Feedback wanted: read-only Kubernetes evidence graph for diagnostics and AI agents

Hi Kubernetes community, I open-sourced a small project called `kubernetes-ontology` and would appreciate feedback from people who work on troubleshooting, platform engineering, SRE tooling, or agent-assisted operations.

Repo: GitHub - Colvin-Y/kubernetes-ontology: Read-only Kubernetes ontology database for diagnostics and AI-agent consumption. · GitHub
Landing page: https://kubernetes-ontology.vercel.app

The project builds a read-only in-memory graph from Kubernetes objects, then exposes typed entity/relation queries, diagnostic subgraphs, CLI/HTTP APIs, and a local topology viewer.

The current MVP focuses on:

  • Pod and Workload diagnostic entrypoints
  • ownerReference chains, including Pod → ReplicaSet → Deployment
  • Service selector matches
  • Pod links to Node, ConfigMap, Secret, ServiceAccount, Image, and PVC
  • PVC → PV → StorageClass → CSIDriver storage paths
  • ServiceAccount to RoleBinding / ClusterRoleBinding evidence
  • Kubernetes Event and admission webhook evidence
  • HelmRelease / HelmChart evidence from standard Helm labels and annotations
  • stable JSON output for downstream AI agents

The safety model is intentionally conservative: it does not mutate observed workloads, does not install workload CRDs, does not require a graph database, and does not perform runtime writes to the Kubernetes resources being diagnosed.

I am trying to validate the graph-first shape more than promote a finished product, so concrete feedback would be very helpful:

  1. Which Kubernetes failure mode should the graph explain next? CrashLoopBackOff, Pending scheduling, DNS / Service routing, volume attach / mount, webhook admission failures, RBAC, Helm upgrade failures, or something else?
  2. Are the graph edges and provenance fields useful enough for real incident workflows?
  3. What would make the quickstart easier to try in private or air-gapped clusters?

If this overlaps with your troubleshooting workflow, I would love to hear what context you usually wish you had in one place.