typescript-refactoring（TypeScript refactoring 搜索）

TypeScript Refactoring

Refactoring is changing structure without changing behavior. Assess before changing. Small steps, verified continuously. The typescript skill defines what good TypeScript looks like — this skill defines how to get there from messy code.

1. The Iron Rules

These are non-negotiable. Violating them turns refactoring into "changing stuff and hoping."

1. Never change behavior and structure in the same commit. Refactoring commits change structure only. Feature commits change behavior only. Mixing them makes rollback impossible and code review meaningless.
2. Never refactor without tests. If tests don't exist, write characterization tests first — tests that capture current behavior, even if that behavior is wrong. Fix the behavior later, separately.
3. One smell at a time, codebase-wide. Fix all any types, then all long functions, then all enums. Don't fix everything in one file — fix one thing across all files. This produces consistent, reviewable diffs.
4. Every step must be independently verifiable. Tests pass after each commit. If a refactoring breaks tests, it's too big — break it into smaller steps.
5. Prefer boring, obvious transformations. Rename a variable. Extract a function. Inline a pointless wrapper. These are safe, reviewable, and reversible. Clever restructuring is risky.

2. Assessment: Read Before You Cut

Before changing anything, understand what you have. Present findings to the user before starting work.

Run the basics:

# Does it build? Does it pass?
npm run build
npm test

# What's the shape of the codebase?
find src -name '*.ts' -o -name '*.tsx' | wc -l          # file count
find src -name '*.ts' -o -name '*.tsx' -exec wc -l {} + # line counts

Identify the hot paths — code that changes most often has the highest refactoring ROI:

# Most frequently changed files in the last 6 months
git log --since="6 months ago" --pretty=format: --name-only -- '*.ts' '*.tsx' \
  | sort | uniq -c | sort -rn | head -20

Categorize what you find:

• Read entry points and trace the dependency graph
• Note the biggest files, deepest nesting, widest functions
• Scan for any, enum, namespace, barrel files, dead code
• Check tsconfig strictness settings — are strict, noUncheckedIndexedAccess enabled?
• Look at test coverage — which areas are safe to touch?

Present your assessment:

## Assessment Summary
- X files, Y total lines
- Build: passing/failing
- Tests: X passing, Y failing, Z% coverage
- Strictness: strict=true/false, noUncheckedIndexedAccess=true/false
- Top smells: [list with counts]
- Hot paths: [most-changed files]
- Recommended priority: [what to fix first and why]
- Estimated scope: [1-hour cleanup / multi-day migration / needs discussion]

3. Priority Order

Fix in this order. Each level makes the next level easier and safer.

Level 1: Type Safety

Eliminate any, add missing return types, enable strict flags. This is the highest-leverage change — the compiler catches bugs for you after this.

Level 2: Dead Code

Remove unused imports, exports, unreachable branches, commented-out code. Less code = less to read, less to maintain, fewer false signals for AI agents.

Level 3: Naming

Rename unclear variables, functions, and files to reveal intent. This is the single biggest readability win. data → userProfile, res → apiResponse, handle → handlePaymentWebhook.

Level 4: Structure

Extract functions from long bodies. Split God files. Flatten deep nesting with early returns and guard clauses. Target: functions under 20 lines, files under 300 lines, nesting under 3 levels.

Level 5: Patterns

Migrate anti-patterns to modern TypeScript: enums → union types, classes → factory functions (where appropriate), boolean soup → discriminated unions, barrel file chains → direct imports.

Level 6: API Boundaries

Add branded types, parse at boundaries, tighten function signatures. This is the final polish — making the type system enforce domain rules.

For specific smells and how to detect them, see references/smell-catalog.md. For step-by-step transformation recipes, see references/transformation-playbook.md.

4. Safe Transformation Workflow

For each refactoring, follow this loop:

1. Verify tests pass                    (baseline)
2. Make ONE structural change           (extract, rename, move, inline)
3. Verify tests still pass              (behavior preserved)
4. Commit with descriptive message      (checkpoint)
5. Repeat

For risky changes, use the strangler fig pattern:

1. Create new implementation alongside old
2. Route one caller to the new implementation
3. Verify behavior matches
4. Migrate remaining callers incrementally
5. Delete old implementation when all callers migrated

For large migrations (e.g., enabling strict: true on a non-strict codebase):

1. Enable the flag
2. Fix errors file-by-file, starting with leaf modules (no dependents)
3. Work inward toward the core
4. Commit after each file or small group of files

Commit messages during refactoring:

refactor: rename UserData → UserProfile across auth module
refactor: extract calculateDiscount from processOrder
refactor: eliminate any in api/client.ts
refactor: migrate Status enum to union type
chore: remove 12 unused imports in src/utils/

5. What Makes Code AI-Friendly

AI agents read code to understand it, then generate new code that fits. These patterns directly improve both capabilities:

Explicit types reduce guessing. When a function has ): Promise<User | null>, the AI knows exactly what it returns. Without it, the AI must infer — and may infer wrong.

Descriptive names communicate intent. AI agents infer purpose from names more than humans realize. getUserById tells the AI everything. get tells it nothing. Rename aggressively.

Small functions fit in context windows. AI agents have limited context. A 200-line function forces the AI to hold all of it in memory. Five 20-line functions let the AI focus on the relevant one.

Flat control flow is easier to trace. Deeply nested if/else/try/catch blocks confuse AI agents the same way they confuse humans. Early returns, guard clauses, and extracted conditions keep the happy path obvious.

Consistent patterns teach by example. If every API handler follows the same shape — validate input, call service, return response — the AI learns the pattern and replicates it correctly. Inconsistency forces the AI to guess which pattern to follow.

Types replace comments. A type like EmailAddress (branded type) or AsyncState<T> (discriminated union) communicates more than any comment. Types are verified by the compiler; comments rot.

Strict tsconfig catches AI mistakes. Enable strict: true and noUncheckedIndexedAccess: true. These flags catch the most common AI-generated bugs: unsafe index access, missing null checks, implicit any.

Colocated code reduces jumping. When types, utilities, and components live in the same directory, AI agents find them faster. Scattered code across distant directories increases the chance the AI misses context.

6. What NOT to Do While Refactoring

• Don't "big bang" rewrite. It always takes 3x longer, loses domain knowledge, and can't be deployed incrementally. Use strangler fig instead.
• Don't abstract prematurely. Wait for the 3rd use case before extracting a shared abstraction. Two similar functions are fine — a wrong abstraction is worse than duplication (AHA principle).
• Don't refactor and add features simultaneously. Separate commits, separate PRs if possible. Mixed diffs are unreviewable.
• Don't "improve" code you're not asked to touch. Scope creep during refactoring creates unexpected breakage. Fix what's in scope, note what's out of scope.
• Don't optimize performance without measurements. Refactoring for readability is always justified. Refactoring for performance requires profiling evidence.
• Don't fight the existing architecture. If the codebase uses classes everywhere, don't convert everything to factories in one pass. Propose a migration plan and get buy-in first.

Companion Skills

• typescript — the target state. What good TypeScript looks like. Reference this for specific patterns (discriminated unions, branded types, factory functions, etc.)
• karpathy-guidelines — behavioral constraints. Think before coding, simplicity first, surgical changes, goal-driven execution.
• agent-first-repo — repo-level improvements. AGENTS.md, ARCHITECTURE.md, documentation structure, mechanical enforcement of architecture.
• systematic-debugging — when refactoring reveals bugs. Follow root-cause investigation, don't patch symptoms.

For the full smell catalog, see references/smell-catalog.md. For step-by-step transformation recipes, see references/transformation-playbook.md.