Agent Skills › gnomeria/usbtree › refactor

refactor

GitHub

指导行为不变的重构流程，强调安全网、小步验证及无功能变更。适用于清理代码、提取函数、重命名或拆分模块等场景。

.agents/skills/refactor/SKILL.md gnomeria/usbtree

触发场景

重构代码结构清理重复代码提取或重命名函数/模块使代码更易于测试

安装

npx skills add gnomeria/usbtree --skill refactor -g -y

更多选项

非标准路径

npx skills add https://github.com/gnomeria/usbtree/tree/main/.agents/skills/refactor -g -y

不安装直接使用

npx skills use gnomeria/usbtree@refactor

指定 Agent (Claude Code)

npx skills add gnomeria/usbtree --skill refactor -a claude-code -g -y

安装 repo 全部 skill

npx skills add gnomeria/usbtree --all -g -y

预览 repo 内 skill

npx skills add gnomeria/usbtree --list

SKILL.md

Frontmatter

{
    "name": "refactor",
    "description": "Behavior-preserving restructuring done safely — test net first, small verified steps, no mixed-in feature changes. Use when the user says \"refactor\", \"clean up\", \"restructure\", \"extract\", \"rename\", \"split this file\/function\", \"reduce duplication\", or \"make this testable\"."
}

Refactor

Structure changes, behavior never. Behavior changes = feature or bug fix — separate change.

Step 0 — Worth it?

Purpose required: enables concrete upcoming change, kills duplication that bit someone, or makes untestable testable. "Could be cleaner" alone = say so, stop.
Name target shape in one or two sentences before touching code. No target = wandering diff.

Step 1 — Safety net

Baseline: cargo test green, cargo clippy -- -D warnings clean. Note it.
Uncovered code: characterization tests first — pin what it does now (odd behavior included). Extra check: --demo --dump output before vs after must be byte-identical.
Working tree clean before starting; git checkout . is the escape hatch.

Step 2 — Small steps, verify each

One named refactoring at a time: extract fn, inline, rename, move, split module. Finish, verify, next.
Verify each step: cargo test + clippy. Red → undo last step only. Never push forward on red planning to "fix at the end".
Rename via rust-analyzer / cargo tooling where available — tool can't miss call site, regex can.
Commit-sized steps. Long refactor → stop at green checkpoints. 40-file big-bang diff "should work" = failure.

Step 3 — Hard rules

No behavior changes. Bug found mid-refactor → note it, checkpoint, fix separately (see debug skill).
No fmt churn. Repo bans blind cargo fmt — hand-aligned blocks (theme, match tables). Diff only what the refactor touches.
Respect // ponytail: comments — known deliberate tradeoffs. Don't "clean up" without reason.
No test rewrites to fit refactor. Tests fail after "behavior-preserving" change → change wasn't behavior-preserving. Exception: tests coupled to now-gone internals; update minimally, say why.
No scope creep. Step 0 shape is boundary. New ideas → list at end, don't do.

Step 4 — Finish

Full cargo test + clippy green. --demo --dump unchanged. Dead code deleted, not commented out.
Report: shape achieved, steps taken, things found-not-done.

Direction

Extract functions over introducing traits/structs; trait only at genuine seam (multiple impls exist or test needs one).
Rule of three: two similar blocks may stand; third occurrence earns abstraction — right shape now visible.
Don't abstract for testability when passing plain value/fn does it.