Code Slimming

Subtraction only. Delete and inline in place. Never move code between files, never rename files, never restructure directories — a reviewable diff matters more than a tidy one.

Out of scope, by design: bug hunting, design-pattern advice, adding comments, cross-file refactoring, architectural "you should have fixed the root cause" judgments.

References (this skill's directory, read on demand):

• references/redlines.md — what must never be deleted. Read before proposing any deletion.
• references/checks-mechanical.md — M-tier checks (决定可删无需理解业务).
• references/checks-judgment.md — J-tier checks, grouped by pass, with batching order.
• references/scope-creep.md — signals S1–S12 for "AI added what nobody asked for".
• references/audit-checks.md — the --audit checklist. Only for --audit; never mix with the M/J checks.
• references/lang-java.md · lang-go.md · lang-typescript.md · lang-python.md — per-language checks and framework-boilerplate redlines. Read only the ones the diff touches.

All user-facing questions and reports are written in Chinese. Rules stay English.

Arguments

Arg	Effect
(none)	List findings → user confirms → edit → verify → report
--review	List findings only. Touch nothing.
--auto	Apply M-tier without asking; still confirm J-tier and scope creep.
--audit [<rev>]	Review a cleanup that already happened. Skips Steps 2–7 entirely; runs references/audit-checks.md instead.
path(s)	Use these files instead of the git diff.

--audit is not --review. --review dry-runs the M/J checks against code; --audit checks a finished cleanup — what it deleted that it should not have, what it missed, whether it stayed inside the subtraction-only contract. Re-running the M/J checks in audit mode produces the same verdicts as the original pass and adds nothing; use the audit checklist and only that.

Step 1 — Gates

Run in order; stop at the first hard block.

1. Rebase/merge in progress (git status shows conflicts) → refuse, say why.
2. Generated files — path or header matches DO NOT EDIT, @generated, Code generated by, .pb.go, *_pb2.py, openapi/, **/generated/** → drop them from scope, name them in the report.
3. Unrelated uncommitted work — the diff spans files with no plausible relation to one task (different modules, mixed intents). Report the split and ask 继续 / 只处理指定文件 / 先去 commit.
4. Project CLAUDE.md — read it. Any check that contradicts a documented project rule is disabled, not overridden. State how many were disabled and which, at the top of the report.

Step 2 — Scope

Default: git status --porcelain + git diff HEAD → modified + untracked source files. Not in a git repo → ask the user for paths. Explicit paths given → use those, skip the diff.

>20 files → stop, report the count, ask 继续 / 分批（10 个一批）/ 缩小范围.

Step 3 — Scan

Six passes, strictly in order. The order is causal: deleting dead code first makes its comments, abstractions, and tests vanish with it, so later passes shrink dramatically.

1. Dead code → 2. Duplicate implementation (reuse) → 3. Redundant abstraction →
2. Defensive paranoia → 5. Comments → 6. Naming / filler

Never mix categories inside one pass. Every candidate passes the redline filter before it reaches the list.

Reuse search radius (pass 2): the diff's own directories plus any utils / util / common / shared / helper / lib module in the repo. Not the whole tree. Reference counting (dead-code checks) is the exception: take the symbols from the diff and grep the whole repo for them. Approximate counting here means deleting live code.

Step 4 — The list (always before editing)

Present four sections in Chinese. Under --auto, apply 将删 immediately and still present the rest.

• [将删] M-tier. Each: file:line · what · one-line reason.
• [待定] J-tier. Each: file:line · what · the judgment call the user has to make. Collect answers with AskUserQuestion, recommended option first, in small batches.
• [越界] Scope creep. List only. Never delete, not even with --auto. Each entry cites its signal numbers (e.g. S1 + S4 + S7) so the user can re-check the reasoning.
• [跳过] Everything skipped, fully expanded, one line each: file:line · check · why (redline name, or "误报: …"). This is the user's only feedback channel for calibrating the redlines — never collapse or summarize it.

Then a final section [附注] 非冗余问题（超出本 skill 职责）: swallowed exceptions, catch that only logs and returns success, as any / force-unwrap, rules asserted by a comment the code doesn't implement. Type, location, count, and one line: 建议跑 code-review-deep. Those code blocks are excluded from simplification — editing them would mask the problem.

Step 5 — Edit

Delete and inline in place. One pass at a time; verify between passes (Step 6). If a pass breaks verification, revert that whole pass and move on — do not bisect inside it.

Never: move code across files, rename files, restructure directories, rename symbols with callers outside the current file, rewrite for style, or "fix the behavior while you're in there". A change that alters observable behavior gets reverted even if the new behavior looks better.

Step 6 — Verify

Prefer the project's own commands (package.json scripts, Makefile, pyproject.toml, pom.xml). Fall back to what is globally available here:

	Static first	Then tests
Python	ruff check, mypy	pytest (or uv run pytest)
Go	go vet ./..., go build ./...	go test ./...
TS	npx tsc --noEmit	npm test
Java	mvn -q compile	mvn test

If nothing runnable exists, say 未验证 and name what was missing. Never imply verification that did not happen.

Step 7 — Report

Terminal only — no report files. Give the diff summary (files touched, lines removed), the skipped list, the 附注 section, and any pass that was reverted with the reason.