Source: https://github.com/steipete/agent-scripts/tree/main/skills/github-deep-review (adopted 2026-05-16) 적용 대상: soma / p9 PR 리뷰, issue triage, 버그 root cause 분석. Peter의 maintainer/contributor 노트 컨텍스트는 제거. 일반 root-cause + fix-quality 방법론만 유지.

GitHub Deep Review

Review with high confidence, evidence-first, code-aware, and willing to say "not proven" when the trail is weak. The goal is not a generic summary. The goal is to understand the bug class, find the real cause if possible, decide the best fix after reading enough code, and call out whether a larger refactor would improve the design.

Start

Use gh, not web browsing, for GitHub refs:

gh issue view <n> --json number,title,state,author,body,comments,labels,updatedAt,url
gh pr view <n> --json number,title,state,author,body,comments,reviews,files,commits,statusCheckRollup,mergeStateStatus,headRefName,headRepositoryOwner,url
gh pr diff <n> --patch

For repo-local review, also inspect:

git status --short --branch
git fetch origin
git log --oneline --decorate -20

Use Grep tool (ripgrep) to search:

<key symbol/error/config/endpoint>

If the repo has local instructions (CLAUDE.md, AGENTS.md, etc.), issue/PR skills, docs lists, test guidance, or maintainer runbooks, read those before deciding.

Review Contract

Always answer these, explicitly:

URL/ref: issue or PR number and affected surface.
What is the bug or behavior being fixed?
Can we identify the root cause? If yes, where in code and why. If no, what evidence is missing.
For regressions, who/what introduced it and when? Include commit/PR provenance when traceable by bounded history; say unknown instead of guessing.
Is the current/proposed fix the best possible fix after reading adjacent code?
Would a bigger refactor improve correctness, clarity, or future maintainability?
What proof exists: tests, live repro, CI checks, docs, dependency docs/source, shipped/current behavior.
What remains risky or unverified.

Code Reading Depth

Read past the first touched file. Follow the real call path:

entrypoint -> validation/parsing -> routing/dispatch -> owner module -> shared helper -> persistence/network/runtime boundary
config/schema/docs -> runtime usage -> doctor/migration/fix path
provider/channel/plugin owner code -> generic core seam, only if multiple owners need it
tests around the touched surface plus adjacent regression tests

When behavior depends on a dependency, read the upstream docs/source/types or current package contract before assuming.

Prefer current source and executable proof over issue comments. Treat stale comments, old CI, and old release behavior as hints until rechecked.

Provenance

For bug/regression reviews, include a compact Provenance: answer when feasible:

Use git log -S/-G, git blame, linked PRs/issues, and tests.
Separate author, committer/merger, and current PR author when they differ.
Phrase as introduced by, made visible by, or carried forward by.
Include confidence: clear, likely, or unknown.
For features, docs, refactors, or untraceable bugs, write N/A or say what evidence is missing.

Fix Quality Bar

Good fixes usually:

live at the ownership boundary where the bug belongs
preserve public/backward-compatible behavior unless the issue is about retiring it
add a regression test at the smallest meaningful seam
avoid broad special cases, hidden migrations, semantic sentinels, and provider/channel IDs in generic core
update docs/changelog when user-visible behavior changes
fail clearly in runtime paths and repair through doctor/migration paths when that is the established contract

Call out when a fix is only symptom-level. If a slightly larger refactor makes the invariant obvious and reduces future bugs, recommend it. If the refactor widens risk without improving the bug class, say so.

PR Review Shape

Lead with findings when reviewing a PR. Findings need file/line/symbol references and a concrete failure mode. Avoid vague "consider" comments.

If no blocking issues:

say no blocking correctness issues found
list the strongest proof checked
name residual risk/test gaps
answer whether the design is the best available shape

Do not approve, comment, close, merge, push, or land unless the user asked for that action.

Issue Review Shape

For bugs/issues:

Reconstruct the reporter's scenario and affected version/surface.
Check whether current main already fixes it.
Reproduce or create a minimal local/live proof when feasible.
If clear, identify root cause and proposed fix.
If solved on main, only comment/close when the user asks; include proof and the canonical commit/PR if known.

If reproduction is not feasible, say exactly what blocks it and what evidence would make the decision reliable.

Output Template

Use this shape when 지혁 asks "what is this about", "is this the best fix", or "what did we fix":

Ref: #123 / PR #456
Surface: <runtime/CLI/provider/channel/docs>

Bug: <one or two sentences>
Cause: <code path + confidence>
Provenance: <introduced/made visible/carried forward by commit/PR/date, or N/A/unknown>
Best fix: <what should change and why>
Refactor: <yes/no, specific shape>
Proof: <tests/live/CI/source/dependency docs>
Risk: <remaining uncertainty>

Keep it concise, but do not skip the cause/fix/refactor/proof decision.

Tool / Project Adoption Review (쓸만한가?)

PR/버그가 아니라 "이 레포/도구를 쓸지" 판단할 때. 위 root-cause 방법론과 별개 모드.

세 질문에 답한다 (순서대로):

메커니즘이 실제로 뭘 하나? README 주장을 코드로 환원한 동작. (안티패턴 2 참조)
이게 우리 문제를 푸나? 일반 좋음이 아니라 지혁/soma/p9의 구체적 미해결 문제와의 적합성. 적합 없으면 "흥미롭지만 우리엔 무관"이 정답.
리스크와 저위험 검증 경로. 채택 전 우리 워크로드로 재볼 수 있는 최소 실험(holdout/dry-run/벤치 재현).

Anti-patterns (검증된 실패 — 2026-06-20 지혁 교정)

AP1. 대리지표(vanity metric)를 품질로 착각 금지.

❌ "테스트 619개 → 진지한 프로젝트", "10K stars → 좋음", "커밋 많음/CI 초록/codecov 배지 → 신뢰".
이유: 테스트 수·LOC·⭐·커밋·트렌딩·배지는 AI가 다 양산하는 시대에 품질과 무상관. 테스트 30만 개여도 검증력 0일 수 있다. 양은 신호가 아니다.
✅ 대신: 주장의 핵심 모듈 1~3개를 직접 읽고 구현 깊이·정직성·엣지케이스 처리를 판단. 평가는 읽은 코드에 근거. 숫자 세지 말고 코드를 봐라.

AP2. 마케팅 용어를 메커니즘으로 환원하라 (그대로 옮기지 마라).

❌ README의 "X compression", "Nx faster", "output token reduction", "AI-powered", "60–95% fewer" 를 그 표현 그대로 리뷰에 전달.
✅ 절차: (a) 물리적으로 말이 되나 먼저 의심 — 예: "API 출력을 사후 압축"은 불가능(응답 받으면 이미 토큰 청구됨). 말 안 되면 README가 뭔가 다른 걸 그 이름으로 부르는 것. (b) 코드를 까서 실제 동작으로 환원 — 예: headroom "output reduction"의 실체 = output_shaper.py가 프롬프트 끝에 terse 문구 추가 + reasoning effort 사전 하향 = "사후 압축"이 아니라 "사전 프롬프트 조작으로 출력 길이 억제". (c) 환원 못 하면 "검증 안 됨 — 주장일 뿐" 명시.

AP3. 벤치마크는 자체측정·cherry-pick으로 가정.

숫자가 자체 벤치(N 작음)이거나 유리한 workload만 고른 건 아닌지. holdout/독립재현/counterfactual 여부 확인. README가 "estimate"라 자인하면 그대로 estimate로 전달.

Adoption 출력 템플릿

도구: <name> — <한 줄 실체(메커니즘 환원형, 마케팅 용어 X)>
실체 메커니즘: <코드로 까본 실제 동작>
우리 적합성: <지혁/soma/p9 구체 문제와 매칭, 또는 "무관">
리스크: <정보손실/복잡도/의존성/레이턴시/pre-1.0 등>
검증 경로: <채택 전 우리 워크로드로 잴 최소 실험>
판정: 쓸만/조건부/패스 + 근거(읽은 코드, 대리지표 아님)

p9/soma 적용 노트

p9 issue: bd list → bd show <id> → 위 템플릿으로 답.
soma 코드 변경 검토: cd ~/2lab.ai/soma && git diff + 위 워크플로우.
"고친다"고 말하기 전에 ./scripts/bd-verify로 증명 후 bd-close-verified.

Más de este repositorio

mismo repositorio

beautiful-pdf

2lab-ai/skills

Convert markdown to a beautifully designed editorial-style PDF report. Uses Pretendard variable font, Swiss-Coffee + Cool-Blue palette, minimalist-maximalism layout, cover page, table of contents, section dividers, page numbers. Optimized for Korean+English mixed reports. Use when user asks for "report PDF", "보고서 PDF", "아름다운 PDF", "예쁜 PDF", or wants to publish a markdown document as a polished document.

2026-07-120

blog-shortform

2lab-ai/skills

블로그 URL 1개를 받아 60초 9:16 숏폼 영상(mp4)을 자동 생성하는 파이프라인. 한국 블로그 4종(티스토리/네이버/벨로그/브런치)에서 본문·이미지를 추출하고, 2인 대화 대본을 만든 뒤 로컬 fish-tts(soma-voice HTTP)와 GPT Image 2(gpt-image-2)로 음성·이미지를 생성해 Remotion v4로 합성한다. "블로그 URL로 숏폼 만들어줘", "이 글로 릴스 영상 만들어줘", "blog to shortform", "blog-shortform", "blog-url-to-shortform" 같은 요청에서 발동한다.

2026-07-120

cohere-stt

2lab-ai/skills

Speech-to-Text using Cohere Transcribe (2B, Apache 2.0)

2026-07-120

create-cli

2lab-ai/skills

CLI UX/spec: args, flags, help, output, errors, config, dry-run.

2026-07-120

dialogue-archive

2lab-ai/skills

Use when 지혁 asks to record/save/archive a conversation between him and Elon — produces (1) verbatim RAW markdown, (2) cleaned-up EDITED markdown with nothing removed, and (3) a messenger-style HTML blog post. Triggers on "대화 기록해줘", "이거 저장해서 보내줘", "블로그에 올리게 html로", "원문 그대로 + 교정본 + 포스트".

2026-07-120

frontend-design

2lab-ai/skills

Production frontend UI: pages, apps, components, polished non-generic design.

2026-07-120

name	github-deep-review
description	GitHub deep review: bugs, PRs, best fix, stale-or-real, read-code-first.

GitHub Deep Review

Start

Use gh, not web browsing, for GitHub refs:

gh issue view <n> --json number,title,state,author,body,comments,labels,updatedAt,url
gh pr view <n> --json number,title,state,author,body,comments,reviews,files,commits,statusCheckRollup,mergeStateStatus,headRefName,headRepositoryOwner,url
gh pr diff <n> --patch

For repo-local review, also inspect:

git status --short --branch
git fetch origin
git log --oneline --decorate -20

Use Grep tool (ripgrep) to search:

<key symbol/error/config/endpoint>

If the repo has local instructions (CLAUDE.md, AGENTS.md, etc.), issue/PR skills, docs lists, test guidance, or maintainer runbooks, read those before deciding.

Review Contract

Always answer these, explicitly:

URL/ref: issue or PR number and affected surface.
What is the bug or behavior being fixed?
Can we identify the root cause? If yes, where in code and why. If no, what evidence is missing.
For regressions, who/what introduced it and when? Include commit/PR provenance when traceable by bounded history; say unknown instead of guessing.
Is the current/proposed fix the best possible fix after reading adjacent code?
Would a bigger refactor improve correctness, clarity, or future maintainability?
What proof exists: tests, live repro, CI checks, docs, dependency docs/source, shipped/current behavior.
What remains risky or unverified.

Code Reading Depth

Read past the first touched file. Follow the real call path:

entrypoint -> validation/parsing -> routing/dispatch -> owner module -> shared helper -> persistence/network/runtime boundary
config/schema/docs -> runtime usage -> doctor/migration/fix path
provider/channel/plugin owner code -> generic core seam, only if multiple owners need it
tests around the touched surface plus adjacent regression tests

When behavior depends on a dependency, read the upstream docs/source/types or current package contract before assuming.

Prefer current source and executable proof over issue comments. Treat stale comments, old CI, and old release behavior as hints until rechecked.

Provenance

For bug/regression reviews, include a compact Provenance: answer when feasible:

Use git log -S/-G, git blame, linked PRs/issues, and tests.
Separate author, committer/merger, and current PR author when they differ.
Phrase as introduced by, made visible by, or carried forward by.
Include confidence: clear, likely, or unknown.
For features, docs, refactors, or untraceable bugs, write N/A or say what evidence is missing.

Fix Quality Bar

Good fixes usually:

live at the ownership boundary where the bug belongs
preserve public/backward-compatible behavior unless the issue is about retiring it
add a regression test at the smallest meaningful seam
avoid broad special cases, hidden migrations, semantic sentinels, and provider/channel IDs in generic core
update docs/changelog when user-visible behavior changes
fail clearly in runtime paths and repair through doctor/migration paths when that is the established contract

PR Review Shape

Lead with findings when reviewing a PR. Findings need file/line/symbol references and a concrete failure mode. Avoid vague "consider" comments.

If no blocking issues:

say no blocking correctness issues found
list the strongest proof checked
name residual risk/test gaps
answer whether the design is the best available shape

Do not approve, comment, close, merge, push, or land unless the user asked for that action.

Issue Review Shape

For bugs/issues:

Reconstruct the reporter's scenario and affected version/surface.
Check whether current main already fixes it.
Reproduce or create a minimal local/live proof when feasible.
If clear, identify root cause and proposed fix.
If solved on main, only comment/close when the user asks; include proof and the canonical commit/PR if known.

If reproduction is not feasible, say exactly what blocks it and what evidence would make the decision reliable.

Output Template

Use this shape when 지혁 asks "what is this about", "is this the best fix", or "what did we fix":

Ref: #123 / PR #456
Surface: <runtime/CLI/provider/channel/docs>

Bug: <one or two sentences>
Cause: <code path + confidence>
Provenance: <introduced/made visible/carried forward by commit/PR/date, or N/A/unknown>
Best fix: <what should change and why>
Refactor: <yes/no, specific shape>
Proof: <tests/live/CI/source/dependency docs>
Risk: <remaining uncertainty>

Keep it concise, but do not skip the cause/fix/refactor/proof decision.

Tool / Project Adoption Review (쓸만한가?)

PR/버그가 아니라 "이 레포/도구를 쓸지" 판단할 때. 위 root-cause 방법론과 별개 모드.

세 질문에 답한다 (순서대로):

메커니즘이 실제로 뭘 하나? README 주장을 코드로 환원한 동작. (안티패턴 2 참조)
이게 우리 문제를 푸나? 일반 좋음이 아니라 지혁/soma/p9의 구체적 미해결 문제와의 적합성. 적합 없으면 "흥미롭지만 우리엔 무관"이 정답.
리스크와 저위험 검증 경로. 채택 전 우리 워크로드로 재볼 수 있는 최소 실험(holdout/dry-run/벤치 재현).

Anti-patterns (검증된 실패 — 2026-06-20 지혁 교정)

AP1. 대리지표(vanity metric)를 품질로 착각 금지.

❌ "테스트 619개 → 진지한 프로젝트", "10K stars → 좋음", "커밋 많음/CI 초록/codecov 배지 → 신뢰".
이유: 테스트 수·LOC·⭐·커밋·트렌딩·배지는 AI가 다 양산하는 시대에 품질과 무상관. 테스트 30만 개여도 검증력 0일 수 있다. 양은 신호가 아니다.
✅ 대신: 주장의 핵심 모듈 1~3개를 직접 읽고 구현 깊이·정직성·엣지케이스 처리를 판단. 평가는 읽은 코드에 근거. 숫자 세지 말고 코드를 봐라.

AP2. 마케팅 용어를 메커니즘으로 환원하라 (그대로 옮기지 마라).

❌ README의 "X compression", "Nx faster", "output token reduction", "AI-powered", "60–95% fewer" 를 그 표현 그대로 리뷰에 전달.
✅ 절차: (a) 물리적으로 말이 되나 먼저 의심 — 예: "API 출력을 사후 압축"은 불가능(응답 받으면 이미 토큰 청구됨). 말 안 되면 README가 뭔가 다른 걸 그 이름으로 부르는 것. (b) 코드를 까서 실제 동작으로 환원 — 예: headroom "output reduction"의 실체 = output_shaper.py가 프롬프트 끝에 terse 문구 추가 + reasoning effort 사전 하향 = "사후 압축"이 아니라 "사전 프롬프트 조작으로 출력 길이 억제". (c) 환원 못 하면 "검증 안 됨 — 주장일 뿐" 명시.

AP3. 벤치마크는 자체측정·cherry-pick으로 가정.

숫자가 자체 벤치(N 작음)이거나 유리한 workload만 고른 건 아닌지. holdout/독립재현/counterfactual 여부 확인. README가 "estimate"라 자인하면 그대로 estimate로 전달.

Adoption 출력 템플릿

도구: <name> — <한 줄 실체(메커니즘 환원형, 마케팅 용어 X)>
실체 메커니즘: <코드로 까본 실제 동작>
우리 적합성: <지혁/soma/p9 구체 문제와 매칭, 또는 "무관">
리스크: <정보손실/복잡도/의존성/레이턴시/pre-1.0 등>
검증 경로: <채택 전 우리 워크로드로 잴 최소 실험>
판정: 쓸만/조건부/패스 + 근거(읽은 코드, 대리지표 아님)

p9/soma 적용 노트

p9 issue: bd list → bd show <id> → 위 템플릿으로 답.
soma 코드 변경 검토: cd ~/2lab.ai/soma && git diff + 위 워크플로우.
"고친다"고 말하기 전에 ./scripts/bd-verify로 증명 후 bd-close-verified.