@chaoyuzhang/dsh-auto-router

A configurable Auto model-routing tier for DeepSeek Harness. It registers a synthetic provider and delegates each request to a model on an existing provider route.

What it adds

The model picker receives an Auto Router provider with six choices:

• Auto · 智能路由: classify each request
• Auto · 快速: force light
• Auto · 均衡: force medium
• Auto · 深度: force heavy
• Auto · 代码: force code
• Auto · 视觉: force vision

The default classifier routes images to vision, code-like requests to code, longer/deeper tasks to heavy or medium, and short requests to light. With sticky: true, a session only upgrades its tier; this avoids oscillating between models during a task.

Installation

Install the exact package version into a DSH Desktop profile through the DSH Community Market or the managed plugin command. The package includes a dsh.bundle.patch entry and does not contain credentials.

The target provider must already exist. The default configuration expects a route named ciyuan2api with these model IDs:

gpt-5.6-luna, gpt-5.6-terra, gpt-5.6-sol,
kimi-k2.7-code, gemini-3.5-flash

Configuration

Override the inserted auto-router row in a profile patch:

- id: auto-router
  config:
    providerId: auto
    displayName: My Auto Router
    targetProvider: my-openai-gateway
    tiers:
      light: fast-model
      medium: balanced-model
      heavy: reasoning-model
      code: coding-model
      vision: vision-model
    sticky: true
    mediumThreshold: 300
    heavyThreshold: 1200
    contextWindow: 262144
    defaultMaxTokens: 32768

The router registers providerId as its synthetic route and delegates to targetProvider. The target route must advertise the configured model IDs.

The package does not store, read, or transmit API keys. Credentials belong to the target provider's normal DSH settings and credentials configuration.

Compatibility

• Node.js >= 22.13.0
• DSH rc.7-compatible LLM seam
• @deepseek-ai/cordis ^4.0.1
• @deepseek-ai/dsh-llm ^0.1.0-rc.7
• @deepseek-ai/schemastery ^3.18.1

Security

This package inspects recent user text locally to choose a target model. The selected model receives the same request that the user intentionally sent through DSH. No telemetry, key handling, lifecycle scripts, shell commands, or external endpoints are included.

See SECURITY.md.

License

MIT.