Get it for macOS

Show tokens per second on every completed turn, so a slow answer is visible without opening a log. Each finished turn gets a badge under the assistant message with output tokens, wall-clock duration, and the resulting rate, and the turn header carries the same figure in short form. Reads usage the provider reports where BB has it, and the local OMP session file for ACP and OMP threads, which report none. The rate divides output tokens by the full turn duration, so tool calls and provider waits count against it; it is not a model decode benchmark. Needs a completed turn, and the ACP and OMP path needs that thread's OMP session file on the machine running BB.


Overview

View source

What you get

Every completed turn shows a badge under the assistant message: output tokens, wall-clock duration, and the resulting tokens per second. The turn header carries the same rate as a short label, with the division spelled out on hover.

How the rate is measured

The rate divides reported output tokens by the full turn duration, from turn/started to turn/completed. Tool execution, retries, and time spent waiting on the provider count against the number, so it reads lower than raw model decode speed. Reasoning tokens are counted once, never added again.

Active turns gain their badge when the turn completes. A turn with no usable usage or lifecycle timestamps shows nothing rather than an estimate.

Where the numbers come from

BB providers that report usage in the event stream, such as Codex, Antigravity, and Exo, supply the token counts directly. ACP and OMP threads report no usage, so the badge reads the local OMP session file for that thread and labels the figure as native generation speed.

Requirements

BB 0.42 or later. The ACP and OMP path needs that thread's OMP session file on the machine running BB; without it, those turns show no badge. Badges require an exact thread and turn match, so a change to BB's timeline markup hides a badge instead of displaying a wrong number.

More from Daniel Walker

More in Token Usage & Limits 18

Understand or control token, context-window, and provider-quota use.

View all