dsh-skills
August 25, 2026 · View on GitHub
面向 DeepSeek Harness / Claude Code / Kimi Code / Codex 等多 Agent 生态的轻量通用技能包。 纯文件形态、零依赖、零安装脚本——放进各自技能发现根即生效。
Reusable, lightweight skills for multiple AI coding agents: DeepSeek Harness, Claude Code, Kimi Code, Codex, and more. Plain files, zero dependencies.
兼容矩阵 / Compatibility
| 目标环境 | 技能目录 | 说明 |
|---|---|---|
| DeepSeek Harness | ~/.dsh/skills 或 ~/.agents/skills | DSH 原生,支持 watcher 热加载 |
| Claude Code | ~/.claude/skills | 标准 SKILL.md 格式,复制即用 |
| Kimi Code | ~/.agents/skills | 标准 SKILL.md 格式,复制即用 |
| Codex | ~/.codex/skills | 标准 SKILL.md 格式,复制后按环境调整路径 |
技能本体统一为标准
SKILL.md+ YAML frontmatter;不同编辑器仅“技能发现根目录”不同。 DSH 专用 CLI(dsh-skill.sh)仍可用于 DSH 环境,其他环境直接复制技能目录即可。
上传必看:Git 发布前兼容性检查
以后任何项目上传到 Git 前,必须先按此清单检查;不通过禁止上传。
1. 路径必须可移植
- 禁止出现本机绝对路径:
/Users/<用户名>、C:\Users\...、/home/... - 路径统一使用相对路径
./data/...、$HOME、Path.home()或环境变量 - 代码、文档、配置中不得出现私人目录名
2. 凭据必须清零
- 不得出现真实
password/secret/token/api_key/ 私钥 - 已泄露的账号密码必须删除
- 新增
.env.example,真实配置通过环境变量注入
3. 私有依赖必须剥离
- 不得 import 外部私有项目模块,例如
backend.*、etl.common.* - 不得出现内部私有项目名:
agent-harness、retro-skills-registry、Xj-rules、个人名等 - 仓库必须能独立
clone后按 README 安装并运行
4. 无关文件必须忽略
-
data/、__pycache__/、*.pyc、.DS_Store、node_modules/、*.tgz、*.zip不入库 - 数据库文件、模型权重、大体积临时文件不入库
5. 上传前必须执行校验
# 绝对路径扫描:应无命中
grep -RIn --exclude-dir=.git --exclude-dir=data -E '/Users/|C:\\Users|/home/' .
# 凭据扫描:不应出现真实密码/密钥
grep -RIn --exclude-dir=.git -iE 'password|secret|token|api[_-]?key' .
# 语法/配置校验
python3 -m py_compile $(find . -name '*.py' -not -path './.git/*')
- Python 语法检查通过
- YAML / JSON 解析通过
- 在全新目录
git clone后,按 README 跑通最小示例
技能清单 / Skills
| 技能 | 说明 |
|---|---|
parallel-dispatch | 并行调度与子分身机制总规则。≥2 个无依赖子任务默认主动并行 spawn 子分身;规模轴(免评估轻分身 / S 档 / M·L 档引擎级)× 数量轴(2~5 subagent 扇出 / ≥10 workflow 编排)双维决策;含场景自动匹配表、最小探针、母体合并校验时点分层、禁止清单。 |
archmap | 架构测绘 Agent(含 Python 引擎,自包含分发)。零参自动分流 full/lite;需求文本→精准影响面(文件/函数/路由级);diff 零 LLM 行级影响面 + 导入闭包 + 测试选择 + 变更台账;sync 增量同步基线并刷新 01~09 报告;ETL 规则注册表项目级可覆盖。以确定性计算替代全库通读,显著节约 tokens。 |
gate-switch | 通用概率执行门禁骨架(实证族 L2 引擎,零依赖)。治 LLM 三类顽疾:该做的没做 / 缺斤短两 / 伪造声称——把"声称 X 已满足"写成 spec JSON,引擎逐项机械核验,A 放行 / B 阻断列违例,判定权从模型移交脚本。7 检查原语(file_exists/json_field/glob_count/grep_count/mtime_after/script_exit 等),自带 8 个通用门禁实例(验收 verdict、测试证据、部署准入、模式分流等)+ L3 框架闸模板。新场景 = 写新 spec,引擎零改动。与 parallel-dispatch 的 dispatch_switch(路由族)互补。 |
Xj-agent(PM 全流程工作流)
通用、自包含的 PM 全流程研发调度骨架(13 节点:pm_bootstrap→spm→pm_prd_confirm→dpm→[ui_designer ∥ test_lead_design]→fe→be→pm_quality_gate→test_lead_full→ops→qa→process_audit→retro),默认引擎接线为同仓库 Xj-engine(engine.kernel.et / xj-engine CLI),引擎可插拔。各节点角色由 flow.yml 的 invoked_skills 指向对应技能(适配方按自身生态替换)。含:
pm/SKILL.md— 流程入口与引擎接线说明pm/flow.yml— 13 节点编排(节点拓扑 / 分支 / 状态机 / 交付物模板)pm/requirements.txt— 脚本运行依赖(PyYAML / jsonschema)pm/scripts/flow_kernel.py— 节点流转内核(规则全入参,success 回执才流转)pm/scripts/engine_preflight.sh— 引擎健康检查(默认xj-engine health)pm/scripts/verify_experience_writeback.sh— 经验固化机械校验
pip install -r Xj-agent/pm/requirements.txt
# 查看节点出口
python3 Xj-agent/pm/scripts/flow_kernel.py routes --rules Xj-agent/pm/flow.yml --node be
角色 agent 集(11 个,Xj-agent/agents/)
pm 工作流 11 个角色技能随包分发,下载即拥有完整 pm 工作流:
| 角色技能 | 承接节点 |
|---|---|
senior-pm-agent | spm / pm_prd_confirm(需求分析+PRD确认+任务拆解) |
detail-product-manager | dpm 交互设计 |
ui-designer | ui_designer |
frontend-development | fe |
backend-engineer | be |
operation-deployment | ops |
test-lead | 测试设计/冒烟/全量/验收/合规 |
whitebox-coverage | 白盒路 |
api-test-engineer | 接口路 |
ui-test-engineer | UI 路 |
retro-skill-dispatcher | retro 收尾复盘 |
均做了私有依赖剥离(无 agent-harness / retro-skills-registry / 绝对路径),可独立安装复用。
安装 / Install
一键安装(推荐,无需先 clone)/ One-liner
# 列出全部技能
curl -fsSL https://raw.githubusercontent.com/xu-jin-cs/dsh-skills/main/scripts/dsh-skill.sh | bash -s -- list
# 安装指定技能(默认符号链接进 ~/.dsh/skills,DSH watcher 热加载即生效)
curl -fsSL https://raw.githubusercontent.com/xu-jin-cs/dsh-skills/main/scripts/dsh-skill.sh | bash -s -- install archmap
# 安装全部技能 + 自动装依赖
curl -fsSL https://raw.githubusercontent.com/xu-jin-cs/dsh-skills/main/scripts/dsh-skill.sh | bash -s -- install --all --with-deps
首次运行会自动把发布仓浅克隆到 ~/.dsh/dsh-skills(可用 DSH_SKILLS_HOME 改位置),之后所有命令在本地仓执行。
已 clone 仓库 / Already cloned
仓库根目录自带安装入口,无需记忆任何命令:
git clone https://github.com/xu-jin-cs/dsh-skills.git
cd dsh-skills
./install.sh # 交互式选择(列清单,输序号即可)
./install.sh archmap # 安装指定 agent(引擎类技能)
./install.sh parallel-dispatch # 安装指定规则(规则类技能)
./install.sh archmap parallel-dispatch # 一次装多个
./install.sh --all # 全部安装
支持 --copy(拷贝模式)、--target DIR(换发现根,如项目级 .dsh/skills)。本质是 scripts/dsh-skill.sh 的友好外壳:
scripts/dsh-skill.sh 子命令:
| 命令 | 作用 |
|---|---|
list | 列出发布仓全部技能 |
install <技能...|--all> | 安装(符号链接进发现根);--copy 拷贝模式;--target DIR 切换目标根(如项目级 .dsh/skills);--with-deps 自动装 requirements |
uninstall <技能...> | 卸载 |
update | git pull 同步上游(符号链接模式即时生效) |
doctor | 体检:发现根、断链、SKILL.md 完整性、依赖环境 |
DSH 按以下顺序发现技能(命中任意一级即生效):
项目/.dsh/skills → 项目/.agents/skills → ~/.dsh/skills → ~/.agents/skills → bundled
手动安装(不用 CLI)
git clone https://github.com/xu-jin-cs/dsh-skills.git ~/dsh-skills
ln -s ~/dsh-skills/parallel-dispatch ~/.dsh/skills/parallel-dispatch
ln -s ~/dsh-skills/archmap ~/.dsh/skills/archmap # 含 Python 引擎的技能
pip3 install -r ~/dsh-skills/archmap/requirements.txt # 可选,缺失时自动回退本地哈希向量化
无需重启:DSH 的技能 watcher 会热加载新技能。之后命中"并行 / 分身 / 批量 / 多任务"等场景即自动触发,也可显式 /parallel-dispatch 调用。
No restart needed — DSH's skill watcher hot-reloads new entries.
设计原则 / Principles
- 轻量化:规则类技能为单文件
SKILL.md+ YAML frontmatter(name+description),无代码、无依赖;引擎类技能(如archmap)自包含分发,依赖显式声明于各自requirements.txt; - 通用:不含任何引擎私有逻辑,不绑定特定后端;标准
SKILL.md格式,可被 DSH / Claude Code / Kimi Code / Codex 等加载;他人的 engine 零冲突; - 自动触发:触发词与场景写在
description中,由 DSH 注入会话目录做场景匹配,命中即主动加载,无需显式指令。
方法论 / Methodology
本仓库技能的治理哲学与 27 个实战案例复盘:《给 LLM 的口头承诺装上机械门禁》(English Version)——强制填充门元方法、L1/L2/L3 三档门禁、骨架冻结纪律、举一反三泛化闸。
License
MIT
自动发布由 launchd WatchPaths 驱动,变更后约 60~90s 自动 commit+push
dsh-skills (English)
Reusable, lightweight skills for multiple AI coding agents: DeepSeek Harness, Claude Code, Kimi Code, Codex, and more. Plain files, zero dependencies — drop them into the skill discovery root of your agent and they work.
Compatibility
| Agent | Skill directory |
|---|---|
| DeepSeek Harness | ~/.dsh/skills or ~/.agents/skills |
| Claude Code | ~/.claude/skills |
| Kimi Code | ~/.agents/skills |
| Codex | ~/.codex/skills |
Skills
| Skill | Description |
|---|---|
parallel-dispatch | Master rules for parallel dispatch & sub-agent clones. ≥2 independent subtasks trigger parallel fan-out by default; two-axis decision matrix (scale: light clone / task-breakdown / engine-level × count: subagent / grouped / workflow); includes scene auto-matching, minimal probe, merge checkpoints, and a mechanical SPDT-style dispatch_switch (A/B verdict, no handwritten decisions, full audit log). |
archmap | Architecture cartography agent (self-contained Python engine). Zero-arg full/lite auto-routing; requirement text → precise impact analysis (file/function/route level); diff mode: zero-LLM line-level impact + import-closure + test selection + change ledger; sync incremental baseline refresh. Deterministic computation instead of full-repo reading — massive token savings. |
gate-switch | Universal probabilistic-execution gate (evidence-family L2 engine, zero deps). Cures three LLM chronic failures: skipped steps / half-done checklists / fabricated "done" claims. Write what must be true as a spec JSON; the engine mechanically verifies each check — A passes, B blocks with violations as the reason. Judgment moves from the model to a script. 7 frozen check primitives (file_exists / json_field / glob_count / grep_count / mtime_after / script_exit …), 8 ready-made generic gates (acceptance verdict, test evidence, deploy admission, mode routing, …) + an L3 framework-gate template. New scenario = new spec, zero engine changes. Complements dispatch_switch (routing family). |
Install
One-liner (recommended, no clone needed)
# List all skills
curl -fsSL https://raw.githubusercontent.com/xu-jin-cs/dsh-skills/main/scripts/dsh-skill.sh | bash -s -- list
# Install a specific skill (symlinked into ~/.dsh/skills, hot-reloaded by DSH's watcher)
curl -fsSL https://raw.githubusercontent.com/xu-jin-cs/dsh-skills/main/scripts/dsh-skill.sh | bash -s -- install gate-switch
# Install everything + auto-install dependencies
curl -fsSL https://raw.githubusercontent.com/xu-jin-cs/dsh-skills/main/scripts/dsh-skill.sh | bash -s -- install --all --with-deps
First run shallow-clones the repo to ~/.dsh/dsh-skills (override with DSH_SKILLS_HOME); all later commands run locally.
Already cloned
git clone https://github.com/xu-jin-cs/dsh-skills.git
cd dsh-skills
./install.sh # interactive picker
./install.sh archmap # install a specific skill
./install.sh --all # everything
CLI subcommands (scripts/dsh-skill.sh): list / install (--copy, --target DIR, --with-deps) / uninstall / update / doctor.
DSH discovers skills in order (first hit wins):
<project>/.dsh/skills → <project>/.agents/skills → ~/.dsh/skills → ~/.agents/skills → bundled
No restart needed — DSH's skill watcher hot-reloads new entries.
Principles
- Lightweight — rule-type skills are a single
SKILL.md+ YAML frontmatter, no code, no deps; engine-type skills (archmap, gate-switch) are self-contained with explicitrequirements.txt. - Universal — no private engine logic, no backend lock-in; standard
SKILL.mdformat works with DSH, Claude Code, Kimi Code, Codex, and more; zero conflicts with your own engine. - Auto-trigger — triggers live in each skill's
description; DSH injects them into the session catalog for scene matching.
Methodology
The governance philosophy behind these skills, plus a 27-case battle retrospective: Mechanical Gates for LLM's Verbal Promises (中文版) — the Mandatory-Completion Gate meta-method, L1/L2/L3 gate levels, skeleton-freeze discipline, and the "1 proven case + N named siblings" generalization gate.
License
MIT