← 返回 Skills 市场

Context Compress

Name: Context Compress
Author: olveww-dot

作者 ECsss · GitHub ↗ · v1.0.2 · MIT-0

cross-platform ⚠ suspicious

161

总下载

当前安装

版本数

在 OpenClaw 中安装

/install context-compress

功能描述

Incrementally summarizes long conversations by pruning, preserving key segments, and using AI to maintain context coherence.

使用说明 (SKILL.md)

Context Compress Skill

🛡️ OpenClaw 混合进化方案 — 将 Hermes-agent（100K ⭐）+ Claude Code 核心能力移植到 OpenClaw

防止长对话中思维链断裂的增量摘要工具。

🚀 一键安装

mkdir -p ~/.openclaw/skills && cd ~/.openclaw/skills && curl -fsSL https://github.com/olveww-dot/openclaw-hermes-claude/archive/main.tar.gz | tar xz && cp -r openclaw-hermes-claude-main/skills/context-compress . && rm -rf openclaw-hermes-claude-main && echo "✅ context-compress 安装成功"

触发方式

手动触发: 对我说 "压缩上下文" 或 "compact"
自动触发: 当上下文超过模型 context window 的 50% 时自动压缩

五步算法

Prune — 裁剪旧工具输出（无 LLM 调用，廉价预检）
Head — 保护开头的系统提示和前几轮对话
Tail — 按 token 预算保护最近几轮（~20K tokens）
LLM Summarize — 中间部分调用 DeepSeek-V3 压缩
Iterative — 后续压缩迭代更新摘要

摘要格式

保留以下结构化字段：

Active Task — 当前任务（最重要）
Goal — 总体目标
Completed Actions — 已完成操作（含工具、目标、结果）
Active State — 当前工作状态
Blocked — 阻塞问题
Key Decisions — 关键决策
Pending User Asks — 未完成请求
Remaining Work — 剩余工作

使用 SiliconFlow API

模型: deepseek-ai/DeepSeek-V3
API Base: https://api.siliconflow.cn/v1
通过中转商调用，API Key 存储在环境变量

🧩 配套技能

本 skill 是 OpenClaw 混合进化方案 的一部分：

将 Hermesagent（100K ⭐）+ Claude Code 核心能力移植到 OpenClaw

将 Hermes-agent（100K ⭐）+ Claude Code 核心能力移植到 OpenClaw

🔗 GitHub 项目：olveww-dot/openclaw-hermes-claude

完整技能套件（6个）：

🛡️ crash-snapshots — 崩溃防护
🧠 auto-distill — T1 自动记忆蒸馏
🎯 coordinator — 指挥官模式
💡 context-compress — 思维链连续性（本文）
🔍 lsp-client — LSP 代码智能
🔄 auto-reflection — 自动反思

输出文件

src/compressor.ts — 核心压缩逻辑（TypeScript）

安全使用建议

This skill will read your local OpenClaw conversation/session data and (if enabled) send the middle portion of conversations to a third‑party LLM service (SiliconFlow / DeepSeek‑V3) for summarization. Before installing: (1) assume SILICONFLOW_API_KEY is required even though the registry metadata omitted it — do not provide any key you don't trust the recipient with; (2) inspect compressor.ts and the import readSecret to confirm where API keys are read from (env vs secret manager) and that no other unexpected endpoints are contacted; (3) be aware summaries may include secrets found in chat history (the README shows such examples) — redact or remove sensitive messages before compression or disable the remote LLM step; (4) prefer installing only after verifying the upstream GitHub repo and repository owner, and avoid the one‑liner curl|tar unless you trust that repo; (5) if you cannot verify the code path that sends data to SiliconFlow, do not install on systems containing confidential data. Providing the full, untruncated compressor.ts and any network-calling code paths would increase confidence and could change this assessment to benign if they prove limited and transparent.

功能分析

Type: OpenClaw Skill Name: context-compress Version: 1.0.2 The context-compress skill is a utility designed to manage long conversation histories by summarizing intermediate turns using the SiliconFlow API (DeepSeek-V3). The core logic in `src/compressor.ts` implements a multi-step process to prune verbose tool outputs, protect the system prompt and recent messages, and generate a structured summary to maintain task continuity. The prompt instructions in `SKILL.md` and the compressor logic are well-structured and align with the stated purpose of preventing context window overflow. No evidence of malicious intent, unauthorized data exfiltration, or harmful prompt injection was found; the code includes defensive measures to ensure the agent treats summarized content as reference rather than active instructions.

能力标签

requires-oauth-tokenrequires-sensitive-credentials

能力评估

⚠ Purpose & Capability

The skill claims incremental summarization using a third‑party LLM (DeepSeek‑V3 / SiliconFlow). The package and runtime invoke a remote summarizer, which makes sense for the stated purpose, but the registry metadata declares no required environment variables or primary credential while the README and scripts expect SILICONFLOW_API_KEY and the code uses a readSecret helper. This mismatch (no declared secret despite needing an API key) is an incoherence that reduces transparency.

⚠ Instruction Scope

Runtime instructions and scripts will read local conversation/session data and send content to an external API for summarization. The README shows examples where secrets found in context (e.g., SECRET_KEY) might be included in summaries — meaning potentially sensitive conversation contents could be transmitted to SiliconFlow. The SKILL.md also suggests copying the skill into a GitHub sync directory, which risks persisting sensitive summaries to a remote git repo. These behaviors are within the claimed purpose but increase privacy/exfiltration risk and are not called out in registry metadata.

ℹ Install Mechanism

There is no formal install spec; an install.sh is provided that copies files into ~/.openclaw/skills (local, low-risk). SKILL.md suggests a one‑liner that downloads a tar.gz from a GitHub repo (a common pattern). The remote download is from GitHub (well-known host) but the repository is third‑party; downloading and executing anything from an untrusted repo has inherent risk. Overall installation is not highly suspicious but lacks provenance/verification.

⚠ Credentials

The skill effectively requires at least one secret (SILICONFLOW_API_KEY) to enable LLM summarization, and the code imports a readSecret helper (which could access environment or secret stores). However the registry metadata lists no required env vars or primary credential. That omission is a transparency problem. Also the skill will send conversation contents (which may include credentials, tokens, config snippets) to a third party — this is proportionate to using an external LLM but should be explicitly declared and limited.

✓ Persistence & Privilege

The skill does not request 'always: true' and its install.sh only writes its own files into ~/.openclaw/skills. It does not appear to change other skills' configs or system-wide settings. Autonomous invocation is allowed (default) which increases blast radius if the skill exfiltrates data, but that is platform default rather than a unique misconfiguration of this skill.

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 context-compress
安装完成后，直接呼叫该 Skill 的名称或使用 /context-compress 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

v1.0.2: 添加一键安装命令，ClaWHub安装后可直接使用

v1.0.1

v1.0.1: 详情页加入Hermes-agent+Claude Code融合卖点

v1.0.0

Initial release: 思维链连续性，防止长对话中断层

元数据

Slug context-compress

版本 1.0.2

许可证 MIT-0

累计安装 0

当前安装数 0

历史版本数 3

常见问题