model-router

Thinking/execution model routing: turn step 1 uses a configurable thinking model, later steps use the default model.

AI 分析

核心用途是实现多轮对话中的模型分流，第一步请求使用高强度思考模型，后续步骤自动切回普通模型。适合希望平衡推理深度与响应成本的 DSH 用户。

パッケージ

model-router

バージョン

0.1.0

ライセンス

MIT

最終更新

2026/08/14

Configuration

The thinking model is a model-router settings namespace (persisted in settings.yaml):

Field	Meaning	Default
`thinkingProvider`	Provider route for step-1 requests	`deepseek-official`
`thinkingModel`	Model id for step-1 requests	`deepseek-v4-pro`
`reasoningEffort`	Adapter-owned reasoning effort (`off` / `high` / `max`, empty = provider default)	`high`

Ways to change it:

UI — the 🧠 pill in the composer: model dropdown (grouped by provider, from the live catalog) + reasoning-effort dropdown; selection applies immediately.
Settings — the namespace may also appear in the Settings → Plugins section (schema-driven card), depending on the deployment's settings UI.
Settings file — edit settings.yaml:

model-router:
  thinkingProvider: deepseek-official
  thinkingModel: deepseek-v4-pro
  reasoningEffort: max

Composition — change the row config: (the base layer; user settings win over it).

The execution model is not configured here: it always follows the agent's default model selection (the Web model selector / agent-default-model settings). To make execution fast, set the default model to a cheap model such as deepseek-v4-flash.

model-router

AI 分析

インストール

ドキュメント

Configuration