Mingxi2077/dsh-plugin-review: Equip DSH with a code review discipline

Preface¶

When writing code using an agent coding workspace like DeepSeek Harness (DSH), the model completes tasks very quickly. The problem lies in the quality after changes: most of the time, one has to manually open an editor again, or write a line in the system prompt like “Please also check the code” — the result is an inconsistent report format, no quantified score, and no audit record left behind.

The Mingxi2077/dsh-plugin-review introduced below turns this into a plugin: performs multi-dimensional health scoring on projects or code, uses procedurally generated SVG radar charts to display results, review history includes categorization, pagination, and a trash can, and supports AI actively triggering a review after completing a module.

What is it¶

dsh-plugin-review is a DSH plugin running on the dsh web platform (client platform is web), created by Mingxi2077, with an MIT license. Its positioning can be summed up in one sentence: turning DSH into a workspace with built-in code quality audit discipline — the scores are procedural metrics, not dependent on the model’s subjective descriptions.

The plugin requires DeepSeek Harness version 0.1.0-rc.6 or higher (depends on Cordis injection specification and client __ModuleLoader__ registration specification). It supports Windows / macOS / Linux. On Windows, if the HOME environment variable is missing, it falls back to the USERPROFILE path.

Core Features¶

The plugin’s capabilities are divided into several parts:

1. Review mode toggle. There is a “Review” button next to the input box’s + icon. When enabled, you can review by entering a project path or pasting code directly.

2. AI Active Review. The plugin injects system rules; AI actively calls the review_audit tool after completing a task, module, or round of fixes; when it finds error-level issues or scores below the passing threshold, it will actively fix and re-audit.

3. 8-Dimensional Health Score. Structure, maintainability, consistency, robustness, testing, documentation, performance, and security are all procedural metrics.

4. Procedurally Generated SVG Radar Charts. The review result card in the conversation and the review history popup will render in real-time. The radar chart is generated purely by SVG using the RadarChart component, with no external chart dependencies.

5. Review History. Supports categorization by time, project, and score. Supports pagination and deletion. Deleted records go to the trash first, defaulting to a 7-day retention period, automatically cleared after expiration.

6. Export and Statistics. Can export CSV, view project statistics, and perform one-click re-audit.

7. Settings Page Configuration. Dimension toggles, passing threshold, scan limit/depth, ignored directories, and trash retention days can all be adjusted in “Settings → Review”.

Installation & Enablement¶

It is recommended to install via the GitHub source. Execute the following command, then restart dsh web and refresh the page:

dsh plugin --profile web add "github:Mingxi2077/dsh-plugin-review#main"

If you have the source code locally, you can also install from the directory:

dsh plugin --profile web add /path/to/plugin-review

To uninstall, use the corresponding remove command, or remove the entry from the web profile’s bundles and restart:

dsh plugin --profile web remove plugin-review

Typical Usage¶

After installing and restarting dsh web, follow these steps:

1. A “Review” button appears at the bottom left of the input box. Click it to enable review mode, and a banner will appear above the input box.

2. Enter the target to review, for example review /path/to/project, or paste the code directly.

3. After sending, the model calls review_audit, and a radar chart score card and review report appear in the conversation.

4. You don’t have to click the button: AI will actively trigger a review when a module is completed; or saying “review” or “evaluate” in the conversation will also trigger it.

There are two usage details to note: review_audit requires an absolute path; when reviewing a single file, please use the text parameter to paste the code; if the radar chart does not display, first confirm that dsh web has been restarted and a hard refresh (Ctrl+F5) has been performed.

Configuration & Persistence¶

The configurable items on the settings page and their defaults are as follows:

Configuration and historical data are stored under $DSH_HOME: review-config.json stores the configuration, review-history.json stores the review history (max 50 entries, deleting this file clears history), and review-trash.json stores the trash. Entries expire and are automatically cleared based on the retention days.

For the target project, scanning only reads the first 40KB of files for statistics; it does not write or modify any content. The conversation URL only stores the session ID when the review occurs; it does not store content.

Implementation¶

The plugin consists of two parts: the Node half (lib/index.mjs) implements the review_audit tool, scanning/scoring logic, and /api/review/* interfaces, following the rc.6 injection specification, injecting webServer, tools, systemPrompt; the Browser half (lib/index.js) registers the shell via window.__ModuleLoader__.load({ id, factory }), implementing the review button, banner, settings page, radar chart, history, and trash can interface.

There are several unfinished plans on the roadmap: exporting review reports as Markdown/JSON, trend charts by project average score, and a comparison view of old and new scores for re-audit results.

Applicable Scenarios & Notes¶

This plugin is suitable for developers who use DSH as their daily coding workspace and want a quantified audit record for every change, especially scenarios where they want the AI to consciously go through the “Review — Fix — Re-audit” process when delivering modules.

Please note before installation: The plugin runs with the permissions of the current dsh process. Scanning reads files from the target project. Before installing, you should check the source code and license. This project uses the MIT License and has no affiliation with DeepSeek official. The plugin name and description are for community distribution use only.

Conclusion¶

A brief summary: dsh-plugin-review turns “code quality” into a quantifiable, archiveable, and re-checkable process using procedural metrics and radar charts, filling the missing link in agent coding workspaces. The plugin page is in the community directory skillhub.cn: https://www.skillhub.cn/plugins/Mingxi2077/dsh-plugin-review . The source code is on GitHub: https://github.com/Mingxi2077/dsh-plugin-review .