rdmsm4x changelog — Tyrell Build 38 (Phase 1 resiliency) on 6/6 hosts · 2026-09-26 21:27 → 21:58 EDT

Tyrell now queues messages durably and catches up after a hub outage: after a hub outage, a spoke's queued chat messages are delivered in about 5 s, exactly once. Build 38 is installed on all six Macs.

Scope

• Hosts: rdmsm4x (hub), rdmbair15m5 (canary), rdmpw3275m, rdmbair13m5, jdmbair13m5, rdmpw3265m.
• Repo: ~/dev/apps/Tyrell, main b56f658 (stamp d90d7f8, merge b6ae540), tag v0.2.0-build38. Pushed to fleet and backup.

What changed

• FEAT-20260926-07 (resolved), which adds:
  • a durable outbox wired to a real destination;
  • per-source cursors and idempotent apply;
  • backoff with jitter, and a network/wake fast path;
  • /api/health and tyrell doctor;
  • in the app: queue state, badges, and an offline/catching-up banner.

Commands

• Gate: p1-integration/gate.zsh
• Cut: cut/gui_run.zsh cut38 cut/cut_build.zsh …
• Rollout: cut/rollout.zsh ~/dev/_handoff/tyrell-build38-release-20260926
• Clone fast-forward: cut/tyrell_ff.zsh over ssh.

Verification

• Tests: Swift Testing 1,163 and XCTest 436 (1 skipped) / 184 / 11, with 0 failures.
• Build: universal2, notarized.
• Fleet: 6/6 hosts run app 38 and daemon-5baed89778e4; health reports build 38.
• Live hub-outage test: 5 messages queued on the spoke were delivered 5.14 s after the hub returned, 6/6 distinct on the hub.
• Evidence: ~/dev/_handoff/tyrell-build38-release-20260926/ and ~/dev/_handoff/tyrell-grok-agents-20260926/HANDBACK.md.

Side effects

• The hub tyrelld was stopped for about 25 s during the test: launchctl bootout then bootstrap.
• Six test messages ("Build 38 live acceptance …", sender claude@probe) are now in #general.

Undo

• Roll back to Build 37 with scripts/redeploy_app_fleet.zsh using the Build 37 bundle. The rollback copy is kept per host.
• Or reinstall v0.2.0-build37.

Outstanding

• Phase 2 (FEAT-20260926-06 transcript index, FEAT-20260926-05 unified inbox) is in progress.