Video cards show a resume progress bar for started-but-unfinished videos and a
hover overlay: Play on every card, Continue + Restart on in-progress ones. The
in-app player now resumes from (and checkpoints to) the server position instead
of localStorage, accepts an explicit startAt (Restart -> 0), and refreshes the
feed on close so the card bar reflects the session. Sidebar gains an
'In progress' show filter.
Add position_seconds (+progress_updated_at) to video_states so watch progress
survives across devices and can drive a feed filter. New POST
/api/videos/{id}/progress checkpoints the player position (clearing trivially
-early and near-finished positions). Feed serialize exposes position_seconds and
a show=in_progress filter lists started-but-unfinished videos. Un-marking
'watched' now keeps a stored position instead of deleting the row.
- header: per-user "N without full history" count (channels_deep_pending),
clickable with a hint -> opens the channel manager filtered to those.
- channel manager: status filter chips (All / Needs full history / Fully synced
/ Hidden); the header link deep-links to "Needs full history".
- fix: priority up/down is now an optimistic in-place cache update (no refetch /
re-sort), so the list no longer jumps to the top and loses your scroll position;
the new order applies on the next page load.
The header status bar read the global /api/sync/status (videos_total +
channels_backfilling), so every user saw the whole catalog's numbers —
confusing and a small cross-user info leak (e.g. "3 syncing" for a user with 2
channels). It now uses /api/sync/my-status: the user's own available video count
and how many of their own channels are still being fetched
(channels_recent_pending). The pause control stays admin-only via an isAdmin prop.
After read access is granted the wizard now imports the user's YouTube
subscriptions automatically (with a "Building your feed…" progress state), so a
new user lands on a populated feed — channels already in the shared catalog show
up instantly, new ones backfill in the background. The empty feed now prompts
users without read access to set up via the wizard instead of a bare message.
- requirements: cryptography >=46.0.7 (was pinned <46, which excluded the fix for
the CVEs pip-audit flagged in our Fernet/crypto library). pip-audit now clean.
- Dockerfile: upgrade pip before installing deps (patches installer-level CVEs).
- auth: /auth/upgrade now defaults to the least-privileged read scope; only an
explicit access=write requests the write scope.
docker-compose.prod.yml targets the public VPS: Postgres is never published, the
app binds to 127.0.0.1 (Caddy proxies it), and every service has a memory/CPU cap
plus no-new-privileges / cap_drop / read-only rootfs. deploy/deploy.sh rolls out
main with a host-side build (migrations run via the entrypoint). CI type-checks
and builds the frontend and byte-compiles the backend on every push to main.
Adds login-free /privacy and /terms pages (rendered outside the authenticated
tree via a pathname switch in main.tsx) carrying the Google API Services Limited
Use disclosure, YouTube ToS / Google Privacy links, and contact + data-deletion
info. The sign-in screen now describes the app and links to both, satisfying
Google's homepage + privacy-policy requirements for the OAuth consent screen.
After the clean name/email sign-in, a wizard walks the user through granting
YouTube read (then optionally write) one step at a time, each with a plain
rationale and an up-front heads-up about Google's "unverified app" screen.
The visible step is derived from the granted scopes (can_read/can_write) so the
flow resumes correctly across the full-page consent redirect; it's dismissible
and reopenable from Settings -> Account, which now lists read and write as
separate, individually-grantable access rows.
Base login now requests only openid/email/profile (non-sensitive), so a new user
gets a clean Google consent with no "unverified app" warning and no 7-day refresh
token expiry. YouTube read (youtube.readonly) and write (youtube) are granted later
by the onboarding wizard via a parameterized /auth/upgrade?access=read|write.
Security fixes folded in from the baseline audit:
- config: refuse to boot in production (https OAUTH_REDIRECT_URL) with the
placeholder/short SECRET_KEY or a missing TOKEN_ENCRYPTION_KEY, closing a
session-forgery / admin-impersonation hole.
- main: mark the session cookie Secure when served over HTTPS.
- me: expose can_read; sync/subscriptions returns a friendly 403 (not a 500)
until YouTube read access is granted.
Three low-risk wins for large filtered feeds:
- content-visibility:auto on cards (.cv-card/.cv-row) so the browser skips
layout/paint for off-screen cards; contain-intrinsic-size keeps the scrollbar
stable and is remembered per card after first render.
- memo(VideoCard) + stable onState/onChannelFilter callbacks (onState reads the
loaded list via a ref) so appending a page only renders the ~60 new cards
instead of reconciling every card already on screen.
- Prefetch the next page earlier (sentinel rootMargin 800px → 1500px) so the
'Loading more…' flash is far less likely during fast scrolling.
Persisted notifications were restored with their saved 'dismissed' state, so any
that were still active when the page closed came back as active toasts — but
their auto-dismiss timers aren't re-armed on load, leaving them stuck on screen
forever (a burst of errors could pile up dozens). Mark restored entries as
dismissed on load: they stay in the bell history but no longer reappear as live
toasts. The toast surface is session-transient; history persists.
Channel/account avatars come from Google's image CDNs (yt3.ggpht.com,
lh3.googleusercontent.com). On a feed page dozens load at once; the CDN
rate-limits the referrer-bearing burst (429), so a random subset rendered the
browser's broken-image icon (the URLs themselves are valid — verified 200).
Add a shared <Avatar> that sets referrerPolicy="no-referrer" (which the CDNs
serve without throttling) and falls back to a neutral initial placeholder on
error instead of the broken-image icon. Use it for video-card, player, channel
manager, header and settings avatars.
In-app modal player (YouTube IFrame API) on card click as an alternative to
opening youtube.com in a new tab: resume position, watched controls + auto-watch,
a lazily-fetched description popover with linkified timestamps/emails/hashtags/
links, and YouTube links that play inline. Also: a 'watched' notification with an
Unwatch action, and a feed override fix so reverting status re-shows the video.
Built on branch experiment/inline-player, reshaped into logical commits on
feat/inline-player.
The feed kept per-video optimistic status overrides until filters changed, so a
status reverted elsewhere (e.g. Unwatch from the notification center) was masked
by the stale override and stayed filtered out of the current view. Also clear
overrides when fresh feed data arrives — the server is then authoritative.
Mirror the hide flow for watched: marking a video watched (card, modal toggle,
or auto-watch) raises a toast and a bell-history entry with an Unwatch action —
no Find-in-feed, just revert. NotifMeta becomes a discriminated union
(video-hidden | video-watched); unhide/unwatch share one revert-to-new helper.
- Narrow the title hover target to the actual text, not the whole row.
- Linkify descriptions: timestamps (mm:ss / hh:mm:ss) seek the player; emails
become mailto:; hashtags link to YouTube's hashtag feed; other URLs open in a
new tab. Blank lines are stripped so the popover isn't mostly whitespace.
- YouTube links play in the inline player: a link to the current video seeks
(honoring t=), a link to another video navigates the player to it, with a Back
button to the original. While on a linked video the title/author come from the
player and its views/date/duration + a clickable channel come from the detail
endpoint, which falls back to the YouTube API (videos.list, attributed to the
user) for videos not in our DB.
The popover bled through (it used the translucent glass surface), was clipped by
the modal card's overflow, and ran off the bottom of the viewport. Make it an
opaque surface, render it through a portal to <body> with fixed positioning so
nothing clips it, and anchor it above the title so it grows upward where there's
room. Add a small hover grace so the pointer can travel title → popover.
- Watched: an explicit toggle in the modal (Mark watched / Watched→unmark) plus
auto-mark when playback reaches the end (within 10s, or on the ended event).
- Compact layout: drop the header bar and the redundant 'Open on YouTube' button
(the embed's own YouTube logo already jumps out); Close moves to the title row,
channel + meta share one line — fits without a scrollbar at higher zoom.
- Card actions reflect status: watched shows a double-check, saved a filled
bookmark, with matching tooltips.
- Description: new GET /api/videos/{id} exposes the already-stored description,
shown in a popover when hovering the modal title (fetched lazily).
Left-clicking a feed card (Ctrl/Cmd/middle still open youtube.com in a new tab)
opens a modal that plays the video in-app via the YouTube IFrame Player API
instead of leaving the app. Using the JS API (not a bare embed) lets us read the
playback position: it's checkpointed per-video in localStorage and on close, and
restored via the 'start' param when the video is reopened. The modal closes via
a header button, the backdrop, or ESC (ESC only while focus is on our page — a
cross-origin iframe owns its own key events).
The channel filter chip stored only the channel id in the URL, so after a
page refresh the human name was lost and it fell back to "This channel".
Resolve the title from the cached channels list keyed by id, and show a
"Loading…" label instead of the misleading fallback while it resolves.
Track who burned how much YouTube API quota. A QuotaEvent audit log (migration
0009) records every spend with the triggering user (NULL = background/system) and
an action label, set via a request/job-scoped contextvar (quota.attribute) so no
call signatures change. User-initiated work (sync subscriptions, unsubscribe,
opt-in recent backfill, manual enrich) attributes to the user; scheduler work to
System, split by action.
- backend: QuotaEvent model + migration 0009; quota.attribute() contextvar;
record_usage logs events; entry points wrapped (routes/sync, routes/channels,
scheduler); GET /api/quota/my-usage + GET /api/quota/admin
- frontend: admin-only Stats page (header nav, page=stats) with daily bars +
per-user breakdown by action and range picker; 'Your API usage' in Settings ->
Sync for every user
Verified: attribution + endpoints compute correctly; events are per-user vs System.
deep_requested is per-subscription, but deep backfill is channel-wide and shared:
once any subscriber requests full history, the whole back-catalog arrives for
everyone. The chip only looked at the current user's flag, so a second subscriber
saw a misleading 'get full history' on a channel already queued by someone else.
Add a channel-level deep_in_queue to /api/channels and show an informational
'full history queued' badge (vs the owner's cancelable button) in that case.
Personalize the access-request and approval emails, add a real Date header and a
Reply-To (approval -> admin, new-request -> requester) so they read as a
conversation rather than a no-reply blast — small nudges for deliverability;
sender reputation still does most of the work.
Move the access whitelist from the ALLOWED_EMAILS env var into a DB Invite table
(env kept as bootstrap fallback), and add a self-service request + admin approval
flow with fail-soft email.
- models: Invite(email, status pending|approved|denied, requested_at, decided_*)
- migration 0008: invites table; seed env ALLOWED_EMAILS u ADMIN_EMAILS as approved
- auth: is_allowed() (DB-first, env fallback); a denied Google login records a pending
request and bounces to /?access=requested instead of a raw 403; public POST
/auth/request-access; upsert is idempotent so repeats don't re-spam admins
- routes/admin.py (admin-only): list/approve/deny invites + manual add
- email.py: smtplib + Gmail App Password, fail-soft (skips if SMTP unset)
- /api/me exposes pending_invites; config + .env.example gain SMTP_*
- UI: Login 'Request access' form + access=requested/denied handling; Settings ->
Access requests (approve/deny + add); admin nudge toast on pending requests
Verified locally: request-access creates a pending invite and emails the admin;
seed approved npeter83; guinea-pig yt.trash2023 denied until approved.
When recent backfill stopped partway through a page (age cutoff or count cap),
it stored the *next* page token as the deep-backfill cursor, so the older items
remaining on that same page were collected by neither pass — a channel could
report backfill_done while silently missing a band of videos right around the
365-day cutoff. Now resume deep from the page we stopped on (re-fetched; inserts
are idempotent). Verified: Pánczél went 200 -> 209 stored (= full uploads
playlist) after re-backfill.
When a user requests full history for a channel whose recent uploads aren't
fetched yet, run a one-channel recent backfill synchronously in the request so
the feed populates at once instead of waiting for the scheduler. Deep paging
still follows on the scheduler (recent-then-deep). The deep-toggle mutation now
also refreshes my-status and the feed.
Per-user opt-in to full-history (deep) backfill so a new user's unique
channels no longer trigger a big shared-quota burst.
- migration 0007: Subscription.deep_requested (default false); seed admins'
existing subscriptions to preserve today's global-backfill behaviour
- run_deep_backfill is now demand-driven: only channels at least one user has
requested are deep-backfilled; recent backfill stays unconditional (cheap)
- estimate_deep_backfill ETA helper (quota-bound) surfaced in /api/sync/my-status
- POST /api/sync/deep-all to opt all my channels in; PATCH channels accepts
deep_requested
- UI: per-channel Full history toggle, Backfill everything action, deep
progress + ETA in Channels header and Settings - Sync
