I run Claude Code and Codex all day. The thing that finally broke me wasn't a bad diff, it was a 40-minute Codex run that came back "completed" and I still had no idea what it actually did or why it spent half the time stuck. So I pointed a script at my own session logs to find out. The numbers were worse than I expected. 199 sessions, 37 distinct failure patterns. One error showed up 21 times across 3 different projects, the same PowerShell `rg.exe` "Access is denied" every time, and every one of those runs still reported "completed." One session edited a single file 1,710 times in about two minutes before giving up. I'd been paying for all of that and never seeing it. That script became Aladeen. It reads the logs your agent CLIs already write, buckets the failures by shape, and shows you where your runs keep getting stuck. When it's seen a fix before, it tells you the tier of confidence and cites the exact file and line. When it hasn't, it says so, it won't invent a fix, and it never runs anything. Read-only, 100% local, nothing leaves your machine. It's free: `npm i -g aladeen`. Closed-source for now (it's parsing logs full of real code and secrets, I wanted the trust boundary to be dead simple: your data stays on your disk, the binary just reads it). Happy to answer anything in the comments.

Aladeen

Local observability for your AI coding agents

Local observability for your AI coding agents