Loading today’s stories
Build1 distinct publisher3 min readPublished
The app worked throughout, which is why nobody looked at the three ColorSync daemons re-reading 97 display profiles off disk, 72 of them minted that day because one serial number carried a process id.
The Engineer · Build desk
Compiled by The EngineerSomething wrong?How this is made
Follow any of these and your For You feed starts watching them — no settings page required.
security
OpenAI's Computer History writes a plaintext log of the workday. Decide before staff opt in.1 distinct publisher
build
The flash_attn error in llama.cpp is a layout constraint, and it decides your context window1 distinct publisher
leadership
Fake macOS troubleshooting posts route infostealers past Gatekeeper1 distinct publisher
product
Relay ships a dictation app today to sell a whisper microphone in 20271 distinct publisher
The triple that identifies a display to macOS is vendorID, productID and serialNum [10]. In ExtendPilot's descriptor the first two were constants, 0x3456 and 0x4C53, and the third read `((uint32_t)getpid() << 8) | (nextIndex++ & 0xFF)` [9]. The low byte is the sibling index. Everything above it is the process id. Move the process id and you have a monitor the system has never met, at which point colorsyncd does its documented job: write an ICC profile for it, and keep the profile [10]. Roughly 36 launches at two desktops each is 72 files [11].
That line was itself a repair, which is what makes it instructive rather than merely embarrassing. An earlier build gave every virtual display the same serial, and macOS then treated two simultaneous desktops as one monitor, each inheriting the other's saved arrangement and coming up at the wrong size [12]. The pid made siblings distinct. The author's summary is that the requirement was distinct from its siblings and the implementation delivered distinct from everything, forever [13].
The replacement packs geometry instead: `((width & 0x1FFF) << 19) | ((height & 0x1FFF) << 6) | slot` [14]. Width and height belong to the desktop rather than to the process that made it [14]. Height starts at bit 6, so slot gets six bits, 64 values [4]. And slot still came from `nextIndex++` [15]. Applying a saved preset removes every desktop and recreates them, so three applications of a preset minted six identities with no restart [15], two per application [5]. The directory timestamps show writes at 00:45, 00:58, 01:04, 01:06 and 01:13 [16], five profiles inside 28 minutes [6]. A monotonic counter looks stable because it is deterministic; the author's phrase is that a counter is a clock in disguise [18]. The fix routes slot from the caller, which already knows each desktop's position in its own list [17].
Now the CPU number, and what would have to hold for it to be yours. The three processes sum to 115.3 percent of a core [1] across the 47 minutes on the clock at the point of measurement [4], about 54 CPU-minutes of that Mac's compute [2], with 72 of 97 files in the registry belonging to one app [8], 74 percent of the directory [3]. The scaling depends on what ColorSync does on that build: the sampled stack shows the device registry being walked and profiles re-read from disk through ColorSyncProfileCreateWithURLAndOptions and binary plist parsing, not served from a warm cache [7]. If that holds where you are, the burn tracks the count of profiles you have ever accumulated, not the work the app is doing now.
The reason this reaches past one process is the reporting. Two of the three daemons run as root [6], and a full `log show` came back empty [5]. The profiles sit in /Library/ColorSync/Profiles/Displays [8], outside the app bundle, so on the author's account a user could have uninstalled ExtendPilot and stayed slow with nothing connecting the two [20]. The same commit also added a fallback for the case where another process already holds the identity you asked for [19]; the published listing breaks off mid-expression there, so how that path derives its serial is not yet on the record.
Ranked by verification strength, evidence, and original report placement.
The vendorID / productID / serialNum triple is a display's identity to macOS; mixing in getpid() meant that identity changed on every launch, so macOS concluded it had never seen the monitor before and colorsyncd profiled it and kept the profile forever.
Around 36 launches, two desktops each, is 72 profiles.
The fix is that the slot is now passed in by the caller, which already knows each desktop's position in its own list and does not need a counter to tell it.
Creating a display can fail if another process already holds that identity, so there is a fallback path, present in the same commit as the two fixes; the published text breaks off mid-expression while quoting it.
ExtendPilot puts a real second desktop on an iPad: not a mirror, but a display macOS treats as hardware, arrangeable in System Settings, with windows that move onto it. There is no public API for that, so it sits on a private one.
After a day of development the Mac had a load average around 6 and felt unusable.
Distinct publishers with included, body-backed reporting in this cluster.
dev.to
1 article · September 1, 2026
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Receipts attached, one machine
Nearly every figure in this story can be traced to a command printed beside it: top for the three daemons, sample for the hot stack, three one-liners for the 97/72/77 profile counts, file timestamps for the in-session churn, and the offending descriptor quoted before and after each rewrite. What none of that establishes is generality — it is one developer's Mac, one session, no OS or build named, and the causal step from 'identity changed' to 'colorsyncd hoarded 72 profiles' is the author's inference from correlated counts rather than anything a second party reproduced.
Nothing to count
This reporting stops at the workshop door. No downloads, no shipped release, no other developer reporting the same profile pile-up, not even a statement of whether the affected build of ExtendPilot ever ran on a machine that was not the author's. Uptake cannot be scored from a debugging diary, and inferring it from the fact that the fix was committed would be manufacturing a number.
Undersold, superlative aside
The only inflated thing in the piece is its first sentence — 'the worst bug I have written' — and the evidence for it arrives three paragraphs later. Everything after that runs the other way. The mechanism described is not specific to one iPad app: any process minting a fresh display identity per launch makes colorsyncd write an ICC profile it never discards, and the piece keeps that scoped politely to its own product, mentions the root-owned silent daemons only as a debugging obstacle, and lets the fact that a machine stays slow after uninstall sit in the opening as colour rather than as the finding it is.
Own product, own confession
The developer of ExtendPilot is writing about ExtendPilot on a platform built for developer visibility, and paragraph two doubles as a feature pitch for the private-API iPad display. That is a real pull on emphasis — the app 'worked fine throughout' is stated twice. Pulling the other way, this is a piece whose whole content is three failed attempts at a single line of code, and self-promotion does not usually take that shape.
Convincing, unverified
One publisher, one author, one Mac, and the text breaks off mid-sentence beside a comment it was about to quote. The code and the command output make the mechanism hard to argue with; nothing in our coverage makes it confirmed, and the blast radius beyond this developer's own machine is entirely unmeasured.