woda/laravel-ralph

Claude Code agent loop runner for Laravel

dev-master 2026-03-04 21:19 UTC

Requires

• php: ^8.2
• illuminate/console: ^11.0 || ^12.0
• illuminate/process: ^11.0 || ^12.0
• illuminate/support: ^11.0 || ^12.0

Requires (Dev)

• driftingly/rector-laravel: ^2.1
• larastan/larastan: ^3.0
• laravel/pint: ^1.0
• orchestra/testbench: ^9.0 || ^10.0
• pestphp/pest: ^3.0 || ^4.0
• pestphp/pest-plugin-laravel: ^3.0 || ^4.0
• phpstan/phpstan: ^2.0
• rector/rector: ^2.0
• systemsdk/phpcpd: ^8.0

MIT

• David Mondok

• laravel
• Agent
• loop
• claude

README

Claude Code agent loop runner for Laravel. Runs Claude CLI iteratively in a GNU screen session, tracking progress, detecting completion, and logging everything.

Requirements

• PHP 8.2+
• Laravel 11 or 12
• Node.js
• Claude CLI (claude in PATH)
• GNU screen (not needed with --once)

Installation

composer require woda/laravel-ralph

Setup

php artisan ralph:init

Configures .claude/settings.json with sandbox permissions so Claude doesn't hang waiting for bash approval.

Usage

# Interactive — choose from open GitHub issues, PRDs, or enter a prompt manually
php artisan ralph:start

# Work on a GitHub issue
php artisan ralph:start --issue=42

# Inline prompt
php artisan ralph:start --prompt="Refactor the payment service"

# Prompt from file
php artisan ralph:start --prompt=prompts/task.md

# Single iteration, foreground (no screen)
php artisan ralph:start --once --prompt="Fix the failing tests"

Options

Flag	Description
name	Session name (argument, optional)
--issue=N	Work on a GitHub issue
--prompt=TEXT|FILE	Inline prompt or path to prompt file
--iterations=N	Max iterations (default: 30)
--model=MODEL	Override Claude model
--budget=USD	Max spend per Claude invocation
--fresh	Don't resume — each iteration is independent
--resume	Resume a previous session
--attach	Auto-attach to screen after start
--once	Single iteration in foreground

Managing Sessions

# List running sessions
php artisan ralph:status
php artisan ralph:status --json
php artisan ralph:status --clean    # Remove dead entries

# Attach to a session
php artisan ralph:attach
php artisan ralph:attach my-feature

# Kill a session
php artisan ralph:kill my-feature
php artisan ralph:kill --all --force

# View logs
php artisan ralph:logs
php artisan ralph:logs --tail
php artisan ralph:logs --lines=100
php artisan ralph:logs --all        # List all log files

Configuration

php artisan vendor:publish --tag=ralph-config

config/ralph.php

'loop' => [
    'default_iterations'       => env('RALPH_LOOP_ITERATIONS', 30),
    'permission_mode'          => env('RALPH_PERMISSION_MODE', 'acceptEdits'),
    'model'                    => env('RALPH_MODEL'),
    'completion_marker'        => '<promise>COMPLETE</promise>',
    'max_consecutive_failures' => env('RALPH_MAX_CONSECUTIVE_FAILURES', 3),
],

'prompt' => [
    'default_file'  => env('RALPH_PROMPT_FILE'),
    'prd_path'      => 'prd/backlog',
    'suffix'        => 'Focus on one task at a time. Run tests after changes. ...',
    'continuation'  => 'Continue working. Check the PRD and progress files ...',
],

'screen' => [
    'prefix' => 'ralph',
    'shell'  => env('RALPH_SCREEN_SHELL', 'zsh'),
],

'tracking' => [
    'file' => env('RALPH_TRACKING_FILE', '.live-agents'),
],

'logging' => [
    'directory'               => storage_path('ralph-logs'),
    'non_json_warn_threshold' => 50,
],

'script_path' => null,

Prompt Customization

Ralph appends prompt.suffix to every initial prompt and uses prompt.continuation as the prompt for iterations 2+. There is no base system prompt — Ralph passes your content directly to claude -p and relies on Claude's defaults and your project's CLAUDE.md.

PRD Mode

Place PRDs at prd/backlog/{name}/project.md. Optionally add a progress.md alongside it. Ralph discovers these in the interactive menu and passes them as @file references to Claude.

How It Works

1. You provide a prompt (issue, PRD, file, or text).
2. Ralph writes the prompt to a temp file, appends the configured suffix.
3. A Node.js loop runner (ralph-loop.cjs) spawns claude -p with --output-format stream-json.
4. On iteration 1, a new session is created. On iterations 2+, the session is resumed with the continuation prompt.
5. Each iteration's output is scanned for the completion marker (<promise>COMPLETE</promise>).
6. The loop stops when: the marker is detected (exit 0), max iterations reached (exit 2), or consecutive failures exceed the threshold (exit 1).

All of this runs inside a GNU screen session so you can detach and reattach freely.

License

MIT