XR 3.1 — the agentic runtime

Notes on the agentic runtime: the skill layer, the model router, and the security gate.

XR 3.1 is the agentic runtime release: a typed skill layer, an explainable model router, and a fail-closed security gate.

The skill layer is manifest-governed: every skill declares its id, version, publisher, permissions, and categories, so what you install is what you audit.

The model router explains why a model was chosen — capability, cost, latency, and locality are all inputs, and sensitive work never silently routes to the cloud.

The security gate canonicalizes policy decisions first, then evaluates them, and denies what it cannot canonicalize. Unparseable reviewer output is treated as a request for changes, never as approval.

Measured baseline: CLI fast path cold start ~36 ms p95, doctor readiness ~456 ms p95, retrieval over 100k items ~33 ms p95. Budgets and regression gates are enforced in CI.

Next: faster local models, richer multi-agent orchestration, and deeper editor integrations.