新着情報

Major version bump for the AI SDK powering all readings (saju, cosmic-question, birthday posts). v2.0 breaking changes are scoped to the Interactions API (Live multi-turn conversations) which we don't use. Tested locally: generateContent returns identical shape (.text, .usageMetadata, .candidates[0].finishReason), generateContentStream iteration pattern unchanged. All call sites in gemini-client.ts work without code changes.

Patch/minor bumps: next 16.2.4→16.2.6, react/react-dom 19.2.5→19.2.6, next-intl 4.11.0→4.11.1, stripe 22.1.0→22.1.1, zod 4.4.2→4.4.3, zustand 5.0.12→5.0.13, resend 6.12.2→6.12.3, @fal-ai/client 1.10.0→1.10.1, @biomejs/biome 2.4.14→2.4.15, @types/node 25.6.0→25.6.2. Plus tailwindcss + @tailwindcss/postcss 4.2.4→4.3.0 (minor). Type-check + build green.

pnpm v10 → v11 changed the build-script approval mechanism. package.json `pnpm.onlyBuiltDependencies` extended with @google/genai, protobufjs, sharp (their native bindings need to compile on install). New pnpm-workspace.yaml carries the per-package allowBuilds map. Without this, `pnpm install` would block native modules on fresh machines / CI.

Each /[locale]/blog/[slug] page now queries the BlogPost table for all PUBLISHED locale versions of the same slug and emits them as alternates.languages in Next metadata. EN gets the x-default reference (or current locale if no EN exists). Google was treating each locale as a separate page without language relation — now they share authority across locales.

Four variant families optimized for users who already know the idol: (1) Score-breakdown hook — '{score}/100 hides 3 sub-scores that change the verdict'. (2) Hidden tension/shadow hook — 'The one tension fans never see'. (3) Timing hook — 'The exact 3 weeks per year your chart syncs with {idolName}'s peak'. (4) Inversion hook — 'What {idolName}'s chart says they'd actually need from a partner like you'. Each has 3 sub-sections (main + 2 follow-ups). EN/ES/KO/JA fully written, no English fallback.

pickTeaserIntro now wraps pickTeaserIntroDetailed which returns both the teaser variant AND a stable variantId for analytics ('idol-match-bias-v2-N'). Bias mode (ctx.type === 'idol-match' && !ctx.locked) routes to IDOL_MATCH_BIAS_INTROS; soulmate + group keep using the shared IDOL_MATCH_INTROS pool which works at 7.9% CTR. teaser_viewed events now include variant: variantId when present.

funnel-snapshot.ts emits a new "Bias-pool teaser-copy iteration" section breaking down v1 (legacy, pre-0.6.10) vs v2-N variant counts. launchkit fetch_metrics parser reads this into per-product variant metrics. The funnel_worst_volume_product insight checks for ≥50 v2 impressions and downgrades priority 1→2 with a mitigation note pointing at 7-14d data accumulation before re-flagging. Generic over product slug — works for future copy iterations on other products too.

Ref-gated useEffect in idol-match/page.tsx fires teaser_viewed with product="idol-match-group" once per locked-result session, reset when user returns to input. Closes the funnel chain — group ranking reads as one continuous chain (teaser → unlock_visible → unlock_clicked → checkout_opened → payment_succeeded) instead of breaking at hop 0. Side-effect: idol-match-bias 2.8% CTR was masquerading partly because of this gap.

Title now leads with exact-match "Saju Coreano Gratis" (was generic "Calculadora Saju Gratis"). Description front-loads "calculadora de saju coreano gratis online" + 30s value prop. Keywords expanded with all four high-CTR query variants ("saju coreano calcular" pos 4.3 30.8% CTR, "saju coreano calculadora" pos 5.7 13.1%, "saju coreano online gratis" pos 4.4 29.7%). Goal: pos 8.8 → top-3 = 2-3× clicks.

When metrics show funnel_sticky_visible ≥ 50 (sticky CTA bar deployed and gathering data) and the biggest leak is teaser→unlock_visible, priority drops from 1 to 2 and a mitigation note lists the sticky bar's visible/clicked/paid counts plus a pointer to watch sticky_bar_lift insight before rewriting teaser copy. Other projects without sticky data keep the original priority-1 behavior.

Two-line layout in the pink sticky bar — small uppercase deal phrase ("BUY 1, GET 1 FREE" / "1+1 무료" / "1+1無料" / "1+1 GRATIS") on top, "Unlock Reading" below, price badge + chevron unchanged. Tracks variant: 'v2_deal_hook' on sticky_unlock_visible and sticky_unlock_clicked so we can compare conversion vs v1 (events without variant).

New trackPaymentWebhook() helper in /api/stripe/webhook. On every payment_intent.succeeded event we write an AnalyticsEvent with flow='webhook' + paymentIntentId + currency + method + pack. Catches the ~9-event gap between Stripe (25 charges/30d) and funnel (16 events) — those were Korean redirect payments (8940 KRW = 6× try pack) where the user never returned to the page after KakaoPay/NaverPay/Toss.

When client AND webhook fire for the same PI we keep the client event (richer metadata: viaStickyBar, flow=express_top vs inline). When only webhook fires we keep that. When only client fires we keep that. Drops duplicates so the count matches Stripe exactly.

Pinned to the bottom of the viewport on mobile (md:hidden), respects iPhone safe-area-inset-bottom. Pink gradient + 3px black border + spinning sparkle icon + price badge + chevron. Visible only when: card off-screen AND user needs to pay AND no modal open AND no payment in flight. Tap → smooth scroll-into-view (block: 'center') so the user sees the full unlock card with pack picker + wallet button. Pure scroll-rescue, not a hijack.

New cameViaStickyRef in StardustUnlock flips true on sticky tap, gets included as viaStickyBar metadata on every downstream event: unlock_clicked, checkout_opened, payment_succeeded (across all 4 flows: inline, express_top, express, redirect). Express path uses a getViaStickyBar callback prop on the child components. Lets us compute true lift: % of sticky-tappers that paid vs % of all unlock-clickers that paid.

scripts/funnel-snapshot.ts now emits a 'Mobile sticky CTA funnel' section: visible/clicked counts plus 4 attribution lines (→ unlock_clicked / → checkout_opened / → payment_succeeded / baseline conv %). Launchkit fetch_metrics parser reads all 6 fields into the metrics DB. New sticky_bar_lift insight in launchkit/scripts/insights.py wired into insights_for() — auto-shows in /init with verdict (✅ positive / ⚠️ negative / ≈ neutral / 0-conv warning).

The IntersectionObserver on the unlock card was previously one-shot — fired unlock_visible once then disconnected. Now it stays mounted and continuously updates a cardInViewport state used by the sticky bar. Same observer fires unlock_visible exactly once per mount (visibleFired ref) so funnel tracking is unchanged.

Cleared 13 biome lint errors in stardust-unlock.tsx and funnel-snapshot.ts: 3× useButtonType (added type="button" to non-submit buttons), 3× useOptionalChain (replaced (x ?? {}).y with x?.y), noUnusedImports/Variables, organizeImports, useless empty fragment. No behavior change — pre-existing errors that had been deferred.

Eleven hooks (engagedRef, handlePaymentChange, loadedRef, handlePaymentReady, expressAvailable, showPaymentMethodsLocal, setShowPaymentMethods, expressVisibleFiredRef, handleExpressReady, handleExpressClick, handleExpressConfirm, handleTogglePaymentMethods) were declared after the if (phase === 'succeeded') / if (phase === 'failed') early returns. Once a payment succeeded the render skipped them, React's hook count check tripped, the page crashed.

Earlier handleTopExpressSuccess polled credits and called onUnlock() directly. Guests paid via Apple Pay never saw the email-collection prompt, so their credits were stranded on a guest cookie they'd lose. New initialPhase prop on InlinePaymentForm lets the parent open the modal pre-set to 'succeeded' — same create-account UI the card flow has always shown.

NextAuth events.signIn handler reads the saju_guest_id cookie and calls migrateGuestCredits(guestId, user.id) directly. Earlier this only ran when the unlock card later mounted and called /api/credits — if the user navigated to settings instead they saw a stale 0 balance.

The card flow has always called /api/checkout/confirm after success as a fallback for the Stripe webhook. The Express paths (top-level wallet button + modal wallet) didn't — on localhost without webhook configured, Apple Pay payments left the guest balance at 0. Both paths now fire confirm too, idempotent with the webhook in production.

The .catch(() => { /* webhook will handle as backup */ }) pattern hid every confirm failure. Replaced with explicit console.error so the dev console actually shows when /api/checkout/confirm returned a non-2xx. Three call sites updated (card flow + both express flows).

scripts/funnel-snapshot.ts now emits a 'Payment flows (split by metadata.flow)' block and a 'One-click (top-level Express) funnel' block. Launchkit's fetch_metrics parser was extended to read both — daily-check / insights now surface express_top click-rate and wallet-vs-card revenue split.

Year/month/day/hour fields used to be free-text inputs with a tiny chevron button at the right edge. On phones the chevron tap target was too small — most taps just opened the keyboard. Now the entire field is a button that opens the drawer/popover. Removed typing mode entirely.

Added min-h-[55px] to the wallet wrapper. The pulsing skeleton now occupies the same area as the future Pay button so users see a controlled loading state instead of a collapsed empty area.

Currency-picker click in the modal now sets isPackSwitching too. The wallet area shows the skeleton during the brief gap when clientSecret is wiped + new PI is being fetched.

New isPackSwitching state set true at the start of handleSwitchPack, cleared in onChecked once the new ExpressCheckoutElement reports ready. Renders a pulsing semi-transparent block over the wallet area during the gap.

Previously: setSelectedPack ran immediately after track() so React re-rendered before the fetch landed. Now: await updatePaymentIntent first, then setSelectedPack. ~200ms delay between dropdown click and chip update, but the wallet sheet now always opens with the correct amount.

Added key={`${selectedPack}-${currency}`} to <Elements>. Forces a full remount of the Stripe Element tree when the user switches pack so ExpressCheckoutElement cannot show a stale amount. fetchUpdates() approach kept as a defensive backup.

TopLevelExpressCTA now receives selectedPack as a prop and calls Stripe elements.fetchUpdates() whenever it changes (after first mount). The wallet sheet now opens with the actual selected pack price, not the cached default.

Stripe ExpressCheckoutElement replaces the Reveal button when a wallet is available on the device. PaymentIntent is pre-fetched on mount so the wallet sheet opens instantly on tap. After payment confirms, credits are polled and the reading auto-unlocks — no second click.

The deal chip (1+1 BUY 1 GET 1 FREE €1.49) is now a real <select>-style dropdown. Tap it → 3 packs slide down inside the unlock card. Picking a different pack updates the PaymentIntent in-place; the wallet button reflects the new amount immediately. Modal is now only for "I want to pay with card" cases.

New unlock-card layout: pulsing READING READY indicator + "Your Reading is Waiting" headline + subline + dropdown + Pay button + trust badges + readings counter, all inside a single pink edge-to-edge container with the brutalist black border. Loss-aversion framing ("your reading is already prepared, claim it") boosts urgency without being aggressive.

The old badge promised refunds we never issue (project rule: never refund) — that was misleading advertising and chargeback-bait. New badge "Instant Access" reflects the actual product: digital reading delivered immediately, no waiting.

FREE signup link is now the same height as the dropdown chip (40px) so primary→secondary→tertiary hierarchy reads correctly. Trust badges + readings counter moved INSIDE the pink zone right under the Pay button — payment + trust signals stay together.

While Stripe initializes (~500ms), a skeleton with the same shape as the upcoming layout shows so there's no layout shift. Held minimum 500ms even on cached loads to avoid a 50ms flash. 4s fallback timeout — if Stripe doesn't respond, gracefully shows the Pay button.

New events: express_top_visible (wallet rendered), express_top_clicked (user tapped wallet button), payment_form_loaded (Stripe finished mounting), payment_engaged (user touched the form), pack_picker_expanded, payment_methods_expanded. Funnel-snapshot.ts registers them all so launchkit picks them up automatically.

Earlier the threshold of 0.5 (50% of card in viewport) never fired for tall unlock cards on small mobile viewports — that produced unlock_clicked > unlock_visible, a logically impossible drop-off. Now threshold 0 + rootMargin fires on any pixel in view, and handleClick has a fireVisible() fallback so the invariant visible ≥ clicked always holds.

Threshold was 0.5 (50% of card in viewport) — never fired for tall unlock cards on small mobile viewports. Now threshold 0 fires on any pixel in view, and handleClick has a fireVisible() fallback so the click can never beat the observer. Without this, the funnel-snapshot script's drop-off analysis was reporting a phantom 90% leak on a logically-impossible step.

Stripe PaymentElement.onReady fires once when the form has fully mounted and is interactive. Combined with the existing checkout_opened and payment_engaged events, the 66% drop from opened-to-engaged now splits into 'didn't wait for load' vs 'saw form, didn't try'. Different fixes for each diagnosis.

Added payment_form_loaded to the event-name filter and the per-product display order in scripts/funnel-snapshot.ts. Launchkit picks it up automatically on the next /init via fetch_metrics → insights.py.

Added 308 redirect in next.config.ts. GSC will transfer the /daily index signals to /my-fortune. The route file is removed entirely so no stale page renders.

Removed the duplicate "Daily Check" tile from the free-products section on the home page and the duplicate footer link. Single CTA per concept — less decision fatigue.

scripts/credit-buyers-journey.ts, page-views-24h.ts, page-views-compare.ts, test-ranking.ts, snapshot-misc.ts, and the cover-image fixers are now in the repo so any future Claude session can re-run them without re-deriving them from scratch.

IntersectionObserver fires unlock_visible once per session when ≥50% of the unlock card enters the viewport. Combined with the existing teaser_viewed and unlock_clicked, the 16.5% click-rate on saju-profile (last 7d) splits into two diagnoses: how many people actually saw the CTA vs. how many saw it and bounced. Without this we were guessing.

Stripe PaymentElement.onChange fires payment_engaged once per checkout open with the chosen method (card, kakao_pay, paypal, etc.). Tells us if users who opened checkout actually interacted with the form. Saju-profile dropped 51 → 2 from checkout_opened to payment_succeeded last week — this event splits that into "didn't engage" (form/UX problem) vs. "engaged but didn't submit" (method/redirect problem).

scripts/funnel-snapshot.ts annotates every product-funnel step with its percentage of teaser_viewed, so the biggest absolute losses are visible at a glance instead of buried in raw counts. New scripts/funnel-7d.ts script gives a session-level rolling 7-day funnel with locale split for Saju Profile (ES vs. EN vs. KO/JA) — exactly the format we use for daily check-ins.

Every payment-related event (checkout_opened, payment_engaged, payment_submitted, payment_succeeded, payment_failed) carries the product slug so the funnel can split per-product without "unknown" buckets. Scripts ignore irrelevant noise and surface what matters per reading type.

Three more readings join the teaser experiment. Tarot weaves the spread type ("three-card spread", "Celtic Cross") and card count into a punchy intro, locked sections cover card-by-card breakdown / hidden influences / lucky timing. My Compatibility uses the partner's name and the score directly. Yearly Reading personalizes by day master, target year, and zodiac animal of the year, with a 12-month structured ToC underneath. Five rotating variants per locale (EN/ES/KO/JA), one randomized pick per visit.

The /height/[cm], /mbti/[type], and /groups/[slug] pages bring hundreds of visitors per week from Google but had effectively zero conversion into a reading because the only CTA sat at the bottom of a long idol list. New brutalist gradient block sits prominently after the H1 with two buttons — "Reveal My Saju" + "Find My Idol Match" — context-tagged with the page topic ("165cm idols", "INFJ idols", "BTS"). Tracking fires pseo_cta_clicked so the bridge effectiveness is measurable.

Kakao Pay, iDEAL, PayPal, Klarna, Amazon Pay all confirm payment by redirecting back to our page with redirect_status=succeeded in the URL. The inline Stripe flow fired payment_succeeded immediately, but the redirect-return path skipped it — so the funnel was undercounting real revenue by roughly two-thirds. Now both flows tag with `flow: "inline" | "redirect"` so the two paths stay separable in analytics.

Twelve StardustUnlock callsites (mbti-reading, tarot, palm-reading, yearly-reading, my-fortune, my-compatibility, blood-type-reading, spirit-animal, seimei-handan, cosmic-questions, soulmate-sketch, settings) had no product prop — their unlock_clicked / checkout_opened / payment_succeeded events showed up as product=unknown in analytics and were uncountable per reading type. All twelve are now tagged. Plus the idol-match teaser_viewed event is aligned with the downstream events (uses idol-match-soulmate / idol-match-bias suffix) so the funnel reads as one continuous chain per sub-mode instead of breaking at the first hop.

Quick DB query against the AnalyticsEvent table that prints event counts and per-product funnel since the last release. Removes the dashboard round-trip when checking "how is the new variant doing" — also feeds the LaunchKit cross-project analyzer (events_sync.py) for daily reports.

@google/genai 1.50.1 → 1.51.0, @stripe/stripe-js 9.3.1 → 9.4.0, zod 4.3.6 → 4.4.2, @biomejs/biome 2.4.13 → 2.4.14.

Unauthenticated visitors now see a personalized 3-section preview before the unlock CTA — section 1 fully visible (with name + day master / sun sign / matched idol baked in via slot replacement), sections 2 and 3 collapsed with a lock icon, then a compact "+ N more sections in your full reading" hint that lists the remaining ToC items. Five rotating variants per locale (EN/ES/KO/JA), one randomized pick per visit. After unlock the user gets the full personalized AI reading, so the teaser sets accurate expectations without spoiling content.

On /idol-match in soulmate mode (find your #1 match), the teaser still appears above the locked match cards but with idol-identifying slots replaced by anonymous placeholders — "your soulmate" instead of the idol's name, "the universe" for the group, "??" for the score. The text reads naturally and communicates the full reading depth without revealing which idol the user is about to unlock.

Below the ranking list, soulmate mode now also shows the same Score Card / Element Synergy / Manseryeok layout that appears post-reveal — but with the idol's photo, name, group, score, level, and pillars all blurred. Users see exactly what they get (chart side-by-side, harmony summary, score box) without seeing who.

Every step from page view to paid reading is now tracked: teaser_viewed → unlock_clicked → reading_unlocked_with_credits or checkout_opened → payment_submitted → payment_succeeded / payment_failed → reading_started → reading_completed. Each event carries product (saju-profile, horoscope, idol-match-bias, idol-match-soulmate, idol-match-group) and contextual fields like dayMaster, sunSign, idolName, score, currency, locked-state, so the funnel is segmentable per product and per variant. Events flow into both Vercel Analytics (dashboard) and our own /api/events log (DB).

Stripe used to pick the Korea PaymentMethodConfiguration (Naver Pay / Kakao Pay / Payco) whenever the saju_currency cookie was KRW — even on /en/* pages, which left non-Korean users staring at unfamiliar Korean methods and 14% of attempted payments expiring at "payment_intent_payment_attempt_expired". Routing is now driven by URL locale instead of cookie currency: /ko/* uses the Korea PMC, every other locale uses the Default PMC. Locale-mismatch payment failures should drop to near-zero.

The client used to override locale to "ko" whenever currency was KRW, which corrupted the locale field in Stripe metadata (a /en/horoscope checkout would log as locale=ko). Locale now follows URL strictly — analytics and Stripe metadata stay consistent.

The primary CTA on /saju-profile (Reveal Saju Profile) was visually outweighed by the form fields above it. Padded up to py-5 with text-xl font and a 5px brutalist shadow so it dominates the form like a real conversion button should.

Teaser intros, locked-section titles, section meta (icon + label + score), and slot-filling logic all live in /lib/teaser-intros.ts. Adding a new reading type (MBTI, Tarot, Yearly, Compatibility, …) means extending the discriminated TeaserContext union and adding 5 variants × 4 locales — no UI changes needed. The component picks up new product types automatically via the type discriminator.

@stripe/react-stripe-js 6.2.0 → 6.3.0, lucide-react 1.11.0 → 1.14.0, next-intl 4.9.1 → 4.11.0, @mediapipe/tasks-vision 0.10.34 → 0.10.35.

Bonus credits no longer dilute every pack. The €1.49 try pack stays "buy 1, get 1 free" forever, while the 5- and 15-credit packs are plain volume options. The bonus regains anchor power because the other packs don't copy it.

The "Reveal Full Reading" button used to expose a per-credit "from €0.49" anchor that read as expensive in some currencies and abstract in others. It now shows a brutalist "1+1 Buy 1, get 1 free" chip right under the title — value framing instead of unit pricing at the conversion point.

Removed the 48-hour countdown timer (a permanent deal needs no urgency) and the duplicate tilted "1+1" sticker that was competing with the try-pack badge. Single bonus marker per element now.

On Korean Won prices like ₩495, the unlock button text would visually crash into the price tag. Button content now stacks vertically with proper wrapping so labels and chips never collide, in any currency.

Replaces the scattered (and partially non-working) privacy@/dmca@/legal@/support@/contact@idolsaju.com mailto links with a single brutalist form in EN/KO/JA/ES. Topic dropdown covers general questions, privacy, DMCA, billing, bugs, and image attribution. Honeypot anti-spam baked in.

Every contact submission triggers an instant confirmation in the user's inbox — yellow IdolSaju background, white card with black borders and hard shadow, brand chip, status sticker, summary of their submission, and a CTA back to their language's homepage. Localized in 4 languages.

New /releases page (linked in the footer as "What's new") lists every IdolSaju update with category chips and dates. This very entry is the first one visible. Built with the same brutalist look as the rest of the site.

The Explore footer column now points into the previously orphaned SEO categories (~256 pages with no internal links from the global nav). Each link leads to a representative entry that cross-links to its siblings, so Google has an actual entry path.

A subtle "✦ 1,247 readings unlocked this week" line now appears under the trust badges, showing real activity from the last seven days. Hidden when the count is too low to be impressive — no fake numbers.

Pulled minor/patch updates across Stripe, Prisma, Biome, lucide-react, Cloudinary, and the Stripe JS SDKs. Stripe API version moved to 2026-04-22.dahlia. Tests and build green throughout.

Cleared an empty `/api/reading-teaser` folder and corrected our internal documentation: the teaser/preview UI is fully client-side and writes nothing to the database until a paid reading runs. No more phantom "cleanup cron" debt on the backlog.