Fujiang Po Jia Installation and Usage Guide: Configuring Security Skills, Knowledge Bases, and Workflows for DeepSeek Harness

Project Overview

Fujiang Po Jia is a custom prompt and security assessment assistant plugin for DeepSeek Harness. The version currently listed in the README is 0.20.12. The project installs @deepseek-ai/dsh-prompt-custom into the Harness configuration directory and stores the plugin program separately from the prompt-custom.json prompt configuration.

The plugin provides custom prompts, a skills library, a built-in CVE/PoC knowledge base, vulnerability reports, and penetration testing workflows. It is suitable for organizing assets, matching candidate vulnerabilities, recording verification steps, and generating reports in environments where explicit authorization has been obtained. Knowledge base candidates and automated progress must not be treated directly as confirmed vulnerabilities.

The project name and references to DeepSeek do not imply official endorsement. Read the disclaimer in the repository before use, and ensure that the test target, scope, and operating methods are all authorized.

Key Features

• Prompt injection: A “Fujiang Po Jia” status switch and settings entry are provided near the input box. Disabling injection restores the official opening while retaining the user's own prompt body.
• Vulnerability reports: Persistently stores vulnerability records written by artificial intelligence, which can be viewed by critical, high, medium, low, and informational severity levels.
• Skills library: Downloads, enables, or disables REA and AboutSecurity on demand, allowing artificial intelligence to discover registered skills.
• Built-in knowledge base: Provides nine categories of CVE/PoC knowledge with the plugin; enabling it requires no additional download.
• Penetration testing workflow: Manages tasks, assets, verification steps, attempt records, stages, and follow-up recommendations.
• Local fingerprint matching: Ranks candidate knowledge base entries by port, service identifier, or product name and provides matching reasons.

Custom Prompt Interface

The custom prompt configuration is stored separately in prompt-custom.json and may contain personal settings. Before reinstalling, backing up, or distributing it, confirm that its contents are suitable for sharing.

Skills and Knowledge Base

The skills library supports two upstream projects: REA reverse analysis tool and the AboutSecurity security assessment skills library. Each third-party project is subject to its own license terms.

The built-in knowledge base covers Java deserialization, Java Web, perimeter devices, Exchange and domain environments, Web and CMS, Linux privilege escalation, cloud and containers, business interface risks, and vulnerability triage. Entries record products, versions, ports, keywords, severity, prerequisites, remediation recommendations, and reference sources.

Prerequisites

1. Confirm that you are using a supported version of DeepSeek Harness. The plugin declares compatibility with DeepSeek Harness and dsh-system-prompt versions >=0.2.1-rc.5 <0.3.0.
2. Confirm that the Cordis version is within the >=4.0.5 <5 range.
3. Launch the official DeepSeek Harness Desktop at least once so the program can initialize the desktop configuration.
4. Fully exit Harness before installation, including selecting “Exit” from the Windows system tray icon in the lower-right corner.
5. Download the repository from a trusted source and inspect the file contents before running the script.

Using the Installation Script

On the repository page, select Code → Download ZIP and extract it, or clone the repository with Git. After fully exiting Harness, double-click 点我执行脚本.cmd in the extracted directory.

# Run after fully exiting Harness
.\点我执行脚本.cmd

# View the installation plan without modifying files
.\install.ps1 -DryRun

-DryRun is useful for checking the operations the script is prepared to perform before the actual installation. The installer uses the desktop configuration by default and prioritizes the DSH command-line tool bundled with Desktop. To customize the environment, use -Profile, -DshCommand, and -DshHome to specify the configuration name, command-line tool, and data directory.

If DSH_HOME is not set, the default data directory is $env:USERPROFILE\.dsh. The installation script copies the plugin to $DSH_HOME\local-plugins\dsh-prompt-custom, registers it with the DSH plugin manager, and then copies the included prompt-custom.json.

If the existing prompt configuration differs from the new configuration, the installer first creates a backup beside the original file. If registration fails and a backup exists, the script restores the original prompt configuration. After a successful installation, the script also attempts to write install.log beside it, but this file is not a complete execution log.

If the installer detects that Harness is still running, it prompts you to exit and stops without modifying files. The installer does not automatically close or restart Harness and does not support the -NoRestart parameter.

When Windows Blocks the Script

The installer is a readable PowerShell script that mainly copies files and invokes DSH commands. The README states that it does not launch hidden windows, forcibly terminate processes, create scheduled tasks, or create processes through WMI, but this does not constitute a security product's detection result.

If Windows blocks the script because of a download-source mark, first verify the download source and review the script contents. You can also calculate the SHA256 checksum and compare it with the value provided by the publisher:

Get-FileHash .\install.ps1 -Algorithm SHA256

After confirming that the file is trusted, use “Unblock” in the Windows file “Properties” to remove the download-source mark.

Manual Installation

If you do not want to run the automatic installation script, you can manually copy and register the plugin after fully exiting Harness. The following commands use the default data directory and the DSH command-line tool bundled with Desktop:

$pluginTarget = Join-Path $env:USERPROFILE '.dsh\local-plugins\dsh-prompt-custom'
New-Item -ItemType Directory -Force -Path $pluginTarget | Out-Null
Get-ChildItem -LiteralPath .\plugin -Force | ForEach-Object {
  Copy-Item -LiteralPath $_.FullName -Destination $pluginTarget -Recurse -Force
}

& "$env:LOCALAPPDATA\Programs\DeepSeek Harness\resources\runtime\cli\bin\dsh.cmd" plugin --profile desktop add $pluginTarget

If you also want to use the prompt configuration included with the repository, first back up the existing $env:USERPROFILE\.dsh\prompt-custom.json, then copy the new file. After installation is complete, restart Harness manually.

Basic Usage Workflow

Step 1: Enable Prompt Injection

After restarting Harness, find the “Fujiang Po Jia” status entry near the input box. When enabled, the plugin replaces the official harness:identity opening with an empty section; when disabled, it restores the official opening. Your custom prompt body is not deleted.

Step 2: Configure the Skills Library

Open “Skills Library” in the settings panel and download and enable REA or AboutSecurity as needed. AboutSecurity downloads an archive of the current master branch from GitHub and extracts only its skills/ directory. The download supports progress display and cancellation.

REA requires a compatible Node.js version of ^22.23.4 || ^24.22.1 || >=26.12.1. When installing REA through npm, the plugin disables lifecycle scripts, so optional native bindings for Frida are not built through this installation method. Whether specific tools can run still depends on the local runtime environment and analysis engine.

AboutSecurity mainly consists of Markdown skill documents and does not require additional native runtime components. A successful skill download only means that the files were obtained and registered; it does not mean that every tool has completed runtime verification.

Step 3: Use the Knowledge Base for Assisted Matching

The knowledge base is enabled by default, with its state stored in $DSH_HOME/fujiang-skills/knowledge-state.json. When disabled, the plugin no longer provides the relevant knowledge to artificial intelligence, but the original files remain stored locally.

In an authorized test, you can first tell artificial intelligence the target product, version, port, service identifier, and test scope, then ask it to retrieve relevant knowledge entries. For example:

This is an authorized test environment. Based on the confirmed product name, version, and open ports, search the knowledge base for candidate entries and list the matching reasons, applicable prerequisites, evidence that still needs to be verified, and remediation recommendations. Do not describe candidates as confirmed vulnerabilities.

Knowledge entries are compilations of public information and typically use metadata such as verified: false and sources: to mark verification status and sources. In actual use, they must be reviewed against the target version, configuration, and prerequisites.

Step 4: Create a Penetration Testing Workflow

The current implementation requires users to create a task in the plugin panel first. Only then can artificial intelligence use the existing task to record assets, steps, and results. Workflows are organized into the following stages:

1. Information gathering
2. Fingerprint identification
3. Vulnerability matching
4. Feasibility verification
5. Exploitation
6. Lateral movement
7. Report output

Each task can be bound to one session, and each session can be bound to at most one task. The plugin provides the bound artificial intelligence with the task, stage, asset count, and follow-up recommendations through campaign:context.

When creating a task, clearly specify the authorized targets, permitted scope, prohibited actions, testing period, and stop conditions. Task progress and stages are organizational information only and cannot replace vulnerability verification evidence.

Step 5: Record Vulnerability Reports

Only after the verification results support a conclusion should you ask artificial intelligence to write a vulnerability record. The report should include affected assets, reproduction conditions, evidence, impact assessment, severity, remediation recommendations, and reference sources. Unverified candidates should remain in a “Pending Verification” state rather than being marked directly as vulnerabilities.

Local APIs and Native Tools

The plugin provides local fingerprint matching and task ledger APIs. Matching calculations and ledger writes are performed locally, but skill downloads, networked model calls, and actual target verification may still generate network requests.

• POST /api/dsh-prompt-custom/match: Returns ranked knowledge base candidates and matching reasons based on ports, service identifiers, or product names. Weights include port 40, product 30, CVE 20, keyword 20, and severity 5.
• /api/dsh-prompt-custom/campaigns: Manages test scope, asset fingerprints, execution steps, attempt records, and task stages. The ledger is stored in $DSH_HOME/vuln-report/campaigns.json.
• GET ?id=: Reads the specified task and next-step recommendations.

The task interface also provides operations such as bind, unbind, and kick. Native tools include kb_match, campaign_upsert, finding_upsert, and pipeline_next; skill discovery entry points include skill_search and skill_read. Native tool registration depends on a resolvable DSH tool service.

Advanced Usage Recommendations

Check the Installation Plan with DryRun First

When installing for the first time or changing the configuration directory, run .\install.ps1 -DryRun first. This confirms the target path, configuration name, and expected operations while avoiding immediate file changes.

Back Up the Plugin and Personal Prompts Separately

prompt-custom.json is separate from the plugin program and may contain personal configuration. Back up and review this file separately before upgrading or redistributing the plugin to avoid overwriting customizations or accidentally sharing sensitive information.

Distinguish Candidate Matches from Valid Vulnerabilities

Fingerprint matching only ranks candidates based on ports, products, CVEs, keywords, and severity levels. A high ranking does not mean the target is necessarily affected. Continue verifying the version, deployment method, patch status, prerequisites, and actual evidence.

Use Automatic Continuation with Care

Automatic continuation runs for a maximum of 8 rounds by default and includes stagnation detection and next-step recommendations. Set clear authorization boundaries and stop conditions, and maintain ongoing human review to prevent the model from performing inappropriate follow-up actions due to insufficient context.

Understand How the Knowledge Base Is Updated

Knowledge entries are located in plugin/knowledge/. The files are reread each time an entry is loaded, so content updates can take effect in subsequent calls. Category definitions are located in lib/knowledge.js; changing categories requires reloading the plugin.

Troubleshoot Input Shortcut Issues

The README documents a known issue: after the disclaimer dialog is closed, the Enter or Esc listeners may continue to affect input and keyboard shortcuts. When abnormal behavior occurs, first check whether it is related to this known issue, then reload the plugin or restart Harness.

Compatibility and Verification Boundaries

The plugin works through the public systemPrompt, skills, tools, agent lifecycle, and webServer services. It does not install by modifying Harness program files. After installation, you must fully exit and restart Harness.

The README states that the current version's file structure and installation logic have been checked, but that complete functional runtime verification has not been performed again. Therefore, it cannot guarantee that all skills, CVE entries, or automation workflows are available. Before deploying to a production environment, verify each item in an isolated, authorized test environment.

Conclusion

Fujiang Armor Breaker integrates prompt management, security skills, a built-in knowledge base, vulnerability reporting, and task tracking into DeepSeek Harness. The proper approach is to check compatibility and the installation plan first, then configure skills and the knowledge base, and finally create tasks, verify candidates, and preserve evidence within clearly defined authorization boundaries. Always prioritize human review, least privilege, and reliable evidence over automated workflows.