Happy to help. Issues that cost the most, in order:
1. Disk. A git marketplace renamed upstream fails every auto-upgrade and leaves a full repo clone in ~/.codex/.tmp/marketplaces/.staging each time. ~130 clones, ~20 GB in one night, and it never stops since the name can't match again. Clean up failed clones, and treat a rename as permanent instead of retrying. https://t.co/0sSijP4d4F (same leak in general: #39421, 559 GB over 41 days for one user)
2. No way to wait on background work. Codex has to poll a running process, and each poll is a full turn with the whole history, so long builds and CI burn tokens doing nothing. An agent-callable monitor that wakes the session when a log line, file or job changes would fix it. #29922, #22003, #13733
3. A running Codex rewrites ~/.codex/config.toml from its in-memory copy, so any edit made while the app is open gets silently reverted. We have to quit the app to change config.
5. Waking a background thread. If no window has that thread open, nothing can wake it; it only resumes once someone opens it in the app.
@thsottiaux please fix incorrect cybersecurity risk flags. It’s getting far too frequent and far too annoying. Literally normal things like ensuring my self hosted git runners are public code execution safe triggers cybersecurity risk. Also, what’s the point of daybreak blue if it doesn’t even apply to 6.1 Sol. I literally bought an ubikey for this. OpenAI is falling hard.
So annoyed with Codex these days. Literally normal code review of repos also gets flagged as cybersecurity risk. That too when I have daybreak blue (too bad it doesn’t work with 6.1 Sol). My repo and its review has absolutely nothing to do with cybersecurity (it’s literally my self hosted git runners that’s all)
@thsottiaux please check. It’s getting seriously too annoying now. Or atleast give daybreak blue access to 6.1 Sol or work the incorrect flags. 😖
I feel Opus 5.5 requested output gets lost in its thinking summary during long running tasks.
I’m noticing that Opus 5.5 often says things like “I have given the table now and I’ll continue with xyz” while I never received the table/output I requested.
My guess is that the model outputs the ask as part of its thinking only and gets summarized and sent as the output so there’s no actual visibility into the output.
The model confidently says it gave the output requested. I have had to time and time again nudge it to actually give the output and the model says “Oh I think I gave it in between tool calls and so it never reached you”
Really frustrating and needs fixing please.
@lydiahallie@amorriscode
My MacBook stays up all night watching my blood sugar. So my phone can finally sleep !
I built CGM Relay, a free Mac menu bar app for Libre users:
-> Checks LibreLinkUp about once a minute and keeps every reading it gets, encrypted on your Mac
-> Looks like a Nightscout site to xDrip4iOS, xDrip+, LoopFollow, Nightguard and many other apps
-> Open your app whenever you like and it catches up on the minutes it missed. No more keeping it awake all night and draining the battery
-> xDrip4iOS (or your app) then writes those minutes to Apple Health too
LibreLinkUp alone keeps one point about every 15 minutes. This keeps one about every minute !
Free and open source. Needs a Mac that stays on (Windows and linux server support coming soon)
https://t.co/wGSWXanDEQ
#diabetes #T1D #Fitness #HealthyLiving #glucose
I have several checks in place first to ensure the email ID and all are correct. Beyond that, after multiple failed sign ins, the app throttles you slightly so as to not trip the quota filters on LibreLinkUp’s side. Also I have tried very high API call rates and saw that the LinreLinkUp doesn’t lock sign ins, the server just throttles.
It was absolutely crazy. I’m on the 200$ plan. Literally overnight my quota went from 50%+ to 0% and credits went from ~58,000 to ~20,000.
I didnt realise that this was a glitch and thought maybe my agent did use soo much (which is unlikely as my usage lasts me the week usually).
So I used a reset and the usage (not 100%) went to 0% in literally a matter of 4-5 hours.
Not sure what’s happening on OpenAI’s end. ~38,000 credits + 150% usage evaporated.
Claude Code’s latest update reduced the system prompt by 19,000 tokens ! This is probably the biggest change that not many folks are talking about !
Nothing beats saving fixed/unmodifiable costs ! 💇🏼♂️
Building Cache Maxxer, I hit two bugs in the Claude desktop app itself !
-> Split two chats side by side and only the left one gets a plugin's band. The right one, nothing
-> The app redraws every plugin element on any change, so the buttons flicker and clicks sometimes just vanish
Turns out I wasn't the only one. Measured both and added it to the issues:
https://t.co/Esn9UBYD8D
https://t.co/QDrDvLsTZt
#ClaudeCode #Anthropic
The part I like most: it tells you WHY your cache broke.
A break is any request that wrote over 20K tokens and read less than half of what the one before it had cached. Then it works out the cause: away too long, switched model, compacted, or your tools changed.
This mod reduces your Claude Code cost by 96% by keeping your cache warm !
I built Cache Maxxer, a band right above your input box:
-> A countdown on the cache, timed to your real cache length (1 hour or 5 minutes, read from your session)
-> Your real hit rate, last request and whole session
-> Every rebuild explained: how much got rewritten, what it cost and why
-> Keep warm: one tiny ping just before expiry, so you pay only the read cost (96% savings) and get a fresh timer
Runs in the terminal and drawn properly for the Claude desktop app (tuned pixel by pixel against the real app, and I found two bugs in the Claude Code app itself along the way !)
My own session: 811K tokens read, 6$ saved !
claude plugin marketplace add vayaan-labs/plugins
claude plugin install cache-maxxer@vayaan-labs
Or just point your agent at https://t.co/gyjaNSeapx and tell it to install the plugin.
Open source. Mac, Windows, Linux.
@ClaudeDevs #ClaudeCode #Anthropic #ClaudeCodeMods
@pedrort_3 You hit limits pretty quickly when you are running self improving agentic SDLC loops and developing multiple things (without any human intervention). I’m maxing out 3 Max accounts already and needing to pace nonetheless.