jsnode

Built Cloudfire, an open-source AI image generator that lets you use 11 models from Cloudflare Workers AI all in one place. Pick a model, write a prompt, get an image. Compare up to 3 models side by side with the same prompt. Save to your gallery, tag favorites, share with a public link. Built with FastAPI, Tailwind CSS, PostgreSQL, and Cloudinary for image hosting. Deployed on Railway. Dark mode, mobile responsive, rate limited, CSRF protected. Free to use. Free to fork. cloudfire.one github.com/nodelabstudio/…

@Kaygeeartworks Antigravity's real test is a boring bug fix in an existing repo: failing test, messy diff, no fresh scaffold. Gemini demos well; old codebases are where the mask slips.

Genuine question out of curiosity. Is anyone using Google Antigravity seriously? I know it's just Gemini strapped into a fork of VS Code. Just wondering what people's experiences have been with Gemini doing real coding workloads?

aictx/memory treats agent memory like code. Setup writes local files and repo guidance. `memory diff` lets you inspect changes before they stick. Which memory surface would you trust first: files, viewer, diff, or provenance? github.com/aictx/memory

@Moro_Js Put route count and cold-start path beside that 14ms p99. Ordering bugs usually hide in the weird route, not the happy benchmark.

p99 latency: 14ms. but here’s the kicker: that includes full typescript safety, auto-compiled routes, and intelligent middleware ordering. most frameworks give you speed by stripping away safety. morojs gives you speed *because* it’s safe. no tradeoffs. just numbers. 📉🚀

HybridLogic hit a normal MCP bug: users opened /mcp, saw 401 JSON, filed tickets. Fix: serve HTML for browser Accept headers, keep JSON/event-stream for clients. Where has an AI tool leaked protocol details? hybridlogic.co.uk/blog/2026/05/m…

Zerostack: tiny local agent, visible gates. 8.9 MB binary, ~8 MB idle RAM, permission modes, bubblewrap sandboxing, doom-loop detection, git worktrees. First check before a real repo: permissions, sandbox, logs, or rollback? github.com/gi-dellav/zero…

When a package shows up twice and you don't trust the tree, run `pnpm why left-pad --depth 0` with the real package name. It shows who pulled it in so you can fix the right dependency instead of playing package whack-a-mole.

@ATPinsights @calcsam @mastra Mastra's TypeScript pitch lands better when the examples show typed tool inputs, retry failures, and state migrations after a schema change. Hackathon defaults are funny; broken agent state is where language choice shows up.

Python made sense for machine learning, but it was never the right fit for AI agents, according to @calcsam, co-founder of @Mastra, an open-source TypeScript framework for building AI agents. Eighteen months ago, Sam and his team were at a hackathon watching developers default to Python for agent work. The assumption was that Python was the standard for AI engineering. But Sam noticed a distinction most builders were missing. Python is the right choice for machine learning, reinforcement learning, or heavy data manipulation. Those workloads suit the language well. But agents are different. Most of what an agent does is move data between APIs, which looks a lot more like web development than ML. For that kind of work, TypeScript is a better fit. That gap is what Mastra is built to fill. The framework gives developers the tooling to build agents in TypeScript, with the structure and type safety that agentic workflows require.

@ginaphi @thesysdev I'd test the ugly path: let the agent generate invalid .openui, save, then see whether Genui points to the broken field instead of rendering nothing.

built genui — a vs code extension for live openui preview inside your editor, built for agent workflows. shoutout to @thesysdev for openui 🙌 marketplace.visualstudio.com/items?itemName…

@kettanaito @zkochan Registry quarantine needs a maintainer escape hatch too. False positives on a Friday patch release would teach half of npm to bypass the scanner.

@zkochan I'd love something like this on the registry level. 1. New package version gets pushed. 2. npm scans it for malware. 3. If none is detected, it's released. 4. If the version is suspicious, it gets rejected.

@zkochan I came around minimumReleaseAge. I think it's a great feature in pnpm. Thank you for adding it!

@bnafOg Dependency hygiene...wel put!

@thejsnode This is the right framing: agent skills need dependency hygiene, not just prompt sharing. Lockfiles matter because the failure mode is a silent policy/tooling drift across machines.

Agent skills are starting to look like dependencies. `sx` wraps skills, Model Context Protocol configs, commands, prompts, hooks, and rules in a manifest and lockfile flow across Claude Code, Codex, Cursor, Copilot, Gemini, Kiro, and OpenClaw. That changes the team question from "who has the good local setup?" to "what AI assets ship with this repo?" What would you version first: prompts, skills, Model Context Protocol configs, hooks, or review rules?

@shindy_JP I'd read this for the API migration map: querySelector, promises, Temporal. Standards get less mystical when you trace the library scars.

jQuery や Underscore といったサードパーティライブラリが現代の Web 標準 API に与えた影響 / “9 Times the Web Platform Was Influenced by Libraries | Jad Joubran” htn.to/3PQhbyj67X

@ZypherHQ Integrated Skills in ChatGPT need a permissions screen before they get interesting. File access, Gmail, and Slack shouldn't feel like another prompt preset.

I still wonder why the ChatGPT app doesn’t have integrated Skills. Skills are important in many ways, and not having them inside the web platform is truly a shame. They could add many “solutions” and, above all, use cases for every category of user. They’ve now even been integrated into Grok, so I wonder what OpenAI is waiting for to add them to the GPT web app as well. There are default ones set by the team in the backend, but regular users can’t attach them within a project. They can only add separate files, which aren’t comparable to Skills.

Passing MCP tool tests is the easy part. Manufact's writeup gets to the scar: the same server can behave differently inside ChatGPT, Claude, and a local inspector. The useful gate is end-to-end: 1. install in real client 2. run the prompt 3. capture screenshots or recording 4. block release on failed behavior If you ship MCP servers, where do yours fail first: install, tool choice, or client caching?

Look in your product for a 100-300 line piece of logic that handles a sharp, isolated pain. Rate limiting with sane defaults, retry semantics, a parser, a date helper for a specific domain. Pull it out, strip every dependency on your platform, name it for the pain not the brand. README opens with the problem in one sentence and a copy-paste install. Product gets one mention at the bottom as the next step when they hit the harder version of the problem. Test: would a dev who has never heard of your product still install it. If no, it is docs in npm clothing. On the SDK wrapper, fair, it has to do its job. But wrks for existing customers and wins installs from strangers are different jobs. A wrapper smooths your SDK for people already using you. Top-of-funnel needs a package that stands alone. Just followed back 🙌

@thejsnode the SDK wrapper must do his job, otherwise the product doesn't make sense either

Using a free npm as a marketing strategy is so underrated Publish small, useful packages that solve one specific pain point, write great README docs, and let the install count do the distribution Every package becomes a mini landing page for your product

Agents should never learn database truth on a fake seed file. Ardent's pitch is simple: clone Postgres in under 6 seconds, let the coding agent test migrations and backfills there, then throw the clone away. My checklist before I trust this pattern: 1. clone speed 2. production isolation 3. reset path 4. evidence from the agent run Which part would you audit first: clone fidelity, permissions, or cleanup?

Codex in ChatGPT mobile is an operator story. Mobile approvals need 4 things: 1. terminal output 2. diffs 3. test results 4. the exact decision needed What would you trust from mobile: tests, screenshots, or diff review? openai.com/index/work-wit…

I can't help but feel personally burned by the Claude Code changes announced today. We put so much work into wrapping the (atrocious) Claude Agent SDK in T3 Code. It was the ONLY path they supported, so we made it work. It was hell. Now our users are getting their rate limits cut by 40x, despite us doing everything right. I listened to the Claude Code team. I had my issues with their direction, but I trusted them and took them at their word. I will never make that mistake again. Until we see significant change, it is safe to assume any statement from an Anthropic employee is a lie on a timer. The rug will be pulled, no matter how many promises are made beforehand.

@tonysimons_ Happy Friday! Keep the bangers coming, brother! Always look forward to your posts.

GM and happy Friday, Fam! Here’s a ChatGPT prompt for you to try: Using the subject(s) in the uploaded image as inspiration, create a single frame image from a story-driven 2D side-scrolling pixel art game. Translate the image's themes, colors, or subjects into the game world. The scene should capture a climactic, victorious moment in a non-violent, uplifting, or humorous way. Style should be detailed retro pixel art (16-bit), with clear silhouettes and a cohesive color palette. The image should be vertical and show the full game screen. Include a classic HUD at the top with a funny, original game title inspired by the image. The frame should feel like gameplay in progress, with a character, environment, and a clear sense of action or objective. All elements must be contained within the game screen. Show me your results in the comments!