@arhen/pi-core-tps-stats

Install

Requires the pi coding agent — install it first: npm install -g @earendil-works/pi-coding-agent.

pi install npm:@arhen/pi-core-tps-stats

Live token-per-second stats for the active model. Status bar shows median t/s + median TTFT; /tps-stats shows full stats (samples, avg/median/min/max, TTFT). Stats reset on model change.

One rate, deliberately. t/s is all output tokens (thinking + text + tool-call arguments) divided by the whole turn, prefill and queue latency included. It reads lower than a provider's marketing number because it is the rate you actually wait for.

There is no separate "streaming t/s", because SSE arrival times measure the gateway's flush schedule rather than the model. Measured against vantis/deepseek-v4-flash-0731-fast (median inter-chunk gap: 0.01ms — chunks land in instant batches separated by long pauses):

| prompt | turn | window-based | this extension | |---|---|---|---| | Say OK. | 1.40s | 263 t/s | 41 t/s | | List 3 fruits. | 8.81s | 801 t/s | 14 t/s | | 900-word essay | 18.49s | 89 t/s | 55 t/s |

The window-based column swings 9x on one model within a minute. Versions before v1.3.0 shipped that math and reported four-digit rates no local model can reach.

TTFT stays a direct observation — first streamed token minus turn start. On reasoning models that first token is usually thinking, not visible text.

npm test --workspace @arhen/pi-core-tps-stats

License

MIT.