agent-orchestrator

---
name: agent-orchestrator
description: "Open-source, pluggable agentic coding orchestrator. Manages durable coding agents (Claude Code, Codex, OpenCode) through a simple interface — spawn agents, track progress, and let feedback loops like PR reviews and CI failures automatically route to the right agents. Use for fixing bugs, building features, working on GitHub issues, checking status, and managing agent sessions."
metadata: {"openclaw": {"emoji": "🤖", "requires": {"bins": ["ao", "gh"], "anyBins": ["node", "npm"], "env": ["ANTHROPIC_API_KEY"]}, "os": ["darwin", "linux", "win32"]}}
---

# 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 |