dsh-model-router

Model Router & Cost Optimizer for DeepSeek Harness (dsh). Routes every agent step to the cheapest sufficient model, degrades gracefully on transient provider failures, and shows live per-session token / cache-hit / cost figures right under the composer.

The harness is in developer preview and iterates quickly — expect compatibility-breaking changes.

Features

• Heuristic tier routing — at turn start the newest user message is classified (agent/pre-step: task keywords + payload size; history is deliberately ignored so a long agentic conversation can still route a trivial follow-up cheap). Trivial prompts (是什么 / 解释 / 翻译 / hello …) go to the cheap catalog model; agentic work (实现 / 重构 / 调试 …) stays on the strong one. The chosen tier is sticky within a turn.
• Switch-cost hysteresis — every model flip pays a full-prefix cache miss on the target model (a long cached session is cheapest staying put). So cheap routing engages only after two consecutive cheap-classified turns, while heavy turns switch back immediately. Manual mode (/router, route_model, dock buttons) bypasses the hysteresis.
• Automatic fallback — transient failures (RATE_LIMIT, SERVER, TIMEOUT, EMPTY_RESPONSE) degrade the turn to the cheap model and retry once; anything else delegates to the provider's own retry policy.
• Real usage metering — a session projection folds the durable log: real adapter token usage (input / output / cache read / cache write / reasoning), per-model breakdown, and estimated cost from a model-class price table. Projection-based, so the numbers are replay-safe and survive cold sessions.
• Composer dock panel — tier buttons (Auto / 省 / 强), current model, miss/out/cache%/≈$ line, and a per-model usage breakdown. Reactively driven by useProjection; buttons reuse the built-in commands remote — no custom wire protocol.
• Manual control — /router auto|cheap|strong slash command, and a model-visible route_model tool so the agent itself can request a tier mid-task.

Install

Prerequisites

The dsh CLI must be on your PATH. If you only ever ran the harness through npx, dsh is not installed and you will get — install it globally first: