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.