Under the hood
A dozen specific decisions about how to play sixteen videos at once on a Mac without setting it on fire. Mostly invisible. Quietly the entire point.
The pipeline runs progressive — playback fires from stage four before stage two has finished. The scanner keeps working in the background, appending to the live grid as it goes. There is no “loading complete” moment because there is no waiting for it.
MemoryMonitor pulls live numbers straight from the Mach kernel —
task_info for resident memory,
thread_info for per-thread CPU. The same data
Activity Monitor uses.
When the room warms up, flexGrid tunes itself across four graduated tiers. Caches get trimmed, off-screen players get recycled, pricey Vision passes get paused. The grid stays smooth. The fan stays quiet. You stop having to think about settings.
Live memory pressure
MemoryMonitor pulls live numbers straight from the Mach kernel — the same data Activity Monitor uses — and trims caches before macOS would.
MemoryMonitor.swift · task_info, thread_info
Naïve looping seeks the player back to frame zero — a tiny black flash, an audio click, every time. AVPlayerLooper keeps a second media replica running across the boundary. The seam isn't fixed; it isn't there.
Naïve looping seeks back to frame zero. The decoder catches up, the player blinks, the audio clicks. You stop noticing it; you never stop seeing it.
AVPlayerLooper keeps a second media replica running across the loop boundary. There's no seek, no rebuffer, no flash. The seam isn't fixed — it isn't there.
Every Vision-dependent surface in the app — filters, Smart Collections, Editorial Layout's hero picker, Aspect Match — checks data coverage before it appears. If the data isn't in yet, the control isn't there. No half-empty filter menus, no controls pretending to work before they actually can.
A small "Analyzing N/M" pill on the status bar reports progress. The features arrive when they're real. It's a small architectural rule with a calming surface effect.
In the status bar, right now
Most of these never come up in conversation with the app. They're the reason the app feels calm. That's not bullet-list material — it's the result of the bullets.
The scanner runs progressively. A cultivated first batch lands in ~1.5 seconds even on a folder of 25,000 files, and the rest streams in behind it. Instant playback isn't gated on scanning being done.
You don't sit through a loading bar. Drop the folder, the grid plays, the library catches up underneath.
Every cell uses AVPlayerLooper, which keeps a second media replica running across the loop boundary. No seek, no buffer, no black flash.
Looping a video most places gives you a tiny stutter you stop noticing. Here, you literally cannot see it happen.
Click a tile and the fullscreen overlay borrows the player that was already running in the cell. The same engine just changes where it's drawing — implemented as an in-window ZStack overlay, not a system panel.
No rebuffer, no poster frame, no spinner. The grid keeps its place. You walk away and back.
MemoryMonitor pulls live numbers straight from the Mach kernel: task_info for resident memory, thread_info for per-thread CPU. The same data Activity Monitor uses.
flexGrid tunes itself across four graduated tiers — Normal, Elevated, High, Critical — trimming caches, recycling off-screen players, and skipping the pricey stuff. You don't have to babysit a settings panel.
A PlayerPool keeps a small set of media players warm and recycles them across cells. When a tile changes content, the least-recently-used off-screen player has its URL swapped in — same engine, new file, no setup cost.
Spinning up a fresh AVPlayer drags a whole decode pipeline behind it. Recycling is why a sixteen-tile grid feels like four.
Every scanned file's metadata persists across launches — aspect ratio, duration, codec, frame rate, Vision tags, saliency rectangles, pixel dimensions. Change detection compares (fileSize, dateModified) and only re-reads what actually changed.
Big folders that took seconds the first time load in milliseconds the next. The library remembers itself.
For images, flexGrid reads EXIF dimensions and orientation straight out of CGImageSource — including all eight EXIF rotation cases — without decoding a single pixel.
Layout math runs before the photos are even drawn. Grid composition is instant and never wrong about which way is up.
Every Vision-dependent surface checks coverage before it appears. Filters, Smart Collections, Editorial templates, and Aspect Match stay hidden until the data they need is in. A small "Analyzing N/M" pill on the status bar reports progress.
The app never shows you a control that wouldn't work. Plus you never get a half-empty filter list pretending to be complete.
Press R and the next batch is already cooked. The engine pre-computes the next three batches of unseen items in the background after every shuffle, and pre-warms the thumbnails too.
The second shuffle is faster than the first. The third is faster than the second. You stop waiting for randomness.
SmartShuffle scores every candidate against a sliding window of what just played — kind, aspect, source folder, date, and duration — and bag-draws every item once before any repeat.
A naïve shuffle clusters similar items by accident. This one feels hand-curated because the math is doing the curating.
Composite detection, brightness classification, monochrome detection, and letterbox detection share a single CGContext render — one pixel buffer, four analyses.
Four separate GPU-to-CPU copies would be the normal way. One copy is the considered way.
FileRelinker resolves moved or renamed media by matching filename and file size, with sibling-folder inference for the cases where the rename took the parent with it.
Your saved collections survive a reorganisation. The wall stays whole.
None of these are flagship features on their own. The flagship is that you don't have to know about any of them — you just notice that the wall plays, the loops are clean, the seams don't show, and your fan stays quiet.
A closing note
The engine notes are here because they should be. The app is here because that's where they stop being notes and start being the experience.