qwen-agent

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

qwen-agent

Offload menial, self-contained tasks to a Qwen model running inside a headless Claude Code instance (

claude-9arm

). Keeps expensive Claude reasoning for work that needs it.

将繁琐、独立的任务交给在无头Claude Code实例（

claude-9arm

）中运行的Qwen模型。让昂贵的Claude推理能力用于真正需要它的工作。

The command

命令

claude-9arm

is a shell alias →

claude --model qwen3.6-35b-a3b

routed through the 9arm gateway. Run it headless with

-p

bash

claude-9arm -p "<self-contained task prompt>" --allowedTools Bash Read Edit Write Glob Grep

This is the default invocation. The flag list scopes which tools the subagent may use without a prompt, so it can finish a menial job unattended. Without it the subagent stalls waiting for approval on the first edit or command.
The alias bakes in
```
--allowedTools '*'
```
, which Claude Code silently ignores with a warning (
```
Wildcard tool name "*" is not supported
```
). That warning is expected and harmless — the
```
--allowedTools
```
you append is what takes effect.
For edit-only, lower-risk tasks you may instead use
```
--permission-mode acceptEdits
```
(auto-accepts file edits, but Bash still prompts — don't use it for verification/build/test runs).

claude-9arm

是一个Shell别名 → 通过9arm网关路由到

claude --model qwen3.6-35b-a3b

。使用

-p

参数以无头模式运行：

bash

claude-9arm -p "<self-contained task prompt>" --allowedTools Bash Read Edit Write Glob Grep

这是默认调用方式。标志列表限定了子代理无需提示即可使用的工具，这样它就能无人值守地完成繁琐任务。如果没有这个列表，子代理会在第一次编辑或命令时停滞等待批准。
该别名内置了
```
--allowedTools '*'
```
，但Claude Code会静默忽略并给出警告（
```
Wildcard tool name "*" is not supported
```
）。这个警告是预期且无害的——你附加的
```
--allowedTools
```
才会生效。
对于仅编辑、低风险的任务，你可以改用
```
--permission-mode acceptEdits
```
（自动接受文件编辑，但Bash仍会提示——不要用于验证/构建/测试运行）。

Writing the task prompt (most important step)

编写任务提示（最重要的步骤）

The qwen subagent has zero context from this conversation. A vague prompt is the #1 failure mode. Every prompt must be standalone:

Absolute paths for every input and output file (
```
/Users/tpatinya/proj/src/foo.ts
```
, not
```
foo.ts
```
).
Explicit inputs, outputs, and acceptance criteria — what to change, what "done" looks like.
No references to "the file we discussed", "above", or prior turns.
Treat qwen as a capable-but-literal junior: spell out the steps, keep scope tight.

Bad:

clean up the imports

Good:

In /Users/tpatinya/proj/src/api.ts, remove unused imports and sort the remaining import statements alphabetically. Do not change any other code. Confirm the file still parses.

qwen子代理完全没有本次对话的上下文。模糊的提示是导致失败的头号原因。每个提示必须是独立的：

每个输入和输出文件都使用绝对路径（
```
/Users/tpatinya/proj/src/foo.ts
```
，而不是
```
foo.ts
```
）。
明确输入、输出和验收标准——要修改什么，“完成”的标准是什么。
不要引用“我们讨论过的文件”、“上面”或之前的对话内容。
把qwen当作能力尚可但字面执行的初级开发者：详细列出步骤，保持范围紧凑。

错误示例：

clean up the imports

（清理导入）正确示例：

In /Users/tpatinya/proj/src/api.ts, remove unused imports and sort the remaining import statements alphabetically. Do not change any other code. Confirm the file still parses.

（在/Users/tpatinya/proj/src/api.ts中，移除未使用的导入并将剩余的导入语句按字母顺序排序。不要修改任何其他代码。确认文件仍能正常解析。）

Mind the context window (128k)

注意上下文窗口（128k）

Qwen runs with a 128k-token context window — much smaller than Claude's. The whole job (your prompt + every file it reads + its own reasoning and edits) has to fit inside it. Size each delegated task to the model, not just to "is it menial":

Estimate the footprint before delegating: roughly the bytes of files it must read + open + write, ÷ 4 ≈ tokens. If a single task would pull in large files or many files at once, it won't fit.
Break large jobs into independent chunks that each touch a bounded slice — e.g. one file (or a few small ones) per run, one directory per run, one log segment per run. Run the chunks as separate
```
claude-9arm
```
invocations (foreground, or background-parallel per the Return contract section).
Don't make it read what it doesn't need. Point it at the exact files/paths required; never tell it to "scan the repo" or read a whole large tree.
Watch for context-exhaustion symptoms when verifying: truncated edits, ignored later instructions, or a summary that omits files it was told to touch usually mean the task overflowed — split it smaller and retry.

When a job is inherently too big to slice cleanly (it needs whole-codebase context to do correctly), that's a sign it isn't a qwen task — keep it yourself.

Qwen运行在128k令牌的上下文窗口中——比Claude的小得多。整个任务（你的提示 + 它读取的每个文件 + 它自己的推理和编辑）必须容纳在这个窗口内。要根据模型的能力来调整每个委托任务的大小，而不仅仅是看“是否繁琐”：

委托前估算占用量：大致计算它必须读取、打开、写入的文件字节数 ÷4 ≈ 令牌数。如果单个任务会引入大文件或同时处理多个文件，可能无法容纳。
将大型任务拆分为独立的块，每个块只处理有限的部分——例如每次运行处理一个文件（或几个小文件）、一个目录、一段日志。将这些块作为单独的
```
claude-9arm
```
调用运行（前台运行，或根据“返回约定”部分后台并行运行）。
不要让它读取不需要的内容。指向它所需的确切文件/路径；永远不要让它“扫描仓库”或读取整个大型目录树。
验证时注意上下文耗尽的症状：截断的编辑、忽略后续指令，或总结中遗漏了它被告知要处理的文件，通常意味着任务超出了上下文窗口——将其拆分为更小的任务并重试。

如果某个任务本质上太大而无法清晰拆分（它需要整个代码库的上下文才能正确完成），这表明它不适合交给qwen处理——你自己来做。

Working directory

工作目录

The Bash tool's

cd

resets between calls and

cd &&

can trip permission prompts. Don't rely on cwd:

Put absolute paths in the prompt, or
Pass
```
--add-dir /abs/path
```
to grant the subagent access to a directory.

Bash工具的

cd

命令在每次调用后会重置，

cd &&

可能会触发权限提示。不要依赖当前工作目录：

在提示中使用绝对路径，或者
传递
```
--add-dir /abs/path
```
以授予子代理访问某个目录的权限。

Return contract

返回约定

Default (text): qwen's final message prints to stdout — read it directly.
Need to parse the result: add
```
--output-format json
```
and extract the
```
result
```
field.
Background / parallel (run several at once): redirect to a log and run with the Bash tool's
```
run_in_background: true
```
, then read the log when it finishes:
bash
```
claude-9arm -p "<task>" --allowedTools Bash Read Edit Write Glob Grep > /tmp/qwen-<label>.log 2>&1
```
Launch independent tasks as separate background runs; collect each log on completion. Use this when delegating 2+ unrelated menial jobs.

默认（文本）：qwen的最终消息会打印到标准输出——直接读取即可。
需要解析结果：添加
```
--output-format json
```
并提取
```
result
```
字段。
后台/并行（同时运行多个任务）：重定向到日志文件并使用Bash工具的
```
run_in_background: true
```
运行，完成后读取日志：
bash
```
claude-9arm -p "<task>" --allowedTools Bash Read Edit Write Glob Grep > /tmp/qwen-<label>.log 2>&1
```
将独立任务作为单独的后台运行启动；完成后收集每个日志。当委托2个或更多不相关的繁琐任务时使用此方法。

Workflow checklist

工作流程检查清单

Confirm the task is menial and low-risk (see description). If it needs design judgment or this chat's context, do it yourself — don't delegate.
Check it fits qwen's 128k context window — estimate the file footprint and split large jobs into bounded per-file/per-dir chunks (see "Mind the context window").
Write a fully self-contained prompt with absolute paths and acceptance criteria.

Run

claude-9arm -p "..." --allowedTools Bash Read Edit Write Glob Grep

(foreground), or background-redirect for parallel jobs.

Verify the output yourself — qwen is cheaper and less reliable. Check the file/result actually meets the acceptance criteria before reporting success.

确认任务是繁琐且低风险的（参见描述）。如果需要设计判断或本次聊天的上下文，自己处理——不要委托。
检查它是否适合qwen的128k上下文窗口——估算文件占用量并将大型任务拆分为有限的按文件/按目录的块（参见“注意上下文窗口”）。
编写一个完全独立的提示，包含绝对路径和验收标准。
运行
```
claude-9arm -p "..." --allowedTools Bash Read Edit Write Glob Grep
```
（前台运行），或重定向到后台以并行运行任务。
自行验证输出——qwen更便宜但可靠性更低。在报告成功之前，检查文件/结果是否确实符合验收标准。

One-time setup (optional, removes repeated prompts)

一次性设置（可选，消除重复提示）

To stop per-call permission prompts on delegated runs, add a Bash allow rule for the command (via the

update-config

skill, or by editing settings):

json

{ "permissions": { "allow": ["Bash(claude-9arm:*)"] } }

要在委托运行时取消每次调用的权限提示，为该命令添加一个Bash允许规则（通过

update-config

技能，或编辑设置）：

json

{ "permissions": { "allow": ["Bash(claude-9arm:*)"] } }

When NOT to delegate

不要委托的情况

Architecture/design, debugging that needs reasoning, security-sensitive changes, anything requiring this conversation's context, or tasks where a wrong cheap-model edit is costly to catch. When in doubt, keep it.

架构/设计、需要推理的调试、安全敏感的更改、任何需要本次对话上下文的内容，或者错误的廉价模型编辑会导致高昂修复成本的任务。如有疑问，自己处理。