Advertisement.species existed in the schema but was never wired up —
the admin form never let you set it and the feed's ad query ignored it
entirely, showing every active ad to every pet regardless of species.
Adds a new Advertisement.breeds field alongside it and wires both
through the whole path: a targeting UI in the ad create/edit dialog
(species checkboxes, each expandable into its own breed checklist),
and a filter in admin.getActiveAds (now scoped to the viewing pet) so
cat food ads stop reaching dog owners. Empty species/breeds = shown to
everyone, unchanged from today's behavior.
Replaces the emailed Supabase signed URL with a bespoke DataExportToken
DB record, delivered via a new /api/gdpr-export/[token] route that
streams the file server-side. The storage backend is now an internal
implementation detail; every link the user sees reads pawfeed.org.
Aggregates everything held about an Owner (gdpr-export.ts: pets, posts,
stories, comments, reactions, reposts, follows/blocks, health records,
messages, notifications, reports, ban/consent history, Clerk profile)
into a single JSON bundle, uploads it to a private Supabase Storage
bucket (data-export-storage.ts — separate from the public pet-avatars
bucket used for media), and emails the owner a 7-day signed download
link via the new SMTP module. Triggered from a new export button per
user in /p/[secret]/users, logged to ModerationLog.
Deliberately not a zip with embedded media bytes: avoids new archiving
dependencies, the risk of a slow synchronous request for owners with
many posts, and the fact that Mux video assets aren't trivially
re-downloadable — media is referenced by its existing CDN URL instead.
Ad interactions and issued invite codes are not yet included, flagged
in the export's own scope note rather than silently omitted.
Verified end-to-end against the live shared DB before committing:
built a real export, uploaded it, generated a signed URL, and sent a
real test email — all succeeded.
Gitea #6 (parts 1+2): admin.listUsers now batch-fetches email/name from
Clerk (users only stored the Clerk userId locally, everything else
lives in Clerk) instead of showing a bare userId. Adds
admin.startTeamConversation, which lazily creates a "PawFeed Team"
system Pet under ADMIN_OWNER_ID and reuses the existing pet-to-pet
Conversation/Message model instead of a parallel admin-inbox system —
the admin panel switches ActivePet to Team and opens the ordinary
/messages/[conversationId] thread, so no new UI is needed on the
recipient's side.
Known limitation: only the literal ADMIN_OWNER_ID account can act as
Team, since assertPetOwnership requires an exact owner match.
Extending this to every AdminRole holder is left for a follow-up.
Part 3 (GDPR Art. 15 data export) is deliberately not included —
research only per the issue's own "erst prüfen dann bauen" note.
Two new chart panels on the admin dashboard, each with a day/week/month
granularity toggle:
- "Posts over time" — bar chart of new posts per bucket (trailing 30
days / 12 weeks / 12 months, zero-filled so gaps show as zero, not
missing points).
- "Owner growth" — line+area chart of the cumulative total registered
owners, seeded from a pre-window baseline count so the curve starts
at the correct absolute total instead of zero.
Backend: admin.getPostsOverTime/getOwnerGrowth bucket timestamps in
plain JS (UTC-consistent day/ISO-week-Monday/month-1st truncation) —
deliberately not raw SQL date_trunc, to avoid DB-session-timezone
ambiguity and keep this testable without hitting Postgres. Data volume
is small enough (Alpha-stage app) that fetching bare createdAt columns
and bucketing in memory is the simpler, lower-risk choice over adding
the codebase's first raw-SQL query.
Frontend: new src/components/admin/ (TimeSeriesChart, GranularityToggle)
following the dataviz skill — thin 2px marks, rounded bar corners,
recessive gridlines, hover crosshair+tooltip on both variants, single
brand-orange hue since each chart is one series (no legend needed).
Admin surface is fixed dark (zinc-950), so colors are hardcoded rather
than light/dark-aware, unlike the app-wide WeightChart this borrows its
SVG approach from.
Added 2 regression tests covering zero-fill bucketing and the
cumulative-with-baseline math. prisma-mock.ts gained owner.findMany/
owner.count (previously missing).
- feed.ts: Postgres fallback path (used when Redis is down) never
filtered blocked pets, unlike the primary Redis path — blocked pets'
posts could reappear during a Redis hiccup.
- feed-helpers.ts/feed.ts: getFeedPage always paginated by a hardcoded
PAGE_SIZE=20, silently ignoring the client's requested (zod-validated
1-50) limit whenever Redis was up.
- milestones.ts: create accepted any MilestoneType including the
follower-threshold values (FOLLOWERS_500..FOLLOWERS_1M), which are
meant to be exclusively auto-awarded — restriction existed only in
the MilestoneForm picker, not server-side. Now validated against the
existing USER_INITIATED_MILESTONE_TYPES constant.
- admin.ts: deletePost/deleteReportedPost didn't decrement the original
post's repostCount when the deleted post was itself a repost, unlike
reposts.ts's user-facing delete — counter drifted upward permanently.
- PawButton.tsx: count was optimistically guessed off a `reacted` flag
that can be stale (callers pass initialReacted={false} unconditionally),
permanently corrupting the displayed count on click. Count is now
computed from the server's actual toggle result instead of guessed.
- PostDetailDialog.tsx: Report/Block dropdown items had no onClick and
neither ReportSheet nor BlockDialog were rendered — wired up to match
PostCard.tsx's existing pattern.
- VideoCard.tsx: invalidated the entire query cache on every video-ready
transition; scoped to feed/posts/explore router queries instead.
deletePost, banUser, unbanUser, dismissReport, and deleteReportedPost each
ran the mutation and the moderationLog.create as two separate sequential
awaits. The audit flagged this as "should be Promise.all", but that would
have introduced a real bug: if the delete/ban call failed, a concurrently
fired log write could still succeed, recording an action that never
happened. Wrapped both in $transaction instead (matching the pattern
deleteComment already used) — atomic, and the log can only be written if
the mutation actually succeeded.
Reviewed the audit's other Cluster H finding (a few include-vs-select
over-fetches) and found nothing actionable — the flagged spots already
narrow relations with select where it matters; the unnarrowed ones are
small lookup/relation tables used in full by their call sites.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Bundles the 2026-07-12 code-audit session (Clusters A/B/C/D partial/E/F):
denormalized Post/Advertisement reaction/comment/repost counters synced
transactionally instead of live _count queries; real cursor-based pagination
for followers/following/blocks lists; assertPetOwnership + formatRelativeTime
centralized; dead r2.ts + AWS SDK deps removed; missing DB indexes added;
account-deletion flow, mention notifications, pull-to-refresh feed, and
mobile UI/i18n fixes from the surrounding sessions.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>