dsh-model-router

A serial model-orchestration layer for the DeepSeek Harness: a delegate tool that lets the main agent split a request into subtasks and hand each one to the model that fits it best, chosen against a built-in catalog of current Chinese-market models.

The main agent keeps the orchestrator role. Each delegate call starts a child agent on the chosen route with a fresh context, waits for its result, and returns it inline — one subtask at a time, serially, so results are reviewed before the next step.

Requirements

• DeepSeek Harness with the Web UI. The package declares a dsh.bundle manifest, is Host-only (no browser half), and its engines.dsh is >=0.2.0-rc.1 || >=0.2.1-rc.1 || >=0.2.2-rc.1.
• Model selection must be allowed by the Session policy: open Plugins → Subagent → Model selection, enable it, tick the routes you want, then start a new Session (the policy is recorded per new Session). Without a recorded policy, only the parent Session's current route may be used — delegate refuses and says exactly that.

Install

dsh plugin add yunxiao11xie/dsh-model-router

Built-in catalog (calibrated 2026-11)

Model	Provider	Fits
deepseek-flash (V4.1-Flash)	deepseek-account	frontend, UI, fast iteration, vision
deepseek-v4-pro	deepseek-account	hard reasoning, complex analysis
glm-5.3	zai-coding-cn	backend, engineering rigor, review
glm-5.3-flash	zai-coding-cn	multimodal, fast frontend work
doubao-seed-evolving	volcengine	Chinese writing, agent work (evergreen id)
hunyuan-turbos-latest	tencent	Chinese documents, low cost (evergreen id)
MiniMax-M3.1-Flash	minimax	long context, batch work
mimo-v2.6-pro-ultraspeed	alibaba (Bailian)	high-volume light tasks
kimi-k3.1-pro / kimi-k3	moonshot	long-context research, Chinese writing
qwen3.8-max / qwen3.8-flash	alibaba	reasoning, presentations; cheap documents
gpt-6-astra / gpt-6.1-sol	openai	general reasoning, English writing

Entries whose provider is not configured are simply refused with guidance until you add the provider. Override any entry or add your own in the plugin config:

- id: model-router
  name: dsh-model-router
  config:
    defaultModel: deepseek-flash
    models:
      - id: glm-5.3
        provider: zai-coding-cn
        strengths: [coding, backend, review]
        notes: ...
        cost: medium

How the agent uses it

The plugin injects a short system-prompt section listing the catalog and a routing method: split the request, judge what each subtask needs (kind of work, language, rigor, speed), match against strengths, and state the match in the required reason field. Any work type fits — code, documents, slides, translation, research.

Behavior rules:

• Serial: one delegation at a time; the tool waits for the result.
• Auditable: every call records why that model was chosen.
• Lenient fallback: a model id outside the catalog falls back to defaultModel, and the result says so.
• Policy-checked: routes must be allowed by the Session model-selection policy and have a mounted provider; refusals explain what to do.

License

MIT