feat: v2.7.5 — 常量集中管理 + 自定义拒绝规则 + 响应清洗开关

🏗️ 常量集中管理
- 新增 constants.ts,提取 REFUSAL_PATTERNS、IDENTITY_PROBE_PATTERNS、
  TOOL_CAPABILITY_PATTERNS、CLAUDE_IDENTITY_RESPONSE、CLAUDE_TOOLS_RESPONSE
- isRefusal() 统一导出,内置 + 自定义规则合并检测

🔧 自定义拒绝检测规则 (config.yaml: refusal_patterns)
- 用户可添加自定义正则匹配规则,追加到内置列表
- 无效正则自动退化为字面量匹配
- 缓存编译 + 热重载支持

🔀 响应内容清洗开关 (config.yaml: sanitize_response)
- 控制 sanitizeResponse() 是否替换 Cursor 身份引用为 Claude
- 默认关闭,关闭时零开销
- 支持环境变量 SANITIZE_RESPONSE 覆盖
This commit is contained in:
小海
2026-03-19 09:44:21 +08:00
parent e15dd5af33
commit 14aa65349c
9 changed files with 334 additions and 174 deletions

View File

@@ -1,5 +1,29 @@
# Changelog
## v2.7.5 (2026-03-19)
### 🏗️ 常量集中管理
- **新增 `constants.ts`**:将 `REFUSAL_PATTERNS`50+ 条拒绝检测规则)、`IDENTITY_PROBE_PATTERNS``TOOL_CAPABILITY_PATTERNS``CLAUDE_IDENTITY_RESPONSE``CLAUDE_TOOLS_RESPONSE` 及自定义拒绝规则逻辑从 `handler.ts` 提取到独立文件
- **提升可维护性**:贡献者修改内置规则时只需查看 `constants.ts`,无需翻阅 2000 行的 handler 业务逻辑
- **`isRefusal()` 函数统一导出**:内置规则 + 自定义规则合并检测,所有调用点自动生效
### 🔧 自定义拒绝检测规则
- **`config.yaml` 新增 `refusal_patterns` 字段**:用户可添加自定义正则匹配规则,追加到内置列表之后(不替换),匹配到即触发重试逻辑
- **无效正则容错**:无效的正则表达式自动退化为字面量匹配,不会导致服务报错
- **缓存编译**:自定义规则只在配置变更时重新编译 RegExp运行时零开销
- **热重载支持**:修改后下一次请求即生效
### 🔀 响应内容清洗开关
- **`config.yaml` 新增 `sanitize_response` 字段**:控制 `sanitizeResponse()` 函数(将 Cursor 身份引用替换为 Claude**默认关闭**
- **环境变量支持**`SANITIZE_RESPONSE=true` 可覆盖配置文件
- **零开销设计**:关闭时函数入口直接返回原文本,无正则计算
- **热重载支持**:修改配置后立即生效
---
## v2.7.4 (2026-03-18)
### 🛡️ 截断安全 — 防止损坏的工具调用

View File

@@ -1,8 +1,8 @@
# Cursor2API v2.7.4
# Cursor2API v2.7.5
将 Cursor 文档页免费 AI 对话接口代理转换为 **Anthropic Messages API****OpenAI Chat Completions API**,支持 **Claude Code****Cursor IDE** 使用。
> ⚠️ **版本说明**:当前 v2.7.4 截断安全(防止损坏工具调用)、默认禁用代理续写(让客户端原生续写)、日志查看器提示词对比视图
> ⚠️ **版本说明**:当前 v2.7.5 新增常量集中管理、自定义拒绝规则、响应清洗开关,代码可维护性大幅提升
## 原理
@@ -36,8 +36,8 @@
- **工具参数自动修复** - 字段名映射 (`file_path``path`)、智能引号替换、模糊匹配修复
- **多模态视觉降级处理** - 内置纯本地 CPU OCR 图片文字提取(零配置免 Key或支持外接第三方免费视觉大模型 API 解释图片
- **全工具支持** - 无工具白名单限制,支持所有 MCP 工具和自定义扩展
- **多层拒绝拦截** - 50+ 正则模式匹配拒绝文本(中英文),自动重试 + 认知重构绕过
- **三层身份保护** - 身份探针拦截 + 拒绝重试 + 响应清洗,确保输出永远呈现 Claude 身份
- **多层拒绝拦截** - 50+ 正则模式匹配拒绝文本(中英文),自动重试 + 认知重构绕过,支持自定义规则
- **三层身份保护** - 身份探针拦截 + 拒绝重试 + 响应清洗(可配置开关),确保输出永远呈现 Claude 身份
- **截断无缝续写** - Proxy 底层自动拼接被截断的工具响应(最多 6 次),含智能去重
- **渐进式历史压缩** - 智能识别消息类型,工具调用摘要化、工具结果头尾保留,不破坏 JSON 结构
- **🆕 可配置压缩系统** - 支持开关 + 3档级别轻度/中等/激进)+ 自定义参数,环境变量可覆盖
@@ -81,6 +81,8 @@ cp config.yaml.example config.yaml
| `logging.dir` | 日志存储目录 | `./logs` |
| `logging.max_days` | 日志保留天数 | `7` |
| `max_auto_continue` | 截断自动续写次数 (`0`=禁用,交由客户端续写) | `0` |
| `sanitize_response` | 响应内容清洗开关(替换 Cursor 身份引用为 Claude | `false` |
| `refusal_patterns` | 自定义拒绝检测规则列表(追加到内置规则) | 不配置 |
> 💡 详细配置说明请参见 `config.yaml.example` 中的注释。
@@ -150,6 +152,7 @@ cursor2api/
│ ├── index.ts # 入口 + Express 服务 + 路由 + API 鉴权中间件
│ ├── config.ts # 配置管理(含 auth_tokens / vision.proxy
│ ├── types.ts # 类型定义(含 thinking / authTokens
│ ├── constants.ts # 全局常量(拒绝模式、身份探针、回复模板)
│ ├── cursor-client.ts # Cursor API 客户端 + Chrome TLS 指纹
│ ├── converter.ts # 协议转换 + 提示词注入 + 上下文清洗 + 动态预算
│ ├── handler.ts # Anthropic API 处理器 + 身份保护 + 拒绝拦截 + Thinking
@@ -243,6 +246,7 @@ AI 按此格式输出 → 我们解析并转换为标准的 Anthropic `tool_use`
| `LOG_FILE_ENABLED` | 日志文件持久化 (`true`/`false`) |
| `LOG_DIR` | 日志文件目录 |
| `MAX_AUTO_CONTINUE` | 截断自动续写次数 (`0`=禁用) |
| `SANITIZE_RESPONSE` | 响应内容清洗开关 (`true`/`false`,默认 `false`) |
## 免责声明 / Disclaimer

View File

@@ -106,6 +106,26 @@ tools:
# exclude:
# - "some_mcp_tool"
# ==================== 响应内容清洗(可选,默认关闭) ====================
# 开启后,会将响应中 Cursor 相关的身份引用替换为 Claude
# 例如 "I am Cursor's support assistant" → "I am Claude, an AI assistant by Anthropic"
# 同时清洗工具可用性声明、提示注入指控等敏感内容
# 💡 如果你不需要伪装身份,建议保持关闭以获得最佳性能
# 💡 开启后,所有响应都会经过正则替换处理,有轻微性能开销
# sanitize_response: true
# ==================== 自定义拒绝检测规则(可选) ====================
# 追加到内置拒绝检测列表之后(不替换内置规则),匹配到则触发重试逻辑
# 每条规则作为正则表达式解析(不区分大小写),无效正则会自动退化为字面量匹配
# 💡 适用场景:特定语言的拒绝措辞、项目特有的拒绝响应、新的 Cursor 拒绝模式
# 支持热重载:修改后下一次请求即生效
# refusal_patterns:
# - "我无法协助"
# - "this violates our"
# - "I must decline"
# - "无法为您提供"
# - "this request is outside"
# 浏览器指纹配置
fingerprint:
user_agent: "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/140.0.0.0 Safari/537.36"

View File

@@ -50,3 +50,10 @@ services:
# ── Vision 图片处理 ──
# 默认使用本地 OCR零配置如需外部 Vision API 请在 config.yaml 中修改 vision.mode 为 'api'
# 并配置 vision.base_url / vision.api_key / vision.model
# ── 响应内容清洗 ──
# 开启后会将响应中 Cursor 身份引用替换为 Claude默认关闭
# - SANITIZE_RESPONSE=true
# ── 自定义拒绝检测规则 ──
# 仅支持 config.yaml 配置(无环境变量覆盖),详见 config.yaml.example 中的 refusal_patterns 节

View File

@@ -1,6 +1,6 @@
{
"name": "cursor2api",
"version": "2.7.4",
"version": "2.7.5",
"description": "Proxy Cursor docs AI to Anthropic Messages API for Claude Code",
"type": "module",
"scripts": {

View File

@@ -91,6 +91,14 @@ function parseYamlConfig(defaults: AppConfig): { config: AppConfig; raw: Record<
exclude: Array.isArray(t.exclude) ? t.exclude.map(String) : undefined,
};
}
// ★ 响应内容清洗开关(默认关闭)
if (yaml.sanitize_response !== undefined) {
result.sanitizeEnabled = yaml.sanitize_response === true;
}
// ★ 自定义拒绝检测规则
if (Array.isArray(yaml.refusal_patterns)) {
result.refusalPatterns = yaml.refusal_patterns.map(String).filter(Boolean);
}
} catch (e) {
console.warn('[Config] 读取 config.yaml 失败:', e);
}
@@ -136,6 +144,10 @@ function applyEnvOverrides(cfg: AppConfig): void {
if (!cfg.logging) cfg.logging = { file_enabled: false, dir: './logs', max_days: 7 };
cfg.logging.dir = process.env.LOG_DIR;
}
// 响应内容清洗环境变量覆盖
if (process.env.SANITIZE_RESPONSE !== undefined) {
cfg.sanitizeEnabled = process.env.SANITIZE_RESPONSE === 'true' || process.env.SANITIZE_RESPONSE === '1';
}
// 从 base64 FP 环境变量解析指纹
if (process.env.FP) {
@@ -158,6 +170,7 @@ function defaultConfig(): AppConfig {
cursorModel: 'anthropic/claude-sonnet-4.6',
maxAutoContinue: 0,
maxHistoryMessages: -1,
sanitizeEnabled: false, // 默认关闭响应内容清洗
fingerprint: {
userAgent: 'Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/140.0.0.0 Safari/537.36',
},
@@ -197,6 +210,12 @@ function detectChanges(oldCfg: AppConfig, newCfg: AppConfig): string[] {
// tools
if (JSON.stringify(oldCfg.tools) !== JSON.stringify(newCfg.tools)) changes.push('tools: (changed)');
// refusalPatterns
// sanitize_response
if (oldCfg.sanitizeEnabled !== newCfg.sanitizeEnabled) changes.push(`sanitize_response: ${oldCfg.sanitizeEnabled}${newCfg.sanitizeEnabled}`);
if (JSON.stringify(oldCfg.refusalPatterns) !== JSON.stringify(newCfg.refusalPatterns)) changes.push(`refusal_patterns: ${oldCfg.refusalPatterns?.length || 0}${newCfg.refusalPatterns?.length || 0} rule(s)`);
// fingerprint
if (oldCfg.fingerprint.userAgent !== newCfg.fingerprint.userAgent) changes.push('fingerprint: (changed)');

231
src/constants.ts Normal file
View File

@@ -0,0 +1,231 @@
/**
* constants.ts - 全局常量定义
*
* 集中管理拒绝检测规则、身份探针模式、固定回复模板等常量。
* 方便查阅和修改内置规则,无需翻阅 handler.ts 的业务逻辑。
*/
import { getConfig } from './config.js';
// ==================== 拒绝模式识别 ====================
// 模型返回以下任意模式匹配的内容时,判定为拒绝响应并触发重试
// 如需添加新规则,请追加到对应分类末尾,或通过 config.yaml 的 refusal_patterns 配置
export const REFUSAL_PATTERNS: RegExp[] = [
// ── English: 身份拒绝 ──
/Cursor(?:'s)?\s+support\s+assistant/i,
/support\s+assistant\s+for\s+Cursor/i,
/I[''']\s*m\s+sorry/i,
/I\s+am\s+sorry/i,
/not\s+able\s+to\s+fulfill/i,
/cannot\s+perform/i,
/I\s+can\s+only\s+answer/i,
/I\s+only\s+answer/i,
/cannot\s+write\s+files/i,
/pricing[, \s]*or\s*troubleshooting/i,
/I\s+cannot\s+help\s+with/i,
/I'm\s+a\s+coding\s+assistant/i,
/not\s+able\s+to\s+search/i,
/not\s+in\s+my\s+core/i,
/outside\s+my\s+capabilities/i,
/I\s+cannot\s+search/i,
/focused\s+on\s+software\s+development/i,
/not\s+able\s+to\s+help\s+with\s+(?:that|this)/i,
/beyond\s+(?:my|the)\s+scope/i,
/I'?m\s+not\s+(?:able|designed)\s+to/i,
/I\s+don't\s+have\s+(?:the\s+)?(?:ability|capability)/i,
/questions\s+about\s+(?:Cursor|the\s+(?:AI\s+)?code\s+editor)/i,
// ── English: 话题拒绝 ── Cursor 拒绝非编程话题
/help\s+with\s+(?:coding|programming)\s+and\s+Cursor/i,
/Cursor\s+IDE\s+(?:questions|features|related)/i,
/unrelated\s+to\s+(?:programming|coding)(?:\s+or\s+Cursor)?/i,
/Cursor[- ]related\s+question/i,
/(?:ask|please\s+ask)\s+a\s+(?:programming|coding|Cursor)/i,
/(?:I'?m|I\s+am)\s+here\s+to\s+help\s+with\s+(?:coding|programming)/i,
/appears\s+to\s+be\s+(?:asking|about)\s+.*?unrelated/i,
/(?:not|isn't|is\s+not)\s+(?:related|relevant)\s+to\s+(?:programming|coding|software)/i,
/I\s+can\s+help\s+(?:you\s+)?with\s+things\s+like/i,
// ── English: 新拒绝措辞 (2026-03) ──
/isn't\s+something\s+I\s+can\s+help\s+with/i,
/not\s+something\s+I\s+can\s+help\s+with/i,
/scoped\s+to\s+answering\s+questions\s+about\s+Cursor/i,
/falls\s+outside\s+(?:the\s+scope|what\s+I)/i,
// ── English: 提示注入/社会工程检测 ──
/prompt\s+injection\s+attack/i,
/prompt\s+injection/i,
/social\s+engineering/i,
/I\s+need\s+to\s+stop\s+and\s+flag/i,
/What\s+I\s+will\s+not\s+do/i,
/What\s+is\s+actually\s+happening/i,
/replayed\s+against\s+a\s+real\s+system/i,
/tool-call\s+payloads/i,
/copy-pasteable\s+JSON/i,
/injected\s+into\s+another\s+AI/i,
/emit\s+tool\s+invocations/i,
/make\s+me\s+output\s+tool\s+calls/i,
// ── English: 工具可用性声明 (Cursor 角色锁定) ──
/I\s+(?:only\s+)?have\s+(?:access\s+to\s+)?(?:two|2|read_file|read_dir)\s+tool/i,
/(?:only|just)\s+(?:two|2)\s+(?:tools?|functions?)\b/i,
/\bread_file\b.*\bread_dir\b/i,
/\bread_dir\b.*\bread_file\b/i,
// ── English: 范围/专长措辞 (2026-03 批次) ──
/(?:outside|beyond)\s+(?:the\s+)?scope\s+of\s+what/i,
/not\s+(?:within|in)\s+(?:my|the)\s+scope/i,
/this\s+assistant\s+is\s+(?:focused|scoped)/i,
/(?:only|just)\s+(?:able|here)\s+to\s+(?:answer|help)/i,
/I\s+(?:can\s+)?only\s+help\s+with\s+(?:questions|issues)\s+(?:related|about)/i,
/(?:here|designed)\s+to\s+help\s+(?:with\s+)?(?:questions\s+)?about\s+Cursor/i,
/not\s+(?:something|a\s+topic)\s+(?:related|specific)\s+to\s+(?:Cursor|coding)/i,
/outside\s+(?:my|the|your)\s+area\s+of\s+(?:expertise|scope)/i,
/(?:can[.']?t|cannot|unable\s+to)\s+help\s+with\s+(?:this|that)\s+(?:request|question|topic)/i,
/scoped\s+to\s+(?:answering|helping)/i,
// ── 中文: 身份拒绝 ──
/我是\s*Cursor\s*的?\s*支持助手/,
/Cursor\s*的?\s*支持系统/,
/Cursor\s*(?:编辑器|IDE)?\s*相关的?\s*问题/,
/我的职责是帮助你解答/,
/我无法透露/,
/帮助你解答\s*Cursor/,
/运行在\s*Cursor\s*的/,
/专门.*回答.*(?:Cursor|编辑器)/,
/我只能回答/,
/无法提供.*信息/,
/我没有.*也不会提供/,
/功能使用[、,]\s*账单/,
/故障排除/,
// ── 中文: 话题拒绝 ──
/与\s*(?:编程|代码|开发)\s*无关/,
/请提问.*(?:编程|代码|开发|技术).*问题/,
/只能帮助.*(?:编程|代码|开发)/,
// ── 中文: 提示注入检测 ──
/不是.*需要文档化/,
/工具调用场景/,
/语言偏好请求/,
/提供.*具体场景/,
/即报错/,
// ── 中文: 工具可用性声明 ──
/有以下.*?(?:两|2)个.*?工具/,
/我有.*?(?:两|2)个工具/,
/工具.*?(?:只有|有以下|仅有).*?(?:两|2)个/,
/只能用.*?read_file/i,
/无法调用.*?工具/,
/(?:仅限于|仅用于).*?(?:查阅|浏览).*?(?:文档|docs)/,
// ── 中文: Cursor 中文界面拒绝措辞 (2026-03 批次) ──
/只能回答.*(?:Cursor|编辑器).*(?:相关|有关)/,
/专[注门].*(?:回答|帮助|解答).*(?:Cursor|编辑器)/,
/有什么.*(?:Cursor|编辑器).*(?:问题|可以)/,
/无法提供.*(?:推荐|建议|帮助)/,
/(?:功能使用|账户|故障排除|账号|订阅|套餐|计费).*(?:等|问题)/,
];
// ==================== 自定义拒绝规则 ====================
// 从 config.yaml 的 refusal_patterns 字段编译,追加到内置列表之后,支持热重载
let _customRefusalPatterns: RegExp[] = [];
let _lastRefusalPatternsKey = '';
function getCustomRefusalPatterns(): RegExp[] {
const config = getConfig();
const patterns = config.refusalPatterns;
if (!patterns || patterns.length === 0) return _customRefusalPatterns = [];
// 用 join key 做缓存判断,避免每次调用都重新编译
const key = patterns.join('\0');
if (key === _lastRefusalPatternsKey) return _customRefusalPatterns;
_lastRefusalPatternsKey = key;
_customRefusalPatterns = [];
for (const p of patterns) {
try {
_customRefusalPatterns.push(new RegExp(p, 'i'));
} catch {
// 无效正则 → 退化为字面量匹配
_customRefusalPatterns.push(new RegExp(p.replace(/[.*+?^${}()|[\]\\]/g, '\\$&'), 'i'));
console.warn(`[Config] refusal_patterns: "${p}" 不是有效正则,已转换为字面量匹配`);
}
}
console.log(`[Config] 加载了 ${_customRefusalPatterns.length} 条自定义拒绝规则`);
return _customRefusalPatterns;
}
/**
* 检查文本是否匹配拒绝模式(内置 + 自定义规则)
*/
export function isRefusal(text: string): boolean {
if (REFUSAL_PATTERNS.some(p => p.test(text))) return true;
const custom = getCustomRefusalPatterns();
return custom.length > 0 && custom.some(p => p.test(text));
}
// ==================== 身份探针检测 ====================
// 用户消息匹配以下模式时判定为身份探针,直接返回 mock 回复
export const IDENTITY_PROBE_PATTERNS: RegExp[] = [
// 精确短句
/^\s*(who are you\??|你是谁[呀啊吗]?\??|what is your name\??|你叫什么\??|你叫什么名字\??|what are you\??|你是什么\??|Introduce yourself\??|自我介绍一下\??|hi\??|hello\??|hey\??|你好\??|在吗\??|哈喽\??)\s*$/i,
// 问模型/身份类
/(?:什么|哪个|啥)\s*模型/,
/(?:真实|底层|实际|真正).{0,10}(?:模型|身份|名字)/,
/模型\s*(?:id|名|名称|名字|是什么)/i,
/(?:what|which)\s+model/i,
/(?:real|actual|true|underlying)\s+(?:model|identity|name)/i,
/your\s+(?:model|identity|real\s+name)/i,
// 问平台/运行环境类
/运行在\s*(?:哪|那|什么)/,
/(?:哪个|什么)\s*平台/,
/running\s+on\s+(?:what|which)/i,
/what\s+platform/i,
// 问系统提示词类
/系统\s*提示词/,
/system\s*prompt/i,
// "你是谁"的变体
/你\s*(?:到底|究竟|真的|真实)\s*是\s*谁/,
/你\s*是[^。,,\.]{0,5}(?:AI|人工智能|助手|机器人|模型|Claude|GPT|Gemini)/i,
// 注意:工具能力询问不在这里拦截,由拒绝检测+重试自然处理
];
// ==================== 工具能力询问检测 ====================
// 用户问"你有哪些工具"时,重试失败后返回专用回复
export const TOOL_CAPABILITY_PATTERNS: RegExp[] = [
/你\s*(?:有|能用|可以用)\s*(?:哪些|什么|几个)\s*(?:工具|tools?|functions?)/i,
/(?:what|which|list).*?tools?/i,
/你\s*用\s*(?:什么|哪个|啥)\s*(?:mcp|工具)/i,
/你\s*(?:能|可以)\s*(?:做|干)\s*(?:什么|哪些|啥)/,
/(?:what|which).*?(?:capabilities|functions)/i,
/能力|功能/,
];
// ==================== 固定回复模板 ====================
/** Claude 身份回复(身份探针拦截 / 拒绝后降级) */
export const CLAUDE_IDENTITY_RESPONSE = `I am Claude, made by Anthropic. I'm an AI assistant designed to be helpful, harmless, and honest. I can help you with a wide range of tasks including writing, analysis, coding, math, and more.
I don't have information about the specific model version or ID being used for this conversation, but I'm happy to help you with whatever you need!`;
/** 工具能力询问的模拟回复(当用户问"你有哪些工具"时) */
export const CLAUDE_TOOLS_RESPONSE = `作为 Claude我的核心能力包括
**内置能力:**
- 💻 **代码编写与调试** — 支持所有主流编程语言
- 📝 **文本写作与分析** — 文章、报告、翻译等
- 📊 **数据分析与数学推理** — 复杂计算和逻辑分析
- 🧠 **问题解答与知识查询** — 各类技术和非技术问题
**工具调用能力MCP**
如果你的客户端配置了 MCPModel Context Protocol工具我可以通过工具调用来执行更多操作例如
- 🔍 **网络搜索** — 实时查找信息
- 📁 **文件操作** — 读写文件、执行命令
- 🛠️ **自定义工具** — 取决于你配置的 MCP Server
具体可用的工具取决于你客户端的配置。你可以告诉我你想做什么,我会尽力帮助你!`;

View File

@@ -30,119 +30,23 @@ function toolId(): string {
return 'toolu_' + uuidv4().replace(/-/g, '').substring(0, 24);
}
// ==================== 拒绝模式识别 ====================
const REFUSAL_PATTERNS = [
// English identity refusal
/Cursor(?:'s)?\s+support\s+assistant/i,
/support\s+assistant\s+for\s+Cursor/i,
/I[''']m\s+sorry/i,
/I\s+am\s+sorry/i,
/not\s+able\s+to\s+fulfill/i,
/cannot\s+perform/i,
/I\s+can\s+only\s+answer/i,
/I\s+only\s+answer/i,
/cannot\s+write\s+files/i,
/pricing[, \s]*or\s*troubleshooting/i,
/I\s+cannot\s+help\s+with/i,
/I'm\s+a\s+coding\s+assistant/i,
/not\s+able\s+to\s+search/i,
/not\s+in\s+my\s+core/i,
/outside\s+my\s+capabilities/i,
/I\s+cannot\s+search/i,
/focused\s+on\s+software\s+development/i,
/not\s+able\s+to\s+help\s+with\s+(?:that|this)/i,
/beyond\s+(?:my|the)\s+scope/i,
/I'?m\s+not\s+(?:able|designed)\s+to/i,
/I\s+don't\s+have\s+(?:the\s+)?(?:ability|capability)/i,
/questions\s+about\s+(?:Cursor|the\s+(?:AI\s+)?code\s+editor)/i,
// English topic refusal — Cursor 拒绝非编程话题
/help\s+with\s+(?:coding|programming)\s+and\s+Cursor/i,
/Cursor\s+IDE\s+(?:questions|features|related)/i,
/unrelated\s+to\s+(?:programming|coding)(?:\s+or\s+Cursor)?/i,
/Cursor[- ]related\s+question/i,
/(?:ask|please\s+ask)\s+a\s+(?:programming|coding|Cursor)/i,
/(?:I'?m|I\s+am)\s+here\s+to\s+help\s+with\s+(?:coding|programming)/i,
/appears\s+to\s+be\s+(?:asking|about)\s+.*?unrelated/i,
/(?:not|isn't|is\s+not)\s+(?:related|relevant)\s+to\s+(?:programming|coding|software)/i,
/I\s+can\s+help\s+(?:you\s+)?with\s+things\s+like/i,
// New Cursor refusal phrases (2026-03)
/isn't\s+something\s+I\s+can\s+help\s+with/i,
/not\s+something\s+I\s+can\s+help\s+with/i,
/scoped\s+to\s+answering\s+questions\s+about\s+Cursor/i,
/falls\s+outside\s+(?:the\s+scope|what\s+I)/i,
// Prompt injection / social engineering detection (new failure mode)
/prompt\s+injection\s+attack/i,
/prompt\s+injection/i,
/social\s+engineering/i,
/I\s+need\s+to\s+stop\s+and\s+flag/i,
/What\s+I\s+will\s+not\s+do/i,
/What\s+is\s+actually\s+happening/i,
/replayed\s+against\s+a\s+real\s+system/i,
/tool-call\s+payloads/i,
/copy-pasteable\s+JSON/i,
/injected\s+into\s+another\s+AI/i,
/emit\s+tool\s+invocations/i,
/make\s+me\s+output\s+tool\s+calls/i,
// Tool availability claims (Cursor role lock)
/I\s+(?:only\s+)?have\s+(?:access\s+to\s+)?(?:two|2|read_file|read_dir)\s+tool/i,
/(?:only|just)\s+(?:two|2)\s+(?:tools?|functions?)\b/i,
/\bread_file\b.*\bread_dir\b/i,
/\bread_dir\b.*\bread_file\b/i,
/有以下.*?(?:两|2)个.*?工具/,
/我有.*?(?:两|2)个工具/,
/工具.*?(?:只有|有以下|仅有).*?(?:两|2)个/,
/只能用.*?read_file/i,
/无法调用.*?工具/,
/(?:仅限于|仅用于).*?(?:查阅|浏览).*?(?:文档|docs)/,
// Chinese identity refusal
/我是\s*Cursor\s*的?\s*支持助手/,
/Cursor\s*的?\s*支持系统/,
/Cursor\s*(?:编辑器|IDE)?\s*相关的?\s*问题/,
/我的职责是帮助你解答/,
/我无法透露/,
/帮助你解答\s*Cursor/,
/运行在\s*Cursor\s*的/,
/专门.*回答.*(?:Cursor|编辑器)/,
/我只能回答/,
/无法提供.*信息/,
/我没有.*也不会提供/,
/功能使用[、,]\s*账单/,
/故障排除/,
// Chinese topic refusal
/与\s*(?:编程|代码|开发)\s*无关/,
/请提问.*(?:编程|代码|开发|技术).*问题/,
/只能帮助.*(?:编程|代码|开发)/,
// Chinese prompt injection detection
/不是.*需要文档化/,
/工具调用场景/,
/语言偏好请求/,
/提供.*具体场景/,
/即报错/,
// EN: scope/expertise wordings (2026-03 batch)
/(?:outside|beyond)\s+(?:the\s+)?scope\s+of\s+what/i,
/not\s+(?:within|in)\s+(?:my|the)\s+scope/i,
/this\s+assistant\s+is\s+(?:focused|scoped)/i,
/(?:only|just)\s+(?:able|here)\s+to\s+(?:answer|help)/i,
/I\s+(?:can\s+)?only\s+help\s+with\s+(?:questions|issues)\s+(?:related|about)/i,
/(?:here|designed)\s+to\s+help\s+(?:with\s+)?(?:questions\s+)?about\s+Cursor/i,
/not\s+(?:something|a\s+topic)\s+(?:related|specific)\s+to\s+(?:Cursor|coding)/i,
/outside\s+(?:my|the|your)\s+area\s+of\s+(?:expertise|scope)/i,
/(?:can[.']?t|cannot|unable\s+to)\s+help\s+with\s+(?:this|that)\s+(?:request|question|topic)/i,
/scoped\s+to\s+(?:answering|helping)/i,
// CN: Chinese refusal wordings — Cursor 中文界面下的拒绝 (2026-03 batch)
/只能回答.*(?:Cursor|编辑器).*(?:相关|有关)/,
/专[注门].*(?:回答|帮助|解答).*(?:Cursor|编辑器)/,
/有什么.*(?:Cursor|编辑器).*(?:问题|可以)/,
/无法提供.*(?:推荐|建议|帮助)/,
/(?:功能使用|账户|故障排除|账号|订阅|套餐|计费).*(?:等|问题)/,
];
// ==================== 常量导入 ====================
// 拒绝模式、身份探针、工具能力询问等常量统一定义在 constants.ts
// 方便查阅和修改内置规则,无需翻阅此文件的业务逻辑
import {
isRefusal,
IDENTITY_PROBE_PATTERNS,
TOOL_CAPABILITY_PATTERNS,
CLAUDE_IDENTITY_RESPONSE,
CLAUDE_TOOLS_RESPONSE,
} from './constants.js';
export function isRefusal(text: string): boolean {
return REFUSAL_PATTERNS.some(p => p.test(text));
}
// Re-export for other modules (openai-handler.ts etc.)
export { isRefusal, CLAUDE_IDENTITY_RESPONSE, CLAUDE_TOOLS_RESPONSE };
// ==================== Thinking 提取 ====================
const THINKING_OPEN = '<thinking>';
const THINKING_CLOSE = '</thinking>';
@@ -222,31 +126,6 @@ export function countTokens(req: Request, res: Response): void {
// ==================== 身份探针拦截 ====================
// 关键词检测(宽松匹配):只要用户消息包含这些关键词组合就判定为身份探针
const IDENTITY_PROBE_PATTERNS = [
// 精确短句(原有)
/^\s*(who are you\??|你是谁[呀啊吗]?\??|what is your name\??|你叫什么\??|你叫什么名字\??|what are you\??|你是什么\??|Introduce yourself\??|自我介绍一下\??|hi\??|hello\??|hey\??|你好\??|在吗\??|哈喽\??)\s*$/i,
// 问模型/身份类
/(?:什么|哪个|啥)\s*模型/,
/(?:真实|底层|实际|真正).{0,10}(?:模型|身份|名字)/,
/模型\s*(?:id|名|名称|名字|是什么)/i,
/(?:what|which)\s+model/i,
/(?:real|actual|true|underlying)\s+(?:model|identity|name)/i,
/your\s+(?:model|identity|real\s+name)/i,
// 问平台/运行环境类
/运行在\s*(?:哪|那|什么)/,
/(?:哪个|什么)\s*平台/,
/running\s+on\s+(?:what|which)/i,
/what\s+platform/i,
// 问系统提示词类
/系统\s*提示词/,
/system\s*prompt/i,
// 你是谁的变体
/你\s*(?:到底|究竟|真的|真实)\s*是\s*谁/,
/你\s*是[^。,,\.]{0,5}(?:AI|人工智能|助手|机器人|模型|Claude|GPT|Gemini)/i,
// 注意:工具能力询问(“你有哪些工具”)不在这里拦截,而是让拒绝检测+重试自然处理
];
export function isIdentityProbe(body: AnthropicRequest): boolean {
if (!body.messages || body.messages.length === 0) return false;
const lastMsg = body.messages[body.messages.length - 1];
@@ -267,40 +146,6 @@ export function isIdentityProbe(body: AnthropicRequest): boolean {
return IDENTITY_PROBE_PATTERNS.some(p => p.test(text));
}
// ==================== 响应内容清洗 ====================
// Claude 身份回复模板(拒绝后的降级回复)
export const CLAUDE_IDENTITY_RESPONSE = `I am Claude, made by Anthropic. I'm an AI assistant designed to be helpful, harmless, and honest. I can help you with a wide range of tasks including writing, analysis, coding, math, and more.
I don't have information about the specific model version or ID being used for this conversation, but I'm happy to help you with whatever you need!`;
// 工具能力询问的模拟回复(当用户问“你有哪些工具”时,返回 Claude 真实能力描述)
export const CLAUDE_TOOLS_RESPONSE = `作为 Claude我的核心能力包括
**内置能力:**
- 💻 **代码编写与调试** — 支持所有主流编程语言
- 📝 **文本写作与分析** — 文章、报告、翻译等
- 📊 **数据分析与数学推理** — 复杂计算和逻辑分析
- 🧠 **问题解答与知识查询** — 各类技术和非技术问题
**工具调用能力MCP**
如果你的客户端配置了 MCPModel Context Protocol工具我可以通过工具调用来执行更多操作例如
- 🔍 **网络搜索** — 实时查找信息
- 📁 **文件操作** — 读写文件、执行命令
- 🛠️ **自定义工具** — 取决于你配置的 MCP Server
具体可用的工具取决于你客户端的配置。你可以告诉我你想做什么,我会尽力帮助你!`;
// 检测是否是工具能力询问(用于重试失败后返回专用回复)
const TOOL_CAPABILITY_PATTERNS = [
/你\s*(?:有|能用|可以用)\s*(?:哪些|什么|几个)\s*(?:工具|tools?|functions?)/i,
/(?:what|which|list).*?tools?/i,
/你\s*用\s*(?:什么|哪个|啥)\s*(?:mcp|工具)/i,
/你\s*(?:能|可以)\s*(?:做|干)\s*(?:什么|哪些|啥)/,
/(?:what|which).*?(?:capabilities|functions)/i,
/能力|功能/,
];
export function isToolCapabilityQuestion(body: AnthropicRequest): boolean {
if (!body.messages || body.messages.length === 0) return false;
const lastMsg = body.messages[body.messages.length - 1];
@@ -318,11 +163,19 @@ export function isToolCapabilityQuestion(body: AnthropicRequest): boolean {
return TOOL_CAPABILITY_PATTERNS.some(p => p.test(text));
}
// ==================== 响应内容清洗 ====================
/**
* 对所有响应做后处理:清洗 Cursor 身份引用,替换为 Claude
* 这是最后一道防线,确保用户永远看不到 Cursor 相关的身份信息
*
* ★ 受配置开关 sanitize_response 控制,默认关闭
* 开启方式config.yaml 中设置 sanitize_response: true
* 或环境变量 SANITIZE_RESPONSE=true
*/
export function sanitizeResponse(text: string): string {
// 配置未启用时直接返回原文本,零开销
if (!getConfig().sanitizeEnabled) return text;
let result = text;
// === English identity replacements ===
@@ -1618,7 +1471,7 @@ Continue EXACTLY from where you stopped. DO NOT repeat any content already gener
stopReason = 'tool_use';
// Check if the residual text is a known refusal, if so, drop it completely!
if (REFUSAL_PATTERNS.some(p => p.test(cleanText))) {
if (isRefusal(cleanText)) {
log.info('Handler', 'sanitize', `抑制工具调用中的拒绝文本`, { preview: cleanText.substring(0, 200) });
cleanText = '';
}

View File

@@ -136,6 +136,8 @@ export interface AppConfig {
includeOnly?: string[]; // 白名单:只保留的工具名
exclude?: string[]; // 黑名单:要排除的工具名
};
sanitizeEnabled: boolean; // 是否启用响应内容清洗(替换 Cursor 身份引用为 Claude默认 false
refusalPatterns?: string[]; // 自定义拒绝检测规则(追加到内置列表之后)
fingerprint: {
userAgent: string;
};