GitHub · Open Source

A Rust VM That Cold-Starts in Under a Second, a Python Scraper That Survives Redesigns, and More

By HelloGitHub · Jun 29, 2026

Read original on juejin.cn ↗ Google Translate ↗ Alt translation

The thread connecting these projects is reducing the gap between "it works" and "it works without babysitting." Scrapers that survive redesigns, VMs that start like containers, and auth middleware that drops in with zero code changes all remove recurring operational toil that teams in any language ecosystem still pay for.

Summary

The latest HelloGitHub digest surfaces tools that solve concrete, annoying problems. smolvm delivers VM-level isolation with container-like speed, cold-starting in under a second and packaging machines into single executables. Scrapling is a Python scraping framework whose parser learns a site's structure and re-locates target elements after a redesign, so scrapers don't break when pages change. On the AI side, vllm-omni extends vLLM into a full omnimodal inference engine handling images, video, and audio alongside text, while a Tsinghua team's OpenMAIC platform turns any document into an interactive multi-agent classroom with an AI teacher and student agents.

Several projects target the friction between ecosystems. winpodx runs Windows apps as native Linux windows via a background container and FreeRDP, with taskbar pinning and Alt-Tab support. tinyauth bolts OAuth login onto any web app through a Docker-deployed middleware, no code changes required. keyd remaps keyboards at the kernel level on Linux, working identically across X11, Wayland, and TTY.

A handful of skill packs for AI coding assistants stand out. ponytail encodes a senior engineer's instinct to avoid over-engineering, cutting code volume by roughly 54% and token spend by 20%. stop-slop strips AI-generated clichés from prose. academic-research-skills offloads literature checks, citation verification, and logic consistency to Claude Code so researchers stay focused on questions and interpretation.

Takeaways

— smolvm cold-starts virtual machines in under one second and packages them as single executables, giving VM isolation with a container-like workflow.

— Scrapling's parser learns a website's DOM structure and re-locates target elements after a redesign, so scrapers don't break on layout changes.

— vllm-omni extends vLLM to handle image, video, and audio inputs and outputs, supporting both autoregressive models and Diffusion Transformers.

— winpodx runs Windows applications as native Linux windows through a background Windows container and FreeRDP, with taskbar pinning and Alt-Tab support.

— tinyauth adds OAuth login (Google, GitHub, etc.) to any web app via a Docker-deployed middleware that requires no code changes in the target application.

— keyd is a kernel-level keyboard remapping daemon for Linux that works uniformly across X11, Wayland, and TTY, supporting per-keyboard configurations.

— ponytail, a skill pack for AI coding assistants, reduces generated code volume by about 54%, cost by 20%, and time by 27% by enforcing a minimal-code philosophy.

— stop-slop is a writing skill pack that makes language models avoid cliché openings, business jargon, and formulaic transitions during generation and editing.

— OpenMAIC, from a Tsinghua University team, generates an interactive multi-agent classroom from any document, with an AI teacher and AI student agents that discuss and answer questions.

— Echo-Loop turns English audio into structured listening and speaking drills with sentence-by-sentence intensive listening, shadowing scoring, and flashcard generation from context.

— CapsWriter-Offline provides fully offline speech-to-text triggered by holding CapsLock, with LLM-based polishing and USB plug-and-play portability.

— llmfit auto-detects local CPU, GPU, and memory to recommend and rank models suitable for local inference across Ollama, llama.cpp, MLX, vLLM, and LM Studio.

Conclusions

Scrapers that survive DOM changes attack the single largest maintenance cost in web scraping: selector rot. If the parser genuinely learns structure rather than falling back to fuzzy heuristics, it changes the economics of long-running scraping jobs.

Sub-second VM cold starts blur the line between containers and VMs in a way that matters for multi-tenant workloads. The isolation story of VMs becomes viable for short-lived jobs that previously demanded containers.

OAuth-as-middleware (tinyauth) is a pattern that should be more common. The fact that it still requires a dedicated project in 2025 suggests reverse proxies haven't absorbed this responsibility as thoroughly as they should have.

ponytail's claimed 54% code reduction is an implicit critique of how AI coding assistants default to over-engineering. The skill-pack approach — encoding a specific engineering philosophy as a prompt — may prove more practical than fine-tuning for controlling AI code quality.

Offline speech-to-text with LLM polishing (CapsWriter-Offline) sidesteps the privacy and latency trade-offs of cloud-based dictation. The CapsLock trigger is a clever affordance that avoids wake-word awkwardness.

vLLM's expansion into omnimodal inference signals that the serving layer is consolidating around a few frameworks. Teams building on vLLM for text now get image, video, and audio without switching infrastructure.

An interactive multi-agent classroom (OpenMAIC) inverts the usual AI-education dynamic: instead of a single tutor bot, it simulates a social learning environment. Whether the pedagogical value holds up is an open question, but the architecture is novel.

Concepts & terms

evdev/uinput

Linux kernel input subsystems. evdev reads input events from hardware; uinput allows userspace programs to create virtual input devices. keyd uses both to remap keyboards at the kernel level, below any desktop environment.

ADS-B

Automatic Dependent Surveillance–Broadcast, a radio protocol used by aircraft to broadcast their position, velocity, and identification. The micro-radar project fetches this data from the OpenSky API rather than receiving radio signals directly.

FreeRDP

An open-source implementation of the Remote Desktop Protocol (RDP). winpodx uses it to render individual Windows application windows as native Linux windows from a background Windows container.

Diffusion Transformer (DiT)

A class of generative models that replace the U-Net backbone in diffusion models with a Transformer architecture. vllm-omni supports DiT alongside traditional autoregressive models for image and video generation.

MCP (Model Context Protocol)

A protocol that lets AI coding tools like Claude Code read and write external data sources. tolaria exposes its Markdown knowledge base as an MCP server so AI assistants can query and update notes directly.

URDF / SRDF / SDF

Robot description file formats. URDF (Unified Robot Description Format) describes robot kinematics and geometry; SRDF adds semantic configuration; SDF (Simulation Description Format) extends URDF for Gazebo simulation. text-to-cad can generate all three from natural language.

Source: juejin.cn ↗ Google Translate ↗ Backup ↗