deepseek-harness-orchestrate

When a model faces a complex goal it naturally wants to decompose the work into subtasks with explicit ordering dependencies and then execute them step by step. Until now, DeepSeek Harness lacked a structured entry point that let the model declare an entire task graph in a single call, have it validated, and then executed safely. The orchestrate plugin fills that gap: it exposes an orchestrate_tasks tool so the model can supply a task array, dependency edges, and optional per-task provider, model, and outputSchema descriptions all at once. Before anything starts running the graph is fully validated for acyclicity and bounded size.

Once validation passes the graph is handed off to the existing ctx.workflowEngine and executed in deterministic topological layers. Each completed child returns either plain text or a structured value; a failed child causes its direct downstream tasks to be marked skipped with a deterministic blockedBy list while unaffected branches continue. Final results preserve the original task-array order and carry an aggregate status of completed, partial, or failed. All model-authored identifiers, prompts, dependency IDs, and upstream results travel strictly as JSON data and are never interpolated into JavaScript source, keeping the injection surface closed.

This plugin is well suited for developers building multi-step agent pipelines on DeepSeek Harness who want the model itself to decide the task breakdown and dependency structure rather than hard-coding each step in orchestration code. The host deployment must already provide ctx.workflowEngine and a subagent provider. Credentials are inherited from the existing LLM route and the plugin stores no API keys of its own.