Capabilities
A capability is a stable contract for one kind of work. The contract describes portable inputs and outputs. A live implementation enters later through explicit host composition, a capability adapter, or the runtime integration selected by your application.
This separation lets application code depend on what a component does without depending on the provider or framework that performs it.
| Import | Responsibility |
|---|---|
@geekist/llm-core/contracts | Identity, versioning, schemas, resources, evidence references, and invocation context |
@geekist/llm-core/model | Model requests, responses, profiles, content, and media |
@geekist/llm-core/tools | Tool specifications, bindings, strict argument validation, and action digests |
@geekist/llm-core/control | Policy, approval, cancellation, concurrency, and controlled tool execution |
@geekist/llm-core/evidence | Redacted execution events and storage-neutral receipt journals |
@geekist/llm-core/state | State lifetimes, intervention contracts, and resume compatibility |
@geekist/llm-core/context | Scoped, budgeted context manifests |
@geekist/llm-core/artifacts | Portable output identity and provenance |
@geekist/llm-core/evaluation | Evidence-bound cases, evaluators, and results |
@geekist/llm-core/agent | Portable agent intent and normalized execution facts |
The capability pages follow the same path you use in an application:
- Define portable contracts.
- Supply qualified live implementations.
- Supply intent and capabilities to a qualified runtime integration.
- Preserve redacted evidence and explicit state.
Retrieval, indexing, storage, memory, and media are independent capability fronts at /retrieval, /indexing, /storage, /memory, and /media. /model owns provider-neutral model requests, responses, profiles, and content contracts. None of these fronts selects a hosted service, executes an agent loop, or exposes provider-native clients.