Introducing the DeepSeek Harness dsh-orchestrate Plugin

Introduction¶

The philosophy of DeepSeek Harness (DSH) is “everything is a plugin.” The official system already provides complete subagent low-level capabilities (such as spawn, continuable, toolFilter), but the model currently only has the ability to “delegate,” lacking an orchestration layer for deciding when to actively delegate, how to break down tasks, and how to wait for convergence. dsh-orchestrate is a third-party plugin built on top of the official subagent capability family. It fills the gap in the decision layer, orchestration pattern layer, and convergence layer, enabling the main agent to proactively and actively schedule sub-agents to complete complex workflows.

Installation and Enablement¶

This plugin depends on the official subagent capability family. When installing, the corresponding provider version must be specified. The following command installs the verified combination from npm:

dsh plugin --profile demo add \
  @deepseek-ai/dsh-subagent-spawn-in-process@0.1.0-rc.6 \
  dsh-orchestrate@0.3.1

After installation, the plugin automatically injects orchestration decision capabilities. To disable automatic injection, set auto: false in the configuration.

Core Capabilities¶

The plugin provides the following orchestration tools for task decomposition, review, exploration, convergence, and management:

• orchestrate_delegate: Delegate tasks in parallel. Supports a single task or an array of task briefs, supports referencing predefined task briefs from a template library, and supports background execution (run_in_background) and convergence modes (all/any).
• orchestrate_review: Adversarial review. Uses multiple read-only sub-agents to critique a task proposal and returns a mechanical summary of risks, assumptions to challenge, boundaries, and recommendations.
• orchestrate_explore: Diverse exploration. Explores along independent directions and returns a findings report with confidence levels.
• orchestrate_converge: Converge background orchestration. Blocks by runId until sub-sessions complete and returns a summary.
• orchestrate_status: Query task and sub-session status.
• orchestrate_stop: Abort an in-progress orchestrated task.
• Template library: Built-in 10 task brief templates (such as research comparison and code review), support parameterized invocation via the template parameter.

Usage Examples¶

In DSH interactions, the main agent can directly express parallel requirements. For example:

并行调研这三个方向：① SQLite vs DuckDB 存储，② RAG chunk 策略，③ 索引方案。

The agent will invoke orchestrate_delegate and return a structured summary. For code review scenarios, you can directly invoke a template:

调用 orchestrate_delegate，template=code-review-multi，templateArgs={ module: ['src/a', 'src/b'] }

Configuration and Limitations¶

Configuration Items¶

Usage Limitations and Notes¶

• Third-party plugin: This plugin is not officially released. Check the source code and license before installation.
• Permissions: It relies on the default read-only sandbox in official deployments. If the deployment is changed to workspace-write, sub-agents inherit write permissions.
• State persistence: Orchestration state is stored in-process and must be re-orchestrated after restarting.
• Cost estimation: costTokens is an estimated value, not an exact official count.
• Tool isolation: Sub-agents automatically hide orchestrate_* tools to prevent recursive calls.
• Task decomposition: The path in the task parameter is not automatically split. It must be manually converted into an array of task briefs.

Conclusion¶

dsh-orchestrate engineers multi-agent collaboration methodologies through templated and standardized tools, making it suitable for scenarios that require parallel processing, multi-perspective review, and decomposition of complex logic. To view more templates or contribute code, visit:
* Plugin Catalog
* GitHub Repository