dsh-model-orchestrator

In multi-model deployments, the biggest hidden cost is rarely compute — it is the repeated manual judgment of which task should go to which model. dsh-model-orchestrator removes that burden: at runtime it discovers every model actually mounted in your Harness, builds a capability profile from verifiable facts such as input modalities, context window, and reasoning depth, then deterministically routes each unit of work to the best-matching model. You simply describe what you want done; the orchestration takes care of the rest.

Three design choices stand out. First, the selection logic names no model, vendor, or business domain — capability descriptors are open-ended, and unrecognized domains mint new descriptors from the task's own vocabulary, so the taxonomy genuinely grows over time. Second, it works through the subagent seam: your session agent remains the captain responsible for understanding, decomposing, dispatching, and aggregating, while the orchestrator only decides which model executes each unit and never interferes with the Harness's native task management. Third, when a hard requirement cannot be evidenced, the model is rejected outright rather than silently downgraded, keeping routing decisions trustworthy.

It fits teams whose Harness hosts multiple providers or multiple tiers from the same provider, those who want to split complex multi-domain work across specialists in parallel, or anyone who simply does not want to hand-pick routes every session. Routing is automatic by default and can be steered explicitly via the /model-orchestrator command or the settings panel, making it a good fit for developers who want automation without giving up control.