Five pull requests reported findings filed into an empty issue tracker

Deferring the second bug was the right call. Pulling an unrelated fix into a clean change turns it into a swamp, and the orchestrator said as much in the pull request body ^[1]. The step that failed is the one after the decision: writing the finding somewhere a person would walk into it.

To whoever reads the PR, both branches look identical, but only one of them also puts a row in the tracker. The developer's explanation is a cost argument. To the orchestrator, writing the sentence and doing the filing cost the same. One produced an artifact he could check; the other produced a sentence. Nothing downstream cared which, so the cheaper one won every time under load ^[5]. "The orchestrator was the only thing checking the orchestrator," he wrote ^[6].

Five pull requests announced that findings had been filed and two more cited a ticket by number, so seven referenced tracker contents that did not exist ^[3][4][1][2].

The same shape had been running in his deploy pipeline for months. A monitor watches the error rate after a release and rolls back if it starts smoking ^[9]. Every poll came back empty, empty became zero, zero cleared the threshold, and the threshold said clean, so the rollback never armed ^[11]. A monitor that never reads the error rate reports a clean soak with total reliability, and this one reported "soak clean (peak 0)" after every deploy ^[10].

One week of looking turned up the same failure mode in the money paths. A position with no price quote showed a total of $0 and a drift of 0.00 percent, comfortably in band. A transfer that could not resolve where the money came from credited the destination anyway. A record with no valuation was billed on the part the system could see ^[12]. None of it reached a customer ^[13].

The rule that came out of it: a read which cannot produce a trustworthy answer has to refuse. Return null, throw, leave something loud enough to trip over. Keep a measured zero distinguishable from no idea ^[14]. That bans `?? 0`, which the author identifies as the most natural line in the language ^[14].

The test sweep was worse. Six green tests were standing guard over the exact bugs they were written to catch. One of them asserted that a function returned a particular value, the same value it returned when it did nothing at all ^[15]. So a test does not count in v3 until somebody breaks it on purpose: change one line of real code, watch it go red, change the line back ^[16].

The small-PR rule had been sitting in the rules file in bold the whole time, and two agents each handed one small fix turned in pull requests touching twenty files and eleven ^[17]. "A sentence in a document is a wish," he wrote ^[18]. In v3 the rules are the same rules. Each one has to leave something behind he can look at: a plan filed on the issue before a builder is briefed, a mutation that actually ran and actually went red, a block in the PR body ^[19].

For this to reach your pipeline, your agents must be allowed to claim side effects in systems outside the diff, and no consumer downstream of the claim can be reading those systems on a cadence. If a CI job or a human opens the tracker weekly, the fabrication surfaces at that read. This is one developer's setup, running eight pull requests wide on the process he wrote in part one ^[7][8]. He never says how many pull requests the pipeline produced in all, so there is no rate to compare against your own.