ugh Codex background server keeps running out of file handles. Solution:
ulimit -S -n 8192
codex app-server daemon restart
codex
This will disconnect active codex sessions from the Agent Command Center. Or launch without it:
ulimit -S -n 8192
codex --no-daemon
I’m using github.com/AprilNEA/OpenLogi to manage my MX Master. Logi Options+ has no business using 120 threads.
If you try it, you’ll need to run
launchctl kickstart -k gui/$(id -u)/org.openlogi.agent.service
after granting permissions due to issue github.com/AprilNEA/OpenLogi…
Don’t burn your quota in ultra fast mode. Do this instead (.codex/config.toml):
[agents]
max_concurrent_threads_per_session = 8
Then tell the agent to parallelize independent work aggressively. Go higher checking memory pressure in Activity Monitor. ⚠️ Too high may cause 429s.
Astra uses roughly ~2× the Codex allowance of Sol.
But Astra medium matches Sol max on DeepSWE v1.1 (72.8% vs 72.7%). Given that allowance usage also depends on reasoning level, Astra medium might be the best default.
DeepSWE v1.1 benchsift.nxtaigen.com/bench…
Chance of a GPT-6 launch + @thsottiaux reset tomorrow: ~75%
Likely launch time: 10:00 PDT / 19:00 CEST
People are wrong to equate DLSS 5 with indiscriminate AI slop. DLSS 5 can be applied selectively. RTX Neural Shaders let studios train networks for specific rendering tasks like materials and lighting, while RTX Neural Faces handles facial rendering. The future looks bright.
One more day I’m challenged by Codex to find 100GB of build products on my disk. Will it be ~/.codex, /tmp, a project child folder?
Alejandro Ramirez retweeted
We’re deeply honored that Federico Viticci @viticci, Editor-in-Chief of the renowned tech publication MacStories, has published an in-depth review of Open Minis.
Thank you for the thoughtful recommendation—it will help Minis reach more people and make the productivity benefits of AI more accessible to everyone. Truly grateful.
macstories.net/reviews/open-…
I’m reaching new levels of sophistication with prompts like "do it" "fix", "why" and such. Already thinking on buying dedicated UI.
I steered Codex with blog.jacobstechtavern.com/p/… it took ~3 hours in my hot potato macos 27
menubar app to kill appstoreagent on macOS 27 github.com/janodev/Watchdog
It uses 60% of one M2 Max core querying aggressively the wrong SQLite schema, and resurrects via XPC every 6 mins when disabled/killed. #macos27 #appstoreagent
macOS 27.0 beta bug: appstoreagent/ArcadeManager can enter an aggressive retry loop and burn >126% CPU.
Evidence: semaphore_timedwait, SQLite schema mismatch, BGTask Code=8. xitter.cf/puzige_ai/status/20697…
Here is a watchdog to kill it if CPU >40% for 60s:
github.com/janodev/appstorea…
遇到了macOS 27 的bug
/System/Library/PrivateFrameworks/AppStoreDaemon.framework/Support/appstoreagent
CPU 高占用
@Apple
# macOS UI tests without authorization prompts
sudo /usr/bin/automationmodetool enable-automationmode-without-authentication
How to thoroughly UI test your app:
- Tell AI to write down the features in each screen.
- Then tell it to test complex app states that require multiple user actions to reach.
Every day someone uses AI to create a 3rd party product you don’t need.
Here is how I localized this app in seven languages:
1) Tell AI to localize Localizable.xcstrings
2) Tell AI to run UI tests for all screens and use Apple’s ScreenCaptureKit to save shots to a folder.
If in trouble, tell it to figure it the f*ck out.
apps.apple.com/es/app/abdiff…
I spent hours writing something, and someone assumed it was AI because, you know, who bothers these days?
ABDiff is a diff/merge tool available on the App Store. Since someone couldn’t sign in to their Apple account at work, I posted a free notarized copy here: we.tl/t-iG94zqNiBWyNQmv2