9. What's Next

The pattern is complete and runnable today. These are the directions that make it stronger, roughly in priority order. Several are good first contributions, see Contributing.

Near term #

• A recorded end-to-end demo in the README, so the flow is visible without a setup.
• Published evaluation results across several models, with the failures called out. See Evaluation.
• Upstream proposals filed from docs/upstream-notes.md: long-running operations in MCP, richer ManifestWork failure feedback in OCM, and contributing the ManifestWork-envelope policy pattern to the Kyverno catalog.

Mid term #

• OCM cluster-proxy transport so the server host never holds spoke credentials directly.
• An authenticated HTTP transport (SSO/OIDC) so the in-cluster Deployment can serve clients standalone rather than over attached stdio.

Shipped: a signed container image (ghcr.io, SBOM + provenance + Cosign) and a Helm chart with a Restricted pod, PVC, NetworkPolicy, and PDB.

Longer term #

• More chaos classes: node pressure, network partitions, noisy neighbors, to broaden what the eval harness covers.
• Placement-aware fleet actions: proposals that target a scored subset of clusters, not one at a time.
• Industry policy presets: ready-made guardrail packs (finance, healthcare) layered on top of the base set.

How this list is chosen #

Priorities follow evidence. The eval harness surfaces where agents actually fail; those failures decide what safety work matters next. If you hit something the current design cannot express safely, open a feature request with the scenario and it becomes a candidate.

Next: Contributing.

Source: wiki/Roadmap.md — this page is generated from it, so edit the markdown, not the site.