Skip to content

perf(images): WebP covers with srcset, JPEG share images, hero and thumbnails through astro:assets - #247

Merged
hotlong merged 32 commits into
mainfrom
claude/pm-dispatch-objectos-ju9td1
Oct 15, 2026
Merged

hotlong merged 32 commits into
mainfrom
claude/pm-dispatch-objectos-ju9td1

Conversation

@objectstack-fleet

Copy link
Copy Markdown
Contributor

Fixes #220

Branch. This PR carries 5 commit(s) from the seat's linear stack, replayed onto main @ 1408a0c. The replay is patch-equal to git diff 0b84524 lin/issue-220, with byte-identical diffs. Head: 85474eb. The staged SHAs quoted below are the pre-replay commits; their content is unchanged.

Images were almost the whole page weight: 1.6-1.9 MB PNG covers as the LCP element of 80 posts, a 1.7 MB hero loaded by 120 marketing pages, and an SVG og:image on 209 posts that X, LinkedIn, Facebook and Slack cannot draw. This PR fixes all three. The PNG-cover posts go from about 10-11 s LCP to 1.7 s on a phone under the card's profile, the marketing hero from 10.6 s to 2.0 s at 1366, one scroll of /en/blog/ downloads 1.2 MB instead of 18.8 MB, and a cluster page downloads 0.28 MB instead of 1.18 MB. Every post now has a 1200x630 JPEG share image.

What changed

1. Covers (BlogPost.astro, ArticleCard.astro, MarketingHome.astro, ClusterPage.astro). A PNG or JPG cover now goes through the astro:assets Image component as WebP, with a width list and a sizes value for each slot:

  • Post header: 480-1600w, (min-width: 908px) 860px, calc(100vw - 2rem). It stays eager with fetchpriority=high because it is the LCP element.
  • Card thumbnails: 400-1280w. sizes gives the width the image is actually drawn at. A 16:10 thumb crops a 16:9 cover with object-fit: cover, which draws it 1.11x the card's width.
  • Feature card: its cover is eager, because it opens the blog index above the fold.
  • Home latest-posts grid: 360-1080w.
  • Cluster reading-path cards: 400-1080w on the 40 cluster pages. They loaded the original 240-370 KB JPGs into 180px slots. Above 640px the image column is 180px (150px at 641-900px) and as tall as the card's text: I measured 190-227px. object-fit: cover fills it by height, so the cover is drawn 304-364px wide and sizes says 400px. On a phone the card stacks and the 16:9 image is drawn at the column width.

An SVG cover is a few KB and scales itself, so it keeps a plain img element on the original. The ai-agent-workbench cover at 1280w is 58 KB; the original is 1.62 MB. A 1:1 crop against the original shows no visible loss (PSNR 39 dB), and neither do the close-ups and thumbnails listed under Browser check.

2. Hero (src/assets/objectos-home-hero.png, moved from public/images/). It is served through the Image component as WebP, 400-1586w. The hero box crops the 1.6:1 image with object-fit: cover, so sizes gives the drawn width: 545px on a desktop, and about 1.2x the box on a phone.

  • Marketing evidence pages: I measured all six viewports. The hero's top is at 172px on a desktop, where it is the LCP element, and at 713-781px on an 800-844px phone. It is in the first screen at every viewport, so it stays eager, now with fetchpriority=high.
  • Home: the hero sits below the hero copy and the video at every viewport, so it stays lazy.
  • Dark mode: a muted image. No dark variant of the render exists, so the image itself is dimmed with filter: brightness(0.6) contrast(1.12). I compared three strengths side by side in the panel. No layout CSS was touched.
  • Original dropped from dist: the components no longer read the import's width, so Astro drops the 1.7 MB original from dist/.

3. Share images (scripts/og-images.mjs, scripts/lib/og-image.mjs, scripts/lib/og-fonts.mjs). A build step runs after astro build and renders one 1200x630 JPEG per cover into dist/og/blog/POST/.

  • Page tags: BlogPost.astro names that JPEG in og:image, twitter:image and BlogPosting.image. It also sets og:image:width/height to 1200x630 and uses the post title as og:image:alt.
  • One URL rule: the page and the renderer share it (scripts/lib/og-image.mjs). The URL carries a hash of the cover's bytes, so an edited cover gets a new URL that the networks re-fetch.
  • Fails the build when:
    • a page names a share image that no cover maps to;
    • a page declares a size that is not the file's;
    • an og:image of this site is not in dist/.
  • SVG text uses bundled fonts only. sharp renders SVG text with the machine's fonts. The covers ask for -apple-system, Segoe UI, Inter, Arial and Menlo, and CI's ubuntu-latest has no CJK font, so the Chinese covers would have shipped as boxes. The rasterizer now sees only bundled OFL faces:
    • Arimo and Cousine: these have the Arial and Courier New metrics that pnpm svg:lint measures with, so a line that passes svg:lint has the same width in the raster.
    • Noto Sans SC, Noto Sans Symbols 2, Noto Sans Math, Noto Emoji.
    • Why wawoff2: Fontsource ships WOFF2 and WOFF. The FreeType in sharp's libvips reads neither (every glyph drew as a box), so wawoff2 decodes them to TrueType for a generated fonts.conf.
  • Glyph gate: every character in an SVG cover's text elements is checked against the faces' cmaps before rendering. A missing one fails the build and names the cover, the code point and the line.
  • Cropping:
    • SVG covers are drawn whole at the card's height and widened by repeating their edge columns. Two covers put a caption in the bottom 23px, which a centre crop would cut.
    • PNG/JPG covers are illustrations with no text, so they are centre-cropped.
  • Worker processes: rendering SVG text on several libvips threads in one process crashed it (SIGSEGV or "SVG rendering failed" in 4 of 6 runs). The step renders in 3 worker processes, one image at a time in each, and passed 5 of 5 runs.
  • Default share image (Base.astro): it was the hero by its public path, which the move removes. It is now a 1200x630 JPEG of the moved asset (94 KB, generated with getImage) instead of the 1.7 MB PNG.

Before / after

Measured with Playwright Chromium (desktop rows at 1366x625), 390x844 at 3x, 4x CPU throttle and DevTools "Slow 4G" (562.5 ms RTT, 1.44 Mbps down). Cache was cold. LCP is the median of 3 runs.

Bytes are transfer sizes from a local static server that does no gzip, so HTML and CSS are counted uncompressed:

  • No scroll: everything the load starts, including lazy images inside Chrome's load-in distance.
  • Full scroll: after scrolling to the bottom.

Before is staged/issue-219 @ 5d685cf; after is bec5251. Both were measured back to back under one lock, on a shared 4-core box with a load average around 10. The cluster rows compare bec5251 (ClusterPage unchanged from 5d685cf) with 95ada62, also back to back under the lock.

Page LCP before LCP after Bytes, no scroll: before → after Bytes, full scroll: before → after
home /en/ 1,944 ms (text) 1,916 ms (text) 1,821,615 → 170,412 1,835,942 → 184,739
pricing /en/pricing/ 1,560 ms (text) 1,548 ms (text) 46,504 → 46,534 46,504 → 46,534
platform /en/platform/ 1,796 ms (text) 1,808 ms (text) 1,805,784 → 129,837 same
blog index /en/blog/ 2,708 ms (SVG cover) 1,808 ms (SVG cover) 142,787 → 148,368 18,826,110 → 1,167,398
post /en/blog/ai-agent-workbench/ (PNG) 9,956 ms (PNG) 1,700 ms (WebP) 1,685,249 → 120,959 same
post /en/blog/ai-expense-audit/ (PNG) 11,120 ms (PNG) 1,752 ms (WebP) 1,903,685 → 135,214 same
platform @1366x625 2,060 ms (h1) 1,928 ms (h1) 1,805,784 → 92,303 same
ai /en/ai/ @1366x625 10,620 ms (hero PNG) 1,964 ms (hero WebP) 1,804,910 → 91,429 same
cluster /en/ai-native-app-platform/ 1,688 ms (text) 1,680 ms (text) 903,617 → 230,655 1,182,014 → 282,059
cluster /en/ai-native-app-platform/ @1366x625 1,848 ms (h1) 1,828 ms (h1) 1,182,014 → 124,226 same

The small increases on pricing and the blog index (+30 B and +5.6 KB) are the longer og:image URL and the srcset attributes.

Build and dist, both with a cold Astro image cache on the same shared box:

before after
pnpm build wall time 87.5 s 108.4 s (90.5 s with a warm image cache at 95ada62)
astro image generation 0.6 s 6.3 s
og-images.mjs n/a 14.3 s (87 images, 3 workers)
dist/ 57,657,630 B, 1,082 files 68,235,078 B, 1,321 files

The dist increase is:

  • +6.05 MB of cover WebP variants (143 files);
  • +5.46 MB of share JPEGs (87 files: median 52 KB, max 152 KB);
  • +0.4 MB of hero variants;
  • −1.73 MB for the hero PNG.

Share-image checks

  • All 87 are 1200x630 JPEG. The step checks this, and a separate metadata pass confirmed it.
  • The glyph gate covered 70 SVG covers and 721 distinct characters. Its first run on this tree found U+21B6 (↶) and U+2298 (⊘), which is why Noto Sans Math is in the set.
  • The gate can fail:
    • A synthetic cover with Hangul and Thai reported those 6 code points. It did not report CJK, ✓ or an entity-encoded emoji.
    • On a copy of dist/, a page with og:image:width 1201 and a page with an unmapped share-image URL each failed the step. It exited 1 with 3 issues.
  • Looked at in full:
    • zh-Hans and zh-Hant (one shared Chinese cover per post): ai-wrote-your-app-dare-to-merge (🤔 ✓ →), beyond-agentforce-copilot-open-runtime (✓ ⊘), natural-language-to-app-metadata (curly quotes), airtable-omni-vs-governed-ai-app-platform (↶ ✓), self-hosted-ai-app-platform (1.6:1 viewBox).
    • ja and ko (these use the English cover-en.svg, like de, es and fr): why-ai-agent-pilots-fail-four-layers (①-④, U+3000), when-ai-agent-deletes-production-database (⛔ 💥 ✓).
    • The 10 PNG and 7 JPG illustrations, which ja and ko share.
    • All 87 rasters in six contact sheets.
    • No missing glyph anywhere.
  • No SVG cover today has Japanese or Korean text. Kana and Traditional Chinese are covered by Noto Sans SC (checked). Hangul is not, so a future Korean-text SVG cover fails the build with its code points named.
  • Emoji render in monochrome (Noto Emoji). The COLRv1 colour emoji font decodes, but cairo draws nothing for it.

Browser check

There are 36 page states (plus the cluster page, below): the six viewports (1366x625, 1440x750, 1512x830, 1920x945, 390x844, 360x800) × {PNG-cover post, blog index, home} × {light, dark}. Each has a first-screen and a full-page screenshot, taken after scrolling every lazy image into view.

  • Every state reports 0 unloaded images and 0 px horizontal overflow.
  • Close-ups:
    • the ai-expense-audit cover at 1920 and at 390@3x;
    • raster card thumbnails at 1366 light, 1920 dark, 390 dark and 360 light;
    • the 1366 thumb before and after, side by side;
    • the home hero and the marketing hero panel in light and dark.
  • /en/ai-native-app-platform/ at 1366x625 and 390x844, in light and dark, full page plus the reading-path section, all after scrolling every lazy image into view:
    • 0 unloaded images and 0 px horizontal overflow;
    • the card image boxes are unchanged (180x227 on a desktop, 356x200 on a phone);
    • a before/after close-up of the first card's image at each width.
  • No visible quality loss.

Gates (at 95ada62)

  • pnpm check: Result (150 files): 0 errors, 0 warnings, 0 hints. Run under the lock, VERDICT command-exit 0.
  • pnpm content:lint: ✓ content lint passed (339 files, 44 glossary terms checked)
  • pnpm content:lint --published: ✓ content lint passed (339 files, 44 glossary terms checked)
  • pnpm build: BUILD_EXIT=0, run under the lock. It includes:
    • svg-text-fit: OK — … (112 SVGs, 2244 text boxes, tolerance 2px, 1 allowlisted);
    • ✓ share images: 87 rendered at 1200x630 (5.2 MB, 13.3 s, 3 workers) for 879 pages; 125 bundled font files, no uncovered glyph;
    • ✓ content lint --dist passed (478 built blog pages checked);
    • ✓ content lint --dist links passed (93199 internal links in 880 built pages resolve to built files).
  • pnpm seo:smoke: SEO smoke test passed (870 HTML pages checked; 569 sitemap URLs against 301 noindexed pages; …)
  • pnpm svg:lint: svg-text-fit: OK — no text box crosses the shape under it or runs under a shape painted over it (112 SVGs, 2244 text boxes, tolerance 2px, 1 allowlisted)

New dependencies

All are build-time devDependencies. sharp was already a dependency.

  • wawoff2 (MIT): decodes WOFF2 to TrueType, because the FreeType in sharp's libvips cannot read WOFF2 or WOFF.
  • Six Fontsource packages (OFL-1.1): @fontsource-variable/arimo, @fontsource/cousine, @fontsource-variable/noto-sans-sc, @fontsource/noto-sans-symbols-2, @fontsource/noto-sans-math, @fontsource-variable/noto-emoji. These are the faces above; each one is needed by at least one character in the current covers.

Acceptance notes


Generated by Claude Code

🤖 Generated with Claude Code

https://claude.ai/code/session_01FeA1nwBz1ohH65dvffUGKr

claude added 30 commits October 5, 2026 12:39
… definition post (slot 2026-10-08)

Post 1 of the positioning wave: the executable business ontology defined
from the objectstack README and the Business Ontology concepts page,
quoted rather than paraphrased — the six-kind core mapped onto the
bundled CRM example, the four projections with a removal test to tell
core from projection, the three things it is not, where code goes, and
the four promises led by Executable. ObjectOS appears only as the
canonical T4 + E1. Eight locales (zh-Hant generated by gen:zh-hant), one
cover and two diagrams, in English text and in Chinese text for the
zh-Hans/zh-Hant pair.

Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FeA1nwBz1ohH65dvffUGKr
The post was written for the Thursday 2026-10-08 slot and dated for it,
but it was never merged that day, so the slot passed with the post still
unpublished. Under the publishing-cadence rules in AGENTS.md, a post's
`date` is its planned slot and merge time is publish time; the
forward-only rule protects the dates of posts that are already live,
and this one is not. The seat therefore moves it to the next free slot,
today, Thursday 2026-10-15, and merges it today.

Only the `date` frontmatter line changes, in the seven hand-maintained
locales (en, zh-Hans, ja, de, es, fr, ko), bare-date format kept and no
`updated` added; zh-Hant is regenerated from zh-Hans with
`pnpm gen:zh-hant`. No other file changes.

Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FeA1nwBz1ohH65dvffUGKr
Certification, AI governance tiers, MCP setup, templates, Cloud data,
Business Self-Managed, and two locale overstatements, each re-sourced
from objectos and objectstack docs at origin/main:

1. Enterprise card: "SCIM + compliance (SOC 2 / ISO)" -> "SCIM, compliance
   attestations" (license.mdx:72-73; security.mdx:142 says ObjectOS itself
   is not certified).
2. AI page: governance claims written per edition (license.mdx:42-43):
   a taste on Free, in-app AI on Team, AI action audit / guardrails /
   permission scoping from Business. The "0 bypasses" tile, the Ask audit
   item, the "never how it is governed" item, the edition table and the
   two FAQs now say that.
3. MCP page: the npx/--url snippet (no bin, no such flag) is replaced by the
   documented `claude mcp add --transport http` command and the
   `{"type":"http","url":".../api/v1/mcp"}` client config
   (connect-mcp.mdx:26-32, 85-94).
4. Templates page: the real eight-template catalog (todo, contracts,
   procurement, compliance, helpdesk, content, hr, project;
   templates.mdx:19-28) replaces the six invented starters; the decision
   table now carries the documented install / fork / publish paths
   (templates.mdx:33-62). llms.txt follows the page description.
5. Trust center: the lead quotes E1 (Cloud / self-managed split), the
   "Self-hosted" tile becomes "2 editions", the data-residency and
   licence-check items follow index.mdx:63-77 and license.mdx:150-152,
   and the model-training FAQ says only what the Privacy Policy says
   (legal.ts:414-415).
6. Pricing: Business Self-Managed (single node, 10-25 AI seats,
   $540 / AI seat / yr, bring your own model; license.mdx:41, 64-69) is a
   labelled option under Business, the billing note no longer says
   self-managed is Enterprise-only, and the Business governance line names
   the AI governance row. Team/Business headline prices untouched (#214).
7. zh home tile "数据不出网" -> "数据驻留" (= en "Data residency"); ja
   security tile "0 既定の外部テレメトリ" -> the en "Agent inherits user
   identity" tile.

zh-Hans changed with en; zh-Hant for marketing pages and pricing derives
at build time (registry deepS2t / toHant), so no generated file changes.

Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FeA1nwBz1ohH65dvffUGKr
Article tables on a phone no longer run each row onto one line. Cells
wrap, each column keeps a 6.5rem floor (three columns fit the 328px
column of a 360px phone), and overflow-wrap: anywhere keeps an
overflowing table at exactly its column count x 104px. The widest
English table goes from 2,035px to 328px at 360x800; the 12 tables with
four to six columns still scroll, to at most 625px, with the always-on
scrollbar and trailing-edge fade they already had.

Code blocks get the same two scroll affordances at every width, since
they overflow at desktop sizes too. The marketing decision table drops
its 760px floor on a phone and takes the same wrap, floor and
affordances.

Every scroll container is keyboard-reachable and named: post tables get
tabindex="0" and aria-labelledby their section heading (the article
title when no heading precedes them); code blocks keep Shiki's tabindex
and gain role="group" with a "Code (lang)" name; the decision-table
wrapper becomes a named, focusable group. group, not region, because a
region is a landmark and duplicate names fail axe landmark-unique.

Body images become links to their own asset: one tap opens the diagram
in the browser, where pinch zoom takes its smallest text past 11px. The
link's name is the image's alt text. No SVG changes; desktop geometry of
tables and images is unchanged.

Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FeA1nwBz1ohH65dvffUGKr
The marketing artifact code block is a scroll container; it gets
tabindex="0" and role="group" named by its section heading, like the
other scroll containers, which clears the last axe
scrollable-region-focusable finding on the marketing pages.

The security page's data-residency table had the same 620px floor as the
decision table, which put its third column ("Leaves boundary?") wholly
off-screen on a phone. It takes the same phone block: cells wrap, a
6.5rem column floor, tighter padding, type size and tracking, and the
scrollbar and trailing-edge fade if it still overflows. Its wrapper is
focusable and named by the section heading. Desktop is unchanged.

Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FeA1nwBz1ohH65dvffUGKr
…e build

Six visible defects from the 2026-10-06 site audit, plus the gate that
would have caught the broken links.

- Home "Latest insights": the cover's height attribute became CSS
  height:500px and beat aspect-ratio, cutting each 16:9 cover to a 344x500
  sliver. height:auto, a 16:9 box (the covers' own ratio), contain for the
  odd 16:10 cover.
- Blog index: the feature card spanned three rows beside three cards
  (~1470px) with ~670px of content, leaving an ~800px band. It now takes
  its own row, cover beside text from 1024px, the other three below.
- 404: a top-level 404.html (Cloudflare Pages otherwise runs the site as
  an SPA and answers every missing URL 200 with the language redirect) and
  a localized one per locale at <locale>/404.html, moved there by a small
  build hook. noindexed, no canonical, out of the sitemap.
- Cluster pages linked topic hubs not built in de/es/fr/ja/ko. They now
  link the hub really built (this locale's, else English), and
  content-lint --dist fails on any internal link to a page the build did
  not produce.
- Footer links and the drawer's close label come from the i18n copy.
- Alignment: .item > span; align-content:start on capability cards and
  next-page links; no min-height floor on template/compare cards; pricing
  cards on a subgrid; footer brand nowrap; the phone rule now clears the
  odd capability items' border; non-breaking space before card arrows.

Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FeA1nwBz1ohH65dvffUGKr
The browser check of the German drawer showed a visible English "Theme"
button beside the localized menu, the same defect as the footer links and
the drawer's "Close" label: hard-coded English in the shared chrome. The
theme toggle's name and tooltip, the language switcher's "Language:"
prefix, the three nav landmark names and the home page's topic-cluster nav
name move into the i18n copy for all eight locales. English output is
byte-identical.

Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FeA1nwBz1ohH65dvffUGKr
…e to 4.5:1, fix three near-misses

The 11 text lines the audit measured running past their cards, now inside
them with room left for wider system faces:

- self-hosted-runtime-en: the Runtime and AI Agent subtitles and the
  "Only approved minimum context is sent" line wrap to two lines (the
  Runtime line no longer runs over the green connector).
- ai-wrote-your-app-dare-to-merge cover (en and zh): the two captions beside
  the Merge pills wrap to two lines centred on the pill; "+8,142 -0" is
  right-aligned 22px inside its panel in both locales.
- ai-ontology-open-protocol cover-en: the two dark pills widen to 264px and
  the lock moves to the panel's lower right.
- tool-call-lifecycle-en: card 1 rewrapped ("change as Zhang, not as / an
  anonymous superuser").
- natural-language-to-app-metadata cover (en and zh): the spec card widens
  to 350px with 282px pills and sits 14px lower, so it no longer clips the
  subtitle's descenders; the arrows shorten to keep their gaps.
- one-definition-four-surfaces-en: the audit line is reworded to fit.

Contrast: defect-layers-en .none goes from #8a9c95 (2.7:1) to the house
.soft #5d736b.

Near-misses: review-gate-doors-en gives the gap card 12px more height and
the 29/16 numerals 6px more air; self-hosted cover paints the ObjectOS label
after the connectors (both locales), as the runtime diagram already does;
review-loop centres its card row (40px each side, both locales) so card 4's
shadow is no longer clipped.

Also objectos-automation-engine cover-en: the run's meta line ran 40px
under the "selected" tile, hiding "runAs: user"; it drops the "duration:"
label to clear the tile.

Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FeA1nwBz1ohH65dvffUGKr
…their alt text

Four English and five zh-Hans posts used AI-illustration webp images whose
alt text was only "Concept diagram" / 「概念示意图」. Each alt now says what
the picture shows (customer hub, read-only connection into a governed object
layer, tangled vs. modelled architecture, factory systems feeding an object
layer, legacy stack to structured model). The images are unchanged.
zh-Hant regenerated with pnpm gen:zh-hant.

Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FeA1nwBz1ohH65dvffUGKr
…ed over it

scripts/svg-text-fit.mjs measures every <text> in content/blog's SVGs and
fails when its box runs more than 2px past the filled shape under its centre,
or when a filled shape painted after it covers more than 2px of it both ways.
The #193 sweep only caught overflow that also dropped contrast, and a
tightest-rectangle measure sees an overflowing line's whole background as
its container; this gate measures each line against the shape it sits on.

It runs without a browser, since CI installs none: text is laid out with
Arial advance widths (scripts/lib/svg-text-metrics.mjs, generated from the
metric-compatible Liberation faces by scripts/gen-svg-text-metrics.mjs) and
the browser's text-box rules. Against headless Chromium on the same files
it agreed to within Chromium's whole-pixel rounding and found the same
crossings. A deliberate crossing goes in the documented allowlist; a stale
entry fails the gate.

Wired into `pnpm build` after content-lint --published (so CI's Build step
runs it) and exposed as `pnpm svg:lint`.

Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FeA1nwBz1ohH65dvffUGKr
Measured with DejaVu Sans (the default sans-serif on Debian/Ubuntu
desktops, about 15% wider than Arial), two of the refitted lines still ran
past their shapes: the spec pills on the natural-language cover widen to
290px (both locales), and tool-call-lifecycle card 1 keeps its original
words, rewrapped as 'change as "Zhang"—not / an anonymous superuser'.

The gate's header states its Chromium agreement as "2px", not "2.0px".

Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FeA1nwBz1ohH65dvffUGKr
…E1 attribution across 18 posts

Post-side update sweep for the 2026-10-03 positioning (work-order item 3):

- CTA: the stale `npm i -g @objectstack/cli && os start` block in 15 en
  posts and their authored locales becomes the objectstack README path —
  `npm create objectstack@latest my-app && cd my-app`, point a coding
  agent at it, `npx os dev --ui`. `how-to-self-host-objectstack` keeps
  `os start` (the production server), as does the mid-article operator
  line in `enterprise-ontology-race-open-vs-closed`.
- Descriptor: "the commercial production platform" / 「生产平台」 and the
  ja/de/es/fr/ko equivalents in ai-ontology-open-protocol,
  beyond-agentforce-copilot-open-runtime,
  enterprise-ontology-race-open-vs-closed, forward-deployed-engineer-tools
  and self-hosted-ai-app-platform become positioning[locale].descriptor
  verbatim ("the commercial runtime environment built on ObjectStack").
- Attribution: extend-existing-systems-with-ai scopes "ObjectOS runs in
  your environment" to self-managed (ObjectOS Enterprise) or the open
  runtime per E1 and objectos license.mdx (Cloud operates the model);
  objectos-action-tools attributes the Action mechanism (opt-in exposure,
  confirmation, permissions, audit) to the open ObjectStack runtime and
  positions ObjectOS with the descriptor. objectos-agent-permission-
  boundaries already attributed everything to ObjectStack; untouched.
- `updated: 2026-10-06T10:00:00+08:00` on every corrected file; no `date`
  touched, no slug changed. zh-Hant regenerated with gen:zh-hant.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FeA1nwBz1ohH65dvffUGKr
…ullet matches the FAQ

The bullet ended "Business data and prompts never leave your perimeter",
which is false once an external AI provider is configured — the FAQ
answer in the same post already says so. It now states the same claim
as the FAQ: business data stays in your database; prompts go only to the
AI provider you configure (on ObjectOS Cloud the model is operated for
you); permissions decide what the model sees. en and zh-Hans carried the
false sentence; ja, de, es, fr and ko had no sentence about prompts and
now carry the corrected claim too. zh-Hant regenerated.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FeA1nwBz1ohH65dvffUGKr
…re menus, tap targets

Contrast. --muted is #626c66 light (5.43:1 on --surface) and #8a968d dark
(5.42:1); the light accent moves to #3d58f0 (4.67:1 on --surface-sunk);
the focus ring is the solid accent (5.44:1 light, 7.06:1 dark) instead of
a 28% tint (1.5:1). The home, evidence, security, glossary and cluster
accents get dark-theme overrides (green #3fbf8f, blue #8fa3ff, amber
#e3a54a, red #f08a8a, all >= 6.9:1 on the dark surface) and the light
amber darkens to #93600f (5.34:1). The status pill reads theme tokens
(published 5.03:1 light, 6.19:1 dark on its tint), the pricing Team CTA
is --surface on --accent (7.06:1 dark, was 2.36:1), and code blocks use
Shiki's github-dark-default (comments 6.15:1, were 3.04:1).

Keyboard. The mobile drawer moves focus to its close button, makes the
page behind it inert, keeps Tab inside it and returns focus to the menu
button on close. The header dropdowns drop role=menu/menuitem for the
disclosure pattern (aria-controls, plain links, Escape returns focus,
focus leaving a menu closes it).

Tap targets. The hamburger, drawer close, header CTA and wordmark are
44px tall below 900px; footer links 24px (44px on a phone); the glossary
and post back links 44px without moving the text; card title links cover
their cards.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FeA1nwBz1ohH65dvffUGKr
…posts

`[ObjectOS](https://www.objectos.ai)` in business-app-in-16k-tokens (1 per
locale) and forward-deployed-engineer-tools (2 per locale) pointed at the
root language-redirect page, so every click took a meta-refresh hop before
landing. Each locale now links its own home (`/en/`, `/de/`, ...); zh-Hant
regenerated with gen:zh-hant (it rewrote `/zh-Hans/` to `/zh-Hant/`). 24
links in 16 files; no `date` or `updated` touched — a link target is not a
substantive revision.

Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FeA1nwBz1ohH65dvffUGKr
…hubs, OG

The site-wide audit on #219 measured, on a build of this tree's base:

- Sitemap: 190 noindexed fallback pages listed. The sitemap filter now
  reads each built page and lists it only when it is indexable and its own
  canonical (astro.config.mjs) — one rule for every fallback producer
  instead of a second copy of each registry.
- Pricing/privacy/terms/refunds in ja/de/es/fr/ko rendered the English
  copy, self-canonical and in every locale's hreflang (20 pages); security
  in those five locales inherits five English sections. They now follow
  the marketing fallback rule through src/lib/fallback.ts: noindex,
  canonical to /en/, hreflang only for authored locales, the switcher
  still offering every locale. Authored-ness is derived, not listed: a
  copy table entry that is the English object is a fallback, and a
  security locale that inherits any English field is one.
- Base takes `contentLocale`: a fallback body gets `<main lang="en">`
  (marketing, glossary, pricing, legal; 210 pages had English under
  lang="de"/"ja"/...), og:locale follows it. The security page marks each
  inherited English part `lang="en"`; glossary fallbacks mark the
  reader-locale article links with their locale.
- The 35 non-English cluster pages are one template per locale: noindex,
  canonical to the English cluster, out of hreflang and llms.txt.
- Pillar pages: every footer links the five clusters (FooterClusters),
  and each post a cluster lists as its reading path links back to it.
- Topic hubs listing fewer than 3 posts (51) are noindexed and out of
  hreflang; every hub gets BreadcrumbList and ItemList JSON-LD and a
  "<topic> · Articles · ObjectOS" title.
- Root `/`: absolute hreflang, canonical /en/, no noindex beside the
  meta refresh. SecurityPage links its post with the trailing slash.
- public/_headers caches the hashed /_astro/* files as immutable.
- og:locale uses en_US/zh_CN/zh_TW/...; og:image:alt everywhere and
  width/height when known; Organization gets @id and sameAs (the GitHub
  organization only); SoftwareApplication gets offers from the pricing
  page's own dollar plans.
- Titles: marketing titles that already name ObjectOS no longer append
  " · ObjectOS" (27 indexable doubles); the blog index takes the home
  page's heading for it instead of "Articles · ObjectOS".

Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FeA1nwBz1ohH65dvffUGKr
… lint warns at 60/160

seo:smoke reads every built page once and asserts invariants over the
whole build instead of per-registry URL lists, so a new fallback producer
that gets the rule wrong fails without this script learning about it:

1. every sitemap URL is a built, indexable, self-canonical page;
2. every canonical and hreflang target is indexable and self-canonical,
   and a page that canonicalizes elsewhere is noindexed;
3. every block of prose in <main> is written in the language its nearest
   `lang` declares (scripts/lib/text-lang.mjs: script for CJK, distinctive
   stopwords for en/de/es/fr; too little text is "cannot tell", never a
   pass), with a per-locale control so it cannot measure nothing;
4. no internal link, canonical, og:url or hreflang redirects (missing
   trailing slash, `/`, `.../index.html`, a `_redirects` source).

Plus the root page (canonical /en/, absolute hreflang, no noindex beside
the refresh), `_headers` (rule present, every /_astro/ name hashed),
og:locale values, og:image:alt, the Organization @id/sameAs, hub
BreadcrumbList/ItemList, and the clusters: English in llms.txt and the
sitemap, the templated locales noindexed and canonical to English.

On the base build this reports 190 noindexed sitemap URLs, 215 pages
whose <main> is in another language than declared (25 indexable), and 9
redirecting link targets on 24 pages; on this branch's build it passes.

content-lint: a title wider than 60 columns or a description wider than
160 (CJK counted double) is a warning. Posts first published before the
check are counted in one summary line; `--snippets` lists them all.

Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FeA1nwBz1ohH65dvffUGKr
…uled

The home page's SoftwareApplication listed Offers derived from the pricing
page ($20/$45 per AI seat billed annually). That price is under a pending
maintainer decision (www#214: $20/$45 annual on www vs $24/$54 monthly on
docs), so structured data must not assert either until it is ruled. The
rest of the item-10 changes stay: the publisher still references the
Organization @id.

Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FeA1nwBz1ohH65dvffUGKr
…ith srcset

The post header cover is the LCP element on a phone, and 80 posts used a
1.5-1.9 MB PNG for it (ai-agent-workbench: 9.96 s LCP at 390x844, 4x CPU,
slow 4G). Card thumbnails on the blog index, topic hubs and the home page
downloaded the same full-size file, about 18.8 MB for one scroll of
/en/blog/.

A PNG or JPG cover now goes through <Image> as WebP with a width list and
`sizes` per slot:

- post header: 480-1600w, `(min-width: 908px) 860px, calc(100vw - 2rem)`;
  still eager with fetchpriority=high, since it is the LCP element;
- ArticleCard: 360-1280w, sized for the three-column grid, the two-column
  tablet grid and the phone column; the feature card's cover is eager, as it
  opens the blog index above the fold;
- the home page's latest-posts grid: 360-1080w.

The ai-agent-workbench cover at 1280w is 58 KB (the original is 1.62 MB);
a 1:1 crop of the 1280w WebP against the original resized to 1280 shows no
visible loss (PSNR 39 dB). SVG covers are a few KB and scale themselves, so
they keep a plain <img> on the original.

Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FeA1nwBz1ohH65dvffUGKr
209 posts (41 English) named an SVG as og:image, twitter:image and
BlogPosting.image, and X, LinkedIn, Facebook and Slack draw no card for an
SVG; the PNG-cover posts named a 1.5-1.9 MB 1672x941 file.

`scripts/og-images.mjs` runs in `pnpm build` after `astro build` and
renders one 1200x630 JPEG per cover into dist/og/blog/<post>/. The URL rule
lives in scripts/lib/og-image.mjs, shared with BlogPost.astro, which now
names that file in og:image, twitter:image and BlogPosting.image, with
og:image:width/height 1200x630 and the post title as og:image:alt. The URL
carries a hash of the cover's bytes, so an edited cover gets a new URL the
networks re-fetch. The step is driven from the built HTML and fails the
build when a page names a share image no cover maps to, or declares a size
that is not the file's.

SVG text: sharp draws it through librsvg/Pango/fontconfig, which use the
machine's fonts. The covers ask for -apple-system, Segoe UI, Inter, Arial
and Menlo, and CI's ubuntu-latest has no CJK font, so the Chinese covers
would have shipped as rows of boxes. The rasterizer now sees only bundled
OFL faces (scripts/lib/og-fonts.mjs): Arimo and Cousine (the Arial and
Courier New metrics `pnpm svg:lint` measures with), Noto Sans SC, Noto Sans
Symbols 2, Noto Sans Math and Noto Emoji. Fontsource ships them as WOFF2,
which the FreeType in sharp's libvips cannot read (every glyph drew as a
box), so wawoff2 decodes them to TrueType in a temp dir that a generated
fonts.conf points at. Before rendering, every character in each SVG cover's
text is checked against the faces' cmaps; one that no face has fails the
build with the cover, code point and line. On this tree that check first
found U+21B6 and U+2298, which is why Noto Sans Math is in the set.

An SVG cover is drawn whole at 1120x630 and widened by repeating its edge
columns (two covers put a caption in the bottom 23px a centre crop would
cut; no cover draws in its outer 40px). A PNG/JPG cover is an illustration
without text and is centre-cropped. Rendering SVG text on several libvips
threads in one process crashed (SIGSEGV or "SVG rendering failed" in 4 of 6
runs), so SVGs are rendered in three worker processes, one at a time each.

Result: 87 images, 5.2 MB, 12 s of the build; zh-Hans, zh-Hant, ja and ko
rasters looked at, no missing glyph.

New devDependencies, all build-time only: wawoff2 (MIT) and six Fontsource
packages (OFL-1.1). sharp was already a dependency.

Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FeA1nwBz1ohH65dvffUGKr
…, dim it in dark mode

The 1,732,355-byte hero PNG was loaded eagerly by the 120 marketing
evidence pages, where it renders in a 415x340 box, and lazily by the home
page; it was also the default og:image of every page without its own.

- src/assets/objectos-home-hero.png (moved from public/images/) goes
  through <Image> as WebP with srcset (400-1586w). The box crops the 1.6:1
  image with object-fit: cover, so `sizes` asks for the drawn width, not the
  box's: 545px on a desktop, about 1.2x the box on a phone.
- Marketing evidence pages: measured at all six AGENTS.md viewports, the
  hero's top is at 172px on a desktop (the LCP element there) and at
  713-781px on an 800-844px phone, so it is in the first screen everywhere
  and stays eager, now with fetchpriority=high. Home: it sits under the
  hero and the video at every viewport, so it stays lazy.
- Dark theme: the white render was a glaring block on the dark page. There
  is no dark variant of the render, so the image itself is dimmed
  (brightness 0.6, contrast 1.12; three strengths compared side by side in
  the hero panel). No layout CSS changes.
- Base.astro: the default share image was the 1586x992 PNG by its public
  path, which the move removes. It is now getImage() of the same asset, a
  1200x630 JPEG (94 KB). og-images.mjs now checks every og:image that is a
  file of this site against the size its page declares, the default
  included, and fails on one that is not in dist/.

The hero component code no longer reads the import's width, so Astro drops
the 1.7 MB original from dist/ (nothing references it).

Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FeA1nwBz1ohH65dvffUGKr
…m at

A standard card's 16:10 thumb crops the 16:9 cover with object-fit: cover,
which draws the image 1.11x the card's width: a 343px card at 1366 needs
381px of image, and `sizes` said 360px, so the 360w candidate was
stretched and visibly soft next to the original. `sizes` now names the
drawn width (400px on a desktop, 55vw - 2rem on the two-column tablet
grid; the phone's 2:1 thumb is drawn at the card's width) and the list
starts at 400w.

Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FeA1nwBz1ohH65dvffUGKr
The reading-path cards on the 40 cluster pages still used the original
cover: /en/ai-native-app-platform/ loaded four original JPGs (240-370 KB
each) into 180px-wide slots. A raster cover now goes through Image as
WebP, 400-1080w. `sizes` is the drawn width, the same way as the
ArticleCard thumbs: above 640px the image is a 180px (150px at 641-900px)
column as tall as the card's text, measured 190-227px, and object-fit:
cover fills it by height, so a 1.6:1 or 16:9 cover is drawn 304-364px
wide; `sizes` says 400px. On a phone the card stacks and the 16:9 image
is drawn at the column width. SVG covers keep the plain img.

Only the frontmatter (the import and the width list replacing coverSrc)
and the card markup change; no CSS.

Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FeA1nwBz1ohH65dvffUGKr
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

2 participants