Preface

DeepSeek Harness (DSH) adopts a plugin-based architecture. Plugin repositories in its ecosystem usually lack llms.txt files that AI agents prioritize when discovering projects. When an agent enters a repository, llms.txt is the fastest entry point for understanding the project structure and key information. The dsh-llms-forge plugin aims to solve this problem. It extracts data from package.json and README.md, and deterministically generates a discovery file compliant with the llms.txt specification.

What Is This

dsh-llms-forge is a DSH plugin maintained by zoahdev, used to automatically generate llms.txt for DeepSeek Harness plugin repositories. It is a zero-runtime-dependency command-line tool and also supports being invoked by agents inside Harness.

The plugin runs in read-only mode by default, ensuring that local files are not modified unless write permissions are explicitly enabled. The generated file content includes the title (from the package name), summary (from the description or the first paragraph of the README), and key facts such as installation commands, license, homepage, keywords, and entry points.

Core Features

  1. Automatic generation: Reads package.json and README.md, then automatically extracts metadata to generate llms.txt.
  2. AI-readable: Outputs a format compliant with the llms.txt specification, making it easier for AI agents to discover and index quickly.
  3. Zero runtime dependencies: The tool itself does not rely on a complex runtime environment, making installation and use simple.
  4. CLI tool: Provides a standard command-line interface.
  5. Agent-invokable: Supports use as a tool invocation inside Harness.
  6. Safe and controllable: Read-only by default; write operations must be explicitly enabled with --write and --overwrite.
  7. Reproducible and CI-friendly: Generation results are reproducible, supports JSON output and exit-code checks, and is suitable for integration into CI pipelines.

Installation and Enabling

Install this plugin in a DSH environment using the official command:

dsh plugin add dsh-llms-forge

After installation, it can be invoked inside Harness or run as a standalone CLI tool:

npx dsh-llms-forge .

Typical Usage

Command-Line Usage

The plugin provides the following CLI parameters:

dsh-llms-forge [dir] [--write] [--overwrite] [--json]

Common examples:

  1. Print preview (read-only by default, no file writing):
    npx dsh-llms-forge .
  1. Write file:
    npx dsh-llms-forge . --write --overwrite
  1. JSON output:
    npx dsh-llms-forge ./my-plugin --json

Using Inside Harness (Agent Invocation)

In the interactive Harness environment, you can call the llms_forge tool to generate the file. The agent needs to set dir (directory) and write options:

  • Directory setting: Point to the root directory of the target plugin repository.
  • Write options: Set write: true and overwrite: true.

The tool returns a result in the dsh-llms-forge/v1 format, including the generated content and warning information. Disk writing is performed only when options.write is true.

Use Cases and Considerations

  • Maintenance: Handwritten llms.txt files can easily become out of sync with the source code. Files generated by this plugin are code-determined, making them easier to maintain.
  • CI checks: You can use the --json output and exit codes (0 clean, 1 warning, 2 error) for checks in CI pipelines.
  • Permission check: Before installation, check the source code and license (MIT) to ensure it meets local environment requirements.
  • Environment requirement: Node.js version >= 18 is required.
  • Data sources: Generated content is strictly based on package.json and README.md.

Conclusion

dsh-llms-forge solves the problem of missing AI discovery files for plugins in the DSH ecosystem. Deterministic generation ensures documentation accuracy and maintainability. For developers, it is a foundational tool for integrating into the AI ecosystem and improving plugin discoverability.