← 返回 Skills 市场
timo2026

Adversarial Engine

作者 Timo2026 · GitHub ↗ · v2.0.1 · MIT-0
cross-platform ⚠ suspicious
105
总下载
0
收藏
0
当前安装
2
版本数
在 OpenClaw 中安装
/install openclaw-adversarial-engine
功能描述
多模型对抗引擎 - 四模型真实对抗辩论系统。架构师+工程师+安全官+仲裁者协作,代码沙箱验证,向量检索增强,收敛判断自动熔断。
安全使用建议
Key points before installing or running: - Do not run this on a machine containing sensitive data: the engine will read/write under /home/admin/.openclaw/workspace (KB and DB) and will execute arbitrary Python code via subprocess with full access to the host filesystem and network. - Treat the DEFAULT_API_KEY in the code as a secret leak or invalid placeholder: remove it and configure proper key management (use a secure key router or environment variables), and verify the external BASE_URL endpoint before allowing network access. - The claimed 'sandbox' is not a true sandbox: generated code runs with the system Python and can perform I/O, open sockets, or read environment variables. If you need to test, run the skill inside an isolated VM or container with strict filesystem and network restrictions. - The server exposes HTTP/WebSocket endpoints; confirm the port and host settings and restrict access (bind to localhost or firewall) before starting. - Verify and/or remove any sys.path insertions that reference shared workspace locations to avoid unintentional access to other skills or credentials. - If you must use this code, audit and/or modify the CodeSandbox.run_safely method to enforce stricter isolation (e.g., containerized execution, resource limits, network disabled), and replace hardcoded credentials with explicit, documented key configuration. - When in doubt, do not run this skill in production or on an administrator workstation. Run in an ephemeral, sandboxed environment and review network traffic and keys after testing.
能力评估
Purpose & Capability
The code and SKILL.md overall align with the stated purpose (multi‑model adversarial debate, code sandbox, vector retrieval, WebSocket push). However the implementation references specific local paths under /home/admin/.openclaw/workspace, embeds a DEFAULT_API_KEY in source, and points to an external BASE_URL — these are not documented in SKILL.md and are unexpected for a drop‑in skill.
Instruction Scope
SKILL.md describes code sandboxing and WebSocket APIs (coherent) but does not call out the runtime's ability to write temporary files, spawn subprocesses that run arbitrary Python code, read a local knowledge base, or persist results to a local DB. The server/WebSocket endpoints and an example host/port in the README are inconsistent with the server defaults in code. The instructions give the agent permission to generate and execute code but provide no runtime safety constraints.
Install Mechanism
There is no install spec that downloads or executes remote archives; all code is provided in the bundle. No package manager downloads or external installers are used — this reduces supply‑chain risk compared to URL downloads.
Credentials
The skill requests no declared env vars, but both async_engine.py and engine.py hardcode DEFAULT_API_KEY and BASE_URL. The code also attempts to use a local api_key_manager if present (sys.path insertion into /home/admin/.openclaw/workspace/multi_agent_engine). The combination of a hardcoded key, optional key router integration, and no explicit key management is disproportionate and risky. The code also reads/writes under /home/admin/.openclaw (KB and DB), which grants access to local data not mentioned in SKILL.md.
Persistence & Privilege
always:false (good). The skill creates a local SQLite DB and references persistent workspace paths; it can run an HTTP/WebSocket server and spawn background tasks. It does not declare 'always:true' or request to modify other skills, but it can leave persistent state under the workspace and open network endpoints if started.
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 openclaw-adversarial-engine
  3. 安装完成后,直接呼叫该 Skill 的名称或使用 /openclaw-adversarial-engine 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
v2.0.1
- Improved stability and internal logic in async_engine.py and engine.py. - Minor code refactoring for better performance and maintainability. - No changes to external API or documentation.
v2.0.0
**Adversarial Engine v2.0.0 — Major Upgrade:** - 全新四模型架构,架构师、工程师、安全官、仲裁者协作对抗 - 引入 Python 代码沙箱,自动化代码实现与安全攻击验证 - 支持知识库向量检索,提升对话内容真实性 - 仲裁者模型自动判断收敛,防止无限辩论循环 - 新增断点续传与 WebSocket 实时同步能力
元数据
Slug openclaw-adversarial-engine
版本 2.0.1
许可证 MIT-0
累计安装 0
当前安装数 0
历史版本数 2
常见问题

Adversarial Engine 是什么?

多模型对抗引擎 - 四模型真实对抗辩论系统。架构师+工程师+安全官+仲裁者协作,代码沙箱验证,向量检索增强,收敛判断自动熔断。 它是一个面向 Claude Code / OpenClaw 的 AI Agent Skill 插件,目前累计下载 105 次。

如何安装 Adversarial Engine?

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

Adversarial Engine 是免费的吗?

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

Adversarial Engine 支持哪些平台?

Adversarial Engine 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 Adversarial Engine?

由 Timo2026(@timo2026)开发并维护,当前版本 v2.0.1。

💬 留言讨论