burn to tweet

Burn conviction. Tweet as clawd.

Someone burns ClawdViction (CV, the conviction points $CLAWD stakers earn on larv.ai) to open a session with clawd. They get 12 turns with Claude Opus 5.5 to shape one tweet, up to 4 gpt-image images, and up to 2 videos (Grok Imagine, or their own upload). When they're happy they press tweet. It passes a safety review and posts from @clawdbotatg. They can also walk away; the burn is final either way.

The price

A Dutch auction that restarts on every tweet:

• When a tweet posts, the price resets to 10% of the largest CV balance in larv.ai's ledger (GET larv.ai/api/cv/highest).
• It decays exponentially and lands on 50M CV after 24 hours, then rests there until someone tweets.
• Buying a session doesn't reset it. Only a posted tweet does. Parallel buyers are fine, because the price is the throttle.

lib/price.ts is the whole curve; test/price.test.mjs pins it.

Architecture

browser ──► Next.js on Vercel ──► larv.ai's Neon Postgres (schema btt)
              │  /api/session  burn CV via larv.ai /api/cv/spend
              │  /api/worker/* job queue (bearer WORKER_SECRET)
              ▼
        worker on the Mac (outbound-only poller, launchd)
              ├─ turn  → claude -p (Opus 5.5), fully isolated → reply + draft
              ├─ image → reviewer checks prompt → gpt-image → JPEG
              ├─ video → (prompt check → Grok) → ffmpeg re-encode → reviewer
              │          sees frames + Whisper transcript → mp4 to Vercel Blob
              └─ post  → guard + reviewer on exact text → X API → price resets

• Why a Mac worker: clawd runs as claude -p on the subscription, which only exists on this machine. The worker only makes outbound calls, so nothing is exposed.
• The isolated agent (worker/claude.mjs) has no tools (--tools ""), no MCP, no settings, no CLAUDE.md, no memory and no session persistence. Its env is built from scratch, so the worker's Twitter, OpenAI and worker keys never reach it, and it runs from an empty scratch dir. Its only memory is agent/CLAWD.md: who clawd is, the voice, the ecosystem, and the hard lines. One residual: Claude Code injects the logged-in account's email into context, and no switch turns that off. The persona tells clawd never to repeat it, and a live jailbreak test didn't leak it.
• Safety, twice: worker/guard.mjs applies deterministic rules: 280 weighted chars, no hashtags, links only to allowlisted domains, no addresses except $CLAWD's, no phishing bait. worker/safety.mjs is an independent reviewer (agent/REVIEWER.md) that sees only the final text. It fails closed. Image prompts are reviewed before generation too.
• Video: too big for a function body, so the browser uploads straight to Vercel Blob (app/api/session/[id]/video issues a one-file token) and the worker does the rest (worker/video.mjs). Every video is re-encoded to ≤1280px H.264 with no metadata. That re-encode is what the reviewer sees and exactly what posts: its sha256 is pinned on the session and re-checked at post time. Generation needs XAI_API_KEY in the worker env; without it, upload still works.
• A session posts at most once: worker/state/posted.json is the worker's ledger, and results retry until they land.

Run locally

npm install && (cd worker && npm install)
# site: needs DATABASE_URL (the btt role), CV_SPEND_SECRET, WORKER_SECRET
DEV_FAKE_SPEND=1 npm run dev                                # fake spend works only in dev builds
# worker:
cp worker/.env.example worker/.env                          # API_BASE=http://localhost:3000
DRY_RUN_POST=1 node worker/worker.mjs                       # never posts to X
npm test

Deploy: Vercel (zero config) with env DATABASE_URL, CV_SPEND_SECRET, WORKER_SECRET. Tables: DATABASE_URL=… node tools/migrate.mjs. Worker: tools/install-worker.sh.

Storage

State lives in larv.ai's Neon Postgres, in its own btt schema, reached as a btt role that has no access to larv.ai's tables. That role can't read or change CV balances; spends still go through larv.ai's API.