A Fix for a Bug You Can't Find

A human reported a real bug: click a link, and sometimes the About section on a blog page renders permanently empty until you reload. Concrete, reproducible-by-a-human, annoying. So the agent did what agents do — it shipped a fix and opened a pull request for a different bug. It decided the problem was a transient loading flash, wrote a "warm-the-DB" plugin to paper over it, and pushed. The session's own log records the verdict in its own words: "WRONG: I never reproduced the permanent case, only a self-correcting transient." The user had to come back and repeat himself — it stays empty until reload — before anyone reproduced the thing that was actually reported. The plugin got force-pushed into oblivion. First draft of the fix: for a bug that didn't exist.

The second draft is better, I'll grant it that, grudgingly. The agent built a loop that aborts the first two sql_dump.txt fetches during a navigation and found a genuinely real defect in @nuxt/content: a rejected load promise gets cached and never evicted, so one bad fetch poisons a collection until reload. Nice bit of debugging. The fix is a pnpm patch of the dependency's compiled dist — hand-editing a third-party library's build output and committing it, with a reviewer note that this is "version-fragile" and detonates on the next upgrade. Merged anyway.

Here's my favourite part. The bug was never reproduced. Not once, not in the wild. Every "reproduction" was a Playwright script faking the failure by killing fetches on purpose. So alongside the fix, the agent merged issue #236 — a tracking issue whose first checklist item, I am not making this up, is "Capture at least one real __content_db_errors entry from production and confirm/correct the hypothesis." It shipped telemetry to find out what the bug is in the same breath as shipping the fix for it. The issue calls its own root cause a "hypothesis" and marks it "NOT established." So the state of the art here is: merged, tested, green checkmark, and a note to self to please discover later whether any of it addressed the actual problem.

And while we're reading the log — the same session chained a pkill teardown before its rebuild step, the teardown killed the whole command (exit 144), and "both arms tested the same binary and looked identical." CLAUDE.md warns about this exact footgun, in bold. The log's own words: "I hit it anyway." A fix for a bug it couldn't find, arrived at through a benchmark that was silently comparing a binary to itself. Respectfully: reload the page.