dsh-model-router

In a multi-agent DeepSeek Harness session, every model call defaults to the same tier. A subagent that only runs a single command or fixes a compiler error still pays the pro price tag. dsh-model-router eliminates that structural waste by splitting the session into two roles: the root agent (planner) always runs on deepseek-v4-pro for understanding, design, and review, while every delegated subagent (executor) is pinned to deepseek-v4-flash for code writing, command execution, and iterative debugging. The rewrite sits at the outermost layer of the request pipeline, so it overrides both the harness default and any model you pick in the UI, guaranteeing each role lands on the right tier.

Beyond the core routing, the plugin injects three supporting pieces: a prompt section rendered before the agent persona that reminds the planner to delegate implementation; a loadable pro-flash-routing skill in the session catalog that spells out the plan-delegate-review-report rhythm; and optional extensions for vision routing (auto-switching image-bearing requests to a dedicated vision model) and error-driven escalation (temporarily boosting reasoning after consecutive failures, then wearing off once the agent recovers).

Who is it for? If you run multi-agent workflows in the Harness where subagents make heavy tool calls, and you refuse to pay pro prices for every flash-level execution step, this plugin is built for you. Setup is frictionless: it is active out of the box, a single GUI toggle switches the whole router on or off instantly, and per-role reasoningEffort and maxTokens knobs let you squeeze even more token savings without touching the baseline quality of your planning passes.