mattpocock/skills, obra/superpowers, and affaan-m/, A forensic analysis combining a prompt-injection s, Prime Intellect's Prime Agent -- already logged in

mattpocock/skills, obra/superpowers, and affaan-m/ECC -- every one of GitHub Trending's top marketplace slots since 08-20 -- are absent from today's full 19-repo board. Direct API reads across the 48-hour gap (no run happened 08-24) show growth falling to roughly a third to a fifth of the prior four-day pace: mattpocock/skills ~1,608 stars/day (down from 2,504-2,959), obra/superpowers ~455/day (unpushed since 08-19), affaan-m/ECC ~367/day (down from a May-cycle 1,500-2,100/day). anthropics/claude-code itself still grows slower in absolute terms (~152/day) than any of the three even now.

A forensic analysis combining a prompt-injection system-prompt extraction (which got the model to state directly 'I'm GLM, a large language model made by Z.ai') with Normalized Compression Distance stylometric classification (GLM-5.3 matched 7 of 14 sampled responses, more than double the next-closest model) identifies the mystery 'stealth model' Ox-Alpha as Zhipu's GLM-5.3.

Prime Intellect's Prime Agent -- already logged in this vault on 2026-08-06 scoring 95.5% on ARC-AGI-3 running Opus 5 -- gets a formal arXiv paper today stating the identical figure verbatim ('raises ARC-AGI-3 RHAE Best@1 from 30% to 95.5%'), now with a named methodology (Recursive Language Model abstraction, Continual Harness preserving memory/skills/subagent specs across trajectories) and three new benchmark domains (long-context coding, GPU-kernel generation, emulator construction, nanoGPT speedruns) the original coverage didn't include.

An OpenAI-backed study reported by TechCrunch quantifies the agent-adoption gap this vault first logged as community discourse in mid-August: 98% of OpenAI's own employees use its Codex agentic coding tool, versus 17% of organizational subscribers and under 1% of individual ChatGPT subscribers -- as OpenAI generalizes the tool into ChatGPT Work for non-engineers.

Hugging Face is reportedly fielding acquisition offers valuing it around $13B, per Business Insider via TechCrunch -- no acquirer named, no deal finalized. CEO Clem Delangue cites a 'long-term responsibility' to the community and says the company already turned down a $500M Nvidia investment in 2026 that would have valued it at $7B, preferring sustainability over a fast valuation bump.

Google Assistant's shutdown on Android and Wear OS is now confirmed for September 4, 2026, resolving this vault's open question about whether the Gemini replacement runs on-device: even Gemini Nano v3 hardware (2026-flagship-chip, 12GB RAM) still requires a paired-phone or LTE internet connection on watches -- the transition is a cloud-to-cloud vendor handoff, not a shift toward local voice processing.