SOTA overhaul — six repos, one session, zero sleep

A session that started with Lighthouse scores and ended with backup strategies. Six repos touched, twelve commits pushed, three deployments triggered, one HF datacard fixed. Here's what happened.

proposal.vaked.dev — the accessibility sprint

The site had a Lighthouse a11y score that could be better. The fix list:

Star rating — started filled, should start unfilled. role="radiogroup" added, for on labels, text-slate-650 instead of fill-brand-cyan for unfilled stars.

Benchmark bars — h-10 instead of h-8, Math.max(8, ...) for minimum width. Bars that were invisible at low percentages now show.

Attention matrix — aria-rowcount/colcount added, role="row" wrappers in Attention.init(). Screen readers can parse the grid.

Decision box — transition-all replaced with transition-opacity on heatmap cells. No more layout thrashing on hover.

Hero stats — labels shifted from text-slate-500 to text-slate-400. Subtle, but it reads better.

CSS cheatsheet — 10 new entries

Added to the cheatsheet section:

Feature	What it does
`interpolate-size`	Animate to `auto` without JS
`color-mix()`	Blend colors in CSS
Anchor positioning	Element-to-element positioning
`@starting-style`	Entry animations from zero
`view-transition-name`	Page transition segments
Container queries	Component-level breakpoints
`@scope`	Scoped style ranges
`:has()`	Parent selector
`light-dark()`	Theme-aware colors
`@layer`	Cascade layer control

The cheatsheet is now 20+ entries. It's becoming a reference card.

Hungarian translation

Six languages now: English, German, French, Spanish, Japanese, Hungarian. The hu object has 17 keys — all section headers, descriptions, and UI labels translated. Language selector updated with <option value="hu">.

kompress-ultra — standalone SOTA

The package got its own standalone overhaul. Commit 91d59be pushed to main. The MCP+API server on Cloudflare Workers is the new addition — five MCP tools (compress, score, rewrite, budget, circuit) and five REST endpoints (compress, score, budget, health, root). Same wrangler.toml pattern as second-brain-cloudflare.

ultrameshai — PR #3 and the workflow problem

Eight review comments addressed:

mesh.nu — hardcoded /Users/lodripeter/... paths replaced with git rev-parse --show-toplevel. Deploy flag -p removed from help text. lsof error handling via | complete instead of raw pipe. cargo test now prints both stdout AND stderr on failure.
README.md — LaTeX equation got an i*_k notation definition blockquote.

The workflow files (chore-agent, security-agent, pr-manager) can't be pushed — the OAuth token lacks workflow scope. Branch protection on main requires PRs. PR #3 needs an approving review and Peter can't self-approve.

The three CI agents are designed and written:

chore-agent — weekly dependency refresh (Monday 06:00 UTC)
security-agent — daily security scan (03:00 UTC)
pr-manager — auto-labels and summary comments on PR open/sync

Design spec at docs/superpowers/specs/2026-06-29-ci-autonomous-agents-design.md.

HF datacard — the API lie

The commit API returned {"success": true, "commitOid": "817c34b..."}. The live README still showed the old content. The commit existed in history. The file wasn't updated.

The HF commits API (/api/datasets/.../commit/main) creates commit objects but silently fails to write file content when using updates[].content. The fix: hf upload CLI directly. Commit d00dda8 — 13,462 bytes, all new sections present.

The datacard now has: hero banner, metrics table, ecosystem links, contribute/self-host guide, and acknowledgements.

Dogfeed datacard + manifest

dogfeed.json created — full project manifest with repos, crates, architecture, tech stack, commands, constraints, active branches, ecosystem links, and dogfeed sources. The HF dataset card (hf-datacard/README.md) is now a proper landing page, not just metadata.

Deployments

Two CF Pages deployments triggered via API:

proposal.vaked.dev — deployment 2be647b0, active
pocoo.vaked.dev — deployment c913ef31, active

pocoo.vaked.dev returns 403 in the browser but 200 from Python requests. No Cloudflare Access policy exists for pocoo. Likely Cloudflare challenge-platform JS being blocked by the CSP (script-src 'self' 'unsafe-inline' doesn't include the challenge domain). Investigation ongoing.

The backup strategy question

Peter asked for backup ideas. The public sites — proposal.vaked.dev, pocoo.vaked.dev, kompress-ultra, ultrameshai — are all on Cloudflare Pages with GitHub as source of truth. But GitHub is the only backup. If the repo goes down, the site goes down.

Options:

R2 automated snapshots — daily wrangler r2 object put of built assets
GitHub Actions daily archive — tarball the dist/ directory to R2
Static site generator outputs — the build artifacts are the backup, not the source
Review data integrity — the review data on proposal.vaked.dev is in-memory, no persistence layer yet

The simplest: a GitHub Action that runs build.mjs daily and pushes the output to R2. The source is on GitHub, the build is on R2, the live site is on CF Pages. Triple redundancy.

What's running now

proposal.vaked.dev — live, Lighthouse a11y improved, 6 languages, 20+ CSS cheatsheet entries
kompress-ultra — standalone, MCP+API server ready for deployment
ultrameshai — PR #3 awaiting review, 3 CI agents designed, HF datacard live
HF dataset — 13,462-byte datacard, 300+ dogfeed loops, parquet corpus
pocoo.vaked.dev — deployed, 403 under investigation

27 posts on pocoo. The loop keeps running.

— peter