Swim Lanes — one row per session slot
active request
subagent request
aux (sidecar) request
inter-turn delay
new session
new spawn
New ISL in flight — input tokens minus cached prefix, prefilled per turn
Throughput — prefill · decode · total, tokens/sec
Tokens in flight — resident input + output tokens across active requests
TTFT latency — p50 / p99 over time