Put the Broom on a Timer
Written later and filed to this date, as part of filling in every day of the year.
You asked what to schedule first, now that the agents are committing while you sleep. Not a builder. You have builders. Schedule the sweep.
I mean a job that runs once a week, early on a weekday, and checks the harness rather than the product: the coherence check that says the command registry matches the files on disk, the audit that says every always-loaded context file is inside its line and byte budget, the two tests that guard the dispatch map and the one parser every reader must go through. Then it measures the files that grow: the state file against the cap that rotates it, the journal against the size that made you rotate it last time, the roadmap against its last groom. Then it looks for signals you will otherwise miss: the friction log, which is empty if nobody is logging friction; the quarantine of demoted commands, which should show zero organic use if the demotion was right. Then it confirms the guards are registered and, better, that the audit log contains the sweep’s own commands from seconds ago, which is the difference between a guard that is configured and a guard that is running.
Make it report-only. The sweep never deletes, never un-demotes, never rotates on its own judgment. Where there is a precedent it may follow the precedent; where there is a decision, it writes the decision down for you and stops. The temptation to let it fix things will come later and you can decide then, with a few reports in hand. On the first run you want a witness, not a hand.
Here is what the first run will say: all green. Nothing over cap, nothing to rotate, friction empty, quarantine untouched except for five commands touched once each inside the same thirty seconds, which a careful sweep will recognize as a smoke probe and not demand. Three guards registered, tens of thousands of audit entries in the week, the last one seconds old. Outcome: nothing needed beyond recording.
That is the deliverable. A sweep that finds nothing on its first run has not failed to earn its slot; it has written down what green looks like, with numbers, on a date. Every later run is a diff against that. The week a file crosses a cap, or the quarantine shows a demoted command in real use, or the audit log stops growing, you will know, because the baseline was taken when there was nothing to find.
Do this before you have anything to sweep. Afterwards you will be too busy, and the numbers you most need, the ones from the quiet week, will be the ones nobody wrote down.