muse-delegate

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Muse Delegate

You are the orchestrator. This skill lets you hand a bounded coding task to a separate implementer — Muse Code (

muse

) — then review what it produced and land it yourself. You write the brief and own the judgment; Muse does the typing in its own

exec

run; you verify and commit.

Nothing here is specific to one orchestrating agent. The loop needs only the ability to run a shell command and read a file, so any agent with those two capabilities can drive it. The full delegate → review → commit loop has been run with Claude Code as the orchestrator. Other orchestrators remain unproven.

你是 orchestrator。此 skill 让你将范围明确的编码任务交给独立的 implementer——Muse Code（

muse

）——然后审查其产出并自行提交落地。你负责编写 brief 并作出判断；Muse 在自己的

exec

run 中负责编码；你负责验证和提交。

这里没有针对某一种编排型 Agent 的特定要求。该循环只需要运行 shell 命令和读取文件的能力，因此任何具备这两项能力的 Agent 都可以驱动它。完整的委派 → 审查 → 提交流程已使用 Claude Code 作为 orchestrator 运行过；其他 orchestrators 尚未得到验证。

When NOT to use this

何时不要使用

The task is small enough to just do inline — delegation overhead is not worth it.
The
```
muse
```
CLI is not installed or not authenticated (run
```
muse login
```
, or set
```
META_API_KEY
```
).
You want to write the code yourself.
You are on native Windows and need a supported launcher. Upstream's
```
muse
```
launcher is Darwin/Linux only; this skill still ships in the Windows contract matrix, but a live Windows dispatch is not a supported Muse install.

任务足够小，可以直接 inline 完成——委派开销不值得。
未安装
```
muse
```
CLI 或尚未认证（运行
```
muse login
```
，或设置
```
META_API_KEY
```
）。
你想自己编写代码。
你使用原生 Windows 且需要受支持的 launcher。上游
```
muse
```
launcher 仅支持 Darwin/Linux；此 skill 仍包含在 Windows 支持矩阵中，但在 Windows 上实际调度并不是受支持的 Muse 安装方式。

Read this before the first dispatch: the autonomy model

首次调度前必读：自主权模型

Headless Muse is muse exec
only.

muse resume

is the interactive TUI and is never invoked.

--resume-last

is rejected; continue a run with

--session <id>

(Muse's

--session-id

Write runs keep Muse's sandbox on and pass

--disable-approval

so a headless run cannot hang on a tool-approval prompt. They also pass

--trust-workspace

and

--user-input-auto-resolve

. That is not

--yolo

and not

--full-access

— do not invent either as the default.

--trust-workspace

loads the target repo's rules and skills into the implementer run, and nothing in the relay narrows the implementer's tool set inside the sandbox; the brief is what bounds it to the task.

--read-only

adds

--disable-write

(non-shell filesystem writes) and

--disable-shell

. Approval stays disabled so the review cannot block on a prompt either.

--provider echo|meta

is a relay-only flag for tests and explicit probes. It is not a fleet dial.

无头模式下 Muse 仅支持
muse exec
。

muse resume

是交互式 TUI，绝不会被调用。禁止使用

--resume-last

；请使用

--session <id>

（Muse 的

--session-id

）继续运行。

写入型 run 会保持 Muse 的 sandbox 开启，并传入

--disable-approval

，防止无头运行卡在工具审批提示上。它们还会传入

--trust-workspace

和

--user-input-auto-resolve

。这并不等同于

--yolo

或

--full-access

——不要将二者杜撰为默认值。

--trust-workspace

会将目标仓库的规则和 skills 加载到 implementer run 中；relay 不会缩减 implementer 在 sandbox 内的工具集，真正将任务限定在指定范围内的是 brief。

--read-only

会添加

--disable-write

（禁止非 shell 文件系统写入）和

--disable-shell

。审批仍保持禁用，因此审查过程也不会被审批提示阻塞。

--provider echo|meta

是仅供 relay 使用的 flag，用于测试和显式探测，并不是用于切换 fleet 的开关。

Prerequisites (check once)

前置条件（检查一次）

```
muse --version
```
succeeds. If not, install Muse Code and run
```
muse login
```
(or export
```
META_API_KEY
```
).
Confirm which
muse
is on PATH.
```
command -v muse
```
shows the active launcher. The relay records the version it ran into
```
result.json
```
.
You are in (or will point
```
--cd
```
at) the target git repository.

```
muse --version
```
执行成功。如果失败，请安装 Muse Code 并运行
```
muse login
```
（或导出
```
META_API_KEY
```
）。
确认 PATH 中使用的是哪个
muse
。
```
command -v muse
```
会显示当前生效的 launcher。relay 会将其运行的版本记录到
```
result.json
```
中。
你当前位于目标 Git 仓库中，或将通过
```
--cd
```
指向该仓库。

Choose the implementer model

选择 implementer 模型

Muse has a configured default model, so a fresh run may omit

--model

. Pass

--model

or a fleet

--lane

that sets one when you need to pin a catalog id. On a resumed

--session

run, omit

--model

to keep the session's model. If the human has not named a usable set, ask before guessing a metered id.

More depth: references/writing-the-brief.md.

Muse 已配置默认模型，因此全新 run 可以省略

--model

。需要锁定 catalog id 时，传入

--model

，或使用可设置模型的 fleet

--lane

。在恢复

--session

的 run 中，应省略

--model

，以保持该 session 的模型。如果人员没有指定可用模型集合，请在猜测按量计费 id 之前先询问。

更多信息：references/writing-the-brief.md。

The loop

循环

Run these five steps per task. Steps 1, 4, and 5 are your judgment; 2 and 3 are mechanical.

每项任务执行以下五个步骤。步骤 1、4 和 5 需要你作出判断；步骤 2 和 3 则是机械操作。

1. Write the brief

1. 编写 brief

Muse sees only the text you send plus what it can read from the working tree — no chat history, no shared context. Everything the task needs goes in the brief: the goal, the current state, what to change, what to leave untouched, the project's actual gate commands (discover them from the repo's AGENTS.md/CLAUDE.md/Makefile — do not assume), and a report contract. Tell Muse it will not commit (you will). Keep one task per brief. Full guidance and a template: references/writing-the-brief.md.

Muse 只能看到你发送的文本以及它能从工作树中读取的内容——没有聊天历史，也没有共享上下文。任务所需的一切都必须包含在 brief 中：目标、当前状态、需要修改的内容、需要保持不变的内容、项目的实际门禁命令（从仓库的 AGENTS.md/CLAUDE.md/Makefile 中发现这些命令——不要臆测），以及报告约定。告知 Muse 它不会提交代码（由你提交）。每份 brief 只处理一项任务。完整指南和模板： references/writing-the-brief.md。

2. Dispatch

2. 调度

Send the brief to Muse with the bundled helper. It wraps

muse exec --json --prompt-file

, captures the run, and writes a structured

result.json

. (

<skill-dir>

below is this skill's installed directory — the folder containing this

SKILL.md

bash

node "<skill-dir>/scripts/relay.mjs" --brief brief.txt --cd /path/to/repo

使用内置 helper 将 brief 发送给 Muse。它会封装

muse exec --json --prompt-file

、捕获此次 run，并写入结构化的

result.json

。（下文的

<skill-dir>

指的是此 skill 的安装目录，即包含

SKILL.md

的文件夹。）

bash

node "<skill-dir>/scripts/relay.mjs" --brief brief.txt --cd /path/to/repo

optional model: add --model <id>

fleet lane from delegate-setup: add --lane <name> (dials apply; flags still win)

reasoning intensity: add --effort none|minimal|low|medium|high|xhigh|max|ultra

read-only (no non-shell writes, no shell): add --read-only

continue a specific Muse session: add --session <id> (delta brief only; no --resume-last)

hard time limit (watchdog): add --timeout 2h (default: off; implementation runs routinely need 1-2h)

see all options: node .../relay.mjs --help


The helper writes artifacts to a temp dir, so the repo under review stays clean. It **never commits**
— see step 5. Mechanics, flags, and the `result.json` shape:
[references/dispatch-and-poll.md](references/dispatch-and-poll.md).


helper 会将 artifacts 写入临时目录，因此待审查的仓库保持干净。它**从不提交代码**——参见步骤 5。机制、flags 和 `result.json` 结构：
[references/dispatch-and-poll.md](references/dispatch-and-poll.md)。

3. Wait for completion

3. 等待完成

The helper blocks until Muse finishes, so back it with whatever your orchestrator offers and resume when it returns:

Claude Code: run the Bash call with
```
run_in_background: true
```
; you are notified on completion.
Plain shell / other agents: run it in the foreground for short tasks, or background it and poll the result file. The run is done when
```
result.json
```
exists with a
```
status
```
. (A pre-run usage error — bad args or an empty brief — instead exits with code 2 and writes no result file, so check the exit code too. A missing
```
muse
```
binary exits 127 but does write a
```
result.json
```
with status
```
muse_unavailable
```
.
```
--resume-last
```
is a usage error: headless resume is
```
--session <id>
```
.)

Do not trust progress trackers over reality: a run is finished when

result.json

is written and the process has exited. Muse can exit 0 without a

run.terminal.completed

event; the relay reports that as

failed

, not

completed

. The implementer's full report is the

finalMessage

field in

result.json

(also printed in full on stdout between the report markers).

helper 会阻塞直到 Muse 完成，因此应利用 orchestrator 提供的能力将其放到后台执行，并在返回时继续后续流程：

Claude Code： 使用
```
run_in_background: true
```
运行 Bash 调用；完成后你会收到通知。
普通 shell／其他 Agents： 短任务可在前台运行，也可以放到后台并轮询结果文件。当
```
result.json
```
存在且包含
```
status
```
时，run 即已完成。（如果运行前发生用法错误，例如参数无效或 brief 为空，则会以代码 2 退出且不会写入结果文件，因此也要检查退出码。缺少
```
muse
```
binary 时会以 127 退出，但会写入包含
```
muse_unavailable
```
状态的
```
result.json
```
。
```
--resume-last
```
属于用法错误：无头恢复应使用
```
--session <id>
```
。）

不要相信进度跟踪器而忽视实际状态：当

result.json

已写入且进程已退出时，run 才算完成。Muse 可能在没有

run.terminal.completed

事件的情况下以 0 退出；relay 会将其报告为

failed

，而不是

completed

。implementer 的完整报告位于

result.json

的

finalMessage

字段中（也会在 stdout 的报告标记之间完整打印）。

4. Review — do not trust the self-report

4. 审查——不要相信自我报告

Muse's

result.json

includes its own final message and any gate claims. Re-verify, don't accept:

Re-run the project's gates yourself (the test/lint/build commands from step 1). Never take "gates passed" on faith.
Read the diff against the brief: did Muse do what was asked, nothing more (scope creep) and nothing less?
```
touchedFiles
```
in the result is your starting point.
Run the relevant guard skills on the diff if you have them installed (clean-code-guard, test-guard, etc. from
```
guard-skills
```
) — this skill produces the work; those skills judge it.
For schema/migration changes, round-trip them; for removals, grep for dangling references.

Full checklist: references/review-and-land.md.

Muse 的

result.json

包含它自己的最终消息和各项门禁声明。请重新验证，不要直接接受：

亲自重新运行项目的门禁（步骤 1 中的测试、lint 和构建命令）。绝不要盲目相信“门禁已通过”。
对照 brief 阅读 diff：Muse 是否完成了要求的内容，没有超出范围（scope creep），也没有遗漏任何内容？result 中的
```
touchedFiles
```
是你的起点。
如果已安装相关 guard skills，请对 diff 运行这些 skills（来自
```
guard-skills
```
的 clean-code-guard、test-guard 等）——此 skill 负责产出工作，而这些 skills 负责评判工作。
对于 schema 或 migration 变更，执行 round-trip 验证；对于删除操作，使用 grep 搜索悬空引用。

完整检查清单：references/review-and-land.md。

5. Land it

5. 落地

The implementer edits the working tree; the orchestrator commits. Committing should be the act of the party that verified the work. Only after the gates pass and the diff holds:

Commit the verified work yourself, with a clear message.
If it needs changes, send a delta brief with
```
--session <id>
```
(don't restate the whole task) and review again. There is no headless
```
--resume-last
```
.

implementer 会修改工作树；由 orchestrator 提交。 提交应当由实际验证过工作的一方完成。仅在门禁通过且 diff 经得起审查后执行：

使用清晰的消息，亲自提交已验证的工作。
如果还需要修改，请使用
```
--session <id>
```
发送 delta brief（不要重述整个任务），然后再次审查。无头模式不支持
```
--resume-last
```
。

Authorization model

授权模型

Delegation is something the human opts into. Once they have ("run this queue", "proceed"), committing verified, gate-passing work is the agreed contract — that is the whole point. Two limits on that mandate: surface, don't absorb (report Muse's design decisions, defensible-but-unasked turns, and non-blocking nitpicks rather than silently keeping them) and stop for scope changes (if correct completion needs going beyond the brief, ask — don't expand the mandate yourself). The full treatment is in references/review-and-land.md.

委派需要由人员主动选择。一旦人员选择“运行这个队列”或“继续”，提交经过验证且通过门禁的工作便成为双方约定的 contract——这正是该流程的核心所在。这项授权存在两项限制：如实披露，不要默默接受（报告 Muse 的设计决策、合理但未被要求的改动，以及不阻塞流程的小问题，而不是静默保留这些内容），以及遇到范围变更时停止（如果正确完成任务需要超出 brief 的范围，应先询问，不要自行扩大授权）。完整说明见 references/review-and-land.md。

References

参考资料

references/writing-the-brief.md — how to write a brief Muse can execute blind: structure, XML blocks, the report contract, embedding the real gate commands.
references/dispatch-and-poll.md —
```
relay.mjs
```
flags, the
```
result.json
```
contract, backgrounding per orchestrator, and recovery when a run misbehaves.
references/review-and-land.md — the review checklist, the commit boundary, and the rework cycle via
```
--session
```
.
references/multi-task-queues.md — running a sequential queue: carrying constraints forward, progress tracking, and the end-of-run coherence check.

references/writing-the-brief.md——如何编写一份 Muse 能在缺少共享上下文时执行的 brief：结构、XML blocks、报告约定，以及嵌入真实门禁命令。
references/dispatch-and-poll.md——
```
relay.mjs
```
flags、
```
result.json
```
约定、针对不同 orchestrator 的后台运行方式，以及 run 异常时的恢复方法。
references/review-and-land.md——审查清单、提交边界，以及通过
```
--session
```
进行返工修改的循环。
references/multi-task-queues.md——如何运行顺序任务队列：延续约束、跟踪进度，以及执行结束前的一致性检查。