llm-adaptive

Managing multiple LLM backends often leads to manual model selection, which is inefficient and error-prone. llm-adaptive plugin offers a smart solution for DeepSeek Harness: it automatically classifies each request's complexity (low, medium, high, critical) via a flash classifier and routes it to the matching backend based on configured chains. This streamlines workflows while improving resource utilization and response quality.

The plugin's core capabilities lie in its flexibility and robustness. It features context-aware judging, incorporating session goals and recent turns for better classification; sticky level protection prevents unintended downgrades during tasks; and configuration-driven routing chains from pool.json allow immediate adjustments. Decision caching and observable logs enhance transparency, and a fail-open mechanism ensures requests are never blocked.

llm-adaptive is ideal for developers, data scientists, or teams using DeepSeek Harness who need to manage multiple LLM backends and seek automated, efficient routing. Whether in development or production, it helps users dynamically optimize model selection, letting them focus on their core work.