Smart Compact

Smart context compaction for OpenClaw agents. 4-phase progressive strategy: Scan, Extract, Check, Compact. Before running /compact, this skill scans tool out...

Security Scan

The skill's purpose (pre-check and preserve important conversation data before /compact) is coherent, but its instructions permit persistent storage of potentially sensitive data (including an explicit example of '登录凭据' / credentials) while simultaneously claiming redaction — an internal inconsistency that elevates risk.

Purpose & Capability

Name and description match the instructions: a 4‑phase pre-compact flow that scans tool outputs and writes extracted items to memory files. Asking the agent to inspect tool outputs (exec/read/web_fetch/web_search) and write to memory/YYYY-MM-DD.md is proportionate to the stated goal. However, the README/SKILL explicitly lists credentials/login data as an example of 'must save', which is unexpected for a compaction helper and inconsistent with other claims about redaction.

Instruction Scope

Instructions direct the agent to scan all recent tool call outputs (exec, read, web_fetch, web_search, etc.) and append extracted items to memory files. This is within the stated scope, but it also authorizes persisting potentially sensitive items (addresses, file paths, and explicitly '登录凭据s'). The SKILL promises redaction but does not define a verifiable redaction mechanism or thresholds, so the agent could persist secrets if redaction is imperfect or absent. The instructions also rely on an 'edit' append tool — behavior and permissions of that tool are not specified here.

Install Mechanism

Instruction-only skill with no install spec or code files; lowest install risk. README suggests optional cloning from GitHub or curl to download SKILL.md, which are normal but require verifying the repository source before fetching code.

Credentials

The skill declares no required env vars or credentials (good), but it writes persistent files containing extracted data. The explicit examples show it may store network addresses, config values, and even credentials. Persisting credentials is disproportionate for a context compaction helper and raises risk if redaction fails. No mechanism is provided to scope which categories are saved automatically vs require user approval beyond the final compact confirmation.

Persistence & Privilege

The skill does persistent writes to memory/YYYY-MM-DD.md (append-only). It does not request always: true or other elevated runtime privileges, and it claims not to auto-compact. Append-only behavior and user confirmation before /compact reduce some risk, but persistent storage of sensitive data still increases blast radius if misused or if file permissions are lax.

What to consider before installing

This skill mostly does what it says (scan → extract → checklist → optionally compact) and is low-risk in install footprint, but there is a clear inconsistency: the docs both promise 'sensitive data will be redacted' and list '登录凭据 (credentials)' as an example of items to save. Before installing or using: 1) Inspect the memory/YYYY-MM-DD.md files it creates and their filesystem permissions; 2) Run it in 'compact check' mode only and verify that secrets are not saved; 3) Confirm what 'redaction' actually does (sample inputs and outputs); 4) Avoid running it on conversations containing real credentials or secrets until you are confident redaction works; 5) If you plan to install via the suggested GitHub repo, review that repo (or clone via a secure channel) rather than blindly curl'ing raw files. If redaction is unreliable or credentials are being persisted, do not use the skill for sensitive contexts.

Smart Compact — 智能压缩增强

四阶段渐进式压缩策略，在 /compact 前先把重要信息救出来。

什么时候用

• 用户说"智能压缩"、"smart-compact"、"压缩检查"
• 在手动执行 /compact 之前先跑一遍
• 对话上下文快满时，主动触发
• Heartbeat 检测到 context 接近 80% 时自动建议

核心理念

传统的上下文压缩是一刀切——整个对话被浓缩成一段摘要，大量细节在过程中丢失。

Smart Compact 采用四阶段渐进式策略，在 /compact 之前插入一个"预处理"阶段：

1. 扫描：识别对话中的大块工具输出和关键信息
2. 提取：把值得保留的信息写入记忆文件
3. 检查：生成压缩前检查清单，标记风险项
4. 压缩：用户确认安全后才执行压缩

核心原则：先救再压，宁可多存也不能漏存。

执行流程

Phase 1 — 扫描工具输出

1. 回顾当前对话中所有的工具调用结果
2. 识别大块输出（超过 50 行或 2000 字符的工具结果）
3. 对每个大块输出评估：
   • 是否包含关键信息（决策、配置、错误信息、地址等）
   • 是否已经被后续对话引用或总结过
   • 是否是重复或冗余的（如多次 ls、git status）

Phase 2 — 提取记忆

1. 从工具输出和对话中提取值得持久化的信息：

   • 新发现的事实：地址、配置值、端点、文件路径
   • 决策和原因：为什么选了方案 A 而不是 B
   • 错误和解决方案：踩坑记录
   • 用户偏好：明确表达的喜好或要求
   • 任务进度：哪些做完了，哪些还没做

2. 将提取的信息追加写入 memory/YYYY-MM-DD.md

   • 使用 edit（追加模式），绝不覆盖已有内容
   • 每条记忆附带简短的来源说明

Phase 3 — 生成压缩前检查清单

输出一份结构化的检查清单：

📋 Smart Compact 检查清单
━━━━━━━━━━━━━━━━━━━━━━

📊 扫描统计：
- 工具调用总数：N 次
- 大块输出（>50行）：N 个
- 已引用/总结过的：N 个
- 可安全压缩的：N 个

💾 已提取到记忆：
- [+] 新事实：简要描述...
- [+] 决策记录：简要描述...
- [+] 错误解决：简要描述...
（共 N 条写入 memory/YYYY-MM-DD.md）

⚠️ 需要注意：
- [!] 某某工具输出包含重要数据但尚未被引用
- [!] 某某配置值只出现在工具输出中

✅ 建议：可以安全执行 /compact

Phase 4 — 执行压缩（可选）

• 如果检查清单显示"✅ 可以安全压缩"，提示用户确认
• 用户确认后，执行 /compact
• 如果有 ⚠️ 警告项，先处理完再压缩

规则

必须遵守

• 绝不丢弃未被记录的关键信息：宁可多存也不能漏存
• 追加写入：只用 edit 追加到 memory 文件，绝不覆盖
• 不自动压缩：除非用户明确确认，否则只生成检查清单
• 透明：每一步操作都告知用户

信息分类标准

• 必须保存：重要配置、地址端点、文件路径、错误解决方案
• 建议保存：决策原因、用户偏好、任务进度
• 可以丢弃：重复的 ls 输出、已被总结的搜索结果、中间调试过程

与 Dream Skill 的配合

Smart Compact 和 Dream 是互补的：

• Smart Compact：实时的，在压缩前抢救信息 → 写入日记
• Dream：定期的，把日记整合到长期记忆 → 更新 MEMORY.md

推荐工作流：

1. 对话中随时触发 Smart Compact 保护信息
2. 每天凌晨 Dream 整合日记到长期记忆
3. 形成完整的记忆保护链条