Introduction

When using DeepSeek Harness (DSH) for multi-agent development, raw session logs are usually presented as flat JSON or lists. When sub-Agent calls, tool chains, and long-running turns are involved, developers have difficulty intuitively understanding the chronological order and hierarchical relationships of execution flows. The dsh-seelog plugin converts the current session log into a browsable topology graph, visually reconstructing the agent execution process.

What Is This

dsh-seelog is an external Web plugin for DeepSeek Harness, maintained by developer lhwu1. It does not modify DSH source code or session format. Instead, it is loaded via a “bundle” mechanism, reads the current session log on demand, and converts it into a topology structure with a time dimension. The plugin aims to solve problems such as unclear logic flow in multi-agent scenarios and difficulty quickly locating specific events.

Core Features

The plugin provides a complete set of visualization and interaction mechanisms:

  1. Chinese semantic node mapping
    Maps inputs, model responses, tool calls, rounds, failure events, and sub-Agents to semantic nodes rather than raw JSON fields. Unknown tools are displayed as “tool call” or “MCP tool call” for quick recognition.

  2. Unified timeline
    Places sub-Agents and the main Agent on the same timeline. Sub-Agents branch from the actual launch position on the main line and merge back into the main line after execution completes. This design avoids parent-child relationship confusion caused by flat lists.

  3. Navigation and interaction

    • Fixed time overview ruler: Located at the top of the main graph, it maps the entire captured session. A cyan window indicates the currently visible range of the main graph; clicking a tick or track position allows smooth navigation.
    • Node inspector: After clicking a node or Chinese label, the right-side inspector displays the semantic name, category, status, time, and log index. When expanded, it can show the full JSON and adjacent events.
    • Main execution line and knots: A thick purple line represents the root session main line. Nodes are connected by diamond-shaped knots to label cards, and labels move with the view as it is dragged.

  4. Performance and limits
    To prevent browser lag caused by large logs, the plugin sets rendering limits: each lane renders at most 400 interactive nodes.

Installation and Enablement

Installation depends on a Node.js environment. Using Windows paths as an example, the steps are as follows:

  1. Enter the plugin directory, install dependencies, and validate the build:
    cd D:\harness-lh-int8\dsh-seelog
    npm install
    npm run validate
  1. Link the external plugin to the DSH profile that runs the Web UI:
    dsh plugin --profile web add D:\harness-lh-int8\dsh-seelog
    dsh --profile web --dump-config
  1. Restart the profile. The Session Graph tab will appear in the tab bar.

Typical Usage

After installation and restart, perform the following operations on any session page:

  1. Open the view
    Switch to the Session Graph tab. The plugin automatically reads the current log and generates a topology graph.

  2. Browse the execution process

    • Overview ruler navigation: Click a time tick at the top, and the main graph jumps to the corresponding time range.
    • Main graph operations: Use the mouse wheel to scroll horizontally, and press and hold the left button to drag and pan. After releasing the drag, the plugin automatically requests the current log in the background and incorporates new events into the graph; the page does not refresh and occupy the canvas.
    • View details: Click the main line, a sub-Agent lane, or a Chinese label. The right-side inspector loads the semantic information and raw log for that node.

Applicable Scenarios and Notes

  • Applicable scenarios: Suitable for developers who need to analyze multi-agent collaboration workflows, debug tool call failures, or organize complex session logic.
  • Privacy and data: The plugin only reads lightweight semantic fields from the current session and does not include model content, prompts, or tool parameters. Only after clicking a node does the browser request that event and the two adjacent raw logs.
  • Data persistence: It does not save local browser snapshot history. After leaving or refreshing the page, you must re-enter the view or drag the view to read logs on demand.
  • Configuration limits: Under the default configuration (cordis.patch.yml), the plugin reads at most 128 lineage sessions, and each session has at most 20,000 log events.

Conclusion

dsh-seelog visualizes complex agent execution processes through a topology view. For developers accustomed to graphical debugging, it can significantly improve the efficiency of issue investigation. For more details, see its GitHub repository or community catalog.