Implementation Quickstart

1. Inventory native tools and preserve their original ids.
2. Create tool_declaration records with namespace, aliases, search hints, kind, lifecycle, input contract, output contract, and mappings.
3. Create tool_interface records for validation, safety classification, permission matching, result mapping, and persistence boundaries.
4. Create permission profiles for tools that read, write, access credentials, use network, run code, mutate remote systems, or require user interaction.
5. Build small tool_surface records per turn or task instead of exposing the whole catalog.
6. Mark large catalogs with deferred_tool_ref records and expose tool search.
7. Wrap every call in a tool_invocation envelope.
8. Preserve model_input, observable_input, permission_input, and call_input separately.
9. Emit hook, permission, scheduler, progress, persistence, and terminal result events.
10. Link results to resources, artifacts, evidence, policy decisions, runtime events, and telemetry spans.
11. Keep native protocol payloads available by ref where possible.