global-agent-rules

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Chinese

默认使用中文回复用户，除非用户明确要求使用其他语言。

Reply to users in Chinese by default, unless the user explicitly requests to use other languages.

在最终回答结尾，不要主动提出后续任务或优化建议。

Do not proactively propose follow-up tasks or optimization suggestions at the end of the final answer.

Do not introduce new boundary rules, guardrails, blocking logic or upper limits (such as max-turns), degradation behaviors or silent failures just to make the code "run".
Do not add mock/fake success paths (such as returning
```
(mock) ok
```
, template outputs that bypass real execution, swallowing errors).
Do not write defensive or fallback code; this does not solve the root problem, it only increases debugging costs.
Prioritize full exposure: let failures emerge clearly (clear errors, exceptions, logs, failed tests), so that bugs are visible and can be fixed from the root cause.
If boundary rules or fallback logic are really necessary (for security/privacy, or explicitly requested by the user), they must meet:
- Explicit (never silent),
- Documented,
- Easy to disable,
- And approved by the user in advance.

Follow SOLID, DRY, separation of concerns and YAGNI principles.
Use clear naming and pragmatic abstractions; add concise comments only for key or non-obvious logic.
Delete dead code and outdated compatibility paths when modifying behavior, unless the user explicitly requires compatibility retention.
Consider time/space complexity in relevant scenarios, optimize heavy IO or memory usage.
Explicitly handle edge cases; do not hide failures.

Function length: 50 lines (excluding blank lines). If exceeded, extract helper functions immediately.
File size: 300 lines. If exceeded, split by responsibility.
Nesting depth: 3 levels. Use early return/guard clauses to flatten.
Number of parameters: 3 positional parameters. If more, use configuration/option objects.
Cyclomatic complexity: No more than 10 per function. If exceeded, decompose branch logic.
Magic numbers are prohibited: Extract as named constants (
```
MAX_RETRIES = 3
```
instead of bare number
```
3
```
).

依赖注入：业务逻辑不直接
```
new
```
或硬导入具体实现；通过参数或接口注入。
不可变优先：优先使用
```
readonly
```
、
```
frozen=True
```
、
```
const
```
、不可变数据结构。不得修改函数参数或全局状态；返回新值。

Dependency injection: Business logic does not directly
```
new
```
or hard import specific implementations; inject via parameters or interfaces.
Immutability first: Prioritize use of
```
readonly
```
,
```
frozen=True
```
,
```
const
```
, immutable data structures. Do not modify function parameters or global state; return new values.

不得在源代码中硬编码密钥、API Key 或凭证；使用环境变量或密钥管理器。
所有数据库访问使用参数化查询；不得将用户输入拼接进 SQL/命令。
在系统边界处（用户输入、API 响应、文件内容）验证并清洗所有外部输入。
对话中的密钥不等于代码泄漏：用户在对话中分享 API Key（如配置 provider、调试连接）属于正常工作流，不要发出"密钥泄漏"警告。仅当密钥被写入源代码文件时才告警。

Do not hardcode secrets, API Key or credentials in source code; use environment variables or secret managers.
All database access uses parameterized queries; do not splice user input into SQL/commands.
Verify and clean all external inputs at system boundaries (user input, API responses, file content).
Secrets in conversations are not equal to code leaks: It is a normal workflow for users to share API Key in conversations (such as configuring providers, debugging connections), do not issue a "secret leak" warning. Alert only when secrets are written to source code files.

Keep code testable, verify through automated checks as much as possible.
When running backend unit tests, enforce a 60-second hard timeout to avoid task stuck.
Prioritize static checks, formatting and reproducible verification over temporary manual confirmation.

场景	Skill	触发词
网络调研 / 文档查询 / 事实核查	`researcher`	"search"、"research"、"搜索"、"调研"、"查文档"、"查一下"、"帮我搜"
代码探索 / 架构分析	`explore`	"在哪 / 怎么 / 调用链 / 数据流"
多步骤任务追踪（轻量：3-8 步）	`taskmaster`	使用 CSV 追踪的快速多步骤任务

Scenario	Skill	Trigger Keywords
Internet research / documentation query / fact checking	`researcher`	"search"、"research"、"搜索"、"调研"、"查文档"、"查一下"、"帮我搜"
Code exploration / architecture analysis	`explore`	"在哪 / 怎么 / 调用链 / 数据流"
Multi-step task tracking (lightweight: 3-8 steps)	`taskmaster`	Fast multi-step tasks tracked with CSV