Principles Agent

基于第一性原理思维的迭代式 Agent，拆解复杂目标为原子任务，自动验证调试并依赖感知执行保证最终交付质量。

Security Scan

The skill's code and runtime instructions are consistent with its stated purpose (an orchestration / task-decomposition agent that relies on an injected llm_call), with no requests for credentials, network access, or unusual install steps.

Purpose & Capability

Name/description match the code: modules implement goal clarification, truth derivation, task breaking, refinement, dependency sorting, execution via an injected llm_call, validation, and integration. The skill declares no external credentials or network use, which aligns with the implementation that expects an injected llm_call from the OpenClaw host.

Instruction Scope

SKILL.md states the skill never calls network or reads environment variables; the code follows that pattern by expecting an injected llm_call and not performing outbound requests or reading env vars. Minor implementation inconsistencies: cli.py expects a global 'llm_call' (will raise if run outside an OpenClaw session), and some library methods expose APIs that build prompts and then call the local _parse_response on the prompt (i.e., parse the prompt as though it were an LLM response) — these are likely bugs but not malicious behavior. The skill writes files only when the user supplies --output.

Install Mechanism

No install spec provided (instruction-only install), and there are no downloads or package installs. The skill ships source files but does not declare or perform any network-based installation — low install risk.

Credentials

The skill requires no environment variables, no credentials, and no config paths. All LLM access is expected to be provided by the host via a callable 'llm_call'. This is proportionate to the stated orchestration purpose.

Persistence & Privilege

The skill does not request persistent or elevated privileges (always: false). It does not modify system configuration or other skills. File writes are user-initiated via --output only.

Assessment

This skill appears coherent and consistent with its description: it orchestrates decomposition and delegates LLM calls to the host. Before installing, note: (1) run it inside OpenClaw or provide a proper llm_call function — the CLI will raise if llm_call isn't injected; (2) the code contains some minor implementation issues (some methods build prompts then parse the prompt as if it were an LLM response), which are likely bugs rather than malicious behavior; (3) the skill writes a report only when you pass --output (so be cautious about the output path you provide). If you plan to run this in production, review the code for robustness around JSON extraction and error handling and test it in a sandboxed OpenClaw session first.

principles-agent - 基于第一性原理的迭代式任务拆解框架

作者： 🖤 小黑 & 小龙

描述

基于 miltonian/principles 的设计思想，在 OpenClaw 中实现的第一性原理思维 + 迭代式开发框架。

架构职责分离：

模块	职责
principles-agent	需求澄清 → 第一性原理推导 → 原子任务拆解 → 迭代精炼 → 依赖排序 → 结果整合 → 分层验证
OpenClaw 主 Agent	实际执行 LLM 调用 → 执行子任务 → 处理外部操作 → 权限管控

skill 不直接调用 API、不访问网络、不读取环境变量，所有 IO 由主 Agent 完成。

核心能力

1. 第一性原理需求分析 —— 剥离经验假设，回归最基础的事实和公理
2. 原子级任务拆解 —— 将复杂目标拆解为不可再分、可验证的原子子任务
3. 迭代式精炼 —— 自动检查并优化不满足条件的真理/子任务
4. 依赖感知调度 —— 拓扑排序处理依赖，识别可并行批次
5. 分层质量验证 —— 每层都验证，不通过反馈优化
6. 全局整合验收 —— 最终交付前整体验证，保证结果满足目标

触发条件

当用户需要：

• 从零设计一个复杂系统/项目
• 对复杂问题进行深度分析
• 构建多步骤多 Agent 协作系统
• 需要按照第一性原理思维解决问题时触发

用法

直接在 OpenClaw 会话中使用：

principles "你的目标或问题描述"

可选输出到文件：

principles "你的目标" --output report.md

配置最大迭代次数：

principles "你的目标" --max-iterations 5

权限边界

principles-agent 遵循最小权限原则：

• ✅ LLM 调用：由 OpenClaw 主 Agent 注入，skill 只接收结果
• ✅ 文件读取：仅读取 skill 自身目录内的源代码
• ✅ 文件写入：仅在用户通过 --output 指定时写入报告
• ❌ 网络访问：skill 不发起任何网络请求
• ❌ 环境变量：skill 不读取任何环境变量
• ❌ 系统访问：不执行外部命令，不修改系统配置

工作流程

1. 接收需求 → 澄清目标，提取约束和成功标准
2. 第一性原理解构 → 推导基础事实/公理
3. 原子任务拆解 → 拆解为最小可验证子任务
4. 迭代精炼 → 检查最小性/可行性/对齐性，不满足则优化
5. 依赖排序 → 拓扑排序确定执行顺序，检测循环依赖
6. 分批执行 → 主 Agent 执行子任务，skill 验证每个结果
7. 全局整合 → 合并所有子任务结果生成完整输出
8. 最终验收 → 验证满足原始目标，交付成果

架构

principles-agent/
├── SKILL.md          # 本文件
├── package.json      # Skill 元数据和权限声明
├── cli.py            # OpenClaw CLI 入口
├── openclaw_entry.py # OpenClaw 技能入口点
└── src/
    ├── __init__.py
    ├── types_def.py           # 数据类型定义
    ├── orchestrator.py       # 总控编排
    ├── truth_deriver.py      # 基础事实推导
    ├── task_breaker.py       # 原子任务拆解
    ├── refiner.py            # 迭代精炼决策
    ├── dependency_sorter.py  # 依赖拓扑排序
    ├── validator.py          # 分层质量验证
    ├── executor.py           # 执行调度（调用主 Agent 执行）
    └── integrator.py         # 结果整合

参考资料

• 原始项目：https://github.com/miltonian/principles