📦 jaineet17/causal-agent-replay

Find the Exact Step That Broke Your Agent

Causal agent replay lets you pinpoint exactly which step caused an LLM agent to fail by rewinding the decision and running it again. Instead of just watching a trace, you intervene on specific moments to see if the bad outcome changes. If changing that step fixes the result while later steps do nothing, you’ve found the true cause. It’s like time-travel debugging for AI, using math to prove where the error actually happened so you can fix it for real.

🆔 @hackernewsgithubprojects

📦 estebanpdn/zelda-alttp-3ds

Zelda A Link to the Past on 3DS

Play a classic Nintendo adventure on your handheld by using this clever port that brings the beloved Zelda sequel to life on the Nintendo three dimensional system. The repository takes the original two-dimensional masterpiece and adapts it for the console’s unique dual screen layout so the gameplay feels fresh yet familiar. You get to explore the same magical world with enhanced visuals that fit perfectly in your hands without losing the charm of the original. It is a neat piece of fan-made engineering that proves old games can still feel brand new when given the right hardware.

🆔 @hackernewsgithubprojects

📦 canivibecodeit/canivibecodeit

Can I Vibecode It?

Canivibecodeit is the directory that finally kills your SaaS bills with one AI prompt. It answers whether you can replace paid subscriptions like scheduling apps or analytics tools with a single, copy-paste instruction for an AI coding agent. Instead of vague advice, it gives you a concrete verdict, the exact code to build a personal version, and an honest list of what functionality you actually lose in the process. Think of it as a utilitarian death list for monthly expenses, where traffic-light badges tell you instantly if a service is truly replaceable.

🆔 @hackernewsgithubprojects

📦 pfcdev/taskbarwidgets

Taskbar Widgets

Taskbar Widgets places live, customizable mini-apps directly on your Windows 11 taskbar without replacing the system shell. It solves the frustration of cluttered desktops by letting you add real-time system monitors, weather updates, and media controls right where your mouse lives. You can drag these widgets to any position, arrange them side by side, or make them rotate automatically. The best part is the Parking Lot feature, a tiny shelf where you drop files, links, or text to carry them between apps later. It keeps your workflow close and your desktop clean, giving you quick access to the data you actually need without opening heavy windows or extensions.

🆔 @hackernewsgithubprojects

📦 qwen-code-dev-bot/oh-my-cli

oh-my-cli: The Self-Driving Terminal Assistant

oh-my-cli is a terminal assistant that actually fixes code for you instead of just suggesting it. It connects to an open ai compatible model to autonomously read your files and run shell commands to complete tasks like refactoring or debugging. You simply give it a goal in plain english and watch it navigate your project directory, edit files, and run tests without you lifting a finger. It is fascinating because it handles the messy parts of development while keeping you in control through a clear approval system. You get to see exactly what it changes and can hit undo if it goes off track.

🆔 @hackernewsgithubprojects

📦 genspark-ai/genoffice

GenOffice AI Office Suite

GenOffice lets you edit Word, Excel, PowerPoint, and PDF files while an AI agent actively rewrites your content instead of just chatting at you. The coolest part is how it handles your original documents. It opens a file, keeps the untouched parts exactly the same bytes, and only regenerates the specific paragraphs the AI changed. This means you can save and reopen the file later without breaking the layout or losing your formatting. It works the same way for spreadsheets and slides, patching only what needs updating while leaving the rest of your data alone.

📰 https://news.ycombinator.com/item?id=49157441

🆔 @hackernewsgithubprojects

📦 wie-project/kakehashi

Run macOS Apps on Linux ARM64

Kakehashi lets you run macOS command line tools directly on Linux ARM64 without needing a Mac. It works like a translator that reads macOS binaries and makes them behave on Linux, handling the differences in how files and systems work. You can install tools like curl or 7zip and run them on a Linux machine as if they were native. This is super useful because you can use cheaper Linux servers for tasks that usually need expensive macOS hardware. It is not perfect speed-wise but it opens up new ways to use your Linux setup for macOS workflows.

📰 https://news.ycombinator.com/item?id=49145937

🆔 @hackernewsgithubprojects

📦 notatestuser/ds4-control

Run Massive AI Models on Mac

Control a local DeepSeek V4 model right from your Mac menu bar without touching the command line. This app acts as a smart pilot for your computer’s brain, letting you pick between a super-smart but heavy Pro variant or a faster Flash version based on how much RAM you actually have. It automatically handles downloading the massive model files, keeps an eye on your GPU and memory usage with cute little widgets, and even lets you chat or write code with it instantly.

🆔 @hackernewsgithubprojects

📦 patchy631/time-to-first-token

The 10-Week Roadmap to LLM Inference

The Time to First Token project transforms abstract machine learning theory into a concrete production skill in ten weeks. It guides you to build a single OpenAI-compatible inference service that grows stronger each session rather than running disjointed experiments. You deploy a model, instrument it with real-time metrics, and stress-test it against thousands of simultaneous users. The roadmap then teaches you how to optimize that same service using quantization, speculative decoding, and cost-aware routing. This structured approach saves setup time and leaves you with a reproducible benchmark and a serving stack you actually configured. It is the clearest path from reading about AI to shipping it.

📰 https://news.ycombinator.com/item?id=49158330

🆔 @hackernewsgithubprojects

📦 yan5xu/codexloom

CodexLoom: AI Agents That Remember

CodexLoom is the local-first AI environment that finally gives your code assistants permanent memory and long-term responsibility. Instead of losing context every time you start a new chat, CodexLoom turns a single conversation thread into a durable Domain Agent that keeps its history, personality, and accumulated knowledge forever. You can then organize multiple specialized agents into a governed team where they collaborate through bounded messages and handoffs, much like colleagues in a real company. It also lets these agents interact with you directly in Slack or Feishu through an Interface Agent, bridging your AI workspace with your daily tools.

📰 https://news.ycombinator.com/item?id=49159174

🆔 @hackernewsgithubprojects

📦 xlang-ai/qwen-cua

Qwen-CUA: An AI That Actually Uses Your Computer

Take a screenshot, ask a question, and watch Qwen-CUA actually use your mouse and keyboard to get the job done. This open-source project creates an AI agent that operates your computer just like a human does, relying entirely on what it sees on the screen rather than hidden code or system access. It reasons through visible interfaces, clicks buttons, types text, and drags elements without needing special APIs or developer tools. The standout feature is its remarkable visual memory, which tracks dozens of recent screenshots to handle complex, multi-step tasks without losing its place.

🆔 @hackernewsgithubprojects

📦 ace-step/ace-step-1.5

Ace-Step 1.5: The Local Music Generator

Ace-Step 1.5 is the open-source music model that finally lets you generate radio-quality songs right on your own laptop without relying on cloud subscriptions. Think of it as a digital composer sitting inside your computer, capable of turning a simple text prompt like "upbeat jazz with saxophone" into a full, structured track with lyrics and instrumentation in just a few seconds. It handles everything from short loops to ten-minute compositions and even lets you tweak the mood or rewrite the lyrics after the fact.

🆔 @hackernewsgithubprojects

📦 pierpaolo28/awesome-fde-roadmap

Awesome Forward Deployment Engineering Roadmap

Awesome Forward Deployment Engineering Roadmap is the definitive career guide that turns complex software deployment into a strategic mission. It is not a standard coding tutorial. It teaches you how to act as a technical diplomat who bridges the gap between perfect code and messy real-world client needs. The repository maps a structured path covering data engineering, cloud architecture, and AI strategy, inspired by elite roles at major tech firms. Its most compelling feature is the focus on the delta, the specific work required to make enterprise software function in hostile environments. You will learn to navigate political roadblocks, audit legacy data, and design secure cloud systems that actually work.

📰 https://news.ycombinator.com/item?id=49161191

🆔 @hackernewsgithubprojects

📦 texpile/texpile

Texpile: The Offline LaTeX Editor

Texpile is the offline LaTeX editor that lets you type in formatted text without ever seeing a single line of code. It solves the headache of learning complex markup by rendering your document as you write, so you see the math, tables, and figures appear in real time while the app updates the preview instantly. Your files stay completely private on your own machine with no cloud storage required, making it perfect for academics who want a simple, focused writing tool.

📰 https://news.ycombinator.com/item?id=49162204

🆔 @hackernewsgithubprojects

📦 real-stanford/iphumi

iPhumi: The iPhone Robot Gripper

iphumi is the handheld iPhone gripper that turns a simple phone into a powerful robot training tool. Most robot training requires expensive cameras and complex setups, but iPhumi swaps that out for an iPhone mounted to a standard 3D printed gripper. You just grab the handle, move the robot’s hands through motions, and the app automatically records high-resolution video, depth data, and even your voice instructions in real time. The phone tracks the fingers using visual markers, so you get precise movement data without any sensors on the robot itself.

🆔 @hackernewsgithubprojects

📦 cloudflare/computer

Computer: Give Agents a Real Computer

Cloudflare Computer lets artificial agents interact with a real filesystem stored in a cloud database. It creates a virtual disk that lives inside a distributed storage service, keeping everything in sync even when the agent is running in different places. You can give an agent a full Linux environment with real binaries and networking, or just a simple coding space in the cloud. This means your code assistant can actually write files, run commands, and see the results directly, rather than just pretending to. It bridges the gap between chatting with a bot and letting it do real work.

📰 https://news.ycombinator.com/item?id=49155280

🆔 @hackernewsgithubprojects

📦 filipj9/hermes-control-deck

Hermes Control Deck: Your Phone as an AI Command Center

Hermes Control Deck turns your smartphone into a real-time command center for managing local AI coding assistants without needing a desktop monitor. This project acts as a mobile web app that connects to your local code tools, letting you send prompts, watch the AI work in real time, and approve tasks right from your phone screen. It works by bridging your local code command line and web interface into one simple dashboard that updates instantly. You do not need special desktop software or private tools to use it. Instead, it just needs your existing code setup and a secure local network connection.

🆔 @hackernewsgithubprojects

📦 sb-ai-lab/sirin

Spot AI Lies Instantly with Sirin

Sirin lets you catch AI hallucinations by spotting when a language model contradicts its own source material. Instead of guessing, this Python tool scans your text to highlight exactly which phrases are unsupported or completely made up. It works by analyzing how confident the AI feels about its answers, flagging moments when it is uncertain or stretching the truth. You get both a simple yes or no on the whole response and a detailed map showing the specific words that are problematic.

🆔 @hackernewsgithubprojects

📦 meshy-dev/meshy-t2

Meshy T2: Fast Native Mesh Generation

Meshy T2 is the open source project that brings fast, native 3D mesh generation using flow matching technology. It solves the slow and complex process of turning simple ideas into detailed 3D models by offering a much quicker way to create these digital shapes. The core feature is its ability to generate clean mesh structures directly, avoiding the messy, jagged results common in other AI tools. This matters because creators can now iterate on 3D assets in real time without waiting minutes for each render or struggling to clean up broken geometry. It is a significant step forward for anyone who needs reliable 3D content quickly.

🆔 @hackernewsgithubprojects

📦 legostin/cull

Cull

cull is a terminal tool that instantly turns your cluttered hard drive into an interactive map, letting you hunt down and delete space-eating files without ever leaving your keyboard. Instead of guessing which folders are heavy, it builds a visual treemap right in your terminal where every rectangle’s size matches the actual bytes it consumes. You can click or use arrow keys to zoom into directories, spot the biggest offenders, and safely move junk to the trash or delete it permanently. It even spots hidden culprits like old app caches, build artifacts, and invisible system backups, warning you away from critical system paths.

📰 https://news.ycombinator.com/item?id=49164541

🆔 @hackernewsgithubprojects

📦 deepexperience/vad_multimodal_opd

VAD Multimodal OPD: Better AI Vision Training

VAD Multimodal OPD is the open-source training framework that teaches AI to learn from its own mistakes by comparing clear and blurry versions of the same image. It solves the problem of how to make vision models smarter without needing endless expensive human feedback. Instead of just copying a teacher model, it creates three views of each picture: the original, a sharp crop, and a deliberately degraded version. The AI then learns by spotting the differences between the clear and blurry images, using those gaps to improve its own understanding. This counterfactual approach helps the model correct its own errors more effectively.

🆔 @hackernewsgithubprojects