web devSun 15 Nov, 09:00

Sources8 newsletters
  1. Runs on Sundays at 09:00 London time.
  2. It is shown its own papers of the last 8 days, to know what it has already said.
  3. Sonnet writes the paper, opening a few of the linked articles.
  4. Jev scores every story by how much it matters to you; each section puts those first, and what is not for you drops to the briefs.

8 newsletters

Every new issue, each delivered once, looking back up to 10 days.

๐Ÿง‘โ€๐Ÿ’ป Web dev weekly, November 15, 2026

Worth your time6

Must read

๐Ÿ”’ Vite critical RCE: update if your dev server is exposed

Unescaped input in two Vite virtual modules lets a plain HTTP request inject code. Vite trusts it and runs it on the server, and files outside the server.fs allow list can be read. It needs no login. Only dev servers exposed with --host or server.host are affected.

Why it matters: Phone testing with Expo web or Vite on a shared or office network is exactly this setup.

Takeaway: Upgrade Vite (vite-plus 1.5.1). Until then, don't bind the dev server to the network.

Read it fully? No, the summary covers it.

Must read

๐Ÿงฉ Chrome's Modern Web Guidance 1.0 steers coding agents to modern HTML/CSS

It is a set of skills that teaches agents current web platform practice, tuned to your Baseline browser targets. It covers accessibility, performance, security (CSP, sanitization, passkeys) and native alternatives to custom JS. Version 1.0 adds a modernize workflow that statically finds safe upgrades in HTML, CSS and JS. It installs with npx modern-web-guidance@latest install, or through the Claude Code, Codex, Cursor and Copilot plugin marketplaces.

Why it matters: Agents tend to write dated web code. This is a cheap way to fix that.

Takeaway: Run the install command in a project and try the modernize workflow on one folder.

Read it fully? No, the summary covers it.

AI & software engineering

๐Ÿ”‘ Coding-agent 'don't ask again' approvals leak secrets to GitHub

Apiiro found 2,539 secrets in 1,994 public repos, saved in local agent settings files. Check your .gitignore for them.

Must read

๐Ÿค– Anthropic ships Fable 5.2 and Mythos 5.2, 20% cheaper

Fable and Mythos are one model with different safeguards: Fable is public, and Mythos is gated for cyber and life-science teams. Generation is up to 30% faster. In Advisor mode, Haiku, Sonnet or Opus escalate hard problems to Fable. In Orchestrator mode, Fable delegates to the smaller models. Anthropic reports 49.1% on FrontierCode 2.0 against 45.5% for Opus 5.5, and says Fable matches Opus at about half the cost per task. These are self-reported benchmarks. Anthropic still recommends Opus 5.5 as the daily default.

Why it matters: Routing between big and small models is now built in, which affects agent costs.

Takeaway: Keep Opus as the default. Use Fable for long, ambiguous jobs.

Read it fully? No, the summary covers it.

Must read

๐Ÿ”’ Node.js security releases: 7 high-severity HTTP/2 and TLS fixes

The November release fixes 7 high, 6 medium and 3 low issues. Most are HTTP/2 use-after-free and leak bugs. Others are TLS problems: one leaks server heap into HTTPS bodies, and another lets TLS 1.2 renegotiation bypass per-host client-certificate checks. Several were found with Claude, and one with OpenAI-linked Trail of Bits. Undici fixes are included for v24 and v26.

Why it matters: If you run Node servers with HTTP/2 or TLS (SSR, BFFs, APIs), this is a patch-now release.

Takeaway: Move to 26.13.1, 24.22.3 or 22.23.5. End-of-life lines get no fix.

Read it fully? Yes, skim the CVE table for the features you use.

Web & mobile

๐Ÿ› ๏ธ Storybook 11 beta: CSF 4, React component metadata, agent skills

CSF 4 adds type-safe story factories, and React metadata moves to react-component-meta. A new headless storybook tools CLI runs docs, tests and review, and official ui-workflow and component-docs skills ship for agents. Try npx storybook@next upgrade.

โญ Must read1

๐Ÿงฑ shadcn and Vercel's UI SDK: headless behaviors built for agents

UI SDK is a low-level library of behaviors: roving focus, typeahead, popups, drag and drop, with the ARIA handled for you. It has 50+ unstyled components and a definePrimitive API. A createAgent helper exposes components to agents, and WebMCP tools expose them to a browser's built-in agent. It is in beta.

Why it matters: It is a possible base for accessible design systems that doesn't tie you to one framework.

Takeaway: Try the 'Build a Component' guide before you commit. APIs may change.

Read it fully? Yes, if you maintain a component library.

๐ŸŒ Web & mobile8

๐Ÿ”Š React Native Audio API 1.0 is stable

Software Mansion's Web Audio API for RN adds spatial audio, and FFmpeg is now opt-in.

๐Ÿง  CSS system colors can fail contrast pairs

HighlightText on Highlight hits only 1.98:1 in Chrome on Android dark mode. Aaron's study shows which pairs hold up.

โšก X rebuilds its web app on React, Relay, Vite and Fastify

Couldn't open the article (403), summary from the newsletters. X is replacing its client-rendered app with streaming SSR on Fastify and Node 26 worker threads. They fixed event-loop stalls and startup slowdowns from hundreds of tiny JS files. Lab tests show profiles showing posts much sooner.

In brief

๐Ÿงช Next.js 16.4.5 fixes a cross-tenant cache disclosure

This is a new fix in 16.4.5 and 15.5.29, which patch seven issues including a high-severity cache leak between tenants.

๐Ÿชถ Parcel 3 alpha, Playwright 1.65, react-virtualized 10 and more

Playwright 1.65's test runner can now run under Bun, and has a new failures reporter.

๐Ÿ“ฑ Expo: an agent-friendly iOS keyboard extension case study

Five web developers built an iOS keyboard extension with Expo.

๐Ÿ“š react-virtualized 10.0 is back, in TypeScript

It is rebuilt on function components and needs React 18+.

๐Ÿงต Node 'green threads' PR: 1,000 workers in ~1GB

An experimental lightweight worker type shares a few OS threads. It uses about 1GB for a thousand, against 6.7GB for regular workers.

๐Ÿค– AI & software engineering6

โšก Google's Gemini 4 Hydrogen: fast model with Antigravity Orchestrator

Summary from the newsletter. Google says it is 2.5x faster than Gemini 3.8 Flash and scores 57% on Terminal-Bench 4.1 against 18.2% for Flash (company benchmarks). Antigravity's new Orchestrator switches between Carbon and Hydrogen mid-session, or has Hydrogen agents build and verify in parallel.

๐ŸŽฏ Gemini Decisions API: probabilities instead of text

Summary from the newsletter. It scores up to 1,000 bounded questions (yes/no, up to 255 choices, rubrics) and returns confidence scores. It suits routing, policy checks and tool selection. OpenAI and TypeSafe's Jev offer similar models.

In brief

๐Ÿงฐ Claude adds Meetings, Sheets, Whiteboards; interactive connectors reach all plans

Also this week: JPMorgan describes staged Claude Code rollouts where agents can't approve their own work.

๐Ÿ” OpenAI Daybreak Red Team and Defense Factory

Agent suites that find exploitable code flaws and propose Codex patches. Red Team is in early access, and Defense Factory is in the Codex Daybreak plugin.

๐Ÿ—๏ธ AWS releases an official local AWS emulator

A Rust emulator for hundreds of AWS APIs, aimed at agents, and built by agents.

๐Ÿ“ˆ JetBrains study: double the tokens, about 10% more commits

A correlation across 404 engineers over five weeks, not proof of cause.

๐Ÿ—‘๏ธ Skipped

  • Nvidia's Jensen Huang exit: covered by nearly every AI newsletter. It is business news, not worth your time.
  • Anthropic IPO, OpenAI revenue and funding rounds: finance, no engineering impact.
  • Google AI glasses, Ring drone, World ID, Adobe Firefly: consumer or sponsored.
  • NativeWeekly: no issue this week.
  • MyClaw agent case studies: mostly vendor-reported numbers.