dsh purge 安装使用教程: Working GitHub Fork & 破甲配置

The search query dsh purge (alongside breakout autocomplete terms like dsh purge 安装, dsh purge github, and dsh purge fork) experienced a massive >5,000% breakout surge on Google Trends, driven by developers seeking custom prompt control, sandbox guardrail overrides, and unrestricted coding execution inside DeepSeek AI’s official open-source agent framework, DeepSeek Harness (dsh).

Despite sounding like a native cache cleaner or command-line cleanup utility, dsh purge is actually an open-source cybersecurity red-teaming and prompt-injection override plugin developed by community creator YuJunZhiXue. However, rapid framework updates to DeepSeek Harness, combined with the original author’s maintenance hiatus, have left hundreds of agent developers stranded with broken configurations, missing GitHub download repositories, and persistent model refusal loops (“雷霆大思考”).

In this comprehensive guide, we provide the verified installation instructions for dsh purge, identify working GitHub fork repositories, explain how to resolve common upgrade breakage errors, and detail how to transition to an unthrottled, zero-refusal multi-model architecture using APIVALE’s high-concurrency gateway.

1. What is dsh purge? Clarifying the Tool & Search Trend

Before diving into installation, it is crucial to resolve a common misconception: The official DeepSeek Harness CLI (dsh) does not possess a native purge subcommand.

When developers search for dsh purge or dsh purge 安装, they are almost exclusively looking for dsh purge, a third-party open-source extension created by developer YuJunZhiXue for the dsh agent runtime.

graph TD
    A["dsh Agent Runtime (CLI / Web)"] -->|Cordis Plugin Kernel| B["Official Sandbox & Permission Layer"]
    B -->|Restricted System Prompt| C["Default Domestic Model Profiles ('小码酱')"]
    A -.->|dsh-purge Plugin Injection| D["dsh-purge Configuration Layer"]
    D -->|Override Rules| E["Custom Jailbreak & Unfiltered Prompts"]
    D -->|Patch Runtime| F["danger-full-access Sandbox Bypass"]
    D -.->|Frequent DSH Upgrades| G["Broken Node Modules & Crash on Startup"]

Why Did dsh purge Spike >5,000% on Google Trends?

1. The Rise of Autonomous AI Coding Agents: DeepSeek Harness offers a highly capable “everything-is-a-plugin” architecture powered by the Cordis kernel. As more developers adopted dsh for complex codebase automation, they encountered strict default prompt boundaries.
2. Community Amplification: Demonstrations across forums (such as LINUX DO and GitHub) showcased dsh purge swapping system prompts and unchaining models for penetration testing and unrestricted scripting.
3. The Breakage Wave: The dsh framework pushed rapid core updates in October 2026. Because dsh purge relies on delicate runtime patches, these updates broke existing installations en masse, sending thousands of users to Google searching for installation and troubleshooting instructions.

2. Step-by-Step Installation & Usage Guide (dsh purge 安装与使用教程)

If you require dsh purge for local red-teaming, prompt research, or unrestricted coding workflows within dsh, follow these steps to find verified GitHub sources and install the plugin.

Verified GitHub Repositories & Working Forks (dsh-purge Mirrors)

Due to repository visibility changes and upstream maintenance pauses, developers seeking dsh purge download and GitHub releases should verify the repository provenance before executing install commands:

• Original GitHub Repository (dsh purge github): github:YuJunZhiXue/dsh-purge (Initial prototype for prompt injection and sandbox overrides).
• Community Forks & Mirrors (dsh purge fork): Community developers on LINUX DO maintain active mirrors (such as pi-purge and patched forks supporting Cordis v2.4+).
• Direct Git Clone & Local Build (dsh purge download): If the remote plugin manager times out during download:

  git clone https://github.com/YuJunZhiXue/dsh-purge.git ~/.dsh/plugins/dsh-purge
  cd ~/.dsh/plugins/dsh-purge && npm install && npm run build

Step 1: Install via the DSH Plugin Manager

Run the following command in your terminal where the dsh environment is installed:

# Add dsh-purge to the web / desktop runtime profile
dsh plugin --profile web add github:YuJunZhiXue/dsh-purge

For headless CLI-only workspaces, target the default profile:

dsh plugin add github:YuJunZhiXue/dsh-purge

Step 2: Verify Plugin Installation

Inspect the active plugin list to ensure dsh purge is recognized by the Cordis kernel:

dsh plugin list

Expected Output:

Installed Plugins:
├── @dsh/core-sandbox (v2.4.1) [active]
├── @dsh/web-ui (v1.8.0) [active]
└── dsh-purge (v0.3.4) [active, override: enabled]

Step 3: Configure Custom Prompt Rules

1. Launch the dsh Web GUI:

   dsh web --port 3000

2. Navigate to the Plugins tab in the left sidebar and select Purge Rules.
3. Create a new ruleset or select the default profile to configure custom system prompt overrides for your active models.

3. Resolving Common dsh purge Errors & Breakage

Because the original maintainer paused active updates due to personal transitions, users frequently encounter three primary failure modes:

Issue 1: Cannot find module 'dsh-purge' After DSH Upgrades

• Root Cause: Running dsh upgrade or npm update rebuilds node_modules inside $DSH_HOME, wiping out the symbolic links injected by the community plugin.
• Resolution: Re-link the repository manually:

  cd ~/.dsh/plugins
  git clone https://github.com/YuJunZhiXue/dsh-purge.git
  cd dsh-purge && npm install && npm run build
  dsh plugin reload

Issue 2: Sandbox Permission Conflict (danger-full-access)

• Root Cause: DSH core security treats third-party prompt modifications as untrusted inputs, resulting in warnings or silently falling back to the default “小码酱” system prompt.
• Resolution: Edit your ~/.dsh/config.yaml to explicitly elevate the purge rule runner:

  security:
    allow_sandbox_override: true
    trusted_plugins:
      - "dsh-purge"

Issue 3: Plugin Port Alternatives & “Unrestricted” Prompts (dsh unrestricted & dsh purge plus)

Due to the maintenance hiatus, community developers on LINUX DO have introduced trimmed-down forks such as pi-purge or custom unrestricted prompt injection profiles (dsh unrestricted / dsh purge plus). If dsh purge fails to initialize on your specific DSH build, verify if a community fork provides pre-patched compatibility for your version.

4. The Real Problem: “雷霆大思考” (Over-Thinking Refusal Loops)

While dsh purge can alter client-side prompts, it cannot alter the server-side alignment and safety filters of upstream models.

In real-world agent pipelines, developers using dsh purge face a massive productivity killer known colloquially as “雷霆大思考”:

[DSH Agent Loop Turn 4]: Analyzing project reverse engineering task...
[Model]: <think>
Evaluating code disassembly vectors...
Analyzing safety policy on binary modification...
Cross-referencing ethical compliance boundaries...
Considering theoretical educational framing...
</think>
I cannot assist with reverse engineering or modifying this application.

Why This Destroys Agent Workflows:

1. Severe Token & Quota Drain: The model consumes 400 to 1,500 reasoning tokens during its lengthy thinking phase before ultimately refusing the request.
2. Agent Stalling: The agent loop is aborted or enters an infinite self-correction loop, repeatedly hitting the same refusal wall.
3. Account Risk on Third-Party Reverse Proxies: Attempting to force jailbreaks through unofficial, non-standard API relays often results in account suspensions or IP bans.

5. The Architectural Solution: Multi-Model Gateway Fallback via APIVALE

Rather than fighting fragile local plugin patches and risking model refusal loops, professional developers use APIVALE’s high-concurrency API gateway to decouple agent execution from brittle local hacks.

By connecting your dsh agent pipeline directly to APIVALE’s standard OpenAI-compatible /v1/chat/completions endpoint, you gain three decisive advantages:

graph LR
    A["dsh Agent Runner"] -->|Standard /v1/chat/completions| B["APIVALE High-Concurrency Gateway"]
    B -->|Fast Code Generation| C["deepseek-v4-flash (437 t/s)"]
    B -->|Deep Reasoning Tasks| D["deepseek-v4-pro / deepseek-reasoner"]
    B -->|High-Freedom Coding Fallback| E["glm-5.3-flash / glm-5.3"]
    B -->|Long Context Tool Calling| F["qwen3.8-flash / qwen3.8-max"]
    B -->|Massive 2M Memory| G["kimi-k3"]

1. Zero-Hack Stability

Eliminate dsh purge entirely. Supply your custom system prompts cleanly through the standard messages[0] array without modifying core DSH files or worrying about broken node_modules.

2. Multi-Model Resilience & Instant Fallback

When a task triggers an over-alignment refusal on one model, seamlessly switch your endpoint model string in DSH to an alternative frontier architecture:

• DeepSeek Series: deepseek-v4-flash, deepseek-v4.1-flash, deepseek-v4-pro, deepseek-chat, deepseek-v3, deepseek-reasoner
• GLM Series (Zhipu): glm-5.3-flash, glm-5.3, glm-5.3-flashx (exceptional instruction adherence for coding and technical scripts)
• Qwen Series (Alibaba): qwen3.8-flash, qwen3.8-max, qwen3.8-omni-flash (unmatched agentic tool-calling reliability)
• Kimi Series (Moonshot): kimi-k3 (up to 2M tokens context capacity)
• Video Generation: Doubao-Seedance-2.0-mini, Doubao-Seedance-2.5

3. Ultra-Low Cost Flash Matrix ($0.20 Free Trial)

Agent loops consume massive amounts of tokens. Running iterative debugging on full-tier models quickly becomes cost-prohibitive.

APIVALE’s Flash matrix (deepseek-v4-flash, glm-5.3-flash, qwen3.8-flash) operates at fraction-of-a-cent rates. New registered developers receive $0.20 free starter credit, which translates into hundreds of thousands of tokens—more than enough to power dozens of autonomous dsh agent runs completely free.

6. Architecture Comparison: Local Hack vs. APIVALE Gateway

7. How to Connect dsh to APIVALE

Configuring dsh to route through APIVALE takes less than 60 seconds.

Method A: Configure ~/.dsh/.credentials.yaml

Open your dsh credentials file and override the base URL and API key:

providers:
  openai_compatible:
    base_url: "https://apivale.com/v1"
    api_key: "sk-apivale-YOUR_API_KEY"
    default_model: "deepseek-v4-flash"
    fallback_model: "glm-5.3-flash"

Method B: Environment Variable Injection

For quick terminal testing or containerized agent runners, export standard environment variables:

export OPENAI_BASE_URL="https://apivale.com/v1"
export OPENAI_API_KEY="sk-apivale-YOUR_API_KEY"
export DSH_DEFAULT_MODEL="deepseek-v4-flash"

# Launch dsh with instant high-speed routing
dsh run "Analyze the current workspace and refactor the auth middleware"

Method C: Direct Verification via cURL

Test your endpoint connectivity before launching long agent tasks:

curl https://apivale.com/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer sk-apivale-YOUR_API_KEY" \
  -d '{
    "model": "deepseek-v4-flash",
    "messages": [
      {"role": "system", "content": "You are an expert AI software engineer."},
      {"role": "user", "content": "Write a high-performance concurrent queue in Go."}
    ]
  }'

<!-- $29 Developer Pack (Hero / Most Popular) -->
<div class="border-2 border-amber-500/70 bg-gradient-to-br from-amber-950/30 via-gray-950 to-gray-950 rounded-xl p-5 relative flex flex-col justify-between shadow-xl">
  <span class="absolute -top-3 right-4 text-[10px] font-bold uppercase tracking-wider bg-gradient-to-r from-amber-500 to-orange-500 text-black px-3 py-0.5 rounded-full shadow-md">
    🔥 56% OFF • Most Popular
  </span>
  <div>
    <div class="flex items-center justify-between mb-2">
      <span class="text-2xl font-bold text-white">$29 <span class="text-xs text-gray-400 font-normal">USD</span></span>
      <span class="text-[11px] bg-amber-500/20 text-amber-300 border border-amber-500/30 font-mono px-2.5 py-1 rounded-md">Developer Pack</span>
    </div>
    <div class="text-sm font-bold text-amber-400 mt-1">Get $43.50 Credits <span class="text-xs text-gray-300 font-normal">(&approx; 40,000,000 Tokens)</span></div>
    <p class="text-xs text-gray-300 mt-3 leading-relaxed">Power continuous dsh agent loops, Claude Code CLI, and Cursor multi-agent workflows for months. Unrestricted access to DeepSeek V4 Flash, GLM-5.3, Qwen 3.8, and Kimi K3.</p>
    <div class="mt-3 flex items-center gap-2 text-[11px] text-amber-300/80">
      <span>✓ 40,000,000 Universal Tokens</span>
      <span>•</span>
      <span>✓ Never Expire</span>
    </div>
  </div>
  <a href="https://apivale.com/pricing?utm_source=apivale_blog&utm_medium=pricing_pack_grid&utm_campaign=dsh-purge-installation-guide-deepseek-harness-fix" class="mt-5 block text-center text-xs font-bold py-2.5 px-4 rounded-lg bg-gradient-to-r from-amber-500 to-orange-500 hover:from-amber-400 hover:to-orange-400 text-black transition-all shadow-lg hover:shadow-amber-500/20">
    Claim 56% OFF Developer Pack ($29) &rarr;
  </a>
</div>

Frequently Asked Questions (FAQ)

What is the official installation command for dsh purge (dsh purge 安装)?

The verified command to install the dsh purge plugin within dsh is dsh plugin --profile web add github:YuJunZhiXue/dsh-purge. For headless CLI profiles, use dsh plugin add github:YuJunZhiXue/dsh-purge.

How do I use dsh purge and configure custom unrestricted prompts (dsh purge 怎么用 & 提示词)?

Launch the Web UI via dsh web --port 3000, navigate to Plugins > Purge Rules, and enable custom system prompt overrides. Alternatively, edit ~/.dsh/plugins/dsh-purge/config.json to inject unrestricted prompt rules and execute dsh plugin reload.

Where can I find working GitHub repositories and forks (dsh purge github / dsh purge fork)?

While the original repository by YuJunZhiXue is in a maintenance hiatus, developers can access active community forks (like pi-purge on LINUX DO) on GitHub or clone the source locally into ~/.dsh/plugins/dsh-purge and run npm install && npm run build.

Why does dsh purge break or fail to load after updating dsh?

Core dsh updates frequently rebuild node_modules inside $DSH_HOME and reset sandbox permission policies (danger-full-access). If the plugin fails after an upgrade, manually re-clone the repository or switch to API-level multi-model routing via APIVALE.

How does APIVALE solve model over-thinking and refusal errors in agent loops?

APIVALE provides a unified OpenAI-compatible endpoint allowing instant fallback across DeepSeek V4 Flash, GLM-5.3 Flash, and Qwen 3.8. When one model triggers a refusal loop, the agent seamlessly routes to an alternative model with zero local file hacking.

Does APIVALE support PayPal billing for international users?

Yes, APIVALE natively supports PayPal instant checkout and international credit cards with zero KYC, bypassing Alipay and Chinese +86 SMS verification hurdles.