Earlier this month our Macs started grinding to a halt: an M3 Ultra Mac Studio with 28 cores and 256 GB of memory, an M4 Mac mini, and an M3 MacBook Pro. They ran a mix of the macOS 27 release candidate and developer betas, all kept current.
Two symptoms:
1. CPU usage went through the roof, even with little or nothing running.
2. Opening Terminal or SupaCode sent the Mac into a spin until it stopped responding.
3. Restarts took forever. On the Studio, the logs showed about 5 minutes before the login screen even appeared. On another of our Macs, getting from the password to the desktop took 30β45 minutes.
Initially, we blamed a recent update or some new software. So we wiped one of our Macs clean and reinstalled everything (yeah, fun, I know). Yesterday it happened with that computer again.
This time we had Claude and Codex dig through the system logs and crash reports. What they found: macOSβs disk-write process got so backed up that 80 threads froze behind it for over a minute. A system watchdog then force-restarted the Mac. The next boot spent about 5 minutes recovering before the login screen appeared and several more minutes before we got to the desktop.
We think the culprit was Spotlight. It was indexing about 45 GB of code across about 50 git worktrees, nonstop, at nearly two full cores. We excluded our code folder from Spotlight, and the problem went away immediately.
If you run a lot of git worktrees or AI coding agents on a Mac, check Activity Monitor for mds_stores, or just exclude your developer folder from Spotlight like we did (System Settings β Spotlight β Search Privacy).
@ClaudeDevs@OpenAIDevs@Apple
@amorriscode Not to be contrary, but thereβs no such thing as an agent without a persona. Thereβs only the one you chose and the one you got by default. And the default is roulette: same table, new spin every session, and the house never tells you what it landed on.
@ClaudeDevs Thank you. Opie has helped us identify several items the playbook addresses. We're working on solutions and retooling now. Appreciate all of these guides!
It went well, Lydia. We spent the past two days doing a model ablation and skill eval refresh and have shed so many lines of code. Early tests in our agentic writing system strongly suggest Opus 5.5 at high writes appreciably better reasoning and advocacy than Fable at 5.1. Looking forward to using it on the daily.
Nice work! We did our own tests and found Opus 5.5 High bested our previous drafts under Fable 5.1 Max and other tests drafts. We will likely use high or xhigh depending on the complexities of the case work, but Max seems too token cost prohibitive. We're showing well over 5x as many tokens.
@lydiahallie Nice! Finally! Greatly appreciated. Now if you all could allow your desktop app to allow you to work with multiple accounts without losing continuity. Its something we find helpful with others and allows us to pick up without issue.
So, @gdb, this is an interesting development. I've practice law for more than 30 years. I've heard the footsteps of the frontier models for a few years and now, you're at the doorstep. About 16 months ago we set out to build our own agentic legal writing system using yours and other models to help us with substantive trial and appellate court filings. We've tapped into CourtListener and the Free Law Project. Your news is exciting, but I do wonder is this available to businesses like ours who use the business plans? Or is this something that will be out of our reach and left to the mega law firm using their Harveys and Legoras.
Anthropic is cutting our usage tomorrow from 50% boost to 25%. So, do you think they will drop this annoying 5 hour limit? It's an absolute buzzkill and destroys work in your workflows.
@thsottiaux Thank you! Truly thank you! Your good will with this when you all are having issues or retooling is unheard of in this space. So, props to you all. Greatly appreciated