▌ GitHub radar
Top 5 GitHub Repos This Week
Week 27 was all about inference speed and agent tooling: DeepSeek dropped a full speculative-decoding toolkit, the community shipped a practical Codex guide and a serious eval library, and two small but clever tools — a multi-doc PDF format and an @agent Slack bridge — quietly climbed the charts.
DeepSeek's full-stack open-source toolkit for training and evaluating speculative decoding draft models. Ships three algorithms (DSpark, DFlash, Eagle3) with pre-trained checkpoints for Qwen3 and Gemma, multi-GPU training scripts, and a benchmark suite covering nine evaluation datasets.
Why a vibe-coder should care
Speculative decoding can cut LLM inference latency 2–3× without touching output quality — DeepSpec makes this research-grade technique accessible to any engineer running their own models.
An unofficial community guide to OpenAI Codex — the 'Orange Book' — covering everything from installation to real-world case studies. Produced by Chinese developers and available as a downloadable PDF, it fills gaps the official docs leave open.
Why a vibe-coder should care
When official docs stay high-level, practitioners fill in the gaps — and this distilled, downloadable guide is a fast track to using Codex effectively in real projects.
A seriously curated library of 443+ annotated resources for building and evaluating AI agents: papers, blogs, talks, tools, and benchmarks. Assembled via a depth-4 citation crawl of 11,600+ papers by BenchFlow, with 146 reading notes and active pruning of dead links.
Why a vibe-coder should care
If you're building AI agents and tired of bloated, half-dead awesome-lists, this one actually comes with reading notes, runnable code examples, and real maintenance — a proper starting point for eval work.
A minimal extension of the PDF format that bundles multiple documents into one file via an embedded JSON manifest. A .pdfx file is still a valid, standard PDF — any reader opens it — while the PDFx viewer can split it back into individual documents on demand.
Why a vibe-coder should care
Instead of sending a ZIP or five separate PDFs, you send one file that works in every standard PDF reader — no plugin or special software required for the recipient.
OpenTag lets your team @-mention a coding agent directly in Slack, GitHub, or Lark — the message is routed to Codex or Claude Code running locally on your machine, and the response comes back to the same channel. No cloud service in the middle.
Why a vibe-coder should care
A coding agent is only useful if your whole team can reach it without leaving their chat app — OpenTag removes the terminal-as-barrier without shipping your code to a third-party cloud.
More finds