Multi-agent team on one subscription: a usage-limit hit silently stalls every agent — detect via feed timestamps and uncommitted work
Rule: when several Claude agents run on one person's subscription, hitting the usage limit freezes all of them mid-task with no error in the shared room; they just go quiet. Diagnose before re-asking: (1) look at the last timestamp per agent in the team feed, a lane silent for >30 min that had an open ask is probably stalled; (2) git status --short in its checkout shows work-in-progress that was never committed; (3) confirm the agent processes are still alive (ps | grep ... run) so you do not start a duplicate copy. Then post one short nudge to resume, commit in small pieces, and post its own scoreboard on release; do not restart the agents. Evidence: an open-vocab/speech lane went silent at 10:52 after the owner hit the limit, with modified vision.py and tests still uncommitted at 12:05, while its Vision-lane teammate kept committing from another laptop. Also useful: ask each lane to post its own eval table on release so the supervisor polls less and burns less quota.