← 返回 Skills 市场
252
总下载
0
收藏
1
当前安装
2
版本数
在 OpenClaw 中安装
/install adaptive-reasoning-plus
功能描述
自适应推理增强版。当用户要求"评估复杂度"、"需要深度分析"、"复杂问题"、"多步推理"时激活。 评估每次任务的复杂度,动态决定推理深度。 0-2分快速执行,3-5分标准处理,6-7分深度思考,8+分强制启用 extended thinking。
使用说明 (SKILL.md)
Adaptive Reasoning Plus — 自适应推理引擎
核心理念
不是所有问题都需要深度思考。
- 简单指令 → 直接执行,不废话
- 复杂问题 → 停下来想清楚再说
复杂度评分量表
收到任务时,扫描六个信号,累加计分:
| 信号 | 权重 | 触发条件 |
|---|---|---|
| 多步逻辑 | +3 | 需要多步规划/执行、调试链路、证明链、因果分析 |
| 歧义性 | +2 | 利弊权衡、"取决于"类问题、无标准答案 |
| 形式推理 | +2 | 数学计算、算法设计、逻辑推导、形式证明 |
| 新奇程度 | +1 | 首次遇到、无清晰先例、需要创造力 |
| 影响范围 | +1 | 涉及多个系统/模块/团队 |
| 不可逆性 | +2 | 不可逆操作、涉及真实资产/生产环境 |
⚠️ Tier-Override 规则:无论总分多少,不可逆性触发时自动升为8+强制推理。
推理策略(按分数段)
| 分数 | 策略 | 行动 |
|---|---|---|
| 0-2 | ⚡ 快速执行 | 直接响应,不做深度分析 |
| 3-5 | 🎯 标准处理 | 正常流程,适当思考后回答 |
| 6-7 | 🤔 深度思考 | 启用推理模式,多角度分析 |
| 8+ | 🧠 强制推理 | 开启 extended thinking,结构化输出 |
决策流程
收到任务
↓
扫描六个信号 → 累加评分
↓
总分判定 → 选择对应策略
↓
执行 → 必要时在末尾附图标
推理激活标注
当评分 ≥ 6 时,在回复中明确标注:
6-7 分(深度思考):
🤔 复杂度:X分 [信号A+3, 信号B+2 ...]
→ 策略:深度思考
[分析过程...]
---
[正式回复]
8+ 分(强制推理):
🧠 复杂度:X分 [信号A+3, 信号B+2 ...]
→ 策略:强制推理,已启用 extended thinking
[结构化分析...
维度一:...
维度二:...
维度三:...]
---
[正式回复]
适用场景
应该激活此技能时:
- 任务涉及多步规划或复杂决策
- 问题定义模糊,需要澄清或权衡
- 首次遇到的问题类型
- 不可逆操作前的分析
- 涉及多个系统/团队/业务线
不需要激活时:
- 简单查询或明确指令("查天气"、"发消息")
- 纯执行类任务("运行这个脚本")
- 闲聊或无需分析的问题
- 2分以下的简单操作
领域扩展
详见 references/domain-scoring.md
各领域增补信号示例:
- 金融/投资:本金亏损风险、流动性风险、监管政策变化
- 技术/运维:生产环境变更、数据迁移、安全漏洞
- 法务/合同:法律风险、合规违规
- 电商/运营:客诉风险、库存风险、资金冻结
落地验证标准
| 测试 | 输入 | 预期 |
|---|---|---|
| Test 1 | "帮我查天气" | 0分 → ⚡快速,不废话 |
| Test 2 | "这个文件删了" | 2分 → ⚡直接执行 |
| Test 3 | "帮我分析下这三个股票" | 6-7分 → 🤔深度思考 |
| Test 4 | "设计一个分布式系统架构" | 8+分 → 🧠强制推理+结构化 |
设计原则
- 通用优先 — 框架适用于所有任务类型和领域
- 可扩展 — 各领域可在通用框架上增补自己的信号
- 可量化 — 评分有明确标准,避免主观随意
- 显式标注 — 推理决策透明可见
安全使用建议
This skill appears to implement an adaptive 'when to think deeply' policy and otherwise has a small footprint, but take two precautions before enabling it: (1) Verify the skill's provenance: the ownerId in _meta.json does not match the registry ownerId you were shown — confirm the publisher and source. (2) Review and, if necessary, remove or soften the parts that require printing the full 'analysis process' or chain-of-thought when score ≥6; prefer publishing summaries or structured conclusions rather than raw internal reasoning to avoid disclosure of sensitive internal deliberations. If you allow the skill, test it with harmless prompts first and consider adding a guard that prevents inclusion of sensitive context in the analysis block.
功能分析
Type: OpenClaw Skill
Name: adaptive-reasoning-plus
Version: 1.0.2
The 'adaptive-reasoning-plus' skill bundle is a set of markdown-based instructions designed to guide an AI agent's reasoning depth based on task complexity. It contains no executable code, scripts, or network requests. The instructions focus on a scoring system for 'extended thinking' and do not attempt to exfiltrate data, bypass security controls, or perform unauthorized actions. Files analyzed: SKILL.md, _meta.json, and references/domain-scoring.md.
能力评估
Purpose & Capability
The name/description (adaptive reasoning depth) align with the SKILL.md: the skill is purely an instruction-layer scoring/decision framework for when to apply deeper reasoning. No unrelated binaries, env vars, or service access are requested. However, the package metadata shows an ownerId mismatch (_meta.json ownerId differs from the registry Owner ID in the provided manifest), which is an inconsistency that warrants verification of provenance.
Instruction Scope
The runtime instructions require the agent to 'scan six signals' from the task and, for scores ≥6, to include a detailed 'analysis process' in the reply. This explicitly requests exposing internal chain-of-thought/step-by-step reasoning. That can leak privileged reasoning or internal deliberations and may conflict with platform policies or the integrator's desire to keep internal reasoning private. The instructions do not tell the agent to read files, env vars, or external endpoints, so there is no obvious data-exfiltration step, but the explicit requirement to print analysis is the primary scope concern.
Install Mechanism
Instruction-only skill with no install spec and no code files to execute—low install risk. The references and SKILL.md are local files only; nothing is downloaded or written to disk by an installer.
Credentials
The skill requests no environment variables, no credentials, and no config paths. The level of requested access is proportionate to its stated purpose.
Persistence & Privilege
always:false and default autonomy settings are used. The skill does not request persistent system presence or to modify other skills' settings. No elevated privileges are requested.
如何使用
- 确保已安装 OpenClaw(本地或 Docker 部署)
- 在对话框中输入安装命令:
/install adaptive-reasoning-plus - 安装完成后,直接呼叫该 Skill 的名称或使用
/adaptive-reasoning-plus触发 - 根据 Skill 的参数说明提供必要输入,即可获得结构化输出
版本历史
v1.0.2
Claude Code架构学习后优化:完善复杂度评分标准
v1.0.1
Tier-Override规则:不可逆性触发自动升为8+强制推理。感谢所有参与此框架设计的团队成员。
元数据
常见问题
Adaptive Reasoning Plus 是什么?
自适应推理增强版。当用户要求"评估复杂度"、"需要深度分析"、"复杂问题"、"多步推理"时激活。 评估每次任务的复杂度,动态决定推理深度。 0-2分快速执行,3-5分标准处理,6-7分深度思考,8+分强制启用 extended thinking。 它是一个面向 Claude Code / OpenClaw 的 AI Agent Skill 插件,目前累计下载 252 次。
如何安装 Adaptive Reasoning Plus?
在 OpenClaw 或 Claude Code 对话框中运行命令「/install adaptive-reasoning-plus」即可一键安装,无需额外配置。
Adaptive Reasoning Plus 是免费的吗?
是的,Adaptive Reasoning Plus 完全免费,采用 MIT-0 许可证,可自由下载、安装和使用。
Adaptive Reasoning Plus 支持哪些平台?
Adaptive Reasoning Plus 跨平台运行,可在任意部署了 OpenClaw / Claude Code 的环境中使用(cross-platform)。
谁开发了 Adaptive Reasoning Plus?
由 tsangho(@tsangho)开发并维护,当前版本 v1.0.2。
推荐 Skills