📦 ggjy/awesome-agent-engineering

Understanding the Foundation of AI Agent Engineering

Large language models are shifting from simple question answering to active task completion, a transition that fundamentally relies on execution harnesses. This repository offers an organized collection of academic papers, systems, and benchmarks that define the agent as the sum of a foundation model and a runtime infrastructure. It identifies six key harness responsibilities, including observation interfaces, control loops, and verification systems, which determine how agents perceive environments and recover from errors. By exploring four engineering paradigms ranging from prompt design to agent-native training, these resources provide a comprehensive framework for building reliable, autonomous, goal-directed systems.

🆔 @hackernewsgithubprojects

📦 astroautomata/thousandworlds

ThousandWorlds: A Climate Emulation Benchmark

ThousandWorlds is the climate emulation benchmark that simplifies predicting exoplanet atmospheres for habitability studies. It offers seventeen hundred sixty simulations across five general circulation models, capturing complex variables like temperature and wind speeds on a structured grid. Researchers can leverage its nested datasets and baseline methods to train and evaluate models against standardized physics-based metrics. Whether you are performing parameter-to-field regression or assessing inter-model uncertainty, this repository provides the necessary tools and precomputed results to advance your work. Download the dataset and start testing your climate emulation models on this robust, expert-validated platform today.

🆔 @hackernewsgithubprojects

📦 orange-3dv-team/car

CaR: Revolutionizing Video World Model Memory

CaR is the video generation framework that finally solves long-term memory challenges in world models through efficient implicit retrieval. This project addresses the difficulty of maintaining visual consistency across complex camera movements by using an attention-driven mechanism to retrieve historical context. By combining a lightweight compression network with specialized retrieval attention, it enables smooth scene exploration and extended video synthesis from single images. The team also introduces the SceneFly dataset, which provides extensive synthetic video data for training these models. This work effectively bridges the gap between static generations and dynamic, consistent video environments for future simulations.

🆔 @hackernewsgithubprojects

📦 spacepxl/comfyui-vae-utils

Unlock Advanced VAE Control in ComfyUI

ComfyUI-VAE-Utils is the custom node toolkit that finally enables advanced VAE loading and processing options unsupported by standard ComfyUI. This repository provides dedicated nodes for loading VAEs with automatic input and output channel detection, specifically tailored for modern models like Wan. You gain access to specialized decoding capabilities, including automatic upscaling and tiling, alongside a neural latent upscaler that significantly outperforms basic interpolation methods. By bridging these functional gaps, the project simplifies complex generation workflows, allowing you to effectively manage latent manipulation and high-quality output decoding directly within your existing pipeline for better image and video results.

🆔 @hackernewsgithubprojects

📦 david-magdy/cvil

Master Computer Vision Interviews with This Checklist

Prepare for computer vision and machine learning interviews by using a structured, phase-based map designed to track your knowledge from core concepts to advanced production techniques. This repository provides a clear, organized list of essential topics, covering everything from fundamental mathematics, statistics, and CNN architecture to complex subjects like vision transformers, object detection, and tracking systems. It emphasizes developing a deep intuition for why specific models exist and the tradeoffs they introduce, helping you bridge the gap between academic theory and practical application. Leverage this checklist to systematically identify your study gaps and confidently master your next technical interview.

🆔 @hackernewsgithubprojects

📦 h-freax/termhub

Build Your Developer Portfolio in Minutes with TermHub

TermHub is the terminal-style portfolio template that turns your resume into a professional developer website without writing code. It simplifies the entire process by using a clever pipeline where you provide your resume to an AI, which then generates the necessary markdown and configuration files for your site. The project features a sleek, responsive design with a classic terminal aesthetic and supports bilingual content. You can even use its built-in server tools to automate site creation, allowing you to manage publications, projects, and work history simply by editing text files. It is the fastest way to showcase your work online.

📰 https://news.ycombinator.com/item?id=47378150

🆔 @hackernewsgithubprojects

📦 xingzhejun/d-opsd-code

Understanding d-OPSD for Diffusion Large Language Models

d-OPSD is the specialized training framework that introduces on-policy self-distillation for diffusion large language models. This project provides the essential implementation code and environmental setup required to train and evaluate models using this new approach. By modifying standard trainer files to enable specific batch generation configurations, it allows researchers to effectively apply self-distillation techniques across tasks like math and logic puzzles. The repository includes clear shell scripts for seamless training and performance evaluation, offering a structured way to improve model reasoning capabilities. This framework is a valuable resource for anyone working on optimizing advanced generative language models.

🆔 @hackernewsgithubprojects

📦 trickygo/moonsplat

MoonSplat: Real-Time 3D Reconstruction from Monocular Video

MoonSplat enables robust real-time 3D reconstruction and camera tracking using only standard monocular video sequences. By integrating global Sim(3) optimization with voxelized Gaussian splatting, the framework addresses common challenges like fragile pose estimation and optimization inefficiency in long-sequence scenarios. The implementation includes a color residual learning strategy to accelerate convergence and boost rendering quality, making it a powerful tool for applications like robotics and virtual reality. Users can configure parameters to balance speed and performance while processing custom data. This research project serves as an efficient, high-fidelity alternative for capturing dense 3D scenes from simple camera inputs.

🆔 @hackernewsgithubprojects

📦 xlang-ai/finevla

Steer Robot Behavior with Fine-Grained Instructions

Enhance robot control by aligning language with the specific physical actions required to complete tasks. This framework solves the problem of coarse instructions by generating fine-grained, action-aligned descriptions across ten dimensions like contact, trajectory, and orientation. By mixing these detailed instructions with standard goal-level goals, you can train vision-language-action policies that offer far superior steerability in both simulation and real-world manipulation. The project includes a complete pipeline for data construction, a specialized model for robotic video understanding, and comprehensive training tools. Adopting this fine-grained approach leads to significantly more precise and reliable robotic execution.

🆔 @hackernewsgithubprojects

📦 cjermain/pi-less-yolo

Secure Your AI Coding Agent with pi-less-yolo

pi-less-yolo is the tool that adds a crucial layer of security to your AI coding agents by wrapping them in an isolated Docker sandbox. By running agents like the pi-coding-agent inside a restricted container, it prevents them from accidentally modifying or deleting files outside your current project directory. It drops Linux capabilities and prevents privilege escalation, ensuring your development environment stays safe while you work. With support for tools like mise, you get simple, command-line control over these sandboxed sessions. Use this practical safety net to confidently run powerful AI coding tools without exposing your entire filesystem to risks.

🆔 @hackernewsgithubprojects

📦 andreicscs/honeywire

Build an Enterprise-Grade Cyber Canary in 60 Seconds

Defend your internal network by transforming any Linux machine into a high-fidelity deception trap in just sixty seconds. This platform uses specialized sensors like network scan detectors and file monitors to replace noisy traditional alerts with actionable intelligence. By deploying these decoys, you can instantly identify unauthorized lateral movement and suspicious activity without the typical headache of false positives. It features a centralized management dashboard for fleet control and uses a universal event standard that lets you integrate custom sensors in any language. Secure your infrastructure by catching intruders the moment they touch an unauthorized path.

🆔 @hackernewsgithubprojects

📦 tonybanters/oxwm

OXWM: A Modern Dynamic Window Manager

OXWM is a dynamic window manager written in the Zig programming language that prioritizes a highly flexible, scriptable experience over traditional compile-time configuration. By embedding a Lua interpreter, it allows users to modify their window management settings and hot-reload changes instantly without restarting the X session. The project provides native support for dynamic tiling, multi-monitor setups, and a modular status bar system with customizable information blocks. It is designed to offer a modern, accessible alternative to classic window managers while retaining essential features like tag-based workspaces and complex keychord support to enhance overall desktop productivity.

📰 https://news.ycombinator.com/item?id=48670511

🆔 @hackernewsgithubprojects

📦 vila-lab/activesam

ActiveSAM: Faster Open-Vocabulary Segmentation

Accelerate open-vocabulary semantic segmentation by using ActiveSAM, a training-free inference framework that transforms the Segment Anything Model 3 into a highly efficient active-vocabulary segmenter. Instead of processing full-resolution data for every potential class, it uses a low-resolution preview to identify only the relevant objects present in an image. By skipping unnecessary computations and utilizing bucketed prompt multiplexing, this method achieves significant speedups while improving accuracy across diverse benchmarks. It requires no target-dataset training or weight updates, making it a robust, high-performance solution for complex visual tasks like autonomous driving and real-world AI applications.

🆔 @hackernewsgithubprojects

📦 ml-from-scratch-book/code

Build 10 Core ML Algorithms from Scratch

You can demystify complex machine learning models by building ten core algorithms from the ground up using only NumPy. This repository serves as a practical companion for understanding exactly how these models function beyond their standard library wrappers. Each algorithm follows a five-stage framework that moves from plain-English intuition and mathematical formalization to clean code implementation and comparative testing. By validating these custom models against industry standards like Scikit-learn and PyTorch, you gain a deeper understanding of how systems actually learn. It is a highly effective way to master the fundamentals and transition from using black-box tools to truly building them.

📰 https://news.ycombinator.com/item?id=48461956

🆔 @hackernewsgithubprojects

📦 studio-saelix/sencho

Self-Hosted Docker Compose Management with Sencho

Sencho provides a graphical interface for managing Docker Compose stacks across one machine or an entire fleet without requiring SSH or a VPN. This tool acts as a single container running on your hardware, allowing you to deploy services, edit files, and monitor logs directly through a web dashboard while keeping your compose files as the primary source of truth. By using authenticated WebSocket tunnels, Sencho handles remote nodes securely, even those behind strict firewalls or NAT. It offers a clean way to oversee your infrastructure, helping you maintain control and visibility over your containerized services with ease.

📰 https://news.ycombinator.com/item?id=48408061

🆔 @hackernewsgithubprojects

📦 coaxk/subarr

Stop Guessing: The Subtitle Manager You Didn't Know You Needed

Subarr serves as a coordination and quality layer for subtitle stacks, operating alongside Bazarr to bring precision to media library management. It resolves subtitle gaps by verifying file content through calibrated audio-language detection rather than relying solely on metadata. The project includes features like a Tuning Lab for Whisper optimization, provider success leaderboards, and an interface to manually queue transcriptions for specific files. By integrating with Sonarr, Radarr, and Tautulli, it prioritizes tasks based on real playback data. This service ensures your subtitle library is accurate and intelligently managed, allowing you to focus on what matters most.

📰 https://news.ycombinator.com/item?id=48460990

🆔 @hackernewsgithubprojects

📦 diekarotte/spatial-omni

Spatial-Omni: Enhancing AI Audio Understanding

Spatial-Omni is the multimodal integration framework that finally enables large language models to process complex spatial audio cues. Standard models often treat audio as simple monaural signals, but this project introduces a lightweight spatial encoder that injects first-order ambisonic data as independent tokens. This allows models to perform sound localization and advanced spatial reasoning by turning 3D audio inputs into actionable context. With a comprehensive dataset containing millions of question-answering pairs, this tool empowers developers to build AI systems that truly understand where sounds originate in a scene. Try implementing this architecture to unlock deeper environmental awareness in your next audio-language model.

🆔 @hackernewsgithubprojects

📦 upb-cn/vllm-startup-profiler

Analyzing vLLM Cold Start Latency

Startup latency is a critical bottleneck for large language models, but this toolkit lets you measure and predict that delay with precision. The project provides an automated testing framework to profile vLLM startup times across various configurations and hardware environments like different GPUs and CPUs. Beyond just benchmarking, it includes a predictor that estimates startup latency based on specific model architectures and system infrastructure. By automating the extraction of graph sizes and model configurations, it helps developers understand performance overhead before deployment. This repository serves as a valuable resource for identifying and mitigating cold start performance challenges.

🆔 @hackernewsgithubprojects

📦 shanraisshan/claude-code-hooks

Give Your AI Coding Assistant a Voice

Enhance your AI coding sessions by adding distinct audio feedback for every action your terminal assistant takes. This project provides a robust system that triggers custom sound notifications across thirty different hook events, such as session starts, tool usage, or task completion. It bridges the gap between silent command-line interactions and audible status updates, making it easier to track progress without constantly monitoring your screen. By integrating these sound hooks, you can create a more responsive and immersive development environment. Dive into the repository to explore the setup and bring a new layer of interactivity to your workflow.

🆔 @hackernewsgithubprojects

📦 green-pt/honey-for-devs

Slash AI Coding Costs with Honey

Honey is a cross-tool coding skill that significantly cuts token usage and API costs for AI coding agents by reducing both code and prose output without sacrificing quality. It works by applying three reflex-style levers: prioritizing the simplest functional code, removing conversational filler, and utilizing dense, token-efficient formats for agent-to-agent communication. Compatible with major platforms like Claude Code, Cursor, and GitHub Copilot, it maintains correctness through safety carve-outs while providing a reproducible benchmark to verify performance. Adopting this writing style allows developers to optimize their agentic workflows, ultimately leading to faster, more cost-effective coding sessions and smarter resource management.

🆔 @hackernewsgithubprojects