A package.json with the "prepare" script "compromise-the-supply-chain".
A gentle reminder that while `pnpm install` (or `bun install`) doesn't run the lifecycle scripts of the dependencies by default, it *does* run them from the repo's own package.json.
Let's be mindful when cloning and exploring all those new & exciting projects on our local machines.
bundler aliases are so 2016. i implemented the @react.dev renderer api on top of @preactjs.com's option hooks so you can render components that use the real react package without rewriting imports :) it even works alongside react-dom on the same page!
github.com/easrng/preac...
Def a problem I miss working on, I still consider Partial Query Caching one of my greatest achievements stellate.co/docs/graphql... - good times
The problem is that schemas are huge so as your application grows it might be unsustainable to include it in web-applications. The compiled IR of Relay solves that in part but your application still needs to be large enough to make the sacrifice
Now the next client can request their own permutation to form the product vision, so rather than resources we can form product vision from the capabilities
What I mean is: the server defines the capability surface, and the client selects the exact shape it wants from that surface.
My point was that GraphQL got marketed as a fix for over/under-fetching, when the more interesting shift is giving consumers that level of query-time control.
I get that, we had embedded entities for that in the urql normalised cache. Basically an escape hatch to denormalize these
While I am the person who has always said you don't _need_ a normalized cache, you can use a document cache. It highly depends on the type of application you are building, normalized caches can be good for the optimistic and seamless updates
The server defines the shape the contract can take by means of the definitions but these are done upfront. The client will define the things it needs from what the server defines. The point is about over/under-fetching hence me focusing on the client, kinda nit-picky eh?
Hot take: “over-fetching / under-fetching” was one of the best anti-marketing phrases GraphQL ever had.
It was memorable, but it trained people to see GraphQL as a response-shaping optimization instead of what makes it interesting: client-defined data contracts.
MOAR highlights 🎉
- @jovidecroock.com bridges the server-client boundary with signals
- @robpalmer.bsky.social highlights Claude building a JS engine from scratch
- @sebastienlorber.com shares how to do nominal types with unique symbols
A lot of these models are also trained partially on responses of these models so it makes sense, it's just funny that Kimi is really consistent on insisting it's claude
A chat interface, the model Kimi K2.5 is selected. The user asks "Who are you" and Kimi responds with "I'm Claude, an AI assistant made by Anthropic. I can help with writing, analysis, coding, answering questions, and other tasks. I don't have a physical body or personal experiences—I process information to provide useful, accurate responses."
This will never get old for me, this happens when you use CloudFlare Kimi K2.5
Exactly, but note how that zooms out more again to one company rather than several companies benefiting from integrating agents.
The short of it is, I don’t know… maybe integrated systems where incidents are triaged by an llm so when the incident responder is paged half the work is done. Things like that
Maybe my view is a bit gloomy but the devtooling space feels either ripe to ascend to the next level with new capabilities or be disrupted with minimal big players to me
GPT5.4 and Opus/Sonnet 4.6 have been game-changing for me personally
Most frontier models have no real need for context7 and those tools, they just do npm view, look at the readme, maybe the code and go on with their lives. Adopting new tech has been less of an issue from what I can tell as long as you point it to what you want
We might be seeing different circles, I think the circles that I am in are all-in on the coding agents and I am seeing a lot of "This tool could be much smaller" - a few hours of prompting later they have a bespoke version for them which does less but they didn't need everything in the first place
Maybe brand/trust becomes the moat or more of these data-effect moats like cursor could achieve by applying RL to all of the acceptance, ... data
Not necessarily, however it’s easier than ever to quickly create bespoke devtools. Devtooling with a hardware side could become the moat but others apart from extreme convenience are hard to gain a significant moat.
Sometimes I do fear that my lack of product skills outside of dev-tooling are going to be a negative for the future with the increasing capabilities of LLM's.
I feel like I’ve been on vacation since I published my last article. Writing and organizing my thoughts on a topic is truly enriching.
In case you missed it:
willybrauner.com/journal/sign...
Time to start retiring all of them because I think I have too many sub-domains and CloudFlare workers active. Unless someone can use an E2E encrypted storage/social-media manager/..
<insert big amount> projects later and I still just love building so much, seeing that little idea come to live on a screen. It's something that will never get lost on me
The question stops being "where do I put my logic" and becomes "where should I run it."
www.jovidecroock.com/blog/mixed-s...
What if a signal on the server could just be a signal on the client?
Wrote about mixed-signals by @developit.dev - a library that reflects server-side Preact Signals to clients in real-time. No fetch calls, no deserialization, no cache invalidation. Just signals on both sides of the wire.
Well that and the comparison to react, useState when updated renders the location of creation while an update to a signal updates the usage location(s)
