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
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
| Field | Value |
|---|---|
| Author | 7DGroup |
| Version | 0.1.0-rc.5 |
| Runtime | Node `^22.19.0 |
| Peer dependencies | @deepseek-ai/cordis · @deepseek-ai/dsh-skill · @deepseek-ai/dsh-invariants |
| Skill name | 7d-code-reviewer |
| Repository | github.com/7dgroup-ai/dsh-skill-7d-code-reviewer |
| License | MIT |
🖼️ Plugin Effect
Skill invocation in a dsh session:

Sample HTML report generated from the pure-placeholder template:

Core Capabilities:
- Template-driven mode — separation of concerns:
SKILL.mddecides 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 (
preparehook, 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:
| Step | What happens |
|---|---|
| 1. Accept the task | Take the submitted code or file paths; determine the language and business context |
| 2. Quick scan | Classify the change (new feature / bugfix / refactor); locate the core files and key logic |
| 3. Line-by-line review | Load 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 generation | Fill 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:
| Level | Marker | Definition | Handling |
|---|---|---|---|
| Critical | 🔴 | security vulnerability, crash risk | must fix |
| Medium | 🟡 | performance hazard, logic flaw | should fix |
| Minor | 🟢 | naming, comments | optional improvement |
Dimension scoring (each on a 1–10 scale):
| Dimension | Excellent (8–10) | Good (6–7) | Needs work (4–5) | Poor (1–3) |
|---|---|---|---|---|
| Code quality | clear naming, clean structure, no duplication | mostly compliant, minor issues | confusing naming or high complexity | violates coding standards |
| Security | no risk, parameterized queries, full validation | basically safe, small flaws | security hazards | severe vulnerabilities |
| Performance | efficient algorithms, caching, no N+1 | acceptable | obvious problems | severe defects |
| Maintainability | documented, modular, high test coverage | maintainable | missing comments/tests | hard 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
- The provider contributes one fixed skill; no runtime customization.
- Report quality depends on the model following the placeholder-filling and HTML-escaping rules; nothing validates the generated report.
- The prepared build ships no type declarations; the dsh Loader loads the runtime entry only.
- The build is transpile-only (
dts: false) with no lint or typecheck scripts — type errors surface in the editor/IDE. - 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.yaml → allowBuilds, 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