K8S Verify

/skill:k8s-verify — Kubernetes Post-Deploy Verification

Purpose: Systematic post-deploy health check for Kubernetes deployments. Verifies pods are running, services have endpoints, no error events, and applications are responding. Project-agnostic — adapts to any namespace.

When to Use

• After deploying to Kubernetes (kubectl apply, helm install, kustomize, etc.)
• After a rollout restart
• When something feels wrong but you're not sure what
• As a smoke test before handing off to users

Usage

/skill:k8s-verify                          # Auto-detect namespace from context
/skill:k8s-verify --namespace myapp        # Specific namespace
/skill:k8s-verify --namespace myapp --full # Include log analysis

Arguments

• --namespace <ns> — Kubernetes namespace to check (auto-detected from kube context or CLAUDE.md if omitted)
• --full — include pod log analysis (slower, more thorough)
• --json — output results as JSON for scripting

Step 1: Namespace Discovery

If --namespace was not provided:

1. Check CLAUDE.md for a namespace reference
2. Check for a kustomization.yaml with namespace
3. Fall back to kubectl config view --minify -o jsonpath='{..namespace}'
4. If still empty, ask the user

Step 2: Pod Health

kubectl get pods -n <namespace> -o wide --no-headers

For each pod, check:

Count: total pods, healthy, warning, failed.

Step 3: Rollout Status

For each deployment and statefulset:

kubectl rollout status deploy/<name> -n <namespace> --timeout=10s 2>&1

Report whether each rollout is complete or still progressing.

Step 4: Service Endpoints

kubectl get svc -n <namespace> -o wide --no-headers
kubectl get endpoints -n <namespace> --no-headers

For each service:

• Check that it has at least one endpoint
• If a service has no endpoints: FAIL — pods may not be matching the selector

Step 5: Recent Events

kubectl get events -n <namespace> --sort-by=.lastTimestamp --field-selector type!=Normal 2>/dev/null | tail -20

Flag any Warning or Error events from the last 10 minutes. Common patterns:

• FailedScheduling — resource pressure
• Unhealthy — probe failures
• BackOff — container crash loop
• FailedMount — volume issues
• FailedCreate — controller issues

Step 6: HTTP Health Checks (if applicable)

If CLAUDE.md or the deployment specs define health endpoints (common paths: /health, /healthz, /ping, /ready):

# Port-forward and check (timeout 5s)
kubectl port-forward -n <namespace> svc/<name> <local>:<remote> &
PF_PID=$!
sleep 2
curl -sf http://localhost:<local>/health
kill $PF_PID 2>/dev/null

Only attempt this for services that appear to be HTTP servers (ports 80, 443, 3000, 8080, 8443, etc.).

Step 7: Log Analysis (with --full)

For each failed or warning pod:

kubectl logs -n <namespace> <pod> --tail=50 --timestamps

Scan for:

• error, fatal, panic, exception (case-insensitive)
• Stack traces
• Connection refused / timeout patterns
• Auth failures

For CrashLoopBackOff pods, also check previous container:

kubectl logs -n <namespace> <pod> --previous --tail=30

Step 8: Report

## Deploy Verification: <namespace>

| Check | Result | Details |
|-------|--------|---------|
| Pods | PASS/WARN/FAIL | N/N healthy |
| Rollouts | PASS/FAIL | N/N complete |
| Endpoints | PASS/FAIL | N/N services have endpoints |
| Events | PASS/WARN | N warnings in last 10m |
| Health HTTP | PASS/FAIL/SKIP | N/N endpoints responding |
| Logs | PASS/WARN/SKIP | N error patterns found |

**Verdict: HEALTHY / DEGRADED / UNHEALTHY**

[If DEGRADED or UNHEALTHY: list specific issues with remediation steps]

Verdict Logic

• HEALTHY: All checks PASS or INFO
• DEGRADED: Any WARN, no FAIL
• UNHEALTHY: Any FAIL

Key Rules

1. Non-destructive. This skill only reads. It never modifies, restarts, or deletes anything.
2. Timeout everything. Kubectl commands can hang. Always use --timeout or wrap with a timeout.
3. Adapt to the cluster. Not all clusters have metrics-server or ingress. Skip checks gracefully when resources aren't available.
4. Port-forward cleanup. Always kill port-forward background processes, even on error.
5. Sensitive data. Never print secret values. Only show secret names and whether they exist.