Apple opens Siri to every developer

Hi there, this is your daily ☕️ Devshot.

Apple opens Siri AI to all developers LINK

• Apple has opened Siri AI to all third-party developers through a new preview, letting apps handle complex spoken requests like booking restaurant reservations, checking deliveries, and placing orders that were previously restricted to a narrow set of supported actions.
• Using the App Intents framework, developers can build custom actions and chain multiple steps into one "skill", such as finding a product and sharing it via Messages, while Siri taps Spotlight's semantic index to surface app content outside predefined categories.
• The iOS 27.3 beta already lets Siri organize Safari tabs, find visited places, and archive files across 15 Apple apps, but third-party integrations remain in developer testing, with customer features expected in a few months and App Store submissions in spring 2027.

Self-hosting frees Revolut from Claude's costs LINK

• Revolut cut its AI coding inference costs by about 90% in under a year by switching its default from commercial services like Claude Code to self-hosted open-weights GLM models, now running GLM-5.3 across its own GPU fleet.
• Self-hosting GLM-5.1 cost roughly $0.30 per million tokens, about a tenth of Codex or Claude Code, and the removal of token budgets let its ~2,000 engineers hand debugging and Jira and Confluence tickets to the models, driving a 60% coding-time improvement.
• Serving GLM-5.3 on 56 Nvidia B200 GPUs, Revolut hit sub-3-second time to first token at 256K context and ~200 tokens/second by tuning hyperparameters and switching to NVFP4 quantization, but keeps routing vision and UI queries to Gemini or GPT since GLM is text-only.

Anthropic builds tool to verify AI models LINK

• Anthropic's Frontier Red Team released FlashProve, an Apache-licensed library that cryptographically proves an API response came from a specific set of model weights without exposing those weights, using the same hash-based proofs behind Ethereum L2s like Starknet.
• The provider publishes a fingerprint of its weights once, then attaches a proof to each reply; on the 400B-parameter Llama 4 Maverick, the prover hits ~110 tokens/sec on one H100, proofs run ~19 MB, and a CPU verifies one in roughly 20 seconds.
• The catch is steep: proving costs 70-95x the GPU time of normal inference, each request waits about five minutes for batching, and context is capped at 1,024 tokens, Anthropic calls it a research prototype and won't maintain the repo.

Xcode beta adds GitHub Copilot support LINK

• Apple released Xcode 27.3 beta 1 today, adding GitHub Copilot to the built-in coding assistant and expanding the roster of AI agents that developers can run directly inside the IDE.
• Copilot joins ChatGPT, Claude, and Google Gemini, which Apple has integrated across recent releases, with agentic capabilities via Codex and Claude Agent arriving back in Xcode 26.3.
• Apple frames this as a win for developers at companies standardized on Copilot, who can now use their org's preferred tool in Xcode; the stable 27.3 release is expected in December.

Vitest 5 cuts test time up to 25% LINK

• Vitest 5.0 shipped from VoidZero with speed as its headline, having inline projects share a single Vite server while quietly tightening several long-standing defaults.
• Across a benchmark suite spanning a five-file package to a sizeable monolith, the team reports roughly 8-25% gains, with VM pool scenarios hitting the top end, from fewer main-to-worker round trips and reused compiled code.
• It now requires Vite 6.4 and Node.js 22.12, makes unawaited async assertions fail outright, and the dangerous migration case is suites that pass locally on Vitest 4 then go red in CI over call-count assertions leaking mock history.

Claude's React port proves nothing broke LINK

• Maple rebuilt its 287,000-line, 131-route web app off React onto Foldkit, an Elm Architecture layer on Effect, with Claude doing the port in three days, verified by a custom harness proving the two apps behaved identically.
• The React version scattered state across hundreds of atom hook calls over many files; Foldkit centralizes all state in one Model mutated by a single function, though losing route-based code splitting roughly doubles the startup bundle from 703 KB gzipped.
• A Playwright harness compared both apps across 2,600+ deterministic variants on pixels, structure, request logs and ARIA; it surfaced old React bugs and let agents fix only the port, while later reviews still caught ~180 inter-screenshot defects the diffs missed.

Tutorials & Advice

> A Pile of MCP Connectors Is Not a Context Engine: Pre-process scattered company knowledge into one reconciled, permission-aware layer so AI agents get accurate answers instead of slowly stitching together contradictory sources themselves.

> The hidden latency inside your database: PlanetScale Metal runs databases on local disks instead of network storage, cutting read latency while keeping durability through replicas across separate zones.

> How software engineering is changing: a tale of two eras: Shipping 11x more code per month comes from letting AI agents build features solo while dropping human code reviews and accepting that things sometimes break.

> Re-Implementing :nth-child() With Modern CSS: Explains how to reproduce the nth-child selector using modern CSS style queries and conditions so that its arguments can be adjusted dynamically with variables.

> OpenShell on Kubernetes: A guide for platform engineers: Run each AI agent in a locked-down sandbox where an outside supervisor approves every request, so agents holding real credentials can't misuse them.

Other news you might like

• Bridging Alloy and OpenTelemetry: the OTel engine is now in Public Preview LINK

• vercel-labs/scriptc: TypeScript-to-Native Compiler LINK

• CheerpJ 5.0: Java 21/25 in the browser, full OpenJDK, and a new WasmGC engine LINK

• GitHub App installation tokens are now 520 characters: What to check before November 30 LINK

• Open source project renders your Windows 11 desktop using the NPU rather than the GPU — ‘npu-driver’ can only run at 2 to 3 Hz, targets 30 Hz performance LINK

• Fedora Linux 45 arrives with GNOME 51 and a big catch for Apple M3 users LINK

• macOS 27.3 adds native Markdown rendering to Quick Look LINK

Trending tools

Die With Me: a Mac app that tracks Claude and Codex usage, compares it with friends, and opens a shared chatroom during AI cooldown periods. LINK

beacon.​sh: captures agent traces across Claude Code, Cursor, Codex, and 20+ other harnesses, turning activity into operational visibility so each agent starts smarter LINK

AgentKit by Scalekit: connect AI agents to 30,000+ tools across 600+ apps through one MCP gateway, handling auth, scoped permissions, custom connectors, and event triggers. LINK

Inferock: a local proxy for OpenAI, Anthropic, Gemini, and OpenRouter that tracks token usage, failures, and retries to catch billing overcharges. LINK

Onepin: a QA layer for AI voiceover that fixes mispronounced names, prices, and dates across 30+ voice engines without re-rendering takes. LINK

Beam.win: streams individual windows from a remote Mac to your local machine, giving each app its own Dock icon and ⌘Tab slot over an encrypted connection. LINK

Trending papers & reports

Python type labeling automatically fills in the missing data-type notes that keep code reliable, hitting top accuracy while cutting AI costs up to ~79% versus a strong coding-assistant baseline. LINK

Code-to-architecture mapping sorts source files into a system's intended building blocks, and surprisingly a plain folder-structure rule hits 89-100% accuracy, matching fancier graph methods and questioning when extra machinery is worth it. LINK

Smart contract failures get a shared vocabulary and five-level severity ranking that describes what deployed Ethereum code actually does wrong when it breaks, helping teams prioritize and communicate the irreversible, costly mishaps on-chain. LINK

GitHub automated test failures cluster into streaks, with three quarters of failures following another and a third starting with no code change, meaning the culprit is often network issues, not the latest commit. LINK

Automated test generation gets a fix for the number-one failure of AI-written tests, broken imports, guaranteeing every generated test imports correctly where rival methods leave up to 45% of import errors unresolved. LINK

See you tomorrow for a new dose of ☕️ Devshot!