Dev Digest — April 8, 2026

🔥 HOT RELEASES

Reth 2.0 Paradigm's Rust Ethereum client hits 2.0 — now achieving gigagas/second execution speeds. Major rewrite with synthetic benchmarks and an auto-optimization harness inspired by Karpathy's autoresearch approach. 🔗 X: https://x.com/gakonst/status/2041932369792516561 💻 GitHub: https://github.com/paradigmxyz/reth

GitHub Copilot CLI "Rubber Duck" Agent GitHub Research shipped a second-opinion review agent for Copilot CLI — automatically gets a review from a different model family. Their SWE Bench data shows Rubber Duck closes 75% of the gap between Sonnet and Opus. Also: Copilot CLI now supports BYOK and local models for air-gapped environments. 🔗 X: https://x.com/burkeholland/status/2041640366189748339 🔗 Blog: https://github.blog/ai-and-ml/github-copilot/github-copilot-cli-combines-model-families-for-a-second-opinion/

Amazon S3 Files AWS announced S3 Files — the first cloud object store with fully-featured, high-performance file system access. Huge deal for anyone working with S3. 🔗 X: https://x.com/awscloud/status/2041611271565381968 🔗 Info: https://go.aws/4tw17Zg

Goose (Block/Square) — Still Trending Jack Dorsey's open-source AI coding agent continues gaining traction at 35K+ GitHub stars. Model-agnostic, works with any LLM. Community debate: the real moat isn't the CLI but the instruction layer (CLAUDE.md, skills, subagents). 🔗 X: https://x.com/jack/status/2039897468775604711 💻 GitHub: https://github.com/block/goose

🧪 INTERESTING REPOS

Career-Ops — AI job search tool that helped its creator evaluate 740+ applications and land a job. 12K+ stars in two days, then open-sourced. 💻 GitHub: https://github.com/santifer/career-ops 🔗 X: https://x.com/DataChaz/status/2041772318209536028

Graphify — Turn anything into a knowledge graph. 10K+ stars in days. People are pairing it with Karpathy's LLM Knowledge Base for visualization. 💻 GitHub: https://github.com/safishamsi/graphify 🔗 X: https://x.com/meta_alchemist/status/2041871110971777175

Awesome AI Apps — 70+ working AI agent projects covering every framework and pattern. Copy-paste ready. 💻 GitHub: https://github.com/Arindam200/awesome-ai-apps 🔗 X: https://x.com/hasantoxr/status/2041848347204710806

Vibe-Trading — AI-powered multi-agent swarm that transforms natural language trading ideas into executable strategies. 💻 GitHub: https://github.com/HKUDS/Vibe-Trading 🔗 X: https://x.com/huang_chao4969/status/2041831977565987323

Learn Claude Code — 20 chapters, 20+ runnable Python agents, 4 progressive stages. Build your own Claude Code from scratch. 💻 GitHub: https://github.com/shareAI-lab/learn-claude-code 🔗 X: https://x.com/dunik_7/status/2041796861825061102

LinuxPDF — Full Linux OS running inside a 6MB PDF via RISC-V emulator. Open it in your browser and get a functional shell. 💻 GitHub: https://github.com/ading2210/linuxpdf 🔗 X: https://x.com/HowToAI_/status/2041802339129106541

Antigravity Awesome Skills — 1381+ coding skills for Claude Code, Gemini CLI, Codex, Cursor, Copilot, and more. 31K+ stars. 💻 GitHub: https://github.com/sickn33/antigravity-awesome-skills 🔗 X: https://x.com/techNmak/status/2041870512633344227

GitNexus — Zero-server code intelligence engine that builds knowledge graphs from your codebase entirely client-side in the browser. 💻 GitHub: https://github.com/abhigyanpatwari/GitNexus 🔗 X: https://x.com/trending_repos/status/2041802689978491228

StackOne Defender — Prompt injection firewall for AI agents. 22MB, no API/GPU needed, 88.7% detection at ~4ms latency. 10x faster than Meta Prompt Guard. 💻 GitHub: https://github.com/StackOneHQ 🔗 X: https://x.com/Suryanshti777/status/2041815323314417913

Superpowers — Forces AI coding agents into real dev workflows instead of freestyling. 40K+ stars. 🔗 X: https://x.com/shedntcare_/status/2041807647465492776

FlipOff — Turn any TV into a retro split-flap display (like old train stations). Vanilla JS, zero dependencies. 💻 GitHub: https://github.com/magnum6actual/flipoff 🔗 X: https://x.com/CamilleRoux/status/2041808273272102977

🎥 WORTH WATCHING

TikTok Clone with AI Feed Algorithm Built a TikTok clone where the feed ranking algo is recursively updated by a coding model on each pull-to-refresh. Videos sourced from Wikipedia. Wild experiment in self-improving recommendation systems. 🔗 X: https://x.com/sdand/status/2041588956421353746

Google AI: Vibe Coding Explainer Google AI published a video explaining vibe coding for beginners (and for explaining it to your parents). 🔗 X: https://x.com/GoogleAI/status/2041637733634912321

💡 TECHNIQUES & IDEAS

Adversarial AI Code Review Workflow Emerging pattern: Claude writes code → fresh Claude context reviews → Codex reviews → Copilot reviews PR. Multiple developers report that Codex finds more logic errors than Claude reviewing its own output. The "Rubber Duck" approach now has official GitHub support. 🔗 X: https://x.com/StratoLens/status/2041787455569826297

Garry Tan on OpenClaw Config Pitfalls Even YC's president hit the classic footgun: asking an AI agent to modify its own config file, resulting in invalid config that prevents restart. A reminder to guard your agent's self-modification capabilities. 🔗 X: https://x.com/garrytan/status/2041587235120935157

Apple Blocks Vibe-Coded Apps Apple banned Anything (an app that lets people build apps with AI and test on their phone) under guideline 2.5.2. The platform gatekeeping vs. AI-generated apps tension is heating up. 🔗 X: https://x.com/karankendre/status/2041755037769281869

Design Systems for Vibe Coders With everyone shipping AI-generated products faster than ever, they all look identical. Strong argument that founders need a design system more than ever to differentiate. 🔗 X: https://x.com/rahulluthra22/status/2041890080273657957

🔮 EMERGING TRENDS

The Instruction Layer is the Moat — Community consensus forming: the CLI/agent binary is becoming a commodity. The real value is in CLAUDE.md, skills, subagents, and tool guardrails. Goose can swap models but can't replicate the scaffolding. Same pattern across all AI coding tools.

Cross-Model Review as Standard Practice — GitHub officially shipping multi-model review (Rubber Duck). Community already doing it manually. Expect this to become default workflow: write with one model, review with another.

Vibe Coding Hits Mainstream Resistance — One year after coining, vibe coding is everywhere. But Apple is blocking it, designers warn everything looks the same, and the "real engineers don't vibe code" debate has flipped — now everyone does it, even the skeptics.

Agent Security Gets Real — StackOne Defender, prompt injection firewalls, and data isolation tools signal that agent security is moving from academic to production concern.

Full digest: https://digest.99.cooking