DSH | DSH SuperAdvisor | A second opinion for your main model, approved by you

[Di-Chai]

Unofficial project, independently developed and maintained by community members.

Project: kvmem/dsh-super-advisor

Renamed in 0.1.7: formerly DSH Advisor (dsh-tool-advisor), now DSH SuperAdvisor (dsh-super-advisor). Existing users: follow the upgrade steps to replace the old package while keeping settings and audit records.

DSH SuperAdvisor lets the main model request a focused second opinion through ask_advisor. You review and approve the complete request before it reaches the advisor. The advisor returns text advice, and the main model continues to verify and carry out the task.

• Review before sending: inspect the question and selected evidence, edit the request, remove evidence, or reject it.
• Reuse your DSH configuration: select an existing service and model in Settings → Plugins → DSH SuperAdvisor. Endpoints and credentials stay in Models settings.
• Control the shared context: only the reviewed text and explicitly selected evidence are sent; the plugin does not automatically attach the full conversation or workspace. The advisor has no execution tools.
• Read the result: expand the advice, inspect the original text, and follow the request ID. Send markers and cached results prevent automatic duplicate sends after interruption.

The plugin uses DSH's native tool, model-adapter, approval, user-question, and Web UI extension services. It also adds guidance about when to seek advice; the main model still decides whether to call the tool.

By default, output follows the model service's token budget and the plugin adds no received-text size cap. Model/service limits and the default 120-second timeout still apply. Approved calls may incur your model service's normal charges.

Typical setups

• Local main model + cloud advisor: keep routine analysis and execution with a local model connected to DSH and capable of calling tools. Approve focused requests to a stronger cloud model when help is needed; the local model verifies the advice and continues.
• Flash main model + Pro advisor: use deepseek-v4-flash for the main task and deepseek-v4-pro for difficult questions or focused reviews, with user approval for each consultation.

The aim is to approach the quality of using a stronger model throughout while reducing stronger-model usage and cost. Actual quality and savings depend on the task, models, context size, and consultation frequency. No comparative quality or cost benchmark has been performed.

Screenshot

Actual DSH settings UI with a local test configuration. The settings card currently uses Chinese labels; documentation is available in English and Chinese.

Install

Tested with DSH 0.1.3-alpha.2, Node 24, Linux, and a local POSIX filesystem. DSH's installer also requires pnpm. Other versions and operating systems have not been verified.

Download and install the latest GitHub release without entering a version number:

curl -fL https://github.com/kvmem/dsh-super-advisor/releases/latest/download/dsh-super-advisor.tgz -o dsh-super-advisor.tgz
dsh plugin --profile web add ./dsh-super-advisor.tgz
dsh web

This downloads the latest release when you run the command; it does not enable automatic background updates.

Configure the service in Settings → Models, including its saved Base URL, credentials, and model. Then select the advisor under Settings → Plugins → DSH SuperAdvisor and save. Main-model and advisor-model selections are independent; saving settings does not run inference.

The current release passes 51 automated tests, build checks, real DSH Loader checks, and Chromium settings checks. This release uses controlled model fixtures for testing. Earlier acceptance included a real DeepSeek Flash → Pro → Flash flow. GLM/Qwen tests used local mock endpoints, not their cloud services. See ACCEPTANCE.md for the full scope.

中文简介

这是一个非官方 DSH 顾问插件：主模型遇到难题时，通过 ask_advisor 整理问题和证据；你先查看、编辑或拒绝，逐次批准后才发送给顾问模型。顾问返回文字建议，主模型继续验证和执行。

0.1.7 起名称改为 DSH SuperAdvisor，包名为 dsh-super-advisor。旧版用户请参考中文升级说明，替换安装包并保留原设置和审计记录。

模型服务、地址和密钥复用 DSH 的 Models 设置，顾问在 Plugins → DSH SuperAdvisor 中选择。默认沿用模型服务的输出预算，插件接收文本不设上限。模型自身限制及请求超时仍有效。

两个典型场景：

• 本地主模型 + 云端大模型顾问：具备工具调用能力、已接入 DSH 的本地模型处理日常任务；遇到难题，经用户批准后向云端大模型求助。
• Flash 主模型 + Pro 顾问：Flash 负责主要执行，按需请 Pro 分析难题或复核，收到建议后由 Flash 继续验证和执行。

目标是在接近全程使用更强模型效果的同时降低成本。实际效果与节省幅度取决于任务、模型、上下文长度和求助频率，目前尚未进行质量与费用的对照评测。

English README · 中文说明 · 反馈问题