dsh-smart-router

The sharpest cut for DSH users is rarely a wrong calculation; it is the invoice. After model price hikes, a simple formatting pass still burns full thinking tokens, and the monthly bill multiplies overnight. Swapping to a cheaper model fixes the cost but tanks the reasoning you actually need. dsh-smart-router takes a different approach: instead of replacing one model with another, it lets a single conversation route automatically by task difficulty—architecture, refactoring, and gnarly debugging go to a strong model; chit-chat, cleanup, and small edits go to a fast, cheap one; medium tasks sit in the middle, each tier with its own thinking effort.

Vision is folded into the same router. If a message contains an image it lands on the vision tier automatically; pure text follows the three difficulty levels. No separate vision plugin required. A free OVHcloud Qwen2.5-VL-72B anonymous endpoint ships as the default, so it works out of the box, and you can swap in any vision model you have configured. Vision evidence follows a structured scheme—summary, OCR, layout order, entities, and uncertainty markers—if parsing fails the image degrades to a placeholder and the request keeps going.

Routing lives at the LLM adapter layer. It does not patch the host, does not hijack existing provider routes, and coexists with other vision plugins. Two classifier modes are available: a zero-cost heuristic (keywords, code volume, file references) and an LLM-based classifier with a 120-second cache. Every tier is optional; unconfigured tiers fall back to your default model with a fail-open guarantee.

Built for developers who run DSH coding agents daily, feel the output-pricing pinch, need both text and vision in one loop, and refuse to pay premium rates for tasks a lighter model could handle.