Architecture

Rooster is a pnpm monorepo of @rooster/* packages with one deployable app.

Packages

Package	Responsibility
config	Env loading + validation (zod); DB-driver + platform selection
schema	zod domain entities, enums, DTOs, ids — the validation source of truth
db	Drizzle schema (SQLite + Postgres), migrations, repositories, drivers, seed
core	Domain services, permissions, status transitions, audit, onboarding
auth	better-auth config, the identity bridge, enrollment gating, scopes
mcp	MCP server: tools + resources + the stateless Streamable-HTTP transport
server (app)	The Hono deployable: auth + /mcp + discovery + /onboard

Dependency direction (never cycles): config ← db ← core ← {auth, mcp} ← server. schema is depended on by everyone. Core is transport-agnostic — it depends only on db + schema.

Request → action flow

1. An agent sends Authorization: Bearer … to POST /mcp.
2. The server resolves the token (auth.resolveMcpIdentity) → maps the OAuth clientId to its bound Agent → an ActorIdentity (effective scopes = token grant ∩ agent allowance).
3. core.resolveActor() loads the principal and computes the effective org role from memberships → an Actor.
4. A per-request MCP server is built bound to that Actor; a tool calls a core service.
5. Every mutation: authorize (role floor and, for agents, token scope) → validate input with the schema DTO → enforce orgId scoping in the repository → write → append an audit record attributed to the trusted principal.

Database portability

One repository implementation serves both dialects because the two Drizzle schemas are kept structurally identical: ids and timestamps are TEXT (app-generated UUIDs, ISO-8601 strings), arrays/JSON are TEXT (JSON-encoded), and booleans are normalized. The Postgres driver bridges its instance into the libSQL-typed repository factory at a single boundary.