QwenLM / QwenLM/qwen-code

QWEN.md 多 agent 禁令被系统默认 Explore 指引覆盖

Open
#7,679 3 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

category/core priority/P2 scope/core type/bug
Dominant language
TypeScript
Stars
28k
Forks
3.1k
Avg merge
1d 2h
Merged PRs (30d)
714

Description

What happened?

这是主代理违反了 QWEN.md 规则,不是规则没生效。原因有三层:

1. 规则确实存在且清晰

你的 ~/.qwen/QWEN.md 里写得很明确:

 │ 默认不使用多agent;仅当用户明确要求 subagent、委派或并行 Agent 时启用,不为简单任务增加额外用量。

而你的原话只是"再次分析一下付款失败相关逻辑是否还有需要补充优化的"——既没说 subagent,也没说委派/并行。按规则就不该 spawn。

2. 为什么还是 spawn 了——系统提示的默认倾向压过了 QWEN.md

Qwen Code 的内置系统提示里有几条很强的默认指引:

 │ For broader codebase exploration and deep research, use the 'agent' tool with subagent_type=Explore.
 │ Agent 工具说明里也鼓励"复杂多步任务自主处理"。

模型在权衡时,把"全仓库审计 PayFail 逻辑"判定成了"broad exploration → 该用 Explore",没有先回头检查 QWEN.md 的禁用约束。这是典型的系统默认行为盖过用户级 rule 的失败案例,不是规则被忽略,而是权重排序错了。

3. 还有一个加重因素

"再次分析"+"是否还有补充优化"是一个开放式、全仓库扫描型问题,正好命中 Explore 子代理的描述触发词("multi-step"、"broad codebase")。模型走捷径直接委派,而不是自己用 grep_search/read_file 串行做完。

---
What did you expect to happen?

遵守QWEN.md约定,不要被系统默认指引覆盖

Client information

Qwen Code: 0.20.1 (305b04910)
Runtime: Node.js v22.23.1 / npm 11.13.0
LSP: disabled
OS: darwin arm64 (24.6.0)
Auth: API Key - anthropic
Base URL: http://127.0.0.1:4000
Model: Peach-07-17-DogFooding-litellm
Fast Model: anthropic:Peach-07-17-DogFooding-litellm
Session ID: 07a97941-6081-42d4-b6df-f3c844c70f4c
Sandbox: no sandbox
Proxy: no proxy
Memory Usage: 178.7 MB

Login information

No response

Anything else we need to know?

No response

Contributor guide

Open the contributing guide

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Start by tracing how QWEN.md rules and the built-in system prompt are assembled, then inspect the agent tool's Explore guidance and reproduce the supplied "再次分析" scenario. Done means a user-level prohibition on multi-agent use is honored instead of being overridden by the default exploration guidance.

Written by the indexing model from the issue text.

Assessment

Tech stack
node.js, typescript
Domain
ai, cli
Issue type
Bug
Difficulty
4/5
Estimated time
3-5 days
Activity status
Quiet
Clarity
Mostly clear
Newbie friendliness
45/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.