Introduction to the DeepSeek Harness Automatic Model Routing Plugin
The `dsh-auto-model-router` plugin addresses the inconvenience of manually switching model tiers in DeepSeek Harness (DSH). Using session state machines, keyword matching, and a manual-priority mechanism, it automatically routes requests to the Chat (cost-saving) or Work (deep reasoning) tier based on context. The plugin supports automatic switching, cooldown limits, and logging of first-round selections, significantly improving the user experience. It becomes active after re-entering DSH following installation, although a few known limitations remain in scenarios such as IM first-round popups.
Read Moredsh-auto-model-router: L0–L3 Model Routing for DeepSeek Harness
dsh-auto-model-router is a third-party model routing plugin for DeepSeek Harness (DSH) designed to address the trade-off between cost and effectiveness caused by significant disparities in model capability requirements across different requests. By registering a virtual `auto` provider, this plugin intelligently routes requests to four capability tiers (L0–L3) and supports three user modes: heuristic lock, fixed tiers, and fully-auto. Its core features include zero-latency heuristic locking based on text width, configurable-weight keyword scoring, an optional LLM classifier, a failure fallback chain, and budget control strategies that prioritize cost, quality, or balance. When a real model is explicitly selected, the plugin does not interfere with routing, allowing the request to pass through unchanged. This plugin requires Node version ^22.19.
Read More