Ralph Review

Orchestrating coding agents for code review, verification, and fixing via the Ralph loop.

Table of Contents

• How It Works
• Reviewer and Fixer Flow
• Workflows
• Prerequisites
• Installation
• Quick Start
• Commands
• Supported Coding Agents
• Configuration
• License

How It Works

Ralph Review now uses a batch-first workflow:

1. rr run performs review only.
2. The reviewer runs in a disposable session worktree and reports structured findings.
3. Findings are deduplicated across review iterations and persisted as a session artifact.
4. If findings exist, you choose which ones to fix with rr fix.
5. The fixer handles the selected findings in a separate batch remediation phase.
6. Resolved fixes are handed back to your working tree, either automatically or as a pending handoff to apply manually.

This keeps review and remediation separate by default. The reviewer can focus on finding real issues, and the fixer treats those findings as input for a later, explicit remediation step.

Use rr run --auto when you want Ralph Review to run remediation immediately after review. Add --priority P0,P1 to auto-fix only selected priority levels.

Reviewer and Fixer Flow

flowchart TD
    A[Your repository] --> B[rr run]
    B --> C[Preflight checks]
    C --> D[Start tmux session]
    D --> E[Create disposable review worktree]
    E --> F[Reviewer agent]
    F --> G{New findings?}
    G -- Yes --> H[Merge findings into inventory]
    H --> I{Max iterations reached or no new findings?}
    I -- No --> F
    I -- Yes --> J[Persist findings artifact]
    G -- No --> K[Clean review result]
    J --> L{Fix now?}
    L -- Later --> M[rr fix --session SESSION]
    L -- "rr run --auto" --> N[Select findings automatically]
    M --> O[Select findings by prompt, all, priority, or ID]
    N --> P[Create disposable fix worktree]
    O --> P
    P --> Q[Fixer agent batch remediation]
    Q --> R{Selected findings resolved?}
    R -- Yes --> S[Create handoff]
    S --> T{Auto-apply succeeds?}
    T -- Yes --> U[Fixes applied to working tree]
    T -- No --> V[Pending handoff]
    V --> W[rr apply or rr prune --discard]
    R -- No --> X[Retain remediation worktree for review]

Reviewer

The reviewer analyzes the selected review scope for correctness, security, reliability, and maintainability issues introduced by the change. It outputs structured findings with stable IDs such as F001, priorities P0 through P3, and source locations.

Reviewer iterations continue until no new findings are discovered or maxIterations is reached. By default the run stops early when an iteration finds nothing new. Use --force to run the full iteration count.

Fixer

The fixer runs only after findings have been persisted and selected. It receives the selected finding inventory, works in a disposable fix worktree, and returns a per-finding result: resolved or unresolved.

When all selected findings are resolved, Ralph Review creates a handoff. Depending on the working tree state, the handoff may be applied automatically or left pending for rr apply. If selected findings remain unresolved, Ralph Review keeps the remediation worktree available for inspection.

Workflows

Ralph Review fits into a few common loops. Each workflow below shows the scenario first, then the exact commands.

1. Review what you are about to commit

You just finished a feature on a working branch. Before you stage and push, you want a second pair of eyes on staged, unstaged, and untracked changes.

rr run --uncommitted

Findings are printed and persisted. You can keep coding while the reviewer runs in the background and come back to triage when it is done.

2. Review your branch against the base

You are preparing a pull request against main and want the reviewer to look at the full diff, not just your last commit.

rr run --base main

Add a focus instruction when the diff is large:

rr run --base main "focus on auth boundaries and input validation"

3. Re-review a single commit

CI flagged something on a specific commit, or you are auditing a hotfix in isolation.

rr run --commit 9f3a2b1

4. Review now, fix later (the default loop)

rr run only reviews. Findings get a stable ID (F001, F002, ...) and a priority (P0 through P3). Triage them, then fix exactly what you want:

rr run
rr fix --session SESSION --priority P0,P1
# or pick by ID
rr fix --session SESSION --id F003 --id F007

When the fixer is done it either applies the patch to your working tree automatically, or leaves a pending handoff that you apply explicitly:

rr apply

5. Review and auto-fix in one shot

Trusted change, low risk, you want the loop to close itself:

rr run --auto --priority P0,P1

The reviewer runs first, then the fixer immediately remediates only matching priorities in a disposable worktree and hands the result back to you.

Auto Setup picks the reviewer model based on Factory.ai's public code review benchmark. See Why these models? below.

6. Pre-PR / team review workflow

Before opening a PR, run a base-branch review and let auto-fix clean up the obvious stuff. Anything left becomes review notes you can paste into the PR description.

rr run --base main --auto --priority P0
rr log -n 1            # grab the latest review log
rr log --json          # or pipe into your own tooling

For an org-wide loop, persisted review logs (rr log --json --global) make it easy to feed findings into dashboards or follow-up issues.

7. Triage with Interactive Mode

Run rr with no arguments to open Interactive Mode. It shows active sessions, recent history, review output, findings, fix results, and handoff status in a single view, which is convenient when several reviews are in flight.

Why these models? (Auto Setup model selection)

rr init with Auto Setup chooses your reviewer and fixer's model. That list is informed by Factory.ai's public code review benchmark, which scored 13 frontier and open-weight models against a golden set of 50 human-curated bugs across five real-world repositories (Sentry, Grafana, Keycloak, Discourse, and cal.com).

The reviewer model priority (highest first) is currently:

1. GPT-5.2
2. Claude Opus 4.6
3. Claude Sonnet 4.6
4. Claude Opus 4.7
5. GLM 5.1
6. GPT-5.3 Codex
7. Gemini 3.1 Pro Preview
8. Kimi K2.6

Models near the top scored highest on the benchmark for finding real bugs at a reasonable cost. The fixer priority is tuned separately and favors models that are strong at code edits rather than at finding issues.

For the methodology and full results, see Factory.ai's writeup: Which Model Reviews Code Best?.

You can always override the auto selection by running rr init and choosing Customize Setup, or by editing reviewer.model and fixer.model in your configuration directly.

Prerequisites

• Bun (runtime)
• tmux (background sessions)
• At least one supported agent CLI installed and authenticated

Ralph Review is a Bun-only TypeScript CLI. Use Bun for development and script execution.

Installation

# Homebrew (install or update)
brew install kenryu42/tap/ralph-review

# npm (install or update)
npm install -g ralph-review

# Or let Ralph Review detect the install method and update itself
rr update

For update checks without installing, run:

rr update --check

If install-source detection is ambiguous, force the package manager:

rr update --manager npm
rr update --manager brew

Quick Start

# Configure reviewer and fixer agents
rr init

# Start Interactive Mode
rr

# Start a review-only background session
rr run

# Review against a base branch
rr run --base main

# Review staged, unstaged, and untracked changes
rr run --uncommitted

# Review a specific commit
rr run --commit SHA

# Fix findings after review completes
rr fix --session SESSION_ID --all

# Review and immediately fix P0/P1 findings
rr run --auto --priority P0,P1

rrr is a shorthand alias for rr run. It starts a non-interactive review run without launching Interactive Mode.

Commands

| Command | Description | |---------|-------------| | rr | Launch Interactive Mode | | rrr | Alias for rr run | | rr init | Configure reviewer and fixer agents | | rr init --global | Write the user-global config | | rr init --local | Write repo-local overrides to .ralph-review/config.json | | rr run | Run review only and persist findings for later fixing | | rr run --base main | Review changes against a base branch or ref | | rr run --uncommitted | Review staged, unstaged, and untracked changes | | rr run --commit SHA | Review changes introduced by a specific commit | | rr run --max N | Set max review iterations | | rr run --force | Run all configured iterations even if no new findings appear | | rr run --auto | Run remediation immediately after review completes | | rr run --auto --priority P0,P1 | Auto-fix only findings with matching priorities | | rr run --sound | Play a completion sound for this run | | rr run --no-sound | Disable the completion sound for this run | | rr fix --session SESSION | Fix selected findings from a persisted review session | | rr fix --session SESSION --all | Select all persisted findings for remediation | | rr fix --session SESSION --priority P0,P1 | Select findings by priority | | rr fix --session SESSION --id F001 --id F003 | Select findings by ID | | rr apply | Apply a pending review handoff | | rr apply --session HANDOFF | Apply a specific pending handoff | | rr prune | Prune orphaned review session artifacts | | rr prune --dry-run | List prunable artifacts without deleting them | | rr prune --discard --session HANDOFF | Discard a pending handoff | | rr list / rr ls | List active review sessions | | rr stop | Stop a running review session | | rr stop --all | Stop all running review sessions | | rr log | View the latest review log for the current project | | rr log -n 5 | View the last 5 review logs | | rr log --json | Print current-project review logs as JSON | | rr log --json --global | Print review logs across all projects as JSON | | rr doctor | Run environment and configuration diagnostics | | rr doctor --fix | Auto-resolve supported diagnostic issues | | rr update | Check for and install a newer version | | rr update --check | Check for a newer version without installing |

You can append one positional custom instruction to rr run when an explicit review target is selected:

rr run --base main "focus on security boundaries"

Supported Coding Agents

| Agent | Link | |-------|------| | Claude Code | https://code.claude.com/docs/en/overview | | Codex | https://openai.com/codex/ | | Droid | https://factory.ai/ | | Gemini CLI | https://geminicli.com/ | | OpenCode | https://opencode.ai/ | | Pi | https://pi.dev |

You can assign different agents and models to the reviewer and fixer roles. For example, Codex can review while Claude or Gemini fixes.

Configuration

Run rr init to create configuration. Ralph Review supports a user-global config and repo-local overrides:

• Global config: ~/.config/ralph-review/config.json
• Repo-local overrides: .ralph-review/config.json

By default, rr config show displays the effective merged configuration for the current project.

# View effective configuration
rr config show

# View raw JSON
rr config show --json

# View only repo-local overrides
rr config show --local

# View one value
rr config get reviewer.agent

# Update global configuration
rr config set maxIterations 8

# Update repo-local configuration
rr config set --local defaultReview.branch main

# Edit configuration in $EDITOR
rr config edit
rr config edit --local

Useful settings include:

| Key | Purpose | |-----|---------| | reviewer | Agent, model, and reasoning used for review | | fixer | Agent, model, and reasoning used for remediation | | maxIterations | Maximum reviewer iterations per run | | iterationTimeout | Per-agent timeout in milliseconds | | defaultReview | Default review target, such as uncommitted changes or a base branch | | notifications.sound.enabled | Completion sound preference |

Run rr doctor to verify that your environment and configuration are valid. Add --fix to let it auto-resolve supported issues.

License

MIT