AgentSkillsCN

ralph-run

启动 Ralph Ultra 自主开发循环。系统会智能处理失败场景,基于技能进行验证,并设置检查点,不断迭代 PRD 中的各项故事。当您准备好以自主方式执行 PRD 计划时,即可启用此功能。

SKILL.md
--- frontmatter
name: ralph-run
description: Start the Ralph Ultra autonomous development loop. Iterates through PRD stories with smart failure handling, skill-based verification, and checkpointing. Use when ready to execute a PRD plan autonomously.
allowed-tools: Read, Write, Edit, Bash, Grep, Glob

Ralph Ultra Autonomous Loop

Start the smart adaptive development loop that iterates through PRD stories.

What this does

  1. Pre-flight checks — Runs environment-doctor and deep-codebase-navigator
  2. Story selection — Picks next story respecting dependencies and failure history
  3. Prompt construction — Builds context from CLAUDE.md, skills, progress, and error history
  4. Execution — Invokes Claude/AMP with appropriate security flags
  5. Verification — Runs post-iteration skills (security-auditor, performance-guardian)
  6. Failure handling — Circuit breaker at 3 failures, error injection for self-correction
  7. Checkpointing — Saves state after each iteration for resumability

Usage

code
/ralph-ultra:ralph-run [--max-iterations N] [--security sandbox|standard|yolo] [--tool claude|amp] [--resume]

Options

OptionDefaultDescription
--max-iterations10Maximum loop iterations
--securitystandardSecurity mode (sandbox/standard/yolo)
--toolclaudeAI tool to use
--resumefalseResume from last checkpoint

Loop Flow

code
Pre-flight Skills → Pick Story → Build Prompt → Execute →
  ↓ Success: Mark done, run post-iteration skills → Next story
  ↓ Failure: Record error, inject context → Retry (max 3) → Block story

Security Modes

  • sandbox: Read-only analysis, no file modifications
  • standard: Allowlist-based, only permitted commands
  • yolo: Full permissions (requires explicit confirmation)

Monitoring

While the loop runs:

  • View live dashboard: /ralph-ultra:ralph-dashboard
  • Check status: /ralph-ultra:ralph-status
  • Session logs: .ralph-ultra/sessions/current/log.jsonl