@liuzhengdongfortest

cs-refactor-ff enables fast-track refactoring for low-risk, single-function changes by skipping design docs and checklists, requiring only three entry checks (behavior preservation, scope <100 lines, test coverage) and validation through test runs.

cs-refactor guides code optimization through a gated three-phase process (scan, design, apply) that maintains behavior equivalence while improving structure, performance, or readability. it routes away architectural changes, new features, and bugs to dedicated workflows.

cs-issue-analyze reads problem reports and codebase to identify root causes, assess impact scope, and propose 2-3 repair strategies for user approval before implementation. it enforces code inspection over speculation and defers execution decisions to the user.

cs-issue-report captures user-reported problems into structured issue records and routes them to fast-track or standard analysis based on code inspection. focuses on observable symptoms, defers root cause analysis to stage 2.

cs-issue-fix executes targeted bug repairs following confirmed root cause analysis, enforcing scope discipline and verification before documenting fixes in {slug}-fix-note.md. two entry paths: standard from analysis stage, fast-track from report with oral confirmation.

cs-decide formalizes tech selections, architecture decisions, constraints, and coding conventions into permanent decision records. guides users through identification, extraction, drafting, and archival with explicit supersede vs update semantics.

cs-explore codifies repository investigation into searchable evidence trails. it structures code reading into question/module-overview/spike types, enforces evidence-first documentation with file:line citations, and prevents exploratory work from becoming throwaway analysis.

cs-req maintains capability vision documents under .codestable/requirements/ across three modes (draft, backfill, update), triggered at design inception or acceptance completion to separate requirement intent from implementation detail.

generates reference documentation for library public surfaces (components, functions, commands) by reading source code directly, with checklist tracking for single or batch entries. focuses on extracting type signatures and defaults rather than usage patterns.

cs-arch maintains a living, read-first system map in .codestable/architecture/ documenting current state across three modes (update, check, backfill). triggered by architecture refresh requests, the skill enforces that design stays grounded in implemented code rather than speculative futures.

cs-note appends project micro-knowledge to .codestable/attention.md via fixed sections, capturing short-lived-but-essential facts like build steps, service startup requirements, and command aliases that ai needs on every launch.

cs-feat routes new feature requests to appropriate sub-skills (brainstorm, design, fastforward, implement, acceptance) based on artifact state. pure router, no code or doc generation.

cs-feat-impl orchestrates feature implementation in paradigm-dimensional steps derived from design phase, enforcing checkpoint discipline through mandatory plan validation, reflection checks, and structured reporting to prevent scope creep and undocumented decisions.

cs-trick prescribes a workflow for capturing reusable code patterns, library usage, and techniques into a searchable reference indexed by type (pattern/library/technique). emphasis on code archaeology before documentation.

系统审计——从代码中主动发现 bug 隐患、安全漏洞、性能问题、可维护性债务和架构偏离，产出批量发现清单。触发：用户说"审查系统"、"审计代码"、"扫描问题"、"找找 bug"、"有什么可以优化的"。

把新仓库或有零散文档的仓库接入 CodeStable 体系，两条路径自动判断：空仓库从零搭骨架，已有文档走审计 + 迁移映射。触发：用户说"在这个项目里用 CodeStable"、"搭 CodeStable 结构"、"初始化 CodeStable"、"迁移到 CodeStable"。

feature 流程阶段 3——验收闭环：对照 design 核实现 + 回写 architecture / requirement / roadmap，最后产出 {slug}-acceptance.md。触发：用户说"功能写完了验收一下"、"做最后检查"、"准备 merge"、"出验收报告"。前置依赖…

想法还模糊时的讨论入口，做分诊后路由到 feature-design / feature-brainstorm / roadmap。AI 是思考伙伴不是记录员。触发：用户说"有个想法还没想清楚"、"先 brainstorm 一下"、"聊一聊这块"、"方向还在摇摆"。不处理 bug 和重构。

CodeStable 工作流根入口，介绍体系全貌并把诉求路由到对应 cs-* 子技能。触发：用户只输入 `cs`、说"介绍一下 codestable"、"该用哪个技能"、"不知道用哪个"，或诉求还很开放未收敛。本技能只做路由不做事。

通过 Chrome 扩展控制真实浏览器。需要访问网页、抽取网页数据、点击按钮、填写表单、执行浏览器自动化、提取渲染后的组件证据，或以程序方式操作页面时使用。通过 DOM diff、简化 HTML 和 component evidence pack 返回节省 token 的结构化结果。适用于 browser…

feature 流程的超轻量通道——不写 design / checklist 直接动手，但先指引 AI 查 CodeStable 知识库再开工。触发：用户说"快速模式"、"fastforward"、"别那么多步骤"、"直接开干"，且需求小到不值得走 design 流程。

修 bug 的子流程入口，把"发现问题"走到验证修复闭环，留下 report / analysis / fix-note 三份文件。触发：用户说"修 bug"、"有个问题"、"修复 XX"。只做路由，根据已有产物走 report / analyze / fix。简单问题走快速通道。

把踩过的坑或好做法沉淀成可检索的 learning 文档，两条轨道 pitfall（坑）/ knowledge（默认做法）。触发：用户说"沉淀知识"、"learning"、"把这次经验记下来"，或 acceptance / fix 收尾时推送。

写或更新对外指南文档——开发者指南（dev-guide）和用户指南（user-guide），产物在项目 docs/ 目录。任务导向（怎么用 X 做 Y），与 libdoc 的零件参考不同。触发：用户说"写文档"、"开发者指南"、"用户指南"，或 feature-acceptance 收尾时推送。

把"大到塞不进单个 feature"的需求做成完整事前规划：概设 + 接口契约 + 子 feature 拆解清单，放在 `.codestable/roadmap/{slug}/`。两种模式 new / update。触发：用户说"我想要一个 X 系统"、"帮我把这块需求拆一下"、"开一份 roadmap"，或…

feature 流程阶段 1——为新功能起草 {slug}-design.md 作为后续实现和验收的唯一输入，拍板后抽出 checklist。触发：用户说"开始设计方案"、"写 design doc"、"准备实现 XX"，前提是已知道做什么、为谁、怎么算成功。