Introduction

The plugin mechanism of DeepSeek Harness (DSH) allows extending functionality outside the host environment. dsh-dev-crew is a workflow plugin whose core role is to distribute work to subagents bound to different models by responsibility within the same session. A common use case is assigning requirement organization and review to high-capability models and bulk implementation to low-cost models, so complex development tasks can be completed within the same conversation flow.

Core Features

The plugin mainly addresses multi-role collaboration and model routing configuration:
* Multi-role Subagents: Supports defining multiple roles, each of which can be bound to one or more models.
* Built-in Roles and Skills: Provides the three built-in roles implementer (Implementer), reviewer (Reviewer), and researcher (Researcher), and embeds four methodology skills: crew, crew-brainstorm, crew-plan, and crew-converge.
* Configuration: Supports configuration through the Web interface or YAML files, and supports hot updates (some configuration items require a reload).
* Artifact Management: Automatically creates and manages artifact directories such as specs, plans, reviews, and reports.

Installation

Before installing, ensure the host DSH version is 0.1.0-rc.7 or newer. The plugin does not depend on DSH’s built-in package declaration. Use the following installation command:

dsh plugin --profile <你的 profile> add dsh-dev-crew

To install from a local path:

dsh plugin --profile <你的 profile> add <本仓库绝对路径>

Configuration

The plugin supports both Web interface and YAML configuration. In the Web interface settings page, under the Dev Crew section, you can configure roles, models, convergence rounds, and discipline gate switches.

For YAML configuration, define roles in the profile’s cordis.patch.yml. Each role can be bound to multiple models, and each model corresponds to an independent delegation tool.

- id: dsh-dev-crew
  config:
    roles:
      - id: implementer
        enabled: true
        models:
          - alias: default
            provider: kimi-coding
            model: k3
      - id: reviewer
        enabled: true
        models:
          - alias: ds
            provider: deepseek-official
            model: deepseek-v4-flash

Tool naming follows the rules below:
* Single-model role: subagent_<role_name>
* Multi-model role: subagent_<role_name>_<alias>

Embedded Methodologies and Artifact Directories

Embedded skills are shipped with the plugin package and require no user-side configuration. The main skills include:
* crew: end-to-end development pipeline.
* crew-brainstorm: requirement discussion and spec finalization.
* crew-plan: decomposing specs into implementation plans.
* crew-converge: review convergence protocol (visible only to models).

Use the /crew-init command or the crew_init tool to create artifact directories (default: docs/specs, plans, reviews, and reports). The command is idempotent and does not overwrite existing files.

Subagent Result Return Mechanism

Subagent results are returned through two channels:
1. Completion Notice: After a subagent finishes, the host returns a message beginning with Background subagent <id> finished, which includes the subagent’s last message. This is the primary indicator for determining whether the subagent has completed its work.
2. Report Tool: A subagent can proactively call the report tool to send content. This is not limited to the last message and is suitable for interim reports or longer conclusions.

Note: The job_output, job_list, and job_kill tools are used to manage background shell jobs and are unrelated to subagents.

Notes

Keep the following restrictions and rules in mind when using the plugin:
* Date Accuracy: Dates in artifact file names may be inaccurate because they depend on the model’s training data.
* File Modification Restrictions: The reviewer role can only modify files using edit or cp+edit; it cannot use write (except for saving a review report).
* Gate Hot Updates: gate.enabled only takes effect when the plugin is mounted; hot updates require a reload.
* Goal Loop Conflicts: The goal loop must pause after a subagent completes; otherwise, the pipeline can get stuck as goal_round messages continuously reset it.
* Result Isolation: Subagents cannot see other delegation tools; the plugin automatically adds them to the deny list.