| name | talos-inspect |
| description | Read-only triage skill for Anton's Talos OS layer. Use to check cluster health, inspect node disks (install disks, serials, NVMe, volumes), or diagnose node networking (interfaces, routes, bonds, VIP, inter-node reachability, DNS, MTU). Keywords — talos health, cluster healthy, healthcheck, check nodes, node triage, etcd members, etcd quorum, talosctl get disks, NVMe, install disk, disk inventory, volume status, talos network, interface, link, route, bond, VIP, MTU, duplicate IP, resolvers, clock skew. |
| allowed-tools | Read, Bash |
Talos inspect
One-stop read-only diagnostic skill for Anton at the Talos / OS layer. Reach for this before debug-flux-reconciliation: if the kernel, disks, or node network is broken, application symptoms are downstream noise.
Read-only. This skill never runs apply-config (without --dry-run), upgrade, reset, bootstrap, or etcd remove-member. If triage reveals a fix that needs one of those, stop and hand off to the matching skill (upgrade-talos-or-k8s, add-or-replace-node, etc.).
Common invocation shape
Use the repository wrapper for the fleet pulse:
mise exec -- task talos:health
For direct queries, consume scripts/cluster-targets.py instead of copying
node endpoints into this skill or relying on the generated LAN defaults:
TALOS="mise exec -- talosctl --talosconfig ./talos/clusterconfig/talosconfig"
NODES="$(python3 scripts/cluster-targets.py resolve --format addresses --show-addresses)"
ENDPOINT="${NODES%%,*}"
$TALOS --endpoints "$ENDPOINT" --nodes "$NODES" <read-only-command>
Decision tree — which reference file?
| You are asking… | Start with |
|---|
| "Is the cluster healthy?" / nothing specific | references/health.md |
Node just rebooted, NotReady, or crash looping | references/health.md (services + dmesg) |
| Etcd quorum, members, leader, slow apply | references/health.md (etcd section) |
| Clock skew / NTP / time drift | references/health.md (clock section) |
| Planning a node replacement; picking an install disk | references/disks.md |
| Disk full, NVMe errors, volume not mounted | references/disks.md |
Enumerating block devices or serials for talconfig.yaml | references/disks.md |
| VIP unreachable, CNI flapping, node-to-node silence | references/network.md |
| Interface/link state, MTU, bond status, routes, DNS | references/network.md |
When in doubt, run the 60-second pulse below first, then jump.
Golden 60-second health pulse
A yes/no "is anything obviously broken at the OS layer":
mise exec -- task talos:health
mise exec -- kubectl get nodes -o wide
mise exec -- kubectl get pods -A | rg -v 'Running|Completed'
If the Talos pulse and both Kubernetes reads are clean, the OS layer is fine —
the problem is upstream (Flux, app, ingress). Hand off to
debug-flux-reconciliation.
Hard rules
- Read-only, always. Never run
apply-config without --mode=no-reboot --dry-run. This skill does not mutate machine config.
- Never paste
talosconfig, controlplane.yaml, worker.yaml, or any *.sops.* contents into the conversation. File paths are fine; contents are not.
- Never write the real tailnet name into a committed file — use the placeholder
<tailnet-name>.ts.net per repo AGENTS.md. Keep fallback node addresses in scripts/cluster-targets.json only.
- Don't improvise a fix. If inspection reveals a mutation is needed (sysctl patch, version bump, bad disk, dead node), stop and hand off to the right skill.
- Cite the docs for anything unfamiliar. Talos
v1.12 docs at https://docs.siderolabs.com/talos/v1.12/ — WebFetch before guessing at resource schemas or flag names.
Related skills
anton-remote-access — how talosctl targets and kubeconfig contexts are resolveddebug-flux-reconciliation — app-layer triage once the OS layer looks fineupgrade-talos-or-k8s — if inspection reveals a version skew that needs fixingadd-or-replace-node — if inspection reveals a bad disk or a dead nodetalos (global, user-invocable) — Talos v1.12 docs lookup for unfamiliar resources
