pi-tps-live

中文文档 (Chinese)

Live tokens-per-second in pi's footer — right-aligned at the end of the first line, next to your working directory.

~/dev/project (main) • refactor-auth                          ⚡ 42.3 tok/s
↑1.2M ↓45k R980k CH82.1% $4.210 37.4%/200k (auto)     claude-opus-4-5 • high

While the model is streaming you get a live sliding-window rate in your theme's accent color; when it stops, the number switches to the exact average of the turn just finished (dim).

Features

• Right-aligned in the footer's first line — doesn't collide with token stats, cost, context usage, model name, or other extensions' status lines
• Live while streaming — sliding-window tok/s (~5 s window) updated in place as deltas arrive
• Exact at rest — after message_end, the last turn's average from usage.output / decode time
• Excludes time-to-first-token — decode timing starts at the first streamed delta, so a long thinking phase doesn't drag the displayed rate down
• Provider-aware — providers that stream cumulative usage (Anthropic output_tokens, …) get exact live numbers; OpenAI-compatible providers that only report usage in the final chunk get a self-calibrating character estimate, marked with ~
• Keeps pi's built-in footer — pwd/branch/session, ↑↓RW token counts, cache hit rate, cost, context percentage, model + thinking level, extension statuses: all still there, with the speed added on line 1

Install

pi install npm:pi-tps-live

Or from git:

pi install git:github.com/adamcjm/pi-tps-live

Then run /reload (or restart pi). The readout appears after the first assistant stream of the session.

To try it without installing: pi -e npm:pi-tps-live.

Display

State	Line 1 right side	Meaning
Streaming, provider streams usage	⚡ 42.3 tok/s	exact, cumulative usage ÷ decode time
Streaming, provider hides usage	⚡ ~42.3 tok/s	estimate from characters (self-calibrating)
Idle	⚡ 42.3 tok/s	exact average of the last finished turn
No data yet	(nothing)	after /new, before the first response

The speed readout is dropped automatically when the terminal is too narrow to show both it and the working directory.

Commands

Command	Effect
/tps	Report the current/last reading in a notification
/tps on	Enable the footer readout (default)
/tps off	Restore pi's untouched built-in footer
/tps reset	Clear the current reading and the learned calibration

How the number is measured

• Decode window starts at the first streamed text/thinking/toolcall delta, not at message start — TTFT is excluded by design.
• Exact path — if the provider streams a cumulative output-token count, the live value is usage ÷ elapsed and the final value is usage.output ÷ decode seconds.
• Estimated path — otherwise characters are counted in ASCII-equivalent units (CJK/full-width ≈ 2.5× an ASCII char) and divided by a chars-per-token ratio. The ratio starts at a code-leaning prior of 4 and is recalibrated on every cleanly finished turn from the real usage.output, so it converges to your content mix (code, prose, Chinese, …).
• Aborted/errored streams are still measured, but never used for calibration.

Notes and limitations

• The extension replaces the built-in footer with a faithful copy plus the speed element. If a future pi version changes the footer layout, the copy may lag behind until updated.
• (auto) next to the context percentage reflects pi's default auto-compaction setting (the extension API does not expose the live setting).
• The (sub) suffix appears for pi's subscription-backed kimi-coding provider (other subscription providers are not detectable from an extension).
• Only installed in TUI mode; in -p / JSON / RPC modes the extension stays loaded and counts streams but does not touch the footer.

Development

bun install          # dev dependencies (pi-coding-agent, pi-ai, pi-tui, typescript)
bun test             # 22 unit checks + 1 end-to-end check on pi's faux provider
bun run typecheck

The layout/measurement logic lives in extensions/pi-tps-live/tps.ts and footer.ts and is framework-free; register.ts contains the event wiring and is tested through a fake pi API. The e2e test runs a real AgentSession against pi's in-process faux provider (no network, no key).

License

MIT