09-26 | Fade built (B9) | `fade-below` on sticky toolbars and table headers; the scroller mask where the band sits on the backdrop | A painted band over the backdrop | The backdrop is the colour under the band, so fading the content into it is the same edge |
| 2026-09-26 | Viewport units | `dvh` everywhere (the Plex player stage with S3f) | `vh` / `h-screen` | `100vh` hides the last rows under a tablet browser's toolbar |
| 2026-09-26 | Video actions built (C1) | One list per video (menu, right-click, long-press, "…"); the card keeps watched / save / add to playlist on its picture on hover; download, details, copy, hide, reset in the menu | The six-button row under the card and the hover ⓘ | The menu is the one door every device has; the card loses a line of height and the touch card its row of 28px buttons |
| 2026-09-26 | Plex actions built (C1) | One list per Plex title (play/open, watched, add to playlist, media info, copy title, pin); the poster keeps watched / add / "…" on its picture on hover; seasons get open / watched / add / copy | The hover ⓘ, the pin-only menu, an episode's add button at rest, a collection's "Remove" X on the card | The same door as a video's; the collection grid goes to two columns on a phone so the touch "…" leaves the title its line |
| 2026-09-26 | Playlist actions built (C1) | A playlist row's "…" (play, pin, copy title / link, move, order, remove), Delete + an undo toast; one list per playlist (rename, pin, YouTube open / push / revert, delete) from its rail row and its header's "…" | The row's X column and its confirm; the header's pencil, YouTube link, push, revert and trash buttons | Remove is reversible, so it needs no confirm; the header keeps Play all; a narrow table drops "#" and Channel so the title keeps a width |
| 2026-09-26 | Channel + Plex list actions (C1) | A channel row's "…" (open, pin, YouTube, copy, full history, hide + undo, unsubscribe); a Plex playlist's title / show / season "…", Delete + undo; its header's "…" (rename, delete); an action toast lasts ≥ 8 s | The channel row's full-history, hide and unsubscribe buttons; the Plex list's red X on every row, show and season; the header's pencil and trash | Hiding and removing are reversible, so they need no confirm and the undo needs time; unsubscribe and deleting a playlist keep theirs |
| 2026-09-26 | Download / notification / Scheduler actions (C1) | A download row's "…" (play, share, save, edit, details, re-download, manage sharing, copies, cancel / delete / remove); a library row's Share stays in sight, ahead of save; a notification row's "…" and hover dismiss; the inbox band's and the Scheduler strip's "…" | The download row's five rare buttons and its cancel / delete / remove; the source link's copy button; an activity notice's Find / Undo links; the band's Clear everything; the strip's Purge discovery | The destructive rows keep their confirm, and Clear everything gains one; a row link keeps the browser's own menu; Find in feed drops the tag / search / age filters that could hide its video |
| 2026-09-26 | Channel menu: Tags + Priority (C1) | Tags A–Z as in-place checks; Priority as fixed steps 10 · 5 · 3 · 2 · 1 · 0, the current one ticked; a pick re-sorts at once and the page + focus follow the row; one drawn `MenuCheck` for every checkable row that cannot be an `<input>` | Relative steps (+1 / +5); straight to the top; the stepper only; seven hand-drawn check marks (three box sizes, a tinted box, a × for "on", a trailing tick) | Bubbling a channel up is one pick; fixed levels keep the scale the reader already uses; the list moving where the reader can see it is the feedback |
| 2026-09-26 | Reorder without a drag (WCAG 2.5.7) | Row menus: Top / Up / Down / Bottom (YouTube + Plex playlists); saved views: four buttons with the place beside the selected view (owner: A); side panels in edit mode: an Up / Down pair per group (owner: A), the name wrapping there | A "…" on every saved-view row; a "…" on every panel group; drag only | A visible, one-click path where the list is edited; the edge moves shown off so the buttons never shift |
| 2026-09-26 | Titles in click surfaces | The surface stays a button, flagged `selectable`: its text selects, and the click that ends a selection does nothing; every plain `Pressable` pinned with its reason | Rebuilding each card and row so the title sits outside the button; exceptions for master-detail lists | Chrome selects inside a button once it is `user-select: text` (measured), so one flag on the primitive keeps the keyboard, focus and list walking every surface already has |
| 2026-09-26 | Right-click on a selection | The browser's own menu opens on selected words; beside them the object's menu | A "Copy selection" row in the app menu | The reader selected the words to do something with them; the native menu copies, searches and translates, and changes nothing else |
| 2026-09-26 | Keyboard ring on a flush row | Inside the row (a table row by its tag, a log line, a show header in its clipping card, with the card's corners) | The ring outside, as on a card | Outside it ran under the table head and past the table, and a clipping card cut it away entirely (invisible focus) |
| 2026-09-26 | Fade at rest | The sticky header's fade appears with the first 1rem of scroll (a scroll-driven animation; no fade where unsupported) | A fade that is always drawn | At rest it dimmed the first row and hid its ring's top edge; nothing is under the header yet |
| 2026-09-26 | Narrow playlist table | "#" and Channel hide by the table's own width (under 42rem), not the window's | Hiding below the `sm` window width | Beside the playlist rail a desktop window left the table 432 px and the title 0 px |
| 2026-09-26 | Background menu built | Five list pages (feed, Plex grid, channels, playlist, notifications); the toolbar's sort / view / filters and the band's actions; no destructive row | Every page with a toolbar; the "…" menus copied whole, delete included | The tab-only pages have nothing a menu would add; a blank-area right-click must not be one click from delete |
| 2026-09-26 | Media card anatomy (C) | One frame; the meta line pinned to its bottom; the length a badge on the picture; the Plex card on the same primitive as the video card | A: the year as an overline over the title; B: the year inline after it; a fixed two-line title slot | A one-line title left a hole the eye reads as a missing line; C puts nothing under a variable-height title |
| 2026-09-26 | One Play | `MediaPlayGlyph` (the media button variant) and only where the click plays | The Plex card's hand-painted accent circle; a playable dot on every card | One Play for every card; a dot that only a colour and a tooltip explain says nothing |
| 2026-09-26 | Filtering values on the info page | A value that filters (year, certificate, studio, director) is an edged chip; a fact is flat text | Everything flat, the clickable ones only hover-tinted | The S3d rule: edged = pressable, so the reader sees which values do something |
| 2026-09-26 | Shelf density | On any strip: one title line over one meta line | The grid's two lines and full meta on a strip | The tallest card set the strip's height and left holes in the cards with less to say |
| 2026-09-26 | Compact list row (B) | The title leads; a faint "…" at the end; the toggles float over the date on hover | A leading 128px action column | That column sat empty at rest, a hole at the start of every row |
| 2026-09-26 | Media-info layout (B) | A main column (collections first, crew, cast) beside a facts sidebar; an empty column is not drawn | Every section full width, stacked | The collection is the most used; the keywords' column left a void beside a short details box |
| 2026-09-26 | Mixed Home shelf (A) | Plex titles as wide art beside videos | Posters, with the videos not stretched | Stretching a video card to a poster's height left an ~80px gap |
| 2026-09-27 | Home hero | A card (hover row + Play on the picture, the whole menu, a touch "…"; owner: A); three quarters wide and centred on a wide window; one fact line; the summary at the measure | The hero's pin row alone; full width; an 18rem picture column | The one answer stands in focus, and it offers what the same card offers anywhere |
| 2026-09-27 | Home shelf tiles | A channel / playlist tile: Play, ↗ open, "…" with Open; a show acts on its next episode, ⓘ opens it; the shelf slot 232px | The tile's pin row alone; 200px | A tile without its object's actions read as broken; at 200 the Play touched the hover row |
| 2026-09-27 | Synopsis preview (C) | A ~0.9 s mouse rest opens a passive panel beside the card | The synopsis as the picture's `title` tooltip; an ⓘ to hover; none | One long tooltip line fired anywhere on the art and clashed with the cut title's own tooltip |
| 2026-09-27 | Reading measure | Prose capped at ~65 characters a line (`measure`) | The container's full width | At 110–130 characters a line the eye loses the next line's start |
| 2026-09-24 | Chip family | One `Chip` (filter / token / tag / suggest / ghost) + `DismissibleChip`; nothing under 24; the lint rule "no raw `<button>` outside ui/" | Five chip geometries in five files | The applied-filter row, the tag cells, the year picker and the saved-view editor each drew their own chip; a one-letter tag was 21px wide |
| 2026-09-24 | Button families | `Button` / `IconButton` / `Pressable` / chips / `DragHandle`; `lg` 40 for a hero CTA; pills and a 48 Play on media (from C5) | Every button hand-painted | 268 hand-rolled buttons in 79 files drifted into as many looks; the ladder only holds if every button stands on it |
| 2026-09-27 | C4 as built | `initialFocus` required on every `Modal`; the commit row in `ModalFooter`, portalled under the body from any depth; the fades driven by the body's scroll timeline; `interactive-widget=resizes-content` | An optional prop defaulting to the ✕; a `footer` prop (the issue composer's buttons live three components down); always-on fades | A default is how 21 dialogs came to open on the ✕; lifting the composer's state to reach a prop is the larger change; a resting fade dimmed the first line |
| 2026-09-27 | C3 + C5 as built | A tap toggles the bar through `pointerType` from `pointerdown`; a double-tap takes back its first tap's toggle; the card's pair is the video card's `MediaResumeButtons` (Continue / Restart), a sibling over the picture; the YT description is an ⓘ click popover | A delayed single tap (wait out the double-tap window); new "Resume / From the start" words for Plex; the pair inside the picture's button | A delayed tap feels dead; one media card says one pair of words; a button inside a button is invalid and the hover row already sits beside the picture |
| 2026-09-27 | C2 as built | The side panel's drawer is `ui/Modal` `drawer-start`; drawer = a coarse pointer or below 768; the pin offered in its header ≥ 1024 | A drawer of its own inside `SidePanel` | Esc, Back, the scrim, the trap and the scroll lock are the dialog's; a second implementation is how two shells drift |
| 2026-09-30 | Composition and the reader (E30 S6) | A chapter and a Conformance step: the space is used, one focal point, worst real data, where am I, a click answers at once, zoom and text size first-class; the person page as two columns (facts and bio left, "Known for" right) | Rules for pieces only; the bio in a full-width panel; a narrower panel | The person page passed every piece rule with half its panel empty; the owner reads zoomed in, so zoom is a requirement, not a nicety |
| 2026-09-30 | Hungarian words (E30 S5) | The D3 list: Feed, Shorts, Admin, Token, Port stay; stream, Prio, Job, Audit, Demo, Auto, Info, Frontend/Backend, Backfill, transcode get the Hungarian word; one name per thing (kérelem, behúzás, one "Test connection") | Keep every loanword; translate Feed/Shorts too | The owner chose the list (D3); the final sweep found the same feature under two names in three places |
| 2026-09-29 | Numbers and order in the UI language (E30 S3) | Every number through `lib/format.ts` in the UI language (decimal mark, grouping, compact, units); A→Z by the shown name in the UI language's alphabet, client and server alike; the language comes off `useTranslation`, never the singleton | `toFixed`, hand-built "1.2K", `localeCompare` in the browser's locale, byte-order server sorts | A Hungarian reader saw "8.3", "1.2K", "Életrajzi" after "Z"; the singleton read also left values in the old language after a switch (React Compiler) |
| 2026-09-27 | Narrow rows by container | The download row, the playlist table and the Channels tag row stack by a container query; a fact line never wraps | Breakpoints on the window | The iPad zoomed in, split screen and a column beside a rail are narrow while the window is not; the zoomed download row broke "18.9 / MB" apart |
| 2026-09-27 | Phones deferred | A phone-only layout (the hamburger nav, a folded feed band) waits for its own epic; E28 stays desktop + tablet | Fix the phone views inside E28 | Principle 2; the first phone attempts made the feed worse, and the epic's review grows with every sprint |
| 2026-09-27 | Conformance check (S3g) | A four-step check before a UI change is done: primitive at a named rung, the chapter re-read, the rendered page measured by kind and look, every state seen | A reviewer's eye per sprint | The S1 probe found 69 button looks; after S3 every control kind is drawn by its primitive, and the check keeps a new look from arriving unnoticed |
| 2026-09-27 | Segmented rung | `SegmentedToggle` defaults to the toolbar rung (28, `text-xs`); 24 only for a panel's captioned match toggle | The small default | Two toolbars had inherited 10px type from the default beside 28px buttons |
| 2026-09-27 | Copy rule (S3g) | One name per feature (the rail may shorten it); four verbs for four meanings; sentence case with the app's places capitalised; "…" when a command asks first; one-sentence messages without a stop; curly quotes; dates relative in lists with the exact minute on hover, three absolute forms; units from Intl, short | Word-for-word names everywhere; the verbs as they were; all lower case; no "…"; always a stop; always absolute dates; long or compact units | The copy survey found 7 EN and 6 HU double names, 22 Delete against 27 Remove, five date shapes and three spellings of a minute; the owner picked every recommended option |
| 2026-09-27 | One glyph, one meaning (S4) | Play only for playback or resuming a paused transfer; Retry = `RefreshCw`; reset / revert = `RotateCcw` | Retry drawn with Play | The first UI-review run (Downloads) found a failed row whose Retry showed ▷ in the row and its menu |
| 2026-09-27 | Fjord kept | Fjord stays a fifth scheme beside Midnight; the trial wording goes | Fold it into Midnight; remove it | The owner lived with it and likes it; the vivid Midnight stays for those who want it. B2's own-hue rule still binds any new scheme |
| 2026-09-27 | Tablet columns | A table hides `tabletHidden` columns below 64rem of its own width; Channels drops Prio, About, Length | Hide About only; give the name a bigger share | At 1024 the channel name had 60 px; the owner chose to drop all three there |
| 2026-09-27 | Glyph pager | First / previous / next / last as glyph buttons round the page box, names as tooltips | Words beside the arrows (hidden only on a phone) | The owner's call: the words widened the row; four arrows are the shared convention |
| 2026-09-27 | Match switch on the panel | `ui/MatchToggle` above the picker from two picks, everywhere; one wording (Any / All) | The picker's popover footer (Feed, Plex collection) | A closed picker hid an active "all"; six pickers had drawn it two ways |
| 2026-09-27 | Missing picture, bare link | A glyph of the kind for no picture; a download with no name is titled by its short link, once | An empty box; the URL as title, channel and link | The owner's picks 4A and 3A: an empty box read as loading, the URL showed three times |
| 2026-09-27 | Copy on a selection | Right-click on selected words keeps the browser's menu; no app Copy row | An app menu with Copy first | A real browser's menu has copy, search and translate; only the desktop pane's is thin |
| 2026-10-01 | Zoom target 125–133% | Screens are held at 125% and 133% browser zoom (1536 × 729, 1444 × 685 on a 1080p window) and the largest text size | 200% browser zoom | The owner: 200% was a guess; 125–133% is what they read Siftlode and most sites at, so 200% would be an upper bound nobody uses |
| 2026-10-01 | E31 Nav rail at short heights | One "Admin" rail item opening a panel of the admin pages | A denser rail when short; a visible "+4" hint | The rail fits at 133–150% without a hidden scroll; the admin pages are rare and admin-only |
| 2026-10-01 | E31 Configuration sections | A vertical, grouped section list, as Settings has | Group tabs + sub-tabs; one long page with a contents list | Sixteen tabs wrapped to three rows at 133%; one pattern for both settings pages |
| 2026-10-01 | E31 Statistics overview | Number tiles, and the per-operation counts as horizontal bars | Three narrow cards; a 30-day chart | The label sits by its value; what spends the quota is seen at a glance |
| 2026-10-01 | E31 Channel page header | The full banner, folding to one row once the grid scrolls | Always compact; no banner | The header is fixed and took 40% of a 685px window; the channel keeps its face at first sight |
| 2026-10-01 | E31 Channel page watch state | A fresh channel page shows the unwatched videos; a picked state is remembered | Show all, forgotten on reload | The owner: open a channel to see what is new; a choice made there should stick |
| 2026-10-01 | E31 Audit log runs | A job's successful runs fold into one row that opens, whatever ran between; only that job's failure (never folded, danger colour) breaks it | A filter defaulting to admin actions; both | Nothing hidden, the noise collapsed, the failure seen |
| 2026-10-01 | E31 Users row | Role and state as badges, the actions in the row's one menu | An inline role toggle; icon-only buttons | Action placement: the row shows state, rare and destructive actions live in the menu |
| 2026-10-01 | E31 YouTube player | Video beside an info panel at 1280px and wider (the video as large as the height allows); stacked below | Today's stack with a one-line bar; full-window like Plex | The owner's eyesight: about 1460px of video on 1080p instead of 1008, the actions still visible |
| 2026-10-01 | E31 Plex info and show pages | A second column for the facts, keywords and crew, as the person page; a collection's hero sized to its content | Full-width text; a larger poster and fanart | The right ~40% was empty; one composition for every media page |
| 2026-10-01 | E31 Nav rail rows are links | Each page row is a real link: a middle-click or Ctrl-click opens the page in a new tab | A button | The owner: open a page in a tab of its own, as on any site |
| 2026-10-02 | E31 Plex detail second column (refined) | The facts over the key crew beside the text; the keywords, chips that filter like the genres, under the text; the show page gets all three too | All three in the column, as the mockup drew it | At 20 keywords and 12 crew the column ran to twice the text's height (849 vs ~400 px at 133%): a hole under the text instead of beside it |
| 2026-10-02 | E31 Plex detail poster and crew | The movie and show poster keeps its 2:3 and grows a rung in a wide hero (w-56 at a 64rem panel); the crew shows its first 6 names and "Show N more" (a list of 8 or fewer shows whole) | Poster stretched to the row; a "More" on the keywords | The row is as tall as the crew column, so the poster drew ~175×590 and cropped; the crew, not the keywords, set the hero's height |
| 2026-10-04 | E32 `field` focus skips a mode switch | A form opens on its first field; a control above it that is not a field (the composer's Edit / Preview) is passed over by marking the field `data-initial-focus` | The first text field everywhere | That moved the focus in the download dialog off its profile choice and in the upload dialog off its drop zone; the mark changes only the form that needs it |
| 2026-10-04 | E35 Lint for z-levels and edged labels | Both rules stay; the two eslint selectors (`Z_RAW`, `EDGED_LABEL`) that refused them go | Keep the selectors | 0 hits since the E28 sweep that fixed the sites: a check that never fires costs a probe test and a snapshot each, and the UI review holds both rules |
| 2026-10-06 | `player` focus | A dialog that is a player opens focused on its panel (the YouTube player's rule, now `Modal`'s too); on the media element itself, Space and the arrows stay the element's own | `close`, or the media element | On the ✕ Space closed the player; on the element the browser's own controls acted on the same key as the player's (one Space played and paused, measured in Chrome) |
| 2026-10-08 | Captions on touch | A `Tooltip` opens for a mouse and a key's focus only, never for a finger | A caption on tap; a long-press caption | On an iPad the tap that showed the caption was taken as a hover and lost its click (every rail button took two taps), and no finger sends the mouseleave that hides it; the rail's labels setting names the buttons on touch |
| 2026-10-08 | Header search on touch | No key cap on a touch screen; the pill grows to 52 px so a ring shows round the 44 px Go; Go drops to its icon by the column's width | Keep the key cap; a 44 px Go overhanging a 40 px pill | On a tablet the 44 px key cap and Go sat on the pill's edge and cut the placeholder; an overhanging Go breaks the nested radii |
| 2026-10-08 | Plex subtitles and the control bar | The player draws the cues itself; while the bar shows they rise above it (200 ms), then settle back | The browser's own cues, moved by `line` | On an iPad the bar covered the text; a native cue cannot be animated or measured, and our layer behaves the same on every engine |
| 2026-10-08 | Hover on a button over a picture | The `media` variant's hover is a 2px accent ring over a darker body (Plex + YT cards, rail arrows; option B) | An accent fill (reads as "selected"); a white wash (weak on a bright poster) | The 60 → 75% black step was invisible, and over a card a rail arrow could not be told from the card's own Play under it |
| 2026-10-08 | Plex "Add to collection" | A movie or show card's menu row opens a panel beside the card: its library's manual collections tick in place, a name + Enter creates one with the title in it (admin; option A) | An Add submenu with a "New…" dialog; the per-title editor modal from the menu | A collection took a trip to the Collections page, a name, a title search, then a second search per title; the panel is the video card's "Add to playlist" |
| 2026-10-08 | YouTube modal chrome | No prev/next strips beside the card; Close an icon (the panel's top right; stacked, the title row's end); the description clamped to three lines + More; the panel shuts to a 2.75rem strip of close, expand, prev/next (option D1) | Keep the strips; a "Close" text button; the full description in a scroller; a panel that hides outright | The strips cost the video 7rem of width for a second copy of the panel's buttons; a long description pushed the channel and the actions out of sight; a hidden panel left no way to step the queue without the keyboard |
| 2026-10-08 | The wheel over the YouTube modal | Over the video: volume. Over a panel with something to scroll: scroll only, never volume, not even at its end. Over a panel with nothing to scroll: volume (owner) | Volume everywhere; a scroll that hands over to the volume at its end | Reading a long description turned the volume down once it hit the end |
| 2026-10-09 | A Plex show card plays | Its picture plays the show's next episode everywhere (an episode with a resume point → Continue + Restart, else the first unwatched, else S1·E1), resolved on the click; the preview names it once up; title and ⓘ open the show (option A + hover ask) | The picture opens the show page (as before); every card carrying `next` | A show card on Discover had no Play while Home's had one; carrying `next` costs a grid page up to 180 queries for the few cards anyone plays |
| 2026-10-10 | Clip mode in the players (E38) | Scissors in the control row + `C`, `[` `]`; the panel in the player with the watched video as preview; overview + zoomed strip on `ui/RangeTrack` (ends are sliders); a toast + the Ready notification after Cut | A ⋯ menu entry; permanent In / Out buttons; a big dialog with its own preview (a second Plex transcode / YouTube iframe) | Owner's picks A, P1, a; a film is hours and a clip seconds, so only a zoomed strip gives a hand the precision. The Downloads editor's cuts moved onto the same track (keyboard-steppable now) |
| 2026-10-10 | Crop in clip mode (E38) | A Crop toggle in the clip panel: a frame over the playing picture (`ui/CropFrame`, the editor's crop too), Free / 16:9 / 9:16 / 1:1, Whole picture; maximised YouTube keeps the panel in the bar's place; the wide side panel shuts while clipping | A frame always on the picture in clip mode (K2); a free shape only | Owner's K1 + a: a crop is rare, so it never covers the picture unasked; a phone or square clip needs a held shape. The iframe hides its picture's size: drawn on the 16:9 box, the cut fits it to the real picture |
| 2026-10-10 | Marks on the player chrome | `--media-mark` (amber) and `--media-danger` (light red), the same in both themes; `media-marks` points the focus ring (`--focus`) and the chosen mark (`--chosen`: a track's tick, a tab's underline; the label stays white) at the amber on the player and every chrome surface; `media-scope` (a form) also the errors at the red | The theme's accent ring, chosen text, warn and danger | On the chrome (#171717) a light theme's accent ring (Slate 2.84:1), warn (2.99:1) and danger (2.77:1) all fall under 3:1; the accent as chosen text under 4.5:1 |
| 2026-10-11 | Files page (one library) | List rows by default, the media grid a toggle remembered per folder; a sidebar of automatic source folders, own folders, Shared with me and the storage meter; the paste-a-link box in the toolbar | A media grid by default | The library holds videos AND plain files: rows read both, show more per screen and scan by size and date |
| 2026-10-11 | Library list columns | Every kind of value in its own column (name, source, phone-safe, retention, size, added, actions), one grid with a subgrid per row so they line up; an action that does not apply keeps its slot empty; a right-click menu on every row (the ⋯ opens the same menu) | Badges flowing after the title; icons that shift when one is missing | Owner: values of one kind must fall in one vertical line; a row's every action reachable by right click |
| 2026-10-11 | Running downloads | An app-wide tray, bottom right, shown only while something runs or failed (reason + retry), foldable to a pill | A rail icon with a count and a popover | Downloading became an action you start anywhere; a running one must be visible without going to look |
| 2026-10-11 | Download done | The notification says whether it plays on a phone and carries Link másolása + Megnyitás (the folder, the item focused); never an automatic jump | A plain "ready" notice; auto-navigation to the item | The daily path is paste → done → copy the link; a jump would pull the user out of what they were doing |
| 2026-10-11 | Library row quick icons (owner A) | On a wide list only the ⋯ holds a place; Copy link, Share and Save come up over the date while the row is hovered or holds the keyboard; a card (phone) keeps them in its own line; Keeps narrower | Three icons always in sight (the name kept ~150px at 1200); dropping Added below 64rem | Owner: the name is what a row is read by; every action stays in the ⋯ and the right-click menu |
| 2026-10-12 | Files preview (owner ★) | A file that is not media opens from its name in the dialog shell, wide: an image, a PDF in the browser's own viewer, text as numbered plain text (HTML and SVG too) with a wrap toggle, else a no-preview card with Save; ← / → page the list's such files in its order; an image under 8 MB is its own thumbnail | A side panel beside the list; a new browser tab; pdf.js; syntax highlighting; a click that saves a file without a preview | Owner: one way to open anything, paged like the list; the browser's PDF viewer is free and isolated; a page is never rendered on our origin |
| 2026-10-12 | A running download's step (owner ★) | The tray row names the step beside its bar ("Downloading audio · 12%", "Merging" with a pulse); a paused-only tray shows a pause mark, never the done check | A bare bar and % | A download fetches video and audio as two streams: unnamed, the second fill reads as a restart; the tray is the only place a download is followed |
| 2026-10-14 | API tokens (owner A/A/A) | Settings → API tokens: a row per token in three lines (name + prefix, its folders and rights, last use and expiry), revoked from its ⋯; a new token asks the kind first (Hub exchange or Custom: rights as chips per folder), expiry 1 year by default; the token shows once in the same dialog, with Copy and a curl line | A table; one form with a preset link; the token in a banner above the list | A table truncates the rights at this width; the common token never shows a matrix to get wrong; closing the dialog is plainly when the token is gone |
| 2026-10-15 | Folder links (owner A/A/A) | An own folder's menu has Copy link and Share with link… as a file row does (the dialog is the link half of the file's); the link opens a list of the folder's files as it is now, Play opening the watch player above it with the playing row lit, a document only downloading; the rail marks a folder with a live link beside its count | Share with link… alone; cards; a download-only list; no mark | Mostly videos from a phone, read on a phone: a list gives a long name its line and the next one is a tap away; the folder link is the one place where moving a file in publishes it, so where it is must show |
| 2026-10-17 | Copy-only changes (owner ★) | Step 4 of the Conformance check alone (not 1–3 or 5): every state the changed text renders in, seen in both languages | The whole five-step check; no screenshot at all (DOM test + key parity) | Steps 1–3 and 5 judge controls, patterns and composition, which a text change does not touch; the one risk left is a longer Hungarian string wrapping or clipping, and only a screenshot shows that |
| 2026-10-17 | Text on glass (owner ★) | Light `--muted` 15% toward `--fg` in every scheme (≥ 6:1 on the bare page); the item count, the badges and the current Files row on opaque tints mixed into the surface | Only the two failing components; the item count's number in `--fg` | Pixel-measured over every scheme × mode × backdrop, the count fell to 3.8:1 (dark too: thumbnails scroll under the toolbar's glass) and Channels' tabs to 4.1:1; an `--fg` count outshouts the filters beside it |
| 2026-10-18 | Scheduler run history (owner D4, E1, E2 ★) | A row's last three runs as dots, newest first (interrupted: a hollow warn ring, set apart from skipped and failed), and a history button that opens the job's runs on the Activity page; a user's ad-hoc runs in their own block under the jobs, with no interval and no run button | An inline expandable history per row; a dark-grey interrupted dot; ad-hoc runs only on the Activity page | Three dots answer "did it run, did it work" at a glance; the whole history needs filters, which the Activity page has |
| 2026-10-18 | Activity page (owner D1, D2, G1–G3 ★) | One page, "Napló" / "Activity": four tabs (Runs, Admin actions, Server log, Quota) under one filter bar (text, a period of 1 h · 24 h · 7 d · 30 d · 90 d defaulting to 7 days, an account), the view in the URL; runs and admin actions as dense rows that open in place as a log line does, filtered and paged on the server; a run's or a request's log lines one pivot away, as a history entry | One merged timeline; a page per source; a default period per tab; card rows; the audit log's client-side table | Each source keeps its own table and retention, read in one place; one bar keeps what you narrowed across tabs; about 1.8k runs a day need a log's density and a server-side filter |
| 2026-10-24 | Related videos (owner U1–U3 ★, D1 A, D3) | On demand: a tab beside the description (wide), a drawer (stacked); the narrow thumbnailed row, its menu alone, a followed check; plays in place and the list stays the opened video's; an on-topic / Hungarian-trending switch kept on the account | Always shown; the end-of-video grid alone; a dialog; a big card or a text row; a list that follows the playing video | The modal is full on a small screen; a picture is what a choice among strangers rests on; a list that stays put is tried one by one, and spends no more of the day's cap |
| 2026-10-25 | Related: one played; off (owner D1 a, D2 a) | Played from the list, it is the player's item: its title, controls and progress, marked watched at its end, the queue left where it was; a daily limit of 0 takes the tab and the button away | A linked video's in-place play (no controls, nothing kept); a tab that only shows a refusal | What is watched should count as watched, wherever it was started from; a control that can only be refused is not offered |
| 2026-10-25 | Player panel glyphs (owner 1A, 2A, 3) | In the side panel the queue row is glyph-only on one line, and Mark watched is a glyph toggle among the actions (✓ / ✓✓); the docked bar keeps its labels; the related switch at the default size, "Trending" | Labelled buttons wrapping onto three lines; glyphs on the docked bar too; Mark watched on its own line | The panel is narrow and the list controls are the same few every time; each glyph keeps its full name and shortcut in its tooltip; the docked bar has the width for words |
| 2026-10-25 | Auto-advance glyphs (owner E A, H a) | Auto-advance wears list glyphs (`ListEnd` next and off, `ListStart` previous, `Shuffle` random), in the panel and the docked bar; the auto-advance and loop cycles are `lit`, not `pressed` | The skip arrows for the mode; an X'd list for off; `aria-pressed` on a cycle | Beside Previous / Next the skip arrows read as a second Next; a cycle of three or four modes is no two-state switch, and its name already says which |
| 2026-10-25 | Restart and Reset glyphs (owner L A) | A menu's Play from the beginning wears `RotateCcw`, as the card's own Restart does; Reset (clear progress and status) wears `Eraser`; Redownload wears `RefreshCw`, the retry arrow | The skip arrow ⏮ on the menu's Restart; `RotateCcw` on both Restart and Reset; `TimerReset` for Restart | ⏮ is Previous beside the player; one action wore two glyphs, and two actions in one menu would have worn one |
| 2026-10-28 | Notification order (owner A-D1 a, A-D2 a, ★ a) | What asks for the reader first, from either source ("Needs you · n"), then one stream newest first; read rows older than a week fold under "Earlier (n)", closed, its label in the section-heading style with a chevron | One stream with no band; this browser's activity above the server's (the two fixed sections); a server-side retention that deletes old rows | An admin's access requests sat under up to a hundred server rows; week-old read notices crowded what is new; folding keeps the rows (nothing is deleted) and an unread row never folds, whatever its age |
| 2026-11-01 | Docked bar glyphs (owner D8 a) | The maximised view's docked bar is glyph-only too (tooltips carry the names and keys): transport, auto-advance, loop, Mark watched (✓ / ✓✓); its controls scroll on a narrow window while Fullscreen and Exit stay pinned at its right end | Labels from `xl` up | On a tablet the labelled bar ran past the right edge and took Exit with it; the side panel had already lost its labels |
| 2026-11-01 | Toolbar group glyphs (owner D9 a, D10 a, D11 a) | The feed's, channel page's and Plex Browse's filter pills name their group by a glyph (`Eye` watch state, `Shapes` content types, `ArrowDownUp` sort), the word in the tooltip and the accessible name; Sort + View end the toolbar's row, the chip row shows only with a chip; a channel page has no count pill | The value alone; Sort + View on a row of their own | "Watch state:" and "Sort:" cost a tablet row its room; the value says enough once the glyph names the group; the channel header already counts the videos |
| 2026-11-01 | Channel header on a tablet (owner E2) | Below `xl` the action row drops its two words: "Prio" (the stepper's tooltip names it) and "Subscribed" (a ✓ glyph; its name in the aria-label, Unsubscribe in the tooltip); the tags keep the row's rest, folding into "+N" | The tag lane on a full line of its own below ~12rem; the tags under the channel name | On an iPad (820 px, 44 px targets) the lane got 29 px, a sliver; now 130 px, and every tag shows at 1180 |
| 2026-11-01 | Clear watch progress (owner E a) | A video menu's Reset reads "Clear watch progress" in English (HU "Megtekintés visszaállítása" stays); "Reset view" is only a page's view reset (sort, filters, grouping) | "Reset view" for both; the playlist page's view reset renamed instead | On a playlist page the two stood one right-click apart under one name, and one of them erases the video's progress with no undo |
| 2026-11-01 | Clear watch progress clears only that (owner A) | The row shows only on a video with a resume point and clears that alone: a hidden video stays hidden, a watched one watched. Un-hiding and un-watching have rows of their own | The old Reset, which also un-hid and un-watched | A hidden video stays hidden: what the row's name says is all it does |
| 2026-11-01 | Selecting a list's rows (owner D2 a) | An always visible box first in the row and one over the column for the rows shown; Shift+click a range, `x` the row under the keyboard, Esc clears; the selected row wears `accent/12` with its tick. Rows a filter hides stay selected, an action covers those shown; Watch later offers no selecting. | A "Select" button that shows the boxes on demand | Found at once and works on touch, with no mode to leave; costs about 2 rem a row |
| 2026-11-01 | Selection band (owner D1 a, D3 a, D4 a) | The toolbar's second row while rows are selected, no floating bar: move (the last target as one click) and remove, no copy; both undo back to the places, a video the target held before stays there; a single row's move undoes too. | A floating bar at the bottom; copying to a list too; no undo on a move | It covers no row, and the player and the notices live at the bottom; twenty videos sent to the wrong list need a way back to their places |
| 2026-11-01 | Grip on a phone (owner A) | Below a 24rem table a playlist row has no grip: the title keeps its width, and the row menu keeps the moves. | The grip and its drag kept, the title at 44 px | The title gets its width back, and the rail is shut on a phone, so a drag had nowhere to go |
| 2026-11-01 | Selection band on a tablet (owner A) | Below `xl` the band is glyphs on one row, the names in tooltip and aria-label, as on the channel header; the replaced row stays in the same grid cell, invisible, so the first tick moves no row under the second tap. | The buttons in words, wrapping to a second row | The first tick pushed the list from under the second tap |
| 2026-11-01 | Playlist heading on a narrow window (owner A) | The list's name had no width left: below `xl` Unsynced is its dot (the word read out, the hint in the tooltip); on a phone the cover is gone and Play all is its glyph, named in the tooltip. | Every word kept, the list's name at no width | A heading without its name says nothing |
| 2026-11-02 | Dragging onto a list (owner D5 a, D6 a) | The grip carries a row, or every selected row with a selected one, onto a list in the rail; the toast undoes. Many dropped inside the list stay; a coarse pointer's drawer never peeks, so its rail is a target only when pinned. | Only the grabbed row going along; the grip off under a view | One row moving out of three lit ones misleads, and a filter is how the rows to move are found |
| 2026-11-02 | Delete with every tick hidden (owner a) | It removes nothing and a warning says the filter hides them | Taking the row under the keyboard | While the band shows a selection, Delete means that selection: a row no one ticked must not go |
| 2026-11-02 | Undo of a video gone further (owner a) | A video the move put in the target comes back only if the target still holds it, and a warning says how many | Everything back, whatever it did since | A move skips a video that left the source; the same rule back keeps a video out of lists the reader never put it in |
| 2026-11-04 | Tag chip counts (owner D5a) | A Channels tag chip shows how many followed channels it holds: a muted number after the name, no badge | The number in brackets after the name | The owner's pick; the number is a detail of the name, so it takes the name's own text, muted, rather than a `CountBadge` |
| 2026-11-05 | Picked tag chips lead the row (owner c) | On the Channels tag row the picked chips come first, in the tags' order, then the rest | Keep the order and fold the unpicked first; mark the "+N" when it holds a pick | The row folds from its end, so a picked tag late in the list sat only inside "+N" with nothing on the row saying a filter was on |
| 2026-11-05 | A toggled tag chip keeps the focus (owner a) | A chip that moves when toggled takes the keyboard focus with it: to its new place, or to "+N" when it went into the closed fold | Keep a picked chip in the "+N" popover until it closes | The row always shows the filters as they are; the focus, not the order, was what got lost |
| 2026-11-06 | Tag chip count after a dot (owner a) | The count keeps the chip's own colour, set off from the name by a "·" | The number dimmed (it fell under the 4.5:1 text floor, resting and pressed); the number in a tinted inner badge | Contrast holds in every scheme without a new token, and the dot still reads the number as a detail of the name |
| 2026-11-06 | A pick in the open fold closes it (owner a) | Toggling a chip inside the "+N" popover closes the popover; the focus goes with the chip | Keep the popover open and focus the chip now in the picked one's place | The focus never sits outside a popover left open, and a picked chip leads the row anyway |
| 2026-11-06 | An expired share link offers no Copy (owner a) | An expired row in the share dialog has a warn badge "expired <date>" and its address struck through; Open and Revoke stay, Copy goes | The warn badge alone, Copy kept | An expired link opens nothing, so it is no link to hand out; Open still lets the owner check it |
| 2026-11-06 | The shared page's unlock button is a noun in Hungarian (owner a) | "Megtekintés" for a video, "Megnyitás" for a file or folder | Both verbs, "Megnézem" / "Megnyitom" | The app's Hungarian buttons are nouns; one button should not change grammatical form by link kind |
| Date | Decision | Chosen | Over | Why |
| ---------- | ----------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| 2026-09-24 | Floating button | Back to top: a round 48 glass FAB; "round" = floats over something (media or the page) | A Material 3 rounded square; a labelled pill; the old 64 circle | The common web convention for back to top; round tells it apart from every toolbar button; 64 was oversized |
| 2026-09-24 | Device contract | Tablet portrait + landscape and up; phones out | A 1024px floor with a "too small" message | Portrait tablet use is real, and the app already renders at 820 portrait |
| 2026-09-24 | Card actions on touch | Fewer visible, the rest in a "…" menu | 44px buttons that no longer fit the card | Six 44px targets need more width than a card has |
| 2026-09-24 | Action placement | One menu with several triggers, never gesture-only, a desktop background menu that mirrors the toolbar | Visible buttons for everything | Fewer non-critical buttons on screen, without hiding any function |
| 2026-09-24 | Text selection | Text stays selectable; a click that ends a selection does nothing | Click-to-play on titles | The user must be able to copy any text the app shows |
| 2026-09-24 | A1 Button shape | r8 + a rung per context; glass is a Button variant | r12; a pill | Nested corners (r8 in r12 in r16); the pill already means "a filter or a choice" |
| 2026-09-24 | A2 Icon buttons | Two sizes, r8; a round button only on media | One size; all round | A lone toolbar icon at the row size looks lost; round rows read as tags |
| 2026-09-24 | A3 Touch size | 44 for every control primitive | 36 + gaps; Button only | One rule (HIG 44) instead of two sizes nobody remembers |
| 2026-09-24 | A4 Input | Bordered, a 55% fg-into-bg edge (≥ 3:1) | Filled; today's look as a primitive | The only option that clears WCAG 1.4.11 without a heavy fill on glass |
| 2026-09-24 | A5 Dropdown trigger | Pill in a toolbar, r8 in a form | Pill everywhere; r8 everywhere | Matches the neighbours in each context |
| 2026-09-24 | A6 Checkbox, switch, slider | Native, accent-tinted, the row is the target | Drawn primitives | Free keyboard / screen-reader / forced-colours support; only 3 sites |
| 2026-09-24 | B1 Density | Compact 32 / 28 / 24 by explicit height | Standard 38 / 32 / 26 | The owner's read of the lab; touch is covered separately (A3), and one toolbar row lines up at 28 |
| 2026-09-24 | B2 Scheme set | Midnight, Forest→cyan, Slate→fuchsia, Matrix→lime | Today's six | The accent collided with danger (YouTube), warn (Slate) and success (Forest, Matrix); Midnight and Starship were one scheme twice; 4 of 6 light accents failed 4.5:1 |
| 2026-09-24 | B3 State layers | Text colour at 10% / 16% over anything | Material's 8 / 12%; today's opacity + bg-card | Visible in every scheme and mode (≥ 1.2:1); today's row hover was 1.00:1 on a card |
| 2026-09-24 | B4 Page chrome | One sticky band + fade; filters fold when short | A floating glass bar; everything scrolls | Tabs and filters always in reach, in one place; a keyboard-up tablet keeps its content |
| 2026-09-24 | B5 Surfaces | One class + radius per role | Today's per-site mix | Nested corners and no per-site alpha drift |
| 2026-09-24 | B6 /watch | The app's surfaces, native video controls | Flat and minimal | The page strangers see should look like the product; native controls work on any device |
| 2026-09-24 | B7 Nav rail | Soft tint + edge bar, optional labels (a user setting), slim | Solid accent (recommended); a Material rail | The owner's combination: calmer than a solid fill, labels for who wants them, no width cost |
| 2026-09-24 | B8 Loading | Skeleton by shape, spinner in a control | A spinner everywhere | Nothing jumps when data lands; a long wait doesn't look stuck |
| 2026-09-24 | B9 Fade | Toolbars / headers, panel tabs, dialog edges, chip rows | Thumbnails; buttons | The north-star depth where content scrolls under chrome; not on content or small controls |
| 2026-09-24 | C1 Action classification | The table (lab C1): state + primary at rest, everyday toggles on hover, "…" on touch, destructive into the menu | Visible buttons per site, as today | One action list per object, nothing gesture-only, fewer buttons without losing a function |
| 2026-09-24 | C2 Side panels | A tap opens an overlay drawer; pin is a separate button ≥ 1024 | A tap pins and the content reflows | Below 768 the filters were unreachable; on touch a tap pinned and pushed the grid |
| 2026-09-24 | C3 Players on touch | Tap = show the controls; 44px seek hit area; 44 buttons; the YT description behind ⓘ | Today (tap also pauses; 7.5px seek bar) | Measured live: you couldn't reach the seek bar without pausing, and the hover-only description never opened on a tap |
| 2026-09-24 | C4 Dialogs | Focus by purpose; a pinned footer | Focus on Close; the footer scrolls | A form starts in its first field, and Save stays on screen with the keyboard up |
| 2026-09-24 | C5 Plex player controls | The owner's lean set: no Stop / Restart / ±10 s buttons, Back by position, Next only with a next item, Resume + From the start on the card, touch double-tap ±10 s | Every control in one row (it ran off-screen below 1024); Back by a press counter; ±10 s keyboard-only | The row fits a portrait tablet; the decision to restart is made on the card; a tablet keeps the 10 s skip |
| 2026-09-24 | Glass tuner | Retired; the schemes are the only colour choice | Keep it with ranges clamped to the rules | Free tuning is the drift the constitution exists to stop (owner) |
| 2026-09-24 | Matrix toned down | Moss #86a05a at Midnight's lightness, near-neutral greys; every white-on-accent part → `--accent-fg` | The B2 lime #a3e635 | The owner found the lime neon and tiring; it was the loudest accent (L 0.85, C 0.21), and fixed white parts sat on it at 1.5:1 |
| 2026-09-24 | Field primitives | `Input` / `Textarea` / `InnerInput` + `fieldClass`, `SelectTrigger`, the native controls; the lint rule "no raw `<input>`" | `inputCls` and ~40 hand-styled boxes | Six edges, four radii and two focus looks for one control; the ≥ 3:1 edge held nowhere |
| 2026-09-25 | Surface roles | `SURFACE` roles in `ui/surfaces.ts`; the lint rule "no bare glass class"; menus p-1 with r8 rows | Three tiers paired ad hoc with five radii and three menu paddings | ~140 sites drifted; nested corners only hold if the radius comes with the role |
| 2026-09-25 | Static labels | Flat tonal `Badge`, no edge; a filtering tag is an edged chip | A glass pill with an edge, like the chips | Edged = pressable at a glance; a static tag that looks like a chip gets clicked and does nothing |
| 2026-09-25 | Header search | A 40 glass pill with the field edge (`lg`); what sits inside it is round | 44, r16, a glass box with a faint edge | The one rounded box among pills, and its edge under 3:1; the hero search is the YouTube/Gmail pill |
| 2026-09-25 | Switch row (F12) | The whole settings row toggles (`SwitchRow`; a hand-laid row points a label at the switch) | Only the 36×24 switch toggles | The iOS / Material settings row; the text was a dead click beside the control it names |
| 2026-09-25 | Edges and hovers (F14) | The ≥ 3:1 edge on the field family (fields, the form trigger); labelled pills keep the faint edge and the accent-edge hover | The ≥ 3:1 edge and a state-layer hover on every control | 1.4.11 asks an edge only where nothing else identifies the control; a heavy edge on every chip and tab weighs the whole UI down |
| 2026-09-25 | Rating and live (F16) | Own tokens: `--rating` (theme-scoped gold), `--live`/`--live-fg`, `--rating-media`, `--media-chrome`; palette classes lint-refused | Map them to `--warn` / `--danger` | A gold star and a red LIVE are what the eye expects; as warn/danger they would read as a status. The old amber score was 2:1 on light |
| 2026-09-25 | Fjord trial (B2 exception) | A fifth scheme, temporary: Midnight's neutrals, Playnite's steel accent (#337599 lifted to #5aa1cb) and steel rules; exit = fold into Midnight or remove | Retune Midnight straight away; borrow only the qualities | The owner wants to live with the calmer palette app-wide before it replaces the default; it breaks "own hue family" on purpose, and only until that call |
| 2026-09-25 | Whole-card row (owner) | A card row's B3 layer is the card's `::before` under its content; the open button's hit layer spans the card, the actions sit above it | The name button's own hover, ending at the first badge or button | A row that half-answers the pointer reads as broken; a badge click did nothing |
| 2026-09-25 | Badge tones (F29) | `Badge tone` neutral / accent / warn at 6%; light `--warn` #b45309 → #a14a07; edged label spans lint-refused | Keep edged status pills (a private Badge in Users, Settings "Granted", the suggestions count) | Edged = pressable at a glance; amber-700 was 4.58:1 as text and fell under 4.5 on its own tint |
| 2026-09-26 | Page band built (B4) | Controls only in the band (intros scroll); the fold shows the same row in place; one `h1` per page (`PageHeading` steps the hidden one aside); `SectionLabel` headings; Skip to content | A popover of the folded filters; the page title kept visible | A folded row is still the same controls, nothing duplicated; the rail and the tab title already name the page |
| 2026-09-26 | Fade built (B9) | `fade-below` on sticky toolbars and table headers; the scroller mask where the band sits on the backdrop | A painted band over the backdrop | The backdrop is the colour under the band, so fading the content into it is the same edge |
| 2026-09-26 | Viewport units | `dvh` everywhere (the Plex player stage with S3f) | `vh` / `h-screen` | `100vh` hides the last rows under a tablet browser's toolbar |
| 2026-09-26 | Video actions built (C1) | One list per video (menu, right-click, long-press, "…"); the card keeps watched / save / add to playlist on its picture on hover; download, details, copy, hide, reset in the menu | The six-button row under the card and the hover ⓘ | The menu is the one door every device has; the card loses a line of height and the touch card its row of 28px buttons |
| 2026-09-26 | Plex actions built (C1) | One list per Plex title (play/open, watched, add to playlist, media info, copy title, pin); the poster keeps watched / add / "…" on its picture on hover; seasons get open / watched / add / copy | The hover ⓘ, the pin-only menu, an episode's add button at rest, a collection's "Remove" X on the card | The same door as a video's; the collection grid goes to two columns on a phone so the touch "…" leaves the title its line |
| 2026-09-26 | Playlist actions built (C1) | A playlist row's "…" (play, pin, copy title / link, move, order, remove), Delete + an undo toast; one list per playlist (rename, pin, YouTube open / push / revert, delete) from its rail row and its header's "…" | The row's X column and its confirm; the header's pencil, YouTube link, push, revert and trash buttons | Remove is reversible, so it needs no confirm; the header keeps Play all; a narrow table drops "#" and Channel so the title keeps a width |
| 2026-09-26 | Channel + Plex list actions (C1) | A channel row's "…" (open, pin, YouTube, copy, full history, hide + undo, unsubscribe); a Plex playlist's title / show / season "…", Delete + undo; its header's "…" (rename, delete); an action toast lasts ≥ 8 s | The channel row's full-history, hide and unsubscribe buttons; the Plex list's red X on every row, show and season; the header's pencil and trash | Hiding and removing are reversible, so they need no confirm and the undo needs time; unsubscribe and deleting a playlist keep theirs |
| 2026-09-26 | Download / notification / Scheduler actions (C1) | A download row's "…" (play, share, save, edit, details, re-download, manage sharing, copies, cancel / delete / remove); a library row's Share stays in sight, ahead of save; a notification row's "…" and hover dismiss; the inbox band's and the Scheduler strip's "…" | The download row's five rare buttons and its cancel / delete / remove; the source link's copy button; an activity notice's Find / Undo links; the band's Clear everything; the strip's Purge discovery | The destructive rows keep their confirm, and Clear everything gains one; a row link keeps the browser's own menu; Find in feed drops the tag / search / age filters that could hide its video |
| 2026-09-26 | Channel menu: Tags + Priority (C1) | Tags A–Z as in-place checks; Priority as fixed steps 10 · 5 · 3 · 2 · 1 · 0, the current one ticked; a pick re-sorts at once and the page + focus follow the row; one drawn `MenuCheck` for every checkable row that cannot be an `<input>` | Relative steps (+1 / +5); straight to the top; the stepper only; seven hand-drawn check marks (three box sizes, a tinted box, a × for "on", a trailing tick) | Bubbling a channel up is one pick; fixed levels keep the scale the reader already uses; the list moving where the reader can see it is the feedback |
| 2026-09-26 | Reorder without a drag (WCAG 2.5.7) | Row menus: Top / Up / Down / Bottom (YouTube + Plex playlists); saved views: four buttons with the place beside the selected view (owner: A); side panels in edit mode: an Up / Down pair per group (owner: A), the name wrapping there | A "…" on every saved-view row; a "…" on every panel group; drag only | A visible, one-click path where the list is edited; the edge moves shown off so the buttons never shift |
| 2026-09-26 | Titles in click surfaces | The surface stays a button, flagged `selectable`: its text selects, and the click that ends a selection does nothing; every plain `Pressable` pinned with its reason | Rebuilding each card and row so the title sits outside the button; exceptions for master-detail lists | Chrome selects inside a button once it is `user-select: text` (measured), so one flag on the primitive keeps the keyboard, focus and list walking every surface already has |
| 2026-09-26 | Right-click on a selection | The browser's own menu opens on selected words; beside them the object's menu | A "Copy selection" row in the app menu | The reader selected the words to do something with them; the native menu copies, searches and translates, and changes nothing else |
| 2026-09-26 | Keyboard ring on a flush row | Inside the row (a table row by its tag, a log line, a show header in its clipping card, with the card's corners) | The ring outside, as on a card | Outside it ran under the table head and past the table, and a clipping card cut it away entirely (invisible focus) |
| 2026-09-26 | Fade at rest | The sticky header's fade appears with the first 1rem of scroll (a scroll-driven animation; no fade where unsupported) | A fade that is always drawn | At rest it dimmed the first row and hid its ring's top edge; nothing is under the header yet |
| 2026-09-26 | Narrow playlist table | "#" and Channel hide by the table's own width (under 42rem), not the window's | Hiding below the `sm` window width | Beside the playlist rail a desktop window left the table 432 px and the title 0 px |
| 2026-09-26 | Background menu built | Five list pages (feed, Plex grid, channels, playlist, notifications); the toolbar's sort / view / filters and the band's actions; no destructive row | Every page with a toolbar; the "…" menus copied whole, delete included | The tab-only pages have nothing a menu would add; a blank-area right-click must not be one click from delete |
| 2026-09-26 | Media card anatomy (C) | One frame; the meta line pinned to its bottom; the length a badge on the picture; the Plex card on the same primitive as the video card | A: the year as an overline over the title; B: the year inline after it; a fixed two-line title slot | A one-line title left a hole the eye reads as a missing line; C puts nothing under a variable-height title |
| 2026-09-26 | One Play | `MediaPlayGlyph` (the media button variant) and only where the click plays | The Plex card's hand-painted accent circle; a playable dot on every card | One Play for every card; a dot that only a colour and a tooltip explain says nothing |
| 2026-09-26 | Filtering values on the info page | A value that filters (year, certificate, studio, director) is an edged chip; a fact is flat text | Everything flat, the clickable ones only hover-tinted | The S3d rule: edged = pressable, so the reader sees which values do something |
| 2026-09-26 | Shelf density | On any strip: one title line over one meta line | The grid's two lines and full meta on a strip | The tallest card set the strip's height and left holes in the cards with less to say |
| 2026-09-26 | Compact list row (B) | The title leads; a faint "…" at the end; the toggles float over the date on hover | A leading 128px action column | That column sat empty at rest, a hole at the start of every row |
| 2026-09-26 | Media-info layout (B) | A main column (collections first, crew, cast) beside a facts sidebar; an empty column is not drawn | Every section full width, stacked | The collection is the most used; the keywords' column left a void beside a short details box |
| 2026-09-26 | Mixed Home shelf (A) | Plex titles as wide art beside videos | Posters, with the videos not stretched | Stretching a video card to a poster's height left an ~80px gap |
| 2026-09-27 | Home hero | A card (hover row + Play on the picture, the whole menu, a touch "…"; owner: A); three quarters wide and centred on a wide window; one fact line; the summary at the measure | The hero's pin row alone; full width; an 18rem picture column | The one answer stands in focus, and it offers what the same card offers anywhere |
| 2026-09-27 | Home shelf tiles | A channel / playlist tile: Play, ↗ open, "…" with Open; a show acts on its next episode, ⓘ opens it; the shelf slot 232px | The tile's pin row alone; 200px | A tile without its object's actions read as broken; at 200 the Play touched the hover row |
| 2026-09-27 | Synopsis preview (C) | A ~0.9 s mouse rest opens a passive panel beside the card | The synopsis as the picture's `title` tooltip; an ⓘ to hover; none | One long tooltip line fired anywhere on the art and clashed with the cut title's own tooltip |
| 2026-09-27 | Reading measure | Prose capped at ~65 characters a line (`measure`) | The container's full width | At 110–130 characters a line the eye loses the next line's start |
| 2026-09-24 | Chip family | One `Chip` (filter / token / tag / suggest / ghost) + `DismissibleChip`; nothing under 24; the lint rule "no raw `<button>` outside ui/" | Five chip geometries in five files | The applied-filter row, the tag cells, the year picker and the saved-view editor each drew their own chip; a one-letter tag was 21px wide |
| 2026-09-24 | Button families | `Button` / `IconButton` / `Pressable` / chips / `DragHandle`; `lg` 40 for a hero CTA; pills and a 48 Play on media (from C5) | Every button hand-painted | 268 hand-rolled buttons in 79 files drifted into as many looks; the ladder only holds if every button stands on it |
| 2026-09-27 | C4 as built | `initialFocus` required on every `Modal`; the commit row in `ModalFooter`, portalled under the body from any depth; the fades driven by the body's scroll timeline; `interactive-widget=resizes-content` | An optional prop defaulting to the ✕; a `footer` prop (the issue composer's buttons live three components down); always-on fades | A default is how 21 dialogs came to open on the ✕; lifting the composer's state to reach a prop is the larger change; a resting fade dimmed the first line |
| 2026-09-27 | C3 + C5 as built | A tap toggles the bar through `pointerType` from `pointerdown`; a double-tap takes back its first tap's toggle; the card's pair is the video card's `MediaResumeButtons` (Continue / Restart), a sibling over the picture; the YT description is an ⓘ click popover | A delayed single tap (wait out the double-tap window); new "Resume / From the start" words for Plex; the pair inside the picture's button | A delayed tap feels dead; one media card says one pair of words; a button inside a button is invalid and the hover row already sits beside the picture |
| 2026-09-27 | C2 as built | The side panel's drawer is `ui/Modal` `drawer-start`; drawer = a coarse pointer or below 768; the pin offered in its header ≥ 1024 | A drawer of its own inside `SidePanel` | Esc, Back, the scrim, the trap and the scroll lock are the dialog's; a second implementation is how two shells drift |
| 2026-09-30 | Composition and the reader (E30 S6) | A chapter and a Conformance step: the space is used, one focal point, worst real data, where am I, a click answers at once, zoom and text size first-class; the person page as two columns (facts and bio left, "Known for" right) | Rules for pieces only; the bio in a full-width panel; a narrower panel | The person page passed every piece rule with half its panel empty; the owner reads zoomed in, so zoom is a requirement, not a nicety |
| 2026-09-30 | Hungarian words (E30 S5) | The D3 list: Feed, Shorts, Admin, Token, Port stay; stream, Prio, Job, Audit, Demo, Auto, Info, Frontend/Backend, Backfill, transcode get the Hungarian word; one name per thing (kérelem, behúzás, one "Test connection") | Keep every loanword; translate Feed/Shorts too | The owner chose the list (D3); the final sweep found the same feature under two names in three places |
| 2026-09-29 | Numbers and order in the UI language (E30 S3) | Every number through `lib/format.ts` in the UI language (decimal mark, grouping, compact, units); A→Z by the shown name in the UI language's alphabet, client and server alike; the language comes off `useTranslation`, never the singleton | `toFixed`, hand-built "1.2K", `localeCompare` in the browser's locale, byte-order server sorts | A Hungarian reader saw "8.3", "1.2K", "Életrajzi" after "Z"; the singleton read also left values in the old language after a switch (React Compiler) |
| 2026-09-27 | Narrow rows by container | The download row, the playlist table and the Channels tag row stack by a container query; a fact line never wraps | Breakpoints on the window | The iPad zoomed in, split screen and a column beside a rail are narrow while the window is not; the zoomed download row broke "18.9 / MB" apart |
| 2026-09-27 | Phones deferred | A phone-only layout (the hamburger nav, a folded feed band) waits for its own epic; E28 stays desktop + tablet | Fix the phone views inside E28 | Principle 2; the first phone attempts made the feed worse, and the epic's review grows with every sprint |
| 2026-09-27 | Conformance check (S3g) | A four-step check before a UI change is done: primitive at a named rung, the chapter re-read, the rendered page measured by kind and look, every state seen | A reviewer's eye per sprint | The S1 probe found 69 button looks; after S3 every control kind is drawn by its primitive, and the check keeps a new look from arriving unnoticed |
| 2026-09-27 | Segmented rung | `SegmentedToggle` defaults to the toolbar rung (28, `text-xs`); 24 only for a panel's captioned match toggle | The small default | Two toolbars had inherited 10px type from the default beside 28px buttons |
| 2026-09-27 | Copy rule (S3g) | One name per feature (the rail may shorten it); four verbs for four meanings; sentence case with the app's places capitalised; "…" when a command asks first; one-sentence messages without a stop; curly quotes; dates relative in lists with the exact minute on hover, three absolute forms; units from Intl, short | Word-for-word names everywhere; the verbs as they were; all lower case; no "…"; always a stop; always absolute dates; long or compact units | The copy survey found 7 EN and 6 HU double names, 22 Delete against 27 Remove, five date shapes and three spellings of a minute; the owner picked every recommended option |
| 2026-09-27 | One glyph, one meaning (S4) | Play only for playback or resuming a paused transfer; Retry = `RefreshCw`; reset / revert = `RotateCcw` | Retry drawn with Play | The first UI-review run (Downloads) found a failed row whose Retry showed ▷ in the row and its menu |
| 2026-09-27 | Fjord kept | Fjord stays a fifth scheme beside Midnight; the trial wording goes | Fold it into Midnight; remove it | The owner lived with it and likes it; the vivid Midnight stays for those who want it. B2's own-hue rule still binds any new scheme |
| 2026-09-27 | Tablet columns | A table hides `tabletHidden` columns below 64rem of its own width; Channels drops Prio, About, Length | Hide About only; give the name a bigger share | At 1024 the channel name had 60 px; the owner chose to drop all three there |
| 2026-09-27 | Glyph pager | First / previous / next / last as glyph buttons round the page box, names as tooltips | Words beside the arrows (hidden only on a phone) | The owner's call: the words widened the row; four arrows are the shared convention |
| 2026-09-27 | Match switch on the panel | `ui/MatchToggle` above the picker from two picks, everywhere; one wording (Any / All) | The picker's popover footer (Feed, Plex collection) | A closed picker hid an active "all"; six pickers had drawn it two ways |
| 2026-09-27 | Missing picture, bare link | A glyph of the kind for no picture; a download with no name is titled by its short link, once | An empty box; the URL as title, channel and link | The owner's picks 4A and 3A: an empty box read as loading, the URL showed three times |
| 2026-09-27 | Copy on a selection | Right-click on selected words keeps the browser's menu; no app Copy row | An app menu with Copy first | A real browser's menu has copy, search and translate; only the desktop pane's is thin |
| 2026-10-01 | Zoom target 125–133% | Screens are held at 125% and 133% browser zoom (1536 × 729, 1444 × 685 on a 1080p window) and the largest text size | 200% browser zoom | The owner: 200% was a guess; 125–133% is what they read Siftlode and most sites at, so 200% would be an upper bound nobody uses |
| 2026-10-01 | E31 Nav rail at short heights | One "Admin" rail item opening a panel of the admin pages | A denser rail when short; a visible "+4" hint | The rail fits at 133–150% without a hidden scroll; the admin pages are rare and admin-only |
| 2026-10-01 | E31 Configuration sections | A vertical, grouped section list, as Settings has | Group tabs + sub-tabs; one long page with a contents list | Sixteen tabs wrapped to three rows at 133%; one pattern for both settings pages |
| 2026-10-01 | E31 Statistics overview | Number tiles, and the per-operation counts as horizontal bars | Three narrow cards; a 30-day chart | The label sits by its value; what spends the quota is seen at a glance |
| 2026-10-01 | E31 Channel page header | The full banner, folding to one row once the grid scrolls | Always compact; no banner | The header is fixed and took 40% of a 685px window; the channel keeps its face at first sight |
| 2026-10-01 | E31 Channel page watch state | A fresh channel page shows the unwatched videos; a picked state is remembered | Show all, forgotten on reload | The owner: open a channel to see what is new; a choice made there should stick |
| 2026-10-01 | E31 Audit log runs | A job's successful runs fold into one row that opens, whatever ran between; only that job's failure (never folded, danger colour) breaks it | A filter defaulting to admin actions; both | Nothing hidden, the noise collapsed, the failure seen |
| 2026-10-01 | E31 Users row | Role and state as badges, the actions in the row's one menu | An inline role toggle; icon-only buttons | Action placement: the row shows state, rare and destructive actions live in the menu |
| 2026-10-01 | E31 YouTube player | Video beside an info panel at 1280px and wider (the video as large as the height allows); stacked below | Today's stack with a one-line bar; full-window like Plex | The owner's eyesight: about 1460px of video on 1080p instead of 1008, the actions still visible |
| 2026-10-01 | E31 Plex info and show pages | A second column for the facts, keywords and crew, as the person page; a collection's hero sized to its content | Full-width text; a larger poster and fanart | The right ~40% was empty; one composition for every media page |
| 2026-10-01 | E31 Nav rail rows are links | Each page row is a real link: a middle-click or Ctrl-click opens the page in a new tab | A button | The owner: open a page in a tab of its own, as on any site |
| 2026-10-02 | E31 Plex detail second column (refined) | The facts over the key crew beside the text; the keywords, chips that filter like the genres, under the text; the show page gets all three too | All three in the column, as the mockup drew it | At 20 keywords and 12 crew the column ran to twice the text's height (849 vs ~400 px at 133%): a hole under the text instead of beside it |
| 2026-10-02 | E31 Plex detail poster and crew | The movie and show poster keeps its 2:3 and grows a rung in a wide hero (w-56 at a 64rem panel); the crew shows its first 6 names and "Show N more" (a list of 8 or fewer shows whole) | Poster stretched to the row; a "More" on the keywords | The row is as tall as the crew column, so the poster drew ~175×590 and cropped; the crew, not the keywords, set the hero's height |
| 2026-10-04 | E32 `field` focus skips a mode switch | A form opens on its first field; a control above it that is not a field (the composer's Edit / Preview) is passed over by marking the field `data-initial-focus` | The first text field everywhere | That moved the focus in the download dialog off its profile choice and in the upload dialog off its drop zone; the mark changes only the form that needs it |
| 2026-10-04 | E35 Lint for z-levels and edged labels | Both rules stay; the two eslint selectors (`Z_RAW`, `EDGED_LABEL`) that refused them go | Keep the selectors | 0 hits since the E28 sweep that fixed the sites: a check that never fires costs a probe test and a snapshot each, and the UI review holds both rules |
| 2026-10-06 | `player` focus | A dialog that is a player opens focused on its panel (the YouTube player's rule, now `Modal`'s too); on the media element itself, Space and the arrows stay the element's own | `close`, or the media element | On the ✕ Space closed the player; on the element the browser's own controls acted on the same key as the player's (one Space played and paused, measured in Chrome) |
| 2026-10-08 | Captions on touch | A `Tooltip` opens for a mouse and a key's focus only, never for a finger | A caption on tap; a long-press caption | On an iPad the tap that showed the caption was taken as a hover and lost its click (every rail button took two taps), and no finger sends the mouseleave that hides it; the rail's labels setting names the buttons on touch |
| 2026-10-08 | Header search on touch | No key cap on a touch screen; the pill grows to 52 px so a ring shows round the 44 px Go; Go drops to its icon by the column's width | Keep the key cap; a 44 px Go overhanging a 40 px pill | On a tablet the 44 px key cap and Go sat on the pill's edge and cut the placeholder; an overhanging Go breaks the nested radii |
| 2026-10-08 | Plex subtitles and the control bar | The player draws the cues itself; while the bar shows they rise above it (200 ms), then settle back | The browser's own cues, moved by `line` | On an iPad the bar covered the text; a native cue cannot be animated or measured, and our layer behaves the same on every engine |
| 2026-10-08 | Hover on a button over a picture | The `media` variant's hover is a 2px accent ring over a darker body (Plex + YT cards, rail arrows; option B) | An accent fill (reads as "selected"); a white wash (weak on a bright poster) | The 60 → 75% black step was invisible, and over a card a rail arrow could not be told from the card's own Play under it |
| 2026-10-08 | Plex "Add to collection" | A movie or show card's menu row opens a panel beside the card: its library's manual collections tick in place, a name + Enter creates one with the title in it (admin; option A) | An Add submenu with a "New…" dialog; the per-title editor modal from the menu | A collection took a trip to the Collections page, a name, a title search, then a second search per title; the panel is the video card's "Add to playlist" |
| 2026-10-08 | YouTube modal chrome | No prev/next strips beside the card; Close an icon (the panel's top right; stacked, the title row's end); the description clamped to three lines + More; the panel shuts to a 2.75rem strip of close, expand, prev/next (option D1) | Keep the strips; a "Close" text button; the full description in a scroller; a panel that hides outright | The strips cost the video 7rem of width for a second copy of the panel's buttons; a long description pushed the channel and the actions out of sight; a hidden panel left no way to step the queue without the keyboard |
| 2026-10-08 | The wheel over the YouTube modal | Over the video: volume. Over a panel with something to scroll: scroll only, never volume, not even at its end. Over a panel with nothing to scroll: volume (owner) | Volume everywhere; a scroll that hands over to the volume at its end | Reading a long description turned the volume down once it hit the end |
| 2026-10-09 | A Plex show card plays | Its picture plays the show's next episode everywhere (an episode with a resume point → Continue + Restart, else the first unwatched, else S1·E1), resolved on the click; the preview names it once up; title and ⓘ open the show (option A + hover ask) | The picture opens the show page (as before); every card carrying `next` | A show card on Discover had no Play while Home's had one; carrying `next` costs a grid page up to 180 queries for the few cards anyone plays |
| 2026-10-10 | Clip mode in the players (E38) | Scissors in the control row + `C`, `[` `]`; the panel in the player with the watched video as preview; overview + zoomed strip on `ui/RangeTrack` (ends are sliders); a toast + the Ready notification after Cut | A ⋯ menu entry; permanent In / Out buttons; a big dialog with its own preview (a second Plex transcode / YouTube iframe) | Owner's picks A, P1, a; a film is hours and a clip seconds, so only a zoomed strip gives a hand the precision. The Downloads editor's cuts moved onto the same track (keyboard-steppable now) |
| 2026-10-10 | Crop in clip mode (E38) | A Crop toggle in the clip panel: a frame over the playing picture (`ui/CropFrame`, the editor's crop too), Free / 16:9 / 9:16 / 1:1, Whole picture; maximised YouTube keeps the panel in the bar's place; the wide side panel shuts while clipping | A frame always on the picture in clip mode (K2); a free shape only | Owner's K1 + a: a crop is rare, so it never covers the picture unasked; a phone or square clip needs a held shape. The iframe hides its picture's size: drawn on the 16:9 box, the cut fits it to the real picture |
| 2026-10-10 | Marks on the player chrome | `--media-mark` (amber) and `--media-danger` (light red), the same in both themes; `media-marks` points the focus ring (`--focus`) and the chosen mark (`--chosen`: a track's tick, a tab's underline; the label stays white) at the amber on the player and every chrome surface; `media-scope` (a form) also the errors at the red | The theme's accent ring, chosen text, warn and danger | On the chrome (#171717) a light theme's accent ring (Slate 2.84:1), warn (2.99:1) and danger (2.77:1) all fall under 3:1; the accent as chosen text under 4.5:1 |
| 2026-10-11 | Files page (one library) | List rows by default, the media grid a toggle remembered per folder; a sidebar of automatic source folders, own folders, Shared with me and the storage meter; the paste-a-link box in the toolbar | A media grid by default | The library holds videos AND plain files: rows read both, show more per screen and scan by size and date |
| 2026-10-11 | Library list columns | Every kind of value in its own column (name, source, phone-safe, retention, size, added, actions), one grid with a subgrid per row so they line up; an action that does not apply keeps its slot empty; a right-click menu on every row (the ⋯ opens the same menu) | Badges flowing after the title; icons that shift when one is missing | Owner: values of one kind must fall in one vertical line; a row's every action reachable by right click |
| 2026-10-11 | Running downloads | An app-wide tray, bottom right, shown only while something runs or failed (reason + retry), foldable to a pill | A rail icon with a count and a popover | Downloading became an action you start anywhere; a running one must be visible without going to look |
| 2026-10-11 | Download done | The notification says whether it plays on a phone and carries Link másolása + Megnyitás (the folder, the item focused); never an automatic jump | A plain "ready" notice; auto-navigation to the item | The daily path is paste → done → copy the link; a jump would pull the user out of what they were doing |
| 2026-10-11 | Library row quick icons (owner A) | On a wide list only the ⋯ holds a place; Copy link, Share and Save come up over the date while the row is hovered or holds the keyboard; a card (phone) keeps them in its own line; Keeps narrower | Three icons always in sight (the name kept ~150px at 1200); dropping Added below 64rem | Owner: the name is what a row is read by; every action stays in the ⋯ and the right-click menu |
| 2026-10-12 | Files preview (owner ★) | A file that is not media opens from its name in the dialog shell, wide: an image, a PDF in the browser's own viewer, text as numbered plain text (HTML and SVG too) with a wrap toggle, else a no-preview card with Save; ← / → page the list's such files in its order; an image under 8 MB is its own thumbnail | A side panel beside the list; a new browser tab; pdf.js; syntax highlighting; a click that saves a file without a preview | Owner: one way to open anything, paged like the list; the browser's PDF viewer is free and isolated; a page is never rendered on our origin |
| 2026-10-12 | A running download's step (owner ★) | The tray row names the step beside its bar ("Downloading audio · 12%", "Merging" with a pulse); a paused-only tray shows a pause mark, never the done check | A bare bar and % | A download fetches video and audio as two streams: unnamed, the second fill reads as a restart; the tray is the only place a download is followed |
| 2026-10-14 | API tokens (owner A/A/A) | Settings → API tokens: a row per token in three lines (name + prefix, its folders and rights, last use and expiry), revoked from its ⋯; a new token asks the kind first (Hub exchange or Custom: rights as chips per folder), expiry 1 year by default; the token shows once in the same dialog, with Copy and a curl line | A table; one form with a preset link; the token in a banner above the list | A table truncates the rights at this width; the common token never shows a matrix to get wrong; closing the dialog is plainly when the token is gone |
| 2026-10-15 | Folder links (owner A/A/A) | An own folder's menu has Copy link and Share with link… as a file row does (the dialog is the link half of the file's); the link opens a list of the folder's files as it is now, Play opening the watch player above it with the playing row lit, a document only downloading; the rail marks a folder with a live link beside its count | Share with link… alone; cards; a download-only list; no mark | Mostly videos from a phone, read on a phone: a list gives a long name its line and the next one is a tap away; the folder link is the one place where moving a file in publishes it, so where it is must show |
| 2026-10-17 | Copy-only changes (owner ★) | Step 4 of the Conformance check alone (not 1–3 or 5): every state the changed text renders in, seen in both languages | The whole five-step check; no screenshot at all (DOM test + key parity) | Steps 1–3 and 5 judge controls, patterns and composition, which a text change does not touch; the one risk left is a longer Hungarian string wrapping or clipping, and only a screenshot shows that |
| 2026-10-17 | Text on glass (owner ★) | Light `--muted` 15% toward `--fg` in every scheme (≥ 6:1 on the bare page); the item count, the badges and the current Files row on opaque tints mixed into the surface | Only the two failing components; the item count's number in `--fg` | Pixel-measured over every scheme × mode × backdrop, the count fell to 3.8:1 (dark too: thumbnails scroll under the toolbar's glass) and Channels' tabs to 4.1:1; an `--fg` count outshouts the filters beside it |
| 2026-10-18 | Scheduler run history (owner D4, E1, E2 ★) | A row's last three runs as dots, newest first (interrupted: a hollow warn ring, set apart from skipped and failed), and a history button that opens the job's runs on the Activity page; a user's ad-hoc runs in their own block under the jobs, with no interval and no run button | An inline expandable history per row; a dark-grey interrupted dot; ad-hoc runs only on the Activity page | Three dots answer "did it run, did it work" at a glance; the whole history needs filters, which the Activity page has |
| 2026-10-18 | Activity page (owner D1, D2, G1–G3 ★) | One page, "Napló" / "Activity": four tabs (Runs, Admin actions, Server log, Quota) under one filter bar (text, a period of 1 h · 24 h · 7 d · 30 d · 90 d defaulting to 7 days, an account), the view in the URL; runs and admin actions as dense rows that open in place as a log line does, filtered and paged on the server; a run's or a request's log lines one pivot away, as a history entry | One merged timeline; a page per source; a default period per tab; card rows; the audit log's client-side table | Each source keeps its own table and retention, read in one place; one bar keeps what you narrowed across tabs; about 1.8k runs a day need a log's density and a server-side filter |
| 2026-10-24 | Related videos (owner U1–U3 ★, D1 A, D3) | On demand: a tab beside the description (wide), a drawer (stacked); the narrow thumbnailed row, its menu alone, a followed check; plays in place and the list stays the opened video's; an on-topic / Hungarian-trending switch kept on the account | Always shown; the end-of-video grid alone; a dialog; a big card or a text row; a list that follows the playing video | The modal is full on a small screen; a picture is what a choice among strangers rests on; a list that stays put is tried one by one, and spends no more of the day's cap |
| 2026-10-25 | Related: one played; off (owner D1 a, D2 a) | Played from the list, it is the player's item: its title, controls and progress, marked watched at its end, the queue left where it was; a daily limit of 0 takes the tab and the button away | A linked video's in-place play (no controls, nothing kept); a tab that only shows a refusal | What is watched should count as watched, wherever it was started from; a control that can only be refused is not offered |
| 2026-10-25 | Player panel glyphs (owner 1A, 2A, 3) | In the side panel the queue row is glyph-only on one line, and Mark watched is a glyph toggle among the actions (✓ / ✓✓); the docked bar keeps its labels; the related switch at the default size, "Trending" | Labelled buttons wrapping onto three lines; glyphs on the docked bar too; Mark watched on its own line | The panel is narrow and the list controls are the same few every time; each glyph keeps its full name and shortcut in its tooltip; the docked bar has the width for words |
| 2026-10-25 | Auto-advance glyphs (owner E A, H a) | Auto-advance wears list glyphs (`ListEnd` next and off, `ListStart` previous, `Shuffle` random), in the panel and the docked bar; the auto-advance and loop cycles are `lit`, not `pressed` | The skip arrows for the mode; an X'd list for off; `aria-pressed` on a cycle | Beside Previous / Next the skip arrows read as a second Next; a cycle of three or four modes is no two-state switch, and its name already says which |
| 2026-10-25 | Restart and Reset glyphs (owner L A) | A menu's Play from the beginning wears `RotateCcw`, as the card's own Restart does; Reset (clear progress and status) wears `Eraser`; Redownload wears `RefreshCw`, the retry arrow | The skip arrow ⏮ on the menu's Restart; `RotateCcw` on both Restart and Reset; `TimerReset` for Restart | ⏮ is Previous beside the player; one action wore two glyphs, and two actions in one menu would have worn one |
| 2026-10-28 | Notification order (owner A-D1 a, A-D2 a, ★ a) | What asks for the reader first, from either source ("Needs you · n"), then one stream newest first; read rows older than a week fold under "Earlier (n)", closed, its label in the section-heading style with a chevron | One stream with no band; this browser's activity above the server's (the two fixed sections); a server-side retention that deletes old rows | An admin's access requests sat under up to a hundred server rows; week-old read notices crowded what is new; folding keeps the rows (nothing is deleted) and an unread row never folds, whatever its age |
| 2026-11-01 | Docked bar glyphs (owner D8 a) | The maximised view's docked bar is glyph-only too (tooltips carry the names and keys): transport, auto-advance, loop, Mark watched (✓ / ✓✓); its controls scroll on a narrow window while Fullscreen and Exit stay pinned at its right end | Labels from `xl` up | On a tablet the labelled bar ran past the right edge and took Exit with it; the side panel had already lost its labels |
| 2026-11-01 | Toolbar group glyphs (owner D9 a, D10 a, D11 a) | The feed's, channel page's and Plex Browse's filter pills name their group by a glyph (`Eye` watch state, `Shapes` content types, `ArrowDownUp` sort), the word in the tooltip and the accessible name; Sort + View end the toolbar's row, the chip row shows only with a chip; a channel page has no count pill | The value alone; Sort + View on a row of their own | "Watch state:" and "Sort:" cost a tablet row its room; the value says enough once the glyph names the group; the channel header already counts the videos |
| 2026-11-01 | Channel header on a tablet (owner E2) | Below `xl` the action row drops its two words: "Prio" (the stepper's tooltip names it) and "Subscribed" (a ✓ glyph; its name in the aria-label, Unsubscribe in the tooltip); the tags keep the row's rest, folding into "+N" | The tag lane on a full line of its own below ~12rem; the tags under the channel name | On an iPad (820 px, 44 px targets) the lane got 29 px, a sliver; now 130 px, and every tag shows at 1180 |
| 2026-11-01 | Clear watch progress (owner E a) | A video menu's Reset reads "Clear watch progress" in English (HU "Megtekintés visszaállítása" stays); "Reset view" is only a page's view reset (sort, filters, grouping) | "Reset view" for both; the playlist page's view reset renamed instead | On a playlist page the two stood one right-click apart under one name, and one of them erases the video's progress with no undo |
| 2026-11-01 | Clear watch progress clears only that (owner A) | The row shows only on a video with a resume point and clears that alone: a hidden video stays hidden, a watched one watched. Un-hiding and un-watching have rows of their own | The old Reset, which also un-hid and un-watched | A hidden video stays hidden: what the row's name says is all it does |
| 2026-11-01 | Selecting a list's rows (owner D2 a) | An always visible box first in the row and one over the column for the rows shown; Shift+click a range, `x` the row under the keyboard, Esc clears; the selected row wears `accent/12` with its tick. Rows a filter hides stay selected, an action covers those shown; Watch later offers no selecting. | A "Select" button that shows the boxes on demand | Found at once and works on touch, with no mode to leave; costs about 2 rem a row |
| 2026-11-01 | Selection band (owner D1 a, D3 a, D4 a) | The toolbar's second row while rows are selected, no floating bar: move (the last target as one click) and remove, no copy; both undo back to the places, a video the target held before stays there; a single row's move undoes too. | A floating bar at the bottom; copying to a list too; no undo on a move | It covers no row, and the player and the notices live at the bottom; twenty videos sent to the wrong list need a way back to their places |
| 2026-11-01 | Grip on a phone (owner A) | Below a 24rem table a playlist row has no grip: the title keeps its width, and the row menu keeps the moves. | The grip and its drag kept, the title at 44 px | The title gets its width back, and the rail is shut on a phone, so a drag had nowhere to go |
| 2026-11-01 | Selection band on a tablet (owner A) | Below `xl` the band is glyphs on one row, the names in tooltip and aria-label, as on the channel header; the replaced row stays in the same grid cell, invisible, so the first tick moves no row under the second tap. | The buttons in words, wrapping to a second row | The first tick pushed the list from under the second tap |
| 2026-11-01 | Playlist heading on a narrow window (owner A) | The list's name had no width left: below `xl` Unsynced is its dot (the word read out, the hint in the tooltip); on a phone the cover is gone and Play all is its glyph, named in the tooltip. | Every word kept, the list's name at no width | A heading without its name says nothing |
| 2026-11-02 | Dragging onto a list (owner D5 a, D6 a) | The grip carries a row, or every selected row with a selected one, onto a list in the rail; the toast undoes. Many dropped inside the list stay; a coarse pointer's drawer never peeks, so its rail is a target only when pinned. | Only the grabbed row going along; the grip off under a view | One row moving out of three lit ones misleads, and a filter is how the rows to move are found |
| 2026-11-02 | Delete with every tick hidden (owner a) | It removes nothing and a warning says the filter hides them | Taking the row under the keyboard | While the band shows a selection, Delete means that selection: a row no one ticked must not go |
| 2026-11-02 | Undo of a video gone further (owner a) | A video the move put in the target comes back only if the target still holds it, and a warning says how many | Everything back, whatever it did since | A move skips a video that left the source; the same rule back keeps a video out of lists the reader never put it in |
| 2026-11-04 | Tag chip counts (owner D5a) | A Channels tag chip shows how many followed channels it holds: a muted number after the name, no badge | The number in brackets after the name | The owner's pick; the number is a detail of the name, so it takes the name's own text, muted, rather than a `CountBadge` |
| 2026-11-05 | Picked tag chips lead the row (owner c) | On the Channels tag row the picked chips come first, in the tags' order, then the rest | Keep the order and fold the unpicked first; mark the "+N" when it holds a pick | The row folds from its end, so a picked tag late in the list sat only inside "+N" with nothing on the row saying a filter was on |
| 2026-11-05 | A toggled tag chip keeps the focus (owner a) | A chip that moves when toggled takes the keyboard focus with it: to its new place, or to "+N" when it went into the closed fold | Keep a picked chip in the "+N" popover until it closes | The row always shows the filters as they are; the focus, not the order, was what got lost |
| 2026-11-06 | Tag chip count after a dot (owner a) | The count keeps the chip's own colour, set off from the name by a "·" | The number dimmed (it fell under the 4.5:1 text floor, resting and pressed); the number in a tinted inner badge | Contrast holds in every scheme without a new token, and the dot still reads the number as a detail of the name |
| 2026-11-06 | A pick in the open fold closes it (owner a) | Toggling a chip inside the "+N" popover closes the popover; the focus goes with the chip | Keep the popover open and focus the chip now in the picked one's place | The focus never sits outside a popover left open, and a picked chip leads the row anyway |
| 2026-11-06 | An expired share link offers no Copy (owner a) | An expired row in the share dialog has a warn badge "expired <date>" and its address struck through; Open and Revoke stay, Copy goes | The warn badge alone, Copy kept | An expired link opens nothing, so it is no link to hand out; Open still lets the owner check it |
| 2026-11-06 | The shared page's unlock button is a noun in Hungarian (owner a) | "Megtekintés" for a video, "Megnyitás" for a file or folder | Both verbs, "Megnézem" / "Megnyitom" | The app's Hungarian buttons are nouns; one button should not change grammatical form by link kind |
| 2026-11-07 | A slow link says so, with both rates (owner b1) | After 5 s of stalling on a link measured slower than the stream: "Slow connection" and "This file needs about X, about Y is arriving" under the spinner; no number measured, no claim | The line without numbers; a pill at the bottom | The viewer can tell their connection from a broken player, and by how much it falls short |
| 2026-11-07 | Text over the picture sits on a dark panel (owner a) | The player's loading, stall and error text, the subtitle notice and the track OSD sit on one dark panel (`OVER_VIDEO`, black 75 %, a white 15 % edge) | A shadow behind the letters; bare text as before | Bare white or red text was unreadable over a bright frame (seen live); the panel is the style the notices already had |
| 2026-11-07 | The panel over the picture is darker, its error the media red (owner a) | `OVER_VIDEO` is black 85 % under `media-scope`: an error on it is the media red in both themes | Only the red changed (3.75:1 on a white frame); black 90 % | A light theme's red was 1.6:1 on the 75 % panel over a white frame; now 5.47:1 at worst, white text too |
| 2026-11-07 | A total's left-out videos are "without length" (owner b) | A summed length names what it left out as "+N without length" / "+N hossz nélkül" | "+N live" as before; "+N stream" | The count holds archives and upcoming streams too, which a card calls a stream or upcoming: "live" there named only one of them (One name per feature) |
| 2026-11-08 | The Channels table's Length is the sum alone (owner) | The table's Length cell shows the summed hours only; the "+N without length" stays on a playlist's total | The line wrapped in the column; the column widened to 6.5 rem | The 4.5 rem column is sized for the sum: the longer line spilled into Tags (88 px in 56), and the owner wants no such line there |
frontend/src/content/release-notes/en.json
+12
View File
@@ -1,4 +1,16 @@
[
{
"version": "0.118.0",
"date": "2026-11-08",
"new": [
"Surprise me now leans on the channels and genres you watch, and one pick in three stays a blind draw, so it remains a surprise.",
"On a connection too slow for a Plex video, playback waits and goes on where it was, and says how fast the video needs and how fast it arrives."
],
"fixed": [
"A 24/7 stream's archive no longer adds years to a list's total: a total names the videos it could not count as \"+N without length\". The Feed sorted by Surprise me opens faster.",
"Plex films the server's GPU cannot decode (VC-1, older DivX or XviD) start in a few seconds instead of about 20. Text over the Plex picture is readable on any frame."
]
},
{
"version": "0.117.0",
"date": "2026-11-06",
frontend/src/content/release-notes/hu.json
+12
View File
@@ -1,4 +1,16 @@
[
{
"version": "0.118.0",
"date": "2026-11-08",
"new": [
"A Meglepetés most a nézett csatornák és műfajok felé hajlik, de minden harmadik választás vak húzás marad, így továbbra is meglepetés.",
"Ha a kapcsolat túl lassú egy Plex-videóhoz, a lejátszás kivár és ott folytatódik, ahol tartott, és kiírja, mennyi kellene és mennyi érkezik."
],
"fixed": [
"Egy 24/7-es közvetítés archívuma már nem ad éveket a lista összegéhez: az összeg „+N hossz nélkül” néven mondja meg, mit nem tudott beszámolni. A Meglepetés szerint rendezett Feed gyorsabban nyílik.",
"A Plex-filmek, amelyeket a szerver GPU-ja nem tud dekódolni (VC-1, régebbi DivX vagy XviD), pár másodperc alatt indulnak a korábbi kb. 20 helyett. A Plex-kép fölötti szöveg bármilyen képkockán olvasható."
]
},
{
"version": "0.117.0",
"date": "2026-11-06",
frontend/src/features/video/withoutLength.test.ts
+19
View File
@@ -0,0 +1,19 @@
import { describe, expect, it } from "vitest";
import type { TFunction } from "i18next";
import { isWithoutLength, withLeftOutCount } from "./withoutLength";
const t = ((key: string, o: { count: number }) => `${key}:${o.count}`) as unknown as TFunction;
describe("a total and the videos it cannot hold", () => {
it("counts every video the server sent no length for", () => {
expect(isWithoutLength({ duration_seconds: null })).toBe(true);
expect(isWithoutLength({ duration_seconds: 5400 })).toBe(false);
expect(isWithoutLength({ duration_seconds: 0 })).toBe(false);
});
it("names them beside the sum, or alone when nothing summed", () => {
expect(withLeftOutCount("12 hr 5 min", 2, t)).toBe("12 hr 5 min card.withoutLength:2");
expect(withLeftOutCount("", 1, t)).toBe("card.withoutLength:1");
expect(withLeftOutCount("12 hr 5 min", 0, t)).toBe("12 hr 5 min");
});
});
frontend/src/features/video/withoutLength.ts
+19
View File
@@ -0,0 +1,19 @@
// A total of video lengths and the videos it cannot hold. The server decides what a length is (a
// YouTube "length" past a day is a 24/7 stream's archive reporting the years it has run, and the API
// sends none for it: `catalog.models.LENGTH_MAX_S`), so this side keeps no copy of that limit. It only
// names what a sum left out: every video with no length — live, an archive, or not fetched yet — as the
// server's `Video.without_length` counts it for the totals it sums itself. The two rules are the same,
// so the count and the sum never disagree about what was left out.
import type { TFunction } from "i18next";
import type { Video } from "../../lib/api/feed";
export function isWithoutLength(v: Pick<Video, "duration_seconds">): boolean {
return v.duration_seconds == null;
}
/** "12 hr 5 min +2 without length": `total` is the formatted sum ("" when nothing summed), `leftOut`
* the count it left out. Empty when both are. */
export function withLeftOutCount(total: string, leftOut: number, t: TFunction): string {
const suffix = leftOut > 0 ? t("card.withoutLength", { count: leftOut }) : "";
return [total, suffix].filter(Boolean).join(" ");
}
frontend/src/i18n/locales/en/card.json
+2
View File
@@ -16,6 +16,8 @@
"restart": "Restart",
"play": "Play",
"stream": "stream",
"withoutLength_one": "+{{count}} without length",
"withoutLength_other": "+{{count}} without length",
"views": "views",
"info": "Details",
"noDescription": "No description",
frontend/src/i18n/locales/en/playlists.json
+1
View File
@@ -89,6 +89,7 @@
"watch_later_not_pushable": "The Watch later list can't be sent to YouTube",
"watch_later_undeletable": "The Watch later list can't be deleted",
"youtube_editing_off": "Enable YouTube editing in Settings first",
"push_cooldown": "You've synced to YouTube many times just now — give it a few minutes",
"quota_exceeded": "Not enough YouTube quota left today for this sync",
"sync_busy": "This playlist is already syncing with YouTube — try again when it finishes",
"sync_all_busy": "Your playlists are already syncing with YouTube — try again when it finishes"
frontend/src/i18n/locales/en/plex.json
+3
View File
@@ -191,6 +191,9 @@
"errNotFound": "This file can't be played — the media isn't reachable on the server (the Plex media mount is missing or the path map is wrong). Ask the admin to check the Plex configuration.",
"errTranscode": "This file's format needs transcoding, which isn't supported yet.",
"errGeneric": "Playback couldn't start. Please try again.",
"errSlow": "Couldn't play: the connection is too slow for this file",
"slowLink": "Slow connection",
"slowLinkRates": "This file needs about {{need}}, about {{got}} is arriving",
"help": "Keyboard shortcuts",
"infoBtn": "Media info",
"more": "More",
frontend/src/i18n/locales/hu/card.json
+2
View File
@@ -16,6 +16,8 @@
"restart": "Elölről",
"play": "Lejátszás",
"stream": "közvetítés",
"withoutLength_one": "+{{count}} hossz nélkül",
"withoutLength_other": "+{{count}} hossz nélkül",
"views": "megtekintés",
"info": "Részletek",
"noDescription": "Nincs leírás",
frontend/src/i18n/locales/hu/playlists.json
+1
View File
@@ -89,6 +89,7 @@
"watch_later_not_pushable": "A „Megnézem később” listát nem lehet a YouTube-ra küldeni",
"watch_later_undeletable": "A „Megnézem később” lista nem törölhető",
"youtube_editing_off": "Előbb engedélyezd a YouTube-szerkesztést a Beállításokban",
"push_cooldown": "Az imént sokszor szinkronizáltál a YouTube-ra — várj pár percet",
"quota_exceeded": "Mára nem maradt elég YouTube-keret ehhez a szinkronhoz",
"sync_busy": "Ez a lista épp szinkronizál a YouTube-bal — próbáld újra, ha végzett",
"sync_all_busy": "A listáid épp szinkronizálnak a YouTube-bal — próbáld újra, ha végzett"
frontend/src/i18n/locales/hu/plex.json
+3
View File
@@ -191,6 +191,9 @@
"errNotFound": "Ez a fájl nem játszható le — a média nem érhető el a szerveren (hiányzik a Plex media-mount, vagy hibás az útvonal-leképezés). Kérd meg az adminisztrátort, hogy ellenőrizze a Plex beállításait.",
"errTranscode": "Ennek a fájlnak a formátuma átkódolást igényel, ami még nem támogatott.",
"errGeneric": "Nem sikerült elindítani a lejátszást. Próbáld újra.",
"errSlow": "Nem sikerült lejátszani: a kapcsolat túl lassú ehhez a fájlhoz",
"slowLink": "Lassú a kapcsolat",
"slowLinkRates": "A fájlhoz kb. {{need}} kell, most kb. {{got}} érkezik",
"help": "Billentyűparancsok",
"infoBtn": "Média infó",
"more": "Továbbiak",
frontend/src/lib/api/channels.ts
+1
View File
@@ -76,6 +76,7 @@ export interface ManagedChannel {
/** When the follow reached this list: the subscription import that first saw it, or a
* subscribe made here. */
added_at: string;
/** The lengths it knows; a video with none (a live one, a 24/7 archive) is left out. */
total_duration_seconds: number;
priority: number;
hidden: boolean;
frontend/src/lib/api/playlists.ts
+2
View File
@@ -17,6 +17,8 @@ export interface Playlist {
dirty: boolean; // linked local playlist has edits not yet pushed to YouTube
item_count: number;
total_duration_seconds: number;
/** The videos the total left out for having no length: the "+N without length" beside it. */
without_length: number;
cover_thumbnail: string | null;
updated_at: string; // ISO; when the list last CHANGED (rename, add/remove/move/reorder)
has_video?: boolean; // only set when listed with ?contains=<video_id>
frontend/src/lib/hlsEngine.test.ts
+184
View File
@@ -46,6 +46,7 @@ class FakeHls implements HlsPort {
loaded: string | null = null;
destroyed = false;
audioTrack = 0;
resumes = 0;
constructor(readonly handlers: HlsHandlers) {}
load(url: string) {
this.loaded = url;
@@ -53,6 +54,24 @@ class FakeHls implements HlsPort {
destroy() {
this.destroyed = true;
}
resume() {
this.resumes++;
}
/** hls.js gave up on a request that never answered in time: no response, so no status. */
timedOut(details = "fragLoadTimeOut") {
this.handlers.onError({
fatal: true,
network: true,
status: null,
timeout: true,
type: "networkError",
details,
});
}
/** A fragment arrived: media is getting through. */
arrived() {
this.handlers.onFragLoaded({ type: "main", bytes: 1_000_000, durationS: 6 });
}
getAudioTrack() {
return this.audioTrack;
}
@@ -72,6 +91,7 @@ class FakeHls implements HlsPort {
fatal: false,
network: true,
status: 410,
timeout: false,
type: "networkError",
details: "fragLoadError",
});
@@ -82,6 +102,7 @@ class FakeHls implements HlsPort {
fatal: false,
network: true,
status,
timeout: false,
type: "networkError",
details: "fragLoadError",
});
@@ -92,6 +113,7 @@ class FakeHls implements HlsPort {
fatal: true,
network: true,
status,
timeout: false,
type: "networkError",
details: "levelLoadError",
});
@@ -579,6 +601,7 @@ describe("a server that is restarting (a deploy)", () => {
fire();
await asks[1]!.resolve(hls("A"));
hlses[1]!.handlers.onManifestParsed();
hlses[1]!.arrived(); // back: media gets through again
advance(AWAY_WINDOW_MS + 1);
hlses[1]!.lost(502); // a second deploy, long after the first
expect(engine.getSnapshot().error).toBeNull();
@@ -608,6 +631,167 @@ describe("a server that is restarting (a deploy)", () => {
hlses[0]!.lost(500);
expect(engine.getSnapshot().error).toBe("generic");
});
it("takes a direct stream's answer for the server being back: it has no fragments to show it", async () => {
const { engine, asks, fire, advance } = setup();
engine.start("A", 0);
await asks[0]!.reject({ status: 502 });
fire();
await asks[1]!.resolve(direct("A"));
advance(AWAY_WINDOW_MS + 1);
engine.start("A", 100); // a second deploy, long after the first
await asks[2]!.reject({ status: 502 });
expect(engine.getSnapshot().error).toBeNull();
});
it("ends in an error when the server answers every session but none of them delivers", async () => {
// Every answer used to start the absence over, so this asked for a new session forever.
const { engine, asks, hlses, fire } = setup();
engine.start("A", 0);
await asks[0]!.resolve(hls("A"));
hlses[0]!.handlers.onManifestParsed();
hlses[0]!.lost(0);
for (let i = 1; i < 40 && fire(); i++) {
await asks[i]!.resolve(hls("A", 0, `s-${i}`));
hlses[i]!.handlers.onManifestParsed();
hlses[i]!.lost(0);
}
expect(engine.getSnapshot().error).toBe("generic");
expect(asks.length).toBeLessThan(40);
});
});
describe("a link too slow for the stream", () => {
// Measured on prod (2026-11-04, a work network at ~4 Mbit/s against 11–16 MB remux segments): an
// audio segment waited behind a video one past hls.js's time to first byte, five times, and gave up
// with no answer. That was taken for a server away: a new session, the same segment lost again, a
// restart every half minute until the viewer closed the player.
it("resumes loading in the same session, and asks for no new one", async () => {
const { engine, asks, hlses, timers } = setup();
engine.start("A", 453);
await asks[0]!.resolve(hls("A", 453));
hlses[0]!.handlers.onManifestParsed();
hlses[0]!.timedOut();
expect(asks).toHaveLength(1);
expect(timers.filter((t) => t.armed)).toHaveLength(0);
expect(hlses[0]!.resumes).toBe(1);
expect(hlses[0]!.destroyed).toBe(false);
expect(engine.getSnapshot()).toMatchObject({ ready: true, error: null });
});
it.each(["fragLoadTimeOut", "levelLoadTimeOut", "audioTrackLoadTimeOut"])(
"treats a %s the same way",
async (details) => {
const { engine, asks, hlses } = setup();
engine.start("A", 0);
await asks[0]!.resolve(hls("A"));
hlses[0]!.timedOut(details);
expect(asks).toHaveLength(1);
expect(hlses[0]!.resumes).toBe(1);
},
);
it("keeps going for as long as media still arrives, however slowly", async () => {
const { engine, asks, hlses, advance } = setup();
engine.start("A", 0);
await asks[0]!.resolve(hls("A"));
hlses[0]!.handlers.onManifestParsed();
for (let i = 0; i < 10; i++) {
hlses[0]!.timedOut();
advance(60_000);
hlses[0]!.arrived();
}
hlses[0]!.timedOut();
expect(engine.getSnapshot().error).toBeNull();
expect(asks).toHaveLength(1);
});
it("says the connection is too slow once nothing has arrived for the whole window", async () => {
const { engine, asks, hlses, advance } = setup();
engine.start("A", 0);
await asks[0]!.resolve(hls("A"));
hlses[0]!.handlers.onManifestParsed();
hlses[0]!.timedOut();
advance(AWAY_WINDOW_MS + 1);
hlses[0]!.timedOut();
expect(engine.getSnapshot().error).toBe("slow");
expect(hlses[0]!.resumes).toBe(1);
expect(asks).toHaveLength(1);
});
it("asks for the session again when its playlist never answered", async () => {
// Loading cannot resume what never parsed: hls.js reloads no manifest on startLoad, so a resume
// here sent nothing, raised nothing, and left the spinner up for good.
const { engine, asks, hlses, fire } = setup();
engine.start("A", 300);
await asks[0]!.resolve(hls("A", 300));
hlses[0]!.timedOut("manifestLoadTimeOut");
expect(fire()).toBe(true);
expect(asks).toHaveLength(2);
expect(asks[1]!.startAt).toBe(300);
await asks[1]!.resolve(hls("A", 300));
hlses[1]!.handlers.onManifestParsed();
expect(engine.getSnapshot()).toMatchObject({ ready: true, error: null });
});
it("says it cannot play once the playlist has not answered for the whole window", async () => {
const { engine, asks, hlses, fire, advance } = setup();
engine.start("A", 0);
await asks[0]!.resolve(hls("A"));
hlses[0]!.timedOut("manifestLoadTimeOut");
advance(AWAY_WINDOW_MS + 1);
expect(fire()).toBe(true);
await asks[1]!.resolve(hls("A"));
hlses[1]!.timedOut("manifestLoadTimeOut");
expect(engine.getSnapshot().error).not.toBeNull();
});
});
describe("what a fragment reports", () => {
it("is passed on from the session asked for only, never from one it replaced", async () => {
const seen: number[] = [];
const { engine, asks, hlses } = setup({ onFragLoaded: (f) => void seen.push(f.bytes) });
engine.start("A", 0);
await asks[0]!.resolve(hls("A"));
hlses[0]!.arrived();
engine.seek("A", 2000);
await asks[1]!.resolve(hls("A", 2000));
hlses[0]!.arrived(); // the replaced session's last fragment lands late
expect(seen).toEqual([1_000_000]);
hlses[1]!.arrived();
expect(seen).toEqual([1_000_000, 1_000_000]);
});
});
describe("an absence when no hls.js plays the stream", () => {
it("ends with an answer the browser plays itself (native HLS)", async () => {
// Without MSE an HLS answer plays natively and no fragment is ever reported, so the answer is
// the only sign the server is back: a later absence gets its own window.
const { engine, asks, advance, fire } = setup({ createHls: () => null });
engine.start("A", 0);
await asks[0]!.reject({ status: 503 });
expect(fire()).toBe(true);
await asks[1]!.resolve(hls("A"));
advance(AWAY_WINDOW_MS + 60_000);
engine.seek("A", 900);
await asks[2]!.reject({ status: 503 });
expect(engine.getSnapshot().error).toBeNull();
expect(fire()).toBe(true);
});
it("starts afresh when the viewer tries again after the error was shown", async () => {
const { engine, asks, advance, fire } = setup();
engine.start("A", 0);
await asks[0]!.reject({ status: 503 });
advance(AWAY_WINDOW_MS + 1);
expect(fire()).toBe(true);
await asks[1]!.reject({ status: 503 });
expect(engine.getSnapshot().error).toBe("generic");
engine.start("A", 0); // the viewer presses play again
await asks[2]!.reject({ status: 503 });
expect(engine.getSnapshot().error).toBeNull();
expect(fire()).toBe(true);
});
});
describe("play intent while a session is asked for", () => {
frontend/src/lib/hlsEngine.ts
+61 -6
View File
@@ -30,14 +30,22 @@ export const RECLAIM_WINDOW_MS = 120_000;
// gone with it, but the buffer keeps playing — so the engine asks again where the viewer is, less and
// less often, and says it cannot play only once the server has been away this long. It used to say
// so at the first 502, over a picture still playing from the buffer, and never asked again.
//
// Under hls.js the absence ends when MEDIA arrives again (a fragment loads), not when a session
// request is answered: a server that answers every session and delivers none of them is still not
// serving, and ending the absence at each answer restarted the stream forever. What hls.js does not
// play (a direct file, HLS the browser plays natively) reports no fragments, so there the answer
// ends it (`answered`); and the viewer trying again after the error starts a window of its own
// (`start`). The same window bounds a link too slow for the stream (see `slowLink`).
export const AWAY_WINDOW_MS = 180_000;
const AWAY_RETRY_MS = [2_000, 4_000, 8_000, 15_000];
const SERVER_AWAY = new Set([0, 502, 503, 504]);
// The server's answer for a session it no longer has; its 404 means a segment not there yet.
const SESSION_GONE = 410;
/** Why a session could not play, in the three buckets the player has a sentence for. */
type EngineError = "transcode" | "notFound" | "generic";
/** Why a session could not play, in the buckets the player has a sentence for. `slow`: the link
* never delivered the stream in time for the whole away window. */
type EngineError = "transcode" | "notFound" | "slow" | "generic";
/** What the player renders from the engine. A new object whenever any of it changes. */
export interface EngineSnapshot {
@@ -55,14 +63,27 @@ export interface HlsFailure {
fatal: boolean;
network: boolean;
status: number | null;
/** A request that got no answer in time (no response, so `status` is null). */
timeout: boolean;
type: string;
details: string;
}
/** A loaded fragment, as the meters read it. `loading`: when its download started and ended
* (performance.now() ms), when known. */
export interface FragLoaded {
type: string;
bytes: number;
durationS: number;
loading?: { start: number; end: number };
}
/** One hls.js instance, narrowed to what the engine uses (hlsPort.ts binds the real one). */
export interface HlsPort {
load(url: string, media: HTMLMediaElement): void;
destroy(): void;
/** Start loading again where it stopped, after a fatal network error (same session). */
resume(): void;
getAudioTrack(): number;
setAudioTrack(track: number): void;
levelInfo(): HlsLevelInfo;
@@ -71,7 +92,7 @@ export interface HlsPort {
export interface HlsHandlers {
onManifestParsed(): void;
onAudioTracksUpdated(): void;
onFragLoaded(frag: { type: string; bytes: number; durationS: number }): void;
onFragLoaded(frag: FragLoaded): void;
onError(e: HlsFailure): void;
}
@@ -106,7 +127,7 @@ export interface EngineDeps {
statusOf(e: unknown): number | undefined;
/** A request for a session went out: the player clears its stall spinner. */
onLoadStart?(): void;
onFragLoaded?(frag: { type: string; bytes: number; durationS: number }): void;
onFragLoaded?(frag: FragLoaded): void;
onHlsError?(e: HlsFailure): void;
}
@@ -299,6 +320,9 @@ export class HlsEngine {
if (quiet) {
this.set({ error: null });
} else {
// The viewer trying again after the error was shown: a new attempt, with a whole window of its
// own — the spent one turned its first failure straight into the error again.
if (this.snapshot.error) this.away = null;
this.live = false;
this.deps.onLoadStart?.();
// Clear the prior session's transcode flag up front, so the cold-start overlay cannot show a
@@ -358,7 +382,6 @@ export class HlsEngine {
}
const media = this.deps.media();
if (!media) return;
this.away = null; // the server answered: a later absence is a new one
const { intent, startAt, id } = req;
// The stream that kept playing is replaced from here: the new one loads like any other.
if (req.quiet) this.set({ ready: false });
@@ -391,6 +414,10 @@ export class HlsEngine {
// Named for the stop on the way out whoever plays it: hls.js, or Safari's native HLS below.
if (sess.mode === "hls" && sess.id) this.attachedSession = { id, sessionId: sess.id };
const hls = sess.mode === "hls" ? this.deps.createHls(this.hlsHandlers(req, media)) : null;
// What hls.js does not play reports no fragments — a direct file, or HLS the browser plays
// natively (no MSE) — so its answer is the only sign the server is back. Under hls.js an
// absence ends when a fragment arrives (see `AWAY_WINDOW_MS`).
if (!hls) this.away = null;
if (hls) {
this.hls = hls;
hls.load(`${sess.url}?t=${Math.floor(startAt)}`, media);
@@ -456,7 +483,12 @@ export class HlsEngine {
/* out of range — ignore */
}
},
onFragLoaded: (frag) => this.deps.onFragLoaded?.(frag),
onFragLoaded: (frag) => {
// A superseded session's late fragment is not this stream's measurement.
if (!this.current(req)) return;
this.away = null; // media gets through: whatever was away is back
this.deps.onFragLoaded?.(frag);
},
onError: (e) => this.onHlsError(req, media, e),
};
}
@@ -489,6 +521,12 @@ export class HlsEngine {
return;
}
if (!e.fatal) return;
// A playlist that never answered is not resumed: hls.js reloads no manifest on startLoad, so a
// resume sent nothing and the spinner stayed. It is asked for again, as a server away is.
if (e.network && e.timeout && e.details !== "manifestLoadTimeOut") {
this.slowLink();
return;
}
if (e.network && SERVER_AWAY.has(e.status ?? 0) && this.serverAway(req)) return;
if (e.network && e.status === 404) {
this.reclaim(req, media, false);
@@ -538,6 +576,23 @@ export class HlsEngine {
return true;
}
/** A request got no answer in time: the link is too slow for what the stream asks of it, not the
* server away (measured on prod: a work network at ~4 Mbit/s against 11–16 MB remux segments; an
* audio segment waited behind a video one past hls.js's time to first byte). A new session would
* throw away what had arrived and lose its first segment the same way — a restart every half
* minute, never playing. So loading resumes in the same session, for as long as media still
* arrives; the away window says it is too slow once nothing has for that long. */
private slowLink(): void {
const now = this.deps.now();
const away = (this.away ??= { since: now, tries: 0 });
if (now - away.since > AWAY_WINDOW_MS) {
this.set({ error: "slow" });
return;
}
away.tries += 1;
this.hls?.resume();
}
/** Where the viewer is: the ticked position, or — before playback has ticked (a session still
* loading, or a resumed one that has stayed paused) — where the session was ASKED to start.
* Anything that restarts "here" must use this: reading the ticked 0 restarted a resume at 20:00
frontend/src/lib/hlsPort.test.ts
+97 -15
View File
@@ -1,42 +1,124 @@
// The one place hls.js is built for the Plex player, held against a stand-in hls.js: what config it is
// built with, and what it hands the engine from hls.js's events.
//
// hls.js empties the cues of every text track on the <video> when a manifest loads or the media
// attaches — the Plex player's own subtitle <track> included, so a VTT that loaded first showed
// nothing for the rest of the session. Its subtitle and caption controllers are not built at all.
import { describe, expect, it, vi } from "vitest";
import type { FragLoaded, HlsFailure } from "./hlsEngine";
const built: Record<string, unknown>[] = [];
type Listener = (e: string, data: unknown) => void;
interface Built {
config: Record<string, unknown>;
on: Record<string, Listener>;
starts: number[];
media: { currentTime: number } | null;
}
const built: Built[] = [];
vi.mock("hls.js", () => {
class FakeHls {
static isSupported = () => true;
static Events = {};
static ErrorTypes = {};
static Events = {
MANIFEST_PARSED: "hlsManifestParsed",
AUDIO_TRACKS_UPDATED: "hlsAudioTracksUpdated",
FRAG_LOADED: "hlsFragLoaded",
ERROR: "hlsError",
};
static ErrorTypes = { NETWORK_ERROR: "networkError" };
readonly self: Built;
constructor(config: Record<string, unknown>) {
built.push(config);
this.self = { config, on: {}, starts: [], media: null };
built.push(this.self);
}
get media() {
return this.self.media;
}
on(event: string, fn: Listener) {
this.self.on[event] = fn;
}
startLoad(at: number) {
this.self.starts.push(at);
}
on() {}
}
return { default: FakeHls };
});
const { createHlsPort } = await import("./hlsPort");
function build() {
const frags: FragLoaded[] = [];
const errors: HlsFailure[] = [];
const noop = () => {};
const port = createHlsPort({
onManifestParsed: noop,
onAudioTracksUpdated: noop,
onFragLoaded: (f) => void frags.push(f),
onError: (e) => void errors.push(e),
})!;
return { port, hls: built[built.length - 1]!, frags, errors };
}
describe("createHlsPort", () => {
it("builds hls.js without the controllers that clear every text track's cues", () => {
const noop = () => {};
createHlsPort({
onManifestParsed: noop,
onAudioTracksUpdated: noop,
onFragLoaded: noop,
onError: noop,
});
const config = built[0] ?? {};
const { hls } = build();
for (const key of [
"subtitleTrackController",
"subtitleStreamController",
"timelineController",
]) {
// Set, as undefined: hls.js spreads the config over its defaults, so an absent key keeps them.
expect(key in config).toBe(true);
expect(config[key]).toBeUndefined();
expect(key in hls.config).toBe(true);
expect(hls.config[key]).toBeUndefined();
}
});
it("gives a fragment 30 s to its first byte: a small audio one waits behind a big remux one", () => {
const { hls } = build();
const policy = hls.config.fragLoadPolicy as { default: { maxTimeToFirstByteMs: number } };
expect(policy.default.maxTimeToFirstByteMs).toBe(30_000);
});
it("names a request that got no answer in time a timeout, and an answered failure not", () => {
const { hls, errors } = build();
const err = (details: string, code?: number) =>
hls.on.hlsError!("hlsError", {
fatal: true,
type: "networkError",
details,
response: code ? { code } : undefined,
});
err("fragLoadTimeOut");
err("levelLoadTimeOut");
err("fragLoadError", 404);
expect(errors.map((e) => [e.timeout, e.status])).toEqual([
[true, null],
[true, null],
[false, 404],
]);
});
it("hands on when a fragment's download ran", () => {
const { hls, frags } = build();
const frag = (loading?: { start: number; end: number }) =>
hls.on.hlsFragLoaded!("hlsFragLoaded", {
frag: { type: "main", duration: 6, stats: { total: 16_000_000, loading } },
});
frag({ start: 1000, end: 32_200 });
frag(undefined);
frag({ start: 0, end: 0 }); // not timed
expect(frags.map((f) => f.loading)).toEqual([
{ start: 1000, end: 32_200 },
undefined,
undefined,
]);
});
it("resumes loading where the media is — at 0 for one that never ticked, not the live edge", () => {
const { port, hls } = build();
hls.media = { currentTime: 0 };
port.resume();
hls.media = { currentTime: 312.5 };
port.resume();
expect(hls.starts).toEqual([0, 312.5]);
});
});
frontend/src/lib/hlsPort.ts
+26 -4
View File
@@ -17,10 +17,23 @@ export function createHlsPort(handlers: HlsHandlers): HlsPort | null {
// a <track> of ours, out of band, and hls.js empties the cues of EVERY text track on the element
// when a manifest loads or the media attaches/detaches — ours included, so a VTT that loaded before
// the manifest showed nothing for the rest of the session.
//
// A fragment gets 30 s to its first byte, not hls.js's 10: a remux segment is the source's own
// bitrate (11–16 MB measured on prod), and on a slow link (a work network at ~4 Mbit/s) the small
// audio segment beside it waits behind it longer than 10 s — five such waits ended the load. The
// rest is hls.js's default policy.
const hls = new Hls({
maxBufferLength: 30,
enableWorker: true,
startPosition: 0,
fragLoadPolicy: {
default: {
maxTimeToFirstByteMs: 30_000,
maxLoadTimeMs: 120_000,
timeoutRetry: { maxNumRetry: 4, retryDelayMs: 0, maxRetryDelayMs: 0 },
errorRetry: { maxNumRetry: 6, retryDelayMs: 1000, maxRetryDelayMs: 8000 },
},
},
subtitleTrackController: undefined,
subtitleStreamController: undefined,
timelineController: undefined,
@@ -28,19 +41,24 @@ export function createHlsPort(handlers: HlsHandlers): HlsPort | null {
hls.on(Hls.Events.MANIFEST_PARSED, () => handlers.onManifestParsed());
hls.on(Hls.Events.AUDIO_TRACKS_UPDATED, () => handlers.onAudioTracksUpdated());
// The delivered bitrate is measured per fragment: the master playlist's own BANDWIDTH is not a
// measurement (see plexStats.ts).
hls.on(Hls.Events.FRAG_LOADED, (_e, data) =>
// measurement (see plexStats.ts). When its download ran says how fast the link delivers.
hls.on(Hls.Events.FRAG_LOADED, (_e, data) => {
const l = data.frag.stats?.loading;
handlers.onFragLoaded({
type: data.frag.type,
bytes: data.frag.stats?.total || data.frag.stats?.loaded || 0,
durationS: data.frag.duration,
}),
);
loading: l && l.end > l.start ? { start: l.start, end: l.end } : undefined,
});
});
hls.on(Hls.Events.ERROR, (_e, data) =>
handlers.onError({
fatal: data.fatal,
network: data.type === Hls.ErrorTypes.NETWORK_ERROR,
status: data.response?.code ?? null,
// hls.js names every no-answer-in-time failure `…TimeOut` (fragLoadTimeOut, levelLoadTimeOut,
// audioTrackLoadTimeOut, manifestLoadTimeOut, …).
timeout: data.details.endsWith("TimeOut"),
type: data.type,
details: data.details,
}),
@@ -51,6 +69,10 @@ export function createHlsPort(handlers: HlsHandlers): HlsPort | null {
hls.attachMedia(media);
},
destroy: () => hls.destroy(),
// hls.js's own recovery after a fatal network error: load again from where the media is. The
// position is passed: with none, a session that never ticked would start at the live edge (see
// `startPosition: 0` above).
resume: () => hls.startLoad(hls.media?.currentTime ?? 0),
getAudioTrack: () => hls.audioTrack,
setAudioTrack(track) {
hls.audioTrack = track;
frontend/src/lib/plexPlayerHooks.slowLink.dom.test.ts
+60
View File
@@ -0,0 +1,60 @@
// @vitest-environment jsdom
//
// The player says the link is too slow only when a stall has lasted and the measured arrival rate is
// below what the stream needs (measured on prod: ~4 Mbit/s arriving against a ~20 Mbit/s remux).
import { act, renderHook } from "@testing-library/react";
import { afterEach, beforeEach, describe, expect, it, vi } from "vitest";
import { SLOW_LINK_AFTER_MS, slowLinkRates, useSlowLink, type LinkRates } from "./plexPlayerHooks";
describe("slowLinkRates", () => {
it("names both rates only when the link delivers slower than the stream needs", () => {
expect(slowLinkRates(20e6, 4e6)).toEqual({ need: 20e6, got: 4e6 });
expect(slowLinkRates(20e6, 25e6)).toBeNull();
});
it("says nothing before both are measured", () => {
expect(slowLinkRates(undefined, 4e6)).toBeNull();
expect(slowLinkRates(20e6, undefined)).toBeNull();
});
});
describe("useSlowLink", () => {
beforeEach(() => vi.useFakeTimers());
afterEach(() => vi.useRealTimers());
const slow: LinkRates = { need: 20e6, got: 4e6 };
it("shows nothing for a short stall, the rates once it has lasted", () => {
const { result } = renderHook(({ stalled }) => useSlowLink(stalled, () => slow), {
initialProps: { stalled: true },
});
act(() => void vi.advanceTimersByTime(SLOW_LINK_AFTER_MS - 1));
expect(result.current).toBeNull();
act(() => void vi.advanceTimersByTime(1));
expect(result.current).toEqual(slow);
});
it("follows the measurement while the stall lasts", () => {
let rates: LinkRates | null = slow;
const { result } = renderHook(() => useSlowLink(true, () => rates));
act(() => void vi.advanceTimersByTime(SLOW_LINK_AFTER_MS));
expect(result.current).toEqual(slow);
rates = null; // the link caught up
act(() => void vi.advanceTimersByTime(2000));
expect(result.current).toBeNull();
});
it("goes when the stall ends, and the next stall waits its own 5 s", () => {
const { result, rerender } = renderHook(({ stalled }) => useSlowLink(stalled, () => slow), {
initialProps: { stalled: true },
});
act(() => void vi.advanceTimersByTime(SLOW_LINK_AFTER_MS));
expect(result.current).toEqual(slow);
rerender({ stalled: false });
expect(result.current).toBeNull();
rerender({ stalled: true });
expect(result.current).toBeNull();
act(() => void vi.advanceTimersByTime(SLOW_LINK_AFTER_MS));
expect(result.current).toEqual(slow);
});
});
frontend/src/lib/plexPlayerHooks.ts
+44
View File
@@ -1,6 +1,7 @@
import {
useCallback,
useEffect,
useEffectEvent,
useRef,
useState,
type RefObject,
@@ -47,6 +48,49 @@ export function useWallClock(): Date {
return now;
}
/** How long a stall lasts before the player says the link is too slow: a short rebuffer is ordinary. */
export const SLOW_LINK_AFTER_MS = 5000;
export interface LinkRates {
/** Bits per second the stream needs (its own bitrate, measured from what arrived). */
need: number;
/** Bits per second the link delivered it at. */
got: number;
}
/** Both rates when the link delivers slower than the stream needs; null when it keeps up, or before
* anything has been measured — the player never says "slow" without a number behind it. */
export function slowLinkRates(need: number | undefined, got: number | undefined): LinkRates | null {
return need && got && got < need ? { need, got } : null;
}
/** The slow-link rates to show while `stalled` (loading, or rebuffering mid-play) has lasted
* `SLOW_LINK_AFTER_MS`, re-read every 2 s while it lasts; null otherwise. `read` is called from the
* timers only (it reads the fragment meter, a ref). */
export function useSlowLink(stalled: boolean, read: () => LinkRates | null): LinkRates | null {
const [rates, setRates] = useState<LinkRates | null>(null);
// Every stall starts from nothing: what the previous one measured is not shown before its 5 s.
const [wasStalled, setWasStalled] = useState(stalled);
if (wasStalled !== stalled) {
setWasStalled(stalled);
setRates(null);
}
const readNow = useEffectEvent(() => setRates(read()));
useEffect(() => {
if (!stalled) return;
let iv = 0;
const t = window.setTimeout(() => {
readNow();
iv = window.setInterval(readNow, 2000);
}, SLOW_LINK_AFTER_MS);
return () => {
window.clearTimeout(t);
window.clearInterval(iv);
};
}, [stalled]);
return stalled ? rates : null;
}
// Fullscreen toggle for `el` (the player wrapper) + a reactive `fs` flag.
// Via lib/fullscreen, not the raw API: WebKit (Safari/iPadOS) ships only the prefixed spelling, so
// reading `document.fullscreenElement` alone made this player believe it is never fullscreen — the
frontend/src/lib/plexStats.test.ts
+40
View File
@@ -6,6 +6,7 @@ import {
buildStatRows,
createFragMeter,
fmtBitrate,
fmtBitrates,
fmtBufferSecs,
fmtClock,
fmtCodec,
@@ -28,6 +29,13 @@ describe("plexStats formatters", () => {
expect(fmtBitrate(4_200_000, "hu")).toBe("4,2 Mbps"); // the UI language's decimal mark
});
it("fmtBitrates says rates side by side in the unit the largest takes", () => {
expect(fmtBitrates([1_400_000, 800_000], "en")).toEqual(["1.4 Mbps", "0.8 Mbps"]);
expect(fmtBitrates([1_400_000, 800_000], "hu")).toEqual(["1,4 Mbps", "0,8 Mbps"]);
expect(fmtBitrates([900_000, 400_000], "en")).toEqual(["900 kbps", "400 kbps"]);
expect(fmtBitrates([2_000_000, 0], "en")).toEqual(["2.0 Mbps", DASH]);
});
it("fmtBufferSecs shows one decimal, dashes negatives/non-finite", () => {
expect(fmtBufferSecs(12.34, "en")).toBe("12.3 sec");
expect(fmtBufferSecs(0, "en")).toBe("0.0 sec");
@@ -180,6 +188,38 @@ describe("measured bitrate", () => {
expect(m.bitrate()).toBe(8_000_000 + 200_000);
});
it("measures how fast the bytes arrive, beside how fast the stream needs them", () => {
// The prod case: a 16 MB, 6 s video segment that took 31.2 s to come through.
const m = createFragMeter();
m.add({ type: "main", bytes: 16_000_000, durationS: 6, loading: { start: 0, end: 31_200 } });
expect(m.bitrate()).toBeCloseTo((16_000_000 * 8) / 6);
expect(m.arrival()).toBeCloseTo((16_000_000 * 8) / 31.2);
});
it("counts downloads that overlap once: the renditions share one link", () => {
// Measured live: summing the per-rendition rates read 2.3 Mbit/s through a 1.6 Mbit/s link.
const m = createFragMeter();
m.add({ type: "main", bytes: 1_800_000, durationS: 4, loading: { start: 0, end: 10_000 } });
m.add({ type: "audio", bytes: 100_000, durationS: 4, loading: { start: 2_000, end: 8_000 } });
m.add({ type: "audio", bytes: 100_000, durationS: 4, loading: { start: 9_000, end: 12_000 } });
m.add({
type: "main",
bytes: 2_000_000,
durationS: 4,
loading: { start: 15_000, end: 25_000 },
});
// 4 MB over the 22 s something was downloading (0–12 s and 15–25 s): not the 29 s the downloads
// add up to, nor per type.
expect(m.arrival()).toBeCloseTo((4_000_000 * 8) / 22);
});
it("has no arrival rate for fragments whose download time is unknown", () => {
const m = createFragMeter();
m.add({ type: "main", bytes: 6_000_000, durationS: 6 });
expect(m.bitrate()).toBe(8_000_000);
expect(m.arrival()).toBeUndefined();
});
it("has nothing to say before the first fragment, and after a reset", () => {
const m = createFragMeter();
expect(m.bitrate()).toBeUndefined(); // an honest dash — never the manifest's stand-in number
frontend/src/lib/plexStats.ts
+32
View File
@@ -64,6 +64,16 @@ export function fmtBitrate(bps: number | undefined, lang: string): string {
return `${formatNumber(Math.round(bps / 1000), lang)} kbps`;
}
// Rates read side by side ("needs 1.4 Mbps, 0.8 Mbps arrives"): one unit for all of them, the one
// the largest would take alone, so they compare at a glance — never "1.4 Mbps" beside "800 kbps".
export function fmtBitrates(rates: number[], lang: string): string[] {
const largest = Math.max(0, ...rates);
if (largest < 1_000_000) return rates.map((bps) => fmtBitrate(bps, lang));
return rates.map((bps) =>
bps > 0 ? `${formatDecimal(bps / 1_000_000, lang)} Mbps` : fmtBitrate(bps, lang),
);
}
// Seconds of buffer ahead of the playhead → "12.3 sec" / "12,3 mp".
export function fmtBufferSecs(s: number, lang: string): string {
if (!Number.isFinite(s) || s < 0) return DASH;
@@ -188,12 +198,18 @@ interface FragSample {
type: string; // hls.js fragment type: "main" (video, and muxed audio) / "audio" / "subtitle"
bytes: number;
durationS: number; // MEDIA seconds the fragment carries (not how long it took to download)
loading?: { start: number; end: number }; // when its download ran (ms), when hls.js knew
}
export interface FragMeter {
add(sample: FragSample): void;
/** Bits per second of media, or undefined while nothing has been measured yet. */
bitrate(): number | undefined;
/** Bits per second the link delivered them at, or undefined while no download time is known: every
* timed fragment's bytes over the time ANY of them was downloading. Not per type and summed like
* `bitrate` — the renditions share one link, and their downloads overlap (measured: summing read
* 2.3 Mbit/s through a 1.6 Mbit/s link). */
arrival(): number | undefined;
reset(): void;
}
@@ -222,6 +238,22 @@ export function createFragMeter(window = 10): FragMeter {
}
return total > 0 ? total : undefined;
},
arrival() {
const timed = [...byType.values()]
.flat()
.filter((f) => f.loading && f.loading.end > f.loading.start)
.sort((a, b) => a.loading!.start - b.loading!.start);
let busyMs = 0;
let until = -Infinity;
for (const { loading } of timed) {
const { start, end } = loading!;
if (end <= until) continue;
busyMs += end - Math.max(start, until);
until = end;
}
const bytes = timed.reduce((n, f) => n + f.bytes, 0);
return busyMs > 0 ? (bytes * 8) / (busyMs / 1000) : undefined;
},
reset() {
byType.clear();
},
frontend/src/lib/theme.contract.test.ts
+11
View File
@@ -80,6 +80,17 @@ describe("the media colours", () => {
expect(field).toMatch(/\bfocus-within:ring-focus\b/);
expect(field).not.toMatch(/\bfocus-within:(border|ring)-accent\b/);
});
// Text over the picture sits on a black panel the frame shows through: its worst ground is a white
// frame. The error line on it is the media red (a light theme's danger was 1.6:1 there). [WCAG 1.4.3]
it("an error over the picture holds 4.5:1 on its panel over a white frame", () => {
const player = readFileSync(new URL("../modules/plex/PlexPlayer.tsx", import.meta.url), "utf8");
const over = /const OVER_VIDEO =\s*"([^"]+)"/.exec(player)?.[1] ?? "";
expect(over.split(" ")).toContain("media-scope"); // --danger is the media red under it
const alpha = Number(/\bbg-black\/(\d+)\b/.exec(over)?.[1]) / 100;
const ground = mixHex("#000000", "#ffffff", alpha);
expect(hexContrast(media()["media-danger"]!, ground)).toBeGreaterThanOrEqual(4.5);
expect(hexContrast("#ffffff", ground)).toBeGreaterThanOrEqual(4.5);
});
it("every surface painted the player chrome names its marks: the focus ring and what is chosen", () => {
const block = (name: string) => {
frontend/src/modules/channels/channelColumns.dom.test.tsx
+13
View File
@@ -185,6 +185,19 @@ describe("buildChannelColumns — Added", () => {
render(<>{added!.render(ch({ added_at: "2026-10-12T08:00:00+00:00" }))}</>);
expect(screen.getByText(formatDate("2026-10-12T08:00:00+00:00", "en"))).toBeTruthy();
});
// The column is sized for the sum: a count of videos without a length beside it spilled into
// Tags, so the table shows the sum alone — even from a server that still sends the count.
it("Length shows the sum alone", () => {
const length = cols.find((c) => c.key === "length");
const row = {
...ch({ total_duration_seconds: 7200 }),
without_length: 2,
} as ManagedChannel;
const { container } = render(<>{length!.render(row)}</>);
expect(container.textContent).not.toContain("card.withoutLength");
expect(container.querySelectorAll("span")).toHaveLength(1);
});
});
describe("tagFilterChips", () => {
frontend/src/modules/channels/channelColumns.tsx
+2
View File
@@ -235,6 +235,8 @@ export function buildChannelColumns(deps: ChannelColumnDeps): Column<ManagedChan
sortValue: (c) => c.total_duration_seconds,
render: (c) => (
<span className="text-muted tabular-nums">
{/* The sum alone: the column is sized for it, and a count of the videos it left out
(no length) did not fit beside it. */}
{formatTotalHours(c.total_duration_seconds, lang)}
</span>
),
frontend/src/modules/playlists/Playlists.dom.test.tsx
+17 -4
View File
@@ -134,6 +134,7 @@ const vid = (id: string, position_seconds: number): Video =>
status: "new",
position_seconds,
duration_seconds: 3000,
live_status: "none",
published_at: "2026-01-01T00:00:00Z",
watch_url: `https://www.youtube.com/watch?v=${id}`,
}) as unknown as Video;
@@ -989,13 +990,25 @@ describe("Playlists · the header", () => {
)
).textContent;
it("sums what is known, and says nothing about a runtime nothing sums to", async () => {
it("sums what is known, and names what it left out beside it", async () => {
// Not fetched yet, and not live: left out of the sum all the same, so it is named.
mount({ 1: [vid("A", 0), untimed("B")] });
expect(await meta()).toBe("playlists.itemCount · 50 min");
expect(await meta()).toBe("playlists.itemCount · 50 min card.withoutLength");
cleanup();
mount({ 1: [untimed("C")] });
expect(await meta()).toBe("playlists.itemCount");
mount({ 1: [vid("A", 0)] });
expect(await meta()).toBe("playlists.itemCount · 50 min");
});
it("names a live video the sum cannot hold, alone when nothing summed", async () => {
// A 24/7 stream's archive comes with no length (the server's rule): it is named, not summed.
const archive = (id: string) => ({ ...untimed(id), live_status: "was_live" }) as Video;
mount({ 1: [vid("A", 0), archive("B")] });
expect(await meta()).toBe("playlists.itemCount · 50 min card.withoutLength");
cleanup();
mount({ 1: [archive("C")] });
expect(await meta()).toBe("playlists.itemCount · card.withoutLength");
});
it("offers the view controls only once there is more than one row to arrange", async () => {
frontend/src/modules/playlists/Playlists.tsx
+6 -2
View File
@@ -68,6 +68,7 @@ import {
playlistViewParamUpdates,
} from "../../lib/playlistsUrlState";
import { formatRuntime } from "../../lib/format";
import { isWithoutLength, withLeftOutCount } from "../../features/video/withoutLength";
import { applyView, sameOrder } from "../../lib/playlistView";
import { useOpenLinkedPlayer } from "../../features/player/useOpenLinkedPlayer";
const PlayerModal = lazy(() => import("../../features/player/PlayerModal"));
@@ -565,11 +566,13 @@ function PlaylistDetail() {
}, [items]);
// Total playtime of the whole playlist (sums every item, matching the itemCount beside it — not
// the filtered `shown` set). Items with an unknown duration count as 0; hidden when nothing sums.
// the filtered `shown` set). An item with an unknown duration is left out of the sum and named
// beside it instead ("+2 without length"); hidden when there is neither.
const totalSeconds = useMemo(
() => items.reduce((acc, v) => acc + (v.duration_seconds ?? 0), 0),
[items],
);
const leftOutCount = useMemo(() => items.filter(isWithoutLength).length, [items]);
// The row's "…" opens the row menu under itself. Which row's menu is open is part of the
// cell (aria-expanded), so the columns rebuild when it changes — a cheap array of closures.
@@ -1619,7 +1622,8 @@ function PlaylistDetail() {
>
<span className="text-xs text-muted">
{t("playlists.itemCount", { count: items.length })}
{totalSeconds > 0 && ` · ${formatRuntime(totalSeconds, lang)}`}
{(totalSeconds > 0 || leftOutCount > 0) &&
` · ${withLeftOutCount(formatRuntime(totalSeconds, lang), leftOutCount, t)}`}
{filterActive &&
` · ${t("playlists.filteredCount", { count: shown.length })}`}
</span>
frontend/src/modules/playlists/PlaylistsRail.drop.dom.test.tsx
+1
View File
@@ -75,6 +75,7 @@ const pl = (id: number, name: string) =>
yt_playlist_id: null,
item_count: 1,
total_duration_seconds: 0,
without_length: 0,
cover_thumbnail: null,
dirty: false,
}) as unknown as Playlist;
frontend/src/modules/playlists/PlaylistsRail.rename.dom.test.tsx
+1
View File
@@ -80,6 +80,7 @@ const alpha = {
source: "local",
item_count: 3,
total_duration_seconds: 0,
without_length: 0,
cover_thumbnail: null,
dirty: false,
} as unknown as Playlist;
frontend/src/modules/playlists/PlaylistsRail.tsx
+7 -2
View File
@@ -8,6 +8,7 @@ import { YouTubeIcon } from "../../lib/brandIcons";
import { api, type Playlist } from "../../lib/api";
import { compareText } from "../../lib/collate";
import { formatRuntime } from "../../lib/format";
import { withLeftOutCount } from "../../features/video/withoutLength";
import { msg, notify } from "../../lib/notifications";
import { notifyYouTubeActionError } from "../../lib/youtubeErrors";
import { LS } from "../../lib/storage";
@@ -378,8 +379,12 @@ export default function PlaylistsRail() {
</div>
<div className="text-2xs text-muted">
{t("playlists.itemCount", { count: pl.item_count })}
{pl.total_duration_seconds > 0 &&
` · ${formatRuntime(pl.total_duration_seconds, lang)}`}
{(pl.total_duration_seconds > 0 || pl.without_length > 0) &&
` · ${withLeftOutCount(
formatRuntime(pl.total_duration_seconds, lang),
pl.without_length,
t,
)}`}
</div>
</div>
</Pressable>
frontend/src/modules/plex/PlexPlayer.hls.dom.test.tsx
+73 -1
View File
@@ -10,7 +10,12 @@ vi.mock("react-i18next", async (importOriginal) => {
const actual = await importOriginal<typeof import("react-i18next")>();
return {
...actual,
useTranslation: () => ({ t: (key: string) => key, i18n: { language: "en" } }),
// The slow line's rates are shown with the key: they are what its test reads.
useTranslation: () => ({
t: (key: string, o?: { need?: string; got?: string }) =>
key === "plex.player.slowLinkRates" ? `${key} ${o?.need} / ${o?.got}` : key,
i18n: { language: "en" },
}),
};
});
@@ -279,3 +284,70 @@ describe("an HLS segment that never arrives (a fatal 404)", () => {
expect(session).toHaveBeenCalledTimes(3);
});
});
describe("the slow-connection line", () => {
it("says both rates in one unit, so they compare at a glance", async () => {
vi.useFakeTimers({
shouldAdvanceTime: true,
toFake: ["Date", "setTimeout", "clearTimeout", "setInterval", "clearInterval"],
});
try {
const session = mount();
session.mockResolvedValueOnce({
mode: "hls",
url: "/hls/A/index.m3u8",
start_s: 600,
} as never);
await waitFor(() => expect(hlsInstances).toHaveLength(1));
// 1.05 MB of a 6 s fragment (1.4 Mbit/s) that took 10.5 s to arrive (0.8 Mbit/s): under a
// megabit, which on its own would read in kbps.
act(() =>
hlsInstances[0]!.emit(Hls.Events.FRAG_LOADED, {
frag: {
type: "main",
duration: 6,
stats: { total: 1_050_000, loading: { start: 0, first: 10, end: 10_500 } },
},
}),
);
act(() => void vi.advanceTimersByTime(5_100));
expect(screen.getByText("plex.player.slowLinkRates 1.4 Mbps / 0.8 Mbps")).toBeTruthy();
} finally {
vi.useRealTimers();
}
});
it("measures only the stream being asked for, not the one before it", async () => {
// A restart that a direct file answers builds no hls.js, so a meter reset only there kept the
// slow HLS session's numbers on screen under the next stream.
vi.useFakeTimers({
shouldAdvanceTime: true,
toFake: ["Date", "setTimeout", "clearTimeout", "setInterval", "clearInterval"],
});
try {
const session = mount();
session
.mockResolvedValueOnce({ mode: "hls", url: "/hls/A/index.m3u8", start_s: 600 } as never)
.mockResolvedValueOnce({ mode: "direct", url: "/media/A", start_s: 0 } as never);
await waitFor(() => expect(hlsInstances).toHaveLength(1));
// 6 MB of a 6 s fragment (8 Mbit/s) that took 12 s to arrive (4 Mbit/s).
act(() =>
hlsInstances[0]!.emit(Hls.Events.FRAG_LOADED, {
frag: {
type: "main",
duration: 6,
stats: { total: 6_000_000, loading: { start: 0, first: 10, end: 12_000 } },
},
}),
);
act(() => void vi.advanceTimersByTime(5_100));
expect(screen.getByText("plex.player.slowLink")).toBeTruthy();
act(() => hlsInstances[0]!.emit(Hls.Events.ERROR, lost404)); // restarted: answered direct
await waitFor(() => expect(session).toHaveBeenCalledTimes(2));
act(() => void vi.advanceTimersByTime(2_100));
expect(screen.queryByText("plex.player.slowLink")).toBeNull();
} finally {
vi.useRealTimers();
}
});
});
frontend/src/modules/plex/PlexPlayer.tsx
+44 -10
View File
@@ -53,7 +53,7 @@ import ClipPanel, { ClipCropFrame, type ClipCrop } from "../../features/player/C
import { markInAt, markOutAt, openAt, type ClipRange } from "../../lib/clipRange";
import { PlexPlayerSettings } from "./PlexPlayerSettings";
import { PlexStatsOverlay } from "./PlexStatsOverlay";
import { createFragMeter } from "../../lib/plexStats";
import { createFragMeter, fmtBitrates } from "../../lib/plexStats";
import { episodeLine } from "../../lib/plexDetail";
import { HlsEngine } from "../../lib/hlsEngine";
import { createHlsPort } from "../../lib/hlsPort";
@@ -70,7 +70,9 @@ import {
usePlayerVolume,
usePlexProgress,
usePlexTracks,
useSlowLink,
useWallClock,
slowLinkRates,
} from "../../lib/plexPlayerHooks";
import { canFullscreenElements, fullscreenElement } from "../../lib/fullscreen";
import {
@@ -116,6 +118,12 @@ const BACK_TO_PREV_S = 3;
const PLAYER_MENU =
"rounded-xl border border-white/15 bg-media-chrome/95 media-marks text-sm text-white shadow-2xl";
const PLAYER_ROW = "block w-full rounded-lg px-2 py-1.5 text-left hover:bg-white/10";
// Text over the picture (the loading/stall/error overlay, the subtitle notice, the track OSD) sits on
// its own dark panel: bare white or red text was unreadable over a bright frame (measured live on a
// white-lit scene). Black 85 % and `media-scope` (its --danger is the media red) keep the white text
// and an error line over 4.5:1 even on a white frame (theme.contract.test.ts): at 75 % with a light
// theme's dark red the error was 1.6:1. The shape (pill or panel) is each site's own.
const OVER_VIDEO = "media-scope border border-white/15 bg-black/85 text-white shadow-2xl";
// A track choice: white like the rest, the chosen one marked by the tick in the leading slot (DESIGN
// "one of several"), which `media-marks` paints the media amber — a light theme's accent as the
// label's colour was dark on the chrome (Slate 2.84:1), and the highlight's white washed it out.
@@ -254,7 +262,12 @@ export default function PlexPlayer({
// `req` rethrows fetch's own TypeError when nothing answered at all: that is a status of 0.
statusOf: (e) =>
e instanceof HttpError ? e.status : e instanceof TypeError ? 0 : undefined,
onLoadStart: () => setBuffering(false),
// Every stream asked for (another item, a seek, a restart) is measured from nothing: a
// direct file answering it builds no hls.js, and must not show the previous one's numbers.
onLoadStart: () => {
setBuffering(false);
fragMeterRef.current.reset();
},
onFragLoaded: (frag) => fragMeterRef.current.add(frag),
// Fatal ones are what the viewer saw fail; the recoverable ones hls.js retries by itself only
// matter to a detailed diagnosis.
@@ -286,9 +299,19 @@ export default function PlexPlayer({
? t("plex.player.errTranscode")
: engineError === "notFound"
? t("plex.player.errNotFound")
: engineError
? t("plex.player.errGeneric")
: null;
: engineError === "slow"
? t("plex.player.errSlow")
: engineError
? t("plex.player.errGeneric")
: null;
// A stall that lasts, on a link measured slower than the stream: said under the spinner, with both
// numbers, so the viewer can tell their connection from a broken player.
const slowRates = useSlowLink((!ready || buffering) && !loadError, () =>
slowLinkRates(fragMeterRef.current.bitrate(), fragMeterRef.current.arrival()),
);
const [slowNeed, slowGot] = slowRates
? fmtBitrates([slowRates.need, slowRates.got], i18n.language)
: [];
// Refs for values read inside stable callbacks / the keyboard handler (avoid stale closures).
const barAutoHideRef = useLatest(prefs.barAutoHide);
// The two language prefs are read only at ITEM LOAD (the detail effect below), so they go through
@@ -1319,14 +1342,23 @@ export default function PlexPlayer({
className="absolute inset-0 grid place-items-center px-6 text-center text-sm pointer-events-none"
>
{loadError ? (
<span className="max-w-md text-danger">{loadError}</span>
<div className={`max-w-md rounded-xl px-4 py-3 ${OVER_VIDEO}`}>
<span className="text-danger">{loadError}</span>
</div>
) : (
<div className="flex flex-col items-center gap-3 text-white/80">
<div className={`flex flex-col items-center gap-3 rounded-xl px-5 py-4 ${OVER_VIDEO}`}>
<Loader2 className="w-10 h-10 animate-spin" aria-hidden />
{/* Announce the stall to assistive tech. Initial load shows visible text; a mid-playback
rebuffer/seek (ready, no error) has no visible label, so give it an sr-only one via the
aria-live region so a screen reader distinguishes "buffering" from a frozen frame. */}
{!ready ? (
{slowRates ? (
<>
<span>{t("plex.player.slowLink")}</span>
<span className="text-xs text-white/70">
{t("plex.player.slowLinkRates", { need: slowNeed, got: slowGot })}
</span>
</>
) : !ready ? (
<span>
{t(sessionIsTranscode ? "plex.player.preparing" : "plex.player.loading")}
</span>
@@ -1350,7 +1382,9 @@ export default function PlexPlayer({
aria-live="polite"
>
{selectedSub && subNotice && (
<div className="flex max-w-full items-center gap-2 truncate rounded-full border border-white/15 bg-black/75 px-4 py-2 text-sm text-white shadow-2xl">
<div
className={`flex max-w-full items-center gap-2 truncate rounded-full px-4 py-2 text-sm ${OVER_VIDEO}`}
>
{subNotice !== "preparing" ? (
<span className="text-danger">{t(`plex.player.subUnavailable.${subNotice}`)}</span>
) : (
@@ -1376,7 +1410,7 @@ export default function PlexPlayer({
aria-live="polite"
>
{osd && (
<div className="max-w-full truncate rounded-full border border-white/15 bg-black/75 px-4 py-2 text-sm text-white shadow-2xl">
<div className={`max-w-full truncate rounded-full px-4 py-2 text-sm ${OVER_VIDEO}`}>
{osd}
</div>
)}