It appears that Safari 17.5 (as well as the current Safari Technology Preview, "Safari 18.0, WebKit 19619.1.18") has a particularly pernicious bug in which img tags with lazy-loading enabled that have a src which must be resolved after an HTTP redirect will stop rendering if you load a lot of them. But only sometimes. And then continuously for, like, 5 minutes.

Suppose I have a bunch of images like this in list:

<img loading="lazy" src="/a/redirect/to/some.webp">

Seems reasonable, right? Well, here's what I'm seeing:

Weirdly, when the bug is encountered:

• Safari won't "wake up" to load the image in response to scrolling or resizing the window
• Nothing is printed to the development console and no errors appear in the Network tab
• The bug will persist after countless page refreshes for at least several minutes (almost as if a time-based cache expiry is at play)
• It also persists after fully quitting and relaunching Safari (suggesting a system-wide cache or process is responsible)

I got tripped up on this initially, because I thought the bug was caused by the fact I was loading WebP files, but the issue went away as soon as I started loading the static file directly, without any redirect. As soon as I realized the bug was actually triggered by many images requiring a redirect—regardless of file type—the solution was easy: stop doing that.

(Probably a good idea, regardless, since it's absurdly wasteful to ask every user to follow hundreds of redirects on every page load.)

So why was I redirecting so many thumbnail images in the first place? Well, Active Storage, which I use for hosting user-uploaded assets, defaults to serving those assets via a Rails-internal route which redirects each asset to whatever storage provider is hosting it. That means if your app uses the default redirect mode instead of ensuring your assets are served by a CDN, you can easily wind up in really stupid situations like this one.

Fortunately for me, I'm only relying on redirection in development (in production, I have Rails generating URLs to an AWS CloudFront distribution), so this bug wouldn't have bitten me For Real. Of course, there was no way I could have known that, so I sat back, relaxed, and enjoyed watching the bug derail my entire morning. Not like I was doing anything.

Being a programmer is fun! At least it's Friday. 🫠

From a new Apple support document explaining what they train from the open web:

Apple applies filters prior to training our models to remove profane and other low-quality website content.

Well, shit.

I use an SSL2+ interface to connect my XLR microphone to my Mac via USB-C when I record Breaking Change. When it launched, the SSL2 and SSL2+ interfaces could monitor your computer audio (as in, pipe it into the headphones you had plugged into it, which you would need if you were recording during a discussion with other people), but there was no way to capture that audio. Capturing computer audio is something you might want to do if you had a Stream Deck or mix board configured to play certain sounds when you hit certain buttons during a stream or other audio production. And up until the day you stumbled upon this blog post, you'd need a software solution like Audio Hijack to accomplish that.

Today, that all changes!

Last year, Solid State Logic released v1.2 firmware update for the SSL2 and SSL2+ to add official loopback support. (At the time of this writing, the current version is 1.3.)

Impressively, the firmware update process couldn't have been easier. Even though the custom UI looks a bit janky, it only required one button press to run and it was all over in about 30 seconds:

Once installed, I had no idea what to do. It appears nobody updated the device's documentation to explain how to actually record looped back audio. I'm brand new to audio engineering (and reluctant to get too deep into it), so it's a small miracle I figured the following out on my own without too much hassle:

• The SSL2/+ ships with 2 physical channels (one for each XLR input), but what's special about the v1.2 firmware update was that it added two virtual channels (3 and 4, representing the left and right channels of whatever audio is sent from the computer to the interface)
• In your DAW (which audio people like me know stands for "Digital Audio Workstation", which in my case is Logic Pro), you can set up two tracks for recording:
  • Input 1 to a mono channel to capture your voice (as you probably have been doing all along)
  • Input 3 & 4 to a new stereo channel to capture your computer audio via loopback
• Both tracks need to be enabled for recording (blinking red in Logic) when you hit the record button.

Here's what that all looks like for me:

Previously, I've been recording the podcast with QuickTime Player, but I'll have to give that up and start recording in Logic (or another DAW) capable of separating multiple channels streaming in from a single input and recording them to separate tracks. Even if you could get all channels recording through a single channel in another app, you probably wouldn't like the result: voices recorded by microphones require significant processing—processing you'd never want to run computer audio through.

Anyway, hopefully this post helps somebody else figure this little secret out. Nothing but love for Rogue Amoeba, but I sure am glad I don't need to add Audio Hijack to my ever-increasing stack of audio tools to get my podcast out the door!

Yesterday, Gruber broke what, in my opinion, is the most important news story regarding Apple Vision Pro since its launch in February. Emphasis mine:

VisionOS 2 is not getting any Apple Intelligence features, despite the fact that the Vision Pro has an M2 chip. One reason is that VisionOS remains a dripping-wet new platform — Apple is still busy building the fundamentals, like rearranging and organizing apps in the Home view. VisionOS 2 isn't even getting features like Math Notes, which, as I mentioned above, isn't even under the Apple Intelligence umbrella. But another reason is that, according to well-informed little birdies, Vision Pro is already making significant use of the M2's Neural Engine to supplement the R1 chip for real-time processing purposes — occlusion and object detection, things like that. With M-series-equipped Macs and iPads, the Neural Engine is basically sitting there, fully available for Apple Intelligence features. With the Vision Pro, it's already being used.

Not being able to run Apple Intelligence would be a devastating blow to any role Vision Pro might serve as Apple's halo car—an expensive gadget most people won't (and shouldn't) buy, but which plays an aspirational role in the lineup and demonstrates their technology and design prowess.

Now, couple this with rumors that work on Vision Pro 2 has been suspended, and it starts to look like we won't see any Apple Intelligence features on the visionOS platform until late 2026 at the earliest. How dated and limited will Apple Vision Pro seem in late 2026 if most new features coming to Apple's other platforms—including, one imagines, updated Watch, Apple TV, and HomePod hardware—don't find their way to Vision Pro, putting its user experience further and further behind?

At launch, I heard a lot of people jokingly refer to Vision Pro as, "an iPad strapped to your face." Recently, as it's become clear most people are using it to watch TV and for Mac screen sharing, Marco Arment said it was more like a mere Apple TV strapped to your face. But if the hardware really can't support Apple Intelligence and isn't going to be updated for several years, how long before Vision Pro feels like an original HomePod strapped to your face?

Outed

Learning Tailwind was the first time I've felt like I had a chance in hell of expressing myself visually with precision and maintainability on the web. I'm a fan. If your life currently involves CSS and you haven't tried Tailwind, you should spend a few days with it and get over the initial learning curve.

One thing I like about Tailwind is that it's so extensible. It ships with utility classes for dealing with screen sizes to support responsive designs, but out-of-the-box it doesn't include any vertical screen sizes. This isn't surprising as they're usually not necessary.

But, have you ever tried to make your webapp work when a phone is held sideways? When you might literally only have 330 pixels of height after accounting for the browser's toolbar? If you have, you'll appreciate why you might want to make your design respond differently to extremely short screen heights.

Figuring out how to add this in Tailwind took less time than writing the above paragraph. Here's all I added to my tailwind.config.js:

module.exports = {
  theme: {
    extend: {
      screens: {
        short: { raw: '(max-height: 400px)' },
        tall: { raw: '(min-height: 401px and max-height: 600px)' },
        grande: { raw: '(min-height: 601px and max-height: 800px)' },
        venti: { raw: '(min-height: 801px)' }
      }
    }
  }
}

And now I can accomplish my goal of hiding elements on very short screens or otherwise compressing the UI. Here's me hiding a logo:

<img class="w-6 short:hidden" src="/logo.png">

Well, almost…

Unfortunately, because of this open issue, defining any screens with raw will inadvertently break variants like max-sm:, which is bad. So in the meantime, a workaround would be to define those yourself. Here's what that would look like:

const defaultTheme = require('tailwindcss/defaultTheme')

module.exports = {
  theme: {
    extend: {
      screens: {
        short: { raw: '(max-height: 400px)' },
        tall: { raw: '(min-height: 401px and max-height: 600px)' },
        grande: { raw: '(min-height: 601px and max-height: 800px)' },
        venti: { raw: '(min-height: 801px)' },

        // Manually generate max-<size> classes due to this bug https://github.com/tailwindlabs/tailwindcss/issues/13022
        'max-sm': { raw: `not all and (min-width: ${defaultTheme.screens.sm})` },
        'max-md': { raw: `not all and (min-width: ${defaultTheme.screens.md})` },
        'max-lg': { raw: `not all and (min-width: ${defaultTheme.screens.lg})` },
        'max-xl': { raw: `not all and (min-width: ${defaultTheme.screens.xl})` },
        'max-2xl': { raw: `not all and (min-width: ${defaultTheme.screens['2xl']})` },
      }
    }
  }
}

Okay, yeah, so this was less of a slam dunk than I initially suggested, but I'm still pretty happy with it!

Last week, when David declared that system tests have failed, my main reaction was: "well, yeah." UI tests are brittle, and if you write more than a handful, the cost to maintain them can quickly eclipse any value they bring in terms of confidence your app is working.

But then I had a second reaction, "come to think of it, I wrote a smoke test of a complex UI that relies heavily on Turbo and it seems to fail all the damn time." Turbo's whole cloth replacement of large sections of the DOM seemed to be causing numerous timing issues in my system tests, wherein elements would frequently become stale as soon as Capybara (under Selenium) could find them.

Finally, I had a third reaction, "I've been sick of Selenium's bullshit for over 14 years. I wonder if I can dump it without rewriting my tests?" So I went looking for a Capybara adapter for the seemingly much-more-solid Playwright.

And—as you might be able to guess by the fact I bothered to write a blog post—I found one such adapter! And it works! And things are better now!

So here's my full guide on how to swap Selenium for Playwright in your Rails system tests:

Turns out, there's more to it…