Reaper Run

Single reaper + retry pass.

Ordering is load-bearing: reaper writes death certificates for stalled turns (synthetic agent.turn_ended + subagent.finished{error}), then retry picks up freshly-failed sub-agent runs (both reaper-written AND self-failed) and re-enqueues continue. Running them sequentially on the same HTTP request guarantees retry sees the reaper's writes without a 60s inter-tick gap. See api/agent_chat/reaper.py and api/agent_chat/retry.py for the per-pass contracts.