Sixteen items written from real use: eleven fixed, one measured and closed without a code change
(the micro-stutter is a 23.976 fps film on a 120.0148 Hz display, reproduced identically outside
Siftlode), two parked for a reproducible case, one scoped out to its own sprint. Two of the eleven
were destroying user state — a YouTube video that rewound itself on reopen, and a "clear resume
point" that told Plex nothing about an in-progress item.
Then seven deep-review rounds over the fixes and over each other's fixes: 32 → 24 → 43 → 31 → 22 →
26 → 18 raised. Minor, not patch: item 14 added an admin knob, item 10 a sort control, and items
2/4/5/7 are features.
Round 6 converted three of four `is_demo` 403s to a structured detail and wrote a docstring saying it
had converted all the sites there are. The account-delete route in `routes/me.py` still answered a
hand-written English sentence, which `notifyActionError` discards in favour of the caller's
translated fallback — so the demo user gets no hint that the demo account is the reason. It is
UI-gated today, so the string is unreachable; what was reachable was the false completeness claim,
which is precisely what stops the next reader from looking. The site is converted, and the list is
now DERIVED from the source: a new test walks `app/` for every `is_demo`-guarded 403, asserts each
raises a structured detail, and fails when the count moves — so this file cannot fall behind the code
the way it just did.
The same round's frontend contract guard was failing open. Its pattern required a bare identifier
before `.detail` and excluded `.detail` used as a last call argument, so `setErr(e.detail)`,
`(e as HttpError).detail`, `String(e.detail)` and four more idiomatic forms were invisible — and
because the file set is compared with `toEqual`, a file the pattern cannot see is simply never listed
and the gate stays green over a fresh copy of the contract. Widened to cover cast receivers, optional
chains and call-argument reads; the pinned set is unchanged, so the narrowing had removed no false
positive. A destructured `detail` stays out of reach (a pattern loose enough for it fires on
react-query's own `const { data: detail }`), and that hole is written into the test rather than
papered over with two innocent files.
Its anti-vacuity guard also did not guard anything: the comment says the glob must have matched
something, but the assertion under it read a sibling file by its own URL, which resolves whether or
not the glob enumerated anything. That only mattered while the pinned list was non-empty — the moment
the deferral sweep lands and it becomes [], the whole file would pass over a glob matching zero
files. The scan size is asserted now.
`commit`'s success handler writes the dragged list into the detail cache — round 6's fix, so a
re-entry inside the 30s staleTime renders the accepted order without a request — but then invalidated
only the LIST key, while the YouTube sibling it cites invalidates both. `next` is the membership of
the render that started the drag, and `reorder_playlist` only touches the rating_keys it is handed,
so it never ratifies that list: anything that landed meanwhile was clobbered by the write with
nothing scheduled to correct it. `commit` sets no busy flag either, so a removal really can complete
while a reorder is in flight, and with staleTime 30s and refetchOnWindowFocus off the phantom row
stayed on screen for as long as the view was mounted — and the next drag persisted an order built
from it.
Round 6's test had to be re-pointed, because it was written in a world with no refetch: its mock
answered the pre-reorder order forever, which a server that just accepted the write cannot do. It now
holds the refetch open instead, so it still pins what round 6 actually fixed — the cache is right
WITHOUT waiting for a request — and would fail if the write were deleted. Both halves are
mutation-checked against each other.
Round 6 widened dialog focus ownership to count focus inside ANY `aria-modal` subtree as ours, so a
dialog and the dialog stacked on it closing together would still hand the focus back. That clause has
no way to tell a layer closing WITH us from one that stays open, so a dialog closing on its own — an
in-flight mutation's `onSuccess`, not a click, since a scrim covers it — under a modal raised above it
pulled the focus onto its own pre-open trigger, behind that scrim, where Enter still activated it.
Three verifiers confirmed the mechanism and two corrected the severity down: it self-heals on the
next Tab, which is why this is a containment violation rather than a broken end state.
`lib/layerStack` cannot refine it. Measured against the real hook: at the instant the lower dialog's
focus cleanup runs, the both-closing case and the sibling-stays-open case give identical readings —
`anyLayerAbove` true, the upper layer topmost — because passive cleanups run parent-first and
`useDialogChrome` is declared before the effect that calls `removeLayer`. Gating on the stack turns
round 6's own test red. What separates them is the DOM: React detaches a deleted subtree before
passive cleanups, so a layer closing with us is already disconnected while one that stays open is
not. Both tests are green, and each was mutation-checked against the other.
Verified in Chrome on localdev: opening the cheat sheet from a sidebar button and closing it restores
that button; with the command palette stacked over the cheat sheet, closing the top one leaves the
focus neither on the background trigger nor on <body>, and closing both returns it to the trigger.
Welcome's Lightbox got the trap and the restore in round 6 but not the layer membership the other
shells have, so its Escape and Tab fired from a raw window listener whatever was on top — the shape
the wizard's comment records being bitten by, unreachable today only by a mount ordering nothing
enforces. It joins the stack. Its comment also read as though the four-shell sweep were finished:
`PlayerModal` is still an `aria-modal` surface with no name, no trap and no restore, which is on
ROADMAP.md now rather than implied to be done.
Two ways round 6's own fixes did not bound what their comments said they bounded.
The reclaim budget's window anchor was written only where the window RESET, so it marked the window's
start rather than the last loss — a tumbling window, not the rolling one the constant is named for.
Any loss cadence slower than half the window satisfied the reset on every second loss, so the budget
never reached its bound: measured over 200 loss cycles, 61/70/90/119s apart all gave 200 reclaims and
never an error, while 5s and 30s apart correctly gave two and then the honest answer. The failing
band is exactly the case the ERROR handler names — a transcode that falls behind realtime — and both
committed tests spaced their losses 0s and 10min apart, i.e. outside it. The anchor now moves with
every loss.
And `finish`'s refusal handler lowered the play-through latch, which is the only gate on the 10s
interval, the page-hide beacon and `settle`. All three post `absRef`, which on the play-to-the-end
path IS the end of the media, and the backend reads a position within `_FINISH_MARGIN_S` of the
duration as a finish — so within ten seconds of the toast saying the mark had failed, the item was
marked watched, its position cleared and a scrobble sent to Plex. Round 6 closed the explicit
rollback write and left the other three open; the test it added set `upNext: false`, which tears the
effect down before any of them can fire. The refusal now takes back the CLAIM rather than the latch:
a second flag stops `settle` re-asserting watched, and playing again clears both.
`FINISH_MARGIN_S` duplicated the backend's constant in a second language with nothing relating them,
and its comment had the asymmetry backwards — if the BACKEND's margin grows, this guard starts
admitting a rollback the backend reads as a finish, i.e. re-opens the defect above. Pinned by a
contract test that reads the Python, and the comment corrected.
Round 6 turned `artifact_may_arrive` into a positive rule, but keyed it to the HLS vocabulary in
general rather than to what THIS session produces: the playlist regex accepted master/index/stream_*
for every session and the `seq == 0` branch accepted any variant token unconditionally. The two argv
branches are disjoint — a single-audio session is given `seg_%d.ts` + `index.m3u8` and no `%v` at
all, so `master.m3u8`, `stream_*.m3u8` and every `seg_<variant>_<n>.ts` are names it can never write,
and vice versa — so each of those was read as "an HLS name, so wait and see" and held a threadpool
worker, plus its pooled DB connection, for the full timeout. That is verbatim the round-5 outage the
route's own comment claims it prevents.
Measured through the real route against a live single-audio session, before: master.m3u8 404 after
30.0s, stream_q7z.m3u8 404 after 30.0s, seg_a0_0.ts / seg_zzzz_0.ts / seg_0_0.ts 404 after 20.1s
each. After, on the running localdev API: 404 in 44-316ms for all of them, while index.m3u8,
seg_0.ts and seg_1.ts still answer 200 — nothing is over-refused.
The session now carries the rendition names its ffmpeg was told to write, and `entry` is derived from
them instead of passed. And "one place" now means the argv BUILDS its patterns from the same
functions the parser reads back (`segment_name("%d", "%v")`, `variant_playlist_name("%v")`): the
layout was still spelled out independently in six places, with `_VIDEO_VARIANT` read by no consumer,
which is the drift that caused round 6's HIGH. A test asserts the argv's own patterns are what those
functions produce, and that every name they expand to is one the guard accepts.
The load-bearing MEASURED comment cited ffmpeg 8.1.1 — the Windows host's. The binary that writes
these files is the API container's 7.1.5; re-measured there, the conclusion holds.
The YouTube list got this rule in round 5 — one accepted write, one cache entry
that matches it — and the Plex list got only the rollback half of that fix. Its
success path invalidated `qk.plexPlaylists()`, a disjoint key root, and never
touched `qk.plexPlaylist(id)`: the key `putBack` rolls back to, and the one a
re-entry renders from without refetching.
So after an accepted drag the detail cache still held the pre-reorder order.
Leaving the playlist and coming back inside the 30 s staleTime showed the
reorder undone, and the next drag from there persisted that — silently undoing
an order the server had taken. Measured in jsdom: server 2,1,3, cache 1,2,3, no
refetch.
1 test, mutation-checked.
Round 5 taught `notifyActionError` to discard a plain-string `detail` in favour
of the caller's translated fallback, because `req` leaves 403/404/429 to the
caller and those routes hand-write English. Three places did not get the
message.
**The demo account lost its explanation.** `require_human` refused with a plain
string, so from that commit on the shared demo account got its caller's generic
"couldn't do that" and no hint that the demo account was the reason — and the
gate is not backed by a UI gate everywhere: the comfort shelf has no `is_demo`
check, so the pin button is live and simply fails. A structured `{code: "demo"}`
and an `errors.demo` string in both locales. The two auth routes that hand-roll
the same check were swept with it.
**`ChatThread` carried a second copy of the contract.** It filtered 404/429
itself and then did `e.detail || t(…)`, the pre-fix order, so the central fix
could not reach it and "That member hasn't set up messaging yet." stayed English
under a Hungarian UI. It goes through the shared helper now, and a contract test
pins the remaining raw-`detail` consumers as an explicit SET — never a count —
so a new copy fails the gate instead of shipping. The five on that list are
inline surfaces (a form field's error line, a setup banner) and are deferred
with the ~60 backend routes whose details they print.
**The Plex "Add to playlist" dialog was still silent.** Same shape as the
collection editor round 5 fixed and the YouTube twin that already had the fix:
no catch, so a playlist deleted on another device answered a structured 404 that
nothing reported — the tick never flipped, the gone row stayed, and the floating
promise became an unhandled rejection.
Also: the `ABORTED` sentinel's VALUE was asserted nowhere. The consumer test
hand-built the error it expected, so replacing `xhr.onabort`'s rejection with a
friendlier message left the lane green — and every upload the user cancels would
then raise a red "the upload failed" toast, which is what the sentinel prevents.
Both producer tests assert through `isAborted` now.
And two backend tests were pinning fewer sites than their own docstrings claimed:
the channel-404 loop said "all five" and issued three requests (explore and
subscribe unexercised, the excuse for subscribe naming a different call site
with a different reason slug), and the recipient/job test named three refusals
and requested two — `share_download`'s `not_found("recipient")` had no test in
the suite at all, nor did the locale keys added beside it. Every site named now
has a request behind it.
6 tests, mutation-checked against 7 mutations: the demo detail at both the gate
and an auth route, a re-hand-rolled ChatThread, the silent Plex dialog, the
abort sentinel, and each unpinned 404 — all red on the test that names them.
`MAX_SESSION_RECLAIMS` was cleared by the first fragment of any session, and a
reclaimed session essentially always delivers one — `start_session` does not
return until its first segment exists. So the budget only ever bit where a
session parsed its manifest and then died before a single fragment, which is not
either case its own comment names. Measured: 40 loss cycles produced 41 sessions
and never an error. A transcode falling behind realtime, or two viewers of one
rating_key taking turns, respawned ffmpeg indefinitely — the round-5 behaviour,
still there behind the guard added to stop it.
It counts within a rolling window now. Two losses inside two minutes is a
session that keeps being lost; one an hour apart is not. And the window is reset
per item, because the player stays mounted across a queue advance — otherwise a
budget spent on one episode would deny the next its first legitimate reclaim.
Separately: `finish`'s rollback re-performed the write it was taking back. On
the play-to-the-end path the position it restores IS the end of the media, and
the backend reads a position within its finish margin as a finish — so the
branch named "take the finish back" set watched, cleared the position and
scrobbled to Plex through a second route, while the caller's error toast was
already on screen saying the mark had failed. There is no resume point at the
end of the media to restore; the cache drop is the whole of the rollback there.
The credits-skip caller captures a mid-episode position and still restores it,
which is the case that needs it.
7 tests, mutation-checked: the FRAG_LOADED reset, the missing window, and the
unguarded rollback each red on the test that names them.
Round 5 extracted `ui/dialogFocus` so no shell would hand-roll the focus
contract, then introduced four ways to lose the focus anyway — all four measured
in jsdom against the real components.
**Ownership was narrower than the trap.** `ui/Modal` traps Tab over the whole
dialog on purpose, so a toast's Undo or ✕ stays reachable; ownership was tracked
over the panel alone. Tabbing to that ✕ therefore read as "the user left the
dialog", and since removing a toast fires no `focusin` to re-arm, closing
afterwards skipped the restore and left the focus on <body>. Before this hook
existed the restore was unconditional, so it was a regression.
**A dialog stacked above us is still ours.** A nested `ui/Modal` portals into its
own host, so it is outside the outer dialog's root however wide that scope is
drawn — widening alone could not fix it. Both closing in one commit left the
focus on <body>: the outer skipped its restore, and the child's own target (a
control inside the outer panel) was already detached, because passive cleanups
run parent-first. Ownership now means "inside my dialog, or inside a modal layer
above it", which is what the rule always intended.
**The wizard disagreed with `ui/Modal` on its first line.** `trapTab(e, panel)`,
with neither the dialog-root scope (so its own toast host was outside the cycle
— the connection notice never auto-dismisses, and its ✕ was unreachable from the
keyboard) nor the `container` argument the JSDoc exists for: `panel.contains(
panel)` is true, so focus sitting ON the tabIndex=-1 panel matched no branch and
a Shift+Tab walked into the app behind the scrim. Both call sites now pass the
same three arguments.
**The toast handover was a render-phase read-modify-write.** A `useMemo` read a
module global that the next line overwrote; React re-invokes a component body
and every `useMemo` factory in it on a StrictMode double render, and the LAST
invocation commits — so the second pass re-read what the first had just written,
the takeover host classified its OWN new toast as inherited, and the dev build
behaved exactly like the code round 5 replaced. Snapshotted through a ref keyed
on `isTop`, so a repeated invocation of one render is a no-op.
And the fourth `aria-modal` shell, `Welcome`'s screenshot lightbox, is finally
swept into the same contract: it declared itself a modal dialog while leaving
the focus on the thumbnail behind an 80%-black scrim and letting Tab walk the
landing page underneath.
The headline fix of the previous commit — StrictMode's re-mount taking the focus
off an `autoFocus` field — had no test at all; deleting it left all 1009 green.
It has one now.
9 tests, mutation-checked against 8 mutations, each red on the test that names it.
Round 5 added a guard so a request for a segment beyond the produced region
answers instead of polling out the full timeout — the route is a SYNC one, so
every such wait parks a threadpool worker and, through `current_user`, a pooled
DB connection. Measured against real ffmpeg, the guard could not fire where it
mattered.
`-hls_segment_filename seg_%v_%d.ts` substitutes the VARIANT first: a dual-audio
session writes seg_0_0.ts (video) and seg_a0_0.ts / seg_a1_0.ts (the `name:aN`
renditions). The regex read those as sequence-first, so every video segment
parsed as segment 0 and every audio segment did not parse at all — the guard was
inert on every multi-audio item, which is most of this library. `_seg()` carried
the same inversion: the warm-up waited 8 s for `seg_1_0.ts`, a name ffmpeg never
writes, on every multi-audio session start AND every seek-restart.
The layout is now one constant plus one parser (`segment_name` /
`_SEG_NAME_RE`), with a test pinning them to each other, instead of three
comments that could drift — and they had drifted, all three ways.
The guard is also a positive rule now. It was a blocklist, so anything that was
not a single-audio segment name was waved through: `GET …/anything.m3u8` still
parked a worker for 30 s, while the route's comment claimed a flood of hopeless
names was what it prevented. A session produces its playlists and its segments
and nothing else. `.vtt` leaves `_HLS_CT` with it — ffmpeg runs with `-sn` and
subtitles are a separate route, so that extension could only ever name a file
that will not arrive.
Also here:
- A 5000-digit segment index reached the client as a 500: `int()` on an
unbounded digit run raises past 4300 digits and only OSError was caught. The
name is bounded on the route and the index in the regex — the same bug class
the subtitle routes were bounded for.
- A session torn down between the wait and the read is a 404, not a 500. `_kill`
rmtrees the directory and no lock spans those two steps; the player's recovery
is keyed on the STATUS, so a 500 skipped the reclaim and ended playback with
"couldn't play" instead of restarting at the playhead. The trigger is ordinary
— `loadSession` awaits the new session before destroying the old hls instance.
- Two comments claimed an open player checkpoints every 10 s "whether or not it
is playing", which is the stated reason the 30 s finished-idle reap is safe.
`flush` sends nothing once the finish latch is up, or before a position is
known. The states it does cover — playing or paused mid-item — are the ones
the rule needs; the comments now say that instead of more.
- The route's tests leave the DB lane: the route reads nothing from the database
since the auto-start was removed, and there those four regressions SKIP on a
machine with no Postgres while pytest prints green.
14 tests, mutation-checked: the old parse, the old warm-up name, the blocklist,
both bounds off the absurd name, the unguarded read, and the route dropping
`proc` — each red on the test that names it.
Both found by UAT in a real browser, after the tests for them were already green — which is the
whole reason this repo settles focus and timing claims by running them.
- **The autoFocus fix worked, and then the dialog took the focus back.** React 18's StrictMode runs
every effect mount→cleanup→mount in development: the cleanup handed focus back to the trigger, and
the second mount, finding the focus outside the dialog, moved it to the first focusable — off the
`autoFocus` field a beat after the browser had put it there. Chrome measured it: INPUT (commitMount)
→ BODY (cleanup) → ✕ (re-mount). The effect is idempotent now — it remembers what the dialog was
focused on and restores THAT on a re-mount — and the close-time restore is conditional on the
dialog still holding the focus, which it tracks as it changes rather than computing at cleanup,
where the panel is already out of the document. A dialog no longer yanks focus back from wherever
the user has since moved it.
- **Passing ffmpeg's handle to the segment wait only shortened the FINISHED case.** Measured against
the running localdev API: a request for `seg_99999.ts` on a live session still took 20 195 ms,
because that ffmpeg has not exited and the wait has no other reason to stop. ffmpeg writes segments
in order, so a segment whose predecessor does not exist is not the next one and cannot arrive —
and that is knowable at once. Same request now: 45 ms.
Verified on localdev against the real library (Westworld S01E06, remux, multi-audio): Mod+K opens the
command palette with the focus in its search field and typing reaches it; playback runs on the new
session-scoped HLS URL; a request naming a stale session id is a 404; and a viewer paused for longer
than the finished-session rule keeps their 2.4 GB session across two reaper passes — the progress
checkpoint touches it — and resumes cleanly.
vitest 1009, pytest 1127.
Deep-review round 5. The assertion behind `subtitles._exists` was only red where pathlib re-raises
ENAMETOOLONG: on Linux it does, on Windows the same stat fails with ERROR_INVALID_NAME, which
pathlib's `exists()` already swallows — so the assertion passed with the fix reverted. It forces the
errno now instead of relying on the filesystem. (The gate's pytest lane is the Linux container, so
the branch was never actually unguarded; the test was just dishonest about why it passed.)
Deep-review round 5, over round 4's own fixes.
- Round 4 made a failed reorder roll the rows back to the CACHED detail, because a resync cannot do
it (the server's list is unchanged, so the refetch is deep-equal, react-query keeps the same
`data` object and the effect keyed on it never runs). But nothing wrote an ACCEPTED order into
that cache — only the refetch that follows a successful reorder did. So a second reorder rejected
while that refetch was still in flight rolled the rows back past the first one, which the server
had accepted; the membership guard cannot correct an order (its key is the sorted id set), so the
wrong order stayed on screen for good and the next drag persisted it. `onApply` now writes the
accepted order into the cache before invalidating.
- `PlexPlaylistView.rename` had a `catch` no test could reach, and `commit` (reorder) was not gated
by `busy` the way its siblings are — a reorder in flight while a removal completes is what makes a
rollback ambiguous in the first place. The rename failure path is now covered; `commit` defers.
- `putBack` read `q.data` from the render that bound the failing handler instead of the cache at
rollback time, unlike the sibling `restoreFromCache`. Aligned. Stated plainly: I could NOT drive
that staleness in jsdom — a mutation that restores the old shape leaves every candidate test green
— so this one ships as an alignment with the proven sibling, not as a fix with a red test behind
it. The scenario and the harness limit are on ROADMAP.md.
Two regression tests, both mutation-checked red (the cache write, the rename catch). The
drag-reorder failure path still has no test: dnd-kit's pointer and keyboard sensors do not activate
under jsdom even with geometry and scrollIntoView stubbed, so a test there would have pinned the
harness, not the code. Logged, not faked.
vitest 1009.
Deep-review round 5, over round 4's own fixes.
- `notifyActionError` preferred the server's `detail` over the caller's translated fallback, and
`reportFailure` deliberately leaves 403/404/429 to the caller — so for exactly those statuses the
dozen call sites round 4 converted started printing whatever English a route hand-wrote into an
otherwise Hungarian UI ("No user with that email.", "Not subscribed to this channel", "Unknown
job"). A STRUCTURED detail is still preferred: `req` has already run it through `localizeDetail`,
so it is both localized and more specific. A plain string is not.
- The other half of the same defect, on the backend: round 4's `get_or_404(what="channel")` sweep
reached exactly ONE of the five identical sites in channels.py and shipped the HU/EN string with
it, so the Block button said it in Hungarian and the four controls beside it in English. All five
are converted, and the subscription / share-recipient / scheduler-job 404s with them (two new
reasons in both locales).
- `PlexCollectionEditor`'s add/remove, take-over and delete were try/finally with no catch — in the
very file round 4 edited to give `createNew` one. 403 and 404 are the caller's to report, so the
tick simply never appeared and a CONFIRMED "delete collection" did nothing, silently.
- `UploadDialog` is the only `uploadReq` caller and kept an empty catch whose comment still claimed
the global error modal had it. It had not since `reportFailure` became shared: a proxy 502 raised
only the self-clearing connection notice and a 413/403/404 nothing at all, so the dialog snapped
back to its pre-upload state with no error named. The user's own Cancel stays silent — and since
a cancel and a network drop both carry status 0, that difference is now a named `isAborted`
rather than a string literal at each end.
Regression tests for each (5 vitest, 2 pytest, plus two existing assertions corrected), all
mutation-checked: 6 mutations, each red on the assertion it is named after. vitest 1007,
pytest 1127.
Deep-review round 5, over round 4's own fixes.
- Sessions are keyed by rating_key alone while a session's DIRECTORY is per (key, offset, tag), and
the artifact URL carried no session identity — so once a seek-restart or a second viewer of the
same item took the registry slot, the old viewer's request for `seg_12.ts` was answered 200 with
the new session's twelfth segment: bytes from a different `-ss` offset under `-copyts`, on a
timeline the client is not on. Wrong content with a success status is the one failure a player
cannot notice, and round 4's reclaim-on-404 could never fire on it. The id is in the path now (so
the playlist's relative names resolve back into the same session), a mismatch is a 404, and the
player reclaims at the playhead. The entry playlist's auto-start is gone with it: it could only
ever begin at 0, silently discarding the viewer's position.
- Round 4 exempted paused viewers from the reaper by keeping EVERY finished session for the full
ten-minute idle window — one whole segmented copy of a film per item, on the volume the download
center shares. The signal it needed already existed: an open player checkpoints its position every
10 s whether or not it is playing, so that route now touches the session and the periodic pass is
back on the short rule.
- POST /session's opportunistic reap was never scoped to the key being started: it walked the whole
registry with the 30 s finished rule, so pressing play on ANY item — even a direct-play one, the
call ran before that early return — killed another viewer's paused session. It is gone; the
periodic pass owns reaping.
- A segment that can never arrive no longer polls out the full 20 s: the producer's handle is passed
to `wait_for`, so a finished ffmpeg answers at once. This is a sync route, so each such wait parked
a threadpool worker AND a pooled DB connection, and an ordinary seek past the produced region was
enough to do it.
- The DataError handler's re-raise branch — the half that keeps a genuine DataError a 500 — is
pinned, so simplifying it to an unconditional 422 cannot pass in silence.
Six regression tests (two route-level, in the DB lane), each mutation-checked red on the assertion
it is named after. pytest 1124, vitest 999.
Deep-review round 5, over round 4's own fixes.
- Round 4 added a `finish` latch so a played-through item records no checkpoint at the credits.
On a NATURAL play-through it was raised a beat too late: per the HTML spec the browser sets
`paused` and fires 'pause' BEFORE 'ended', so the pause checkpoint went out while the latch was
still down and raced the watched mark — both requests read "not watched yet" from their own
snapshot, so Plex counted one watch twice, and where Plex's reported duration runs more than the
finish margin past the real media the checkpoint wrote a resume position AFTER the scrobble. The
pause that ends the media no longer flushes. The regression test that was meant to cover this
modelled the end as a bare 'ended' event — a browser that does not exist — and is now the real
order.
- The watched mark was fired and forgotten (`.catch(() => {})`) while the latch patched the cached
detail to watched / position 0 regardless, and round 4's once-per-id start guard meant the
corrected detail arriving later was never used: an episode whose mark failed reopened at 0:00
with a Watched badge the server disagreed with. `finish` now takes the in-flight write and takes
itself back if the server refuses it — the real position is reported, the lying cache entry is
dropped, and the failure is said out loud instead of swallowed. Both callers (the natural end
and the credits auto-skip) hand it over.
- The reclaim of a 404'd HLS session had no bound. `parsed` cannot provide one — it belongs to one
hls instance and every restart builds a fresh one — and the server answers 404 both for "the
session is gone" and for "that segment is not there yet", so a transcode falling behind realtime,
or two viewers of one item taking turns, respawned ffmpeg forever behind a spinner that never
became an error. Two consecutive reclaims, reset by the first fragment that actually loads, then
the honest failure.
Five regression tests, each mutation-checked red on the assertion it is named after.
Deep-review round 5, over round 4's own fixes.
- Round 4 gave `ui/Modal` a persistent portal host so a fullscreen transition would stop tearing
the dialog down. The host was attached in a `useLayoutEffect` — one commit phase AFTER React
runs `autoFocus` in `commitMount`, and `focus()` on a disconnected node is a no-op. Every
autoFocus in every dialog was dead: Mod+K opened the command palette with the focus on the ✕,
so typing searched nothing and Enter closed it; the same for the download rename field and
every non-danger confirm. The host is attached in a `useInsertionEffect` now, a whole phase
earlier, and the whole re-parent — including the focus save/restore across the move, which
`appendChild` needs because it is remove-then-insert — lives in one `useFullscreenPortalHost`.
`ui/AnchoredPopover` had a second copy of that re-parent WITHOUT the focus restore, and a
comment claiming focus survived it; both now come from the hook.
- `ui/Toaster` classified a toast raised in the SAME commit as a host takeover among the ones it
inherits, because it read the store live at the takeover render. A dialog that notifies and
closes in one handler (UploadDialog) therefore put its toast up with no role, into a region
still `aria-live="off"` — announced by neither path. "Inherited" is now what the outgoing host
last had on screen, and a toast that is NOT inherited is announced even inside the unarmed
window.
- The onboarding wizard renders its own shell, and Escape ordering was all of the dialog contract
it had: no role, no label, no initial focus, no trap, no restore. A keyboard user tabbed out of
the first-login wizard into the app behind the scrim. The focus half of the contract is now one
shared `ui/dialogFocus` that both `ui/Modal` and the wizard use.
Every fix has a regression test, each mutation-checked (6 mutations, all red on the assertion
they are named after).
Deep-review round 4, the overlays.
- Toast hosts were ranked by mount time, and PlayerModal swaps its host on every fullscreen change:
leaving true fullscreen with the download dialog open put the player's host back on top, so toasts
were drawn under that dialog's scrim and a click on one closed the dialog. Each overlay takes a
rank once, when it opens (`useOverlayOrder`), and hands it to every host it renders.
- A dialog open while fullscreen is left keeps a stable portal container (moved, not re-created):
React tore the portal down when `fsRoot` changed, resetting a half-filled form and dropping focus
to <body>. Focus is put back after the move.
- The onboarding wizard is a layer, not a raw document listener: one Escape used to dismiss the
error dialog over it AND end onboarding, skipping the write-scope step.
- `navigateTo` (a toast's "View", clicked inside the player's own host) goes over the overlays it
unmounts, so the player's Back-to-close cleanup no longer pops the page it just opened — the same
fix Shift+W already had. The access-requests toast action too.
- The live region arms itself a task after it takes over inherited toasts: React flushes effects in
the same task as that render, so the "silent inheritance" arming was a no-op.
- The overlay-host guard matches class tokens in any order (PlexPlayer's root was never checked),
globs `ui/` too (ui/Modal was not checked either), and requires the rank.
Deep-review round 4, the playlist surfaces.
- A failed move or reorder relied on the resync to put the rows back — but the server's list is
unchanged, so the refetch comes back deep-equal, react-query keeps the same `data` object, and the
re-seed effect keyed on it never ran: the moved row stayed gone, the rejected order stayed on
screen. Both restore from the cached detail now.
- A failed removal puts its row back itself: when the server is unreachable the resync fails too,
and round 3's "keep the list on a failed refetch" guard then kept showing it as removed.
- The Plex playlist page (remove / rename / delete / reorder), the add-to-playlist popover and the
tag manager had no failure handling at all: a list or tag deleted elsewhere (a 404, now a
structured `not_found` the request layer leaves to the caller) failed in silence and an
optimistic removal stayed removed. Each says it and resyncs; deleting a Plex playlist that is
already gone just leaves the page.
- Playlists' own loading/error branches use the shared `LoadingState` / `StateMessage` (the loading
line had no live region).
- The "player stays mounted" tests count mounts: a remount ended with the same flag as staying up.
Deep-review round 4, the "one report per failure" contract round 3 introduced.
- `req` marked gateway (502/503/504) and network failures "already reported" because they raise
the connection notice. But that notice is about connectivity and removes itself on the very next
answer — which every caller's own resync refetch fetches at once — so a delete that hit a
keepalive reset flashed the notice and was then reported nowhere, the list still there. Only the
durable reports (the error dialog, the 401 sign-out) count now; the caller says the action failed.
- A connection notice the user closed swallowed every later one for the rest of the visit; it is
raised again.
- `req` and `uploadReq` share one failure policy (`reportFailure`). The upload copy had drifted: a
502 opened the blocking "Server error (502)" dialog and cleared a notice that was still true.
- A taken Plex collection name is a structured 409, localized, and the create/rename calls are
quiet: it was an English dialog plus a Hungarian toast. `notifyCollectionEditError` is gone.
- The mutation onErrors that toasted unconditionally after a non-quiet request (invite approve /
deny / add, demo whitelist, audit clear, scheduler, share, full history, edit, config tests, tag
attach) go through `notifyActionError`; the playlist delete's hand-rolled copy of it too (it takes
a level now). The user role / suspend / delete actions had no onError at all: a user deleted in
another tab failed silently and stayed listed — now said, and the list refetched.
Deep-review round 4, the Plex player.
- A 404 on an HLS session that WAS playing means the server reaped it (a long pause, a deploy);
hls.js never retries a 4xx, so playback ended with "couldn't play". A new session at the playhead
now takes its place (the frontend half of the paused-viewer reap fix). A session that 404s before
its manifest is still a plain failure, so it cannot restart itself in a loop.
- Leaving an item that ended, or that the credits auto-skip left, sent a final checkpoint at the
credits (racing the watched mark, and pushing a resume point to Plex after the scrobble) and
patched "new, at 49:55" into the cache — Prev reopened it at its last seconds. `finish` settles it
as watched at 0, said before the pause whose own checkpoint could land first.
- An audio switch, an A/V offset change and every seek control read `abs`, which is 0 until
playback ticks: a resume at 20:00 restarted from 0:00. They use the shared `playhead`.
- A seek while a restart is pending always restarts (the old media's `seekable` described the old
session, and the pending answer overrode the newer seek).
- A superseded direct-play session's loadedmetadata/canplay are generation-gated and detached (they
seeked the next item to the previous one's resume point).
- The session starts once per item: a reconnect refetch, the info panel's watched toggle or a
language switch no longer restarts the film at the last checkpoint.
- Tests for the session generation's failure paths and the item-change bump, which had none.
Deep-review round 4, backend half.
- The periodic HLS reaper (round 3) applied the "finished and unfetched for 30 s" rule that only a
NEW session request may apply. A remux finishes long before the film, and a paused player with a
full buffer fetches nothing, so a one-minute pause killed the session and playback died ~30 s after
resuming. The periodic pass is `reap_abandoned` now (the long idle limit, finished or not).
- A subtitle `publish` that could not keep its output (full / read-only cache volume) returned
quietly, so every poll queued the window and the whole-container pass again. It reports failure
and the pass is backed off like a failed read; a cached VTT that cannot be READ is answered
`unavailable` (`cache_unreadable`) instead of scheduling reads that can never land.
- The subtitle routes bound `rating_key` / `sub_ord` (they become a cache file name: a 250-digit
one was ENAMETOOLONG, which pathlib's exists() re-raises -> a 500), and the cache helpers no longer
throw on such a name.
- One app-wide handler turns psycopg's "text fields cannot contain NUL" DataError into a 422: round 3
guarded only the by-rating_key lookup, and `contains=`, `contains_group=`, reorder/bulk key lists
and the collection search still 500'd.
- The hand-written `db.get(...) is None -> 404 "Unknown ..."` siblings use `get_or_404` with a
`what` (channel / video / user), so the Hungarian UI no longer shows English for a gone row.
Deep-review round 3 of bug hunt 2026-09.
- The YouTube and Plex progress bodies are bounded by POSITION_MAX_S (the int4 position columns): a
finite position of 2^31 or more was a DataError, i.e. a 500 — the sibling of round 2's non-finite
fix.
- plex/stream.py promised "a periodic reaper", but reap_idle ran only inside the NEXT session
request: after the last playback an abandoned session (its ffmpeg still working, its directory a
whole segmented copy of the film) lived until someone pressed play again. A lifespan task reaps
every 60 s.
- Tests for what round 2 left untested: the 422's body keeps the list shape the UI localizes, and
stream_session refuses non-finite start/aoff.
Deep-review round 3 of bug hunt 2026-09 — the sibling of the collection page's fix (d349bc90).
Discover's toggle flips the Surprise draw too, but its resync refetches only the rails (refetching the
draw would reshuffle it under the viewer), so a refused toggle stayed flipped there. toggleWatchState
takes an optional `revert`, called on a failure only; the queryKeys comment that claimed the watched
toggle invalidates the Surprise key now says what the code does.
Deep-review round 3 of bug hunt 2026-09.
- The detail's error branch applies only when there is nothing to show (`isError && <no data>`, the
QueryState convention): one failed refetch — on focus, after a move or a removal — replaced a
loaded list with "couldn't load". The error line is announced (role=alert).
- Tests for the two branches of "the player survives a change of selection" that round 2 left
untested (a list that fails to load, no selection).
Deep-review round 3 of bug hunt 2026-09.
- get_or_404 / owned_or_404 answer a structured {code: "not_found", reason: <what>}
(errors.notFound.* in HU+EN, generic fallback). The structured-404 fix of round 2 had gone into the
ad-hoc `_own_playlist` (CLAUDE.md's "Not" pattern) while the shared helpers stayed string-only, so
views, tags, pins, Plex playlists, invites and users still 404'd in English. `_own_playlist` now
goes through owned_or_404; admin's hand-rolled user lookups use get_or_404.
- A move into a list that is gone names the TARGET (reason "move_target") — "this playlist doesn't
exist any more" was read as the list on screen, which still exists.
- Every refusal in the playlists router is a {code: "playlist", reason} (playlists.errors.* HU+EN)
instead of a bare English string shown verbatim in the Hungarian UI.
- `req` / `uploadReq` mark the errors they already put on screen (the error dialog, the connection
notice, the 401 sign-out), and notifyActionError skips those. The watched toggles, the saved views
and a playlist delete toasted a fault the dialog was already showing; Playlists' two hand-written
status filters now go through the same check.
Deep-review round 3 of bug hunt 2026-09 (the round-2 overlay toast host, finished).
- Inside an overlay the toast stack sits above the bottom band (bottom-24, the UpNextCard rung) and
lets clicks between cards through: at the app's own spot it covered the Plex transport row and
PlayerModal's docked bar and took their clicks while a toast was up (measured: play/pause, ±10,
stop, prev/next, mute all hit the toast).
- ui/Modal and the onboarding wizard mount an overlay host too, so a toast raised under any dialog
is readable and a click on it no longer closes the dialog; the Modal's Tab trap spans the whole
dialog, so a toast's action and dismiss are keyboard-reachable. A test now enforces the
convention: a component that covers the viewport at z-overlay must mount <Toaster overlay />.
- PlayerModal's host moves into the stage in true fullscreen (outside the element holding the lock it
was painted under it).
- A host that takes over shows the toasts it inherits silently (no second role=alert, its region
aria-live=off until they are in), and each countdown bar starts as far in as its toast already is
instead of restarting while the real dismiss timer does not.
- The overlay-host registry uses lib/store instead of a hand-rolled listener triad.
Deep-review round 3 of bug hunt 2026-09.
- loadSession checks, after its await, that it is still the newest call and the player is still
mounted. A late answer used to replace what was playing (Next twice quickly: the slower transcode
landed last and played the middle item under the last one's title, recording its progress there)
or attach a stream to the detached <video> after Back, never destroyed.
- The <video>'s clock is trusted only once the new session has answered. go()'s pause() fires a
timeupdate AFTER the reset to 0 (measured in Chrome), so the previous item's position was sent as
the next item's subtitle position and saved as its progress (final, pushed to Plex).
- Next settles the item being left (final checkpoint + reopen-cache patch) before switching: the
progress cleanup ran with the position already reset and skipped, so every Next dropped the last
stretch watched.
- A seek-restart puts its target into the position state too; the ref alone was re-synced back to
the pre-seek spot by the restart's own render.
- The unmount teardown captures the <video> at mount: React detaches refs before an unmount's
cleanups, so the stop/detach never ran.
- The tracks panel exposes the selected audio/subtitle track via aria-pressed.
Verified in real Chrome on localdev (Westworld S1E6 → Next): the item left was settled at 783 s, the
next asked for subtitles at its own start (0), and no progress was written for it.
Deep-review round 3 of bug hunt 2026-09.
- `failed` is now final only for a track that can never be text (bitmap, no item/track/file). A Plex
hiccup or a stretch ffmpeg could not read answers `unavailable`: said, polled again every 5 s, and
the <track> stays mounted when cues exist elsewhere (gen > 0). Re-picking the selected track asks
the server again at once.
- Within LOOKAHEAD_S of a covered stretch's end the next window is read ahead, instead of only once
the playhead has fallen off it (subtitles dropped out at every edge while the full pass was backed
off or slower than a window lasts).
- Windows and full passes run on separate pools with separate in-flight caps, so a window never
queues behind (or is refused by) minute-long full passes — the pool comment promised that and
nothing enforced it.
- A pass that times out is TERMinated and keeps what it reached (published as partial) instead of
being killed and discarded.
- Lapsed failure memos are swept on every write (they were only pruned when the same key was read
again, so polling fresh keys grew the process without bound).
- `publish` no longer merges into a sidecar with no VTT behind it, or claims a region whose VTT write
failed, or overwrites an unreadable track with one window.
- A NUL in a rating_key names no row (it was a DataError, a 500, on every Plex route) and no longer
500s the status poll.
Tests: 16 new, each mutation-checked (18 mutations, all red on the test that names them).
The player's "still alive" flag — every play() is gated on it, so a late one cannot start a detached
<video> after Back — was only ever LOWERED, in an effect's cleanup. React may run a cleanup and mount
again (StrictMode does exactly that in development), and then the flag stayed false for the life of
the player: it never started by itself, and a seek back past the loaded region, which restarts the
session, left it paused. Found on localdev (Westworld S01E06: playing → seek back → the picture
settles, paused). Production does not double-mount, which is why only dev showed it — and why
several earlier localdev measurements "stayed paused".
The effect now raises the flag in its setup. A new test renders the real player under StrictMode and
checks it plays once the media is ready; red without the fix. Verified on localdev: it starts by
itself, and keeps playing through a backward session-restart seek.
The feedSortOptions test compared each option's default direction against SORT_DEFAULT_DIR — the
implementation's own table, so a wrong entry passed — and that import was the export's only user
outside its module, which kept knip from seeing it had become dead after F3. The test now states the
directions, the table is private again.
C1 moved the optimistic flip onto the key the page reads — and so made visible that nothing corrects
it: the resync only invalidated `plex-library`, so a failed toggle stayed "watched" beside its own
error toast, and the collection's other cached sort orders stayed stale after a good one. The
resync now takes this page's `detail` prefix too, as the Discover and person pages do for theirs.
The C1 test holds the request open, so it still pins the optimistic write on its own.
- D3 snapshotted the removal's TARGET, but whether the player offered Remove still followed the
selected list: launched from Watch later (no Remove, by design), selecting another list under it
made Remove appear — and delete from Watch later. The snapshot now carries the launch list's
editability too.
- Selecting another list under an open player (the command palette works over it) put the page into
"Loading…", and the player was rendered only in the loaded branch — it vanished mid-playback and
came back at its launch video. It now sits at one place in the tree whatever the page is doing.
- A removal takes the video out of the CACHED list as well as the rows: the player's progress
checkpoint rewrote that entry before the refetch landed, handing the membership guard a list that
still held the video — the row came back and the undo history went.
- The watched-state callback refreshed the selected list, not the launch list (D3's sibling).
- `_own_playlist`'s 404 is structured (`{"code": "playlist", "reason": "not_found"}`) and localised:
the bare "Unknown playlist" reached the Hungarian UI in English, which also made round 1's new
`removeFailed` string unreachable. Both Remove and Move report through `notifyActionError` now.
- The "viewer moved on under the confirm" test stepped back only once, which lands on the same video
whether or not the removed one left the queue; it now takes the second step that tells them apart.
The app's Toaster lives in the content column, which is a stacking context (its `@container` implies
layout containment), so both full-screen players — portalled to <body> at the same z-index — painted
over every toast. Round 1's "a failed Remove is reported" fix was therefore still silent inside the
player: its toast sat under an 80% black, blurred scrim (elementFromPoint returned the scrim), and a
click on it closed the player.
An overlay now mounts its own `<Toaster overlay />`, and exactly one host shows the toasts: the most
recently mounted overlay, else the app's. Same component, same look — and inside the aria-modal
dialog, where a screen reader can still reach it. A click on a toast inside an overlay stays there.
Both players get one (the Plex player's also works in native fullscreen, where nothing outside the
fullscreen element is drawn). Verified on localdev: the removal failure reads above the scrim,
bottom-left, and clicking it leaves the player open.
- The <track>'s `error` never reached `onSubTrackError`: it does not bubble, and React 18 attaches no
listener of its own to a <track> (measured in Chrome and jsdom — `onError` 0 calls, `onErrorCapture`
1). Round 1's remount-and-give-up was dead code behind it. The hook now hands the player
`onErrorCapture`, and the test fires a real non-bubbling `error` at a rendered <track>.
- Once reachable, the retry burned its whole budget within milliseconds of an outage (a remount
fetches at once) and the give-up outlived turning subtitles off and on. Retries now wait (2 s,
doubling), running out is reported at once, and picking the track again starts a fresh budget.
- Until the first timeupdate the poll sent `sessionK` — still 0 while the session request was out,
and the PREVIOUS item's start after a move to the next episode. It now sends where the session was
asked to start, and the track choice belongs to the item it was made for, so the next item is not
asked about the old ordinal at the old position.
- A failing poll over a stale `partial` answer was silent: react-query keeps the last good data, the
track stayed "showable", and the notice only rendered for an unmounted track. The hook now decides
what the player says (`subNotice`), independently of whether a window is mounted.
The JSON parser under FastAPI accepts `Infinity`/`NaN`, and both progress checkpoints truncate the
position with `int()` — a hand-written checkpoint came back as a 500 with a traceback. Their bodies
now refuse non-finite floats, and the app's validation handler — FastAPI's own shape — echoes a
non-finite input by name: the default one failed to serialise its own error (JSON has no NaN), which
was a 500 by another route. Sibling of the subtitle routes' non-finite query parameters.
Deep-review round 2 of the 2026-09 bug hunt, over the round-1 fix diff.
- A window was LATE by the distance to the keyframe its input seek landed on, and then doubled for
good once the full pass merged in: a cue already running at the window's start came out at a
negative time, and the webvtt muxer's default shifted the whole output so it sat at 0. Measured on
the real Westworld S01E06 at 25:00: every window cue +5.005 s, 682 cues merged instead of 608.
`-avoid_negative_ts disabled` keeps the times exact (the parser reads ffmpeg's signed-field
spelling of a negative time); now 608/608 on that file, and a real-ffmpeg test pins it.
- A full pass that failed was re-armed every minute while the viewer watched what it did publish —
a whole-container read per minute. It now backs off for an hour; windows keep running meanwhile.
- A failed WINDOW set the track's failure memo, so any uncovered position read "failed" — a final
answer that ends the player's poll, even with the full pass about to deliver, and one any account
could plant with an absurd `at`. It now marks only the span it failed to read; "failed" is given
once the full pass AND a window there have both failed.
- Windows join by what they COVER, not a bucketed identity: a backward seek no longer waits behind a
window that cannot cover it, and a playhead crossing a bucket edge no longer queues a second read.
The queue is bounded per track and overall.
- shutdown() sets a flag a worker checks before it spawns ffmpeg (one still in its Plex lookup used to
start a read after the kill sweep); queued-job cancellation and the lifespan call are now tested.
- A window starts where ffmpeg is told to seek (rounded to the ms it gets), a rounded-up millisecond
carries into the minute (no more `00:00:60.000`), a key's first publish claims nothing before its
VTT lands, and `at`/`offset`/`start`/`aoff` refuse non-finite values (`at` also an absurd one).
The D1 regression test patched the cache and clicked at once; react-query delivers a cache write to
its observers on a setTimeout(0), so under the gate's parallel load the click could beat the page's
re-render and read the old row (748, not 830). Test-only race — a human reopen comes seconds later.
Stable across three full-suite runs; the D1 mutation still turns it red.
Deep-review R6 — the small items:
- The video card's channel name was still a <button> (pre-existing): ctrl/middle-click did nothing
there while the title above it and the player's channel name honoured them. It renders through
ChannelNameLink now, which grew a forwarded ref (the clamp measurement), pass-through attributes,
a `tooltip`, and an opt-in `stopPropagation` so no click reaches the card's row.
- ChannelNameLink rendered an empty, still-tabbable <a> for a channel with no title yet (a stub first
seen through a live search; pre-existing at this site). It falls back to the translated "Channel".
- The feed's SortKey→SortOption mapping existed twice in Feed.tsx; one `feedSortOptions` in
feedSort.ts serves both modes (DIRECTIONLESS is module-private again).
- The backfill refresh's comment claimed a one-page cost. It refetches every loaded page of the
grid, and that is the right price: backfilled videos are older, so they land below the head.
Tests for call sites that could be deleted with every lane green: the channel page running the
backfill refresh (knip had been blinded by the hook test's import), Feed handing the grid the live
results in the chosen key AND direction, the card using the link. Plus the alt-click arm of the shared
click predicate, and a plex filter-chip test that claimed it made no request while firing a real one.
Verified on localdev: the feed's channel names are anchors (same 14 px muted look), a plain click
opens the channel in-app. 7 mutations, each red on its assertion.
Deep-review R5 — five defects in the Remove button f76a2afb put inside the YT player:
- It removed from whichever playlist was selected by the time it was pressed, not the one the player
was launched from; the selection can change under an open player (the command palette opens over
it). The launch list is now part of the player snapshot, and the optimistic row removal only
touches the list actually on screen.
- A 403/404 failed in total silence: the catch claimed the global dialog had reported it, but `req`
deliberately leaves those to the caller. It now notifies, as moveItemTo already did (the silent
catch predates this round; the full-screen consumer did not). New key playlists.removeFailed, HU+EN.
- It computed the next item from the cursor as it was when the button was pressed, but prev/next and
auto-advance keep working under the confirm dialog: the viewer was thrown back onto a video they had
moved past. It reads the live cursor, and if they moved on it just drops the item from the queue.
- The button was disabled for the confirm's lifetime, so the dialog's focus restore hit a disabled
element and a keyboard user who cancelled was left on <body>, outside the aria-modal player. A ref
guards the double press instead.
- Its accessible name was a bare "Remove"; it now names its object, like the row control.
The call-site wiring (step on, drop from the queue, close on the last item, leave everything alone
when the removal did not happen) finally has tests. Verified on localdev in the real YT player: a row
reopened at the 136 s it had just stored; Cancel returned focus to the Remove button; a confirmed
removal stepped 1/3 -> 1/2. 10 mutations, each red on its assertion. Scratch list and positions reset.
Deep-review R4. 1630a3df wrote the YT player's stored position into the react-query cache so a
reopen could not rewind it — but only the feed actually read it back:
- Playlists renders from its own undoable copy of the list, re-seeded only when the SET of ids
changes, so the patched position never reached it: close a video at 830 s, reopen it from the
playlist, and it resumed at the old spot and the next checkpoint wrote that over the 830. A detail
with the same membership now rebases its fresh rows into every snapshot, keeping order and undo
history. (The ids-only guard is older than this round; the write that depended on it was new.)
- Live YouTube search results carry the same per-user position and open the same player, but their
cache was never patched. It is now — patched only, never invalidated, since refetching an
API-sourced search page spends 100 quota units.
Tests for what had none: a Playlists harness (PlayerModal stubbed to record its launch props) and a
PlayerModal harness with a fake IFrame API that records where each item is started — the latter pins
that stepping back to an item watched earlier in the same session resumes at THIS session's position.
3 mutations, each red on its assertion.
Deep-review R3. When 889d2acf put the sort into the collection detail's query key, two things on
the page did not follow it:
- The watched toggle's optimistic write kept the key's old spelling. setQueryData matches exactly,
so it landed on an entry nothing reads, and its resync only invalidates the library root — the
click succeeded server-side and changed nothing on screen (badge, icon, aria-label) until a
reload; a second click toggled it back. One `detailKey` now serves the read and the write.
- Every sort change was a key with no data yet, so the whole page — hero, grid, and the Sort control
with its open panel — was swapped for a spinner, and focus fell to <body>: key and direction could
not be set in one visit. The previous page now stays on screen while the new order loads, for the
SAME collection only (stepping to another must not show this one's hero meanwhile).
Verified on localdev (Rocky Collection): the toggle flips the poster; IMDb rating then Descending set
in one open panel, the hero never blanked, order reversed. 3 mutations, each red on its test.
Deep-review R2, plus one defect found while verifying it on localdev.
- The status poll re-armed every second after a FAILED request: `data` stays empty on an error,
so it read as "still preparing" for as long as the player was open — and the request was not
quiet, so a persistent 500 re-opened the global error dialog every second. It now backs off to
10 s (a blip still heals on its own) and is quiet; the player already shows the failure inline.
- A <track> whose own VTT fetch failed never fetched again: with the server's answer unchanged its
key was unchanged, so React kept the dead element — silently blank subtitles. The key (now built
in the hook, where it is tested) carries a failed-load count; after two remounts the track is
reported as unavailable instead.
- Found in UAT: until the first timeupdate `abs` is 0, so after a resume every poll said at=0 and
the first window was cut around the head of the film (a resume at 20:00 polled at=0 for as long
as it stayed paused). The poll reports the session's start until playback ticks. And the window's
dedup identity is the track plus a window-length bucket of its start — keyed on the track alone
(R1), the right window queued behind that wrong one.
The poll's start/stop rules finally have tests (every earlier one resolved a single fixed answer).
Verified on localdev against Westworld S1E6 resumed at 20:00: the polls carried at=1197, one
subtitle line on screen at 21:02, the poll stopped at `ready`; the real file's window (56 cues) +
full pass (608) merged to 608 — the old merge would have served 664. 11 mutations, all red.
Deep-review of the bug-hunt round (R1). The background extractor from ed4ab4e1 held, but its seams
did not:
- Window + full pass doubled every cue in the first hour. merge_vtt de-duplicated on the raw timing
text, and the two passes spell a timestamp differently (shift_vtt writes HH:MM:SS.mmm, ffmpeg's
webvtt muxer drops the hour below 1 h; shift_vtt(-0.0) never normalised the full pass). Timing is
now parsed and re-emitted canonically, keyed with the cue text so two cues sharing a time survive.
- A full pass that exited non-zero was published as complete: "ready" forever, subtitles ending
where ffmpeg died, and no eviction to ever retry. It now covers only up to its last cue, is never
complete, and records the failure memo.
- The window's dedup identity held the live playhead, so every 1 s poll queued another read of the
file. It is keyed on the track now, and a queued job re-checks coverage before reading anything.
- A sidecar track waited behind an unrelated embedded full pass (the full job is per item). Its
window job now fetches it whole — one small request.
- The <track> fetch ran the scheduler at position 0: an unrequested window over the head of the file,
and a refusal of a window that was on disk. It is a pure read of the cache now.
- publish() is serialised (a lost update became permanent once complete was set); the pool is shut
down and running ffmpeg children killed at app shutdown (they held interpreter exit for minutes);
?at=inf is a 422, not a 500; and both subtitle routes use current_user instead of the cookie-only
poster gate, which let a suspended or epoch-revoked session keep scheduling reads (pre-existing).
_extract_and_publish finally has tests of its own — every earlier test stubbed it out. 19 mutations,
each red on the assertion that names it.
Items 2, 4, 5 and 7 of the bug hunt, one round. They are independent, but they are all the same
kind of thing: an action that exists somewhere else in the app and is missing where you actually
want it.
**5 — the channel name in the YT player is a link again.** It was a `<button>`, so middle-click and
Ctrl/Cmd-click did nothing on a thing that looks exactly like a link (the YouTube icon beside it IS
an anchor and opens a tab). It is now a real `<a>` whose href is the same `?chan=<id>` URL
`openChannel` pushes, with a plain left click still taking the in-app path — that path does more
than navigate, since the player must close and reclaim its own Back-to-close history entry first.
The "is this click ours or the browser's" test is now ONE predicate (`isPlainLeftClick`), which
VideoCard's card link also uses, so a second such link cannot forget the middle-click arm.
**2 — "Remove" from inside the player.** The parent owns the removal (the SAME `removeItem` the row
menu and the inline ✕ use: one confirm, one optimistic update, one undo-stack rebase), and the
player owns what happens next, because it is the only side that knows which item is playing — its
launch queue is frozen at mount. It drops the item, plays what followed, and closes when the list
runs out. That forced a real fix in Playlists: the player's mount was derived from `shown` every
render, so anything that dropped the launch row yanked the player off the screen mid-playback. That
was a latent edge (change the view and it happens) and the normal case for this feature. The parent
now snapshots what it launched with, and only onClose takes the player down.
**4 — "Move to <the last list you moved to>".** The submenu is two levels deep and moving a run of
videos into the same list is the shape this gets used in. The remembered id is a per-account pref;
the row is derived from `moveTargets`, which already excludes the list you are looking at, so it
cannot offer a move to where the video already is. Remembered only on a SUCCESSFUL move — a
shortcut to a move that failed is a shortcut to failing again. Absent, not disabled, when there is
nothing to repeat: there is nothing to explain.
**7 — the live-YouTube results can be sorted.** Client-side, over the page already loaded. Not a
new `order=` param, and this is the decision worth recording: the owner's case is "even one page of
results", i.e. reorder what is on screen; a re-query would REPLACE the results, which is not what
"sort these" means; and the zero-quota scrape source has no order parameter at all, so a
server-side control would silently do nothing (or quietly cost 100 quota units) depending on which
source served the page. Only keys a search hit carries are offered — `subscribers`/`priority`
describe a subscription, which a search hit is not, and `shuffle` on a ranked list is noise. Its
own pref, so sorting a search page never re-sorts the library behind it.
Verified in the browser on localdev against the real account: the channel name renders as
`<a href="…&chan=…">` and a ctrl-click / middle-click is left to the browser while a plain click
still opens the channel in-app; removing from the player stepped 1/3 → 1/2 onto the next video and
closed on the last one; the row menu grew "Move to “…”" only after a move and only on a list that
is not the target; and the results sort reordered by views 41.7M → 29.8M → 24.2M → 15.8M. (The
pane's own ctrl-click arrives without the modifier — a harness limit, so the modifier cases were
driven as real MouseEvents.) Scratch playlists and the touched prefs were cleaned up afterwards.
22 vitest, mutation-checked against 14 mutations (no href at all · href without the rest of the
query · every click hijacked in-app · the middle-click arm dropped from the predicate · plain click
no longer in-app · always jumping to the top of the list · the removed item left in the queue · no
clamp when the last item goes · an empty queue not reported as "nothing left" · sort-then-reverse
flipping ties · unknown values first · titles compared numerically · shuffle offered on search
results · direction ignored) — each red on exactly the test that names it. Item 4 has no test of
its own: its rule is enforced by construction, not by a check.
Two halves of the same report ("Collectionben nem működik a rendezés?"). The pasted browse URL
sorts correctly — that part was working. What did not:
The Collection INFO page had no sort control at all and a hardcoded release-year-ascending order.
It now shows the browse grid's Sort control, with the browse grid's vocabulary, because a second
sort vocabulary on a neighbouring surface is exactly the kind of drift this repo keeps paying for:
`PLEX_SORT_KEYS` is now one tuple in `plex/query.py`, and a test pins BOTH implementations to it —
browse's key→column map (extracted from the handler to be reachable) and the collection page's
key→member map. A key added to the vocabulary without teaching one of them fails the gate instead
of putting a dead option in the menu.
Ordering a collection cannot be a single ORDER BY: its two legs are different tables, so the merged
member list is sorted in Python. Members with no value for the key sink to the END in both
directions (a movie with no rating is not "the worst rated" — the browse grid's `nullslast()`), and
ties always resolve title-ascending, so flipping the direction reverses the ranking rather than the
alphabet inside it. `lastwatched` needed the show leg's `resumed_at`, so the page now takes ONE
rollup pass that yields both the badge status and the recency instead of the status alone — without
it that menu entry would have ordered every show as never-watched. The default stays release-year
ascending: a franchise reads in release order, and that is what the page did before.
The other half: a shared Plex link carries the collection filter as bare rating_keys, so a link
opened cold had no name for them — the sidebar token rendered `17045` and the chip could only say
"Collection". `GET /api/plex/collections` takes a `keys=` CSV now, and a display-only hook resolves
just the missing ids. It deliberately does NOT write them into `collectionTitles`: that map is
hashed into the Plex query keys, so storing a resolved label there would split the cache from the
identical state and refetch the grid for a piece of text.
Verified in the browser on localdev against the real Plex library: the movie collection's menu
offers Length and the show collection's does not (the same scope rule as browse); rating-ascending
reorders 8.1/7.4/7.2/6.6/6.5 correctly; `lastwatched` on a show collection brings the in-progress
shows up; the choice survives a reload (it is a server pref, on its own key so opening a collection
never re-sorts the library behind it); and the reported URL now reads "Pirates of the Caribbean
Collection" in both the sidebar and the chip.
12 pytest + 8 vitest, mutation-checked against 14 mutations (direction ignored, nulls first, no
stable tiebreak, unknown key untouched, wrong default, show recency dropped, empty `keys=` falling
through to everything, a key missing from either sort map, every id refetched, ids unsorted,
fetched titles unmerged, the chip ignoring them, the hook's result never reaching the builder) —
each red on exactly the test that names it.
"Get full history" on a channel page left the list on the cards the channel already had. The
count above it climbed into the thousands, the grid did not move, and only a browser F5 brought
the rest in.
The mechanism: the backfill runs in a BackgroundTask, so the PATCH that starts it returns before
a single row is written — and the mutation invalidates `qk.feed()` on that RESPONSE, which just
refetches the OLD catalogue. Nothing afterwards refetches it: the feed's infinite query has no
interval, and the app-wide defaults are `staleTime: 30s` + `refetchOnWindowFocus: false`. The job
emits no event either (the realtime layer only carries messages today), so the channel detail's
4s live poll is the only thing in the app that sees the job progress — and it refreshed only
`qk.channel`.
So the poll now drives the grid too: on a tick where the stored count moved (or the backfill
finished), it invalidates `qk.feed()` and the `qk.feedSummary()` behind the count pill. The window
is the poll's own (`deep_requested && !backfill_done` on the previous reading), so a detail
refetch caused by some other mutation — each of which invalidates the feed itself — cannot come
through here twice.
Measured on localdev against the real account, same instrument, `document.visibilityState`
forced visible (the poll deliberately pauses on a hidden tab, and a headless pane reads hidden —
without that the first measurement was about the harness, not the fix):
before — CASCADA 19 -> 176 stored: the header updated live to "162 normal", the grid and its
count pill sat on "5 items" for 50s.
after — fern 54 -> 139 stored: "54 items" -> "139 items" by ~12s, no reload.
6 vitest, mutation-checked against six mutations (no invalidation at all; grid only, count pill
forgotten; predicate ignored; no channel-identity guard; completion with an unchanged count
dropped; poll-window guard dropped) — each one red on exactly the test that names it.
Bug hunt 2026-09, item 14 — "how does the multichannel downmix work, and is it
normalised?" It isn't: `plex/stream.py` has always encoded `-c:a aac -ac 2 -b:a
192k`, i.e. fixed stereo with ffmpeg's default downmix matrix and no loudness
handling at all, which is why a film mastered for a cinema arrives with whispered
dialogue and deafening action.
New admin knob `plex_audio_loudnorm` (Configuration → Plex), DEFAULT OFF: it makes
a film markedly louder and narrows its dynamic range, which is a taste call rather
than a fix. It sits at the end of the Plex group, after the software-transcode
knobs — it is an audio concern, not one of the libx264 ones.
-af aformat=channel_layouts=stereo,loudnorm=I=-16:TP=-1.5:LRA=11,aresample=48000
Three things about that chain are load-bearing and each has a test:
* the downmix is FORCED FIRST rather than left to `-ac 2` — measuring 5.1 and
folding it down afterwards would normalise something other than what is heard;
* the trailing resample is not decoration: `loudnorm` runs its own resampler at
192 kHz internally, which the AAC encoder will not take;
* one `-af`, no stream index, so EVERY audio rendition of a multi-audio file is
normalised — not just the first.
The knob is also part of the session directory name (`…_multi-ln_+0.00`), because
it changes the audio the segments contain, and a session left over from before the
admin flipped it must not be mistaken for one that honours it. The suffix is
appended with a DASH: `_SESSION_DIR_RE` treats the tag as one underscore-free
field, and that regex is the only thing keeping the orphan sweep off what may be a
shared scratch path. A test pins both halves — the name changes, and the sweep
still recognises it.
Measured end-to-end on the app's own HLS segments (a real 5.1 AC3 file, same item,
same 160 s offset, same ten segments): OFF -35.2 LUFS, ON -17.2 LUFS. On a 60 s
bench: I -27.6 → -16.6 LUFS with LRA 13.3 → 9.8 LU. A/V sync is untouched — under
this pipeline's own `-copyts` the audio start PTS is identical with and without the
filter, so the 3 s lookahead desyncs nothing.
Direct-played files are served untouched and stay so; a 5.1 mix is always `remux`,
so the reported case is fully covered.
Bug hunt 2026-09, items 8 and 12 — the two things the Plex player was measurably
lying about.
Item 8: picking a subtitle track showed nothing for a long time, and the reported
workaround was to stop/start until it suddenly worked. Measured on prod against a
17.06 GB MKV with two embedded SubRip tracks:
full pass, one track 154 s
full pass, both tracks in ONE run 111 s
-c:s copy (no WebVTT conversion) 143 s
from 1:00:00 to the end of file 66 s
a 630 s window at 1:00:00, both 8 s
So the conversion was never the cost: ffmpeg streams every byte of the container
between its seek point and its stop point, and the array delivers ~140 MB/s. The
only lever is reading less. Three changes follow from that:
* a window around the playhead is extracted and published FIRST (seconds), and
the full pass completes behind it;
* every embedded text track rides that same single read — picking Hungarian used
to cost a full pass and then switching to English cost another one;
* none of it happens inside a request any more. The endpoint answers 202 while a
background worker extracts, the player polls /status (carrying its position, so
the window follows the viewer), and one ffmpeg run finally has a timeout.
-copyts is deliberately absent from a windowed pass: with it ffmpeg ignores -t and
reads the whole container anyway (measured — a [0, 600 s] window still took the
full 145 s), so a window comes back on a zero-based clock and is shifted to
absolute time here. Cached tracks written by older builds have no coverage sidecar
and count as complete, so nothing is re-extracted after the upgrade.
The <track> is no longer mounted before the server has something to put in it; a
"Preparing subtitles: …" notice stands in, and a track that can never be shown says
why instead of leaving the viewer to conclude the picker does nothing.
Item 12: "Bitrate" read a constant 211 kbps on every remux, because it was the
master playlist's BANDWIDTH — which ffmpeg derives from the one stream whose rate
it knows, the -b:a 192k AAC track, plus 10%. It is a measurement now: bytes per
second of MEDIA from hls.js FRAG_LOADED, accumulated per fragment type and summed
(a multi-audio session's video and audio renditions cover the same media seconds,
so pooling them would report half the truth), and an honest dash until the first
fragment lands. The panel also finally names the audio and subtitle track in play.
Verified in the browser against the real Plex library, and every new test
mutation-checked.
The button deletes the local row and mirrors the un-mark to Plex, but it decided whether
to push from the previous STATUS alone (prev_status != "new") — and an in-progress item's
row is status="new" carrying a position, which is exactly how the importer maps a bare
viewOffset. So for the single case the button exists for, nothing was ever sent to Plex:
the offset stayed, and the next pull imported it straight back. The user cleared Protector
twice and it returned to Continue watching at 92s both times.
Push now when Plex holds anything to clear — a view count, or a resume point. "hidden"
stays Siftlode-only and pushes nothing, which is what this endpoint's docstring always
claimed and the status-only guard contradicted.
One unscrobble clears BOTH viewCount and viewOffset — measured against the live server from
a watched+offset state and an offset-only one, so the push itself needed no change. The
client docstring that claimed "clears viewCount" is what sent the first diagnosis down a
wrong path (a timeline write with time=0, which does not clear an offset at all); it now
records what the server actually does.
Verified end to end on the real Plex server: the button leaves viewCount=None viewOffset=None,
and a full 44,885-item re-import brings back no row for it.
The player checkpointed every 5s and the server stored it, but nothing wrote the new
position back into react-query: the close path only marked qk.feed() stale with
refetchType "none", and nothing refetches the feed until it remounts (staleTime 30s,
refetchOnWindowFocus off). So the resume read the position the cache had been FETCHED
with, and the next checkpoint then persisted that older value over the newer one.
Measured on localdev: played to 830s, closed (server had 830), reopened the same card
without a reload -> the row was back to 751s. 79 seconds of progress destroyed by a
reopen, and the card's progress bar never moved either.
The Plex player already solved this for its detail query, with a comment describing the
identical failure; this is the feed-side twin that was missed in that sweep. patchCachedProgress
writes the position the SERVER stored (the endpoint clears trivially-early and near-finished
checkpoints to 0, so mirroring our own reading would disagree with the row) into every cached
feed page and playlist detail, keeping array identity where nothing changed. A session map
covers the second half: the launch queue is frozen at mount, so stepping next->prev back to an
item played earlier in the same session resumed from that snapshot.
Verified in the browser: 811 -> close 816 -> reopen 825, and the card now shows Continue/Restart.
Round 4 reviewed round 3's commit. Nothing in production behaviour was wrong; what was wrong
was the round's own test and its comments, which is the failure class this loop keeps
finding.
- The new NavigationProvider case named after the `&& !channelViewRef.current` clause never
reached it. With storage cleared and a bare URL the provider lands on "home" (E24's own
default), not "feed", so asking for "feed" was an ordinary page change and the guard's
first conjunct short-circuited. Deleting the clause left all 828 tests green — a dead nav
control (the rail's own module stops closing an open channel page) with no red anywhere.
The test now asserts the precondition and calls `setPage` with the page that IS current.
- Both new Home cases rejected a promise and then asserted after one fixed macrotask.
React-query notifies through `notifyManager`'s own `setTimeout(0)`, registered after any
timer the test schedules, so the assertion raced it; two reviewers measured intermittent
reds. They poll for a signal that the rejection LANDED instead — the reroll button
releasing, which is also the honest one for the empty-answer case, where the empty state is
already on screen while the press is in flight and asserting it early would pass either way.
- `buryUnder`'s comment still opened with "one rule for both entry points" ten lines above the
paragraph saying the rule is complete only for one of them, listed the non-reporting palette
rows without `openChannel` — which has its own `replaceState` no-push branch and is reached
from PlayerModal too — and pointed at a ROADMAP.md this repository does not contain.
- `history.dom.test.ts`'s header claimed its layers close by genuine Backs only, which two of
its own tests contradict; and `CommandPalette.ordering.dom.test.tsx` still claimed to be
"the only assertion in the suite" doing what the file corrected in round 3 also does. Round
3's message said that pair was handled and only one side had been.
The provider test was re-mutation-checked (dropping the clause now goes red), and both
hardened Home tests still go red against their own reverts. Seven consecutive full runs green.
Gate: ruff, ruff format, 3x tsc, eslint (0 errors), prettier, knip, vitest 828, pytest 1005.
Round 3 reviewed the round-2 commit across six dimensions. Fifteen findings, five survived
three adversarial refuters each; three died on 2+ refutations (a keyboard-focus claim that
is a property of `ui/Button` at all six of its `loading` call sites, a contract-test
false-pass that needs a deliberate revert plus a self-incriminating comment, and a
MAX_EXCLUDE coupling whose slice is unreachable from the real client).
- `rerolling` was still wrong for a FAILED press. `data === undefined` is true forever for a
key whose request failed (`retry: false`), so the next unrelated mutation re-disabled the
reroll button — the same defect round 2 reported, one branch over. `!isFetched` is the
flag that remembers a key settled; `!isError` would not have worked, because query-core
resets a data-less query to "pending" for the duration of a refetch, hiding the error
exactly when it matters.
- The hero fallback used `??`, which cannot tell "this budget's last pick resolved to NO
hero" from "nothing remembered". So a press that failed after an empty one resurrected the
card the page had already excluded as seen and just said it had nothing beyond. The tagged
slot is now authoritative for its own budget.
- `setPage`'s boolean — the input the whole burial rule reads — was asserted by nothing.
The type fences the SHAPE, so deleting a `return` fails tsc, but swapping the two values
passed typecheck and all 822 tests. NavigationProvider gets its first render test, on the
value AND on the effects the early return skips.
- `closeThenNavigate`'s door into `buryUnder` was equally unpinned: reverting that one call
site to the old unconditional burial left the suite green. It has the mirror case now.
- The comment claiming one rule serves both entry points was overstated. `buryUnder` fails
OPEN, and `CommandItem.run` is typed `() => void`, so only the palette's bare
`() => setPage(page)` rows forward the boolean at all. Nothing regressed — that path buried
unconditionally before — but the prose said more than the code does, so it now says what is
true and points at the deferred fix. Same for a "the only assertion in the suite" claim
that the file it cites as precedent also makes.
Every new test mutation-checked red for the right reason, both edges where there are two.
Gate: ruff, ruff format, 3x tsc, eslint (0 errors), prettier, knip, vitest 828, pytest 1005.
Nine of the round's ten surviving findings. The tenth (a failed shelf reorder leaving an
order the server rejected on screen) is on ROADMAP.md: its obvious remedy reintroduces the
overlapping-move defect the code already documents, so it wants a design, not a review edit.
- `navigateOverOverlays` buried the live overlay entries before knowing whether the
navigation pushed one. `setPage` no-ops when the page is already current and Home is the
default landing page, so Shift+W over a player ON Home marked its entry buried under a
push that never happened: the unmount consumed the mark instead of reclaiming the entry,
and the next browser Back was a press that did nothing. `setPage` now reports whether it
navigated, and one `buryUnder` rule serves both entry points — `closeThenNavigate` takes
the same no-op navigation from a palette command for the page you are already on.
- The hero could answer a new time budget with the previous one's pick. The chip row resets
seed/seen/press but cannot reset an answer that already resolved, and `keepPreviousData`
chained through the disabled seed-0 query — so a reroll under "10 perc" showed the film
picked for "Bármennyi" while it loaded, and kept it if the request failed. `lastPick` now
carries the server's own budget echo, and the placeholder is scoped to the budget.
- The reroll button spun and swallowed clicks after mark-watched, bookmark, pin and shelf
moves: the pick key sits under the invalidated `discover` prefix on purpose, so raw
`isFetching` was true for background refetches. It now means "a press of mine is
outstanding" — which still covers the first press, where there is no placeholder.
- Failed writes on Home toasted the page's LOAD-failure string over a fully loaded page;
they get their own key (HU+EN). The empty-state reroll had no `loading`, so the press was
unacknowledged and a second click skipped a tier. `danger` moved into the shared unpin row
so the same command is not red on the shelf and neutral in the rail below it.
- Two comments claimed things the code does not do: the `exclude` cap guards no "unbounded
scan" (it never reaches SQL), and the route test that named the cap sends three keys.
Tests: the burial rule gets its own file (it is the one assertion that lets a reconcile
issue a real programmatic back()); App's keymap wiring gets a source contract, because
dropping it left all 818 tests green; `add_pin`'s IntegrityError arm was reached by nothing.
Every new test mutation-checked red, both edges where the predicate has two.
Gate: ruff, ruff format, 3x tsc, eslint (0 errors), prettier, knip, vitest 822, pytest 1005.
`add_pin`'s IntegrityError handler assumed the unique constraint was the only way in and
re-raised otherwise, so a target row pruned between `resolve_ref` confirming it and the
INSERT (a Plex sync dropping a removed item) tripped the FK and handed the user a 500.
Semantically that is the same "no such thing" the pre-check already returns. Untested: the
race needs two concurrent transactions.
The five per-kind unique constraints had no test at all — the route's own dedupe returns
before reaching the DB, and the migration drift check filters constraint diffs out as noise,
so they could go missing with every lane green.
Frontend, all from the same review: the shelf's unpin row is now `danger` like every other
destructive row in the app; the hero's artwork button says what it will actually do and
names the item instead of always announcing "Play"; the empty-budget message offers a reroll
rather than `StateMessage`'s "Retry", which promised a re-run of something that had not
failed; `setState` and the watch-later toggle catch their rejections instead of leaving an
unhandled promise and a click that silently does nothing.
`StateMessage` gained a `compact` rung, which retires the frontend's only two `!important`
overrides — they existed to beat the primitive's own padding from the call site, and were
the copyable idiom for the next surface that wanted a compact empty state.
Dead vocabulary removed or made load-bearing: `DISCOVER_BUDGETS` was exported, re-exported
and consumed by nothing while `Home.tsx` re-typed the same four bands (a `readonly
DiscoverBudget[]` accepts any subset, so a fifth band would have typechecked green and
rendered no chip); `PIN_KINDS`' value export had no consumer at all; `useComfortPin` no
longer publishes `pinIdOf`/`toggle`, which nothing outside the module calls.
Three comments corrected against the code they describe: a pointer to a `resolve.py` that
was never written, the file header's claim that Home defines no card component of its own,
and the `seen` window still quoting the 8 that was measured and abandoned for 32.
The rest of the review's survivors are logged on ROADMAP.md under "Deferred — E24 epic-end
deep-review (2026-09-08)" — the ones whose dissent was substantive, and the ones whose fix
is wider than the finding.
`test_only_a_show_row_pays_for_a_resolution` seeded no watch state, so every rail query
inner-joined its way to zero rows and the loop body — with it the assertion — never ran. It
now seeds a video, a movie and an episode, and asserts all three kinds are present before
the loop, so it cannot go vacuous again.
The Plex half of the surprise rail had no test for either of its promises: every test that
reached `plex_surprise` passed budget="any" against a library with no state rows, so the
duration bounds and the already-started exclusion could both be deleted with the suite
green. The YouTube half has had tests for the same two rules since S1 — the asymmetric half
of one principle.
`plex_item` and `plex_show` — two of the five pinnable kinds — were pinned by no test at
all, leaving the rating_key → internal-id translation and both shelf arms uncovered,
including the self-reference cycle that once 500'd the shelf. And `GET /api/discover/pick`
was exercised only through its domain function, so the route's own query-string translation
could regress silently; the new test excludes all three candidates, which is what proves the
whole list parsed rather than one unsplit string.
`PIN_COLUMNS`/`PIN_KINDS` claimed in a comment to be the one place the pin vocabulary is
spoken and had no readers anywhere: `resolve_ref` re-spelled the column names and the route
re-typed the kind list. Both now drive off the constants.
The `_NO_TS` sentinel was introduced as a crash fix and is not one. The key it replaced,
`(ts is not None, ts)`, cannot raise — tuple comparison finds the first UNEQUAL element via
`==`, so two null keys compare equal and `None < None` is never reached; three independent
runs confirmed it over every permutation. The rationale now says what the sentinel actually
buys, lives once in `plex.rollup`, and the third rail that was still on the older spelling
has been swept to match. Its test is renamed and re-scoped to the mutation it does catch.
Frontend: the reroll's second press asserted `expect.any(Array)`, which cannot tell an
accumulating exclusion window from a replacing one — both that and SEEN_WINDOW = 1 (each
reinstating the measured nine-card cycle) left the suite green. And `onInfo`'s jump type was
untested while `onPlay`'s was pinned twice, so pointing the ⓘ affordance at the player
typechecked and shipped green.
All eight fixes mutation-checked; every mutant died on its intended test.
Shift+W was dead whenever anything was open. The global handler called `setPage` directly,
so the page change unmounted the player (it lives inside the module column) and that
unmount's reconcile reclaimed the entry the navigation had just pushed — the player closed
and the page snapped back. `closeThenNavigate` cannot serve here: a global key does not
know which overlay is open, or whether one is. `navigateOverOverlays` is the same fix one
step shorter — bury the live overlay entries first, so the unmount consumes the marks
instead of issuing `history.back()`.
`q.isError` had no `&& !data` guard, alone among 33 call sites and against the convention
`ui/QueryState` documents: react-query flips status to error on a failed REFETCH while
keeping the payload, and every mutation here invalidates the discover prefix, so one blip
while marking a card watched blanked a fully populated landing page.
The reroll's key rooted at "discover-pick", outside the prefix every mutation invalidates,
so a rerolled hero outlived the page: play it, finish it, and it was still the answer with
a zeroed progress bar. It now sits under the same root. A FAILED reroll also rewound the
hero to the card the page loaded with, several presses back, because a new key per press
means `keepPreviousData` (pending-only) cannot cover an error; the last resolved pick is
held instead.
The bookmark on a rail tile never updated — `VideoCard` reads `video.saved` from this
payload and keeps no optimistic state — so the star stayed empty and the next click
re-sent the same call.
The shelf reorder wrote a captured snapshot back on failure rather than re-asking the
server, which its own comment promised: two overlapping moves each reverted to their own
snapshot and settled on an order neither side held, and since `before` is the whole
payload a rollback also discarded any rail refresh that had landed meanwhile. It now
invalidates the discover PREFIX on settle, which also fixes the sibling budgets keeping
the pre-move order.
The hero/rail de-duplication matched one key, so for a Plex series — the rail row is the
show, the hero is the episode it resolves to — the same series printed twice; it now
matches both keys, as `hero.pick`'s own exclusion already did. And when the hero empties
the resume rail, the rail is hidden rather than announcing "nothing half-finished right
now" under a hero captioned "where you left off".
Ten regression tests, all mutation-checked. Three of them were rewritten after the first
mutation round showed they could not go red: two asserted before the state under test had
settled, one matched a title the Plex tile does not render.
The resume rail built its own SELECT instead of going through `filtered_query` and copied
only three of its predicates, so it dropped the per-user channel blocklist the module
docstring claims it inherits. Blocking a channel never touches `video_states`, so a
half-watched video from it kept its resume position and headed the rail — and, at the
default budget with the resume tier leading, became the landing page's hero: the app
answering "what should I watch" with the thing the user had just banned. The clause is now
`feedquery.not_blocked_cond`, shared with `filtered_query` rather than copied a second
time; the feed's SCOPE clauses stay deliberately unapplied, for the reason `yt_video_row`
already gives.
`next_position` returned the shelf's row COUNT while nothing ever renumbers positions —
`remove_pin` deletes, and every FK cascade removes a pin when its target disappears. Past
the first hole the count under-counts, so a new pin landed in the middle of the shelf and,
past the second, on a position another pin already held. `max(position) + 1`, the idiom
`routes/playlists` already uses.
`next_episode` took the FIRST in-progress episode in broadcast order while `plex.rollup`,
which drives the show page's Resume button, takes the last. One show therefore had two
different "continue watching" answers, with the Home tile ranking it by its most recent
touch and then offering the episode abandoned weeks earlier. The new test asserts the two
implementations against each other rather than against a literal.
All three mutation-checked: each new test goes red for the defect it names.
S4's reachability half. Home already led the nav order; what was missing was every
way IN other than the rail.
- Settings → Appearance → Startup: "Start on Home", the documented escape hatch for
the landing override. Deliberately still localStorage and not a server pref —
`loadInitialPage` answers before /api/me lands, and a landing module decided after
the first paint is a visible jump. Its two stored values are now ONE exported
constant, so the writer and the boot-time reader cannot drift on a magic string.
- The command palette gets "Meglepetés" / "Surprise me": it navigates to Home and
rerolls the hero on arrival. No second "Mit nézzek?" row beside it — the Home
module row already matches that phrase through its descriptive title.
- A global Shift+W jumps to Home. A chord, not a bare `w`: the Plex player binds bare
letters and `w` sits on `s`, so a blind reach for the subtitle key would have thrown
a viewer out of the film.
The palette's request crosses a lazy module boundary, so it is a token that waits
(lib/homeIntent) rather than an event fired into an empty room. Browser-measured:
taken with a compare-and-clear, because StrictMode's double-invoked mount effect
otherwise rerolled twice and landed on the page two picks deep.
HU/EN parity kept; UAT on localdev in both languages, including the empty shelf.
All three were measured before they were touched, and the diagnosis moved twice.
**Back from a Plex title played off Home landed on the Plex module's root.** One
click pushed three history entries: the page navigation, the sub-view opened by
the `jumpTo` hand-off, and — under StrictMode, whose effect body runs twice
before `setJumpTo(null)` re-renders — that same sub-view a second time. So the
first Back visibly did nothing and the second stranded the user on a Discover
page they had never visited.
A cross-module hand-off is the CONTINUATION of the navigation that brought you
there, not a second navigation, so it now REPLACES the page entry instead of
stacking on it (`useHistorySubview.open(view, {replace})`, asked for by the jump
target). Replacing is also idempotent, which is what disarms the StrictMode
double-run. The intent rides on `PlexJumpTarget` rather than being inferred,
because the header typeahead jumps from INSIDE the module, where the entry being
replaced would be a sub-view the user really was looking at.
**"Mást kérek" cycled nine cards and never left the resume tier.** Measured:
twelve presses, nine distinct titles, period nine, every one `tier: "resume"` —
the shelf and the surprise tier were unreachable by pressing the button. Two
causes multiplied: the tiers were tried in a fixed priority order over a 20-row
resume rail, so the button just walked down it; and the client's rolling
exclusion window of 8 guaranteed the wrap.
The press count now rotates which tier answers first (`start`), so each press
offers a different KIND of answer — the user's call, given the choice between
that, a wider window, and a surprise-only reroll. The fallthrough is kept, so a
press does not come back empty because the tier it led with had nothing. The
window is matched to the server's own cap rather than a smaller number of ours.
**A half-watched series on Home opened the show page.** A movie already went
straight to the player; only shows did not — and the comfort shelf and the hero
both resolved a show to its next episode while this rail alone did not, handing
back six seasons to choose between. The rail row now carries `next`, and the
client already played `next ?? card`, so this is a backend fix.
Affording it needed `queries.next_episode` to stop loading every episode of the
show into Python to pick one of them (236 rows for Friends, 227 for JAG) — three
ordered LIMIT 1 probes instead. The show→episode rule now lives in ONE function
(`rails.playable`), which is what the shelf, the hero and the rail all call; two
implementations are how they came to disagree.
Ten mutations run, all ten red for the right test — including a first pass where
the press count reached `api.discoverPick` and stopped there, invisible because
every Home test mocks that module. The slice now has its own URL test.
Verified in a clean browser tab against the live account: one Back returns to
Home, the rotation gives resume/comfort/surprise across nine presses with no
repeat, and Westworld opens at S1E1 rather than at its show page.
S3 wires the shelf's editor. The gesture is the E10 right-click menu, so the
decision "this is comfort media" can be made where the thing is — the feed, a
Plex grid, a channel row, a playlist — rather than only on the page that lists
the shelf.
One shared piece, `components/comfortPin.tsx`: the five pin-kind constructors,
the shelf query, and ONE `ActionMenuItem` that switches Pin/Unpin honestly (and
says "not known yet" with a disabled row until the shelf has answered, since a
confident "Pin" over something already pinned is exactly the quiet lie the house
rules forbid). Surfaces that already have a row menu compose the item in
(`Playlists`, `PlaylistsRail`); the rest mount `ComfortPinMenu`.
Two structural notes:
- The feed's menu lives in `VirtualFeed`, not in `Feed` — that is the list, and
it is also what gives the channel page the gesture for free, since it renders
`<Feed channelScoped>` all the way down.
- Cards take `menuProps` as a FUNCTION of the row, not the built props object.
`triggerProps` is stable across renders while its return value is not, and a
fresh object prop would blow past `VideoCard`'s memo for every card in the
virtualized feed on every parent render.
`DataTable` and `ChannelLayoutGrid` grow a generic `rowProps` so a channel row
answers the same gesture in both of the manager's layouts; the table stays
domain-free and never learns what a menu is.
On Home the shelf tile's menu is the whole editor: four move rows (front / left
/ right / end, reusing the playlist menu's ordering idiom) plus the unpin, over
menu rows rather than drag & drop — dragging inside a horizontal, scrolling
Carousel fights the scroll, and menu rows are keyboard-reachable for free. The
reorder is optimistic on the cached payload and re-asks the server on failure
rather than showing an order that was never accepted.
The exhausted shelf tile (a pinned channel with nothing unwatched left) is
`aria-disabled`, not `disabled`: measured in Chromium, a disabled button still
receives contextmenu and pointerdown, but it loses focus — and with it Shift+F10,
the only keyboard route to the menu it most needs.
Verified end to end in the browser against the live account: all five pin kinds,
the reorder (persisted across a reload), the unpin, and the honest Unpin label on
a resume tile that is already on the shelf.
E24 S2. A new `home` module in layout C+ (the mockup the owner picked): the
time-budget chips on top, one large hero that answers them, then the comfort
shelf and a cross-source resume rail. Home is the landing — deliberately OVER
the remembered last module, since landing on whatever you happened to close the
app on is the habit this epic exists to replace; `LS.startOnHome = "0"` opts an
account back out (Settings toggle lands in S4).
No third card component: each unified card carries the native feed / Plex
payload, so the rails render through the existing VideoCard (`cardsSmall`) and
PlexPosterCard, and `aspect` decides the tile shape without branching on source.
Plex playback hands off to the Plex module via a new `jumpTo` type — the player
lives there with its lazy hls.js and its provider tree, and mounting a second one
on the landing page to save one hand-off would be the expensive way round.
Four defects found by driving it in a real browser against the live account, each
now pinned by a mutation-checked test:
- **The reroll gave the same card back.** A new seed only re-draws the SURPRISE
tier; the resume tier deterministically returns the most recently progressed
thing that fits, so anyone with something half-finished pressed a dead button.
`/pick` now takes an `exclude` list of `<source>:<id>` keys and the client keeps
a rolling window of 8. Excluding a SHOW also excludes the episode it resolves
to — otherwise the same suggestion comes back wearing a hat.
- **"Nothing to suggest yet" flashed over a full library** whenever a reroll was
in flight: the new query key has no data for a beat. An empty state standing in
for a loading one is the dishonesty the house rules name outright. The hero now
falls back to the loaded payload while a pick is pending, and only trusts a null
once the pick has RESOLVED.
- **A budget change replaced the page — chips included — with a spinner**, so the
control you just clicked vanished under your cursor. `keepPreviousData`, the
same reason the feed uses it.
- **The hero printed twice**, once as itself and once as the first resume tile.
Two smaller UI calls made from the same session: the budget caption sits NEXT to
its chips rather than in SegmentedToggle's justify-between mode, which on an 80rem
page throws the answers to the far edge; and the feed takes the YouTube mark its
own toolbar already uses, since Home now owns the house icon.
Gate green: ruff, ruff format, 3x tsc, eslint, prettier, knip, vitest 780,
pytest 989. Verified end to end on localdev against the real account: budget
chips re-answer (matched on REMAINING time — a 51-minute video with 2:46 left is
a ten-minute answer), the reroll walks three distinct heroes, and a Plex show
tile opens its drill-down in the Plex module.
E24 S1. The anti-paralysis answer layer, served from our own Postgres with no
upstream call: one hero suggestion, one CROSS-SOURCE resume rail, and the user's
own comfort shelf.
Survey first, because the roadmap line was wrong about the starting point: Plex
already ships the whole trio (E5's Discover rails, including a reshuffling
Surprise me), the YouTube side ships none of it, and nothing anywhere answers
the question across both. The data already existed on both sides — VideoState
and PlexState are structurally identical, the feed has an `in_progress` filter
and a seeded shuffle, and the YT player already resumes position.
- app/discover/budget.py — the time bands. "How much time do you have"
(10 perc / 30 perc / egy este / mindegy) replaces 15/D's mood questionnaire:
deterministic, answerable from a column we have, and it picks the source for
free. Matched against REMAINING time, so a film with 8 minutes left is a
ten-minute answer. An unknown runtime fits only the unbounded band.
- app/discover/cards.py — a thin unified card that carries the NATIVE feed /
Plex payload, so the page renders with the existing card components instead of
a third one. `aspect` is the one new field: a rail needs to know poster from
16:9, and that is a property of the artwork, not of the module.
- app/discover/queries.py — the SQL, reusing feedquery and plex.rollup rather
than the Plex Discover ROUTE, whose LibraryFilters narrowing belongs to a
surface that has a sidebar. One seed drives both halves so a reroll moves
them together.
- app/discover/hero.py — resume -> comfort -> surprise, with the chosen tier in
the payload so the client labels it honestly instead of guessing.
- comfort_pins (0083): five nullable CASCADE-ing FKs under num_nonnulls(...) = 1
rather than a flag or a polymorphic string ref. A flag cannot order a shelf or
hold a channel; a string ref would let a deleted library item leave a dead tile
on the one surface whose whole job is to be reliable.
Two honest-state rules the tests pin: the surprise draw excludes anything
already started (the resume tier may have skipped it for not fitting, and
offering it back as "Meglepetés" would contradict the question), and which
source leads the draw follows the seed, so the Plex library can actually answer
"egy este".
Gate green: ruff, ruff format, 3x tsc, eslint, prettier, knip, vitest 770,
pytest 987.
The Continue-watching and Surprise rails sorted on `(ts is not None, ts)`. That
key LOOKS null-safe and is not: two rows that both lack a timestamp compare
equal on the flag and then compare `None < None`, which raises — so a show whose
rollup has no `resumed_at`, or a movie state written before
`progress_updated_at` existed, could 500 the rail instead of sorting last.
Replaced with an explicit epoch sentinel, and named it beside the other rail
constants so the sibling rail E24 adds cannot re-derive the same broken key.
The browse scope is a pure DB query with no request state, but it lived in
routes/plex/_shared.py. E24's Home surface needs the same scope from a domain
package, and a domain package importing from routes/ inverts the one-way
dependency CLAUDE.md draws.
Moved to app/plex/query.py and renamed to content_libs (public — it is a domain
helper now). Swept all five call sites and both tests in one go rather than
leaving a re-export shim: the canonical-patterns table already names the
routes.admin re-export as the anti-pattern.
Round 3 reviewed round 2's diff. Trend across the loop: 32 findings over
the whole epic, 15 over round 1's fixes, 11 over round 2's. What it caught
is mostly the same shape as before — a guard narrowed until it was wrong
in the other direction — plus one item this branch had now got wrong
three different ways.
Backspace ownership over the player, third and final answer. Round 1
deferred to nothing (the cheat sheet took the key and the player stopped
underneath it). Round 1's fix deferred to `anyLayerAbove` (the player's
own gear/tracks popovers push layers too, so it stopped closing them).
Round 2's fix named the sheet and palette (the app-wide ErrorDialog is
neither, so the player tore down under it). The rule that actually holds
is not a list of overlays: the player's own panels answer first, because
handleBack's cascade is what closes them, and anything else layered above
takes the key.
The rail's inline rename is settled by measurement rather than argument.
Two rounds reasoned about whether a React passive effect lands before or
after floating-ui's return-focus microtask and reached opposite
conclusions; a jsdom harness built to decide it reproduced no failure for
ANY shape, because focus never enters the menu there — so it was deleted
rather than kept as a test that discriminates nothing. Checked in a real
browser instead: the field opens and holds focus. The shipped fix works,
and the comment now says what was measured rather than what was inferred.
The same check found that round 1's OTHER rail fix never worked at all.
Escape does run its own handler (typing then Escape leaves the name
unchanged, so it cancels rather than blur-commits) and the row button is
focusable when focused from outside — yet focus still lands on <body>.
Three explanations, none of which survived. So the refocus machinery is
reverted to the shape that shipped before this epic and the gap is
documented where the next person will hit it, with the evidence and an
instruction to instrument before theorising. A fourth guess would have
been worse than a known bug.
Also: the two comments in Playlists.tsx that still claimed the disabled
menu rows explain themselves — untrue since round 2 reverted `inert`, and
exactly the comment-vs-code drift CLAUDE.md calls the source of the worst
bugs here. And the move-failure toast printed HttpError.detail verbatim,
which for the reachable 404 is the backend's bare English "Unknown
playlist"; only a structured detail has been through localizeDetail, so
only a structured detail reaches the toast now.
Round 2 reviewed round 1's own fix diff, which is where a fix round does
its damage. It found 15 issues in 792 lines — a much higher density than
round 1 found in the whole epic — and most were regressions round 1
introduced. So this follows the rule rather than fighting it: revert the
fixes whose mechanism was wrong, narrow the ones whose guard was too
broad, and stop reaching for a third variant.
REVERTED, because the mechanism does not work:
- The `inert` menu item. Round 1 swapped a native `disabled` for
`aria-disabled` so a row's "why is this off" tooltip could be reached.
floating-ui's list navigation skips `aria-disabled="true"` exactly as it
skips `disabled` (isListIndexDisabled) — and AnchoredPopover's own
comment, written before this branch, already said so. The swap fixed
nothing and made the mouse worse: without native `disabled` the row took
pointer events, the full active highlight and DOM focus, then no-opped
on click. The underlying finding is logged where the prop was, unfixed
and honest: the reason has to go somewhere a disabled row is not.
- `useContextMenu` owning the focus restore. Turning floating-ui's
returnFocus off platform-wide traded one bug for two: Playlists spreads
triggerProps on a bare <tr>, so restoring to it is a spec no-op and
Escape stranded focus on <body>; and with dismissOnScroll now on, the
restore's plain focus() scrolled the row back into view, undoing the
scroll that closed the menu. Patching that needs three more guards on
one fix. The original bug — the rail's rename field dying on arrival —
is fixed at the field instead: it focuses from an effect, which is
scheduled after the microtask that was blurring it.
NARROWED, because the fix was right and the guard was not:
- PlexPlayer's Backspace deferred to `anyLayerAbove`, but the player's own
gear and tracks menus register layers via useDismiss — so Backspace
stopped closing them and did nothing at all. Only the cheat sheet and
the palette own that key, so it names them.
- `historyDepth = 0` also desynced the counter from overlays the
navigation does NOT unmount (About and ReleaseNotes live outside the
page column), leaving one Back to both navigate and close them. The
entries are marked buried and consumed silently by reconcile instead.
- The move's error toast fired for every status, so a transport fault
surfaced twice on top of req's own dialog. Scoped to 403/404.
- The palette's moved-pointer guard could not judge the FIRST move of a
session and went stale across a close. `e.movementX/Y` asks the browser
directly and needs no state.
- `useKeyScope("contextMenu")` runs once per hook, and the Playlists page
mounts two — so the cheat sheet rendered the chords twice under a
duplicated React key. The scope stack dedupes.
- The `enabled:hover:` sweep missed the half that lives in CSS, so a
disabled glass button still lifted. `.glass-hover:not(:disabled):hover`.
- ShareDialog's migrated buttons passed a colour className that the
variant's own colour class outranks under clsx, so Revoke hovered
neutral and a password-protected link looked unprotected. The colour
goes on the icon, as the Copy button beside them already did.
- `matchesShortcut` took an untyped id and ids are only unique within a
scope, so `back` would have answered for the wrong player. Keyed on
scope + id, both literal.
And the systemic one: round 2 reverted eight of round 1's behaviour fixes
as a single mutation and all 763 tests still passed. Round 1 tested the
three fixes that were easy to test and none of the rest. So the guards
with the widest blast radius are pinned now, each mutation-checked to
fail for its own reason: the global keymap's repeat/allowInInput/disabled
handling (a new dom test), the cheat sheet bound to the real
useContextMenu rather than a scope stub, the scope dedupe, the history
burial, and the Button accessible-name union — whose @ts-expect-error
fixtures turn a widened type into a red typecheck lane.
The chipLane contract test that round 1 rewrote to count per lane counted
the import line as a lane, giving every file one free mirror — so the
exact hole it was rewritten to close was still open. Mutation-checked
both ways this time.
Finishes round 1. Part 1 took the runtime bugs; this takes the prose that
had gone false and the three "the comment says so" gaps behind it.
CLAUDE.md is auto-loaded into every agent session and ui/README.md is
not, so an architecture paragraph that still described a flat components
bag was the most expensive stale sentence in the repo: the next agent
adding a popover reads it, sees no platform layer, and writes the generic
component into components/ — the invisible-layer failure this epic exists
to end. It now names components/ui/ and says the boundary is lint-enforced.
The canonical-patterns table gained the three rows the epic established:
ui/Button over a hand-rolled <button>, a keymap row over a private keydown
listener, and disabled-when-busy vs inert-when-inapplicable.
ui/README.md's inventory was written in this branch's first commit and
never learned about CommandPalette, ContextMenu/useContextMenu,
KeyboardHelp or gridColumns — in a section whose own closing rule is that
a second component painting the same thing is how the last inconsistency
happened. It also claimed "everything is rem-derived" while VirtualGrid's
minCol/GRID_GAP are deliberately px; that is now stated as the exception
it is, and a new section records which promises here are tests rather
than sentences.
Three contracts stopped being prose:
- A text-less Button must carry a name. "Required for iconOnly" was a doc
comment, which is why the migration shipped two nameless icon buttons
through eight green lanes. It is a union now, covering both spellings
(iconOnly, and size="icon"/"icon-sm"), satisfied by label, aria-label or
title. It caught one more on the way in: Banner's dismiss took an
optional dismissTitle, so it could render unnamed — it falls back to
common.close now.
- gridColumns' gutter arithmetic. Every existing assertion survived
replacing floor((w + GAP) / (minCol + GAP)) with the naive
floor(w / (minCol + GAP)): the fixed widths agree, the capped cases are
cap-bound, and the loops compare columnsFor against itself. Pinned at
the exact-fit boundaries where the two first disagree (536 and 828).
- ShareDialog's link row. Only Copy was migrated in S1, leaving one 2rem
button between two 1.75rem ones — and on a coarse pointer .touch-target
grew only the migrated one to 2.75rem while the destructive Revoke
stayed small.
Also: the maxCols docstring said 8 columns of 384px where the test it
shipped with asserts 369.
The epic-end review found 32 verified issues. This is the first fix round:
the three highs, the user-facing mediums, and the tests that were passing
for the wrong reason. Docs and the remaining low-severity tail are next.
The boundary rule was the worst of them. Both `no-restricted-imports`
regexes were written in plain "..." literals, where `\.` is an unknown
escape that collapses to the any-char `.` — so the rule read like a
boundary and matched far less than one. `../../lib/api` (the barrel every
feature component imports) and `../../components/Feed` (the same file
`../Feed` names) both sailed through a green gate. String.raw, a third
pattern for the second spelling, and a vitest over ESLint.lintText so a
rule that silently matches nothing can never look like a clean codebase
again.
closeThenNavigate waited for exactly one reclaim popstate. With a second
Back-closable layer still mounted — a palette opened over the player —
the navigation unmounted that layer too, and its reconcile backed over
the entry the navigation had just pushed: "go to Channels" closed the
player and landed on the feed. It now waits for the LAST programmatic
back, then adopts the entries the navigation buried, and falls back after
250ms rather than hanging if no popstate is ever coming.
The rail's inline rename died on arrival. With no domReference, floating-
ui's returnFocus falls through to a module-global stack and restores some
unrelated element from an earlier popover — blurring the field into its
own commit-and-close, from the second popover of a session onward. The
context menu now owns the restore: back to the row, but only if the row
is still there and nothing else has claimed focus.
Also:
- Backspace under the cheat sheet stopped Plex playback and dropped to
the grid; `back` now defers to any layer above, like `escape` beside it.
- A move invalidated only the source list, so the target kept serving
pre-move contents for 30s while the rail already counted N+1.
- Every move error a user can reach is a 404, which `req` leaves to the
caller — so the row snapped back in silence and the shipped moveErrors
strings were unreachable.
- swallowClick armed on long-press but was disarmed only by the click it
was armed for; any press that produced no click ate the next tap on the
whole list.
- The context menu stayed parked over whatever row scrolled under it.
- overflow-clip on the chip lanes ate the focus ring's top and bottom.
- "Why is this off" rode on disabled rows, which show no tooltip and take
no focus: a new `inert` item keeps the reason readable.
- Shift+F10 was advertised app-wide but wired only on Playlists; it has
its own scope now, and useContextMenu matches through the shared parser
instead of its own literals.
- Mod+K could not close the palette, `h` machine-gunned on auto-repeat,
and both were bound above the branch that mounts their overlays.
- Palette options were in the tab order under aria-activedescendant, and
a scroll-synthesised mousemove stole the keyboard selection.
- The OSD live region was created together with its own text.
- Ending a rename stranded focus on <body>.
- The Button variants' hover halves are `enabled:` — the one migrated
site that had guarded its own lost the guard on the way in.
- HU builtinLocked named the list "Watch later" in a Hungarian sentence.
Tests: the chip-lane contract counted per FILE, so a second hand-rolled
lane in an already-compliant file was invisible; the closeThenNavigate
ordering test only passed on a popstate left behind by its neighbours and
failed alone, so it moved to its own file.
The viewport contract has two ends, and only the floor was enforced. An auto-fill grid
answers "more room" with more columns, never bigger ones, so it settles just above its
minimum column at any width: measured at 3440px, the feed laid out 11 columns of 264px.
A feed diluted across an ultrawide fails the contract as surely as a broken one at the
floor.
Multi-column feed views now carry a `maxCols`; past it the leftover width flows into
the items. At 3440: cards 11x264 -> 8x370, cardsSmall 15 -> 11x264 (still visibly the
dense card, and never as wide as cards gets), tiles 7 -> 6x498. Each cap is set so the
item tops out near 1.5x its own minCol, which is what keeps a view recognisably itself.
A cap can only bind ABOVE the width at which that many columns already fit, so it is a
pure ultrawide affordance — 1024, 1920 and 2560 lay out exactly as before. Verified in
the browser at all three widths and across all five views.
The Channels grid needs no cap: the page's own 96rem --content-mw already bounds it.
Also renames VirtualGrid's `maxCol` to `colMaxWidth`. It was never a column count — it
is the single-column reading-width cap — and `maxCol`/`maxCols` one letter apart is the
kind of drift that has bitten this repo before.
The column maths moves to ui/gridColumns.ts as a pure function so both ends of the
contract are unit-testable without a layout engine; the tests assert the cap's numbers,
that it never binds at the floor, and that the dense card cannot grow into the standard
one.
The four chip lanes render an intentionally unbounded hidden "mirror" so the shared
fit core can read each chip's natural width. An absolutely positioned box still
contributes to its ancestors' SCROLLABLE overflow, and `visibility: hidden` does not
exempt it — and nothing between the toolbar band and <html> clipped it.
At the 1024x768 floor that meant the channel manager's document scroll width was
1258px: the whole app scrolled 249px sideways into dead space and dragged the fixed
header to x=-177. The same mirror was also 228px wide inside the Tags column's 144px
cell, which pushed the table to 990px inside its 906px container — i.e. it broke
DataTable's "the table never exceeds its container" contract and truncated channel
names to "The...". One cause, two symptoms.
The lane now carries a shared `chipLaneClass` = `relative overflow-clip`, so the
contract lives in one constant instead of four hand-rolled class strings. `clip`
rather than `hidden` because it creates no scrollport, so nothing can ever scroll the
mirror into view. Measurement is untouched: offsetLeft/offsetWidth are layout values
and a clip does not change layout.
Measured after: document 1258 -> 1024, sideways scroll 249 -> 0, table 990 -> 921
with no cell overflow, and the "+N" collapse still fits exactly as before.
A source-level test asserts every component rendering a mirror uses the lane class —
the drift a fifth hand-rolled lane would introduce, which no DOM test can catch since
jsdom has no layout engine to overflow.
Reaching a channel three modules away cost a rail, a page and a scroll.
Now it costs a keystroke and half a word: Mod+K searches modules,
channels, playlists, the saved views applicable where you stand, and the
app-wide actions.
Scope agreed with the product owner: navigation + actions, but only the
reversible local ones (the shortcuts sheet, the theme). "Sync from
YouTube" is deliberately absent — it is a long-running job against a
quota, it wants the progress and error surface the Channels page already
gives it, and one mistyped Enter is a worse failure than the walk it
would have saved.
Matching is word-wise and accent-folded, because this is a bilingual app
and nobody reaches for the accent keys in a hurry: "pod thevr" finds
TheVR Podcast and "penzugy" finds pénzügy. The arrows walk one flat
result list so crossing a group heading is invisible, and the field
carries the combobox/listbox aria tree that virtual selection needs.
ui/CommandPalette is domain-free (items are a prop, per the ui/ import
rule); AppCommandPalette supplies them. The shell is Modal again, with a
new `top` placement so a search-first dialog grows downwards instead of
jumping as results change.
Also fixes a bug the palette hit on its first run, and that PlayerModal
had already hit once: closing a dialog reclaims its Back-to-close
history entry with a programmatic history.back(), which undoes any
pushState made before it lands — so a command that navigates only closed
the palette. The rule is now closeThenNavigate() in history.ts with a
test, and PlayerModal's inline copy is gone.
The global `h` moves out of KeyboardHelp into useGlobalKeymap beside
Mod+K, so one listener owns the whole `global` scope and dispatch's
exhaustiveness keeps meaning what it says.
The Plex player had a shortcuts overlay nobody else did, so the YouTube
player's Space/F/arrows and the Plex browser's Backspace were documented
nowhere. ui/KeyboardHelp is that sheet for every surface, opened by a
global `h` - not the conventional `?`, which is a Shift chord on a
Hungarian layout, while `h` is the key this app's reader already has in
their fingers from the player.
It renders the shared table filtered to the scopes that are LIVE: a
surface registers its scope with useKeyScope while mounted, and only
while its keys really do something - PlexBrowse deactivates its own
scope inside the player, where its Backspace no-ops, so the sheet cannot
promise a "Back one level" the handler declines to perform. Same rule
for the `when` rows: ⇧←/⇧→ are listed only inside a series, an answer
the player publishes because the sheet cannot ask.
No new overlay primitive, after checking: Modal already portals into
document.fullscreenElement ?? document.body (so this works over a
fullscreen player), registers an Escape layer, traps focus and handles
browser Back. Third time this epic that the thing was already built.
Shift+F10 is in the table as `handledLocally` - documented and sharing
the chord parser, but matched by useContextMenu on the focused row.
dispatch neither routes it nor demands a handler for it, so the
exhaustiveness check stays honest about what it actually covers.
Also: A/S have cycled the Plex audio and subtitle tracks since E5, but
silently - the only way to see where you landed was to open the tracks
menu, the exact round-trip the shortcut exists to avoid. cycleAudio and
cycleSubtitle now return the track they moved to, and the player flashes
it over the video.
Six surfaces each owned a private keydown listener with its own key
literals and its own idea of "am I typing?", and they had drifted:
PlexPlayer matched `case "a"` without checking modifiers, so Ctrl+A
(select all) cycled the audio track, and its typing guard missed
<select> and contentEditable where the other two surfaces covered both.
lib/keymap.ts is a TABLE, not a dispatcher. Each surface keeps its own
listener and closures - PlexPlayer's refs are not worth re-plumbing -
but the chords, the modifier rules and the typing guard live in one
place, and dispatch() binds a scope's rows to handlers through a mapped
type, so a row without a handler (or a handler without a row) is a tsc
error rather than a comment someone forgets.
Matching is now exact on modifiers, which is what kills the Ctrl+A bug
and lets Shift+arrow be its own row beside the bare arrow. Escape opts
out of the typing guard (allowInInput) so it stays the way out of a
player's own subtitle-offset field.
The player's shortcut sheet renders helpLines() instead of a hand-copied
array beside the handler - the two used to be edited separately, which
is how a sheet ends up documenting a key that no longer exists. Labels
move to the new `keys` i18n namespace (EN + HU), and a test asserts
every row is named in both.
S2's headline was never a new Menu: `ActionMenu` already sits on `AnchoredPopover`
(arrows, typeahead, Escape, anchor-flip). What was missing is an INPUT PATH into it, so
this adds `useContextMenu` + `<ContextMenu>` over the same `ActionMenuItem[]` and the same
`MenuItemList` — a row cannot grow a right-click menu that looks or behaves unlike its own
⋯ menu. Three ways in: right-click anywhere on the row, long-press on touch/pen, and
Shift+F10 / the Menu key. The LEFT button is untouched and keeps the row's own action.
One popover per list, anchored to a virtual element at the pointer (native behaviour, and
a table of 200 rows would otherwise mount 200 floating contexts to keep 199 shut).
Platform bits it needed: `hoverAlso` on AnchoredPopover (a submenu falls open under a
mouse and is tapped by a finger — mouse-only, so a tap's synthetic mouseenter can't race
the click toggle) and `elementRef` on PopoverItem (the submenu anchors to the very
menuitem that opens it, where aria-haspopup/expanded belong). `ActionMenuItem` becomes a
discriminated union — a row either fires or opens a submenu — plus `danger` and
`separatorBefore`.
First consumers, both in Playlists:
- item rows: play · move to another list (a submenu of every other playlist, one gesture)
· to top / up / down / bottom · remove. Reordering goes through the same undoable store
as the drag handle and is disabled under a view for the same reason, with the same
explanation rather than silence.
- rail rows: rename (in the row that was right-clicked, keeping its cover so the rail
doesn't jump) · delete. Both now run through `usePlaylistActions`, which also takes over
the detail header's pencil and trash — the delete dialog's YouTube branch and the
"what gets selected next" rule exist once, not twice.
- the rail gains a "recently changed" sort key off the new `updated_at`. NOT the default:
a rail that reshuffles itself while you look at it costs more than it saves.
The one-step move S2 needs has no honest client-side shape: "add to the target, then
remove from the source" leaves the video in BOTH lists whenever the second call fails,
and the reverse order loses it entirely. So it is one endpoint in one transaction —
POST /api/playlists/{id}/items/{video_id}/move — which re-parents the existing item row
(so `added_at` survives), recomputes `dirty` on both sides, and is idempotent when the
target already holds the video.
Also publishes `updated_at` on the playlist summary, and bumps it explicitly for the
edits that only write `playlist_items` (add / remove / reorder / move). UpdatedAtMixin's
`onupdate` fires on an UPDATE of the playlist row, so without `_touch` the rail's coming
"recently changed" sort would have seen renames and nothing else.
S1c, the last part of S1. The contract now lives in components/ui/ and
in one CSS recipe instead of in ~160 call sites:
- Minimum target 2.75rem, on coarse pointers only (.touch-target). A
32px icon button is fine with a mouse -- precise pointer, hover to
confirm the aim -- and not fine with a thumb, so it grows exactly where
there is no precise pointer, which on this app's real second device is
the tablet. It grows the real box rather than painting an invisible
::after hit area, because two expanded hit areas on icon buttons 4px
apart would overlap and the overlap fires the WRONG button. `link` is
excluded: it is text inside a sentence.
- 2.75rem rather than 44px so the target obeys the text-size slider like
everything else; verified that the type/geometry ladder was already
rem-derived off --font-scale, so nothing else was needed there.
- The ladder is documented as a closed set in Button.tsx, with the site
counts it was derived from, and ui/README.md states the contract.
Measured in the browser: fine pointer 36x36 and 34px-tall buttons
(unchanged), coarse pointer the same buttons at 44x44 and 44px tall.
1024x768 renders identically. 3x tsc, eslint, knip, prettier and 690
vitest green.
Third sweep, 20 more sites: the rest of AdminUsers (the access-request
and invite clusters), Scheduler, SettingsPanel, SetupWizard, About, Feed,
KeyGate, OnboardingWizard -- and, importantly, four platform components
that were still hand-rolling their own buttons: Modal's close, Banner's
dismiss, SidePanel's two toggles and QueryState's retry. A platform layer
that bypasses its own primitive is the same defect one level up.
Hand-rolled buttons with their own geometry: 110 -> 39; Button call sites
37 -> 106. What is left is deliberate, and it is the list Button's own
docstring excludes plus two families this pass refused to force: video
overlay controls painted on white/N, drag handles, chip removers,
accent-solid pills, the bg-card family on the signed-out screen, and
sites whose paint is conditional (`suspended ? warn : neutral`) or
otherwise has no variant yet. Forcing those through a className override
would be unsafe without tailwind-merge, which is exactly the hazard this
sprint documented.
3x tsc, eslint, knip, prettier, 689 vitest green; UAT on Settings and the
About dialog (its close button and full-width CTA are both migrated),
clean console.
Second sweep, 34 more sites: NotificationsPanel (14 -- eight identical
accent text links, three ghost icon buttons, two ghost text actions),
Welcome's five auth links, KeyGate, Messages, NavSidebar, ChatThread,
ShareDialog, ChannelPage, PlexPlaylistView, PlexPlaylistAdd,
PlexCollectionEditor and plexViews.
Moves the font-weight out of Button's BASE and into each variant, so the
`link` variant can be semibold (which is what every text link in this app
looks like) while the action variants stay medium, and no element can end
up with two font-* utilities.
Welcome is the one signed-out surface this session cannot see, so only
its byte-identical link swaps were taken; its primary submit and the
bg-card button family are left for a pass that can UAT the screen.
3x tsc, eslint, knip, prettier, 689 vitest green; UAT on Notifications,
Messages and the nav rail, clean console in a fresh tab.
Counting the un-migrated call sites showed the primitive and the app had
drifted into two different looks: the majority of hand-rolled buttons
hover to accent (outline) and fill with card (ghost), while Button
hovered to fg and filled with surface. That split is the inconsistency
this epic exists to remove, so the primitive follows the majority rather
than the other way round -- 16 existing call sites shift slightly, ~90
hand-rolled ones stop being special.
Adds the variants the call sites prove are missing (outline-accent,
outline-danger, outline-warn, link), an icon-sm size for the p-1.5 icon
family, and a size=none so a link owns no button padding; renames
danger-outline to danger-soft so it cannot be confused with
outline-danger (red at rest vs red on hover). Two new tests lock the
link-size default and the rule that a variant never adds a second
font-weight utility -- without tailwind-merge, two font-* utilities are
decided by stylesheet order, not class order, so that is a real hazard.
Migrates AdminUsers, AuditLog, Channels, TagManager, Playlists,
PlaylistsRail and SettingsPanel (Button call sites 37 -> 52). Left alone
deliberately: conditional-paint sites, video-overlay white buttons, drag
handles, chip removers and card-click targets -- the families Button's
own docstring excludes.
3x tsc, eslint, knip, prettier and 689 vitest green; visual UAT on Users,
Access requests and Playlists at 1280x900, no console errors.
The app HAS a platform layer -- AnchoredPopover (22 importers, with
arrows/typeahead/Escape/anchor-flip), QueryState (21), Modal (17),
PageShell, DataTable, Tabs, Tooltip and ~26 more -- but it sat flat among
~160 feature components with nothing marking it as one. Several of those
files describe themselves as 'the one place' or 'canonical' and were
still bypassed by hand-rolled markup, because a layer nobody can see is a
layer nobody uses; three separate inventory passes of this codebase
misread it the same way.
Moves the 33 domain-free modules into components/ui/, rewrites the
imports in 89 files, and makes the boundary a lint rule rather than a
convention: inside ui/ a one-level-up import (a feature component) and
any lib/api slice are errors. The rule was verified to fire on both and
to leave generic lib/ helpers alone. README.md states which side of the
line a component belongs on and why.
No behaviour change: 3x tsc, eslint, knip, prettier and 687 vitest tests
all green; no backend file touched.
Round five. 47 -> 34 -> 15 -> 7 -> 6 raw findings, and this round's sharpest was the
loop's own blind spot: **the branch did not pass the gate.** `prettier --check` is one of
its eight lanes and I had been running the other seven, so an unformatted `onError` from
round one sat on the branch for four rounds — `siftlode publish` would have refused it,
with no --skip flag to get past.
**The restart hint could still strand the restored data.** Round four said the scripts no
longer print a command they cannot resolve; they did. The COMPOSE_FILE fallback only fires
when `docker-compose.yml` is ABSENT, so in a source checkout it never fires and the hint
expanded to a bare, pasteable `docker compose up -d` — which reads the other file, whose
volumes are `pgdata`/`downloads` rather than `siftlode_pgdata`/`siftlode_downloads`, and
recreates the database on an empty one. When the file is unresolved the scripts now print
`-f <the compose file you started with>`: not runnable, which is the point.
**A comment naming the wrong owner.** `formats.py` still said a zero budget is "a refusal
for enqueue to make" — the premise round one disproved, three files away from the two
docstrings that now say the opposite. Anyone acting on it would delete the worker's refusal
and re-open the original inversion, so it now says where the refusal actually lives.
Also: the reset-flag rule was rewritten in three consecutive rounds and pinned by nothing —
two tests now cover a fragment losing its token, both by a pasted verify link and by
navigating away entirely; `backup.ps1` deletes a partially-copied dump before throwing, the
guard round four gave the docs but not the script it mirrors; and one comment that read as
if the asset were untouched by a line four below it.
Gate green, all eight lanes: ruff, ruff format, 3x tsc, eslint, prettier, knip,
vitest 687, pytest 950.
Round four, on round three's diff. 47 -> 34 -> 15 -> 7 raw findings.
**The printed restart command could strand the restored data.** Round three replaced
`start` with `up -d` because `start` ignores a newly pinned tag — but printed it bare. In
the git-clone layout BOTH compose files exist, so the scripts' COMPOSE_FILE fallback never
fires and a bare `docker compose up -d` reads `docker-compose.yml`: different volume names,
locally built images, no IMAGE_TAG at all. `stop` and `exec` had been fine there (they
resolve running containers by project and service), but `up -d` RECREATES — so the operator
pastes the hint straight after a successful restore and gets `db` on a new empty volume and
a first-run setup wizard, with their data in an orphaned volume. Verified: the two files
resolve `siftlode_pgdata` vs `pgdata` and different images.
The scripts no longer print a command that recreates from a file they did not resolve. They
name the compose file when they resolved one, and otherwise say to start it the way the
operator normally does — with the caveats on their own lines rather than in a parenthetical
at the end of a very long one.
**The reset flag is now decided where the live hash is already parsed.** Round three cleared
it in the token effect, which only runs when a NEW verify token arrives — so the back
button, or re-pasting an already-consumed verify link, still left the banner promising a
recovery whose token was gone. Setting it from the hash on every fragment change is both
smaller and closes those.
Also: the doc said `start` "still carries the old tag" four lines after "pin the older
IMAGE_TAG", where "old" then reads as the tag being rolled back TO — it now reuses the
scripts' unambiguous wording; the Windows recipe removes a partially-copied dump before
throwing, so a mid-stream `cp` failure cannot leave a truncated file under the final name;
and `$WarningPreference` is pinned beside `$ErrorActionPreference` so an inherited
SilentlyContinue cannot drop the half-restored warning, with the comment corrected to say
what the warning stream actually is.
Gate green: vitest 685, pytest 950.
Round three, on round two's diff. The trend across the loop is 47 -> 34 -> 15 raw
findings; the ops half kept producing the sharpest ones because the gate cannot run it.
**`docker compose start` ignores a newly pinned IMAGE_TAG.** Round two removed the
automatic restart because it re-ran `alembic upgrade head` — and then told the operator to
run that exact command by hand, in both scripts and the docs. `start` relaunches the
container that already exists, which still carries the tag being rolled back FROM, so
pinning 0.79.0 and following the printed instruction brings 0.80.0 back up and re-applies
its migrations. Verified against a real daemon: after changing the pinned tag, `start` kept
the old image and only `up -d` picked up the new one. Both scripts and the doc now say
`up -d`, with the pin-first caveat where the operator reads it.
**The reset flag outlived its token.** Reverting round one's "each link clears the other"
went one step too far: a `#verify=` link pasted over a standing reset notice replaces the
whole fragment, so the reset token is gone — but `resetPending` stayed true, and being
checked first it hid the confirmation while promising a recovery that signing out could no
longer complete. The flag is now cleared when the fragment it describes is gone, which is
narrower than what was reverted and does not blank the signed-out card.
**A test that could not fail.** The shared-asset regression test seeded the asset as
`pending`, and `_reset_asset_pending` only acts on `downloading` — so the release half was a
no-op and deleting it kept the suite green. Seeded as the worker actually leaves it now.
Also: the malformed-URL guard answers with urlguard's structured detail rather than a bare
English string, so both halves of one typo class get the same localized copy, and the third
site in that module handing user input to `urlparse` was swept with it; the asset is
released after the job is terminal, not before; the Stats 403 toast gets the wizard handle
its own comment claimed it had; the pasted Windows recipe sets `$ErrorActionPreference`
inside the block (a scriptblock inherits the caller's, so a `SilentlyContinue` profile
swallowed every throw) and cleans up the container-side dump on the failure path too; the
restore warning moved to the warning stream so a `> log` redirect cannot file it away; and
two comments narrowed to what the code does.
Gate green: vitest 685, pytest 950.
Round two of the review loop, on round one's own diff. Two of its findings were
regressions the fixes themselves introduced, which is exactly what the loop is for:
**A per-user refusal was writing onto a shared row.** MediaAssets are deduped across
users, and the generic failure handler puts the message on the asset — right for an
upstream failure, wrong for "your storage is full". Two users queuing the same video share
one asset, so the first user hitting their cap failed the second user's job with the first
one's private reason, telling someone with an empty library to free up space. The refusal
now raises `_OwnerRefused`, caught separately: the job fails, the asset goes back to
`pending` for whoever else holds it. Pinned by a test that drives `_download` directly —
the refusal happens before extraction, so nothing needs stubbing.
**The restore scripts restarted the app, undoing the restore.** `docker compose start`
re-runs the container command, and that runs `alembic upgrade head` — so on the rollback
path this documents as recommended (restore the dump, THEN pin the older tag) the newer
image immediately re-applied the migrations the restore had just undone, and on a FAILED
restore it put the app back on a half-restored schema. Both are the opposite of why the
stop was added. Nothing is restarted automatically now; the scripts say what to run and
when, and the docs' snippets say the same.
**Reverted rather than patched:** the OnboardingWizard toast (the wizard already renders
that failure inline, next to a Retry button — three recoveries for one failure), and
`useVerifyEntry`'s "each new link clears the other". The latter blanked the signed-out
card's only banner, where nothing renders the reset flag to replace it. The two states are
independent again; what changes is that `dismiss` peels ONE layer, which is what the
original finding was actually about.
Also: `Stats` uses the YouTube helper rather than the bare one, so the route's English 403
detail no longer beats the translated fallback under a Hungarian UI; `urlparse` guarded
alongside `.port` (an unclosed bracket raised just as late), the reason renamed `malformed`
since "bad port" misdescribes it; three docstrings narrowed to what the code actually does
(the worker's byte check is per-job, not an aggregate ledger, and not on the edit path;
the redirect residual leaks a reachability oracle, which the note had denied; the
invalidation helper does not cover suspend); the pasted Windows recipe wrapped so `throw`
can stop it; the bash one-liner's failure branch no longer exits 0; and the tests that were
missing for the round-one fixes — the account panel's error paths, the invalidation set,
and the shared-asset case above.
Gate green: vitest 685, pytest 950.
The deep review of the branch raised 47 findings, 18 of which survived adversarial
verification. Grouped by what they actually were:
**The size ceiling inverted at its boundary.** `remaining_bytes` clamped to 0 for a user
already at their storage cap, and `build_ydl_opts` reads 0 as "no limit" — so the one
user over quota was the only one downloading with no ceiling at all. `check_enqueue`
could not catch it: it ran while the job was queued and the user was still under the cap.
The arithmetic moves into `quota.remaining_bytes` (where every other max_bytes decision
lives, and where it can be tested), and the worker now fails the job outright rather than
proceeding unbounded — as a non-retryable `quota_exceeded`, since retrying cannot help
until space is freed.
**Silent failures in the new account panel.** All three actions went through `quiet: true`
endpoints, which suppress the global error dialog on the promise that the caller reports
inline — with no `onError`, that promise was broken and a refusal rendered as nothing. On
the broken-SMTP instance this panel exists for, an admin would read silence as "the mail
went out". The co-invalidation set and the email-off predicate move to `lib/adminUsers`,
shared with the pending-access queue, so the two copies of these mutations cannot drift
apart again — which is how half of this was lost in the first place.
**A rollback recipe that destroyed data without working.** `alembic downgrade -1` steps
back one migration, but a release routinely ships several (0.80.0 shipped four): it drops
the newest one's columns and leaves the schema at a revision the older image also cannot
resolve, i.e. the crash loop it was meant to escape, minus the data. The dump restore is
now the recommended path, and the alembic route asks the target version for its own head
revision instead of counting backwards.
Also: redirects named in the urlguard docstring beside the DNS-rebinding residual they
share a fix with (SEC-17); a malformed port refused as `bad_port` instead of escaping as
a 500; the pull cooldown extended to the sibling route spending the same QuotaAction;
restore stops the app before dropping its tables; the PowerShell cleanup can no longer
decide the script's verdict on 5.1, nor report the wrong exit code; the documented
one-liners hardened to match the scripts; the registry's retention window stated where
pinning is advertised; the two banners made mutually exclusive so neither can shadow the
other; `hover:bg-hover` (a utility this project never generated) replaced by the
`selectable` recipe; the clipboard helper reused; and the sync cooldown's copy swept to
the two callers that discarded it.
Verified: gate green (vitest 678, pytest 949), and the panel, the row hover and the reset
banner UAT'd on localdev.
Design option C, chosen by the user 2026-09-05. Nothing in the admin UI
ever showed an account's GATE STATE: an admin could see a "Pending
approval" chip but not whether the address was confirmed, whether SEC-10
had cleared the password, or which Google account was attached — so
every support conversation started by guessing. Clicking a row now opens
a drawer that says which gate the person is actually stuck behind, with
the recovery actions under that readout rather than in the row.
The drawer is `Modal placement="drawer"`, not a second dialog: the layer
stack, ESC, the focus trap, Back-to-close and the scroll lock are the
hard parts, and duplicating them is how two shells drift apart. It
slides in, and honours prefers-reduced-motion like the rest.
The link itself (SEC-15, SEC-16, and the everyday "no SMTP and someone
forgot theirs") reuses the existing token machinery and the SAME
`#reset=` entry the mailed link uses, so the signed-out card consumes it
with no new flow. It is a distinct token KIND for one reason: a mailed
reset arrives in the mailbox, so completing it proves ownership and
confirms the address; an admin-issued link proves nothing about the
mailbox — the admin hands it over by whatever means they choose — so it
must not satisfy the email gate. Claiming otherwise would be the same
false attestation SEC-15 is about.
Issuing does not revoke the current password: a misclick must not lock
anyone out. USING it rewrites the hash and bumps the session epoch,
which is what evicts a squatter — SEC-16's actual remedy. Spending
either kind burns the siblings of both, so a completed reset cannot
leave a live admin link behind. The non-consuming probe learned the new
kind too, or every admin link would have opened onto "this link is
spent". The audit log records that a link was issued, by whom, for whom
— never the link, which is a live credential-setting secret and the last
thing that belongs in a long-lived log a human reads.
Not gated on account state on purpose: it is useful for a pre-empted
address, a revoked password, a Google-only account that wants one, and a
plain forgotten password. Narrowing it would only make it refuse exactly
when the admin needs it. The shared demo account is the one exception.
Verified on localdev, both languages, end to end: issue from the panel
-> the signed-in tab shows the honest "reset links open in a signed-out
window" banner with the token still in the address bar -> signing out
draws the "Set a new password" form, which is only possible because the
probe recognises the new kind. Gate: pytest 925 -> 942, vitest 676.
DX-17, three one-liners the epic's canonical sweep did not reach:
`routes/setup.py::setup_admin` still resolved the wizard's admin by
literal `User.email ==`. The wizard can be re-run, and E9.5's premise is
that two spellings of one Gmail mailbox are one person — so a second run
under the other spelling would have created exactly the duplicate admin
row the epic exists to prevent.
`routes/admin.py::_decide` hand-rolled `db.get` + 404 where the file's
newer endpoints use `get_or_404`. There were TWO such sites, not the one
the review named; both are converted, per the house rule that a
principle gets fixed everywhere in one sweep.
`useVerifyEntry`'s `hashchange` listener recognised `#verify=` but not
`#reset=`, so pasting a reset link into a tab already running the app
was silently ignored — the same silence that made the verify half a bug.
It now notices, and says out loud that reset links need a signed-out
window; the fragment is still deliberately left in the address bar
(review round 4), so signing out from that tab recovers it.
DX-16: `.catch(() => setOutcome("invalid"))` could not tell a spent
token from a 502 mid-deploy or a dropped connection, and by then the
fragment is cleared — so a network blip was reported as "that link is no
longer valid", pushing the user to request a fresh one on the strength
of a false statement. Only a 400 carrying {code:"verify"} is the
server's verdict now; everything else gets its own outcome, with no
resend CTA, because the token is untouched and retrying it works. The
rule was already written one file away for the reset-link probe: "a
network blip must never be cached as 'this link is dead'".
Two existing tests asserted the old conflation by rejecting with a bare
Error("400"); they now reject with the real HttpError. The errors are
built inside the rejection, not hoisted: mountAt calls vi.resetModules(),
so an HttpError imported earlier is a different class identity and
`instanceof` would quietly miss.
Gate: vitest 671 -> 676, pytest 924 -> 925.
SEC-12: the download box accepts an arbitrary http(s) URL and yt-dlp fetches
it from the WORKER, so any approved user could aim the server at the compose
network, the LAN or a cloud metadata endpoint — and read the result off
job.error, where "connection refused" vs "HTTP Error 401" vs "timed out" is a
port scanner. downloads/urlguard.py refuses a URL whose host resolves to a
loopback/private/link-local/reserved address, at enqueue AND again in the
worker (DNS can move in between, and the worker is where the socket opens).
A network-class failure on a `url` job now reports one fixed sentence instead
of the raw transport text; the error CODE is still classified from the
original, so the localized explanation is unchanged. DNS rebinding is
explicitly NOT closed — that needs the connection gated inside yt-dlp — and
the module says so rather than implying more than it does.
SEC-13: `noplaylist` only demotes a video-inside-a-playlist URL to that video,
so a URL that is purely a playlist or channel still expanded into every
entry — one enqueue, unbounded downloads. `playlist_items` pins it to one.
Size was unbounded too: `check_enqueue` compares the CURRENT footprint against
the cap, and a download's size is unknown until it lands, so one job could
overshoot by any amount; the owner's remaining budget is now handed to yt-dlp
as `max_filesize`, applied AFTER the adapter merge so no source adapter can
raise a limit the quota layer set. A join's cut-list is capped at 100: an
accurate join emits two filters per segment into one ffmpeg filter_complex.
SEC-14: the YouTube budget is instance-wide, so one user holding down a
manual sync spent the day for everyone. Per-user (not per-IP — the cost
belongs to the account) 3-per-10-minutes limiters on the two bulk pulls,
subscriptions and playlists, with separate budgets so one does not close the
other. The review also listed /rss, /backfill and /enrich; E3 deleted them,
and search stays unlimited on purpose — it is interactive, not a bulk pull.
Gate: pytest 896 -> 924 (none skipped), vitest 671.
DX-10: the repo invited people to clone, build and self-host while carrying
no LICENSE at all — which legally means all rights reserved, so none of that
was permitted. Adds the verbatim GNU AGPL-3.0 text, the SPDX id in
package.json and a README section. AGPL rather than MIT deliberately: this is
a server application, so the case worth covering is someone running a
modified copy as a network service.
DX-12: .env.example documented PLEX_MEDIA_HOST_PATH -> /plex-media and the
README advertised the Plex library, but only the localdev and home compose
files carried the mount — on both PUBLIC files the module could read Plex's
catalogue and then fail to open a single file, because it streams the media
itself rather than proxying Plex. Both api and worker now mount it read-only
with an empty named volume as the fallback, and self-hosting.md gains the
section that explains the server URL/token, the path map and the mount
together (it had never mentioned Plex).
DX-15: the worker had no init: true on either public file, though localdev
and home both carry tini with the reason written out — without PID 1 reaping,
a watchdog hard-kill of a download's process group orphans its ffmpeg/yt-dlp
grandchildren.
Verified: 'compose config' resolves on both files, the Plex mount comes out
read_only: true on api and worker, and init lands on worker (not just the
bgutil-pot sidecar). Gate green: vitest 671, pytest 896.
DX-13: both shell scripts ran bare 'docker compose', which has no config
file to find in the layout docs/self-hosting.md tells people to use (only
docker-compose.selfhost.yml is downloaded) — they now fall back to it, while
a source checkout keeps resolving docker-compose.yml as before. The two
PowerShell scripts moved the dump through the PS pipeline, which corrupts
binary on 5.1 and cannot express -Encoding Byte at all on 7 ('Byte' is not a
supported encoding name), so they simply failed there; they now write inside
the container and 'docker compose cp' it across. backup.sh also wrote its
output straight into backups/, leaving a truncated file that looks like a
real backup when pg_dump fails — it writes a .part and promotes on success.
Both restores report a non-zero pg_restore rather than printing 'Restored'
over it. Docs now state that a backup is the dump PLUS .env: without
TOKEN_ENCRYPTION_KEY every encrypted secret in the restored database is
unreadable, which the old one-line 'back up the volume' never said.
DX-14: 'latest' moves with every release and migrations run on every boot,
so self-hosters had no way to choose when they move and no way back. Docs
gain IMAGE_TAG pinning, a rollback recipe (image-only vs schema-bearing) and
the migration-failed-at-boot case; CHANGELOG.md is generated from
releaseNotes.ts so the file cannot drift from what the app shows, and
'siftlode release' refuses on a stale one rather than regenerating it into
the clean tree it just demanded.
Verified: dump + restore round-trip on both platforms against localdev into
a scratch database (8 users, 262822 videos, schema 0082_invite_lang), and
IMAGE_TAG=0.80.0 resolves on both services in 'compose config'.
Release notes for the epic: bilingual mail, one-mail-one-click approval, order-independent gates
with named states, canonical Gmail identity, the admin lifeline for sign-ups stuck outside, and the
four-round closing review that converged 12 -> 4 -> 3 -> 2.
Round 4 came back converged — nothing above low, and both findings in one family. This is the half
worth fixing, because the branch introduced it.
Only the signed-out auth card consumes a `#reset=` token, and S2 moved the address-bar cleanup up to
a hook App mounts unconditionally. So opening a password-reset link in a browser that already holds a
session erased the token from the URL while nothing had read it: the page loads, nothing is said, and
clicking the link again does exactly the same, because the token is untouched and the session is
still there. That is the dead end this epic exists to remove, and `main` did not have it — the strip
used to live inside the card, which a signed-in visitor never renders.
An unspent reset fragment now stays, which restores the recovery that shape always had: signing out
reloads with the fragment intact and the form is waiting. Once the card has spent the token a later
reload meets S4's named dead-link screen, which is true. The affirmative version — a banner telling a
signed-in visitor that reset links open in a signed-out window — is logged with DX-15, along with the
`hashchange` half that still only recognises `#verify=`.
Three low findings, all in what the review rounds themselves wrote.
The comment on `test_approving_an_unverified_account_keeps_its_password` still asserted the argument
rounds 1 and 2 refuted — that dropping the approval-time revoke was safe because `password_login`
reads `email_verified` first — while the docstrings on the two functions it describes now say the
opposite. Two contradictory "why" records for one rule is how the round-1 takeover primitive got
reasoned into existence in the first place, so the test comment now carries the corrected account
and names SEC-16.
`test_a_later_registration_leaves_a_link_already_in_flight_alive` asserted a subset relation that is
vacuously true on an empty set, and the guard that made it non-empty had gone out with the revert.
It now pins both halves: the old link survives, and exactly one new one appears.
Round 2's own fix — activation moved to after a successful send — was unpinned; putting it back
above the send left the suite green. The failed-invitation test now also asserts the account was not
activated by a request the UI reported as a failure.
Round 1 fixed the address pre-emption chain by letting a later registration replace the password on
an unverified account and delete its outstanding verify tokens. Round 2 read that code and showed it
was worse than what it fixed, so it goes back out.
It closes nothing. The rule is symmetric and an attacker can simply register last — on a timer,
inside the five-per-five-minutes budget — while the owner registers once. What it does add is two
primitives an unauthenticated caller did not have: overwrite a real registrant's chosen password,
and delete the confirmation link already sitting in their inbox (the approval mail's 7-day token is
an ordinary verify token, so the blanket delete took that too). Worst of all it fired on the rows
SEC-10 left stripped — password_hash NULL, password_revoked_at set, active — installing an
attacker's password where the entire state means "there is none", and clearing the stamp that is the
only reason password-reset still accepts them. That is a straight takeover which did not exist
before, in the code written to prevent one.
The owner's recourse needed no new primitive and already worked: password-reset accepts a
pre-empted account, proves the mailbox, rewrites the hash and bumps the session epoch. What remains
is the bounded window between a confirmation click and that reset, which is SEC-16 — logged with
this attempt written down, so the next person does not reach for it again.
Two smaller round-2 findings go with it. `upsert_pending_invite`'s new exact-spelling tie-break ran
before any status preference, so it could deterministically pick a DENIED sibling over a pending one
and reopen it — the very outcome its comment claimed to prevent; the status preference now comes
first. And `send_invitation` activated the account before the send it might have to report as
failed, so a request the UI called a failure had quietly let someone in; activation now follows a
mail that actually left. The verify token minted to render that mail still commits either way — it
is single-use, expires on its own and was mailed nowhere, and the note says so rather than leaving
the next reader to wonder whether it was missed.
Eleven findings survived adversarial verification; these are the fixes.
**Address pre-emption.** S1 dropped the approval-time password revoke, arguing `password_login`
refuses an unverified account before it reads `is_active`. Two verifiers showed that holds only
while the address stays unverified — and S6's approval mail exists to get the real owner to end that
state, at which point the squatter's password works. While an address is unverified the NEWEST
registration now owns it: the password is replaced and every outstanding verify token deleted, so
the owner takes the row back simply by registering, and a link already in flight cannot confirm it
for anyone else. The residual (an owner who confirms and neither registers nor resets) is logged as
SEC-16 beside SEC-15, where the re-attestation path it needs already lives. The comments that
asserted the old, false invariant now say what actually guards this.
**The entry URL.** S2 moved `clearEntryUrl()` out of the signed-out auth card into a hook App mounts
unconditionally, and it blanket-stripped the whole query string, the fragment and `history.state` on
the FIRST commit. That broke first-run setup outright — the wizard mounts only after its query
resolves and then finds no `?token=` — and silently ate `?link=`/`?oauth=` (no toast, no refetch
after connecting Google), `?v=`/`?chan=` deep links, and scroll restoration. It now clears only the
fragment and the two one-time flags it owns, and preserves `history.state`.
**The env allow-list.** S3 canonicalised the Invite branch of `is_allowed` and left the
ALLOWED_EMAILS/ADMIN_EMAILS fallback matching literally — the 2026-09-03 lockout with a different
source of truth, plus a shadowing trick where an unauthenticated request under a sibling spelling
permanently masked an env grant. Canonical on both sides now, and the same for the delete dialog's
warning and the ADMIN_EMAILS role bootstrap.
**Honest states.** `send_invitation` stamped "I already wrote to them" before a fail-soft background
send; it sends inline and records only what left. The admin-side resend answered `{ok: true}` for a
dropped mail; it now says so. Both also activate/serialize consistently with the other decide paths.
**Correctness sweep.** The 7-day approval link said "expires in 168 hours" — durations render in
days once there is more than one. `_register_account`, `upsert_pending_invite` and `add_invite`
picked among canonical siblings with no ORDER BY and no exact-spelling tie-break, so which account
got the re-send, and whether a denied row was reopened over a pending one, was up to Postgres.
`purge_user` deleted every canonically-equal invite, revoking a surviving sibling account's access.
`ConfirmProvider` wrapped its ReactNode message in a `<p>`. Approving no longer leaves the
stuck-sign-ups list stale. The refusal panels announce and take focus.
**Tests.** Three mailers rewritten this epic never rendered in any test; every `send_*` now renders
in both languages. The orphan-key guard could not fail (it scanned the catalog that defines the
keys) and had no reverse direction. Newly pinned: canonical activation, the surviving sibling's
grant, a failed invitation and resend, an expired reset link, a throttled login's silence, the
per-language admin fan-out, and that the entry URL keeps what it does not own.
Twelve mails carried hardcoded English while the app itself has enforced HU/EN parity since E5.
A mail is the one user-facing surface nobody can re-render by switching the UI language: it arrives
once, in whatever language we picked, and stays that way. That made the epic's own headline mail —
"one mail, one click" — English for a Hungarian reader.
The text moved to a catalog (`mailtext.py`) with a `text(key, lang, **vars)` renderer, and the parity
rule is now a test rather than a habit: every key must carry both languages AND the same
`{placeholders}`. A translation that drops `{link}` sends a mail with no link in it; one that invents
a placeholder raises inside a background task nobody is watching. Both are silent in production.
Every `send_*` takes the language as its second argument, resolved by the REQUEST handler and never
inside the mail — these all run as background tasks, after the request's session is gone, which is
the same reason the approval mail computes its link in the handler.
Where the language comes from, in order: the account's own preference; failing that, the language the
person asked us in; failing that, English. The middle one is new — `invites.lang`, migration `0082` —
because the mails an access request produces go to an address that usually has no account yet, and
the signal does exist: the page their form was submitted from, or the locale Google reported for a
sign-in we refused. Captured rather than derived, for the third time in this epic (`google_email` in
S3, `preferences.language` in S6 taught the same lesson): a signal that exists only in one moment
cannot be backfilled later. So the admin is asked in exactly one place — the invitation to an address
they typed in themselves, where nobody has ever told us anything.
That send is now a dialog rather than a bare click, which is also where the "you already wrote to
them" warning belongs — it used to live only in a tooltip the admin may never hover, guarding the one
action whose failure mode is mailing a real person twice.
S6's registration `lang` field turned out to be dead on arrival: the backend accepted it, the
frontend never sent it, so every account has been storing nothing. The request builders now stamp the
UI language themselves — a signal that has to be passed in by every caller is one every caller can
forget, which is exactly how it shipped unset.
Admins are notified one mail per language group instead of one blast to everyone. They are people
with their own UI language; a single message would have to pick one of them and be wrong for the rest.
Also removes a duplicated `invited_at` column declaration on `Invite` (two identical definitions
landed in S6).
The funnel's common ordering made a person open TWO messages: one saying their access was approved,
another asking them to confirm their address. The approval mail now branches on what they actually
still have to do — carry the confirmation link, point at sign-in, or say a confirmation mail follows
once they register — so the usual case is one mail and one click. A mail that guesses wrong is worse
than a plain one: telling someone to sign in while their address is unconfirmed contradicts the
screen that will refuse them.
That embedded link lives 7 days rather than the usual 24 hours. Nobody sits waiting for an approval —
it can easily be read on Monday after a weekend, and an expired link spends the one-mail promise and
puts them back on two. The risk is small: the token only ever confirms an address; it signs nobody in
and sets no password.
The most likely root cause of the 2026-09-03 report that a confirmation mail "never arrived" is here
too, and it was never a delivery failure: re-registering an address that already had an unconfirmed
account returned in SILENCE — same cheerful "check your inbox", no mail sent at all. It now re-sends
the link. Nothing new leaks: the response and its timing are unchanged, and the public verify/resend
already does exactly this for the same input.
Whitelisting stays silent, and that is a decision, not an omission. The operator opens a door as a
possibility — usually already agreed with the person — and a row can sit approved for weeks with
nobody registering; a mail fired automatically there would be nudging someone who never asked.
Telling them is its own act on its own button, `invited_at` records it so a second click is a
decision rather than an accident, and the button only appears where nobody is behind the address yet:
"the door is open, come in" is meaningless to someone already inside.
Two things UAT caught that no test would have. Every fresh registration appeared in BOTH admin lists
at once, because the stuck-sign-ups filter keyed on "unconfirmed" — the rule is now "the address may
sign in, and this account still can't", which leaves an undecided request with the queue that owns
it. And the invitation button was offered on every approved row, half of which are existing users.
Also: the confirmation mail finally says what its link's rules are (once, 24 hours) — the reset mail
always has, while this one left the reader to discover both by meeting a dead link. The card's
"didn't get the email?" moves up under the form, and the three waiting messages name the spam folder,
which is where a first message to a fresh recipient sometimes lands.
Registration now records the language the sign-up page was in. Nothing reads it yet — the bilingual
templates are S7 — but it is the only moment that choice exists anywhere the server can see, and a
preference nobody recorded cannot be backfilled later.
UAT on localdev: re-registering re-sent the link (one token became two, one account); approval issued
a 7-day token and moved the account into the stuck list with only a re-send; whitelisting mailed
nobody; the invitation recorded itself and the row then warned that sending again would write twice.
Test data removed.
Approve an address BEFORE the person registers and their sign-up produces no queue row at all —
`_decide` only takes a still-pending invite — so the Access-requests screen sat empty, the tab badge
read zero, and the only working move was re-typing the address into the whitelist box. Nobody would
guess that, and on 2026-09-03 nobody did.
This is roadmap UX-25, where two earlier attempts were reverted after review. Both moved AUTHORITY:
auto-activating on registration turns a standing whitelist entry into a bearer grant for whoever
registers the address first, and reopening the approved invite to `pending` lets an anonymous sign-up
REVOKE a standing approval. What was actually missing was visibility, so this changes nothing about
what an unauthenticated caller can cause — the admin could already do exactly this by re-entering the
address; they just had no way to know they had to.
The section lists every account that cannot sign in yet and says WHICH gate is shut, because the two
dead ends look the same from the admin's chair: waiting for activation, or waiting on a confirmation
mail. Naming the unconfirmed ones is not decoration — on a no-SMTP instance that is the difference
between the owner and whoever registered the address first (SEC-15), and an admin about to let
someone in should see it. A person the queue above already covers is deliberately NOT listed here;
two buttons for one person is worse than one.
Activate is gated on `is_allowed` — the same function the sign-in paths consult, in core, so the
button that appears and the badge that counts cannot disagree. Without that gate it would
manufacture an inconsistent account: password sign-in only checks `is_active` and would let them in,
while their Google sign-in would be turned away and silently reopened as a fresh access request.
Letting in someone with no approved invite stays the whitelist box's job, which says so out loud.
The badge counts activations SEPARATELY from pending invites rather than widening an existing
field's meaning, and only counts what genuinely waits on the admin: an unconfirmed account waits on
its owner, so it is listed with a re-send but does not nag. The admin-side resend says plainly what
happened, including refusing when the instance has no working SMTP — mail is fail-soft everywhere
else, and that silence is exactly what left the 0.78.5 sign-ups with nothing to fall back on.
`send_verification_link` is now one function with three callers (registration, the public resend,
this one), so the link's shape and TTL cannot drift by who triggered it.
UAT on localdev, both languages: the UX-25 shape reproduced end to end (approve the address, then
register), the badge counting it, Activate letting them in and dropping the badge, the row losing
its Activate but keeping the re-send, a fresh token issued by the re-send, and both actions in the
audit log. Test data removed.
Reported from UAT: delete an account, then open its password-reset link. The card drew the full "set
a new password" form — because a `#reset=` fragment was present, and nothing more — and only after
the user had typed a password did the server answer 400. Measured first: no security hole. A user's
tokens cascade away with the row, so nothing can be set on a deleted or a re-created account. What
was left is a dead end of exactly the kind this epic exists to remove, and it is not rare: a reset
link lives one hour, mail is slow or lands in spam, and whoever got two mails clicks the older one.
password-reset/check answers whether a link is still usable WITHOUT spending it, so asking can never
cost the user their one real attempt — the property the tests pin hardest. It gets its own limiter,
because opening a link must not eat the budget for requesting one, and a throttled caller is told
"ok", never "dead": this check may only ever save someone a wasted form, so when it cannot answer it
degrades to the old behaviour instead of declaring a live link finished. The confirm endpoint stays
the authority, and its failure raises the same named screen, since the token can die between the two
calls or the probe itself may have failed.
A too-short password is deliberately NOT that screen. The length check runs before the token is
spent, so a typo costs the link nothing, and the card tells the two 400s apart by `reason` so only
one of them takes the form down.
Both email-link confirmations (verify, password-reset/confirm) now answer in the {code, reason}
shape S4 introduced. They were the last refusals a user meets by simply clicking an old link while
still being answered in hardcoded English under a bilingual UI. The seven remaining bare strings in
this module (register, request-access, the demo guards, set-password) sit behind client-side
validation the UI does first; they are logged as a follow-up rather than swept in silently here.
One bug found by the UAT itself, in the new code: the check went out FOUR times per opened link —
StrictMode double-invokes effects and the shell remounts the card as the session probe settles.
Harmless per request, but against a per-IP limit it would exhaust the budget in two reloads and then
silently fall back to drawing the form unchecked: nothing errors, the feature just stops working.
Deduped in a module-level map holding the in-flight promise (a ref dies with the remount that causes
it); a failed probe drops itself so a blip is never cached as "dead". The test renders the card
twice, because with one render the guard would rot undetected.
UAT on localdev: the reported scenario end to end (delete the account, open the link) shows the named
screen and no form; a live link still opens the form and completes; the CTA lands on the request-a-
link form; one probe per opened link. Throwaway accounts removed.
"Your account is awaiting admin approval." was true and useless. It never said WHICH mailbox the
account was on, and on 2026-09-03 that silence hid a Gmail alias mismatch for a day: a registration
as np.temu@gmail.com met a Google claim of nptemu@gmail.com, and every screen in the funnel agreed
to say nothing about it.
Every refusal password-login can answer with now uses the codebase's structured {code, reason}
detail, built in one helper. The three that arrive AFTER the password is proven carry the address on
the account plus both gate flags; the two that answer before it (invalid, rate_limited) carry
nothing, so a wrong password and an unknown account stay byte-identical. The frontend renders the
first three as a named screen in place of the form — there is nothing to retype once the password is
correct — with an icon, the two-gate checklist, the address in monospace, and a resend CTA where one
helps. Reporting each gate from its own flag is the payoff of S1: approved-but-unverified is a real
state now, and a screen that assumed a sequence would send that user off to wait for an admin who
has already decided.
The Google access-request notice names the address too, and says so when the request was already on
file under ANOTHER spelling of the same mailbox — the honest version of "we didn't open a second
one", stated only when it applies. It moved from the query string to the fragment: the notice now
carries an address, and a query string lands in every reverse-proxy access log. Same reasoning that
put the verify/reset tokens there, and the same owner already clears it. urlencode, not an f-string,
because URLSearchParams reads a bare + in a fragment as a space and would mangle exactly the
+tag addresses this epic exists to protect.
upsert_pending_invite returns the row it settled on beside the is-new flag, since the row's stored
address is what the notice has to name and what mail will go to.
Deviation from the PendingScreen mockup, deliberately: no "Sign out" button and no separate "Not
your address?" link. The mockup drew the panel in the app shell; the accepted decision (a) keeps the
pending user session-less, so there is no session to end, and the address is named directly above a
"Back to sign in" that answers the same question.
UAT on localdev, both languages: all three refusals with the typed spelling deliberately differing
from the stored one (the panel named the stored one), the resend CTA carrying the typed address into
the resend form, and both access-request banners with a +tag address surviving the fragment round
trip. Throwaway accounts and the language switch reverted.
An account can now hold two true addresses: the one its owner typed, which a Google
sign-in no longer overwrites, and the one Google answers with. Only the first was
stored, so Settings could say 'Connected' without saying connected AS WHOM -- exactly
what someone who registered with a +tag needs to know.
users.google_email (migration 0080) records Google's answer at sign-in and at an
in-app link. /api/me exposes it ONLY when it differs from the account address, so the
comparison lives in one place and the majority, whose two addresses are the same
string, see no change at all. Rendered in the account menu as a muted third line under
the identity, and in the Settings sign-in row under the existing hint.
A short 'Google' label rather than a brand glyph: the app draws one icon set (lucide),
and a generic mail icon next to an email address says nothing. It also reads to a
screen reader, which an unlabelled logo would not.
No backfill is possible -- Google's answer is only known at sign-in -- so accounts
linked earlier stay NULL until their owner next signs in with Google.
Two halves of one rule: the address someone typed is theirs to keep, and it must be
the one that gets them back in.
A plus-tagged or dotted Gmail is often a deliberate choice -- registering as
me+siftlode@gmail.com is how you find out later which site leaked your address. The
Google callback overwrote it with the account's primary, which can never contain a
'+', so the tag vanished on the first Google sign-in and, because sign-in matched the
stored address, took their chosen address with it. The rewrite line is older than
this epic, but S3 made it reachable here by matching adoption across spellings. It
now runs only when the mailbox GENUINELY changed, and lives in a named helper rather
than inline, so the rule has one home.
The three lookups that resolve a typed address -- password login, password reset,
verify resend -- go through find_user_by_email: exact spelling first, then the unique
canonical candidate, and None when several accounts share a mailbox rather than
guessing which one a password should open. Reset and resend previously answered a
different spelling with silence (they reply uniformly by design), which was a dead end
with no way out. Their mail now goes to the address on the account, never the one
typed into the form.
Gmail delivers x.y@, xy@ and xy+tag@ to one inbox; the invite gate only lowercased.
On 2026-09-03 that turned an approved user's Google sign-in into a silent SECOND
access request -- OAuth returns the account's stored primary, which need not be the
spelling they typed -- and left prod with two user rows for one mailbox (ids 13, 14).
The +tag case fails every time: a Google account cannot contain +, so OAuth can only
ever answer with the base address.
Mail still goes to the address AS TYPED; only identity matching uses the canonical
form. utils.canonical_email is the single definition, and a model validator derives
users.email_canonical / invites.email_canonical on every write, so no call site can
forget it. Scope is narrow on purpose: dots are significant on most providers and the
Gmail rule does not hold for Workspace custom domains.
Swept every sibling of the same principle, not just the reported one: is_allowed,
upsert_pending_invite, the Google-callback adoption, _register_account (so no new
duplicate pair can be created), _activate_user_by_email, admin _accounts_for (an
invite backing a live user under another spelling used to look deletable) and
add_invite, plus purge_user, which now erases every canonically-equal invite instead
of leaving one behind that still grants access.
Migration 0079 backfills both columns and indexes them NON-uniquely: prod already
carries canonical duplicates (approved invite pairs since 2026-06-16), and merging
accounts is a judgement call about people's data, not something a migration does on
its own. A test pins the backfill SQL against the Python helper.
Found in UAT. Navigating a tab that already runs the app to a URL differing only in
the fragment is a same-document navigation: nothing reloads, so the import-time entry
snapshot never sees the token. It sat unconsumed in the address bar and the previous
outcome stayed on screen -- the same silence this sprint set out to remove, wearing a
different hat. Pasting the emailed link into a running tab does exactly this, and our
verification mails were plain text until 0.78.8, which makes pasting the normal way to
use one.
A hashchange listener now picks up the new token, the guard tracks consumed tokens
rather than a single fired flag (still exactly one POST per token), and a fresh link
clears the standing outcome instead of leaving the old banner up.
The token was only ever consumed inside Welcome.tsx, which App renders ONLY when
there is no session. Opening the emailed link in a browser that already held a
Siftlode session posted nothing, the fragment was then stripped, and the user got
no signal at all -- the link simply did nothing. Incognito worked, which is why it
survived. Confirmation now happens in App, above that branch: signed out the outcome
still drives the auth card's banners, signed in it drives a top-of-app bar.
The entry URL is snapshotted once at import time (lib/entryUrl) instead of per
component, so no future URL writer can strip a token before whoever needs it has
read it -- that ordering race is what broke this in the first place. POST /auth/verify
now names the address it confirmed, which matters when the browser holds a session
for a different account; the same reasoning as password_reset_required already
covers the disclosure.
Also here, from the 2026-09-02 review tail: the dead GET /auth/verify goes (both
senders emit the fragment form, and the last query-token link expired 24h after the
0.78.7 deploy), a spent #verify=/#reset= fragment is cleared while signed in, and
the auth-card banners get role=status. The verifyOkReset copy no longer blames
admin approval, which E9.5 S1 made false.
SEC-10 revoked any password set while an account's email was unverified, on both
the Google-adoption and the admin-approval path. On the approval path that was
redundant: password_login refuses an unverified account before it ever reads
is_active, so approval hands a pre-emptor's password nothing it could use. All the
revoke achieved there was destroying the password a legitimate registrant had just
chosen -- and since the admin normally decides before the user reads their mail,
'your password was cleared for security' was the DEFAULT experience, not an edge
case. The guard stays on the Google-adoption path, where the real owner does attest
ownership.
password_reset_required and the ok_reset banner stay: accounts stripped on prod
before this change have no other way back in.
Two existing tests pinned the old rule, so they are rewritten rather than deleted,
and the SEC-10 recovery pack now builds the revoked state through the surviving
producer.
E9 Playlist module: total playtime in both playlist headers, the header
compacted to two rows, and the sticky column header sitting flush. Plus the
six defects the epic-end deep review confirmed.
Round 2 mutation-checked round 1's own fix and found the one line it left
uncovered: every table row was an integer, so `Math.floor(sec)` in the new
seconds tier could be deleted or swapped for round/ceil with the suite still
green — exactly the "later cleanup removes it as redundant" failure the fix's
comment claims to prevent. Two fractional rows kill all three mutants.
Epic-end deep review, round 1. Six findings survived verification; five of them
were one defect wearing different hats.
formatRuntime floored 1-59s to "0m". Harmless while it only ever formatted a
Plex movie's length, but 12D repurposed it for an aggregate, and a total is the
only input that can land under a minute: a playlist of Shorts reported "0m"
while the rail beside it read "0:45". Fixed in the function, not behind a
`>= 60` guard at each of the seven call sites — the invariant belongs where it
cannot be forgotten. Its contract comment claimed "Plex movie/episode lengths"
and was two epics stale.
The rail rendered the same playlist total as formatDuration (1:23:45) while the
new header rendered formatRuntime (1h 23m) — two shapes of one number, side by
side on one screen. The rail now uses formatRuntime too; formatTotalHours had
already recorded that H:MM:SS is wrong for an aggregate.
The rename input kept min-w-0 without flex-1 in the compacted row, so it was
the only shrinkable child beside a shrink-0 action cluster. The Plex half of
this same epic already had the fix; this restores the symmetry.
plex.browse.items had no plural forms, so a one-item Plex playlist read
"1 items". Pre-existing, but the new header is the call site where count=1 is
ordinary. Both languages get both forms, no key drift.
formatRuntime had no unit tests at all across seven call sites. Added a tier
table plus a sweep asserting no non-zero second ever renders as "0m".
Dropped after 3/3 adversarial refutations: the meta-line segment ordering
(the filtered count is labelled, and count-then-runtime is the convention) and
the null-duration lower bound (the server's own SUM has skipped NULLs since
long before this branch; changing only the client would split them).
Parity with the YouTube playlist header (E9 12D): the Plex playlist detail
now carries a meta line under the title — "N items · Xh Ym", summed from the
items' duration_seconds. No backend change: leaf_card already serializes
duration_seconds for both movies and episodes, and no new i18n key either —
the count reuses plex.browse.items, the same key the collection header uses.
The title/rename input moved into a min-w-0 flex-1 column so the meta line
sits under both; the line is hidden on an empty playlist.
The header portals into the shell's FIXED toolbar band, so its height is
list height the user never gets back — and it was spending four stacked
blocks on it: title, meta line, an action row, then a separate view-control
row under the cover. Option B collapses that into two tight rows beside a
smaller cover: title + rename/YouTube-link on the first, with Play-all,
push/revert/delete and the unsynced badge pushed right; the meta line
(count · playtime · filtered) and the view controls (group-by, reset view,
save order, view status, undo/redo) sharing the second.
Every control and its gating is unchanged — the view-control group still
appears only for `items.length > 1 || viewActive`, so a one-item playlist
shows just the meta line.
Measured on a 32-item playlist: the band goes 170px -> 88px at rest
(122px while a view is active, where the status sentence wraps), which is
~80px more visible list.
The playlist content wrapper carried `pt-3`, so its sticky column header sat
12px below the scroller top at rest and jumped up to `top:0` on the first
scroll. Drop the top padding so the header is flush from rest and never
shifts — the fixed toolbar band above supplies the breathing room, matching
the channel table's already-correct pattern.
Surface the playlist's total playtime beside the item count in the detail
header (12D) — sums every item's duration (formatRuntime "h m", matching the
approved mockup), hidden when nothing sums. Data is already client-side; no
backend change.
Pass fadeTop:false for the playlists page (12B) so the shell's top edge-fade
no longer sits over the already-sticky column header (tableHeader.tsx:257),
matching the channels/audit pages.
The bodies are text/plain only, so whether the URL in a verification or
password-reset mail is a link at all was the client's decision. Gmail, Outlook and
Apple Mail linkify plain text; lighter web clients render the body verbatim and leave
the link as dead text — in a message whose entire purpose is that one link.
_send now attaches an HTML part generated from the SAME plain body, rather than
per-template HTML that would have to be kept in sync with it: one converter, all
eleven templates, and text/plain stays the primary part and the source of truth.
Escaping runs before linkification — bodies interpolate untrusted values (a
requester's address, the operator contact) — so no caller can inject markup, and a
query string's `&` is already `&` when it reaches the href, which is what an HTML
attribute wants. Sentence-final punctuation is trimmed out of the href while
url-safe token characters (`-`, `_`) are left alone. Inline styles only, no images,
no external CSS, no tracking.
Review found the backfill's fourth predicate `AND email_verified = false` strands
exactly the subset that needs it most. The revoke does not invalidate the account's
outstanding verification token, so clicking the emailed link is the natural next
thing such a user does — and it sets email_verified=True while password_hash stays
NULL. That row then has no route back at all: reset ineligible, resend refuses an
already-verified address, set-password needs a session the account cannot obtain,
and nothing re-stamps it because the revoke itself is guarded on being unverified.
The remaining three predicates already pin the victim set exactly: registration and
the setup wizard always set a hash, every Google row carries a google_sub (adoption
assigns it before revoking), and nothing in the codebase ever clears google_sub — so
a local, non-demo, passwordless row can only have come from the revoke. Widening is
therefore safe, and it is the argument the docstring was already making while the
SQL contradicted it.
Also corrects that docstring: set-password is unreachable because the account cannot
authenticate at all, not because the epoch bump killed a session — an unverified
account is refused at login before a session is ever minted, so it never had one.
Two independent bugs made an email+password registration unrecoverable in prod.
The verification link did nothing. urlWithParams/clearParams rebuilt the URL as
pathname+query and dropped the fragment — but the one-time auth tokens ride the
fragment (#verify=/#reset=, SB3), and the providers that mirror nav params
(useUrlMirror in FeedFilters/Playlists/Plex/Channels) sit ABOVE the lazily-loaded
<App> in main.tsx. Their mount effect replaceState'd the token away before the auth
card's chunk had loaded, so the card read an empty hash: the link opened a bare
sign-in page, no request, no banner. Both helpers now carry the fragment over.
The account was then locked out for good. SEC-10 nulls password_hash when an
unverified account is attested (here: admin approval landing before the user could
verify — which the bug above made impossible). That is indistinguishable from a
Google-only account, and reset eligibility keyed off password_hash alone, so no
reset link was ever issued: no password to sign in with, and no way to get one.
users.password_revoked_at records the revocation, reset eligibility honours it, and
setting a password again (reset-confirm or set-password) clears it. An account that
never had a password stays ineligible on purpose — its owner adds one from Settings.
/auth/verify now also answers password_reset_required so the card can say why the
registered password stopped working, instead of leaving the user at the uniform
"Invalid email or password". Telling the holder of a valid verify token leaks
nothing: they have already proved they own the mailbox.
The 0078 migration backfills the stamp for rows carrying the exact lockout signature
(local, non-demo, no password, unverified), so existing victims recover unaided.
Round 4's only finding: the file header and one comment still said Deny stops a
refused person re-asking. It does not — upsert_pending_invite reopens a denied
row as pending — and the delete_invite docstring in the same branch already says
so. The real difference is that Deny records an auditable decision and mails the
requester; Delete forgets the address.
Round 3's correctness/security pass came back clean; what was left said more than
the code does.
- emailOff promised the access-request list is where sign-ups show up. A sign-up
on an address that is already approved or already pending adds nothing to it
(roadmap UX-25), and "only" was falsified by this release's own inbox notice.
- deleteConfirm claimed the admin was dropping "an approval you already gave" —
it is the confirmation for pending and denied rows too.
- sendFailed blamed the mail server for rejecting a message; the reason covers
connect, DNS, TLS and login failures alike, so it now describes the symptom and
leaves the diagnosis to the log.
- ADMIN_EMAILS re-promotion happens in the Google callback only, not on any
sign-in.
Tests: the decided-list copy of the delete affordance was never rendered (both
new cases fail if that branch is dropped), and the throttle release — a
deliberate behaviour the notes advertise — was asserted by nothing; a drop with
no admin yet, or with a failing write, must not silence the following hour.
Round 2 on the feature diff (15 → 12):
- still_allowed_by_env now rides the invite LIST as well as the delete response,
so the confirmation warns while the admin can still change their mind. It
mattered most for a DENIED row on an env-allow-listed address: deleting it
silently converted a refusal into a grant, and the only warning arrived after
the irreversible click.
- Both inbox bodies claimed to name "the most recent" dropped message. The
throttle records the FIRST of each hour and never updates it, so they now say
"one of them", and the not-configured body finally names where to fix it.
- settings.invites.emailOff told the admin they "get no notice when someone
signs up" — rendered directly above the queue that shows exactly that. Only
the email is lost, and it now says so.
- The disabled delete buttons put their reason out of reach: `disabled` drops a
control from the tab order and suppresses its title. Switched to
aria-disabled + preventDefault, the pattern Playlists.tsx already established
for the same reason.
- deleteInvite is quiet, so the 409 the UI handles with a translated toast no
longer also raises the server's English detail as a modal.
- After a completed password reset, Back landed on the reset form again with a
consumed token — the history change made that state reachable. The token is
cleared on success so the reset view stops being the card's root.
- The hu deleteHasAccount named a tab that renders under a different label.
Tests: a dom test for the Access-requests panel covering both has_account
variants and all three delete outcomes, /api/me flag cases, an ADMIN_EMAILS
case, and the env-allowlist test now patches the raw setting so the parsing
stays under test.
Round 1 on the feature diff found 15; the substantive ones:
- The delete control VANISHED on any invite whose address had registered, which
is the ordinary sign-up shape (register writes the User and the pending invite
together). The guard stays broad — that row is the account's approval gate, and
orphaning it strands someone who can then neither sign in nor be approved — but
the control is now a disabled button carrying the reason, on both lists, so an
admin never sees a row silently missing a button its neighbours have.
- Deleting an entry is not always a revoke: is_allowed falls back to the env
ALLOWED_EMAILS / ADMIN_EMAILS lists once no row exists, and an ADMIN_EMAILS
address is force-promoted on its next Google sign-in. The route now reports
still_allowed_by_env and the UI says so instead of claiming access was removed.
- "Deny keeps the row so they can't immediately re-ask" was simply false —
upsert_pending_invite reopens a denied row. Corrected everywhere it was stated.
- The Configuration warning reads me.email_configured, which save() did not
invalidate, so it kept warning about fields the admin had just filled in.
- The read-only variant was an aria-hidden span explained only by a Tooltip that
the hints toggle can switch off; now a disabled button with title + aria-label.
- The delete error path threw away the 409's reason (the stale-list case) and
left the stale row on screen.
- The throttle window was spent even when no notice was written, silencing the
next hour on the path whose job is to stop silent drops.
- email_enabled() opened a second session and decrypted the SMTP password on
every /api/me poll; it now takes the caller's session.
- welcomeMode.dom.test.ts never rendered the card, so reverting the history
wiring to useState left it green. A render test now drives the real links and
fails on that revert (verified).
Plus /api/me flag tests, the registration-shaped pending case, the env-allowlist
case, and a release note that matches what the code does.
Three operator gaps, all found the same way — a person could not sign up and the
app showed nobody anything.
- DELETE /api/admin/invites/{id} with a delete control on both the pending
requests and the previously-decided list. Deny keeps the row so the person
cannot immediately re-ask; delete forgets the address. Refused with 409 while
a User row exists for it, because is_allowed reads that row and removing it
would revoke a live user's access with no other sign — the list ships
has_account so the UI withholds the button instead of offering one that fails.
- Email is fail-soft by design, which had come to mean invisible. _send() now
records an email_undelivered notification for every admin when a message is
skipped or rejected, naming the message that was lost, throttled to one an
hour. /api/me carries email_configured, driving a standing warning on
Configuration → Email / SMTP and above the access-request list.
- The sign-in card's Create account / Forgot password / Didn't get the email
views are history entries (useHistorySubview), so Back returns to the previous
view instead of only moving the address bar.
Also: the per-test rate-limiter reset sweeps every loaded app.* module rather
than a hand-kept list of three — this change added a limiter outside it.
Round 6's two remaining nits. The operator note said "Configuration → Email"
but the group renders as "Email / SMTP" in both locales, and
test_an_admin_can_activate_a_pre_approved_registrant requested make_user without
using it — a leftover from the reverted rounds that made the test look like it
seeded an admin when it did not.
Round 5's correctness/security pass came back clean; what was left was text that
said more than the code does.
- The register-path notify comment (and its test) claimed the admin "is still
told". The call fires, but on a no-SMTP instance _send drops it — which is the
open half of the problem, not something this branch fixed. Both now say so.
- The release notes tell admins what to actually do: an account whose address was
whitelisted in advance has no pending request to approve, so the whitelist box
is what activates it. That instruction only existed in the roadmap and in
chores, which the dialog does not render.
- "That one line is the whole production change" now accounts for the test-only
RateLimiter.reset() that also ships.
- The rate-limiter test claimed to pin that blocked attempts are not counted, on
an assertion that held either way. Replaced with one that drives the clock
across the window boundary and fails if hammering extends it.
Round 4 showed the `inv is not None` gate added in b5a1002 removed a signal that
matters: for an address the admin had already approved, that email is their only
notice that someone has now created an account on it — the invite is not pending
so nothing lands in the Access-requests queue either. Suppressing it left the
registrant at "awaiting admin approval" with the admin unaware they existed. The
duplicate it avoided (Request access, then sign up) is mild noise by comparison.
The production change is now one db.commit(), which is all the reported bug ever
needed. Release notes corrected again: a previously-denied address was also
unaffected, the pending-request case never lost its queue entry (only the
account behind it), and the test inventory was missing a guard.
Round 3 of the review killed the reopen from b5a1002: flipping an approved
invite back to pending let an unauthenticated POST /auth/register REVOKE a
standing approval for any whitelisted-but-unclaimed address, which then fails
is_allowed() and locks the real owner out of Google sign-in — the primary way
people get in here. It also re-armed the pre-emption window it was meant to
close, because with no SMTP email_verified is set True and the SEC-10 guard the
design leaned on never fires.
Two attempts at fixing the stranded-account UX in the same change both came back
worse than the wart. What remains is the narrow, real bug: the account was
silently rolled back and is now committed. An admin still binds the approval to
the account, exactly as before this branch. UX-25 in the roadmap records the
follow-up and both dead ends so they are not retried.
Also from round 3: the upsert docstring no longer claims the admin UI cannot
reach these accounts (re-entering the address in the same panel's whitelist box
does, and a test now pins it), the release notes are scoped to what the diff
actually changes with the operator note moved where the dialog renders it, and
RateLimiter has unit tests so reset() is not production code only a fixture
calls.
Replaces the activation added in ea931dc, which round 2 of the review showed was
a real weakening: on a no-SMTP instance it made a standing whitelist entry a
bearer grant — the first anonymous party to register an approved-but-unclaimed
address got an immediately usable account under that identity, with no
attestation, no admin action and no signal anywhere. The verify-time activation
had the same shape one step removed, and neither site ran the SEC-10 preset-
password guard or left an audit trail.
Keep a human in the loop instead. An approval granted before any account existed
was a decision about the ADDRESS, so registration reopens the invite
(upsert_pending_invite gains revive_approved, used only by the register worker).
The request lands in the Access-requests queue, approving it runs
_activate_user_by_email — and with it revoke_preset_password_if_unverified — and
the account activates. One mechanism, already audited, no new grant path, and
the "awaiting admin approval" copy is honest again.
The explicit commit from the previous commit still matters: with an invite
already pending, upsert returns without committing and _issue_token is gated on
SMTP, so a no-SMTP registration was still being rolled back.
Also from round 2: gate the admin notice on a fresh decision (matching the two
other upsert callers) so a retry no longer re-pings every admin, and reset the
module-level rate limiters between tests — the suite pooled every login into one
300s window and was two logins short of failing an unrelated test with a 429.
Follow-up to the commit fix, from the deep-review. Persisting the account was
only half the story: when the admin had approved the address BEFORE the person
registered, the account landed is_active=False against an invite that is no
longer pending — absent from the Access-requests queue, and approve_invite 409s
on an already-decided invite. The sign-up worked and still could not sign in.
Derive activation from is_allowed() at the moment ownership is attested:
- verify_email_confirm activates a pre-approved account when the emailed link is
confirmed (only the mailbox owner can follow it), so SEC-10 stays intact;
- _register_account activates only when there is no SMTP to verify with, where
the admin whitelist is by design the sole gate — the same one add_invite
already activates on.
Tests: the pre-approved no-SMTP case and the verify-activates case are real
regression tests (both fail without this change); guards cover the unknown-email
and denied->pending branches, verification of an unapproved account, and the
first end-to-end coverage of the /auth/register route and its
allow_registration gate. Release notes rewritten — the previous entry miscounted
its regression tests and advertised an email fix this release has no code for.
_register_account() flushed the new User but relied on someone else to commit:
upsert_pending_invite() commits only when it actually writes an invite row, and
_issue_token() runs only when SMTP is enabled. For an email an admin had ALREADY
approved, on an instance where email is unavailable, neither fired — the session
closed and the account was rolled back. The would-be user got the uniform "ok"
and no account, could not sign in, and re-registering was a silent no-op.
Commit explicitly after the invite upsert (a no-op when upsert already committed
the same transaction). Three DB-lane regression tests cover the pre-approved
no-SMTP path (fails without the fix), the unknown-email path that always worked,
and the pre-approved path with SMTP on.
hls.js startPosition:0 so the transcode/remux HLS session starts VOD-style
at the requested position instead of chasing the live edge of the growing
EVENT playlist.
Our HLS is a growing EVENT playlist with no ENDLIST, which hls.js reads as
LIVE. Its default startPosition (-1) then starts at the live edge and hugs it
(no forward buffer → rebuffer spinners), and on a resume/seek session the
segments' absolute PTS (-copyts, offset by the server's -ss) don't line up
with the live-edge timeline → 0 fragments load → permanent stall. The
seek-restart model makes fragment 0 the requested content position, so start
VOD-style at 0 and buffer forward. Fixes the transcode playback stall/rebuffer
found in 0.78.2 UAT.
Deep-review round 1 on bug/download-mediakind-probe found a real HIGH the prior
review missed:
- resolve_media_facts used a per-field `probe or info` fallback, so a probe that
RAN and deliberately found no video track (cover-art-only audio: attached_pic
skipped; a still image: format-guarded) had its None video overridden by
yt-dlp's stale vcodec. An audio-mode download from a progressive source (info
carries the source's video codec, embed-thumbnail on) was then classified VIDEO
-> a Play button + a broken in-app <video> with no video track. Only the
download path regressed (upload calls probe_media directly).
Fix: gate precedence on whether ffprobe RAN, not per-field. If the probe read
the file, its codec verdict wins outright (a None means 'no such track'); only a
full ffprobe failure falls back to info. Dimensions still gap-fill from info.
Refactored the parsing into _facts_from_probe so resolve_media_facts can see the
raw probe-ran signal; probe_media keeps its 5-tuple (upload + TestProbeClassifier
unchanged).
- tests: the cover-art-audio and still-image resurrection cases, a differing
populated-info precedence case (the prior-flagged gap), and the ffprobe-failed
info fallback — now stubbing run_ffprobe_json so the real parsing + precedence
run together.
Backend gate green (ruff + full pytest).
Deep-review round 2 on the round-1 fixes (all low):
- part_key churn (state): round 1's part-key reset re-probed EVERY sync for a
multi-version item (an original + a Plex 'Optimized' copy), because
_media_facts reads media[0]/part[0] whose order Plex doesn't guarantee stable,
oscillating transcode_hw between the two versions. Gate the part-key signal on
a single-version item (len(Media)<=1); codec-change reset still applies to all.
A multi-version same-codec re-encode not self-healing is an acceptable rare
edge; a per-sync re-probe loop is not.
- dead i18n key: this diff's playableClass() collapses to transcode_hw/sw, so the
bare plex.playable.transcode key it orphaned is now unused (knip can't see
dynamic keys). Dropped from both locales, parity kept.
- test: cross-facet narrowing by an active playable class (a co-facet's year
range collapses to the selected subset) — the flip side of self-exclusion,
previously unpinned.
The aria-label+title double-announcement on the card dot is a deliberate accepted
trade-off (reliable SR name + mouse tooltip). Backend + FE gates green.
Deep-review round 1 on feature/plex-playback-class-filter:
- Sticky watermark (medium): _apply_item reset the bit-depth probe watermark
only on a codec-string change, so a same-codec re-encode (8-bit->10-bit HEVC)
kept a stale video_bit_depth/transcode_hw forever (the backfill only re-probes
media_probed_at IS NULL rows). Also reset when the Plex Part key changes (the
file was replaced), not just the codec.
- i18n parity (medium): the HU playback-class labels were mixed-language (only
'direct' translated). Translated remux/transcode_hw/transcode_sw and the
tooltips consistently to Hungarian (Kozvetlen streameles / Atkodolas).
- Layering (low): _PLAYBACK_CLASSES was module-private yet imported across into
routes.plex.browse; renamed to public PLAYBACK_CLASSES (it is the shared
source of truth for facet order + zero-fill).
- a11y (low): the browse-card playback-class dot conveyed meaning only via a
mouse-hover title; added role=img + aria-label so SR/touch users get the class.
- tests: the NULL/unprobed transcode_hw -> SW fold (filter + facet), facet
self-exclusion under an active playable filter, the migration codec-set ==
HW_DECODABLE_CODECS drift guard, the probe-batch bound + oldest-first order,
the transient probe-failure retry path, and the same-codec re-encode re-probe.
Backend gate green (ruff + full pytest); FE gate green (tsc/lint/knip/prettier/
vitest). NOTE: the facet-vs-runtime disagreement for 10-bit HEVC (C1) is the
known merge-order constraint — this branch's runtime is codec-only; the facet
becomes truthful once BUG1/tonemap's bit-depth-aware routing lands with it. Left
as a sequencing constraint, not code-changed here.
Deep-review round 3: the only surviving finding — a comment-drift nit. The
fallback's cull-yield comment named the route's 'Not available yet' path, but a
culled-key segment request actually returns 'No active playback session' (the
substantive invariant — never a None that 404s as file-not-found — is unchanged).
Reworded to match the route. Concurrency + correctness came back clean.
Deep-review round 2 on the round-1 fix:
- Race (medium): the round-1 identity guard did the supersede check and the
respawn in TWO separate _lock sections. A newer same-key session installed in
the gap was still torn down by _spawn's unconditional pop+kill (the round-1
comment even had the direction backwards). Fold the check into _spawn via a
new replace_only= param: the identity test and the pop+replace now share ONE
lock, so the respawn is a true no-op when superseded/culled.
- Honest states (low): the superseded branch returned current_session(key),
which is None when the key was cap-culled during our out-of-lock wait; the
route turns None into a dishonest 'Media file not found on disk' 404 for a
file that exists. Yield 'current_session(key) or s' so a culled session
degrades to the route's retryable 'not available yet' path instead.
- Hygiene: _want_hw docstring corrected (it is a test-only shim now; production
decides via _select_encode) and the two stale _want_hw references refreshed
(stream.py header, config.py plex_transcode_hw).
- tests: the cap-cull no-respawn/no-None sub-case; the _opencl_present except
OSError branch (glob raising, not just empty); tightened a tautological -vf
assertion. Backend gate green (ruff + full pytest).
Deep-review round 1 on feature/plex-hdr-hw-tonemap:
- F1 (correctness): _select_encode gated tonemap on bit-depth alone, so a
10-bit *SDR* (bt709) source got the HDR hable tone curve applied, visibly
altering an SDR picture where only a 10->8-bit downconvert was needed. Split
the ffprobe gate into a 3-way classifier (_probe_video_kind: hwsafe / hdr /
sdr10); only true-HDR (PQ/HLG) takes the OpenCL tonemap graph, 10-bit SDR
drops to software libx264 (an exact downconvert). Applies to both auto and on.
- F2 (race): the runtime HW->SW fallback popped+killed _sessions[key] with no
identity check. Because wait_for blocks OUTSIDE _lock, a concurrent same-key
_spawn (rapid seek-restart) or _enforce_cap cull could kill this session; the
fallback then respawned software over the NEWER session and breached the
transcode cap. Guard the respawn: only fall back if _sessions[key] is still
this session, else yield the currently-registered one.
- tests: 3-way classifier (incl. HLG + empty-probe), real _opencl_present glob,
real wait_for proc fast-fail branch, tonemap-variant fallback, the F1
regressions (auto + on), the supersede/no-respawn race, and the on/opencl
matrix cell. Backend gate green (ruff + full pytest).
Upgrade mode-3's 10-bit/HDR branch from "route to software libx264" (BUG 1's
behaviour) to a full-GPU pipeline: VAAPI decode → OpenCL tonemap (hable,
HDR10/HLG → bt709 SDR nv12) → h264_vaapi. Gen 9.5 (UHD 630) can't tonemap in
the VAAPI VPP (tonemap_vaapi fails), so OpenCL is the only HW route — proven on
the prod iGPU at ~4.2x realtime, ~1 CPU core.
- stream.py: `_select_encode` returns (hw, tonemap); a 10-bit/HDR HW-decodable
source picks the tonemap graph when the OpenCL runtime is present, else drops
to software. `_opencl_present` gates on a registered ICD. `_venc_opts` /
`_hwaccel_in_opts` gain the tonemap variant (named vaapi+opencl devices,
`-filter_hw_device ocl`, hwmap→tonemap_opencl→hwmap reverse). `_want_hw` kept
as the hw-half wrapper. The runtime HW→SW fallback still backstops a session
that emits no segment.
- Dockerfile: add the Intel NEO OpenCL runtime (intel-opencl-icd). It was removed
from Debian trixie, so it's pulled from bookworm (22.43 still supports Gen 9.5;
the .deb runs on the trixie base — verified live). Harmless without /dev/dri.
- route + FE: thread `tonemap` through the session response; the stats overlay
mode row reads "transcode · GPU · HDR→SDR" for the tonemap path.
- tests: tonemap command branch, the (hw, tonemap) selection matrix (auto/on ×
8-bit/10-bit × opencl), start_session tonemap wiring, and the fmtMode tag.
Adds a movie-scoped "Playback" filter to the Plex feed so a user can isolate items by how
they play — the groundwork for collecting HW-transcodable (or no-transcode) versions.
Four classes derived from the stored `playable` + a new `transcode_hw` facet:
- Direct Play (direct), Direct Stream (remux)
- Transcode · HW — playable=transcode AND the iGPU can hardware-transcode it
- Transcode · SW — playable=transcode, software libx264 only
The HW/SW split is precise, not codec-only: a 10-bit/HDR HEVC is HW-decodable yet breaks the
VAAPI nv12 pipeline (the same case plex/stream special-cases), so it must count as SW.
Backend:
- New PlexItem columns (migration 0077): transcode_hw, video_bit_depth, media_probed_at.
Migration backfills transcode_hw codec-only so the filter works immediately.
- `plex.sync.transcode_hw_capable` is the single classifier; the HW-decodable codec set moves
to sync as `HW_DECODABLE_CODECS` and plex.stream imports it, so the catalog facet and the
runtime encoder choice can't disagree by codec.
- Bit-depth backfill (`_probe_media_batch`, wired into sync): probes only HW-decodable transcode
items' video Stream for bit-depth + HDR and refines a 10-bit/HDR HEVC to software. Bounded
per run, watermarked by media_probed_at. A codec change resets the watermark (stale probe).
- Filter + self-excluded facet on a shared `playback_class_expr` CASE (movie-leg only — a show
is an aggregate with no single class). transcode_hw threaded onto the feed/detail/card JSON.
Frontend:
- Multi-select "Playback" Combobox in the sidebar (movie scope, facet-empty drop-out), active
chips, 4-way corner-dot badge (green/blue/amber/red) + tooltip, and an info-page playback chip
with a codec · bit-depth subtitle. New `bg-info` (blue) token, i18n HU/EN parity.
- Fix a regression where the card's playback dot lost its hover tooltip: the full-card reveal
overlay swallowed the cursor; the dot now sits above it (z-20).
Gate: backend full lane + frontend typecheck/lint/knip/prettier/vitest all green. Browser-UAT'd
on localdev (filter narrows to the SW set, facet counts, badge colour + tooltip, info chip).
A 10-bit/HDR HEVC source (e.g. an HDR10 x265 file) died at scale_vaapi=format=nv12
on the UHD 630 (Gen 9.5 VPP: "VAProfile not supported"); ffmpeg produced 0 segments
and the item never played, with no fallback — a retry just re-picked the failing HW
path. An 8-bit HEVC and the same file on libx264 both work.
Two defenses:
- `_want_hw` auto mode now also gates on an 8-bit-SDR source via a new
`_probe_video_hwsafe` ffprobe (bit depth from pix_fmt/bits_per_raw_sample, HDR from
color_transfer) — a 10-bit/HDR source takes the software libx264 path, which
downconverts to yuv420p cleanly. The Plex mirror stores no bit-depth signal, so this
is probed at session start. "on" still trusts the admin and skips the gate.
- `start_session` adds a runtime HW→SW fallback: a HW session that emits no first
segment respawns on libx264 and reports hw=False (the route/stats overlay honestly
show CPU). `wait_for` gained an optional `proc` so a dead ffmpeg fails fast instead
of burning the 25s budget before the fallback fires.
Tests: TestProbeVideoHwsafe (parsing + probe-failure-assumes-safe), the 10-bit auto
gate + "on" skip in TestWantHw, and the no-segment HW→SW respawn in
TestTranscodeStartSession. Backend lane green.
A downloaded video was silently demoted to a download-only "file" (no
in-app play/edit) whenever yt-dlp's info dict omitted vcodec/acodec — which
it does for Facebook and other sources. Finalize derived media_kind from
those nulls (`derive(None, None)` -> file), so a genuine H.264/AAC MP4
became non-playable in the library.
Fix: probe the finished file on disk (the ground truth) at finalize and
derive from that, falling back to the info dict only for a field the probe
couldn't read. This is the same ffprobe the upload ingest already used, now
extracted to a shared downloads/probe.py so the download worker and the
upload path share one classifier (and can't drift). Also backfills the
stored vcodec/acodec/width/height from the real file.
Tests: TestResolveMediaFacts pins probe-wins-over-info (incl. the Facebook
no-codec case and yt-dlp's literal "none"); upload's probe tests repoint to
the shared module.
The runtime image ships ffmpeg with h264_vaapi/libva2 but no VA driver, so
h264_vaapi can't initialise and a Plex transcode silently falls back to
libx264. Add the iHD driver (intel-media-va-driver) so the UHD 630 iGPU can
hardware-encode once /dev/dri is passed into LXC 118 (host + compose side,
tracked in siftlode-ops). Harmless where no render node is present.
Eight verified findings from the sprint deep-review:
- Transcode cap: _enforce_cap now counts only transcode sessions, so a
live remux (stream-copy) stream can no longer be LRU-evicted mid-play —
matching the config comment + the HU/EN Config hint that promise remux
is uncapped.
- Player Back: the browser/mouse Back interceptor (BackClose) now also
mounts for the stats overlay and up-next card, so Back closes them first
instead of ejecting the whole player (keyboard/Esc already did).
- Honest cold-start: loadSession clears sessionIsTranscode up front, so a
transcode->direct/remux switch no longer flashes a stale "Preparing
stream..." during the session request.
- Stats overlay a11y: the close (X) buttons in the stats, help and info
overlays now read as "Close"/"Bezaras" (new plex.player.close key), not
"Back"; dropped aria-live from the per-second stat grid so a screen
reader is not flooded with telemetry.
Tests: route-level transcode routing (transcode item -> HLS re-encode
session, remux stays copy, direct opens none); start_session threading the
preset/crf/max_height knobs + hw/preexec into the ffmpeg argv; UpNextCard
countdown auto-advance, Play-now, Cancel and the S?-E? heading fallback.
Also brings UpNextCard.tsx / plexStats.ts / plexStats.test.ts to prettier.
Surface the transcode session's runtime facts in the player. The session
response's transcode/hw flags now thread through: the stats overlay's mode row
reads "transcode · GPU" (VAAPI) or "transcode · CPU" (libx264) — the at-a-glance
confirmation of which S1 path engaged — and the initial-load overlay says
"Preparing stream…" instead of "Loading…" for a transcode session, since its
first segment is genuinely slower than a remux and shouldn't read as a hang.
fmtMode (pure, unit-tested) appends the encoder only for a transcode with a known
encoder. i18n HU/EN for the new preparing label. Browser-UAT'd on the XviD
fixture: mode row reads "transcode · CPU" live (software, no render node here).
The software-fallback re-encode now runs deprioritized and admin-configurable.
The SW transcode ffmpeg child is niced (preexec_fn → os.nice, applied per-child
because the Popen lives in the long-lived API process, which must NOT itself be
niced — unlike the download worker's whole-process _deprioritize) so a CPU-bound
re-encode can't starve request handling or playback; the VAAPI path needs none of
it.
New sysconfig knobs (plex group, rendered on the admin Configuration page, i18n
HU/EN): plex_transcode_hw (auto/on/off, layered onto _want_hw), plex_transcode_
preset, plex_transcode_crf, plex_transcode_max_height (software-only downscale-to
cap; the GPU path encodes at source resolution). plex_max_transcodes stays the
concurrency cap (comment corrected — the fallback is shipped, not "later").
pytest extends the command-builder + _want_hw coverage (preset/crf/max_height
plumbing, the on/off/auto policy). Browser-UAT'd: max_height=240 drops the live
resolution to 330x240; the SW ffmpeg child runs at nice 10 (ppid=api process).
Non-h264 items (playable="transcode": HEVC/VP9/AV1/MPEG-2/XviD…) now play by
re-encoding the video to H.264 inside the existing HLS session, instead of the
old HTTP 501. VAAPI hardware encode (h264_vaapi, full GPU decode→scale→encode
pipeline) where the render node is present and the source is HW-decodable;
libx264 software fallback otherwise (yuv420p 8-bit, -threads reserve-one-core).
The re-encode mints a keyframe at every segment boundary (-force_key_frames) so
the HLS muxer still cuts independent segments and seeking works; everything else
— -copyts, the HLS tail, multi-audio renditions, the AAC audio path — is the
same machinery as remux, so hls.js/subtitles/seeking are unchanged. The session
route returns transcode/hw so the player can surface the active mode.
pytest pins both command branches (software + VAAPI shape) and the HW-selection
gate (render node present AND source HW-decodable). Browser-UAT'd on the XviD
fixture (Winning London): 501→plays, real 528×384 H.264 frames, 0 dropped, seek
lands clean; stats overlay Mode row reads "transcode".
At episode end, when a next item exists and the pref is on, show an up-next
card (poster + title + ~10s countdown + Cancel) instead of the old instant jump,
so binge stays in the viewer's control (D6-A). Countdown or Play now advances;
Cancel/✕ dismisses and stays on the finished frame. Lives only on the
play-to-the-end path — credits-auto-skip advances earlier and never reaches the
card, so the two reconcile by construction (no double-advance). New pref
'upNext' (default on) with a Playback-tab toggle; the card closes on item change
and joins the Back/Escape cascade. i18n HU/EN.
Add a toggleable stats overlay to the Plex player: active mode
(direct/remux/transcode), transport, resolution, codec, bitrate, HLS level,
dropped frames, buffer-ahead, session offset and position — sampled once a
second off the live videoRef + hlsRef (no data layer). Non-modal corner panel,
aria-live polite, dismissable; toggled with 'd', a gauge control-bar button, or
its X. Number-crunching lives in a pure, unit-tested plexStats module. Doubles
as the S1 transcode-verification surface. i18n HU/EN.
Add discoverable rewind-10 / forward-10 controls to the Plex player control
bar, flanking play/pause (D5-A). New Skip10 glyph: an outline circular arrow
with a centred '10', deliberately distinct from the episode-level
SkipBack/SkipForward and from Restart's plain RotateCcw. Reuses seekTo; mirrors
the keyboard ←/→ (seekStep=10). i18n HU/EN aria-labels.
Bundles the whole E7 epic (library metadata, crop fix, share-by-default, in-app player,
browser-playable guarantee, local upload) plus the deep-review hardening. Bump VERSION and
add the 0.76.0 release-notes entry (RELEASE_NOTES[0] must match VERSION).
Deep-review remediation on the frontend, all findings adversarially verified:
- LibraryPlayer picks audio vs video from the canonical media_kind (single source), not a
hardcoded container whitelist, so an uploaded audio file with an uncommon/absent container
renders in <audio> instead of a broken video box (legacy null media_kind keeps the heuristic).
- UploadDialog aborts the in-flight upload when the dialog is closed by X/ESC/backdrop/Back
(not only the footer Cancel); drag highlight uses an enter/leave depth counter to stop the
flicker; the progress bar gets a stable role=progressbar aria-label with the value on aria-valuenow.
- DownloadCenter's Shared tab renders honest loading/error states instead of a blank panel, and
the library/shared row-action gating moves to a single tested helper (lib/downloadActions).
Tests: downloadActions gating matrix, the uploadReq XHR path (abort/401/422-localize/onerror),
and the media_kind-driven audio/video branch.
i18n: downloads.upload.uploadProgress (HU + EN).
Deep-review remediation for the E7 upload path, all findings adversarially verified:
- Serve every library file with a Content-Type chosen from the ffprobe-verified media_kind
+ a strict container allow-list (storage.served_media_type), never guessed from the
user-controlled file extension. Closes a stored-XSS vector where a media/HTML polyglot
uploaded as "evil.html" was served inline as text/html on the login-free /watch surface.
Applied to both the authenticated /file and the public watch_file endpoints.
- Classify uploaded still images (jpg/png/gif/webp/…) as media_kind='file' so they are
force-downloaded and don't get a broken in-app video player.
- Roll back the on-disk side of an upload when the ingest transaction fails after place_file
(unlink the placed file + its deterministic poster path), so a failed commit can't orphan
bytes in the GC-skipped .uploads tree.
- Floor download_upload_max_bytes at min=1 so a stored 0 can't silently disable the per-file cap.
- Dedupe the ffprobe-JSON invocation into edit.run_ffprobe_json (shared by probe_codecs and
upload._probe) and the upload extension into storage._clean_upload_ext; drop the now-unused
parse_probe_streams.
Tests: public + authenticated force-download and the polyglot-not-served-as-HTML guard, the
real _probe classifier (image/cover-art/dims/duration), the ingest rollback/cleanup path, the
max_jobs branch, the cap-floor invariant, and run_ffprobe_json error handling.
The /watch/{token} SPA gets a third branch: media_kind==='file' renders an
info card (name / TYPE · size) + a Download button, no player. The <video>/
<audio> branches now guard on media_kind!=='file' so legacy links
(media_kind null) still play. Download is gated on the link's allow_download,
mirroring the media branch; the backend serves a file force-download
(attachment + octet-stream) regardless, so the gate only reflects owner ACL.
FE-only (public_meta already emits media_kind/size_bytes/container since S5a).
i18n HU/EN/DE 'file' fallback label. Two DOM tests: file -> info card +
Download, no player; view-only file -> no Download.
UploadDialog (drag-drop + native picker + optional display name) opened
from a new Upload button in the queue add-row. Uploads stream through a
new XHR helper (uploadReq in api/core) so the dialog shows real 0-100%
progress with a cancel — fetch/req can't report upload progress. The
upload lands as an already-done job; the S5b file-kind row gates it to
download+share. i18n HU/EN; upload cap/empty errors localize via the
shared structured-detail path.
Surface the S5a media_kind discriminator in the library UI. A generic
uploaded file (media_kind='file') is download+share only — no player and
no timeline to trim, so Play/Edit/Edit-details are media-only; it shows a
neutral file glyph instead of an (always-empty) thumbnail and no duration
chip (container/size still read). The "Shared with me" tab's file row is
gated the same way. Re-download is now hidden for edit clips and uploads
too (neither has a re-fetchable source — the backend already rejects both
with 409), removing a latent click-to-409 on edit rows.
FE-only: the serializer already emits media_kind (S5a). Adds media_kind to
the DownloadAsset type and "upload" to the job_kind union; DOM tests cover
the file-row presentation (glyph, no img, no duration).
Add the non-yt-dlp ingest path so a user-uploaded file joins the same download
library and inherits its function set. Any file type is accepted: media
(video/audio) gets the full set (metadata, share, play/edit), a generic file
(pdf/zip/docx/…) is download-only.
- MediaAsset.media_kind ("video"|"audio"|"file") — the stored discriminator that
gates library actions. Set at every finalize (download probe, editor carries the
source's kind, upload probe); consolidates the ad-hoc audio-detect that lived only
in links.py into one invariant (downloads.mediakind.derive). Migration 0076 adds
the column + backfills existing rows.
- downloads/upload.py: ingest_upload probes the file (ffprobe → codecs/dims/duration),
places it in a per-user .uploads tree, and creates an already-ready asset + done job
(the same cache-hit shortcut enqueue uses). Uploads never expire (irreplaceable).
- POST /api/downloads/upload: streaming multipart with the per-file cap
(download_upload_max_bytes, default 500 MiB) enforced on actual bytes read; quota
charges the known size up front (quota.check_upload).
- Security: a generic file is always served force-download (attachment +
application/octet-stream) on both the in-app and public endpoints, so uploaded
HTML/SVG can't execute inline. Edit + redownload reject non-media / upload sources.
- python-multipart added (first multipart endpoint in the app).
Play a completed library item in-app: a Play button on Library rows (and
Shared rows with download access) opens a modal that streams the local file
into a native <video controls> (or <audio> for audio-only assets).
Rather than the YouTube IFrame PlayerModal or the Plex HLS PlexPlayer — both
bound to their own domain APIs and unable to point at a local file — this
reuses the shared Modal a11y shell (focus-trap, ESC, layer-stack, portal) and
the range-aware /api/downloads/{id}/file endpoint. S4a's universal-codec
guarantee (mp4/H.264/AAC + m4a/mp3) makes native playback always work, so seek,
volume, fullscreen and speed come for free with zero transcode.
D2-A: the library row's Share button becomes a single action — quickShareLink()
ensures a public link (minting on demand), copies it to the clipboard, and toasts
with an "Open in new tab" action. A copy failure still toasts (warning) with that
action so the link is never a dead end. Advanced controls (password/expiry/revoke/
user-ACL) stay in the existing ShareDialog, reached via a new "Manage sharing" row
button.
D1 (forward-only): CreateLinkIn.allow_download defaults to true, so a newly minted
link is downloadable by default. The default lives in the API, not a migration —
the download_links column default stays false, so no existing link flips. The
ShareDialog create-default is flipped to match.
Tests: TestCreateLinkDefault (default true; explicit false honored); 4 vitest cases
for quickShareLink (mint-vs-reuse, copy ok/fail, create failure). HU/EN parity.
The Download module's purpose is download -> share -> play in-browser, so the
output must play in EVERY browser (iOS/Safari/WebKit included). The only
universally-safe combo is mp4 + H.264 + AAC/MP3.
- formats.normalize() now CLAMPS every spec to that profile (container->mp4,
vcodec->None, audio_format opus->m4a) at the single point all paths flow
through (picker, stored profile, raw API) — not just by hiding UI options. A
source exposing no H.264 is still caught by the worker's browser-safe
re-encode, so playability is guaranteed end-to-end.
- build_ydl_opts drops the now-unreachable non-compat branches (+ _VCODEC_SORT).
- _SIG_VERSION 3->4 so assets cached under an old non-compat spec — or a stale
non-H.264 mp4 that slipped the v3 net (the HEVC Facebook dls) — re-derive.
- ProfileEditor drops the container/codec pickers and Opus; audio is M4A/MP3.
- i18n: remove the now-unused container/vcodec/any keys (HU+EN parity).
The video preview is object-fit: contain (the browser default for <video>), so a
non-16:9 source is letter/pillarboxed inside the element box. The crop marquee was
drawn in fractions of the whole box (bars included) and those fractions were then
multiplied by the source resolution, so the emitted ffmpeg crop diverged from what
the selector showed. Anchor the overlay to the computed content rect (and set
object-contain explicitly) so a marquee fraction is a fraction of the real picture;
extract the display-px<->source-px math into cropGeometry.ts with unit tests, and
clamp the even-rounded window in-bounds so a full-frame crop can't overflow.
Enlarge every non-title text on the card (channel, status/format/size/date meta,
source-ref line, progress readout) one step (xs→sm) plus their inline icons, and
grow the row action buttons (padding + icon) via the actions wrapper so the shared
IconBtn stays smaller in the modals. Buttons keep the title's line — the wrapper is
shrink-0/no-wrap, so the text column shrinks instead.
The completed download rows already surfaced size + duration; add the concrete
container (MP4/M4A/WebM…) as a compact chip so the library reads as a media list
with a visible format — the audio-vs-video distinction the running-only phase pill
can't show at rest. Pure display of the already-serialized asset.container; locked
by a small DOM test (present when ready, absent before the worker fills it).
- Button: `loading` always disables (|| not ??) — stops a double-submit
when a call-site passes an explicit boolean `disabled` (VideoEditor).
- playlists: track un-ingestable-but-live YouTube ids in the sync baseline
so a linked playlist containing a private/deleted video converges instead
of re-classifying it as a pending pull forever; re-attempt ingestion each
sync so a restored (private->public) video heals, and prune ids that
leave YouTube.
- SettingsPanel: download-auth section renders an honest error/loading
state (StateMessage + retry) instead of a bare empty section on a failed
credentials fetch; HU+EN loadError string.
- tests: reorder-order, partial-failure baseline hold, un-ingestable
converge/heal/prune, Button loading-disables, download-auth error/retry.
The confirm dialog spelled out both sides even when one was "+0 / −0", which read
as engineer shorthand. Build the message from one sentence per side that actually
changes and add a short "stays unchanged" line for the untouched side; the zero
side is no longer shown as a bare +0/−0 to decode. Assembled in the component
from granular i18n clauses (add/remove per side, reorder, rename) with count
plurals; HU+EN. Replaces the single pushPlanMerge template.
"Sync to YouTube" was a one-directional destructive push (local wins, deletes
from YouTube every item not in the local list), so items added on YouTube were
lost and a dirty pull was skipped entirely. Replace it with a real 3-way merge:
each item is classified against the last-synced baseline as a local-add, a
YouTube-add, or a genuine delete, so an add or delete on either side survives.
- Persist the ordered baseline id set (new playlists.synced_video_ids, migration
0075); the fingerprint hash alone couldn't drive a merge. NULL baseline (fresh
export / pre-S7 link) unions losslessly on the first sync and records one.
- Pure merge_plan(baseline, local, youtube) -> add/remove per side + merged order
(local order, YouTube-only additions appended). merge_playlist applies the
local side directly and the YouTube side through the failure breaker; the
baseline only advances on full success so a partial failure re-converges.
- plan_merge/push-plan now report what each side gains and loses; the confirm
dialog shows the two-way plan before applying (pushPlanMerge, HU+EN).
- Tests: 10 new (pure classification incl. delete-vs-stale-copy; DB apply proving
a YouTube-add and a Siftlode-add both survive on both sides, a genuine delete
propagates without collateral loss, no-baseline never deletes, create path).
Each user stores their own login cookies (Netscape cookies.txt) for an
auth-walled download host (Facebook, Instagram, X, TikTok, Reddit) in
Settings; a download of an auth-walled link on that host is retried with
the acting user's cookie.
Two load-bearing security controls, both in downloads.credentials as the
single source of truth so enqueue and the worker can't drift:
- cache privacy: a cookie-authed download salts its MediaAsset format_sig
with the user id, so the authed asset is private and never cache-served
to another user. The worker attaches a cookie ONLY when the asset it is
downloading into IS that private, salted identity.
- host scope (SEC-12): a cookie is resolved only for a URL whose netloc
the credential owns (netloc-scoped _host_is); a look-alike/unrelated
host gets none.
cookies_enc is Fernet-encrypted at rest, only decrypted in-worker into a
0600 file under staging/.auth/ (survives the retry wipe + _produced_file
scan) and unlinked in finally. The blob is never returned by a route,
logged, audited, or written to a job's error/snapshot. cookiefile is set
after the adapter's extra merge so an adapter can't clobber it. Settings
UI is write-only (host + set/updated only, never echoed). New
download_credentials table + migration 0074; per-user routes; HU+EN i18n.
Tests: host-scope, cache-privacy (no cross-user serve), encrypt-at-rest,
write-only API, worker attach-gate; FE dom test for the write-only field.
Sweep the rounded-lg solid/ghost/outline action-button family onto the shared primitive:
ConfirmProvider, ErrorDialog, DownloadDialog, DownloadCenter, downloadModals, downloadAdmin,
ProfileEditor, PersonalViewsModal, SettingsPanel, WatchPage, DraftSaveBar. Deliberately left
on their own styles per the agreed scope: login/onboarding rounded-xl CTAs (Welcome/Onboarding/
Setup/KeyGate), glass Plex surfaces, segmented pills/toggles/switches/chips, dropdown triggers,
data-driven Banner/tone maps, text-xs dense buttons, and tiny inline p-1/p-1.5 icon controls —
none map 1:1 to a variant+size without a visual change, and the repo has no tailwind-merge.
Introduce frontend/src/components/ui/Button.tsx — the one action-button primitive
(variants primary|danger|outline|ghost|glass|danger-outline; sizes xs|sm|md|lg|icon;
icon/iconOnly/label/loading/fullWidth/forwardRef). No focus-visible (already centralized
in index.css @layer base). Migrate the cleanest proof sites off btnCls, and flip the
channel Subscribe button to the outline variant so it matches Full history (the user's ask).
The download/edit child (yt-dlp + ffmpeg) shares one LAN box with the API,
Postgres, and Plex playback. A CPU re-encode pegged every core for ~30s and
starved web requests + playback until it finished. Full-scope S4 mitigations:
- Child self-de-prioritizes at startup (inherited by every ffmpeg/yt-dlp
descendant): CPU nice increment + best-effort idle ionice class + an
OMP_NUM_THREADS budget, all config-gated and independently disable-able.
- Cap the browser-safe re-encode's encoder threads (`-threads`) so a transcode
reserves a core for foreground work instead of claiming all of them.
- Throttle the retention GC's whole-tree orphan sweep (yield IO after each
batch of files) so its rglob can't monopolize disk IO in the API process.
- Idle-back-off the Download Center (2s→15s when nothing queued/running) and
UsageBar (2s→15s when running_jobs is 0) polls; both snap back to 2s on the
next fetch after an enqueue (invalidateDownloads refetches immediately).
Tests: encode_thread_cap (pin/reserve-core/edge), -threads flag presence,
_deprioritize (nice/ionice/OMP branches), GC sweep throttle batching.
5a — the public watch page rendered every asset in a <video>, but WebKit (all
iOS browsers) refuses to play an audio-only mp3 in <video>, so shared audio
links were broken on iPad while working on desktop Chromium. public_meta now
flags audio-only assets (codec present, no video stream) and WatchPage renders
a native <audio> in a padded surface card (no artwork exists for these).
5c — a login-walled Facebook post failed as a bare 'unknown', which offered a
misleading Retry. Classify closed-group / cookie-suggestion auth walls as
login_required and genuinely unsupported sites as a new unsupported_host class
(both non-retryable), and drop unknown from the modal's retryable set — an
unclassified failure is most often an unparsed auth wall. The queue row keeps
its own force-retry for real transients.
The asc/desc buttons in the sort popover footer share one component across
Feed and Plex browse, but rendered differently: the Plex popover is narrower
(short option labels), so "Descending" wrapped icon-over-text and made the
active button taller. Add whitespace-nowrap to the direction buttons and bump
the popover min-width (min-w-52 -> min-w-56) so both call-sites render an
identical single-line toggle.
The incremental watch-sync (run_plex_watch_sync) learned a movie's watched
transition only from the history feed + onDeck. When a movie is finished on
Plex (commonly by scrubbing through the end credits) Plex sets viewCount>0,
clears viewOffset and DROPS it from Continue-Watching, so onDeck no longer
lists it (and onDeck can only ever ADD an in-progress row, never resolve one),
while a scrub-to-end frequently emits no history session row. The stale local
status='new', position>0 row then survived until the DAILY full reconcile —
which is why movies kept showing "in progress".
Add _reconcile_inprogress_movies: for each local in-progress MOVIE that Plex's
onDeck no longer lists, read its current per-account viewCount/viewOffset with
one metadata() call and apply it via the shared _pull_apply, flipping to
watched when Plex says so. Bounded by a probe cap (100), most-recently-touched
first; episodes are left to onDeck/history. No-signal rows are preserved (an
un-watch stays the daily reconcile's job).
Tests: watched-flip, onDeck+episode skip, and no-signal-preserved.
Deep-review of the quickfix batch surfaced two data/preference-loss bugs plus
smaller items; all fixed with tests.
- watch_sync _pull_apply: don't let the Plex-authoritative pull overwrite a DIRTY
local row (a change whose immediate push failed) — that stranded the value
forever since _repush_dirty only retries dirty rows. Leave it dirty for re-push.
- PlexProvider / ChannelsProvider: a partial shared link opened in a fresh tab
persisted mount-time DEFAULTS over the account's stored scope/sort/filters
(base read before the account was pinned). Persist/keep only URL-present dims;
re-hydrate the omitted ones from storage.
- playback resume write-back: guard against out-of-order delivery (a superseded
older push now drops) and evict the throttle map so it can't grow unbounded.
- PlexPlayer: announce mid-playback buffering to assistive tech (role=status +
sr-only label); new plex.player.buffering (HU+EN).
- columnSort: hoist the shared key:dir sort codec used by both URL-state modules.
- Tests: dirty-guard + flag-settle + per-key throttle scoping + superseded-push
(backend); hold-repeat/suppress-click/unmount + partial-link no-clobber +
refetch-doesn't-clobber-undo (frontend). Backend pytest 475, FE vitest 519.
The Channels-manager content state now round-trips through the address bar,
so a manager view is bookmarkable/shareable — the S4 pattern applied to the
last content page the audit flagged.
- New lib/channelsUrlState.ts (channelsParamUpdates / channelsParamsToState /
hasChannelsParams) under NAMESPACED keys ctab/csort/cf — channels-unique
like ?v=/?pl/pscope, so no other provider reads or clears them and the
feed's URL-stripping can't delete them. The mixed scalar/array FilterMap
(sync/priority scalars + tags multi) rides in `cf` as a JSON blob so both
shapes round-trip exactly. Only non-default values emit.
- ChannelsProvider mirrors {view, sort, filters} via useUrlMirror, page-gated
on channels (off-page the keys clear; localStorage still holds the state and
it re-mirrors on return). A shared link wins over stored state: readChannels
UrlSeed captured once, applied before paint through the slice setters (so the
account-pin re-hydrate sees them dirty and won't reload storage over the
shared view), then persisted to the account once pinned so it survives a
later plain reload — mirrors PlexProvider's seed/rehydrate. The name search
stays ephemeral (out of the shared contract).
- NavigationProvider's page-switch nulls ctab/csort/cf alongside the plex keys
so switching modules resets to the manager root.
Tests: channelsUrlState.dom.test.ts (9 round-trip) + ChannelsProvider.dom
seed/mirror (2); the existing tab-history test's beforeEach now resets the
query string too (the mirror writes it). FE vitest 503→514. FE-only.
The Personal-views manager's detail pane becomes editable instead of
read-only:
- Captured filters render as REMOVABLE chips. Removing one prunes the set
and writes it straight back to the stored view (PATCH filters), via the
same live-row builders (useActiveFilters / buildPlexActiveFilters) that
already compute the pruned snapshot — so the modal and the live chip row
can't drift. The prune is wrapped in useUndoable + the shared UndoToolbar,
so undo/redo re-persist the same way. The undoable draft is the single
source of truth for the chip row, so a chip-prune refetch never clobbers
the history.
- A "Save current" button beside Apply stores the module's currently-applied
filters INTO the view (the live→stored inverse of Apply). Disabled when the
view already equals the live filters (it's the active one). On save the
draft re-seeds to the saved snapshot and clears the stale undo history.
- Rename gains a visible commit affordance: a check button beside the name
input that enables once the name changes (Enter and blur still commit).
Provider gains two methods mirroring `rename`: updateFilters(id, filters)
and saveCurrent(id). One shared provider serves feed + plex, so both get it.
PATCH filters was already supported end-to-end — FE-only.
New i18n: views.saveCurrent / saveCurrentHint / saveCurrentUpToDate (HU+EN).
Tests: 5 new in PersonalViewsModal.dom.test.tsx (chip-remove persists, undo
re-persists, save-current fires + disables when up-to-date, rename commit).
Item 6 — the tag add-trigger now LEADS the chip row (a stable left anchor)
instead of trailing it, so on the detail page (bounded=false) adding chips
grows the row to its right and the '+'/its editor popover no longer drift.
Compact icon in the width-bounded manager cell; a labelled '＋ Tag' pill
where there's free width (the detail header).
Item 7 — surface the per-subscription 'hidden' flag on the detail GET
(channel_detail was omitting it) and add an Eye/EyeOff hide/unhide toggle
next to Block on the channel page, mirroring the manager. FE ChannelDetail
gains 'hidden'; new channel.* i18n keys (HU+EN).
Tests: +3 backend (channel_detail surfaces hidden, both states + unsubscribed
default). Gate green: backend pytest 471, FE vitest 498.
Redesign the shared PriorityStepper (channel manager cell + detail header)
from a value-plus-stacked-arrows control into a horizontal [−] value [＋]
segmented stepper. The value now sits centred between the buttons and opens
an inline editor in a FIXED-width slot, so click-to-type no longer "pops"
the control wider (the old absolutely-positioned overlay grew w-6 → w-14).
One `size` prop drives the two surfaces: `sm` draws its own bordered pill
for the standalone table cell; `md` drops the outer border and keeps only
the segment dividers, blending into the detail header's lozenge instead of
double-bordering it. Hold-to-repeat, negatives, and the debounced-flush
persistence hook are unchanged.
The manager priority column widens 3.5rem → 5.5rem to fit the horizontal
control under fixed table layout.
Tests: 7 new DOM tests for the stepper interaction contract.
Mirror the Plex-browse content state into the address bar so a browse view is
bookmarkable/shareable, the Plex analogue of the feed's urlState. Keys are
NAMESPACED (pscope/psort/pshow/pf) rather than reusing the feed's bare
scope/sort/show: FeedFiltersProvider actively strips its own KEYS from the URL
whenever off the feed page, so a shared key would be deleted out from under the
Plex page. The namespaced keys are Plex-unique (like ?v=/?pl), so no other
provider ever reads or clears them.
- New lib/plexUrlState.ts: plexParamUpdates / plexParamsToState / hasPlexParams.
The pf filter blob reuses the SAME appendPlexFilters wire serializer (new
public plexFiltersToParams + inverse parsePlexFilters, kept next to each other
so a new wire key can't be added on one side only).
- PlexProvider: URL-seed wins over localStorage at mount; the account-pin
rehydrate persists the seed instead of clobbering it (mirrors feed's
initialHadUrlFilters); useUrlMirror mirrors/clears the 4 keys, gated on
page===plex. q/field stay out of the shareable contract.
- NavigationProvider: a page switch now also nulls the 4 plex keys.
Tests: 8 new round-trip cases in plexUrlState.dom.test.ts (FE vitest 483→491).
Add a 'lastwatched' sort key to the unified Plex browse. Movies rank by
coalesce(progress_updated_at, watched_at); shows by max(coalesce(...)) across
their episodes (rollup.resumed_at, now broadened from progress-only so watched
episodes count too). Never-watched titles sink via nullslast.
Frontend auto-applies lastwatched:desc when entering the in-progress watch-state
view and restores the pre-entry sort on exit; a manual sort change while there
still wins. i18n HU+EN. Backend tests cover both union legs + mixed nullslast.
The incremental/reconcile job docstrings still described last-write-wins and a
30-min cadence; update them to match the S2 behavior (Plex wins on divergence,
15-min incremental) so the "why" doesn't drift from the code.
Make Plex the source of truth for watch state:
- Write-back during playback: item_progress now pushes the resume position to
Plex on the periodic checkpoint too (not only the final flush), throttled to
one push per item every ~10s (in-memory, single-process rate limiter). The FE
now also fires an immediate (debounced) push on seek.
- Faster incremental pull: 30 → 15 min (plex_watch_sync_interval_min).
- Plex-authoritative pull: _pull_apply resolves a genuine divergence to Plex's
state instead of last-write-wins; the _same_state + synced_to_plex short-circuit
still prevents a freshly-pushed row from bouncing. Removed the now-unused
_LWW_TOLERANCE_S / _sift_ts.
Tests: 5 new in test_plex_watch_sync.py (Plex-wins resume/watched, no-bounce,
throttle final vs periodic).
The Plex player stalled ~2s a few seconds into every auto-advanced next
episode. Root cause: each auto-advance spins up a cold ffmpeg HLS remux
session, and for a fresh start (start_s==0) start_session returned
immediately with zero buffer — hls.js drained the first segment, then
rebuffered waiting for the still-warming ffmpeg to write the next one.
Once ffmpeg gets ahead (stream-copy is faster than realtime) it never
stalls again.
- stream.py: before reporting a session ready, wait for the first
_WARMUP_SEGMENTS (3) video segments to exist — the seg0 wait is now
unconditional (was start_s>0 only) and warm-up runs for all starts.
Best-effort and bounded: stop at the first segment not ready in time.
- PlexPlayer.tsx: a loading/buffering spinner. The existing overlay only
covered the INITIAL load (!ready); add a `buffering` state driven by the
video's waiting/stalled/seeking (set) and playing/canplay/seeked/pause
(clear) events so a mid-playback stall or in-flight seek reads as
"loading" instead of a frozen frame, and surface a fatal loadError even
once playback has started. Spinner uses the shared Loader2.
Merges roadmap E6·item1 (spinner while loading/seeking) into this fix,
since the server-side warm-up deliberately holds ready a beat longer.
Tests: 3 new in test_plex_stream.py (warm-up waits for the first few
segments; stops at the first missing one; no warm-up when seg0 never
arrives). Gate: backend pytest 460 (Docker lane), FE tsc/lint/knip/
prettier + vitest 483.
Fixes from several multi-agent deep-review passes over the E8 feed-filtering &
channel-manager epic (S1-S4):
- feed_years buckets by UTC year to match the UTC-pinned year-filter bounds
- tag suggestions: examples ∩ subscribed (drop tagged-then-unsubscribed channels
from the profile denominator); exclude hidden channels as suggestion targets
- channel manager: restore ghost chips + surface an error toast on a failed tag accept
- year filter: debounce the search term so the per-year facet doesn't refetch per keystroke
- channel tags: aria-label the "+N" overflow pills
- extract useTagSuggestions hook; add backend + hook tests (profile boundaries, cap,
language exclusion, unsubscribed/hidden handling, UTC bucketing, bulk-accept)
Bump VERSION to 0.73.0 + release notes.
With 2+ suggestions the narrow tags cell clipped the second ghost chip
mid-word right next to the edit "+", reading as the icon covering it. Route
the ghost chips through the same fit machinery the attached chips use: a
dedicated SuggestedChipsLane fits what the cell holds and collapses the rest
into a "+N" pill whose hover popover still one-tap accepts / dismisses each.
The lane is attached-XOR-ghost (suggestions are untagged-only, and accepting
one now drops the whole channel from the suggestion cache), so each lane owns
the full cell width instead of clipping the other.
Learn, per user tag, which auto TOPIC tags are characteristic of the channels
you've tagged it, then suggest that tag for still-untagged subscribed channels
whose topics cover enough of the profile. Language auto-tags are excluded from
the signal — near-universal within a language, they flooded every channel.
New GET /api/tags/suggestions (read-only, deterministic) backed by
app/sync/tagsuggest.py. Surfaced in the Channel Manager as dashed "ghost" chips
in the tags cell — one tap accepts (reuses the attach-tag endpoint), × dismisses
for the session — plus a header badge and a bulk "Accept N confident" (the
full-profile matches) behind a confirm. No silent writes (D4 = suggest-and-confirm).
Also reflow a pre-existing prettier drift in dateFilter.ts from E8·S2.
Add a Year filter to the feed rail: quick chips for recent years plus a
'More…' popover listing all years with per-year video counts, sortable by
year or count. Selecting a year narrows the feed to that whole calendar
year (a published_after/before range applied in the api layer) and is
mutually exclusive with the Upload-date control.
Backend: new /api/feed/years endpoint + feed_years() — one grouped query
that honours every filter EXCEPT the publish-date narrowing (dropped), so
the year list stays a live drill-down and switching years never zeroes the
others. FE keys the list on the date-stripped filters, so picking a year
re-queries the feed but not the list.
Wires year through FeedFilters, the filterParams/URL (?year) round-trip,
the active-filter chip row, and the panel layout. i18n HU+EN.
Tests: feed_years count query (date narrowing ignored, other filters
honoured); dateFilter year-clearing; urlState year round-trip.
The channel manager's tag-filter row wrapped the trailing "…" catalog-actions
menu onto a second line once the chips + Clear outgrew the width. Rework it as
one non-wrapping row: the TAGS editor trigger (left), a single-line chip lane
that collapses overflow into a "+N" pill whose popover holds the rest (still
toggleable), then Clear and the "…" menu pinned right — so "…" always stays on
the first line.
New TagFilterChips reuses the shared measure core (useChipOverflowFit +
fitChipsWithPill), the same compression the feed's applied-filter row uses.
Filter to channels you haven't personally tagged in the channel manager
(no user-owned ChannelTag; auto language/topic tags don't count):
- Feed: a "No tag" sentinel in the "Your tags" sidebar combobox (exclusive
with real tag picks) + a removable "Untagged" chip; new `untagged` bool
feed param / `notag` URL key, persisted + coerced on every filter path.
- Channel manager: a virtual "No tag" chip in the tag-chip row filters the
table to untagged channels (client-side, shared NO_TAG_SENTINEL).
Backend `untagged` is a plain channel-level narrowing, so feed_summary keeps
it applied in the base pass — the count/facet lockstep holds with no mirror.
Note: item E (per-tag counts) and item D (Any/All toggle) were already
shipped; S1 delivers only item B.
UX-8 layerStack: add an integration test wired through the real useLayer
hook (popover document-listener + stamp) and a PlayerModal-shaped window
listener, driven by a real dispatched keydown. The popover close is wrapped
in flushSync so its layer is popped synchronously between the two listeners
-- the production mid-dispatch race -- making the escapeHandled stamp the
only thing that keeps the player from also closing. Mutation-checked:
removing the guard fails exactly this case.
SavedViews: cover feedCanSeedDefault's "yield to a bookmarked filter URL"
branch (a filter-bearing search string -> no seed) and the per-source
seededRef guard against the 60s refetch (a changed-list refetch re-runs the
seed effect; the guard blocks re-seed). The refetch test waits for the
changed list to reach the component so the effect has its chance to wrongly
re-fire. Both mutation-checked.
Frontend gate green: 464 vitest (+4). Backend untouched.
The S4 Play-all wrapper had no test — a regression inverting the empty-queue
guard or dropping the notice would re-introduce the silent no-op it was added
to fix, and stay green. Adds a PlexCollectionView DOM test (spied api +
notify) asserting: a leaf queues as itself, a show expands to its episodes in
season->episode order, and an empty expansion surfaces plex.collection.playEmpty
instead of no-op'ing.
Extract the show->ordered-episode-keys flatten into one pure, unit-tested
helper (lib/plexQueue.ts) and use it at all three sites — the collection
Play-all (S4) and the show/season play-through pages in plexViews — so the
season-then-episode queue-order invariant lives in one place instead of three
hand-rolled copies. Behaviour-identical (the helper's ?? [] is a safe superset
of the collection site's previously non-null read).
Also: a show-collection whose shows expand to zero episodes had an enabled
Play-all that silently no-op'd — the same 'does nothing' state S4 removed for
the empty-queue case. Surface an honest info notice (plex.collection.playEmpty,
HU+EN) instead.
The S2 seed effect ran on the saved-views query's first resolve, but that
query is ungated and — since SavedViewsProvider mounts above <App>, which pins
the tab-account only after /api/me — resolves before the account is known on a
fresh tab. It then evaluated canSeedDefault() against a null account (which
reports 'stored, do not seed') and burned its one-shot guard, so the starred
default was never applied until a reload: exactly the fresh-tab miss S2 set out
to fix. Gate the effect on the reactive account id so it defers instead of
consuming the guard, and re-fires when the account pins. Covers feed + plex.
Adds a dom test for the defer-then-seed-on-pin path the old test masked by
pre-pinning the account.
A collection's Play-all/Shuffle queue held only playable leaves
(movies/episodes), filtering out shows — so on a show-collection the queue
was empty and Play-all was hidden/dead. Build the queue by expanding each
show member into its episodes (season then episode order), concatenated in
member order; leaves still queue as themselves. Shows are fetched lazily on
click (parallel plexShow fetches), not on page load. Gate Play-all/Shuffle
and the poster on "has any playable member" (leaves OR shows), and disable
them while the queue builds. Scope: option A — no member manual-order
migration. Adds plex.collection.playFailed (HU/EN).
ActiveFilterChips (feed applied-filter row) and ChannelTagsControl both
hand-rolled the same single-line chip-fit measurer: a hidden mirror + a
layout-effect + ResizeObserver reading chip right edges, the real rem gap,
the pill width and the lane width. The policy half was already extracted
(decideChipLayout); the DOM-measurement half was still duplicated.
Promote that half to useChipOverflowFit(rowRef, chipsRef, pillRef, deps,
decide, initial): it owns the effect + observer + initial measure and reads
the common geometry, then hands it to each caller's own policy. The feed
row layers its leading label + trailing "Clear all" reservation on top and
keeps decideChipLayout; the tag cell's chips-only policy is extracted as
fitChipsWithPill (pure, unit-tested) alongside it. Add a jsdom test for the
hook (stubbed offsets, re-measure on deps). Behaviour unchanged — verified
in the browser: feed all/count modes + popover, tag "+N" fit + hover reveal.
The feed applied its starred default via a localStorage mirror
(LS.defaultViewFilters) written by SavedViewsProvider and read by
FeedFiltersProvider on load. That raced the login effect: on a genuinely
fresh browser (no LS.filters and no mirror yet) the feed could seed
DEFAULT_FILTERS and never re-apply the default until a reload. Plex had no
default-application at all.
Replace the mirror with a single path: SavedViewsProvider applies the
starred default via cfg.setFilters on the views query's FIRST resolve, but
only into a genuinely fresh module (no stored filters for the account, and
— feed only — no URL filter override). Applying persists it, so it happens
once and the last-view-wins path takes over, matching the documented
storage semantics. Because it runs off cfg.setFilters it delivers the Plex
port for free. Extract pickDefaultToSeed (pure, unit-tested) and cover the
first-resolve/stored/no-default cases for both modules in a provider test.
Removes the now-dead LS.defaultViewFilters key and its mirror read.
UX-8: one Escape must close only the topmost layer, never both an open
popover AND the player beneath it. The popover's Escape listener is on
`document` and fires before the player's on `window`; closing the popover
flushes its layer off the shared stack synchronously, so the player's
`anyLayerAbove()` lookup read "nothing above" and closed too. Add a
consumed-event flag to layerStack (`markEscapeHandled`/`escapeHandled`):
the acting topmost layer stamps the shared KeyboardEvent, and both players
check the stamp before the (now-mutated) stack. Immune to listener order
and mid-dispatch stack mutation. Cover with layerStack.dom.test.ts.
UX-7: the E4 `.reveal` sweep already covered the card action rows; two
residual hover-only traps remained — invisible-but-tappable/Tab-focusable
on touch. Add a group-agnostic `.reveal-hover` recipe (display:none under
`(hover: none)`) paired with each site's own hover/focus-within reveal:
- VideoCard play/Continue/Restart overlay: also reveal on
group-focus-within (keyboard), and drop it on touch — the card tap
already plays/resumes, so no permanent scrim over every mobile thumbnail.
- Carousel prev/next arrows: gone on touch, where the strip swipes.
The notifyCollectionEditError helper localizes the backend's duplicate-name
409 to plex.collection.nameTaken; it was applied to the New-collection panel
but two sibling collection-write sites still routed 409 through
notifyActionError, surfacing the server's raw English detail under the HU UI:
- PlexCollectionView.doRename (rename collision)
- PlexCollectionEditor.createNew (create-in-editor duplicate name)
Both now use notifyCollectionEditError, matching the create path. Also adds a
direct dom test for the extracted PlexLibraryPicker covering the enabled gate,
emptyHint, results, honest error, and no-results branches.
- CLAUDE.md: new "Collaboration" section (plan-first, options-with-marked-
optimal-and-why, mockups for design decisions, challenge unsound requests,
visual UAT before done, terse otherwise).
- CLAUDE.md: clarify release is user-owned in timing, agent-run in execution.
- CLAUDE.md + docs/development.md: document the ruff format --check gate step
(already enforced in .gitea/workflows/check.yml, was missing from the docs).
- CLAUDE.md: refresh the stale footer to point at the now-existing
docs/development.md and backend/app/README.md.
Backend: create-collection happy path (write-through + editable) and the
create resync-failure 422 (honest reconcile detail), mirroring the rename/remove
tests. Frontend: extract notifyCollectionEditError (the 409->nameTaken vs
generic-fallback mapping that was inline in the New-collection create()) and
unit-test it; wire create() through it.
Renders PageScroller and drives restoreScroll's rAF synchronously with fake
metrics: asserts a forward nav (new entry id) opens at top and Back restores
the offset, and that a new content key within one entry resets to top while
returning restores — the per-(entry,key) contract only the id primitives were
covering before.
AddItemPanel (Collection page) and NewCollectionPanel (browse) had near-
duplicate debounced library-search + result-list rendering. Extract it into
one PlexLibraryPicker that owns the query, the loading/error/empty states and
the row markup; callers keep their own chrome and pass term + renderAction.
No behaviour change.
Adds ruff's formatter to the gate. line-length 100 matches the frontend
prettier printWidth. The reformat itself lands as the next commit (kept
separate so it can go in .git-blame-ignore-revs).
2026-08-23 07:42:42 +02:00
667 changed files with 52213 additions and 5485 deletions
The local gate is **ruff + typecheck (3 tsc projects) + eslint + prettier --check + knip + vitest +
pytest**. It runs on every `publish`, which **refuses to build on failure — there is deliberately no
The local gate is **ruff (`check` + `format --check`) + typecheck (3 tsc projects) + eslint + prettier
--check + knip + vitest + pytest**. It runs on every `publish`, which **refuses to build on failure — there is deliberately no
`--skip` flag.** Run it and get it green **before** any merge, release, or handing work to review.
- Full gate (agents with the private ops dispatcher): `siftlode check`.
- Directly, without the dispatcher: `cd frontend && npm run typecheck && npm run lint && npm run knip && npm run test`, plus backend `ruff check backend` and `pytest` (the DB-lane needs Postgres).
- Directly, without the dispatcher: `cd frontend && npm run typecheck && npm run lint && npm run knip && npm run test`, plus backend `ruff check backend` + `ruff format --check backend` and `pytest` (the DB-lane needs Postgres).
- New DB-backed test? It must actually run, not silently skip — assume `SIFTLODE_TEST_REQUIRE_DB=1`.
## Workflow
@@ -26,43 +26,71 @@ pytest**. It runs on every `publish`, which **refuses to build on failure — th
keep commits logical and Conventional-Commit-styled.
- **No AI attribution in commits.** No `Co-Authored-By` trailer, no "generated by" lines. (This file
being AI-written is the intended, visible exception; commits are not.)
- **Release is a deliberate, user-owned step** (`siftlode release` from the feature branch: merge to
`main` → publish → deploy → verify `/api/version`). Do not self-ship; bring work to a green,
committed state and stop for the user's go-ahead.
- **Release: user-owned in timing, agent-run in execution.** The user decides _when_ to ship; on their
go-ahead _you_ run `siftlode release` from the feature branch (merge to `main` → publish → deploy →
verify `/api/version`). Never self-ship — bring work to a green, committed state and stop for the
go-ahead first, then run it yourself.
- **Public repo.** Keep ops/topology/secrets out — those live in a separate private repo. Nothing here
should assume or leak them.
## Collaboration — how to work with the human
The human is the product owner: they bring ideas and requirements. You are the technical lead — you own
how it gets built. Concretely:
- **Plan before you build.** For anything non-trivial, present a short plan (sprint-level for larger
work) and get a go-ahead before writing code.
- **Offer choices, mark the optimal — and explain the trade-off.** When several solutions are viable,
list them and mark the recommended one. This is the _one_ place to spend extra words: give the _why_
behind each option and the weight of the decision. These decision points are where the product owner
engages with the build and learns from it, so make them understandable — a reasoned menu, not a bare one.
- **Mockups for design/UX decisions.** For any visual or interaction choice, show mockup options to
choose from before building.
- **Challenge unsound requests.** If a request is technically ill-advised and you can back it with real
arguments, push back or propose an alternative — don't comply blindly.
- **Verify UI in the browser before calling it done.** tsc/unit/DOM tests miss visual and interaction
bugs; do a real visual check of any UI change.
- **Be terse everywhere else.** Outside those decision points, lead with the answer or decision. During
implementation and debugging especially: no restating the same point reworded, no filler.
# A COPY, never the card itself: a self-playing pin (a video, a movie) would otherwise make
# `card["next"] is card`, which serializes as a circular reference and 500s the whole shelf. The
# copy is taken before `pin_id`/`next` are attached, so `next` stays a plain playable card.
card["pin_id"]=pin.id
card["next"]=nxt
returncard
defshelf(db:Session,user:User)->list[dict]:
"""The whole shelf, in the user's own order."""
out=[]
forpininlist_pins(db,user):
card=_shelf_card(db,user,pin)
ifcardisnotNone:
out.append(card)
returnout
defplayables(shelf_cards:list[dict])->list[dict]:
"""The `next` of every shelf entry that has one — what the hero may draw from. A pinned channel
with nothing unwatched left, or an empty playlist, simply is not an answer right now."""
return[c["next"]forcinshelf_cardsifc.get("next")]
Some files were not shown because too many files have changed in this diff
Show More
Reference in New Issue
Block a user
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
peter