☕️ Meta rewrites React Compiler in Rust

Hi there, this is your daily ☕️ Devshot.

In today's Devshot:

⚛️ Meta rewrites React Compiler in Rust

⚡ GitHub makes issue navigation 5x faster

🧩 VS Code 1.130 rearchitects its AI agent

📱 Claude Code adds iOS Simulator support

🐍 PyPI rejects new files after 14 days

Plus: 🎁 12 other news you might like, 🧰 6 tools, and 📚 5 papers.

⚛️ Meta rewrites React Compiler in Rust LINK

Meta has merged an experimental Rust port of the React Compiler, the tool that auto-memoizes components and hooks, into the main React repository, targeting faster builds and tighter frontend toolchain integration.

As a drop-in Babel plugin the Rust version runs ~3x faster than the TypeScript original, with isolated transformation logic reaching up to 10x, and all 1,725 test fixtures passing via arena allocation and index-based structures.

Linking the compiler directly into Turbopack rather than running it as a Babel transform cut compilation by over 40% on Vercel's v0, with experimental support arriving in Next.js 16.3 and the public Babel-like API unchanged.

⚡ GitHub makes issue navigation 5x faster LINK

GitHub rebuilt the navigation architecture behind Issues to cut perceived latency, raising the share of instant navigations from 4% to 22% by shifting more work to the client using caching, prefetching, and service workers.

The system takes a local-first, stale-while-revalidate approach, rendering data immediately from IndexedDB and in-memory caches while background syncs fetch fresher information, plus preheating that populates likely cache entries before users request them.

Navigation latency dropped sharply across the distribution: P10 fell from ~600ms to 70ms, P25 from 800ms to 120ms, and median from 1,200ms to 700ms, with P75 and P90 seeing smaller gains.

🧩 VS Code 1.130 rearchitects its AI agent LINK

VS Code 1.130 rearchitects agent sessions around an agent host process that runs harnesses like Copilot, Claude, and Codex over the Agent Host Protocol, letting sessions live in separate processes rendered from multiple windows.

The preview Agents window gains file-level diff stats showing live insertion and deletion counts, a compact multi-file diff gutter, single-line quick chat rows, and worktree isolation now supported across all agent harnesses.

Assisted permissions let the model evaluate each tool call's risk during long-running tasks to auto-run or request approval, while Git diffs now render clickable file links that strip mnemonic prefixes to open the correct file.

📱 Claude Code adds iOS Simulator support LINK

Claude Code's Mac desktop app now supports Apple simulators, letting the tool build, install, launch, and interact with your app across iOS, iPadOS, visionOS, or watchOS simulators directly from within the app.

Opening an existing project surfaces a new iPhone-icon button that spawns a simulator pane on the right, and clicking the simulator name lists every installed option, with multiple simulators runnable at once though performance suffers.

Because the simulator runs inside the Claude app rather than through computer use, no extra windows clutter your workspace, and with live access to a running app it can locate UI bugs, fix them, retest, and commit.

🐍 PyPI rejects new files after 14 days LINK

PyPI now blocks uploads of new files to any release that is more than 14 days old, a measure aimed at stopping attackers from poisoning old releases when a project's publishing tokens or workflows are compromised.

The change followed the March 2026 compromise of the popular LiteLLM and Telnyx packages, which stemmed from a mutable reference in those projects' use of the Trivy GitHub Action, reviving a discussion first raised in PEP 740.

Concerns that projects add newer Python support to published releases proved minor after querying wheels for the top 15,000 packages, since only 56 had shipped a 3.14-compatible wheel beyond the 14-day window.

Other news you might like

• How to use Chrome’s Modern Web Guidance to prevent AI agents from writing legacy frontend codeLINK

• Building Sierra’s MCP Gateway: An engineering icebergLINK

• Eval Engineering Skill: Build Evals From Repo Context and TracesLINK

• What Is Terraform Vendor Lock-In and How to Avoid ItLINK

• Ink & Switch Introduces Bijou64: Canonical Variable-Length Integer Encoding for Safe ParsingLINK

• Codeberg’s programmer user base overwhelmingly votes to ban slopcoded projects form its platformLINK

• APIFlow-Bench: The Enterprise Survival Test for AI AgentsLINK

• Agents keep changing their answers. Harness just built delivery pipelines that don’t care.LINK

• Confidential Containers becomes a CNCF incubating projectLINK

• New Markdown rival: Open-source DGML format aims to turn docs into data that AI (and humans) can trustLINK

• OpenAI unveils Presence, a new platform that lets enterprises launch and manage realtime voice agents and chatbotsLINK

🧰 Trending tools

AnySearch: a search API built for AI agents that pulls filtered, de-duplicated, structured results from trusted sources in parallel, improving reliability.LINK

Zro: routes coding requests to open-source models like MiniMax M3, GLM-5.2, and Kimi K2.7 across multiple regions without retaining any data.LINK

Humalike: gives AI agents behavioral APIs for turn-taking, proactiveness, and social timing, so conversations feel natural instead of robotic or scripted.LINK

Bento: a self-contained HTML file that functions as a full slide deck editor, viewer, and presenter with offline live collaboration and no cloud dependency.LINK

Loomal: adds a paywall to APIs, digital products, or stores so AI agents can pay in USDC, settling in seconds with zero revenue cut.LINK

Codex Micro: a lightweight platform for building and deploying AI-powered products, letting developers integrate leading models without heavy infrastructure overhead.LINK

📚 Trending papers & reports

AI-written unit tests can be made to actually compile and run reliably in real, messy codebases, by explicitly feeding the model project dependencies instead of letting it guess, cutting both generation time and cost versus current AI test-writing tools.LINK

Automated glue code connects plain-English test scenarios to actual software code automatically, delivering ready-to-use code for 46.1% of cases and cutting the manual busywork that slows down product testing as requirements change.LINK

Zero-knowledge blockchain code gets auto-checked for hidden bugs by pairing AI translation with a math verifier, catching over 90% of faulty implementations that could let broken transactions pass as valid.LINK

Automated code migration converts C programs into memory-safe Rust with far fewer errors, cutting a costly manual rewriting task that has slowed software safety upgrades industry-wide.LINK

Timing-based fault testing lets engineers trigger failures at the exact moment in a chain of microservice calls, like right after a database update, hitting the right timing every time across 450 trials while cutting search time by 95.91%, without touching application code.LINK

See you tomorrow for a new dose of ☕️ Devshot!