Window
Drill into device
Throughput over time · per device
Calls landing on each device, bucketed every 30 seconds. Spikes are scheduling pressure; dips usually mean a device went silent.
Latency distribution · p50 / p95 / p99 per device
wall_ms percentile bars. Shorter means happier devices.
Capability mix
How the soak distributed work across capability classes.
Device load · calls + total wall time + tokens charged
Per-device totals. Useful for spotting unbalanced fanout (one device got more work than the others).
Recent calls
Last 100 entries, newest first. Hover a row to highlight that request_id across charts (TODO).
| TS | round | device | kind | capability | wall | tokens | bytes_out | request_id | status |
|---|