Introduction

When running agents in DeepSeek Harness, a single task may include user prompts, agent visible replies, tool calls, sub-agents, file read or write, external resource access, and the relationships between these events. If understood merely as a linear string of messages, the positions of events, trigger relationships, and sources of final outputs are often not intuitive enough.

Below is introduced a DeepSeek Harness client plugin: Rgndmln/dsh-agent-astra. It projects Harness session events into a traversable spatialized view to inspect key nodes, causal channels, and information hierarchies during the Agent execution process.

What is this

Rgndmln/dsh-agent-astra is a client-side plugin for DeepSeek Harness. Verified materials give the product name as AgentAstra, with the package name and Harness identifier as spatial-trajectory.

The core goal of this plugin is to organize an Agent run into a readable, browsable spatialized trajectory: user messages and final replies remain as readable cards, while tools, sub-agents, files, external resources, as well as read/write, call, delegation, and flow relationships are projected into the same view.

Basic information is as follows:

  • Repository path: Rgndmln/dsh-agent-astra
  • Category: Client-side
  • License: MIT
  • README states it is an experimental interactive plugin
  • Target compatibility: @deepseek-ai/dsh 0.1.0-rc.6
  • Uses Harness client module loader
  • react, react-dom, and three dependencies in package.json are set to latest

Core Features

AgentAstra provides the following verified capabilities:

  1. Projects DeepSeek Harness session events into a traversable spatialized view.
  2. Organizes the execution process using semantic time (Z), causal channels (X), and information hierarchies (Y).
  3. Displays Agent run start/end, user prompts, agent visible replies, tools, sub-agents, files, external resources, and read/write, call, delegation, and flow relationships.
  4. Projects user messages and final replies into readable cards, allowing full Markdown content to be opened.
  5. Supports interactions such as keyboard navigation, timeline dragging, and node click/double-click.
  6. Allows dragging precise file snapshots to the native chat box draft.
  7. Views are bound to sessions, rendering only the currently selected conversation, and releasing the WebGL scene when the view is unmounted.
  8. When WebGL is unavailable, surrounding controls and text fallbacks are retained.

Installation and Activation

First, install the plugin to the web profile. You can use the GitHub short format:

dsh plugin --profile web add github:Rgndmln/dsh-agent-astra

You can also use the HTTPS format:

dsh plugin --profile web add https://github.com/Rgndmln/dsh-agent-astra.git

After installation, open a conversation in DeepSeek Harness and select the Spatial view. The view only follows the currently active session; sessions no longer active will not continue to render in the background.

If you encounter file watching limits during local Harness development on macOS, you can temporarily start it using polling:

CHOKIDAR_USEPOLLING=1 dsh web

Typical Usage

After the above steps, you can view and traverse an Agent execution trajectory as follows:

  1. Select Spatial in a DeepSeek Harness conversation.
  2. Use the keyboard to switch between semantic events or messages:
    - ↑ / ↓: Previous / Next semantic event
    - Shift + ↑ / Shift + ↓: Previous / Next visible user or Agent message
  3. Browse history:
    - Pause
    - Drag timeline
    - Use Home / End
  4. View node details:
    - Single click node: Open floating detail card
    - Double click node: Focus camera on that node
  5. Put file content into the chat box:
    - Drag precise file snapshot to native chat box draft

There is a limitation here: historical files can only be pushed to the composer if Harness events provide a precise source snapshot; history with only metadata is unavailable.

Scenarios and Notes

This plugin is suitable for viewing Agent execution trajectories in the web profile of DeepSeek Harness, especially for scenarios requiring observation of the following:

  • Relationships between user prompts, Agent replies, and tool calls
  • Sub-agent delegation process
  • File reading, writing, and external resource access
  • Which events occur first in a run and which are triggered by other events

Pre-use notes:

  • The plugin is described as an experimental interactive plugin, not a formally stable component.
  • The plugin runs with the permissions of the current dsh process. You should check the source code, dependencies, and license before installing.
  • react, react-dom, and three dependencies in package.json are set to latest; it is recommended to confirm the final locked dependency versions in a controlled environment.
  • The browser view does not write to workspace files on its own.
  • The view only renders the currently selected conversation and releases the WebGL scene when unmounted.
  • When WebGL is unavailable, surrounding controls and text fallbacks will be retained rather than making the entire interface unavailable.

Links