Skip to content

abcduyt1243-git/deepseek-cursor-proxy-turbo

BranchesTags
 
 

Repository files navigation

logo
DeepSeek Cursor Proxy Turbo

⚡ 极致提速版 · Performance-Optimized Fork

基于 yxlao/deepseek-cursor-proxy | 原项目 MIT License

🀄 中文版

一、为什么 Cursor 不能直接用 DeepSeek 的思考模式？

DeepSeek 官方的要求（Thinking Mode 文档）：

"Between two user messages, if the model performed a tool call, the intermediate assistant's reasoning_content must participate in the context concatenation and must be passed back to the API in all subsequent user interaction turns."

翻译：当模型执行了工具调用（tool call），后续所有请求都必须把之前的 reasoning_content 原封不动传回去。这是 DeepSeek 的硬性 API 契约——在思考模式下，工具调用的推理链必须保持连续。

Cursor 的实际行为（Cursor 社区论坛 Bug Report）：

Cursor 遵循 OpenAI Chat Completions 标准 schema。reasoning_content 不在 OpenAI 的标准字段里，它是 DeepSeek 的专有扩展。Cursor 在构建后续请求时不会保留这个字段，导致 API 返回 400 错误：

{
  "error": {
    "message": "The reasoning_content in the thinking mode must be passed back to the API.",
    "type": "invalid_request_error",
    "code": "invalid_request_error"
  }
}

正如 APIDog 指南 所总结的："Cursor's chat client follows the OpenAI Chat Completions schema. reasoning_content isn't part of that schema; it's a DeepSeek-specific extension. Cursor would need to add provider-specific handling to pass the field through."

二、deepseek-cursor-proxy 解决了什么？

代理位于 Cursor 和 DeepSeek API 之间：

Cursor → ngrok/Cloudflare Tunnel → Proxy (localhost:9000) → DeepSeek API

核心机制：

  • 缓存 reasoning：每次 DeepSeek 返回响应时，代理用 SHA-256 哈希（基于对话前缀、消息签名、工具调用 ID/签名）作为 key，将 reasoning_content 存入本地 SQLite 数据库
  • 回填 reasoning：Cursor 发来后续请求时，代理检测到缺失的 reasoning_content，从缓存中查找匹配的记录并注入
  • 恢复机制：当缓存未命中时，代理自动裁剪对话历史到最近一次用户消息，追加恢复提示，保证对话能继续

三、原项目存在什么问题？

在 300+ 条消息的深度对话中，原项目每次请求会卡住 50-110 秒。根因有三层：

层次 问题 原因
O(n²) 哈希 turn_context_signature() 在 normalize_messages 循环内每条 assistant 消息都被调用，每次都从头扫描全部前置消息 355 条消息 = ~63,000 次重复哈希
逐 key COMMIT backfill_portable_aliases() 为每个 portable key 分别调用 store.put()，每次 put() 都做一次 SQLite COMMIT（fsync 刷盘） 65 条 patched 消息 × 7 key = 455 次 fsync
无索引全表扫描 _prune_locked() 中 DELETE FROM reasoning_cache WHERE created_at < ? 在 created_at 列上没有索引，每次 COMMIT 前全表扫描数万行 455 次 COMMIT × 每次全表扫描 = 灾难

三层叠加：455 次 fsync × (COMMIT 70ms + 全表扫描 30ms) ≈ 45,000ms

四、本仓库做了什么？效果如何？

🔧 六项核心优化：

# 优化 文件 效果
1 O(n²) → O(n) 增量哈希 reasoning_store.py compute_conversation_scopes() / compute_turn_signatures() 一次遍历生成全部前缀哈希
2 backfill 批量 COMMIT reasoning_store.py 每条消息的所有 portable key 合并为一个事务，一次 COMMIT
3 created_at 加索引 reasoning_store.py CREATE INDEX idx_reasoning_cache_created_at
4 _prune_locked 短路 reasoning_store.py COUNT 预检，无过期行时跳过 DELETE
5 SQLite WAL 模式 reasoning_store.py PRAGMA journal_mode=WAL，写不阻塞读，COMMIT 飞起
6 Windows 静默启动 silent-start.pyw + VBS 开机自启、无命令行窗口、端口检测防重复启动

📊 实测效果（真实 399 条消息对话）：

修复前：┌ context ... (108,584ms)  ← 108 秒
修复后：┌ context ... (119ms)      ← 119 毫秒

快了 912 倍。 从无法忍受的分钟级等待降到完全无感。

🔑 缓存命中率保持不变：

修复前：cache_hit=99.8%
修复后：cache_hit=99.8%（完全一致）

代理由原项目演化而来，reasoning_content 注入逻辑和缓存 key 计算完全一致。验证脚本确认：增量哈希结果与原版等价（verify_hash_optimizations.py: OK: hashes match originals），不会出现缓存 miss 增多或注入错误。

🆕 额外功能：

  • --user-suffix TEXT：可在每条用户消息末尾自动追加文本（用于引导模型行为）
  • 启动即用：双击 start-deepseek-proxy.bat，自动完成 ngrok 隧道、代理启动、URL 配置

🙏 本项目 fork 自 yxlao/deepseek-cursor-proxy，作者 Yixing Lao 的杰出工作使 DeepSeek + Cursor 成为可能。我们在此基础上做了极致性能优化。

🌐 English

1. Why Can't Cursor Use DeepSeek Thinking Mode Directly?

DeepSeek's Official Requirement (Thinking Mode Docs):

"Between two user messages, if the model performed a tool call, the intermediate assistant's reasoning_content must participate in the context concatenation and must be passed back to the API in all subsequent user interaction turns."

In thinking mode, tool-call reasoning chains must remain continuous across all requests. This is a hard API contract.

Cursor's Actual Behavior (Cursor Community Forum):

Cursor follows the OpenAI Chat Completions standard schema. reasoning_content is not part of that schema—it's a DeepSeek-specific extension. Cursor strips this field when constructing follow-up requests, causing a 400 error:

{
  "error": {
    "message": "The reasoning_content in the thinking mode must be passed back to the API.",
    "type": "invalid_request_error",
    "code": "invalid_request_error"
  }
}

As the APIDog Guide explains: "Cursor's chat client follows the OpenAI Chat Completions schema. reasoning_content isn't part of that schema; it's a DeepSeek-specific extension that emerged with the R1 family and stayed in V4-Pro."

2. What Does deepseek-cursor-proxy Solve?

The proxy sits between Cursor and the DeepSeek API:

Cursor → ngrok/Cloudflare Tunnel → Proxy (localhost:9000) → DeepSeek API

Core mechanism:

  • Cache reasoning: Every DeepSeek response's reasoning_content is stored in a local SQLite database, keyed by SHA-256 hashes of conversation prefixes, message signatures, and tool-call IDs/signatures
  • Restore reasoning: On subsequent requests, the proxy detects missing reasoning_content, looks up the cache, and injects the correct value before forwarding to DeepSeek
  • Recovery: On cache miss, the proxy trims history to the last user message and adds a recovery notice, keeping the conversation alive

3. What Performance Problems Existed?

With 300+ message conversations, the original project would stall for 50–110 seconds per request. Three layers of bottlenecks:

Layer Problem Cause
O(n²) hashing turn_context_signature() called per-message inside normalize_messages, re-hashing all prior messages each time 355 msgs = ~63,000 redundant hash ops
Per-key COMMIT backfill_portable_aliases() called store.put() separately for each portable key, each doing a SQLite COMMIT (fsync) 65 patched msgs × 7 keys = 455 fsyncs
Unindexed full scan _prune_locked() ran DELETE WHERE created_at < ? with no index on created_at, scanning all rows each time 455 COMMITs × full table scan = disaster

Compounding effect: 455 fsyncs × (COMMIT 70ms + full scan 30ms) ≈ 45 seconds

4. What This Fork Improves

🔧 Six core optimizations:

# Optimization File Effect
1 O(n²) → O(n) incremental hashing reasoning_store.py Single-pass prefix hash generation for all scopes and turn signatures
2 Batched backfill COMMIT reasoning_store.py All portable keys for one message in a single transaction
3 Index on created_at reasoning_store.py CREATE INDEX idx_reasoning_cache_created_at
4 Prune short-circuit reasoning_store.py COUNT pre-check; skip DELETE when nothing to prune
5 SQLite WAL mode reasoning_store.py PRAGMA journal_mode=WAL — non-blocking writes
6 Windows silent launcher silent-start.pyw + VBS Auto-start on boot, no console window, port-guard

📊 Real-world results (399-message conversation):

Before: ┌ context ... (108,584ms)  ← 108 seconds
After:  ┌ context ... (119ms)       ← 119 milliseconds

912× faster. From unbearable minute-long waits to completely imperceptible.

🔑 Cache hit rate preserved:

Before: cache_hit=99.8%
After:  cache_hit=99.8% (identical)

The reasoning_content injection logic and cache key computation are identical to the original. Benchmark verification confirms: incremental hashes match the originals exactly (verify_hash_optimizations.py: OK: hashes match originals). No increased cache misses, no injection errors.

🆕 Bonus features:

  • --user-suffix TEXT: Automatically append text to every user message before forwarding (useful for guiding model behavior)
  • One-click startup: double-click start-deepseek-proxy.bat for automatic ngrok tunnel + proxy launch

🛠 完整安装指南（中文） · Full Installation Guide (English)

第 1 步：注册并安装 ngrok 隧道

Cursor 屏蔽 localhost 等非公网 API URL，因此代理必须通过公网 HTTPS 可达。ngrok 可免费将本地端口暴露到公网。也可使用 Cloudflare Tunnel。

注册 ngrok 账号后，访问 ngrok dashboard 获取 authtoken：

macOS / Linux:

brew install ngrok
ngrok config add-authtoken <your-ngrok-token>

Windows: 用 winget install Ngrok.Ngrok，或从 ngrok.com/download 下载。安装后在终端运行 ngrok config add-authtoken <your-ngrok-token>。

Step 1: Install ngrok Tunnel

Cursor blocks non-public API URLs such as localhost, so the proxy needs a public HTTPS URL. ngrok exposes your local proxy to Cursor without opening router ports. Cloudflare Tunnel is also supported.

Register an ngrok account, visit ngrok dashboard, get your authtoken:

# macOS / Linux
brew install ngrok
ngrok config add-authtoken <your-ngrok-token>

# Windows
winget install Ngrok.Ngrok
ngrok config add-authtoken <your-ngrok-token>

第 2 步：安装并启动代理

方式 A：使用 UV（推荐）

# 安装 uv（如果没有）
curl -LsSf https://astral.sh/uv/install.sh | sh   # macOS/Linux
# Windows: winget install astral-sh.uv

# 克隆并启动
git clone https://github.com/abcduyt1243-git/deepseek-cursor-proxy-turbo.git
cd deepseek-cursor-proxy-turbo
uv run deepseek-cursor-proxy

方式 B：使用 pip + venv

git clone https://github.com/abcduyt1243-git/deepseek-cursor-proxy-turbo.git
cd deepseek-cursor-proxy-turbo
python -m venv .venv
.venv\Scripts\activate       # Windows
# source .venv/bin/activate  # macOS/Linux
pip install -e .
deepseek-cursor-proxy

方式 C：使用 Conda

conda create -n dcp python=3.10 -y
conda activate dcp
git clone https://github.com/abcduyt1243-git/deepseek-cursor-proxy-turbo.git
cd deepseek-cursor-proxy-turbo
pip install -e .
deepseek-cursor-proxy

启动后代理会在终端打印 ngrok 公网 URL。

如果使用固定 ngrok 域名（reserved endpoint / custom domain），在 ~/.deepseek-cursor-proxy/config.yaml 中设置 ngrok_url，或通过命令行传递：

ngrok: true
ngrok_url: https://your-subdomain.ngrok.dev
deepseek-cursor-proxy --ngrok-url https://your-subdomain.ngrok.dev

首次运行会自动创建：

  • ~/.deepseek-cursor-proxy/config.yaml：配置文件
  • ~/.deepseek-cursor-proxy/reasoning_content.sqlite3：推理缓存数据库

常用命令行选项：

deepseek-cursor-proxy --no-display-reasoning   # 隐藏思考 token
deepseek-cursor-proxy --verbose                 # 详细日志
deepseek-cursor-proxy --no-ngrok               # 不使用隧道（localhost）
deepseek-cursor-proxy --port 9000              # 自定义端口
deepseek-cursor-proxy --user-suffix "请用中文思考"  # 追加用户消息后缀

Step 2: Install & Start

Option A: UV (recommended)

curl -LsSf https://astral.sh/uv/install.sh | sh   # macOS/Linux
# Windows: winget install astral-sh.uv

git clone https://github.com/abcduyt1243-git/deepseek-cursor-proxy-turbo.git
cd deepseek-cursor-proxy-turbo
uv run deepseek-cursor-proxy

Option B: pip + venv

git clone https://github.com/abcduyt1243-git/deepseek-cursor-proxy-turbo.git
cd deepseek-cursor-proxy-turbo
python -m venv .venv
source .venv/bin/activate     # macOS/Linux
.venv\Scripts\activate        # Windows
pip install -e .
deepseek-cursor-proxy

Option C: Conda

conda create -n dcp python=3.10 -y
conda activate dcp
git clone https://github.com/abcduyt1243-git/deepseek-cursor-proxy-turbo.git
cd deepseek-cursor-proxy-turbo
pip install -e .
deepseek-cursor-proxy

Common flags:

deepseek-cursor-proxy --no-display-reasoning   # Hide thinking display
deepseek-cursor-proxy --verbose                 # Verbose output
deepseek-cursor-proxy --no-ngrok               # Run on localhost only
deepseek-cursor-proxy --port 9000              # Custom port
deepseek-cursor-proxy --user-suffix "Think in Chinese"  # Append to user msgs

第 3 步：配置 Cursor

在 Cursor 中添加自定义模型：

  • Model：deepseek-v4-pro（或 deepseek-v4-flash）
  • API Key：你的 DeepSeek API Key（以 sk- 开头）
  • Base URL：代理打印的 ngrok URL + /v1，例如 https://xxx.ngrok-free.dev/v1

代理会原样转发 Cursor 发送的模型名（如 deepseek-v4-pro、deepseek-v4-flash）。config.yaml 中的 model 字段仅在没有传模型名时作为兜底。

按 Ctrl+Shift+0（Windows/Linux）或 Cmd+Shift+0（macOS）可快速切换自定义 API。

Step 3: Configure Cursor

In Cursor, add a custom model:

  • Model: deepseek-v4-pro (or deepseek-v4-flash)
  • API Key: your DeepSeek API key (starts with sk-)
  • Base URL: the ngrok URL printed by the proxy + /v1, e.g. https://xxx.ngrok-free.dev/v1

Toggle with Ctrl+Shift+0 (Windows/Linux) or Cmd+Shift+0 (macOS).

第 4 步：开始对话

在 Cursor 中选择 deepseek-v4-pro，正常使用 Chat 或 Agent 模式即可。

Step 4: Start Chatting

Select deepseek-v4-pro in Cursor and use chat or agent mode as usual.

🔧 调试 · Debugging

deepseek-cursor-proxy --verbose                                    # 详细日志
deepseek-cursor-proxy --no-ngrok --port 9000 --verbose             # 本地测试
deepseek-cursor-proxy --verbose --trace-dir ./trace-dumps          # 完整追踪
deepseek-cursor-proxy --config ./dev.config.yaml                   # 自定义配置
deepseek-cursor-proxy --clear-reasoning-cache                      # 清空缓存

📜 License

MIT — same as the original project.

🙏 Credits

Forked from yxlao/deepseek-cursor-proxy by Yixing Lao — the brilliant original work that made DeepSeek + Cursor possible.

About

Fixes missing reasoning_content for DeepSeek V4

Resources

Readme

License

MIT license
Activity

Stars

1 star

Watchers

0 watching

Forks

0 forks
Report repository

Releases

No releases published

Packages

 
 
 

Contributors

Languages

  • Python 99.9%
  • VBScript 0.1%