openclaw-superpowers

Design-first workflow plugin for OpenClaw — forces your agent to think before it codes.

No more "let me just write that real quick." Every task goes through brainstorming first.

English | 中文

English

What is this?

openclaw-superpowers is an OpenClaw lifecycle plugin that prevents your agent from jumping straight into code. When a task is detected, the agent is forced to:

1. Ask clarifying questions first (one at a time)
2. Propose 2-3 approaches with trade-offs
3. Get approval before writing any code

For bugs, it switches to systematic debugging (root cause first, no guessing).

Inspired by obra/superpowers and thedotmack/claude-mem, adapted for OpenClaw's plugin architecture.

Why?

OpenClaw agents are powerful but impulsive:

• Code before design — jumps into implementation without understanding requirements
• Config disasters — edits openclaw.json without reading it, breaks plugin IDs
• Repeating mistakes — no structured approach to debugging
• soul.md corruption — overwrites personality files without understanding structure

This plugin fixes all of that.

How it works

Inverted detection logic — instead of trying to detect "is this a task?" (fragile), we detect what is clearly NOT a task, and brainstorm everything else:

| Skipped (no brainstorming) | Everything else | |---|---| | Short confirmations (ok, yes, 好) | → brainstorm by default | | Slash commands (/compact) | | | Greetings (hi, 你好) | | | Pure emoji | |

Errors/bugs → systematic debugging mode instead of brainstorming.

Config file mentions → safety overlay injected automatically.

Features

| Feature | Description | |---|---| | Brainstorming | Hard gate: agent CANNOT write code until user approves design | | Systematic Debugging | 5-step root cause process, stops guessing after 3 failed fixes | | Config Safety | Auto-injected rules when modifying openclaw.json / soul.md / SKILL.md | | Verification | "No completion claims without fresh evidence" | | Planning | Structured implementation plans with no placeholders | | Phase Tracking | Auto-detects phase transitions: brainstorm → design → spec → plan → execute → verify | | Compaction Hints | Suggests /compact when conversation gets long (6+ turns in same phase) |

Memory Integration

Smart memory handling based on your setup:

| Setup | Recall | Capture | |---|---|---| | openviking plugin installed | openviking handles it (no duplication) | superpowers captures decisions | | No plugin, but memory.baseUrl set | superpowers handles recall | superpowers captures decisions | | Neither | disabled | disabled |

Installation

openclaw install openclaw-superpowers

Then in ~/.openclaw/openclaw.json:

{
  "plugins": {
    "allow": ["openclaw-superpowers"],
    "entries": {
      "openclaw-superpowers": {
        "enabled": true
      }
    }
  }
}

That's it. No memory config needed — auto-detected from openviking plugin if installed.

Configuration

| Key | Type | Default | Description | |---|---|---|---| | enabled | boolean | true | Enable/disable | | autoDetect | boolean | true | Auto-detect tasks and inject workflow | | memoryEnabled | boolean | true | Capture decisions to OpenViking (auto-detected) | | language | string | "auto" | Prompt language | | specDir | string | "superpowers/specs" | Spec document directory | | planDir | string | "superpowers/plans" | Plan document directory | | skipPatterns | string[] | [] | Additional regex patterns to skip |

Architecture

openclaw-superpowers/
├── index.ts              # Plugin entry: hooks + tools
├── detector.ts           # Inverted detection (skip non-tasks, brainstorm the rest)
├── state.ts              # Per-session phase tracking
├── memory.ts             # OpenViking integration (auto-detect config)
├── context-budget.ts     # Token budget + compaction hints
├── prompts/              # Compact injection prompts (~100 tokens each)
│   ├── brainstorming.ts
│   ├── writing-plans.ts
│   ├── verification.ts
│   ├── debugging.ts
│   ├── config-safety.ts
│   └── spec-reviewer.ts
├── openclaw.plugin.json
└── package.json

Credits

• obra/superpowers — Original superpowers skills by Jesse Vincent
• thedotmack/claude-mem — Memory lifecycle inspiration
• OpenViking — Long-term memory backend

Author

Bill Zhao — LinkedIn | GitHub

License

MIT

中文

这是什么？

openclaw-superpowers 是一个 OpenClaw 生命周期插件，强制 agent 在写代码前先思考。检测到任务时，agent 必须：

1. 先提问 — 逐个确认需求
2. 提方案 — 给出 2-3 个方案和推荐
3. 获得批准 — 用户同意后才能写代码

遇到 bug 时自动切换到系统化调试模式（先找根因，不猜测修复）。

为什么需要？

OpenClaw agent 很强但容易冲动：

• 不问就做 — 跳过需求确认直接写代码
• 配置灾难 — 不读就改 openclaw.json
• 重复犯错 — 没有结构化的调试方法
• 覆盖 soul.md — 不理解结构就改

工作原理

反转检测逻辑 — 不检测"是不是任务"（很难），而是检测"明确不是任务的东西"，其余全部走 brainstorming：

| 跳过 | 其他所有 | |---|---| | 短回复（ok、好、嗯） | → 默认 brainstorm | | 命令（/compact） | | | 打招呼（你好） | |

错误/bug → 自动切换调试模式 涉及配置文件 → 自动注入安全守则

功能

| 功能 | 说明 | |---|---| | 头脑风暴 | 强制：未批准设计前禁止写代码 | | 系统化调试 | 5 步根因流程，3 次修复失败后停下质疑架构 | | 配置安全 | 修改 openclaw.json/soul.md/SKILL.md 时自动注入规则 | | 完成验证 | 没有证据不能声称完成 | | 计划编写 | 结构化实施计划，无占位符 | | 阶段追踪 | 自动检测阶段转换 | | 压缩提示 | 同阶段 6+ 轮后建议 /compact |

记忆集成

智能记忆策略：

| 场景 | 召回 | 捕获 | |---|---|---| | 已装 openviking 插件 | openviking 负责（不重复） | superpowers 捕获决策 | | 没装插件，但配了 memory.baseUrl | superpowers 自己召回 | superpowers 捕获决策 | | 都没有 | 禁用 | 禁用 |

安装

openclaw install openclaw-superpowers

在 ~/.openclaw/openclaw.json 中添加：

{
  "plugins": {
    "allow": ["openclaw-superpowers"],
    "entries": {
      "openclaw-superpowers": {
        "enabled": true
      }
    }
  }
}

不需要配置记忆 — 自动从 openviking 插件检测。

配置

| 键 | 类型 | 默认值 | 说明 | |---|---|---|---| | enabled | boolean | true | 启用/禁用 | | autoDetect | boolean | true | 自动检测任务 | | memoryEnabled | boolean | true | 捕获决策到 OpenViking | | language | string | "auto" | 提示语言 | | specDir | string | "superpowers/specs" | 设计文档目录 | | planDir | string | "superpowers/plans" | 计划目录 | | skipPatterns | string[] | [] | 额外跳过模式 |

致谢

• obra/superpowers — Jesse Vincent 的原版 superpowers
• thedotmack/claude-mem — 记忆生命周期灵感
• OpenViking — 长期记忆后端

协议

MIT