Agent 与会话技能

dsh-skills

arch3rPro/dsh-skills

Generalized development-specification skills distilled from DeepSeek Harness's engineering conventions — defensive patterns, code conventions, testing tiers, prose standards, decision records, review and PR discipline.

查看仓库项目主页

Stars: 5
Forks: 0
Issues: 0
更新: 1 个月前

PROJECT TOPICS

项目标签

agent-orchestration agent-skills claude-code claude-skills codex-skill dsh dsh-plugin dsh-skill

INSTALL REFERENCE

安装参考

未验证

dsh plugin --profile web add github:arch3rPro/dsh-skills

在 DSH 中安装

该命令指向仓库当前默认分支；尚无绑定当前 commit 的完整验证结果。

PROJECT README

README

dsh-skills — coding-agent skills distilled from DeepSeek Harness, organized across architecture, testing, documentation, and process

DeepSeek-Harness-Skills

A set of coding-agent skills distilled from the engineering conventions of DeepSeek Harness — executable standards your coding agent applies while coding, reviewing, testing, or documenting.

English · 简体中文

Get started

Install the set with the skills.sh installer, then pick the skills and agents you want:

npx skills@latest add arch3rPro/dsh-skills

Or just read one — each SKILL.md is a self-contained executable standard. Pick the discipline that matches what you're doing: code-conventions, testing-tiers, prose-standard, or any in the set.

The set

Ten skills across four categories. Every one is model-invoked — the agent reaches for it automatically when a trigger fires.

Skill	Category	What it does
defensive-patterns	architecture	bug-class rules for lifecycle, concurrency, teardown, untrusted I/O
code-conventions	architecture	reversible side effects, explicit-over-implicit, config-over-hardcoding, branded ids
extension-points-and-seams	architecture	build a seam only when a capability has multiple implementations or consumers
testing-tiers	testing	pick the right tier, test the real entry path, verify the world not the self-report
prose-standard	documentation	contract-first prose: preserve propositions, drop reasoning transcripts
documentation-placement	documentation	one home per fact; tutorials vs references; budgets as guardrails
decision-records	process	ADRs with classification, lifecycle, mandatory alternatives
minimal-evidence-checks	process	the smallest check set that covers the diff; fix or explain failures
code-review	process	two axes (standards + spec); substantiated blockers over nits
pr-history-hygiene	process	lease-protected rewrites, deliberate labels, native stack merges

Beta (in development)

Five further disciplines are grounded in DeepSeek Harness and under development. They are npx skills-installable but not yet part of the official plugin bundle.

Skill	Category	What it does
postmortem	process	backward-looking failure record: executive summary, root cause, guardrails
runtime-invariants	testing	assert only authoritative state, never service/method presence; fail loud
responding-to-review-on-a-stack	process	fix on the introducing PR, propagate up-stack, re-audit after rewrites
event-contract-matrix	documentation	who produces and who listens to each event, bypass sites included
module-layering	architecture	derive and enforce the module dependency graph

Where they come from

Every skill is distilled from a real convention in DeepSeek Harness and adapted so it works in any codebase — not only inside the harness it came from.

How it's packaged

Each skill is one <category>/<name> directory carrying three files in a single change:

SKILL.md — the executable standard (Claude Code / any agent harness)
agents/openai.yaml — Codex metadata and invocation policy
docs/<category>/<name>.md — a human-facing docs page

skills/
  architecture/   how code is shaped and how failure is handled
  testing/        how behavior is verified
  documentation/  how prose and docs are written
  process/        the workflow around code — decisions, review, checks, git

Invocation design

Skills are either model-invoked (the agent reaches for them automatically when a trigger fires) or user-invoked (only a human typing the skill's name can fire them — zero context load, but the human is the index). All ten skills here are model-invoked.

Installation

The two routes below are alternatives, not add-ons — pick one.

`npx skills` — any agent (Claude Code, Codex, …)

Install one skill at a time (or the whole set from Get started):

npx skills@latest add arch3rPro/dsh-skills --skill=code-review

npx skills pulls from GitHub — publish the repo first, then run the command above.

Claude Code — the plugin

Add this repo as a Claude Code marketplace, then install the plugin:

claude plugins marketplace add arch3rPro/dsh-skills
claude plugins install dsh-skills@dsh-skills

The plugin manifest lives in .claude-plugin/ (marketplace.json + plugin.json).

Acknowledgments

DeepSeek Harness — an open-source agent harness developed by DeepSeek AI.
mattpocock/skills — Matt Pocock's agent skills for real engineering.

CLASSIFICATION EVIDENCE

分类依据

项目类型技能

功能分类Agent 与会话

规则置信度高

系统优先读取 GitHub Topics，再与站内分类词典和词根规则比对。当前命中： agent-skills、claude-skills、codex-skill、dsh-skill、agent-orchestration、claude-code。