Introduction

DeepSeek Harness (DSH) plugin ecosystem is designed to extend agent capabilities through modularization. In long-term project collaboration, agent sessions often face context-window limits and cognitive fragmentation: tasks are forgotten after completion, and each new session must start learning from scratch. To address this pain point, project-ai-docs provides a general-purpose protocol that aims to persist project knowledge in documentation rather than in session memory.

Plugin Overview

project-ai-docs is a general-purpose project AI documentation protocol skill. Its core logic is to maintain an independent docs/.ai/ knowledge base for each project. Before executing a task, the agent first reads the index of this knowledge base, then selectively reads the relevant documents; after the task is completed, changes are automatically written back to the documentation. This approach aims to prevent context dispersion and preserve the granularity of project knowledge.

Core Features

This plugin ensures the effectiveness of the documentation system through the following protocols and mechanisms:

  1. Map first: index.md serves as the single entry point, building a routing table for the entire documentation set.
  2. Layered budget: Documentation is divided into four layers—core/modules/config/best-practices—to explicitly control token overhead.
  3. Red lines first: Hard constraints (prohibitions) are placed at the top of index.md, taking priority over technical details.
  4. Selective reading instead of full reading: Only relevant slices are read based on task requirements, avoiding interference from redundant information.
  5. Write-back closed loop: Task completion is treated as a documentation update, so knowledge is deposited in the documentation system.
  6. Incremental awareness: A changelog ledger records changes; a new session can understand the current project state by reading the tail.
  7. Human-agent separation: README.md is intended for users, while docs/.ai is intended for agents.
  8. Verifiable facts: Documentation includes only verified paths, versions, and commands, eliminating guesswork.

Installation and Activation

Install it using the plugin command and manage it with a profile:

dsh plugin --profile web add project-ai-docs

After installation is complete, restart the DSH process to activate it.

Local Verification Installation

If you need to test local code that has not been published to npm:

dsh plugin --profile web add file:D:\Desktop\project-ai-docs

How It Works

Route B installation uses the dsh.bundle.patch configuration in the package to point to cordis.patch.yml, which then loads extensions/dsh/index.js. This adapter registers the contents of the skills/ directory as technical sources for DSH.

Other Hosts

For non-DSH hosts such as Claude Code and Codex, you can use it by copying the skills/project-ai-docs/ directory to the corresponding skills path.

Usage

The plugin defines three standard stages, executed in order:

1. New Session Start

When the agent starts, it first reads index.md and changelog, then selectively reads the relevant layered documents based on task requirements to establish context.

2. Initializing Project Documentation

The trigger command is /init. It executes the Init protocol:
* Scanning the project structure
* Building the documentation skeleton
* Populating content through verification

3. After Task Completion

It executes the cleanup and write-back flow:
* Removing redundant information
* Writing back to the corresponding documents
* Recording the change in changelog

Ecosystem and Dependencies

  • Dependencies: This plugin depends on @deepseek-ai/dsh-skill-filesystem as a peer dependency (optional).
  • License: MIT.
  • Directory structure:
    • skills/project-ai-docs/: Runtime skill unit.
    • extensions/dsh/index.js: DSH bundle adapter.
    • cordis.patch.yml: Bundle registration configuration.

Conclusion

project-ai-docs provides a structured documentation management solution, helping developers and agents build a persistent and verifiable project knowledge system in the DSH environment. For more details and source code, refer to its GitHub repository.