📦 epoch-research/mirrorcode

MirrorCode

An artificial intelligence model once spent nineteen straight days working on a single problem in this evaluation suite. This is mirrorcode, a rigorous benchmark designed to test the limits of machine learning on complex programming tasks. Built on the Inspect framework, it forces artificial intelligence to tackle real-world software engineering challenges using actual source code and documentation from popular tools like Ruff and JQ. Instead of simple coding puzzles, it evaluates how systems handle deep, multi-step engineering workloads over long periods. It is a fascinating look at how researchers measure the boundary between human developers and next-generation software agents.

🆔 @hackernewsgithubprojects

📦 fansunqi/videotool

VideoTool

Large multimodal AI models can now analyze video questions with unprecedented accuracy by strategically combining specialized temporal and spatial tools. videotool is an open-source framework from NeurIPS 2025 that equips language models with a collaborative toolkit containing frame selectors, object trackers, and image captioners. Instead of relying on a single AI to guess its way through a video, this project coordinates these specialized tools to progressively narrow down and isolate exactly where and when an action occurs. It dramatically reduces errors and boosts video understanding performance, making it a must-watch helper for building smarter video applications.

🆔 @hackernewsgithubprojects

📦 open-gsd/gsd-core

Gsd Core

The software development companion that actually gets your code out the door without the typical chaos, gsd-core is a framework designed to turn manual terminal work into structured, issue-driven developer workflows. Instead of jumping between endless AI tools, it uses specialized agents and built-in commands to coordinate everything from planning and code reviews to debugging and deployment. It acts like an automated project manager and quality assurance team living right inside your command line, verifying your code changes and updating your documentation dynamically as you work. Give it a spin on your next project to see how easy shipping code can really be.

🆔 @hackernewsgithubprojects

📦 danielmiessler/lifeos

lifeos

Build a personal digital command center that aligns your daily actions with your biggest life goals. This repository, called lifeos, provides the actual framework to run what it calls a personal AI infrastructure. Instead of juggling random productivity apps, it combines your goals, habits, and tasks into a single system powered by a local dashboard and a digital assistant. The core engine analyzes your current state, maps out where you want to go, and helps you take the next step. It is a fascinating way to use modern AI as a highly personalized coach that keeps you focused on what actually matters.

🆔 @hackernewsgithubprojects

📦 cloudflare/skills

Cloudflare Skills

AI coding assistants can now get direct, native instructions to build and manage full cloud applications on Cloudflare. The skills repository is a collection of pre-made capabilities designed for popular AI agents like Claude Code and Cursor. Instead of relying on outdated training data, this project equips your agent with specialized knowledge for databases, AI models, and secure code sandboxes. It even includes remote servers that let your AI browse live documentation and manage active deployments. It is a fantastic way to turn a generic coding assistant into a dedicated cloud engineer that gets things right the very first time.

🆔 @hackernewsgithubprojects

📦 mindmuxai/brain.md

brain.md

Keep your coding agents on the exact same page across every single session by giving them a shared, persistent memory. This zero-dependency command line tool scaffolds a plain markdown brain file directly inside your repository. Instead of losing critical decisions, project constraints, and architectural context the moment a chat ends, agents like Claude Code read and write to this local file to instantly recall why choices were made. By routing all memory updates through a simple command line interface, the project structurally prevents malformed edits, keeping your project's durable knowledge base clean, portable, and completely independent of any specific AI model.

📰 https://news.ycombinator.com/item?id=48717245

🆔 @hackernewsgithubprojects

📦 pelski/ytzero

YT Zero

YT Zero is the self-hosted YouTube subscription reader that finally lets you watch the channels you love completely on your own terms. It pulls public feeds directly into a local database, giving you a quiet inbox to filter, watch, and organize videos without a Google account, API keys, or an annoying algorithm dictating what you see. The coolest part is the smart watch-later bucket system, letting you sort videos into custom schedules like Tonight, Tomorrow, or Weekend so you can tackle your queue when you actually have time. It is an incredibly clean, distraction-free way to reclaim your subscription feed.

🆔 @hackernewsgithubprojects

📦 jeikeilim/tenet

Tenet

Supercharge your coding assistant into a relentless software engineer that builds entire features completely on autopilot. Tenet is a smart wrapper that orchestrates popular AI agents like Claude Code, keeping them on track for continuous twelve-hour development sessions without losing their minds. Instead of letting an AI make lazy assumptions, it interviews you, drafts a spec, visualizes the architecture, and breaks everything down into a dependency graph. The coolest part is the evaluation pipeline: it forces separate, independent AI critics to grill the code, tests, and user interface, rejecting any lazy shortcuts. It is the ultimate tool for turning fragile AI sessions into a robust, self-healing developer workflow.

🆔 @hackernewsgithubprojects

📦 dingwu1021/promsa

Promsa

Most visual assistants get stuck on hard questions because they use a rigid, one-shot search pipeline that cannot recover from a bad first guess. Promsa changes this by turning the AI into a budget-aware researcher that dynamically decides how to search. During a task, it analyzes the clues gathered so far and chooses whether to run a reverse-image search, perform a text query, or stop once it has enough evidence. This makes visual question answering incredibly flexible, letting the assistant chain multiple search steps and switch strategies on the fly until it finds the correct answer.

🆔 @hackernewsgithubprojects

📦 j-c-zhao/structsplat

StructSplat

Imagine creating a fully detailed three-dimensional digital scene from just a few smartphone photos, without even knowing the camera settings or angles. StructSplat does exactly that by predicting three-dimensional gaussians directly from uncalibrated, sparse images. Traditional methods require precise camera positions or heavy per-scene processing to stitch images together, but this system uses a clever structured approach. It separates texture, geometry, and semantic cues to keep everything highly accurate and globally consistent. It is a massive step forward for virtual reality and game design because it lets you reconstruct rich, complex spaces instantly from a handful of casual snapshots.

🆔 @hackernewsgithubprojects

📦 mat-mgm/kb-prolog

kb-prolog

You can now make complex statements about other statements and track their entire history like code using kb-prolog. This local-first knowledge base lets you build nested, hyper-relational graphs where a claim can easily describe another claim, such as recording that one person doubts what another person said. It combines a Prolog runtime with SQLite for storage, automatically saving and tracking version changes so you can travel back in time to see exactly how your data evolved. Plus, it features built-in file storage and an interactive graph visualizer to browse everything easily. It is a brilliant way to bring deep, logical relationships and version control to your local data.

📰 https://news.ycombinator.com/item?id=48697599

🆔 @hackernewsgithubprojects

📦 uw-syfi/tracelab

tracelab

Peek behind the scenes of real-world coding assistants by analyzing the exact logs they generate during development. Under the hood, tools like Claude Code emit highly complex event sequences, from tool calls to split-second timing events. This repository gives you a massive, pre-sanitized dataset of over three hundred thousand developer rounds alongside a local pipeline to clean and inspect your own assistant history. You can run prebuilt data experiments, reconstruct prompt-cache behavior, or drag logs directly into a web app to see interactive charts. It is a fantastic way to understand how agent infrastructure behaves under load, helping you build smarter, more predictable developer tools.

🆔 @hackernewsgithubprojects

📦 clawdbrunner/openclaw-graphiti-memory

openclaw-graphiti-memory

Multi-agent AI teams can now share a unified, evolving memory without running into constant file conflicts. The openclaw-graphiti-memory project introduces a clever three-layer memory system designed specifically for OpenClaw setups. It combines private local files for individual agent tasks, read-only shared files for stable team rosters, and a dynamic, temporal knowledge graph powered by Graphiti and Neo4j. This means agents can query a central brain to recall exactly when a decision was made or what another agent discovered, all while keeping their own working notes separate. It is a practical blueprint for building collaborative, well-coordinated digital workforces.

🆔 @hackernewsgithubprojects

📦 miaai-lab/deepseek-v4-flash-dspark-2x-dgx-spark

deepseek-v4-flash-dspark-2x-dgx-spark

deepseek-v4-flash-dspark-2x-dgx-spark is the multi-node recipe that finally lets you run DeepSeek's fast model at scale without crashing under heavy traffic. Normally, running this model with speculative decoding across multiple hardware nodes causes memory confusion and errors as soon as multiple people ask questions at the same time. This project solves that by applying a clever concurrency patch that links memory slots to specific user sessions, preventing data corruption. Now you can easily run up to sixteen fast, simultaneous conversations with zero errors. Grab the code and scale your local setups tonight.

🆔 @hackernewsgithubprojects

📦 keijiro/audiotailor

Audio Tailor

audiotailor is the Unity editor extension that lets you clean up and loop sound assets without ever leaving your game engine. Instead of jumping back and forth to an external audio editor just to fix a messy file, you can trim dead air, boost volume, and convert files to mono right inside your inspector. The coolest part is how it handles looping. It blends the start and end of a sound together using a smooth crossfade, making it incredibly easy to build seamless ambient tracks and music loops. It is a massive time-saver for keeping your audio workflows simple.

🆔 @hackernewsgithubprojects

📦 plummerssoftwarellc/tinyretropad

TinyRetroPad

Build a fully functioning, Notepad-style Windows text editor that fits into an unbelievably small two-and-a-half kilobytes. That is where tinyretropad comes in. It is a size-obsessed, minimalist text editor written in x86 assembly. By stripping away every piece of unnecessary overhead and using clever compression at build time, it packs a complete set of classic menus, file dialogs, printing capabilities, search and replace, and a real-time status bar into a tiny binary. It is a masterclass in hyper-efficient programming that squeezes maximum utility out of every single byte. Grab the code, compile it yourself, and see how much punch you can pack into a microscopic file.

📰 https://news.ycombinator.com/item?id=48732565

🆔 @hackernewsgithubprojects

📦 feelingai-team/codebrain

CodeBrain

CodeBrain is the model context protocol plugin that finally gives your coding agents real compiler-grade diagnostics and structural code search. Instead of letting an AI agent blindly poke around your codebase and burn through expensive tokens, this tool connects right to your language servers to map your files, find symbols, and instantly validate changes. It catches type errors and broken imports before running a full build, auto-recovering when agents get stuck in loops or crash. It works out of the box with Python, TypeScript, and Go, reducing token costs by over sixty percent. Grab it today to make your coding agents genuinely smart.

🆔 @hackernewsgithubprojects

📦 vercel-labs/knowledge-agent-template

Knowledge Agent Template

This open-source template lets you deploy AI agents that can search and understand your documentation without using any vector databases, chunking pipelines, or embedding models. Instead, the system spins up lightweight, isolated sandboxes and runs familiar command-line searches like grep, find, and cat directly across your files for instant and highly predictable results. It includes an administrative interface where you can monitor usage, sync content from sources like GitHub, and even chat with an AI admin agent that analyzes your server logs. Ultimately, it provides a clean, self-hostable foundation for building smart bots that instantly run on web chat, Discord, and GitHub.

🆔 @hackernewsgithubprojects

📦 gommzystudio/device-activity-tracker

Device Activity Tracker

Device-activity-tracker is the proof-of-concept tool that demonstrates how easily your daily messaging habits can leak to outsiders without your knowledge. By measuring the round-trip time of silent delivery receipts on apps like WhatsApp and Signal, it detects if your device is active, in standby, or offline. When you receive a background message probe, your phone automatically sends back a quick receipt; because your processor responds faster when you are actively using the screen than when it is locked in your pocket, this tool uses those tiny millisecond differences to map your exact daily routines. It highlights a massive, unpatched privacy gap in modern secure messaging.

🆔 @hackernewsgithubprojects

📦 nicobailon/visual-explainer

Visual Explainer

Turn dry terminal outputs and messy text diagrams into gorgeous, interactive web pages and slide decks you actually want to read. When your AI agent tries to explain a system architecture or run a code review, it usually dumps a wall of unreadable monospace text, broken tables, and ASCII art. This tool fixes that by generating self-contained HTML files complete with clean typography, responsive tables, dark mode, and zoomable flowcharts. You can even generate high-quality slide decks instantly. It is a brilliant way to visually audit plans, review code diffs, and map out projects without drowning in terminal clutter.

🆔 @hackernewsgithubprojects

📦 gvclab/personalive

PersonaLive

Animate any portrait photo in real time using a live webcam feed to drive the facial movements, expressions, and poses. It relies on a clever diffusion framework designed specifically for live streaming, letting you broadcast infinite-length animations without running out of video memory. Instead of lagging behind, it processes your expressions instantly, keeping the generated character completely synced to your real-life speaking rhythm. It is incredibly fun because you can suddenly stream as a completely different character, historical figure, or digital avatar with natural, fluid facial movements. Grab a portrait, start up your webcam, and watch your new persona come to life instantly.

🆔 @hackernewsgithubprojects