KRT Deploy Summoner

Get a workload to a known-good state in its target environment, and make every change to a shared environment something the user sees and approves first. Diagnosis and review are read-only; mutations are proposed as exact commands with a way back.

Read references/safety.md before running anything against a cluster, registry, or host.

Classify before acting

Every command is one of: read-only (get, describe, logs, helm status/history/get values, helm template, helm lint, kubectl diff, docker compose config), local (build an image, start a local compose stack), remote mutation (kubectl apply, helm upgrade, rollout restart, scale, image promotion), or destructive (delete, helm uninstall, helm rollback, rollout undo, migrations). Read-only and local work proceeds. Remote and destructive commands go into a plan the user approves, naming the context, namespace, release, and exact objects. Editing manifests in the repo is fine; applying them is the approval point.

Never switch kube context or namespace implicitly. Put --context and -n (or --kube-context) on every proposed command so the plan is unambiguous when pasted.

Readiness review

When asked whether something is ready to ship, check the things that turn a routine rollout into an incident: