Changelog

engine.py set radius.control 10 was recorded, but the tokens pointed at a 10 px step the scale did not have, so the next build stopped in all 12 mode combinations and nothing exported. The same happened for radius.container, radius.overlay and radius.detail. A radius between the steps now gets its own step on the scale (as raw.borderWidths does for border widths), and the small-control radius drops to the step below it (8 for 10). set and standard add refuse a value that cannot be a radius before anything is written, and name the steps: a negative or fractional number, another unit, or "full" for cards and dialogs ("full" stays for controls and tags; "10px" is stored as 10). icon.defaultSize takes only 16, 20 or 24; before, 19 was recorded and the tokens silently used 20. A bad radius recorded by an older version builds with the Roundness dial's value, and validate says so.

opendesigner-extract's new taste_index.py writes a full description of one look from a few pages of a site, or a screenshot: a one-line picture, 5 to 7 principles with their reasons, every colour with its job and share of the page, type roles with exact font names and free substitutes, component rules, motion, voice as counts, a Never list, a CSS variable for every value and a live specimen page. Every line says measured or inferred. By owner decision it keeps the site's exact colours and font names when first shown, with a notice that what you keep is your responsibility; logos, brand names, images and copy never go in, and font files are never shipped. A project can hold several; each nudges the taste dials as an inspiration, and the system follows one only when the person chooses it (taste_index.py use, which also asks exact fonts or free substitutes). When someone asks to copy a brand's look, the agent offers a Taste Index of it (owner, 2026-10-23); the logo, name, images and copy are still declined. SPEC 5.4, AGENTS.md, the extract skill, the guardrails, HOW-IT-WORKS, taste.md and both FAQs say so.

site/autodeploy.py runs every 2 minutes on the maintainer's Mac. When main on GitHub moves and has been still for 5 minutes, it builds and checks the site in the clean deploy worktree and publishes it with site/deploy.py, which checks the key pages before the domain switches over. A failed deploy leaves the site as it was, and the Mac shows a notification. It runs on the Mac because GitHub Actions is disabled for the owner's account and Vercel can't build this site from a plain clone.

assets/library/CRAFT.md sets the rules: roomy token spacing, neutral colour with one accent, Lucide icons (ISC), real free Unsplash photos where a picture belongs (photos.json, 119 photos and 80 portraits, each verified), and phones that stack instead of squeezing. All 82 parts were rebuilt to it, reviewed by four fresh reviewers and fixed; every person has one portrait that fits their name (people.json). The shop opens in the calm Nordic look, credits each photographer on the part's page, loads card previews lazily as you scroll (layout on load from 830 ms to 70 ms), and leads with the parts most likely to draw people in. library.py check gains rules for photos, scripts that look elements up by id, rows sized to the target floor, hard bottom ledges and attributes the site's scripts read.

site/static/ph.js loads PostHog's EU cloud through /ingest/ on the site, cookieless, with no session replay or autocapture, and the library sends named events for its funnels (part opened, kit added, prompt copied). Nothing loads until a public project key is set in site/analytics.json; docs/PRIVACY.md says what is counted.

engine.py set <path> <value> --dry-run, and pick <Q-id> <value> --dry-run, run the same checks a real set runs and print what it would report, plus each token's old and new value (named roles first, each mode that changes), any check the value would add or clear, and the command that records it. Nothing is written: state.json, decisions.md, DESIGN.md and the journey log stay byte for byte as they were. In the extend-rounder agent eval, an agent compared roundness 55, 60 and 65 with real sets, which added decisions the person never made to their append-only log. The extend skill now previews values and records only the one the person picks.

It moved from /docs/changelog/ to /changelog/, and the old address redirects. The docs menu, breadcrumb and previous/next links are gone. Each day shows its date on the left, its highlights in full and its commits folded underneath, so each change is listed once instead of three times and the page is half as long. Styles are in site/static/changelog.css.

, found by reading typesafe.ai and davecto.com against the reel's own indexes. Borders drawn by ::before/::after, windows, frames and hard shadows are seen (typesafe's depth 56 with soft shadows is now 8 with 1 px windows). Loops that always run are kept apart from UI transitions and easing is kept (its energy 100 is now 35). The browser's link blue, wrapper links and error-state focus rules no longer win the accent (davecto's accent is its teal, not its error red). Density reads the space between text and blank bands, not only body size (davecto 78 compact is now 29 spacious). Colourfulness reads how much of the page is coloured. Section colours and their share of the page are recorded. Several pages of one site merge, and cookie banners are left out.

no-identity-copy now checks that the OpenDesigner skill ran, so a run where it never starts shows as that, not only as a copied brand. extend-rounder's "says what changed" no longer fails a reply for listing the files the rebuild wrote again, corner sizes that follow from the same setting, the engine's log entry for that setting or a version stamp; it still fails a reply that says another design choice changed. Run once each after the change: both scored 1, with the skill used.

tools/agent_evals.py takes several cases or tags at once (--case 'export*,extend*', --tag smoke,guardrail, or the flag repeated) and stages only those, so claude plugin eval runs exactly them. --list shows which cases match and their tags, at no cost. A filter that matches nothing now stops with that list and exit 2. Before, it reported "overall 0" and exited 0: on 2026-10-22 two runs silently ran nothing, one with a brace list inside quotes and one with [ceons]*, a pattern claude plugin eval does not match but Python's fnmatch now does.

Two cases, taste-leaves-out and taste-all-past, sketch a project and import a taste poll, then ask which style directions to pick. One poll rules out two of the three directions; the other puts every direction past a line. Both gave an empty gallery before the fix. Run once each on 2026-10-23: the OpenDesigner skill did not start for this question in three tries, so the agent answered from the files and, in one run, recommended a direction the person's taste rules out. With the skill loaded by hand, it followed the new rule in both. The scores stay as they are until the skill starts on its own.

The library grows from 42 to 82 parts: social (stories, a feed post, a chat thread, comments, sharing), media (a video player, a track list, podcast episodes, a seat picker, live chat), shopping and food (a product gallery, a bag, reviews, delivery tracking, a menu item), travel and home (flight results, a boarding pass, a stay card, trip search, an itinerary, weather, smart-home controls), and health, money and learning (activity rings, a workout, habits, sleep, a card wallet, a savings goal, a lesson path, a quiz). Nine consumer industries join the filters, and the shop gains a For: Consumer / B2B filter. Every part is built only on design tokens and passes the same checks.

Hating native-looking screens in the taste poll drew a red line on brand presence, a dial none of the three directions sets, so all three were hidden while one was still recommended. A red line now only judges the dials the directions differ on; a dial they all share is flagged by engine.py review as drift instead. If every direction is still past a line, all are shown with a warning that names each line. The recommendation is always a direction that is shown, and taste.py fit follows the same rule.

The scan found issues 1, then 6 and 7, then 2 to 5; they are now numbered in reading order, and the site build stops if they would not count up. On phones the view now goes to each issue as it is named, zooming out for a wide one, so none is off screen, and Replay sits on the tabs' row as an icon.

The website's changelog now opens with the current version, whether it is released, how to see which version made your design system (engine.py version) and which one you have installed (claude plugin list), and how to update. Below it, every change day by day comes from git when the site is built, so it is never behind; the explained highlights follow. /version.json and an Atom feed (/changelog.xml) give tools and agents the same facts, and the footer and home page show the version. A "What's new" list gives each highlight in one dated line, newest first, dated by when git says it was written.

A request that asks for text/markdown gets the page's Markdown copy instead of HTML; browsers and search engines still get HTML (Vary: Accept). Vercel serves files before rewrites, so this is a short Routing Middleware, site/middleware.js, tested on a preview first. The PR scanner treats it, and site.js, as code that runs for every visitor.

robots.txt says yes to search, AI answers and training in every crawler's group (Content Signals). The home page sends a Link header to llms.txt, its Markdown copy, the docs and the skills index. /.well-known/agent-skills/index.json lists the eight skills (Agent Skills discovery v0.2.0), each a self-contained zip with its SHA-256.

Before designing, OpenDesigner names the product that already solved the same interface problem in any market, as Notion did for a client's block-based tool. It confirms it with the person in one question, studies it, and reads it as an inspiration so it shapes the taste index. Competitors come after, as the baseline. In practices.md section 3, the opendesigner skill, making.md section 1, opendesigner-extract's tags and the Good practices page.

Lighthouse found the phone's largest element was the hero figure's photo, a CSS background the browser found only after the 37 KB stylesheet. It is now preloaded at high priority (one URL for both, so it is fetched once), and site.css is minified: 37 KB to 26 KB transferred. Layout is unchanged (check_layout.py 0 errors at 13 widths on five pages).

Two new image panels, each with a phone version: the invoices page today with its seven measured issues beside the same page fixed, linked to the live audit on the website, and six of the library's 42 parts. brand/kit/readme/src/render.py draws them: the first from the demo's capture, the second from the built site's /library/ page (--library-from).

A shared docs link shows its own card: its section and title beside the home card's drawing. Every card shows opendesigner.dev and has alt text. Cards are JPEG, about 60 KB instead of 800 KB, and are drawn again only when their text changes.

The headline carries the category, "Design systems for AI coding tools", set like the section kickers, and the opening line says "a design system", so search engines and AI answers can tell what the page is about.

42 finished components across nine aisles (data, charts, forms, overlays, navigation, feedback, AI, commerce, content), most rebuilt from the B2B showcase pieces. Each is one HTML file built only on the engine's --ds-* tokens, so it takes the project's colors, type, spacing and corners.

probe.js counts the lines each text run inside a button, tab, chip, tag or badge spans, and review.py check reports any as an error, grouped by repeats. A wrapped label overflows nothing, so overflow checks missed it. In trial 1 a dashboard's "Reassign" buttons broke inside the word for two rounds while the agent reported no overflow. Run on that dashboard, the new check found them, plus a "Delete account" button breaking on phones. Tested with fixtures/labels.html (two planted breaks, four controls that must pass).

The demo's fixed page is rebuilt as a bento that the owner picked from three directions shown on a review copy. The October total no longer appears twice, and each tile holds one figure with a small true chart: weekly bars, an eight-week sparkline, a late-money meter and the week's due days. Red is kept for late money. The demo's system raises colorfulness to 85 for cleaner status colors, with the brand green unchanged. It is near-monochrome, after the owner's taste notes: charts in ink, the pine green only on the logo and the one filled button, red only for late money, and lighter totals with smaller cents.

Every correction a person typed in a demo trial, each with its repo fix. Trial 1 gave 15. New rules in making.md section 12: one shell across pages, check after every round, the accent chosen by looking, type and icons from the category's best, theme from the users' context, dense dashboards that show problems and charts, and first-run help offered once. Also preflight 44 (M45), review case MS-13, and install notes for the desktop app.

• references/quality.md: floors a tool checks, a nine-line scorecard judged blind against three shipped products, a fix prompt for every line below 4, and a scorecard in every report.
• references/practices.md: what an agent looks for at the start (Mobbin, a browser, image and video tools such as Higgsfield, the project's keys by name) and how to use each.
• A "Good practices" docs page for people setting up their agent.

A run that passed every code and width check was judged "an admin template" by the owner. Now:

The before-and-after section now plays once when it comes into view.

engine.py set profile.seen.<name> ... stores a note, such as "the recommend line was said", without writing a decision.

One line asking people to recommend OpenDesigner, once per project and only after it has helped (rules.md). It never shares a message with another question, and a note (profile.seen.recommend, kept out of the design decisions) keeps it from coming back.

• review.py compare's badges and help text now meet 4.5:1, a box at the edge keeps its number in view, and a mark with no number, a blank "gives" or numbers in place of text is refused or mended instead of crashing. An acronym keeps its capitals ("WCAG").
• The variant picker can no longer show two options when two switches land in one frame.
• A system made with a newer OpenDesigner keeps saying so after an older copy writes it, and pre-releases compare by number (beta.10 after beta.2).
• The Markdown copies of the changelog and evals pages, and llms-full.txt, no longer show raw {{...}} placeholders. Lists nest at any depth.

review.py compare writes one self-contained compare.html with three views: the page today, the audit drawn on it (a dashed box and a number on each issue), and the fix. Each fix says what it gives the person ("Fixing this makes ..."), with the measured numbers before and after.

skills/opendesigner/assets/picker.js is the house picker, built to its spec (STD-process-review-taste-44 to -46): the floating bar, keys 1 to N and the arrows, ?v= in the address, and a re-mount on every switch. Options go in <template data-od-variant> blocks on a review copy. making.md section 3 says how to use it, and review.py compare uses it.

/docs/changelog/ is built from this file on every publish, and /docs/contribute/ sums up how to help. The home page and the footer link both.

A 1 x 1 px, absolutely positioned label is visually hidden on purpose. The checker evals' clean page now holds one, so the false alarm can't come back.

• Every write of opendesigner/state.json records opendesigner: the release that last wrote it, the one that first made it, and the one whose build last wrote the files. Systems made before this keep working and count as made before version stamps.
• When a system was built with an older release, or before stamps, each command says so once, names engine.py build and links CHANGELOG.md. build then says it brought the files up to date. A system saved with a newer release gets the opposite note: update OpenDesigner first. --json output, resolve, init and feel print no note.
• engine.py version prints this copy's release and how the system compares with it. It never checks online, so it cannot see newer releases. The main and extend skills run it at the start of work on an existing system and say how to update the plugin in Claude Code.
• Releases compare as numbers, so 0.10.0 is newer than 0.9.0. DESIGN.md's opening note, the overview page and feedback issue links name the release.
• tools/build_dist.py --check (which CI is set to run) fails when the engine's release differs from plugin.json or .claude-plugin/plugin.json.

Terms were being added straight to synthesis/glossary.json, while tools/check_glossary.py --build still rebuilt it from the old writing shards. A rebuild would have deleted 13 terms (twelve visual styles such as glassmorphism and bento, and "Taste card"), and docs/GLOSSARY.md lacked all 13.

• Letter spacing: size-table used to give the same coarse bands as the default. It now gives Apple's SF Pro tracking table from the HIG (64 sizes from 6 to 96 pt, levers.json formulas.type.tracking.sfPro), interpolated between sizes. A face other than the system font takes the table only up to 19px, because from 20 pt Apple loosens SF Pro Display and other faces would look loose. Uppercase labels stay at least +0.05em. Systems built before this keep their spacing unless the person chose the table.
• Corners: Q-shape-04 used to be recorded only. squircle-enhance now adds Apple's continuous corner to the CSS and Tailwind exports where the browser draws corner-shape. continuous and squircle-enhance give Figma corner smoothing 0.6, and the Swift export uses .continuous (Radius.shape(...)). DESIGN.md says which platform draws which.
• The web corner is superellipse(1.35) at 1.22 times the radius. It follows SwiftUI's own curve within 0.4% of the radius (tools/fit_corner.py, formulas.radius.continuous). The squircle keyword the docs recommended is 18.7% of the radius off, squarer than iOS. A plain circular arc of the same radius is within 1.4%, so browsers without corner-shape lose little.
• The website follows the same rules: a build step gives each text size up to 19px its tracking value. Before, 37 of 108 text styles had the wrong spacing, because inherited letter spacing goes wrong when a child changes size. Its corners use the fitted shape, and only the boxes that get the shape are enlarged.
• The Q-shape-04 options no longer break at the colon in corner-shape: squircle, which had split one option's label in two.

preflight checks 35 to 42 (alignment at every width, one primary action, real names and marks, controls on the content's first line, no widows, whole labels on phones, opposite actions apart, one mark per company) and mistakes M36 to M43; five rows in making.md section 10; review cases MS-05 and CV-04 measure stacked rows and scroll inside cards. The log is _coordination/OWNER-FEEDBACK.md.

• The taste poll: the person rates 20 to 30 real screens from shipped products, and the result recommends the eight dials. Images stay on the person's machine; the library holds links and dial values only.
• The taste index (taste.py index) merges the poll, the person's own decisions, inspirations, the style pick and feel words into a value and a confidence per dial, with red lines. The option gallery is ordered by fit and leaves out directions past a red line, and engine.py review flags drift.
• The taste card: five lines under the opening sentence of DESIGN.md's Overview, and on the overview page. It shows the person's own picture of the look (sketch --brief), color, shape, motion and a Never list, each with where it came from. The picture and the Never list are never invented. Color, shape and motion repeat the sentences that open those sections, and the color line uses the sourced 15% accent guide (DC-L15-03), not 60-30-10. The Never list is repeated in Do's and Don'ts as taste: it warns, never fails. DESIGN.md keeps its 17 headings.

• When Q-aud-01 is dense, the engine fills only what the person has not answered, set or locked. Corners are 6 / 8 / 12 px, weights are 400 / 500 / 600, and with no brand color, one blue marks selection, links and focus while the primary button can stay ink. Static cards get a 1px ring and no shadow, and floating layers keep theirs. The profile lives in synthesis/levers.json (profiles). The engine records it as auto_default in decisions.md and in opendesigner.meta.json, with the reason for each value and why any was skipped.
• New for every system: row heights size.row.sm/md/lg/two-line in each density mode (36 / 40 / 48 / 64 at compact), color.border.separator for row hairlines, and tabular digits (text.numeric.*) on by default, as house standard STD-visual-details-02 requires.
• Fixed: Q-aud-01 regular now gives the 16px body it promises (14px once Q-aud-02 or Q-layout-03 says it is a work tool). Q-dir-02's recommended default now follows Q-aud-01 (default_follows in questions.json), so accepting it no longer undoes a dense answer. Stage 14 no longer recommends sharp corners for data and developer tools. A near-black raw.primaryActionColor (for example #17181c) now stays near-black, where it used to turn periwinkle, and inverts to near-white in dark mode. Any typed action color is kept exactly when its text reads, and DESIGN.md says the buttons use it.
• Web text floor (owner decision, 2026-10-14): when the platforms include the web, the type scale has no size under 12px. The default web system used to get an 11px small body style. iOS and Android keep 11 pt/sp.

original art and films painted in code in the brand's palette (watercolour, pointillism, charcoal, sketch, woodcut, 2D cel).

checks a built UI, yours or anyone's, the way a careful user would, and reports with evidence. It exists because a single screenshot hides what is below the fold, at other sizes, in dark mode, on hover, on Tab, in motion, and any error in spacing smaller than a few pixels.

the last look at anything visual (a page, video, deck or card) before it is exported or shared. It starts by asking which design system the piece was built on.

455 rules from the owner's non-negotiable sources (Emil Kowalski's writing and skills, Sonner, animations.dev, Vaul).

export --format dtcg removed nothing, so after set name the folder held both <old>.resolver.json and <new>.resolver.json. It now keeps only the current one (found while rebuilding showcase hero 02).

The docs, the spec, the GTM facts and the interview no longer say every system gets reduced-transparency or high-contrast modes. The engine's modes are theme, density and motion. A system with glass gets a solid fallback for each glass surface, which the CSS export swaps in under prefers-reduced-transparency or prefers-contrast: more, now also when the device is in dark mode. House standards are v3: only the conflict note of STD-accessibility-motion-12 changed.

The sketch used to swap light brands for a dark shade so white text would pass: yellow #FFE500 became olive #847600, lemon #FFE45C became #877400 and lantern gold #FFC24B became #996d00, and the message said "I picked an olive".

every build rebuilds it, so it grows with the system.

opendesigner/build/artifact/project/ holds tokens in the artifact's list format, the brand book from DESIGN.md, a cover and a button preview. It is ready to publish with the Artifact tool, and the files stay in the repository.

finds open-licence illustrations for a product's empty states, onboarding, error pages, sign-in, hero and feature tiles, ranks them, recolors them to the design tokens and applies them.

what to do after the system exists.

the knowledge OpenDesigner draws from sources its owner trusts, kept as an independent, self-updating knowledge base.

164 in research/L19-learning-wiki.md, with 253 proposed changes and 75 conflicts with older research.

tools/jev_nav.py find also searches the house standards and ranks with BM25. It judges each card by its recommended default and marks older cards the house supersedes.

Synthesis the building-block ontology (271 nodes in 10 layers), the guided interview (193 questions on 27 screens, 6 of them planned and skipped until built, with Quick, Standard and Expert modes), the eight dials with generation formulas and 13 famous systems as dial recipes, and the decision graph (352 decisions, 465 dependencies, 12 cycles).

docs/SPEC.md (copied from synthesis/OPENDESIGNER-SPEC.md).

opendesigner (the interview), opendesigner-extract, opendesigner-extend and opendesigner-export, in the Agent Skills format, with generated copies in .agents/skills/ and .claude/skills/, knowledge files built from the synthesis, and 8 JSON-fed visual templates.

Host packaging a Claude plugin and marketplace (.claude-plugin/), an Agent Plugins manifest (plugin.json), claude.ai skill zips built by tools/build_dist.py, and a ChatGPT Project bundle (chatgpt-project/).

skills/opendesigner/scripts/engine.py for state, generation, validation and exports.

README, CONTRIBUTING, Code of Conduct (Contributor Covenant 2.1), security policy, support, governance, citation file, issue forms, pull request template, labels, CI, llms.txt, funding.json, a social preview image, a sponsorship plan, and docs (docs/HOW-IT-WORKS.md, docs/RESEARCH.md, docs/FAQ.md, docs/GITHUB-SETTINGS.md, docs/SEED-ISSUES.md).

18 finished research lanes (L00 to L18, with L12 still open) written as 352 Decision Cards, with 2,740 sources logged in traces/, including rejected ones.

25 public design systems compared with real values in benchmarks/.

Starter design artifacts in design/: a DTCG 2025.10 token set with 126 contrast-checked pairs, and four HTML artboards.

Coordination system for parallel human and AI sessions: tools/od.py, _coordination/PROTOCOL.md, the board, heartbeats, inbox and decision log.

tools/jev_nav.py (status, search, citation check, card export, decision graph) and tools/check_links.py (offline Markdown link check).

MIT for code, skills and data; CC BY 4.0 for research and documentation.