dsh-review-skills
by ben7am1n
DeepSeek Harness 的工程纪律技能包——五个实战检验过的流程以内置 skill provider 交付:代码评审、化简、先计划后执行等
Engineering-discipline skill pack for DeepSeek Harness — five battle-tested procedures delivered as a bundled skill provider: code review, simplification, plan-then-execute,…
安装
dsh plugin --profile web add github:ben7am1n/dsh-review-skillsGitHub 源码安装:首次需按提示配置 allowBuilds 构建授权后重试
安装与环境配置指引、插件开发教程见 DSH 中文社区文档 ↗
安装即在你的机器上以你的权限运行第三方代码——它可读写文件、使用凭据、访问网络,DSH 的工具审批不会为插件代码加沙箱。「检测到 manifest」仅代表发现 dsh.bundle / dsh.plugin 清单,不构成兼容性或安全审查;安装前请审阅源码,不熟悉的插件先在不含密钥的环境试用。
README
目录
Engineering-discipline skill pack for DeepSeek Harness — five battle-tested procedures delivered as a bundled skill provider: code review, simplification, plan-then-execute, test-first, and conflict resolution.
Overview
dsh-review-skills packages the discipline that senior engineers apply by habit into model-visible skills. Once installed, the harness discovers five SKILL.md bundles through the standard ctx.skills registry and can invoke them when a task matches.
The skills
| Skill | When to use |
|---|---|
code-review |
Before merging: review a diff for correctness, test coverage, edge cases, needless complexity, and security. Produces severity-ranked findings. |
simplify |
After a change lands: find and remove accidental complexity — clearer names, fewer layers, less code that does the same thing. |
plan-then-execute |
Before implementing anything non-trivial: write a plan, get it accepted, then execute it. |
test-first |
TDD discipline: red → green → refactor, with honest tests and no fake green. |
resolve-conflict |
When a merge or rebase hits conflicts: understand both sides, resolve deliberately, verify. |
Who is it for?
- Solo developers who want a senior-engineer checklist without hiring one.
- Teams that want consistent review and planning standards across sessions.
- Anyone building on the dsh skill system who wants a reference-quality skill pack to copy.
Compatibility
- Built and verified against
@deepseek-ai/dsh@0.1.0-rc.6/@deepseek-ai/cordis@^4.0.1/@deepseek-ai/dsh-skill@^0.1.0-rc.6. - Last verified: 2026-08-14.
- Requires a dsh composition that mounts the
ctx.skillsservice (thedsh-basebundle ships it). - dsh is in developer preview; the skill frontmatter contract (
name/description/whenToUse/ invocation flags) may evolve. The included unit tests pin the format this pack targets.
Install / Uninstall
Install into a dsh profile (local checkout):
cd /path/to/deepseek-harness
pnpm dsh plugin --profile web add /path/to/dsh-review-skills
From GitHub (source install — pnpm runs the prepare script, so allow it once):
pnpm dsh plugin --profile web add github:<you>/dsh-review-skills
From npm (once published):
pnpm dsh plugin --profile web add dsh-review-skills
Uninstall:
pnpm dsh plugin --profile web remove dsh-review-skills
Quick start
Install the bundle, restart dsh, then simply ask:
Use the code-review skill on the current diff before we merge.
or mention a skill by name in context:
Plan-then-execute: refactor the auth module, then implement.
The harness resolves the skill through ctx.skills and injects its procedure when the model invokes it.
Configuration
None. The plugin registers its provider with no config; the skills and their routing metadata live in skills/*/SKILL.md frontmatter (name, description, whenToUse, disable-model-invocation, user-invocable).
Permissions & data
- Read-only: the plugin reads its own packaged
skills/directory; it writes nothing and touches no user files. - No network: all five skills are static Markdown bundled in the package.
- Model-visible content: skill bodies are injected into the model context only when the skill is invoked, per the harness skill catalog policy.
Troubleshooting
| Symptom | Cause | Fix |
|---|---|---|
| Skills not discoverable | ctx.skills service not mounted in the composition |
Verify the profile loads dsh-base (which ships the skill service) |
missing YAML frontmatter at boot |
A packaged SKILL.md lost its --- block |
Rebuild/reinstall; the tests (pnpm test) catch this before publish |
| Skill does not apply to the conversation | The model decided it was not relevant | Ask explicitly: "use the code-review skill" |
| Stale skill content | Old lib/ after editing skills/ |
Rebuild (pnpm run build) and reinstall; skill bodies load lazily from disk |
Development
pnpm install
pnpm run build # tsc → lib/
pnpm run test # vitest: frontmatter contract + 5 bundled skills present
Structure:
skills/<name>/SKILL.md— one skill per directory; frontmatter drives routing, body carries the procedure.src/skills.ts— frontmatter parser and the bundled provider (lazy body loading).src/index.ts— plugin entry registering the provider onctx.skills.tests/— pins the frontmatter contract and asserts all five bundles are parseable.
Design notes:
- Follow the existing skill-file conventions in the dsh repo (
@deepseek-ai/dsh-skilltypes:SkillCandidate,SkillDefinition,SkillProvider) — this pack mirrors thedsh-skill-badgeprecedent. - Keep each skill body focused: when-to-use, procedure steps, pitfalls, verification. A skill that tries to cover everything covers nothing.
License & security
MIT. Report security issues privately via the repository's security advisory. The pack ships static Markdown only; it executes no code of its own.
原始 README: https://github.com/ben7am1n/dsh-review-skills/blob/main/README.md ↗
同类插件
查看全部 →vibe-skills
vibe 编码技能合集(含快速开始、CI 证明、发布说明与文档)

Aegis
面向编码 Agent 的软件工程方法包,提供基线优先规划、系统化调试、提示词卫生、完成前验证,以及修复/退役双轨跟踪技能。

J-Space-Cognition-Suite-V3.6
A model-agnostic inference-time control suite for deep reasoning, long-horizon work, tool use, verification, and recovery. 面向深度推理、长程任务、工具调用、验证与恢复的模型不可知推理时控制套件。

MisakaNet
失败恢复记忆库:从真实工程会话中搜索和记录失败恢复教训,支持 BM25 + 语义 RAG 检索和知识库管理。

dsh-find-plugins
对 DSH 说一句「有没有插件能……」,它就会从全 GitHub 的 dsh-plugin topic 里找出候选,解释差别,等你选好以后再安装和验证。

humanizer-ru
AI代理技能:识别并清除俄语文本中的机器生成痕迹。38种模式,39个带证据注册的regex标记,盲对偶运行,C2PA/EXIF/XMP文件元数据清理