Skip to content

Releases

Bogen 0.4.3

Choose a tag to compare

@marselnenaj marselnenaj released this 18 Nov 19:15

Agents get on with it.

  • Quick changes and new designs finish sooner. A run no longer has to fill in a formal self-review (a fresh screenshot, written observations and an eight-part rubric) before it may finish. Only objective defects in what it changed (overflow, unreadable text, placeholder copy) and the canvas plan for three or more new screens hold it back. The self-review and "Review like a designer" are there when you ask for a review.
  • "Do it again, it is not good" gets a new design. Asked to redo a design that is not good enough, an agent used to run the checks, find no craft defects and leave it as it was. A request to redo ("redo it", "not good", "a lot better", "komplett neu", "nicht gut") now tells it that checks only find craft defects and that a clean check is never a reason to keep the design: it makes a visibly different, better one and says what changed.

Bogen 0.4.2

Choose a tag to compare

@marselnenaj marselnenaj released this 18 Nov 18:19

The design review speaks your language and acts on one click.

  • A review in your language, in plain words. "Review like a designer" now writes its verdict and every finding in the language you use Bogen in, for someone who may not be a designer: the problem in a short sentence, why it matters to the people using the screen, and what will change. Text on your screens is quoted as it is, in its own language.
  • Fix means fix. Fixing the chosen findings starts the agent at once; a confirm for large fixes used to appear below the list, out of view, so the click seemed to do nothing. The fix shows at the top of the review while it runs, and one step in the chat takes it back.
  • Text, not "Copy". The review's score for the words on a screen, and the matching choice for directions in the chat, were labelled with the same word as the copy button, so translations read "Kopieren". They are labelled Text now.

Bogen 0.4.1

Choose a tag to compare

@marselnenaj marselnenaj released this 18 Nov 17:25

A fix for agents that load Bogen's full tool list.

  • Claude runs work again with every client. The Claude API checks each tool's input schema against JSON Schema draft 2020-12 and refuses the whole request when one fails. Two selector patterns in design_system_css and tokens_export had an unescaped [, and the number pairs of type_scale_set, the web tools and mockup points were written as draft-07 tuples, so any Claude run that loaded the full list stopped with "input_schema: JSON schema is invalid". The patterns are escaped, the pairs are arrays of two numbers with the same limits, and a test now checks every tool schema in every list mode. Thanks to acr-sed for the precise report.

Bogen 0.4.0

Choose a tag to compare

@marselnenaj marselnenaj released this 18 Nov 15:44

I made Bogen something you can talk to: call your agent and it answers in a sentence or two while it works on your file. Files open as tabs, a brand opens on a real brand book, and pictures come from your own Codex CLI.

Call your agent

  • Talk with your agent instead of typing. The phone button in the chat head (or "Call your agent" in the command palette) puts the call on top of the chat as one calm strip: a status line for who has the line and for how long, a quiet line for what the agent is doing, and Mute, More and End. Under it is the last thing said, in a space that is kept, so the chat does not jump; the whole transcript opens from there. The voice, how the call hears, hold to talk, moving with the file, sounds, agent calls and what the call remembers are in Call settings, from More. The agent's answer, its tools and the details stay in the chat underneath.
  • A conversation, not a reading. The voice answers in a sentence or two and never reads the agent's text out: "On it." when your request is in, quiet while the agent works (one soft sound if it takes a while), and a short spoken result at the end, such as "The headline is bigger now. Does that work?"
  • It lets you finish. A pause mid-sentence ("Make the headline bigger and ...") no longer ends your turn, and a "mhm" or "yes" while the voice speaks only lowers it for a moment. Real words over the voice stop it at once; "stop" stops the run, "undo that" takes it back.
  • It answers what you ask, and asks back when you are vague. "Which design are we talking about?", "What's on this page?", "Are you done?" and "What did you just do?" are answered by the voice from what is on screen and from the agent's last answer, without starting a run. "Change something" gets one short question back, about the agent's last offer when there is one. When the agent ends with a question or an offer, the voice relays it, and your "yes" goes to the agent. The voice is told never to claim a change the agent did not make.
  • Who speaks is your choice. A signed-in Claude Code on your plan (Opus at low effort, no tools) speaks by default, because it retells the agent most faithfully. Without it, a small open model on your computer (Qwen3 4B on a GPU, Qwen3 1.7B otherwise, Apache-2.0, downloaded on request and checked by SHA-256), or plain rules.
  • A natural English voice. English is now spoken by tiny-voca (Apache-2.0, trained on the CC BY 4.0 tiny-voca dataset), a 21 MB model with two speakers, F2 and M2, that reads a sentence in about 30 ms on the CPU. Its phonemizer, dizione, is ported to TypeScript, so nothing native is needed. The voice is chosen in the call bar; LJ (Piper) stays selectable. Calls speak English.
  • Heard on your graphics card if you like. With a GPU, Whisper large-v3-turbo (MIT, about 574 MB) stays loaded during the call and writes a sentence down in about 0.2 s, primed with the names of your file, page and artboards; without one, Whisper base runs through Voxtype. Everything runs on your computer and is checked by SHA-256 on download.
  • Every call is kept with the chat. What the voice heard and said, how it decided and how long it took, so a call can be looked at afterwards.
  • Say more while it works. What you add while the agent is working ("more modern, full screen", "use Fable as an advisor") reaches it at once: the run stops where it is, what it changed stays, and it carries on with your words as the next message. Only a task you mark for later ("after that…") waits for the run to end. The voice says what it understood each time instead of the same line, and never the same acknowledgement twice in a row.
  • "Keep me updated" gives a one-sentence update about every minute while the agent works, from what it is doing; "be quiet" stops them.
  • No minutes of silence. In a long run the voice says one short line about every 90 seconds from the agent's latest steps, never the same news twice; "be quiet" stops it.
  • Small requests run faster. When you have not chosen a model, Bogen picks one by the size of the request: Sonnet at medium effort for a small change ("make the headline bigger"), Sonnet at low effort for a question, and the default model for new screens, a redesign or anything unclear. The chat shows the choice next to the agent's name ("Sonnet · small change"). A model or effort you chose in the chat, in run settings or in the call always wins. Agents also check their edits with the layout map or design_lint, look at a picture only first and last, and run design_verify once at the end.
  • No more "Thank you." from nobody. Before anything is written down, a voice activity model (Silero VAD, MIT, under 1 MB, downloaded with listening) checks that there is speech at all, so silence, room noise, typing, a cough or a door never reach Whisper. Lines Whisper is known to make up on noise ("Thank you.", "Thanks for watching", "Untertitel im Auftrag des ZDF") are dropped when little speech was behind them, and the call's own voice heard back through the speakers is recognised and ignored. The call record shows which engine heard each line, and keeps what was taken for noise and why.
  • The call hosts the conversation; the agent gets a brief. What you say for one task is collected until you are quiet (about 2.5 s, or say "go"), then sent as a brief: a goal and conditions in plain words, your own words marked as speech recognition, your corrections ("no, I said not happy" replaces what it corrects), and what else you said to the voice. When a request is unclear or would throw away a lot ("from scratch"), the voice asks before it starts. Only work reaches the agent: thanks are not answered at all, remarks ("that's awesome", "I'm waiting") are only listened to, "what?" and "is it still working?" are answered by the voice, and nothing of that ever stops a run. Changes you say while it works go together after about 3 s of quiet; tasks for after are kept as a list, each sent on its own.
  • The agent keeps what it knew. A call run keeps its Claude Code session: a change ends the turn the way Ctrl+C does, and the next run continues the same session, so the agent does not read the whole file again. The agent knows it is in a call: it never asks you to paste anything, and ends its answer with a short line for the voice, which is spoken as it is. A pause mid-sentence now waits up to 2.8 s, never longer than 12 s for one turn. Replayed against a real call of 16 Nov: 13 messages reached the agent instead of 105, 3 lines that were not requests instead of 89, 5 interruptions of a run instead of 52 (none for a remark instead of 44), and every status question answered.
  • Hears you better, on any computer. NVIDIA Parakeet TDT 0.6B v3 (CC BY 4.0, through sherpa-onnx, Apache-2.0) is now how a call hears: 25 European languages in one 671 MB model on the CPU, about 0.3 s a sentence, with the names on screen as hotwords. On real voices it made about half Whisper turbo's word errors, and lost no sentence to a made-up "Thank you." (Whisper turbo lost 6 of 100). Whisper on the graphics card stays a choice in Call settings.
  • A question changes nothing. Asking the agent to check, review or say what it thinks ("What do you think?", "How does it look?") no longer ends in changes you did not ask for: it gives its view and offers changes instead. Questions written down without their question mark, as speech recognition does, count as questions too.
  • The call goes where you go. Leaving the file, going home or opening another file keeps the call, its microphone and its voice. In its own file the call is a card under the chat's header; everywhere else a bar at the foot of the content column shows the same: a ring and a glyph for who has the line (blue is you, amber is the agent, red only End and problems), a title that says what happens, what the agent does and in which file, the clock, and Mute, More and End in the same places in every state. What you said is in the card's transcript, from More; call settings are in More too. Hanging up leaves a short note with Call again. At phone width the bar spans the screen with 44 px targets.
  • It talks like a person, not a phone menu. With Claude Code as the talking layer, one model now decides and words every turn: it answers small talk and "hello?" in its own words, asks back when a request is unclear, hands only real work to the agent (in its own words, never reading yours back), changes or stops the running task, and moves the call to another file when you name one. What it says about the agent comes from the server's record, so "is it working?" is right also for runs started from the chat. The fixed lines are only a fallback when it does not answer in time. Measured on 11 messy calls, among them a real one of 17 Nov: turns that went to the agent although they were chat, got no answer or a wrong state went from 9 of 58 to 1 (a question the echo filter took for the voice), an Opus judge scored the calls 6.5 of 10 instead of 4.1, and the first word comes after about 1.3 s (1.8 s at most for nine in ten turns).
  • It knows when to keep quiet, and starts sooner. The voice says nothing it already told you during a run: a line that mostly repeats one of its last three is cut to its first part, or left out while the agent works and nobody asked. "Alright", "got it" or "you're amazing" while it works get no answer, and "you don't need to reply to every thanks" is kept for later calls. Half sentences ("Could you...") are let hang, with at most one "Mm?" a minute. A clear request gets a short "Okay." the moment you stop, and goes to the agent with the brief the voice wrote on its decision line, about 1.2 s after your last word instead of 5.5 s. The voice speaks from the first comma, the voice heard back is told from you by when it began (a question is never dropped), talking over it stops it after half a second, and a line you talked over is marked as not heard. Files are ...
Read more

Bogen 0.3.0

Choose a tag to compare

@marselnenaj marselnenaj released this 13 Nov 21:00

Agents get more reach: shorter tool lists and plugins, a script tool, browser capture, DESIGN.md, Mermaid, PDF and PowerPoint, a real round trip to pull requests, variations, visual compare, a pre-publish check and translation memory.

Connecting agents

  • Shorter tool lists for agents with a small context, and every tool still there: bogen mcp --tools compact lists each tool in one sentence (about 117 KB instead of 270 KB), --tools grouped lists 19 area tools that take a tool's name and arguments (about 19 KB). A schema tool and bogen://reference/{tool} give any tool's whole schema, and a tool can be called by its own name in every mode. Also BOGEN_MCP_TOOLS, and ?tools= or X-Bogen-Tools over HTTP. bogen install writes the list that suits each client: full for Claude Code and Codex, which load schemas on demand, compact for OpenCode, grouped for the others; --tools picks another. A shorter tool list.
  • Plugins: a Claude Code plugin and marketplace (/plugin marketplace add marselnenaj/bogen), a Codex plugin (codex plugin marketplace add marselnenaj/bogen), and a Claude Desktop extension, bogen-<version>.mcpb, built and signed with each release. Each starts the bogen command the install script writes, on this computer only. Plugins.
  • Add to agent in Settings, Agents: the MCP clients on this computer, whether Bogen is in each, and Connect, which does what bogen install does. Add to agent.
  • Layer links: bogen open "<file>#<layer>" opens a file with a layer selected and in view, both by id or name, and selection_set and view_focus answer with a link an agent can give you. A link opened in a tab that has the file already moves the selection without a reload. Layer links.
  • A reviewer profile, bogen mcp --profile reviewer: the critique's scope for a client of your own, with read-only tools, comments and checks with fixes off. A write is refused with the reason.

Capture from your browser

  • A browser extension for Chromium and Firefox captures the tab you are on, a region you drag over or an element you pick, including apps you are signed in to and pages on localhost, and sends it to Bogen on this computer, which turns it into layers through the same capture pipeline. You pair it once in the editor by comparing a code; its token adds captures and nothing else, and you revoke it in Settings, Privacy. It asks for activeTab, scripting and storage only and connects only to 127.0.0.1 and localhost. Build it with node --import tsx integrations/browser-extension/build.ts and load it unpacked. From your browser.

References for agents

  • A reference board on this computer: captures from the extension, artboards you keep and images you add, each with tags, its main colours and the type it uses, in Settings, References. Agents search it with reference_search before they choose a direction and keep something for you with reference_add; results are marked as untrusted design data. References.

Your agent

  • script_run: an agent writes a short JavaScript program against the open file (find layers by name, text, kind, style or CSS selector; set styles, text and attributes; rename; create from HTML; move, remove, duplicate; bind properties to tokens; read tokens and components) for edits that would take dozens of calls, such as renaming every button, binding every button colour to a token or building a grid of cards from data. Its changes land as one atomic step that one undo takes back, dry_run shows the plan first, and locked layers, revisions and a brand's strict mode hold. The script runs in QuickJS compiled to WebAssembly in a worker of its own, with no network, files, timers or Node, and limits on time, memory, calls, new layers and output; on any error or limit nothing is applied. Scripts has the threat model.
  • Propose and Directions take a range and aspects: how many copies, how far they go (Refine, Explore, Reimagine) and what they vary (Layout, Colour, Type, Copy), holding the rest fixed. Each copy keeps them, and proposal_list reports what a copy changed outside its aspects. proposal_create takes count, range and vary; races take range, vary and split. Variations.
  • A race can split the page: each lane takes sections of the selected layer instead of a copy of all of it, and accepting a lane takes its sections in. Split the page.
  • In the chat, / lists the design playbooks and your saved skills, and @ mentions layers, components and tokens, which reach the run with their ids. Save these decisions as a skill drafts a section for a skill from a refined artboard and its accepted proposals, for you to review and save. The chat, Save these decisions.
  • Reloading the editor while an agent works brings the chat back attached to that run, with Stop, its elapsed time and the reply still streaming, and a second message waits until the chat knows whether a run is going. When the server refuses a message, the chat shows its reason in your language ("A run is already going on this file") instead of POST /api/runs failed with 409.
  • Every run keeps a trace for bug reports: the prompt, tool calls with their arguments and results cut short, timings and touched layers, with secrets masked and files outside Bogen left out. Download it from the chat, bogen agents trace <run id> or trace_get. The trace of a run.

Compare, check and translate

  • A visual compare of two versions: side by side with scrolling together, an overlay that highlights the changed pixels, a swipe, and the difference alone. Review and merge on a branch lists the screens the merge would change, each opening the screen as it is now next to how it would be after the merge, with Previous and Next through the list. Proposals and race lanes compare with the original they would replace in the same view. Agents see it with design_branch_compare. Compare.
  • A check before publishing, in the Publish dialog and as publish_check: title, meta description, language, canonical address, the social picture and its size, icon, broken links and anchors inside the site, heavy images and images without a size, and short accessibility, responsive and brand results. It runs on this computer with no network, calls only real breakage a problem, and lets you publish anyway. web_page_set sets the title, description, language, canonical address, social picture and icon, which published pages now carry in their head, and published images get their width and height. Publish.
  • Language versions follow their original with a translation memory per file: only new and changed text goes to your agent, words you list are never translated, and each language is included, untranslated or hidden. New: translation_status, translation_apply, translation_rules_set. Keeping language versions up to date.
  • Renaming a token changes every var() reference in the file, also with a fallback and in rules, attributes, text ranges and components, as one undo step; a linked library or brand passes the rename on as an update. New: tokens_rename. Tokens and theme modes.

Formats in and out

  • DESIGN.md, the open format of YAML tokens and prose that many coding agents read: tokens_export and brand_export write one with format: "design-md", design_md_import reads one into the file's tokens and text styles or into the brand as suggestions, with the format's lint mapped onto brand_check, and a linked repository's DESIGN.md counts as a token source, with a token pull request writing one next to the token file. DESIGN.md.
  • diagram_insert draws a Mermaid flowchart, sequence, state or other diagram as native layers in the file's tokens: frames with text labels for nodes, SVG connectors with arrowheads, groups and lines, laid out by Mermaid and editable afterwards. Mermaid diagrams.
  • PDF pages come in as layers: text as text layers, paths as vectors, images as pictures, and only what has no plain form (gradients, masks, Type 3 fonts, annotations) as pictures, with a report of which. Drop a PDF or an SVG on the canvas, or call pdf_import. [PDF and SVG](https://github.com/marselnenaj/bogen/blob/v0.3.0/docs/guide/capture-and-import.md#pdf-and-s...
Read more

Bogen 0.2.0

Choose a tag to compare

@marselnenaj marselnenaj released this 13 Nov 17:09

A chat beside the canvas, fonts your agent draws, more kits and frameworks, accessibility and legal checks, and design guidance that gives agents more to work with.

Your agent

  • A chat with your agent per file, docked between the canvas and the inspector, in place of the floating prompt bar. Ctrl+Alt+3 shows or hides it and the file remembers it; Ctrl+K takes the focus. Answers render as markdown (never as HTML) block by block with the steps as they happen, the list follows new words until you scroll up, and Stop shows the time. Copy, Ask again and Edit sit under the last answer. The selection goes with a message as chips you can take out, and images can be pasted, dropped or attached. The title lists the file's earlier chats, and one opens again to continue. Review modes, runner, model, races, proposals, critiques and Redo this screen work as before. In a narrow window it opens over the canvas. The chat.
  • Speak instead of typing, in the chat and in every comment box on the desktop: your computer writes it out with its own speech model and the words land in the field to be checked. Voxtype, Omarchy's dictation, is found with a Whisper model it already has. Without a model the chat says how to get one, and a comment keeps a voice note. Speak instead of typing.
  • bogen setup finds the agents on this computer, says which ones have Bogen, connects the ones you
    pick and says what to do next. The install script ends with it; BOGEN_AGENTS answers ahead.

Design guidance for agents

  • The MCP instructions open with the design core: a direction written out before the first screen (audience and job, an idea from the subject with two references, type, colour as named hex values, a layout sentence per screen), the system as tokens, building with an ambition bar (the real product large, real content in quantity, a layout idea per section, states), and the checks. The design guide adds worked examples and a direction in full; each discipline skill opens with what strong work in its medium holds. The Claude Code skill no longer repeats the instructions. The social playbook covers ads and campaign images: one picture as a visual metaphor, a short headline that turns on it, the brand inside the scene, one colour treatment. Colour is a decision with roles (ground, ink, supporting colours, an action colour, tinted neutrals and a named contrast), and palette_suggest gives such palettes from a subject, a seed colour or a stored image, with the WCAG 2 contrast ratio of every text pair.
  • design_lint and design_review report twenty-nine craft findings for what reads as generated or unfinished: hard for objective defects, advisory with a reason for matters of taste, such as more than two font families, decorative monospace, equally loud type, competing filled actions, glow, tilt, drawn window dots and text on a picture without a scrim. Design playbooks and verification lists each line.
  • design_verify takes a scored rubric of eight criteria. It does not count while an objective defect stands (overflow, unreadable text, placeholder copy); advisory findings never block or cap, and a kept choice is recorded with its reason; a score above 3 does not hold while measured findings stand against that criterion, and built-in runs do not complete without the rubric.

Brands

  • A file goes to another device with its brand: a .bogen file of a linked file carries the brand's
    whole document and assets unless you clear Include its brand, and opening it adds the brand under
    the same id, or keeps the one that is there and says when the two differ. A brand also goes on its
    own (.bogen on the brand page, brand_export, bogen export <brand>), and design folders carry
    it in brand/. New: file_export, file_import, brand_export, bogen export, bogen import.

Kits, icons and code for more frameworks

  • Six more kits: Mantine, Carbon (dense, with a data table), USWDS (for public services, light only), daisyUI, Web Awesome (web components) and Pico CSS (classless HTML). Their tokens are read from each library's package, a brand maps onto them through the same roles, and none ships a project's logo or a government mark. Component kits.
  • Kits that are markup (daisyUI, USWDS, Pico CSS, Web Awesome) hand off as that markup, one class attribute per tag, with the scripts of custom elements imported; React gets it as JSX.
  • icon_search and icon_insert reach Phosphor, Tabler Icons, Heroicons (outline and solid) and Material Symbols next to Lucide, with names such as tabler:home, and set: "all" searches them together. Logos of other companies are left out.
  • code_get and export_file write Vue (script setup, TypeScript), Svelte 5 and SwiftUI: a file per component with typed props and its variants, instances as usages, tokens in light and dark, motion that honours reduced motion, and what the format could not say. Every Vue and Svelte file is compiled in the tests; SwiftUI is checked with swiftc -parse. Vue, Svelte and SwiftUI.
  • Licences says what Bogen may ship of each candidate design system, icon set and export target, and what it must not use.
  • In a set of a kit, each member's text props now come into code with that member's own text.

Fonts of your own

  • Agents draw fonts: font_create makes a page of glyph cells on one em grid with guides, font_glyphs_set writes glyphs from SVG path data, font_kerning_set kerns, and font_build makes an OTF and a WOFF2 that text in the file uses at once and exports carry. font_specimen sets a waterfall, a pangram, your words and spacing strings to judge it by, and fonts_made lists the fonts made in a file. Fonts of your own has the details.
  • A font can start from a traced drawing (image_vectorize, then from_layer) or from an open font whose own licence allows changes, under a new name that keeps clear of its Reserved Font Names. A proprietary font is refused.
  • On a glyph page the inspector shows the Font panel: Build font, downloads, a preview line in the current build, metrics and every glyph. Built fonts come first in the font list, marked Made here.
  • guide_get fonts tells agents how to make a display face hold together: one stem weight, overshoot, optical corrections and spacing with control characters.

Accessibility and legal requirements

  • compliance_check and the editor's Accessibility and legal check check a page against EN 301 549 V3.2.1, WCAG 2.2 AA, WCAG 2.1 AA, the European Accessibility Act with the Austrian BaFG, or the WZG for public bodies, criterion by criterion: failed with layers and fixes, passed, not applicable, or needs a person with a short test. They run axe-core in the renderer, text and non-text contrast, reflow at 320 CSS px, the text spacing override, language of page and parts, and form labels, errors and autocomplete.
  • The score never claims conformance while a criterion failed or is open; the automated pass rate is reported apart. A person marks what only a person can decide, and the marks are stored in the file with their name and the date.
  • compliance_statement drafts the Erklärung zur Barrierefreiheit in German and English in the EU model's sections, with the FFG's complaint procedure under the WZG and the Sozialministeriumservice under the BaFG, and the fields a person fills in marked. The report exports as HTML or PDF.
  • With legal: true, reminders with sources for what a design can show of other duties: Impressum and privacy links, consent banners, embeds and remote Google Fonts, order button wording, prices, discounts, urgency, green claims, ad and AI labels, picture licences. Accessibility and other legal requirements says what each law asks, as read on 13 November 2026. A technical aid, not legal advice.

Bogen 0.1.0

Choose a tag to compare

@marselnenaj marselnenaj released this 12 Nov 09:45

The first version. Everything below is new. The canvas and the other guides say where a
feature works only in part.

Design guidance for agents

  • Every design skill, the MCP instructions and the design guide open with a quality bar and a six-step work loop: brief and point of view, tokens and scales, sections with screenshots, whole-artboard review, scored rubric, report. The instructions kept their size.
  • design_lint and design_review report nineteen craft findings for the habits that make work look generated or unfinished, such as more than two font families, decorative monospace, equally loud type, competing filled actions, glow, tilt, drawn window dots and text on a picture without a scrim. Design playbooks and verification lists each line.
  • design_verify takes a scored rubric of eight criteria. A score above 3 does not hold while measured findings stand against that criterion, and built-in runs do not complete without the rubric.
  • design_lint also finds hedges, test-report language and headings chopped into fragments.
  • The checks read what a template shows rather than designs by its mark: counter-examples are left out, and specimens such as every button variant or the type scale are spared only the rules about composing one screen. An Omarchy theme file is read as a preview of the system, so its monospace bar, key glyphs and portrait monitors are not judged as a web page or a story. FAQ questions, file names, icon-font glyphs, var() fallbacks and unlike siblings no longer raise findings. What the checks read as something other than a page has each rule.

Large files

  • All cached screens remain visible in large canvas overviews. Actual WebP previews replace empty offscreen boxes after background capture; unchanged previews survive restarts.
  • Automatic live rendering follows expanded layer cost and device memory instead of a fixed four-screen limit. Selecting a screen still opens its editable content.
  • Preview capture runs one screen at a time, pauses under memory pressure and closes its browser context after every capture. The device-local cache has explicit size and record limits.
  • Canvas maps list unassigned screens even before a page has sections. External agents receive fresh organization status in file and context responses; text headings alone do not count as assigned sections.

Proposal decisions

  • An unchanged accepted proposal is recorded as accepted, including decisions synced from another editor. Changes inside a screen preserve layer ids without confusing the result.
  • Choosing one direction reports the other directions as discarded; an unrelated original edit cannot turn a discard into an acceptance.
  • Small decision receipts share the same undo step as the decision; the local preference history remains separate.
  • Completion checks stay inside an assigned revision/race copy when no layer ids changed. Creating an empty page does not recheck unrelated older screens; actual visual changes still require screenshot and verification.

Updates, packaging and checks

  • Large canvases show real cached previews and budget live screens by expanded layer cost. Short isolated render batches reuse one page for at most four screens; tiny thumbnails are derived from the completed raster. This also supports absolute clipped hug-height layouts. Offscreen proxies avoid retaining wallpaper/effects; geometry edits hydrate their real content before measurement. An aarch64 laptop passed the bounded 25-screen opening, focus, pan and edit fixture.

  • npm updates prepare and check a separate installation, retain the previous app and support explicit offline rollback. Preparation/activation failures preserve the current app and documents.

  • The packed runtime uses exact top-level versions and a shrinkwrap derived from the reviewed dependency lock, consistent with dependency notices and the SBOM.

  • Normal editor settings reads no longer start model catalogs. Opening a model picker still queries the selected CLI.

  • CI checks source hygiene, types, builds, documentation, core/editor contracts and signed-update recovery. Test examples use fictitious device identities.

  • Canonical sixteen-chapter brandbooks, sources/open questions and human review; a branch tree/list/checkpoint overview; canvas-only arrangement no longer expands quality checks to unchanged screen content.

The canvas

  • An infinite canvas where every layer is an HTML element with CSS: frames, flex and grid, absolute
    position with constraints, fixed, hug and fill sizes.
  • Tools for frames, shapes, lines and arrows, text, the pen and pencil, images, comments and shaders.
  • Smart guides, snapping, measuring, gap and padding handles, reordering flex children by dragging.
  • Fills with layered gradients and handles on the canvas; borders, shadows, filters, blend modes; any
    CSS property under Other styles, and Simplify styles.
  • Rich text ranges, Google Fonts and local fonts with variable axes; text that merges per character
    when several people type.
  • Vector editing, shape to path and boolean operations.
  • Lucide icons, shader layers, video and iframe embeds.
  • Images: crop with aspect presets and handles on the canvas, brightness, contrast, saturation, blur
    and grayscale as CSS filters, both kept by every export, and Vectorize, which traces an image into an
    editable SVG layer on this computer.
  • Designs from data: data sources of JSON or CSV, pasted, uploaded or read from a URL, kept in the
    file; layers bound to fields for text, image, visibility, props and variant, for a list or a single
    record; refresh with live copies on the canvas; code export that maps over the records. This
    computer's own network only when you allow it on this device.
  • Prototype links, hover styles and transitions, a preview with present mode, and a motion panel with
    keyframes.
  • Prototypes with state: boolean, number and text variables; interactions on click, hover and submit
    that set variables, show and hide layers, switch variants and navigate; layers, variants, fields and
    text that follow variables. The preview runs it without writing to the file.
  • Responsive variants: a 1440 artboard becomes 768 and 390 versions as media rules on the same
    layers or as linked artboards, then a check lists what still sticks out.
  • Command palette, find on canvas, layer search, settings, minimal UI, and a shortcuts dialog.
  • Comments with attachments and mentions; cursors, presence, follow and cursor chat.
  • A home screen with folders, search and thumbnails.
  • Measured with a file of about 10,000 layers: it opens and paints in about half a second, and an
    agent's write shows in the open editor about 30 ms after it is sent (Limits).
  • A welcome file on the first start, made once and never again once deleted (bogen welcome and
    Help open it), and four examples to start from, each with a small brand of its own.

Design systems

  • Tokens as CSS variables, theme modes with one value per mode, and text styles.
  • Components with instances, overrides, props, variants and slots.
  • Component kits: shadcn/ui, Radix Themes, Bootstrap 5, and Material 3 and Ionic in iOS mode as phone
    kits, each with variants, props and slots, tokens in the kit's own names with light and dark modes, a
    showcase page, and code that imports from the kit's library. A brand maps onto a kit's tokens, and a
    kit can be made from a linked repository's components. Kits under Assets, kit_list and kit_use.
  • Shared libraries between files, with updates you accept.
  • Code components: React components from your own project, rendered by its own code, and its
    Storybook stories as variant sets that stay in step.

Brands

  • Brands as documents of their own, synced, with logos, palette, type, layout, voice and rules.
  • Extraction from files, live pages, CSS and Tailwind configs, as suggestions you confirm.
  • The brand check with a score that explains itself, waivers by a person, and strict mode.
  • Versions, impact on linked files, and applying a change as one undo step per file.
  • Templates that lay out a brand system, a landing page or social images from a brand.
  • Memory from decisions: what you accepted and discarded, measured, turns into rule suggestions in
    the Brand panel that you add or dismiss, and a short summary for agents in brand_get.
  • Like this site, in a brand: a captured page redrawn in a brand's colours, fonts and scales, with
    placeholders for its text and images, then checked against the brand (site_restyle).
  • A setting an agent proposes for a brand, such as strict mode or no gradients, binds nobody until you
    confirm it, like any other suggestion, and publishing a version confirms nothing. brand_update
    refuses a value that is not one of a setting's choices instead of reading it as the default, and
    writes nothing of a call that fails.
  • A brand read from CSS keeps the stylesheet's own variable names as tokens and takes family roles from
    names such as --font-display. Looking at a live page and then applying it captures the page once.
  • A file whose brand was deleted says so with the way out, in the Brand tab and in every brand tool. The
    brand page says why an export, an upload or an edit did not go through.
  • Brand languages: a brand speaks a primary language and any others, as tags such as de-AT or
    pt-BR. Words, examples and text rules belong to one of them. brand_check, design_lint and
    voice_check judge copy in the language it is written in, and agents are told which languages the
    brand speaks and which one to write in. Brandbook ...
Read more