Meta open-sources its React design system

Hi there, this is your daily ☕️ Devshot.

In today's Devshot:

⚛️ Meta open-sources its React design system

⚙️ Netflix rebuilds Conductor for scale

⚡ Lambda SnapStart supports container images

🕸️ Cilium 1.20 adds Gateway API ExternalAuth

📦 Jotai 3.0 goes ESM-only

Plus: 🎁 8 other news you might like, 🧰 6 tools, and 📚 5 papers.

⚛️ Meta open-sources its React design system LINK

Meta released Astryx in beta, an open-source, MIT-licensed React design system built internally over eight years that ships more than 150 accessible UI components alongside dedicated CLI and MCP tooling.

The core package distributes pre-compiled, collision-free CSS from StyleX plus typed React components, requiring no compiler setup and interoperating with Tailwind, CSS Modules, or vanilla stylesheets via a centralized design token layer.

A CLI eject command copies a component's full source into your repo for deep customization past exposed internals, though you then own the maintenance; Astryx requires React 19 and exposes an MCP endpoint for AI agents.

⚙️ Netflix rebuilds Conductor for scale LINK

Netflix rebuilt its Conductor workflow orchestration engine to handle scale, now running about 420 million workflows a month across 150 apps while raising the maximum workflow size from roughly 2,500 to 30,000 tasks.

Conductor 4.0 separates workflow metadata from task and user data, loading only the tasks needed for the next decision via a lightweight blueprint, which cut p99 evaluation latency by about 40% versus loading full state into memory.

The redesign drops locking by storing pending and terminal task states separately and reconciling in the app layer, moving evaluation off the synchronous path into Timestone queues and pushing failed lock attempts from ~2,700 per interval to near zero.

⚡ Lambda SnapStart supports container images LINK

AWS extended Lambda SnapStart to container-image functions, letting teams packaging heavy dependencies like pandas and numpy up to 10 GB avoid the multi-second startup penalty they previously accepted when abandoning the 250 MB zip limit.

SnapStart snapshots the initialized environment at deployment and resumes from it, cutting cold starts to sub-second, but AWS base images outside Java 11+, Python 3.12+, and .NET 8+ must add a Dockerfile LABEL or runtime hooks or publishing fails.

Serverless Framework 4.42.0 added support within a week, now rejecting ephemeral storage above 512 MB with SnapStart before deploy; the feature works in all commercial regions except New Zealand and Taipei via API, console, CLI, and CDK.

🕸️ Cilium 1.20 adds Gateway API ExternalAuth LINK

Cilium 1.20 shipped with a broader Gateway API layer, jumping from v1.4 to v1.6 and adding ExternalAuth so authentication and authorization run before requests reach applications, alongside CORS filters, ListenerSets, and TCPRoute/UDPRoute for non-HTTP traffic.

The upgrade lets platform and security teams route more north-south traffic through one API, positioning the CNI many already run as a replacement for standalone Ingress NGINX deployments for traffic management.

The release rounds out other datapath work, including netkit auto mode that probes each host and falls back to veth on older kernels, beta datapath plugins from Google, and beta IPv6 support in AWS ENI IPAM.

📦 Jotai 3.0 goes ESM-only LINK

Jotai, the atomic React state library from Poimandres, shipped its v3.0.0 stable release as an ESM-only package, dropping the CommonJS build with no new APIs and, the team says, no breaking changes.

v3 removes the UMD and SystemJS bundles, older Node, React and TypeScript support, targets ES2020, reads NODE_ENV directly, and swaps Rollup for other tooling, ending Daishi Kato's long-documented dual package hazard for a leaner core.

Several deprecated APIs are gone, atomFamily now lives in jotai-family with an added atomTree helper, and loadable users move to jotai-eager, so migration is mainly a mechanical import swap for teams that heeded v2 warnings.

Other news you might like

• I made a build profiler to understand Bun’s compile timesLINK

• Open-Source Project Brings Full iOS 27 Virtualization to Apple SiliconLINK

• Self-taught tinkerer demos working room-sized Cold War-era supercomputer with vacuum tubes — takes 15 minutes to warm up and smells like burning dust, 8-bit design uses 460 recycled 1950s Soviet tubesLINK

• One Decade of Rustls: Evolution, Benchmarks, and Future RoadmapLINK

• cutting a Rust GPUI launcher’s idle memoryLINK

• What Is an Agent Harness? The Architecture Behind Claude Code, DeepSeek Harness, and Hermes AgentLINK

• ESP32 Bit Pirate: Bridging Modern Microcontrollers and Browser-Based Hardware DebuggingLINK

• Intel revives One Mono font after brief retirement during open-source purge — typeface built to fight coder eyestrain gets reprieveLINK

🧰 Trending tools

SKI: adds voice control to AI coding agents like Claude Code and Codex, letting you dictate commands and hear responses hands-free while codingLINK

Humalike x Hermes: gives AI agents turn-taking, timing, and memory APIs to decide when to speak, wait, or interrupt, useful for game characters and roleplay.LINK

Kilo Code for JetBrains: an open-source AI coding agent for JetBrains IDEs that lets you build, ship, and iterate on code with flexible model choices.LINK

Mastra Factory: an open-source, self-hosted web environment where persistent coding agents handle issue intake, planning, implementation, and pull request reviews across your repositories.LINK

Anysite.io: a B2B data layer for AI agents that returns targeted company and contact lists via MCP or REST, no scrapers or queries needed.LINK

Meridian: open-source, locally-run activity tracker that logs your work with context, revealing where time went and the blockers behind it.LINK

📚 Trending papers & reports

Four-day workweeks can survive long-term at software firms if teams keep redesigning how they coordinate and meet, with employees defending the shorter week through voluntary tweaks even under ownership changes and economic pressure.LINK

Agent instruction tuning lets you fix workflow-building AI by editing its guiding prompts instead of retraining, catching cases where one fix quietly breaks another before the change is kept.LINK

Vendor-tuned coding agents deliver no reliable accuracy edge over neutral third-party setups running the same model, yet still solve each task 1.2 to 1.6 times more cheaply, so the pairing matters for cost, not correctness.LINK

Command-line access lets enterprise AI agents finish office tasks up to ~24 points more accurately than purpose-built tools while using as much as ~72% fewer tokens, meaning simpler setups often beat elaborate ones.LINK

Test-case generation for code spends heavy computation only on the tricky programs it's unsure about, picking better AI-written code while cutting cost versus checking every case the expensive way.LINK

See you tomorrow for a new dose of ☕️ Devshot!