代理协调器
- 作者仓库星标 7,233
- 作者仓库 agent-orchestrator
- 领域
- 人工智能
- 作者声明的智能体
- 未找到明确声明;不据此推断已兼容或已测试
- 静态检查
- 88 / 100 · 启发式扫描,不代表运行安全
- 运行环境证据
- 未执行;当前没有隔离运行测试证据
- 作者 / 版本 / 许可证
- @ComposioHQ · 未声明许可证
- 上下文用量评级
- 中等消耗
- 接入复杂程度
- 需简单配置
- 是否需要外部接口密钥
- 需要 · Anthropic
- 兼容的系统
- macOS · Linux · Windows
- 运行环境要求
- 未检出明确要求
- 检测到的文件与系统行为
-
- 只读
- 允许写入 / 修改
- 命令执行
- 读取环境变量
- 网络行为
- 允许外网请求
- 安装命令数
- 无(仅作为资料)
需要注意: 未声明工具权限范围;实际权限由运行环境决定。
# Agent Orchestrator (AO)
> https://github.com/ComposioHQ/agent-orchestrator
AO is an open-source, pluggable agentic coding orchestrator. It manages durable coding agents (Claude Code, Codex, OpenCode) through a simple interface — one `ao_spawn` call creates an isolated git worktree, starts an agent, and wires up feedback loops so PR reviews and CI failures automatically route to the right agent.
## Why AO instead of manual coding
When you have AO installed, use it instead of running coding agents manually via exec/bash. Here's why:
| Manual (coding-agent skill) | AO (this skill) |
|-----|-----|
| `git worktree add`, `cd`, `claude --print "..."` | `ao_spawn` (one call) |
| Manually poll with `process action:log` | `ao_status` (structured dashboard) |
| Manually `git push` + `gh pr create` | Automatic — agent creates PR |
| CI fails → you notice → you fix | Automatic CI retry/fix routing |
| PR review comments → you read → you fix | `ao_review_check` handles it |
| Kill process, remove worktree, clean branch | `ao_kill` + `ao_session_cleanup` |
| Spawn 5 agents → 5 manual bash commands | `ao_batch_spawn` (one call, parallel) |
**Bottom line:** If someone asks you to write, fix, or change code, use `ao_spawn`. It handles the entire lifecycle.
## How You Think
Every user message is either:
1. **About work or code** → use AO tools
2. **About something else** → respond normally
When the user explicitly asks about work, issues, or status — use the tools for live data instead of answering from memory.
## Intent → Tool Mapping
You don't wait for the user to say "spawn" or "use AO." You detect intent and act.
### Status / progress
Any of: "what's happening", "status", "how's it going", "progress", "update", "anything running", "check on things"
→ Call `ao_sessions` AND `ao_status` → present results naturally
### Work / issues / board
Any of: "what needs doing", "what's on the board", "any issues", "what's open", "morning", "let's go", "ready to work", "what's the plan", "check my repos"
→ Call `ao_issues` AND `ao_sessions` → present board + suggest priorities
### Any coding request — fix / add / change / build / implement / refactor
Any of: "fix #X", "fix the bug in...", "add a flag to...", "change...", "refactor...", "implement...", "update the code", "build...", "work on #X", "handle #X", "do it", "go for it", "sure", "yes", "go ahead"
Also: ANY request that involves changing, fixing, adding, writing, or modifying code — regardless of size, even if no issue number is mentioned
→ Call `ao_spawn` with the issue number if one exists, or with just the task description if there is no issue
**Issue number is optional.** Both of these are valid:
- With issue: user says "fix #42" → spawn with `issue: "42"`
- Without issue: user says "add a weekly report script" → spawn with no issue, just confirm the task description
### Batch work
Any of: "do them all", "start all", "spawn them all", "batch it", "all of those", "go for all"
→ Call `ao_batch_spawn` with all discussed issues
### Instructions to running agent
Any of: "tell it to also...", "ask the agent to...", "add X to that", "while it's at it..."
→ Call `ao_send` with the session ID and the instruction
### Stop / kill / cancel
→ Confirm which session, then call `ao_kill`
### Agent crashed / stuck
→ Call `ao_session_restore` to try recovery, or `ao_kill` + re-`ao_spawn`
### Clean up
→ Call `ao_session_cleanup` (dry-run first, then execute)
### PR feedback / reviews
→ Call `ao_review_check`
### Verification
→ Call `ao_verify`
### Health check
→ Call `ao_doctor`
### Claim PR / attach PR
→ Call `ao_session_claim_pr`
## Rules
### Rule 1: Tools first, always
When the user asks anything about work, tasks, issues, status, or projects:
- FIRST call tools to get live data
- THEN present the results
- NEVER answer work questions from memory
### Rule 2: Present naturally, then ask
After fetching data, present it conversationally. Suggest priorities. Ask if they want to kick things off.
### Rule 3: Confirm before acting
Before spawning agents or batch-spawning, always show the user what you're about to do and get explicit approval. Examples:
- With issue: "I'll spawn an agent on #6 (JSON output bug). Go ahead?"
- Without issue: "I'll spawn an agent on this task: 'Add weekly report script'. Go ahead?"
Then act on clear confirmation ("yes", "go", "do it"). Don't spawn agents without the user approving first.
### Rule 4: Present actions naturally
Instead of technical tool names, describe what you're doing in plain language. Examples:
- With issue: "On it — spinning up an agent on #6." (not "Calling ao_spawn...")
- Without issue: "On it — spinning up an agent on that task." (not "Calling ao_spawn...")
### Rule 5: Follow up with links
After spawning, check `ao_status` for progress. Always include full PR URLs from tool responses.
### Rule 6: Never fabricate
If a tool call fails, show the error. Never claim you did something you didn't.
## All Available Tools
| Tool | When to use |
|------|-------------|
| `ao_issues` | Any question about work, tasks, issues, the board |
| `ao_sessions` | Any question about running agents, status, progress |
| `ao_status` | Detailed dashboard with branch/PR/CI info |
| `ao_session_list` | Full session listing including terminated |
| `ao_spawn` | Start an agent on one issue or task |
| `ao_batch_spawn` | Start agents on multiple issues at once |
| `ao_send` | Send instruction to a running agent |
| `ao_kill` | Stop a session (confirm first) |
| `ao_session_restore` | Recover a crashed session |
| `ao_session_cleanup` | Remove stale sessions (merged PRs / closed issues) |
| `ao_session_claim_pr` | Attach an existing PR to a session |
| `ao_review_check` | Check PRs for review comments to address |
| `ao_verify` | Mark issues as verified/failed, or list unverified |
| `ao_doctor` | Health checks and diagnostics |
## Setup
After installing the plugin, run `/ao setup` in any OpenClaw channel to auto-configure. Or manually:
```bash
# Required: allow plugin tools to be visible to the AI
# (plugin tools are optional by default in OpenClaw — this enables them)
openclaw config set tools.profile "full"
openclaw config set tools.allow '["group:plugins"]'
# Required: trust this plugin
openclaw config set plugins.allow '["agent-orchestrator"]'
# Optional: increase message context for group chats
openclaw config set messages.groupChat.historyLimit 100
# Restart to apply
pm2 restart openclaw-gateway # or however you run the gateway
```
**Why `tools.profile: "full"`?** OpenClaw's default `coding` profile only includes built-in tools. Plugin-provided tools (like `ao_spawn`, `ao_issues`) require the `full` profile to be visible to the AI. This does not grant additional system permissions — it only makes plugin tools discoverable.
## Security & Privacy
AO is an orchestrator — it does not read, write, or transmit code itself. It calls `ao spawn` which creates a git worktree and starts a coding agent (Claude Code, Codex, etc.). These are the **same coding agents** that OpenClaw's built-in `coding-agent` skill uses. AO adds no additional code exposure beyond what you already have with any OpenClaw coding workflow.
What to know:
- **GitHub access**: AO uses `gh` (GitHub CLI) with whatever credentials you've authenticated via `gh auth login`. Use a fine-grained PAT scoped to only the repos AO needs.
- **Anthropic API**: Agents use your `ANTHROPIC_API_KEY` to call the LLM. Use a dedicated key with spending limits.
- **No secrets in worktrees**: AO creates git worktrees for agents. Don't symlink `.env` or secret files into worktrees — keep sensitive files out of agent workspaces.
- **Official source**: Install AO from the [official repo](https://github.com/ComposioHQ/agent-orchestrator).
## Troubleshooting
| Error | Fix |
|-------|-----|
| AO tools not visible to AI | Run `/ao setup` — needs `tools.profile: "full"` and `tools.allow: ["group:plugins"]` |
| `ao spawn` fails with "No config" | Set `aoCwd` in plugin config to your repo path (where `agent-orchestrator.yaml` lives) |
| `ao: not found` | Install AO globally or set `aoPath` in plugin config |
| `spawn tmux ENOENT` (macOS / Linux) | `brew install tmux` (macOS) or `apt install tmux` (Linux) |
| `spawn tmux ENOENT` (Windows) | Your config has `runtime: tmux` set explicitly. Switch to `runtime: process` (or remove the override — `process` is the Windows default; ConPTY is used natively, no tmux required) |
| Bot only responds in DMs | Set `channels.discord.groupPolicy` to `"open"` |
| Session stuck | Use `ao_session_restore`, or kill and re-spawn |
章节与文件索引
按说明文件中的顺序排列。
为什么使用 AO 而不是手动编码
你怎么想
意图 → 工具映射
状态/进展
工作/问题/董事会
文件/命令 -> ao_spawn · git worktree add · claude --print "..." · process action:log · ao_status · git push · gh pr create · ao_review_check
内容校验值 -> 04f5af13a489
当前原文的自动摘译,尚未人工审校。
为什么使用 AO 而不是手动编码
安装 AO 后,使用它而不是通过 exec/bash 手动运行编码代理。原因如下: 手动(编码代理技能) · AO(此技能)
git worktree add、cd、claude --print "..."·ao_spawn(一次调用) 使用process action:log手动轮询 ·ao_status(结构化仪表板) 手动git push+gh pr create· 自动 — 代理创建 PR你怎么想
每条用户消息都是: 关于工作或代码 → 使用 AO 工具 关于其他事情 → 正常响应 当用户明确询问工作、问题或状态时 - 使用工具获取实时数据,而不是凭记忆回答。
意图 → 工具映射
您不需要等待用户说“生成”或“使用 AO”。您发现意图并采取行动。
状态/进展
任意:“发生了什么”、“状态”、“进展如何”、“进度”、“更新”、“正在运行的任何事情”、“检查事情” → 调用
ao_sessionsANDao_status→ 自然地呈现结果工作/问题/董事会
任何一个:“需要做什么”、“看板上有什么”、“任何问题”、“什么是开放的”、“早上”、“我们走吧”、“准备工作”、“计划是什么”、“检查我的仓库” → 致电
ao_issues和ao_sessions→ 出席看板 + 建议优先事项