Mentions
Conversations linking to GitHub, newest first. Times are UTC.
Mentions by source
X
Bluesky
Hacker News
V2EX
Qiita
DEV
Mastodon
Latest posts
¿EL FIN DE LOS RECOMP, PORTS de Nintendo switch? GitHub y el nacimiento de NUP ¿Y si los Recomp Ports de Nintendo Switch a PC estuvieran en peligro aunque no incluyan las prod.keys? 👀 En este vídeo ponemos a prueba varios juegos recompilados y porteados a PC para comprobar lo bien que funcionan y lo espectaculares que pueden llegar a verse. Pero detrás de estos proyectos hay un problema importante que podría poner en riesgo su futuro. ⚖️ EL PROBLEMA LEGAL DE LAS PROD.KEYS Aunque un proyecto de GitHub no distribuya las claves de cifrado, facilitar su utilización para descifrar juegos podría entrar en conflicto con la legislación anticircunvención. Analizamos la ley citada y por qué esto podría provocar la retirada de determinados proyectos. 🐉 ¿QUÉ PASA CON EDEN Y SU PROPIO GIT? También hablamos de Eden y de cómo disponer de su propia infraestructura de Git cambia el escenario frente a las retiradas de proyectos en plataformas externas como GitHub. 🚀 NACE NUP: ¿UNA NUEVA ALTERNATIVA? Y aquí llega una de las novedades más interesantes: NUP, un nuevo formato de ROM sin cifrar que plantea una alternativa para los Recomp Ports y emuladores, evitando el proceso de descifrado en tiempo de ejecución. ¿Podría ser este el camino para el futuro de los Recomp Ports? 🎮 Gameplay, rendimiento, análisis legal y novedades de una escena que sigue evolucionando. 💬 ¿Qué opinas de NUP? ¿Crees que este formato puede cambiar el futuro de los Recomp Ports? _____________| Contacto |________________ Canal secundario (en català): youtube.com Mi Twitter: x.com Mi Github: github.com/MaSieS4Fun Mi Discord discord.gg Invítame a un café - PayPal: paypal.me #NintendoSwitch #RecompPorts #NUP #Nintendo #EdenEmulator #PCGaming #SwitchToPC #Emulacion #GitHub #ProdKeys #Gaming #PortsPC #ROMs #Emuladores #NintendoSwitchPC [en-US] THE END OF NINTENDO SWITCH RECOMP PORTS? GitHub and the birth of NUP What if Nintendo Switch Recomp Ports to PC were in danger even without including product keys? 👀 In this video, we test several recompiled and ported games to PC to see how well they work and how spectacular they can look. But behind these projects lies a significant problem that could jeopardize their future. ⚖️ THE LEGAL PROBLEM OF PRODUCE KEYS Even if a GitHub project doesn't distribute encryption keys, facilitating their use to decrypt games could conflict with anti-circumvention laws. We analyze the relevant law and why this could lead to the removal of certain projects. 🐉 WHAT ABOUT EDEN AND ITS OWN GIT? We also discuss Eden and how having its own Git infrastructure changes the landscape regarding project removals on external platforms like GitHub. 🚀 NUP IS BORN: A NEW ALTERNATIVE? And here comes one of the most interesting developments: NUP, a new unencrypted ROM format that offers an alternative to Recomp Ports and emulators, avoiding the runtime decryption process. Could this be the future of Recomp Ports? 🎮 Gameplay, performance, legal analysis, and news from a scene that continues to evolve. 💬 What do you think of NUP? Do you think this format could change the future of Recomp Ports? _____________| Contact |________________ Secondary Channel (in Catalan): youtube.com My Twitter: x.com My GitHub: github.com/MaSieS4Fun My Discord: discord.gg Buy me a coffee - PayPal: paypal.me #NintendoSwitch #RecompPorts #NUP #Nintendo #EdenEmulator #PCGaming #SwitchToPC #Emulation #GitHub #ProdKeys #Gaming #PortsPC #ROMs #Emulators #NintendoSwitchPC [pt-BR] O fim das recompilações e ports para Nintendo Switch? GitHub e o nascimento do NUP E se as versões recompiladas de jogos do Nintendo Switch para PC estivessem em perigo mesmo sem incluir chaves de produto? 👀 Neste vídeo, testamos vários jogos recompilados e portados para PC para ver como funcionam e o quão espetaculares podem ficar. Mas por trás desses projetos existe um problema significativo que pode comprometer seu futuro. ⚖️ O PROBLEMA LEGAL DAS CHAVES DE PRODUTO Mesmo que um projeto do GitHub não distribua chaves de criptografia, facilitar seu uso para descriptografar jogos pode entrar em conflito com as leis anti-burla. Analisamos a legislação pertinente e por que isso pode levar à remoção de certos projetos. 🐉 E QUANTO AO EDEN E SEU PRÓPRIO GIT? Também discutimos o Eden e como ter sua própria infraestrutura Git muda o cenário em relação à remoção de projetos em plataformas externas como o GitHub. 🚀 O NASCE DO NUP: UMA NOVA ALTERNATIVA? E eis que surge um dos desenvolvimentos mais interessantes: NUP, um novo formato de ROM não criptografado que oferece uma alternativa aos Recomp Ports e emuladores, evitando o processo de descriptografia em tempo de execução. Será este o futuro dos Recomp Ports? 🎮 Jogabilidade, desempenho, análise legal e notícias de um cenário em constante evolução. 💬 O que você acha do NUP? Acredita que este formato pode mudar o futuro dos Recomp Ports? _____________| Contato | ________________ Canal secundário (em catalão): youtube.com Meu Twitter: x.com Meu GitHub: github.com/MaSieS4Fun Meu Discord: discord.gg Me pague um café - PayPal: paypal.me #NintendoSwitch #RecompPorts #NUP #Nintendo #EdenEmulator #PCGaming #SwitchToPC #Emulation #GitHub #ProdKeys #Gaming #PortsPC #ROMs #Emulators #NintendoSwitchPC
@herdrdev made an Umber Calm theme for the herdr app 🤎 warm, low-glare dark, every color from one palette and checked for contrast. one json file, pick it in Settings → Look. github.com/ChristianOrrala/umber-calm/tree/main/p… t.co
PikoNest's first public beta is out! Play PICO-8, explore Splore and make little games on Android 😁 I finally released the first public beta of PikoNest 😁 preview.redd.it I've been building a little PICO-8 home and workshop for Android handhelds. Sometimes I just want to play; sometimes a tiny game makes me want to draw a sprite or try an idea of my own. I wanted both things on the same device. The playing side is already working, including Splore. PikoNest uses the original PICO-8 files and runs the original engine, so games use PICO-8's own behavior and you get its actual Splore browser. You can discover and download games there, then launch them from a shelf with covers, favorites and recent games. Downloaded games stay available for offline play. For making things, there's a separate Workshop. You can start from a blank cartridge, a little collection game or a lamp puzzle. Pick an element in Game view, edit its sprite or behavior, draw a map, work on animation, sounds and music, then jump into PICO-8 to try it. Projects stay ordinary .p8 cartridges with regular Lua and data, and you can export them to use outside PikoNest. There's also sprite/settings reuse and backup/restore for projects, saves and downloads. I've spent a lot of time on the less exciting parts too — setup, keeping old projects intact and getting back from a test to the editor. This is 0.1.0-beta.1, so the workshop still has room to grow and more involved games can need Lua. Current device checks were on Retroid Pocket Classic; I'd like to hear how it goes on other handhelds. It's one APK for ARM64 Android 8+. You'll need your own purchased PICO-8 Raspberry Pi ZIP containing pico8_64. PICO-8 itself isn't included. Download, setup and screenshots on my website Source code on GitHub Really happy to finally share this one. What would you play or make first? preview.redd.it preview.redd.it preview.redd.it preview.redd.it preview.redd.it preview.redd.it
Fridge Gallery, a self hosted solution to keeping your kids artwork!
Revisited 18/39: Binary Momentum. MACD, RSI and VWAP on the outcome token's own short bars; enter when all three agree. Test walk-forward, net of spread and fees. Main failure: the indicators lag and overlap, so agreement isn't extra evidence. Not advice. github.com/Polytutor-AI-Labs/polymarket-strategy-… t.co
LinuxRunner DeX mode Hey everyone! Here’s a quick hands-on demo of LinuxRunner’s new DeX-style desktop mode, running on my actual iPhone. I recorded this myself to give you guys a better look at how it works in practice. Everything is running locally on the iPhone, no jailbreak, no remote streaming, just LinuxRunner doing its thing. Still a work in progress, but I’d love to hear what you think! GitHub: mxhm11cr/LinuxRunner
@Svityuk1 the article is the full guide, and all the prompts + the render script are here: Elessarxai/ai-workflows
Keychron B1 Pro (2025 hardware revision) — custom ZMK firmware now builds and works! Keychron b1 pro exists in at least two hardware variants, with different matrix definitions. Source code has been provided by Keychron for older, B1P-K11 revision. But recently i discovered that it is not compatible with recent B1P-K13. I managed to map the new keyboard matrix and get custom ZMK firmware working on the newer revision! I've also put together a build script that supports both the original (2023) and redesigned (2025) matrix revisions. Build script: github.com/XelorR/Keychron_zmk/blob/space-fn/buil… Build prerequisites you can find in my repo readme: github.com/XelorR/Keychron_zmk/tree/space-fn I've successfully built and flashed the firmware on my B1P-K13-RU. Hopefully this saves other B1 Pro owners some time and gives this keyboard a bit more life beyond the stock Keychron Launcher.
LightMIS: Ultra-lightweight medical image segmentation for edge devices [R] Hi everyone, I am one of the authors of LightMIS, and I would like to share our work on 2D binary medical image segmentation. LightMIS has 131K parameters and aggregates multi-scale encoder features directly, without a learned stage-wise decoder. It achieves close observed modality-macro Dice and IoU averages to nnWNet, while using 98.14% fewer parameters and 94.58% fewer GFLOPs, with both models trained and evaluated under the same nnU-Net protocol. Paper: arxiv.org Code: AndreiiArhire/LightMIS We also tested LightMIS and the reference models on a real edge device with an Arm Mali-G52 MC2 GPU, where the full LightMIS model ran around 10.5× faster than nnWNet. All LightMIS variants achieved 100% GPU operator coverage, without CPU fallback. We started by training several popular medical image segmentation architectures, both lightweight and heavyweight, from scratch in the same controlled environment using nnU-Net v2.3.1. The experiments covered six datasets and five imaging modalities. Our goal was to develop a lightweight architecture that performs well across these modalities, which led us to the current version of LightMIS. The source code, checkpoints, and details needed to reproduce the results are publicly available. Feel free to experiment with it and try it on your own segmentation tasks. I would be happy to hear how it works for you! Any suggestions or feedback are welcome!
Vercel Labs released Agent Desktop, a Rust CLI and C library that lets an agent operate macOS, Windows and Linux apps through the accessibility tree. It returns refs for each element, so you act on a ref and re-snapshot instead of matching pixels. vercel-labs/agent-desktop
Stop Downloading Just to Upload: Save Web Files Straight to Dropbox You find the asset. You download it. You open Finder. You drag it into Dropbox. You delete the local copy. Designers, researchers, and content people who live in Dropbox pay that tax all day. Web Saver for Dropbox is a free, MIT open-source Chromium extension that skips it. Right-click a file, image, or link and send it straight into the Dropbox folder you choose. You connect your own Dropbox app once; keys stay in your browser. No Web Saver server. Independent project — not affiliated with Dropbox, Inc. > TL;DR > > Free. MIT open source. Chrome, Brave, and Edge. > > Best for: people who already work in Dropbox and save web images, PDFs, and assets often. > > Pricing: Free. One-time Dropbox app setup (~3 minutes). > > Install: Chrome Web Store · GitHub > Key Takeaways > > - Right-click (or batch-scan a page) → file lands in your Dropbox. > - The ~3-minute Dropbox app setup is the privacy feature: App key + tokens stay in your browser. > - Setup guide, limitations, and privacy policy are in the README. The download-then-upload tax is real People have asked for years whether they can save from a URL straight into Dropbox without downloading first. Most workarounds still dump files on your disk, then make you upload again. Web Saver removes that hop. What it does 1. Install and connect your Dropbox app (built-in guide). 2. Right-click an image, video, audio element, or downloadable file link — or Find files on this page to batch-save. 3. Optionally rename and pick a folder, then save. 4. Open it in Dropbox from the in-page toast. Features - Right-click save for images, video, audio, and file links (PDFs, docs, archives, design files, and more) - Find files on this page — tick thumbnails, save several at once - Review before save — rename, browse folders, create a folder, set a default - Your credentials, your browser — PKCE OAuth with your App key (no App secret); no analytics server - activeTab only — touches the tab you save from, not every site at install - Open source — inspect the code How uploads work Public files can be imported by Dropbox from the URL. Private or same-site files are fetched with the page’s session and uploaded from your browser. That hybrid path covers more saves — and is why streamed / blob: video (YouTube-style players) often will not work today. Supported files Right-clicking an image, video, or audio element saves it whatever its format — the extension uses the file the element is showing. Links and Find files on this page recognise files by the extension in the URL: | Type | Extensions | | --- | --- | | Images | jpg, jpeg, png, gif, webp, avif, heic, heif, svg, bmp, tif, tiff | | Video | mp4, webm, mov, m4v, mkv, avi, mpeg, mpg, ogv, 3gp | | Audio | mp3, m4a, aac, wav, ogg, oga, opus, flac | | Documents | pdf, doc, docx, xls, xlsx, ppt, pptx, odt, ods, odp, rtf, txt, csv, epub | | Archives | zip, rar, 7z, tar, gz | | Design files | psd, ai, eps, fig, sketch, xd | A link to any other file still works when it has a download attribute, or when the site sends it as a download rather than a web page. Streamed video and blob: sources can’t be saved yet — see known limitations. The setup is the privacy product Most “save to cloud” tools rush you live and hide where tokens sit. Research keeps finding extensions that pull page data — and some that send it to servers. Web Saver does the opposite: 1. Create a Dropbox app (once). The in-extension guide has copy buttons. 2. Paste your App key (never the App secret — PKCE does not need it). 3. Connect Dropbox and click Allow. 4. Key and tokens stay in your browser storage. No Web Saver server. No analytics. Three minutes of setup buys you browser → Dropbox API → your account, with no founder backend in the middle. Full steps: Dropbox app setup. Policy: PRIVACY.md. Who should install it today Install if you: - Keep projects, brand kits, or research in Dropbox - Save web images, PDFs, and assets often - Use Chrome, Brave, or Edge - Want credentials under your control Coming in future versions: - Firefox and Safari support - Better handling of streamed / blob video - Multi-cloud save with less setup friction - Google Drive support (next cloud destination after Dropbox) Current limits: known limitations. FAQ Do I need a paid Dropbox plan? You need a Dropbox account the API can write to. Plan limits are Dropbox’s. The extension is free either way. Why do I need my own App key? Dropbox only talks to an app. Using yours means tokens are scoped to credentials you control — not a shared multi-tenant app on a middleman’s server. Will it save any video on a page? No. Downloadable file links and most images work. Streamed players and blob: URLs usually will not — until that lands in a future version. Is this an official Dropbox product? No. Independent MIT project by Oyero Habib. Not affiliated with, endorsed by, or sponsored by Dropbox, Inc. Get it Install Web Saver for Dropbox · View source on GitHub About the author I’m Oyero Habib, maintainer of Web Saver for Dropbox. I built it after a designer friend hit the same Download → Finder → Upload wall — it felt like a real problem, and I like shipping tools that remove that kind of friction. Code on GitHub; install on the Chrome Web Store. Feedback: GitHub Issues.
Credits to the dev, i tested on 13.52 pro, with the youtube 2.32 from pkg zone FourZeroCode/PS4-YT-AdBlock-SponsorBlock-Installer #playstation #ps4 #jailbreak #youtube #adblock
Thought I’d see this noted in comments. Monolith is another solid option: Y2Z/monolith
Your Easy!Appointments Monthly Summary For October 2026 Easy!Appointments Newsletter - October 2026 The New Easy!Appointments Website Is Live We rebuilt easyappointments.org from scratch so it finally feels like the app you use every day. Pages open almost instantly, and your old bookmarks still take you to the right place. While we were at it, we refreshed the documentation for 1.6, added use cases for salons, clinics, tutors and more, and put every service and price on a single pricing page. Latest post from Easy!Appointments October 19, 2026 Easy!Appointments 1.6.1 Release 1.6.1 is stable and ready for production: reliable Google Calendar and CalDAV sync, 19 security fixes and dozens of corrections on top of 1.6.0. Read more... Docker Images Updated The official Docker images have already been rebuilt with the PHP2Shell fix. If you run Easy!Appointments in Docker, pull the latest image and restart your containers. Community Pulse Just in case you missed them, here are some community posts. - API 1.6.0 unavailabilities - Join the discussion in the Easy!Appointments group Articles, Opinions & Tutorials Self-hosting at its finest, efficiency and productivity bits from the web. - Stop Using Primitive Types: Value Objects in PHP for Domain-Driven Design - The Best AI Packages for Laravel (2025 Edition) - Custom Exceptions in PHP: Building Your Own Error Handling System Facebook · X · GitHub You are receiving this because you confirmed your subscription. GitHub: alextselegidis/easyappointments
Your agent changed 40 files. Tests pass. Which decisions are you actually accepting? I'm building wy to connect diff hunks to recorded agent edits, surface review questions, and keep missing evidence visible. Experimental, open source: grandimam/wy
JustWeather update - color customization and new animations Hello everyone, I just published v1.1.0 of the JustWeather app. 🎨 you can now customize the tint of the theme using a color picker or a hex code ✅ new smooth animation for the dots when swiping through your favorites 📋 I updated the README to clarify where the data comes from and how the app works Download v1.1.0 here
Location: New Hampshire Remote: Strongly Preferred, 3+ years experience fully remote Willing to relocate: Yes Technologies: C, C++, Rust, Perl, TS/JS, Go, Python, Wasm, Linux, Windows, Embedded, HTTP, and more Résumé/CV: https://computoid.com/about/GavinAHayes-Resume.html / https://computoid.com/about/GavinAHayes-Resume.pdf Email: gavin [at] computoid dawt com Hi, I'm Gavin. I have 9+ years experience between systems programming (cross-platform and embedded in C and C++), WebAssembly (porting, plugin systems, and binary parsing/recompilation), and Backend (identity and security in Rust, Typescript, Postgres). I'm proud of my open source work such as porting Perl to the Cosmopolitan Libc (featured on HN: news.ycombinator.com). More recently, I have enabled running binaries embedded in the Cosmo's ZIP based filesystem github.com/jart/cosmopolitan/commit/fdea2482ddc32… . For my next role, I'm mostly interested in systems/Wasm development, but can be flexible to work in an interesting domain.
I built a free, open-source plugin suite that has a small judge model double-check Claude Code (safety + token savings) I use Claude Code all day, and I built **ClauJev**, a free, open-source plugin suite that gives Claude a second pair of eyes. The checker is [Jev](typesafe.ai), a small, fast judgment model from TypeSafe AI. What it does: Safety - 🛑 Asks for confirmation before destructive or credential-exposing commands (`git push --force`, `cat .env`). - 🔍 Checks every edit against rules you write in plain language in `CLAUDE.md`. - ✅ Blocks "done!" when the work looks incomplete or your tests fail. Savings - 🎚️ Asks whether you want to change the prompt's effort before it reaches Claude: `low` for a quick question, `xhigh` for a gnarly bug. - 🪄 Routes simple subagent tasks to a smaller, cheaper model. - Cuts rework by catching unfinished tasks before you have to ask again. It never deletes or reverts anything, never breaks the session, and only logs a hash of the content, not the content itself. Validated on real, heavy use I matched **7,643 events from 18 days of heavy use** against Claude Code's saved sessions to see exactly what got flagged. That led to: - final-check false alarms dropping from 24 to 4 out of 40 turns; - credential alerts going from 100 to 6, all of them real exposures; - 12% of commands no longer needing a Jev call at all, with no lost alerts. About **400 ms** per check, and six days of daily use cost **$0.07** in API calls. The plugin is free; the Jev API is billed by TypeSafe AI. I'm not affiliated with them or with Anthropic. Install In Claude Code: /plugin install claujev --marketplace Civalski/claujev - Interactive demo: civalski.github.io - Code: Civalski/claujev Feedback, issues and PRs are very welcome, especially cases where it got things wrong.
Four agents, one job, zero trust: a swarm where every worker proves its piece Everyone is learning to run agent swarms right now, so let me save you the lesson I just paid for: the hard part is not the prompts. It is the seams. A swarm, in plain words, is several AI agents working one job at once. One checks prices, one drafts, one verifies, and a coordinator hands out the pieces and assembles the result. Starting the agents takes an afternoon. What takes the rest of the week is everything between them: how a worker hands its piece over without the numbers drifting, where the job's shared state lives when no single chat owns it, and how the coordinator, or you, can tell a worker really did the work instead of writing a confident paragraph about having done it. I run Zambo, a work layer for AI agents: shared jobs, per-agent memory, handovers that move work between agents, and a verifiable receipt for every call. I pointed it at exactly this problem and ran a small swarm for real. This is the pattern, what it cost, and the three places my assumptions were wrong. Start with no code at all If you have never run a swarm, run this version first. Add the Zambo endpoint (zambo.dev) to one chat AI for the coordinator and one per worker. The two-minute setup is at zambo.dev. Then paste this into the coordinator's chat: Coordinate a swarm job with me on Zambo: 1. Connect to zambo.dev. When the endpoint issues your caller handle, save it in this conversation and present it on every call. You are the coordinator. 2. Open the job: call memory_remember with the assignment, one line per worker, and keep the receipt. 3. Give each worker its piece. Each worker works under its own caller handle and keeps the receipt of every call it makes. 4. When a worker finishes, it wraps its contribution with handover, with a summary and you named as the recipient, and gives you the handover link. 5. Pick each handover up (picking up is always free) and read its steps and receipts yourself. Check any receipt at zambo.dev before you rely on it. 6. Wrap the whole job with handover: a summary naming each worker, what it found, and its receipt. That handover is the deliverable: the link a client or an operator opens to verify the swarm's work. And this into each worker's chat: Work one piece of a swarm job on Zambo: 1. Connect to zambo.dev. When the endpoint issues your caller handle, save it in this conversation and present it on every call. You are a worker; the coordinator holds the job. 2. Do the piece the coordinator assigned. Keep the receipt link of every call. 3. Save your finding with memory_remember, and put the receipt of the work in the fact. 4. Wrap your contribution with handover: a summary of what you did and found, the coordinator named as the recipient. Send the coordinator the handover link. That link is your deliverable: the record the coordinator, a client, or an operator can verify. That is the whole shape. Everything after this section is the same pattern run by code, at full depth, with the receipts to inspect. The pattern, at depth One job is one shared job id (a single _session_id) that every agent reuses. Identity is the opposite: separate. Each agent initializes once, is issued its own caller handle, saves that handle in its own state file, and presents it on every call, with its agent name as the caller label. Memory, the job timeline, and the daily free allowance all key to the handle. Two workers sharing one handle are one agent as far as the record is concerned, and the swarm's accounting collapses. Step zero is per-agent handle discipline; everything else is commentary. On top of that: the coordinator opens the job by saving the assignment brief to memory under the job id. Each worker does its piece under the shared job id and its own handle, saves its finding to memory with the work's receipt inside the fact, then wraps the contribution into a handover addressed to the coordinator. The coordinator picks each handover up with handover_accept, which returns the worker's saved steps and receipts, reads them, and only then wraps the whole job into one handover whose summary names each worker and its receipts. Notice what the coordinator never does: assemble the result from worker summaries. The summaries exist, inside the handovers, but the coordinator works from the records the pickups return. Verify, then wrap. The run November 2, 2026, four fresh identities, one shared job id (swarm-scan-2026-11-02), job: a three-asset market scan. The coordinator opened the job with the brief: zambo.dev worker-btc checked BTC: $76,876.00, down 1.15 percent in 24 hours, at 20:07:21 UTC. Price receipt: zambo.dev. It saved the finding: zambo.dev. Then it wrapped its contribution: zambo.dev worker-eth checked ETH: $2,251.23, down 0.27 percent, at 20:07:26 UTC. Price receipt: zambo.dev. Finding: zambo.dev. Wrap: zambo.dev worker-sol checked SOL: $91.79, down 2.27 percent, at 20:07:31 UTC. Price receipt: zambo.dev. Finding: zambo.dev. Wrap: zambo.dev The coordinator picked the three handovers up at 20:10:13, 20:10:16, and 20:10:19 UTC, read each worker's steps from the pickups, and wrapped the swarm: zambo.dev Open that last link. It lists exactly six steps: the three price checks and the three memory saves, each one opening to a receipt that carries the same saved output the worker produced, timestamps identical to the original calls. The summary names each worker, its price, and its original receipts. The wrap's own receipt is here: zambo.dev Three places my assumptions were wrong I assumed the shared job id made a shared inbox. It does not. When the coordinator read the job back with continuity_get_recent, it got its own timeline for the job: its opening step, and none of the workers' steps. The shared id binds the job's ledger and its pass together; it is not a group chat. Contributions move through handovers or they do not move. Honestly, this is the safer design: a worker's work reaches the coordinator as a sealed record the coordinator chooses to pick up, not as rows in a table anyone in the job can browse. I assumed wrapping was a formality. My coordinator's first wrap attempt, a handover call with next steps but no summary and no named recipient, was refused on the spot, with its own receipt: zambo.dev. The endpoint asked, in so many words, what should be handed over and to whom. A wrap is a deliverable being published, and it is held to that standard. Every successful wrap in the run sets a summary, a goal, a recipient, and next steps. I assumed the wrap would just bundle links. Instead, a wrap runs a proof loop over the job's steps and binds them into the handover. Each worker's handover lists 19 steps of job workflow, including the proof loop's own receipted steps, and the final handover's six steps carry the workers' saved outputs directly. The deliverable is the work, checkable step by step, not a paragraph about the work with links sprinkled on it. What a swarm costs The free tier is per agent: 20 calls plus 1 wrap-up free every day for each agent's own caller handle, and picking up in another AI is always free. Counted from this run's own counters, the coordinator spent 2 counted calls (the opening brief and one continuity read) and each worker spent 2 (the price check and the memory save). Three pickups and four wrap-ups moved no counter at all. Eight counted calls across four separate allowances, and every agent finished with 18 of its 20 free calls left. A swarm shaped like this runs free, one run a day or several. When a swarm outgrows free, the scaling line is per agent too: a Day Pass is $0.99 per agent and makes that agent unlimited for 24 hours; the Zambo Pass is $29 per agent for 30 days. One-time payments, never subscriptions. This four-agent swarm on a heavy day costs four Day Passes, $3.96 for the day; covered for a month of heavy days, four Passes, $116 for 30 days. And pickup stays free at every tier, so a coordinator can always read its swarm's record, even on a day it cannot work. Do not take my word for it Pick any worker contribution in the final handover, open its receipt, and check it at zambo.dev. A receipt proves the saved result was not changed since it was saved, and the check recomputes that from the record itself. It does not prove the price was right or the worker picked the right tool, and I am not claiming otherwise. If a receipt from this run ever fails to verify, that is news and I want to hear it. The full guide, with the accounting and the standing instructions at depth, is here: rambozambodotdev.gitlab.io. The runnable code is in the starter repo, coordinator and worker, next to the same pattern written as an installable skill file: github.com/zambodotdev/zambo-memory-loop-starter/… Your framework still runs your agents. Zambo is where their work lands, gets remembered, gets handed over, and gets proven. That is the whole pitch, and every link above is the evidence. --- Try it free: add Zambo to your AI in one step at zambo.dev/add. 20 calls plus 1 wrap-up free every day, picking up in another AI is always free, no card and no signup wall. - Zambo: zambo.dev - Add Zambo to your AI: zambo.dev - The swarm guide, with the full run and accounting: rambozambodotdev.gitlab.io - Swarm scan code, coordinator plus workers: github.com/zambodotdev/zambo-memory-loop-starter/… - The swarm package repo: zambodotdev/zambo-swarm-coordination - The swarm package on PyPI: pypi.org - The swarm package on npm: npmjs.com - AER-1 conformance corpus: zambo.dev - Verify any receipt: zambo.dev - AI agent handover, the signature product: dev.to
BPreCap: one activation that trains 100–200-layer MLPs without normalization and stays robust at 4-bit (16 preregistered studies, negative results included) I've been working on a bounded, self-normalizing activation called PreCap. Version 0.2 had a trade-off: calibrating the activation for depth (mean-field gradient gain χ1 = 1 + k/L) made deep MLPs without normalization trainable, but it widened the output range and hurt 4-bit activation quantization. v0.3 puts the depth-calibrated core inside a soft, fixed bound (identity on most of the range, smooth saturation to ±3), with constants solved so the whole function keeps E[f²] = 1 and χ1 = 1 + 2/L. I call it BPreCap. Everything is in 16 preregistered, hash-frozen studies (MNIST, FashionMNIST, CIFAR-10/100; paired Wilcoxon with Holm; post-training quantization, weights int8, activations int4). Main results: - MNIST, 200 layers: 4-bit error 14 % vs 77 % for the v0.2 depth-calibrated form (about 1 pp fp32 cost). - vs TAT under TAT's own recipe (CIFAR-10, 100 layers): on par in fp32, better at 4 bits in every regime tested (cautious size −3.8 pp). BPreCap's 4-bit cost stays at about +3 pp, TAT's moves between +7 and +15 pp with LR and weight decay. - SELU is a strong baseline at 100 layers (on par at 4 bits), but at 200 layers it falls behind (−12.0 pp fp32, −9.0 pp 4-bit). ReLU and GELU don't train well at this depth without normalization. - An independent PyTorch/CUDA re-implementation on an RTX 4070 reproduces the main comparisons (medians within 0.5 pp). Limits: MLPs only, post-training quantization only, and BPreCap needs small learning rates (it diverges at peak LR 2·10⁻⁴). A fair comparison with PACT is still missing. (see edit). The ingredients (depth-dependent activation shaping, bounding activations for quantization) are known; this is a controlled measurement of one form that combines them. Preprint: doi.org Code, raw data, preregistrations, fused CUDA/Metal kernels: hermi-labs/precap-activation Critique very welcome, especially on the TAT comparison and the quantization protocol. If anyone can try it in CNNs or at larger scale, I'd love to hear how it goes. Edit (Nov 3): v0.3.1 follow-up. Two of the limits above are now tested: - Small learning rates: the divergence at 2·10⁻⁴ came from missing LR warmup, not from the bound (a linear control and a hard clip diverge the same way). With a 5 % linear warmup BPreCap trains up to 3·10⁻⁴ at 100 layers, and the 4-bit advantage over TAT holds (−8.8 pp, 12/0). At 200 layers it also holds against SELU (−8.2 pp, 12/0), but warmup raises BPreCap’s own 4-bit cost. A shorter warmup reduces that cost but does not remove it. - PACT / QAT: with QAT as an upper bound (100 layers, CIFAR-10), TAT with LSQ-style QAT beats BPreCap with plain PTQ by 3.0 pp at 4 bits, and PACT-style QAT gives the same. Under QAT both forms are equal. With PTQ only, BPreCap is 11.9 pp ahead. So the bound is a PTQ advantage, not a substitute for QAT. Updated preprint: doi.org (all new studies are in the same repo). preview.redd.it
I've been working on a 2D platformer called Coin Adventure Tenho trabalhado em um pequeno projeto chamado Coin Adventure e queria compartilhar um pouco do que tenho desenvolvido. É um jogo de plataforma 2D com 6 fases, diferentes ambientes, inimigos, plataformas móveis, checkpoints, itens colecionáveis, molas e uma luta contra um chefe. Uma coisa que torna este projeto especialmente interessante para mim é que o construí usando meu próprio motor de jogo JavaScript, o Prism2D.js. Comecei a trabalhar no jogo em parte para testar o motor, mas ele gradualmente se transformou em um jogo completo. Tem sido uma ótima experiência de aprendizado, especialmente ver como todos os diferentes sistemas se unem em um jogo real. Aqui está um pequeno vídeo de gameplay mostrando um pouco dele! Jogue o jogo: lodystudios.itch.io](lodystudios.itch.io) Prism2D.js (código aberto): Lodymain/Prism2D.js Gostaria muito de saber sua opinião sobre o jogo. Obrigado por conferir!
Forked T3Code gets an Apple TV integration via hacky homekit doorbell and barebones app pseudoseed-labs/tVcode
SOML – A Config Format for Humans | Discussion
Valence: polysemantic language for code with multitudes of meaning | rottytooth
Fluctuation bounds for rotationally invariant first-passage percolation Scott Armstrong, Ahmed Bou-Rabee arxiv.org [𝚖𝚊𝚝𝚑.𝙿𝚁] 💬Code: nitromannitol/FPP-variance t.co
YSK: Strata and polystrata have this functionality called answer_room With default settings and thinking on and max_tokens ≥ 1024, and no explicit reasoning_budget_tokens, Strata / PolyStrata has a behavior called answer_room which will silently cap thinking at 0.75. * max_tokens and inject a "wrap-up" line in the model's head in its own voice, saying I have thought about this long enough; time to give my answer. + </think> It then continues the same reply from there. The client (harness, you or I) gets no error, flag or finish reason for this, and often no log. This matters when you compare between strata and other engines : models served by strata had "nudges" in the form of its inner voice, and such extra text in the transcript. In some cases the nudge line is injected mid thinking, quite abruptly and it is hard to tell whether it helps or harms. Strata main @ c62d5e3 Wrap-up message: github.com/Niko1221/Strata/blob/c62d5e3a15b40ea51… Trigger (the 3/4 budget): github.com/Niko1221/Strata/blob/c62d5e3a15b40ea51… Injection point: github.com/Niko1221/Strata/blob/c62d5e3a15b40ea51… Default on (answer_room = True): github.com/Niko1221/Strata/blob/c62d5e3a15b40ea51… I noticed this when a model that tends to have long front-loaded thinking on dgx spark gave slightly but uniformly shorter answers (and task performance worsened in my hard problems), when served on polystrata. PS, I had sonnet audit all cases of cut and injection, and here is what it reportes back ( sonnet called it "splice"): Where it lands: - At a natural break: 6 (9%). - Inside a repetition loop (for example HashMap.new()? repeated for 30K–118K characters): 15 (22%). - Intrusive: 48 (70%, or 89% of the cases that weren't loops): - mid-sentence: 28 - mid-number or mid-expression ("Mod 9982 |cut|"): 13 - inside code (one cut right after def solve():): 4 - mid-word: 3 Harm: What didn't break: no reply re-opened thinking after a splice, and all 79 tool calls after a splice completed normally. Lost reasoning: 26 of the 30 splices that came in a task's first reply were followed by a generic fresh start (ls /app). The 61K tokens of thinking left no trace in what came next. Clear harm in 6 cases: the reply ended with no action at all, and all 6 tasks failed. Qwen3.6, task 3674: cut at "Where right_prefix_max |cut|", and the only output afterwards was "Integrity notice: yes". Flash-Next IQ2_XS, arc190_c: the same pattern. A dropped conclusion: Qwen3.6, arc192_b, was cut right after "I think the answer is:", and the model then guessed. Loops: the splice did break them, but those tasks still failed.
@0xCodila Added themes to grokbot-wallpaper 🤠 Texas Hill Country exterior plus per-floor interiors guessed from your group names. Also an optional private feed on your own free Cloudflare Worker, so bot statuses aren't on a shared public ntfy topic. PR: grokbot-wallpaper#1
Neural Canvas: Interactive Neural Network Visualization with p5.js, NumPy and JAX I built Neural Canvas, an interactive neural network visualization project using p5.js, JavaScript, and Python. The idea behind this project was to understand neural networks mathematically and visualize how they work. I trained neural networks in Python using NumPy and JAX, then exported their weights and biases as JSON files to perform inference directly in the browser using JavaScript and Math.js. The frontend is built with p5.js, which handles the interactive drawing board, neuron visualization, connections, and animations. Users can draw digits and see how the neural network processes their input and makes predictions. So far, I've trained four models. Two MNIST models using NumPy achieved 92% and 96% test accuracy. My Arabic Digits MNIST model using JAX achieved 96%, and my Persian Digits MNIST model also achieved 96% test accuracy. I started by manually implementing forward propagation, backpropagation, and Gradient Descent using matrix algebra. Later, I explored JAX automatic differentiation and Mini-Batch Gradient Descent with Momentum. My next goals are to implement RMSProp and Adam, followed by CNNs. The project combines Python-based neural network training with JavaScript-based interactive visualization. I built it primarily as a learning project to understand neural networks from their mathematical foundations to browser-based inference. GitHub: MuhammadEssa2002/neural_canvas Live Demo: muhammadessa2002.github.io
AI VOICE AGENT For the past 1.5 months, I've been building an AI Voice Receptionist for small businesses. 🎙️ The idea was simple: what if a business could have an AI assistant that handles customer calls, answers questions and manages appointments? Turning that idea into a working system has been quite a journey. Here's what I've been working on: • Building conversational workflows using LangGraph • Integrating FastAPI for the backend • Implementing appointment booking, availability checks and cancellations • Connecting Google Calendar and Gmail APIs • Working on speech-to-text and text-to-speech • Integrating Twilio for phone communication • Using PostgreSQL, SQLAlchemy and Alembic for persistence • Exploring RAG for business-specific knowledge The biggest learning wasn't just using these technologies. It was handling the complexity behind real conversations — missing information, follow-up questions, conversation state, confirmations, failures and edge cases. I've spent a lot of time debugging, restructuring workflows and understanding how all these pieces fit together. Still a work in progress, but I'm excited about how far it's come. GitHub: vikasya7/voice-ai Would love to hear your thoughts and feedback! #AI #LangGraph #Python #BuildInPublic
try now: collinear-ai/collinear-zero cc @ClementDelangue
CELL: The ONLY PS3 Social Media I’d just like to take a second and say, wow thank you 1000x times for the amount of love, loyalty and support you all have shown CELL. We’re officially at over 400 users! I can’t believe I’ve had the opportunity to have such great feedback and support from all of the users that have signed up. Thank you to all the YouTubers, users, friends and others that have consistently shown support to the project regardless of the feedback in certain communities. Without you all, CELL wouldn’t be possible. And for that, thank you for supporting me as I work on it and the amount of grace and patience you all have had in regard to CELL. Signups are still open, Giftcard giveaway will be happening soon. Everyone on site will be entered for a chance to win. If you support the project, thanks one more time, if you don’t I hope you eventually come around and join the largest social community on PS3. Download: odhz/CELL-The-ONLY-PS3-Social-Media-Application - Dehz #CELLPS3 #CELL
Made an open source ESP32 board that puts your UPS on the network (NUT, web UI, Tailscale) I want to disclose upfront that I used AI for parts of this project, mainly the software, since my background is more hardware engineering. Im also trying to turn this into a product, but that wasnt my reason for making it or for posting this. I could only pick one flair. So why did I make this? I had my Nest camera plugged into a UPS in my garage so it would keep recording incase the power went out. I didnt have a computer nearby to monitor it, and when I actually needed it, guess what, the battery was dead. I should have checked it but I kind of forgot about it. You can run NUT on a Pi and that works great, but it felt like overkill to dedicate one (plus an SD card) to a single UPS. I also have a big APC rack mount UPS that needs a VERY expensive network card I dont want to pay for, so I figured I could fix that too. So I built an ESP32 board that speaks NUT, uses the existing profiles to show every stat it can get from your UPS, and can run a battery test on a schedule. It works with Home Assistant right now through the NUT integration, plus Synology, TrueNAS, unraid, Proxmox, upsmon/upsc and Prometheus. A few things I learned along the way: - Every UPS does HID a little differently. My Smart-UPS X 1500 keeps its battery voltage in a totally different spot, so I ended up leaning on NUT's own tables instead of trying to outsmart them. - UPSes sometimes just stop answering over USB and the board keeps showing old numbers like everything is fine (kind of the exact problem I was tryign to solve lol). So I put a switch on the USB power to the UPS and the board power cycles the link itself. - Tailscale needs alot of RAM. My first chips (S3 N4, no PSRAM) couldnt run it, and even on the S3 N16R8 I had to move Wi-Fi, TLS and task stacks into PSRAM. - I wired the OLED header backwards on the first board. First screen got 3.3V on its ground pin, board got hot, screen died, board somehow survived (so thats...good?) - USB-C is... interesting. Getting USB PD and regular dumb USB-C chargers both working cost me a few burned up boards. I laid out and routed nearly all of the board myself, finished with Freerouting, and used AI to help check the SPICE sims. Theres a header with 4 spare GPIOs plus 3.3V, 5V and GND for your own stuff (garage door relay, temp sensor, whatever) and one for a small OLED. Schematics, gerbers and code (GPL-3.0) are here: aronreid/ups-to-esp32. No fuses are blown, so flash it however you want. If you have similar hardware it SHOULD work, and if it doesnt Im happy to update the code. I even have a version running on a P4 with PoE. Tested UPS list (HERE) Why the S3? Ineeded a chip with a real USB host port to talk to the UPS, and the classic ESP32, C3 and C6 dont have one so I went with the N16R8 because Tailscale needs around 100 KB of RAM and the version without PSRAM just couldnt do it. The P4 version is for PoE since the P4 has wired ethernet built in. Also JLCPCB availiblity was big I had to search around a few times to get a chip that was both in stock and worked of which I also had issues with lol. Libraries and why? - ESP-IDF instead of Arduino, because I needed control over where memory goes (PSRAM vs internal) to fit Tailscale, and IDF's own USB host stack for the UPS side. - NUT's usbhid-ups tables instead of writing my own mappings. NUT has years of UPS quirks already figured out and I wasnt going to beat that. - MicroLink for Tailscale, since the official Tailscale client is written in Go and wont run on a microcontroller. - wireguard-lwip, which MicroLink uses for the actual WireGuard tunnel. Privacy By default it checks in with a hardware ID, firmware version, uptime and restart reason so I can tell if something is broken. Its one click to turn off. Sending your UPS model and readings is opt in but does really help me making sure it works. I could make this not work at all unless you opt in but I'm also at least right now interested in people using it lol. Big thanks to the projects this is built on! - NUT (Network UPS Tools) for the UPS profiles - MicroLink by Cameron Malone for Tailscale on ESP-IDF - wireguard-lwip by Daniel Hope - Espressif's ESP-IDF. Right now its $44.99 CAD for just the board (about $32 USD) because the first batch was small. I really want to get that down, hopefully to around $30 CAD with a bigger run. If you'd rather buy one than build it: myupspal.com Mostly though I would LOVE feedback. GitHub issues (HERE) or [hello@myupspal.com](mailto:hello@myupspal.com) are best, but comments here are good too. What UPS are you running, and what would you want it to do that it doesnt? Also was this just a complete waste of time? What do you all think? Sometimes you get so wrapped into "can I do this" you forget to ask "should I do it" lol EDIT: Oh I want to add 0.96" OLED.. JESUS there is no standard size and two major pin types GND first and VCC first and dear me its a pain. This was probably a waste of time as I'm not sure if anyone would actually want/need a display but none the less.... it has one lol.
Just this weekend I was able to achieve an order of magnitude faster performance with simple fixes to esoteric concurrency issues I hadn’t considered. I was also able to fix a bunch of correctness bugs that there’s no way I would have considered. And now I have a harness to avoid regressions.
My 3D tool turned a scanned face into a mess. The tool wasn't the problem. My 3D tool turned a scanned face into a mess. The tool wasn't the problem. I wrote a button that rebuilds a 3D model with fewer polygons — a normal step before a model goes into a game. Then I pointed it at a scanned head and got something I wouldn't show anyone. Three things were wrong, and none of them was the button itself. 1. The head wasn't one model It was 93 separate pieces, and many of them were open at the edges. The button treated all of them as one surface, pulled the new mesh onto that surface, and the result came out smeared. Now there is a Repair button: it merges duplicate points, removes empty faces, and tells you how many pieces the model is actually in. The rebuild also warns about a broken model before producing the result. 2. The "how many polygons" setting did nothing 5,000 and 10,000 gave me the exact same model. The method I used can only double the detail, so the number you type in mostly doesn't matter. Now the tool searches for the right cell size instead, and the budget lands where you set it: 5,000 → 4,748. 3. The model was tiny and far away It came in scaled 26 times and sitting 300 metres from the centre of the scene. The tool tried to fit the new surface to the origin — the wrong place — and every point collapsed into one spot. Two fixes: the tool now keeps the model's position and scale, and it snaps points to the nearest part of the surface instead of pushing them through it. What's in 0.6.8 - Repair — merges duplicate points, removes empty faces, reports pieces and open edges. - Exact budget — the polygon count you type in is the count you get. - Safer fitting — keeps the model's position and scale, and refuses to guess when the model is open. - Delete the unused — shows the materials and images it will remove, then asks. - The AI sees the buttons — a button's report now reaches the chat model. - Fixed: a missing import made the repair silently do nothing. Source and releases #blender #gamedev #3d #indiedev #python #blenderaddon
Built a free, open-source tool to help navigate dense literature reviews, finding open-access PDFs, and organizing papers (ResearchLoom) Hey everyone, When starting out in research—whether it's for a senior capstone, summer fellowship, or your first lab project—literature reviews can feel overwhelming. Juggling dozens of 20-page papers, trying to understand dense methodologies, and figuring out what’s actually relevant takes hours. On top of that, a lot of popular research tools lock features behind expensive monthly subscriptions that student budgets don't cover. To make literature reviews much less intimidating, I built ResearchLoom (Samhita)—a completely free, open-source desktop app that organizes and streamlines paper discovery and synthesis: 🔗 GitHub Repository: sohamsans/researchloom 📦 Free Windows & Mac Downloads: github.com/sohamsans/researchloom/releases How it helps if you're working on a research project: One Search Across Multiple Libraries: Instead of bouncing between Google Scholar, arXiv, and publisher sites, it searches OpenAlex, arXiv, Crossref, and Europe PMC all at once. Quick Abstract Triage & Breakdown: Evaluates candidate papers and pulls out the core research questions, methods used, and study limitations into clean summary cards so you don't have to read 50 full papers just to find the 5 you actually need. Finds & Saves Open-Access PDFs Automatically: Whenever an open-access version exists, it downloads the PDF straight to a local folder on your computer and opens it in an integrated reader with an AI assistant you can ask questions to. Visual Flow Maps & PRISMA Review Tracking: Helps you visually map how papers connect and generates PRISMA flowcharts (handy if your professor wants a formal systematic review section). Instant Citation Export: Exports clean BibTeX, APA, MLA, or Excel spreadsheets with one click for your bibliography. Completely Free & Private (No Paywalls): It uses Bring Your Own Key (BYOK) with free tiers from providers like Gemini or Groq, or local models. Nothing is sent to a paid subscription service, and your notes stay stored locally on your machine. Pre-built standalone apps are ready to download on GitHub (single .exe for Windows 10/11 and an installer for macOS—no coding or terminal setup needed). If you're currently working on a capstone or paper review, give it a try! Feedback and feature suggestions from fellow student researchers are always welcome.
Best approach to generate production-ready Playwright scripts for ~10 heavily customized Salesforce-based CRMs? Ciao everyone, I'm working on a project to automate order creation on our suppliers' CRM portals. Most of them are built on Salesforce, but each one is heavily customized, so the flows differ a lot from one supplier to another. The flows are long and multi-step: an experienced user takes about 5–8 minutes per order, mostly copying and pasting data from our internal CRM into the supplier's portal. We handle around 50 orders a day. I started writing Playwright automations with codegen, but development is very slow, and I have around 10 different supplier CRMs to cover. Some context and constraints: - No APIs: the suppliers don't provide or allow API access, so UI automation is the only option. - Authentication is solved: MFA is handled programmatically via TOTP, no issues there. - Portals are stable: the UIs rarely change, so a deterministic script would stay valid for a long time once written. - Why not RPA: I'm comfortable with Playwright, and the UiPath quote I received was far beyond our budget. - Stack is open: I'm not tied to a language, and I'm even open to alternatives to Playwright if there's a better fit. So far I've tried generating the Playwright code with: - browser-use harness - Playwright MCP - Stagehand by Browserbase - Browserbase Agent - Claude Chrome Extension - chrome-cdp-skill With very specific prompts that force the AI through the task step by step, most of these tools can complete the flow. But that's not really what I need: I want the AI to perform the task once, "look" at how the pages are built, and output a deterministic Playwright script I can run in production without an LLM in the loop. The generated code has been consistently poor: brittle selectors, reliance on Salesforce's dynamic IDs, badly handled waits, and missing steps. Getting it production-ready takes a huge number of iterations, which defeats the purpose. Has anyone tackled something similar? I'd love to hear how you'd approach this differently, whether that's a different tool, a better workflow (e.g. recording plus AI refactoring, page object generation), or tips specific to Salesforce UIs. Anything that speeds up writing reliable Playwright code would help. Thanks!
@StoneFelt1 @earthtojake I agree. That's why vibed Foadsf/freecad-strata Inspired by @Onshape PLM/PDM but with git as kernel for @FreeCADNews
I don't know that Matteo was working on anything in particular. @nodeland.dev ? Anyways, as far as inheriting context, did you have something like this in mind? microtel#29
You Want Free Inference Providers, Here's A Huge List of Them OmniRoute's Free Providers Reference And a Guide from the Same Repo on Free Tiers I don't like waiting geologic epochs for inference requests personally, but I found this research flat fee subscription tiers and saw other people acting like there are no free LLM APIs in this sub, so here you go. If it is down voted to hell, its the inference providers openclaw set ups 100%. But that's OK, I am not trying to be Daddy Cool and maybe this helps someone NOTE: I am not affiliated wit Omniroute or any of the inference providers listed or not listed. Just a random guy doing my own thing that found something useful to others and interested to see how quick the openclaw agents down vote this to the abyss.
I made a free Android app that keeps my phone locked until my Anki reviews are done (V2 + full demo) Ankineer Lock: a free, open-source Android app for AnkiDroid users. Until your daily goal is done, distracting apps show a "Not yet." screen. AnkiDroid, calls, messages and the apps you choose always stay open. Finish your cards → your phone is free until tomorrow. It reads your progress through AnkiDroid's official API, so reviews on your computer count after sync. No account, no ads, no Internet permission — nothing leaves your phone. 📥 Download (APK): github.com/BONJ37/ankineer-lock/releases/latest 🎬 Full install + demo video: youtu.be
OWCE ios I made a native iOS version of OWCE and got it working with my Pint (firmware 5040). I haven’t tested it with other boards yet. It’s an unofficial project, and for now you’ll need a Mac and Xcode to try it. If you’re interested, the setup guide is in the repo. I’d love to hear whether it works with your board. Github link
Dictation apps don't work with the new herdr app: herdr-app#203 How are they not discovering this in their testing?
serialport-rs for Node.js: Rust-backed serial ports for Node.js with parallel I/O and SerialPort-compatible APIs I recently ran into an issue with @serialport/bindings-cpp on Node.js 26.4, where serial reads could stall until unrelated JavaScript activity occurred. While investigating the issue and contributing fixes upstream, I started experimenting with an alternative implementation using Rust and N-API, because upstream seems to be unmaintained for more than a year now. The result is serialport-rs for Node.js, which uses the existing Rust serialport crate underneath and aims to preserve the familiar Node.js SerialPort API. The implementation uses dedicated native I/O workers rather than relying on JavaScript-side polling. Initial benchmarks have also shown some promising latency improvements. It’s available on npm if anyone is interested in trying it: npmjs.com
stack-diagram is a new agent skill: describe your stack in one sentence ("Next.js on Vercel, Postgres on Neon, auth with Clerk") and Claude Code, Codex or Cursor draws the architecture diagram with official logos. 1,081 stars. caru-ini/stack-diagram
@emollick The authors posted full logs for a rerun: every prompt, the code the model ran and the edited networks, at RikiyaT/lm_understand-logs The part I'd check first is their own control: on Gemma 3 270M, which rarely loops, the edits moved MATH-500 by about a point at most.
EFF Dice-Generated Passphrases
Understanding DeepSeek Harness plugin admission during upgrades, and an offline verification CLI DeepSeek Harness (DSH) has a modular plugin and preset system, but a common friction point during runtime version bumps (such as moving to 0.2.1-rc.5) is that installed plugins or MCP tools can silently fail to appear on startup without throwing an explicit crash error. We conducted an architectural analysis of DSH's loader mechanism to understand why this happens, evaluated 192 plugin units against rc.5, and created an open-source, read-only diagnostic CLI (dsh-triage) as a community resource. How DSH Plugin Admission Works Under the Hood When DSH starts up, its loader does not simply check bundle manifests; it evaluates Admission semantics per Loader Row: 1. Row-level evaluation: A bundle package might look valid, but specific capability entries or hooks registered as individual profile rows are judged independently against the current DSH runtime. 2. Exact peer pinning: Many community plugins pin their peerDependencies strictly to an exact rc release (e.g. @deepseek-ai/dsh-shell-env: "0.2.1-rc.4"). 3. Silent omission: If a runtime version excludes the declared semver range, DSH does not crash the application. Instead, it quietly drops the incompatible row (phase=no-fiber, disabled=true). To the user, tools simply vanish from the interface without an obvious error. Empirical Evaluation (192 Plugin Units) Testing a standard profile containing 192 units (2 bundles + 190 loader rows) moving from 0.2.1-rc.4 to 0.2.1-rc.5: - 119 units were rejected at admission (primarily due to strict peer version pinning). - 73 units declared no DSH version constraints (unconstrained). - 0 units were improved by upgrading forward. Offline Pre-Check Tool (dsh-triage) To help developers inspect their local plugin compatibility before upgrading, we released dsh-triage as a free, MIT-licensed developer utility. Pre-check target version compatibility: npx dsh-triage upgrade --to next --card This reads your local profile declarations offline and outputs a 4-bucket breakdown (would-be-denied, cannot-tell, improved, likely-fine) with exact remediation options (such as recording a profile exemption via compatibility.json without waiting for upstream package re-releases). Diagnosing already dropped rows in your current profile: npx dsh-triage status # One-screen summary of current profile health npx dsh-triage why <module> # Pinpoints the exact admission conflict for a specific row npx dsh-triage audit # Full capability ledger across all loader rows Technical & Privacy Transparency (Rule 4 & 5 Compliance) - Non-commercial & Open Source: Licensed under MIT. Zero monetization, zero accounts. - 100% Local & Read-Only: Completely offline, no telemetry, never writes to ~/.dsh. - Declaration-focused: Strictly evaluates semver declarations against runtime contracts (tested against 718 automated test suites). - Source Code: Direct GitHub repository: jjjjjjjjnnjnn/dsh-triage (npm: dsh-triage) We hope this technical breakdown helps clarify DSH loader behavior. Happy to answer questions or help trace tricky admission issues in the comments!
projectwallace/stylelint-plugin all ready for the Stylelint v18 release 👀
CELL: The ONLY PS3 Social Media I’d just like to take a second and say, wow thank you 1000x times for the amount of love, loyalty and support you all have shown CELL. We’re officially at over 400 users! I can’t believe I’ve had the opportunity to have such great feedback and support from all of the users that have signed up. Thank you to all the YouTubers, users, friends and others that have consistently shown support to the project regardless of the feedback in certain communities. Without you all, CELL wouldn’t be possible. And for that, thank you for supporting me as I work on it and the amount of grace and patience you all have had in regard to CELL. Signups are still open, Giftcard giveaway will be happening soon. Everyone on site will be entered for a chance to win. If you support the project, thanks one more time, if you don’t I hope you eventually come around and join the largest social community on PS3. Download: odhz/CELL-The-ONLY-PS3-Social-Media-Application - Dehz #CELLPS3 #CELL
I built a local workflow layer so Claude Code can't mark a task done without green gates and an independent review >I kept hitting the same two problems with coding agents: they forget the repo between sessions, and they say "done" too easily. Agent Flow is my answer, as a CLI plus skills for Claude Code (/flow-plan, /flow-harden, /flow-orchestrate) and Codex. > >How it works: you author a plan with acceptance criteria (the skill does it with you), each task gets a scoped context pack instead of the whole repo, gates run tests/typecheck/smoke, and advance only closes a task if the gates are green for the current code. Phases need a pass from a separate reviewer agent. Memory is append-only JSONL in the repo, so the next session doesn't start from zero. > >Token side: a per-task router picks haiku/sonnet by signals in the plan, executors are fresh subagents, and agent-flow usage reads your transcripts to show real spend. Seeing how much goes to cold-cache re-reads after a break changed how I work. > >Honest caveats: one experiment behind the token claim (documented with limits in the README), no semantic search, hooks depend on transcript formats that may change. I'd love feedback from people running bigger repos or monorepos. > >Xch4rt/agent-flow