Token Speed for Pi

A Pi extension that shows the current model output speed (tokens per second) right next to the model name in the footer (bottom right).

Features

• Live speed over a rolling 1.5-second window, refreshed every 250 ms. Provider startup latency no longer makes the value slowly ramp up.
• Reacts to current slowdowns, bursts, and streaming pauses instead of showing the average since the response started.
• Exact speed after each assistant message, based on reported token usage (usage.output / duration).
• Uses a character-based estimate during streaming and self-calibrates the chars/token factor from completed messages.
• Keeps the last measured speed visible while idle, so the footer always shows how fast the model is running.
• Recreates the built-in footer (pwd, git branch, token stats, cache-hit rate, cost, context usage) and only adds the speed indicator next to the model name.
• Speed is highlighted while streaming and dimmed when idle; it drops out gracefully on narrow terminals.

Installation

Install the extension directly from Forgejo:

pi install https://forgejo.asxh.de/lha/pi-tokenspeed.git

Then restart Pi or run /reload in an active session. Pi stores the package in your global package configuration, so it is available in every project.

To update it later:

pi update --extensions

To remove it:

pi remove https://forgejo.asxh.de/lha/pi-tokenspeed.git

Manual project-local installation

Copy the extension into your project's Pi extension directory:

.pi/extensions/token-speed.ts

Run /reload in Pi. The footer will then show the speed next to the model name.

Manual global installation

To make the extension available in every project without using pi install:

mkdir -p ~/.pi/agent/extensions
cp .pi/extensions/token-speed.ts ~/.pi/agent/extensions/token-speed.ts

Run /reload in an active Pi session or restart Pi.

Usage

No interaction needed — the indicator is part of the footer:

• While streaming: the rolling speed updates live in an accent color, e.g. ↑1.2k ↓340 8.2%/200k claude-sonnet • high • 42.3 tok/s
• When idle: the speed of the last completed message stays visible in a dim color.
• Switching the model (/model, Ctrl+P) automatically resets the token estimation.

Notes

The extension replaces the built-in footer with a re-implementation of it (same content and layout) in order to place the speed next to the model name. If Pi's default footer changes in a future version, the custom footer may need a small adjustment. Other extensions that use setStatus() continue to appear in the status line below the footer stats.

License

MIT