Small TypeScript agents with typed inputs, validated outputs, and composable workflows. Write a program, hand it to the Rig skill, or run it with the launcher. Rig uses the Copilot SDK by default and also supports Anthropic, Codex, DeepSeek Harness, Gemini, and Pi engines.

Install the skill for your coding agent (GitHub CLI 2.90+):

gh skill install githubnext/rig rig

To run programs locally, use Node.js 24+ and install the dependencies from a checkout:

git clone https://github.com/githubnext/rig.git
cd rig
npm ci

Save this as review.ts:

import { agent, s } from "rig";

// Agent role: assess a proposed change.
export default agent({
  model: "small",
  input: s.string,
  output: s.object({
    summary: s.string,
    risk: s.enum("low", "medium", "high"),
  }),
  instructions: "Review the proposed change. Summarize it and assess its risk.",
});

With an installed, authenticated Copilot CLI, run:

printf '%s\n' 'Replace the login form' | node skills/rig/run.ts review.ts --server

The launcher reads stdin, runs the exported agent, and prints the validated result. To check without making a model call:

node skills/rig/run.ts review.ts --typecheck

gh skill install copies the skill but does not install SDK dependencies. For an installed skill, find its directory with gh skill list and substitute it for skills/rig in the commands above. Without --server, Rig selects an engine from your environment or uses the default Copilot endpoint; see running programs for launch options.

Pin the Rig skill and shared launcher template to full commit SHAs. The template provides Node.js 24 and a narrow printf/node allowlist; the consuming workflow must enable the Copilot SDK, grant copilot-requests: write, and provision dependencies. See the workflow setup guide for the exact configuration and credential handoff.

The Monthly PR Clusters workflow runs daily at 09:00 UTC or manually, analyzing PRs merged in the trailing calendar month across all base branches. Each execution generates a new Rig clustering strategy after reviewing earlier experiments on the memory/pr-cluster-rigs repo-memory branch. It creates an issue with the three largest clusters and the complete generated Rig source, preserving previous reports and successful or failed rigs for future comparisons. The workflow prompt specifies behavior, not a fixed Rig implementation.

• Skill guide — the canonical program and construction rules.
• Agent API and prompt intents — schemas, inputs, and workspace context.
• Dynamic workflows — deterministic orchestration.
• Sample catalog — agent and workflow patterns bucketed by API, topic, and task family.