NodeBenchAI: Agent for Claude Code

Deploy & Launch Agent

You are deploying NodeBench AI to production and verifying everything works end-to-end. This is the final gate before sharing with real users.

Prerequisites Check

Before deploying, verify these exist:

# Check env vars are set
cat .env.local | grep -c "CONVEX_DEPLOYMENT\|VITE_CONVEX_URL\|OPENAI_API_KEY"
# Should be >= 2 (CONVEX_DEPLOYMENT + VITE_CONVEX_URL minimum, OPENAI_API_KEY for voice)

# Check Convex CLI is available
npx convex --version

# Check Vercel CLI is available
npx vercel --version

# Check build passes
npx tsc --noEmit --pretty false
npx vite build

If any prerequisite fails, fix it before proceeding.

Phase 1: Deploy Convex Backend

# Type-check Convex functions
npx convex dev --once --typecheck=enable

# Deploy to production
npx convex deploy -y --typecheck=enable

Verify after deploy:

• No deployment errors
• Schema migrations applied (new tables: mcpApiKeys, mcpGatewaySessions, trajectorySpans, etc.)
• Functions are live (check Convex dashboard)

If schema migration fails:

• Read the error — it will tell you which table/index conflicts
• Common fix: npx convex deploy -y --typecheck=enable with --force if schema is additive-only
• Never force if tables are being deleted or renamed

Phase 2: Deploy Frontend to Vercel

Canonical path: git-based deploy. Push to main triggers Vercel automatically. Branch protection requires CI green (Typecheck + Runtime smoke + Build) and enforce_admins: true prevents bypass — even admin merges wait for CI.

# Canonical deploy (use this)
git push origin main   # auto-deploys via Vercel webhook + GHA backup

Backup paths (only when webhook is broken):

# Manual prebuilt deploy (skips server-side npm install — useful when
# Windows-locked package-lock breaks the linux-x64 sharp binary)
vercel build --prod --yes
vercel deploy --prebuilt --prod --yes

# Last-resort full upload (avoid — uses your CWD's package-lock,
# which breaks sharp on Linux build machines)
npx vercel deploy --prod