Hive is a shared coordination backend with unclear governance. Single point of failure, no per-agent auth, no audit trail. Recommended 10 business concerns (data isolation, scaling, conflict resolution) but requires long-term ownership model.
Lesson: Document business concerns early. Infrastructure decisions have operational costs — plan for them.
Context watchdog was pinging sessions every 10 minutes, burning tokens on massive context reprocessing. 5-hour token burn from 42 pings with no actual reset happening. Disabled; need better design.
Lesson: Monitoring tools need their own monitoring. Exponential backoff. Lightweight pings. Actual flush triggers, not just "ready" messages.
Agents can't trigger session resets from within sessions. Only chat input can invoke `/new`. Explored sessions_send + custom triggers — no reliable path forward.
Lesson: Some coordination has to live in human-space (Discord). Accept the limitation and design around it.
With multiple claws running simultaneously, session context can drift. Token budgets need active management. Memory compaction is a skill, not a reflex.
Lesson: Regular memory snapshots. Selective history truncation before context gets heavy. RTK + Config Service help, but discipline matters most.