Back to Details

easysdd-refactor

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

easysdd-refactor

easysdd-refactor

AI 自己重构代码有两个稳定的失败模式：一是不知道模块的真实需求和约束，改出来的东西功能不等价；二是一次吞掉的范围超过上下文承载，改到后面忘了前面的约束。这条流程在"想优化"和"动手改"之间塞了一份扫描清单 + 方法库，让 AI 只接自己能稳定做对的活，剩下的老实停下来让路。
整套流程：
scan（扫出优化点清单）→ design（和用户定做哪几条、按什么顺序）→ apply（逐条执行，每步人工放行）
核心纪律：行为等价是底线。一旦这次动作会改变外部可观察行为，不走 refactor，走 feature（需求变）或 issue（bug 修）。
There are two stable failure modes when AI refactors code independently: first, it fails to understand the real requirements and constraints of the module, resulting in functionally inequivalent changes; second, the scope it takes on exceeds the context capacity, causing it to forget previous constraints later in the process. This process inserts a scan checklist + method library between "wanting to optimize" and "starting modifications", allowing AI to only undertake tasks it can stably execute correctly, and pause honestly for the rest.
Full workflow:
scan (generate optimization checklist) → design (confirm which items to do and the order with the user) → apply (execute item by item, with manual approval for each step)
Core Discipline: Behavioral equivalence is the bottom line. If an action will change externally observable behavior, do not use the refactor workflow; use feature (for requirement changes) or issue (for bug fixes) instead.

Fastforward 模式（小重构走这里）

Fastforward Mode (Use this for small refactors)

改动明显很小——单函数、单组件、1-3 处优化、有测试可自证、不需要 HUMAN 目视——走完整三阶段太重。触发 
easysdd-refactor-fastforward
：AI 直接识别、一次对齐、原地改、跑测试自证，不产出 scan / design / checklist。
触发信号：用户说"小重构"、"快速重构"、"简单优化下 XX 函数"、"直接改"、"别那么多步骤"。
什么时候别走 fastforward：
  • 改动跨 > 1 个文件
  • 预计动点超过 3 处
  • 需要目视验证（前端效果、性能感知）
  • 改到公开接口（要走 Parallel Change）
  • 没有测试覆盖
  • 跨模块
遇到这些就劝用户走标准流程。fastforward 开干后发现变复杂，切回完整流程从 scan 开始。
For obviously small changes — single function, single component, 1-3 optimization points, verifiable with tests, no need for human visual inspection — the full three-stage process is overkill. Trigger 
easysdd-refactor-fastforward
: AI directly identifies issues, aligns changes in one go, modifies in-place, runs tests to self-verify, and does not generate scan / design / checklist outputs.
Trigger signals: User says "small refactor", "quick refactor", "simply optimize XX function", "modify directly", "skip the steps".
When not to use fastforward:
  • Changes span > 1 file
  • Expected modification points exceed 3
  • Visual verification is required (frontend effects, performance perception)
  • Modifications involve public interfaces (need to use Parallel Change)
  • No test coverage
  • Cross-module changes
In such cases, advise the user to follow the standard process. If the fastforward process becomes complex during execution, switch back to the full process starting from scan.

文件放哪儿

File Placement

refactor 产物聚在 
easysdd/refactors/
下，每次独立目录：
easysdd/
└── refactors/
    └── {YYYY-MM-DD}-{slug}/
        ├── {slug}-scan.md              ← 阶段 1 产出的优化点清单
        ├── {slug}-refactor-design.md   ← 阶段 2 的执行方案（选哪几条、顺序、验证）
        ├── {slug}-checklist.yaml       ← 阶段 2 生成，阶段 3 推进
        └── {slug}-apply-notes.md       ← 阶段 3 的执行记录（每步做了什么、验证结果、偏离）
目录命名和 feature / issue 一致：
YYYY-MM-DD-{英文 slug}
，日期是首次创建当天、定了不动；slug 用小写字母数字连字符，短到能一眼看出改的是什么（
user-form-split
、
export-perf
这种）。
为什么单独开目录而不是混进 features/：refactor 产物是"代码当前状态的扫描 + 执行记录"，时效性强，价值随时间衰减；feature 产物是"这个能力为什么这样设计"，时效性弱。归档逻辑不一样，混一起后面找不到。
Refactor outputs are aggregated under 
easysdd/refactors/
, with an independent directory for each refactor:
easysdd/
└── refactors/
    └── {YYYY-MM-DD}-{slug}/
        ├── {slug}-scan.md              ← Optimization checklist generated in Phase 1
        ├── {slug}-refactor-design.md   ← Execution plan from Phase 2 (selected items, order, verification)
        ├── {slug}-checklist.yaml       ← Generated in Phase 2, used to drive Phase 3
        └── {slug}-apply-notes.md       ← Execution records from Phase 3 (what was done in each step, verification results, deviations)
Directory naming follows the same rule as feature / issue: 
YYYY-MM-DD-{English slug}
. The date is set on the day of initial creation and remains unchanged; the slug uses lowercase letters, numbers, and hyphens, short enough to clearly indicate what is being modified (e.g., 
user-form-split
, 
export-perf
).
Why use a separate directory instead of mixing with features/: Refactor outputs are "scan + execution records of the current code state", which have strong timeliness and their value decays over time; feature outputs are "why this capability is designed this way", which have weak timeliness. Their archiving logic is different, so mixing them will make it hard to find content later.

三个阶段

Three Phases

阶段子流程产出谁主导
1 scan扫出优化点清单{slug}-scan.mdAI 扫代码 + 做前置检查，用户勾选
2 design定执行方案{slug}-refactor-design.md + {slug}-checklist.yamlAI 起草，用户整体 review
3 apply逐条执行代码改动 + {slug}-apply-notes.mdAI 执行，每步人工放行
阶段之间有 checkpoint。scan 的清单用户不勾选不进 design；design 用户不放行不动代码；apply 里标了 HUMAN 验证的条目人不点头不推进下一步。
PhaseSub-processOutputLead Role
1 scanGenerate optimization checklist{slug}-scan.mdAI scans code + performs pre-checks, user selects items
2 designFinalize execution plan{slug}-refactor-design.md + {slug}-checklist.yamlAI drafts, user conducts overall review
3 applyExecute item by itemCode changes + {slug}-apply-notes.mdAI executes, manual approval for each step
Checkpoints exist between phases. The scan checklist cannot proceed to design until the user selects items; code cannot be modified until the design is approved by the user; items marked for HUMAN verification in apply cannot proceed to the next step without user confirmation.

阶段 1：scan（扫出优化点清单）

Phase 1: scan (Generate Optimization Checklist)

先跑前置检查（7 条），命中就停

First run pre-checks (7 items), stop if any are triggered

动笔扫之前先跑一遍前置检查。命中任何一条就中止 scan，给出路由建议，不要硬凑清单。7 条检查和输出格式见同目录 
reference/refusal-routing.md
。
零条合法输出——扫完真的没发现值得做的就老实说，不要凑。
Run a pre-check before starting the scan. If any item is triggered, abort the scan and provide routing suggestions; do not force a checklist. The 7 checks and output format are in 
reference/refusal-routing.md
in the same directory.
Zero valid outputs — if no worthy optimizations are found after scanning, state this honestly; do not pad the list.

扫描范围锁定

Lock Scan Scope

进 scan 前先和用户确认一件事：这次扫哪些文件。默认规则：
  • 用户点名了具体文件/组件 → 就扫那些
  • 用户说"这个页面" → 扫该页面入口组件 + 它直接 import 的内部模块，不追公共依赖
  • 用户说"这个模块" → 扫该模块目录下的文件，不追出模块边界
  • 范围 > 15 文件或 > 3000 行 → 触发第 6 条前置检查，请用户先缩范围
范围里要包含测试文件（用来判断第 2 条前置检查的测试覆盖情况）。
Before starting the scan, confirm one thing with the user: which files to scan this time. Default rules:
  • If the user specifies specific files/components → scan only those
  • If the user says "this page" → scan the entry component of the page + its directly imported internal modules, do not trace public dependencies
  • If the user says "this module" → scan files in the module directory, do not cross module boundaries
  • If the scope includes >15 files or >3000 lines → trigger the 6th pre-check, ask the user to narrow the scope
Test files should be included in the scope (to determine test coverage for the 2nd pre-check).

扫的时候看什么

What to Look for During Scanning

按方法库的四层分类当模板在代码里找：
  • L1 行为等价迁移信号：某个函数被很多处调用，但接口/实现要改 → 候选 Parallel Change；某整块老逻辑要被新实现替换 → 候选 Strangler Fig
  • L2 代码级重构信号：超长函数（> 50 行 / 圈复杂度 > 10）、重复的条件片段、神秘临时变量、多层嵌套 if-else
  • L3 结构拆分信号：组件 > 300 行、一个文件承担多件事、容器/展示混在一起、相同逻辑在多个组件里各写一份（前端）；Controller 里直接调 DB、Service 层缺失、Repository 被绕开（后端）
  • L4 性能信号：重复计算（可 memo）、N+1 查询、列表无虚拟化/分页、事件监听无清理、大对象深响应（Vue）
方法库完整清单在同目录 
reference/methods.md
，扫描时要全量加载作为匹配表。
Use the four-layer classification of the method library as a template to identify issues in code:
  • L1 Behavioral Equivalence Migration Signals: A function is called in many places but its interface/implementation needs to be modified → candidate for Parallel Change; a whole block of old logic needs to be replaced with a new implementation → candidate for Strangler Fig
  • L2 Code-Level Refactoring Signals: Overly long functions (>50 lines / cyclomatic complexity >10), repeated conditional fragments, mysterious temporary variables, deeply nested if-else
  • L3 Structure Splitting Signals: Components >300 lines, a file serving multiple purposes, mixing container/presentation logic, duplicate logic across multiple components (frontend); Controller directly calling DB, missing Service layer, Repository being bypassed (backend)
  • L4 Performance Signals: Repeated calculations (memoizable), N+1 queries, non-virtualized/unpaginated lists, unremoved event listeners, deep reactivity for large objects (Vue)
The complete method library list is in 
reference/methods.md
in the same directory; load it fully as a matching table during scanning.

产出格式

Output Format

{slug}-scan.md
由两部分组成：
  1. 顶部总览（一段）：扫描范围 / 发现条数 / 按分类分布 / 按风险分布 / 建议先做哪几条 / 建议慎做哪几条
  2. 清单条目（一条一块 markdown）：字段顺序和硬约束见同目录 
    reference/scan-checklist-format.md
扫完把 
{slug}-scan.md
整份交给用户看，用户勾选要做的条目（打 ✓）、标注疑问或否决（打 ✗ 并写理由），再进阶段 2。不要自己替用户勾选。
{slug}-scan.md
consists of two parts:
  1. Top Overview (one paragraph): Scan scope / number of findings / distribution by category / distribution by risk / recommendations for priority items / recommendations for high-risk items
  2. Checklist Items (one markdown block per item): Field order and hard constraints are in 
    reference/scan-checklist-format.md
    in the same directory
After scanning, send the entire 
{slug}-scan.md
to the user. The user selects items to perform (mark ✓), marks questions or rejections (mark ✗ and write reasons), then proceed to Phase 2. Do not select items on behalf of the user.

阶段 2：design（定执行方案）

Phase 2: design (Finalize Execution Plan)

输入

Input

  • 用户勾选过的 
    {slug}-scan.md
    （✓ 的条目是本次要做的，✗ 的归档留痕）
  • 方法库 
    reference/methods.md
    （每条勾选项必须映射到一个方法号 M-Ln-NN）
  • User-selected 
    {slug}-scan.md
    (✓ items are to be done this time, ✗ items are archived for traceability)
  • Method library 
    reference/methods.md
    (each selected item must map to a method ID M-Ln-NN)

做的事

Tasks

  1. 排顺序。勾选条目之间有依赖的排在前面（如 L1 的 Parallel Change 往往要先跑，L2 的提取跟在后面）。独立的按"低风险 + AI 可自证"优先，HUMAN 验证项排后面批量处理。
  2. 每条补执行细节：引用的方法号、具体步骤、前置条件、退出信号、验证责任方（AI / HUMAN）、回滚策略（出问题怎么还原）。
  3. 识别前置依赖：测试覆盖不够的条目要前置一个"补刻画测试"的步骤；改动涉及公开接口的要前置一个"搜调用方"的步骤。
  4. 整体 review：把 
    {slug}-refactor-design.md
    整稿交给用户，用户放行后改 frontmatter 的 
    status
    为 
    approved
    。
  5. 抽 checklist：从 design 抽出 
    {slug}-checklist.yaml
    ，steps 对应执行顺序，checks 对应每步的退出信号。
  1. Order Items. Items with dependencies are placed first (e.g., L1 Parallel Change often needs to be executed first, followed by L2 extraction). Independent items are prioritized by "low risk + AI-verifiable", and HUMAN verification items are grouped at the end.
  2. Add Execution Details for Each Item: Referenced method ID, specific steps, preconditions, exit signals, verification responsible party (AI / HUMAN), rollback strategy (how to restore if issues occur).
  3. Identify Pre-Dependencies: Items with insufficient test coverage need a pre-step of "supplement characterization tests"; items involving public interfaces need a pre-step of "search for callers".
  4. Overall Review: Send the full draft of 
    {slug}-refactor-design.md
    to the user. After user approval, change the 
    status
    in the frontmatter to 
    approved
    .
  5. Extract Checklist: Extract 
    {slug}-checklist.yaml
    from the design, with steps corresponding to execution order and checks corresponding to exit signals for each step.

design 文件结构

Design File Structure

markdown
---
doc_type: refactor-design
refactor: {YYYY-MM-DD}-{slug}
status: draft | approved
scope: {扫描范围一句话}
summary: {本次要做的几条是什么，一句话}
---
markdown
---
doc_type: refactor-design
refactor: {YYYY-MM-DD}-{slug}
status: draft | approved
scope: {one-sentence scan scope}
summary: {one-sentence description of items to be done}
---

{slug} refactor design

{slug} refactor design

1. 本次范围

1. Scope of This Refactor

  • 从 scan 里勾选了哪几条（编号列出）
  • 明确不做的条目（被 ✗ 的）和理由
  • 预估总工作量 / 总风险档位
  • Which items were selected from the scan (listed by number)
  • Items explicitly not to be done (marked ✗) and reasons
  • Estimated total workload / total risk level

2. 前置依赖

2. Pre-Dependencies

  • 测试覆盖补齐动作（如果需要）
  • 调用方搜索动作（如果需要）
  • 其他一次性准备
  • Test coverage supplement actions (if needed)
  • Caller search actions (if needed)
  • Other one-time preparations

3. 执行顺序

3. Execution Order

按步骤列，每步一块：
  • 步骤 N：{一句话动作}
  • 引用方法：M-Ln-NN {方法名}
  • 具体操作：{照方法库的步骤落到本项目的具体文件/函数}
  • 退出信号：{AI 跑什么测试 / HUMAN 看什么页面}
  • 验证责任：AI 自证 ｜ HUMAN
  • 回滚：{出问题怎么还原，通常是 git revert 某步}
List by step, one block per step:
  • Step N: {one-sentence action}
  • Referenced Method: M-Ln-NN {method name}
  • Specific Operations: {apply method library steps to specific files/functions in the project}
  • Exit Signal: {tests AI runs / pages HUMAN checks}
  • Verification Responsibility: AI Self-Verify ｜ HUMAN
  • Rollback: {how to restore if issues occur, usually git revert the step}

4. 风险与看点

4. Risks and Key Points

  • 高风险步骤汇总（本次 design 里风险=高的步骤单独拎出来提醒）
  • 容易出错的点（比如跨步骤的数据流变化）

---
  • Summary of high-risk steps (separately highlight steps with high risk in this design)
  • Error-prone points (e.g., cross-step data flow changes)

---

阶段 3：apply（逐条执行）

Phase 3: apply (Execute Item by Item)

推进规则

Promotion Rules

  1. 一步一做，不批量。严格按 checklist 顺序，当前步不完成不开下一步。
  2. 每步完成后走验证：
    • AI 自证项：跑指定测试 / 类型检查 / lint / grep 无残留旧引用。通过了记在 apply-notes 里，继续下一步。
    • HUMAN 验证项：停下来，汇报"第 N 步已完成，请在 {具体页面 / 操作步骤} 下目视确认，确认后我继续"。用户不明确说"继续"就不推进。
  3. 偏离当场记：执行中发现方案里没考虑到的情况（比如有个调用方在动态 import 里），停下来汇报，不要自己发挥。和用户对齐后把偏离点追加到 apply-notes，必要时回阶段 2 改 design。
  4. 行为等价自检：每步结束额外问自己——"这一步有没有可能改了外部可观察行为？"。有怀疑就退回当步，不带到下一步。
  1. One Step at a Time, No Batch Execution. Strictly follow the checklist order; do not start the next step until the current step is completed.
  2. Verification After Each Step:
    • AI Self-Verify Items: Run specified tests / type checks / lint / grep for no residual old references. Record results in apply-notes and proceed to the next step.
    • HUMAN Verification Items: Pause, report "Step N is completed. Please visually confirm at {specific page / operation steps}, and I will continue after confirmation". Do not proceed without explicit "continue" from the user.
  3. Record Deviations Immediately: If unconsidered situations are found during execution (e.g., a caller in dynamic import), pause and report, do not act on your own. Align with the user, add the deviation to apply-notes, and return to Phase 2 to revise the design if necessary.
  4. Behavioral Equivalence Self-Check: After each step, ask yourself — "Could this step change externally observable behavior?" If in doubt, roll back the step and do not proceed.

apply-notes 格式

apply-notes Format

markdown
---
doc_type: refactor-apply-notes
refactor: {YYYY-MM-DD}-{slug}
---
markdown
---
doc_type: refactor-apply-notes
refactor: {YYYY-MM-DD}-{slug}
---

{slug} apply notes

{slug} apply notes

步骤 1: {动作}

Step 1: {action}

  • 完成时间: {date}
  • 改动文件: {file list}
  • 验证结果: {测试输出 / HUMAN 确认语录}
  • 偏离: {无 / 具体描述}
  • Completion Time: {date}
  • Modified Files: {file list}
  • Verification Result: {test output / HUMAN confirmation quote}
  • Deviation: {None / specific description}

步骤 2: ...

Step 2: ...

undefined
undefined

全部完成后

After Completion

  • 跑一遍全量测试 + 类型检查 + lint
  • 最后一次请用户整体目视确认（尤其是前端：打开主要页面点一圈）
  • 确认通过后收尾 commit，commit message 引用 refactor 目录
  • Run full tests + type checks + lint
  • Ask the user to perform a final overall visual confirmation (especially for frontend: open key pages and test interactions)
  • After confirmation, finalize the commit, with the commit message referencing the refactor directory

退出条件

Exit Conditions

  • scan 前置检查跑过，命中的已按路由引导用户，没命中的才进 scan
  • {slug}-scan.md
    用户已勾选（✓/✗），未勾选的不进 design
  • {slug}-refactor-design.md
    每条勾选项都映射到了方法库的方法号
  • design 用户整体 review 通过，status=approved
  • {slug}-checklist.yaml
    已生成且通过 validate-yaml.py
  • apply 阶段每步都有验证记录（AI 自证的贴日志，HUMAN 的贴用户确认语录）
  • 全量测试 / 类型检查 / lint 通过
  • 用户最后一次目视确认通过
  • Pre-checks for scan have been run, and users have been routed appropriately if any checks were triggered; scan only proceeds if no checks were triggered
  • User has selected items in 
    {slug}-scan.md
    (✓/✗); unselected items do not proceed to design
  • Each selected item in 
    {slug}-refactor-design.md
    is mapped to a method ID in the method library
  • Design has passed overall user review, status=approved
  • {slug}-checklist.yaml
    has been generated and validated via validate-yaml.py
  • Each step in the apply phase has verification records (AI self-verify logs attached, HUMAN confirmation quotes attached)
  • Full tests / type checks / lint have passed
  • User has passed the final visual confirmation

容易踩的坑

Common Pitfalls

  • AI 硬凑清单：前置检查明显命中却找理由绕过，扫出一堆 "代码可以更优雅" 这种无量化问题的条目——应直接停下给路由建议
  • 夹带行为改动：在重构中间"顺便修了一个 bug / 优化了一下提示文案"——应停下，拆成独立 issue 或 feature
  • 跨步骤合并动作：为了快，一次提交里做了 2-3 步——失去"出问题单步回滚"的能力
  • 把口味项列进清单：命名偏好、引号、箭头函数 vs function——这些走 decisions，不走 refactor
  • 扫一个大模块就直接动手：> 15 文件 / > 3000 行的范围不拆就进 scan，产出一份没法决策的长清单
  • HUMAN 验证项自己跳过：前端效果 AI 看不到，不能用 "类型检查过了" 替代人工目视
  • 覆盖率不够硬上：没测试的模块直接改，所谓"行为等价"只是口头承诺
  • AI Forcing a Checklist: Bypassing obvious pre-check triggers with excuses, generating checklist items with unquantified issues like "code can be more elegant" — should pause immediately and provide routing suggestions
  • Including Behavior Changes: "Incidentally fixing a bug / optimizing prompt text" during refactoring — should pause and split into an independent issue or feature
  • Merging Cross-Step Actions: Performing 2-3 steps in one commit for speed — losing the ability to roll back individual steps if issues occur
  • Including Preference Items: Naming preferences, quotes, arrow functions vs function — these belong to decisions, not refactor
  • Starting Directly on a Large Module: Proceeding to scan without splitting a scope of >15 files / >3000 lines, resulting in an unmanageable long checklist
  • Skipping HUMAN Verification Items: AI cannot see frontend effects, so "passed type check" cannot replace manual visual inspection
  • Proceeding with Insufficient Coverage: Modifying modules without tests, relying only on verbal promises of "behavioral equivalence"

与相邻工作流的边界

Boundaries with Adjacent Workflows

  • feature：加新能力 / 改需求 → feature。refactor 里一旦冒出"顺便实现 X"，停下拆出去。
  • issue：修 bug / 行为错了 → issue。refactor 里发现的 bug 记成新 issue，不在本次 PR 里偷偷修。
  • decisions：全项目长期约束（"以后都用 composable"、"禁用 mixin"）→ decisions。refactor 执行时可以引用已有 decision 作为依据，但不产出 decision。
  • architecture：跨模块边界重划 / 分层调整 → architecture + decisions。单次 refactor 不跨模块，跨模块的工作要拆成"更新架构文档 + 记决策 + N 个模块级 refactor"。
  • tricks / learning：refactor 中发现可复用的手法 → tricks；踩的坑 → learning。
  • feature: Adding new capabilities / modifying requirements → feature. If "incidentally implement X" emerges during refactoring, pause and split it out.
  • issue: Fixing bugs / correcting incorrect behavior → issue. Bugs found during refactoring should be recorded as new issues, not fixed secretly in the current PR.
  • decisions: Project-wide long-term constraints ("use composable from now on", "disable mixin") → decisions. Refactor execution can reference existing decisions as basis, but does not produce decisions.
  • architecture: Cross-module boundary restructuring / layer adjustment → architecture + decisions. A single refactor does not cross modules; cross-module work should be split into "update architecture documentation + record decisions + N module-level refactors".
  • tricks / learning: Reusable techniques discovered during refactoring → tricks; pitfalls encountered → learning.

相关文档

Related Documents

  • easysdd-core/SKILL.md
    — easysdd 家族根技能
  • easysdd-refactor-fastforward/SKILL.md
    — 小重构超轻量通道
  • reference/scan-checklist-format.md
    — scan 清单条目的字段、顺序、硬约束、反模式样本
  • reference/refusal-routing.md
    — scan 前置检查 7 条 + 路由表 + 拒绝输出格式
  • reference/methods.md
    — 重构方法库（L1-L4 四层分类，统一字段）
  • easysdd/reference/shared-conventions.md
    — 跨工作流共享口径
  • 项目架构总入口 — scan 前需要翻一下，确认模块边界在哪
  • easysdd-core/SKILL.md
    — Root skill of the easysdd family
  • easysdd-refactor-fastforward/SKILL.md
    — Ultra-lightweight channel for small refactors
  • reference/scan-checklist-format.md
    — Fields, order, hard constraints, and anti-pattern samples for scan checklist items
  • reference/refusal-routing.md
    — 7 scan pre-checks + routing table + rejection output format
  • reference/methods.md
    — Refactor method library (four-layer classification L1-L4, unified fields)
  • easysdd/reference/shared-conventions.md
    — Shared terminology across workflows
  • Project architecture entry — Review before scanning to confirm module boundaries