dsh-multi-model-orchestrator

When a single complex task demands deep reasoning, code generation, and high-volume templating all at once, throwing everything at one model is neither economical nor fast. dsh-multi-model-orchestrator turns your dsh primary model into a multi-model main brain: it decomposes the task into subtasks, reads the per-route strengths and weaknesses you have described in config, and dispatches each subtask to the best-fit model route (GLM, Kimi, Qwen, or any OpenAI-compatible endpoint) in parallel within a single conversation, then synthesizes the results.

Beyond routing, the plugin tracks per-model token usage: input, output, cache-read, and cache-write tokens plus request counts accumulated per provider/model since the process started, so you always know where compute is going. All you need to do is define your routes and model notes in settings.yaml; from then on the main brain handles the decompose-dispatch-aggregate loop autonomously.

It is designed for dsh users who have configured two or more OpenAI-compatible routes and want a single prompt to fan out across models while keeping granular visibility into per-route cost. The plugin does not touch harness internals; every route and allow-list entry is a plain config value, and removing the plugin restores your setup exactly as it was.