← Home

TAB Project Wiki

CHANGELOG

36 min read

cmds-llm-wiki — Changelog

템플릿 버전 이력. (cmds-system-files 의 CHANGELOG 형식 참고)

v1.10.0 — 2026-07-23 (Paper Ingest Mode — 12-step paper atomization)

Source: 운영 볼트 schema v6.2 의 Paper Ingest Mode 를 sanitize 하여 이식. v1.9.0 에서 “학술 전용 버티컬이라 의도적 제외” 했던 항목을 사용자 결정으로 포함 — RQ 카드 (v1.9.0) 와 결합해 논문 → 질문 → 논증의 전체 파이프라인이 킷 안에서 완결된다.

Added — Paper Ingest Mode (/ingest 자동 감지)

  • 감지: DOI/arXiv/저널 URL · Abstract+References 구조 PDF · 00. Inbox/02. Papers/ → Standard 대신 논문-스코프 12단 분석 (1 Raw Source + 1 Paper Hub + 지식 원자 30~100+ + Wiki 승격). 질문 예산 2개 (P-0 목적/RQ, P-1 유형).
  • P-0 → P-7 파이프라인 리소스: .agents/skills/ingest/resources/paper-ingest.md (progressive disclosure — 논문 감지 시에만 로드). scope guard 는 개인 도메인 레지스트리 대신 Core Context 정렬로 일반화, 모선 단계는 Mode B 옵션.
  • ingest 3-surface 포인터: .claude/commands/ingest.md + .codex/commands/ingest.md (폴백 = 단일 실행 중요도순) + .agents/skills/ingest/SKILL.md.

Added — 12-Step Analysis Schemes + 신규 타입 2종

  • 90. Settings/Templates/12-Step Analysis Schemes.md — 6유형 (quantitative / qualitative / theory-concept / mixed-methods / scale-development / meta-analysis) × 12 좌표 (S01 CITATION ~ S12 WRITING VALUE) 통합 스펙 + Coverage & Exposition Contract + 원자 분해 규칙 + Decomposition Heuristics.
  • type: paper-hub (S00 허브 — paperType·citekey·doi·targetManuscript) + type: paper-analysis (지식 원자 — analysisStep 2~12·analysisStepName·paperHub). 원자는 claimType/evidenceScope 면제 — 검증은 p7 인용 충실도로 대체.
  • 새 폴더 40. Paper Analyses/{citekey}/. Wiki Page layer 확장: theory / method / scale.

Added — Templates ×4 (7 → 11) + 검증 스크립트

  • Template_Paper Hub (Coverage Map·Step Map·Atom Catalog) / Template_Paper Analysis Note (Analysis Context + 쉬운 도입→정밀 해설→원문 근거 verbatim quote) / Template_Atomization SPEC (팬아웃 배포용 작업 명세) / Template_Scale Page (measuredConstruct·itemCount + Psychometrics 표 + 저작권 콜아웃).
  • 90. Settings/Scripts/p7_verify.py — P-7 기계 게이트: YAML·좌표·provenance·인용 전수 verbatim·Coverage·Atom Catalog↔파일 대조. ALL PASS 가 통과 조건 (서브에이전트 자체검증은 게이트 아님).

Added — Paper Ingest Guide (사용자 매뉴얼)

  • 90. Settings/Sharing/Paper Ingest Guide.md — 12단계 (S01~S12) 공통 골격 + 6유형별 변형 + 산출물 구조 + 인용 규율 + RQ 연계 + 실행 전략 + Zotero 옵션 + FAQ 를 사람 독자용으로 해설.

Changed

  • 스키마 (CLAUDE.md ↔ AGENTS.md parity): type 목록 + provenance 대상 (×2곳) + Layer별 프로퍼티 (Paper Hub/Analysis 블록) + 폴더 트리 (40 + Scripts) + File Naming 3행 + Callout 2종 (Analysis Context·quote) + camelCase 키 8종 (paperType·analysisStep·analysisStepName·paperHub·targetManuscript·doi·measuredConstruct·itemCount) + Citation Standard 에 “Paper Mode 는 citekey 항상 사용” 명시.
  • /status 3-surface 에 Paper Analyses 카운트, index.md 에 Paper Analyses (허브만) nav 섹션, qmd config 에 paper_analyses 컬렉션, README/Setup Guide/CLAUDE-Template/kit 내부 문서 트리·카운트 정합 (6→7 폴더, 7→11 템플릿).

Kit adaptations (운영 볼트와 의도적으로 다른 점)

  • P-0.1 도메인 가드: 운영 볼트의 9-도메인 레지스트리 → Core Context §2 재활용 축 정렬 확인으로 일반화.
  • Zotero Tier-0 는 명시적 옵션 (미사용자는 provisional citekey 경로 완결).
  • 운영 실증 사례 참조 (특정 논문 허브 wikilink) 는 수치만 남기고 제거.

Known gaps

  • /autoresearch 는 여전히 운영 볼트 전용 (v2.0 후보). /lint 는 RQ/citation/paper 커버리지 리포팅 미지원 (후속 마이너 후보).

v1.9.1 — 2026-07-23 (docs: RQ/Synthesis usage guide + onboarding next-steps)

Docs-only patch — no schema, template, or harness behavior changes.

Added — Setup Guide 사용법 문서화

  • FAQ 신설: “Research Question / Synthesis 카드는 언제 쓰나” — 두 카드의 용도 구분 (알게 된 것 vs 묻고 있는 것/주장하는 것), 승격 흐름 (Open Question 콜아웃 → RQ 카드, sourceCallout 역추적), status 생애 추적, “둘 다 필수 아님 — 산출물 단계에 도입” 가이드, Citation Standard 연계.
  • “다음 단계” 성장 경로에 RQ 승격·Synthesis 작성 시점 2개 bullet 추가.
  • 헤더 대상 버전 v1.6.2+ → v1.9.0+, 갱신일 표기.

Changed — /onboard wrap-up

  • 온보딩 마무리 추천 단계에 “(성장 후) 반복 Open Question → RQ 카드 승격, 주장 방어 → Synthesis 카드” 항목 추가 (4→5단계).

v1.9.0 — 2026-07-23 (Research Question + Synthesis cards · Citation Standard · CLAUDE.md Operations parity)

Source: 운영 볼트(CMDS_LLM_Wiki) schema v6.1 (2026-07-07) 의 “논문 산출층” 중 일반화 가능한 절반만 선별 이식 — 질문·논증의 1급 객체화. 학술 전용 버티컬(v6.2 Paper Ingest Mode, 12단 논문 분석)은 의도적으로 제외하고 운영 볼트 전용으로 유지한다.

Added — Research Question 카드 (type: research-question, v6.1)

  • 새 폴더 20. Wiki/25. Questions/ + 파일명 RQ-{slug}.md. in-page > [!question] Open Question 콜아웃을 1급 카드로 승격 — 반복 등장하거나 산출물(논문·책·블로그 시리즈·제품 결정)로 이어질 질문을 first-class object 로 추적.
  • 프로퍼티: status(open/investigating/answered/parked/superseded) · questionType(descriptive/causal/comparative/methodological/normative/design) · feedsInto · hypotheses · evidenceFor/evidenceAgainst · cites(권장) · sourceCallout(역추적) · related.
  • RQ 는 claim 이 아니라 질문이므로 v5 claimType/verificationStatus 대신 status 를 쓴다.

Added — Synthesis 카드 (type: synthesis, v6.1)

  • 30. Queries/ 에 query-result 와 같은 폴더, type 으로만 구분 (YAGNI — flat 유지). Query 가 반응형 답변이면 Synthesis 는 능동 논증: thesis + 근거 claim + counters + gap + targetVenue.
  • /query operation 에 Step 5 추가: 능동 논증이 필요하면 type: synthesis 카드로 저장.

Added — Citation Standard (v6.1, 옵션)

  • BetterBibTeX citekey (authYearShorttitle) + Pandoc/CSL [@citekey] inline + cites: frontmatter + ## References + Raw Source citekey 이중 연결. 옵션 규약 — 채택하지 않아도 볼트 운영에 지장 없음.

Added — Templates ×2 (5 → 7)

  • Template_Research Question.md / Template_Synthesis.md — v6 provenance (model/effort) + quoted description: "" 포함, 생성 시점부터 스키마 만족.

Fixed — kit CLAUDE.md Operations parity (v1.7.1 부터 이월된 부채)

  • kit CLAUDE.md Operations 챕터에 §0 Capture Tabs / §5 Verify / §6 Audit 추가 — kit AGENTS.md 와의 비대칭 해소 (Claude-first entrypoint 표기로 적응).

Fixed — CLAUDE.md ↔ AGENTS.md parity 복원 + AGENTS.md 일반화 (릴리스 전 적대적 검증에서 적발된 pre-1.9 drift)

  • Frontmatter Standards 7곳 동기화: author 행 (Claude → Claude/Codex/Grok), effort 문구 중립화, CLAUDE.md category 에 AI Research 누락 보충, mainVaultRelated 문구 (URL_ENCODED_PATH 리터럴 예시 제거 → stat 검증 문구), chapterPart 예시, camelCase 헤딩의 mothership 내부 alias (@CMDS-Guide) 제거, Quality Control v4 의 운영볼트 이력 문장 제거 + Bias Check 콜아웃 표기.
  • Compatibility Matrix: AGENTS.md 헤딩 Codex Compatibility Matrix (v1.3) → Cross-Agent Compatibility Matrix (stale v1.3 태그 제거), 정보량 없는 Status 열 제거.
  • AGENTS.md mothership 옵션화: Vault Overview·Core Context·ingest Step 0-a·Cross-Vault Reference 를 CLAUDE.md 와 같은 (옵션) 프레이밍으로 — standalone Codex 사용자가 모순된 지시를 받지 않도록. 잔존하던 원저자 볼트 구조 예시 링크 2건을 generic placeholder 로 교체.
  • type 목록 완결: 필수 프로퍼티 type 에 실제 사용 중이던 documentation·inbox 추가 (양 스키마).
  • 부속 파일 정합: CLAUDE-Template.md (규칙 문구 + RQ/Synthesis 최소 프로퍼티 블록 + naming 행), index.md (description 큰따옴표 + Questions 반영 + nav stub), Setup Guide.md (템플릿 7종 표기 + 잘못된 brace glob 정리 커맨드 수정), Template_Synthesis 의 미선언 status 키 제거.

Changed

  • Provenance Rule 대상 타입에 research-question·synthesis 추가 (양 스키마 ×2곳).
  • 필수 프로퍼티 type 목록, Folder Structure 트리, File Naming 표 (RQ-{slug}.md 행), camelCase 키 목록 (+11 keys), Parity Contract 문구 (v6/v6.1 키 포함) 갱신.
  • README 트리: 25. Questions/ 추가 + Templates 카운트 stale “4종” → “7종” 정정 (v1.7.1 부터 5종이었음).

Kit adaptations (운영 볼트와 의도적으로 다른 점)

  • cites 는 운영 볼트에서 필수지만 킷에서는 권장 — Citation Standard 자체가 옵션이므로.
  • feedsInto/targetVenue 예시를 중립화 (논문 / 책 챕터 / 블로그 시리즈 / 제품 결정).
  • localDev (v6.1) 는 개인 로컬 경로 패턴이라 미이식.

Known gaps

  • /autoresearch 는 운영 볼트 전용 (Generator-Verifier loop — v2.0 후보).
  • /lint 는 아직 RQ/synthesis/citation 커버리지를 리포팅하지 않는다 (후속 마이너 후보).

v1.8.0 — 2026-07-23 (Description Rule + v6 Provenance schema)

Source: 운영 볼트(CMDS_LLM_Wiki) schema v6 (2026-07-05) 에서 검증된 두 규칙을 sanitize 하여 이식. Schema/template edits were staged in the canonical on 2026-07-07; this release publishes them.

Added — Description Rule (CLAUDE.md + AGENTS.md CRITICAL RULES)

  • description 값은 항상 큰따옴표 — 값에 콜론(: )·엠대시·괄호·#·wikilink 가 들면 unquoted YAML 은 파싱 에러로 노트 전체가 렌더 실패하는 실전 사고 클래스를 규칙화. 빈 값도 description: "".
  • Pre-Flight Checklist, Essential (Post-Compact) #6, 필수 프로퍼티 7개 표가 모두 double-quoted 요구사항을 반영하도록 갱신.
  • 두 스키마 파일의 자기 frontmatter description 도 큰따옴표로 전환 (dogfooding).

Added — Provenance Rule (v6): model + effort

  • 에이전트가 쓰는 모든 콘텐츠 페이지(raw-source·wiki-page·query-result·moc·inbox)는 author 바로 뒤에 model(상세 세션 모델 id) + effort(추론 강도) 를 항상 기록 — Claude Code·Codex·Grok cross-agent provenance.
  • 새 “Provenance 프로퍼티 (v6 신설)” 섹션 + Essential #9 + Pre-Flight 항목 추가. camelCase 허용 키 목록에 model, effort 추가.
  • harness 정의 파일(.claude/commands/*, .codex/commands/*, .agents/skills/*)의 자기 frontmatter 는 provenance 면제 (슬래시 커맨드/스킬 파서 혼선 방지) — Description Rule 만 적용.

Changed — Templates ×5 (schema alignment)

  • Template_Raw Source / Template_Wiki Page / Template_Query Result / Template_MOC / Template_AI Research Capture: description: → description: "" (quoted empty) + author 뒤에 model: / effort: 키 추가 — 템플릿으로 생성되는 페이지가 생성 시점부터 v6 스키마를 만족.

Known gaps (carried over from v1.7.1)

  • Kit CLAUDE.md Operations chapter still lacks the Capture Tabs / Verify / Audit sections that kit AGENTS.md already has (asymmetry — port planned).
  • /autoresearch remains operating-vault-only (Generator half of the Generator-Verifier loop).

v1.7.1 — 2026-07-03 (hook runtime fixes + broken template dependency + sanitization)

Patch release driven by the 2026-07-03 full satellite system-files audit (multi-agent, adversarially verified). No schema or feature changes.

Fixed — hooks (real runtime bugs, both .claude/ and .codex/ copies)

  • qmd-reindex.sh debounce was a no-op on macOS — date +%s -r FILE is GNU argument order; BSD date fails silently (empty output), so the settle-wait loop never ran. Replaced with date -r FILE +%s.
  • Stale-lock permanent lockout — a worker killed mid-run (or aborted by set -e on a qmd failure) left .running behind, silently disabling all future auto-reindexes. Added a 10-minute mtime expiry on the lock + trap EXIT cleanup in the worker.
  • validate-raw-source.sh Check 2 false positives — the line counter stopped at the first ## heading after ## Original Content, but verbatim bodies legitimately preserve the source article’s own H2 headings, so long real bodies were counted as near-empty and would block the next edit. Counter now stops only at a trailing ## Metadata section (or EOF).

Fixed — broken dependency

  • Template_AI Research Capture.md now shipped — /capture-tabs (distributed since v1.6.0) referenced this template as its required shape, but the kit never included it. Added (sanitized), so the kit’s template count is now 5 (Raw Source / Wiki Page / Query Result / MOC / AI Research Capture).

Fixed — templates vs schema drift

  • Template_Wiki Page.md — added explored: false (v4 mandatory) and verificationStatus: unverified (v5 default); pages created from the template no longer violate the schema on creation.
  • Template_Raw Source.md — added v2 keys collectionPurpose / source-vault / mainVaultRelated / mainVaultCmds (the ingest command already required them).
  • Template_Query Result.md — added reusableFor (the query command’s Step 6 already set it).

Fixed — sanitization / genericization leftovers

  • CLAUDE.md / AGENTS.md Cross-Vault Reference: hard-coded author paths (CMDSPACE/40. Docs/..., 03-1. Claude Code (MBP)/...) → {your-mothership-vault-name} placeholders; entry-point note phrased as optional.
  • AGENTS.md Core Context: removed dangling reference to a CLAUDE.md alias table that no longer exists in the kit; the 9-system-files table is now framed as an optional author example (standalone users can skip); stale “91 categories” → 87.

Known gaps (deferred to a future minor)

  • Kit CLAUDE.md Operations chapter still lacks the Capture Tabs / Verify / Audit sections that kit AGENTS.md already has (asymmetry — port planned).
  • /autoresearch remains operating-vault-only (Generator half of the Generator-Verifier loop).

v1.7.0 — 2026-07-01 (v5 Verification schema completion + ingest/lint cross-vault hardening)

Source: 운영 볼트(CMDS_LLM_Wiki)에서 검증·안정화된 v5 검증 스키마와 cross-vault 링크 하드닝을 sanitize 하여 킷으로 이식. schema 는 킷이 아직 v4 수준이었고, ingest/lint 는 링크 rot 방지 로직이 빠져 있던 갭을 메움.

Added — v5 Verification schema completion (CLAUDE.md + AGENTS.md 양쪽)

  • Verification Properties (v5) 3 기준 블록 (지식요건해당성·정합성·확증가능성 + 운영 규칙) 을 Quality Control (v4) 다음에 추가. CLAUDE.md 는 “vs CLAUDE.md policy”, AGENTS.md 는 “vs AGENTS.md policy” 로 각자 자기 스키마 참조.
  • Wiki Page 6 v5 키: claimType, evidenceScope, verificationStatus, verifiedAt, verifiedBy, disputed 추가.
  • camelCase 목록 v5 엔트리 8개: ✅ 6개(위 키) + ❌ 2개(claim_type, verification-status).
  • > [!warning] Disputed Claim 콜아웃 을 Callout Conventions 에 추가 (Contradiction 과 별개 — /verify --resolve 로만 해소).
  • Parity Contract 콜아웃 (CLAUDE.md ↔ AGENTS.md) — 두 스키마 미러가 반드시 동일해야 하는 섹션(Cross-Agent Matrix · Frontmatter Standards · v5 3기준 · Callout Conventions) 을 명문화.

Added — lint v5 coverage reporting

  • .claude/commands/lint.md: Step 8 헤더 v2/v4 → v2/v4/v5, v5 검증 필드(claimType/evidenceScope/verificationStatus) 커버리지 체크 + disputed → /verify --resolve 큐 추가. Output 에 “v5 Verification Coverage” 항목 추가.
  • .codex/commands/lint.md: 이미 v5 verification + cross-vault integrity 항목 보유(변경 불필요).

Added — ingest/lint cross-vault hardening (sanitized)

  • Step 0.5 Format Conversion (binary → markdown): .pdf/.pptx/.docx/.hwp/.epub/image 등을 변환 후 ingest, 원본은 80. References/Attachments/ 로 보존 (.claude + .codex ingest 양쪽).
  • mainVaultRelated URL percent-encoding 필수 검증: qmd 반환 실제 경로 사용 → urllib.parse.quote(safe='') 인코딩 → test -f stat 검증 → obsidian://open?vault=...&file= 구성. 손으로 URL 추측 금지.
  • mainVaultCmds authoritative 목록: find ... -name "📚 [0-9][0-9][0-9] *.md" 로 실제 카테고리 집합을 만든 뒤 정확 매칭. 추측 금지.
  • Cross-vault link integrity 검사 (ingest Step 6/7 + lint Step 10): touched 파일의 obsidian:// URL 과 mainVaultCmds 를 모선 파일시스템에 stat 대조, 깨진 링크·리터럴 placeholder 적발.
  • Index-sync 검증 강화: 신규 Wiki 페이지가 index.md 섹션에 실제로 나타나는지 grep -F, Stats 카운트를 find | wc -l 로 대조.
  • CLAUDE.md mainVaultRelated 설명 을 AGENTS.md 와 정합하도록 obsidian:// 클릭 링크 형식으로 통일 (Parity Contract 준수).

Sanitization

  • 모든 이식분에서 실제 경로/볼트명 제거 → {PATH_TO_YOUR_MOTHERSHIP_VAULT} · {your-mothership-vault-name} placeholder. cross-vault 단계는 “mothership only” 로 표기 — 모선 없이 단독 운영 시 skip.

No breaking changes

기존 v1.6.x 볼트는 갱신 파일만 pull. v5 키는 전부 추가형(신규 페이지 기본 verificationStatus: unverified), 기존 페이지는 /verify//lint 접촉 시 점진 backfill.

v1.6.2 — 2026-06-27 (정합성 audit — sanitization sync + doc accuracy + folder scaffold)

Source: 배포 전 8-차원 정합성 audit (version·inventory·cross-harness·placeholder-PII·frontmatter·doc·zip·hooks), 각 finding 을 적대적으로 검증. blocker 0, confirmed 20 (major 5 / minor 5 / nit 8 + 1 dismissed). 모두 반영.

Fixed — Sanitization sync (Codex 미러가 Claude gold 와 어긋나 있던 부분)

  • .agents/skills/refresh-context/SKILL.md — 하드코딩돼 있던 5개 실제 개인 에세이 제목을 generic 패턴 (Read("<essay path>") + {PATH_TO_YOUR_ESSAYS}) 으로 교체. Claude command 는 이미 genericize 돼 있었으나 Codex skill 만 누락됐던 단일 spot.
  • AGENTS.md CJK Person Naming 예시 — 실명(Yohan Koo)·핸들(johnfkoo951)·제3자(안창현/Changhyun Ahn, Andy Suh)·실제 저자(zhanghandong) 를 CLAUDE.md gold (홍길동/Gildong Hong/johndoe) 로 동기화.
  • collectionPurpose 예시의 실제 클라이언트명 → 기업 임원교육 사례 로 generic 화 (CLAUDE.md + AGENTS.md 양쪽).

Fixed — Doc accuracy (제품과 어긋난 사용자 문서)

  • “7 commands” → “11 commands”: README (×3) + Setup Guide. onboard/capture-tabs/verify/audit 누락분 추가.
  • README 가 v1.6.0 Codex 듀얼-harness 를 누락 → .codex/ · .agents/skills/ · AGENTS.md 를 구조도·파일목록·소개 bullet 에 반영.
  • Setup Guide 의 죽은 vercel 쇼케이스 URL (cmds-llm-wiki.vercel.app, SSO 벽) → llm-wiki.cmdspace.work.
  • 예시 wiki 카운트 10개 → 16 (README·Setup Guide·index.md). index.md Stats 표 자체모순(헤더 8 vs 합계 10) 을 실제 디스크 수치(16: Concepts 9·Entities 3·Guides 2·MOC 2)로 정정.
  • Setup Guide placeholder 파일 수 12개 → “약 22개 (sed glob 이 전부 커버)”. LLM-Wiki-Starter-Kit.md 의 “Codex는 이름만 변경” FAQ + 3-operations 설명을 듀얼-harness/11-command 현실로 갱신.
  • AGENTS.md Codex Compatibility Matrix Notes 칼럼을 CLAUDE.md genericized 표현으로 정렬 (v2/v4/v5, 하드코딩 9 제거) + onboard Claude-전용 안내 추가.

Added — Folder scaffold (문서가 참조하나 미생성이던 카테고리)

  • 00. Inbox/01. Articles/, 00. Inbox/05. AI Research/ (/capture-tabs 기본 저장 위치), 10. Raw Sources/12. Papers/, 10. Raw Sources/16. AI Research/ 에 .gitkeep 추가 — 표준 카테고리 세트 완성. CLAUDE.md 폴더 다이어그램도 05/16 AI Research 포함하도록 AGENTS.md 와 정합.

Housekeeping

  • CHANGELOG v1.0.0 의 “18 Web Clipper” 표기를 17 로 정정 (당시 실제 17개, Stibee 는 v1.4.0 에서 18 로).
  • 정렬: canonical .DS_Store 4개 제거, archives/cmds-llm-wiki-starter-kit.zip (구 v1.3.0 가리키던 stale alias) 삭제 — archives/ 는 versioned ZIP 만 보관.

Deliberately NOT changed

  • .codex/hooks/validate-raw-source.sh 의 stub-skip 순서 (Check 1 뒤) 는 Claude 와 다르지만 의도적·문서화된 설계 (stub 도 heading 은 필요). 양 harness 모두 well-formed stub 을 정상 통과하므로 유지.
  • Karpathy X-thread 데모 wiki 페이지의 실제 공개 인물명 (Karpathy·kepano·Changhyun Ahn·Yohan Koo) 은 의도된 예시 출처라 유지 (sanitization 대상 아님).

No schema/structure breaking changes

기존 v1.6.x 볼트는 갱신된 파일만 pull 하면 됨. 폴더 추가는 비파괴적 (.gitkeep).

v1.6.1 — 2026-06-16 (3-Layer 개념층 vs 물리 폴더 정합 노트)

Source: 운영 볼트의 지식 레이어 재검증 과정에서, 템플릿이 30. Queries/ 물리 폴더를 제공하면서도 3-Layer(Raw/Wiki/Schema) 개념과의 관계를 설명하지 않아 신규 사용자가 “30. Queries 는 4번째 레이어인가?” 혼동하던 지점 발견.

Changed

  • 3-Layer Architecture: “이 볼트에서의 구현” 표 아래에 callout 추가 — 30. Queries/ 는 4번째 구조 레이어가 아니라 Ingest-Query-Lint 의 Query 단계 산출 위치임을 명시 (구조 축 Raw/Wiki/Schema vs 운영 축 Ingest/Query/Lint 구분). 유저별 콘텐츠가 아닌 일반 개념 정합이라 템플릿에 반영.

v1.6.0 — 2026-06-08 (Codex 호환 — dual-harness)

Source: 운영 볼트가 Claude Code + Codex 양 harness 로 정합화(같은 operation 이름을 .claude/ · .codex/ · .agents/skills/ 에 1:1 유지)된 것을 공개 킷에도 반영. 이전 킷은 Claude 전용이었음.

Added

  • AGENTS.md — Codex / 타 에이전트용 schema mirror (CLAUDE.md 와 같은 규칙, audience 만 다름).
  • .codex/ — Codex operation harness: commands/ 10개 (Claude commands 와 1:1 mirror) + hooks/ (validate-raw-source.sh, qmd-reindex.sh) + hooks.json. 코덱스 hook 은 단순 payload shape (tool_input.path / file_path / path) 도 처리하고, Book Ingest status: stub chapter 는 body-length 검사에서 제외.
  • .agents/skills/ — Codex reusable operation skills 10개 ({operation}/SKILL.md).
  • .claude/commands/ 에 capture-tabs, verify, audit 3개 추가 (기존 8 → 11). Capture Tabs = AI research / 브라우저 탭 묶음 → Inbox pre-ingest layer. Verify = 단일 페이지 3-기준 검증. Audit = vault/MOC 전체 점검 + /verify 큐.

Changed

  • CLAUDE.md: Operations 섹션에 Agent Note + Cross-Agent Compatibility Matrix + Codex Compatibility Notes 추가, Folder Structure 에 .codex/ · .agents/ · AGENTS.md 반영, Essential operations 한 줄 갱신. version 1.5.0 → 1.6.0.

Notes

  • Codex hook 경로는 /c/Users/David/Desktop/Vaults/04_GuWiki placeholder — 셋업 시 절대 경로로 치환 (또는 /onboard 가 처리). Claude hook 은 $CLAUDE_PROJECT_DIR 라 치환 불필요.

v1.5.0 — 2026-06-01 (/onboard 인터뷰 셋업 커맨드)

Source: 온보딩이 수동 Setup Guide.md(sed)뿐이라, “온보딩해줘/처음 시작할게” 트리거로 필수 질문을 물으며 placeholder·Core Context를 자동으로 채우는 대화형 단계 신설.

Added

  • .claude/commands/onboard.md — 인터뷰 기반 first-run 셋업 커맨드. pre-flight 감지 → 필수 5문항(위치/이름 · Mode A·B · 모선 경로 · Core Context §1 정체성 · §2 재활용 축) + 옵션(§3·§4) → placeholder 일괄 치환 + Core Context 작성 + status: active. answer-1-of-many · resume · 음성 모드 · passive fallback 포함 (cmds-onboarding 철학 차용).

Changed

  • Setup Guide.md: 상단에 /onboard 대화형 경로 안내 추가, 대상 버전 v1.3.0 → v1.5.0+.

v1.4.1 — 2026-05-31 (Doc Fixes · Live Showcase URL · Placeholder Count · Mothership Count Soften)

Source: cmds-system-files v4.9.0 audit surfaced three documentation bugs in this kit.

Fixed

  • README “Live Showcase” link → https://llm-wiki.cmdspace.work (the prior cmds-llm-wiki.vercel.app returned HTTP 401; the branded apex is live 200).
  • CHANGELOG placeholder count corrected from “10 files” to “13 files carry placeholder tokens” (live count).
  • Core Context.md mothership reference softened from a hard “5 system files” count to durable “multiple audience-specific system files … + semver changelog” phrasing — the mothership is now 9 files / 6 public after the 2026-05-27 8→9 restructure, and an un-counted phrasing won’t re-stale.

Note

  • No schema/structure changes. canonical ↔ DEV mirror ↔ ZIP regenerated together.

v1.4.0 — 2026-05-27 (Missing Files Audit · Hotkeys · Empty Queries Folder · Stibee Clipper)

Source: User audit “왜 옵시디언 웹클리퍼 json 파일들이 빠져있지?” surfaced three template gaps that had silently drifted from the live operating vault. Same release also closes the 3-place sync loop documented at v1.3.0 by promoting the canonical (_starter-kit/cmds-llm-wiki/) as single source of truth.

What’s missing → now included

  • .obsidian/hotkeys.json (9.2 KB · 73 bindings) — Korean PKM-friendly Obsidian hotkey set used in the operating vault. Heading shortcuts (⌘1~4), wikilink/callout inserters (⌘[ / ⌘]), sidebar toggles, etc. Optional — delete if your own keymap suits you.
  • 30. Queries/ (empty folder + .gitkeep) — Was missing entirely from the v1.3.0 template even though /query writes results here and the README structure referenced it. .gitkeep ensures the folder ships through git/zip.
  • clipper-stibee.json — Korean newsletter platform (Stibee) Web Clipper. Operating vault has had this since 2026-04-14; v1.3.0 sanitization excluded it. Re-included for Korean newsletter users. (Sanitization decision retained: clipper-parkjoon-mdshare.json stays excluded — it targets a personal domain.)

Numbers

  • Clipper JSON: 17 → 18
  • Folders shipped: was missing 30. Queries/ and .obsidian/; now complete
  • No schema changes (CLAUDE.md still v1.4 over v1.3 structurally identical except frontmatter version/date modified)

3-place sync canonicalised (operational doc, not template content)

Documented in parent vault’s CMDS_LLM_Wiki/CLAUDE.md “Starter Kit Distribution (3-Place Sync)” section:

  1. Canonical: _starter-kit/cmds-llm-wiki/ (this folder, unversioned)
  2. DEV mirror: /DEV/cmds-llm-wiki/ ↔ GitHub johnfkoo951/cmds-llm-wiki
  3. ZIP archives: _starter-kit/cmds-llm-wiki-{starter-kit,vX.Y.Z}.zip

🚫 ~/Downloads/cmds-llm-wiki-*.zip is no longer used as a distribution surface. External users obtain releases via GitHub Releases.

Placeholder coverage

13 files carry placeholder tokens (~10 require user substitution before operation). Placeholder types: David, David, /c/Users/David/Desktop/Vaults/04_GuWiki, {PATH_TO_YOUR_MOTHERSHIP_VAULT}, {your-mothership-vault-name}, 2026-07-28. Run the sed commands in 90. Settings/Sharing/Setup Guide.md to replace all in one shot per Mode A (standalone) or Mode B (mothership-satellite).

Files changed

  • .obsidian/hotkeys.json (new)
  • 30. Queries/.gitkeep (new)
  • 90. Settings/Sharing/clipper-stibee.json (new)
  • CLAUDE.md — frontmatter version: "1.3" → "1.4", date modified
  • README.md — clipper count 17→18 (with Stibee), hotkeys.json mention, structure section
  • CHANGELOG.md — this entry

v1.3.0 — 2026-05-04 (Quality Control v4 + Tool Outputs + Multi-Harness)

Source: Distilled from one month of operating the parent vault since v1.2.0. Three classes of upgrade: (1) a Book Ingest hook bug uncovered during real ingestion, (2) page-level quality-control properties that emerged from a self-audit (“we have 89 wiki pages but 0% are human-verified”), (3) generalised harness compatibility for non-Claude-Code agents.

Bug fixes

  • validate-raw-source.sh — skip status: stub: Book Ingest chapter stubs intentionally have no ## Original Content until the user reads them and re-invokes /ingest for promotion. The v1.1.0 release added Book Ingest but the verbatim-validation hook still blocked stubs as malformed Raw Sources. Hook now reads frontmatter and exits 0 for status: stub files.
  • lint.md indentation: Step 8 sub-bullets used 2-space indent instead of TAB, violating the project’s body-uses-TAB rule. Fixed.

What’s new

Exploration Gate (v4 frontmatter)

A page-level human-verification track on Wiki pages. Addresses the gap where confidence: high was being set by the LLM at compile time without any human read pass.

  • New optional Wiki page properties: explored: false|true, exploredBy: "[[Name]]", exploredDate: YYYY-MM-DD
  • New > [!note] Bias Check callout pattern (Counter-argument + Data gap) for high-confidence and synthesis-heavy pages
  • New > [!check] Exploration Gate callout for documenting verification status
  • /ingest now emits new pages with explored: false by default and adds Bias Check on confidence: high
  • /lint Step 8 reports explored coverage % and high-confidence pages without bias check
  • /status reports explored: and mainVault*: coverage %
  • /query flags missing mainVaultRelated, explored, or Bias Check during synthesis

70. Outputs/ Tool Output Convention (optional)

A separate folder for external-tool side products (graphify, audio-transcriber, etc.) that should not pollute the Wiki layer. Generic pattern: 70. Outputs/{tool-name}/2026-07-28-{topic}/. Outputs are exempt from the standard schema (frontmatter, naming) — they’re the tool’s format. Insights distilled from outputs go into 30. Queries/ or get absorbed into Wiki pages, so Wiki body never wikilinks into 70. Outputs/. Skip this folder entirely if you don’t run such tools.

CJK Person Naming Rule

Native script only for Korean / Chinese / Japanese person entities; English Romanization moves to aliases. Examples: 홍길동.md, 张汉东.md (not 홍길동 (Gildong Hong).md). Latin handles + real-name combos (kepano (Steph Ango).md) keep their existing form. Reasons: file-name duplication adds wikilink friction, Romanization is transliteration not identity, Obsidian graph/search reads aliases.

Multi-harness comment headers

Every .claude/commands/*.md now carries a YAML-comment line listing the equivalent tool names on Antigravity (Gemini) and similar harnesses — view_file, write_to_file, replace_file_content, list_dir, grep_search, run_command, read_url_content. Helps users running these commands through non-Claude-Code agents identify the equivalent tool surface.

Genericised “7 reuse axes”

/ingest and /query previously hard-coded the parent vault’s specific axes (PhD / 학술 / 강의 / 컨설팅 / CMDS 시스템 / 에세이 / 제품). Now both files reference Core Context §2 directly with neutral examples (학술 / 저술 / 강의 / 컨설팅 / 제품 / 에세이 / 커뮤니티). User-defined axes (5~9 recommended) remain the source of truth.

Setup Guide (deeper personalization manual)

New 90. Settings/Sharing/Setup Guide.md — covers Mode A (standalone) vs Mode B (mothership-satellite) operation, single-pass sed replacement commands per mode, verification grep, 8-step setup procedure, and 7-item FAQ. Complements README.md (which stays as the GitHub-facing 5-step quick start). Convertible to PDF via the same md-to-pdf workflow used to produce the 2026-04-30 standalone PDF.

Files changed

  • .claude/hooks/validate-raw-source.sh — stub skip
  • .claude/commands/{ingest,inbox,lint,query,refresh-context,reindex,status}.md — Antigravity equivalents header
  • .claude/commands/ingest.md — Quality control v4 block + genericised axes
  • .claude/commands/lint.md — Step 8 v2/v4 coverage + indent fix
  • .claude/commands/status.md — explored + mainVault coverage lines
  • .claude/commands/query.md — quality gap flagging + genericised axes
  • CLAUDE.md — 70. Outputs/ section, Frontmatter Standards v4 block, CJK naming rule, Bias/Exploration callouts, version bump 1.0 → 1.3
  • 90. Settings/Sharing/Setup Guide.md — new
  • README.md — pointer to Setup Guide above the 5-step quick start
  • CHANGELOG.md — this entry

Migration notes

No breaking changes. Existing v1.2.0 vaults adopt v1.3.0 by:

  1. Pulling the updated hook (validate-raw-source.sh) — required if you use Book Ingest
  2. Pulling updated commands and CLAUDE.md — adds optional fields, doesn’t break existing pages
  3. (Optional) backfilling explored: false on existing Wiki pages via /lint flag list

v1.2.0 — 2026-04-29 (Cohort Learnings)

Source: Distilled from a real classroom cohort that adopted LLM Wiki Starter Kit as the practical backbone of a regular university course (10 students, 10 distinct domains — systematic review, healthcare AI, anti-aging content, silver-care market analysis, sports analytics, biomedical research, embryology AI, etc.). All cohort-specific names, organizations, and personal contexts have been anonymized — only the generalizable patterns and operational learnings ship in this template.

What’s new

/ingest skill — Inbox cleanup discipline (operational fix)

Previously possible failure mode: writing the Raw Source but forgetting to delete the Inbox original → next /inbox scan re-ingests the same source → duplicate Raw Sources. Now explicit:

  • Step 2 renamed to “Save Raw Source (Move, not Copy)” with prominent warning callout
  • Inbox cleanup section added — when source originated from 00. Inbox/, the Inbox file MUST be deleted after pre-flight checklist passes (with source-origin matrix specifying when delete applies)
  • Step 7 verification — explicit Inbox cleanup check (ls "00. Inbox/{subfolder}/" should not show consumed file)
  • Failure modes section restructured: (1) summarization-instead-of-verbatim, (2) Inbox residue

Edge case documented: delete only AFTER all pre-flight checks pass, otherwise verbatim failure could lose the only copy.

New Concepts (4)

  • Cohort Token Economy — Predictable failure pattern when 80%+ of a cohort relies on Pro plan. Distinguishes cohort runs from individual hobbyist usage. Informs curriculum/onboarding design (token economy must be in week 1, not discovered in week 4).
  • External Pre-processing Pattern — “Claude 가 잘하는 것 만 Claude 에게 시킨다.” Heavy raw conversion (long PDFs, video transcripts, large web scrapes, foreign-language content) goes to GPT Deep Research / Google AI Studio / Whisper STT first; only the distilled output enters LLM Wiki via /ingest. Cohort estimates: 50-80% token savings.
  • Track Classification and Research Gap Detection — Custom command pattern for typology-driven domains. Splits ingested papers into thematic tracks during ingest, computes per-track coverage %, surfaces lowest-coverage track as research gap. Generalizable across systematic review · market analysis · content planning · learning path design. Includes /track skill proposal.
  • Idea Generation Pipeline — Each new raw source triggers automatic idea generation tied to a domain seed context. Distinguishes LLM Wiki from passive learning tools (NotebookLM-style) by treating each new input as a divergence prompt, not just a record. Includes /idea skill proposal.

New Guide (1)

  • LLM Wiki Token Optimization Strategies — 4-step practical guide (1: external preprocessing → 2: model selection → 3: /effort throttling → 4: infrastructure decisions). Designed as a curriculum companion when introducing LLM Wiki to a class, lab, or study cohort. Includes decision-flow diagram, week-by-week curriculum recommendation, and policy-change hedge advice (don’t lock cohorts into a single vendor).

Files changed

  • .claude/commands/ingest.md — Step 2 title + Inbox cleanup section + Step 7 verification + Failure modes restructure
  • 20. Wiki/21. Concepts/Cohort Token Economy.md — new
  • 20. Wiki/21. Concepts/External Pre-processing Pattern.md — new
  • 20. Wiki/21. Concepts/Track Classification and Research Gap Detection.md — new
  • 20. Wiki/21. Concepts/Idea Generation Pipeline.md — new
  • 20. Wiki/23. Guides/LLM Wiki Token Optimization Strategies.md — new
  • CHANGELOG.md — this entry

Migration notes

No breaking changes. Existing v1.1.0 vaults can adopt v1.2.0 by:

  1. Pulling the updated .claude/commands/ingest.md
  2. Copying the 4 new concept pages and 1 new guide page (or letting /ingest auto-generate them as wikilinks resolve)
  3. Reviewing Cohort Token Economy if running a cohort

v1.1.0 — 2026-04-20 (Book Ingest Pattern)

Addition: /ingest operation gains a new variant — Book Ingest Mode (Progressive Stubs) — for multi-page books and documentation sites.

What it solves

멀티 페이지 책 (mdBook, VitePress, GitBook, Docusaurus, ReadTheDocs, Nextra 등 TOC ≥5 챕터) 을 표준 /ingest 로 처리하면 두 가지 실패:

  1. 한 파일에 전부 → 가독성 붕괴, ## Original Content 비대화
  2. 전체 동시 컴파일 → 읽지도 않은 내용이 Wiki 에 심어져 contamination (AI 요약이 human-curated knowledge 자리 차지)

How it works

  • Scaffold: 1 Book Index Raw Source (verbatim preface + TOC with wikilinks) + N chapter stubs (status: stub, placeholder ## Original Content, navigable via chapterPrev/chapterNext)
  • Wiki seed: book entity + author entity + preface-anchor concept (≤3 페이지). 장별 wiki 는 미생성.
  • Promote on read: 사용자가 장을 읽을 때 해당 stub 파일에 /ingest 재호출 → URL 에서 verbatim fetch + ## Original Content 채움 + 장 특화 Wiki 컴파일 + status: stub → completed + Book Index Progress Tracking 업데이트

Karpathy “지식은 스크랩이 아니라 독서 시점에 컴파일” 원칙의 구체화. Progressive Disclosure Pattern 의 ingest-layer 대응.

New frontmatter keys (Raw Source, book stubs only)

  • status: stub | reading | completed (기존 ingested 와 병존)
  • bookIndex — 소속 Book Index wikilink
  • chapterNumber — 정수
  • chapterPart — 편/파트 이름 (원문 언어 보존)
  • chapterPrev / chapterNext — 이전·다음 챕터 wikilink (null 가능)

File naming

  • Book Index: YYYY-MM-DD-{authorSlug}-{bookSlug}-book-index.md
  • Chapter Stub: YYYY-MM-DD-{authorSlug}-{bookSlug}-ch{NN}-{slug}.md

Files changed

  • .claude/commands/ingest.md — “Book Ingest Mode (Multi-Page Sources)” 섹션 추가 (step B-1 ~ B-5 + promotion workflow + hook interaction + visual pattern)
  • CLAUDE.md — Operations §1 Variants callout + Raw Source frontmatter 에 status enum 확장 + Book Ingest 전용 4 키 + 파일 네이밍 테이블 2 행 추가
  • 20. Wiki/21. Concepts/Book Ingest Pattern.md — 신규 self-documenting concept 페이지 (architecture, lifecycle, 적합성 매트릭스, hook interaction)

Hook compatibility

validate-raw-source.sh 는 stubs 도 통과 — 모든 stub 이 ## Original Content 섹션을 placeholder 와 함께 보유. Hook 이 본문 substantiveness 검사로 업그레이드될 경우 status: stub 예외 처리 필요 (document in Book Ingest Pattern concept).

v1.0.0 — 2026-04-14 (Initial Template Release)

Source: Extracted and sanitized from Yohan Koo’s personal LLM Wiki satellite vault, active since 2026-04-10.

Architecture

  • 3-Layer: Raw Sources (immutable) / Wiki (LLM-managed) / Schema (CLAUDE.md)
  • Operations: Ingest · Query · Lint · Refresh-Context
  • Core files: index.md, log.md, CLAUDE.md, Core Context.md

Claude Code Harness

  • 7 slash commands in .claude/commands/:
    • /ingest — purpose-gated source ingestion with mothership cross-link search
    • /inbox — batch scan of 00. Inbox/ with single-axis / per-file / auto-infer modes
    • /query — wiki-grounded Q&A with 7-axis reuse tagging
    • /lint — health check (orphans, broken links, contradictions, v2 coverage, Core Context freshness)
    • /status — at-a-glance stats + coverage metrics
    • /reindex — manual qmd index rebuild
    • /refresh-context — re-snapshot Core Context when mothership drifts
  • 2 PostToolUse hooks in .claude/hooks/:
    • validate-raw-source.sh — enforces ## Original Content verbatim preservation
    • qmd-reindex.sh — debounced auto-reindex on Write/Edit

Schema Standards

  • YAML 2 spaces / Body TAB indentation
  • 7 required properties: type, aliases, description (English), author, date created, date modified, tags
  • Wikilinks in YAML must be quoted: "[[link]]"
  • Mermaid labels in quotes: A["label"]
  • camelCase for new YAML keys: collectionPurpose, mainVaultRelated, mainVaultCmds, reusableFor

미래의 나에게 보내는 편지 Policy (v2 frontmatter)

  • collectionPurpose — mandatory user answer to “왜 수집?” at ingest time
  • mainVaultRelated — 2~5 related notes from (optional) mothership vault, found via qmd
  • mainVaultCmds — mothership CMDS category (if applicable)
  • reusableFor — which of the 7 reuse axes a query answer feeds

Collection Infrastructure

  • 17 Obsidian Web Clipper templates in 90. Settings/Sharing/ (Stibee added in v1.4.0 → 18):
    • Articles (web, tech-blog, news, substack), Social (X, threads, linkedin, reddit, instagram, hackernews)
    • Video (youtube, podcast), Technical (github, arxiv, tech-docs, linkedin-pulse)
    • Selection clipper for ad-hoc highlights
  • qmd config template in 90. Settings/qmd-config-template.yml — BM25 + vector hybrid search

Example Content

  • 2 Raw Sources (Karpathy LLM Wiki gist + X thread)
  • 4 Concept wiki pages (LLM Wiki Pattern, RAG vs Compiled Wiki, 3-Layer Architecture, Ingest-Query-Lint Cycle)
  • 3 Entity pages (Karpathy, Vannevar Bush, Memex)
  • 1 Guide (Obsidian Tooling for LLM Wiki)
  • 2 MOCs (Knowledge Management, LLM Wiki Guide)

Notes

  • Template uses David, David, /c/Users/David/Desktop/Vaults/04_GuWiki, {PATH_TO_YOUR_MOTHERSHIP_VAULT}, {your-mothership-vault-name} placeholders per the cmds-system-files convention.
  • Mothership integration is optional — this wiki operates standalone or as satellite to any Obsidian PKM vault.

그래프 뷰