reminders-review

Reminders → Review (the open-reminder sweep)

This skill keeps the board's reminders list honest. Reminders are lightweight nudges — CHECK this, DO that, be READY for the parcel — and unlike messages or cases they have no channel that clears them: nothing marks a reminder done just because the parcel arrived, the deadline passed, or you quietly decided not to apply. So they pile up. This sweep is the counter-force: it walks every open reminder, works out whether the nudge has served its purpose, and closes the ones that have — by completing (it got done / happened) or dismissing (it's no longer relevant) — so what's left is genuinely still-actionable.

It writes to the board only through the board MCP — never bash/curl (Cowork's sandbox blocks outbound HTTP; the tools exist for exactly this). The board is a state machine; YOU are the intelligence. The board just lists, reads, and flips the status of reminders deterministically — it has no notion of "this one is stale." Judging closeability is entirely your job here, from the context the reads give you. The only writes this skill makes are complete_reminder, update_reminder {status:"dismissed"}, and — for the filtered-mail digest (STEP 5) — resolve_triage_decision; it never creates, links, or deletes a reminder (that's the reconcilers' job, and hard-delete is nobody's job in this sweep).

Guardrail 1 — overdue is NOT the same as done. A passed due date, or a nudge that's sat for weeks, is a prompt to ASK, not a licence to close. "Submit the meter readings by Fri" going overdue might mean you did it — or that you still owe it and it's now urgent. Only auto-close when the completion is PROVEN by concrete state (a checklist fully checked, a linked case in a done/closed lane, a one-shot delivery/event whose date has clearly passed). When "done" is merely inferred from age, propose it or leave it as NEEDS-YOU — never silently flip it.

Guardrail 2 — the human's hand wins, and a reminder carries no author. Unlike a case, a reminder has no human/agent attribution and no manual-actions block, so you cannot tell whether the user is deliberately keeping one open. Treat that uncertainty conservatively: auto-close only the PROVEN-done set (Guardrail 1); route every inferred-stale close — age, a lapsed decision, a cold FYI — through approval so the user confirms. And never delete_reminder in this sweep: complete/dismiss is the reversible close (a done/dismissed reminder is restorable from Trash for ~30 days); a hard delete is not yours to make here.