DeepSeek Harness (DSH) execution is usually presented as logs or trajectory streams. When dealing with complex task chains, quickly locating key steps, error points, or the steps that dominate execution time becomes difficult. The dsh-task-graph plugin turns a single DSH task execution flow into a graph, transforming an Agent’s internal runtime state from a “running account” into an interactive topology diagram.

Plugin Introduction

This plugin is maintained by KevinZhangNothing and belongs to the workflow category. It takes “Task” as the first-level concept, and after entering a task, it first displays the execution graph rather than a long list of logs.

Core Features

  • Task First and Graph First: The graph is parsed from underlying Trajectory data and focuses on task flow.
  • Detail on Demand: The canvas displays only key nodes, while specific execution details (LLM steps, tool calls, inputs and outputs) are moved to a detail panel on the right.
  • Key Node View: By default, the canvas only shows Turns (turns), code changes, Plan/Skill/SubAgent/testing, and failure or retry events. Ordinary steps are hidden to keep the view clean.
  • Multiple Layouts: Supports four layouts—serpentine (default), DAG, timeline, and force-directed—to accommodate different analysis needs.
  • Real-time Execution Status: Incremental refresh via SSE provides second-level timing and LIVE badges on nodes, making the running state clear at a glance.
  • Graph ↔ Trajectory Bidirectional Linking: Clicking a graph node jumps to the trajectory details, and clicking a trajectory event back-selects the graph node.
  • Analysis and Localization: Supports viewing the critical path, error localization, retry aggregation, and task summaries (duration, Token, high-frequency tools, etc.).

Installation and Enablement

  1. Run the installation command:
    dsh plugin --profile web add github:KevinZhangNothing/dsh-task-graph
  1. After installation, restart the engine (or reload the Web UI).
  2. A third “Graph” Tab appears next to the “Chat / Trajectory” tab; click it to enter.

Typical Usage

  • Standalone Demo Test:
    npm run demo

Open the browser to access the Demo page, view built-in data, or load local `$DSH_HOME` sessions.

Technical Details and Notes

  • The server requires Node >= 22.15.
  • It uses the built-in zlib.zstdDecompressSync to decode .zstd session files.
  • The plugin has zero native dependencies.

By turning the abstract execution process into a concrete representation, this plugin helps developers debug and analyze more efficiently. For more details and source code, you can visit the plugin directory or the GitHub repository.