The 7 principles

The Capsule architecture is a way of operating one AI coding agent across many domains. These seven principles define it. The scaffolder generates a repo that embodies them, but the pattern is tool-agnostic — you can follow it by hand.

1. One agent, many domains. One surface routes the agent across all your contexts instead of re-explaining context per project.
2. Thin router, rich capsules. The entrypoint stays minimal (startup, transversal rules, routing, hierarchy); detail lives in composable capsules.
3. On-demand context. Load only the active domain’s capsule and its recent state; don’t front-load everything (focus and token efficiency).
4. Safety gate by default. Plan → review → approve → execute. The agent never executes a plan alone. (Security default, non-negotiable.)
5. Confidentiality isolation. Data never leaks across domains (work ↔ personal ↔ client). (Security default.)
6. Persistent, indexed memory. Facts are recorded as durable, typed memory with a lightweight index loaded each session; the agent learns across sessions.
7. Asymmetric governance + explicit hierarchy. Constitutional files change via ritual (propose → approve → changelog); operational files change directly. An explicit context hierarchy resolves conflicts (current instruction > domain capsule > router > memory).

Principles 4 and 5 are security defaults and ship active in every generated command center. The rest are structure.