State, recovery and improvement

Client-managed state

Keep the adopted skill version and digest, goal, source and target pins, approved baseline, completed slices, outstanding operations, evidence pointers, blockers and next safe action in a project-scoped checkpoint. Store minimal data or references and never credentials.

Before resuming, validate the checkpoint schema, source identity, target revisions and any unknown outcome. A local save and a remote MCP commit are not atomic. Reconcile the original durable request before retrying an effect whose result is unknown.

Feedback and evaluated lessons

Separate raw observations, candidate lessons and accepted improvements. Record evidence, applicability, counterexamples, privacy classification and the decision that accepted or rejected a lesson. Do not turn one customer's transcript or one successful slice into a global rule.

Hooks and runners

A supported hook should capture only a bounded redacted event or review request and a consistent checkpoint. A separate supported runner may claim the work, load the pinned skill and state, evaluate evidence and publish a result. Verify the host's current hook and scheduling interfaces before configuration. If no qualified persistence, hook or runner exists, keep the workflow manual and report the limitation; installation does not activate these components.