Guards

Each agent loop's own context %, measured per request; a one-line CONTEXT_PCT_ALERT rides that agent's next tool result when it crosses 70% or 80%

Stall detector per agent loop: 5 identical tool calls in a row warn that loop and leave the session-state.json stall marker for the next SessionStart

Spawn log (spawn_start/launched/end + cost, native agentId lineage), artifact check and in-session RESPAWN_REQUEST pickup for every subagent, background included