Three Shell repositories encoding software develop, 'Show HN: Needle, CloakHQ/CloakBrowser surges to 1,835 stars/day

Three Shell repositories encoding software development methodologies pulled a combined ~6,200 GitHub stars in a single day — mattpocock/skills at 3,392 stars/day (79,617 total), obra/superpowers at 1,401 stars/day (189,900 total), and K-Dense-AI/scientific-agent-skills at 99 stars/day (21,269 total). Agent skills distributed as `.claude` directory files are now being starred at rates comparable to major framework launches. This is a new category of open-source artifact — not code, but structured interaction patterns — and it has no historical precedent at this velocity. The question shifts from 'will agent skills become mainstream' to 'will platform-native skill stores (Cursor extensions, Copilot extensions) absorb this distribution, or will GitHub remain the primary channel?'

'Show HN: Needle — We Distilled Gemini Tool Calling into a 26M Model' sustains at 655 HN pts / 185 comments — the highest AI technical engagement on today's front page, still climbing from yesterday's launch. Cactus Compute distilled Gemini's tool-calling into a model running at 6,000 tokens/sec prefill that beats all sub-350M alternatives on single-shot function calls. Six months ago, tool calling required frontier-scale models (100B+ parameters); Needle demonstrates it at 26M — a 4,000x compression ratio. Combined with the open-weight cost-arbitrage movement (DeepSeek V4, Kimi K2.6) and the local-first infrastructure surge (OpenClaw at 210K+ stars, openhuman at 6,132), the path to fully local agent execution now covers both reasoning and tool calling.

CloakHQ/CloakBrowser surges to 1,835 stars/day (9,796 total) — stealth Chromium that passes every bot detection test as a drop-in Playwright replacement. 49 source-level patches compiled into the Chromium binary (canvas, WebGL, WebRTC, CDP input mimicking), scoring 0.9 on reCAPTCHA v3 vs 0.1 for stock Playwright — a 9x improvement. On a day where privacy and evasion infrastructure trends across multiple layers (apernet/hysteria proxy at 485 stars/day for censorship resistance, CloakBrowser at 1,835/day for bot detection evasion), the arms race has moved from JavaScript injection to source-level binary patching. The demand signal is clear: developers building AI agent workflows need browsers that don't get blocked.

tinyhumansai/openhuman surges to 1,696 stars/day (6,132 total) — a Rust + Tauri local-first AI assistant that builds a Karpathy-style Obsidian wiki from your connected accounts (118+ integrations), with < 50MB memory footprint vs 100MB+ for Electron. GPL-3.0 licensed. Running concurrently: rohitg00/agentmemory at 1,379 stars/day (8,031 total — '#1 Persistent memory for AI coding agents'), supertone-inc/supertonic at 859 stars/day (on-device multilingual TTS via ONNX), and danielmiessler/Personal_AI_Infrastructure at 435 stars/day. Four independent projects assembling a coherent on-device AI stack with ONNX as the common runtime — no single vendor is driving this; it's emergent.

Linux gaming crosses a structural threshold — 610 HN pts / 411 comments, the highest-engagement story on today's front page. Windows APIs (DXVK, Proton-translated DirectX) have been so extensively reverse-engineered and implemented as proper Linux kernel features that the translation layer overhead is essentially eliminated. 411 comments indicates the community reads this as an inflection point after 10+ years of Wine/Proton development, not incremental progress. Running simultaneously with 'The Emacsification of Software' (244 pts / 164 comments) and 'A History of IDEs at Google' (326 pts / 224 comments), the day's secondary theme is technology that wins by patient accumulation rather than disruptive launch.

Twin brothers wipe 96 government databases minutes after being fired — 368 HN pts / 268 comments. The commentary is focused on access control failure: principle of least privilege, offboarding automation, and the gap between security policy and implementation in government IT. Landing simultaneously with YellowKey BitLocker zero-day bypass (65 HN pts, confirmed on Lobsters), Fragnesia Linux privilege escalation (Lobsters, Tildes ~comp at 49 votes / 52 comments), and an NGINX 18-year-old RCE vulnerability disclosure, security infrastructure across endpoint encryption, access control, and kernel layers is under concurrent stress for the second consecutive week.

Claude for Small Business launches at 82 HN pts / 32 comments (1 hour old at scrape time, still accelerating) — Anthropic's first dedicated SMB offering, expanding from developer API to business platform. On the same day GitHub's spec-kit holds at 1,120 stars/day (98,604 total) and millionco/react-doctor debuts at 604 stars/day ('Your agent writes bad React. This catches it'), the theme is guard rails over capabilities: the community is building agent discipline infrastructure (skills, spec-driven development, validation tools) faster than agent capability infrastructure. The competitive axis for AI platforms has shifted from model quality to workflow integration and trust.