The core design philosophy of DeepSeek Harness is “everything is a plugin.” In the Web client, the default views are usually only “Conversation” and “Trajectory.” When developers or users debug or analyze a session, they often need finer-grained data insights. Below is a plugin that adds a third view tab to display session statistics.
What It Is¶
This plugin is named dsh-stats-board, maintained by PastSheep, and is a client-side plugin. Its core purpose is to add a third tab named “Statistics,” beyond the Conversation/Trajectory views, and provide visual charts and tables to present session data.
Core Features¶
-
View Registration and Suppression
The plugin registers a component into theconversation.viewslot with order 20 and tab name “Statistics.” It also registers an empty component inconversation.composer.dockto suppress the built-in statistics row in the Web client. -
Session Overview KPIs
The top of the page displays a series of key metrics, including total turns, total steps, LLM duration, tool duration, average Time To First Token (TTFT), output speed, input/output token counts, cache hit ratio, and context occupancy. -
Token Composition Donut Chart
It offers two modes to display token composition. One is projection mode, showing the composition of system, tools, and messages; the other is billing bucket mode, showing uncached input, cache reads/writes, and output. -
Turn Trend Chart
It shows a bar chart of input (blue) and output (purple) for the most recent 16 turns, as well as a cumulative line chart for output tokens. -
Data Details
It provides a per-turn detail table listing turn, steps, LLM duration, TTFT, speed, output volume, and tool count. It also includes a tool call table sorted by call count and total duration.
Installation and Enablement¶
First, run the installation command.
dsh plugin --profile web add @deepseek-ai/dsh-stats-board
After installation, restart DeepSeek Harness. A “Statistics” tab appears in the upper-right corner of the Conversation view. Click it to view the statistics dashboard for the current session.
Data Sources and Dependencies¶
The plugin’s data sources mainly rely on session nodes useSession((s) => s.chat.legacy.nodes) to derive steps, durations, and tokens, as well as projections such as useProjection('sessionStats') to obtain context pressure and composition. Before installation, ensure that the peerDependencies are already installed, including @deepseek-ai/dsh-tools, @deepseek-ai/dsh-client-connection, @deepseek-ai/dsh-client-ui-slots, and react.
When not published, you can install it from a local directory:
dsh plugin --profile web add file:C:/path/to/dsh-stats-board
Notes¶
Because the plugin runs with the permissions of the current DSH process, it is recommended to review the source code and license (MIT) before installation. This plugin uses a build-free approach and is implemented with hand-written JS (React.createElement + inline SVG). It can be installed directly using file: or npm.
Summary¶
After the steps above, you have added a visual session analysis tool to DeepSeek Harness Web. It uses a build-free approach, rendering charts directly with React.createElement and inline SVG, providing a convenient way to inspect data for agent development.
Ecosystem directory: https://www.skillhub.cn/plugins/PastSheep/dsh-stats-board
Source code: https://github.com/PastSheep/dsh-stats-board