为什么值得关注

它引入了一种新颖的多智能体协商流程，利用已有的智能体 CLI 安装，无需额外 API 或守护进程，并将所有交互记录在案以供审计。

适合谁使用

使用 Claude Code、Codex、Cursor、Gemini、Grok 或 OpenCode 的开发者
希望在代码更改前获得多种 AI 意见的团队
探索多智能体工作流的进阶用户
评估架构决策或迁移的工程师

典型使用场景

决定是否从一种框架迁移到另一种
获取多个智能体对重构方案的意见
通过安全检查审核复杂实现策略
审计不同智能体 CLI 如何处理同一任务

项目优势

只读成员调用确保安全探索
优雅处理缺失或失败的智能体
无需额外 API 密钥或守护进程，仅需已有 CLI
协商日志存储在 ~/.council/ 中，便于透明审计

使用前须知

至少需要安装并登录两个不同的智能体 CLI
协商过程可能比单一智能体响应更慢
仅支持 Agent Skills 格式或手动指令注入的智能体 CLI

README 快速开始

🏛 council

An agent skill that stops your coding agent from winging it: before giving you a final plan, whichever agent you're talking to must convene the other agent CLIs installed on your machine and deliberate.

Say:

council, 2 turns: should we migrate this app from Express to Hono?

and the lead agent (Claude Code, Codex, Cursor, Gemini, Grok, or OpenCode — whoever you asked) will:

Draft a plan.
Present it independently to every other agent CLI on your machine — each member gives exactly one reply per turn (verdict: SHIP IT / SHIP WITH CHANGES / RETHINK, strongest objections first). Members never see each other's replies.
Revise the plan, telling members what changed — and repeat for as many turns as you asked.
Answer with the whole deliberation as a conversation, plus the final plan.

Everything is one-shot CLI invocations under the hood (no ACP, no daemons, no API keys beyond the CLIs you already have installed and logged in). All member invocations are read-only: they may read your repo, never touch it. Failed or missing members are skipped gracefully and noted in the transcript.

Install

The skill is a single SKILL.md. Clone it where your agent looks for skills:

# Claude Code (personal)
git clone https://github.com/kitze/council ~/.claude/skills/council

# Claude Code (per-project)
git clone https://github.com/kitze/council .claude/skills/council

For other agents that support the Agent Skills format, clone into their skills directory (e.g. ~/.codex/skills/council), or just hand your agent the SKILL.md as instructions — it's self-contained.

Requirements

At least one other agent CLI installed and logged in: claude, codex, cursor-agent, gemini, grok, or opencode. The more you have, the bigger the council.

Notes

Deliberation sessions (plans, per-round prompts and replies) are kept in ~/.council/ so you can audit what every member actually said.
SKILL.md embeds the exact headless flags each CLI needs (cursor --trust, gemini --skip-trust, codex read-only sandbox, and the grok --system-prompt-override workaround for its silent headless cancellation) — found the hard way so you don't have to.

License

MIT

项目描述

🏛 Agent skill: your coding agent must convene the other agent CLIs on your machine and deliberate for X turns before giving you a plan

council