What problem does it solve? Operating a two-node k3s homelab cluster involves many failure-prone areas: BuildStream distributed builds that get preempted, Zot registry sync-prefix changes that silently break image pulls, GPU memory misconfiguration on Strix Halo nodes, and wedged nodes that report Ready while pods hang. This Skill consolidates verified operational runbooks, measured hardware findings, and GitOps-safe procedures so cluster changes are made correctly the first time. ## Core Features & Use Cases - Cluster and registry operations: Covers kubectl/k3s workflows, Zot OCI registry sync-prefix changes with metric-based verification, external-secrets, and K8sGPT MCP-based cluster analysis. - BuildStream and Buildbarn distributed builds: Enforces USB4 admission gating, correct cache configuration (override-project-caches, Buildbarn frontend endpoints), PriorityClass-based preemption avoidance, and RECC pilot evidence handling. - Hardware and storage runbooks: Documents AMD GPU topology and GTT sizing on Strix Halo, NVMe/XFS migration procedures, FCOS memory-limit quirks, and wedged-node recovery without SSH. - Use Case: Before changing a Zot sync prefix, an operator follows the Skill to inventory served repositories via zot_repo_downloads_total metrics, applies the config-version bump through GitOps, and verifies every repository still reports non-zero download counters instead of trusting a misleading skopeo timeout. ## Quick Start Ask the assistant to check the cluster's USB4 link admission state and BuildBarn worker readiness before submitting a Dakota BuildStream build.
