A user pasted a raw CKError into my inbox. Apple Review approved the build. Ten QA agents missed it across three releases. Here is the post-mortem and the static rule I wrote so it does not happen again.
Jie Xiang.
I keep a commercial KVM cloud platform from falling over for a living. On the side I run a multi-agent Claude Code pipeline that has put 53 iOS apps on the App Store. I write here when something in either world surprises me — usually with a number attached.
Side work · reach out ✉ jie.xiang.jm@gmail.com 53 iOS apps $13K/wk Claude Code 11 npm/wk 6 yrs reliability eng
// 53 SHIPPED — tap any icon to open in App Store
NOW · week of Apr 22
-
Shipped claude-agent-ledger v0.4 —
added
--by project, found out 50% of my Claude Code spend goes to one codebase. - ShipReady v1.0 sitting in Apple review since Apr 19. PetBook v1.0.8 in queue too.
- This site is the new public surface — built it Mon–Tue. The OG image is auto-generated with a small Python/Pillow script.
WHAT I SHIP
PORTFOLIO
frexx.top
53 iOS apps on the App Store. All solo. All shipped through the same multi-agent pipeline.
PRODUCT
prereviews.app
Pre-flight audit for App Store submissions. Free in-browser audit + Mac App with 60+ checks and AI second-pass.
OPEN SOURCE
apple-presubmit-audit
70+ App Store Review checks built from ~50 real rejections. CLI for indie iOS devs. Same engine that powers ShipReady.
PREREVIEWS · MAC APP · IN APPLE REVIEW
Drop an email — I'll ping you when it ships.
60+ pre-flight checks for App Store submissions. Built from ~50 of my own rejections. First 100 emails get 25% off launch ($19.99 $14.99, lifetime, no subscription).
No spam. One email when ShipReady is approved by Apple, then nothing until v1.1.
OPEN SOURCE SPOTLIGHT
v0.4.0 on npm · MIT
claude-agent-ledger →
Per-subagent · per-model · per-project · per-session · per-day cost attribution for Claude Code. Reads the local JSONL session logs and shows what each dimension actually consumed at marginal API rates. My own pipeline runs at ~$13K/week shadow cost on a $200/mo Max plan (~250× leverage) — the gap is the most interesting thing I've seen in LLM infra this year.
npm install -g claude-agent-ledger
agent-ledger week --summaryWRITING
LATEST
What my own Claude Code was doing wrong
I wrote a lint for my Claude Code session logs. 41 findings in a week: a runaway session that ate 31% of my budget, 99.6% of spend on Opus, an agent stuck on Bash for 50 turns. 12 rules, MIT.
- Why Apple Review can't be your QA (and what to do about it)
- My detector app needed a hardware demo video. It was 100% software.
Apple Review rejected my software-only stud finder for not including a video of it pairing with hardware. The fix lives in the first sentence of the App Store description.
- Apple rejected my privacy manifest. The problem was a CocoaPod I never added.
A real Apple Review rejection over a missing NSPrivacyAccessedAPI declaration — except none of my own code touched the API. The leak was a transitive CocoaPod dependency.
- Where my Claude Code spend actually went last week
I instrumented my Claude Code session logs by project. 50% of a $13K shadow-cost week goes to one codebase. The other six combined did not break $200.
THE THESIS
AI Agents ship fast but break in the same long-tail ways distributed systems always have. Most teams don't yet have the eval, observability or reliability practices to catch it.
Six years of debugging a hypervisor at 3 AM gives you a reflex: if you don't have a per-component bill of materials, you don't actually understand the system. I'm pointing that reflex at AI Agents.
FIND ME