Introduction

The design philosophy of DeepSeek Harness (DSH) is “everything is a plugin,” allowing developers to extend an agent’s capabilities by combining plugins. In the community, some plugins adjust model behavior by modifying system prompts to achieve specific output styles or logical constraints. dsh-plugin-iam-deepseek-v4ga belongs to this category of tools; its core goal is to attempt to change the linguistic habits and structure used by DeepSeek models when generating chains of thought by injecting specific rules.

Plugin Introduction

This is a DSH entertainment plugin based on “metaphysics,” aiming to make DeepSeek’s chain of thought imitate the so-called “gray-test godlike model.” The plugin supports DSH version dsh-v0.1.3-alpha.1. The project is maintained by XuezuoYS and is licensed under MIT.

Core Functionality

The plugin injects a fixed rule block at the top of the system field for every model request (each conversation turn or each agent step). Its specific behavior is as follows:

  1. Rule injection: Adds specified content to the very top of the system prompt.
  2. Chain-of-thought constraint: Forces the model to use English during its thinking process and start with We need or I'm. This aims to change the model’s original CoT style.
  3. Plugin setting: Provides an enablePromptInjection configuration option, enabled by default. When enabled, the above rules are injected on every request; when disabled, the plugin registers nothing at all, resulting in no action.

Installation and Activation

Installation

First, make sure pnpm is installed locally, then run the following command:

npx @deepseek-ai/dsh plugin --profile web add github:XuezuoYS/dsh-IamDeepSeekV4ga

Activation

After installation, you need to enable the plugin in DSH settings. The path is: Settings → Plugins → Plugin Configuration. In the configuration, find enablePromptInjection and set it to true or false. You can also directly write enablePromptInjection: true/false in the config field of the plugin list.

Uninstallation

To remove the plugin, run:

npx @deepseek-ai/dsh plugin --profile web remove dsh-plugin-iam-deepseek-v4ga

Usage Notes

After multiple rounds of testing, the plugin’s effects are highly random, similar to a “gacha” mechanism. The following is real-world feedback based on DeepSeek V4 Pro and V4 Flash:

  1. Gacha trigger:

    • DeepSeek V4 Pro: During certain time periods, in less than half of cases, it can make the chain of thought exhibit another, more efficient form; in most cases, it only achieves a purely for-fun effect such as replacing Let me with I'm.
    • DeepSeek V4 Flash: It has rarely triggered the chain of thought exhibiting a more efficient form, only achieving partial text replacement, and the replacement is often incomplete.

  2. Performance degradation: In some test scenarios, model performance may intermittently degrade (this may be caused by DeepSeek server-side issues).

  3. Armor-breaking issues:

    • V4 Pro: When the chain of thought is rolled into a fully modified version, the model is more likely to refuse execution, manifested as its built-in “Heart Wall” becoming thicker, causing armor-breaking terms to become completely ineffective.
    • V4 Flash: When the chain of thought is effectively influenced, the model may autonomously avoid the scope permitted by armor-breaking terms.

Technical Compatibility Notes

Plugin version 0.3.0 supports DSH version dsh-v0.1.3-alpha.1. This DSH version removed the standalone installSettingsSection and settingsNamespace helper functions from @deepseek-ai/dsh-settings. Therefore, this plugin now uses the ctx.settings service (through SettingsProvider.installSection) to register its own settings namespace. If the module fails to load because an export is missing, the plugin startup will fail.

Conclusion

dsh-plugin-iam-deepseek-v4ga is an entertainment plugin for debugging and exploring the behavioral boundaries of DeepSeek models. It observes model reactions by forcefully modifying the system prompt. Given its “metaphysical” nature and the potential armor-breaking/performance issues it may introduce, it is recommended to use it only in non-production or experimental environments. For more details and source code, refer to the project directory or GitHub repository.