Architecture

┌──────────────┐   typed RPC + SSE   ┌──────────────┐   Prisma / pg   ┌──────────┐
│ packages/cli │ ──────────────────► │   packages/  │ ──────────────► │ Postgres │
│  OpenTUI TUI │ ◄────────────────── │    server    │                 └──────────┘
│  local tools │   streamText →      │    (Hono)    │ ──────────────► OpenRouter,
└──────────────┘   tool calls       └──────────────┘                 Clerk, Polar

Packages

• packages/cli (@coolcode/cli) — terminal UI via @opentui/react + react-router (memory router). Talks to the server with a typed Hono RPC client (hc<AppType> from @coolcode/server, used only for types) plus DefaultChatTransport for chat streaming. Auth token injection lives in src/lib/api-client.ts.
• packages/server (@coolcode/server) — Hono app exporting AppType. Routes: auth (Clerk), billing (Polar), sessions, chat. All chat/session/billing routes sit behind requireAuth; chat additionally runs requireCreditsBalance, and usage is billed to Polar per message in the chat onFinish handler.
• packages/database (@coolcode/database) — Prisma 7 + @prisma/adapter-pg. Import db from @coolcode/database/client, not the root export. The generated client lives at packages/database/generated/prisma and is gitignored.
• packages/shared (@coolcode/shared) — model registry, modeSchema (BUILD/PLAN), and AI tool contracts (getToolContracts(mode)) shared by the server and the CLI.

Deliberate version split

The CLI pins ai@6 + @ai-sdk/react@3, while the server and shared packages use ai@7. Types cross the boundary through @coolcode/shared contracts. Do not “align” these versions casually.

Server notes

• The server listens on port 3000 (hardcoded).
• idleTimeout: 255 is deliberately high so long LLM and tool runs are not cut off — do not lower it.

Development commands

bun run dev:server   # Hono API on :3000
bun run dev:cli      # TUI with watch
bun run build:cli    # build to packages/cli/dist
bun run link:cli     # build:cli, then bun link