Jev → v0 model routing POC

Demonstrates routing v0 Platform API chats to the right model using Jev (TypeSafe System One models) as a per-chat classifier, plus Vercel Organizations provisioning. Research background: jev-v0-model-routing.md.

Core design decisions

• Route once per chat. Jev classifies the first prompt; the chosen model is fixed for the chat's lifetime. Switching models mid-chat breaks v0 prompt caching (cache read/write was 86% of measured customer spend).
• Fail-safe default is v0-pro — the customer's status-quo model — when confidence is low or Jev is unavailable.
• v0-max-fast is never auto-selected; latency preference is a user choice, exposed through the manual model picker.
• Escalation is one-way and manual ("Escalate to max"), priced as one cold cache write.

Layout

• lib/jev/ — AI Gateway TypeSafe-compatible client and question definitions
• lib/routing/policy.ts — the routing policy (versioned, plain module)
• lib/v0/chats.ts — v0 Platform API chat operations with routing metadata
• lib/orgs/ — Organizations API provisioning (child team + v0 key)
• app/api/ — route handlers (/api/chats, /api/chats/[chatId]/{resume,messages}, /api/orgs/teams)
• app/page.tsx — demo UI: chat, decision panel, orgs provisioning
• data/ — gitignored; holds tenant-keys.json for provisioned teams and the usage export CSV for the savings replay (pending)

Running

cp .env.example .env.local   # fill in the keys
npm run dev

Requires: AI_GATEWAY_API_KEY (Jev via AI Gateway), ROOT_ORGS_ID, VERCEL_ORGS_TOKEN (org access), ROOT_V0_KEY (v0 key for the root org / default team), DEFAULT_TEAM_ID.

Verification status

• Build and lint must pass (run npm run build).
• Live behaviors (Jev classification, chat creation under the commercetools team, orgs provisioning, cache-cost spike) require the real keys and are not yet verified.