Context Booster

Use this skill when managing long-context conversations and tasks. Provides intelligent compression, key extraction, context enhancement, memory retrieval, a...

Security Scan

The skill's purpose (long-context management) is plausible, but the documentation and runtime instructions claim local persistence and 'capture of complete execution context' while there is no installable code provided (despite package.json/CLI references) and the capture semantics are vague — this mismatch and the potential to collect sensitive context make the package suspicious.

Purpose & Capability

Name and description match the documented features (compression, retrieval, SQLite persistence). However, package.json and README advertise a Node.js package and CLI (context-booster / cb) even though no source code or binaries are present in the bundle. That mismatch (instruction-only skill that nevertheless documents an installable package) is incoherent and could lead to runtime failures or surprising behavior if the agent assumes a CLI exists.

Instruction Scope

SKILL.md repeatedly instructs the agent to 'capture the complete execution context' and calls helpers like getEcosystemState(), captureContext(), and SessionStart/Stop hooks. Those terms are vague: ‘complete context’ can include other skills' data, environment variables, and user-sensitive content. The FAQ/README specify a persistent SQLite DB at ~/.openclaw/context-booster/context.db and export commands — meaning the skill will read/write persistent local data and can export it. The instructions do not clearly limit what fields are captured or require explicit consent for sensitive items, creating a risk of unintended data collection/exfiltration.

Install Mechanism

There is no install spec in the registry and no code files shipped, making this an instruction-only skill. Yet README/package.json describe npm install, a main entry (src/index.js), and CLI usage. That inconsistency is a red flag: the runtime instructions assume binaries/JS code exist, but the bundle provides only docs. This could cause the agent to attempt commands that don't exist, or a malicious/incorrect install step might be introduced later to satisfy the documented behavior.

Credentials

The skill declares no required environment variables or credentials (good). However, it asks to persist context to a local SQLite DB in the user's home directory and to read 'ecosystemState' and full execution context, which can include environment/config data. While no explicit secret requests are present, the informal instruction to capture broad context increases the chance of recording sensitive env/config data unintentionally. The documentation's claim 'does not upload to external servers' mitigates network exfil but is self-reported and not enforced by the bundle.

Persistence & Privilege

The skill intends to persist data across sessions (SQLite at ~/.openclaw/context-booster/context.db) and provides CLI cleanup/export commands. It does not set always: true and does not request elevated system privileges in metadata. Persisting a local DB is expected for a memory/context skill, but users should be aware this creates a durable store of potentially sensitive conversations unless encryption/permissions are verified (enterprise version mentions encryption as a paid feature).

What to consider before installing

This package looks functionally coherent with its stated goal but has important inconsistencies and ambiguous capture behavior you should address before installing. Ask the publisher for: (1) the actual source code (src/index.js and any CLI) so you can review what captureContext/getEcosystemState do; (2) an explicit install spec (how the CLI is installed) instead of only docs; (3) a clear data schema showing exactly which fields are captured and persisted; (4) whether the DB is encrypted by default and how to delete or export stored contexts safely. Until you can review the code, avoid using it with sensitive conversations or granting the agent access to system files or other skills' data. If you proceed, restrict filesystem permissions to the stated DB path and require user consent before any export or cross-session capture.

context-booster - 上下文增强器

版本：v0.1.0
定位：L1 增强层 - ECC 式上下文管理，解决 AI 记忆限制

---

📖 技能说明

context-booster 是一个上下文增强引擎，能够：

1. 捕获执行上下文 - 记录技能执行前后的完整上下文
2. 压缩记忆 - 智能压缩历史对话，保留关键信息
3. 语义检索 - 基于语义检索相关上下文
4. 注入上下文 - 在执行前注入相关上下文
5. 状态持久化 - SQLite 存储，支持跨会话持久化

核心价值：

• 解决 AI 上下文窗口限制
• 保持长对话的连贯性
• 提升复杂任务的执行效果

---

🎯 使用场景

场景	示例
长对话	「继续我们昨天讨论的技能进化方案」
复杂任务	「帮我完成这个多步骤的研究任务」
跨会话	「继续上次未完成的技能创建」
上下文丢失	「我刚才说的需求是什么？」

---

🚀 使用方法

方式 1：直接调用

请增强当前上下文：
[任务描述]

方式 2：检索上下文

检索相关上下文：
[关键词或主题]

方式 3：压缩记忆

压缩过去 24 小时的对话记忆

---

📋 核心功能

1. 上下文捕获

// 捕获执行上下文
const context = await captureContext({
  sessionId: 'session-123',
  skillId: 'first-principle-analyzer',
  input: userInput,
  ecosystemState: getEcosystemState()
});

2. 记忆压缩

// 压缩历史对话
const compressed = await compressMemory({
  timeRange: 'last-24-hours',
  retentionRate: 0.3,  // 保留 30% 关键信息
  method: 'semantic'   // 语义压缩
});

3. 语义检索

// 检索相关上下文
const related = await retrieveContext('技能进化', {
  limit: 5,
  minRelevance: 0.7
});

4. 上下文注入

// 注入到执行环境
await injectContext(executionId, related, {
  priority: 'high',
  format: 'structured'
});

---

🔧 技术架构

ECC 模式应用

基于 ECC 的上下文管理模式：

1. 执行前捕获 - SessionStart Hook
2. 执行后总结 - SessionStop Hook
3. SQLite 存储 - 持久化状态
4. 语义索引 - 快速检索

数据结构

interface ContextRecord {
  id: string;
  sessionId: string;
  skillId?: string;
  
  // 输入/输出
  input: any;
  output?: any;
  
  // 压缩后的摘要
  summary: string;
  keywords: string[];
  
  // 语义向量（用于检索）
  embedding?: number[];
  
  // 时间
  createdAt: Date;
  expiresAt?: Date;
  
  // 元数据
  importance: number;  // 重要性 0-1
  accessCount: number; // 访问次数
}

---

📊 效果评估

评估维度

维度	说明	目标
压缩率	原始大小/压缩后大小	≥3:1
检索准确率	检索结果相关性	≥85%
响应时间	检索 + 注入时间	<500ms
上下文完整度	关键信息保留率	≥90%

使用示例

# 上下文增强效果报告

## 压缩效果
- 原始对话：10,000 tokens
- 压缩后：3,000 tokens
- 压缩率：3.3:1 ✓

## 检索效果
- 查询："技能进化"
- 检索结果：5 条
- 相关率：100% ✓

## 执行效果
- 响应时间：320ms ✓
- 关键信息保留：95% ✓

---

⚠️ 使用注意

何时使用

• ✅ 长对话需要保持连贯性
• ✅ 复杂任务需要多步骤上下文
• ✅ 跨会话继续未完成任务
• ✅ 上下文窗口即将耗尽

何时不使用

• ❌ 短对话（<10 轮）
• ❌ 简单任务（单步骤）
• ❌ 敏感信息（需特殊处理）

---

📚 版本演进

当前版本：v0.1.0

• 基础上下文捕获
• 记忆压缩（语义）
• 语义检索
• SQLite 存储

下一版本：v0.2.0 (规划中)

• 自动过期清理
• 多会话关联
• 上下文可视化

---

创建时间：2026-03-30
创建者：王的奴隶 · 严谨专业版
版本：v0.1.0
状态：初始版本

---

💰 购买与授权

本技能版本：v0.1.0（基础版）

授权类型：

• 个人版：$49.9（永久使用，个人非商业用途）
• 商业版：$149.9（永久使用，商业用途）
• 企业版：$499.9（永久使用 + 定制支持）

购买后获得：

• ✅ 技能完整源码
• ✅ 使用文档
• ✅ 1 年免费更新
• ✅ 邮件支持

联系方式：

• 📧 邮箱：[e-mail]
• 💬 微信：CloudShrimpSupport（备注：技能名称）

---

技能版本：v0.1.0
创建时间：2026-03-30
维护者：王的奴隶 · 严谨专业版
授权：商业许可（购买后）