Rein

GUIDE AUTONOMY. SHIP QUALITY.
A Claude Code plugin that lets your AI agent move at full speed — while requiring it to plan, leave evidence, and pass review before code lands.

Report Bug · Request Feature · AGENTS.md Examples

What Rein is

Rein is a Claude Code plugin for teams that let an AI agent write production code but don't want to rely on prompt discipline alone. It adds a small set of repo-level guardrails so the agent has to define the task, leave an evidence trail, and pass review before code is committed.

Use Rein if:

• You maintain a long-lived codebase
• Consistency of AI output matters more to your team than raw speed
• You want code review, evidence, and checkpoints codified as team discipline, not held in people's heads

Don't use Rein if:

• You write throwaway scripts or one-shot projects
• Your team doesn't want process files in the repo
• You run outside POSIX bash / WSL2 (native Windows PowerShell is unsupported)

Tool support: Rein's automatic guardrails are built on Claude Code's hook lifecycle. The same conventions (AGENTS.md, rule files) work as reference documents only in Cursor / Copilot — there is no automatic blocking outside Claude Code.

What changes — a 30-second example

Without Rein — typical flow

Developer: "Add a login feature"
   ↓
AI: writes code → git commit
   ↓
No review. No definition of done.
The next session has no record of what was done or why.

With Rein — same request

Developer: "Add a login feature"
   ↓
Rein: "First, write the completion criteria to trail/dod/dod-2026-04-22-login.md"
   → AI writes the checklist (DoD file)
   ↓
AI edits source code (allowed, because a DoD exists)
   ↓
AI tries git commit
   → Rein: "No review record found. Blocked."
   → AI runs codex review → review record created → commit allowed
   ↓
At session end, decisions and changes are logged to trail/inbox/.
The next session auto-loads trail/index.md (the project state summary)
so context carries over.

Why AI-Native

In an AI-assisted workflow, every fix is a re-prompt. In an AI-native workflow, every fix is a rule that prevents the same mistake next time. Rein makes the second possible.

How Rein differs from Claude Code workflows

Claude Code workflows help agents perform tasks. Rein constrains when those tasks are allowed to proceed.

Claude Code gives agents tools. Rein gives teams control. See docs/architecture.md for the hook lifecycle and docs/policy-model.md for the governance model.

What you get

1. Tasks must be defined before code is edited. A Definition-of-Done file is required before any source edit.
2. Reviews block commits and tests. Until a review record exists, git commit and test runners are blocked.
3. Evidence accumulates and rotates automatically. New work lands in trail/inbox/. The next day, yesterday's inbox merges into a daily/ summary; daily entries older than 7 days merge into weekly/. The next session auto-loads trail/index.md for project state — older summaries are read on demand.
4. Updates are handled by Claude Code's plugin manager. Your customisations are not overwritten because the plugin owns its own files.
5. Two models, one session. Rein routes implementation to Claude and code reviews to Codex (with automatic Claude fallback if Codex isn't installed) — both happen inside the same Claude Code session, so you never have to switch tools.

Install

Two commands inside Claude Code, then run the bootstrap command Rein shows you on first edit. No shell installer.

1. /plugin marketplace add JayJihyunKim/rein
2. /plugin install rein@rein
3. Restart your Claude Code session

After install, the first time the agent tries a source edit (Edit / Write / MultiEdit) or a Bash command, Rein's bootstrap gate detects that trail/ is missing, blocks the operation, and prints a one-line python3 …/rein-bootstrap-project.py command. Run that command to create trail/ and .rein/; subsequent edits pass through normally. The same flow runs after /reload-plugins — it converges on the same path as a fresh session.

Non-git projects are supported — git init is not required. When no git_root is found, bootstrap uses the project directory itself as the root.

To disable the gate, set bootstrap-gate: false in .rein/policy/hooks.yaml. Detailed options are in the troubleshooting docs.

Requirements

Recommended setup — use Rein as intended. Rein is designed around a two-model split: Claude implements, Codex reviews. To get the real experience, install the Codex CLI and sign in with a paid ChatGPT plan — Plus tier or higher. The independent Codex review is a core part of Rein's guardrail. The Claude fallback keeps Rein working without Codex, but that is a degraded mode, not the intended setup.

What gets added to your repo

your-repo/
├── .rein/
│   ├── project.json          ← Rein mode + scope (committed)
│   └── policy/               ← repo-local policy templates
│       ├── hooks.yaml
│       └── rules.yaml
├── trail/                    ← evidence store (auto-rotated)
│   ├── inbox/                ← today's completed work (yesterday's entries merge into daily/ on the next session)
│   ├── daily/                ← per-day summaries (entries older than 7 days merge into weekly/)
│   ├── weekly/               ← per-week summaries
│   ├── dod/                  ← Definition-of-Done files
│   ├── decisions/            ← key architectural decisions
│   ├── incidents/            ← hook-block records (used to evolve rules)
│   ├── agent-candidates/     ← new agent suggestions from recurring incident patterns
│   └── index.md              ← current project state (5–25 lines, auto-loaded on session start)
└── .claude/
    └── settings.json         ← one line: plugin pin for `rein`

That's all. The framework's hooks, rules, agents, and skills ship inside the plugin and live in Claude Code's plugin cache, not in your repo — so plugin updates never overwrite anything. If you want project-specific instructions to ride alongside Rein, add your own AGENTS.md (or .claude/CLAUDE.md) in your repo; Rein reads them but never modifies them.

For repo-local hook policy overrides, .rein/policy/hooks.yaml accepts either <hook-name>: false or <hook-name>: { enabled: false }. A profile: key (one of lean, standard, strict) sets defaults for the heavy gates — lean disables post-edit-plan-coverage, post-write-spec-review-gate, and post-write-dod-routing-check for exploratory work; standard (default) keeps all gates on; strict is reserved for stricter future defaults. Per-hook entries always override the profile.

Recommended: add trail/ and .claude/cache/ to your .gitignore if you don't want session evidence committed (Rein doesn't auto-edit .gitignore).

Platform support

For Windows, WSL2 (Ubuntu) is recommended. See docs/troubleshooting/windows.md for installation and Git Bash diagnostics.

Troubleshooting

/plugin marketplace add JayJihyunKim/rein

claude plugin remove everything-claude-code

Security

Do not file security vulnerabilities as public GitHub issues.

If you discover a vulnerability, please report it privately via GitHub Security Advisories: Report a vulnerability.

Contributing

Issues and pull requests are welcome.

1. Fork the repo
2. Create a feature branch: git checkout -b feat/amazing-feature
3. Commit your changes: git commit -m "feat: add amazing feature"
4. Push the branch: git push origin feat/amazing-feature
5. Open a Pull Request

Before submitting, read AGENTS.md to understand the framework structure and the rules that govern contributions.

Release history

Latest release: v1.5.8 (2026-06-26) — code reviews (/codex-review) now run faster on small changes: review depth scales automatically to the size of your change (lines, files, document length) instead of always running at the deepest, slowest setting. You can still set the depth explicitly to override it. (CHANGELOG)

For prior dev-cycle history (v0.x), see docs/changelog-archive/2026-04-pre-v1.md.

Full release notes: CHANGELOG.md

License

MIT — see LICENSE.

References

• agentsmd/agents.md — AGENTS.md hierarchy
• anthropics/skills — skill definitions
• Shields.io — badges