dsh-agentic-router: A Learning-Based Intelligent Routing Plugin for DeepSeek Harness

dsh-agentic-router is an intelligent routing plugin for DeepSeek Harness (DSH), designed to automatically select the optimal model tier based on task complexity. It employs a four-expert architecture (Rule-based, UCB, LinUCB, and k-NN) with an EXP3 meta-selector, and continuously optimizes its routing policy in real time through a data flywheel mechanism driven by quality signals such as failures, cost, and latency. The new version supports actual model switching and step-level routing while providing various control tools. It defaults to shadow mode and can be switched to active mode via configuration or commands. All decision and reward data are persisted to disk, enabling replay to restore the learning state, and human feedback correction is supported through the Web UI.