Introduction

The core design philosophy of DeepSeek Harness (DSH) is “everything is a plugin.” When building multi-agent collaboration systems, developers often face issues such as chaotic task decomposition, uncontrolled expert delegation, and unpredictable delivery quality. Traditional single-agent or simple chain-call approaches struggle to handle complex projects.

The dsh-expert-orchestrator plugin provides an “expert orchestration” session mode. It introduces PM-first planning, a dependency DAG task board, and tiered delivery gates, delegating control to the orchestration layer and ensuring that implementation is always completed through expert delegation rather than self-implementation by agents.

Core Features

  1. Per-turn triage and anchor mechanism: Every user message re-enters the triage loop. Before any action is executed, a self-check with five anchors is run (review, convergence, drift prevention, collaboration, resources) to prevent task drift across turns.
  2. PM-first planning: Complex tasks are first assigned to a project management expert for decomposition, dependency mapping, and acceptance criteria definition, and then delegated.
  3. Expert delegation and merged roster: Implementation is always delegated. The system maintains a merged expert roster (including core experts and enabled source experts), invoked via summon_expert or summon_experts. Recursive protection restricts subagent behavior through a six-item blacklist.
  4. Dependency DAG task board and message bus: Uses the taskboard.py tool to manage the dependency graph, with the message bus enabling disk persistence and summarized replies for parallel experts.
  5. Tiered delivery gates: Before delivery, work must pass independent review (reviewer ≠ implementer, at most 2 rounds of rework) and a PM checkpoint before submission/delivery.
  6. Experience pool and prompt distillation: Each task captures no more than 3 reusable experiences to inject into future tasks. Supports expert-level prompt distillation (method prefix) and on-demand deep-read pointers.
  7. Safety and fault tolerance: v2.8 introduces delegation safety gates (source-chain marking, artifact ownership registration, write locks); v2.7 implements event-sourcing-based crash-safe orchestration state.

Installation and Enablement

The plugin is integrated as a directory. Place the dsh-expert-orchestrator directory under the DSH plugin loading path, or add it through the DSH plugin manager. Ensure the DSH version satisfies the engines.dsh requirement (>=0.1.7-alpha.2 <0.2.0-0 || >=0.2.0-rc.1 <0.3.0-0 || >=0.2.1-0 <0.3.0-0).

Typical Usage

The plugin provides the following interaction modes:

  1. Browse the expert roster
    list_experts

Supports compact and expanded modes, lists experts in the merged roster, and marks cross-source conflicts and shadowed entries.
  1. Summon a single expert
    summon_expert

Suitable for summoning a single white paper or expert in a single domain. Task size is limited to 8K characters, and a sanitized persona is automatically injected.
  1. Summon experts in batch
    summon_experts

Supports batch summoning, with quantity ≤8 and concurrency 4. Provides partial-success semantics.
  1. Task board and incremental reading
    taskboard.py -- (structured envelopes)
    taskboard.py --since-seq

Uses the dependency DAG task board to process structured envelopes, and supports incremental reading since the last sequence number.
  1. Configuration and switches
    Enable tool budget and automatic idle-edge reclaim (recommended to enable both):
    export DSH_EXPERT_TOOL_BUDGET=1  # 启用工具调用硬预算
    export DSH_EXPERT_IDLE_RECLAIM=1 # 启用空闲边缘自动回收

Applicable Scenarios and Notes

This plugin is suitable for complex development scenarios that require strict workflow management, multi-role collaboration, and delivery quality control.

Notes:
- Permissions and Security: The plugin runs with the current DSH process permissions. Ensure the installation source is trusted (license is MIT).
- Recursive Protection: The subagent tool filter includes a six-item blacklist to prevent infinite nesting.
- UI Pending: The “Clear Deleted” button UI in the settings panel is not yet implemented (the RPC protocol is in place).
- Idle Loop: Automatic idle-edge reclaim is a zero-token operation, but it is visible in the event stream and does not consume model quota.

Summary

By decoupling and standardizing planning, delegation, and delivery, dsh-expert-orchestrator provides DSH with a multi-agent orchestration solution. It does not directly implement code; instead, through PM planning and expert delegation, together with an event-sourced task board and gate mechanisms, it ensures reliability and quality of task delivery.

Plugin Directory | GitHub Repository