DevOps

Docker, Kubernetes, Helm, Argo CD, Terraform, OpenShift, GitHub Actions, GitLab CI. The rules here are about blast radius, because that is what differs from application code: a bad commit is revertible, a bad kubectl apply is an incident.

The one rule

Observe, plan, bound, then mutate. Before any operational write:

1. Resolve the exact target and your authorization for it
2. Capture current state
3. Produce plan / diff / dry-run output
4. State blast radius, health gate, rollback, and stop condition
5. Require explicit approval for production or anything destructive
6. Apply the smallest scope, verify, then expand

That is the plan-before-operational-change instinct. Most infrastructure failures are target, scope or rollback failures — not syntax failures. A green exit code proves execution, not safety.

Safe by default

Tool	Read-only first	Then
Terraform	terraform plan	apply with the saved plan file, never a bare apply
Kubernetes	kubectl diff -f	apply — and --context explicit, every time
Helm	helm diff upgrade	upgrade --atomic --timeout
Argo CD	argocd app diff	sync — or better, let git do it
GitHub Actions	act locally	push to a branch, never edit on main

kubectl context is the classic incident. Make the agent state the context and namespace in the command every time — kubectl --context staging -n api get pods — rather than relying on whatever the last use-context left behind.

GitOps

With Argo CD, the agent should be editing manifests in git, not the cluster. A change that does not go through a commit is a change that vanishes on the next sync and that nobody can review. If the agent reaches for kubectl apply on a GitOps-managed namespace, that is a bug in the plan.

Git hooks

Claude Code hooks run inside the agent loop. Git hooks catch what already got written — belt and braces.

cp ~/Repos/ai-dev-kit/templates/shared/agent/git-hooks/pre-commit-secrets.sh .git/hooks/pre-commit
chmod +x .git/hooks/pre-commit

Blocks commits containing sk-ant-, sk_live_, ghp_, AKIA, Slack tokens, SendGrid keys, JWTs, private key blocks — and refuses .env, *.pem, *.key, credentials.json, id_rsa outright.

For a repo-wide, shared version use pre-commit with gitleaks — .git/hooks/ is not committed, so a local hook protects only you.

CI/CD

• The agent reads CI logs, it does not rerun the pipeline to see what happens. Fetch the failed job's log, diagnose, fix, push once.
• Pin action versions by SHA, not by tag. A moving tag is remote code execution with extra steps.
• Secrets come from the CI secret store. If the agent proposes a secret in a workflow file, that is a stop.
• gh run view --log-failed is the fastest path from a red pipeline to the actual error.

Kubernetes diagnosis

k8sgpt (★8k) scans a cluster and explains failures in plain language — the fastest way to go from "pods are unhealthy" to the actual cause:

k8sgpt analyze --explain --namespace <ns>

Read-only by default, which is what you want an agent pointed at. Pair it with LukasNiessen/kubernetes-skill.

Containers

• Multi-stage builds; the final stage carries no build toolchain
• Non-root user, read-only root filesystem where the app allows it
• Pin base images by digest
• docker scout cves or trivy image before anything ships
• .dockerignore mirrors .gitignore — node_modules in a build context is the most common slow build

What is missing from this kit

No terraform-patterns, k8s-manifests or helm-chart skills yet. See the gap list in the repo's README — they are the top of the backlog for this stack.