cel-contracts

Shared Rust contracts for AI agent actions, planning views, execution receipts, and runtime capabilities.

cel-contracts contains pure data types for the boundary between a planner and the runtime that carries out its actions. It is intentionally small: no LLM clients, no runtime loop, no storage backend, no policy engine.

What It Owns

• PlannedAction — what a planner asks a runtime to do.
• PlanningView — budgeted context shape a planner can consume.
• ExecutionReceipt — what the runtime actually dispatched and observed.
• DispatchRoute — adapter, CDP, accessibility, native input, focus, or other.
• ObservedEffect — whether the expected effect was observed, timed out, contradicted, or not checked.
• RuntimeCaps and related planning/run contracts.

Why It Exists

Agent frameworks and runtimes need a shared language at the boundary:

planner → PlannedAction → runtime → ExecutionReceipt

Any runtime can implement this boundary while keeping the schema open, inspectable, and testable.

Receipt Boundary

ExecutionReceipt proves dispatch and observed effect. It does not prove the whole user goal is complete. Completion still requires post-state evidence such as adapter readback, CDP/AX state, a fresh context snapshot, screenshot, or external system confirmation.

License

Apache-2.0