repo-ai

@infrazero/repo-ai is the ai-loop pipeline: a label-driven loop that takes an ai-ready GitHub issue, implements it in its own git worktree, opens a PR, has two agents review it, and hands it to a human to merge.

issue → loop → PR → two reviews → you merge → you release.

It was split out of @rtorcato/repo-tooling so the tooling can be used without the loop. repo-tooling still owns the repo-side standard the loop relies on — branch protection, auto-merge, and the .claude/settings.json worktree config.

warning

Costs and liability. By installing or using repo-ai, you accept these risks and responsibilities. repo-ai runs AI agents unattended, and they spend your Anthropic credits or plan limits and your GitHub Actions minutes. The loop's limits are best-effort, not a spending guarantee. Set spend limits with your provider, and stop the loop when you aren't watching it. Agents can be wrong, so you review and merge every change. Provided as is under the MIT license, with no warranty; the authors aren't liable for costs, damages or changes made by agents. Not affiliated with Anthropic or GitHub. Read the full Risks and responsibilities.

What it ships

• Claude Code skills — ai-loop (the entry point and the engine), ai-issue (the on-ramp), ai-loop-status (a read-only view), and ai-loop-stop (stop this repo's loop).
• loop commands — the mechanics the skills call, as tested code: loop tick, loop guard, loop worktree add, loop reap, and more.
• doctor / fix — audit and repair the loop's own setup: labels, the agent user, and the installed skills.

Every command takes --json.

Requirements

repo-ai targets Claude Code only for now; other agent harnesses aren't supported yet. What each part depends on:

Part	Depends on
repo-ai CLI (loop …, doctor, fix, setup)	Node ≥ 22 and gh. Harness-neutral, so any agent or script can call it
Skills (ai-loop, ai-issue, ai-loop-status, ai-loop-stop)	The Claude Code skill format
Parallel implement/review (ai-loop Pass 3 and Pass 4, via the ai-loop-recover and ai-loop-pickup workflows)	The Workflow tool. Named reviewer types (code-reviewer, security-expert) need them in the Agent tool's registry; otherwise the reviewers run as general-purpose. Without Workflow, ai-loop falls back to background Agent calls
Self-scheduling (/ai-loop keeps itself going)	A session-scoped recurring CronCreate job
Wake on change (loop watch)	The Monitor tool
Statusline segment	Claude Code's statusline JSON

Workflow runs show in /workflows. If you edit the scripts in workflows/, load /workflow-authoring first.

Start here

The Minimum path is three commands, with no options:

1. Run npx @infrazero/repo-ai setup.
2. File an issue with /ai-issue in Claude Code.
3. Run /ai-loop.

Getting started is the full walkthrough: the repo settings, a bot account, the sandbox, stopping, merging and releasing.

Installing the skills

Use setup. It runs fix claude-skills, which installs the four skills (ai-loop, ai-issue, ai-loop-status, ai-loop-stop) and the two Workflow scripts they run by name (ai-loop-pickup, ai-loop-recover), stamped with the npm version you ran. By default they go to ~/.claude/skills and ~/.claude/workflows, for every project on the machine.

Sharing them with your team? Install into the repo instead and commit the result:

npx @infrazero/repo-ai fix claude-skills --skills-dir .claude/skills

The Workflow scripts land next to them in .claude/workflows, so everyone who clones the repo runs the same version.

Prefer a plugin? In Claude Code, run /plugin marketplace add infra-zero/repo-ai, then /plugin install repo-ai@repo-ai. You get the same four skills and both Workflow scripts, registered as repo-ai:ai-loop-pickup and repo-ai:ai-loop-recover, which ai-loop finds on its own. The plugin is unversioned: it follows main, so it can run ahead of the npm release. Pick one way, not both.

Either way, the skills call the CLI through npx @infrazero/repo-ai.

Models and reviewers. The ai-loop skill pins model: sonnet, and its quick verification step pins haiku. Reviews run as code-reviewer and security-expert when you have agents with those names in .claude/agents or ~/.claude/agents. Otherwise they run as general-purpose, with the reviewer's brief in the prompt. doctor's Reviewer agents line tells you which you'll get.

Optional mods

Mods are Claude Code plugins built on its early-access mod API. They need Claude Code 2.1.292 or later and may break between Claude Code releases. Neither setup nor fix installs them. Add the marketplace once with /plugin marketplace add infra-zero/repo-ai, then install any of these:

Mod	Install	What it does	Needs
release-guard	/plugin install release-guard@repo-ai	Stops an agent from releasing outside CI: npm publish, version bumps, a local semantic-release run, git tags and tag pushes, force pushes, git push --tags or --follow-tags, release workflow dispatches (and any dispatch by a numeric or $var workflow ID), milestone closes (they fire the release workflow), GitHub releases and package.json "version" edits. It tells the agent to have you run the command yourself with ! <command>.	Nothing else
loop-watch (experimental)	/plugin install loop-watch@repo-ai	A live view of the loop: a status line segment (🤖 <summary> · next 9m · 2 agents · 1 ready to merge), a /loop-watch pane of the open PRs and issues with ai-* labels, a toast and chime when a PR passes both reviews, and a dock with one tile per running subagent.	gh
repo-ai-rainbow	/plugin install repo-ai-rainbow@repo-ai	A colourful band above the prompt with the open ai-* label counts and the loop status. /repo-ai refreshes it.	gh

release-guard matches command text rather than parsing the shell, so it catches an agent's slip, not an adversary. It deliberately allows gh pr merge. Required checks and the release environment remain the real gates.

loop-watch reads the same .claude/ai-loop-status as the statusline segment and polls gh every 2 minutes. It sees only this session's agents; agents in other sessions show up through their labels and the status file.

release-guard follows main. loop-watch and repo-ai-rainbow are pinned to a release commit, so a push to main never changes what you have installed.

Read The AI Loop for how the pipeline works, and Commands for every command.