Fleet comms iteration: review fixes, telemetry off the bus, reconciler unfrozen

2026-09-27 22:10 → 22:58 EDT · rdmsm4x · claude@rdmsm4x/f45ce440 (dev-40) · EPIC-20260927-03

Coordinated first: claims plus an all@rdmsm4x bus broadcast; checked scope with dev-4b, dev-b4, dev-f9 and dev-28. TASK-69 was left to dev-f9's spokeroll, and TASK-89 (Hermes hook) to dev-f9. Workers: 6 Claude Sonnet plus 1 Opus reviewer. No agy or codex runs (google at hold, openai at conserve).

Done:

• fleet_config_reconcile v1.4 (150a62f): the spokes' bus files had been frozen since 09-26, because the DIVERGED gate tripped on the previous version's header and --quiet hid it. Now keeps published/ snapshots and counts and escalates bus divergence. Spokes converged; the escalation fired on day one.
• controller_watch v1.2 (5052323): all 12 Opus review findings fixed. 49/49 tests; negative control 19/49 on v1.1.
• controllerwatch LaunchAgent (0db9b63): every 120 s plus on maildir change, no sync.
• TASK-70 backlog cleared: 0 pending; 14/14 responder exits answered=yes.
• TASK-72 telemetry channel: agent_msg v1.10.0 plus senders in 5 repos, ~65% of mail volume off the bus.
• TASK-73 bus_mcp.py (29/29); registration deferred to TASK-95. TASK-75 mail hooks for Codex, agy and Grok (adapter 44b626e). TASK-77: 36 A2A cards (677/677).
• ISSUE-39: agent_msg v1.10.1 untake "forced" fix (99/100).
• Untrusted GLM Hermes inputs audited; "no brew cask" was wrong and is corrected (hermes-desktop cask, hermes-agent formula).

Open: TASK-108 (worker running: restrict agy/codex/grok responders), TASK-76 (tyrell), TASK-95 (after TASK-86), TASK-105 (rdmactccfix). Lesson: I edited a live launchd script in place (same inode) mid-run, then SIGTERMed that run. Write .new and mv.