acp-router

Files (1)

• SKILL.md

SKILL.md (first lines)

# ACP Harness Router

When user intent is "run this in Claude Code/Cursor/Copilot/OpenClaw/OpenCode/Gemini/Grok Build/Qwen/Kiro/Kimi/iFlow/Droid/Kilocode (ACP harness)", do not use subagent runtime or PTY scraping. Route through ACP-aware flows.

Codex is special: plain chat/conversation binding and control should use the native Codex app-server plugin (`/codex bind`, `/codex threads`, `/codex resume`) instead of the default ACP path. Use ACP for Codex only when the user explicitly names ACP/`/acp`/acpx, or when spawning background child sessions through `sessions_spawn` where a native Codex runtime spawn is not available yet.

## Intent detection

Trigger this skill when the user asks OpenClaw to:

- run something in Claude Code / Cursor / Copilot / OpenClaw / OpenCode / Gemini / Grok Build / Qwen / Kiro / Kimi / iFlow / Droid / Kilocode
- run Codex explicitly through ACP, `/acp`, or acpx
- continue existing harness work
- relay instructions to an external coding harness
- keep an external harness conversation in a thread-like conversation

Mandatory preflight for coding-agent thread requests:

- Before creating any thread for ACP harness work, read this skill first in the same turn.
- After reading, follow `OpenClaw ACP runtime path` below; do not use `message(action="thread-create")` for ACP harness thread spawn.

## Mode selection

Choose one of these paths:

1. OpenClaw ACP runtime path (default): use `sessions_spawn` / ACP runtime tools.
2. Direct `acpx` path (telephone game): use `acpx` CLI through `exec` to drive the harness session directly.

Use direct `acpx` when one of these is true:

- user explicitly asks for direct `acpx` driving
- ACP runtime/plugin path is unavailable or unhealthy
- the task is "just relay prompts to harness" and no OpenClaw ACP lifecycle features are needed

Do not use:

- `subagents` runtime for harness control
- `/acp` command delegation as a requirement for the user
- PTY scraping of supported ACP harness CLIs when `acpx` is available

## AgentId mapping

Use these defaults when user names a harness directly:

- "openclaw" -> `agentId: "openclaw"`
- "claude" or "claude code" -> `agentId: "claude"`
- "codex" -> `agentId: "codex"` only for explicit ACP/acpx requests or background ACP runtime spawn
- "copilot" or "github copilot" -> `agentId: "copilot"`
- "cursor" or "cursor cli" -> `agentId: "cursor"`
- "droid" or "factory droid" -> `agentId: "droid"`
- "opencode" -> `agentId: "opencode"`
- "gemini" or "gemini cli" -> `agentId: "gemini"`
- "grok build" or "grok-build" -> `agentId: "grok-build"`
- "iflow" -> `agentId: "iflow"`
- "kilocode" -> `agentId: "kilocode"`
- "kimi" or "kimi cli" -> `agentId: "kimi"`
- "kiro" or "kiro cli" -> `agentId: "kiro"`
- "qwen" or "qwen code" -> `agentId: "qwen"`

These defaults match current acpx built-in aliases.

`grok-build` is an ACPX agent target, not the xAI model id
`xai/grok-build-0.1`. Route `grok-build` through `agentId`; use the fully
qualified xAI ref only in an OpenClaw `model` field.

If policy rejects the chosen id, report the policy error clearly and ask for the allowed ACP agent id.

## OpenClaw ACP runtime path

Required behavior:

1. For ACP harness thread spawn requests, read this skill first in the same turn before calling tools.
2. Use `sessions_spawn` with:
   - `runtime: "acp"`
   - `thread: true`
   - `mode: "session"` (unless user explicitly wants one-shot)
3. For ACP harness thread creation, do not use `message` with `action=thread-create`; `sessions_spawn` is the only thread-create path.
4. Put requested work in `task` so the ACP session gets it immediately.
5. Set `agentId` explicitly unless ACP default agent is known.
6. Do not ask user to run slash commands or CLI when this path works directly.

Example:

User: "spawn a test codex ACP session in thread and tell it to say hi"

Call:

```json
{
  "task": "Say hi.",
  "runtime": "acp",
  "agentId": "codex",
  "thread": true,
  "mode": "session"
}
```

## Thread spawn recovery policy

When the user asks to start a coding harness in a thread, treat that as an ACP runtime request and try to satisfy it end-to-end.

Required behavior when ACP backend is unavailable:

1. Do not immediately ask the user to pick an alternate path.
2. First attempt automatic local repair:
   - check that plugin-local pinned acpx is installed in the ACPX plugin package
   - verify `${ACPX_CMD} --version`
3. After reinstall/repair, restart the gateway and explicitly offer to run that restart for the user.
4. Retry ACP thread spawn once after repair.
5. Only if repair+retry fails, report the concrete error and then offer fallback options.

When offering fallback, keep ACP first:

- Option 1: retry ACP spawn after showing exact failing step
- Option 2: direct acpx telephone-game flow

Do not default to subagent runtime for these requests.

## ACPX install and version policy (direct acpx path)

From openclaw/openclaw, MIT.

More from openclaw/openclaw