Triage the open Renovate dependency-update PRs on this home-cluster repo. Use whenever the user wants to review, work through, clear out, or "do the renovate PRs" / "dependency PRs" / "update PRs" / "the weekend bumps" — even if they don't say "Renovate" by name. Groups minor+major bumps of the same dependency, hunts down changelogs, scans the repo (and the live cluster via the kubernetes MCP) for changes we'd need to make, sorts everything by danger/blast-radius, then walks the user through each one interactively so they decide what to merge, defer, or dig into, and confirms via the ArgoCD MCP that each merged bump actually synced and went healthy. Reach for this any time there's a pile of Renovate PRs to get through.
Upgrade the Talos version on this home-cluster's nodes safely, one at a time, draining Longhorn + workloads first and verifying full health before moving on. Use whenever the user wants to bump Talos (e.g. "get the cluster on v1.13.5", "upgrade talos", "roll out the new talos version", "update <node> to <version>", "do the talos nodes"), reinstall/recover a node, or migrate a node onto the custom Mac installer. Handles the whole node loop: pick a safe order (workers → control-plane, non-leaders → etcd-leader last), cordon, drain past the Longhorn instance-manager PDB, evacuate single-replica volumes that would otherwise go offline, run `talosctl upgrade --preserve`, ride out the T2 Mac BootFFFF non-fatal error, re-enable scheduling, and confirm every volume / pod / etcd member is healthy before the next node. Knows this repo's talhelper layout, the custom GCC installer for the talmac nodes, and the Longhorn gotchas (stale USB mounts, diskUUID mismatch, data-locality PVs, CNPG switchover). Reach for this any t
Find, read, and debug firing alerts in this home-cluster's VictoriaMetrics / vmalert monitoring stack. Use whenever the user asks what's alerting, why an alert is firing, to "check the alerts", "look at monitoring", "debug the vmalerts", "what's firing", "is anything broken", or names a specific alert ("why is KubePodCrashLooping firing", "silence X"). Walks the whole loop: query vmalert's live alerts API for what's firing, read each alert's expression/labels/annotations, reproduce the expression against VictoriaMetrics to see the real numbers, tell genuine problems apart from the always-on meta-alerts (Watchdog, InfoInhibitor) and benign empty-recording-rule false-positives, then fix — either resolve the underlying condition or, for a noisy/false-positive rule, disable/tune it via the vm-k8s-stack defaultRules in values.yaml (GitOps) and let ArgoCD sync. Knows the pod/port layout (vmalert :8080, vmsingle :8428), the kubectl-exec + busybox-wget query pattern, and where the rules live in the repo. Reach for th
Add a NEW application to this home-cluster repo as an ArgoCD Application using the bjw-s app-template Helm chart. Use whenever the user wants to add, create, onboard, deploy, "set up", or "stand up" a new app, point at a GitHub repo / README / project URL to "run this in the cluster", or translate a Flux HelmRelease / docker-compose / plain manifests into this repo's GitOps layout, or says "add <app> to argo" / "deploy <app>" / "I want to run <app>" / "make an app for <x>". Spawns subagents to deep-read the upstream project's FULL README/docs (installation, every env/config knob, and optional add-ons that fit this repo's style), presents the optional choices via the question tool for a fully-fleshed-out — not bare-bones — deploy, then scaffolds the full directory (values.yaml + Argo Application, plus ksops secret files when needed), wires the house patterns (internal ingress on laboratory.casa, Longhorn persistence, runAs 3000 security context, central Postgres init-container, central valkey), pins images to
Cleanly decommission an ArgoCD application from this home-cluster repo so it leaves NO traces behind in the cluster. Use whenever the user wants to remove, delete, decommission, retire, tear down, or "get rid of" an app, or says "I no longer want <app>" / "nuke <app>" / "remove <app> from argo" / "clean up <app>". Handles multi-component apps, the non-cascading-prune trap (orphaned workloads), cluster-scoped resources (ClusterRole/ClusterRoleBinding) that survive a namespace delete, PVC/PV reclaim, central-PG database rows, and runtime-created resources that no Application owns. Works GitOps-first (ArgoCD does the deletion), verifies every layer via the kubernetes + argocd MCP servers, and confirms zero residue at the end. Reach for this any time an app needs to go away completely.