Introduction

The philosophy behind DeepSeek Harness (DSH) is “everything is a plugin.” When reading the open-source code corresponding to a paper, the paper structure is usually fixed, but code repositories vary greatly, often leaving readers unsure where to begin. dsh-code-reading-coach is an Agent preset plugin that uses a “Five-Stage Code Reading Method” to provide interactive guidance, helping you move from “not knowing where to start” to “being able to tell the whole system story.”

Plugin Positioning

This is an Agent preset plugin based on DeepSeek Harness, maintained by TobySUN and released under the MIT license. Its core value lies in structuring the abstract process of code reading and mapping paper claims to code implementations item by item, thereby lowering the barrier to effective code study.

Core Features

The plugin guides reading through a fixed workflow, mainly consisting of the following five stages:

  1. Anchoring: Distill the paper content into 3–5 core claims, forming a list of questions the code needs to answer.
  2. Terrain: Scan the code language, frameworks/architecture, dependency list, directory tree, and test layout to build a terrain map.
  3. Entry: Locate runnable entry points and trace the top-level execution flow.
  4. Core Mapping: Read the core modules, verify the paper claims item by item, and generate a mapping table.
  5. Closure: Run minimal tests, manually trace data flow, modify code to observe results, and perform a Feynman-style explanation.

In addition to workflow control, it provides the following capabilities:

  • Repository-type adaptation: Automatically adjust the reading path based on different repository types, such as training frameworks, kernel libraries, and model repositories.
  • Three-level depth switching: Provide Overview, Trace, and Deep Dive levels. It starts from Trace by default to avoid diving into details too early.
  • Feynman checkpoints: Require users to restate the material in their own words before Deep Dive. If the explanation is unclear, the level is downgraded and re-explained.
  • Explicit contradiction marking: When the code implementation does not match a paper claim, it is explicitly marked.
  • Artifact accumulation: Reading notes are continuously saved to the workspace, generating a <repository-name>-notes.md file and supporting cross-session continued reading.

Installation and Enablement

Use the official installation command to add the plugin. After installation, the dsh process must be restarted for the changes to take effect.

dsh plugin --profile web add github:tobysunsun/dsh-code-reading-coach

After restarting, create a new session and select “Code Reading Coach” from the preset selector to begin.

Typical Usage

After selecting the preset, you can directly enter commands such as:

  • “I want to read the FlashAttention code” — the plugin will first explain the project’s tech stack, then guide you through reading it step by step.
  • Paste a paper link or PDF — the plugin will first distill the paper claims, then verify them against the code one by one.
  • Only have a repository, no paper — simply say “help me understand this repository,” and the plugin will treat this as the learning goal.

Applicable Scenarios and Notes

This plugin depends on the @deepseek-ai/cordis and @deepseek-ai/dsh-tools environments.

Source Code and Maintenance: The plugin source structure is simple, containing only lib/index.js and three files under preset/, so you can review it yourself.

Uninstallation Note: When uninstalling with the remove command, the preset directory is not automatically deleted, to avoid accidentally removing files that users may have modified. If manual removal is needed:

rm -rf ~/.dsh/.agent-presets/code-reading-coach

Ecosystem Background: This plugin is published in the community catalog and has no official affiliation with DeepSeek or Huanfang.

Summary

If you need to systematically read the code behind AI or software papers, dsh-code-reading-coach provides a validated workflow and interaction pattern. Released under the MIT license, you can confidently inspect its implementation details.