Back to home

7dgroup-ai

dsh-skill-7d-code-reviewer

这是一个专业级的 DSH(DeepSeek Harness)代码审查技能插件,由 7DGroup 团队开发,专为 AI 辅助代码审查场景设计。基于 TypeScript + Cordis 开发,以组合包(bundle)形式安装,通过 ctx.skills 注册 7d-code-reviewer 技能:五步审查流程、严重/中等/轻微三级问题分级、四维度评分标准,文本摘要与 HTML 报告双输出。零核心改动——安装即启用,移除 bundle 行即卸载。

Stars
1
Language
HTML
Created
Aug 16, 2026
Updated
Aug 16, 2026

Introduction

7DGroup code review skill plugin banner

npm version license MIT node by 7DGroup

English | 中文

@7dgroup/dsh-skill-7d-code-reviewer

Author: 7DGroup

A professional, template-driven code review skill plugin for DeepSeek Harness (DSH), developed by the 7DGroup team for AI-assisted code review in any dsh session. Built on TypeScript + Cordis, it installs as a composable bundle and registers the 7d-code-reviewer skill with ctx.skills: a five-step review flow, critical/medium/minor severity grading, four-dimension scoring, and dual text + HTML report output. Zero core changes — install to enable, remove the bundle row to uninstall.


📌 Project Info

FieldValue
Author7DGroup
Version0.1.0-rc.5
RuntimeNode `^22.19.0
Peer dependencies@deepseek-ai/cordis · @deepseek-ai/dsh-skill · @deepseek-ai/dsh-invariants
Skill name7d-code-reviewer
Repositorygithub.com/7dgroup-ai/dsh-skill-7d-code-reviewer
LicenseMIT

🖼️ Plugin Effect

Skill invocation in a dsh session:

Invoking the 7d-code-reviewer skill for a code review

Sample HTML report generated from the pure-placeholder template:

Sample code review HTML report

Core Capabilities:

  • Template-driven mode — separation of concerns: SKILL.md decides what to review and how severe it is, templates/ only presents. The HTML report template stays pure placeholders; all placeholders must be filled, and every dynamic value is HTML-escaped.
  • Five-step review flow — accept the task → quick scan → line-by-line review (loading references/ on demand) → severity grading → report generation.
  • Three-level severity grading — 🔴 critical (must fix) / 🟡 medium (should fix) / 🟢 minor (optional polish).
  • Four-dimension scoring — code quality / security / performance / maintainability, each on a 1–10 scale, plus an overall score and an auto-generated summary.
  • Dual output — a text summary for quick reading, plus a full HTML report saved as code-review-report-{timestamp}.html.
  • Built-in knowledge base — coding standards, security checklist (SQL injection, XSS, authentication/authorization, sensitive-data leaks) and worked review examples, loaded on demand instead of bloating the prompt.
  • Zero core changes — pure composable bundle; no patches to the DSH core, safe to install and remove.

Use Cases:

  • Code review before commit / merge request
  • Security audit of existing code
  • Quality assessment before refactoring
  • Enforcing team coding standards
  • Any code quality question inside a dsh conversation

✅ Features

  • ✅ Five-step template-driven review flow
  • ✅ Three-level severity grading with fix suggestions
  • ✅ Four-dimension scoring rubric (code quality / security / performance / maintainability)
  • ✅ Text summary + HTML report dual output
  • ✅ Pure-placeholder HTML report template with mandatory filling rules
  • ✅ Documented HTML escaping rules for all filled content
  • ✅ On-demand knowledge base (coding standards / security checklist / review examples)
  • ✅ No executable scripts ship with the skill
  • ✅ Installable from GitHub (github: shorthand), npm or tarball
  • ✅ Git-install build is self-contained (prepare hook, transpile-only)

📂 Project Structure

dsh-skill-7d-code-reviewer/
├── src/                                # source code
│   ├── index.ts                        # Cordis plugin: registers the skill provider
│   └── invariant.ts                    # companion plugin: package ownership invariant
├── assets/7d-code-reviewer/            # skill resources shipped with the package
│   ├── SKILL.md                        # review logic + template selection
│   ├── references/                     # knowledge base, loaded on demand
│   │   ├── coding-standards.md         # naming rules, code complexity
│   │   ├── security-checklist.md       # SQL injection, XSS, auth, leaks
│   │   └── review-examples.md          # worked review examples
│   ├── templates/
│   │   └── report-template.html        # pure-placeholder HTML report
│   └── scripts/
│       └── html-report-generation.md   # HTML escaping rules for filled content
├── tests/                              # vitest suite
│   └── skill-7d-code-reviewer.spec.ts
├── screenshots/                        # README screenshots
│   ├── skills.png                      # skill invocation
│   └── report-preview.png              # sample HTML report
├── cordis.patch.yml                    # composition patch layer
├── tsdown.config.ts                    # build config (transpile-only)
├── 7dgroup-dsh-skill-7d-code-reviewer-0.1.0-rc.5.tgz   # prebuilt tarball
├── package.json
└── README.md

🚀 Quick Start

Prerequisites: dsh CLI, Node ^22.19.0 || >=24.0.0, pnpm 10+.

Install from within a dsh session (recommended)

The most direct way — just ask the agent in any dsh conversation, and it runs the install for you. Use the GitHub spec — the npm name @7dgroup/dsh-skill-7d-code-reviewer only works after the package is published:

安装插件 github:7dgroup-ai/dsh-skill-7d-code-reviewer

(Or in English: "Install the plugin github:7dgroup-ai/dsh-skill-7d-code-reviewer" — the agent executes the equivalent dsh plugin command through its session shell.)

For a git install the agent will hit the same pnpm allowBuilds gate and print the exact key to add to the profile's pnpm-workspace.yaml; after you add it, ask the agent to retry and the skill is enabled.

Install directly in dsh (CLI)

Run one command directly in dsh — the github: shorthand is the fastest way:

dsh plugin --profile <name> add github:7dgroup-ai/dsh-skill-7d-code-reviewer

The full URL form is equivalent:

dsh plugin --profile <name> add git+https://github.com/7dgroup-ai/dsh-skill-7d-code-reviewer.git

dsh plugin appends the bundle to the profile's dsh.profile.bundles, and the bundle's own patch layer mounts the skill-7d-code-reviewer row over the base composition.

pnpm blocks a git dependency's build scripts until explicitly allowed, so the first add fails. Copy the exact package key pnpm printed into the profile's pnpm-workspace.yaml, then re-run:

allowBuilds:
  '@7dgroup/dsh-skill-7d-code-reviewer@git+https://github.com/7dgroup-ai/dsh-skill-7d-code-reviewer.git#<sha>': true

(With the github: shorthand the key reads @7dgroup/dsh-skill-7d-code-reviewer@github:7dgroup-ai/dsh-skill-7d-code-reviewer#<sha> — always copy the exact key pnpm prints.)

Allowing a build means letting that package's code run on your machine at install time, outside any agent sandbox. Prefer pinning a commit (...#<sha>) so later pushes cannot silently change what runs.

Install from tarball (no build approval)

A prebuilt tarball is committed at the repository root — download it and install directly:

dsh plugin --profile <name> add ./7dgroup-dsh-skill-7d-code-reviewer-0.1.0-rc.5.tgz

Or once published on npm:

dsh plugin --profile <name> add @7dgroup/dsh-skill-7d-code-reviewer

Both forms ship prebuilt code and need no allowBuilds allowance.

Build and test

pnpm install
pnpm build   # tsdown; also runs as the `prepare` hook on git installs
pnpm test    # vitest

💡 Usage

The skill activates whenever you ask for a code review — either with the slash command or in natural language:

/7d-code-reviewer Review this module: ...

The five-step review flow:

StepWhat happens
1. Accept the taskTake the submitted code or file paths; determine the language and business context
2. Quick scanClassify the change (new feature / bugfix / refactor); locate the core files and key logic
3. Line-by-line reviewLoad the matching references on demand; check naming, security, performance and error handling
4. Severity grading🔴 critical — must fix · 🟡 medium — should fix · 🟢 minor — optional improvement
5. Report generationFill the placeholder HTML template; output the text summary plus code-review-report-{timestamp}.html

Output example (text summary)

✅ 优点
- 函数意图明确,返回用户数据

⚠️ 问题
🔴 严重:SQL 注入风险
  位置:get_user() 第 2 行
  描述:直接使用 f-string 拼接用户输入到 SQL 语句
  建议修复:使用参数化查询,如 cursor.execute("SELECT * FROM users WHERE id=?", [uid])

📊 总体评分:3/10
   代码质量: 5/10 | 安全性: 1/10 | 性能: 7/10 | 可维护性: 4/10

The full HTML report is saved to code-review-report-{timestamp}.html and the file path is reported back to you.

📊 Grading & Scoring Standards

Severity levels:

LevelMarkerDefinitionHandling
Critical🔴security vulnerability, crash riskmust fix
Medium🟡performance hazard, logic flawshould fix
Minor🟢naming, commentsoptional improvement

Dimension scoring (each on a 1–10 scale):

DimensionExcellent (8–10)Good (6–7)Needs work (4–5)Poor (1–3)
Code qualityclear naming, clean structure, no duplicationmostly compliant, minor issuesconfusing naming or high complexityviolates coding standards
Securityno risk, parameterized queries, full validationbasically safe, small flawssecurity hazardssevere vulnerabilities
Performanceefficient algorithms, caching, no N+1acceptableobvious problemssevere defects
Maintainabilitydocumented, modular, high test coveragemaintainablemissing comments/testshard to maintain

Overall score bands: 9–10 excellent · 7–8 good · 5–6 fair · 3–4 poor · 1–2 very poor (fix immediately).

📈 HTML Report

  • Score circle — overall score (1–10) with an auto-generated summary
  • Issue statistics bar — critical / medium / minor counts and good points
  • Dimension score cards — code quality / security / performance / maintainability
  • Issues grouped by severity — location, description and fix suggestion (with code sample)
  • Good points & improvement suggestions sections
  • Pure-placeholder template — all placeholders must be filled; every dynamic value is HTML-escaped per scripts/html-report-generation.md
  • Empty sections follow the no-content rule (e.g. "🎉 未发现严重问题!")

⚠️ Notes

  1. The provider contributes one fixed skill; no runtime customization.
  2. Report quality depends on the model following the placeholder-filling and HTML-escaping rules; nothing validates the generated report.
  3. The prepared build ships no type declarations; the dsh Loader loads the runtime entry only.
  4. The build is transpile-only (dts: false) with no lint or typecheck scripts — type errors surface in the editor/IDE.
  5. Commit messages in this repository follow the Simplified Chinese convention: 【类型】简短描述 (nine fixed type tags).

❓ FAQ

Q: Why does the first dsh plugin add fail? A: pnpm refuses to run build scripts of git dependencies until explicitly allowed. Copy the exact package key pnpm printed into the profile's pnpm-workspace.yamlallowBuilds, then re-run.

Q: How do I pin a specific commit? A: Append #<sha> to the spec, e.g. git+https://github.com/7dgroup-ai/dsh-skill-7d-code-reviewer.git#<sha> — later pushes cannot silently change what runs.

Q: How do I uninstall? A: Remove the bundle row from the profile's dsh.profile.bundles; no core patches are left behind.

Q: Can I install without approving builds? A: Yes — use the prebuilt tarball (committed at the repository root) or the npm package (once published); neither needs allowBuilds.

📄 License

MIT · Copyright (c) 2026 7DGroup