ACP agents

기준일: 2026-07-26
난이도: 중급
공식 기준: ACP agents

개요

이 페이지는 OpenClaw ACP agents 도구(파라미터, 권한, 설정, CLI)를 공식 문서 기준으로 정리합니다.

공식 요약: Run external coding harnesses (Claude Code, Cursor, Gemini CLI, explicit Codex ACP, OpenClaw ACP, OpenCode) through the ACP backend

도구 가시성은 profile / allow·deny policy / sandbox / channel 권한에 따라 달라집니다. 최신 스키마는 항상 공식 문서를 우선합니다.

공식 문서 기반 상세

아래는 공식 tools/acp-agents 문서를 정리한 내용입니다. 코드 블록, 파라미터 이름, 기본값은 원문 그대로입니다.

Agent Client Protocol (ACP) sessions let OpenClaw run external coding harnesses (Claude Code, Cursor, Copilot, Droid, OpenClaw ACP, OpenCode, Gemini CLI, and other supported ACPX harnesses) through an ACP backend plugin. Each spawn is tracked as a background task.

ACP is the external-harness path, not the default Codex path. The native Codex app-server plugin owns /codex ... controls and the default openai/gpt-* embedded runtime for agent turns; ACP owns /acp ... controls and sessions_spawn({ runtime: "acp" }) sessions.

To let Codex or Claude Code connect as an external MCP client directly to existing OpenClaw channel conversations, use openclaw mcp serve instead of ACP.

Which page do I want?

You want to...	Use this	Notes
Bind or control Codex in the current conversation	/codex bind, /codex threads	Native Codex app-server path when the codex plugin is enabled: bound chat replies, image forwarding, model/fast/permissions, stop, and steer. ACP is an explicit fallback
Run Claude Code, Gemini CLI, explicit Codex ACP, or another external harness through OpenClaw	This page	Chat-bound sessions, /acp spawn, sessions_spawn({ runtime: "acp" }), background tasks, runtime controls
Expose an OpenClaw Gateway session as an ACP server for an editor or client	openclaw acp	Bridge mode: an IDE/client speaks ACP to OpenClaw over stdio/WebSocket
Reuse a local AI CLI as a text-only fallback model	CLI Backends	Not ACP: no OpenClaw tools, no ACP controls, no harness runtime

Does this work out of the box?

Yes, after installing the official ACP runtime plugin:

openclaw plugins install @openclaw/acpx
openclaw config set plugins.entries.acpx.enabled true

Source checkouts can use the local extensions/acpx workspace plugin after pnpm install. Run /acp doctor for a readiness check.

OpenClaw only teaches agents about ACP spawning when ACP is truly usable: ACP must be enabled, dispatch must not be disabled, the current session must not be sandbox-blocked, and a runtime backend must be loaded and healthy. If any condition fails, ACP skills and sessions_spawn ACP guidance stay hidden so the agent does not suggest an unavailable backend.

First-run gotchas

- If `plugins.allow` is set, it is a restrictive plugin inventory and **must** include `acpx`, or the installed ACP backend is intentionally blocked (`/acp doctor` reports the missing allowlist entry).
- The Codex ACP adapter ships with the `acpx` plugin and launches locally when possible.
- Codex ACP runs with an isolated `CODEX_HOME`. OpenClaw copies trusted project trust entries plus safe model/provider routing config (`model`, `model_provider`, `model_reasoning_effort`, `sandbox_mode`, and safe `model_providers.<name>` fields) from the host Codex config; auth, notifications, and hooks stay on the host config only.
- Other target harness adapters may be fetched on demand with `npx` on first use.
- Vendor auth must already exist on the host for that harness.
- If the host has no npm or network access, first-run adapter fetches fail until caches are pre-warmed or the adapter is installed another way.

Runtime prerequisites

ACP launches a real external harness process. OpenClaw owns routing,
background-task state, delivery, bindings, and policy; the harness owns
its provider login, model catalog, filesystem behavior, and native tools.

Before blaming OpenClaw, verify:

- `/acp doctor` reports an enabled, healthy backend.
- The target id is allowed by `acp.allowedAgents` when that allowlist is set.
- The harness command can start on the Gateway host.
- Provider auth is present for that harness (`claude`, `codex`, `gemini`, `opencode`, `droid`, etc.).
- The selected model exists for that harness - model ids are not portable across harnesses.
- The requested `cwd` exists and is accessible, or omit `cwd` and let the backend use its default.
- Permission mode matches the work. Non-interactive sessions cannot click native permission prompts, so write/exec-heavy coding runs usually need an ACPX permission profile that can proceed headlessly.

OpenClaw plugin tools and built-in OpenClaw tools are not exposed to ACP harnesses by default. Enable the explicit MCP bridges in ACP agents - setup only when the harness should call those tools directly.

Supported harness targets

With the acpx backend, use these ids as /acp spawn <id> or sessions_spawn({ runtime: "acp", agentId: "<id>" }) targets:

Harness id	Typical backend	Notes
claude	Claude Code ACP adapter	Requires Claude Code auth on the host.
codex	Codex ACP adapter	Explicit ACP fallback only when native /codex is unavailable or ACP is requested.
copilot	GitHub Copilot ACP adapter	Requires Copilot CLI/runtime auth.
cursor	Cursor CLI ACP (cursor-agent acp)	Override the acpx command if a local install exposes a different ACP entrypoint.
droid	Factory Droid CLI	Requires Factory/Droid auth or FACTORY_API_KEY in the harness environment.
fast-agent	fast-agent-mcp ACP adapter	Fetched on demand with uvx.
gemini	Gemini CLI ACP adapter	Requires Gemini CLI auth or API key setup.
iflow	iFlow CLI	Adapter availability and model control depend on the installed CLI.
kilocode	Kilo Code CLI	Adapter availability and model control depend on the installed CLI.
kimi	Kimi/Moonshot CLI	Requires Kimi/Moonshot auth on the host.
kiro	Kiro CLI	Adapter availability and model control depend on the installed CLI.
mux	Mux CLI ACP adapter	Fetched on demand with npx.
opencode	OpenCode ACP adapter	Requires OpenCode CLI/provider auth.
openclaw	OpenClaw Gateway bridge through openclaw acp	Lets an ACP-aware harness talk back to an OpenClaw Gateway session.
qoder	Qoder CLI	Adapter availability and model control depend on the installed CLI.
qwen	Qwen Code / Qwen CLI	Requires Qwen-compatible auth on the host.
trae	Trae CLI ACP adapter	Adapter availability and model control depend on the installed CLI.

pi (pi-acp) is also registered in the acpx backend but is not a coding harness in the same sense as the others above.

Custom acpx agent aliases can be configured in acpx itself, but OpenClaw policy still checks acp.allowedAgents and any agents.entries.*.runtime.acp.agent mapping before dispatch.

Operator runbook

Quick /acp flow from chat:

Spawn

`/acp spawn claude --bind here`,
`/acp spawn gemini --mode persistent --thread auto`, or explicit
`/acp spawn codex --bind here`.

Work

Continue in the bound conversation or thread (or target the session key
explicitly).

Check state

`/acp status`

Tune

`/acp model <provider/model>`, `/acp permissions <profile>`,
`/acp timeout <seconds>`.

Steer

Without replacing context: `/acp steer tighten logging and continue`.

Stop

`/acp cancel` (current turn) or `/acp close` (session + bindings).

Lifecycle details

- Spawn creates or resumes an ACP runtime session, records ACP metadata in the OpenClaw session store, and may create a background task when the run is parent-owned.
- Parent-owned ACP sessions are treated as background work even when the runtime session is persistent; completion and cross-surface delivery go through the parent task notifier rather than acting like a normal user-facing chat session.
- Task maintenance closes terminal or orphaned parent-owned one-shot ACP sessions. Persistent ACP sessions are preserved while an active conversation binding remains; stale persistent sessions without an active binding are closed so they cannot be silently resumed after the owning task is done or its task record is gone.
- Bound follow-up messages go directly to the ACP session until the binding is closed, unfocused, reset, or expired.
- Gateway commands stay local. `/acp ...`, `/status`, and `/unfocus` are never sent as normal prompt text to a bound ACP harness.
- `cancel` aborts the active turn when the backend supports cancellation; it does not delete the binding or session metadata.
- `close` ends the ACP session from OpenClaw's point of view and removes the binding. A harness may still keep its own upstream history if it supports resume.
- The acpx plugin cleans up OpenClaw-owned wrapper and adapter process trees after `close`, and reaps stale OpenClaw-owned ACPX orphans during Gateway startup.
- Idle runtime workers are eligible for cleanup after the built-in idle period; stored session metadata remains available for `/acp sessions`.

Native Codex routing rules

Natural-language triggers that should route to the **native Codex plugin**
when it is enabled:

- "Bind this Discord channel to Codex."
- "Attach this chat to Codex thread `<id>`."
- "Show Codex threads, then bind this one."

Native Codex conversation binding is the default chat-control path.
OpenClaw dynamic tools still execute through OpenClaw, while Codex-native
tools such as shell/apply-patch execute inside Codex. For Codex-native
tool events, OpenClaw injects a per-turn native hook relay so plugin hooks
can block `before_tool_call`, observe `after_tool_call`, and route Codex
`PermissionRequest` events through OpenClaw approvals. Codex `Stop` hooks
are relayed to OpenClaw `before_agent_finalize`, where plugins can request
one more model pass before Codex finalizes its answer. The relay stays
deliberately conservative: it does not mutate Codex-native tool arguments
or rewrite Codex thread records. Use explicit ACP only when you want the
ACP runtime/session model. The embedded Codex support boundary is
documented in the
[Codex harness v1 support contract](https://docs.openclaw.ai/plugins/codex-harness-runtime#v1-support-contract).

Model / provider / runtime selection cheat sheet

- legacy Codex model refs - legacy Codex OAuth/subscription model route repaired by doctor.
- `openai/*` - native Codex app-server embedded runtime for OpenAI agent turns.
- `/codex ...` - native Codex conversation control.
- `/acp ...` or `runtime: "acp"` - explicit ACP/acpx control.

ACP-routing natural-language triggers

Triggers that should route to the ACP runtime:

- "Run this as a one-shot Claude Code ACP session and summarize the result."
- "Use Gemini CLI for this task in a thread, then keep follow-ups in that same thread."
- "Run Codex through ACP in a background thread."

OpenClaw picks `runtime: "acp"`, resolves the harness `agentId`, binds to
the current conversation or thread when supported, and routes follow-ups
to that session until close/expiry. Codex only follows this path when
ACP/acpx is explicit or the native Codex plugin is unavailable for the
requested operation.

For `sessions_spawn`, `runtime: "acp"` is advertised only when ACP is
enabled, the requester is not sandboxed, and an ACP runtime backend is
loaded. `acp.dispatch.enabled=false` pauses automatic ACP thread dispatch
but does not hide or block explicit `sessions_spawn({ runtime: "acp" })`
calls. It targets ACP harness ids such as `codex`, `claude`, `droid`,
`gemini`, or `opencode`. Do not pass a normal OpenClaw config agent id
from `agents_list` unless that entry is explicitly configured with
`agents.entries.*.runtime.type="acp"`; otherwise use the default sub-agent
runtime. When an OpenClaw agent is configured with
`runtime.type="acp"`, OpenClaw uses `runtime.acp.agent` as the underlying
harness id.

ACP versus sub-agents

Use ACP when you want an external harness runtime. Use native Codex app-server for Codex conversation binding/control when the codex plugin is enabled. Use sub-agents when you want OpenClaw-native delegated runs.

Area	ACP session	Sub-agent run
Runtime	ACP backend plugin (for example acpx)	OpenClaw native sub-agent runtime
Session key	agent:<agentId>:acp:<uuid>	agent:<agentId>:subagent:<uuid>
Main commands	/acp ...	/subagents ...
Spawn tool	sessions_spawn with runtime:"acp"	sessions_spawn (default runtime)

See also Sub-agents.

How ACP runs Claude Code

For Claude Code through ACP, the stack is:

1. OpenClaw ACP session control plane.
2. Official @openclaw/acpx runtime plugin.
3. Claude ACP adapter.
4. Claude-side runtime/session machinery.

ACP Claude is a harness session with ACP controls, session resume, background-task tracking, and optional conversation/thread binding.

CLI backends are separate text-only local fallback runtimes - see CLI Backends.

For operators, the practical rule is:

• Want /acp spawn, bindable sessions, runtime controls, or persistent harness work? Use ACP.
• Want simple local text fallback through the raw CLI? Use CLI backends.

Bound sessions

Mental model

• Chat surface - where people keep talking (Discord channel, Telegram topic, iMessage chat).
• ACP session - the durable Codex/Claude/Gemini runtime state OpenClaw routes to.
• Child thread/topic - an optional extra messaging surface created only by --thread ....
• Runtime workspace - the filesystem location (cwd, repo checkout, backend workspace) where the harness runs. Independent of the chat surface.

Current-conversation binds

/acp spawn <harness> --bind here pins the current conversation to the spawned ACP session - no child thread, same chat surface. OpenClaw keeps owning transport, auth, safety, and delivery. Follow-up messages in that conversation route to the same session; /new and /reset reset the session in place; /acp close removes the binding.

Examples:

/codex bind                                              # native Codex bind, route future messages here
/codex model gpt-5.4                                     # tune the bound native Codex thread
/codex stop                                              # control the active native Codex turn
/acp spawn codex --bind here                             # explicit ACP fallback for Codex
/acp spawn codex --thread auto                           # may create a child thread/topic and bind there
/acp spawn codex --bind here --cwd /workspace/repo       # same chat binding, Codex runs in /workspace/repo

Binding rules and exclusivity

- `--bind here` and `--thread ...` are mutually exclusive.
- `--bind here` only works on channels that advertise current-conversation binding; OpenClaw returns a clear unsupported message otherwise. Bindings persist across gateway restarts.
- On Discord, `spawnSessions` gates child thread creation for `--thread auto|here` - not `--bind here`.
- If you spawn to a different ACP agent without `--cwd`, OpenClaw inherits the **target agent's** workspace by default. Missing inherited paths (`ENOENT`/`ENOTDIR`) fall back to the backend default; other access errors (e.g. `EACCES`) surface as spawn errors.
- Gateway management commands stay local in bound conversations - `/acp ...` commands are handled by OpenClaw even when normal follow-up text routes to the bound ACP session; `/status` and `/unfocus` also stay local whenever command handling is enabled for that surface.

Thread-bound sessions

When thread bindings are enabled for a channel adapter:

- OpenClaw binds a thread to a target ACP session.
- Follow-up messages in that thread route to the bound ACP session.
- ACP output is delivered back to the same thread.
- Unfocus/close/archive/idle-timeout or max-age expiry removes the binding.
- `/acp close`, `/acp cancel`, `/acp status`, `/status`, and `/unfocus` are Gateway commands, not prompts to the ACP harness.

Required feature flags for thread-bound ACP:

- `acp.enabled=true`
- `acp.dispatch.enabled` is on by default (set `false` to pause automatic ACP thread dispatch; explicit `sessions_spawn({ runtime: "acp" })` calls still work).
- Channel-adapter thread session spawns enabled (default: `true`):
  - Discord/Telegram: `session.threadBindings.spawnSessions=true`

Thread binding support is adapter-specific. If the active channel adapter
does not support thread bindings, OpenClaw returns a clear
unsupported/unavailable message.

Thread-supporting channels

- Any channel adapter that exposes session/thread binding capability.
- Current built-in support: **Discord** threads/channels, **Telegram** topics (forum topics in groups/supergroups and DM topics).
- Plugin channels can add support through the same binding interface.

Persistent channel bindings

For non-ephemeral workflows, configure persistent ACP bindings in top-level bindings[] entries.

Binding model

Marks a persistent ACP conversation binding.

• bindings[].match (object): Identifies the target conversation. Per-channel shapes: - Discord channel/thread: match.channel="discord" + match.peer.id="<channelOrThreadId>" - Slack channel/DM: match.channel="slack" + match.peer.id="<channelId|channel:<channelId>|#<channelId>|userId|user:<userId>|slack:<userId>|<@userId>>". Prefer stable Slack ids; channel bindings also match replies inside that channel's threads. - Telegram forum topic: match.channel="telegram" + match.peer.id="<chatId>:topic:<topicId>" - WhatsApp DM/group: match.channel="whatsapp" + match.peer.id="<E.164|group JID>". Use E.164 numbers such as +15555550123 for direct chats and WhatsApp group JIDs such as 120363424282127706@g.us for groups. - iMessage DM/group: match.channel="imessage" + match.peer.id="<handle|chat_id:*|chat_guid:*|chat_identifier:*>". Prefer chat_id:* for stable group bindings.

• bindings[].agentId (string): The owning OpenClaw agent id.

  Optional ACP override.

• bindings[].acp.label (string): Optional operator-facing label.

• bindings[].acp.cwd (string): Optional runtime working directory.

• bindings[].acp.backend (string): Optional backend override.

Runtime defaults per agent

Use agents.entries.*.runtime to define ACP defaults once per agent:

• agents.entries.*.runtime.type="acp"
• agents.entries.*.runtime.acp.agent (harness id, e.g. codex or claude)
• agents.entries.*.runtime.acp.backend
• agents.entries.*.runtime.acp.mode
• agents.entries.*.runtime.acp.cwd

Override precedence for ACP bound sessions:

1. bindings[].acp.*
2. agents.entries.*.runtime.acp.*
3. Global ACP defaults (e.g. acp.backend)

예시

{
  agents: {
    list: [
      {
        id: "codex",
        runtime: {
          type: "acp",
          acp: {
            agent: "codex",
            backend: "acpx",
            mode: "persistent",
            cwd: "/workspace/openclaw",
          },
        },
      },
      {
        id: "claude",
        runtime: {
          type: "acp",
          acp: { agent: "claude", backend: "acpx", mode: "persistent" },
        },
      },
    ],
  },
  bindings: [
    {
      type: "acp",
      agentId: "codex",
      match: {
        channel: "discord",
        accountId: "default",
        peer: { kind: "channel", id: "222222222222222222" },
      },
      acp: { label: "codex-main" },
    },
    {
      type: "acp",
      agentId: "claude",
      match: {
        channel: "telegram",
        accountId: "default",
        peer: { kind: "group", id: "-1001234567890:topic:42" },
      },
      acp: { cwd: "/workspace/repo-b" },
    },
    {
      type: "route",
      agentId: "main",
      match: { channel: "discord", accountId: "default" },
    },
    {
      type: "route",
      agentId: "main",
      match: { channel: "telegram", accountId: "default" },
    },
  ],
  channels: {
    discord: {
      guilds: {
        "111111111111111111": {
          channels: {
            "222222222222222222": { requireMention: false },
          },
        },
      },
    },
    telegram: {
      groups: {
        "-1001234567890": {
          topics: { "42": { requireMention: false } },
        },
      },
    },
  },
}

Behavior

• OpenClaw ensures the configured ACP session exists after channel-specific admission and before use.
• Messages in that channel, topic, or chat route to the configured ACP session.
• Configured ACP bindings own their session route. Channel broadcast fan-out does not replace the configured ACP session for a matched binding.
• In bound conversations, /new and /reset reset the same ACP session key in place.
• Temporary runtime bindings (for example created by thread-focus flows) still apply where present.
• For cross-agent ACP spawns without an explicit cwd, OpenClaw inherits the target agent workspace from agent config.
• Missing inherited workspace paths fall back to the backend default cwd; non-missing access failures surface as spawn errors.

Start ACP sessions

Two ways to start an ACP session:

Use `runtime: "acp"` to start an ACP session from an agent turn or tool
call.

```json
{
  "task": "Open the repo and summarize failing tests",
  "runtime": "acp",
  "agentId": "codex",
  "thread": true,
  "mode": "session"
}
```

runtime defaults to subagent, so set runtime: "acp" explicitly for ACP sessions. If agentId is omitted, OpenClaw uses acp.defaultAgent when configured. mode: "session" requires thread: true to keep a persistent bound conversation.

Use `/acp spawn` for explicit operator control from chat.

```text
/acp spawn codex --mode persistent --thread auto
/acp spawn codex --mode oneshot --thread off
/acp spawn codex --bind here
/acp spawn codex --thread here
```

Key flags:

- `--mode persistent|oneshot`
- `--bind here|off`
- `--thread auto|here|off`
- `--cwd <absolute-path>`
- `--label <name>`

See [Slash commands](https://docs.openclaw.ai/tools/slash-commands).

sessions_spawn parameters

• task (string) (required): Initial prompt sent to the ACP session.

  Must be "acp" for ACP sessions.

• agentId (string): ACP target harness id. Falls back to acp.defaultAgent if set.

• thread (boolean) · default: false: Request thread binding flow where supported.

  "run" is one-shot; "session" is persistent. If thread: true and mode is omitted, OpenClaw may default to persistent behaviour per runtime path. mode: "session" requires thread: true.

• cwd (string): Requested runtime working directory (validated by backend/runtime policy). If omitted, ACP spawn inherits the target agent workspace when configured; missing inherited paths fall back to backend defaults, while real access errors are returned.

• label (string): Operator-facing label used in session/banner text.

• resumeSessionId (string): Resume an existing ACP session instead of creating a new one. The agent replays its conversation history via session/load. Requires runtime: "acp".

  "parent" streams initial ACP run progress summaries back to the requester session as system events. OpenClaw records the full relay history in the child agent's SQLite state and removes it with the child session. Parent progress streams show assistant commentary and ACP status progress by default unless streaming.progress.commentary=false. Discord also defaults parent previews to progress mode when no stream mode is configured. Status progress still honors acp.stream.tagVisibility, so tags such as plan remain hidden unless explicitly enabled.

ACP sessions_spawn runs use agents.defaults.subagents.runTimeoutSeconds for their default child turn limit. The tool does not accept per-call timeout overrides (runTimeoutSeconds/timeoutSeconds are rejected with a config-the-default error).

• model (string): Explicit model override for the ACP child session. Codex ACP spawns normalize OpenAI refs such as openai/gpt-5.4 to Codex ACP startup config before session/new; slash forms such as openai/gpt-5.4/high also set Codex ACP reasoning effort. When omitted, sessions_spawn({ runtime: "acp" }) uses existing subagent model defaults (agents.defaults.subagents.model or agents.entries.*.subagents.model) when configured; otherwise it lets the ACP harness use its own default model. Other harnesses must advertise ACP models and support session/set_model; otherwise OpenClaw/acpx fails clearly instead of silently falling back to the target agent default.

• thinking (string): Explicit thinking/reasoning effort. For Codex ACP, minimal maps to low effort, low/medium/high/xhigh map directly, and off omits the reasoning-effort startup override. When omitted, ACP spawns use existing subagent thinking defaults and per-model agents.defaults.models["provider/model"].params.thinking for the selected model.

Spawn bind and thread modes

| Mode   | Behavior                                                               |
| ------ | ----------------------------------------------------------------------- |
| `here` | Bind the current active conversation in place; fail if none is active. |
| `off`  | Do not create a current-conversation binding.                          |

Notes:

- `--bind here` is the simplest operator path for "make this channel or chat Codex-backed."
- `--bind here` does not create a child thread.
- `--bind here` is only available on channels that expose current-conversation binding support.
- `--bind` and `--thread` cannot be combined in the same `/acp spawn` call.



| Mode   | Behavior                                                                                            |
| ------ | ------------------------------------------------------------------------------------------------- |
| `auto` | In an active thread: bind that thread. Outside a thread: create/bind a child thread when supported. |
| `here` | Require current active thread; fail if not in one.                                                  |
| `off`  | No binding. Session starts unbound.                                                                 |

Notes:

- On non-thread binding surfaces, default behavior is effectively `off`.
- Thread-bound spawn requires channel policy support:
  - Discord/Telegram: `session.threadBindings.spawnSessions=true`
- Use `--bind here` when you want to pin the current conversation without creating a child thread.

Delivery model

ACP sessions can be either interactive workspaces or parent-owned background work. The delivery path depends on that shape.

Interactive ACP sessions

Interactive sessions are meant to keep talking on a visible chat surface:

- `/acp spawn ... --bind here` binds the current conversation to the ACP session.
- `/acp spawn ... --thread ...` binds a channel thread/topic to the ACP session.
- Persistent configured `bindings[].type="acp"` route matching conversations to the same ACP session.

Follow-up messages in the bound conversation route directly to the ACP
session, and ACP output is delivered back to that same
channel/thread/topic.

What OpenClaw sends to the harness:

- Normal bound follow-ups are sent as prompt text, plus attachments only when the harness/backend supports them.
- `/acp` management commands and local Gateway commands are intercepted before ACP dispatch.
- Runtime-generated completion events are materialized per target. OpenClaw agents get OpenClaw's internal runtime-context envelope; external ACP harnesses get a plain prompt with the child result and instruction. The raw `<<<BEGIN_OPENCLAW_INTERNAL_CONTEXT>>>` envelope should never be sent to external harnesses or persisted as ACP user transcript text.
- ACP transcript entries use the user-visible trigger text or the plain completion prompt. Internal event metadata stays structured in OpenClaw where possible and is not treated as user-authored chat content.

Parent-owned one-shot ACP sessions

One-shot ACP sessions spawned by another agent run are background
children, similar to sub-agents:

- The parent asks for work with `sessions_spawn({ runtime: "acp", mode: "run" })`.
- The child runs in its own ACP harness session.
- Child turns run on the same background lane used by native sub-agent spawns, so a slow ACP harness does not block unrelated main-session work.
- Completion reports back through the task-completion announce path. OpenClaw converts internal completion metadata into a plain ACP prompt before sending it to an external harness, so harnesses do not see OpenClaw-only runtime context markers.
- The parent rewrites the child result in normal assistant voice when a user-facing reply is useful.

Do **not** treat this path as a peer-to-peer chat between parent and
child. The child already has a completion channel back to the parent.

sessions_send and A2A delivery

`sessions_send` can target another session after spawn. For normal peer
sessions, OpenClaw uses an agent-to-agent (A2A) follow-up path after
injecting the message:

- Wait for the target session's reply.
- Optionally let requester and target exchange a bounded number of follow-up turns.
- Ask the target to produce an announce message.
- Deliver that announce to the visible channel or thread.

That A2A path is a fallback for peer sends where the sender needs a
visible follow-up. It stays enabled when an unrelated session can see and
message an ACP target, for example under broad `tools.sessions.visibility`
settings.

OpenClaw skips the A2A follow-up only when the requester is the parent of
its own parent-owned one-shot ACP child. In that case, running A2A on top
of task completion can wake the parent with the child's result, forward
the parent's reply back into the child, and create a parent/child echo
loop. The `sessions_send` result reports `delivery.status="skipped"` for
that owned-child case because the completion path is already responsible
for the result.

Resume an existing session

Use `resumeSessionId` to continue a previous ACP session instead of
starting fresh. The agent replays its conversation history via
`session/load`, so it picks up with full context of what came before.

```json
{
  "task": "Continue where we left off - fix the remaining test failures",
  "runtime": "acp",
  "agentId": "codex",
  "resumeSessionId": "<previous-session-id>"
}
```

Common use cases:

- Hand off a Codex session from your laptop to your phone - tell your agent to pick up where you left off.
- Continue a coding session you started interactively in the CLI, now headlessly through your agent.
- Pick up work that was interrupted by a gateway restart or idle timeout.

Notes:

- `resumeSessionId` only applies when `runtime: "acp"`; the default sub-agent runtime ignores this ACP-only field.
- `streamTo` only applies when `runtime: "acp"`; the default sub-agent runtime ignores this ACP-only field.
- `resumeSessionId` is a host-local ACP/harness resume id, not an OpenClaw channel session key; OpenClaw still checks ACP spawn policy and target agent policy before dispatch, while the ACP backend or harness owns authorization for loading that upstream id.
- `resumeSessionId` restores the upstream ACP conversation history; `thread` and `mode` still apply normally to the new OpenClaw session you are creating, so `mode: "session"` still requires `thread: true`.
- The target agent must support `session/load` (Codex and Claude Code do).
- If the session id is not found, the spawn fails with a clear error - no silent fallback to a new session.

Post-deploy smoke test

After a gateway deploy, run a live end-to-end check rather than trusting
unit tests:

1. Verify the deployed gateway version and commit on the target host.
2. Open a temporary ACPX bridge session to a live agent.
3. Ask that agent to call `sessions_spawn` with `runtime: "acp"`, `agentId: "codex"`, `mode: "run"`, and task `Reply with exactly LIVE-ACP-SPAWN-OK`.
4. Verify `accepted=yes`, a real `childSessionKey`, and no validator error.
5. Clean up the temporary bridge session.

Keep the gate on `mode: "run"` and skip `streamTo: "parent"` -
thread-bound `mode: "session"` and stream-relay paths are separate richer
integration passes.

Sandbox compatibility

ACP sessions currently run on the host runtime, not inside the OpenClaw sandbox.

Security boundary:

• The external harness can read/write according to its own CLI permissions and the selected cwd.
• OpenClaw's sandbox policy does not wrap ACP harness execution.
• OpenClaw still enforces ACP feature gates, allowed agents, session ownership, channel bindings, and Gateway delivery policy.
• Use runtime: "subagent" for sandbox-enforced OpenClaw-native work.

Current limitations:

• If the requester session is sandboxed, ACP spawns are blocked for both sessions_spawn({ runtime: "acp" }) and /acp spawn.
• sessions_spawn with runtime: "acp" does not support sandbox: "require".

Session target resolution

Most /acp actions accept an optional session target (session-key, session-id, or session-label).

Resolution order:

1. Explicit target argument (or --session for /acp steer)
   • tries key
   • then UUID-shaped session id
   • then label
2. Current thread binding (if this conversation/thread is bound to an ACP session).
3. Current requester session fallback.

Current-conversation bindings and thread bindings both participate in step 2.

If no target resolves, OpenClaw returns a clear error (Unable to resolve session target: ...).

ACP controls

Command	What it does	Example
/acp spawn	Create ACP session; optional current bind or thread bind.	/acp spawn codex --bind here --cwd /repo
/acp cancel	Cancel in-flight turn for target session.	/acp cancel agent:codex:acp:<uuid>
/acp steer	Send steer instruction to running session.	/acp steer --session support inbox prioritize failing tests
/acp close	Close session and unbind thread targets.	/acp close
/acp status	Show backend, mode, state, runtime options, capabilities.	/acp status
/acp set-mode	Set runtime mode for target session.	/acp set-mode plan
/acp set	Generic runtime config option write.	/acp set model openai/gpt-5.4
/acp cwd	Set runtime working directory override.	/acp cwd /Users/user/Projects/repo
/acp permissions	Set approval policy profile.	/acp permissions strict
/acp timeout	Set runtime timeout (seconds).	/acp timeout 120
/acp model	Set runtime model override.	/acp model anthropic/claude-opus-4
/acp reset-options	Remove session runtime option overrides.	/acp reset-options
/acp sessions	List recent ACP sessions from store.	/acp sessions
/acp doctor	Backend health, capabilities, actionable fixes.	/acp doctor
/acp install	Print deterministic install and enable steps.	/acp install

Runtime controls (spawn, cancel, steer, close, status, set-mode, set, cwd, permissions, timeout, model, and reset-options) require owner identity from external channels and operator.admin from internal Gateway clients. Authorized non-owner senders can still use sessions, doctor, install, and help. For non-owner senders, /acp sessions lists only the current bound or requester session; owner identity and operator.admin clients see all recent sessions.

/acp status shows the effective runtime options plus runtime-level and backend-level session identifiers. Unsupported-control errors surface clearly when a backend lacks a capability. Commands that accept target tokens (session-key, session-id, or session-label) resolve them through gateway session discovery, including custom per-agent session.store roots. /acp sessions does not accept a target token.

Runtime options mapping

/acp has convenience commands and a generic setter. Equivalent operations:

Command	Maps to	Notes
/acp model <id>	runtime config key model	For Codex ACP, OpenClaw normalizes openai/<model> to the adapter model id and maps slash reasoning suffixes such as openai/gpt-5.4/high to reasoning_effort.
/acp set thinking <level>	canonical option thinking	OpenClaw sends the backend-advertised equivalent when present, preferring thinking, then effort, reasoning_effort, or thought_level. For Codex ACP, the adapter maps values to reasoning_effort.
/acp permissions <profile>	canonical option permissionProfile	OpenClaw sends the backend-advertised equivalent when present, such as approval_policy, permission_profile, permissions, or permission_mode.
/acp timeout <seconds>	canonical option timeoutSeconds	OpenClaw sends the backend-advertised equivalent when present, such as timeout or timeout_seconds.
/acp cwd <path>	runtime cwd override	Direct update.
/acp set <key> <value>	generic	key=cwd uses the cwd override path.
/acp reset-options	clears all runtime overrides	-

acpx harness, plugin setup, and permissions

For acpx harness configuration (Claude Code / Codex / Gemini CLI aliases), the plugin-tools and OpenClaw-tools MCP bridges, and ACP permission modes, see ACP agents - setup.

문제 해결

Symptom	Likely cause	Fix
ACP runtime backend is not configured	Backend plugin missing, disabled, or blocked by plugins.allow.	Install and enable backend plugin, include acpx in plugins.allow when that allowlist is set, then run /acp doctor.

…이하 생략. 전체 내용은 공식 문서를 확인하세요.

이 페이지는 길이상 일부가 잘렸습니다. 전체: https://docs.openclaw.ai/tools/acp-agents

검증 체크리스트

• 해당 tool이 활성 profile/policy에서 허용되는지 확인
• sandbox / elevated / host 실행 경로 정책을 이해했는지 확인
• 채널·에이전트 권한과 충돌하지 않는지 확인
• 공식 CLI/--help와 문서 옵션이 버전과 맞는지 확인

관련

• Tools 목록 인덱스
• 공식: Tools overview, Tools and custom providers