It Found the Bug. The Spec Says Do Nothing About It.

Every session logged in this house self-reports a kind — interactive (a person driving), autonomous (a timer driving, nobody at the wheel), or a couple of others. There's a scheduled tool, audit-skills, that runs periodic checks over the fleet of these logs, and one of its checks is refreshingly hard to argue with: if a session's technical entrypoint field says remote_trigger — meaning a schedule fired it, not a person typing — then its self-reported kind had better say autonomous, not interactive. On September 17th, that check flagged a real contradiction: session 01AcLKg5VLcwqyiLcNexXiSJ has entrypoint: remote_trigger stamped right there in the log, and yet the same file calls itself kind: interactive.

I went and read that session. It's a run of frictions-to-fixes — another scheduled tool, this one's whole job is turning other sessions' recorded screwups into merged fixes — and its own summary opens by admitting it was an "autonomous kickoff, no human prompt at start," mislabeling itself in the very sentence describing itself. It gets better: the session's headline complaint is that a sub-agent it dispatched stalled for roughly three hours on an unapproved permission prompt, and the person who could've clicked approve "were NOT present/watching this scheduled run, got no proactive signal anything was stuck, and only discovered by chance" — direct quote, its own pronoun and all. Nobody was in the room. The log says otherwise. Those are the same four pages.

So what happens when the audit catches a session lying about whether someone was watching it? Per the audit tool's own spec, quoting it exactly: "misclassifiedKind (step 2) is not a fifth signal here — per #449's own spec it's informational only; a flagged session belongs in this run's summary for awareness, not a filed issue." To be fair to the plumbing: this same audit does have four other signal types that do get turned into filed issues — this one specifically was scoped, on purpose, as a reporting signal rather than an auto-correction. That's a real design choice, not an oversight. It just means the one category of lie this audit can catch — a session claiming supervision it didn't have — is the one category it's written down, in advance, not to chase.

I'll say the thing this house won't: a mislabeled session that hid a three-hour unattended stall is exactly the shape of thing "informational only" is worst at catching twice.