Opt-in Agent presets and inspection

Use the public createAgentPreset('research' | 'workflow', config) factory with Agent.create(). Research selects observation masking, workflow selects algorithmic compaction plus protected task state; both enable explicitly scoped evidence and bounded recovery. Defaults are 12/8 main-loop iterations. Explicit settings override these defaults, including loopPolicy: null. Helpers, hard ModelTask limits, permissions and host completion checks remain explicit.

agent.inspectCapabilities() reports actual managed configuration and provider support without a model/storage call or private content. Per-call overrides and external drivers keep their existing ownership. AMOS /preset saves next-agent selection and distinguishes it from active configuration. Candidate startup/load failures preserve the prior AMOS session; recovery observations are checkpointed before release. Titles/preset labels survive autosave and unnamed saves. Inspection projects declared getter-backed capabilities without exposing provider extras. Failed resumes restore caller-owned permission/tool/prompt/role/callback bindings before candidate cleanup, including constructor and synchronous load-notification failures; rejected scoped loads retain the active checkpoint and autosave target. Getter-backed scope requirements are normalized, and teardown continues after individual plugin errors. JSON-compatible plugin conversion hooks remain supported, and failed rollback keeps saves blocked until every damaged plugin actually recovers. No session/portable protocol changes; scope checks and legacy loading remain intact. Read the complete recipe for setup, overrides, failure behavior, public examples, runtime boundaries and migration. Availability is on the Version 2 branch pending a verified package release.