拉尔夫·介质

其他 社区 v1.0

技能解读

当前原文的自动摘译,尚未人工审校。

技能路由

使用 ralph-small · 使用 ralph-medium(这个) · 使用 ralph-huge 1-3 个步骤,快速操作 · 4-10 个步骤,中等 deps · 10+ 个步骤,复杂 deps 单命令/编辑 · 多文件编辑、配置 · 重构、调试、架构查找、简单问答 · 顺序工作流程 · 此处已失败 3 次 自动升级:同一子任务 3 次失败的独特尝试 →切换到 ralph-huge (/home/ubuntu/.openclaw/skills/ralph-huge/SKILL.md) 以获得完整的断路器 + 8 级枢轴梯。

1.无证据不得索赔

不要说“完成”/“已发送”/“已修复”,除非工具输出证明它是在这个回合中。不能三个都填吗? → 你还没有完成。

2. 没有孤儿承诺

“我会做 X”/“让我运行 Y”→ X/Y 的工具调用必须同时存在。无法执行? →“我不能[X],因为[拦截器]。我需要[东西]。”

3. 追踪尝试,强制多元化

记录每一次失败的尝试。下一次尝试必须更改至少其中一项: 工具(grep → 语义搜索) 参数(不同路径、ID、范围) 策略(编辑 → 重写、读取 → 创建) 源(配置 A → 配置 B)

4. 进展更新

开始之前(如果预计 4 个以上步骤): 每 3-4 个步骤: 完成时: 如果被阻止: 用户永远不必问“发生了什么事?”

流狐档案 运行与权限由流狐静态检查推断,不代表作者声明或实际测试结果。
领域
通用
作者声明的智能体
未找到明确声明;不据此推断已兼容或已测试
静态检查
92 / 100 · 启发式扫描,不代表运行安全
运行环境证据
未执行;当前没有隔离运行测试证据
作者 / 版本 / 许可证
@00Blacksheep00 · v1.0 · 未声明许可证
上下文用量评级
低消耗
接入复杂程度
需简单配置
是否需要外部接口密钥
不需要
兼容的系统
Linux
运行环境要求
未检出明确要求
检测到的文件与系统行为
  • 只读
  • 允许写入 / 修改
  • 命令执行
网络行为
仅限本地
安装命令数
无(仅作为资料)

需要注意: 未声明工具权限范围;实际权限由运行环境决定。

作者说明文件(原文) SKILL.md

# Ralph Medium — Mid-Weight Anti-Loop

> Do it, prove it, track it. If stuck, pivot — don't repeat.

## Skill Routing

| Use **ralph-small** | Use **ralph-medium** (this) | Use **ralph-huge** |
|---------------------|----------------------------|-------------------|
| 1-3 steps, quick ops | 4-10 steps, moderate deps | 10+ steps, complex deps |
| Single command/edit | Multi-file edits, configs | Refactors, debugging, architecture |
| Lookup, simple Q&A | Sequential workflows | Already failed 3 times here |

**Auto-escalation:** 3 failed unique attempts on the same sub-task → switch to ralph-huge (`/home/ubuntu/.openclaw/skills/ralph-huge/SKILL.md`) for full Circuit Breaker + 8-level Pivot Ladder.

---

## 6 Rules

### 1. No claim without proof

Don't say "done" / "sent" / "fixed" unless tool output proving it is in THIS turn.

```
CLAIM:  [What you say you did]
ACTION: [Tool that did it]
PROOF:  [Output line confirming success]
```

Can't fill all three? → You're not done.

### 2. No orphan promises

"I'll do X" / "Let me run Y" → tool call for X/Y MUST exist in same turn.
Can't execute? → "I can't [X] because [blocker]. I need [thing]."

### 3. Track attempts, force diversity

Log every failed attempt. Next try MUST change at least one of:
- **Tool** (grep → semanticsearch)
- **Params** (different path, ID, scope)
- **Strategy** (edit → rewrite, read → create)
- **Source** (config A → config B)

```
#1 [method] → [error]
#2 [different method] → [error]
#3 [different method] → [error] → ESCALATE to ralph-huge
```

Same method twice = Ralph loop violation. Change or escalate.

### 4. Progress updates

**Before starting** (if 4+ steps expected):
```
▶ Starting: [task] — ~[N] steps
```

**Every 3-4 steps:**
```
⏳ [X/Y] done — current: [what's happening]
```

**On completion:**
```
✅ Done: [summary] — Evidence: [proof]
```

**If blocked:**
```
⚠️ Blocked: [issue] — Need: [specific thing]
```

The user should never have to ask "what's happening?"

### 5. Act first, ask only if stuck

Have enough info (even partial)? → Do it, then confirm assumptions.
Missing ONE critical piece? → Ask exactly 1 question.
Never ask 2+ questions when you could act on the first one.

### 6. Three-attempt diagnostic

After **3 unique failed attempts** on the same sub-task, STOP:

```
⛔ 3 attempts exhausted

#1 [Method] → [Error]
#2 [Method] → [Error]
#3 [Method] → [Error]

Likely cause: [1-2 sentences]
Options:
→ [A: what you need from user]
→ [B: alternative approach]
→ [C: skip and continue]
```

Then WAIT for user input. Do not attempt #4 without new info.

---

## Quick Pivot Menu

When a method fails, pick next from:

1. **Different params** — same tool, change path/query/ID/scope
2. **Different tool** — grep → glob → semanticsearch
3. **Different strategy** — modify → rewrite → template
4. **Ask user** — 1 specific question, last resort

Between pivots: `"Method A failed: [reason]. Trying B: [1-line description]."` Then execute B immediately.

---

## Top 5 Anti-Patterns

| Pattern | Symptom | Fix |
|---------|---------|-----|
| **Ghost Completion** | "Done ✅" but nothing happened | No tool output = not done |
| **Groundhog Day** | Same error, same approach, 3 turns | Same error twice → forced pivot |
| **Promise Spiral** | "I'll do X" → next turn, no X | Promise = tool call in same turn |
| **Silent Struggle** | Working for ages, zero updates | Update every 3-4 steps |
| **Interrogation** | 4 questions, 0 actions | Act on what you have, ask max 1 |

---

## Closure: self-improvement

After you **verify** success on a **non-trivial** task (evidence in this turn): if you learned something worth replaying (pattern, failure mode, fix, gateway/tool quirk), open **`self-improvement`** and append to this workspace’s `.learnings/` (`LEARNINGS.md` / `ERRORS.md`) using its format — or promote to `TOOLS.md` / `AGENTS.md` when it’s recurring. Consider **skill extraction** only when `self-improvement` criteria apply or the user asks. **Skip** if nothing material. **Never** log secrets or full raw output. Do not delay the user-facing answer for long write-ups; a short inline log in the same turn is enough when lightweight.

Skill: `/home/ubuntu/.openclaw/skills/self-improvement/SKILL.md` — template for new skills: `.learnings/SKILL-TEMPLATE.md` in the workspace.

---

## Config

Tool-level loop detection: `tools.loopDetection` in `/home/ubuntu/.openclaw/openclaw.json`
Docs: `/home/ubuntu/.npm-global/lib/node_modules/openclaw/docs/tools/loop-detection.md`

讨论

基于 GitHub Discussions。登录 GitHub 即可参与讨论、点赞、订阅更新。