Introduction¶
The core design philosophy of DeepSeek Harness (DSH) is “everything is a plugin.” When local session logs have accumulated to a sufficient extent, extracting statistical information from the logs or orchestrating a visual Dashboard requires additional tools. The dsh-dashboard plugin provides a statistical analysis solution by offline scanning local session logs. It builds an orchestrable Dashboard based on Gridstack and ECharts, supporting custom formulas and AI orchestration.
Installation and Activation¶
The installation command is as follows:
dsh plugin --profile web add @weibaohui/dsh-dashboard -w
After installation, restart dsh web for the changes to take effect. Once the plugin starts, it automatically performs incremental scans of ~/.dsh/sessions. The default entry is located at the top of Global panels in the left sidebar.
Core Features¶
Factory Presets and Charts¶
The plugin includes eight built-in factory preset pages, covering Overview, Tokens and Cost, Models and Quality, Commands and Skills, Working Hours and Sessions, Error Analysis, Input and Duration, and Context and Efficiency. These pages directly provide complete Dashboards. It supports 22 chart types, including stacked bar charts, treemaps, box plots, histograms, and more.
Configurability and Interaction¶
The Dashboard supports dragging and resizing through Gridstack. After entering edit mode, users can add cards, create new pages, rename pages, or delete pages, and changes are saved automatically. It supports three-level drill-down: select a day from the Overview heatmap to locate the model leaderboard or session details for that day, then click a session row to view the complete profile of that individual session.
In-depth Analysis¶
The plugin provides multi-dimensional statistical analysis capabilities:
* Context and Efficiency: analyzes context composition (such as system prompts and tool schemas), cache hit rates, TTFT (first response latency), thinking time ratio, and file activity.
* In-depth Error Analysis: classifies errors by provider and error code, and displays Top error clusters, an error sample table, and a tool error leaderboard.
* Output Speed Measurement: calculates precise tokens/s based on streaming chunk timestamps.
* Cost Estimation: estimates costs based on a built-in price list, and supports visually editing the price list (including cache read/write unit prices) and setting a monthly budget.
Intelligence and Formulas¶
Custom formulas are supported, such as pct(a,b) for calculating percentages and perSec(tokens, ms) for calculating rates. AI orchestration is also supported: describe requirements in settings to generate a prompt, pass it to any dsh agent session to generate a configuration JSON, and import that JSON to create a new page.
Performance Notes¶
The plugin includes performance optimization behaviors: it uses multi-frame zstd decompression frame by frame to avoid silent API truncation, and it uses mtime-based incremental scanning, achieving scan speeds measured in seconds for routine operations.
Usage Guide¶
- View Presets: open the dashboard and switch between the eight preset pages to view data.
- Edit Layout: enter edit mode, drag and resize cards, and modify page structure.
- Drill Down: click heatmaps, the model leaderboard, or session rows to view details progressively.
- AI Orchestration: use AI in settings to generate a page configuration and import it.
- Cost Management: edit the price list and monthly budget in settings.
- Scan Control: manually trigger an incremental rescan or a full rescan in settings.
- Formula Editing: use formulas in the card editor to perform custom calculations.
