1960697431/dsh-difficulty-router
1960697431★ 0JavaScript最后同步: 2026-08-14
Difficulty-aware + taste-aware model routing plugin for DeepSeek Harness (DSH): hard/ambiguous tasks → strong model, UI/design → taste model, easy work stays cheap. Provider-agnostic, safe-by-default, calibrated on real bake-offs. MIT.
README 摘要
dsh-difficulty-router Difficulty-aware + taste-aware model routing for DeepSeek Harness (DSH). Hooks the agent/request waterfall and, per request step, decides which model should actually serve it: - Hard / ambiguous tasks → escalate from your cheap default to a configured strong model . - UI / design tasks → route to a configured taste model (design quality matters more than a few cents). - Easy tasks → stay on the cheap default. The goal: spend strong-model money only where it buys quality , and protect your codebase from the places where a weak model guesses wrong and creates cruft. 中文:难度 + 品味双通道的模型路由插件。困难/含糊任务升级到强模型,UI/设计 任务路由到有品味的模型,简单活留在便宜的默认模型上。只在真正需要质量的地方 花强模型的钱。 What it looks like Same session, three requests — the router reads each step and routes it ( logFile JSONL, one line per decision): Every line is a real measured decision from a live session. Install it, run dryRun: true for a day, and watch what would have routed — then flip to false and let it work. Why A real bake-off on a production monorepo (backend security task + frontend redesign, hidden acceptance tests, token usage read from session logs) showed: Task cheap model strong model single-file security fix pass…
在 GitHub 查看完整 README →