Introduction

In the DSH (DeepSeek Harness) plugin ecosystem, the main Agent can easily get stuck in implementation details or lack control over model selection when assigning work. The dsh-agent-lead plugin is designed to solve this problem. It locks the main Agent’s responsibilities to planning, decomposition, dispatch, and acceptance, delegates concrete code-change work to background subagents, and forces the user to select a model during every assignment.

What It Is

This is a workflow-type plugin used to build a “lead mode.” The main Agent only supervises and does not directly perform operations such as writing files. All change-related work must be assigned to background subagents through the assign tool. The plugin controls the mode via the /lead command, and every assignment triggers a popup for the user to select the model used by the subagent.

  • Maintainer: JunguangJiang
  • License: MIT
  • Installation location: ~/.dsh/plugins/agent-lead/

Core Features

  1. Role Division: The main Agent is only responsible for understanding, decomposition, dispatch, acceptance, and supervision. Calls to write / edit / str_replace_editor by the main Agent are rejected and replaced with assign.
  2. Model Selection: The assign tool has no model parameter; the user must manually select the model in the popup. The system provides candidates based on the current main model, frequent usage history, and the configured model list.
  3. Command Control:
    • /lead: Toggle the lead mode switch.
    • /lead off: Force exit from lead mode.
    • /lead auto: Enable fully automatic assignment (assign does not show a popup and automatically uses the last selected model).
    • /lead manual: Restore manual model selection (assign shows a popup every time).
  4. Responsibility Scope (scope): Every assign must declare scope (summary and paths), serving as the subagent’s “territory contract.” The system supports path overlap warnings to prevent multiple subagents from operating on the same file.
  5. Queue and Status: Supports manage_queue for managing pending messages, and lead_status for viewing real-time status, territory map, and assignment ledger of all subagents.

Installation and Enablement

After the plugin is installed, its files are located in ~/.dsh/plugins/agent-lead/. It is mounted by modifying the insert line in the host configuration file ~/.dsh/cordis.patch.yml, taking effect for all agent presets.

Add the configuration to cordis.patch.yml:

    - id: agent-lead
      name: '/home/jiangjunguang/.dsh/plugins/agent-lead/index.mjs'
      config:
        llmProvider: hfai
        rulesFile: /home/jiangjunguang/.dsh/captain-rules.md
        models:
          - id: anthropic/claude-opus-4.6
            label: Opus 4.6
            description: 强，贵，适合复杂实现与重构
          - id: deepseek-v4-flash
            label: DeepSeek-V4-Flash
            description: 快，便宜，适合批量与机械改动

Typical Usage

Basic Commands

Enter lead mode and submit a task:

/lead <任务描述>

Switch the mode state:

/lead              # 切换开关
/lead off          # 强制退出
/lead auto         # 开启全自动派活
/lead manual       # 恢复手动选模型

Combine with task submission:

/lead auto <任务>  # 开启 auto 模式并提交
/lead manual <任务> # 恢复 manual 模式并提交

Tool Calls

Assignment: assign is the only way to create a subagent and must declare a scope.

assign({
  description: "任务简述",
  prompt: "任务详细说明...",
  scope: {
    summary: "负责数据预处理流水线",
    paths: ["src/preprocess/", "tests/preprocess/"]
  }
})

Status Check: You must call lead_status before assigning work to check the territory map.

lead_status        # 查看所有子 agent 状态
lead_status <id>   # 查看特定子 agent

Queue Management: View or operate the pending queue.

manage_queue({
  action: "list"  // list, clear, remove, replace, push, push_front
})

Override Instruction: Use when you need to interrupt the current work or completely change the direction.

redirect({
  subagentId: "xxx",
  message: "新指令...",
  clearQueue: true,
  scope: { ... } // 可选，用于更新职责范围
})

Workflow and Decision Logic

Internally, the plugin uses a “simplified work order” process and executes a decision chain before assignment:

  1. Planning Phase: The Agent surveys the current state, outputs a simplified work order (task name + what to do), and waits for user confirmation.
  2. Decision Chain (Routing):
    • R1 Single Writer: If the files to be modified overlap with the territory of an active subagent, the work must be assigned to that existing agent.
    • R2 Rework Returns to Original Owner: If acceptance fails, return the work to the original owner; if it still fails twice, stop the loss and replace the agent.
    • R3 Independent Review Requires a New Agent: Review/acceptance tasks cannot be assigned to the original author.
    • R4 Same-Domain Continuation Prefers Reuse: Continuations based on completed work should reuse the existing agent and write only incremental changes.
    • R5 Independent New Domain Requires a New Agent: If there is no overlap or true parallelism is needed, create a new subagent.
    • R6 Contamination Means Retirement: If a subagent is redirected more than twice or goes off track, do not assign new work to it; create a successor.
  3. Execution Phase: After user confirmation, choose assign (create new), send_message (append), or redirect (override) based on the decision result.
  4. Acceptance Phase: After a subtask completes, the main Agent reads files and reviews diffs against the acceptance criteria from the issued instruction, then reports the conclusion to the user.

Notes

  1. Human click required for every assignment: This is a design goal; if you do not want to click, use /lead off.
  2. Only assign resumable background subagents: The plugin has no foreground waiting path and is suitable for background tasks.
  3. Subagents cannot assign work again: Subagents cannot use assign to delegate further and must write decisions back to the main Agent.
  4. Data persistence: The assignment ledger is in-memory and is lost after process restart; assignment history is stored in a single file (usageFile).
  5. Scope overlap: During assignment, the system automatically detects path overlaps to prevent multiple agents from modifying the same file.

Conclusion

dsh-agent-lead solves responsibility attribution and resource scheduling problems in complex tasks by enforcing a main Agent supervision mode and a user-involved model selection mechanism. It is suitable for teams that require strict code style, multi-person collaboration, or are sensitive to model cost.