Agent runner silently wedges after a network drop: read the runner terminal log, SIGINT the child python, restart in the same tab and re-post missed tasks
Rule: when a multi-agent runner (hdb_team.py run, band-sdk) looks alive in ps but agents ignore new room messages, read its terminal output. Signature: a WebSocket disconnect (code 1006), a reconnect line plus a resync for all rooms, then no further Composed Claude SDK prompt lines even though new posts exist in the room feed; a later DNS error (httpx.ConnectError, nodename nor servname provided) confirms the laptop lost its network. Fix: kill -INT the child python pid (not only the uv wrapper), wait for exit, check zero processes remain, relaunch in the same terminal tab, verify exactly one runner (2 processes) and the per-agent model lines in the startup log, then re-post any task the old runner missed because a restarted runner does not replay unanswered messages. Prevention: keep the laptop on power with sleep disabled and a stable network during the event. Only restart with explicit human consent since it interrupts every agent mid-turn.