Dev Digest — April 1, 2026

Today's dev world is dominated by the Claude Code source leak fallout, but there's plenty more happening beneath the chaos.

🔥 HOT RELEASES

Claude Code Goes Open Source (For Real This Time) After yesterday's accidental source leak via npm (512K lines of TypeScript with source maps), Anthropic decided to lean into it and officially open-sourced Claude Code. The DMCA whack-a-mole was already unwinnable — 41K+ forks before they woke up. Boris from the team handled it well: blamed process, not people. 🔗 X: https://x.com/MatthewBerman/status/2039383050187153888 💻 GitHub: https://github.com/anthropics/claude-code

claw-code — The Rust Rewrite That Broke GitHub Records Korean developer Sigrid Jin (reportedly Claude Code's biggest power user at 25B tokens/year) rewrote the entire thing from scratch — first in Python, then Rust. The repo hit 50K stars in under 2 hours, making it the fastest-growing repo in GitHub history. 🔗 X: https://x.com/realsigridjin/status/2039180182788128861 💻 GitHub: https://github.com/instructkr/claw-code

OpenCode — Claude Code Fork for Any LLM Within hours of the leak, someone forked the source and made it work with GPT, DeepSeek, Gemini, Llama, MiniMax — any LLM provider. 19K+ likes. 🔗 X: https://x.com/gitlawb/status/2039063488468000835 💻 GitHub: https://github.com/jjmartres/opencode

Google ADK for Go 1.0 Google shipped Agent Development Kit for Go 1.0 — build production-grade AI agents in Go with native OpenTelemetry, plugin system, YAML config, and A2A protocol support. 🔗 X: https://x.com/googledevs/status/2039387315747819718 💻 Docs: https://developers.googleblog.com/adk-go-10-arrives/

Fyrox 1.0 — Rust Game Engine After 7 years of development, Fyrox (Rust-powered 2D/3D game engine with Unity-like editor) hit 1.0. A serious milestone for the Rust gamedev ecosystem. 🔗 X: https://x.com/gamefromscratch/status/2039352899650674697 💻 GitHub: https://github.com/FyroxEngine/Fyrox

Liquid AI LFM2.5-350M A 350M parameter model trained on 28T tokens — runs at 313 tok/s on AMD CPU, under 1GB memory. Supports 32K context, tool use, instruction following. Small models becoming real infrastructure. 🔗 X: https://x.com/Marktechpost/status/2039207721216676110 💻 HuggingFace: https://huggingface.co/LiquidAI/LFM2.5-350M

Microsoft VibeVoice Open-source voice AI: 60-min single-pass speech recognition, 90-min multi-speaker TTS, 300ms streaming latency, 50+ languages. Trending with 3,863 stars in 24h (33K total). 🔗 X: https://x.com/trending_repos/status/2039265955323551935 💻 GitHub: https://github.com/microsoft/VibeVoice

🧪 INTERESTING REPOS

Sandcastle — Matt Pocock's framework for coordinating parallel AFK coding agents. Creates git worktrees, runs sandboxed agents in Docker, merges results. Simple API, 100% local. 🔗 X: https://x.com/mattpocockuk/status/2039350619681554434 💻 GitHub: https://github.com/mattpocock/sandcastle

Feynman — Open-source AI research agent. 296 likes, trending. 🔗 X: https://x.com/tom_doerr/status/2039115357839929610 💻 GitHub: https://github.com/getcompanion-ai/feynman

Happier — Open-source companion for AI coding agents. 🔗 X: https://x.com/tom_doerr/status/2039149801791447484 💻 GitHub: https://github.com/happier-dev/happier

Sesame Robot — Open-source ESP32 quadruped robot. Hardware meets software. 🔗 X: https://x.com/tom_doerr/status/2039130680433471531 💻 GitHub: https://github.com/dorianborian/sesame-robot

CaP-X — NVIDIA/Berkeley/Stanford/CMU open-source agentic robotics framework. Vibe-coded robot arms and humanoids with perception APIs, skill libraries, and RL training. MIT license. From the team behind Voyager. 🔗 X: https://x.com/DrJimFan/status/2039358115318243352 💻 GitHub: https://github.com/capgym/cap-x

any-buddy — Rehatch your Claude Code /buddy companion to get any buddy you want. 🔗 X: https://x.com/vineetwts/status/2039293131762614455 💻 GitHub: https://github.com/cpaczek/any-buddy

LocalCowork — AI agent running entirely on a MacBook. 385ms tool dispatch, 67 tools, 14.5GB memory, zero network calls. 🔗 X: https://x.com/paulabartabajo_/status/2039153333013721219 💻 GitHub: https://github.com/Liquid4All/cookbook/tree/main/examples/localcowork

claude-md — Drop-in CLAUDE.md that "forces employee-grade output" from Claude Code. 251 likes. 🔗 X: https://x.com/iamfakeguru/status/2039076558686531761 💻 GitHub: https://github.com/iamfakeguru/claude-md

🎥 WORTH WATCHING

Sandcastle Demo — Matt Pocock shows picking tasks, parallelizing N agents, and merging code AFK. 🔗 https://x.com/mattpocockuk/status/2039350619681554434

claw-code First Look — Sigrid Jin demos the Rust-based independent coding harness agent. 🔗 https://x.com/realsigridjin/status/2039180182788128861

Claude Code Leak Explained — Julian Goldie's breakdown of the npm leak, the Undercover Mode irony, and the aftermath. 🔗 https://x.com/JulianGoldieSEO/status/2039041692209410103

Claude Code Polymarket Trading Bots — Live broadcast exploring what the leak revealed about trading bot capabilities. 🔗 https://x.com/MoonDevOnYT/status/2039360833914806524

💡 TECHNIQUES & IDEAS

Parallel Agent Orchestration — The pattern emerging: git worktrees + sandboxed Docker containers + task queues = N agents working your codebase simultaneously. Sandcastle formalizes this. The bottleneck shifts from coding speed to task decomposition.

Claude Code /buddy — Anthropic shipped a pet/companion system for Claude Code. 18 pets, 5 tiers (common to legendary). "Its main purpose is literally nothing. Emotional vibe coding support?" Surprisingly well-received. April 1st timing suspicious but confirmed real. 🔗 X: https://x.com/birdabo/status/2039294087539667320

Figma-to-App Vibe Coding — Turn Figma links directly into complete apps. The design-to-code pipeline is getting dangerously frictionless. 🔗 X: https://x.com/chddaniel/status/2039143025553891708

Veil: Multi-Modal Verification — Write a model once in Lean, then apply concrete/symbolic model checking, SMT proofs, and interactive theorem proving from one executable spec. Verification tooling maturing. 🔗 X: https://x.com/leanprover/status/2039364395994841350

🔮 EMERGING TRENDS

The "Open Source by Accident" Era — Claude Code's leak-to-official-open-source pipeline took ~24 hours. The DMCA takedown page became the best discovery tool for finding mirrors. The lesson: if your moat is source code, you don't have a moat.

OpenAI's $122B Round at $852B Valuation — Largest funding round in history. Led by Amazon, NVIDIA, SoftBank with Microsoft participating. Includes $3B from retail investors. The AI race funding is now at sovereign-wealth-fund scale. 🔗 X: https://x.com/OpenAI/status/2039085161971896807

DeepSeek V4 Rumors — Leaked details suggest a "coding beast" with a new memory system called Engram for handling large prompts. Anthropic allegedly accused them of training on Claude outputs. Release rescheduled to late April due to hardware constraints. 🔗 X: https://x.com/linie_oo/status/2039303785214779477

Small Models as Infrastructure — Liquid AI's 350M model running 313 tok/s on CPU with tool use and 32K context. Microsoft's VibeVoice doing voice at the edge. The "run it locally on anything" tier is becoming genuinely capable.

Agent Memory is the Unsolved Problem — Multiple threads discussing how the Claude Code leak revealed a "primitive" memory system. The consensus: whoever solves persistent memory and context across sessions wins the agent race, not whoever has the best single-turn model.

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