Concepts
| Term | Meaning |
|---|---|
| Capability package | A directory containing capability.yaml, schemas, Provider code, and optional configuration assets. |
| Provider | A Python object that implements the VCP lifecycle and Actions. |
| Action | A versioned, schema-defined operation declared in the Capability manifest. |
| Manifest | Portable declaration of identity, compatibility, resources, configuration, dependencies, secrets, and Actions. |
| Backend adapter | Capability-owned code that isolates vendor or robot APIs from Provider logic. |
Why Capabilities exist
Capabilities let a team add one domain-specific function without modifying Vega Runtime. A speech Capability owns its audio integration; a vision Capability owns its camera and inference dependencies; an HTTP Capability owns its service client. Each package is versioned, tested, deployed, and upgraded independently.
Responsibilities and boundaries
- Declare what each Action can, cannot, requires, guarantees, and limits.
- Validate integration-specific preconditions and use injected context services.
- Keep vendor dependencies inside the Capability package and report physical uncertainty accurately.
- MUST NOT import vega.* Runtime internals.
- SHOULD NOT make Runtime scheduling, mission retry, resource arbitration, or recovery decisions.
- MUST NOT claim an external or physical effect it cannot establish.
Relationship with Runtime, SDK, and VCP
The Runtime validates, loads, initializes, invokes, and shuts down a Provider. The SDK is the current Python binding used by Provider code. VCP defines the public manifest and Provider contract surface. Current VCP has no remote transport; the Runtime invokes loaded Python Providers through the SDK contract.
Lifecycle and workflow
- Define a narrow outcome and its operational boundaries.
- Design input, output, and optional progress schemas.
- Create the manifest and package structure.
- Implement a Provider with a separate backend adapter.
- Validate, test, package, and deploy in controlled stages.
Example
The repository's unitree_go2_speak package implements speech.speak. It declares a Unitree Go2 compatibility boundary, an exclusive speaker resource, non-idempotent execution, progress support, and no restart recovery. Its Provider delegates work to a backend adapter and accurately reports uncertain physical termination for the Go2 bridge.