The most common scenario for me was: I implement backup, done by a daily cron calling pg_dump. It worked and I moved on, until one day I needed to restore and the backups did not work. Typically at some point the VM failed to run the commands to clear temporary backup files, the disk would fill up and the new backups would be truncated mid way.
psql is pretty awful at identifying this which must be a very common scenario. Instead of a header check leading to "Corrupted backup, aborting" or something, it simply executes the SQL commands until it reaches the very last line, which will be truncated midway and thus invalid SQL.
Anyway, it took some time to figure out, but later I did basically what your tool does, probably simpler: another daily cron job that downloads the file and checks if it ends with the correct pg_dump success signature. It has worked so far.
Why do LLMs like this kind of writing? Why did you need to “plainly” name the competition as opposed to “flamboyantly” naming them?
This is a cool idea but the README is absurdly not-to-the-point.
“I should not pretend other implementations don’t exist. I should name them plainly.” Unable to tell itself that it puts this in output so that it can tell itself there.
Same with the “not X but Y” style. It is self-steering introspection leaking into output for one reason or the other. Recent poor language use by model is probably an attempt to minimize this by being concise. But it really needs more thinking to solve problem and so it does some of it in output.
I'm willing to use (semi-)vibecoded projects under certain circumstances, but I like to know up front what I'm getting. The github contributors list makes that easy to see as long as people don't try to hide it.