Introduction

When developing with DeepSeek Harness (DSH), code review is a key procedure for ensuring code quality. Traditional review methods often require manually writing Git commands to obtain a Diff, or vaguely describing the review scope through natural-language prompts. This typically leads to low efficiency or imprecise reviews.

The dsh_review plugin aims to solve these pain points. Inspired by the Codex CLI workflow, it provides a CLI-like code review experience through interactive menus. Users do not need to manually type complex Git commands; they can quickly select review targets and assign explicit review tasks to the current Agent.

Plugin Overview

dsh_review is a third-party plugin that provides code review capabilities for DeepSeek Harness (DSH). It is maintained by the developer CodeIdeal and is licensed under the MIT license.

The core value of this plugin lies in interactive target selection and structured review. It reads local and remote Git information through a Web-based menu, converts the user’s selection into a specific review request, and finally allows the DSH Agent to read the Diff and generate a review report that includes severity, evidence, and remediation suggestions. It is not an official OpenAI Codex component, but an independent tool for the DSH ecosystem.

Installation and Enablement

Before installation, ensure that Node.js and npm are installed locally and that the DeepSeek Harness Web profile is running correctly.

  1. Clone the plugin repository and install dependencies:
git clone https://github.com/CodeIdeal/dsh_review.git
cd dsh_review
npm install
npm run build
  1. Add the built plugin to the specified DSH profile (the example uses the web profile):
dsh plugin --profile web add ./
  1. Restart the corresponding DSH session. The plugin will automatically apply cordis.patch.yml and register the review tool for the current Agent.

Features

The plugin mainly provides the following capabilities:

  1. Interactive target selection: Select the review scope through the /review menu on the Web interface, supporting uncommitted changes, specific Commits, or specific Branches.
  2. Commit / Branch second-level menus: Dynamically read the most recent commits in the current session repository (up to 100), local branches, and remote branches.
  3. Uncommitted changes review: Covers staged, unstaged, and related untracked files, using HEAD as the comparison baseline.
  4. Custom review requirements: Users can directly input natural language to describe the review scope, background, acceptance criteria, or key areas of focus.
  5. Direct Agent review: The plugin handles target selection and task organization, while the Agent handles reading the Diff, deduplication, severity assessment, and conclusion output.
  6. Structured results: Each review issue must include severity, exact location, evidence, impact scope, and remediation suggestions.

Usage

The plugin provides three ways to use it: command line, interactive menus, and natural language.

1. Interactive Menu

Type / in the DSH Web input box and select review, then choose through the multi-level menus:

/review
├── Uncommitted changes  审查当前工作树相对 HEAD 的改动
├── Commit diff          选择最近的 commit，仅审查该 commit 引入的差异
├── Branch base diff     选择 branch，从 merge-base 审查到当前 checkout
└── Custom requirement  输入任意审查范围或要求

2. Direct Command Entry

You can also bypass the menu and enter commands directly to specify the review target:

/review uncommitted
/review commit <commit-ref>
/review branch <branch-name>
/review custom 重点检查认证逻辑、异常处理和测试覆盖

3. Tool Interface

In addition to commands, the plugin registers a structured review tool for the Agent to call. Key parameters include:
* target (required): Git target, such as uncommitted or working tree.
* targetKind: Git target type; optional values are commit, branch, or range.
* base: Comparison baseline, such as the parent baseline of a commit or the baseline for uncommitted changes.
* task: Original task intent.
* spec: Acceptance criteria.
* focus: Array of additional focus points.

The tool returns diffCommand and structured instructions, and the Agent uses this information to complete the actual review.

Limitations and Notes

  1. Commit menu limitation: The Commit menu displays at most the most recent 100 commits.
  2. Model and context limitations: Review quality and the size of the Diff that can be processed are limited by the current Agent model’s performance and context window.
  3. Separation of responsibilities: The plugin only handles target selection and task organization; specific code reading, Diff analysis, and report generation are performed by the current Agent.
  4. Third-party nature: This plugin is developed by a third party, is not an official OpenAI Codex component, and is not an official app store plugin. Please check the source code and license before installation.