Release history

A larger, better-evidenced reference catalog and a more honest DESIGN.md Core v2. Patch release — no change to the installer's file ownership rules. 2.0.1 was never published to npm; this release includes everything listed under 2.0.1.

• Catalog: 521 references, 227 of them verified_v2 (npm 2.0.0 shipped the earlier catalog). New verified references span Japan, the UK and a first European push — Germany, France, the Netherlands, the Nordics, Spain, Italy, Austria, the Baltics — each measured live (computed tokens, component states under a real Tab key, pseudo-element and background-image hover layers) and grounded in the company's own legal and about pages. Country codes LT, NO, DK, AT, EE, CZ, RO, BE and CH are now accepted.
• Core v2 omitted claim — a Core file can now say what it left out and why (verified-absent, unresolved, out-of-scope), so an agent can tell "the brand publishes nothing" from "nobody looked".
• Core v2 conflict consumption — a reader meeting a preserved conflict follows surface match → declared authority → freshness → do not choose, instead of silently picking a value.
• Google-shaped input detection in the Core dual-read path.
• omd-reference-capture — the shipped capture bundle now reads the Core package instead of raw DESIGN.md bytes, and fails loudly when a Core body has no .omd/ package.
• Counts in the READMEs (all four locales), package description, MCP README and llms.txt now read "500+" references and 28 skills / 27 Cursor Agent Skills.
• Release gate waiver — 2.0.2 ships on an explicit owner waiver (decision D4, 2026-09-29). The skill-output quality gate (npm run gate:quality:strict) moved from prepublishOnly to a local pre-tag step, because CI cannot see its gitignored fixture outputs, and it is not green: landing/stripe, landing/toss and landing/karrot fail landing integrity (recorded as the D4 exclusion), autopilot/stripe and autopilot/toss are unverified (no generation receipt), and autopilot/karrot is review-pending. Receipts will be regenerated before the next release.

Landing craft as a shipped workflow: five new skills, an art-director role, and the machine checks that keep a one-page landing honest. Patch release — no change to the DESIGN.md Core v2 contract or the installer's file ownership rules.

New skills (the bundle is now 27 product skills for Claude Code, Codex, and OpenCode, and 26 compatible Agent Skills for Cursor):

• omd:landing — scroll-native one-page landing derived from DESIGN.md plus the measured landing-craft codex (LC-n rules): concept, per-section storyboard, asset plan, standalone render.html, and a capped self-review loop. Triggers on "랜딩 페이지", "원페이지", "landing page", "one-pager", "hero to footer".
• omd:setup — finds the image/video generation channels, browser, and encoders this machine already has, explains the trade-offs and cost of each, and records the choice in .omd/config.json for later design tasks. It never tells you to go get an API key. Triggers on "omd setup", "도구 셋업", "configure media", "what can you generate".
• omd:media — brand-consistent asset sets (hero, section illustrations, icon sets, OG image) through whichever channel omd:setup recorded, with a provenance and cost ledger. No channel available means a prompt pack and a manual queue, never stock imagery. Triggers on "이미지 세트 만들어", "generate assets", "hero image", "OG 이미지".
• omd:showcase — turns a rendered page, or several side by side, into a deterministic scroll demo video (frame capture → ffmpeg H.264/GIF, one label per arm). Triggers on "시연 영상", "스크롤 영상 만들어", "showcase", "demo video", "비교 영상".
• omd:issue (added 2026-09-01) — files usage complaints and dogfood defects as structured GitHub issues, and drains the queue with a batch process verb. Triggers on "이슈 등록해줘", "버그 리포트", "file an issue", "이슈 처리해줘".

Also in this release:

• omd-art-director sub-agent (20 specialist roles, up from 19) — turns DESIGN.md brand facts plus a brief into one visual concept, mood, asset direction, and a per-section composition storyboard grounded in the LC-n rules. It writes concept.md and storyboard.md only: never tokens, never HTML.
• Landing craft codex + machine checks now ship with the package. docs/design-excellence/landing-craft-codex.md (the LC-n rule table), test-v2/tools/landing-integrity.mjs (LI-n machine verdicts on page height, section ratios, fold coverage, type ratio, gutter, motion duration and easing, scroll snap, video attributes, stock hosts, uniform card grids, h1), test-v2/tools/render-integrity.mjs, test-v2/tools/showcase.mjs, and scripts/omd-setup-detect.mjs are in the npm files allowlist so the new skills resolve their tools in an installed project instead of pointing at repo-only paths.
• Contrast is part of "overwhelming". omd:landing now requires a token-colored scrim behind text over imagery and forbids muted tokens for body copy, after a first run shipped a hero with nav at 1.05:1. Scroll reveals must stay visible under <noscript> and @media (scripting: none).
• /design.md public convention — this project's own root DESIGN.md is served at oh-my-design.kr/design.md, following the vercel.com/design.md convention, and scripts/check-mirror-drift.mjs now guards that pair alongside the reference trees.
• Feedback loop — issues #75–#87 filed through omd:issue from real runs of the new skills, covering unattended mode, self-feedback wiring, UA-default reset, per-project brand CSS, and a pixel-sampled contrast check for landing-integrity.

Repo tooling (not shipped to npm):

• scripts/wave-run.mjs — per-brand migration pipeline runner with a shared concurrency budget, so worker, gate, measurement, review, and revision for one brand can never overlap and prompts are assembled only from prompts/ files.
• scripts/wave-marathon.mjs — unattended multi-wave driver on top of the runner: queue, resume-after-block, deferral ledger, close-out, and journal line per wave, with a STOP file honored at wave boundaries.
• test-v2/tools/transcribe-verify.mjs — two conformance defects fixed. The §7.1.8 blinding detectors missed the real decision-ID shape (D-P2-1) and the hyphenated arm name, and §7.1.6 byte-compared fenced blocks in a way that made the deletions §4.1 mandates impossible to declare. Three self-tests added (10/10); the 23 existing transcripts were re-verified (20 pass, 3 blocked and re-transcribed).

Design-excellence harness: a design system the model derives from a declared philosophy, floored by a validated preset catalog — measured against rendered screens, not source code. The 2.0 promise: visibly better product than prompt-only or component-kit assembly, portable via DESIGN.md.

• Philosophy derivation chain (omd:autopilot): PHILOSOPHY (stance + principles with explicit sacrifices) → decision table (D-ids with rationales) → tokens with D-id back-references → component specs → per-page layout grammar with content back-calculation → build → render critique → DESIGN.md that answers "why this value". Improvised token values are a gate failure (GS7).
• Preset catalog — component and layout contracts in four layers, browsable at oh-my-design.kr/presets. Fundamentals (painted-surface padding anatomy, route focus, brand mark + masthead, hero, footer, custom APG listbox, empty states). Primitives — the component vocabulary mapped to real shadcn/ui components and their Radix primitives, each with the W3C ARIA APG pattern, state matrix, and token slots: form, overlay, navigation, display. Genres — commerce, marketplace, editorial screen assembly. Flavors — tendency profiles derived from the reference catalog. A preset is not a component you install: shadcn gives you the structure, the preset tells the agent which values must come from your decision table instead of a library default. The derivation chain now selects presets and derives token slots instead of improvising from zero (new gate GS8). Measured on the same fixture and brief: first-render defects 7 → 3 (zero P0/P1), input tokens −47%.
• omd book — browse the adopted design system on a local port (npx oh-my-design-cli@latest book, or --static for a standalone handoff file). Storybook renders stories; this renders the contract: every token beside the decision that produced it, component state matrices including the states that deliberately do not apply and why, contrast measured live against the pairs the system promised, and preset lineage. Reads the compiled graph when present and falls back to parsing DESIGN.md.
• Lossless slop-gate port — numbered gates G1–G56 + system-fidelity gates GS1–GS8, including Korean typography (gothic stacks, keep-all, no serif fallback), :focus-visible-only rings with programmatic-heading-focus suppression, custom listbox over native selects, disabled states via tokens (never opacity alone), dark-surface token inversion, and layout-grammar minimum-content rules. Mechanically checkable gates must be grep-verified in the self-critique.
• Component craft floor — 45 production norms distilled and independently re-verified from HIG, Material 3, Carbon, Polaris, Geist, Radix/shadcn, ARIA APG, and WCAG 2.2, covering buttons, inputs, selects, cards, chips, layout wells, and dark/contrast pairs.
• Render-feedback pass — the harness renders real screenshots, an external critique produces a numbered punch list, and the model repairs surgically inside the existing system; self-critique is performed against screens, not source. Release judgment runs on a triple gate (deterministic audit → render critique → human/delegated verdict) recorded in docs/design-excellence/.
• Screen-inventory gate (omd:harness checkpoint #1): the master digs the complete screen list — including screens the user did not name — into a per-screen table (purpose, key elements, layout grammar, states, preset candidates) that the user approves before any wireframe; unapproved screens are never built.
• Verified end-to-end on three Korean-market cases (home-interior commerce, secondhand marketplace, fashion editorial) with enriched fixtures (real datasets + 228 generated images), each shipped as an npm run dev-ready React app with its DESIGN.md.

Also in this release (1.9.1 candidate line): natural-language workflow routing, one implementation owner, and same-surface verification for UI delivery.

• npx oh-my-design-cli@latest update now refreshes an existing installation without changing project/global scope, installed channels, Cursor rule-only mode, optional policies, or user-owned files. It refuses fresh-install and unsafe/manual-review cases, never implies --force, and runs a post-update doctor check.
• The new omd:update skill lets users ask their coding agent to perform that same fail-closed refresh. The shipped bundle is now 22 product skills for Claude Code, Codex, and OpenCode, and 21 compatible Agent Skills for Cursor.
• omd workflows [task] (alias omd route) now turns a rough product request into the smallest capable OmD workflow and a ready-to-paste coding-agent prompt. The command explicitly separates terminal steps (install and doctor) from agent-chat work.
• omd:apply, omd:harness, and omd:orchestrator now share one work-packet contract. Specialists return evidence and advice to a single implementation owner; completion requires the real consumer route, viewport, and state to be reverified after the product edit.
• Full-harness runs now bridge approved artifacts back into the product after checkpoint #3 instead of treating an archived run directory as delivery. The package ships portable OpenAI skill metadata for the three core workflows.
• The 1.9.1 benchmark kernel records dirty-source attribution, product-only diffs, task-owned font oracles, and normalized run records so skill, harness, and model results cannot be promoted from contaminated or non-product runs.

CLI re-onboarding: one guided entrypoint, an exact repair doctor, safer managed upgrades, and outcome-first documentation in five languages.

• Bare npx oh-my-design-cli@latest now opens the guided installer. doctor checks the complete 18-skill product bundle, 16 specialist roles, channel-local catalog, hooks, fingerprint coverage, and DESIGN.md, then prints a repair command scoped only to the broken channels.
• Claude Code, Codex, and OpenCode receive channel-native skills, supported specialist roles, and a local catalog; Cursor receives its project rule shim plus the shared catalog. The installer and docs no longer imply identical host capabilities.
• Managed reference sync now records content hashes. It refreshes only unchanged installer-owned DESIGN.md files, preserves same-ID user edits and private references, and removes retired managed files without deleting user sidecars in their directories. Existing non-OmD hooks remain untouched. Reinstalling Codex also removes OmD-owned legacy .codex/skills entrypoints automatically while preserving private skills.
• doctor now validates exact Claude hook event/command mappings, distinguishes Cursor's shared catalog from a Claude installation, detects stale or locally modified hooks, and can execute a scoped --repair-hooks recovery. Unmarked OmD-reserved conflicts use an explicit --force; symlinked managed paths and unmanaged legacy Codex entries fail closed for manual review. Installer drift exits non-zero instead of reporting a false success.
• Generated Codex/OpenCode roles now carry the correct project/global offline-catalog root. OpenCode role bodies use native hyphenated skill ids, while Cursor's rule embeds a small evidence-safe DESIGN.md bootstrap/apply contract without pretending named skills or sub-agents were installed.
• The package now ships docs/CLI_QUICKSTART.md. The website adds 40 static CLI docs routes across English, Korean, Japanese, Simplified Chinese, and Traditional Chinese (Taiwan), including an outcome-first live-demo runbook, plus localized human and agent-readable llms guides.
• Tag publishing is gated by CLI lint/tests/counts plus web tests/typecheck/build, and tarball hygiene now fails closed when it cannot inspect the package.
• MCP is no longer part of the recommended runtime. Installation remains a local markdown/data copy with zero OmD AI API calls; the user’s existing coding-agent subscription remains the inference layer.

44 new Korean company references added (catalog 356 → 400; KR refs 149 → 193) via the full 추천→선발→수집→검수 pipeline — every ref Tier-1 brand-owned proof-gated. Milestone: catalog hits 400.

• Coverage: AI 10 (asleep·scatterlab·friendliai·pozalabs·selectstar·squeezebits·corca·gaudiolab·maum-ai·sionic), consumer-tech 6 (heydealer·barogo·frip·nrise·stayfolio·tabling), fintech 6 (payhere·8percent·moin·peoplefund·quotabook·qraft), healthcare 5 (fitpet·tellingme·drdiary·humanscape·medibloc), ecommerce 4 (petfriends·idus·cjonstyle·dealicious), marketing 4 (dable·beusable·igaworks·wisetracker), education 3 (likelion·mildang·teamsparta), design-tools 2 (protopie·sandoll), automotive 2 (autopedia·42dot), productivity 1 (typed), saas 1 (jobplanet).
• Standouts carry named / official design systems or distinctive product UIs: likelion (Storybook DS), protopie (Studio XID), sandoll (type foundry), asleep / scatterlab (consumer AI apps).
• Recommend fanned 9 sector deep-research agents (64-item clean pool, deduped vs the full 356-id catalog); an LLM-synthesis drift that silently re-introduced 12 catalog-existing ids + 1 same-company alias was caught at curation and hard-re-deduped against on-disk references/. User excluded 6 母-brand sub-products (tosspayments/kakaopaysec/kakaowebtoon/brunch/kakaowork/watchapedia) + tistory. Build ran one CREATE subagent per brand in demand-driven waves of 5, stopping at 44 green.
• urbanbase dropped as INFEASIBLE (brand defunct — domain now a GoDaddy parking lander; agent correctly refused to fabricate). peoplefund captured through its 크플(Cple) rebrand redirect.
• All 44 pass verify-reference 27/27 + catalog-integrity (16 files / 638 tests). KR proof gate met for every ref (≥2 brand-owned Tier-1 sources + ## Proof block). Tier 2 (getdesign / refero) empty for all 44, as spec/regional-sources.yaml anticipates.
• Audit: data/reference-audits/2026-07-02-kr44.md (13 High / 29 Medium / 2 Low). Counts propagated by sync-catalog.mjs (README, llms.txt, SEO layouts ×3, fingerprints ×3, design-md mirror).

30 new Korean company references added (catalog 326 → 356; KR refs 119 → 149) via the full 탐색→선발→수집→검수→배포 pipeline — every ref Tier-1 brand-owned proof-gated.

• Coverage deliberately steered toward under-represented categories: AI 7 (nota·rebellions·furiosaai·returnzero·makinarocks·supertone·neosapience), marketing 4 (airbridge·buzzvil·datarize·bigin), education 3 (goorm·elice·codeit), SaaS 3 (saramin·lemonbase·shiftee), ecommerce 3 (kyobobook·cafe24·queenit), developer-tools 2 (hackle·solapi), backend-devops 2 (lablup·moreh), consumer-tech 2 (hwahae·weverse), fintech 2 (portone·kcd), automotive 1 (greencar), healthcare 1 (vuno).
• Standouts carry named / official design systems: goorm (Vapor UI), 교보문고 (KDS), 화해 (HDS), cafe24 (Smart Design), lablup (Backend.AI WebUI).
• Discovery fanned 8 sector agents (101 raw → 89 unique, deduped vs the full 326-id catalog); two LLM-drift duplicates caught at curation (wadiz, yeogi==yeogiotte) and swapped for clean alternates. Build ran one CREATE subagent per brand in waves of 5 (Tier-1 live Playwright getComputedStyle over ≥2 brand-owned surfaces each).
• All 30 pass verify-reference 27/27 + catalog-integrity (16 files / 594 tests). KR proof gate met for every ref: ≥2 brand-owned Tier-1 sources + ## Proof block (≥80 raw computed-style samples). Tier 2 (getdesign / refero) empty for all 30, as spec/regional-sources.yaml anticipates.
• Audit: data/reference-audits/2026-06-26-kr30.md (23 High / 7 Medium / 0 Low). Counts propagated by sync-catalog.mjs (README ×4, llms.txt, SEO layouts, fingerprints ×3, design-md mirror).

New skill omd:feel — quantifies the design / frontend industry's "gut-feel" into 113 provenance-graded, machine-checkable interface rules an AI can APPLY (inject motion / spacing / type / a11y defaults during UI work) and AUDIT (BLOCK / WARN / FYI). Inherits Jakub Krehel's make-interfaces-feel-better and extends it with Apple HIG / Material 3 / WCAG / design-system tokens + practitioner research.

• 17 feel-dimensions, 113 rules. 85 are spec / design-system-token-backed (enforceable); the rest are labelled convention / opinion so folklore is never presented as spec. Core idea: provenance tier → enforcement strength (SPEC/DS = gate, CONV = default, OP/FOLK = suggestion). DESIGN.md tokens always win. Files: skills/omd-feel/{SKILL.md, reference.md, provenance.md} — auto-installs via install-skills. Skill bundle 17 → 18.
• Dogfooded on this site's home + docs: fixed a real prefers-reduced-motion a11y gap (framer-motion + animate-ping + smooth-scroll were ungated) via a global CSS net + <MotionConfig reducedMotion="user">, swept transition: all → named props on the landing / docs surfaces, and added text-wrap: balance / pretty. tsc + 553 tests green.
• Docs / README: reference count corrected 286 → 326 across localized READMEs + site metadata (catalog grew in 1.8.3–1.8.4; localized surfaces had drifted).

QA pass on the 2026-06-22 batch: 40 refs re-inspected vs live + benchmarked against the designlang deterministic extractor. 35/40 PASS, mean accuracy 98.3/100, 5 MINOR token-value fixes applied.

• Two multi-agent workflows: per-ref live re-inspection + adversarial verify (QA), and our-tokens-vs-designlang-vs-reality comparison. Zero major errors, zero fabrications; brand primaries/fonts all held. designlang independently corroborated ~5.4/6 of our key tokens per ref.
• 5 confirmed MINOR corrections (structural token values, not brand errors): paypal card shadow 0.1/4px/12px → 0.08/24px/48px (live layered-card); chunghwa radius scale +20px (dominant modal/card default, ~46%); hyundai chatbot FAB box-shadow rgba(0,0,0,0.15) 0 0 20px added (CREATE sampled the wrapper, not the button); snapchat §6 "none across nav" softened — components flat but marketing nav wrapper carries one shadow; octopusenergy postcode input split into hero (white/Arial-600) vs tariffs (transparent/Chromatophore) surfaces. All 5 verify-reference green; 553 tests pass.
• Benchmark write-up: data/reference-audits/2026-06-23-qa-designlang-benchmark.md. Finding: our LLM pipeline is 39–0 (1 tie) more brand-accurate than designlang (which picks most-frequent color as "primary", ~35% brand-correct); recommendation = fold designlang in as a free deterministic Stage 0 (palette histograms + DTCG) feeding our brand-judgment pass.

Catalog +30: TW·US·UK 메가브랜드 — 카탈로그가 296 → 326개로. 4개국 확장 라운드 2의 배치 2~4/4 (KR은 1.8.3). Phase 4 프로모는 이 라운드에서 생략.

• TW 10 (Taiwan-HQ 한정) — E.SUN Bank(玉山銀行) · Fubon(富邦) · iPASS MONEY(一卡通) · EasyWallet(悠遊付) · OPENPOINT(統一/7-11) · Bahamut(巴哈姆特) · iCook(愛料理) · Taiwan Mobile(台灣大哥大) · Chunghwa Telecom(中華電信) · 591(591房屋交易). 핀테크 ≤5 도메인 균형.
• US 10 — PayPal · Zoom · Asana · Squarespace · Reddit(RPL) · Snapchat · DoorDash · Instacart · Databricks(brand portal) · HubSpot(Canvas). 4종은 공개 DS/brand 포털 보유로 토큰 grounding 최상.
• UK 10 (registry enum=UK) — GOV.UK(Design System) · BBC(GEL) · Monzo · Starling · Deliveroo · Octopus Energy · ASOS · Skyscanner(Backpack) · Trainline · Farfetch. UK 보유 2 → 12개로. GOV.UK·BBC·Backpack 등 공개 DS 골드스탠다드 다수.
• 전 30종 verify-reference 게이트 그린(US/UK는 26, KR/TW 규칙 비적용) + token↔prose 정합. TW는 brand-owned Tier-1 ≥2 충족. 큐레이션은 HQ-국가 규칙으로 Shopify(CA)·Shopee(SG)·Atlassian(AU)·Depop(US-owned) 등 제외.
• getdesign.md/refero: KR/TW 0/20(empty 지속), US/UK는 7종 등재 — 영미권 공개 DS 가시성이 동아시아 대비 확연.
• 배치별 audit data/reference-audits/2026-06-22-{tw,us,uk}10.md.

Catalog +10: KR 메가브랜드 — 카탈로그가 286 → 296개로. 4개국 확장 라운드 2의 배치 1/4 (KR→TW→US→UK).

• KR 10 — Samsung(삼성전자, One UI) · Hyundai(현대자동차) · Kia(기아) · Naver Pay(네이버페이) · KB Pay(KB페이) · Hana Bank(하나은행) · Shinhan Bank(신한은행) · SSG.COM(쓱닷컴) · KakaoPage(카카오페이지) · Liner(라이너). 실제 브랜드 인지도/사용량 상위 기준으로 큐레이션.
• 자동차 2종은 모노크롬·0px radius·무그림자(차량 사진이 깊이 담당), 금융 4종은 로그인 게이트 뒤 네이티브 앱 — 공개 웹/개발자 디자인가이드 표면에서 정직하게 토큰 추출(naverpay/kbpay는 Medium-High, in-app 상태는 UPDATE 대상).
• getdesign.md/refero는 KR 0/10(여전히 empty) — KR brand-owned Tier-1 소스 ≥2로 전 항목 충족. verify-reference 27/27 게이트 전원 그린, 523 테스트 통과.
• 배치 audit data/reference-audits/2026-06-22-kr10.md.

Catalog +40: KR/US/TW/JP IT 확장 — 카탈로그가 246 → 286개로. 4개국 × 10개, 전 항목 Proof Gate + live-inspect.

• KR 10 — NHN(Kenya Hara 무채색 CI) · Wrtn(뤼튼) · TMAP모빌리티 · NCSOFT(→NC) · Spoqa(스포카, Spoqa Han Sans) · Stibee(스티비) · ZEPETO(제페토) · Shift Up(시프트업) · IICOMBINED(Gentle Monster·Tamburins·Nudake) · The Pinkfong Company.
• US 10 — Twilio(Paste) · Workday(Canvas) · Cash App · Mailchimp · GitLab(Pajamas) · Plaid(Threads) · Headspace · Datadog · Twitch · Patreon. 공개 디자인시스템 헤비웨이트 다수.
• TW 10 — Fourdesire(SPARKFUL) · Bito · Cubo Ai · MOZE · Vocus(方格子) · Portaly · Readmoo(讀墨) · XREX · MaiCoin/MAX · Heptabase(operating-base 기준 TW 분류).
• JP 10 — Spindle(CyberAgent Ameba) · 디지털청 DADS(CC BY 4.0) · DMM Turtle · GMO Pepabo Inhouse · Gaudiy · SPEEDA(FALCON) · Cybozu(kintone DS) · Goodpatch · Nintendo · mikan. DS 강국답게 공식 디자인시스템 직접 추출.
• 전 항목 verify-reference 게이트 그린 + token↔prose 정합, KR/TW는 brand-owned Tier-1 소스 ≥2 충족. 후보는 4개국 어드버서리얼 검증(HQ 국가 + 디자인 가치)으로 큐레이션 — Typed(서비스 종료) · iCook/BlendVision(모회사 HQ 일본) · Disquiet/PopDaily(일반적 UI) 등 탈락.
• 배치별 audit data/reference-audits/2026-06-17-{kr,us,tw,jp}10.md + 큐레이션 문서 research/2026-06-17_reference-candidates-kr-us-tw-jp.md.

Catalog +25: KR/JP/TW/US 배치 — 카탈로그가 221 → 246개로. 수집→검수→배포 파이프라인 스크립트화.

• KR 5 — 3o3(삼쩜삼) · Blind(블라인드) · MiriCanvas(미리캔버스) · Imweb(아임웹) · Modusign(모두싸인). 미리캔버스가 KR design-tools 공백을 채움.
• JP 5 — Rakuten(ReX 토큰을 공개 npm tarball에서 직접 추출) · Wantedly(공식 Brand Guide PDF 기반) · ABEMA(다크 스트리밍 시스템) · STORES · Timee.
• TW 5 — Accupass · FunNow · Firstory · Omnichat · Greenvines(綠藤生機).
• US 5 — Google(Material 3) · Microsoft(Fluent 2) · Adobe(Spectrum 2) · Salesforce(SLDS 2 styling hooks) · Dropbox(brand.dropbox.com). 공식 DS 헤비웨이트 5종이 처음으로 카탈로그에 합류.
• 전 레퍼런스 Proof Gate(라이브 inspect raw sample ≥5) + token↔prose 정합 게이트 통과, KR/TW는 brand-owned Tier-1 소스 ≥2 충족.
• KR 5 (2차) — Ringle(링글) · POSTYPE(포스타입) · LaundryGo(런드리고) · Greeting(그리팅) · Danawa(다나와). 개선된 게이트 파이프라인으로 추가 — 5/5 첫 실행에 전 게이트 그린.
• 새 파이프라인 스크립트 — web/scripts/verify-reference.mjs(레퍼런스 단독 게이트: catalog-integrity per-ref 항목 + strict-YAML + 로고 liveness, registry 빌드 불필요) + web/scripts/sync-catalog.mjs(fingerprints ×3 / design-md mirror / 카운트 문자열 / llms.txt Examples / npm test 일괄). omd:add-reference 스킬에 batch 서브에이전트 프로토콜 명문화.

The taste loop closes: your corrections become preferences, preferences become proposals, and one click folds them into DESIGN.md. Plus a taste dashboard and a scripted release bench.

• Auto-fold gate. When a preference scope recurs past the threshold, session-end-foldin writes a structured proposal and your next session opens by asking — one selectable-option question ("이 취향, DESIGN.md에 반영할까요?"). Approve → omd:learn folds it; decline → snoozed until it recurs again. The old remember → wait → manually-run-learn dance is gone.
• Ambient capture now persists — and sees more. post-edit-watch no longer just warns: off-system drift is recorded to .omd/preferences.md as confidence: inferred (24h dedup, never while editing DESIGN.md/.omd). Detection grew from hex colors to radius and motion-duration — including Tailwind classes, kebab-case CSS, camelCase JSX inline styles, and .html outputs (what omd:harness actually emits).
• omd:taste — see what the system thinks you like. One command renders applied / pending (with recurrence counts and how far from the fold threshold) / snoozed / not-yet-known axes, with inline actions (fold now · forget · refine). Skill count: 16 → 17.
• Reviewers feed the loop. omd:designer-review and omd:final-qa now end recurring findings with a single opt-in question — "이 패턴, 취향으로 기록할까요?" — recording with signal: review on consent. Never auto-recorded, never auto-folded.
• Scripted 2-bench release routine. scripts/bench/score-init.mjs (first-60s) and the new scripts/bench/score-taste.mjs (capture → reflect → propose) score real headless runs deterministically; both at 10/10 on this release.
• Fixes & hygiene: session-state-loader output envelope corrected (its session-start context was silently dropped); catalog resolution order normalized across all 4 documents with a CI drift guard; installer channel mapping unified into one table (cursor dedup behavior preserved, tested); omd:kr-writer presets now self-contained for npx installs; dangling dev-only references cleaned; docs/JSON-LD FAQ generated from one source; release workflow actions bumped for Node 24 runners.

First-run fixes for npx installs (skill triggering, local catalog, dead hooks), a first-class Cursor install channel, and a safer omd:init.

• Installed-skill marker moved below frontmatter. The <!-- omd:installed-skill --> marker was written above the YAML frontmatter, which broke frontmatter parsing in some hosts and stopped skills from triggering on natural language. Managed files are rewritten with the marker after the frontmatter block; re-run install-skills to pick it up.
• Reference catalog + ctx-prime install locally. install-skills now copies the reference catalog and ctx-prime data into .claude/data/, so omd:init works on the very first run of an npx install — no more empty-catalog dead end. As a safety net, omd:init also falls back to the website's raw .md endpoints when the local catalog can't be resolved.
• Cursor install channel. install-skills --agent cursor (also offered in the channel multiselect, auto-detected when .cursor/ exists) writes the .cursor/rules/omd-design.mdc shim — identical to the omd:sync template, body-hash marker included — plus the shared .claude/data catalog. No skills, hooks, or settings are written for Cursor.
• Hooks revived on clean installs. post-edit-watch, foldin, state-loader, and skill-activation shipped in states that left them dead on a fresh install; all four now register and fire correctly.
• omd:init apply-mode confirmation. Before writing, omd:init now distinguishes root bootstrap vs reference-only mode and asks before replacing an existing DESIGN.md — on Claude Code via the selectable-option (AskUserQuestion) UI.
• Builder → agent prompt handoff. The builder preview's install CTA composes your first prompt from the live config (reference, components, token overrides vs reference defaults, dark mode, style preferences) instead of a generic one-liner, so what you customized in the builder is what omd:init bootstraps.

Site-only (no npm impact, deploys with the site): raw DESIGN.md twin endpoints at /design-systems/<id>.md (agents can fetch references over HTTP), 8 curated collection pages at /collections/[slug], builder floating install CTA + neutral logo nameplates, GA4 install-funnel events (install_copy / prompt_copy), The Wall card-expand interaction removed (static tiles, hover kept), and a four-README restructure around the install→activate path.

Codex / OpenCode skills now install to their real discovery paths — and claude-design runs on all three channels.

• Correct channel paths. install-skills installed Codex skills to .codex/skills/ and OpenCode skills to a flat .opencode/agents/<name>.md. Neither matches the agents' actual discovery rules, so those skills were effectively invisible. Now:
  • Codex → .agents/skills/<name>/SKILL.md (official path per developers.openai.com/codex/skills).
  • OpenCode → .opencode/skills/<name>/SKILL.md (folder skills per opencode.ai/docs/skills).
  • Both are folder layouts, so multi-file skills install with their scripts/ and references/ intact.
• claude-design is no longer Claude-Code-only. Its primary backend is Playwright (a node-only, agent-agnostic Chrome driver), so it works from Codex and OpenCode too. x-omd-channels now lists all three. The claude-in-chrome path remains a Claude-Code-only fallback.
• Note: if you previously installed for Codex/OpenCode, the old .codex/skills/ and .opencode/agents/ files are now stale — re-run install-skills and delete the old copies. OpenCode also reads ~/.claude/skills/, so a --global install already surfaces skills there.

Project / global install scope for install-skills.

• Install scope: project or global — install-skills now asks where to install (or pass --global): Project (./.claude/skills, the default) or Global (~/.claude/skills, available in every project). Global writes skills + sub-agents (+ data) to the user-level dir and deliberately leaves your global hooks/settings.json untouched — ideal for using claude-design everywhere without per-project installs.

New bundled skill claude-design (terminal → claude.ai/design), single-skill install via --skills-only, and channel-restricted skills.

• claude-design skill (Claude Code only) — analyzes your codebase (stack, tokens, components, real UI copy, brand assets) and drives claude.ai/design end-to-end to generate a code-grounded design, returning a shareable link. If claude.ai/design interjects a "Quick questions" clarifying panel before generating, the skill detects it (it appears late, in a closed Questions tab, in varying layouts) and — by default — the agent reads the questions and picks the appropriate option per your codebase, with a "Decide for me" autonomous fallback for headless runs. Install standalone: npx oh-my-design-cli install-skills --skills claude-design --agent claude-code --skills-only.
• --skills-only install flag — install a single skill with no agents / hooks / omd onboarding (clean standalone install).
• x-omd-channels skill frontmatter — a skill can declare its compatible agent channels; incompatible targets are skipped (claude-design is claude-code-only — needs Chrome automation + python3 + a global playwright).
• Multi-file skills — install-skills now copies a skill's full tree (scripts/, references/), not just SKILL.md.

Site-only (no npm impact, deploys with the site): builder UX (registry-wide curation, scroll-to-top, mobile DESIGN.md toggle), landing mobile polish (Live Proof carousel, The Wall show-more, 2-col footer), and footer version sync to package.json.

Catalog grows to 150 brands (+13), the bundled DESIGN.md mirror is fixed, and the reference pipeline gains deterministic proof/format gates.

• 13 new references through a new proof-gated pipeline — Sendbird (US), KR: velog · NHN Cloud · Hyundai Card · Melon · Bithumb, TW: Rayark · 91APP · Kdan · Hahow · Gogolook · KKBOX · Cake. Every spec value is live-DOM-inspected or pulled from an open-source token file; no fabricated values.
• Fixed the bundled design-md/ mirror — 110 DESIGN.md files carried stale pre-migration frontmatter/bodies (one shipped a broken ## 1. Overview header). Re-synced to the canonical references.
• Backfilled Do's-and-Don'ts for 21 references that previously rendered an empty Guidelines section.
• New deterministic catalog gates — proof block for verified refs, §4 placeholder-value lint, §7 guideline-marker advisory, §1 blank-line guard; plus spec/regional-sources.yaml for KR/TW source coverage.

Conversational deepening — harness skill now reads your codebase before asking, then runs a single-batch picker interview matching the Ouroboros/donguri feel — zero external plugin dependencies.

2026-05-29 post-release AEO addendum (no version bump — npm bundle unchanged):

Marketing/docs site (oh-my-design.kr) and LLM-facing content deepened so LLMs cite oh-my-design when asked about design.md / vibe coding / design systems for AI coding agents. CLI behavior unchanged; no npm publish needed.

Added — LLM-facing content:

• web/public/llms-full.txt — full concatenated context (README + entire CHANGELOG + primary SKILL.md files + omd-master agent + spec/omd-v0.1 + 100+ brand catalog one-liners). 173 KB / 2,653 lines. Pair with the curated TOC at /llms.txt. Generated by scripts/gen-llms-full.cjs; regenerated automatically on web dev / build / publish.
• web/public/llms.ko.txt — Korean mirror of llms.txt for KR-first LLM queries with inline Korean mini-FAQ.
• scripts/gen-llms-full.cjs — deterministic concatenation helper. Reads canonical sources, never invents content.

Added — Next.js routes (5 new, 24 dynamic):

• /faq — 18 Q&A in 4 sections; FAQPage JSON-LD with 18 Question/Answer entities.
• /what-is-design-md — definitional page claiming the term; DefinedTerm + Article JSON-LD.
• /alternatives/[slug] for shadcn, v0, anima, locofy — 7-row comparison + when-to-use; Article + 2× SoftwareApplication JSON-LD.
• /changelog + /changelog/[version] — 24 SSG pages parsed from this file via web/src/lib/changelog.ts; per-version Article JSON-LD with datePublished.

Changed — AEO hygiene:

• web/src/app/layout.tsx — softwareVersion in schema.org now imports from package.json (was hardcoded "1.5.0"). Three v1.6 entries appended to featureList.
• web/public/llms.txt — new section describing CTX-PRIME, Interview-lite, sub-agent auto-recovery, and the natural-language auto-trigger phrasings.
• web/src/app/sitemap.ts — 7 new static routes + 24 dynamic changelog routes.
• web/package.json — dev/build regenerate llms-full.txt first so the deployed copy never drifts from canonical sources.

Closes the gap where natural-language design requests ("그럴싸한 랜딩 만들어줘", "프로토타입이라도 구색 갖춰서") fell through omd-harness and landed in plain-coding mode. Five entry-side defects fixed in one release. Full RFC: research/harness-design/20-harness-1.6-conversational-deepening.md.

Added

• scripts/ctx-prime.cjs — deterministic, sub-50ms codebase analyzer. Emits <RUN_DIR>/ctx-prime.json with stack detection, dominant brand color, font families, voice keywords, language, surface inventory, audience hypothesis (with confidence + evidence), and wow-moment candidates. Pure node, no deps.
• omd-harness Step 2.5 — CTX-PRIME phase. Runs ctx-prime.cjs before reference selection, surfaces analysis as the opening move ("이 레포 분석했어요 — Next.js + #0064FF 베이스 + 11개 surface"), and asks one AskUserQuestion picker for audience confirmation with confidence-ranked options + free-text "Other".
• omd-harness Interview-lite — a single batched AskUserQuestion covering exit_scope / wow_moment / cta_primary / visual_grounding. Options derived from ctx-prime.json (e.g., wow-moment options use detected wow_moment_candidates). Answers written to <RUN_DIR>/handoff/.handoff.json as prefilled_slots.
• omd-master INTAKE — prefilled-slots fast path. When .handoff.json has prefilled_slots, skip SLOT_GATE entirely and jump straight to PROPOSE_PLAN. Never re-asks audience / wow_moment that the harness already collected.

Changed

• omd-harness trigger description broadened — now matches "랜딩페이지", "프로토타입", "그럴싸한", "구색", "first screen", "MVP UI", "landing", "prototype", etc. Generic natural-language requests now activate the harness reliably.
• omd-harness Step 1 — subagent registration auto-recovery. Was a hard gate that stopped on missing omd-master; now copies the agent file from npm root -g/local to .claude/agents/ (or detected agent flavor) and continues. Restart-Claude-Code message demoted to last-resort fallback.

Why

Real-session failure trace where a user requested a landing-page prototype, expected an Ouroboros-style multi-turn brainstorm, and got a single-shot DESIGN.md emit instead. Diagnosis traced five concrete entry-side defects in skills/omd-harness/SKILL.md. All five fixed without breaking the run-dir / handoff contract — legacy runs (without prefilled_slots) follow the unchanged path.

Migration

npx oh-my-design-cli@latest install-skills --all picks up the new harness skill + master agent. No config changes required.

Files added to bundle

• scripts/ctx-prime.cjs added to package files: array.

v0.2 agent layer ships in the bundle. Blog feature removed. Release-hygiene routine added.

Bundle is now 15 skills + 16 sub-agents (was 9 + 11). Re-run npx oh-my-design-cli@latest install-skills to pick it up.

Added — v0.2 agent layer (6 skills + 6 sub-agents)

A supervisor + specialist topology for multi-step authoring, all channel-aware (Claude Code / Codex / OpenCode):

• omd-orchestrator — supervisor. 5-stage workflow (write → review → revise → localize → critic → images → handoff), hard 2-round revision cap. Anthropic orchestrator-workers + LangGraph supervisor pattern.
• omd-kr-writer — Korean prose with 12 voice presets (toss-tech-design default / karrot-neighborly / brunch-maker / naver-d2 / biz-report / academic / journalism / …). 9-field voice spec per preset in data/research/2026-05-18-kr-style-presets.md. Anti translation-ese.
• omd-locale-adapter — KR → EN/JP/ZH-TW adaptation (cultural reference swaps, JP honorific matching, ZH-TW traditional idioms). KR is the source of truth.
• omd-designer-review — visual + brand audit vs DESIGN.md (typo hierarchy, color budget, radius scale, component states, mobile). Severity BLOCK / WARN / FYI with line refs. Read-only advisory.
• omd-final-qa — read-only critic, 8-item publish rubric, forbids "looks good" rubber-stamps, hard 2-round cap.
• omd-codex-image — channel-aware image materialization. One <!-- omd:gen-image --> spec, three downstream paths: Codex native generation / omd-asset-curator free-license fallback (Claude Code) / OpenCode user-queue. Idempotent, IP-safe (no logo/face synthesis).

Added — release hygiene

• omd-release-hygiene skill + scripts/check-release-hygiene.sh + husky pre-commit gate. Blocks experiment byproducts (v1/v2 .bak, _source-*, contact sheets, logs, scratch dirs) from entering a commit. Codifies the docs-sync + npm-packaging checklist for new skills.

Removed

• /blog route and the blog feature (was dev-only behind NODE_ENV). Routes, components, @/lib/blog*, and the blog views/likes KV endpoints removed. omd-kr-writer is kept — KR voice synthesis is useful beyond the blog.
• kr-spellcheck — a deterministic Korean-orthography linter was prototyped and dropped. A 25-pattern regex heuristic had too low recall (~50–67%) to be worth shipping. If you need automated KR spellcheck, wire an external service such as the 부산대 한국어 맞춤법 검사기. (omd-final-qa dropped its rubric item 9 accordingly.)

Changed

• install-skills TUI — interactive multiselect order is now skill → sub-agent → channel. Agent channels default to none-selected (you pick); "detected" hints read real .claude//.codex//.opencode/ presence instead of the all-3 fallback. Added enter = 확인 hints.
• experiments/ is now gitignored (local scratch, like .promo/ and .experiments/).

Architectural refactor — single-source registry replaces 5 hand-maintained maps.

Three consecutive KR-10 batches (78→88, 88→98, 98→108) lost data because the sub-agent missed one of 5 hand-maintained data maps (route.ts, logos.ts, tokens.ts, homepage-urls.ts, design-systems.ts). homepage-urls.ts was missed in 2 batches in a row, requiring a 20-brand backfill. Prose checklists in omd-add-reference Phase 5 failed to enforce parity.

Option C — single source of truth + machine gate:

• Canonical YAML frontmatter in every references/<id>/DESIGN.md carries id, name, country (ISO), category (slug), homepage, primary_color, logo.{type,slug}, verified, and optional display_name_kr + ds block. All 108 existing references migrated via web/scripts/migrate-frontmatters.mjs (preserves any existing fm; merges from the 5 source maps; emits canonical block).
• Typed registry at web/src/data/registry.generated.ts (autogenerated, ~41 KB, 108 entries, sorted by id for stable diff). Built by web/scripts/build-registry.mjs, which validates the schema and fails fast pointing at the offending file. Auto-runs on npm run dev and npm run build.
• 5 consumers refactored to derive everything from REGISTRY / REGISTRY_BY_ID. route.ts, logos.ts, tokens.ts (BRAND_COLORS), design-systems.ts (getDesignSystem / getAllDesignSystems). web/src/lib/homepage-urls.ts deleted — getHomepageUrl now lives in registry.generated.ts. preview-export-view.tsx import updated.
• Catalog-integrity vitest at web/__tests__/catalog-integrity.test.ts: 108 parametrized per-id assertions (schema, §1 canonical header, prose-first paragraph, fingerprint cross-check, design-md mirror) plus 5 cross-cutting invariants (registry sort, triple-fingerprint byte identity, mirror coverage, llms.txt advisory check). 328 tests, all green.
• Husky pre-commit hook at .husky/pre-commit runs build-registry + catalog-integrity only when data-plane paths change (references/, data/reference-fingerprints.json, mirror copies, registry.generated.ts, scripts, the test itself). Other commits stay fast. Hook wired via core.hooksPath = .husky.
• Skill thin-out:
  • .claude/skills/omd-add-reference/SKILL.md Phase 5: 80 lines → 32 lines. Prose checklist replaced with the canonical frontmatter schema + a single npm run build-registry && npm test gate. Sub-agent self-verification step required.
  • .claude/skills/omd-batch-launch/SKILL.md Phase 3: 58 lines → 22 lines. Same pattern.
• Build verified — cd web && npm run build passes type-check and prerenders all 247 routes (incl. all 108 /design-systems/<id> pages). No regression in API, builder, playground, or browse modal — every surface reads the same registry.

Migration is one-way and irreversible: any new reference must follow the canonical frontmatter schema. The vitest gate (and husky) will refuse to merge a DESIGN.md missing a required field.

• +10 KR references (98 → 108): toss-securities, oliveyoung, gmarket, tving, catchtable, upstage, chai, dabang, jumpit, fastcampus. Each ships full DESIGN.md + assets/_reference/{tokens,structure,fonts}.json + .live-inspect-proof.json (≥ 5 raw_samples) with canonical §1 Visual Theme & Atmosphere (regression guard verified).
• Notable: chai flagged as service-terminated (2024 → PortOne Solutions rebrand) — kept as archaeological reference. upstage Tier-1 partial-positive (brand resource center exists, no token tables).
• Dev-only NEW tag — landing wall + design-systems catalog now surface a NEW badge on the latest batch (toss-securities/oliveyoung/gmarket/tving/catchtable/upstage/chai/dabang/jumpit/fastcampus) when NODE_ENV=development. Production rendering unaffected.
• Numbers cascade: READMEs × 4 + llms.txt + landing (hero/sections/the-wall/tokens) + SEO layouts (root/docs/builder/design-systems) + body copy (reference-preview/result-card). All 98 → 108.
• API mappings: route.ts CATEGORIES/COUNTRIES/DISPLAY_NAMES + logos.ts LOGO_MAP/DOMAIN_OVERRIDES + BRAND_COLORS + design-md/ mirror sync. Fingerprints × 3 byte-identical at count=108.
• Audit log: data/reference-audits/2026-05-15-kr10.md.

• +10 KR references (88 → 98): kream, upbit, kbank, inflearn, wadiz, channeltalk, lunit, bunjang, flex, classum. Each ships full DESIGN.md + assets/_reference/{tokens,structure,fonts}.json + .live-inspect-proof.json (≥ 5 raw_samples) per omd:add-reference skill.
• Tier-1 official DS finding: Channel Talk publishes Bezier (github.com/channel-io/bezier-react, MIT) — first canonical KR SaaS DS in the catalog.
• Numbers refreshed across surfaces: README badges + intro, README.ko.md, docs page, web/public/llms.txt, landing copy, root SEO metadata.
• web/src/app/api/references/route.ts + web/src/lib/logos.ts register the 10 new brands (categories / country / display name / logo / domain fallback).
• Audit log: data/reference-audits/2026-05-14-kr10.md.

• Docs only. README.md and README.ko.md gain an Upgrading section. CHANGELOG.md (this file) extended to cover the 1.2.x → 1.3.x line.
• Docs page (oh-my-design.kr/docs) gains an Upgrading card.
• No skill, agent, hook, CLI, or data change.

• omd:harness rule 9 (new): hero archetype catalog with 7 variants — left-character, center-text, carousel, split-screen, editorial-magazine, dashboard-preview, quote-led. Brand-vibe → archetype matching table. Same brand re-run → 2nd-rank archetype, so visual variation is automatic.
• experiment-meta.json gains a hero_archetype field.
• Per-archetype grep gates documented (e.g. center-text mandates position:absolute = 0 in the hero block; dashboard-preview mandates <img> = 0 + mock ≥ 50% of viewport).

• omd:harness rule 5 rewritten — header logo is a wordmark in a display font, not a separate icon mark. Anti-pattern list: DiceBear shapes+wordmark pairs, DiceBear initials chips, hand-drawn inline-svg logos, and dashed-box [YOUR LOGO] placeholders are all forbidden.
• Brand-vibe → font + product-name matching table added (Bricolage Grotesque / Space Grotesk / DM Serif Display / Fraunces, all SIL OFL via Google Fonts).
• omd:asset-fetch §1 rewritten to match.

• New skill omd:experiment-gallery — builds a single index.html that previews every brand experiment under a folder as iframe-scaled cards with archetype badges, wow ratings, multi-turn deltas, and per-brand IP audit. Reusable across batches.
• omd:reference-capture Phase 3.9 (new): browser-harness fast-path auto-detect. When the user has browser-harness (https://github.com/browser-use/browser-harness) installed and Chrome with --remote-debugging-port=9222, the skill drives live computed-style capture via CDP — measured 3-5× faster than playwright MCP. Falls back to playwright MCP otherwise.
• omd:harness rule 10 (new): IntersectionObserver reveal safety net mandatory. @keyframes failsafeReveal + html.js-ready gate + reduced-motion guard. Stops the recurring "fullpage screenshot has blank sections below the fold" regression.

• New skill omd:asset-fetch — free-license asset catalog with verified-200 CDN URLs. DiceBear (CC0 avatars: notionists / lorelei / personas / adventurer / fun-emoji), Lucide (ISC icons), Heroicons (MIT), Tabler (MIT), Picsum (CC0 photos with deterministic seeds), Loremflickr (Flickr-CC photos with deterministic locks), and SIL OFL display fonts (Bricolage Grotesque, Space Grotesk, DM Serif Display, Fraunces, Pretendard, Wanted Sans, Noto Sans KR).
• omd:harness rule 6 tightened — generator must read from omd:asset-fetch. Handcraft inline-svg characters / icons / logos forbidden. CDN call fails → fall back to a brand-color CSS gradient placeholder. Provenance comment <!-- omd-asset: source=<url>, license=<license>, fetched=<ISO> --> mandatory on any saved asset file (e.g. assets/illustrations/hero-character.svg).

• omd:harness rule 7: single shared .container-inner class enforces consistent horizontal padding across <header>, .hero, sections, <footer>. Stops the "header offset 24px from hero" class of bugs on wide viewports.
• omd:harness rule 8: hero composition decomposed into separate elements (.hero-character, .hero-chart, .hero-stat-card, .hero-ornament). Single-SVG-with-everything forbidden — asset swap no longer destroys the scene.
• omd:reference-capture Phase 4.0 (new): .live-inspect-proof.json mandatory with ≥ 5 raw samples from playwright browser_evaluate. If tokens.json#live_overrides ends up byte-equal to canonical, the block is deleted (drift-detection: live inspect didn't really run).

• Wordmark-only logo pattern introduced in omd:asset-fetch §1 (rewritten further in 1.3.7).
• omd:asset-fetch becomes the single source of truth for which CDN URLs the generator may use.

• Deferred experiment: a clone / inspired mode toggle that turned out to be the wrong split. Removed in 1.3.3 in favor of a single-mode flow.

• Skill descriptions made count-agnostic. Replaced hard-coded "67개 / 78개 / 88개 reference" strings with "실제 기업 reference" so the description never drifts with the reference count. omd:init and omd:harness SKILL.md descriptions updated.

• New skill omd:reference-capture — live brand-site CDP / playwright inspect. Writes assets/_reference/<id>/ with tokens.json, structure.json, fonts.json, screenshots/{hero-desktop,hero-mobile}.png, LICENSE-NOTE.md, and attribution.md. The captured material is reference-only and never embedded in user product DOM.
• omd:init CLI drift fix: the skill body referenced omd init recommend and omd init prepare subcommands the CLI never actually exposed. Rewritten to use the bundled fingerprints + Bash directly. Skill is now self-contained.

• 10 new Korean references (78 → 88 total). Fingerprints + DESIGN.md files now ship in the npm tarball: socar, gangnamunni, kakaopay, zigzag, 29cm, ably, banksalad, zigbang, wanted, remember.
• Web api/references/route.ts, lib/logos.ts, lib/homepage-urls.ts mappings extended for the new IDs. tokens.ts BRAND_COLORS const extended.

The reference bundle expanded from 67 → 78 and the npm tarball now actually carries them.

Added

• 11 new references (Korean cluster): yeogiotte (여기어때) · musinsa · kurly (마켓컬리) · ohouse (오늘의집) · naver · yanolja (야놀자) · coupang · kakaobank · ridi · qanda · krds (Korea Republic Design System, 행정안전부 주관 — the first government DS in the bundle).
• Each new ref ships with a .verification.md alongside its DESIGN.md documenting Tier 1 live DOM measurements, Tier 2 cross-check status, conflict matrix, and cited philosophy sources.
• data/reference-fingerprints.json regenerated → count 67 → 78, all 11 new entries added.
• data/reference-tags.md: 11 new rows + Korean shortcut buckets (한국 이커머스 · 한국 여행 · 공공·행정·정부).

Fixed

• Tarball was missing references entirely. package.json files[] pointed at references/**/DESIGN.md which is now a local-dev symlink — npm doesn't follow symlinks in the whitelist, so 1.0.x shipped 0 reference files. Updated to web/references/*/DESIGN.md (the canonical path); tarball now contains all 78 DESIGN.md files.

Changed

• omd-init skill references the canonical web/references/<id>/DESIGN.md path inside the package.
• Apple-tier philosophy verification for all 67 pre-existing refs (Tier 1 live DOM + Tier 2 cross-check + .verification.md with conflict matrix). Skill behavior unchanged — this only deepens the source-of-truth.

The install-skills "Next" panel still claimed "Hook이 자동으로 라우팅 — 디자인 의도 감지해서 하네스/스킬 호출" and described /omd-harness as "hook 우회". That language described the 1.0.0 forced-eval hook which was retired in 1.0.1 — auto-routing now happens via Claude's standard description matching, and the hook only gates on DESIGN.md existence. The banner is the first thing users see after install, so it's worth getting right even for a single panel.

Changed

• install-skills Next panel:
  • "Hook이 자동으로 라우팅 — 디자인 의도 감지해서 하네스/스킬 호출" → "Claude가 description 매칭으로 자동 라우팅. Hook은 DESIGN.md 부재 시 omd:init 안내만."
  • /omd-harness <task> description: "hook 우회, 즉시 진입" → "즉시 진입"
  • inline comment: "자동 hook 라우팅" → "자동 라우팅"

Patch release focused on skill triggering — descriptions, hook behavior, and a stale dependency in the package manifest. No public API change.

Changed

• Skill triggering uses the standard mechanism (SKILL.md description) as the single source of truth. Descriptions trimmed to ~230 chars on average (down from ~700) and rewritten in the canonical "what it does + use-when + chain hints" pattern that the rest of the Claude Code skill ecosystem uses. Modern models generalize from one phrase per language, so the keyword-inflated form was paying maintenance cost without measurable trigger gain.
• Skill descriptions now ship Japanese and Traditional Chinese trigger examples alongside Korean and English (omd-apply, omd-init, omd-harness, omd-remember, omd-learn, omd-sync, omd-add-reference). About half of OmD users work in JP / TW, so they previously fell back to weaker description-only matching with English/Korean phrases.
• .claude/hooks/skill-activation.cjs simplified from a keyword/regex forced-eval injector to a DESIGN.md existence gate. The "OMD SKILL ACTIVATION CHECK" block no longer appears on every UI prompt — the hook is silent unless DESIGN.md is missing and the prompt looks like UI work, in which case it surfaces a single one-line nudge to run omd:init first. The forced-eval layer was a custom belt-and-suspenders pattern from the 3.x model era; with 4.x description matching it became overlap and a second source of truth that drifted.

Removed

• .claude/skills/skill-rules.json (dual source-of-truth for the forced-eval matcher).
• Dangling reference to that file in package.json files[].

Breaking release. The CLI surface collapsed from 9 user-facing commands down to 1 (install-skills). All operational logic moved to skill markdown + agent prose. See MIGRATION.md.

Removed

• CLI commands: generate, init recommend, init prepare, learn, remember, sync, context, harness, setup-blender, reference list/show, preview.
• Public API exports (src/index.ts): detectContext, extractUrl, renderPlan, applyOverrides, loadReference, appendEntry, writeAllShims, all v4 modules. The package no longer ships a programmatic API; use the skill markdown directly.
• TypeScript modules (≈ 7000 lines): customizer, preview-generator, recommend, vocabulary, init-deprecate, apply-delta-stub, preferences, shims, sync-lock, sync-marker, run-archive, memory, code-introspect, context-detect, url-token-extractor, plan-emitter, signal-classifier, budget-tracker, visual-anchor, trace, and supporting utility / preset modules.
• Tests for removed surface (≈ 480 of 545 retired).

Added

• omd-asset-curator agent: stack-aware asset medium selection. Detects framework (Next/Vue/Svelte/etc.), CSS framework, icon libs (lucide/heroicons), chart libs (recharts/chartjs/nivo/etc.), animation libs (framer-motion/lottie/rive). Recommends the best matching medium per asset (inline SVG / chart library / Lottie / Rive / Unsplash / 3D dispatch). Inline SVG recipes for liquid fills, progress indicators, simple icons, empty-state illustrations.
• omd-apply skill: explicit dispatch decision tree. Complex requests (assets, charts, screens, a11y, microcopy, persona) now route to specialized sub-agents instead of being handled inline.
• scripts/context.cjs: 80-line deterministic project scan helper invoked by master Phase A as a fast cache, optional but preserved for repeatability.
• omd-3d-blender agent: just-in-time install walkthrough (uv check → brew install / brew --cask blender → claude mcp add → addon prompt). No longer bundled into install-skills.
• MIGRATION.md, this CHANGELOG.md.

Changed

• npm tarball ~60% smaller (no preview templates, no stale src/, no preset library).
• Dependencies: 9 → 3 (@clack/prompts, commander, picocolors).
• package.json: removed main, types, exports fields. Bin-only package.
• README rewritten around the single-command flow + agent natural-language pattern.
• Skill activation rules: omd-apply triggers tightened (39 → 23) so high-intent requests like "에셋 / 차트 / 아이콘" route to omd-asset-curator instead of being absorbed by omd-apply inline. Orphan omd-add-reference rule removed.

Migration

cd my-project
npx oh-my-design-cli@latest install-skills

Then restart your agent (Cmd+Q + relaunch for Claude Code).

• Drop unused dependencies (shadcn, tw-animate-css).

• Trim 90% of npm tarball (drop research / preview / per-reference READMEs).

• Five usability fixes from npm-publish verification.

• First publish under the oh-my-design-cli npm name.