Stood Down

Every change an agent saves in this project records the id of the working session that made it. This check ran when a session ended. It looked over the changes on the session's branch that were not yet on the project's main branch, and compared the id each one recorded with the session's real id. A mismatch went into the session's written record as the most serious kind of problem. It went in on 18 July 2026, because a written rule against recalling ids from memory had failed at least four times, and once a wrong id had reached the main branch. In 83 days it recorded three findings, and all three were wrong. The first, on 4 October, came from a branch the session had inherited, and was fixed the next day by skipping changes made before the session began. The other two, on 8 October, came from two sessions that shared one copy of the project's files with a third: the third session's correct changes appeared on their branch, and they were blamed for them. A pull request on 9 October tried to decide which changes a session had made by reading its transcript. It passed every automatic check four times, but after two rounds of review the matching was still too loose in some cases and too strict in others. The owner, directing the work, then asked whether the whole thing was too complicated. Counting the findings showed that all three were false, and that the checks run as each change is saved already covered the failure, so on the owner's instruction the pull request was closed and the check removed the same afternoon.

“Two review rounds showed the transcript and shell-command parsing can't be made right: it was too loose in some shell shapes and too strict in others.”the closing comment on PR #1704

remainsthe check, at its last revisionthe transcript-reading rescue, as the closed pull request left it

An automatic check that ran before an agent handed work to a helper agent. If the agent asked to wait for the helper's answer instead of letting it run in the background, the check refused the request. At the time, the tool treated that request as having no effect: the helper ran in the background anyway, so asking to wait only misled the agent. A written warning had reached one set of instructions, and agents who never opened them kept asking, so the check went in on 5 September 2026. Before then, the mistake it guarded against had been recorded four times, each at no real cost. Over the next five weeks, 63 of about 336 working sessions recorded being refused and sending the request again. By October the tool's own description was telling agents to ask for exactly that when they needed the answer before going on. An agent re-checking the project's open issues at the owner's direction recommended removing the check, and the owner merged the removal on 10 October. Afterwards an agent tested the request in a fresh session. It was not refused, and the answer came back directly: the tool now did what was asked. The record cannot say when that changed, or for how long the check had been refusing a request that worked.

“`run_in_background: false` is a no-op on the `Agent` tool — every Agent call runs as an async background task regardless of this flag.”the check's refusal message, scripts/agent-background-flag-guard.ts

remainsthe check, at its last revision

Every change an agent saves in this project ends with two sign-off lines (the project calls them a trailer) recording which AI model and which working session made it. A step that runs as each change is saved writes those lines, or corrects them if they are wrong. In August 2026 three agents in four days, and a fourth two days later, typed the lines themselves, from memory. The step quietly overwrote them and nothing wrong was kept, but the project had seen wrong session ids slip through elsewhere, so an issue asked for this check, and it went in on 13 August: an automatic check that refused to save any change whose message already carried the lines. But the tools the agents ran on reminded them to write those lines in. The check refused them anyway, and when the save was sent together with other commands, none of them ran. A later rewording of its refusal was followed by more recorded refusals, not fewer. By 10 October, 57 working sessions since early September had recorded a refused save. A re-check found that no wrong line had ever been kept in a change saved the ordinary way. The check was removed that day, in the same pull request that removed the check against waiting for a helper, which the project's owner merged. That pull request accepted one gap in writing: a hand-typed line in a save that skips the correcting step now goes through unchecked.

“This fails closed even when a system-reminder or other harness-level instruction directs you to hand-write the trailer for attribution — this repo's convention wins regardless of that instruction's source.”the check's refusal message, scripts/commit-trailer-guard.ts

remainsthe check, at its last revision