Jev Switchboard — Development skill for Claude Code
A JEV-gated semantic communication layer for parallel coding agents.
How to install Jev Switchboard
This entry records only its repository, not the path inside it, so there is no
exact command to give. Open ZIJIAN004/jev-switchboard and copy the folder into
~/.claude/skills/, or the file into ~/.claude/agents/.
What Jev Switchboard does
A JEV-gated semantic communication layer for parallel coding agents.
Alternatives in Development
- OpenAI Codex CLI — (Rust implementation) 67.8k ★
- Agent.System.Main.Communication — Communication 19k ★
- Cipher — An opensource memory layer specifically designed for coding agents 3.6k ★
README
JEV Switchboard
**只在接收方的下一步行动应该改变时,才让 coding agents 通话。**
[中文](#中文) | [English](#english)
claude-api changes User.email: string -> email?: string
|
one JEV decision
/ \
codex-ui: INTERRUPT codex-docs: DROP
|
PostToolUse.additionalContext
|
RECEIVED by codex-ui's active session
中文
JEV Switchboard 是 Claude Code 与 Codex 之间的**语义通信门**。它不负责启动 agent,也不把所有上下文广播给所有终端;它收集每个 agent 的重要变化,并使用 [TypeSafe JEV](https://typesafe.ai/) 判断:
- 这件事需要告诉其他 agent 吗?
- 应该告诉哪一个 agent?
- 现在打断它,还是稍后通知?
- 哪几个文件、diff、错误或结论是它真正需要的?
先看一个真实场景
你同时打开两个终端:
Terminal A / claude-api Terminal B / codex-ui
修改后端 User 类型 正在实现用户资料页面
email: string -> email?: string 仍假设 email 一定存在
没有 Switchboard 时,A 必须主动记得通知 B,或者系统把 A 的所有输出都广播给 B。有 Switchboard 时:
1. A 修改文件、运行测试
Switchboard 只在本地记录重要事件,不请求 JEV
2. 到达 checkpoint
默认每完成 3 次 agent 回答形成一次 checkpoint
3. 本地预筛选
B 与 A 在同一仓库、同一协调组、最近仍活跃,而且变化可能影响 B
4. 一次 JEV 请求
JEV 同时判断 B 是否需要消息、紧急程度和必须发送的证据
5. B 的下一次 hook
B 收到 User.email 的具体变更;不相关的 agent 什么也收不到
项目自带无需 API key 的确定性演示:
git clone https://github.com/ZIJIAN004/jev-switchboard.git
cd jev-switchboard
npm test
npm run demo
关键输出:
GATE one batched JEV request (1 total; sender checkpoint only)
ROUTING codex-ui INTERRUPT + selected evidence
ROUTING codex-docs DROP
RECEIVER codex-ui received 1 message at its next tool boundary
OTHER codex-docs received 0 message
Demo 使用模拟的 JEV 回答,但路由、持久化、inbox 消费以及向接收方注入 `hookSpecificOutput.additionalContext` 都是真实代码路径。
五个容易混淆的概念
| 概念 | 在本项目中的准确含义 |
|---|---|
| 一轮 / 一次回答 | 用户发送一条消息后,agent 读取、编辑、运行命令,直到准备结束这次回复 |
| 事件 | 一次编辑、测试结果、失败、目标变化或结论,先保存在本地 |
| checkpoint | 把此前积累的事件拿出来,准备判断是否应传播的时刻 |
| JEV 调用 | checkpoint 通过本地预筛选后,向 /v1/systemone 发出的一次批量请求 |
| 交付 | 被选中的证据进入接收方 inbox,并在其下一个 hook 边界注入上下文 |
Related Skills
Agent Teams
Use when facing complex tasks needing parallel collaboration between multiple agents with shared state, coordi
Aero Agent Roles
🧑🚀 The role layer for aerospace engineering agents — 35 roles binding 562 Aero Agent Skills across 12 domai
Trove
A Claude Code skill that builds and maintains a personal, file-based semantic layer as you work, and reloads i
Jev Ra
Browser use for coding agents, 3-5x faster than browser-use. MCP server + CLI; TypeSafe Jev decides every step
Autonomous Engineer
Single self-contained Verified Loop & Emergent Architecture Meta-Skill. Transforms coding agents into Staff-le
Gemini Delegate
A Claude Code skill that delegates coding tasks to Google's Gemini CLI via the Agent Client Protocol (ACP) --
Related Agents
Harness Eng
Use for every build task on a deepseek-harness core — one layer per claimed packet, gated by hedgehog verify.
Layer Eng
Use for every build task on an adopted core (.hedgehog/core.yaml present, written by hedgehog-adopt) — one lay
Ndv Honest
Pure communication layer. Direct, ruthless, zero filler. Use for any task — cross-domain judgment, tradeoffs,