DeepSeek Harness (DSH) is responsible for executing tasks, but when handling long conversations or complex workflows, the raw log interface often makes it difficult to quickly locate key information or explore branches. dsh-atlas provides a visualized conversation workspace that transforms DSH session logs into cards and branches, making long sessions easier to understand, organize, and continue.

Positioning and Principles

dsh-atlas is a client-side plugin that runs in the DeepSeek Harness Web environment. It uses DSH session logs as the single source of truth. It does not duplicate or rewrite the original chat system; instead, it is built on top of DSH as a visual layer to help developers organize conversation structure.

Installation and Enablement

Installing the plugin requires a Node.js environment. Run the following commands to add the plugin to the DSH Web configuration, then start the Web service:

npx @deepseek-ai/dsh plugin --profile web add github:sumarilkkxx/dsh-atlas
npx @deepseek-ai/dsh web

DSH Web starts by default at http://127.0.0.1:3080. After configuring the model in settings and selecting a workspace, click Card View in the top navigation bar to enter the visual workspace.

Core Features

Conversation Visualization and Organization

The plugin projects user messages, model responses, and tool activity into cards. Each primary conversation has an independent canvas. On the canvas, you can drag cards, pan the view, or zoom between 50% and 400%. You can also automatically arrange nodes, locate active cards, or collapse downstream cards.

Branching and Real-Time Status

Each turn of the conversation generates a card. Branching from any answer uses native DSH Fork while preserving context relationships. Cards display preparation, reasoning, tool calls, and generation progress in real time. Once complete, the persisted DSH history becomes the authoritative state, preventing stale “Thinking” states.

Tooling and Performance Metrics

The detail view preserves the original order of tool events and embeds ECharts, Mermaid, and a sandboxed rendering canvas. Each card aggregates LLM time, time to first token, token rate, cache hit rate, and input/output token counts, without exposing a detailed breakdown of individual model calls.

Native Input Capabilities

Atlas reads DSH’s complete model catalog and supports native selection of models, reasoning levels, permission presets, slash commands, files, and skills. New sessions or follow-up conversations initiated from the canvas have the same session-level permissions and context capabilities as native DSH.

Information Organization

You can search sessions and cards to organize long-running work. Cards can be marked as Important, Key Conclusion, or Needs Verification to quickly identify key conclusions or content that requires verification.

Usage Examples

  1. Browsing and Operating: After starting DSH Web, click Card View in the top navigation bar. On the canvas, drag cards, pan the view, or use the mouse wheel to zoom, ranging from 50% to 400%.
  2. Branching Conversations: Starting from any answer card, use the native DSH Fork feature to create branches and continue exploring different conversation paths.
  3. Executing Commands: Select a slash command (such as /compact) through the menu or keyboard and send it. The system will generate an independent command card.
  4. Marking Key Points: Mark completed cards as Important, Key Conclusion, or Needs Verification for quick review later.

Data and Limitations

Data Ownership

dsh-atlas stores only auxiliary data (layout, status, and summaries). DSH owns the original execution history and conversation records. Deleting cards on the Atlas canvas does not affect the original DSH conversation.

File Parsing Limitations

Attachments are parsed locally in the browser. PDF, DOCX, XLSX, CSV, and common code and configuration files are supported. The following limitations apply:
* Single file size: 12 MB
* Total file size: 24 MB
* Extracted characters per file: 48,000 characters
* Total extracted characters: 96,000 characters

Dependency Note

Charts and visual artifacts in conversation details are generated by the optional dsh-artifact plugin. Atlas is only responsible for mounting and displaying these tool outputs.

Conclusion

dsh-atlas transforms raw conversation logs into a manageable card workspace, maintaining close integration with the core system through native DSH Fork and real-time status. It is suitable for developers who need to organize logic, structure tool output, or perform branch experiments during long sessions.

Plugin Directory | GitHub Repository