Skip to content

AGENTS.md

Latest commit

 

History

History
264 lines (210 loc) · 10 KB

AGENTS.md

File metadata and controls

264 lines (210 loc) · 10 KB

AGENTS.md

Agent guidance for the Ralph CLI project.

Project Overview

Ralph CLI is a command-line tool for managing AI development workflows using the "Ralph Wiggum" technique - autonomous coding loops with AI agents.

Build & Test Commands

# Install dependencies
bun install

# Run in development mode
bun run dev

# Type check
bun run typecheck

# Link globally for testing
bun link

Code Conventions

  • TypeScript with strict mode
  • ES modules ("type": "module")
  • Use bun runtime (not Node.js)
  • Minimal comments, self-documenting code
  • No .js extensions in imports (Bun + moduleResolution: "bundler" handles this)

Architecture Overview

The codebase follows a layered architecture:

┌─────────────────────────────────────────────────────────┐
│                    CLI (index.ts)                       │
│              Command parsing via Commander              │
└─────────────────────┬───────────────────────────────────┘
                      │
┌─────────────────────▼───────────────────────────────────┐
│                   Commands Layer                        │
│     init.ts, plan.ts, build.ts, status.ts, stop.ts     │
│         Thin wrappers that delegate to services         │
└─────────────────────┬───────────────────────────────────┘
                      │
┌─────────────────────▼───────────────────────────────────┐
│                   Services Layer                        │
│  Planner, Builder, AgentRunner, HookDispatcher, etc.   │
│        Business logic and orchestration                 │
└─────────────────────┬───────────────────────────────────┘
                      │
┌─────────────────────▼───────────────────────────────────┐
│                    Domain Layer                         │
│   Workspace, Session, Implementation, Spec, Task       │
│           Core data models with behavior                │
└─────────────────────┬───────────────────────────────────┘
                      │
┌─────────────────────▼───────────────────────────────────┐
│                   Agents Layer                          │
│       BaseAgent + implementations (claude, amp, etc.)   │
│            AI agent command builders                    │
└─────────────────────────────────────────────────────────┘

Project Structure

src/
├── index.ts              # CLI entry point (commander setup)
├── types.ts              # TypeScript type definitions
├── constants.ts          # Markers, file names, defaults
│
├── commands/             # CLI command handlers (thin wrappers)
│   ├── init.ts           # ralph-wiggum-cli init
│   ├── plan.ts           # ralph-wiggum-cli plan
│   ├── build.ts          # ralph-wiggum-cli build
│   ├── status.ts         # ralph-wiggum-cli status
│   └── stop.ts           # ralph-wiggum-cli stop
│
├── services/             # Business logic and orchestration
│   ├── planner.ts        # Plan mode orchestration
│   ├── builder.ts        # Build mode orchestration (task loop)
│   ├── agent-runner.ts   # Spawns agent process, parses output
│   ├── hook-dispatcher.ts# Event system for listeners
│   ├── console-listener.ts # Console output for hook events
│   ├── notification-service.ts # Telegram notifications
│   └── logger-service.ts # Per-session/task file logging
│
├── domain/               # Core data models with behavior
│   ├── workspace.ts      # Project config, session manager
│   ├── session.ts        # Single run session (plan or build)
│   ├── session-manager.ts# Session CRUD operations
│   ├── implementation.ts # implementation.json wrapper
│   ├── spec.ts           # Spec entry with tasks
│   └── task.ts           # Individual task with status
│
├── agents/               # AI agent implementations
│   ├── base.ts           # Abstract BaseAgent class
│   ├── claude.ts         # Claude Code agent
│   ├── amp.ts            # Amp Code agent
│   ├── droid.ts          # Factory Droid agent
│   ├── pi.ts             # Pi agent
│   ├── opencode.ts       # OpenCode agent
│   ├── cursor.ts         # Cursor agent
│   ├── codex.ts          # Codex agent
│   ├── gemini.ts         # Gemini CLI agent
│   └── index.ts          # Agent registry
│
├── templates/            # Prompt templates
│   └── prompts.ts        # PROMPT_PLAN, PROMPT_BUILD, etc.
│
└── utils/
    ├── paths.ts          # Path utilities (.ralph-wiggum/ dirs)
    └── telegram.ts       # Telegram API client

Key Concepts

Domain Models

  • Workspace: Root model representing an initialized project. Loads/saves config.json, manages sessions.
  • Session: A single execution run (plan or build mode). Tracks status, iteration count, PID.
  • Implementation: Wraps implementation.json. Contains specs and provides nextPendingTask.
  • Spec: A specification with tasks. Tracks completion status.
  • Task: Individual unit of work. Has status (pending, in_progress, completed, blocked, failed).

Services

  • Planner: Orchestrates plan mode. Validates state, runs agent with plan prompt, handles signals.
  • Builder: Orchestrates build mode. Loops through tasks, generates per-task prompts, tracks progress.
  • AgentRunner: Spawns agent subprocess, pipes prompt via stdin, parses output for completion markers.
  • HookDispatcher: Event bus. Emits events like onLoopStarted, onIterationSuccess, onTaskBlocked.
  • ConsoleListener: Prints formatted output to console for hook events.
  • TelegramListener: Sends notifications for hook events.
  • LoggerService: Creates timestamped log files per session/task.

Hook Events

interface HookListener {
  onLoopStarted?(payload): void;
  onIterationStarted?(payload): void;
  onIterationSuccess?(payload): void;
  onIterationFailure?(payload): void;
  onTaskBlocked?(payload): void;
  onSpecCompleted?(payload): void;
  onLoopCompleted?(payload): void;
  onLoopStopped?(payload): void;
}

Task Completion Markers

Agents signal completion via stdout:

  • <TASK_DONE> - Task completed successfully
  • <TASK_BLOCKED reason="..."> - Task blocked

Supported AI Agents

Claude Code (Anthropic)

claude -p --dangerously-skip-permissions --output-format=stream-json --model <model>

# Models: sonnet, opus, or full model names
# Docs: https://docs.anthropic.com/en/docs/claude-code/cli-reference

Amp Code (Sourcegraph)

amp --execute --stream-json --mode <smart|rush>

# Modes: smart (full power), rush (faster/cheaper)
# Docs: https://ampcode.com/manual

Factory Droid

droid exec --auto high -o stream-json -m <model>

# Autonomy levels: low, medium, high
# Docs: https://docs.factory.ai/reference/cli-reference

Pi

pi --print --mode json --thinking high --model <model> --provider <provider>

# Docs: https://github.com/mariozechner/pi-coding-agent

Gemini CLI

gemini -p --auto

# Docs: https://ai.google.dev/gemini-api

Adding a New Agent

  1. Create src/agents/<name>.ts extending BaseAgent
  2. Implement buildCommand(), checkInstalled(), getInstallInstructions()
  3. Register in src/agents/index.ts
  4. Add to AgentType union in src/types.ts

Example:

// src/agents/myagent.ts
import { BaseAgent } from "./base";
import type { AgentCommand, AgentOptions } from "../types";

export class MyAgentAgent extends BaseAgent {
  type = "myagent" as const;
  name = "My Agent";

  buildCommand(options: AgentOptions): AgentCommand {
    const args = ["--auto", "--json"];
    if (options.model) args.push("--model", options.model);
    return { command: "myagent", args };
  }

  async checkInstalled(): Promise<boolean> {
    try {
      await Bun.$`which myagent`.quiet();
      return true;
    } catch {
      return false;
    }
  }

  getInstallInstructions(): string {
    return "npm install -g myagent";
  }
}

Project Directory Structure

All Ralph data is stored in the project's .ralph-wiggum/ folder:

.ralph-wiggum/
├── config.json          # Project config and session history
├── PROMPT_plan.md       # Planning mode prompt (customizable)
├── implementation.json  # Structured task tracking (generated by plan)
├── GUARDRAILS.md        # Compliance rules (before/after checks)
├── PROGRESS.md          # Audit trail (what was done)
├── specs/               # Specs with tasks + acceptance criteria
│   └── example.md       # Example spec template
└── logs/                # Session logs (gitignored)

When You Are Writing Specs

Always ask me questions so that you can strengthen your understanding of my requirements and ensure that the specifications are clear and accurate.