A self-healing scraper that must prove its repair against twelve records that cannot move

When a site redesigns, a selector breaks and the data quietly goes to zero or, worse, goes wrong ^[5]. Zero gets noticed. Twelve tidy records with the wrong fields attached pass every smell test available to a human skimming output, and the harness in patchwright is built for that case: a fake listings site holding twelve records rendered through a template, where the mutation that breaks it only ever touches the template ^[6][7]. Verification then collapses into a comparison against something that cannot drift, so a patch returning twelve well-formatted but wrong records is scored as a failure ^[8]. The author's own summary is that every other guarantee in the project leans on that one ^[9].

The arithmetic on the real system is worth doing. 380 ingestion runs in fourteen days is about 27 a day ^[21], averaging just under two fetches per run ^[22], roughly 39 fetches per scraper across the fortnight ^[23]. A wrong patch approved on a Monday keeps two dozen runs a day writing plausible output until somebody reads it line by line.

Why a person sits at the end of the chain is visible in the failures the write-up is candid about. One agent was named s3_diagnose, read its own name, decided it was working with Amazon S3, and diagnosed a missing bucket permission on a problem that had nothing to do with storage ^[13]. Separately, Gemini 3.7 spends tokens thinking before answering, and with the output budget set too low the answers came back chopped off mid-word, looking exactly like a serialization bug ^[14]. Gemini 3.7 Flash does the diagnosis and the rewrite; the reviewer's brief is written by Gemma 4 ^[11]. The adversarial review catches errors the first model makes alone. It is not evidence about errors both are prone to.

The sandbox history says the same thing about proof. A bare node:vm context turned out to carry globals the author had not asked for, one of which would have corrupted the channel the sandbox uses to report its results the first time a patch tried to log ^[15]. A corrupted result channel is worse than a crash: nothing stops, and the verifier grades numbers it has no reason to distrust. The second failure was louder, a patch importing a Node module raising the error in the parent process rather than inside the sandbox, which would have taken the whole service down instead of failing one repair ^[16]. The threat named is not an attacker but a model that writes an infinite loop or reaches for the network to check something, and both have to fail safely every time ^[18].

None of this showed up in the test suite. Over a hundred tests passed before deployment and deployment still found bugs, because the tests stub the models and never touch the real cloud services ^[19]. One of them was a field the author called severity colliding with Cloud Logging, which reserves that name and uses it to set a log entry's level ^[20]. Commercial self-healing scrapers already exist, which he credits as prior art ^[10]. The part worth copying is the harness discipline, and its price is a fake site whose contents you refuse to change. The spending cap on Cloud Run belongs in the same category ^[12]: a repair loop that never terminates is a billing failure before it is a correctness one.