Reality check: Claude Code's own bug tracker

Today’s post is some support for those who, amid the news from the AI world, have started actively digging their own grave. In demos we’re shown how easily you can hand Claude Code a bug description, and it will find the root cause by itself and prepare a PR with the fix. Every task Claude Code does either test-first or with full test coverage. And, of course, it keeps regressions to a minimum by running all kinds of guardrails. And obviously the agent SDK is wired into GitHub to triage new bugs on intake, filter out the junk and merge duplicates.

A few facts from the real world about the GitHub repository of Claude Code itself:

1. As of today there are 3,789 open defects (issues with the bug label).
2. Minimal triage is done only for defects that are four or more days old (extra labels are added).
3. Spam and junk aren’t filtered out automatically on intake, and some of the filed “bugs” are pretty funny.
4. The number of regressions is off the charts. I run into them myself with almost every update. And anything can break, at random.
5. On many defects people make an effort and describe the cases in detail, and in return there’s silence. Only the good old github-actions bot occasionally marks duplicates or adds a stale label.

But this is purely for support, don’t get distracted from your agent consoles. ;)