In the development workflow of DeepSeek Harness (DSH), understanding how the AI executes step by step and produces results is essential for debugging prompts and optimizing model behavior. Although DSH’s built-in “Trajectory” tab records events, it is primarily a plain-text list, making it difficult to intuitively grasp the overall time flow or quickly locate the specific context of a request. The dsh-trajectory-traceview plugin renders this data source into a visual horizontal timeline, making every operation and each raw payload sent to the LLM clearly visible.

Features

The plugin adds a new “Trajectory View” tab next to the “Trajectory” tab in the DSH Web GUI. It renders a horizontal timeline with a minimap overview at the top and step details and raw request panels at the bottom.

  • Timeline and Overview: The timeline is ordered by actual clock time. The minimap at the top uses colors to distinguish site types (user=blue, step=green, tool=orange, compression=purple, error=red, retry=yellow, command=cyan). Clicking a color block in the minimap jumps to the corresponding site.
  • KPI Statistics: The top displays the number of turns, steps, tool calls, errors, total tool duration, request and generated token counts, and the current model name.
  • Search and Playback: Press / to focus the search box. Enter a tool name, parameter, content, or error name to filter; matched items have a gold outline. Keyboard shortcuts (◀/▶) and the playback button support step-by-step browsing or automatic playback at supported speeds (1x/2x/4x).
  • Step Details: Click any site to display the reasoning content, tool parameters, result, sub-call chain, duration, and token usage in the bottom panel.
  • Raw Request Analysis: Press r to switch to the “Raw Request” tab at the bottom and view the complete context sent to the LLM for that request, including the system prompt, tool catalog, and reconstructed message list. The panel also shows prompt changes (added/removed tools) and token usage breakdown.
  • Export: Use the Markdown icon in the header toolbar to export the entire trajectory as a Markdown file.

Installation and Activation

Install the plugin with the following command:

dsh plugin --profile web add github:AFAP/dsh-trajectory-traceview

After installation, restart dsh web to activate it.

Typical Usage

  1. Open any session containing an AI execution trajectory.
  2. Click the new “Trajectory View” tab.
  3. Browse the timeline and click a site to view reasoning and tool call details.
  4. Use / to search for a specific tool name or error message.
  5. Press r to view the detailed context of the current request for prompt learning.
  6. Use the 📋 button in the header to export the entire trajectory as Markdown.

Notes

This plugin is a read-only tool. It only reads the TrajectorySnapshot snapshots already generated in DSH Web sessions and renders them. It does not modify any files, access the network, or write to local storage. All data processing is performed locally in the browser, with no persistent storage.

For more information, see the project repository: https://github.com/AFAP/dsh-trajectory-traceview