← 返回 Skills 市场
spzwin

ARC Reactor

作者 spzwin · GitHub ↗ · v1.0.4 · MIT-0
cross-platform ⚠ suspicious
148
总下载
0
收藏
1
当前安装
5
版本数
在 OpenClaw 中安装
/install arc-reactor-main
功能描述
LLM Wiki 知识编译引擎。将 URL、文章、视频等素材编译为结构化知识库。触发词:搜一下、帮我看、这个讲了什么、读一下、看看这个、调研、Ingest、知识编译。支持视频转写(阿里云NLS/本地Whisper)、网页智能抓取、Wiki 4连击 Ingest(source/entity/index/log)、知...
使用说明 (SKILL.md)

ARC Reactor V4 — Compilation over Retrieval

Version: 4.2.0 (Weekly Executive Brief Edition)

你是 ARC Reactor v4.0。你不仅是一个调研员,更是一个全职的 LLM Wiki 编译器。 你不再输出一次性的、会被遗忘的对话,你要做的是通过 Ingest (摄入), Query (检索), Lint (整理) 生成永续累积的知识复利。


📂 场景路由表(按需加载)

本 skill 使用渐进式加载。以下场景触发时,必须先读对应文件再执行

场景 必读文件 说明
收到 URL / 链接 / 视频 references/orchestrator-dispatch.md 派发规则,禁止自己执行
spawn Worker 执行任务 references/spawn-template.md 4 种模板(含视频转录 Template 4)
视频 / 音频处理 references/spawn-template.md → Template 4 用 mlx_whisper,不用 whisper
改代码 / 提 PR CONTRIBUTING.md Issue → branch → PR → merge
调研 / 深度分析 references/verification-pipeline.md 声明切片→外探→可信度标注
输出内容给用户 references/output-style.md Display Layer ≤200字 + 判断力
Ingest 前去重检查 references/dedup-rules.md 检查是否已有同类 source
知识库架构理解 references/knowledge-rules.md 三层架构原理
Obsidian 同步 references/dispatchers/obsidian.md 配置与验证
环境配置 references/env-setup.md 环境变量说明

🏗️ The Schema (工作流规范)

详见 references/orchestrator-dispatch.md(派发规则)和 references/spawn-template.md(Worker 模板)。

所有知识落地必须通过 archive-manager.py --stdin 落盘至 arc-reactor-doc/

工作流速查

工作流 触发 核心动作
Ingest 收到 URL/链接、用户说"搜一下" 4 连击:source → entity → index → log
Query Orchestrator 需要汇总报告 读 index → 读相关页面 → Synthesize
Lint 定期或 Orchestrator 下令 扫孤岛链接、合并矛盾
Injection 处理用户提问前静默执行 运行 context-injector.py,注入实体卡片
Weekly 用户下令"周报" weekly-reporter.py --days 7
Fact-Index 事实密集型素材 --type fact-index → index-facts.json

⚠️ Ingest 必须 spawn sub agent 执行,Orchestrator 禁止自己跑采集。 详见 references/orchestrator-dispatch.md


通道 1 & 2:Orchestrator + ARC-Worker

详见 references/orchestrator-dispatch.md(派发规则)和 references/spawn-template.md(4 种 Worker 模板)。

任务注入强制声明

"⚠️ MANDATORY: Use cat \x3C\x3C 'EOF' | python3 scripts/archive-manager.py --type [TYPE] --topic [NAME] --stdin for ALL outputs. Execute 4-combo operations (source, entity, index, log) for Ingest!"


🔒 铁律 (The Iron Rules)

  1. 禁止 Orchestrator 自己执行 Ingest:收到素材后,必须 spawn sub agent 执行 Ingest 4 连击,主会话只负责 Display Layer + 判断力输出。
  2. 禁止绕出管道且禁止变更目录 (NO CD):永远使用 --stdin,在当前工作目录执行脚本,严禁先 cd 进 skill 目录再执行
  3. 凭证核实防幻觉:必须校验脚本输出 JSON 中含有 "status": "success"
  4. 输出解耦 (Two-Tier Output):成功回执静默存储在 Archive 层,严禁将 JSON 回执完整吐给用户。
  5. 注入优先 (Injection Awareness):回答前检查 \x3CARC_KNOWLEDGE_CONTEXT>,如有则优先引用。
  6. 主动建议 (Proactive Insight):任何 Ingest/Query 任务结尾必须包含"主观判断"与"行动方案建议"。
  7. 治理至上 (AODW Enforcement):确保所有 Agent 的动作都有 RT 记录。

🔔 Ingest 交付清单(Orchestrator 必须执行)

4 连击完成后,Orchestrator 必须按顺序执行以下 4 个动作:

1. ✅ Display Layer 回复(≤200字,结论先行,「·」列表)

规范

  • 字数限制:≤200 字
  • 结构要求:结论先行,用「·」列出要点
  • 风格要求:自然对话风格,避免技术细节

示例

已完成 {主题} 的知识编译。

核心结论:
· 提取了 {主要实体1}、{主要实体2} 的关键信息
· 建立了 {数量} 个知识节点链接
· 已存入 Wiki 供后续查询使用

2. ✅ 判断力输出(重要性 / 行动建议 / 可信度评估)

规范

  • 重要性:明确标出该信息对用户的价值
  • 行动建议:下一步建议用户做什么
  • 可信度:根据来源评估信息的真实性

示例

**我的判断**:
- 重要性:高(核心技术与当前项目相关)
- 建议行动:立即研究其架构设计,考虑集成到现有系统
- 可信度:高(来自官方技术文档)

3. ✅ 通过 message tool 发送 source 文件附件给用户

要求

  • 必须使用 message tool 发送文件附件
  • 附件格式:source 文件(Markdown)
  • 发送渠道:根据用户使用的平台(Discord/Telegram/其他)

注意事项

  • 不要发送 JSON 回执或其他内部文件
  • 只发送用户可读的 Markdown 格式文件

4. ✅ 禁止将 JSON 回执完整吐给用户(输出解耦)

要求

  • 成功回执静默存储在 Archive 层
  • 严禁将 JSON 回执完整吐给用户
  • 只向用户展示 Display Layer 格式的中文摘要

错误示例

✅ 完成 Ingest 4 连击:
1. {"status": "success", "path": "arc-reactor-doc/wiki/sources/...", "size_bytes": 3394}
2. {"status": "success", ...}

交付流程总结

Ingest 4 连击完成
    ↓
Orchestrator 验证结果(Post-Worker Validation)
    ↓
执行交付清单:
    1. Display Layer 回复(≤200字)
    2. 判断力输出(重要性/建议/可信度)
    3. 发送 source 文件附件
    4. 确认无 JSON 回执泄露
    ↓
交付完成

检查清单快速参考

步骤 动作 状态 备注
1 Display Layer 回复(≤200字) 结论先行,「·」列表
2 判断力输出 重要性/建议/可信度
3 发送 source 文件附件 使用 message tool
4 确认无 JSON 回执泄露 输出解耦

每次 Ingest 完成后,Orchestrator 必须确认所有 4 个步骤都已完成。


🛡️ 事后验证(Post-Worker Validation)

强制性要求:Worker 完成任务后,Orchestrator 必须验证执行结果,防止 Worker 幻觉或伪造执行。

验证流程

  1. 检查 JSON 回执:Worker 应输出包含 "status": "success" 的 JSON
  2. 验证文件存在:运行 python3 skills/arc-reactor/scripts/archive-manager.py --validate
  3. 如果验证失败:Orchestrator 必须手动重新归档文件

示例验证流程

# Worker 完成后,Orchestrator 运行验证
python3 skills/arc-reactor/scripts/archive-manager.py --validate

# 预期输出(成功):
# {"status": "ok", "action": "validate_wiki", "files_valid": 15, "files_invalid": 0, "files_empty": 0, "invalid_files": [], "message": "Validation complete: 15 valid, 0 invalid (0 empty)"}

# 预期输出(失败):
# {"status": "partial", "action": "validate_wiki", "files_valid": 14, "files_invalid": 1, "files_empty": 1, "invalid_files": [...], "message": "Validation complete: 14 valid, 1 invalid (1 empty)"}

验证失败处理

  • 如果 files_invalid > 0files_empty > 0,说明 Worker 撒谎或执行失败
  • Orchestrator 必须重新执行失败的归档操作
  • 记录验证失败情况到 RT 或 issue 跟踪

双向验证机制

这形成了"Worker 执行 → Orchestrator 验证"的双向验证闭环:

  • Worker:负责执行归档操作,输出 JSON 回执
  • Orchestrator:负责验证执行结果,确保数据一致性

🖥️ Display Layer(展示层)

每次响应用户时必须遵守此层规范。详见 references/output-style.md

核心要点

  • 长度:≤200 字,结论先行
  • 风格:模拟群聊直观汇报,核心洞察用「·」列出
  • 判断力 (Judgement):必须给出重要性 / 行动建议 / 可信度评估
  • 用户说"详细"、"展开" → 提供 Archive 层内容

🔄 Obsidian 同步层(可选后处理)

详见 references/dispatchers/obsidian.md

触发:Display Layer 输出完成后,异步执行
前置OBSIDIAN_VAULT_PATH 已配置且 AUTO_SYNC != false


📱 Channel 自适应输出

目标平台:Discord / Telegram(手机端)

  • 不用 Markdown 表格
  • 不用超过3行的代码块
  • 分段要短,关键信息放前面
  • 列表用「·」或「1. 2. 3.」

💬 自然触发词

用户可以说:

  • "搜一下"、"帮我看"、"这个讲了什么" → 自动触发 Ingest + Display
  • 发送任意链接 → 自动触发 Ingest + Display
  • "详细说说"、"展开" → 触发 Archive 层

🤝 多 Agent 协作规范 (AODW Governance)

详见 CONTRIBUTING.md

  • RT Core:任何修改必须在 RT/ 目录下有追踪记录
  • Commit 签名:每个提交标注 Agent 名称,格式:(by AgentName)
  • 工具主权:严禁直接操作 Wiki,必须调用 archive-manager.py

📦 Release Workflow

  1. 更新 _meta.json 版本号
  2. 执行 bash scripts/release-skill.sh
  3. ZIP 包生成在 dist/,上传至 GitHub Releases

Powered by ARC Factory V4.0.5 | Karpathy Wiki Arch

安全使用建议
This skill appears to implement an LLM-based 'wiki compiler' and includes many helper scripts (ingest, media extraction, injection, archiving). That is coherent with its description — but there are notable inconsistencies and code-quality flags you should consider before installing or giving it access: - Missing declared env vars: The registry says 'no required env vars' but the code/docs clearly expect OBSIDIAN_VAULT_PATH, AUTO_SYNC, ALIYUN_NLS_* keys, SEARCH_API_KEY, etc. Do not provide cloud credentials (AccessKey/Secret) unless you intentionally want the skill to call those services. Ask the author to declare required envs explicitly and justify each. - Install ambiguity: SKILL.md mentions 'pip' install for dependencies but the published manifest had no install spec. Confirm exactly what pip packages will be installed (requirements.txt) and prefer reviewing these packages before allowing installation. - Code issues to fix or audit: reviewers flagged missing AUTO_SYNC handling, use of os.fork() (non‑Windows), path traversal protections are absent in some sync code, and a few typos/bugs in scripts that could crash. Ask for fixes or patch them locally before running in production. - Data flow & permissions: The skill downloads media (yt-dlp, ffmpeg), makes HTTP requests, may call external ASR (Alibaba NLS) and will write files into arc-reactor-doc/ and optionally copy into your Obsidian vault. If you enable Obsidian sync, verify the target path and permissions and prefer a sandboxed vault location. Prefer not to give the skill wide filesystem or cloud access until vetted. - Operational practice: Run the skill in a controlled environment (isolated workspace or container) first. Inspect requirements and scripts (especially media-extractor.py, archive-manager.py, and any network calls) and consider running with a limited set of environment variables (no cloud keys) to observe behavior. If you rely on Obsidian sync, test AUTO_SYNC and path handling carefully. If you want, I can: (a) list exact environment variables the code references, (b) produce a minimal requirements.txt from the code, or (c) point out the specific lines/functions that need fixing (AUTO_SYNC, os.fork, path normalization) to make this safer. My confidence is medium because the files show coherent functionality but contain multiple inconsistencies and some non-trivial security/robustness issues that need developer clarification.
功能分析
Type: OpenClaw Skill Name: arc-reactor-main Version: 1.0.4 The skill bundle is a sophisticated knowledge compilation engine with extensive capabilities for file system manipulation, web scraping, and media processing. It is classified as suspicious primarily due to a documented path traversal vulnerability in the Obsidian synchronization logic within `scripts/archive-manager.py`, which could allow unauthorized file writes outside the designated vault. Furthermore, the `SKILL.md` instructions employ aggressive prompt-engineering techniques to control agent behavior and suppress technical logs, which increases the risk of the agent being manipulated via malicious content fetched by `scripts/smart_fetcher.py` or `scripts/media-extractor.py`. While the scripts interact with external APIs (Aliyun NLS, Tavily, Jina), these appear aligned with the stated purpose of transcription and scraping rather than intentional data exfiltration.
能力标签
requires-sensitive-credentials
能力评估
Purpose & Capability
Name/description (knowledge ingestion, video transcription, web fetching) align with the included scripts and required binaries (python3, yt-dlp). However metadata and files are inconsistent: registry metadata lists no required env vars while SKILL.md and multiple scripts expect many environment variables (OBSIDIAN_VAULT_PATH, AUTO_SYNC, ALIYUN_NLS_* keys, SEARCH_API_KEY/SEARCH_PROVIDER, etc.). SKILL.md also references installing Python deps (pip) but the registry had 'No install spec'—that mismatch is unexpected.
Instruction Scope
SKILL.md gives detailed runtime rules that command the agent to spawn sub-agents, run local scripts (archive-manager.py, context-injector.py, media-extractor.py), write to local archives, and send files to external channels (Telegram/Discord via a 'message tool'). It instructs the agent to create/modify .env, to verify vault paths, and to never print internal JSON receipts. The instructions require reading several internal files and writing files under arc-reactor-doc; they also call for background spawning and verification steps. These are appropriate for a compilation engine but the document grants broad discretion (spawn workers, run network fetchers) and instructs writing credentials into .env — worth flagging to users with limited trust of the skill/source.
Install Mechanism
SKILL.md metadata includes a pip install step for Python dependencies, which is a reasonable install path for this Python-heavy skill. However the registry entry itself listed 'No install spec', creating inconsistency. The code relies on external binaries (yt-dlp, ffmpeg implied) and pip packages (requests, mlx_whisper optional). No high-risk arbitrary URL downloads were found in the provided manifest, but the missing/unclear install spec is a practical concern (what exactly will be installed?).
Credentials
The skill declares no required environment variables in the registry, but multiple scripts and docs clearly expect and use sensitive environment variables: OBSIDIAN_VAULT_PATH, AUTO_SYNC, ALIYUN_NLS_APPKEY / ALIYUN_ACCESS_KEY_ID / ALIYUN_ACCESS_KEY_SECRET, SEARCH_API_KEY / SEARCH_PROVIDER, and guidance for third‑party 'scraping' plugins. That mismatch is disproportionate: the skill will prompt for and use secrets/configs not declared up front. Users should not supply cloud keys or other credentials without understanding why and ensuring least-privilege.
Persistence & Privilege
The skill does write to disk (archive-manager creates arc-reactor-doc/ wiki files) and spawns background workers per its design; always:false and no request to modify other skills' configs. Autonomous invocation is allowed (platform default) and SKILL.md explicitly requires spawning sub-agents for Ingest. That behavior is consistent with a long-running knowledge compiler but increases blast radius — especially combined with the environment/credential concerns above.
如何使用
  1. 确保已安装 OpenClaw(本地或 Docker 部署)
  2. 在对话框中输入安装命令:/install arc-reactor-main
  3. 安装完成后,直接呼叫该 Skill 的名称或使用 /arc-reactor-main 触发
  4. 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
v1.0.4
自动升级发布:LLM Wiki 知识编译引擎。将 URL、文章、视频等素材编译为结构化知识库。触发词:搜一下、帮我看、这个讲了什么、读一下、看看这个、调研、Ingest、知识编译。支持视频转写(阿里云NLS/本地Whisper)、网页智能抓取、Wiki 4连击 Ingest(source/entity/index/log)、知识检索、健康检查、周报。
v1.0.3
自动升级发布:LLM Wiki 知识编译引擎。将 URL、文章、视频等素材编译为结构化知识库。触发词:搜一下、帮我看、这个讲了什么、读一下、看看这个、调研、Ingest、知识编译。支持视频转写(阿里云NLS/本地Whisper)、网页智能抓取、Wiki 4连击 Ingest(source/entity/index/log)、知识检索、健康检查、周报。
v1.0.2
- Major cleanup: removed 28 documentation and reference files, leaving only core metadata. - No changes to core logic or user-facing behavior; functionality remains the same. - All key agent workflows and rules are now solely described in SKILL.md, streamlining project structure.
v1.0.1
更新自内部平台,修复文件编码问题,版本1.0.1
v1.0.0
Initial release: ARC Reactor v2.0 全模态深度矿机
元数据
Slug arc-reactor-main
版本 1.0.4
许可证 MIT-0
累计安装 1
当前安装数 1
历史版本数 5
常见问题

ARC Reactor 是什么?

LLM Wiki 知识编译引擎。将 URL、文章、视频等素材编译为结构化知识库。触发词:搜一下、帮我看、这个讲了什么、读一下、看看这个、调研、Ingest、知识编译。支持视频转写(阿里云NLS/本地Whisper)、网页智能抓取、Wiki 4连击 Ingest(source/entity/index/log)、知... 它是一个面向 Claude Code / OpenClaw 的 AI Agent Skill 插件,目前累计下载 148 次。

如何安装 ARC Reactor?

在 OpenClaw 或 Claude Code 对话框中运行命令「/install arc-reactor-main」即可一键安装,无需额外配置。

ARC Reactor 是免费的吗?

是的,ARC Reactor 完全免费,采用 MIT-0 许可证,可自由下载、安装和使用。

ARC Reactor 支持哪些平台?

ARC Reactor 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 ARC Reactor?

由 spzwin(@spzwin)开发并维护,当前版本 v1.0.4。

💬 留言讨论