Introduction

DeepSeek Harness (DSH) extends its capabilities through a plugin mechanism. Building complex, stateful workflows requires solving engineering problems such as state management, evidence tracing, and human feedback.

Plugin Overview

Name: captain-dsh-plugin
Maintainer: Angel2518975237
Positioning: A resumable, multi-agent DSH workflow plugin for job-seeking company screening.
Problem Addressed: Provides a resumable, verifiable, and reproducible agent workflow for handling career direction and company screening. Through multi-agent collaboration and evidence tracing, it turns vague job-seeking decisions into item-by-item traceable reports.

Core Capabilities

The plugin implements the following capabilities:
- Agent Capabilities: Agentic Workflow, Agent Skill, Multi-Agent Orchestration, Agent Task Contracts, Agent Reviewer, Concurrency, Dynamic Planning
- Engineering and Security: Context Engineering, Single CLI Entry Point, Checkpoint Restoration, Versioned Long-Term Memory, State Routing, Idempotency, Deduplication/Conflict Checks, Permission Guardrails, Privacy Guardrails, Budget Guardrails, Deterministic Rules, Boundaries Between Agents and Deterministic Programs, Guardrails, Evals
- Data Processing: LLM Understanding of Unstructured Content, Structured Output, Prompt Engineering, Source Tracing, Business Modeling, Entity Resolution, Long-Term Memory, Time Awareness, Multimodality
- Other: Tool/MCP, Human-in-the-loop

Installation and Enablement

The plugin code is hosted in a GitHub repository. Make sure the required dependencies are installed:
- @deepseek-ai/cordis (4.0.2)
- @deepseek-ai/dsh-tools (0.1.1-rc.2)
- react (18.3.1)

Enable it via the “Captain Console” action in the DeepSeek Harness sidebar.

Usage Examples

  1. Console and Real-Time Progression: After uploading materials, the voyage trajectory advances in real time with the actual tasks.
  2. Conversation and Global Control: The conversation box truly occupies the entire column, while the right-side voyage trajectory continuously follows the state machine.
  3. Material Auto-Fill: Upload existing materials (such as a resume or job preferences) and let the agent fill in information for you.
  4. Template Download: Download a standard template, fill it out, and upload it again.
  5. Version Management: A new version is generated only upon explicit submission.

Important Notes

  • Factual Boundaries: A resume can support only professional facts; it must not be used to automatically infer salary floors, work arrangements, or leadership preferences.
  • Privacy Protection: Sensitive file contents are never written to logs or messages; only opaque source references are quoted.
  • State Management: Progress is never inferred from chat; it is restored from the persisted current_node. Ordinary chit-chat does not needlessly “light up nodes.”
  • Identity Confirmation: Company identity is not resolved here; that is handled in later stages.
  • Entity Disambiguation: Never conclude based solely on “the most well-known company with the same name”; perform lookups only by the confirmed company_id.
  • Evidence Reuse: Evidence that has already been accepted and is reusable is not re-investigated unless review is required.
  • Access Control: Workers have write permissions limited to Draft and cannot review their own results; a failed task does not cancel successful tasks; expansion stops when the budget is exceeded.
  • Inference Labeling: Agent inferences must be labeled as hypothesis and must never be presented as your original words.

Conclusion

This plugin uses multi-agent orchestration and rigorous evidence verification to provide traceable support for career decisions. Related code and documentation can be obtained from the following links:
- GitHub Repository
- SkillHub Directory