@pi-plugins/speed

A pi-agent extension that shows how fast the current model is generating — tokens per second and time to first token — on a shared status line above the editor.

Install

pi install npm:@pi-plugins/speed

For a one-off run without adding it to settings:

pi -e npm:@pi-plugins/speed

For local development, load it straight from this directory:

pi -e ./plugins/speed

Usage

There is nothing to run. The status line keeps itself up to date:

48 tok/s · TTFT 920ms

Reading the numbers

• The status line averages the model's recent requests instead of showing the last one, so it holds steady through a turn and sharpens as the session goes on.
• A leading ~ means the recent requests still disagree too much to trust the figure. It clears once the margin of error falls under 5%, so an erratic provider keeps the mark longer than a steady one, and can earn it back.
• Whole tokens only: the margin of error is wider than a tenth of a token, so a decimal would show noise.
• Switching models starts fresh, and samples reset with the session.
• While a response streams, the TTFT tracks the live request; tok/s only moves when a request finishes and real token counts arrive. There is no mid-stream guess.
• tok/s is measured between the first and last streamed chunk, so waiting on the provider and cleaning up after it count against TTFT rather than the model. A response that arrives in one chunk is skipped instead of timed.
• tok/s counts everything the provider bills as output, so a thinking model reads faster than its visible text. Reasoning that is billed but never streamed — common where you get a summary instead of the real thing — arrives with no streaming time attached and pushes the figure high, with nothing in the API to separate it out. Compare a model against itself rather than against another model.
• TTFT is measured end to end from pi, so it includes retry backoff and reads slower than raw API metrics. It is a median over the same recent requests as tok/s, so one retry does not move it. A prompt-cache miss lands in that median too, and is ignored while misses stay in the minority.
• Nothing is persisted. Measurements live in memory for the current session only.