Frontend Frontier Radar Workflow

Daily frontend/design-engineering radar scans for signals that should change Kevin's stack, taste library, demos, or implementation habits.

The frontend-frontier-radar automation narrows signal radar to surfaces Kevin ships on: CSS, Tailwind, animation, R3F, shadcn, design systems, agent-browser workflows, and product-feeling frontend craft. Source: automations/frontend-frontier-radar.md, 2026-05-31

Routing

Use this page for frontend and design-engineering discovery. Use Signal Radar Workflow for broad tech pulse, SEO GEO Radar Workflow for search/AI visibility, and Content Pipeline Workflow when the output is a publishable post plan.

Trigger

• Daily automation from automations/frontend-frontier-radar.md.
• Manual request after a notable frontend trend, launch, library release, or design-system discussion.
• Before building a high-visibility frontend demo or landing surface.

Inputs

• wiki/design/frontend-frontier.md
• wiki/design/frontier-stack-2026.md
• wiki/skills/frontend-design-taste.md
• Public X discussion, GitHub Trending, Show HN, Product Hunt, and high-signal design-engineering creators.
• Design Inspiration Galleries as the job-shaped discovery router. Use editorial galleries such as VARCHIVE as queues of products and credited tools rather than quality verdicts.

Focus Areas

• Animation and scroll: GSAP, Motion, CSS scroll-driven animation, view transitions.
• Generated scroll cinematics: Scroll World only when the product is a pre-rendered camera journey rather than DOM choreography or runtime 3D.
• React and Next.js performance patterns.
• Design systems, registries, primitives, and component distribution.
• Icon systems as product dependencies: real-concept coverage, one family/style grammar, exact free/Pro asset rights, owned wrappers and semantic maps, accessibility/RTL/state behavior, direct imports, and measured bundle output.
• Browser automation for visual QA: Browser Testing Skills, Complementary Browser Testing.
• Perception-grounded visual QA: deterministic rendered state, atomic visible facts, repeated model observations, and separate design judgment.
• Anti-slop frontend taste: Frontend and Design Skills, Taste Enforcement - Anti-Slop Frontend Rules, CSS UI Enforcement.
• Direction systems that make greenfield/redesign choices reproducible, such as Impeccable, without turning inspiration worlds into copyable templates.
• Rich interaction demos that can become portfolio-quality proof.

Runbook

1. Read the context pages before scanning so the run has a baseline.
2. Name the artifact being designed—product surface, social post, OG image, deck, logo, brand system, motion detail, or broad scan—then route through Design Inspiration Galleries to the narrowest matching source. When the source is an editorial gallery, open the underlying product and its primary implementation or creator evidence; a feature card alone cannot promote a design rule, dependency, performance claim, or security claim.
3. Expand each source through its links, media, thread/replies, and correlated Discord or user-intent context before judging it. When a source claims to turn a DOM snapshot into a design surface, inspect the actual capture and source-resolution implementation. Require stable capture identity, semantic/framework hierarchy, source location with method and confidence, computed token/style state, route/viewport/fixture revision, an isolated non-mutating artboard, SSR/hydration/non-framework fallbacks, and exact license authority. A layer screenshot, catalog description, or React-only happy path cannot establish a general design-tool route. Compare critic and fixer authority separately; the worker applying a finding cannot certify its own result. Source: X/@lukaivanovic 2079178687409279303; benjitaylor/agentation@8158a97; cocopon/tweakpane@50a18c3, reviewed 2026-08-12 When a saved frontend source is itself an agent skill, resolve and pin the repository before accepting its procedure. For a generated scroll cinematic, distinguish the output from CSS/GSAP and runtime 3D; inspect the actual scrub engine, provider/model/price path, license, source-asset rights, tests/CI, mobile and reduced-motion behavior, cleanup lifecycle, and failure modes. Require neighboring clips to preserve both actual-frame position and camera velocity, review the seam in both scroll directions, and treat a native 9:16 chain—not a crop—as the mobile product. A paid run or silent connector degradation requires separate user approval. Source: X/@the_cyw 2075033878352494824; oso95/scroll-world@71cc36d3, reviewed 2026-08-12 When the source is a hosted component or animation registry, freeze each authority surface independently: saved post/media, live preview, catalog metadata, MCP/CLI client, exact delivered payload, package dependencies, and component terms. A catalog count, engagement total, copy button, preview, marketing “open source” label, or client-package license cannot certify the delivered source. Inspect representative code only to form hypotheses; the exact selected component still needs semantics, keyboard/touch, reduced-motion, teardown, offscreen/visibility, remote-import, frame/GPU, mobile, and rights proof. Originkit is the current reference: 160 live components plus 48 sections on 2026-08-12, not a verified 250+ catalog, with an MIT CLI but no public component-code license found. Source: X 2081620717775237573; Originkit; npm originkit@0.2.23, reviewed 2026-08-12 A machine-readable registry is a stronger intake surface than a social card, but it is not self-authenticating. Enumerate the complete catalog, free and locked cohorts, framework variants, access method, exact rights per operation, and broken canonical fields before selecting one payload. AIcss currently exposes 14 agent-UI components (nine free) through /r and llms.txt; its registry emits localhost:3000 URLs, its free code permits commercial app use but not as-is catalog redistribution, and its 25-orb gallery keeps hundreds of CSS animations alive offscreen. Route it as a discovery shelf, then compare the chosen job with its primary package or repository. For continuous agent status, the canonical thinking-orbs@0.3.1 package outranks the stale six-state third-party skill because the current runtime has nine states plus hidden-tab/offscreen suspension. Social post < imported skill < canonical package/source < host-app proof. [Sources: X 2085012911798292828; AIcss registry/pricing; boraoztunc/skills@11dd317f; Jakubantalik/thinking-orbs@e04f3e87, reviewed 2026-08-12]
4. Resolve unnamed interface vocabulary before evaluating or implementing it. Run ui-vocabulary on the screenshot, rendered page, DOM, code clue, or abstract description; preserve observed facts separately from unproved trigger, dismissal, focus, modality, and keyboard behavior. A public generic clue may use NameThatUI discovery, but no private product language may be sent to its logged search API. Verify the leading name and exact semantic/platform symbol against primary documentation, distinguish the nearest confusable pattern, and attach the vocabulary receipt to the source or implementation review.
5. For an interactive reference, extract its information contract before its style: entities, shared attributes, primary user verb, narrowing controls, overview-to-detail transition, provenance, and failure/empty behavior. Name whether the representation is a normalized grid, comparison atlas, physical consequence metaphor, collection room, graph, or something else—and reject the metaphor when the target product has a different data shape or job. Then run the behavioral contract from Design System: identify input and permission, local/remote processing, transformation and visible state, output custody, share/publish boundary, community governance, and keyboard/reduced-motion/error behavior. Exercise at least one reversible happy path in the rendered product. Inspect network activity for sensitive inputs. Stop before purchase, account/repository connection, submission, vote, or publication unless the external effect was explicitly authorized. For a living gallery, personal graph, or self-organizing surface, also trace the source objects and relationships through each classify, group, merge, surface, nudge, or action decision. Verify a complete list/search/detail alternative to spatial, hover, canvas, or mosaic navigation; subject consent and visibility for people graphs; distinct authority for read, summarize, edit, send, delete, and publish; target-bound approval and recovery; and visible connected scopes, processors, retention, and revocation. Measure the initial transfer and confirm selected-object lazy loading instead of accepting spectacle or automation as a performance exemption. For a generative or stateful reference, also preserve its replay recipe: exact input or data revision, seed/state, implementation version, viewport, selected mode, and output settings. Restore that state, exercise the quiet or reduced-motion path, and record where preview, save/export, and share diverge. A screenshot-only reproduction cannot promote a stateful interaction rule. When one bookmarked creative site is actually a family of generators, enumerate every mode before evaluating it and choose one exact mode for the behavioral fixture. Freeze the mode page and local asset hashes, dependency versions, controls, parameters, viewport, export path, and replay state. Keep four rights questions separate: implementation source, font or embedded assets, generated output, and redistribution. A license header on some files cannot license the unheadered remainder, and an enabled Save/Record control proves capability rather than commercial-use authority. Space Type Generator is the reference case: 23 current modes, 139 local script/style assets, eight visible export modes, p5.js versions spanning 0.7.2–2.0.5, and no corpus-level reduced-motion or lifecycle suspension. Source: X 2084270201802244140; Space Type Generator live corpus; golanlevin/p5-single-line-font-resources@bd08c4fe, reviewed 2026-08-12 When a bookmark is a font roundup, retain every named face as a signal but review each one as an artifact dependency, not a moodboard vote. Resolve the maker and primary distribution; pin the repository revision or hosted files; record internal version, hashes, bytes, formats, styles/axes, glyph/codepoint and required-script coverage, OpenType features, and upstream maintenance; then classify the exact rights as libre, free-commercial proprietary, commercial, personal-only, or unresolved. A social “free” label, zero price, trial, starter pack, CDN, or downloadable binary is not a commercial web or redistribution grant. Assign one literal product job, compare the incumbent, and run the actual host fixture with body/headings/controls/tables, numerals, punctuation, italics, required languages, truncation, missing-font reflow, subset/preload bytes, font-display, CLS, and host-derived metric overrides. High engagement earns deep review; only the artifact, rights, and host proof earn adoption. Source: X 2082553396259647652, 2083123338365595712, 2084182576823206011, 2084184286463402110; raw/web/source-evidence/font-roundups-deep-audit-2026-08-12.json When a bookmark is an illustration or stock-asset roundup, translate and enumerate every named library, inspect every saved board, then follow each destination to the primary asset/product and current terms. Keep CC0, libre/open-source, free-commercial proprietary, attribution-required, paid/plan-bound, personal-only and rights-unresolved states separate. Record creator, exact asset/product URL, terms revision/hash when available, local asset hash, format, price/count limit, attribution, modification, client-transfer, redistribution, trademark and ML-training boundaries. A zero price, download button, “free to use,” social engagement, or one site's license cannot authorize the cohort. Missing operative terms preserve a useful visual signal but fail closed on shipment. Source: X 2087346831248539978; ten primary sites; reviews/adoption-evidence/illustration-library-terms-2026-08-12.json When a saved artifact is a brand board, identity reel, or studio case, recover the creator, client/product, project status, deliverables, primary site, and expression/font/media/code rights before extracting style. Decompose reels by project and artifact class. Reconstruct the process as evidence brief, category anti-references, materially distinct directions, selection receipt, codified system, and cross-surface proof; one overview board cannot prove a complete brand world. For technical products, test whether the visual grammar derives from real mechanisms, objects, stakes, and failure modes rather than transferable AI-category decoration. Source: X 2083230040133964214, 2084293211141570717, 2084343196994224458; Drapz, Pixel Orb, and Runline AI primary sites; complete local media review, 2026-08-12 For reference-based reproduction, separate observation from copying: preserve creator and URL, asset/font/code licenses, allowed extraction path, and an originality plan; rebuild the target information contract and behavior with owned tokens and assets, then compare at frozen routes and viewports. Chrome or browser access can inspect public behavior but does not grant rights to extract protected code, media, fonts, or brand expression. A one-prompt clone, design-arena rank, or attractive one-shot remains candidate evidence until accessibility, responsive behavior, performance, interruption, source rights, and independent visual/product review pass on the owned artifact. For games and simulations, preserve state-machine boundaries, deterministic seeds, fixed-step/update semantics, render ordering, pause/resume/restart, content-schema scale, tail-frame time, input/focus, save/reset, and failure fixtures. A short demo reel cannot establish runtime or architecture quality. For an AI-native component gallery, enumerate every live pattern and inspect its actual source/behavior rather than accepting the launch count or demo reel. Distinguish ask, recommend, and authority-bearing approve; replace timers with named runtime events and receipts; keep product progress separate from private reasoning; resolve repository/license/dependencies before any copying; and prove representative keyboard, focus, semantic, mobile, light/dark, error, cancel, and reduced-motion states. A site-wide motion override does not prove that copied component source remains accessible in a different host app. For a creative/design-app starter, do not review it as a component catalog. Inspect the complete application architecture: shell versus product-source boundary, typed extension points, schema/control ownership, canvas and panel interaction ownership, reference-study procedure, persistence, timeline and animation intent, export intent, generated-source custody, functional/visual/artifact proof, performance envelopes, and failure recovery. Hash repeated embedded skills before counting them, sample tests across repository areas and conventional *.test.* / *.spec.* suffixes, and run one representative generated product before making it a default. A polished demo reel or large source tree does not prove that ordinary UI should inherit an editor runtime. Toolcraft is the retained route when the actual job is a focused canvas-and-controls creative application. When one object or character image should become editable runtime 3D, route through skills/engineering/img2threejs/SKILL.md before generic image-to-code prompting. Preserve the rights-cleared image hash, suitability and hidden-view uncertainty, intended use, local state, quality contract, sculpt spec, pass-by-pass code/render/comparison evidence, rejected corrections, and residual mismatch. The generated factory is only the asset layer: ordinary Three.js/R3F integration, meaningful subject-specific controls, semantic fallback, responsive behavior, and runtime budgets still need independent product proof. Never reuse a gallery's generic controls or final label as acceptance evidence. For an icon-family source, separate authoring procedure, accepted bitmap, procedural reconstruction, and host-product integration. Preserve source license/hash and never treat public model JavaScript as a reuse grant. Run preservation utilities twice on an already-accepted fixture before allowing them into the default route; quarantine any utility that changes the second pass. Keep polished presentation captures separate from fixed clean and map-stripped gate captures. Before GLB delivery, clone the named asset root, remove texture maps and live runtime references from userData, retain only portable metadata, and probe nodes, meshes, materials, bytes, and hash. The controlled voxel-router route is the current proof: 17 named meshes, 204 triangles, 0.9104 silhouette IoU, four non-degenerate views, reduced-motion behavior, and a sanitized 71,004-byte GLB; it proves the route, not arbitrary image reconstruction. Source: X 2084217575412826219; reviews/fixtures/voxel-icon-owned-specimen/, reviewed 2026-08-13 When a source demonstrates a complete procedural Three.js world, decompose it before adoption: route architecture through threejs-towers, the stage through threejs-landscape, and atmosphere through threejs-weather. Inspect the skill repository, license, demos, geometry/state seams, mobile behavior and performance receipts separately from the showcase application. A public repository is not necessarily reusable: Meng To's skill trio is MIT, while the Towers showcase explicitly grants no license for its code or artwork. Rebuild from the procedural rules and owned assets; do not copy the showcase. Prove clipping caps and winding, draw calls and disposal, camera-frustum particle density, transition interruption, audio/source rights, semantic fallback, reduced motion, mobile input and target hardware.
6. Extract atomic implementation, design, preference, task, and counterexample signals. Bind each signal to the current workflow, skill, design rule, tool route, project, or knowledge owner it could affect.
7. Interrogate those owners: state what the evidence reinforces, refines, replaces, contradicts, or leaves on watch. A no-change decision must name the current owner bytes compared and explain why they still hold.
8. Apply or propose the bounded owner changes, then test retrieval and routing. A source ID in a report or inspiration hub is not completion.
9. For greenfield products and substantial redesigns, run the six-gate AI-native pre-build contract: problem evidence, owned and attributed design direction, comparable key-screen variants, one readable product/design/technical spec, complete flows and states, and an ambiguity review before code. Keep the accepted spec and design authority synchronized during implementation. For an addition or scoped refinement, verify that the incumbent world stays authoritative rather than restarting the whole design process.
10. Before a visual critic scores the result, freeze route/state/viewport and ask only the applicable atomic perception questions. Corroborate localization, attributes, counts, relationships, text, deltas, details, context, and absent elements against DOM/data/scene evidence; repeat any stochastic observation that controls promotion. Then run the separate taste and product-fit review. Use agent-browser diff screenshot or committed Playwright image assertions when a machine delta is the proof. Use pinned @vercel/before-and-after only when the deliverable is a human-readable paired image/Markdown artifact for a PR or handoff. Freeze URLs/selectors, route state, viewport, full-page policy, labels, and image custody; keep captures local by default. Its --markdown path uploads to 0x0.st unless replaced, so private or authenticated screenshots must never take that default path. A side-by-side image proves a visible difference, not DOM, interaction, accessibility, performance, or correctness. Source: vercel-labs/before-and-after@1785942, reviewed 2026-08-12
11. Pick one "demo bet" only when there is a concrete implementation path.

Skill and tool composition

Do not install every high-signal design skill as another global authority. Route the job through one primary owner, then load the smallest complementary source:

The current primary-source replay reinforces consolidation. emilkowalski/skills contains ten focused skills, ibelick/ui-skills is a seven-skill registry/router, and Hallmark, Jakub Krehel, shadcn/improve, Remotion, and Diffusion Studio each carry useful but overlapping procedures. Preserve their distinct signal and revision, merge stronger rules into the incumbent owner, and create a new executable route only when the trigger, dependencies, authority, and held-out behavior are genuinely distinct. Source: commit-pinned repository receipts, 2026-08-12

The broader current cohort reinforces the same rule. MengTo/Skills is MIT at 3f4c22d1 with 127 skill files and no detected tests; buildermethods/design-os is MIT at 9dae0ba1 with one skill and two workflows but no detected tests; and cathrynlavery/diagram-design is MIT at da45d4a7 with one skill and no detected tests. They remain valuable procedures and source maps. Their size, stars, or polished examples do not grant global write authority: run them project-scoped, inspect generated files and dependencies, preserve reference rights, and measure behavior in the target app. Source: commit-pinned repository receipts, 2026-08-12

Architecture and performance claims

Treat social architecture prescriptions and framework backlash as failure hypotheses. Before changing the stack, reproduce the claimed failure in the target shape and measure server/client boundaries, cache keys, request fan-out, hydration, layout reads, bundle output, navigation state, and recovery. A chat navigation-trail component is useful only when anchor identity, streaming insertion, virtualization, keyboard behavior, reduced motion, and restoration remain correct. For text-heavy streaming or virtualized surfaces, evaluate Pretext only against the exact fonts, scripts, whitespace, line-breaking modes, browser ground truth, and bundle/runtime budget; it is a narrow layout engine, not a replacement for CSS or accessible DOM content. When the product genuinely needs 3D/canvas/XR text or one layout plan across renderers, retain pmndrs/text@d970744d as the current frontier reference: inspect its shaping and deterministic bake contract, but do not install its private 0.0.0 package as a stable dependency. Promotion requires a public versioned package or pinned source integration plus project fixtures for exact fonts/scripts, clusters and bidi, line breaking, editing semantics where applicable, atlas/memory lifecycle, Worker failure, WebGL/WebGPU parity, disposal/context loss, and a semantic accessible mirror.

For a multi-site collection, one completed review must account for every linked destination. Text-extraction failure routes to a rendered browser pass rather than a rejection. Keep the cluster in one stable owner when the durable output is design vocabulary or a shared rule; create a capability page only when one item has an installable/callable job that agents must route independently. A broken live state, inaccessible control surface, ambiguous license, or unsafe publication boundary is retained as a counterexample and test case, not erased from the evidence graph.

For performance-bearing changes, run the same evidence loop on the production URL rather than a development impression. Inspect the LCP element, client boundaries, shipped JavaScript, responsive image sizes, font preload/weight budget, hydration-time layout reads, route-level isolation of expensive experiences, cache headers, reduced motion, and layout stability. Prefer native HTML or server-rendered SVG when a component does not need interaction. Record the measured regression or improvement; a PageSpeed number is evidence, not a permanent badge. Source: X/@pontusab, 2026-07-21

Historical frontend rule lists are signals, not executable standards. The popular bendc/frontend-guidelines repository is a useful 2019 HTML/CSS/JS reference, but it has no declared license, no tests or CI, and includes era-specific absolutes such as avoiding loops. Preserve its concerns—semantic HTML, small dependency surface, progressive enhancement, readable selectors, cached expensive work—then verify every proposed rule against current platform standards, project constraints, accessibility, browser behavior, performance, and the installed local design/frontend skills. Do not copy the repository into a global skill or override an incumbent system merely because it is the “OG.” Source: X 2055291623727091903; bendc/frontend-guidelines@a6e4857, reviewed 2026-08-12

Specialized component proof

When a source's value is a difficult behavior—email editing, skeleton generation, localization, charts, calendars, drawing, maps, computer use, canvas/WebGL composition, native rendering, or vector/image export—evaluate that behavior directly instead of awarding points for visual breadth. Freeze framework/browser/platform versions, representative data and size, input methods, language/direction, theme, viewport, dependency/bundle budget, and expected output. Exercise keyboard, assistive technology, touch, zoom, loading/error/empty/recovery, reduced motion, and source/export round trips.

Experimental render targets never replace semantic truth by default. Keep DOM content and controls accessible when canvas or WebGL is decorative; provide a semantic table/list/control route for charts, maps, archives, and canvas-heavy interfaces; and require static/fallback behavior on unsupported or low-power devices. Native/HTML-in-canvas claims are promoted only on target-platform fidelity, accessibility, performance, editability, source ownership, and failure recovery—not on “no webview” or spectacle alone.

Security announcements follow the same current-authority rule. Resolve the target's installed Next.js/React versions and feature use, check current maintainer advisories, upgrade to the highest applicable fixed maintained line, then test the affected routes and lockfile. The saved May 2026 16.2.6 / 15.5.18 announcement is historical; later official advisories require at least 16.2.11 / 15.5.21 for the maintained lines they cover. Never encode a social release number as a timeless doctor threshold. Source: X 2052489312944759202; vercel/next.js advisories/releases, 2026-08-12

For archive and editor sources, inspect request fan-out after selecting one object or page. Record initial versus interaction-triggered requests, payload size, cancellation of stale work, cache behavior, and whether the product preloads unrelated pages or media. Promotion requires a bounded initial load and selected-object lazy loading; visual polish does not excuse fetching an entire archive path for one opened item.

For an icon-system source, popularity and screenshots create challengers, not a global replacement rule. Inspect the incumbent product imports, render its real high-frequency concepts at shipping sizes, and verify every named challenger's official repository/package/docs, exact revision, style coverage, and license. Distinguish open-source, free-with-attribution, free-but-not-open, and paid assets. Pin one family/style behind a project wrapper and semantic map, use direct imports, exercise accessibility/RTL/states, measure the production bundle, and preserve a no-change or rollback receipt. Brand marks route through The SVG instead of the interface-glyph decision. Source: X 2030310771326861336; eight-family exact-source replay, 2026-08-11

Promotion

Every signal remains represented in the evidence graph. Write it into the operating graph when it changes or materially reinforces one of:

• Frontend and Design Skills or Frontier Stack 2026
• a design or style rule
• a tool ranking or Skill Resolver route
• a reusable pattern that belongs in wiki/design/, wiki/style/, or wiki/tools/
• a project acceptance criterion or review surface such as Fieldwork — Visual Design IDE

Routine launches and inspiration links may remain evidence-only, but they must still bind to an object and record the comparison. Old reports, generic hub mentions, and engagement thresholds cannot independently close review.

PerceptionBench is a candidate offline diagnostic for the visual critic, not an installed frontend dependency and not a demo-quality score. Fetch its pinned dataset only for a declared evaluation run with storage, license, model, fallback, prompt, repeat, cost, and held-out/public-set receipts. Its ten axes can seed product-specific checks, but Fieldwork failures—not a frozen external leaderboard—govern the local visual-doctor suite. Source: PerceptionBench paper; repository; dataset

Validation

Each completed item needs source-revision identity, inspected artifacts, atomic signals, object bindings, an interrogation decision, writeback when behavior or knowledge changed, and retrieval/routing proof. If the run cannot access a required artifact, keep the review blocked instead of turning missing evidence into a terminal no-op.

Run Contract

This workflow follows Workflow Run Contract: name the sources, write output or no-op proof, promote only durable facts, update state only when the run really completed, refresh generated surfaces when durable pages or skills change, and log user-visible work.

---

Timeline

• 2026-08-12 | Added the illustration/stock-asset roundup contract after the saved ten-site list resolved into CC0, custom commercial, proprietary bounded, attribution/plan-bound, per-asset price and rights-unresolved states. Future reviews retain every signal but authorize shipment only from an exact per-asset rights receipt. Source: X 2087346831248539978; ten current primary terms surfaces

• 2026-08-12 | Added the font-roundup artifact contract after expanding ten named specimens into current repositories, foundry products, hosted files, internal metadata, five distinct rights classes, semantic product jobs, and fallback/host fixtures. Saved names remain useful even when their current disposition is licensed, personal-only, or rights-unresolved. Source: X 2082553396259647652, 2083123338365595712, 2084182576823206011, 2084184286463402110; deep font evidence and exploration receipt

• 2026-08-12 | Deep-reviewed vercel-labs/before-and-after@1785942: retained it as the paired screenshot/Markdown/PR artifact route, below agent-browser and Playwright for actual verification. Added frozen state/viewport/selector requirements and blocked its default 0x0.st upload for private or authenticated captures. Source: X/@hugorcd 2086471205998264699; pinned README, skill, source, and 31-test receipt

• 2026-08-12 | Added the hosted component-registry expansion rule after the Originkit deep review: freeze social claim, preview, catalog, MCP/CLI, payload, dependencies, and terms separately; audit the exact selected effect; and never promote on engagement, count, copy affordance, marketing label, or client license alone. Source: X 2081620717775237573; Originkit; npm originkit@0.2.23

• 2026-08-12 | Reopened X 2075033878352494824 after its prior cohort review failed to resolve the truncated GitHub URL. Exact repository and engine inspection promoted Scroll World as a distinct installed skill and added paid-provider, source-rights, position-plus-velocity seam, native mobile, reduced-motion, teardown, and independent playback gates to this workflow. Source: oso95/scroll-world@71cc36d3; local CLI and skill admission proof

• 2026-08-12 | Added the semantic DOM-to-design admission pass after the article's anonymous-layer failure was verified against Agentation's actual React/source-location implementation and PolyForm Shield license. Future reviews must prove capture identity, source confidence and fallbacks, isolated non-mutating state, tool scope, critic/fixer authority, and license from primary evidence before promotion. Source: X 2079178687409279303; benjitaylor/agentation@8158a97; cocopon/tweakpane@50a18c3

• 2026-08-12 | Replayed the final design/tool cohort into existing owners: recovered the full JoePro frontend-design article through the FxTwitter fallback, retained its context-aware design-skill signal without installing a second global skill, and routed Shader Lab, pmndrs UI, Dither Kit, Blume, photography, graphic-design, DOM-editor, critique, and Vercel-design evidence through the design-system contract. One deleted/private X Article remains an explicit evidence gap rather than reconstructed text. Source: compiled from X bookmarks 2076877282312954311, 2077122916227920288, 2043771441297150366, 2075091550258815080, 2075780400199446966, 2076714329404268576

• 2026-08-12 | Retained Open Lovable, Open Generative AI, Astryx, OfficeCLI, Saguaro, OpenAuth, micro-transition, slide, and interface-reference sources as candidates under the existing reference-reproduction and specialized-output contract. Authorized target, creator/code/asset rights, prompt-injection and secret isolation, originality, editable output custody, accessibility, performance, deterministic comparison, and recovery govern use; popularity or a one-prompt clone does not authorize copying. Source: X 2077698254566011070, 2041546716806357435, 2071735241786675703, 2077202205682348032, 2080778117690581353, 2077541589048541526, 2075899704693666282, 2081360169750151463; current repository receipts

• 2026-08-12 | Added specialized-component proof for email editors, skeletons, localization, charts/calendars, drawing, maps, computer use, canvas/WebGL, native renderers, and export tools. Experimental render targets must preserve semantic fallback, input/accessibility, source ownership, round trips, performance, and recovery. Source: 86-source frontend/design cohort; 17 current repository receipts

• 2026-08-12 | Reconciled a 45-source frontend cohort through consolidated owners: current Native SDK, Lightpanda, Graphify, OpenPencil, Mapcn, Extend UI, DESIGN.md/Stitch, MengTo, Design OS, UI Skills, Vercel skills, Matt Pocock skills, and Diagram Design sources. Added project design-authority, reference-to-prompt, and diagram routes while retaining live sites, games, motion demos, and catalogs as rights-aware visual candidates rather than multiplying global instructions. Source: 45 saved X records; 15 commit-pinned repository receipts

• 2026-08-12 | Added reference-reproduction and simulation gates from the website-clone, model one-shot, design-arena, and Princeton Tower Defense cohort: creator/asset/code rights, owned originality plan, frozen visual comparison, accessibility/performance/interaction proof, and deterministic state/render/failure fixtures. Source: 14 saved X records; Kimi K3 primary report/source; current frontend proof contract

• 2026-08-12 | Reconciled the browser/artifact, design-skill, frontend-architecture, Pretext, React Doctor/ReactBench, Lighthouse/browser-to-API, and motion-source cohort against current primary repositories. Added a single-owner composition matrix, explicit adapter/security boundaries, and target-app proof requirements instead of multiplying global skills. Source: 37 saved X records; 15 commit-pinned repository receipts; official Lighthouse agentic-browsing documentation

• 2026-08-12 | Retained the original bendc frontend guideline corpus as a dated reference, promoted its durable concerns through current standards rather than installing a competing global skill, and added current-authority handling for framework advisories. Source: X 2055291623727091903, X 2052489312944759202; pinned repository and current Next.js advisories

• 2026-08-11 | Added the AI-native component-gallery review pass after Beautiful UI's saved “17 patterns” became 19 live patterns. Runs now enumerate the actual revision, separate ask/recommend/approve authority, replace demo timers with runtime receipts, protect private reasoning, resolve source/license/dependencies, and test representative keyboard/mobile/theme/reduced-motion behavior before adoption. Source: X 2082479500944904432; exact live-site, source-dialog, and media replay

• 2026-08-11 | Added the icon-system source contract after replaying Om Patel's complete post and four images against nine repositories, eleven npm records, and six official docs. Reviews now distinguish useful visual signals from adoption, require incumbent/product coverage and one family/style grammar, resolve exact free/Pro/set licensing, enforce owned wrappers/direct imports/accessibility/RTL/state proof, measure bundle output, and keep no-change/rollback evidence. Source: X 2030310771326861336; exact capture manifest

• 2026-08-11 | Added replay-state and archive-fan-out proof. Stateful/generative references now require exact input/data revision, seed/state, version, viewport, mode, output settings, restored-state proof, quiet/reduced behavior, and separate preview/export/share boundaries. Archive/editor reviews must measure selected-object request fan-out, cancellation, and caching rather than accepting a fast initial paint alone. Source: X 2077488853401620893; Retrowave Road, Paint Pomodoro, and clipart.studio rendered/source replay

• 2026-08-11 | Added the unnamed-interface vocabulary gate: preserve visual versus behavioral clues, resolve at most three candidates, verify the winning platform semantics/API against primary sources, record the decisive confusable difference, and prevent private project language from entering NameThatUI's logged public search. Source: X/@argofowl 2076284369883615555; NameThatUI methodology, API, focus-ring entry, and complete launch video

• 2026-08-11 | Added a rendered behavioral pass for interactive sources: one reversible happy path, sensitive-input network boundary, output-versus-publication state, governance, and keyboard/reduced-motion/error proof. Multi-link bookmarks now remain incomplete until every destination is accounted for; extraction failures route to browser inspection, while broken states and license/accessibility failures remain counterexamples instead of becoming rejections or disappearing. Source: X 2070242834985431293; complete eight-site browser replay

• 2026-08-11 | Added an information-contract pass for interactive references: agents must extract entities, attributes, primary verb, controls, overview/detail transition, provenance, and edge behavior before borrowing a grid, atlas, physical metaphor, room, or graph. Source: X/@jonaasw1 2079246871063568437; Natural Selection; Humanity's Last Machine; Ciggie Index; sneak peek; Shape of Minds; local image review

• 2026-08-11 | Added perception-grounded visual QA before taste review: freeze rendered state, isolate applicable atomic facts, corroborate against deterministic evidence, and repeat verdict-bearing model observations. PerceptionBench remains a conditional verifier diagnostic with explicit storage/license/held-out boundaries. Source: X/@Kimi_Moonshot 2081813202514681878; arXiv:2607.24957v1; repository/dataset and local image

• 2026-08-11 | Replaced undirected inspiration browsing with artifact-first discovery routing. A run now names the output, selects the narrowest gallery through Design Inspiration Galleries, preserves the original creator/source, and keeps interactive proof independent of gallery selection. Source: X/@tranmautritam 2080583043929223469; nine primary gallery sites; local video review

• 2026-08-11 | Added VARCHIVE as a product-and-taste discovery input after the source video proved the bookmark's media was an editorial AI-assisted-work gallery, not a demo of its first link. Gallery inclusion now creates a review lead; the underlying product still needs direct design, behavior, accessibility, performance, rights, and provenance inspection. [Sources: X/@csaba_kissi 2068961003799196098; VARCHIVE; local 22-second video review]

• 2026-08-10 | Added Impeccable v4's job-shape and direction-proof gate to frontend promotion: greenfield/redesign bets need reproducible direction and pre-build visual intent, while additions/refinements preserve the incumbent world. [Sources: X/@pbakaus 2079984598105358828; Impeccable skill-v4.0.4]

• 2026-08-10 | Replaced report-centric source promotion with the evidence-to-operating-delta contract. Every frontend signal now binds to current owners and interrogates them; source mentions and old reports no longer prove integration. The DialKit plus Discord replay demonstrated the route into Fieldwork's animation-testing contract. [Sources: plans/005-evidence-driven-wiki-rebuild.md; X/@joshpuckett 2079014450712203291; private Discord capture, 2026-07-20]

• 2026-08-10 | Added a production critical-path review for frontend signals: LCP visibility, server/client boundaries, JavaScript, responsive images, fonts, startup layout work, route isolation, caching, accessibility, and layout stability. Source: X/@pontusab X Article, 2026-07-21

• 2026-07-01 | Rebuilt as a frontend-specific radar contract with routing, trigger, inputs, runbook, promotion criteria, and validation against current design/frontier pages. Source: User request, 2026-07-01

• 2026-05-31 | Workflow page created from frontend-frontier-radar automation. Source: automations/frontend-frontier-radar.md, 2026-05-31