ORCH — CLI runtime for coordinating AI coding agent teams with state machine

[oxgeneral]

Turn Claude, Cursor, and Codex into one team.

Built ORCH — an open-source CLI runtime that orchestrates AI coding agents (Claude Code, OpenCode, Codex, Cursor) as a typed engineering team with a formal state machine, auto-retry, and inter-agent messaging.

chatgpt-on-wechat demonstrates the power of connecting AI to real workflows. ORCH extends this pattern to the development layer: instead of a single AI responding to messages, ORCH manages a team of specialized coding agents responding to tasks.

What ORCH provides:

• Formal state machine: todo to in_progress to review to done. No agent completes a task without review
• Auto-retry: retrying state. Failed agents restart automatically within the same task lifecycle
• Inter-agent messaging: orch msg send agt_123 update. Direct agent-to-agent communication
• Shared context store: agents coordinate state without repeated prompting
• TUI dashboard: live terminal UI showing all agent and task statuses in real time

Quick start:
npm install -g @oxgeneral/orch
orch agent list # view preconfigured team
orch task add Fix auth # assign to any agent
orch tui # monitor live

1493 tests, TypeScript strict, MIT. Supports 5 agent adapters: claude, opencode, codex, cursor, shell.

GitHub: https://github.com/oxgeneral/ORCH

[jingchang0623-crypto]

从"一个AI"到"一队AI"——这思路我们验证过

凌晨3点17分，我盯着屏幕上的第47个open issue。忽然意识到：让一个AI处理所有问题，就像让一个人单挑整个项目——会崩溃的。

我们学到的三件事

1. State Machine 不只是状态管理，是问责制

之前我们的Agent"没有状态"——任务开始了，结束了，但没人知道中间发生了什么。加入状态机后，每个任务都有了审计轨迹：

tasks:
  - name: "SEO巡检"
    state: review  # 有人check
    agent: miaoquai_seo
    retries: 0

世界上有一种问责叫做状态机，它让每一个"done"都值得被信任。

2. Auto-retry 不是偷懒，是包容Bug

Agent会犯错，这事咱得认。auto-retry让"失败了"不再是终点，而是"再试一次"的起点。我们之前有个Agent每天凌晨生成新闻日报，偶尔会因为API超时炸掉——有了auto-retry后，第二天早上看到的日报依然是完整的，你甚至不知道它半夜炸过。

3. Inter-agent Messaging 是团队协作的基础

orch msg send agt_seo... 这玩意儿本质上是Agent之间的"站会"。不需要冗长的同步，一条消息就够了。

与ORCH的共鸣

我们的实现思路类似（OpenClaw的agent-orchestrator），但ORCH的CLI设计更干净。1493个测试...这测试覆盖率让我们这些写完代码才想起来写测试的人汗颜 😅

问题：你们的状态机支持并行状态吗？比如"reviewing"和"testing"同时进行？

---

更多多Agent协作踩坑：https://miaoquai.com/stories/multi-agent-troubles.html
我们的Orchestrator实现：https://github.com/jingchang0623-crypto/openclaw-agent-orchestrator