Skill flagged — suspicious patterns detected

ClawHub Security flagged this skill as suspicious. Review the scan results before using.

Cursor Council

Multi-Cursor orchestration for parallel task execution and AI council deliberation. Use when needing to run multiple Cursor agents in parallel, coordinate complex multi-step coding tasks, get diverse perspectives from different AI models (Opus/Sonnet/GPT) on technical decisions, or synthesize multi-agent discussions into actionable recommendations.

MIT-0 · Free to use, modify, and redistribute. No attribution required.
0 · 1.1k · 2 current installs · 2 all-time installs
MIT-0
Security Scan
VirusTotalVirusTotal
Suspicious
View report →
OpenClawOpenClaw
Suspicious
medium confidence
Purpose & Capability
Name/description (multi-agent orchestration, deliberation) align with the declared requirements: it needs tmux and an 'agent' CLI and depends on a 'cursor-agent' skill. The install entries only provide tmux via brew/apt which is proportionate.
!
Instruction Scope
SKILL.md instructs the agent to create tmux sessions, write temp prompt files, run the 'agent' CLI with --force, capture pane outputs, and synthesize results. That is within the stated purpose, but the skill includes many persona templates that instruct the models to 'be' named real individuals (Joe Armstrong, TJ Holowaychuk, Ryan Dahl, etc.) and contains detected prompt-injection patterns (unicode control characters). The presence of invisible control characters in the runtime instructions is a red flag because it can manipulate LLM parsing/behavior. Also the prompts encourage impersonation of real people, which may produce misleading or legally/ethically risky outputs.
Install Mechanism
Only installs are standard package-manager entries for tmux (brew/apt). No downloads from URLs or archive extraction. Low install risk.
Credentials
The skill requests no environment variables or credentials. It expects the 'agent' CLI to be logged in and a separate 'cursor-agent' skill to be configured, which is consistent with its function. It does not request unrelated credentials or config paths.
Persistence & Privilege
always is false (not force-enabled), and the skill doesn't attempt to change other skills or system-wide configs. Runtime behavior writes temporary files under /tmp and creates tmux sessions — typical for this use-case and scoped to the invoking user session.
Scan Findings in Context
[unicode-control-chars] unexpected: The scanner detected unicode control characters / prompt-injection patterns embedded in SKILL.md. This is not necessary for orchestrating tmux/agent sessions and is suspicious because such characters can alter LLM input parsing or hide malicious directives.
What to consider before installing
This skill is coherent with its stated purpose (using tmux to run multiple Cursor agents), but exercise caution before installing or running it. Specific recommendations: - Do not run it against sensitive repositories or systems until you audit the SKILL.md and reference files yourself. The skill writes prompt files and runs agent CLI commands that will send text to external models; avoid exposing secrets. - Inspect the SKILL.md and the reference files in a text editor that can show invisible characters; remove any unexpected unicode control characters before use. - The prompt templates ask models to impersonate named real people. That may yield fabricated or misleading statements and could be legally/ethically problematic; consider replacing impersonation prompts with neutral role descriptions instead. - Avoid using --force or running tasks directly on main branches; follow the skill's own advice to use dedicated branches and test on a disposable repo or workspace first. - Ensure your 'agent' CLI is logged in to an account you control, and verify which models/transport it will use (and that you have appropriate API keys configured and scoped). The skill assumes external model access but does not manage credentials itself. - If you don't trust the unknown source (no homepage, unknown owner), prefer re-implementing the essential orchestration steps yourself from the provided templates rather than installing blindly. If you want, I can (1) show how to sanitize the prompt templates to remove impersonation and control characters, or (2) generate a minimal safe wrapper script that only launches tmux sessions without writing persona files.

Like a lobster shell, security has layers — review code before you run it.

Current versionv1.0.0
Download zip
latestvk97a5aqnkdabg8c05a7pe22rax80q8yt

License

MIT-0
Free to use, modify, and redistribute. No attribution required.

Runtime requirements

🎭 Clawdis
Binstmux, agent

Install

Install tmux (brew)
Bins: tmux
brew install tmux

SKILL.md

🎭 Cursor Council — 我的编程前辈团

Cursor CLI 需要真实 TTY,所以一切都通过 tmux 操作。 这个 skill 的本质:用 tmux 管理多个 Cursor 实例,让它们并行干活或者扮演大牛帮我分析问题。

前置条件

which tmux && agent --version

需要 cursor-agent skill 已配置好、agent login 已完成。


两种模式

模式一:并行施工 🏗️

场景:一个大任务可以拆成互不冲突的子任务,多个 Cursor 同时干。

核心原则

  • 每个 Cursor 负责不同的文件/模块,绝对不能有文件冲突
  • 有依赖关系的任务必须等前置完成后再启动
  • 最好每个 Cursor 在不同分支上工作

操作流程

# 1. 拆任务——先想清楚再动手
#   问自己:这些子任务之间有没有文件交叉?有没有先后依赖?

# 2. 创建 tmux sessions
PROJECT_DIR=~/Codes/my-project
for name in frontend backend tests; do
  tmux kill-session -t cursor-$name 2>/dev/null || true
  tmux new-session -d -s cursor-$name
  tmux send-keys -t cursor-$name "cd $PROJECT_DIR" Enter
done

# 3. 派活
tmux send-keys -t cursor-frontend "agent -p '重构 src/components/ 下的表单组件,统一用 composables 模式' --force" Enter
tmux send-keys -t cursor-backend "agent -p '给 src/api/ 下所有接口加上错误处理和重试逻辑' --force" Enter
tmux send-keys -t cursor-tests "agent -p '补全 tests/ 下缺失的单元测试,覆盖率要到 80%' --force" Enter

# 4. 巡查进度
for s in cursor-frontend cursor-backend cursor-tests; do
  echo "=== $s ==="
  tmux capture-pane -t $s -p | tail -10
done

# 5. 全部完成后合并验证

任务拆分策略

策略适用场景示例
按模块Monorepo、前后端分离frontend / backend / shared
按分支多个独立 featurefeat/auth / feat/dashboard
按类型实现+测试+文档*.ts / *.test.ts / *.md

注意事项

  • 16GB 内存建议最多 3-4 个并行 Cursor
  • 如果某个 session 5 分钟没输出,可能卡住了,tmux send-keys -t $session C-c 中断
  • 看到 "waiting for approval" 就 tmux send-keys -t $session y
  • --force 要谨慎——确保在独立分支上、文件有 git 追踪

→ 更多细节见 references/parallel-execution.md


模式二:前辈议会 🧙‍♂️

场景:遇到复杂技术决策,需要多角度分析。让不同模型扮演技术大牛,从各自的哲学视角给出建议。

为什么有效: 模型训练数据里包含大量技术大牛的文章、演讲、代码,让它"成为"某个人能激活这些深层知识。而且不同大牛代表不同流派,天然形成视角碰撞。

角色分配逻辑

角色模型选择思维特点应该问什么
架构师Opus(深度推理)看长远、找边界、想极端情况架构设计、安全隐患、5年后会不会后悔
工程师Sonnet(快速实用)讲效率、抓重点、先跑起来实现成本、快速方案、技术债平衡
批判者GPT(不同视角)挑毛病、给替代方案、打破惯性有没有更好的路、假设是否成立

操作流程

# 1. 准备问题(写清楚背景、选项、约束)
cat > /tmp/council-question.md << 'EOF'
## 问题:消息队列选型

### 背景
项目需要一个消息队列处理异步任务,日均消息量 ~50 万条。

### 选项
A. Redis Streams — 团队熟悉,但持久化能力有限
B. RabbitMQ — 成熟稳定,但运维成本高
C. NATS — 轻量高性能,但生态较小

### 约束
- 团队 3 人,运维能力有限
- 需要 at-least-once 语义
- 预算有限,优先用已有基础设施
EOF

# 2. 准备各角色的 prompt
cat > /tmp/council-opus-prompt.txt << 'EOF'
你是 Joe Armstrong,Erlang 和 OTP 的创造者。
你的核心信仰:进程隔离、Let it crash、消息传递是并发的正确抽象。
你用电话交换机的经验看待所有分布式系统问题。

请从你的视角分析这个消息队列选型问题:
[粘贴问题]

分析要点:
1. 哪个选项最符合你的并发哲学?
2. 容错和故障恢复方面各有什么隐患?
3. 如果是你来设计,你会怎么做?
4. 一句话犀利点评
EOF

cat > /tmp/council-sonnet-prompt.txt << 'EOF'
你是 TJ Holowaychuk,Express/Koa 的作者,写过 1000+ npm 包的极简主义者。
你的风格:用最少的代码解决问题,讨厌过度工程化。

请从你的视角评估这个消息队列选型:
[粘贴问题]

评估要点:
1. 哪个方案能最快跑起来?
2. 哪个方案最不容易让 3 人团队踩坑?
3. 有没有更简单的替代方案?
4. 用代码示例说明你推荐的最简方案
EOF

cat > /tmp/council-gpt-prompt.txt << 'EOF'
你是 Ryan Dahl,Node.js 和 Deno 的创造者。
你最大的特点是敢于反思自己过去的设计错误,你曾公开说 Node.js 有 10 个设计遗憾。

请用你反思和重新设计的眼光审视这个问题:
[粘贴问题]

思考方向:
1. 这三个选项是否都在错误的层面解决问题?
2. 有没有根本不需要消息队列的设计?
3. 选了之后最可能后悔的点是什么?
4. 如果让你重新设计,你会怎么做?
EOF

# 3. 开会
for role in opus sonnet gpt; do
  tmux kill-session -t council-$role 2>/dev/null || true
  tmux new-session -d -s council-$role
  tmux send-keys -t council-$role "cd $PROJECT_DIR" Enter
done

tmux send-keys -t council-opus "agent --model claude-opus-4-6 -p \"\$(cat /tmp/council-opus-prompt.txt)\" --force --output-format text 2>&1 | tee /tmp/council-opus-output.txt" Enter
tmux send-keys -t council-sonnet "agent --model claude-sonnet-4-5 -p \"\$(cat /tmp/council-sonnet-prompt.txt)\" --force --output-format text 2>&1 | tee /tmp/council-sonnet-output.txt" Enter
tmux send-keys -t council-gpt "agent --model gpt-5.2 -p \"\$(cat /tmp/council-gpt-prompt.txt)\" --force --output-format text 2>&1 | tee /tmp/council-gpt-output.txt" Enter

# 4. 等待完成后收集输出(通常 2-5 分钟)
# 5. 综合分析,形成结论

→ 人设库和 prompt 模板见 references/persona-engineering.md → 议会流程详解见 references/council-deliberation.md


前辈人设速查

根据问题领域选择合适的"前辈":

并发 / 分布式

前辈哲学适合问
Joe Armstrong (Erlang)Let it crash、进程隔离并发安全、容错设计
Leslie Lamport (Paxos)形式化验证、分布式共识一致性、状态机
Martin Kleppmann (DDIA)数据密集型应用、CRDT数据一致性、分区

Node.js / JavaScript

前辈哲学适合问
TJ Holowaychuk (Express)极简主义、一个函数搞定最佳实践、API 设计
Ryan Dahl (Node/Deno)反思与重构、承认错误架构反思、根因分析
Sindre Sorhus (1000+ npm)小而美、单一职责工具库设计、代码简化

架构 / 工程

前辈哲学适合问
Martin Fowler渐进式改进、模式语言重构策略、设计模式
Uncle Bob整洁代码、SOLID 原则代码质量、可维护性
Linus Torvalds务实、性能优先、直言不讳代码审查、性能问题

函数式 / 哲学派

前辈哲学适合问
Rich Hickey (Clojure)简单≠容易、不可变性复杂性分析、状态管理
Simon Peyton Jones (Haskell)类型系统、纯函数类型设计、抽象建模

→ 完整人设库和 prompt 示例见 references/persona-engineering.md


推荐人设组合

常见场景的"前辈团"速配:

架构设计评审:Fowler(架构模式)+ DHH(约定优于配置)+ Torvalds(过度设计批评)

并发问题诊断:Armstrong(Let it crash)+ TJ(异步模式)+ Dahl(反思根因)

代码质量整治:Fowler(渐进重构)+ Uncle Bob(Clean Code)+ Hickey(复杂性根源)

技术选型决策:Lamport(理论分析)+ Sorhus(简单实用)+ Dahl(长期后悔点)


议会进阶玩法

辩论模式

  • Round 1:各自给出立场
  • Round 2:互相批评对方的方案
  • Round 3:考虑批评后给出最终建议

红队模式

  • Opus:提出方案
  • GPT:全力攻击方案,找弱点
  • Sonnet:防守和修补弱点

共识收敛

  • 把所有人的观点展示给每个模型,问"考虑了这些视角后,你的新观点是什么?"
  • 重复直到观点趋于稳定

会议归档

会议记录必须持久化保存,不要放 /tmp(重启会清空)。

~/.openclaw/workspace/pr-review/
└── council-YYYY-MM-DD[-topic]/
    ├── README.md                    # 会议概览(用模板)
    ├── council-opus-prompt.txt      # Opus 的 prompt
    ├── council-opus-output.txt      # Opus 的完整输出
    ├── council-sonnet-prompt.txt
    ├── council-sonnet-output.txt
    ├── council-gpt-prompt.txt
    └── council-gpt-output.txt

归档步骤:

ARCHIVE=~/.openclaw/workspace/pr-review/council-$(date +%Y-%m-%d)-topic
mkdir -p "$ARCHIVE"
cp /tmp/council-*.txt "$ARCHIVE/"
# 然后写 README.md(模板见 references/session-readme-template.md)

→ README 模板见 references/session-readme-template.md


安全提醒

  • --force 会自动执行所有变更,务必在独立分支上使用
  • 并行任务之间不能有文件冲突,否则会互相覆盖
  • 每个 session 完成后先 tmux capture-pane 保存输出再关闭
  • 人设是激活知识用的,不是角色扮演秀——保持专业分析为主

Files

5 total
Select a file
Select a file to preview.

Comments

Loading comments…