Introduction¶

When working with agents or daily conversations in DeepSeek Harness (DSH), assistant replies are often a large block of Markdown: tables require horizontal scanning, simple data can only be viewed as ASCII, and complex explanations lack clickable and draggable controls. The DSH ecosystem emphasizes “everything is a plugin,” with community plugins hosted on an independent directory site, having no official affiliation with DeepSeek or 幻方. If you want to make formal replies more readable while preserving the ability to view the original text at any time, without breaking DSH’s original page structure, you can learn about the community-maintained client plugin dsh-reply-beautify (repository ErSu137/dsh-reply-beautify).

What Is This¶

DSH Reply Beautify (package name dsh-reply-beautify) is an unofficial community plugin, MIT licensed, maintained by ErSu137, and categorized as a client plugin on SkillHub. The current README indicates 0.6.5 pre-release; it is designed for Windows DSH Desktop, with interface compatibility based on 0.2.1-rc.2. The SkillHub directory page description matches the repository description: it beautifies the layout of Harness assistant replies and provides interactive charts, tables, and sandboxed HTML widgets.

The plugin only enhances the assistant’s formal replies; user messages, reasoning processes, and tool outputs do not participate in component transformation, and regular code blocks are not automatically executed.

Core Features¶

Reading Layout¶

Three presets are provided: Default, ChatGPT, and Claude. You can adjust body font size, code font size, line spacing, and paragraph spacing; some interface fonts, labels, and accent colors change according to the style. The README states that in Claude style, the regular accent color is orange, while semantic colors such as traces retain distinctions like purple, green, and gold.

Visualization and Computation¶

When the model follows the read-only formatting guidelines provided by the plugin, it can present reliable ASCII field diagrams, relationship diagrams, some Mermaid TD/LR flowcharts, data charts, and explicit formulas; it also supports interactive charts, parameter sliders and curves, self-test questions, forms, etc. Since version 0.6.1, there are more proactive visualization guidelines: regular proportional graphics are fully adapted, extreme proportions retain scrolling; charts support mouse wheel zooming, double-click or “Auto” to restore; line charts with short numbers on the horizontal axis are more compact while preserving all data. The model may still not follow the guidelines, and old ASCII replies will not be automatically redrawn.

General Tools and Workshop Space¶

The capabilities listed in the README include: searchable/sortable/editable tables with CSV export, calculator, weighted scheme comparison, parameter scenarios, timeline/Gantt/calendar, checklist/kanban, hierarchy, step-by-step playback, branch wizard, diff, source cards, linked reports, etc.; see the repository documentation《Formats and Limitations》for specific quantities and format limits.

Workshop Space can run HTML/CSS/JavaScript according to the documentation《Workshop Format》, supporting inline replies, sidebar, fullscreen preview, and local version editing; inputs using state API can be restored. Workshops run in a sandbox iframe with double-layer opaque origin, where CSP restricts external resources, connections, and subframes; no file, Node/Electron, chat history, or tool bridge is provided. Sandboxing does not guarantee absolute safety for arbitrary model-created workshops, nor does it provide CPU/memory hard limits.

In streaming replies, complete and validated blocks can be presented first, with subsequent content continuing to output; invalid or unsupported content retains the original text.

Installation and Activation¶

Currently, no npm package has been published, nor has it been submitted to custom plugin markets; installation is based on GitHub Release and install.ps1. Please read the source code and license yourself before installing; the plugin runs locally with the DSH process, and data boundaries are described below in “Use Cases and Cautions.”

1. Download dsh-reply-beautify-0.6.5.zip and SHA256SUMS-0.6.5.txt from Release v0.6.5.

2. In PowerShell, verify and extract to a long-term directory (do not move or delete this directory after installation), and enter the inner directory containing install.ps1 and package.json:

Get-FileHash .\dsh-reply-beautify-0.6.5.zip -Algorithm SHA256

3. Ensure DSH has been started at least once, then completely exit via “App → Exit”. Execute the following in that directory:

powershell.exe -NoProfile -ExecutionPolicy Bypass -File .\install.ps1 -Check
powershell.exe -NoProfile -ExecutionPolicy Bypass -File .\install.ps1

-Check is a read-only check; the installer adds the plugin through the official CLI using link:, --offline --ignore-scripts, without needing to download dependencies. The default profile is %USERPROFILE%\.dsh\profiles\desktop; if DSH_HOME is set, it follows that variable. For custom DSH installation paths, you can pass -Installation 'D:\DeepSeek Harness' to the installation script; see the repository《Installation, Update, and Recovery》for details.

4. Restart DSH, open the paintbrush in the bottom-left account row, confirm version 0.6.5, and try it in a new test conversation.

If DSH is installed on drive D or another directory, check, install, and uninstall operations can all pass -Installation, using the same set of commands as the default path.

Typical Usage¶

No need to memorize special command prefixes; just describe the task directly to the assistant. Examples given in the README include:

• “Compare budgeted hours of A, B, C (12, 18, 9) with actual hours (10, 20, 8), and display with charts”
• “Use a calculator with adjustable inputs to compare investment costs”

The plugin provides formatting guidelines to the model, and the model autonomously chooses components; generation is not guaranteed every time. Sources, scores, and question answers come from the model; the plugin does not verify facts.

After downloading the source code, you can open docs/demo/index.html locally to browse demo graphics, or open docs/demo/interactive.html to switch series, modify inputs, and run workshop samples; no account is needed, and no model connection is required.

Use Cases and Cautions¶

Who it’s for: Users who primarily work on Windows DSH Desktop (interface compatibility based on 0.2.1-rc.2) and want to improve reading experience, need tables/charts/small interactive pages. Other platforms and DSH versions, long conversation performance are marked as pending verification in the README; it is recommended to validate first in a dedicated test conversation.

Disabling and Uninstalling:

These operations do not actively delete local preferences, interactions, and workshop records; please do not upload the backups directory.

Data and Runtime Boundaries: Preferences, interaction data, workshop code, and API states are stored in local DSH browser storage, which may contain content you input. The README provides upper limit examples (e.g., up to 80 entries/approximately 280,000 characters for regular persistent state, etc.); exceeding limits will evict old records, and cross-device synchronization is not promised. The plugin itself does not automatically send messages or upload this data.

ChatGPT/Claude layouts are only style references, with no official association, and do not package corresponding application fonts or resources. Complex Mermaid, loop/timing diagrams, etc. may still display the original text; formulas reuse DSH’s native rendering.

Conclusion¶

DSH Reply Beautify adds three layout presets, interactive charts/tables, and sandboxed HTML workshops to assistant replies, while preserving DSH’s page structure and original text viewing capability. It is suitable for on-demand enabling on the Windows desktop version of DSH, with validation first in test conversations.

• SkillHub Directory: ErSu137/dsh-reply-beautify
• Source Code and Release: github.com/ErSu137/dsh-reply-beautify