dsh-effort-router

In daily DeepSeek Harness use, the most annoying cost problem is rarely that the model is too weak. It is that every single turn burns the most expensive brain on the lightest question, or that a hard problem is shoved into a cheap model and the quality visibly drops. dsh-effort-router automates the decision of how much reasoning a turn actually needs: it classifies each request, overrides the provider, model, and reasoning effort for that one call only, then hands everything back. No session-level or global settings are ever touched.

At its core is a zero-token rule classifier that sorts input into four tiers: simple, standard, hard, and vision. Only the ambiguous gray zone triggers a single 8-token call to a cheap model for arbitration, so there is no keyword whitelist to maintain. The vision tier is independent, with reasoning effort scaled from low to max by difficulty, so a holiday photo and a stack-trace screenshot no longer share the same level. Before every send, the client syncs the turn plan onto the composer seat, so the model box always reflects what this turn will really invoke. A seat you picked by hand is never overridden; the plugin logs skip:manual and steps aside.

Ideally suited for DeepSeek Harness users who bounce between models, want to save tokens without sacrificing quality on hard problems, and dislike plugins that silently rewrite global settings. One toggle right in the composer takes effect on the very next turn, or switches off just as easily. Zero overhead.