dsh-auto-router

When multiple model routes are connected in DeepSeek Harness, users often switch between light, heavy, code, and vision models by hand, disrupting flow and risking a wrong pick. dsh-auto-router抽象ifies this decision into a configurable auto-routing layer: the model picker gains an Auto Router provider with six choices, and a single click routes the request to the best-fit model automatically.

By default, the classifier sends images to vision, code snippets to code, long or complex tasks to heavy or medium, and short queries to light. With sticky mode enabled, a session can only upgrade its tier and never oscillate back, keeping multi-turn tasks on a consistent model. Users can also override the tier to force a specific level and skip classification entirely. The package stores, reads, or transmits no API keys; credentials remain in the target route standard DSH settings.

Ideal for developers who already run multiple model routes in DSH and want to cut down manual switching, as well as users who prefer context-aware routing over picking a model every time. It requires only one existing route to work with, with no extra credentials or external services needed.