acp

name	acp
description	Delegate sub-tasks to remote ACP (Agent Client Protocol) peers such as Claude Code or Codex via the built-in acp_* host actions. Use when the user asks to hand work to another agent, wants a second opinion, or needs long / sandboxed execution that shouldn't block this chat.

/acp — Delegate to remote agents

Five host-side actions are registered automatically when the AgentLite host has ACP peers configured:

• acp_list_remote_agents
• acp_new_session
• acp_prompt
• acp_cancel
• acp_close_session

These are host actions, not direct MCP tools. Reach them through mcp__agentlite__call_action (and mcp__agentlite__search_actions to discover / introspect). If search_actions with query "+acp" returns nothing, no peers are configured and this skill does not apply.

When to use

• User says "ask codex", "hand this to claude code", "have do X".
• A coding task is long, needs a sandboxed runner, or benefits from a second agent's judgment.
• You want parallel work on a sibling directory without blocking this conversation.

Do not use ACP for anything you can finish in this turn — delegation has a round-trip cost and the result arrives as a later chat message, not inline.

Flow

Each step is a call_action invocation: call_action({ name: "<action>", payload: {...} }).

1. Discover — call_action({ name: "acp_list_remote_agents" }) returns { agents: [{ name, description?, agent_info? }] }. Skip if the user already named a peer.
2. Open — call_action({ name: "acp_new_session", payload: { peer, cwd? } }) → { session_id }. cwd defaults to the caller group's workdir; override only when the user points at a specific path.
3. Prompt — call_action({ name: "acp_prompt", payload: { session_id, prompt: ContentBlock[] } }) → { ok: true }. Returns immediately. The peer runs in the background; AgentLite writes an artifact and injects a completion notice into this chat when it finishes. Do not poll.
4. Cancel (optional) — call_action({ name: "acp_cancel", payload: { session_id } }) if the user changes their mind while a prompt is in-flight. You'll still get a terminal notice with stop_reason: "cancelled".
5. Close — call_action({ name: "acp_close_session", payload: { session_id } }) when the conversation is over. The peer child process stays alive for reuse; only local tracking is dropped.

Acknowledge the hand-off in one line ("Sent to codex — I'll ping when it's back") and stop. The completion notice arrives as its own message.

ContentBlock shape

prompt is the ACP spec array. Most of the time you just want:

[{ "type": "text", "text": "Summarize README.md in three bullets." }]

Other block types (image, audio, resource_link) are only valid if the peer's agent_info advertised the matching promptCapabilities.

Pitfalls

• Session / peer mismatch — a session_id belongs to exactly one peer. Don't reuse it across peers.
• Polling — there is no acp_get_result. The notice will arrive on its own; asking the user "any update?" is wrong.
• Forgetting cwd — if the user names a subproject, pass its absolute path as cwd in acp_new_session; otherwise the peer runs in the group folder and can't see the code.
• Synchronous assumptions — acp_prompt returning { ok: true } means "queued", not "done".
• Skipping search_actions — if you're unsure whether ACP is wired up on this host, search_actions({ query: "+acp" }) is the authoritative check.

See also: /capabilities for the full MCP tool surface; /status for a quick health check.