DSH Cursor Toolkit

Three small pieces that let DeepSeek Harness (DSH) use Cursor:

package	what it is
@qyjqio/dsh-cursor-cli	DSH host plugin adding the cursor_agent tool — hands a coding task to the local cursor-agent CLI and returns its final message plus chat id (resumable).
@qyjqio/cursor-openai-bridge	OpenAI-compatible HTTP server backed by cursor-agent, so Cursor models can be registered as a normal DSH LLM provider.
@qyjqio/dsh-llm-probe	Diagnostic tool that lists the provider routes a running DSH process registered and runs one real streamed call — useful because a Session's subagent policy is frozen at start.

They are independent: install the bridge + probe if you only want Cursor models inside DSH's model picker; install the CLI tool if you want an agent that edits files in a workspace.

Why a bridge instead of direct API access

Cursor exposes no OpenAI-compatible endpoint (/v1/models → 404); inference runs over a private connect/protobuf API. Wrapping the official CLI keeps auth, model routing and the agent loop in Cursor's own hands, and keeps this repo small and dependency-free.

Install

From the plugin market

Add this index source in DSH → Plugins → Discover → Add index source:

https://raw.githubusercontent.com/qyjqio/dsh-cursor-toolkit/main/index.json

Then install Cursor CLI Bridge and/or LLM Probe from the market UI.

From Git

plugin_manager install_bundle target="git+https://github.com/qyjqio/dsh-cursor-toolkit.git#<commit>&path:/packages/dsh-cursor-cli"
plugin_manager install_bundle target="git+https://github.com/qyjqio/dsh-cursor-toolkit.git#<commit>&path:/packages/dsh-llm-probe"

Pin <commit> to a full 40-hex SHA; main also works for local experiments.

The bridge

npm i -g @qyjqio/cursor-openai-bridge   # once published to npm
cursor-openai-bridge                    # listens on 127.0.0.1:8791

Requirements: cursor-agent installed and logged in (cursor-agent login), Node.js ≥ 22.

License

MIT — see LICENSE.