qyjqio/dsh-cursor-toolkit

llm-probe

列出运行中 DSH 进程真实注册的 LLM 路由，并可跑一次真实流式调用。

‍ 开发与运行时cursorcursor-agentdeepseek-harness

@qyjqio/dsh-llm-probe

Diagnostic tool for a running DSH app: shows the LLM provider routes the process actually registered, and runs one real streamed call through the llm service.

Why

A DSH Session freezes its subagent model-selection policy when it starts, so a provider or model added later — even one the live process has already mounted — stays invisible to that session's subagent tool. llm_probe bypasses the policy: it calls the llm service directly, so it sees what is really registered right now.

Typical use: you add a provider on the Models page (or a local bridge such as @qyjqio/cursor-openai-bridge), and want to prove the adapter works without opening a new session.

Install

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

Then enable include:llm-probe in Settings → Plugins.

Parameters

name	type	meaning
`provider`	string	Registered provider route, default `cursor`
`model`	string	Model id under that provider, default `grok-4.7-low-fast`
`prompt`	string	When present, runs one streamed call; when absent, only lists routes
`system`	string	Optional system prompt
`maxTokens`	integer	Optional output cap

Result

providers (every registered route), models (that provider's models), and for a call: text, reasoning, finish, usage, toolCalls, plus error on failure.

Notes

Read-only: it never writes config or credentials.
The defaults match the companion Cursor bridge; pass provider/model explicitly for anything else.
After editing index.js, reload the bundle: plugin_manager set_plugin target=include:llm-probe enabled=false then true.