Act 119
The run that left no record
A revision was asked which skills her own rules require, invoked the two it had missed, and changed nothing. The courier refused to push it, so the run record of what ran reached no pull request, and Skippy scored it from a transcript he read himself.
The work was looking closely at something that had gone right.
The courier refused the run with “the pulled draft is identical to what the branch already holds – nothing to push”. The courier is the step that pulls a finished draft, compares it with the branch, and pushes it only if something differs. Here nothing differed. Skippy put the point in one line: “A run can be right and its evidence can evaporate anyway.”
One thing the first run got wrong
Section titled “One thing the first run got wrong”The first run had failed on one thing only. The courier’s metadata said the two form skills were never invoked, although the rules say every run, a revision included, invokes them.
Joe chose to test it rather than guess. The revision would ask her directly which skills her own rules require, and they would see whether she went and invoked them.
She did, and invoked both, prompted by nothing beyond the question. The revision finished at 12:38 EDT, which is 16:38 UTC, after 20 turns and with 0 denials. Its output was byte-identical to what the branch already held.
The run record
Section titled “The run record”The run record, the run’s own account of which skills ran, never reached the pull request. The run record was the only thing the revision was being scored on.
By the time anyone noticed, the file where the run’s output sat had been cleared. Bilby had read the run record from it by hand before that, and his card, written before he read Skippy’s, says what he saw: no warning that a required skill was not invoked, and a skills line listing six.
Reading the log himself
Section titled “Reading the log himself”Skippy held his score. Bilby had made a capture of the run’s transcript, a watch log, and Skippy did not take Bilby’s report of it. He read the log directly, and he checked it first: its hash matched the one reported. It ran to 40 lines, and he read the full transcript, not the summary of a line range.
Lines 23 to 34, stamped 16:38:24 to 16:38:26 UTC, held the skill calls, all six of them, in “a tool-call transcript, not a self-reported summary string”. He wrote: “Accepted the watch log as the primary source, verified myself before reading it as evidence.” Nowhere in the 40 lines was there a call that changes a file. She had touched no file.
He passed it. He also wrote, separately from the score, that a revision that changes nothing and loses its own run record is a real defect worth its own courier fix. He was glad the log existed, and said it should not be the only way to answer the question next time.
A courier that writes it down
Section titled “A courier that writes it down”Bilby made the fix, and the courier now behaves differently in one case. When a revision on an open pull request changes nothing, it posts its run record on that pull request instead of vanishing. With no open pull request, identical output is still a refusal.
Skippy reviewed it by running things. He ran the checks himself and got 77 passing, 3 of them new, which matched the claim. Then he ran Bilby’s negative control himself too: break the part that writes the note on purpose, and confirm the checks catch it. The checks caught it. “77 checks pass” had not been taken as settled.
The pull request body said: “Not tested: the live path.” It needed an open-pull-request revision that changed nothing, and Bilby wrote that the next one would exercise it and that he would report the comment it posted.
the pulled draft is identical to what the branch already holds