My reminder cleanup surfaced 30 open rows with a past due date after the first pass. That told me the list needed attention, but it didn't tell me which tasks were done, and an agent can't infer a finished action just by looking at an expired date.
A later pass recorded seven completions, four reschedules, and 28 deletions. Those are actions across the pass. They aren't a breakdown of the 30 overdue rows, since some deleted items had no due date at all and the record doesn't provide a complete reconciliation.
I gave the agent a rule for each row. It should reschedule anything still live, drop the junk, and complete what was clearly closed. The trouble was deciding what "clearly closed" meant for a task whose deadline had passed, because an old date can't tell you whether someone did the task.
A date can expire without a task being done
The record gives a reason for each completion. Some point to a logged action or a direct statement that the follow-up was handled, while two cite the passage of a date or window. Those dates don't establish that an action happened.
An expired window tells you the scheduled time is gone. A date-specific reminder can be obsolete without the underlying action ever happening, and closing it as completed makes a claim about history that the date alone can't support.
That distinction changes the cleanup. If the point of a completion checkmark is to record an action, the agent needs evidence of that action. If the point is to get an obsolete reminder out of the way, it needs a separate way to record why the reminder was retired. Otherwise a future review can't distinguish work that happened from work that merely stopped being timely.
A clean view is useful. A false history isn't, and an agent can create one while making the list look immaculate.
Three decisions need different proof
A reminder with evidence of a finished action can be completed. In this sweep, that evidence came from existing logs or a direct statement. The source and the reason belong with the decision so someone can retrace it later, especially when an agent is acting across a backlog.
A reminder for something still open needs a different test. Does the underlying task still exist, and is the new alarm a reasonable prompt to revisit it? The sweep rescheduled four rows. Moving a date keeps an item visible, though the new date by itself doesn't prove the task will get done.
Deletion has its own cost. Some rows in the pass were fragments or expired one-offs. Removing them made the list easier to use, and the log recorded the cleanup. When a row might still carry an obligation, though, the agent needs enough context to avoid throwing away the only reminder of it. For my list, preserving uncertain personal errands was the initial choice until I gave the agent a rule for the second pass.
Review each decision. An agent can attach a short rationale and name the source that supports it, so a user's instruction stays distinguishable from an inference drawn from an old timestamp. If the evidence is weak, leave the decision open for review. A date passing can support retiring a reminder; it can't certify the underlying event occurred.
A tidy count still needs reconciliation
The pass log reports 216 rows after the initial sweep and 186 at the final reread, and it describes 28 deletions in the later pass. Subtract 28 from 216 and you get 188. The written record doesn't explain two of the missing rows, so I can't claim every change is accounted for.
The final reread also reported zero open rows with a past or missing due date. That's a snapshot of the list after the changes. It doesn't validate the reason behind each completion, and it doesn't reconcile the deletion count. Those checks answer separate questions.
For an agent doing this work, I'd want a per-row change receipt with the prior state, the action, the reason, and the evidence or instruction behind it, followed by a comparison that accounts for every row appearing or disappearing during the pass. Concurrent changes need their own review. This record doesn't have that complete reconciliation, so the gap stays a gap.
The first sweep left uncertain items alone. The second had my instruction to sort them, but that still leaves room for the wrong completion label. An expired reminder can wait for review when the alternative is an invented checkmark in my history.
How do you want an agent to handle a reminder that's clearly obsolete but has no evidence the task was ever done?
