AI Agent Hub
Back to plugins
⚙️

dsh-consult

Workflow Updated 2026.08.25

Run the following command in DeepSeek Harness:

dsh plugin install 1339190177/dsh-consult

Paste the following prompt into your AI chat to install this plugin:

Install it in DeepSeek Harness by running dsh plugin install 1339190177/dsh-consult; the source repository is available at https://github.com/1339190177/dsh-consult.

About this plugin

When you rely on a single model vendor, edge cases often leave you stuck in trial-and-error loops. dsh-consult adds a cross-vendor second-opinion channel to the dsh ecosystem: native agents can registered a consult tool to request structured advice from a different provider mid-session, while external harnesses (ZCode, etc.) gain a lightweight local OpenAI-compatible advisor endpoint that reuses the existing web port with no extra service.

The design centers on default-official routing, single-call semantics, and one authoritative timeout. The default route points at official DeepSeek Flash; a three-level selector (provider → model → reasoning tier) lets you switch freely. Multimodal requests accept multiple data-URL images (png/jpeg/webp/gif) and automatically upgrade to a vision model. The scout investigation mode layers real-time web search (up to 8 rounds) on top of the official route for high-difficulty problems. Every consultation is audited as a single JSONL line with actual route, latency, and image count, and the response carries an x-model-upstream header for provenance. Structured output returns recommendation, reasons, confidence, risks, and alternatives by default; the HTTP endpoint also supports json_schema constraints.

Built for engineers running multi-agent dsh workflows who want an independent perspective on critical decisions, and for teams already using external harnesses like ZCode that need a low-cost local advisor without standing up a full LLM gateway.

Use Cases

  • An agent hits a tricky edge case mid-session and calls the consult tool to get structured advice from a different provider
  • External harnesses like ZCode hit a local loopback endpoint for lightweight consulting without standing up their own LLM gateway
  • Paste screenshots or charts into a multimodal request and the advisor auto-upgrades to a vision model for cross-vendor interpretation

Best For

  • Engineers running multi-agent dsh workflows who want an independent perspective on critical decisions
  • Teams with external harnesses like ZCode that need a low-cost local advisor without a full LLM gateway
  • Teams prioritizing auditability, provenance tracing, and structured logging of model outputs