📦 vectorize-io/agent-memory-benchmark

Agent Memory Benchmark: The Honest Test for AI Memory

Agent Memory Benchmark is the open evaluation harness that finally exposes whether an AI agent actually remembers or just guesses. Built by Vectorize, it solves the problem that older tests were too easy for modern models to pass by accident. Instead of just checking if an answer is correct, it measures the real cost and speed of finding that information. The most surprising part is that it treats memory like a strict exam where returning wrong details is just as bad as forgetting them entirely. If you are building agents that need to hold knowledge over time, this is the tool that tells you the truth.

🆔 @hackernewsgithubprojects

📦 ninjahawk/livenerf

livenerf: The Nerf Detector

Track whether a frontier AI model gets quietly worse right after it launches. Months of online reports claimed that major AI companies degrade their tools days after release, but without a fixed starting point, everyone was just guessing. This project, called livenerf, fixes that problem by running the exact same set of difficult math and science questions every single day for thirty days. It locks down the software version and the questions so nothing else can change. By comparing every daily result against the first week, it uses simple statistics to spot a real drop in performance.

📰 https://news.ycombinator.com/item?id=49809555

🆔 @hackernewsgithubprojects

📦 makson179/bello

bello

AI software developers are great until they wander off and make mistakes on long tasks. Bello is a local tool that acts as a manager for your AI coding agents, keeping them on track and preventing them from running unsafe commands. It wraps your existing coder in up to four specialized roles, including a runtime supervisor to stop logical drift, an independent reviewer to catch bugs, and even an adversary that tries to break the code with edge cases. Best of all, it includes a local model that shrinks command logs so your AI reads less text, saving you up to sixty-six percent on model subscription limits.

🆔 @hackernewsgithubprojects

📦 pixaroma/comfyui-pixaroma

Pixaroma: The Visual Editor for ComfyUI

ComfyUI users can now build animated video loops that pulse and shake in perfect sync with music beats. Pixaroma turns the standard node editor into a full creative suite, adding features like audio-reactive motion and text overlays that were previously missing. You can load a video, apply visual effects driven by the soundtrack, and export the result directly as an MP4 file. It bridges the gap between raw AI generation and polished post-production. If you want your generated images to actually move and feel alive, this is the tool that makes it happen without leaving your current workflow.

🆔 @hackernewsgithubprojects

📦 mlhher/late-cli

Late: The AI Agent That Keeps Your Context Clean

Late is the AI coding agent that keeps your model sharp by physically isolating its working memory. Standard AI tools get dumber as they work because every error log and file read piles up in one messy memory space. Late solves this by acting like a project manager who hires temporary workers for specific jobs, then fires them immediately after the task is done. Only the final results go back to the manager, keeping its mind clean. This means a small model can handle massive coding projects without forgetting what it was doing.

🆔 @hackernewsgithubprojects

📦 yappologistic/spun

Spun: The Music Player That Looks Like Hardware

Spun turns your Linux computer into a room full of physical music hardware, projecting album artwork onto spinning CDs, vinyl records, cassettes, and even a retro tape recorder. It solves the problem of sterile digital interfaces by giving your music collection tactile, animated depth that responds to every beat. You can plug in local files, connect to self-hosted libraries, or stream anonymously from YouTube without needing an account. The interface even includes a working mixing board for adjusting audio levels right on your desktop screen.

🆔 @hackernewsgithubprojects

📦 anomalyco/opencode-bench

How to Benchmark AI Coding Agents

Measure exactly how well artificial intelligence agents write real code. OpenCode Bench runs those agents against actual production commits from popular open source projects, then compares their work against the official changes. It uses three independent large language models to grade the output on five specific qualities like logic and test coverage. This framework solves the problem of guessing which AI is actually useful for serious software development. It gives you concrete, mathematically rigorous scores based on real world performance rather than vague marketing claims. If you are choosing an AI coding assistant, this is the tool that tells you who actually delivers.

🆔 @hackernewsgithubprojects

📦 netresearch/agent-rules-skill

Agent-Rules-Skill: The AI Onboarding Cheat Sheet

Agent-Rules-Skill is the project context generator that finally gives your AI coding assistant the exact manual it needs to write good code. Instead of guessing how your project works, this tool scans your repository to automatically build a dedicated instruction file specifically for artificial intelligence. It looks at your build scripts and directory structure to create a concise guide that tells the AI exactly what to do, what to avoid, and which commands to run. This ensures your code assistant understands your unique setup immediately, saving you from the frustrating cycle of correcting bad code suggestions.

🆔 @hackernewsgithubprojects

📦 bilibili/index-translate

Bilibili's Index-Translate Family

Bilibili just open-sourced a translation model family that can actually keep a character's name consistent across a thirty-two thousand token fantasy novel. While standard models often forget names or swap them out for generic titles, Index-Translate uses full-document context to lock in specific terms, ensuring the protagonist stays the same from the first page to the last. It also handles speech dubbing and even adjusts translation length to fit specific subtitle timings. This is a massive step forward for anyone trying to localize complex fiction or media without manual editing. If you are working on long-form localization, this is the tool that finally solves the consistency nightmare.

🆔 @hackernewsgithubprojects

📦 corvofeng/atv-core

Control Your Mac With Your iPhone

Turn your Mac into a target for the built-in Apple TV Remote widget on your iPhone. This Rust-based project, called atv-core, speaks the same private protocols as real Apple hardware, so you can pair your phone and control your computer without buying a remote or jailbreaking anything. The coolest part is that it turns your iPhone's touchpad into a precise mouse cursor for your desktop, letting you swipe to move the pointer and tap to click. It runs in the background with a tiny footprint and even includes a web page where you can watch your finger movements in real time.

🆔 @hackernewsgithubprojects

📦 edenfunf/reelmimic

reelmimic

Take a single video clip and receive a brand new video in that exact same visual style, all through the power of collaborative artificial intelligence. This project acts as a bridge between your creative taste and actual production, letting you point at a video you love and have a team of AI agents plan, build, and review a fresh piece for you. Instead of struggling with complex software interfaces or technical scripts, you simply describe the vibe you want, and the system handles the heavy lifting of animation and editing. It turns the process of making high-quality motion graphics into a conversation rather than a technical hurdle.

📰 https://news.ycombinator.com/item?id=49906508

🆔 @hackernewsgithubprojects

📦 internlm/intern-decision

Intern-Decision: The AI That Chooses For You

Intern-Decision is a specialized AI model that looks at a screenshot and answers specific questions about what to do next, all in under fifty milliseconds. It is not a general chatbot; it is a decision engine designed to pick the right action, whether that is moving a mouse, clicking a button, or choosing a game move. By freezing the visual parts of the model and only training the language logic, it stays incredibly fast while still being accurate. The project even includes a test to check if the model’s confidence is honest, which is rare in this field.

🆔 @hackernewsgithubprojects

📦 nyaomaru/is-kit

is-kit: The TypeScript tool that makes type checking feel like normal code

is-kit is the TypeScript toolkit that finally makes runtime validation feel like normal coding. TypeScript checks your types when you write code, but it cannot check the actual data arriving from an API or a file at runtime. is-kit solves this by letting you build tiny, reusable functions that verify data as it enters your app. Instead of wrestling with massive schema libraries, you compose small checks using simple logic like and and or. This keeps your code clean and fast, while ensuring your program never crashes on unexpected data. It is the lightweight way to trust your data without slowing down your development.

🆔 @hackernewsgithubprojects

📦 nanaism/yomiyasu

Yomiyasu: The Anti-AI Filter for Japanese

Rewrite robotic AI drafts into natural, human-sounding Japanese. Yomiyasu fixes the subtle structural quirks that make machine-generated text feel stiff, replacing vague metaphors with clear, direct language. It doesn't just swap words; it reconstructs the underlying sentence logic to ensure the meaning stays accurate while the flow becomes effortless. This is the fix for that distinct, uncanny vibe you get from large language models. If you publish in Japanese, this is your secret weapon for sounding like a real person instead of a robot. It turns digital noise into clean, readable prose.

🆔 @hackernewsgithubprojects

📦 cybersecurity/halo-ce-universal

Halo CE Universal Brings Classic Xbox Shooter to Android

Halo Combat Evolved, the definitive original Xbox shooter, now runs natively on Android, Windows, and Linux through a single unified codebase known as halo ce universal. This project isn't just a simple port; it is a massive reverse engineering effort that reconstructs the game's internal logic from scratch to work across totally different operating systems. The coolest part is that it takes a console exclusive from 2001 and makes it playable on your phone, preserving the original gameplay feel without relying on clumsy emulation. This means you can grab your phone and jump back into the fight against the Flood, anywhere you go.

🆔 @hackernewsgithubprojects

📦 edgcpp/compiler

compiler: The C++ Front End

The EDG compiler front end reads C plus plus code and converts it into plain C, letting developers swap out the complex back end of major compilers for a simpler, fully open source engine. It focuses entirely on parsing and bug emulation, ensuring that any source code that compiles in Clang, G++ or Microsoft Visual Studio will behave identically here. This open source project provides the critical first step of translation, complete with a prelinker for automatic template handling and a minimal runtime library.

📰 https://news.ycombinator.com/item?id=49915484

🆔 @hackernewsgithubprojects

📦 otakit/otakit

OtaKit: Ship Updates Without the App Store

OtaKit lets you push fresh code straight to your phone without waiting for the app store. Instead of the usual weeks of review, you just build your web assets, run one command, and the new version is live. Your app checks for it quietly in the background, downloads it safely, and swaps it in on the next restart. If something breaks, it automatically rolls back to the last good version so your users never see a crash. It is fully open source and self-hostable, so you own the entire pipeline. No more waiting on gatekeepers for hot fixes. Just ship.

📰 https://news.ycombinator.com/item?id=48759570

🆔 @hackernewsgithubprojects

📦 agent-ecosystem/afdocs

AFDocs: The Linter for AI-Readable Docs

Hand your documentation to an AI agent and watch it stumble over your site. That is exactly what AFDocs does for you before your users do. Think of it as a linter, but instead of catching syntax errors, it checks if your docs are actually readable by machines. It crawls your site to see if AI agents can easily find and digest your content, giving you a specific score and a list of exactly what is broken. This is huge because if your docs are a mess for bots, your AI-powered development workflows are going to fail too. It is the quality assurance step for the agent economy.

🆔 @hackernewsgithubprojects

📦 vast-ai-research/mira-scene

Mira-Scene: 3D Worlds From One Photo

Mira-Scene is the 3D reconstruction tool that places objects exactly where they belong in a single photo. It solves the tricky problem of turning flat images into realistic 3D spaces by using pixel-aligned layouts to guess exact object positions. The tool analyzes your photo to detect depth and object boundaries, then builds a complete scene with accurate spatial relationships. This is useful for developers who need detailed 3D environments without expensive scanning equipment. The surprising part is how accurately it can predict where furniture sits on floors or how objects hang from walls. You get a fully interactive 3D scene ready for games or simulation.

🆔 @hackernewsgithubprojects

📦 githubsecuritylab/gh-secure

Gh Secure: Fix GitHub Security in 2 Minutes

Lock down your GitHub project in just two minutes using Gh Secure, a tool built by the GitHub Security Lab. It automatically turns on five essential safety features that most people forget to enable, protecting your code from leaked secrets, broken dependencies, and unwanted changes. The most impressive part is that it works directly through the command line without needing you to be a security expert or know how to code. You simply run one command, and it handles the complex settings for you, ensuring your project follows industry best practices.

🆔 @hackernewsgithubprojects