Skip to content

Extensions

The Keystone engineering platform includes two workload reference implementations and a standalone evaluation harness:

  • keystone-engage implements governed conversational-agent workflows.
  • keystone-counsel implements authorization-first retrieval.
  • keystone-verify evaluates compatible HTTP endpoints through profiles.

These repositories contain related engineering mechanisms, but they are separately composed implementations. They do not currently consume one composed Keystone runtime.

keystone-engage

Engage's default served path uses a single EngageOrchestrator. An optional multi-agent Coordinator is available behind configuration and is not the default route. Engage also contains severity routing, corpus-scope authorization, task-state mechanisms, OpenTelemetry instrumentation, and an unkeyed SHA-256 audit chain.

Some implemented mechanisms are outside the normal served path. NATS JetStream belongs to the optional multi-agent and event integration, and public deployment repositories do not establish it as a shared deployed service. Budget, tempo, model, token, and cost fields exist, but real cost-based selection and complete budget enforcement are not demonstrated end to end.

keystone-engage

keystone-counsel

Counsel's served path checks role and classification access, applies client-relationship constraints during retrieval, handles confidence separately, and generates locally when authorized evidence meets the configured threshold. Records excluded by the retrieval predicates are not returned through that query.

Client isolation is implemented and regression-tested, but the currently published corpus contains only global records with client_id NULL. This is not evidence of a production multi-client corpus. Counsel has no dedicated evaluation baseline retained in keystone-ledger.

keystone-counsel

keystone-verify

Verify is a standalone HTTP evaluation harness, not a workload running on the Engage or Counsel runtime. Profiles describe how to call and interpret compatible endpoints. The current CLI writes structured results and run metadata.

Verify does not own historical Ledger retention, and the current CLI should not be attributed as the producer of historical keystone-core artifacts without specific lineage evidence. A passing internal run describes the evaluated commit, cases, and configuration; it is not independent validation or evidence of general production suitability.

keystone-verify

Relationship to the architecture

The repositories provide engineering examples relevant to shared-substrate design concerns such as authorization, task state, evidence, and resource metadata. Their mechanisms have different coverage and are not one universal audit, authorization, event, or dispatch implementation.

The broader system architecture documents those boundaries. The substrate model is a research abstraction whose candidate dimensions remain hypotheses rather than a complete ontology or an implemented service stack.