/fleet-wrap — the quiet-down protocol

Fleet law: canon #7 owns guard/DONE/harvest-close. This skill keeps the operational wrap procedure: stop monitors, publish one final artifact/message, then stand down.

When the fleet wraps: ZERO polling crons, ONE final dashboard + ONE message, then SILENT.

If the planned overnight mechanism breaks, choose one of two paths only: delegate real approved work to a live owner, or stand down silent. A polling loop with zero workers is the inverted fleet-wrap failure.

When this fires

• All workers are done/idle AND their claims are verified — no queued approved work remains.
• The overnight/sprint queue is exhausted.
• Etan said "wrap the fleet", "stand down", "we're done", or approved the final state.

Disambiguation vs /session-handoff: "wrap up" aimed at YOUR session (context high, session ending, successor needed) routes to /session-handoff — that skill owns single-session continuation. THIS skill fires on FLEET-wide close: all workers done, nothing queued, the channel goes quiet. When both apply, run /session-handoff FIRST — then this protocol governs the outgoing agent's exit.

Retire protocol (outgoing agent — pairs with /session-handoff)

Succession is routine (gen-12 evening: orc → gen-13 path, vl LEAD v1→v2, bl LEAD outgoing→successor — three successions one day). Before you compact or go silent:

1. Hand off completely — live-watch transfer list written (see /session-handoff); successor brief posted; grill answered (monitors still running during transfer).
2. Stop monitors — kill every cron/monitor you own so compaction doesn't fire on stale loops (71a8e3f5 [5418]).
3. Final dashboard (if fleet-wide wrap via this skill).
4. Kill your own pane — do not linger as a zombie orc writing duplicate handoffs.

Order: handoff → stop monitors → final dashboard (if fleet-wide wrap) → close pane.

The protocol (in order)

1. VERIFY done. Per worker: state check + independent claim verification (/never-fabricate R8). A fleet is not wrapped because it looks quiet.
   • Live guards: run bash "$GOLEMS_DIR"/scripts/hooks/install-hooks.sh --host <mbp|m1> --status ($GOLEMS_DIR = the golems clone) (read-only) and parse every line: missing/unregistered exit 0 with the gate off, so the exit code alone is not the check. Wrap-blocking: hooks-live=absent; drift not 0 (incl. ?); a hook line other than ok/external(registered) (including retired-registered: … and dangling golems link: …); golems-fail-open not ok; any E1 … PRESENT; a nonzero exit or no hooks-live= header at all (the installer died before reporting). Quote each blocking line on the dashboard and in the one message. Never run --apply/--update during a wrap: that is the hooks owner's (golemsLead).
2. KILL ALL POLLING. Follow /collab-monitor to stop every owned watch; TaskStop native Monitor task IDs and stop packaged fallback producers; report the stopped IDs in the wrap-up (orc REF9). Zero exceptions — a "harmless" 5-minute status cron is exactly how the WhatsApp-all-night failure happened.
3. ONE final dashboard (/html-dashboard, cloned template — never from scratch): sprint state, merged PRs, open loops, blockers, money/tokens spent.
4. ONE final message (WhatsApp / the agreed channel): dashboard link + a 2-3 line summary. ONE message. Not one per event.
5. brain_store the wrap — sprint summary, stopped cron/monitor IDs, open loops. Tags: milestone, <project>.
6. SILENT. No further messages, no re-armed loops, no "still quiet" heartbeats — until Etan speaks or new approved work arrives.

What silence does NOT mean

• Blockers found during wrap go ON the dashboard and IN the one message — not as a follow-up stream.
• A real incident (data loss in progress, security) justifies ONE alert. Everything else waits for the next session.
• Silence ≠ abandonment. Open loops are written down (dashboard + collab + brain_store) so the next session picks them up — the channel goes quiet, the state does not evaporate.

Anti-patterns

Integration