Preface

DeepSeek Harness (DSH) supports plugin-based extensions. When debugging agents or complex workflows, relying only on the built-in text or table views to inspect trajectories is often not intuitive enough and makes it difficult to quickly locate tool-calling logic or model decision paths. The TangRj-Git/trajectory-graph plugin adds a Graph tab next to the default Chat and Trajectory views in DSH, projecting structured trajectory snapshots into an interactive left-to-right execution graph.

The plugin runs entirely locally, does not upload trajectory data, and does not modify session history. It is intended to help developers understand execution flow in a graphical way.

Plugin Overview

  • Plugin Name: dsh-plugin-trajectory-graph
  • Repository: https://github.com/TangRj-Git/trajectory-graph
  • Maintainer: TangRj-Git
  • License: MIT
  • Core Value: Convert DSH structured trajectory snapshots into a visualized execution topology graph, with support for node-level detail inspection and interaction.

Core Features

The plugin provides the following capabilities to assist with trajectory analysis:

  1. Turn Navigation and Selection: Supports independently selecting the graph for a specific turn, with the latest turn selected by default. The Turn navigator can isolate a single user request and avoid interference from prior history.
  2. Semantic Node Display: Displays semantic node types, including User, Context, Model, Tool, Tool Group, Subtool, Agent, Final, Retry, and Error nodes.
  3. Tool Call Visualization: Tool calls are presented as vertical branches and converge at later Model or Final nodes.
  4. Interactive Operations: Supports search, type filtering, collapse/expand, zoom, and fit-to-view.
  5. Node Details: Clicking any node reveals structured details. Tool nodes specifically display data such as Payload, Result, Schema, and Timing.
  6. Real-Time Updates: The graph updates in real time while a session is running, and the layout remains stable when streaming text changes.
  7. Privacy and Security: Visualization is entirely local; no trajectory data is uploaded, and no external analytics services are called.

Installation and Enablement

Before installing, make sure your environment meets the following requirements:

  • Node.js 20 or later
  • pnpm 10 or later
  • DeepSeek Harness 0.1.0-rc.6 or 0.1.0-rc.7
  1. Download the dsh-plugin-trajectory-graph-0.3.0.tgz file from GitHub Releases and keep it in a local directory.
  2. Run the following command to install the plugin into the default Web preset (replace with the actual file path):
npx --yes @deepseek-ai/dsh@0.1.0-rc.7 plugin --profile web add "C:\path\to\dsh-plugin-trajectory-graph-0.3.0.tgz"

After a successful installation, restart the session for the change to take effect.

Typical Usage

  1. Start the default Web preset:
npx --yes @deepseek-ai/dsh@0.1.0-rc.7 web --port 3080
  1. Open http://127.0.0.1:3080 in your browser and select a session that contains trajectory data.
  2. In the top tabs, switch to Chat | Trajectory | Graph.
  3. Use the Turn navigator on the left to select a specific turn, or click and drag an empty canvas area to pan, and scroll the mouse wheel to zoom.
  4. Click any node to view information such as Payload, Result, or Timing in the details panel.

Considerations and Limitations

  • Viewer Scope: This is an execution viewer, not an editable workflow designer. It does not modify prompts or session history.
  • Agent Inference Limitation: The current version does not infer Agent branches. Because Harness does not provide persistent identities, the plugin cannot infer Agent paths from tool names or model text.
  • Performance Consideration: With very large history, the graph may become too wide; use turn navigation, search, or collapse features to inspect it.
  • Data Dependency: Sessions without structured trajectory snapshots cannot generate graph nodes.
  • Privacy: The plugin runs entirely locally and does not upload trajectory data.

Conclusion

The TangRj-Git/trajectory-graph plugin addresses a key pain point for DSH users analyzing complex tool calls and model decision paths by providing a visualized execution topology graph. Combined with Turn navigation and node-level detail inspection, it can significantly improve debugging efficiency. For more details and source code, refer to the GitHub repository.