01
Conversation, Workbench, configuration and authorization experiences. The Knowledge Agent (KA) is the conversational layer that turns human intent into governed operations.
ARCHITECTURE / TWO RUNTIMES · ONE SOURCE OF TRUTH
Cap2Layer deliberately separates the server that owns product rules from the server that renders the user experience.
SYSTEM BLUEPRINT / LIVE BOUNDARIES
Conversation, Workbench, configuration and authorization experiences. The Knowledge Agent (KA) is the conversational layer that turns human intent into governed operations.
Policy, identity, orchestration and the single source of operational truth.
01 Product meaning flows inward 02 Infrastructure remains replaceable at the edges
THE TWO-SERVER MODEL
Owns authentication, workspace/project scope, KA semantics, Workbench read models, preview/apply policy, MCP transport, OAuth, vault references, persistence profiles, LLM readiness and every persisted product rule.
Owns the interaction surface: KA conversation, Workbench navigation, configuration dialogues, provider setup and human authorization experiences.
HEXAGONAL LAYERS
Hosted MCP, HTTP/UI, OAuth and future clients
Transport adapters enter through explicit contracts.Records, cuts, certified prompts and result certification. A Work Item is a unit of work with its scope and definition of done. A Step is an executable part of a Work Item. A cut is one bounded execution of a Step, on its own branch, limited to declared paths. A certified prompt is the exact instruction the AI programmer executes, bound to the authorized scope. Queueing for parallel work is in progress. A lease (in progress) is a reservation that will let parallel cuts avoid colliding.
Coordinates behavior without knowing engines or transports.Records, identity, authorization, profiles, execution semantics and policy
Pure product meaning, independent from frameworks.Persistence, LLM inference, vault, provider and repository capabilities
Stable contracts point outward without importing infrastructure.Supabase/Postgres, OpenAI-compatible inference, hosted MCP and provider integrations
Concrete engines implement ports and remain replaceable.REPLACEABLE EDGES
Future: direct Postgres, Aurora, SQLite/local, object storage, Elasticsearch, Notion or Airtable through the same boundary.
Customer-configured providers, model compatibility probes, worker roles and safe credential references—no single hardcoded LLM.
Web, IDE, mobile, REST, gRPC or approval clients can call the same application services through new interface adapters.
WHY IT MATTERS
Domain and application never depend on Supabase, FastAPI, Next.js or a specific LLM SDK.
Multiple provider profiles are explicit; no silent fallback when several choices exist.
Secrets remain behind vault references and never enter prompts, UI payloads or logs.
Boundary violations are treated as product bugs, not stylistic disagreements.