The Wikimedia feed's thumbnail is 330px, which upscales blurry in our hero. Use
originalimage.source instead — it's reliably sharp. (Can't just request a bigger
thumbnail width: for very large source images Wikimedia only serves pre-generated
bucket sizes and 400s on arbitrary widths — e.g. 500px ok, 800/1024px fail.)
- onthisday._best_image() prefers originalimage, falls back to the thumbnail.
- scripts/otd_image_upsize_backfill.py re-fetches each stored MM-DD and upgrades
image_url in onthisday_pool + daily_onthisday in place (ran on host: pool + 6
daily rows now sharp; today's hero verified 200). Only the /onthisday hero
loads this image (home card is text-only), so larger files are a single-page,
one-time load.
- test_best_image locks the prefer-original/fallback behavior.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- daily_art gains blurb + palette columns (idempotent migration).
- art._palette: Pillow median-cut to ~5 hex colors from the cached image (best-
effort → [] on any failure). art._blurb: a warm 2-3 sentence "what you're
looking at" note grounded in the Met catalogue (title/artist/bio/date/medium/
classification/culture/tags). Prompt leans on context/significance and the
title+tags for subject — explicitly NOT asserting literal composition (figure
counts/poses) it can't see, since the model can't view the image. Markdown
stripped from the output.
- pick_daily generates both (client optional → blurb skipped when absent); cycle
+ art CLI pass an LLM client. /api/art/today exposes blurb + palette.
- Backfilled the last 3 days on host (Veteran / Magnolia Vase / Bierstadt).
- scripts/art_blurb_palette_backfill.py for in-place backfill (no re-pick).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Replaces the gist-based read-time with the SOURCE article's full read time — the
contrast that sells the gist ("calm 1-min version here; ~10 min for the deep dive").
- goodnews/readtime.py: word_count_from_html (strips script/style/nav/header/
footer/form/button/aside furniture before counting) + source_read_minutes
(~225 wpm, 200-word floor, None when extraction looks failed/too thin).
- articles.source_words + read_checked_at columns (count only, never the body;
fits the privacy posture). Idempotent migration.
- enrich.fetch_source_words + enrich_read_times: a bounded, retry-guarded cycle
step (mirrors the image enrichers) that counts words for recent accepted
articles. Only ever writes a real count; never overwrites good with zero. Wired
into the cycle after recent-image enrichment.
- queries: source_words flows through _ARTICLE_COLUMNS; api exposes
source_read_minutes on Article (null when unknown).
- home3: News card shows "Full story · ~N min", hidden entirely when null (no
misleading "1 min").
- Tests: furniture stripping, threshold/rounding, enrich idempotency + no
zero-overwrite, API null handling. 412 backend.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Per Codex audit: only accept a polish when there's a gloss AND at least one
example sentence that actually contains the word (case-insensitive). Examples
that don't use the word are dropped; if none remain, fall back to the raw
dictionary def/examples instead of shipping a gloss with empty/irrelevant usage.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Content quality ("LLM polishes, dictionary anchors"):
- New wotd._polish: rewrites the real dictionary gloss into ONE warm plain
sentence + two clear everyday example sentences, grounded in the real
definition (no invented meanings). Stored in new wotd_pool/daily_wotd columns
gloss + usage, alongside the raw definition/examples which stay the anchor.
- harvest() polishes each new word; pick_daily() lazily polishes + caches back
any older pooled word that lacks a gloss (client threaded through run_daily).
- Admin word-add polishes on insert; re-pick passes an LLM client so quote
meaning / word gloss fill on a forced fresh pick.
- /api/word/today now prefers gloss + usage, falling back to the raw dictionary
def/examples when polish is absent (so it's always safe).
- db._migrate adds gloss/usage to wotd_pool + daily_wotd (idempotent ALTER).
Frontend — /word redesigned to CD's "Editorial Asymmetric": faded oversized
initial bleeding off the right, vertical part-of-speech rail, big Newsreader
word, airy definition, left-ruled italic example sentences, outline Listen
button + date. (Uses our self-hosted Newsreader/Hanken stack rather than the
mockup's Google fonts; the made-up syllable respelling is omitted since we only
have real IPA.)
Tests: _polish parse/trim/cap, harvest stores gloss/usage, pick lazy-polishes
older words, admin gloss flows through to /api/word/today. 403 backend + 27 fe.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- Admin joy item route moved to /api/admin/joys/{kind}/items/{item_id} so the
/add and /repick verbs resolve to their own routes instead of 422-ing as a
non-int item id (the launch blocker). Frontend mutate URL updated to match.
- Re-pick now excludes the currently-shown item: the endpoint reads today's
daily pool_id and passes it as `avoid`, so "Re-pick today" yields a different
item. Added `avoid` to pick_daily/_candidates across wotd/quote/onthisday.
- WOTD sense selection: the LLM now proposes word + intended part of speech, and
_lookup prefers that sense (fixes "serene" returning the archaic noun).
- On This Day tone prompt tightened to favor genuinely uplifting events and
exclude merely procedural/political-administrative ones.
- Caddy @hidden now also noindexes /word /quote /onthisday /admin (+ .html).
- Regression tests: add/repick resolve (401 not 422), add/feature/block/delete,
re-pick excludes current; WOTD pos-preference + proposal parsing units.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- /api/art/image/{id} now answers HEAD as well as GET (was 404 on HEAD) — mirrors the
/a/{id} fix. Added tests/test_art_api.py (GET+HEAD+size=full fallback + today payload).
- /textures/* served immutable (long cache) instead of no-cache; excluded from the
revalidate matcher. Live Caddyfile + repo snapshot both updated.
- Lightbox: Escape closes it, and focus moves to it on open (keyboard-friendly).
- Trimmed the gallery's top padding so "Daily Art" sits closer to the bar.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- Virtual frames (Walnut/Gold/Silver/None), selectable + remembered in localStorage,
built as a beveled moulding around a cream museum mat.
- Header uses the real /logo.svg wordmark; the "No ads" pill is replaced by an
account icon (the pill doesn't need to follow every page).
- Lightbox now opens a full-resolution copy that fills the screen: art._download_image
caches a hi-res {id}-full copy alongside the web-large display copy, served via
/api/art/image/{id}?size=full (image_url_large in /api/art/today).
- Centered the placard bullet separators (explicit .sep spans, equal margins).
- Image no longer shifts on hover; a quiet "Click to expand" affordance sits on the art.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Hardening before it runs further on the cycle:
- DB-lock/network: all HTTP (metadata + image) happens before any write; the write txn
opens only at the brief INSERT and commits immediately. Images download to a temp file
then atomic os.replace into cache (a reader never sees a half-written file).
- Site-timezone "daily" already used local_today() (same rhythm as the Brief) — confirmed.
- Attribution from day one: store + return title/artist/date/medium/department/credit/
source_url/object_id/source + museum name + is_public_domain license marker + the full-
res source URL (for a richer /art view later). UI can show: Title · Artist · The Met.
- "highlight != always beautiful": added a manual `blocked` flag on art_pool (excluded
from picks) as the cheap curation lever; a featured override can follow.
Schema migrated (existing art tables get the new columns). 373 tests green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The engine for the /art room (design-independent; deploy held for Codex review).
- goodnews/art.py: harvest a curated pool of public-domain HIGHLIGHT artworks from the
Met (isHighlight+isPublicDomain+hasImages -> masterworks, never potsherds; CC0). Daily
deterministic pick from the least-recently-shown (no soon-repeats, same for everyone),
fetch metadata + download the image to OUR cache (data/art_cache) so the homepage never
waits on or hotlinks the museum. Bulletproof: bad object/image falls through candidates;
a failed day keeps the last piece (room never empty). Injectable HTTP for tests.
- Schema: art_pool + daily_art. /api/art/today (edge-cacheable) + /api/art/image/{id}
(served from cache, immutable). CLI `art [--harvest] [--force]` + a non-fatal cycle step.
- Tests (5, mocked HTTP) + verified live against the Met: harvested 1641 works,
picked/cached "Repose" by John White Alexander. 371 tests green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Per the brand-name standard (camelCase, one word). Updated the SMTP From default and
the digest email body/subject strings. Live env From values (auth.env + goodnews.env)
updated to match. (Web/OG brand strings in share.py + app.html are the remaining sweep.)
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- Hero constraint: _pick_lead now runs only within the CLOSEST non-empty section of a
personalized Brief, so a "gentler" wider-region/world story can never be floated into
the hero slot above a local one. Only widens if the closest section is empty.
- Dial gains a visible Clear (alongside Change) so a reader never feels locked into
personalization; "World" stays the keep-home-but-go-global option.
366 tests green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Codex-approved evolution: the reader controls the "emotional radius" of the landing.
- Census-region "Regional" grain (geo.region_of / region_states). Scope-aware tiering
(queries.home_tiers): closest->widest lead, confidence-gated on state + region, never
a hard filter — blends outward so the set is always full. 'world' = the global brief.
- queries.home_brief takes a scope; /api/brief gains a scope param (nearby|region|
country|world). Country-only / non-US homes collapse to country.
- Homepage dial replaces the 2-button toggle: adaptive stops (4 with a US state, else
Country/World), persisted scope, "Good news closest first" framing. Concrete, soft
section labels (Around New Jersey / Across the Northeast / Across the US / Around the
world) so the reader sees the dial worked.
Backend 366 + frontend tests green. (Latest feed still on v1 local-first; aligning it
to the dial is the immediate follow-up.)
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Per the owner's call (overrides the earlier "Brief sacred" stance): when a home is
set, the homepage opens with local good news first, not global. This is the hook —
you land and see awesome stories from YOUR corner first.
- queries.home_brief: local-first highlights (high/medium-confidence near, blended
out to country then world so it's always a full, strong set), preferring already-
summarized stories so the calm read stays rich. Recent window, ranked within tier.
- /api/brief gains a `home` param: private/no-store when set; over-fetches + caps so
dismissal/boundary filtering never thins it; falls back to global top-up if needed.
- Landing UI: a Local <-> Global toggle ("📍 Near you / 🌍 Everywhere") when a home
is set, the calm picker invite when not (dismissible), and Change. Default leads
local; one tap back to the global brief. No home set => exactly today's behavior.
Backend + frontend tests green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Codex polish before deploy: anything elevated as Near you / Close to home must have
geo_confidence in (high, medium) — the feature's promise is relevance. Country-only
mode now gates "near" too; since it has no "country" tier, the "world" scope is
widened to absorb low-confidence home-country stories so they surface there instead
of vanishing between tiers (the same edge-case class, fixed). State mode unchanged.
364 tests green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Completes the server side of "Closer to Home". /api/feed gains a `home` param
('US' or 'US-NY'); when set the response is private (like prefs) and sectioned:
- Near you (+ Elsewhere in your country when a state is set) is a ONE-TIME lead
block on page 0; the world is the paginated body. next_offset tells the client
where to continue, so the lead block never skews world paging.
- Thin tiers fold down (MIN_TIER=3) so a header is never shown empty (lead, don't trap).
- State match counts only on high/medium geo confidence; the "country" tier excludes
exactly what went to "near", so a low-confidence home-state story still surfaces
(it doesn't vanish between tiers — caught + tested).
- Items carry a `section` tag; paywalled sort is now within-section. No home => exact
prior behavior (section null, default/edge-cached feed unchanged), Brief untouched.
364 tests green. Frontend next: Home picker + sectioned feed rendering.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
"Closer to Home" foundation (audit greenlit by Codex). Durable geography, kept
decoupled from volatile scoring.
- Schema: article_geo (breadth/confidence/rationale/geo_version) + article_places
(0..N ISO-coded places), separate from article_scores so re-runs/audits never
disturb scoring or acceptance. "local" is never stored — it's relative to the
reader; the UI computes "Near you" later.
- geo.py: LLM proposes place NAMES, code disposes to ISO codes (country alpha-2,
US state 2-letter); region words like "Europe" can never become a country.
'global'/placeless is first-class, not failure. Confidence calibrated so 'high'
needs an explicit location. Geo is its OWN LLM pass, not merged into the scoring
prompt (durable metadata, re-runnable, keeps the sensitive prompt untouched).
- store_geo replaces places (geo is re-derivable, unlike scores). tag_articles is
idempotent by geo_version, only touches accepted non-duplicate articles.
- CLI `geo` command (cycle-locked, --limit/--reclassify) for backfill, plus a
bounded geo step in the cycle (--geo-limit 60, --no-geo). scripts/geo_audit.py
is the prototype audit tool.
360 tests green; live smoke tagged real articles correctly (Gaza->PS, London->GB,
placeless science->global). No UI / SEO pages yet — ranking/personalization only.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Sharpen the existing daily-game share loop into something measurable (per Codex's
"instrument what you have, then feed people into it" plan), ahead of a Show HN launch.
Analytics:
- Per-game funnel events <game>_{arrival,started,completed,shared} (article_id=0).
arrival = landed via a shared link (utm_source=game_share); started = first move
(guess/find/flip); completed = solved/cleared/Full Bloom; shared = on share success.
- trackVisit() moved into the global layout so direct /play landings count; the
server-rendered /a/ share page now creates a visitor token + sends a daily visit
beacon (first-time /a/-only visitors were previously dropped).
- Admin "Games funnel" panel: arrivals / engaged / completed / shared, per game.
Sharing:
- Memory Match gains a Share button (it was the only game without one).
- All shares deep-link to the exact game+variant with a full https:// URL +
utm_source=game_share (gameShareUrl helper), instead of a bare /play.
- "shared" is counted only after navigator.share()/clipboard.writeText() succeeds.
/play social metadata:
- /play served homepage canonical/OG (static SPA, ssr=false). postbuild script
patches build/play.html's head to /play canonical/title/description/OG; fails the
build if the homepage tags drift. Caddy try_files now serves {path}.html so /play
is served from the patched file (snapshot in deploy/caddy/).
Tests: backend 352, frontend 27.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The deploy pipeline runs from the working tree, so a wave of shipped features
had never been committed. This snapshots git to what's actually running.
SEO impression recovery (live + verified):
- Duplicate /a/{id} now 301-redirect to their canonical twin instead of 404
(a hard 404 silently dropped already-indexed URLs and tanked impressions).
- Dedup representative selection reworked: accepted/serveable -> established
rep (URL stability) -> quality score, so an accepted page never retires to a
rejected rep and an indexed canonical doesn't churn when a newer twin arrives.
- HEAD /a/{id} returns the same status as GET (api_route GET+HEAD) instead of
falling through to the static mount and 404ing.
- `dedup --force-recluster`: cycle-locked, model-free re-cluster to re-apply the
policy to the existing corpus (shared cycle_lock context manager).
- CLI honors GOODNEWS_DB for its default --db (was silently ignored).
Publishing Desk (admin tool to post highlights to X via Web Intents):
- publishing.py queue/rank/handle-resolution; admin UI; full searchable emoji
picker (bundled data, no CDN) for the blurb editor.
Play games + site:
- Bloom (word-wheel), Memory Match, daily ritual set, Zen Den (dev-gated).
- English-only language gate; source prospecting; paywall + dedup hardening.
Tests: full suite green (349). Ignores tightened (node_modules, data/*.db).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The Articles inspector revealed paywall is domain-coarse: nytimes.com is flagged,
so NY Times Learning's free Word-of-the-Day inherits 🔒 — and that flag isn't
cosmetic, it deprioritizes the content in feed sort + lead selection. Add a
per-source override so admins can correct it after inspecting.
- sources.paywall_override: NULL (domain rule) | 'free' | 'paywalled'.
- paywall.py: keep low-level is_paywalled(url) (domain); add is_paywalled_for_source
(url, override) for the EFFECTIVE decision — never patched the domain helper
globally (per Codex), so "domain says X" stays distinguishable from "overridden".
- Threaded everywhere ranking/UI touches paywall, via src.paywall_override on the
shared _ARTICLE_COLUMNS + the source-aware helper: feed sort, /api/since, replace,
lead selection, Article badge, brief composition (briefs.py), digest, source_health
(table 🔒), the Articles inspector, and the review/attention check — so ranking and
UI always agree.
- Endpoint POST /api/admin/sources/{id}/paywall {override}; admin UI: a select in the
inspector header (Use domain rule / Treat as free / Treat as paywalled) + the basis
("ON (domain)" / "OFF (override)"), optimistic so the panel stays open.
Test: domain rule → paywalled in table+inspector+feed badge; 'free' → off in all
three; validation 422 + 404. 242 pytest + 11 vitest.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
New per-row "Articles" button on the Sources table expands a read-only inline
panel of the source's ACTUAL ingested articles — so the automated metrics
(paywall/image/acceptance/duplicate) can be verified against evidence instead of
trusted blind. Distinct from "Check" (which re-samples the LIVE feed for
would-pass quality); this shows what's already in the DB, which is what the table
metrics are computed from.
- Backend: GET /api/admin/sources/{id}/articles?filter=&limit=&offset= (admin,
read-only). queries.source_articles + source_articles_summary — per article:
title, url, date, accepted, reason (the "why"), topic/flavor, paywalled
(domain rule), has_image, duplicate. Summary = counts + source-level paywall
rule.
- Frontend: expandable panel with a summary header ("27 ingested · 18 accepted
· … · paywall rule: ON (domain)"), filter chips (All/Accepted/Rejected/No
image/Duplicates), compact rows with title→link + badges + reason, Load more.
So "100% paywall" or "0% images" becomes clickable evidence: open two articles
to tell a real paywall from a mis-flagged domain, or a true image gap from an
enrichment failure. Test: test_source_articles_inspector. 241 pytest + 11 vitest.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Don't trust client JSON at the storage layer:
- sanitize_game_state() runs before merge AND on the merged result (heals legacy
rows). Word Search: keep only finds whose cells actually spell a real word in
that day's grid (validated when the puzzle exists, shape-only 4-12 alpha +
cell-length otherwise), dedupe, renumber ci. Word: validate status enum, guess
count/length/alpha, colour-row shape, terminal answer/why.
- Completion is now derived from the real puzzle word count (foundWords ==
expected), not a client-sent `ms` — so stats can't be inflated by junk.
- Date validated as YYYY-MM-DD at the API (400 otherwise) — no junk/future rows.
Tests: sanitizer-rejects-junk + bad-date 400; existing tests updated to use
real-shaped data (the sanitizer is a good forcing function). 237 pytest + 11
vitest green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Two game polish items:
- Word Search: overlapping cells now multiply-blend the crossing words' colours
(deepening to a darker shade with readable text) instead of the newest colour
stomping the rest — matches the new interlocking grids.
- Cross-device game-state sync (signed-in): per-puzzle progress + stats now
follow you between devices. New game_state table; server-side merge on every
save so two devices converge regardless of push order, tailored per game:
* Word Search → UNION of finds (monotonic; can't un-find), earliest start,
best completion time.
* Word → furthest-progress wins (terminal beats in-progress; more guesses
beats fewer) — picks one device's game whole, never splices guesses.
Stats (streak/distribution/best) derived server-side from the synced states,
so they're consistent instead of per-device counters. Endpoints GET/PUT
/api/games/state + GET /api/games/stats (signed-in; size-capped). Frontend is
local-first: games paint instantly from localStorage, then reconcile in the
background; both game components push debounced on each move and adopt the
merge. Conflict handling unit-tested + an API two-device convergence test.
235→ tests + 11 vitest green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Two concrete latency wins found by measuring (server compute is 2-17ms; the time
is in the path, not the box):
- Admin panel fired its 6 API calls SEQUENTIALLY (await chain) — so it paid the
uncached origin round-trip six times back-to-back. Now one Promise.all batch.
This is the admin lag.
- /api/brief (the home "Gathering the good news…" content) wasn't edge-cached, so
a distant anonymous visitor triggered a Cloudflare→residential-origin pull.
Same global/shareable boundary as /api/feed: public s-maxage=45 when no
prefs/exclude, else private,no-store. (Needs /api/brief added to the CF cache
rule path list to take effect at the edge.)
Tests: test_brief_cache_boundary. 228 pytest + 11 vitest.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
"Gathering the good news…" waits on the home's startup API calls, which were all
DYNAMIC → a round-trip to the residential origin every load (the occasional 2-3s
linger). These responses depend only on the URL, never the session, so they're
safe to share at the edge:
- /api/moods, /api/categories (static config) → public, s-maxage=900
- /api/lanes, /api/families (global, data-derived counts) → public, s-maxage=120
- /api/feed → public, s-maxage=45 ONLY when shareable (no following / prefs /
exclude); the following feed (reads the session) and personal filters stay
private, no-store.
Hard personalization boundary, explicit per-endpoint (no blanket /api/* rule).
Pairs with a Cloudflare cache rule (added separately) making these paths
eligible. Tests assert the global endpoints are public+s-maxage and the feed
boundary (default/topic public; following/prefs/exclude private). 227 pytest.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Two small server-side tweaks so the endpoint matches the UI policy:
- Rename is refused (409) for promoted/rejected candidates — they're settled
history; the UI already hides Rename for them, now the server enforces it too.
- Name is capped at 160 chars before save, so an accidental pasted paragraph
can't wreck the queue layout.
Tests extended: 300-char name truncates to 160; renaming a promoted candidate
→ 409. 225 pytest + 11 vitest green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
A staged candidate could only be renamed by rejecting and re-adding it, which
churns the queue and discards the preview just to fix a typo. Add an inline
Rename on each candidate: a "Rename" pill swaps the name for an input
(Enter saves · Esc cancels), POST /api/admin/candidates/{id}/rename →
sources.rename_candidate(). Empty clears the name (promote then derives one
from the feed host). Preview is preserved; the fixed name carries into promotion.
Tests: test_candidate_rename (rename in place keeps preview, promotes with the
new name, gated + 404). 225 pytest + 11 vitest green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Three follow-ups from Codex's audit of the deep-preview/search/dedup work:
- Promote-time duplicate guard: promote_candidate() now re-checks
find_existing_feed() and raises DuplicateFeedError → 409, so an
old/CLI/direct-DB candidate or a race can't bypass the add-time check and
silently overwrite a live source's settings via upsert. (sources scanned
first, so a real source collision wins over the candidate matching itself.)
- postJSON/putJSON/delJSON gain opt-in {timeout} (AbortController, default
none so other calls are unchanged); deep preview uses 120s and surfaces a
calm "timed out" message instead of pinning the button on "Deep-checking…"
if the LAN model stalls.
- feed_key() now lowercases the host only, not the whole URL — paths/queries
can be case-significant; scheme/www/trailing-slash/host-case still collapse.
Tests: test_candidate_deep_preview_and_dedup extended — promote succeeds once,
then a re-promote of the same candidate is refused 409. 224 pytest + 11 vitest.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Three admin Sources upgrades:
- Deep preview: a per-candidate "🔬 Deep preview" button runs the REAL
classifier on an 8-item sample (the same model that judges live articles),
versus the fast keyword heuristic the add/Re-preview path uses. Preview now
carries `classified`, surfaced as a "model-checked" vs "quick estimate"
badge — so the acceptance % is no longer ambiguously heuristic. conn is
released during the ~30-60s model pass; postJSON has no client timeout.
- Search: free-text box over the sources table (name / category / feed URL /
homepage), folded into the existing status filter, with a live match count
and empty state. Makes "is this already added?" a glance.
- Duplicate-add guard: sources.find_existing_feed() + feed_key() normalize
scheme/www/trailing-slash/case, so re-adding a feed that's already a live
source or a queued candidate is refused with a 409 naming where it lives
(DB already enforced exact-URL uniqueness; this catches the near-miss
variants and overwrite-on-promote footgun).
Tests: test_candidate_deep_preview_and_dedup (deep flag wires the model +
uses the small sample; exact/www/slash/case variants all 409). 224 pytest +
11 vitest green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Codex's finding: cache-as-you-go would pin files Caddy deliberately serves
no-cache (version.json, manifest, word lists, icons) in the SW cache until the
next SW version — silently defeating the revalidate policy for controlled
clients. version.json is the critical one (it's how the app detects a fresh
deploy); stale word lists could drift from the server's validated answer pool.
New isMutablePath() exclusion: the SW steps aside and the browser HTTP cache
revalidates these per their headers.
Telemetry polish (also Codex): the boot beacon now fills the app_version
column with the entry chunk's hashed filename scraped from the shell's own
modulepreload link (no extra fetch) — deploy-correlated load errors become
obvious. Admin list returns + shows it.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Root cause of the intermittent white screen: the shell HTML is no-cache
(cf-cache-status: DYNAMIC), so every page-open does a synchronous round-trip
to the residential origin before any pixel renders — and the SW's network-first
navigation only fell back to the cached shell on REJECTION, never on slowness.
A stalled fetch meant staring at white with a perfectly good shell in cache.
The boot seatbelt couldn't see it either: it lives inside the HTML that hadn't
arrived yet, so slow boots left no telemetry.
- service-worker: race navigation fetch vs 2.5s grace timer. Network wins →
fresh HTML as before; timer/5xx/failure → cached shell instantly, network
response still refreshes the cache in the background. Safe due to the 14-day
immutable-chunk grace window. Caps the white screen at ~2.5s for repeat
visitors on any network.
- app.html: beacon `boot-slow: Nms (html Nms) on 4g` when mount takes >4s —
the "white screen, then it loaded" glitches finally leave a trace, with
HTML-arrival timing to separate slow-origin from slow-JS.
- admin: bot UAs (HeadlessChrome/bot/spider/crawl/…) excluded from the
headline "Load errors today" count — throttled crawlers trip the 10s boot
check routinely (the one recorded error was HeadlessChrome on X11, not a
phone). Bots stay visible in the list, tagged + dimmed.
Tests: telemetry test extended for bot flag + filtered counts. 223 pytest +
11 vitest green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Two hardening fixes from Codex's audit:
- _pick_answer falls back to the curated baseline if the live pool is empty,
so an admin tombstoning every answer in a variant can't divide-by-zero the
daily picker. Test added (test_picker_survives_empty_live_pool). Chosen over
a minimum-count block: robust without refusing legitimate removals.
- Removal copy is now honest — "Removed from future puzzles (today's answer is
already set)" — since a tombstone doesn't rewrite today's generated
daily_puzzles row. Panel intro updated to match.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Daily Word pool curation, full add/delete/import — no redeploys to fix tone:
- Remove ANY pool word, curated or admin-added, via a word_pool_removed
tombstone table. Runtime pool = (static ∪ added) − removed, so even a
baked-in word can be pulled on negative feedback. Reversible: a "Removed"
list with one-tap Restore lifts the tombstone. Lookup now surfaces a Remove
button when in-pool, Restore when removed.
- Import a vetted list (paste or .txt/.csv upload, read client-side): validates
each word (alpha · 5–6 · in guess dictionary), ignores duplicates, and reports
rejects with reasons. Re-adding/importing a removed word lifts its tombstone.
- Word Search theme delete already existed (Edit/Remove per theme) — verified.
Pool stays the clean 251/224; today's noisy LLM enrichment is discarded.
Tests: +tests/test_pool_admin.py, extended test_word_pool_admin. 222 pytest +
11 vitest green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* New "Word Search themes" panel in the Games tab: enter a theme name + words,
with live validation (4–8 letters, alpha, deduped) and a count vs the 28 needed
to fill all three sizes. An "✨ Suggest a word" button asks the LLM for one
fresh word that fits the theme. Save/edit/remove; authored themes join the daily
fallback rotation alongside the curated ones (wordsearch_themes table). The
system still handles word distribution across sizes + placement.
* Daily Word pool's added-word chips now scroll within a bounded area so the
console stays tidy as the list grows.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* client_error details, not just a count: new client_errors table + POST
/api/client-error (reason/path/user-agent/time) + GET /api/admin/client-errors.
The boot-seatbelt beacon now sends the reason + path (once per page); the admin
Overview lists the recent errors so we can tell chunk vs SW vs API vs JS — the
truth meter for the next day as the new SW propagates.
* Deploy warming now also hits the shell, routes (/play /account /admin), SW,
version.json, word lists, and icons/logo/font — not just immutable chunks.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
First games admin tool. A "Games" tab in the operator console for the Daily Word
answer pool.
* Lookup: is a word real (in the guess dictionary), the right length (5/6), and
already in the pool — instant as you type.
* Add: appends to the pool, enforcing the invariant (alpha · 5/6 letters · in the
guess dict) so the daily answer is always guessable. Remove: drops admin-added
words (curated static ones stay).
* Additions persist in a new word_pool table (survives redeploys, unlike the
baked-in JSON); the daily picker reads static pool ∪ DB additions. Guess dicts
shipped with the package (goodnews/data/words-5/6.json) for server-side
validation. Admin-gated endpoints + tests.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Per Codex audit. test_wordsearch_endpoint now asserts the exact promise — small 6,
med 9, large 13, pairwise-disjoint. app.css comment updated to .playing-game (the
class was renamed when the focused viewport was generalized to both games).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Per field feedback.
* Each day is now THREE distinct puzzles: the three sizes draw DISJOINT word
slices from a date-shuffled pool (small/med/large = 6/9/13, sum 28 unique).
Curated fallback themes expanded to 30 words each; LLM proposals accepted only
if they supply >= 28 unique words, else fall back. No more repeats across sizes.
* Word Search is now a focused game screen on mobile (same as Daily Word): body
scroll locked + footer hidden (generalized .playing-game), and the grid sizes
to the largest square that fits between the theme and the palette (container
query) — the whole puzzle is on screen, no page scroll.
* Theme placement: full "Today's theme · <name>" on the size-selection screen;
just the theme name on the puzzle itself, saving vertical space for Large.
* cosy → cozy. 🇺🇸
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Two reported bugs, same root cause: the fixed-cell grid overflowed its wrapper
on Large, so (a) the last column spilled past the border and (b) the pointer→cell
math drifted across the row, recording finds "off by a letter".
* Grid now uses 1fr columns with max-width = n·32px: the board grows with the
grid and can never overflow (shrinks to fit a narrow phone instead).
* cellAt() accounts for the grid padding/border, so selection is exact edge-to-edge.
* restore() now validates each saved find against the CURRENT grid and drops any
whose cells no longer spell the word — clears stale highlights if the day's
puzzle changed.
Codex follow-ups:
* _ws_propose now requires >= large.count + 4 valid words before accepting an LLM
proposal (else falls back to a curated theme), so a thin LLM result can't
underfill Large. Added a thin-LLM fallback test.
* Cleaned Svelte warnings: removed the now-unused .gamecard.soon CSS, added an
ARIA role/label to the grid, declared gridEl with $state. Build is warning-clean.
* Added a stale-load guard in WordSearchGame.load() so rapid size switches can't
let an older request overwrite the newer selection.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
From playtesting findings:
* Pools nearly doubled (115/104 → 228/201) with calm/neutral everyday words
(claps, dance, drench, beach…), not just strictly-upbeat ones — more variety,
~7-month runway. The post-solve "why" prompt reworded to fit neutral words.
* Word Search now stores one theme + word list per day; the grid is built per
request for three SIZE tiers — Small (8×8, 6 words), Medium (11×11, 9),
Large (14×14, 13). Large packs more words = a longer sit ("too fast" fix).
All sizes share the day's theme; every size still code-placed + solvable.
* Word Search themes can now be neutral everyday scenes ("Around the house",
"At the beach", "In the kitchen", "A walk outdoors", "Making music"…), not
only hopeful — same shape as the articles.
* Each found word gets its own colour from a calm palette, in the grid and its
word-list chip. Per-size local progress + best time.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
A calm second daily game, same philosophy as Daily Word — LLM proposes, code
disposes.
* LLM proposes a hopeful theme + ~8 words; code validates (alpha/length/dedup)
and PLACES every word in a date-seeded grid, so the puzzle is always solvable.
Curated fallback themes if the LLM is thin. Only placed words are returned;
the solution cells (placements) are never sent to the client.
* GET /api/puzzle/wordsearch → {theme, words, grid, size}. No answer to hide:
the grid and word list are meant to be seen — the play is finding them, which
the client validates by reading the selected line off the grid.
* WordSearchGame.svelte: pointer-drag selection snapped to the 8 straight
directions (mouse + touch), found-word highlighting, no-fail, no pressure
timer — time is recorded quietly and shown at the end with a personal best.
Spoiler-free share. localStorage progress (restores found cells + timer).
* Hub's Word Search card is now live with today's status; cycle pre-generates
both games with the LLM.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Per Codex's v2 hardening. The GET /api/puzzle/word response no longer carries
the answer at all — guesses POST to /api/puzzle/word/guess and the server
returns the colour pattern, computed against the day's answer. The answer (and
the "why") are revealed only once solved or the guesses are spent. This removes
the "open DevTools, read the answer" issue without pretending to be a fortress
(a deliberate crafted request can still peek; there's no leaderboard or prize,
so that's fine). Client keeps local progress/stats; dict validation stays
client-side. Trade-off accepted: each guess needs the API (the site already
depends on it for today's content).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Per Codex. Pool grown 51/44 → 115/104 hopeful answers (5/6 letter) via the
agreed workflow: LLM proposes themed candidates → code filters to the bundled
guess dictionary (length/alpha/dedup) → human spot-check prunes tone-drift
("growl", "plain", "color"…). ~3.5-month runway before repeats per variant.
test_wordpool.py locks the invariant in CI: every answer must be lowercase
alpha, correct length, unique, and present in words-5/6.json — so no future
addition can become an unguessable puzzle.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
A calm /play space — "after the brief, a small thing to enjoy." Framework-ready
for more games (Word Search next; zen/coloring later).
* Daily Word (5 letters / 6 guesses) + Long Word (6 / 7) — same Wordle mechanic,
Upbeat Bytes flavor (no "Wordle" in the UI). Hopeful answers; after solving, a
one-line "why this word matters."
* LLM proposes, code disposes: answers are picked deterministically by date-seed
from a hand-curated hopeful pool that's pre-validated ⊆ the guess dictionary
(always typeable), avoiding recent repeats; the LLM only adds the optional
"why" (with fallback). daily_puzzles(date, game, variant, payload) stores them
so everyone gets the same daily; the cycle pre-generates with the "why".
* Bundled guess dictionaries (words-5/6.json, ~12.6k/22.4k) for client-side guess
validation — never the LLM. Answer lightly obfuscated (base64) in the payload.
* Private, gentle stats (played/solved/streak, guess distribution); spoiler-free
emoji-grid share. No leaderboard, no timer, no streak-loss drama.
* Play in the bottom nav (replacing Browse, still on the lane rail) + the header.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Two calm returning-reader features.
Since-last-visit (Highlights companion, not a nav lane — per Codex):
* queries.feed gains a `since` filter; GET /api/since?ts= returns the count +
a few accepted/non-dup/visible articles discovered since the reader's last
visit (boundary-respecting; invalid/future ts → 0, no error).
* Home stores last_seen in localStorage (reads prev, then stamps now); on
Highlights, a gentle "Since you were last here, N new calm reads came in"
note with a "See what's new" reveal of a compact inline section. Dismissible.
No badges, no unread counts, no "missed" language.
PWA:
* Real PNG icons (192/512 + full-bleed maskable) rasterized from favicon.svg;
manifest fixed (azure theme to match the brand, PNG icons); apple-touch-icon.
* Minimal service worker: precache the app shell, always-fresh API + /a/ pages.
* Gentle, dismissible install banner (beforeinstallprompt → Install; iOS → the
Share → Add to Home Screen hint). Never nags.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Per Codex: generate_summary treated why_belongs alone as a complete explanation,
but get_explanation requires all three — so a partial older row (e.g. only
why_belongs) would never top up and the page would fall back forever. Now the
fully-cached check requires summary + what_happened + why_matters + why_belongs.
Test covers the partial-row top-up.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Per Codex — make /a/<id> feel like Upbeat Bytes has editorial judgment, not just
a summary wrapper. Trust-building, short, not an essay.
* article_summaries gains what_happened / why_matters / why_belongs (+ migration).
* summarize.explain_article: a separate, fallback-able LLM pass producing three
short notes (parsed from a labelled WHAT/MATTERS/BELONGS format). generate_summary
now stores them alongside the summary, and tops up older summaries on next view.
get_explanation returns them only when all three are present.
* API: share_page + /api/summary expose the explanation.
* share.py: renders the three-part section (accent rule) when complete; otherwise
the single "Why it's here" reason line is the calm fallback. The page polls and
swaps in both the summary and the section as they cache.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Per Codex — give follows an immediate payoff without making the email feel
algorithmic. The editorial brief stays the star: subject + brief count unchanged.
* followed_digest_items(conn, user_id, exclude_ids, limit=3): recent items from
the user's followed sources/tags, same accepted/non-dup/content-visible gate,
excludes anything in the brief, capped to one per source so a single follow
can't dominate. Empty → section omitted (no empty state in email).
* build_digest gains an optional `followed` list → a small "From what you follow"
section AFTER the brief, only when there are items. Item rendering factored into
shared _item_html / _item_text_lines helpers.
* send_due_digests computes the followed items per user (excluding the brief).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
thejayman77