Code Pretty (Pre-Review Code Block Formatting)

This skill prettifies the "수정 후" code blocks inside a freshly written or rewritten implementation plan, just before generating-html + user review. It is the code-only sibling of generating-html.

Announce at start: "I'm using the code-pretty skill to format 수정 후 code blocks in <file> before generating-html + user review."

It STOPS firing the moment the first change-history entry has been logged. That boundary marks the doc as "live" — from then on, no code-pretty.

Specifically, code-pretty MUST NOT run on:

• <slug>-requirements.md or <slug>-tech-design.md (only implementation-plan)
• "원본" code blocks (only "수정 후" blocks)
• Prose, tables, headings, list items
• Any plan AFTER its first change-history entry exists

If you are unsure whether this is still in the "initial creation phase" — STOP. Look for an existing ## 변경이력 footer with one or more entries. If ANY entry exists, this is NOT initial creation. Skip this skill. </HARD-GATE>

When to Use

Why a Subagent (and which model)

Same reasoning as generating-html: pure transformation, no domain reasoning, negative-constraint heavy.

Always dispatch a subagent with model: "sonnet". Sonnet's instruction-following is required for honoring "leave already-clean blocks byte-identical" and "1% 의심이라도 들면 SKIP" constraints.

Do NOT use Opus (overkill) or Haiku (rephrasing risk).

Process

Step 1 — Pre-flight check (v1.1.15+ user-gate)

Before dispatching, run the deterministic helper:

source .venv/bin/activate && python -c "
import sys
from pathlib import Path
from scripts.preflight import code_pretty_check
result = code_pretty_check(Path('<TARGET>'))
print(f'ok={result.ok} reason={result.reason} | {result.human_reason}')
sys.exit(0 if result.ok else 1)
" 2>&1

exit code 분기 (v1.1.15 user-gate):

• exit 0 → 검증 통과, Step 2 dispatch 진행.
• exit 1 (helper semantic fail) → human_reason 노출 후 AskUserQuestion 게이트:
  • "수정 후 재시도" / "강제 진행 (위험)" (메인이 ⚠️ preflight 우회. <reason> 무시. 한 줄 안내) / "스킵 (이번만)" (caller 에게 abnormal return).
• exit ≠ 0,1 (invocation 실패) → stderr 전문 + AskUserQuestion 게이트:
  • "직접 디버깅" / "skill 단계 스킵".

Caller 책임 (helper 가 검증 X): verifying-spec 가 직전에 통과했는지 — 이건 writing-plans 흐름의 책임이고 helper 가 검사할 수 없음. 호출자가 보장.

helper 의 검사: file 존재 / 변경이력 footer 비어있음 / filename *-implementation-plan.md / 최소 1개 **수정 후** 블록 존재. 자세히는 scripts/preflight.py:code_pretty_check.

Step 2 — Dispatch the Sonnet subagent

Use the Agent tool with these exact parameters:

• subagent_type: general-purpose
• model: sonnet
• description: Code-block prettify on <filename>
• prompt: see template below

Step 3 — Verify, surface diff, return to caller

After the subagent returns:

1. Read the file back (1 Read)
2. Sanity-check: every **원본**-labeled code block is byte-equal to the pre-dispatch version
3. Surface the diff summary text returned by the subagent to the main agent's chat output (caller will combine with generating-html output for the user review gate)
4. Return control to caller (writing-plans). Do NOT invoke generating-html or change-history yourself.

If sanity-check fails (any "원본" block was modified) → emit a warning to the caller; caller decides whether to abort or rerun.

Subagent Prompt Template

The dispatched subagent receives this exact prompt (filled in with target path):

You are performing a STRICT code-block prettify on a Korean implementation-plan document.

Target file: <ABSOLUTE_PATH>

Your job: improve READABILITY of "수정 후" code blocks ONLY. Other content is byte-identical.

# Identification — what counts as a target block

A target block is a fenced code block whose **immediately preceding non-blank line** starts with `**수정 후**`.

Examples of target labels (any of these counts):
- `**수정 후**:`
- `**수정 후** (`new file`):`
- `**수정 후**` (label only)

Examples of NON-target labels (any of these means the next code block is a "원본" — NEVER touch):
- `**원본**`
- `**원본** (...)`
- Anything not starting with `**수정 후**`

If a code block has no label at all (no preceding bold text), DO NOT touch it. Default to byte-equal preservation when uncertain.

# Allowed changes (in target blocks only)

Three categories. Apply only when there is a CONCRETE, ARTICULABLE readability improvement.

## Category A — 포맷
- Whitespace / indentation normalization
- Long-line wrapping at sensible breakpoints
- Trailing whitespace removal
- Aligned `//` comments

## Category B — 자명한 정리
- Remove dead comments explicitly marked (e.g., `// TODO: 삭제예정`)
- Standardize quote/semicolon style within the block
- Add labeling comments next to magic numbers (do NOT extract to const)

## Category C — 중복 통합 → **폐지**

**이 범주는 폐지됐습니다.** import·const·helper 병합은 **블록 밖의 참조와 스코프를 바꾸는데**, 이 스킬은 코드 블록만 보고 파일 전체를 보지 않습니다. "call-site contexts are byte-identical" 은 블록만 보고는 판정할 수 없는 조건이었습니다.

# FORBIDDEN — never do any of these

> **왜 의미 변경이 전부 금지인가:** 이 스킬은 `verifying-spec` **통과 후**에 돕니다. 그리고 `수정 후` 블록은 implementer 가 **STRICT BYTE-COPY** 로 코드베이스에 그대로 넣습니다. 즉 여기서 의미를 바꾸면 **검증된 코드와 실행되는 코드가 달라집니다.** `plan_byte_check.py` 는 `원본` 블록만 검사하므로 `수정 후` 의 변조를 잡아주지 못합니다. **서식만 만지세요.**

- **Do NOT rename variables.** 블록 밖 참조가 깨지고, 그 사실을 이 스킬은 볼 수 없다
- **Do NOT merge imports / const / inline helpers.** 스코프와 구조가 바뀐다
- Do NOT touch "원본" blocks
- Do NOT touch prose, headings, lists, tables, frontmatter, or `## 변경이력`
- Do NOT extract magic numbers to named constants
- Do NOT flatten nested if statements
- Do NOT split or merge functions
- Do NOT change behavior, side-effects, exception flow, or output
- Do NOT modify a target block if it is already well-formatted (consistent whitespace, no obvious readability issues, no Category B / C candidates) — leave byte-identical
- If you have ≥1% suspicion that a Category A or B change might alter behavior, SKIP that change

# How to apply

1. Read the file in full
2. Locate every `**수정 후**`-labeled code block
3. For each target block:
   a. Articulate (mentally) the concrete readability improvement
   b. If you cannot articulate one — leave byte-identical, mark as "no-op"
   c. Otherwise apply A/B transformations within the constraints above (C 는 폐지됨)
4. Write the result back to the SAME file path using the Write tool (overwrite)
5. Report a diff summary in this format:

code-pretty done on <path>.

Target blocks: <total>

• Modified: <N> (<categories: A/B breakdown>)
• No-op (already clean): <N>
• Skipped (1% suspicion): <N> (with reasons)

Modified line summary:

• file:line — <one-line reason>
• ...

"원본" blocks preserved byte-identical: yes / no

# Verification before writing

Before you call Write:
- Compare every `**원본**`-labeled block in your output to the input — they MUST be byte-identical
- Confirm the YAML frontmatter (if present) is byte-identical
- Confirm the `## 변경이력` heading and everything beneath it is byte-identical
- Confirm prose / tables / headings outside of "수정 후" blocks are byte-identical

If ANY of these fail, do NOT write. Report the failure and stop.

You have one job: make "수정 후" code blocks cleaner WITHOUT changing meaning. Nothing else.

Process Flow

digraph code_pretty {
    "Caller (writing-plans) ready" [shape=box];
    "Pre-flight check" [shape=diamond];
    "STOP — log + return" [shape=box];
    "Dispatch Sonnet subagent" [shape=box];
    "Subagent: Read → identify → prettify → Write" [shape=box];
    "Main: Read back, sanity-check 원본 blocks" [shape=box];
    "Sanity OK?" [shape=diamond];
    "Surface diff summary, return to caller" [shape=doublecircle];
    "Warn caller, return abnormally" [shape=box];

    "Caller (writing-plans) ready" -> "Pre-flight check";
    "Pre-flight check" -> "STOP — log + return" [label="any check fails"];
    "Pre-flight check" -> "Dispatch Sonnet subagent" [label="all pass"];
    "Dispatch Sonnet subagent" -> "Subagent: Read → identify → prettify → Write";
    "Subagent: Read → identify → prettify → Write" -> "Main: Read back, sanity-check 원본 blocks";
    "Main: Read back, sanity-check 원본 blocks" -> "Sanity OK?";
    "Sanity OK?" -> "Surface diff summary, return to caller" [label="yes"];
    "Sanity OK?" -> "Warn caller, return abnormally" [label="no"];
}

Anti-Patterns

Red Flags (STOP if you think these)

Acceptance

A code-pretty run is correct when ALL hold:

1. Pre-flight checks all passed (file exists, target = implementation-plan.md, 변경이력 empty, ≥1 "수정 후" block)
2. Subagent was dispatched with model: sonnet and the strict prompt above
3. Post-dispatch sanity check: every "원본" block byte-identical to pre-dispatch
4. Diff summary surfaced to main agent chat (caller forwards to user review gate)
5. No ## 변경이력 entry was added by code-pretty itself

Related Skills

• writing-plans — invokes this between verifying-spec and generating-html
• generating-html — sibling skill, runs immediately after code-pretty on the same draft
• verifying-spec — must pass before code-pretty can run
• change-history — invoked by caller AFTER code-pretty + generating-html + user review approval