sileya-ai
/
deepseek-cursor-proxy
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
deepseek-cursor-proxy/README.md

5.1 KiB
Raw Blame History

deepseek-cursor-proxy

A simple proxy that caches and restores DeepSeek reasoning_content across tool-call turns in Cursor, making thinking models like deepseek-v4-pro and deepseek-v4-flash work correctly.

What It Does

  • Caches DeepSeek reasoning_content from regular and streamed responses, then restores it on later tool-call turns when Cursor omits it.
  • Mirrors streamed reasoning_content into Cursor-visible <think>...</think> text so thinking tokens are shown in Cursor BYOK/proxy chats. Cursor currently renders this as normal chat text, not as a native collapsible Thinking block.
  • Provides other compatibility fixes for running Cursor with the DeepSeek official API.

Why This Exists

DeepSeek thinking mode returns reasoning_content separately from final content. After an assistant turn with tool calls, DeepSeek requires that same reasoning_content to be sent back in later requests. Cursor can omit it in custom OpenAI-compatible flows, causing The reasoning_content in the thinking mode must be passed back to the API. This proxy caches reasoning by conversation prefix, message signature, and tool-call IDs, then restores it before forwarding to DeepSeek.

For streamed responses, the proxy also mirrors DeepSeek reasoning_content into Cursor-visible <think>...</think> content while leaving the original reasoning_content field intact. This lets Cursor display the thinking text in OpenAI-compatible BYOK/proxy flows, and the proxy strips those display-only tags from later assistant history before replaying it to DeepSeek.

This repo fixes the following error:

Error 400 - reasoning_content must be passed back

⚠️ Connection Error

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

1. Install

source ~/miniconda3/etc/profile.d/conda.sh
conda activate pytools
PIP_REQUIRE_VIRTUALENV=false python -m pip install -e .

2. Configure

mkdir -p ~/.deepseek-cursor-proxy
chmod 700 ~/.deepseek-cursor-proxy
cp .env.example ~/.deepseek-cursor-proxy/.env
chmod 600 ~/.deepseek-cursor-proxy/.env

.env.example is only a safe template. The proxy loads ~/.deepseek-cursor-proxy/.env automatically, and that file should stay outside this repository because it contains your keys.

Edit ~/.deepseek-cursor-proxy/.env:

DEEPSEEK_API_KEY=sk-your-deepseek-key
PROXY_API_KEY=cursor-local-token
CURSOR_DISPLAY_REASONING=true

Keep PROXY_API_KEY set when using ngrok because the proxy will be reachable from the public internet.

By default, reasoning cache data is stored at:

~/.deepseek-cursor-proxy/reasoning_content.sqlite3

Override it with REASONING_CONTENT_PATH or deepseek-cursor-proxy --reasoning-content-path <path> only when you need a custom location.

3. Set Up Ngrok Once

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

4. Run

deepseek-cursor-proxy --verbose

The proxy prints a line like:

Cursor Base URL: https://example.ngrok-free.app/v1

Use that URL in Cursor. If you do not use ngrok and point Cursor at localhost or 127.0.0.1, Cursor may fail with ssrf_blocked: connection to private IP is blocked.

5. Cursor Settings

  • OpenAI Base URL: the printed ngrok URL ending in /v1
  • OpenAI API Key: the value of PROXY_API_KEY
  • Model: deepseek-v4-pro

Useful Commands

Run without ngrok for local curl testing:

PROXY_NGROK=false deepseek-cursor-proxy --port 9000 --verbose

Disable the Cursor display mirror if you only want raw OpenAI-compatible response fields:

CURSOR_DISPLAY_REASONING=false deepseek-cursor-proxy --verbose

Log full request bodies only when needed:

deepseek-cursor-proxy --ngrok --verbose --log-bodies

This prints the Cursor request body, the normalized DeepSeek request body, DeepSeek error bodies, and the final streamed assistant message.

Use a different env file for development:

deepseek-cursor-proxy --config ./dev.env

Run tests:

PYTHONPATH=src python -m unittest discover -s tests

Development

Pre-commit runs whitespace checks, Black, and Ruff:

PIP_REQUIRE_VIRTUALENV=false python -m pip install -e ".[dev]"
pre-commit install
pre-commit run --all-files

Notes