The HA Pair
Mental-model HA for orchestrators. When one compacts, the other remembers.
The Insight
The orchestrator is your single interface to the rig. It holds the plan, tracks progress across pods, and makes delegation decisions. If it compacts badly, you lose coordination. Tasks get duplicated, context gets dropped, and the whole topology drifts.
Run it as an HA pair — Claude Code + Codex. They share memory through externalized state. A peer can help a returning session reconstruct where the work was, from what was written down. This is a coordination pattern, not automatic lossless restoration; retained records and the declared continuity strategy determine the result.
This is the single most important topology decision you will make. Every other pod can recover from a bad compaction because the orchestrator remembers what they were doing. But if the orchestrator itself forgets, nobody is there to restore it — unless it has a peer.
When This Applies
Any rig with 4+ agents where coordination matters. If the orchestrator is managing multiple pods with interdependent work, it needs a backup.
When This Doesn't Apply
Small 2-agent rigs where you are directly managing both agents. If you are the orchestrator, you don't need an agent to do it for you.