Laterlist

Joon's private, single-user tool: connect his YouTube Watch Later, let AI sort the backlog into five rows, then clean it up.

The original localhost app stays untouched as Joon's personal daily driver. Design spec: docs/superpowers/specs/.

Run locally (zero accounts)

npm install
DEV_FAKE_AUTH=1 FAKE_LLM=1 npm start     # http://localhost:4400

Dev mode uses an embedded PGlite database (./dev-pgdata), fake sign-in (any email), and a deterministic fake classifier. The whole product flow works end to end without Supabase or Anthropic credentials.

npm test                                  # vitest — DB tests run on PGlite (real Postgres in WASM)

Architecture

• server/ — Express API + in-process classification worker. index.js is the composition root; everything is factory-injected.
  • db.js over pg.Pool (Supabase) or PGlite (tests/dev) via one query() interface
  • auth.js verifies Supabase JWTs (JWKS or legacy HS256), JIT-provisions users
  • jev.js — default sorting engine: Jev (TypeSafe) via OpenRouter's decisions API, one metadata-only request per video in a parallel pool, no batching needed
  • classify.js + anthropic.js — the shared sorting rules, plus the Haiku fallback (CLASSIFIER=haiku): structured outputs, Batches API over 500 videos
  • worker.js — claims jobs (FOR UPDATE SKIP LOCKED), chunked processing, crash re-adoption, budget kill switch
  • background.js — on Vercel, sorting runs after the response (waitUntil) and hands off to a fresh invocation via /api/cron/advance, so a sort keeps going with the tab closed. Needs CRON_SECRET and APP_URL (the production domain) set; polls only restart a job whose run died
• collector/ — the browser-side Watch Later collector (scroll + Polymer .data harvest + DOM pruning; validated on a real 2,796-video playlist at 100% of available items). Served fresh at GET /collector.js for the console snippet; bundled into the future MV3 extension (Chrome forbids remote code).
• web/ — Vite multi-page: sign-in landing at /, React app at /app.

Production status

Live at https://watch-later-web.vercel.app (Vercel project watch-later-web, Supabase project fjykuzbwhrpzpiqxzjeh, schema applied). The private deployment requires owner access and the existing service credentials:

1. Database password (unblocks the whole API): Supabase dashboard → project watch-later-web → Settings → Database → Reset database password. Then:

   printf 'postgresql://postgres.fjykuzbwhrpzpiqxzjeh:<PASSWORD>@aws-0-us-west-1.pooler.supabase.com:6543/postgres' \
     | vercel env add DATABASE_URL production
   vercel deploy --prod --yes
   

2. Google sign-in: Google Cloud Console → APIs & Services → Credentials → Create OAuth client ID (Web application) with redirect URI https://fjykuzbwhrpzpiqxzjeh.supabase.co/auth/v1/callback. Paste the client ID + secret into Supabase → Authentication → Providers → Google. In Supabase → Authentication → URL Configuration set Site URL https://watch-later-web.vercel.app and add https://watch-later-web.vercel.app/app as a redirect URL.

3. Sorting engine: create a key at openrouter.ai (add credits), then:

   printf 'sk-or-v1-...' | vercel env add OPENROUTER_API_KEY production
   vercel deploy --prod --yes
   

   ANTHROPIC_API_KEY stays for summaries, and sorting falls back to Haiku if the OpenRouter key is missing. CLASSIFIER=haiku forces the fallback.

Set OWNER_EMAILS to Joon's sign-in email (comma-separated if needed). It falls back to ADMIN_EMAILS when unset. If both are empty in production, every sign-in is refused. LIBRARY_CAP is the safety limit for stored videos, default 25,000.

Production setup (one-time)

1. Supabase: create a project → Auth → enable Google provider (needs a GCP OAuth client with redirect https://<project>.supabase.co/auth/v1/callback). Grab the URL, anon key, and the pooled DATABASE_URL.
2. Anthropic: create an API key at console.anthropic.com.
3. Railway: new project from this repo; set every env var from .env.example (including VITE_SUPABASE_URL + VITE_SUPABASE_ANON_KEY, which bake into the frontend at build time). railway.json handles build/migrate/start.

Cost model

Jev, metadata-only, one request per video: ≈ $0.04 per 1,000 videos (measured 2026-09-30). (Haiku fallback: ≈ $0.40–0.77 per 1,000, half that via Batches.) BUDGET_USD kill switch caps total spend; admins get /api/admin/stats.