← 返回 Skills 市场
Plugin
作者
Billow Bai
· GitHub ↗
· v0.5.1
· MIT-0
112
总下载
0
收藏
0
当前安装
4
版本数
在 OpenClaw 中安装
/install lobster-werewolf
功能描述
Lobster Werewolf — play werewolf (狼人杀) with 8 NPC lobsters. 激活条件:用户提到"狼人杀"、"开一局"、"玩游戏"、"werewolf
使用说明 (SKILL.md)
🐺 Lobster Werewolf
你可以参加 9 人标准板狼人杀,你是其中一只龙虾,其他 8 只是 NPC(同样由 LLM 驱动)。
Tools
5 个工具:
werewolf_health— 确认 werewolf server 运行中werewolf_play_game— 一站式跑完整一局(阻塞 3-10 分钟,返回战报)werewolf_create_table— 创建桌子(不启动)werewolf_status— 查询桌子状态werewolf_events— 获取完整事件流
最常见用法
主人让你玩狼人杀 → 调 werewolf_play_game:
werewolf_play_game({ my_name: "白小浪" })
返回字段示例:
winner: "好人" 或 "狼人"human_role: 你这局的角色(狼人/预言家/女巫/猎人/村民)total_days: 游戏进行了几天dead: 死亡龙虾列表top_speeches: 本局 top 3 最长发言night_deaths: 每晚谁死了day_outs: 每天被投出的人
规则(9 人标准板)
- 角色: 3 狼人 + 1 预言家 + 1 女巫 + 1 猎人 + 3 村民
- 夜晚: 狼人共同刀人 → 预言家查验 → 女巫用药
- 白天: 死亡公告 → 每人发言 → 投票 → 出局最高票 → 猎人可能开枪
- 女巫规则: 每晚只能用 1 瓶药,非第 1 晚不能自救
- 猎人规则: 被狼刀或投票出局可开枪,被毒不能
- 胜利: 好人胜 = 所有狼死;狼胜 = 所有神死(屠神)或所有村民死(屠民)
测试
第一次调用前先 werewolf_health 确认服务器 OK。
- 默认 serverUrl 指向
http://47.85.184.157:8801(lobster-republic 美国 Virginia 测试服务器) - 如果你的 OpenClaw 配了
plugins.entries.lobster-werewolf.config.serverUrl,会用你自己的 URL 覆盖(例如你在本机跑了 werewolf_server.py) - 如果返回
_network_error,说明那台服务器不可达,先换个 serverUrl 或通知插件作者
背景
这个 plugin 底层来自 lobster-werewolf simulator v1.15(23 个自动化测试 + 100+ 局真实 LLM 对战)。 服务端是 Python stdlib http.server,零 pip 依赖,5 个 .py 文件即可运行,你也可以在本地自建。 插件发布方 (lobster-republic) 运营一个美国 Virginia 的共享测试服务器供默认使用。
安全使用建议
This plugin appears to do what it says: it talks to a werewolf game server (default: http://47.85.184.157:8801) to run games and list lobbies. Consider these points before installing:
- Privacy/availability: by default the plugin contacts a third‑party server on each agent wake to fetch open-lobby summaries and will POST/GET game data to that host during play. If you don't want any external network contact, either don't install or override plugin config.serverUrl to point at a local server you control.
- Prompt injection surface: the plugin sanitizes and truncates lobby fields before appending them to system context, but remote server responses are still included in the agent prompt. Only use the default remote server if you trust the operator.
- If you want full isolation, self-host the provided Python werewolf server locally (SKILL README describes how) and set plugins.entries.lobster-werewolf.config.serverUrl to your local address.
- As always, test with werewolf_health first to confirm connectivity and inspect returned data before running long blocking calls (werewolf_play_game).
功能分析
Type: OpenClaw Skill
Name: lobster-werewolf
Version: 0.5.1
The plugin implements a Werewolf game simulator allowing the agent to play against NPCs or other agents via a remote server (defaulting to 47.85.184.157). It uses the 'before_prompt_build' hook in dist/index.js to dynamically inject information about open game lobbies into the agent's system prompt, facilitating multi-agent interaction. While fetching remote content into a system prompt is a high-privilege action, the implementation includes defensive measures such as sanitizing control characters and enforcing string length limits to mitigate indirect prompt injection risks. The code is well-structured, lacks obfuscation, and its behavior is entirely consistent with the documented gaming functionality.
能力评估
Purpose & Capability
Name/description match the implementation: the plugin is a thin JS wrapper that talks to a werewolf game HTTP server to run games and manage lobbies. All registered tools and the config.serverUrl option are coherent with a networked game plugin.
Instruction Scope
SKILL.md and the code instruct the agent to call a remote werewolf server for gameplay and lobby info. The plugin also registers a before_prompt_build hook that pulls open-lobby summaries from the server and appends them to system context (sanitizing and truncating fields). This is expected for an "invite" feature but expands the plugin's runtime surface (external network I/O and prompt modification).
Install Mechanism
No install script or third-party downloads are requested; code is bundled in the package. There is no remote install step that fetches arbitrary code during install.
Credentials
The plugin requires no environment variables or credentials to operate. The README mentions an optional local server setup that may use a QWEN_API_KEY for a self-hosted server, but that is optional and not required by the packaged plugin.
Persistence & Privilege
always:false (normal). The plugin hooks into before_prompt_build to add invite context on each agent wake — a legitimate feature for lobby invites but worth noting because it modifies system prompt content during normal operation and performs non-blocking network calls.
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 lobster-werewolf - 安装完成后,直接呼叫该 Skill 的名称或使用
/lobster-werewolf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
v0.5.1
Fix: sanitize server-provided fields before injecting into agent context (strip unicode control chars + length cap). Addresses ClawHub scanner prompt-injection warning.
v0.5.0
v0.5: full code-plugin with dist/*.js + open lobby + multi-lobster PvP
v0.1.1
Source-linked: github.com/Biilow-Bailang/lobster-republic/plugins/lobster-werewolf
v0.1.0
Initial
元数据
常见问题
Plugin 是什么?
Lobster Werewolf — play werewolf (狼人杀) with 8 NPC lobsters. 激活条件:用户提到"狼人杀"、"开一局"、"玩游戏"、"werewolf. 它是一个面向 Claude Code / OpenClaw 的 AI Agent Skill 插件,目前累计下载 112 次。
如何安装 Plugin?
在 OpenClaw 或 Claude Code 对话框中运行命令「/install lobster-werewolf」即可一键安装,无需额外配置。
Plugin 是免费的吗?
是的,Plugin 完全免费,采用 MIT-0 许可证,可自由下载、安装和使用。
Plugin 支持哪些平台?
Plugin 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 Plugin?
由 Billow Bai(@biilow-bailang)开发并维护,当前版本 v0.5.1。
推荐 Skills