EXAMPLES / THREE REVIEW BOUNDARIES
From proposal to permission.
These three walkthroughs show how the architecture treats action, context, and learning. All use synthetic identifiers. Each is illustrative and performs no execution.
01 / Proposed action / ILLUSTRATIVE
Review the effect before execution.
- Propose
An agent proposes changing a configuration value in demo.review. The workflow produces a diff artifact and a content hash.
- Review
A human authorizes this run, configuration-change action, exact resource scope, artifact reference, and content hash. If the artifact changes, the approval must be renewed.
- After approval
Only an appropriately bounded executor may apply the approved change. Idempotency, retry and timeout limits, an audit record, and an outcome receipt are required.
Without matching authorization, remain advisory or dry-run. A production external executor is not validated in this reference.
Inspected evidence at 7897a80: approval.schema.json; orchestrator.py; test_foundation.py.
02 / Context update / ILLUSTRATIVE
Review what the next run may know.
- Propose
A run identifies a stale source reference and proposes a new context version. It supplies provenance, applicability, and a content hash.
- Review
A reviewer checks the proposed sources, their approval status, the exact namespace, and allowed use. The proposed context stays separate from the approved version.
- After approval
After review, a subsequent run may pin the approved context version. Reuse remains namespace_only. The earlier run and its context reference remain attributable.
This sequence is illustrative. The current context schema represents these fields; it does not establish a complete automated ingestion and context-review service.
Inspected evidence at 7897a80: context.schema.json; sqlite_state.py; test_state.py.
03 / Knowledge proposal / ILLUSTRATIVE
Review learning before promotion.
- Propose
Outcome evidence supports a proposed reusable instruction. The proposal identifies source runs, evidence, applicability, and intended reuse scope.
- Review
A human reviews the knowledge proposal and the exact artifact. Local promotion checks require a matching decision for knowledge.promote and the knowledge record’s scope.
- After approval
Reviewed promotion can create approved knowledge within its authorized reuse boundary. Production activation and any broadened reuse need their own governance decisions.
Artifact-bound promotion is locally tested. Production identity, immutable audit infrastructure, and autonomous production-rule activation are not established.
Inspected evidence at 7897a80: knowledge.schema.json; sqlite_state.py; test_state.py.
EXECUTABLE / LOCAL ONLY
Move from a diagram to a local check.
The documentation includes a verified Python reference that validates a synthetic run, uses a pure handler, and rejects a parent namespace grant. It requires an existing authorized engine checkout.
Open the local reference