DeepSeek Harness (DSH) uses a plug-in architecture. When using it, managing conventions and instructions for a specific project is a common requirement. dsh-learning-bundle is designed to address this. It allows project conventions to be taught to agents in natural language within the project while preserving the integrity of the project instruction files.

Core Features

  • Natural-language conventions: Supports teaching agents project conventions in natural language.
  • Instruction maintenance: Maintains existing project instruction files and applies them in subsequent sessions.
  • Native instructions: Uses DSH v0.4 native instructions instead of creating an additional memory database or correctness gating.
  • Legacy recall support: Retains the original store/CLI/core API as a legacy recall feature.

Installation and Enablement

Installing this plugin requires the following:
* Node.js version ^22.19.0 or >=24.0.0
* pnpm
* A supported DSH version
* The headless profile (using the standard instruction loader and file tools)

Run the following commands in the terminal to install it:

git clone https://github.com/B1lli/dsh-learning-bundle.git
cd dsh-learning-bundle
dsh plugin --profile headless add -w "$(pwd -P)"

After installation is complete, it should work normally in the project.

Typical Usage

The plugin manages conventions through natural-language instructions in conversations. The following are typical operation examples:

  1. Permanent convention:
    > From now on, use pnpm to install dependencies in this project. New conversations should follow this as well.
  2. Temporary override:
    > Use npm this time.
  3. Update convention:
    > From now on, use yarn instead.
  4. Revoke convention:
    > Revoke this convention.

These operations are behavioral validation logic for agents, not deterministic guarantees.

Notes

  • Runtime permissions: The plugin runs with the permissions of the current DSH process.
  • Effect evaluation: The plugin states that “Improved net user value is UNPROVEN.” Natural interpretation and compliance still depend on the model’s own capabilities.
  • No additional models: The plugin does not introduce an additional review model.
  • No special commands required: It does not require record, adopt commands, keywords, or item IDs.
  • License and attribution: The plugin uses the MIT license. It is a community project and has no official affiliation with DeepSeek.

Summary

dsh-learning-bundle provides a way to maintain conventions and instructions in DeepSeek Harness projects using natural language. It persists and applies conventions by maintaining native instruction files, making it suitable for users who want to add project-level teaching capabilities to existing DSH workflows.