Finishing a half-working script, scored against a sealed list by Jago WesterduinFinishing a half-working script, scored against a sealed list by Jago Westerduin

Finishing a half-working script, scored against a sealed list

Portfolio project
Jago Westerduin
The job: a previous developer left a script that mostly works. Make it reliable without rewriting it. Here a booking-to-invoice sync for a fictional climbing gym: one invoice per confirmed booking, a credit note when one is cancelled, run every 15 minutes.
Done blind. I wrote the spec and 25 tests from it before the script existed. Someone else (a separate AI agent playing the previous developer) wrote the script, hid problems in it and sealed a list of them. I ran the tests, traced every failure to its cause, read the code for what the tests could not reach, and patched it: 59 lines added, 41 removed of 261, so the client can follow the diff.
Then the sealed list was opened: 9 of 9 hidden problems found and fixed, 8 of them by the tests written before I read the code, 1 (a rate limit that never gives up and hangs the job) only by reading. Tests: 14 of 27 passing as handed over, 27 of 27 after. A lost response no longer makes a second invoice, 12 people are 12 customers instead of 23, and a booking whose sync failed is picked up by the next run instead of lost.
Stated plainly: the problems were planted by an AI agent, not found in a real client's script, and one of my own tests was wrong at first and was corrected. The client gets the fixed script, the offline tests and a runbook that says what each log line means.

0 likes · 0 views · Nov 21, 2026

A half-working sync script made reliable without a rewrite. Bugs hidden by someone else: 9 of 9 found and fixed, tests 14/27 to 27/27.

Tags