In DeepSeek Harness (DSH), a single agent handling complex tasks often encounters serialization bottlenecks, leading to long task queues, full context windows, and quality drift. dsh-swarm-orchestrator is designed to address these issues by turning a goal into a supervised agent team: the architect reviews and consolidates the plan into PLAN.md, parallel builders execute the task DAG, reviewers gate quality, and the integrator merges and delivers the results. The entire process is visible on a live Kanban board.

Core Features

  • Role-model binding and fallback chains: Each role (architect, builder, reviewer, integrator) can be bound to a specific model. The plugin supports model fallback chains and reasoning-effort ladders to ensure task completion.
  • Parallel task DAG execution: Tasks are defined as a directed acyclic graph (DAG) with dependencies (blockedBy). Independent tasks run in parallel, subject to a global concurrency limit.
  • Review loops with feedback: Tasks flagged for review are evaluated by reviewer agents. Rejected tasks loop back to builders with feedback. Review gating can be set to human for manual approval.
  • Live Kanban board: Provides a Live Kanban Dashboard for tracking the running status and task progress of each chat workspace.
  • Evidence checks: Supports machine verification of deliverables through evidence commands or file checks.
  • Memory-safe concurrency: Agents run as in-process sessions within the DSH host. A global limit (maxTotalConcurrentAgents) prevents heap memory exhaustion.
  • Scoped workspace execution: The Swarm tab for each chat tab shows the run status for that workspace.
  • One run per goal: Unless the user opts out, building will not start unless the plan has been reviewed by the architect.

Typical Usage

  1. Dispatch a goal: The user assigns a group goal to the chat agent in the chat interface.
  2. Architect review: After the agent generates a plan, the architect agent reviews and integrates it into PLAN.md.
  3. Parallel building: Parallel builder agents execute the task graph based on task dependencies (DAG).
  4. Quality control: Reviewer agents evaluate deliverables, reject nonconforming results and provide feedback, or set the gate to human to wait for manual approval.
  5. Integration and delivery: The integrator agent merges all parts, validates the result, and generates the final report.

Use Cases and Considerations

  • Runtime environment: This plugin runs as an in-process session in the dsh web host, without starting a daemon or secondary process.
  • Concurrency control: A global concurrency limit exists (default is 8) to prevent heap memory exhaustion. This limit can be adjusted through maxTotalConcurrentAgents in Runtime settings.
  • Tool access: By default, modlens tools are denied to swarm roles. If needed in a headless environment, add them back by role in the Roster.
  • Evidence commands: Evidence commands are run under PowerShell (in the root directory).
  • Task status: Adopted tasks must have status "completed".
  • License: MIT.

Summary

dsh-swarm-orchestrator provides DSH with a structured multi-agent orchestration layer. It addresses the limitations of a single agent by ensuring parallel workflows, model assignment, and machine verification. For installation and documentation, please visit the GitHub repository.