REIN

English | Español

From a coding request to a change you can verify.

REIN gives Claude Code and Codex a shared method for inspecting a change, planning when needed, checking the result, and getting independent review. Work stays in your current session by default; Orca runs delegated workers when coordination helps.

See an example · Try a check · Get started · Reference

flowchart TB
    task["Your task"] --> inspect["Inspect code<br/>and tests"]
    inspect --> native["Current session<br/>default"]
    inspect -->|"delegation helps"| workers["Orca workers<br/>Quick / Ship"]
    native --> verify["Checks + evidence<br/>review when required"]
    workers --> verify
    classDef entry fill:#e0e7ff,stroke:#6366f1,color:#1e1b4b
    classDef local fill:#dcfce7,stroke:#16a34a,color:#14532d
    classDef delegated fill:#fef3c7,stroke:#d97706,color:#78350f
    class task,inspect entry
    class native,verify local
    class workers delegated

Start with the code. Add only the process the change needs. Discover resolves questions that could change the solution. Plan organizes requirements and dependencies. Audit adds an independent reviewer for sensitive changes. Quick and Ship provide explicit delegation through Orca.

What changes in practice

One request, three different engineering decisions

“Add EUR support.”

The same request needs different work depending on the code. In the repository's recorded entry-policy evaluation, REIN-guided sessions inspected three development fixtures:

The report includes the fixture checks and remaining work. The simple native control also completed its change; these bounded cases demonstrate observed behavior, without establishing general superiority over an unassisted session.

Try three controls without starting an agent

From this checkout, with Python 3.9+, run:

python3 -m unittest -v \
  tests.test_gate_decision.DecideGateTestCase.test_nothing_ran_is_never_green \
  tests.test_gate_decision.DecideGateTestCase.test_uninvocable_test_is_setup_not_red \
  tests.test_gate.TestCheckReview.test_editing_a_reviewed_file_makes_the_approval_stale

Expected: three passing tests. They demonstrate that an empty gate cannot pass, a missing test tool is reported as a setup problem, and an edit invalidates an approval of the previous state. They run in temporary fixtures and require no model turn or Orca worker.

How review stays attached to the work

flowchart TB
    reviewed["Independent reviewer<br/>approves a code state"] --> current["Approval is current"]
    current -->|"a reviewed file changes"| stale["Approval is stale"]
    stale -->|"review the new state"| reviewed
    classDef approved fill:#dcfce7,stroke:#16a34a,color:#14532d
    classDef changed fill:#fef3c7,stroke:#d97706,color:#78350f
    class reviewed,current approved
    class stale changed

Review is required by the method for changes to security, permissions, money, persistence, concurrency, or service contracts. A blocking finding needs a confirmed defect and evidence. The reviewer must be independent of the implementer. Audits can still miss defects; approval applies to the reviewed state and scope.

What runs where

Explicit Discover is investigation-only. Plan previews the draft before writing. Quick delegates a small change when requested. Ship coordinates a plan when parallel work or recovery justifies it. Answer and Watch surface decisions back to the initiating session. Operational details.

Getting started

1. Make the skills available to your agent

For Claude Code, install from the marketplace:

claude plugin marketplace add luisfarfan/rein-agentic-kit
claude plugin install rein@rein-agentic-kit --scope project

To update that installation:

claude plugin marketplace update rein-agentic-kit
claude plugin update rein@rein-agentic-kit

You can also load the plugin directly from a local checkout:

claude --plugin-dir /absolute/path/to/rein/plugins/rein

For Codex, this repository exposes the same seven skills through .agents/skills and the plugin manifest. To use them in another project, follow the local installation and provider guide. Both providers share the same core in plugins/rein, including the same scope, verification, and independent-review rules.

2. Adopt the project entry guidance

Merge one copy of ENTRY.md into the project's AGENTS.md, preserving its existing instructions. For Claude, add @AGENTS.md to the neighboring CLAUDE.md when needed. Start a fresh session in the correct checkout or worktree and describe the task normally.

Installing the plugin exposes the skills; adopting the guidance enables the shared approach to ordinary tasks. REIN does not edit those files automatically. The guidance is an instruction to the agent. Compliance also depends on the provider and the project's other instructions.

You can invoke a capability directly:

Standalone installation may expose names without the prefix. Use the skill picker to resolve ambiguity.

3. Use the CLI and, for delegation, Orca

The CLI requires Python 3.9 or newer and has no runtime dependencies. To install from this checkout in an isolated environment:

python3 -m venv .venv
.venv/bin/pip install .
.venv/bin/rein doctor .

You can also run the source directly:

python3 plugins/rein/bin/rein doctor .

For ordinary terminals, install the CLI with uv tool install rein-agentic-kit or pipx install rein-agentic-kit. Your project's test and build tools are installed separately.

Quick and Ship require Orca, its CLI, and repositories registered there. Ship runs inside an Orca terminal; rein ship <change> --detach opens that terminal. Local checks such as doctor, detect, and gate work without Orca.

Configuration and checks

REIN respects this precedence: flow.config.json → project task runner → stack detection. It recognizes Python, Node/TypeScript, Rust, and Go, along with frontend and infrastructure subtypes. Monorepos with multiple candidates require an explicit subproject selection.

The example configuration and operational reference cover commands, per-worktree environments, execution limits, and protections.

rein doctor .                            # Inspect stack, commands, and configuration
rein gate .                              # Run checks and decide whether they pass
rein plan-check tasks.md --root . --gate  # Check a plan mechanically
rein review check                        # Check whether the review is still current
rein done my-change --native closure.json --root /plan/repo --dry-run
                                         # Preview native/multirepo closure; see docs/native-closure.md

rein gate returns 0 for a pass, 1 for a failure, and 126 when it cannot verify the result. rein verify diagnoses whether commands can run: its exit code does not certify that tests passed. See the command reference for flags and complete workflows.

Evidence and current limits

The repository includes regression tests, Orca fixtures, and evaluations with reports and traces. The entry-policy validation records bounded cases of native execution and selective investigation in Codex. The acceptance matrix distinguishes automated checks from real integrations and retains pending Claude and cross-provider Quick checks.

A comparison with native execution produced the same functional behavior on a small task, with extra process work when reports and routing were added. The planning, audit, and orchestration evaluations inform selective use of those layers. Available evidence does not establish a general improvement in quality, speed, or savings.

Keep these limits in mind:

• Audits can miss defects. Approval applies to the reviewed state and scope.
• Worktree protections observe watched checkouts through Git. Ignored files and changes outside those repositories are outside that detection.
• Ship ends in review. Publishing, opening a PR, merging, and deploying remain outside its execution. rein done --push allows only an optional push of the plan-archive commit, with its own restrictions.
• Quota checks depend on usable Orca observations. REIN can resume dispatch when valid data arrives, but cannot guarantee that the source refreshes or that enough capacity remains to finish a task.

Consumption measurement remains available through token-report, ledger, and baseline. Token reports read Claude Code transcripts; they are not complete accounting for all Codex sessions. Historical studies are linked from the research documentation.

Documentation and development

• Installation, providers, and project adoption.
• CLI, configuration, and delegated execution reference.
• Shared core structure.
• Project decisions.
• Evolution and evidence record.

Questions or bug reports: GitHub issues.

Run the regression suite from the repository root:

python3 -m unittest discover -s tests -q

MIT license.