dsh-orchestrate

DeepSeek Harness already ships a complete subagent foundation—spawn, await, structured output—but the model itself lacks an orchestration layer that answers when to delegate, how to split, and how to converge. A single agent can only push long tasks forward serially, sees solutions through one lens, and guesses at multi-direction exploration. dsh-orchestrate adds a decision layer, a pattern layer, and a convergence layer on top of the official subagent API, enabling the root agent to proactively assemble a team of sub-agents for parallel work.

The plugin organizes collaboration around four modes: parallel delegation splits a parent task into a task-brief array and dispatches them concurrently; adversarial review spawns read-only reviewers who probe risks, assumptions, edge cases, and suggestions; diverse exploration sends researchers down independent directions with confidence-weighted findings; and background delegation returns a runId immediately and lets the agent converge all results later. Each sub-agent can assume a role (executor, reviewer, researcher, or devil) with a matching structured-output schema. The orchestration policy injection is visible only to the root agent, teaching the model when to delegate and when not to—sub-agent prompts never see it.

Built for DeepSeek Harness developers who face complex, multi-step work: splitting a large task into parallel sub-tasks, getting a second perspective on a design, or fanning out exploratory research. Ten built-in task-brief templates cover scenarios like comparative research, code review, competitive analysis, and bug triage, and an external template directory can be mounted for extension. Recursion guards, timeout backstops, soft token budgets, and well-defined failure semantics are all included out of the box.