docs(metrics): clarify projection and latency contracts

This commit is contained in:
imccyu
2026-08-06 00:22:09 +08:00
parent aaca2fa52f
commit a17eebcaca
8 changed files with 13 additions and 13 deletions

View File

@@ -59,8 +59,8 @@ interface TurnFold {
/**
* Fold assistant nodes into per-turn footer metrics.
*
* TTFT is the turn's lowest-step reading — the user-perceived wait before
* output appeared — so it is only meaningful when the turn's start is inside
* TTFT is the turn's lowest-step request-dispatch-to-first-token reading, so
* it is only meaningful when the turn's start is inside
* the loaded window (the caller gates on `turnTimings`, which shares that
* window). Throughput divides summed output tokens by summed decode wall time,
* counting only steps that carry both.