dsh-task-models

By default, DeepSeek Harness subagent and subagent_fork inherit the current route of the caller for model selection and reasoning effort, leaving no way to override per invocation. dsh-task-models extends both tools with optional model and reasoning_effort parameters, letting every subagent call independently specify a provider/model route and an effort id without touching preset or host configuration.

The plugin introduces a clear priority chain: omitting both arguments inherits the caller route; specifying model only selects that route with its default effort; specifying effort only overlays it on the current route; passing default clears any explicit effort from fork history and re-resolves the model default. All parameters are validated via ctx.llm.resolveCallConfig() before a Job, continuable session, or one-shot child is created. A task_models discovery tool also lets callers list available models and their supported reasoning efforts at runtime.

This is ideal for Harness users who need to switch models or reasoning effort across different subtasks. The plugin fully preserves native behavior: background continuable versus one-shot scheduling, fork history, provider add/remove and HMR lifecycle, maxDepth, persona, toolFilter, and system-prompt semantics, so existing subagent and subagent_fork calls can be upgraded simply by adding a model or reasoning_effort field.