daily.dev · Popular · 实时热榜

Engineers using AI fall into three camps: refusers who distrust it, overreliers who stop thinking critically, and multipliers who use AI to amplify their judgment. The piece offers concrete workflows for using AI to onboard into unfamiliar codebases (via Claude Code's /init, architecture mapping, testing strategy discovery, and setup checklists), to learn new languages/frameworks by mapping concepts from a known language, to build hands-on learning projects instead of watching tutorials, and to

A deep architectural walkthrough of OpenSEO, an open-source SEO platform combining keyword research, competitor analysis, backlinks, site audits, and AI-visibility tracking. Built on TanStack Start, React, and Cloudflare bindings (D1, R2, KV, Workflows, Durable Objects), it pulls market data from DataForSEO and model access from OpenRouter. The piece traces how keyword research, rank tracking, the site audit crawler, MCP tool access, the in-app SAM agent, report generation, and Search Console/GA

Cloudflare has open-sourced Void, a three-part toolchain that came out of VoidZero, the company behind Vite, which Cloudflare acquired in June. It is MIT-licensed and installable via npm. VoidZero has reportedly used it in production for several months. What’s in it Void SDK is a composable JavaScript SDK for full-stack Vite apps. It has file-based routing, type safety from the database to the UI, and automatic resource provisioning for things like D1 and AI bindings. Under the hood it uses Hono

Yesterday’s release added a call to an external fraud-check API (about 400 ms) to checkout. Checkout p50 went from 80 ms to 500 ms. You expected that. But checkout p99 at peak went from 300 ms to 3.8 s. CPU is 35%. Memory is 50%. Error rate is 0.1%. Database query p99 is 12 ms. The fraud check runs inside the database transaction. Each pod has a pool of 20 database connections and handles about 45 checkouts per second at peak. Traces show the fraud call itself stays under 500 ms at p99, and GC p

Large open-source projects like tldraw, Godot, and others are closing or restricting external contributions because AI-generated patches are easy to produce but costly to review. The author argues contributors should instead focus on smaller, less famous projects that genuinely need help, citing their own tool Help Wanted (650K+ issues from 96K+ projects) and a personal Hacktoberfest experience where three PRs to small projects were reviewed the same day. The piece frames this as an opportunity

A developer reflects on the difference between getting AI to produce working code and actually understanding it. Drawing on experience building a text editor called Slate, the author discusses a save-race-condition bug (revision-based save tracking) as an example of a problem worth truly understanding versus just accepting AI's solution. The piece argues that experimentation, investigating unexpected results, and remembering problems (not code) are what make AI-assisted building a genuine learni

“Bug fixed. All tests passed.” Then you open the app… and nothing works. AI agents are getting really good at reporting success. Actually verifying the result? That’s another story. 😅 What’s your best “the agent said it was done” moment? #AIAgents #AgenticAI #SoftwareDevelopment

The funny thing is, they migrated off Git and used Claude to build their own version control system, Scion. But Why? Anthropic reports more than 30,000 daily merged changes by September, up from roughly 2,500 in May. The company moved the monorepo from Git to Scion after source control became the bottleneck. Basically, Git was designed for developers who keep a full copy of the repository and sometimes work offline. But today, many users are AI agents and CI jobs that are always online. That’s w

1,170+ free, open-source animated icons, now available beyond React. AnimateIcons brings the same collection of animated icons to React, Vue, Angular, and vanilla JavaScript, with support for Svelte, Solid, and Astro. Add motion and polish to your interfaces without creating animations from scratch. Choose your framework and use the same icons across your frontend stack. Browse the collection: https://animateicons.in Open source: https://github.com/Avijit07x/animateicons

freeCodeCamp.org published a free 3-hour YouTube course teaching system design using a five-step framework for structuring thinking and tackling vague requirements under interview pressure. The course walks through building high-level architectures for real-world systems like Spotify, Booking.com, and WhatsApp, covering bottleneck identification, time management, storage and traffic estimation, and trade-off discussions such as WebSockets versus HTTP or when to add load balancers. It also explai

An experiment with two Spring Boot instances using HikariCP against PostgreSQL shows that default pool sizing (minimumIdle = maximumPoolSize = 10) leaves 20 idle connections open during quiet periods, consuming about 57-58 MiB of PostgreSQL cgroup memory. Setting minimumIdle to 1 with a short idleTimeout let idle connections retire down to 2, cutting PostgreSQL cgroup memory to roughly 28-29 MiB. A control profile confirms that starting with a low minimumIdle alone doesn't help — only enabling i

An investigative breakdown examines Cloudflare's acquisition of Deno, revealing that the Deno runtime will get only one more year of bug fixes before development ends, with Deno Deploy shutting down 6 months after that. Commit data shows contributions collapsed over 90% between June and September 2026, with most remaining work done by two people, well before the October 9th announcement. The piece traces warning signs: a 2025 layoff of roughly half the staff that was never disclosed, shrinking f

An explanation of how Uber's real-time map handles millions of GPS location updates per second by splitting the system around read patterns rather than writes. Covers an edge ingestion layer that validates, deduplicates, and adaptively samples pings before they hit a database; H3, Uber's open-sourced hexagonal geospatial index, which turns 'find nearby drivers' from a full scan into a bounded lookup; a fast in-memory store (Redis-like) for live rider map queries paired with a durable Cassandra-f

A rundown of npm's ongoing supply chain security crisis, covering a fresh malicious package campaign from an account called Dirty Blanket that used typosquatted Express/React packages, a bash worm, and a Tor-based backdoor. It recaps the Shai-Hulud worm family (original, the 700+ package second wave, and Mini Shai-Hulud hitting TanStack), Google's undercover infiltration of the Team PCP group leading to two arrests in Australia, and related incidents at Red Hat, GitHub, and Axios via malicious V

OpenCode 2.0 is a major overhaul of the AI coding agent, introducing a persistent background server shared across TUI, web, and desktop clients, a redesigned API with an @opencode/client package (JavaScript and Effect support), and a new Code Mode that lets agents write JavaScript to call MCP/OpenAPI tools in a single round trip instead of sequential calls. It adds background subagents and async shell commands for parallel orchestration, promotes a stabilized hot-reloadable plugin system (80+ in

Proposes organizing AI agent guidance into three layers: repository-level rules (architecture, code style, exceptions specific to an app), company-level shared workflows (ticket handling, PR process), and a personal developer layer capturing individual expertise and review heuristics. Argues that standardizing collaboration should not flatten developer judgment, and that code review remains the place where diverse experience becomes shared knowledge. References Bacchelli and Bird's study on code