Jina Reader returned 404/garbage for many news sites. Now fetch the page
HTML directly, parse og:title/og:description + author + article<time>
+ <p> paragraphs; Jina only as fallback. Verified: theverge AI page →
real title 'Artificial Intelligence', 8 paragraphs.
- New scripts/source-fetch.mjs: fetch tweet/thread via fxtwitter API +
Jina fallback; news via r.jina.ai (title/author/excerpt/paragraphs)
- New src/remotion/cards.tsx: TweetCard, ThreadCard, NewsCard — mono,
flat, white cards w/ avatar initial, handle, text, stats (views/likes/
replies/retweets) or headline+excerpt; ThreadCard stacks tweets
- ShortsVideo: new scene types tweet/thread/news; source card scene
REPLACES the story segment's visual (duration kept identical so the
audio/video mux stays in sync) — the voiceover keeps playing while the
actual tweet/news is shown on screen
- Worker: fetch sourceUrl when present → sourceToCard() → inject card
- Schema: YoutubeScript.sourceUrl + sourceData (Json)
- UI: approved scripts get a '🔗 Tweet/News URL' input; Render Shorts
saves sourceUrl then queues the render
- Verified: card scene renders (bright%≈20.8 vs caption 3.9 in pixel
histogram); TSC clean; frame extraction OK
User directive: no gradient, no purple (consistent with GMW black-white
monochrome theme). Full redesign of parts.tsx:
- Solid near-black scene backgrounds (5 subtle variants: circle, plus,
barcode, rings, diagonal motifs as faint white SVG line art)
- GlassCard → flat white card (black text) — high contrast inversion
- CaptionKaraoke: black on white card, active sentence full-opacity
- KineticTitle: white on black; accent word = inverted white block w/ black ink
- StatCounter: big number framed by 2 offset rectangle outlines (editorial)
- BrandTag: white outline pill (flat), ProgressBar: solid white fill
- Removed ALL color props/palettes; removed gradient backgrounds/orbs
Perf: lighter than v2 (7.1MB vs 10.3MB); ~2min render per 32s short
- Fix critical bug: scenes were stacked without <Sequence> — every segment
rendered simultaneously, last scene covering the whole video (root cause
of 'too simple/ugly' visual)
- Scene variety: kinetic word-by-word title (hook), karaoke caption card,
big-number stat scene (auto-detected: $12B, 65%, 3.2M), CTA center
- Karaoke captions synced to TTS: edge-tts SentenceBoundary timing via new
scripts/tts-words.py (mp3 + sentences.json per segment); active sentence
bright, others dimmed
- Rich background per scene: 4 color palettes (indigo/pink, cyan/teal,
purple/pink, orange/red), drifting orbs, dot grid, ken-burns motion
- Real fonts: @fontsource Inter 700 + Montserrat 800 (display/captions)
- Glass card captions, progress bar per-scene, safe zone 900x1400
- Word-boundary truncation for stat labels (no mid-word cuts)
- Perf kept: no blur/SVG-noise (software renderer); ~2 min per 32s short
- Committed SVG hero (Calm Luxury: near-black ground, blue accent,
Postgres-bus architecture motif) rendered inline on GitHub
- Shields badges for the stack
- Native [!IMPORTANT] callout at the top so a Hermes handed only the
repo URL knows to clone, read ONBOARDING.md, and guide the install
- "Two ways to install" section (agent-driven vs by hand)
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
A Next.js 16 + Prisma/Postgres dashboard that pairs with a local Hermes
agent over a Postgres message bus: dispatch work, approve side-effecting
actions, browse the agent's memory, and watch it run. Ships with an
agent-onboarding prompt (ONBOARDING.md) so your Hermes can install it for
you step by step. All secrets are env-configured; nothing sensitive is
bundled.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
asepharyana
MythEclipse
MythEclipse
asepharyana
sharbelxyzandClaude Opus 4.8