skill-based-runtime.md

Skill-based Runtime

One-line Positioning

Skill gives AI capabilities. Yutra makes those capabilities run under rules.

Architecture (Text Diagram)

Skill / Tool / API / Function
  -> Action Adapter
  -> Yutra Action
  -> State
  -> Agent
  -> Runtime
  -> Trace / Audit / Certification

Skill and Action in Yutra

• Skill: capability unit and implementation source.
• Action: standardized invocation metadata and execution envelope.
• State: execution scene and transition boundary.
• Guard/Policy: boundary controls for risk, approval, and handoff.
• Runtime: deterministic scheduler/executor.
• Trace/Audit: evidence chain of what ran, when, and under which rule.

Skill is not a peer object of Agent/State. Skill is attached through implementation.type = "skill" on an Action.

Current Support

• Local skill manifest (skill.json) loading and validation.
• Local skill registry and discovery.
• CLI: yutra skill list, yutra skill validate, yutra skill inspect.
• Skill to Action adapter (skillToAction, skillToActionRegistry).
• Runtime execution for implementation.type = "skill".
• Ecommerce support pack skill-based variant (agent.skill.yutra.yaml).

Current Non-support

• Remote registry.
• Marketplace.
• Install workflow.
• Sandbox runtime isolation.
• Skill store.
• Cloud execution.

Why This Milestone Matters

Generic skill calling can invoke capability. Yutra provides governed execution evidence:

• which state called the skill,
• which policy metadata applied (riskLevel, requiresApproval, sideEffect),
• and what trace/audit proof was generated.