Preface¶
When using DSH for data analysis, paper review, or reporting experimental results, a common issue is that the tables and figures provided by the model must be manually copied into a plotting tool to be viewed. ASCII art in tool logs offers a rough visual but lacks style customization and usually disappears after a session refresh.
ylifeonlyonce/dsh-inline-visuals addresses this issue by enabling the agent to directly generate a card-style chart within the conversation flow. These cards persist alongside the session logs, remaining available after refreshing, reopening, or branching. Below, we introduce this plugin.
What is it¶
This is a DeepSeek Harness plugin maintained by YLifeOnlyOnce. Its purpose is to render data into charts embedded directly below the reply body, rather than remaining solely in the tool logs. The “Inline” in the name refers to the chart growing directly within the answer.
The community directory page classifies it under the “Fun Outfit” category (a classification by a third-party directory, unrelated to the actual functionality). The project was initialized on 2026-11-05, the README specifies the MIT license, and the version number in package.json is 1.1.0.
The README and repository description are inconsistent regarding the count of chart types: the README title states “18 kinds of charts…” accompanied by a 18-category badge, while the repository description reads “14 replayable visual kinds.” The actual capabilities are defined by the table below.
Capabilities¶
The chart types table provided in the README is divided into three groups:
| Group | Types |
|---|---|
| Charts | bar · line · area · pie · doughnut · scatter · radar · waterfall |
| ML & papers | training curves · ROC/PR curves · confusion matrix · histogram · model leaderboard · embedding scatter |
| Everything else | compare table · timeline · gantt · heatmap · flow · KPI tiles |
These charts support various presentation styles: horizontal or stacked bars, step lines, data labels, legends, grids, five color schemes, height and density, scatter trend lines, logarithmic axes, EMA smoothing, and error bars.
In terms of size, the host-side bundle is approximately 35 KB, while the browser-side self-contained file is about 217 KB (Chart.js is inlined, and React is shared with the host). There are 38 tests written using node:test, CI runs on every push, and npm publishing is tag-driven with provenance.
Installation and Usage¶
Paste the following line in a DSH session:
plugin_manager({ action: "install_bundle", target: "dsh-inline-visuals" })
No cloning or configuration is required. After installation, show_visual and edit_visual will appear in the agent’s tool list.
The README notes that DeepSeek Harness v0.2.1-rc.3 or higher is required. To install from source, the alternative path provided in the README is github:YLifeOnlyOnce/dsh-inline-visuals, or you can point the Plugin Manager to a local clone; for details, see docs/reference.md in the repository.
Typical Usage¶
After installation, simply communicate your requirements to the agent. The README provides several examples:
Compare three caching strategies — hit rate, memory and complexity — as a chart.
Turn that into a horizontal stacked bar, and add value labels.
Draw our release pipeline as a flow.
Show last week’s traffic as a heatmap.
Changing styles does not require redrawing from scratch; edit_visual patches the original specification.
Scenarios for reading papers and training logs:
Plot train/val loss for 8 epochs on a log scale, smoothed, with error bars.
Draw the ROC curve from these points and mark AUC.
Make a confusion matrix for cat/dog/bird from this table.
Compare these four models on accuracy, F1 and latency — highlight the best.
Cluster these embeddings by class and show a t-SNE scatter.
How it Works¶
The model emits a small declarative specification. The plugin validates it and writes it to the tool result, which the client then renders using the host’s theme. Since the data is already within the session log, replaying the session does not require calling the model again.
A few implementation details:
- Cards exist in the session log (
tool/result.meta), not as separate files. - The plugin makes no network requests.
- The model only sends data, not HTML or JavaScript.
- Canvas charts support clicking data points to automatically insert follow-up questions into the input box.
- Pure text clients receive a compact text summary; no information is lost.
- The stable version of the spec is
version: 3; v1/v2 specs are still renderable. - Export: copy as JSON or CSV, charts export to PNG, hand-drawn cards export to SVG.
Target Audience and Notes¶
It is suitable for those who want to read papers, view training curves, or compare data within DSH and expect the charts to remain in the session. Functionally, it overlaps with dedicated plotting tools, but it wins by keeping everything within the conversation without requiring secondary processing.
Please note the following:
- The plugin runs with the permissions of the current DSH process; you should read the source code before installing to confirm its behavior. The license is MIT (© 2026 YLifeOnlyOnce).
- There is an inconsistency in the chart type count (18 in README vs. 14 in repository description). If you are making a selection, it is recommended to directly check the parameter list in
docs/reference.md. - The repository currently has only 1 star, indicating low attention. If you encounter issues, you will basically need to open an issue or modify the source code yourself.
- As per the README, currently only chart canvases support click-to-ask interaction; other interaction types are not yet implemented.
Summary¶
The value of this plugin lies in turning “charts” into a part of the session log: charts appear below the answer, support click-to-ask, editing, and exporting, and persist even after reopening the session. It requires a single command to install and no configuration.
- Directory page: https://www.skillhub.cn/plugins/YLifeOnlyOnce/dsh-inline-visuals
- GitHub: https://github.com/YLifeOnlyOnce/dsh-inline-visuals
