mmparko/dsh-purge

dsh‑purge – Red‑team plugin for DeepSeek Harness

What it is

• An open‑source plugin that works with the official DeepSeek Harness (DSH) client – both the web version (dsh web) and the desktop executable (v0.2.0‑rc.2).
• It lets you inject custom prompts and policy patches into a locally‑installed DSH instance so you can evaluate how tightly the host‑side safeguards block model responses. In other words, it separates the model refusal layer from the host policy layer for controlled security testing.

Key features

Feature	What it does
dsh‑purge dock	Adds a button next to the chat session title that opens a two‑page UI: Clean and Drill.
Clean page	Shows which patches are applied, lets you apply, restore or uninstall them, edit the injected prompt, and switch between predefined rule‑sets.
Drill page	(stable) After proper authorization, lets you inspect assets, skills and the local environment of the host you manage – useful for offline or lab‑only red‑team exercises.
Host‑policy handling	Copies the default permission policy and tool limits from the official @deepseek‑ai packages, then overwrites only the parts you configure. No second‑hand identities are created.
Zero‑touch install	The plugin never patches the Harness source tree directly; it writes changes only when you click Apply.
Cross‑profile support	Works separately for the web profile and the desktop profile – you must install and apply each one independently.

Supported environments

• DeepSeek Harness 0.2 (official web or desktop client). Older 0.1.x builds require an older plugin version (see the Releases page).
• Node.js ≥ 18 (Windows “no‑flash” builds need Node ≥ 22).
• Works on Windows, macOS and Linux as long as the official DSH client is present.

Installation

1. Via the DSH CLI (recommended)

   # Web profile
   dsh plugin --profile web add dsh-purge
   # Desktop profile
   dsh plugin --profile desktop add dsh-purge
   

   If the repository is already cloned locally, replace dsh-purge with ..
2. One‑click Hub button – click the Install button on the plugin’s page at dsh-plugin.org; it resolves to the same npm install.
3. Manual install (fallback) – clone the repo into $DSH_HOME/plugins/dsh‑purge, add a file: dependency to the appropriate profile’s package.json, and run pnpm install inside that profile directory (see the README for the exact JSON snippet).
4. Finish – quit the DSH host, reopen it, click the new dsh‑purge button beside the session title, and press Apply on the Clean page. The host restarts once to load the patches.

Usage

• After applying, start a new chat. The injected prompt (default for the Chinese model “小码酱”) is now active, allowing you to test how the host reacts to jailbreak attempts or policy changes.
• Use /purge status in the chat to see which patches are active and where the plugin is installed.
• To revert, open the Clean page again and click Restore or Uninstall.

Legal & compliance

• The project is released under the MIT license with a strong attribution requirement: any reuse must credit the original repository.
• It is explicitly non‑profit; commercial resale or selling of the plugin violates the license.
• The README contains a detailed disclaimer stating that the tool is only for local, authorized red‑team research on the user’s own Harness installation. It must not be used against third‑party systems or in violation of Chinese cybersecurity laws.

Community & support

• The author has paused maintenance due to lack of income; no new releases or issue triage are guaranteed.
• Sponsorship addresses (USDT, Solana, Ethereum, Bitcoin) are listed for voluntary support.
• The plugin is listed on the official DSH plugin catalog and on awesome‑dsh‑plugin.com.

Bottom line dsh‑purge is a niche, open‑source utility for security researchers who want to probe the host‑side controls of DeepSeek’s local AI client. It provides a UI for applying, testing, and rolling back prompt‑injection patches, but it does not contain malicious payloads or scan external networks.