ralph-wiggum-loop（拉尔夫·维格姆循环）

Ralph Wiggum Loop - Enterprise Workflow Engine

Build powerful AI-driven development workflows that iterate until tasks succeed. Features a steering packet architecture for real-time workflow control, plugin system, agent orchestration, monitoring, and enterprise-grade workflow patterns including DAGs, state machines, and circuit breakers.

Prompt Architecture Overview

Ralph Loop uses stacked markdowns with a two-level architecture:

Level 1: System Context (Static - Always Loaded)

• AGENTS.md - Global project context for ALL agents (tech stack, conventions, how to run)
• System Prompts - Per-agent identity from config (who the agent is)
• These are NOT modified during execution

Level 2: Steering Prompts (Dynamic - Modified Each Iteration)

• PROMPT.md - Loaded by ALL agents (common execution context: current task, failures, evidence)
• PROMPT-BUILDER.md - Loaded ONLY by builder (builder workflow guide)
• PROMPT-VERIFIER.md - Loaded ONLY by verifier (verifier workflow guide)
• PROMPT-.md - Loaded ONLY by that agent role

How Stacking Works:

Builder loads:  AGENTS.md + PROMPT.md + PROMPT-BUILDER.md
Verifier loads: AGENTS.md + PROMPT.md + PROMPT-VERIFIER.md
Planner loads:  AGENTS.md + PROMPT.md + PROMPT-PLANNER.md

Key Principle:

• PROMPT.md = steering context (CHANGES each iteration)
• PROMPT-.md = workflow guide (OCCASIONALLY updated)
• AGENTS.md = global context (RARELY changes)
• System prompts = agent identity (NEVER changes during loop)

You steer by modifying PROMPT.md, NOT by changing system prompts.

See PROMPT_ARCHITECTURE.md for complete details.

Steering Packets: Core Architecture

Ralph Loop uses steering packets to control agent workflows in real-time. Based on research from Microsoft Azure, AWS Step Functions, and AI orchestration patterns.

What are Steering Packets?

Steering packets are structured state messages that carry:

• Workflow Context: Current state, history, and trajectory
• Agent Instructions: What to do, constraints, and success criteria
• Control Signals: Interrupt, redirect, pause, resume capabilities
• Fresh Evidence: Real-time data from tools, APIs, and external systems

flowchart LR
    A[Steering Packet] --> B[Workflow State]
    A --> C[Agent Instructions]
    A --> D[Control Signals]
    A --> E[Fresh Evidence]

    B --> F[Current Phase]
    B --> G[Iteration Count]
    B --> H[Failure History]

    C --> I[Task Description]
    C --> J[Success Criteria]
    C --> K[Constraints]

    D --> L[Continue]
    D --> M[Redirect]
    D --> N[Pause]
    D --> O[Interrupt]

Steering Packet Structure

{
  "version": "2.0",
  "timestamp": "2024-03-08T10:30:15Z",
  "trace_id": "abc123-def456",

  "workflow": {
    "name": "default",
    "iteration": 5,
    "phase": "building",
    "state": "active",
    "started_at": "2024-03-08T10:00:00Z"
  },

  "control": {
    "signal": "continue",
    "priority": "normal",
    "timeout": 7200,
    "can_interrupt": true,
    "can_redirect": true
  },

  "context": {
    "prior_thread": [
      {"role": "system", "content": "..."},
      {"role": "assistant", "content": "..."}
    ],
    "current_task": {
      "id": "task-5",
      "description": "Implement authentication",
      "from_todo": "line 42",
      "spec_ref": "section-4.2"
    },
    "completed_tasks": ["task-1", "task-2", "task-3", "task-4"],
    "failure_history": [
      {"task": "task-5", "attempt": 1, "error": "..."}
    ]
  },

  "evidence": {
    "files_changed": ["src/auth.ts"],
    "test_results": {"passed": 8, "failed": 2},
    "logs": ["..."],
    "metrics": {
      "duration_ms": 45000,
      "tokens_used": 15000
    }
  },

  "instructions": {
    "primary": "Complete task-5: Implement JWT authentication",
    "constraints": ["Use existing auth library", "Follow security best practices"],
    "success_criteria": ["All auth tests pass", "No security vulnerabilities"],
    "next_steps": ["Update TODO.md", "Write tests"]
  }
}

Real-Time Steering Capabilities

Interrupt: Stop current agent execution mid-flight

control:
  signal: interrupt
  reason: "Human override"
  preserve_context: true

Redirect: Change agent direction without restarting

control:
  signal: redirect
  new_phase: planning
  reason: "Implementation approach flawed"
  keep_evidence: true

Pause: Suspend execution for human review

control:
  signal: pause
  wait_for: human_approval
  timeout: 86400

Resume: Continue from checkpoint

control:
  signal: resume
  from_checkpoint: "iteration-5-phase-building"

Message-Level Signaling

Based on Microsoft Research on AI backend networks, Ralph Loop uses message-level signals to capture global workflow state:

Signal Types:

• heartbeat: Agent health check
• progress: Percentage complete updates
• checkpoint: State persistence point
• decision: Branch point reached
• error: Failure signal with context
• complete: Task finished

Signal Flow:

sequenceDiagram
    participant O as Orchestrator
    participant A as Agent
    participant P as Plugin

    O->>A: steering_packet (signal: start)
    A->>O: heartbeat (every 30s)
    A->>O: progress (25%)
    A->>P: checkpoint (save state)
    P->>O: metrics_update
    A->>O: decision (branch A vs B)
    O->>A: steering_packet (signal: redirect, branch: B)
    A->>O: progress (75%)
    A->>O: complete (success)

State Machine Transitions

Steering packets enable explicit state transitions with guard conditions:

stateDiagram-v2
    [*] --> Planning: Start
    Planning --> Building: Plan Complete
    Building --> Verifying: Build Success
    Building --> Planning: Build Failure [retries < 3]
    Building --> HumanReview: Build Failure [retries >= 3]

    Verifying --> Complete: Verified
    Verifying --> Building: Verification Failed
    Verifying --> HumanReview: Critical Issue

    HumanReview --> Building: Rejected
    HumanReview --> Verifying: Approved with Changes
    HumanReview --> Complete: Approved

    Complete --> [*]

Benefits of Steering Packets

1. Low-Latency Control: Interrupt and redirect without losing context
2. State Persistence: Recover from failures at exact checkpoint
3. Observability: Every packet is logged and traceable
4. Flexibility: Dynamic workflow adaptation
5. Reliability: Circuit breaker integration at packet level
6. Human-in-the-Loop: Pause and wait for human signals

Research Sources

• Microsoft Azure: State machines in AI conversations and dynamic transitions
• AWS Step Functions: State machine data and transitions
• Ably: Realtime steering with low-latency control signals
• Scale AI: State machines for structured workflows
• AWS Strands: Steering agents for modern workflows
• ArXiv SAGE: Steering dialog generation with state-action representations

Quick Start

# Generate a working Ralph loop with full plugin system
bash scripts/generate-ralph-loop.sh ./my-loop

cd my-loop
pip install -r requirements.txt

# Run with basic configuration
python ralph-loop.py loop --commit

# Run with custom workflow
python ralph-loop.py loop --workflow advanced.yaml --config ralph.yaml

What This Skill Provides

1. Core Workflow Engine

File: scripts/ralph-loop-example.py

A production-ready workflow engine with:

• State Machine Architecture: Explicit states with transitions
• Plugin System: Extensible hooks for custom behaviors
• Agent Orchestration: Multi-agent coordination with specialized roles
• Monitoring & Observability: Traces, metrics, and structured logging
• Failure Recovery: Retry policies, circuit breakers, compensating actions
• DAG Support: Complex dependency graphs for parallel execution
• Circuit Breaker: Prevent cascade failures

2. Plugin System

Extend the loop without modifying core code:

# ralph.yaml - Plugin configuration
plugins:
  - name: slack-notifier
    entry: plugins/slack.py
    hooks:
      - workflow.on_start
      - workflow.on_complete
    config:
      webhook: https://hooks.slack.com/services/...

  - name: custom-verifier
    entry: plugins/verifiers/security.py
    hooks:
      - phase.before_verify
      - phase.after_verify
    config:
      security_level: strict

  - name: metrics-exporter
    entry: plugins/metrics/prometheus.py
    hooks:
      - workflow.on_iteration
    config:
      endpoint: http://localhost:9090

Available Hooks:

• workflow.on_start - Loop initialization
• workflow.on_iteration - Each iteration start
• workflow.on_complete - Loop completion
• workflow.on_failure - Loop failure
• phase.before_build - Before build phase
• phase.after_build - After build phase
• phase.before_verify - Before verify phase
• phase.after_verify - After verify phase
• agent.on_invoke - Agent invocation
• state.on_change - State transitions

3. Advanced Workflow Patterns

State Machine Workflows:

# advanced.yaml
workflow:
  type: state_machine
  states:
    - planning
    - building
    - verifying
    - human_review  # Human-in-the-loop
    - merging
    - complete

  transitions:
    - from: planning
      to: building
      condition: plan_complete

    - from: building
      to: verifying
      condition: build_success
      else: planning  # On failure

    - from: verifying
      to: human_review
      condition: verified
      else: planning  # On failure

    - from: human_review
      to: merging
      condition: approved
      else: planning  # On rejection

DAG Workflows:

# Parallel execution with dependencies
workflow:
  type: dag

  nodes:
    - id: lint
      agent: linter

    - id: test
      agent: tester
      depends_on: [lint]

    - id: security
      agent: security_scanner
      depends_on: [lint]

    - id: build
      agent: builder
      depends_on: [test, security]

    - id: verify
      agent: verifier
      depends_on: [build]

Saga Pattern:

# Distributed transactions with compensation
workflow:
  type: saga

  steps:
    - name: update_database
      action: update_schema
      compensate: rollback_schema

    - name: migrate_data
      action: run_migration
      compensate: rollback_data

    - name: deploy_api
      action: deploy
      compensate: rollback_deployment

4. Agent Orchestration

Multi-Agent Coordination:

agents:
  builder:
    prompt: PROMPT-BUILDER.md
    timeout: 7200
    max_retries: 3
    retry_policy:
      strategy: exponential_backoff
      base_delay: 5s
      max_delay: 300s

  verifier:
    prompt: PROMPT-VERIFIER.md
    timeout: 3600
    parallel: true  # Can run in parallel

  planner:
    prompt: PROMPT-PLANNER.md
    timeout: 1800
    on_failure: always  # Called on any failure

  security_scanner:
    prompt: PROMPT-SECURITY.md
    condition: "file_contains: auth"  # Conditional invocation

Agent Chains:

chains:
  security_pipeline:
    - security_scanner
    - dependency_checker
    - secret_detector

  code_quality:
    - linter
    - type_checker
    - coverage_analyzer

5. Monitoring & Observability

OpenTelemetry Integration:

observability:
  provider: opentelemetry
  traces: true
  metrics: true
  logs: true

  exporters:
    jaeger:
      endpoint: http://localhost:16686
    prometheus:
      port: 9090

  dashboard:
    enabled: true
    type: grafana
    url: http://localhost:3000

Metrics Collected:

• Task duration, success/failure rates
• Agent iteration counts
• Token usage and costs
• Phase timing
• Queue depth
• Circuit breaker state

Structured Logging:

{
  "timestamp": "2024-03-08T10:30:15Z",
  "level": "INFO",
  "trace_id": "abc123",
  "span_id": "def456",
  "event": "phase_complete",
  "phase": "build",
  "agent": "builder",
  "duration_ms": 45000,
  "success": true
}

6. Failure Recovery

Retry Policies:

retry:
  default:
    strategy: exponential_backoff
    max_attempts: 3
    base_delay: 1s
    max_delay: 60s
    jitter: 0.1

  transient_errors:
    strategy: linear
    max_attempts: 5
    delay: 5s
    retry_on:
      - TimeoutError
      - ConnectionError

  critical:
    strategy: fixed
    max_attempts: 1  # No retry

Circuit Breaker:

circuit_breaker:
  enabled: true
  failure_threshold: 5
  timeout: 300s
  half_open_max_calls: 3

  # Actions on state change
  on_open:
    - notify_team
    - pause_loop

  on_close:
    - resume_loop

Compensating Actions:

# For saga workflows
compensation:
  on_failure:
    - action: git_reset
      args:
        hard: false

    - action: notify
      args:
        message: "Workflow failed, manual intervention required"

Generator Script

bash scripts/generate-ralph-loop.sh [options] ./my-loop

Options:
  --template <name>      Workflow template: basic, advanced, dag, saga
  --with-plugins         Include example plugin directory
  --with-dashboard       Include monitoring dashboard config
  --with-tests           Include test suite for the loop

Examples:
  bash scripts/generate-ralph-loop.sh ./basic-loop
  bash scripts/generate-ralph-loop.sh --template dag ./data-pipeline
  bash scripts/generate-ralph-loop.sh --with-plugins --with-dashboard ./enterprise-loop

Configuration

Basic Configuration (ralph.yaml)

# Core settings
loop:
  max_iterations: 100
  auto_commit: true
  parallel_agents: false

# Agent timeouts
timeouts:
  builder: 7200s
  verifier: 3600s
  planner: 1800s

# Simple retry
retry:
  max_attempts: 3
  delay: 5s

# Logging
logging:
  level: INFO
  file: .ralph/logs/ralph.log

Advanced Configuration

# Full enterprise configuration
project:
  name: My Enterprise Project
  description: AI-driven development workflow

# Plugin system
plugins:
  directory: ./plugins
  auto_discover: true

  registry:
    - name: core
      url: https://github.com/wojons/ralph-plugins

    - name: custom
      url: https://github.com/myorg/ralph-plugins

  installed:
    - name: slack-notifier
      version: ^1.0.0
      from: core

    - name: security-scanner
      version: ^2.1.0
      from: custom

# Workflow definition
workflow:
  type: state_machine

  states:
    - id: planning
      on_enter: [log_start]
      on_exit: [log_complete]

    - id: building
      timeout: 7200s
      on_timeout: [escalate_to_human]

    - id: verifying
      parallel: true
      agents: [verifier, security_scanner, linter]
      require_all: true  # All must pass

    - id: human_review
      type: human_in_the_loop
      timeout: 86400s  # 24 hours
      reminder_interval: 3600s

    - id: complete
      on_enter: [notify_success]

# Agent orchestration
agents:
  builder:
    prompt: PROMPT-BUILDER.md
    model: claude-3-opus
    temperature: 0.7

  verifier:
    prompt: PROMPT-VERIFIER.md
    model: claude-3-sonnet
    temperature: 0.3

  security_scanner:
    prompt: PROMPT-SECURITY.md
    model: claude-3-opus
    tools: [security_db, vulnerability_scanner]

  human_reviewer:
    type: human
    notification:
      slack: "#code-reviews"
      email: ["lead@company.com"]

# Circuit breaker
circuit_breaker:
  enabled: true
  failure_threshold: 5
  timeout: 300s

# Monitoring
observability:
  provider: opentelemetry

  traces:
    enabled: true
    sampling_rate: 1.0

  metrics:
    enabled: true
    exporters:
      - type: prometheus
        port: 9090

      - type: cloudwatch
        region: us-east-1

  logs:
    level: INFO
    format: json
    rotation:
      max_size: 100MB
      max_files: 10

  dashboard:
    enabled: true
    type: grafana
    config: ./grafana/dashboard.yaml

# Notifications
notifications:
  channels:
    slack:
      webhook: ${SLACK_WEBHOOK_URL}
      channel: "#builds"

    email:
      smtp: ${SMTP_SERVER}
      to: ["team@company.com"]

  events:
    on_start:
      - slack

    on_failure:
      - slack
      - email

    on_complete:
      - slack

    on_human_review_needed:
      - slack:
          mention: "@here"
      - email

# State management
state:
  persistence:
    type: database
    url: ${DATABASE_URL}

  snapshots:
    enabled: true
    interval: 60s
    retention: 30d

CLI Commands

# Basic operations
python ralph-loop.py loop              # Run full workflow
python ralph-loop.py loop --commit     # Auto-commit changes
python ralph-loop.py status            # Show current status
python ralph-loop.py logs              # View recent logs
python ralph-loop.py reset             # Reset loop state

# Workflow management
python ralph-loop.py workflow list                    # List available workflows
python ralph-loop.py workflow validate <file>          # Validate workflow YAML
python ralph-loop.py workflow visualize <file>         # Generate workflow diagram
python ralph-loop.py loop --workflow custom.yaml       # Use custom workflow

# Plugin management
python ralph-loop.py plugin list                      # List installed plugins
python ralph-loop.py plugin install <name>           # Install from registry
python ralph-loop.py plugin enable <name>            # Enable plugin
python ralph-loop.py plugin disable <name>            # Disable plugin

# Monitoring
python ralph-loop.py metrics                          # Show current metrics
python ralph-loop.py traces                          # Show recent traces
python ralph-loop.py dashboard                        # Open monitoring dashboard

# Debugging
python ralph-loop.py debug --iteration 5              # Debug specific iteration
python ralph-loop.py replay --from-state <snapshot>   # Replay from state
python ralph-loop.py inspect --phase build            # Inspect phase execution

Project Structure

my-project/
├── ralph.yaml                    # Main configuration
├── TODO.md                       # Tasks to complete
├── SPEC.md                       # Project specification
├── AGENTS.md                     # Project conventions
│
├── PROMPT-*.md                   # Agent instructions
│   ├── PROMPT-BUILDER.md
│   ├── PROMPT-VERIFIER.md
│   ├── PROMPT-PLANNER.md
│   └── PROMPT-SECURITY.md       # Custom agent
│
├── workflows/                    # Workflow definitions
│   ├── default.yaml
│   ├── ci.yaml
│   ├── release.yaml
│   └── custom/
│
├── plugins/                      # Custom plugins
│   ├── __init__.py
│   ├── slack.py
│   ├── security.py
│   └── metrics/
│       └── prometheus.py
│
├── .ralph/                       # Loop state
│   ├── loop-state.yaml
│   ├── snapshots/
│   ├── logs/
│   └── cache/
│
└── monitoring/                   # Observability config
    ├── grafana/
    ├── prometheus/
    └── jaeger/

Agent Prompts

The skill includes specialized PROMPT.md files at the project root:

PROMPT-BUILDER.md

Implementation agent with:

• Mermaid charts: Build workflow and decision trees
• Lifecycle phases: Discovery → Planning → Implementation → Completion
• Standards: Code quality, testing requirements
• Edge cases: What to do when stuck
• Plugin hooks: How to interact with plugins

PROMPT-VERIFIER.md

Quality assurance agent with:

• Mermaid charts: Multi-layer verification flow
• 5 verification layers: Functional → Code → Spec → Edge Cases → Integration
• Failure patterns: Common issues to catch
• Report formats: Detailed PASS/FAIL reports

PROMPT-PLANNER.md

Strategic planning agent with:

• Mermaid charts: When planner is called
• Root cause analysis: Framework for diagnosis
• Solution patterns: Clarify, restructure, fix context, strategic
• Decision framework: When to split vs clarify

Custom Agents

Create custom agents by adding PROMPT-<NAME>.md:

# ralph.yaml
agents:
  security_scanner:
    prompt: PROMPT-SECURITY.md
    hooks:
      - phase.before_build
      - phase.after_build

Plugin Development

Plugin Structure

# plugins/my_plugin.py
from ralph.plugins import Plugin, hook

class MyPlugin(Plugin):
    name = "my-plugin"
    version = "1.0.0"

    def __init__(self, config):
        self.config = config

    @hook("workflow.on_start")
    def on_start(self, context):
        """Called when loop starts"""
        print(f"Starting workflow: {context.workflow_name}")

    @hook("phase.after_build")
    def after_build(self, context):
        """Called after build phase"""
        if context.success:
            self.send_notification("Build successful!")

    @hook("workflow.on_failure")
    def on_failure(self, context):
        """Called on workflow failure"""
        self.escalate(context.error)

Plugin Configuration

# In ralph.yaml
plugins:
  - name: my-plugin
    entry: plugins.my_plugin.MyPlugin
    config:
      api_key: ${API_KEY}
      endpoint: https://api.example.com

Available Context Objects

Workflow Context:

{
    "workflow_name": "default",
    "iteration": 5,
    "started_at": "2024-03-08T10:30:15Z",
    "config": {...},  # Full ralph.yaml
}

Phase Context:

{
    "phase": "build",
    "agent": "builder",
    "iteration": 5,
    "started_at": "2024-03-08T10:30:15Z",
    "duration_ms": 45000,
    "success": True,
    "output": "...",
    "metrics": {
        "tokens_used": 15000,
        "cost": 0.45
    }
}

State Context:

{
    "from_state": "planning",
    "to_state": "building",
    "trigger": "plan_complete",
    "timestamp": "2024-03-08T10:30:15Z"
}

Monitoring Dashboard

The skill provides a built-in monitoring dashboard:

# Start dashboard
python ralph-loop.py dashboard

# Open in browser
open http://localhost:8080

Dashboard Views:

• Overview: Current status, recent iterations, success rate
• Timeline: Phase execution timeline with traces
• Metrics: Token usage, costs, duration trends
• Logs: Structured logs with filtering
• Agents: Agent performance and health
• Plugins: Plugin status and metrics

Advanced Features

Human-in-the-Loop

workflow:
  states:
    - id: human_review
      type: human

      notification:
        slack: "#reviews"
        email: ["reviewer@company.com"]
        mention: "@here"

      ui:
        type: web
        template: review_form.html

      timeout: 86400s
      reminder_interval: 3600s

      on_timeout:
        action: auto_approve
        notify: true

Conditional Execution

agents:
  security_scanner:
    condition: "file_changed: '*.py' and not file_changed: 'test_*'"

  performance_test:
    condition: "branch == 'main' and tag =~ 'v.*'"

Parallel Execution

workflow:
  type: dag

  nodes:
    - id: lint
      agent: linter

    - id: unit_tests
      agent: tester
      depends_on: [lint]

    - id: integration_tests
      agent: integration_tester
      depends_on: [lint]

    - id: security_scan
      agent: security_scanner
      depends_on: [lint]

    - id: build
      agent: builder
      depends_on: [unit_tests, integration_tests, security_scan]

State Snapshots

state:
  snapshots:
    enabled: true

    triggers:
      - phase_complete
      - iteration_complete
      - before_human_review

    retention:
      count: 100
      age: 30d

Examples

CI/CD Pipeline

# .github/workflows/ralph.yaml
name: AI-Driven CI

on: [push, pull_request]

jobs:
  ralph:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v3

      - name: Run Ralph Loop
        run: |
          pip install ralph-loop
          ralph-loop loop --workflow ci.yaml

Custom Workflow

# workflows/security-first.yaml
workflow:
  type: state_machine

  states:
    - security_scan
    - planning
    - building
    - verifying
    - complete

  transitions:
    - from: security_scan
      to: planning
      condition: security_passed
      else: stop

    - from: building
      to: verifying
      condition: build_success

    - from: verifying
      to: complete
      condition: verified
      else: building

Notes

• Reliability: Durable execution with state persistence
• Extensibility: Plugin system without core modification
• Observability: OpenTelemetry standard for vendor flexibility
• Failure Handling: Classify errors, exponential backoff, circuit breakers
• State Management: Explicit state machines with persistence
• Extensibility: Clear extension points via plugins

Resources

• Documentation: https://github.com/wojons/skills/tree/main/skills/ralph-wiggum-loop
• Plugin Registry: https://github.com/wojons/ralph-plugins
• Examples: https://github.com/wojons/ralph-examples
• Community: https://github.com/wojons/skills/discussions