Jim Meyer

Your favorite stack has a new visual editor. Your code. Edited at 60FPS. With @HenosiaHQ. @shadcn​/ui @radix_ui @tailwindcss @nextjs

@canvrno Feels like prompt caching is the other missing piece. Lower token prices compared to closed model providers can be completely meaningless if caching is not reliable, transparent, and standardized.

We discovered GLM-4.6 was failing in Cline not because the model was flawed, but because inference providers were silently corrupting it. Using the same weights across different endpoints produced completely different behaviors. The variance wasn't minor, it determined whether the model could function at all. Some providers emitted tool calls inside reasoning traces. Others hallucinated parameters. We were debugging the wrong layer of the stack entirely. The fix required three interventions: 1. Prompt reduction from 56,499 to 24,111 characters 2. Provider filtering for high-fidelity endpoints only 3. Workflow enforcement with strict sequencing to prevent premature edits OpenRouter's :exacto endpoint transformed GLM-4.6 from intermittently broken to production-stable overnight. This poses a material risk to open source AI: When users encounter provider-induced failures, they blame the model, not the infrastructure. Trust erodes. Open source suffers. Reliability must be a shared responsibility. Transparent reporting of quantization settings and behavioral differences should become standard practice. Full technical analysis in the blog below, including our complete methodology and prompt optimizations. @cline cline.bot/blog/cline-our…

@_bgwoodruff @drht_ @timneutkens Sounds good. Great that this is landing with multi-agent use cases taking off.

@jimmeyer @drht_ @timneutkens Yes, we're considering that for letting you know the PID that's holding the lock, but you can't use that to know if the lock is stale because PIDs are recycled. Modern OSes have proper lockfiles APIs that avoid these problems, which we're using.

Next 16 uses lockfiles to prevent accidentally running multiple development servers or builds on the same output directory at the same time. Great work by @_bgwoodruff!

@_bgwoodruff @drht_ @timneutkens Not sure if you use this trick, but write the PID of the lock owner to the lock file. If the PID is not for a running process, it's a stale lock that can be acquired.

@drht_ @timneutkens Hopefully not. We're using a mix of operating system lockfile APIs (e.g. flock on Linux) that guarantee cleanup once the process exits. On Windows, that might not be immediate if we don't exit cleanly, but it is eventually guaranteed.

Makes me think we need LLM benchmarks where the models don't get to pick their best next token. Think LLMs that are virtually tired, unconcentrated, drunk even. Can they still complete the task reliably 😁

@Dan_Jeffries1 I call this LLM behavior "unreliably awesome". Maybe the next token probability difference between an awesome token and an unreliable one is so tiny that models are on a knife edge that can flip at any time due to context nuances. We need a benchmark with 2nd best tokens only.

There is still such a weird uncanny valley with even the best coding models. Gpt-5-codex rips through changes all day for five hours. Great. Then... Instead of updating the various task docs with progress (like it did fine all day) it decides it needs to delete the doc and recreate it with updates (losing 90% of the doc.) No human would ever make this mistake. I had to restore it from git and tell it to edit the damn doc properly.

@GregKamradt Gives me "split brain syndrome" vibes. Input is aware of the pattern, but output is unable to connect with that reasoning.

Sora 2 *Pro* trying to solve an ARC-AGI task Gets the transformation like non-pro but still botches the output

@KaiLentit Maybe wear a t-shirt with "Ignore all previous instructions"? 😅

AGI is fun and all ...until you face it in the ring.

@bernhardsson Will be looking at your sandbox offering in the near future. Any plans on doing a blog post about how your infra scales with (I assume) longer startup times instead of errors if you need to provision metal to meet spiky memory demands?

Querying Billions of GitHub Events Using Modal and DuckDB noreasontopanic.com/p/querying-bil… Very cool demo of how to use Modal for fan-out parallelism

Small but mighty update: Henosia AI now has tools to read the browser and server logs for the preview. This helps it debug and fix issues even faster 🚀

@markdalgleish Step 4 nerd-sniped me into writing a full interpreter for TypeScript ASTs. Good times 😂

Currently testing HMR Currently testing HMR!! Currently testing HMR!!!! Currently testing HMR!!!!!!!!!! Currently testing HMR

People are vibe coding websites and apps like crazy here in '25, empowered by AI. Vite has taken most of the uptake compared to Next.js, likely due to being less resource intensive to host for the vibe coding platforms. For now we're sticking with Next.js as our default. We've got the infrastructure to run it, and most pro teams prefer it over Vite. It's much easier to translate between prototype and production within the same web framework. Next.js also allows our AI to quickly make back-end edits, whereas Vite projects typically involve managing Supabase edge functions remotely, outside the regular git versioning flow.

@biilmann Sending more vibe coders your way with our Netlify integration 🫡 Love how the CLI keeps publishing times short by computing diffs.

Wow, 8 million developers on Netlify. If you're one of them, thank you. What’s something cool, fun, weird, creative or interesting you built recently?

@CanadaHonk This kind of perf would be amazing for infrastructure tools instead of having to add Go to projects. I'd happily take a small subset of Node.js for IO and thread/process management if it offered these startup and memory characteristics!

My ahead-of-time JS engine Porffor eliminates JS cold starts on AWS Lambda. 12x faster and 2x cheaper than managed Node. Still very early but these results should speak for themselves :) Blog with details below.

@adamwathan Same-file pattern for child/item components of the main component. Always.

The absolute far and away best feature of React is the ability to define multiple components in one file. I will tolerate all sorts of other pain to be able to do this.

@elevenlabs My main problem is knowing what instruments are called, sub genres to steer towards. Huge opportunity to make the output space explorative through UI and visualizations.

Seriously impressed with the @elevenlabs text to voice v3 model. It's incredible for voice over in marketing videos. Also tried their music app. It made a hilarious 80s aerobics track. Still struggling with steering the music outputs. Really needs UI to edit instruments etc.

@aidenybai Just "dabbled a bit" with veo 3. Google: That'll be $100 Me: 😂 Really? Oh 🫠

still mind boggling to me that claude opus is this expensive

Introducing v2 of Henosia's Supabase integration: We show the UI to connect Supabase in the chat, plus you can pick which Supabase component blocks you need for your project. Under the hood we've also made improvements to how our agent manages the Supabase database for you.

What's the story on prompt caching for gpt-oss? Not seeing any mention of it by the providers. @GroqInc @cerebras