dsh-routed-subagent

In multi-model workflows, subagent delegation often suffers from a lack of precision: you cannot dictate which model handles a given child task or cap how much reasoning effort it expends. dsh-routed-subagent fills that gap by introducing allowlisted model-and-reasoning routing to DeepSeek Harness, requiring every delegation call to select a preconfigured model policy and an exact effort level, so arbitrary provider or model passthrough is never accepted.

Under the hood, the plugin validates the model and effort pair against local policy and ctx.llm.resolveCallConfig() before any child is spawned. It exposes fine-grained controls such as maximum delegation depth, a per-policy list of models children may further delegate to, output-token caps, and foreground or background execution toggles. All of these are editable through the built-in Settings card; saving hot-remounts the tool and automatically revokes recursive authority held by children from the previous settings generation. Every invocation result echoes the effective provider, model, and reasoningEffort for easy auditing.

If you are building multi-model pipelines in DeepSeek Harness and need deterministic routing, enforced reasoning budgets, or auditable delegation rules for a team, this plugin is the control layer you have been missing.