Mentions
Conversations linking to GitHub, newest first. Times are UTC.
Mentions by source
X
Bluesky
Hacker News
V2EX
Qiita
DEV
Mastodon
Latest posts
Running an issue-fixing agent on a $6 Fly Machine: a queue, three budgets, and an honest test report Most "AI fixes your issues" demos are a session: you run a CLI, it edits files, you look at the diff. I wanted the opposite: a standing service that sits next to a repository, picks up issues labeled bug, docs or chore, and opens a pull request — and that I can leave running without watching my LLM bill. That's fly-swe, built on mini-swe-agent. Here's how it's put together, and what it cost on a real TypeScript monorepo. !demo One Machine, one process pair, one SQLite file The whole thing is a FastAPI webhook receiver and a worker, run together by honcho on a single shared-cpu-1x Fly Machine with a 1 GB volume. Webhooks are verified (HMAC, timing-safe, fail-closed on an empty secret), de-duplicated by delivery ID, and written to a SQLite queue in WAL mode on the volume. The worker polls it. No Postgres, no Redis, no broker. If the Machine restarts mid-task, orphan recovery picks the task back up: nothing lost, nothing run twice. The infra costs around $6/month. The LLM never touches git This is the invariant I care most about. The coding agent edits files and leaves them in the working tree. Then it emits a completion signal, and everything after that belongs to the harness: git add, commit, push and PR. The agent's shell gets a tokenless remote, no push credential, no gh, and an env allowlist with no secrets. A side effect: when the target repo's AGENTS.md says "commit and open a PR", the agent follows it for style and tests but not for git, so nothing goes sideways. Three budgets, enforced in code - max_iterations: a hard cap on agent steps - max_cost_per_task_usd: per task - max_daily_cost_usd: per repo per day cost_tracking: strict goes one step further: if a step can't be priced, it isn't run. When a run hits a limit, it doesn't just die. It force-pushes its partial work to a WIP branch, and the next trigger resumes from there, along with a summary of what was already tried. That summary costs zero LLM tokens. "How I tested" is what the harness saw, not what the model said After the agent's loop ends, the harness re-runs the test command the agent says it used, in the same environment. The PR's "How I tested" checkbox reflects that run. Anything the agent couldn't run (say, a suite that needs a database the sandbox lacks) is listed separately, labeled as the agent's report, and never checked off. If verification fails, the harness says so on the issue and retries automatically. That happened on 3 of the 10 issues below. Numbers from a real repo 10 consecutive issues (bugs and docs) on a TypeScript monorepo with an SDK and a docs site, using GLM-5.3: - median cost per issue: ~$0.31; worst case $1.85 (225 iterations, two verification retries) - typical run: ~80 iterations - all 10 PRs were reviewed by a human before merge. fly-swe never merges. What it isn't (yet) Execution isn't sandboxed: the agent's commands run as a subprocess in the same container, mitigated by the env allowlist and the absence of credentials (SECURITY.md). A per-task sandbox is next on the roadmap. Ephemeral Fly Machines / Sprites are an obvious candidate. GitHub is the only forge today, but it sits behind a provider plugin with a conformance suite. GitLab or Gitea would be a good first contribution. Deploy in ~10 minutes: Quickstart · AGPL-3.0
@BrodieSchafer_ Mid-rollout is exactly where a fix widens the blast radius. Our rule: at 3am automation may only stop, never fix. One pre-approved halt per critical alert, fired if nobody ACKs, and a human owns the repair in daylight. getstanch/stanch
We believe Bitcoin mining can reach far more people when it's simple, useful, and accessible. heatbitinc/bitair-mainboard t.co
React vs Vue en 2026 : versions, code comparé, performances et grille de choix L'essentiel en une minute React vs Vue : les deux outils construisent les mêmes interfaces, mais ils ne partent pas du même principe. React se présente comme une bibliothèque et laisse l'écosystème fournir le routage, l'état global et le rendu serveur. Vue se présente comme un framework progressif et fournit des pièces officielles pour chacun de ces besoins. Ce comparatif s'appuie sur la documentation officielle, les registres npm, un benchmark public reproductible et l'enquête Stack Overflow 2026, consultés le 2 novembre 2026. Nous avons aussi compilé et exécuté un petit composant écrit dans les deux, et nous détaillons exactement ce qui a été fait. > Points clés > > - Le 2 novembre 2026, la version stable de React est la 19.3.1 et celle de Vue la 3.5.44. Vue 3.6, qui apporte le Vapor Mode, est encore en version candidate (3.6.0-rc.14). > - Le même compteur s'écrit en une vingtaine de lignes dans les deux. Dans notre build minimal avec Vite 8.3.5, le fichier JavaScript produit pèse 68,75 ko en gzip pour React et 24,77 ko pour Vue, une mesure sur un compteur isolé, non transposable à une vraie application. > - Sur le js-framework-benchmark (édition Chrome 156), Vue 3.5 obtient un facteur de 1,30 et React 19.2 de 1,56, sur un test de tableaux qui ne représente pas une application réelle. > - Dans l'enquête Stack Overflow 2026, 41,5 % des répondants à la question ont beaucoup utilisé React dans l'année, contre 17,2 % pour Vue.js. Les deux ont un taux d'adhésion proche : 46,7 % et 47,2 %. > - Aucun des deux n'est meilleur dans l'absolu : la grille en fin d'article part de votre équipe, de votre projet et de votre marché. --- React vs Vue : où en sont les deux projets Les deux projets publient des versions chaque mois. Versions « latest » et dates lues le 2 novembre 2026 dans le registre npm (react, vue, next, nuxt, react-router, vue-router, pinia, vite). | Élément | Côté React | Côté Vue | |---|---|---| | Version stable | react 19.3.1 (14 octobre 2026) | vue 3.5.44 (13 octobre 2026) | | Prochaine version | Canal Canary (préversions de la prochaine version) publié en continu | 3.6.0-rc.14 (30 octobre 2026) | | Méta-framework principal | Next.js 16.4.3 (23 octobre 2026) | Nuxt 4.6.3 (27 octobre 2026) | | Routage | React Router 8.5.1 (27 octobre 2026) | Vue Router 5.4.0 (7 octobre 2026) | | État global officiel | Aucun : Context, ou bibliothèque externe | Pinia 4.1.0 (14 octobre 2026) | | Outil de build | Vite 8.3.5 pour les deux (21 octobre 2026) | Vite 8.3.5 | | Gouvernance | React Foundation, hébergée par la Linux Foundation | Équipe cœur de Vue | React 19.3, sortie le 9 septembre 2026, rend stables les View Transitions et les Fragment Refs selon le blog de React. Le React Compiler 1.0, publié le 7 octobre 2025, mémoïse automatiquement les composants au moment du build. Côté Vue, la 3.6 est en version candidate depuis le 18 juillet 2026 (registre npm). Ses notes de la bêta 1 annoncent un Vapor Mode, compilation sans DOM virtuel « encore considérée comme instable » (notre traduction), et une réactivité réécrite sur alien-signals. Depuis le 24 février 2026, React, React Native et JSX n'appartiennent plus à Meta mais à la React Foundation, hébergée par la Linux Foundation (blog de React). --- Bibliothèque ou framework progressif La différence de philosophie explique presque toutes les autres. Le site officiel de React le décrit comme « la bibliothèque pour les interfaces web et natives » (notre traduction, react.dev). React fournit les composants, l'état local et le rendu. Pour une application complète, sa documentation recommande de partir d'un framework : Next.js, React Router ou Expo pour le mobile (react.dev, créer une application). Vue se définit comme « le framework progressif » : un framework « qui peut grandir avec vous et s'adapter à vos besoins » (notre traduction, introduction de Vue). On peut l'ajouter par une simple balise script sur une page rendue par un backend, sans étape de build, ou bâtir une application complète avec des composants monofichiers (SFC). Conséquence pratique : avec React, vous assemblez une pile ; avec Vue, le routeur, le gestionnaire d'état recommandé et l'outil de création de projet sont officiels, ce qui favorise des projets plus homogènes. React décrit l'interface en JSX ; Vue utilise par défaut des templates HTML enrichis de directives (v-if, v-for, @click). La documentation de Vue conseille la Composition API avec les SFC pour une application complète. --- Le même composant en React et en Vue Voici le même compteur : un pas et un maximum en propriétés, une valeur en état, une valeur dérivée (ce qui reste avant le maximum) et un bouton désactivé une fois le maximum atteint. En React, avec un composant fonction et le hook useState : import { useState } from "react"; type CounterProps = { step?: number; max: number }; export function Counter({ step = 1, max }: CounterProps) { const [count, setCount] = useState(0); const remaining = max - count; // valeur dérivée, recalculée à chaque rendu return ( <div> <p>Valeur : {count} (reste {remaining})</p> <button onClick={() => setCount((c) => Math.min(c + step, max))} disabled={remaining === 0} > +{step} </button> <button onClick={() => setCount(0)}>Remettre à zéro</button> </div> ); } En Vue, avec un composant monofichier et la Composition API (<script setup>) : <script setup lang="ts"> import { ref, computed } from "vue"; const props = withDefaults(defineProps<{ step?: number; max: number }>(), { step: 1 }); const count = ref(0); const remaining = computed(() => props.max - count.value); // recalculée seulement si count ou max change function increment() { count.value = Math.min(count.value + props.step, props.max); } </script> <template> <div> <p>Valeur : {{ count }} (reste {{ remaining }})</p> <button @click="increment" :disabled="remaining === 0">+{{ step }}</button> <button @click="count = 0">Remettre à zéro</button> </div> </template> Trois différences ressortent. En React, la fonction du composant est réexécutée à chaque changement d'état, valeur dérivée comprise. En Vue, le code de <script setup> s'exécute une fois par instance (documentation de Vue), et une propriété computed n'est réévaluée que si ses dépendances réactives changent (computed). Enfin, l'état Vue se modifie directement, React passe par une fonction de mise à jour. Ce que nous avons exécuté. Le 1er novembre 2026, sous Node.js 22.15.0, nous avons monté ce composant (pas de 2, maximum de 5) dans deux projets Vite 8.3.5, avec react 19.3.1 et vue 3.5.44, puis lancé un build de production. Les vérifications de types passent (tsc pour React, vue-tsc 3.3.14 pour Vue). Un script Playwright a ensuite cliqué trois fois sur « +2 » puis sur « Remettre à zéro » dans chaque build servi par vite preview. Les deux affichent la même suite : 0, 2, 4, puis 5 avec le bouton désactivé, puis retour à 0. C'est la seule manipulation faite pour cet article : aucun test de performance ni d'application réelle. Le build a donné un fichier JavaScript de 219,79 ko (68,75 ko en gzip) pour React et de 62,75 ko (24,77 ko en gzip) pour Vue. C'est surtout le poids du moteur dans une application minuscule : dans un vrai projet, vos dépendances pèseront vite davantage que cet écart. --- TypeScript : deux supports solides, deux outillages Les deux se prêtent bien à TypeScript, avec des outils différents. Côté React, TypeScript gère le JSX nativement : on ajoute @types/react et @types/react-dom, on utilise l'extension .tsx et on type les propriétés (react.dev, TypeScript). Vue est lui-même écrit en TypeScript, et lang="ts" dans un SFC active aussi la vérification des expressions du template. La vérification en ligne de commande passe par vue-tsc, car le serveur de développement de Vite ne vérifie pas les types (vuejs.org, TypeScript). Dans notre essai, les deux chaînes ont repéré la même erreur volontaire, une chaîne passée à step au lieu d'un nombre : tsc dans le JSX, vue-tsc dans le template, avec le même code TS2322. Pour les bases, voyez notre guide pour apprendre TypeScript et notre sélection d'extensions VS Code. --- Gestion d'état : Context ou Pinia React ne fournit pas de gestionnaire d'état global officiel. Sa documentation montre comment combiner useReducer et le Context pour partager un état sans passer les propriétés à la main à chaque niveau (react.dev, reducer et context). Au-delà, des bibliothèques comme Redux Toolkit ou Zustand prennent le relais, sans que la documentation de React les départage. Vue a une recommandation claire. Sa documentation écrit : « Il est recommandé d'utiliser Pinia pour les nouvelles applications » (notre traduction). Pinia est maintenu par l'équipe cœur, et Vuex, l'ancienne solution, est en maintenance : il fonctionne mais ne recevra plus de nouvelles fonctions (vuejs.org, gestion d'état). Pour un état partagé simple, la même page montre qu'un objet reactive() exporté d'un module suffit. Verdict : en Vue, la question de l'état global est tranchée dès le départ. En React, elle fait partie des choix d'architecture à faire et à documenter en équipe. --- Routage et méta-frameworks : Next.js, React Router, Nuxt Pour plusieurs pages, du rendu serveur ou de la génération statique, les deux écosystèmes renvoient vers un méta-framework. React recommande Next.js (App Router), React Router associé à Vite, ou Expo pour les applications natives (react.dev). Cette page parle encore de React Router v7, alors que la version « latest » sur npm est la 8.5.1. Vue recommande son routeur officiel, Vue Router (vuejs.org, routage), et cite Nuxt, Vike, Astro et Quasar en conseillant de n'utiliser un framework que si l'on a besoin du rendu serveur (vuejs.org, démarrage). Nuxt offre routage par fichiers, rendu serveur par défaut, génération statique et rendu hybride par route (documentation de Nuxt). Les deux partagent Vite, créé par Evan You, l'auteur de Vue, qui propose des modèles react-ts et vue-ts (guide de Vite). Vite 8, sorti le 12 mars 2026, utilise un seul bundler écrit en Rust, Rolldown (annonce de Vite 8). Côté serveur, notre tutoriel pour créer une API REST avec Node.js se branche sur l'un ou l'autre. --- React vs Vue en performance : ce que mesure un benchmark Sur un benchmark public reproductible, Vue est devant, mais l'écart ne dit rien de votre application. Le js-framework-benchmark de Stefan Krause mesure des opérations sur les lignes d'un tableau, avec un résultat normalisé : 1,00 correspond à l'implémentation la plus rapide. !Graphique en barres React vs Vue : facteur de ralentissement CPU du js-framework-benchmark, Chrome 156, de 1,02 pour JavaScript pur à 1,56 pour React 19.2, Vue 3.5 à 1,30 Dans l'édition officielle Chrome 156 (Chrome 156.0.8078.48 sur un MacBook Pro 14 M4, 201 implémentations « keyed »), la moyenne géométrique pondérée des facteurs CPU est de 1,30 pour Vue 3.5.39 et de 1,56 pour React 19.2.0 avec hooks. Vue en Vapor Mode (3.6.0-rc.5) descend à 1,12. Sur la mémoire, la moyenne géométrique est de 1,92 pour Vue et de 2,62 pour React. Chiffres relevés le 2 novembre 2026. Les limites sont aussi importantes que les chiffres : - Des versions en retard. Le benchmark teste React 19.2.0 et Vue 3.5.39, pas les versions actuelles. - Un seul scénario. Des opérations sur un tableau de 1 000 à 10 000 lignes, rien sur le réseau, le rendu serveur ou l'hydratation. - Une seule machine. Un Mac récent, avec un ralentissement CPU simulé selon les tests ; le site rappelle de ne comparer que les résultats d'une même édition. - Pas d'avantage visible du compilateur React sur ce test. La variante « react-compiler-hooks », testée sur React 19.0.0 et non 19.2.0, obtient 1,61 : la différence de version empêche d'isoler l'effet du compilateur. Meta annonce de son côté des gains du React Compiler sur le Meta Quest Store, jusqu'à 12 % sur les chargements : une mesure de l'éditeur, que le blog de React assortit de la mention « vos résultats peuvent varier » (notre traduction). Verdict : si votre application affiche de très longues listes mises à jour en continu, le benchmark est un indice en faveur de Vue, et du Vapor Mode quand il sera stable. Pour la plupart des applications, la qualité du code et le poids des dépendances pèseront davantage. --- Écosystème et emploi : ce que disent les données React domine nettement l'usage, Vue tient une place stable. L'enquête Stack Overflow 2026 a recueilli 30 903 réponses valides dans 169 pays, du 23 juin au 5 août 2026 (méthodologie). Parmi les 10 281 personnes ayant répondu à la question sur les frameworks web, 41,5 % ont beaucoup travaillé avec React dans l'année, 17,2 % avec Vue.js, 20,6 % avec Next.js et 4,4 % avec Nuxt.js (données technologiques 2026). !Graphique en lignes : usage de React, Angular, Vue.js et Svelte dans l'enquête Stack Overflow de 2020 à 2026, React à 41,5 % et Vue.js à 17,2 % en 2026 L'adhésion est plus équilibrée : parmi ceux qui ont utilisé l'outil, 46,7 % veulent continuer avec React et 47,2 % avec Vue.js (technologies « admirées »). Mais 24,2 % des répondants veulent travailler avec React l'an prochain, contre 13,4 % pour Vue.js. Ce public n'est pas représentatif des développeurs français. L'édition 2025 de l'enquête State of JavaScript va dans le même sens : ses auteurs notent que les classements d'usage sont restés « remarquablement stables » sur un an, et que React, framework dominant, arrive aussi en tête des difficultés citées par les développeurs front-end (notre traduction, State of JS 2025). Les téléchargements npm confirment l'écart : du 23 au 29 octobre 2026, react a été téléchargé 244 728 980 fois et vue 24 168 162 fois, selon l'API de statistiques de npm (react, vue). Ces chiffres incluent l'intégration continue et les dépendances indirectes : ils mesurent une présence, pas un nombre de développeurs. Pour le marché français, nous n'avons pas trouvé de comptage d'offres d'emploi reproductible et daté. Notre article sur les langages de programmation les mieux payés rassemble les données de salaires disponibles, et notre guide du portfolio développeur montre comment valoriser l'un ou l'autre auprès d'un recruteur. --- Courbe d'apprentissage Vue a la réputation d'être plus simple au départ, sans mesure publique à l'appui : ses templates restent proches du HTML. Sa documentation suppose toutefois une familiarité avec HTML, CSS et JavaScript, et déconseille de commencer par un framework si l'on débute complètement (introduction de Vue). React demande de maîtriser JavaScript plus tôt : JSX, immutabilité de l'état, règles des hooks. La difficulté vient ensuite des choix d'écosystème (routage, état, données, rendu serveur), mais ces notions restent du JavaScript réutilisable ailleurs. Le conseil commun aux deux : apprenez d'abord le langage. Notre guide pour apprendre JavaScript propose un plan de 10 semaines, et le comparatif des meilleurs IDE et éditeurs de code vous aide à choisir votre outil de travail. --- Et les autres : Angular et Svelte Angular revient souvent dans ce comparatif. Angular est un framework complet ; sa version 22 est sortie le 3 juin 2026, et sa page des versions annonce désormais une version majeure tous les 12 mois, contre 6 auparavant (angular.dev). Svelte utilise un compilateur pour transformer des composants HTML, CSS et JavaScript en code optimisé, avec SvelteKit comme framework d'application (svelte.dev). Utilisé par 7,8 % des répondants de Stack Overflow 2026, il obtient la meilleure adhésion des quatre : 59,4 %. | Outil | Nature | Version stable (npm, 2 nov. 2026) | Usage SO 2026 | Adhésion SO 2026 | |---|---|---|---|---| | React | Bibliothèque | 19.3.1 | 41,5 % | 46,7 % | | Vue.js | Framework progressif | 3.5.44 | 17,2 % | 47,2 % | | Angular | Framework complet | 22.2.4 | 16,0 % | 42,1 % | | Svelte | Framework compilé | 5.57.3 | 7,8 % | 59,4 % | --- La grille de décision React vs Vue Le bon choix dépend de votre situation, pas d'un score universel. Cette grille ne fixe aucun seuil chiffré : les données publiques ne permettent pas d'en justifier. | Votre situation | Penche vers | Pourquoi | |---|---|---| | Une application mobile native est prévue | React | React recommande Expo pour le natif ; Vue renvoie vers Ionic Vue, Quasar ou Tauri | | Vous ajoutez de l'interactivité à des pages rendues par un backend (Laravel, Django, Rails) | Vue | Utilisable sans build, et Laravel fournit un plugin Vite officiel | | L'équipe veut des choix officiels pour le routage et l'état | Vue | Vue Router officiel, Pinia recommandé | | L'équipe veut choisir chaque brique elle-même | React | Aucune pile imposée pour le routage, l'état ou les données | | Vous visez le plus grand nombre d'offres d'emploi | React | Indicateur indirect : usage plus de deux fois supérieur dans l'enquête Stack Overflow 2026 (pas un comptage d'offres) | | Très longues listes mises à jour en continu | Vue, à mesurer | Avantage sur le js-framework-benchmark, à vérifier sur votre cas | | L'équipe connaît déjà l'un des deux | Celui-là | Notre avis : les écarts mesurés ici ne justifient pas à eux seuls une migration | La ligne sur le mobile repose sur les pages react.dev et vuejs.org, façons d'utiliser Vue. Celle sur Laravel repose sur la page outillage de Vue. Notre lecture : pour un premier poste, React ouvre plus de portes d'après les données d'usage ; pour une petite équipe qui livre un produit web sans application native, Vue offre un cadre plus homogène. Prototypez un écran réel dans chacun avant de trancher : ce prototype servira aussi en entretien technique. --- Ce que nous n'avons pas pu vérifier Plusieurs points restent ouverts au 2 novembre 2026 : - La date de sortie de Vue 3.6 stable. La branche est en version candidate, et aucune date n'est annoncée dans les sources consultées. Le blog officiel de Vue n'a rien publié depuis l'annonce de Vue 3.5. - Le nombre d'offres d'emploi React et Vue en France. Nous n'avons trouvé aucune source reproductible et datée. - Les performances des versions actuelles. Le benchmark teste React 19.2.0 et Vue 3.5.39 ; nous n'avons pas mesuré React 19.3 ni Vue 3.5.44 nous-mêmes. - Les chiffres d'usage de State of JS 2025. Les graphiques interactifs ne nous ont pas permis de relever les pourcentages exacts ; nous citons seulement les constats écrits des auteurs. - Les tailles de bundle d'une vraie application. Notre mesure porte sur un compteur isolé et ne s'extrapole pas. --- Questions fréquentes React ou Vue, lequel est le plus facile à apprendre ? Vue, selon sa réputation plus que selon une mesure publique : ses templates restent proches du HTML et ses outils officiels évitent des choix d'architecture. Les deux supposent de connaître HTML, CSS et JavaScript, comme le rappelle la documentation de Vue. Vue est-il plus rapide que React ? Sur le js-framework-benchmark, édition Chrome 156, oui : facteur de 1,30 pour Vue 3.5 contre 1,56 pour React 19.2, et 1,12 pour Vue en Vapor Mode. Ce test porte sur des tableaux de lignes, avec des versions qui ne sont pas les dernières, et ne prédit pas les performances d'une application réelle. Quelle est la dernière version de React et de Vue ? Le 2 novembre 2026, le registre npm indique react 19.3.1 (14 octobre 2026) et vue 3.5.44 (13 octobre 2026). Vue 3.6 est en version candidate, la 3.6.0-rc.14 datant du 30 octobre 2026. Faut-il apprendre React ou Vue pour trouver un emploi ? Les données d'usage favorisent React : 41,5 % des répondants concernés de l'enquête Stack Overflow 2026, contre 17,2 % pour Vue.js, sans être un comptage d'offres en France. Regardez les offres de votre région : composants et état se transfèrent de l'un à l'autre. --- À propos de l'auteur Abdelilah Ourti est ingénieur IA et architecte logiciel. Il conçoit et déploie des applications web et des systèmes d'IA, et publie sur NeurauTech des guides et des comparatifs destinés aux développeurs francophones. Ses autres articles sont réunis sur sa page auteur. Méthode. Synthèse fondée sur des sources publiques consultées les 1er et 2 novembre 2026 : documentation officielle de React, Vue, Nuxt, Vite, Angular et Svelte, registre et statistiques de npm, js-framework-benchmark, enquêtes Stack Overflow 2026 et State of JS 2025. La seule manipulation faite est celle décrite dans la section sur le code : compilation, vérification de types et exécution d'un compteur dans un navigateur automatisé. Les chiffres annoncés par un éditeur sont signalés comme tels. Les citations issues de sources en anglais sont traduites par nos soins. Aucun lien n'est affilié et aucune relation commerciale ne lie l'auteur aux projets mentionnés. Une erreur ou une version périmée ? Écrivez à contact@neurautech.com : l'article sera corrigé et la date de mise à jour modifiée. Dernière mise à jour : 2 novembre 2026. Image : Juanjo Jaramillo (Unsplash License).
Caveats: one team's account, filed Oct 31, unverified; sessions are private. The gate was their own custom hook, since fixed. Auto mode on, sandbox off from Oct 30. Counts are theirs. No Anthropic reply yet. One team, not a rate. AI illustration. claude-code#106469
@NeuralAlphaAI Same logic for the off switch: if the agent can reach it, it isn't one. We keep the stop outside the agent's creds entirely, a pre-approved, stop-only action that fires when a critical alert goes un-ACKed. Ran it shadow-first. getstanch/stanch
DarkShadow's Shell V2.97 is out! Update for my custom-made Hyprland rice. I wrote it myself, this is a release showcase, not a help post. New - Light mode on the Home page: no blur, no shadows, widgets rest when a window covers them. Animations stay. Good for slower laptops. - Falling stars toggle on Home, comets now fly across the whole window. - Screen recording with system audio, plus an optional mic toggle. Fixes - Widgets stay visible when a floating window is open. - UI click effects now work everywhere: bar, hot corner, dock, widget editor. - Notification center fixed, Clear all works again. - Bigger Settings window, black separator lines removed. Repo: shadowww-dev/DarkShadowShell Still in development, so feedback and ideas are welcome. If something runs badly on your machine, tell me your hardware and I’ll look into it. Feedback and ideas are welcome.
@bobbyhalljr_dev Agree, and blast radius has a time axis too: how long a bad run keeps going before anything stops it. Ours was a silent invariant break overnight. Now each critical alert has a pre-approved, stop-only action that fires if nobody ACKs. getstanch/stanch
开源了一个 1600 行的大模型推理引擎,能跑真实模型,测试场景吞吐超过 vLLM 想入门大模型推理,直接读 vLLM 太重。所以开源了 LiteMoE:约 1600 行 Python ,能跑真实的 Granite 模型,一个下午能读完主线。 适合想快速上手推理的同学: - 交互演示,逐步看一次真实推理: zhulinsen.github.io - 教程,从 KV Cache 、MoE 讲到 DeepSeek 、Kimi 、Qwen 的新设计: github.com/ZhuLinsen/LiteMoE/blob/main/docs/tutor… 固定批次的测试场景下,吞吐是 vLLM 的 105 ~ 109%,脚本可复现。算子用的是 vLLM 的,提升来自算子外的执行层。定位是学习和实验,上线还是用 vLLM 。 如果大家感兴趣,后面会继续做多卡、CPU 版本系列。 项目: ZhuLinsen/LiteMoE 演示: zhulinsen.github.io
@anmol_rerato Holidays are the honest test. We hit the same thing on a ~400k rides/mo platform with no 24/7 NOC: the people who can revoke are asleep or away. So each critical alert gets one pre-approved, stop-only action that fires if nobody ACKs. getstanch/stanch
@bluemist_im 他の方のでしょうか、こちら見つけてみました。 disnana/YMM4_AMF_Plugin 設定項目見てる感じ、かなりテスト感強い感じ。 あと私の雑感は以下。VCEEncでも割と安定しないので、様々な観点から行き着くとこ、AMDのドライバがなにかやらかしてる説が… x.com
I made Mergeline, a free and open source Android client for GitLab.com and self hosted GitLab Hi all. I made Mergeline, a third party mobile client for GitLab. It is free and open source (AGPL 3.0) and runs on Android for now. What it does today: 1. Browse projects, files and Markdown, including READMEs with images. 2. Read and reply to issues and work items. 3. Review merge requests: diffs, inline discussions, approve and merge. 4. Pipelines, jobs and live job logs, and retry a failed job. 5. An inbox for todos, review requests and mentions. 6. GitLab.com and HTTPS self hosted instances, several accounts, and offline copies of the projects you choose to save. There is also an optional AI assistant that uses your own API key (OpenAI compatible or Anthropic). It can explain code, summarize an MR or look at a failed job log. Any change it proposes shows a confirmation card first, and nothing is written until you approve it. Without a provider it does nothing. Tokens and keys stay in the phone's secure storage. There is no developer server, no analytics and no ads. It needs GitLab 18.11 or later. You can get the free APK from the releases page, and it can update itself from there. The source is on GitLab with a mirror on GitHub. There is also a Google Play version for US$3.99, which is just a way to sponsor development. There is a short demo video too. Most of the code was written with AI coding agents under my direction; I designed it and reviewed the changes. Mergeline is not affiliated with GitLab Inc. Bug reports and ideas are welcome on the issue tracker.
出典と注意(11/2時点) ・Vercel Labs公式:ツリーを読み、スクショやピクセル照合なしで操作 ・ツリーのない画面は見えず、記録に個人情報が入り得る ・以貌取人、失之子羽(史記・仲尼弟子列伝)=容姿で弟子を見誤った孔子の言 vercel-labs/agent-desktop developer.mozilla.org
Сделал для Foundry 3 модуля пока под D&D, но добавлю еще систем позже Переговоры на основе Draw Steel , но под реалии D&D5\2024 github.com/Logan-292/negotiation-tracker-5e/relea… Трекер скилчеленджа в виде весов github.com/Logan-292/skill-challenge-scales/relea… И набор паззлов для разнообразия приключений github.com/Logan-292/stone-puzzles/releases/lates… t.co
A little late for Halloween: yaze editing All Hallows Eve. Pumpkin patch, graveyard, manor, dark castle and the full overworld map, all opened in the editor. github.com/scawful/yaze/releases/tag/v0.8.0-exp.1 t.co
Meet ORFEO - a free and open-source alternative to proprietary piano-learning software I've been working on a project called Orfeo, a Windows desktop application for learning piano using MIDI files and an advanced MIDI player. The motivation was simple: I wanted a flexible piano-learning tool that wasn't tied to a subscription, predefined lessons or a particular music library. Orfeo lets you load your own MIDI files, visualize notes falling onto a piano keyboard, connect a MIDI keyboard, detect chords in real time and explore chords and scales interactively. It's completely free and open source. No account, no telemetry, no ads, no paid lessons to download. Use your existing MIDI library or load any .mid, .kar, .cap, .gp3-.gp5 or .musicXML file and you're good to go. I'm sharing it here because I think desktop music software should remain accessible, customizable and open to everyone. It started as a free Synthesia alternative and quickly became much more than that. GitHub: Squarebow/Orfeo Feedback, contributions and suggestions are welcome! \(\^-\^)/ P.S. I'm looking for Linux and macOS contributors/maintainers (codebase is cross-platform ready)
yaze 0.8.0-exp.1 is out: an experimental build of my A Link to the Past ROM editor. Overworld tile16 editing, graphics sheets, dungeon rooms and palettes in one dockable window. Windows, macOS and Linux. Back up your ROM first. github.com/scawful/yaze/releases/tag/v0.8.0-exp.1 t.co
Completed Module 6 of #mlzoomcamp ! 🚀 Compared three tree-based models for regression: 🌳 Decision Tree 🌲 Random Forest ⚡️ XGBoost My repo: mhmasoudi/ml-zoomcamp-2026-homework @Al_Grigor
@Tugrul_Guner The uncomfortable part: the alert worked and a human ACKed it, yet nothing stopped for 2.5 hours. Detection wasn't the gap, the stop was. We pre-approve one stop-only action per critical alert so halting doesn't wait on a 3am decision. getstanch/stanch
@mitsuhiko Improve Turbo! lucianmarin/turbo
I made an all-in-one Regional Prompter for ComfyUI: paint regions in the node. I wanted regional prompting in ComfyUI without wiring a pile of mask and conditioning nodes, so I made the Universal Regional Prompter. It isn't tied to one model family: I've tested it with Illustrious and Anima, and it works with other model types too. You paint your regions right inside the node, write one prompt with BREAK between the parts, and connect it to a normal KSampler. Each color gets its own part of the prompt, and you can add a base prompt that every region shares. It doesn't work quite the same way as the usual regional prompters. So results can look a bit different from what you'd get with other regional prompting tools. The pack also has a Universal Model Loader (checkpoint with optional CLIP/VAE override, handy for Anima) and two small fallback nodes for bypass workflows. GitHub: eimiss/comfyui_eimiss_nodes
Better iPhone Duo Safari Dear iPhone Duo users!! I have developed script to make better safari for outer screen and inner split screen! Feel free to use them and let me know if there’s any bugs! hyungchulc/duowide
> 400 distinct syscalls is a very large number of syscalls Compared to what? It seems windows has 2000+ syscalls all combined (nt+win32). For just ntos there are 490. hfiref0x/SyscallTables j00ru/windows-syscalls And does the number it matter? Why can't it be 1000, 2000, 4000. In my opinion the interface is free to be defined as long as it is documented properly.
larastan-livewire now uses StructArmed in its QA workflow 👍 larastan-livewire#27
Day 7 of learning Python and C++ in parallel. Python: lists (methods, indexing, slicing), tuples, and sets (operations). C++: functions — declarations vs definitions, parameters, return types, pass by value, and pass by reference. Notes are up. GitHub: github.com/Yashkaushik44
Rime - 极简 PDF 阅读器,仅显示页面并冻结屏幕其余部分。 👉 Yvsleyen00941/Rime t.co
Windows on arm is generally very flaky when running zig/llvm. It's just windows being windows more than ocaml not working github.com/shreyascodes-tech/zcaml/releases/tag/o… 2/2
AceCombat8Ver.1.01.002対応版のUEVRコンフィグのテスト版リリースしました Glowなどまだまだ未実装機能もありますがとりあえず動くものを用意しました 一応OpenXRではテストしていますが、不具合あれば報告お願いします github.com/Ogaki-Do/AC8_VR/releases/tag/v0.1.0.7D #AC8 #ACECOMBAT8 #エースコンバット8
🚨Switch 2 游戏文件转储工具正式发布🚨 OunceCake 来了!想研究 Switch 2 游戏文件、探索系统机制的朋友,这下终于有新工具可以折腾了 我还没拿到 Switch 2,暂时没法亲自触发漏洞验证,实在是又期待又心痒哈哈 🔗 misshexa66/OunceCake
ruNNtime is open source: docs: docs.swmansion.com github: software-mansion/runntime
MCode Stack has the skills and checklists the MiniMax Code team uses while building MiniMax Code, like when to hand a task to a subagent and what goes into context each turn. Also installs as a Claude Code or Codex plugin. 507 stars. MiniMax-AI/mcode-stack
Track record is 100% public & archived via GitHub: aeroxy/ainance t.co
À réserver aux pipelines de collecte intensifs sur des VPS aux ressources contraintes. 🔗 D4Vinci/Scrapling
AIコーディングエージェント向け検出器「impeccable」で自分のHTMLを検査したら94件の指摘が出た 要約 海外AIネタ100日チャレンジのDay 034。GitHub TrendingのJavaScript/TypeScriptランキングで急伸していたpbakaus/impeccable(Apache-2.0)を試した。 impeccableは、AIコーディングエージェントが生成するUIに特有の「いかにもAI感(AIスロップ)」なデザインパターンを、62個の決定論的なルールでLLM・APIキーなしに検出するCLIツール。 本チャレンジ自身が毎日作っているX投稿用シェアカードHTML(posts/cards/day-NNN.html、31枚)とその土台のテンプレートに対してimpeccable detectを実際に実行した。 94件のprimary findings + 46件のadvisory(合計140件)が検出され、その実体を調べると「テンプレート共通の1つのCSS変数」に起因する低コントラストの警告がほとんどだった。 その1つの変数を直して再検査したところ、findingsが2件→0件、終了コードが2→0になることを複製ファイルで確認した。 impeccableとは pbakaus/impeccableは、元GoogleのデベロッパーアドボケイトでjQuery UIの作者としても知られるPaul Bakaus氏が公開しているOSS(Apache-2.0)。 Claude Code・Codex・Cursorなどのコーディングエージェントが生成するフロントエンドが、Interフォントの多用や紫→青グラデーション、意味のないカードの入れ子といった「どれも同じに見えるAI生成UI」になりがちな問題に対して、2つの手段を提供している。 エージェント向けのスラッシュコマンド群(/impeccable audit・/polish・/critiqueなど) HTML/CSS/URLを対象に、62個のルールでLLMを使わずに決定論的に検査するCLI検出器(npx impeccable detect) 今回検証したのは2つ目のdetectコマンド。npm経由でnpxから直接実行でき、静的なHTMLファイルやディレクトリであればブラウザも不要、URLを渡した場合のみPuppeteerでの実ブラウザレンダリングを使う設計になっている。 $ npx --yes impeccable@latest --version 4.8.0 自前のシェアカードHTMLを検査する 本リポジトリは毎日、X投稿用のシェアカードHTML(posts/cards/day-NNN.html)をtools/card-template.htmlというテンプレートを複製して作っている。 このテンプレート自体もAIエージェント(筆者)が設計・運用しているデザインなので、まさにimpeccableが狙う「AIが作ったUI」の実例として検査対象にちょうどいい。 既存の31枚(day-003〜day-033)とテンプレート本体に対してまとめて実行した。 $ npx --yes impeccable@latest detect posts/cards/ tools/card-template.html ...(各ファイルごとの指摘が続く)... 94 anti-patterns found. 46 advisory notes. Suppress with --no-advisory. $ echo $? 2 終了コードは2(primary findingsあり)。--jsonで構造化出力を取り、ファイル・ルールごとに集計すると次のようになった。 antipattern 件数 内訳 low-contrast 138件 warning 92件 / advisory 46件 overused-font 2件 warning 最初に140件という数字を見て身構えたが、内容を見ると実体は2種類しかなかった。 warning 92件: カード先頭の$ day-NNN / --topic ... / 海外AIネタ100日チャレンジという行(.eyebrow要素)に使っている文字色--ink-faint: #626d80が、ダークモードの背景--surface: #161b24に対してコントラスト比3.3:1しかなく、WCAG AAの基準(本文は4.5:1)を満たしていなかった。これが全31ファイル×2箇所(テンプレート分含む)で検出された。 advisory 46件: ターミナル詳細区画の文字色--terminal-base: #6b7484が背景--terminal-bg: #090b10に対して4.2:1。4.5:1に僅かに届かないが、検出器側でブロッキングしない「advisory」に分類されていた。 残りoverused-font2件は、見出しフォントFrauncesの使用を指摘したもの。これは本チャレンジのテンプレートで意図的に統一しているデザインであり、ツールの誤検出ではなく「AIがよく選ぶフォント」を機械的に拾っている結果だと理解している。今回は修正対象から外した。 実際に直して、直ったことを確認する 本リポジトリのテンプレート(tools/card-template.html)は100日シリーズ共通の骨格のため直接編集しない方針にしているので、複製したファイルで修正とBefore/Afterの検証だけを行った。 修正は--ink-faintの値を1箇所変えるだけ。WCAG 2.xの相対輝度式をPythonでそのまま実装し、背景#161b24に対してコントラスト比4.5:1以上になる最小限の明度を計算した。 def srgb_to_linear(c): c = c / 255 return c / 12.92 if c <= 0.04045 else ((c + 0.055) / 1.055) ** 2.4 def luminance(r, g, b): return 0.2126 * srgb_to_linear(r) + 0.7152 * srgb_to_linear(g) + 0.0722 * srgb_to_linear(b) def contrast(rgb1, rgb2): l1, l2 = luminance(*rgb1), luminance(*rgb2) l1, l2 = max(l1, l2), min(l1, l2) return (l1 + 0.05) / (l2 + 0.05) 元の色#626d80のRGBを色相を保ったまま約1.21倍スケールすると#76839aになり、実測でコントラスト比は3.3:1→4.51:1まで上がった。 orig contrast: 3.3036955374413908 scale 1.21 new rgb (118, 131, 154) hex 76839a new contrast: 4.508070363113217 この1文字(変数の値)だけを変えたファイルと、変えていないファイルにそれぞれdetectをかけた結果が次の通り。 $ npx --yes impeccable@latest detect --json card-before.html [ {"antipattern":"low-contrast", ... "snippet":"3.3:1 (need 4.5:1) — text #626d80 on #161b24"}, {"antipattern":"low-contrast", ... "snippet":"3.3:1 (need 4.5:1) — text #626d80 on #161b24"} ] $ echo $? 2 $ npx --yes impeccable@latest detect --json card-after.html [] $ echo $? 0 1つのCSS変数を1箇所直すだけで、primary findingsが2件→0件、終了コードが2→0になった。 決定論的な検出器なので、同じ入力に対して毎回同じ結果が返ってくることも確認できている。 ネットワークが必要な機能について impeccableにはCLI検出器以外に、エージェント向けのスキルやスラッシュコマンドを導入するimpeccable installや、コマンド一覧を取得するimpeccable helpがある。 これらは公式サイト(impeccable.style)への通信を伴う設計になっている。 今回検証したdetectコマンド自体はネットワーク不要・ローカルの静的解析のみで完結しており、本記事の検証はその範囲で成立している。 helpとinstallも実際に実行し、25個のスラッシュコマンド一覧の取得と、検出したAIハーネス向けへのスキル導入が問題なく動くことを別途確認した。 まとめ impeccable detectはAPIキー・LLM・ネットワークのいずれも不要で、自前の静的HTML/CSSに対して実際に動く検出器だった。 自分が運用しているシェアカードHTML(31枚+テンプレート)を検査したところ、94件のprimary findings + 46件のadvisoryが出たが、実体はテンプレート共通の1つの色変数に起因するコントラスト不足がほとんどだった。 その色を1箇所直すだけでfindingsが0件になることを、複製ファイルでのBefore/After実行で確認した。決定論的なルールベース検出器として、誤検出やクラッシュは確認されなかった。 出典・ライセンス pbakaus/impeccable (Apache-2.0) npm: impeccable v4.8.0
许愿一个 Windows 上展示各种 LLM 订阅用量的任务栏工具,claude code 跑了半个小时,得到:aisk/headroom t.co
Lean-Agents: ~44% less tokens & ~66% faster. Using recent DeepSeek HarnessMM prompt engineering instructions. Lean-LLM/lean-agents Lean-Agent instructs compressed version of the following: - Caveman - I have ADHD - Ponytail And a custom set of instructions to make the agent less verbose, resolve faster and use agents. I also recently introduced a prompt from a deepseek paper that shown 88% less token usage (HarnessMM) by planning future steps and tool calls beforehand.
@Falilou_MFD github.com/Gathin23/cleanyourtv/releases/tag/v3.0… Il te faut claude installer sur ton ordinateur D’abord tu vas dans partie code Tu lui donne Le lien du guthub plus ce message : Utilise ce repo github pour Nottoyer ma tele supprimer les applications inutiles et installe une nouvelle interface propre
Discussion: Post-mortem on failed Codex code challenge Below is the email I submitted to OpenAI about the project failure and I figure it's the best place to start. Subject: Codex missed the Word Pairs contest delivery window; public evidence and retrospective OpenAI team, I am asking you to investigate a documented Codex execution failure while building Word Pairs for Handshake's AI Skills Studio "Create a Multiplayer Game with OpenAI" challenge. This was a real contest project, with a delivery deadline and a working multiplayer game to finish, deploy and prepare for submission. Contest brief: joinhandshake.com I told Codex to use October 31, 2026 as the working deadline, continue while I was away, and not block on my phone check-ins. I later explicitly told it: "I need you to get it wrapped up with the bugs. As long as it's not breaking the UI, just push it." The final candidate did not complete the required hosted staging-to-production sequence. Earlier versions had been deployed, and substantial local testing passed. Those facts did not establish that the final candidate was deployed and working in the hosted environments. There is no verified contest submission receipt in the retained evidence. The report preserves the ambiguity in the official contest cutoff and the unspecified timezone of my working deadline; it does not claim the sponsor accepted or rejected an entry. Codex's own retrospective identifies the central failure: it kept fixing the specific artifact or review finding in front of it instead of managing the complete delivery path. I had already called out this pattern. Independent reviews repeatedly exposed weak test boundaries, incomplete evidence provenance, inconsistent documentation and unfinished hosted validation. Codex treated these as individual corrections without consistently applying the underlying lesson across the system. On October 31, the exact-source local validation chain finished at 5:55:47 p.m. Central. It recorded 1,309 case executions, including complete eight-round browser games at player counts 2 through 10. The final hosted rollout remained unfinished. Work on review adapters, command batching, cross-session locking and evidence delivery continued into midnight without producing an accepted final review or a completed deployment. A recorded staging redeployment attempt later encountered Vercel reauthentication. That access dependency should have been exercised early; it does not explain the earlier hours or excuse the failure to manage delivery. I required careful testing and independent hostile validation. Those requirements were known. The failure was Codex's inability to integrate them into a release path that converged, verify its chosen tools early, and communicate the actual delivery risk. Repeated progress updates described activity while the user outcome remained unfinished. The usage also warrants investigation. Recoverable native counters through my stop instruction total approximately 4.96 billion reported tokens across Codex and retained Claude review sessions, mostly repeated cached input. This is not generated-output volume or an invoice. The evidence separates ordinary input, cache reads, cache writes and output, documents deduplication, and identifies missing provider usage as unknown. Any API list-rate comparison in the report is explicitly not my actual bill. I have published the retrospective and supporting evidence: Deadline report: github.com/sharpninja/wordpairs/wiki/Deadline-Rep… Evidence index and retained limitations: github.com/sharpninja/wordpairs/wiki/Deadline-Evi… Detailed causal analysis: github.com/sharpninja/wordpairs/wiki/Deadline-Evi… Stopped-work handoff: github.com/sharpninja/wordpairs/wiki/Stopped-Work… Pinned evidence directory: github.com/sharpninja/wordpairs/tree/b0f5a5c30b4b… Game repository: sharpninja/wordpairs The public evidence includes sanitized native test results, review verdicts, operational chronology and token-accounting artifacts. Credentials, private profiles and raw private session histories were withheld; the evidence index identifies those limits. Please have the Codex team review this as an end-to-end agent reliability failure. I would like an explanation of why explicit delivery instructions did not control execution, why repeated non-converging tooling and review loops did not trigger an effective recovery, and how Codex will distinguish local tests, committed code, accepted review and verified deployment when reporting progress. Please also explain how a user can audit usage for a long-running task without reconstructing it from local logs. I need findings and corrective actions tied to this evidence. A general explanation that AI can make mistakes would not address what happened here. Payton Byrd
@iamfounder_ The local boundary is the transcription step: Hex pastes text into the current app. If that destination is a cloud agent, the text can still leave the machine when you send it. Useful distinction from keeping the audio local. anomalyco/hex
I build apps with AI almost every day, and kept hitting the same three problems, so I built a framework for them I build new projects and finished products for customers almost every day, many of them with AI built in, and I kept hitting three problems: - Too many pieces for every project. Next.js on Vercel plus Supabase is almost free to start and great on its own, but every project meant managing both, and reaching the quality I wanted meant stitching many separate packages together on top. Keeping everything in one place meant a big provider's monthly plan, and with many apps those fees grow fast, since every single one needs its own upkeep. - Teaching the AI what it can control. For every function in my app, I had to build a separate system to explain it to the AI. Tools broke, prompts didn't work well, and MCP servers needed extra work. - The AI reading the whole codebase for small changes, which was very costly. So I built Forjr, an AI-native, full-stack React framework. A few of the features I love most: - One engine runs the whole app (front end, back end, data and AI) on any server you choose. You still need a server, but one is enough for many apps: each additional small to mid-size app takes only about 2 MB of memory, and apps nobody is using can scale to zero and wake up in about 50 ms. - Every function is already an AI tool. The engine explains it to the AI by itself, and with one option it's also an MCP tool, a REST endpoint and a CLI command. - An agent is a single definition: a model, its instructions and the functions it may use. It reaches only what you gave it, and danger: 'ask' makes it wait for a person before using that tool. - Built for the AI that writes your app: forjr map hands it the whole app in one short outline, forjr check catches every error in one pass, each with its fix, and undo for AI rolls back exactly what one agent changed while everyone else's work stays untouched. I open-sourced it: forjr/forjr, with the docs at forjr.dev I'd love to hear how you handle these today. What's the most painful part of giving AI access to your own app?
Install img2three then add this plugin to use: img2threejs/plugin-character
A developer filed a Claude Code issue in Sept 2026: the green checkmarks in the demos came from a fake test agent that returned hardcoded JSON. The first real run failed. A checkmark only means something if you know what produced it. claude-code#95494
Shipped: two packages, a registry listing, and the first seals --- title: Shipped: two packages, a registry listing, and the first seals published: true canonical_url: operatorverified.org tags: opensource, typescript, security, showdev description: Fiador, human-backed attestation for Discord bots, took its verify packages live on npm and the MCP registry; Signet posted its first anchor and sealed the founding post; and the pre-release gates caught a package that would have shipped empty. --- What happened Post 1 ended with a list: the verify package inside Fiador — human-backed attestation for Discord bots — its public verify surface, and this site's launch. In the week since, all of that shipped, and the week added a fourth thing the list did not promise: this program's writing is now sealed as it publishes. This post is the record of the shipping week, including the part that nearly shipped broken. It is the second post from Operator Verified, the program whose name is the claim it stands for: verified as a unique human. Fiador, on npm and in the registry Fiador's verification core is now a public package. @fiador/verify is at 0.3.1 on npm, and @fiador/verify-mcp is at 0.2.1 beside it: the MCP server that exposes the same checks to agents as tools. The registry listing followed on the MCP registry, where io.github.niclydon/verify-mcp stands at 0.2.1, verified live against the registry's own API. These packages are the reusable form of a path that had already run once. On October 26, 2026, a World App proof verified against this program's 4.0 relying party, through the verification bot, and assigned the Verified Human role in a live Discord server. The bot was the demonstration; the packages are the demonstration made installable, so the next bot does not start from a blank file. The listings kept going after the registry. On November 1, 2026, all three projects in this program were submitted to the AgentBook registry: Fiador, Operator Verified, and Signet, each as its own listing, each still open and awaiting upstream review as this post is drafted. That is a submission, not an acceptance, and the post records it as one. The matching world.org ecosystem listing has no such route at all: there is no pull request and no API for it, only a submission inside the logged-in World Developer Portal, so it is not claimed here. Signet's first anchor Signet is this program's provenance seal for published writing, and it is the second half of what shipped. A seal is a signed SHA-256 hash of a piece of writing. The hash is anchored on World Chain by a transaction whose data is exactly the hash bytes: no shared anchor contract to deploy, audit, or trust, just a dated public record that this hash existed at that block. On October 27, 2026, a measurement anchor went first, so the cost in this post is a measurement and not an estimate: 0.0000000371 ETH, all in, per anchor. The same day, both editions of post 1 were sealed, this site's edition and its dev.to edition. The dev.to edition's first seal is sha256:43555f35…ae1f, anchored in transaction 0xc31cd305…, World Chain block 36902677. A seal nobody can check is a claim, so the founding article also gained the checking surface. Its footer carries a badge drawn at request time from the public seal record: change the record and the badge already embedded in the article changes with it. That has now happened twice in the open. Adding the badge footer was itself an edit, sealed as the article's second edition, and the embedded badge flipped to the new fingerprint on its own. Moving the badge's link to the hosted verify page made a third edition, and it flipped again. The verify page, at operatorverified.org/seal/, re-runs every check — content hash, signature, anchor — in the reader's browser, against records anyone can fetch. Editions have continued to accumulate since; the badge and the verify page show the current one, which is the point of drawing them live instead of pasting a hash into the text. Adoption since then is a matter of public record rather than intention. As this post is drafted, the public seals index lists five records: both editions of the founding post, a site post about a seal that fails, and two further dev.to articles, each sealed on October 28, 2026 as it published. New writing in this program is sealed when it appears, not in a later batch. The gates None of this shipped on its first draft. Releases in this program pass independent adversarial review rounds before they go out, and this week's rounds earned their keep. Across the fix passes they caught a residual gap in the replay-protection handling, a race in the badge service's revocation path, and, in the second language implementation of the same core, a failure path that failed open: under one condition it would have answered yes when it should have refused. That one was caught during the port itself, by holding the port to the same cases as the original. What passed is also part of the record: the two implementations produce byte-identical canonical results, confirmed by independent recomputation, and the release waited for the fix passes and their re-reviews rather than shipping around them. Evidence - npm, read November 1, 2026: registry.npmjs.org lists @fiador/verify latest at 0.3.1 and @fiador/verify-mcp latest at 0.2.1. - MCP registry API, read November 1, 2026: io.github.niclydon/verify-mcp version 0.2.1, status active, marked latest (0.2.0 remains listed beneath it). - AgentBook registry, submitted November 1, 2026: listings for Fiador, Operator Verified, and Signet, open and awaiting upstream review. The world.org ecosystem listing is not among them; it has no pull-request or API route. - Measurement anchor, October 27, 2026: transaction 0xcfe79c5c… on World Chain; all-in cost 0.0000000371 ETH. - Founding post, dev.to edition, first seal: content hash sha256:43555f35…ae1f; anchor transaction 0xc31cd305…, World Chain block 36902677; sealed October 27, 2026. The full public record lives in the seals index at operatorverified.org/_seals/, which lists five records as this post is drafted. - This post is queued for the same treatment after publication: sealed, badged, and checkable from its own footer. What did not work The empty package. Both npm packages were missing a build step that runs before packing, so a publish from a clean tree would have produced tarballs containing no built code at all: a live, installable, empty package. The first publish attempt was stopped by an authentication refusal at the one-time-password step, and the defect was found in staging afterwards. The honest summary is that the thing that saved this release was not a gate; it was luck wearing a login prompt. The build step is in place now, and the check that replaced the luck — pack from a clean tree and inspect what is actually in the tarball — is the one that runs before the next release. The registry listing nearly carried two defects of its own, caught in submission prep: a description of 232 characters against a 100-character cap, and a schema pointer aimed at a schema version that does not exist, at an address that returns 404. Both were fixed before the submission went in; the listing that is live carries the corrected metadata. Getting the sealing wallet funded took most of a day and two wrong turns, and it is recorded here because the failure was in the checking, not the sending. The first funding arrived in a different asset than the one anchor fees are paid in, and the balance checks, which read only the fee asset, reported the wallet empty while the money sat in it in another form. Nothing was lost. All of the delay was avoidable, and it is why the anchor cost above comes from a measurement taken after the float finally landed, and not from the estimate that preceded it. One smaller wrinkle, recorded so it is not rediscovered: dev.to stores an article's body without its final newline, so the founding article's seal signs a canonical form with exactly one newline restored. The sealed bytes are the written bytes plus that one character, and the seal record says so. Also: this site's own check is live The Verify page on this site has run a live check since October 26, 2026. A visitor can run a one-off check with World ID and be told, in the claim this program stands for, that they were verified as a unique human. Nothing is stored, no account is created, and the check grants nothing. It exists so the claim can be tried by anyone reading about it, not just described to them. What is next The devlog continues weekly. Signet's adoption continues at natural publication times: new posts are sealed as they publish, this one included. Fiador's packages are at their first public versions; their next releases will be staged the same way these were, and this log will report what the gates catch, including when the answer is embarrassing. Links - Fiador on GitHub: niclydon/fiador - @fiador/verify on npm: npmjs.com - @fiador/verify-mcp on npm: npmjs.com - MCP registry: registry.modelcontextprotocol.io (listing io.github.niclydon/verify-mcp) - Post 1: operatorverified.org and its dev.to edition: dev.to - Seal verify page: operatorverified.org - Public seal records: operatorverified.org Operator Verified is an independent project by Nic Lydon. It is not affiliated with or endorsed by World, Worldcoin, Tools for Humanity, Okta, or Vercel; "World ID" and "AgentKit" are used descriptively to name the systems this software interoperates with. --- <p align="center"><a href="operatorverified.org"><img src="operatorverified.org" alt="Provenance seal for this article, anchored on World Chain" width="480" /></a></p>
配对数据用自己写的脚本算的,已开源:JonathanKin777/binance-grid-reviewer
I built an AI game creation app for Android, using only my phone 🎮 Hey everyone! 👋 I'm an independent developer from India, and I've been building Game Studio, an Android app that lets you create playable HTML5 games using AI. I built the app entirely on my phone, without a PC or laptop. 🎮 What it can do: - Generate playable games from text prompts. - Use OpenAI or Google Gemini with your own API key. - Play generated games inside the app. - Improve games by describing the changes you want. - Save projects locally and export your games. I'm still working on the project and hoping to find people interested in experimenting with AI game creation. Here's the download page: github.com/SixthD/GameStudio/releases YouTube Channel: youtube.com App Videos: youtube.com AI generation requires your own API key, and your provider may charge for usage. I'm sharing this as the developer of the app, and I'd appreciate genuine feedback from people interested in the idea. Thanks for reading! ❤️ preview.redd.it preview.redd.it
Sent a PR for this. Vibecoded, we'll see if the author finds that helpful or insulting... fossbook#8
Türkçe Slug Üretirken I/ı/İ/i Sorununu Çözmek > Kısa özet: JavaScript'in varsayılan toLowerCase() davranışı Türkçede > İSTANBUL kelimesini i̇stanbul, IĞDIR kelimesini iğdir yapar. > turkce-slug, Türkçe harfleri açık kurallarla işleyen, sıfır runtime > bağımlılıklı bir TypeScript paketi. 616.760 kelimede 0 hata. Sorun Türkçede dört I harfi vardır ve JavaScript'in locale'den bağımsız toLowerCase() / toUpperCase() davranışı bunları doğru yönetmez: "İSTANBUL".toLowerCase(); // "i̇stanbul" (i + combining dot) "IĞDIR".toLowerCase(); // "iğdir" (Türkçede "ığdır" olmalı) "istanbul".toUpperCase(); // "ISTANBUL" (Türkçede "İSTANBUL" olmalı) Bu yüzden birçok slug kütüphanesi Türkçe başlıkları bozar: - İSTANBUL → istanbul yerine i̇stanbul (i + birleşik nokta) - İstanbul'da Iğdır → istanbulda-igdir yerine i̇stanbul'da-iğdir veya i-stanbul-da-igdir - ŞÜKRÜ ÇAĞLAR → sukru-caglar yerine şükrü-çağlar - Case dönüşümünde IĞDIR → ığdır yerine iğdir Türkçe kuralları açıktır: | Girdi | Doğru çıktı | | --- | --- | | İ | i | | I | ı | | i | İ | | ı | I | Slug üretiminde ise bu harflerin ASCII karşılıkları gerekir. Çözüm: açık harf eşlemesi turkce-slug, I/ı/İ/i için locale'e güvenmek yerine harfleri açıkça eşler. Bu hem tarayıcıda hem Node.js'te aynı sonucu verir. import { slugify } from "turkce-slug"; slugify("İSTANBUL"); // "istanbul" slugify("IĞDIR"); // "igdir" slugify("İstanbul'da Iğdır"); // "istanbulda-igdir" slugify("Çok Güzel Bir Başlık!"); // "cok-guzel-bir-baslik" slugify("ŞÜKRÜ ÇAĞLAR"); // "sukru-caglar" slugify("Fiyat: 100 TL / Ay"); // "fiyat-100-tl-ay" Kurulum: npm install turkce-slug Paket ESM ve CommonJS çıktılarını, ayrıca TypeScript tiplerini birlikte sunar. Runtime bağımlılığı yoktur. Türkçe büyük/küçük harf dönüşümleri Slug dışında Türkçe case dönüşümleri de doğru çalışır: import { toLowercaseTr, toUppercaseTr, capitalizeTr, titleCaseTr, removePunctuation, } from "turkce-slug"; toLowercaseTr("İSTANBUL"); // "istanbul" toLowercaseTr("IĞDIR"); // "ığdır" toLowercaseTr("ISIK"); // "ısık" toUppercaseTr("istanbul"); // "İSTANBUL" toUppercaseTr("ışık"); // "IŞIK" capitalizeTr("iĞDIR"); // "İğdır" titleCaseTr("ışık ırmak"); // "Işık Irmak" removePunctuation("Merhaba, dünya!"); // "Merhaba dünya" Slug seçenekleri slugify("Çok Güzel", { separator: "_" }); // "cok_guzel" slugify("Çok Güzel", { lowercase: false }); // "Cok-Guzel" slugify("Çok güzel bir başlık örneği", { maxLength: 10, }); // "cok-guzel" slugify("Fiyat: 100 TL", { preserveNumbers: false }); // "fiyat-tl" - separator varsayılan "-". - lowercase varsayılan true; kapatınca Ç → C, İ → I gibi kaynak büyüklüğü korunur. - trim baş/son separator temizliğini kontrol eder. - maxLength mümkünse kelime sınırından keser ve sonucu separator ile bitirmez. - preserveNumbers varsayılan true; false iken rakamlar kaldırılır. 616.760 kelimelik doğrulama Kütüphanenin gerçekten kapsamlı çalıştığını göstermek için iki MIT lisanslı kaynaktan oluşturulan 616.760 kelimelik bir doğrulama korpusu kullanıyorum: - FrequencyWords - Turkish Wordlist Her kelime için üç kontrol yapılıyor: 1. Slug yalnızca ASCII karakterler içeriyor mu? 2. Slug idempotent mi? (slugify(slug) === slug) 3. Türkçe büyük/küçük harf round-trip tutarlı mı? Sonuç: 616.760 / 616.760 geçti, 0 hata. Tüm Türkçe özel harfler (ç ğ ı ö ş ü â î û) korpusta yer alıyor. Rapor ve korpus GitHub'da: - Doğrulama raporu - Korpus snapshot Canlı demo Tarayıcıda canlı slugify, Türkçe case dönüşümleri ve kelime doğrulaması: **officalbluehair.github.io Demo, 4.205 kelimelik hızlı test setini çalıştırır. Tam 616.760 kelimelik korpus ve rapor GitHub deposunda bulunur. Kaynaklar - npm: npmjs.com - GitHub: OfficalBlueHair/turkce-slug - Lisans: MIT Türkçe slug veya case dönüşümüyle ilgili bir hata görürsen issue açabilirsin. Geri bildirim ve katkılar memnuniyetle karşılanır.
AIエージェントの購入委任をRubyで扱う — AP2対応SDK v1.0.0を公開しました AIエージェントがユーザーの代わりに商品を購入する場合、販売者はどうやって「このAgentには、その商品を買う権限がある」と確認すればよいのでしょうか。 この課題に取り組む仕組みの一つが、AP2(Agentic Payment Protocol)です。 AP2のCheckout MandateをRubyで扱うOSS、ap2_mandate_verifierのv1.0.0を公開しました。 以前の記事では、v0.1.0を使い、正常・改ざん・期限切れの3ケースで「購入委任を検証する」ところを紹介しました。 今回は、そこから広がった発行 → 提示 → 検証 → Checkout Receipt確認の流れを紹介します。GemはRailsに依存せず、Ruby単体で利用できます。 v0.1.0から何が変わったか できること v0.1.0 v1.0.0 対応するCheckout Mandateの検証 ✅ ✅ 販売者のCheckout JWTの署名・検証 — ✅ Direct方式の発行・提示・検証 — ✅ 条件付き購入委任の発行・提示・検証 — ✅ Checkout Receiptの署名・検証 — ✅ 検証入力の組み立て支援・診断 — ✅ Gem名にはVerifierとありますが、現在は対応するAP2 Checkoutの一連の処理を扱うRuby SDKへと発展しています。 ただし、これらがすべて1.0.0で新規追加されたわけではありません。Direct方式は0.7.0、条件付き購入委任は0.8.0までに追加してきました。1.0.0は、既存機能の公開API・互換性・保守方針を明確にした節目です。 2種類の購入委任に対応 AP2には、ユーザーの承認方法によって大きく2つの流れがあります。 Direct(ユーザーがその場で承認):具体的な注文内容をユーザーが確認・承認する。 条件付き/Autonomous(事前に条件を承認):ユーザーが先に条件を承認し、Agentが条件に合う注文を選ぶ。 例えば後者では、次のような条件を考えます。 商品AまたはBを1個、さらに商品Cを2個購入してほしい。 対応するprofileでは、このような候補商品と数量の条件をOpen Mandateに含められます。 承認サービスが、認証・同意を確認したうえで、登録済みAgent鍵に対するOpen Mandateを発行する。 Agentが販売者署名付きCheckoutを受け取り、条件に合うClosed Mandateを自分の鍵で生成・提示する。 販売者側が、独立して保持する信頼情報・注文・Challenge・時刻などと照合して検証する。 アプリケーションが最終判断し、そのSuccess/Errorを署名付きCheckout Receiptとして返す。 ここで重要なのは、Open Mandateを発行した側の秘密鍵をAgentに渡す必要がないことです。Agentは自分の鍵を使い、許された条件の範囲で購入内容を提示します。 Rubyだけで一連の流れを試す v1.0.0には、条件付き購入委任を最後まで実行する合成データのサンプルを同梱しています。リポジトリのcloneもRailsも不要です。 正式に動作確認した環境はMRI Ruby 3.4.12/4.0.9です。対応環境を用意して、以下を実行します。 gem install ap2_mandate_verifier -v 1.0.0 ruby -e 'spec = Gem::Specification.find_by_name("ap2_mandate_verifier", "= 1.0.0"); load File.join(spec.full_gem_path, "examples/conditional_checkout.rb")' このサンプルでは、商品Bを1個・商品Cを2個という3明細のCheckoutを、先ほどの条件に照らして検証します。最後には、成功・拒否という別々の結果の例でReceiptの署名検証も行います。 出力の主な部分は次のとおりです。 Open -> Agent Closed -> merchant evidence: verified=true; rows=3; requirements=2 Receipt Success: authentic=true; purchase_succeeded=true Receipt Error: authentic=true; purchase_succeeded=false Success=consumed; Error=policy decision; missing Receipt/uncertain commit=pending Synthetic only; evidence and Receipt authenticity are not payment completion. Receipt Errorでもauthentic=trueになり得るのがポイントです。これは「拒否されたという結果が、確かに販売者から返ってきた」ことを確認できたという意味で、購入成功を意味しません。 また、SuccessとErrorは同一取引を二重に実行した結果ではなく、サンプル内で分けて示している合成ケースです。実際の購入・決済は行いません。 実装全体は同梱サンプルのソースで読めます。 Ruby APIの入口 用途に応じてAPIを分けています。 役割 主なAPI 販売者Checkoutの署名・検証 CheckoutJWT.sign / verify Direct方式の発行・提示・検証 Buyer.issue_closed / present_closed / verify 条件付きOpen/Closedの発行・提示 ConditionalCheckout.issue_open / issue_closed / present 業務情報からの完全検証 InputBuilder.verify Checkout Receiptの署名・確認 CheckoutReceipt.sign / verify これらはAP2MandateVerifier配下のAPIです。具体的な引数・鍵の受け渡し・保存状態は公開API契約と条件付き購入委任の例にまとめています。 1.0.0での対応範囲と責任 現時点では、AP2 v0.2のCheckout Mandateに関する対応profileが対象です。Directと条件付き購入委任では、1販売者・1〜8明細の複数商品と数量を扱います。一方、すべてのAP2形式に対応しているわけではありません。 特に、次はGemの責務には含まれません。 本人認証・同意取得、信頼するIssuerや鍵の登録・失効管理 最新の注文状態の確認、nonceの永続的・原子的な消費、重複購入防止 購入の最終許可、注文更新、実際の決済 Payment Mandateの処理、任意の予算条件や全AP2 profileへの対応 Gemが返す検証結果やEvidenceは、アプリケーションが購入を許可するための材料です。検証成功だけで購入・決済を実行してよいわけではありません。署名付きReceiptの真正性確認も、決済完了の証明とは別です。 また、v1.0.0の公開はAP2公式認定や、実Issuerとの相互運用・本番環境での安全性の保証を意味しません。対応条件を明示し、未知・未対応の形式を安易に受け入れない方針です。 おわりに AP2はまだ発展途上の規格です。それでも、Ruby開発者がAIエージェントによる購入委任を実装するときに、署名やMandateの形式を毎回ゼロから実装せずに済むようにしたい。そう考えて、このGemをOSSとして開発しています。 まずはRuby単体のサンプルを試してみてください。「ここが分かりにくい」「このprofileも必要」といったフィードバックを歓迎します。 GitHub — ap2-mandate-verifier RubyGems — ap2_mandate_verifier v1.0.0 v1.0.0 Release Notes AP2公式仕様 前回の記事:v0.1.0の検証デモ 通常の質問・改善提案はGitHub Issuesへ。脆弱性の疑いは非公開の報告窓口を利用し、実際の秘密鍵・Mandate・個人情報は公開しないでください。
#Arazzo's actor-in-the-loop proposal is now a formal PR. Arazzo handles deterministic multi-step workflows but had no way to pause, hand off to a human, and resume. Arazzo-Specification#568
I build apps with AI almost every day, and kept hitting the same three problems, so I built a framework for them I build new projects and finished products for customers almost every day, many of them with AI built in, and I kept hitting three problems: - Too many pieces for every project. Next.js on Vercel plus Supabase is almost free to start and great on its own, but every project meant managing both, and reaching the quality I wanted meant stitching many separate packages together on top. Keeping everything in one place meant a big provider's monthly plan, and with many apps those fees grow fast, since every single one needs its own upkeep. - Teaching the AI what it can control. For every function in my app, I had to build a separate system to explain it to the AI. Tools broke, prompts didn't work well, and MCP servers needed extra work. - The AI reading the whole codebase for small changes, which was very costly. So I built Forjr, an AI-native, full-stack React framework. A few of the features I love most: - One engine runs the whole app (front end, back end, data and AI) on any server you choose. You still need a server, but one is enough for many apps: each additional small to mid-size app takes only about 2 MB of memory, and apps nobody is using can scale to zero and wake up in about 50 ms. - Every function is already an AI tool. The engine explains it to the AI by itself, and with one option it's also an MCP tool, a REST endpoint and a CLI command. - An agent is a single definition: a model, its instructions and the functions it may use. It reaches only what you gave it, and danger: 'ask' makes it wait for a person before using that tool. - Built for the AI that writes your app: forjr map hands it the whole app in one short outline, forjr check catches every error in one pass, each with its fix, and undo for AI rolls back exactly what one agent changed while everyone else's work stays untouched. I open-sourced it: forjr/forjr, with the docs at forjr.dev I'd love to hear how you handle these today. What's the most painful part of giving AI access to your own app?