Architecture

System layers

flowchart TB
  subgraph clients [Clients]
    AG[Agent / Orchestrator]
    ED[Editor MCP]
    UI[Console UI]
  end
  subgraph runtime [Runtime Phase 1-2]
    MCP[MCP Server]
    API[REST API]
    SDK[Python / TS SDK]
    RTR[Skill Router]
    PRIM[Context Primitives]
  end
  subgraph content [Content]
    SK[skills/ SKILL.md]
    REF[references/]
  end
  subgraph control [Control Plane Phase 3]
    IAM[IAM / RBAC]
    TEN[Tenancy]
    POL[OPA Policy]
    AUD[Audit]
  end
  AG --> MCP
  AG --> API
  ED --> MCP
  UI --> API
  MCP --> RTR
  API --> RTR
  SDK --> RTR
  RTR --> SK
  PRIM --> SK
  API --> IAM
  API --> TEN
  API --> POL
  API --> AUD

Progressive disclosure in the loader

1. Index — list_skills() reads frontmatter only (cheap).
2. Body — get_skill(name) loads SKILL.md.
3. References — get_reference(name, path) loads on demand.

The Skill Router builds a lexical index from the same files (offline).

Single implementation rule

MCP, REST, SDKs, and demos call one loader/router/primitives implementation in runtime/core/context_skills/. No duplicated business logic in surfaces.

Deployment modes (Phase 3)

Mode	Description
SaaS	Shared multi-tenant cluster
Customer VPC	Terraform module + dedicated region
Air-gapped	Offline bundle, embedded OPA, file vault

See Enterprise deployment and deploy/.

Repository layout

skills/           # 28 Agent Skills (content)
runtime/          # MCP, REST, SDKs, primitives
researcher/       # Gates, benchmarks, corpus
control-plane/    # IAM, tenancy, policy, console
deploy/           # Helm, Terraform, air-gapped
docs/             # This documentation site
examples/demo/    # Runnable proof scripts