Cloud-ops agent skills

skills/ in the repo holds agentskills-style skills that a coding or ops agent (Claude Code, Cursor, a CI bot) can load to operate Ramen safely. Each skill is a folder with a SKILL.md (front matter name / description, numbered steps, explicit boundaries).

Skill	Purpose
deploy-gcp	Terraform + images + Helm bring-up on GKE, then first zone/group/deploy through the API
deploy-aws	Same for EKS (the skill demands a dry run first)
rotate-keys	Rotate rmk_ MCP keys, rmn_ API keys, the admin key and the Fernet key without downtime
backup-restore	Versioned JSON backups to local/bucket and restores
scale-zone	Add a zone, scale counts/sizes, rebalance, remove a zone

Point an agent at one by handing it the file. In Claude Code, @skills/deploy-gcp/SKILL.md in a prompt, or copy the folder into your own .claude/skills/. Cursor and other agents take the same file as context. Nothing in a skill is Ramen-specific tooling: each is plain Markdown with numbered steps.

Validation sub-agent convention

Every skill ends with a Validate section. The convention is that the agent running the skill spawns a cheaper, read-only sub-agent with only that section and the URLs/keys it needs; the sub-agent reports PASS/FAIL with the evidence (HTTP status, gRPC status, job status, tools/call result) and never mutates anything. The parent agent may not declare the task done until the validator has passed. skills/README.md has the exact wording to hand to the sub-agent.