dsh-adaptive-model-router

Running both a Flash and a Pro model in DeepSeek Harness but still having to hand-pick which one handles each turn is tedious and error-prone. dsh-adaptive-model-router adds a local, synchronous, fully deterministic per-turn adaptive routing layer: simple work stays on the economy model, while complex work or stalled progress automatically upgrades to the quality model. No external classifier is called, no prompt or hidden instruction is injected, and zero extra tokens or provider requests are added.

Routing is rule-based: the quality route activates when entering text hits a configurable character threshold or matches a configured case-insensitive keyword. Within a single turn, hysteresis ensures the route can only upgrade, never downgrade. Consecutive non-excluded tool failures beyond a set count also trigger the upgrade. A successful tracked tool resets only the failure counter; it does not pull an already-upgraded turn back down. A new user turn starts a fresh decision and may return to the economy route. Each route accepts optional reasoningEffort and maxTokens; every other request field stays unchanged. The package targets DeepSeek Harness 0.1.1-rc.2 and ships as plain ESM JavaScript with no runtime dependencies.

Ideal for developers running multiple models side-by-side in Harness, and for teams that value reproducible inference chains, telemetry-friendly request logs, and per-turn cost optimization. Tune the thresholds to balance quality against spend without modifying application code.