Native automation policies and event producers

Decision

The native SQLite scheduler owns durable failure streaks, cooldown decisions, delayed occurrences, supervised source generations and event receipts. Provider execution remains bound to the saved workspace member/account and live authority. Process/stream events are accepted only from the in-process supervisor's persisted receipts. There is no browser or public callback endpoint for creating events.

Source commands use the existing Bubblewrap launcher, a credential-free home, workspace-only working directory, and the existing public-egress/private-network boundary. Revocation and maintenance stop those children. Unknown process outcomes are retained for review. Streams use bounded buffering and queues, with a visible hold on overflow. Light-context AI runs use a fresh authentication-only home and explicit skill selections and their declared supporting packages; workspace instructions continue to apply. Supporting skill names resolve through the same member/scope-filtered catalog, enabled-state checks and immutable snapshot path. Declarations cannot grant filesystem access or enable a disabled package. See Skill dependencies.

PostgreSQL connector messages reference notification events. Stable handoff keys prevent duplicated sends across crashes, and external delivery rechecks current subscriptions immediately before provider admission. Provider uncertainty stays in the connector outbox. No legacy application-mail or direct SMS fallback remains.

Compatibility and release

SQLite schema 8 and PostgreSQL migration 25 add state without rewriting historical receipts, subscriptions, credentials, or imported definitions. Held jobs require explicit review. Unsupported legacy policies remain held rather than being silently dropped. Operator releases do not certify automatic migration baselines. See Automations for the supported capability table.