A small control layer for capabilities that must earn execution.
Contracts, evaluations, permissions, policy gates, runtime boundaries and evidence are treated as first-class parts of the capability lifecycle.
Capability lifecycle
The V0.1 implementation proves one narrow lifecycle before expanding into team workflows or a hosted control plane.
Skill package
├── skill.yaml
├── instructions.md
├── permissions.json
└── evals.json
Contract
Identity, version, goal, runtime, inputs, outputs and required resources.
Blocking evaluations
Required checks can stop execution before a runtime is invoked.
Permissions
Network and shell access are deny-by-default; approved actions are explicit rather than inferred from model output.
Evidence
Each run records status, checks, timestamps and concise evidence for later inspection.
Runtime model
The current publicly verified runtime is deterministic MOCK. A Claude adapter exists in the private prototype as the planned first live provider path, but live Claude API execution is not claimed yet.
Public repository
The public repository exposes architecture, roadmap, status, security notes and a safe example skill without publishing private implementation code or credentials.