Agent skill

What it does

• Runs the loop: deploy, read the DeployResult and the verification verdict (working, partly_working, broken), apply error.fix, deploy again.
• Asks you once for missing secrets and stores them with env_set, never in the repo.
• Knows every error code and verification issue, with its fix, and uses logs, status, verify, issues and rollback to investigate.
• Stops and hands over when a person is needed: secret values, payment, DNS records, destroy confirmation. Retries are capped, so it never loops blindly.

The skill drives the Rivtor MCP tools when they are connected and the CLI otherwise. Connect the MCP server as well for the best results.

Install with the CLI

This installs for Claude Code in your home directory. Pick another agent with --target codex, --target cursor or --target all, and add --project to install into the current project so your team gets it from git.

Running it again replaces the previous copy. Start a new agent session to load the skill.

Claude Code plugin

Install it from the Rivtor marketplace inside Claude Code:

/plugin marketplace add shyamolkonwar/rivtor-cli
/plugin install rivtor@rivtor

Update it later with /plugin marketplace update rivtor.

Codex

Codex loads skills from ~/.agents/skills and from .agents/skills in your repository, and picks it when you ask to deploy or fix a deploy.

Cursor

Cursor has no user-level skill folder, so this writes a project rule, .cursor/rules/rivtor.mdc, that the agent applies when a request is about deploying. The reference files go next to it in .cursor/rules/rivtor/references/.

Copy it by hand

The skill ships inside the npm package. npx rivtor skills path prints its folder; copy that folder to wherever your agent reads skills:

mkdir -p ~/.claude/skills
cp -R "$(npx -y rivtor skills path)" ~/.claude/skills/rivtor

The source is skills/rivtor in github.com/shyamolkonwar/rivtor-cli: one SKILL.md and a references/ folder that the agent opens only when it needs them.