dsh-delegation-workflow

In a DeepSeek Harness multi-step workflow, the main agent plans, delegates sub-tasks to sub-agents, and then reviews their output. The built-in subagent tool, however, does not let the caller explicitly choose which LLM model or reasoning effort the sub-agent should run on. dsh-delegation-workflow fills that gap: it adds a delegate_with_model tool so the main agent can specify the provider, model, and reasoning_effort for every delegated step, and even run it in the background to receive a manageable job id.

On the interaction side, the plugin ships a /delegate-model slash command for the Web UI—pick a model, then a reasoning level, and the choice is persisted locally in settings.yaml. The General Settings panel also gains a dedicated sub-agent default model entry with separate selectors for model and reasoning effort, showing the provider name next to each model. Reasoning levels are validated against the target model at runtime, so an invalid tier is rejected before any request is sent.

If you regularly orchestrate multi-step tasks in DeepSeek Harness and want to match different sub-tasks to different capability tiers—say, high reasoning for planning steps and low reasoning for simple lookups—this plugin adds exactly the knobs that matter at the delegation boundary.