Introduction

When using DeepSeek Harness (DSH), if the requirement descriptions submitted directly to the background agent are not sufficiently rigorous or lack context, the agent may be unable to accurately locate files or understand the constraints. The vinzelles/dsh-prompt-enhancer plugin aims to solve this problem. By isolating the background agent, it rewrites and enhances user prompts using conversation context and repository knowledge, and appends an existence-verified list of reference files at the end of the result to ensure the instructions are clear and actionable.

Core Features

This plugin provides two action buttons on the right side of the chat input box:
* Enhance: Uses an isolated background agent, combined with the current conversation context and repository knowledge (via glob/grep and other methods), to rewrite the original input into a task prompt with a clear structure and explicit constraints. The enhancement result is not automatically sent, does not enter the main conversation history, and only replaces the draft in the input box. The result ends with an existence-verified list of reference files.
* Restore: One-click restoration of the input box to its original state before enhancement.

Installation and Enablement

After installation, you need to restart dsh web or refresh the page for the icon to appear.

dsh plugin --profile web add github:Vinzelles/dsh-prompt-enhancer

Configuration

Configuration can be performed in Settings -> Plugins -> Plugin Configuration -> the “Prompt Enhancement” card:
* Enhance Strength:
* Low: Only polishes wording and structure, with output length closest to the original input, and includes no file references.
* Medium: Standard optimization, with compact output (approximately ≤1.5 times the original length) and reference files listed by path only.
* High: Fully expands the background, steps, and acceptance criteria, with output length ≤3 times the original length and reference files accompanied by descriptions.
* Model Configuration: Supports specifying a separate model for the enhancement feature. Configuration takes effect immediately without a restart. It is recommended to set a lightweight and fast model to avoid excessive token consumption.
* Convert to English: Supports a “Convert to English” toggle. When enabled, regardless of the input language, the enhancement result is output as an English prompt (the reference files section title becomes ## Reference Files, and the skill references section title becomes ## Skill References). When disabled, it follows the input language.

Usage Flow

  1. Enter your requirements in the chat input box.
  2. Click the “Enhance” button on the right side. The input box will be locked, and the background agent will begin processing.
  3. After enhancement is complete, the input box will display the rewritten prompt and the list of reference files.
  4. Click the “Restore” button to restore the original text.

Notes and Troubleshooting

  • The input box is locked during enhancement and cannot be edited.
  • The plugin runs with the permissions of the current dsh process. Before installation, you should review the source code and license.
  • Zero Pollution: Enhancement runs in an isolated sub-conversation and does not pollute the main conversation history.
  • Path Validation: The list of reference files is validated line by line for existence, and no paths are fabricated.
  • Troubleshooting: If the icon is not displayed or the behavior is abnormal, check the log at ~/.dsh/super-injector/prompt-enhancer-diag.log.

Conclusion

Through its isolated agent execution mechanism and strict path validation, this plugin provides DSH users with a safe and reliable prompt enhancement solution. For developers who frequently collaborate with agents, it can significantly improve the accuracy of instructions and execution efficiency.