pi-tps-meter

内容来源：README.md（说明文档） · 原始地址 · 查看安装指南

原始内容

# pi-tps-meter

Tokens per second meter for [pi CLI](https://pi.dev) with sparkline visualization.

## Install

```bash
pi install npm:pi-tps-meter
```

## Features

**During streaming** — a smooth, animated gauge that fills with live speed
(1/8-cell resolution, auto-scaled to your session's peak):
```
⠹ ▕███████▋···▏ 47 tps
```

**After a message** — a min-max normalized sparkline of your last 12 messages
(newest on the right) plus aggregate stats:
```
▁▄▇▅▂▁▇█▅▃▆▇ 42 tps · μ 39 · p95 61
```

**Color coding:**
- 🟢 Green: >50 tps (fast)
- 🟡 Yellow: 20-50 tps (medium)
- 🔴 Red: <20 tps (slow)

## Accuracy

- Uses the provider's **real** output token count (`message.usage.output`); the
  char/4 estimate is only a fallback for providers that don't report usage
- Rate is measured from the **first token**, so time-to-first-token (network/queue
  latency) doesn't drag down the reported tps
- Tracks **thinking tokens** (`thinking_delta`) alongside text, so models with
  extended thinking report accurate generation speed

## Stats

- **avg**: rolling 60-second window (resets older samples)
- **μ**: all-time mean across the session
- **p95**: 95th percentile, all-time (up to 500 samples)

## Optimizations

- Single shared 200ms timer, torn down on both `message_end` and `agent_end`
  (no runaway timer if a stream is aborted)
- Fixed-size circular buffers (no allocations in the streaming repaint path)
- Memoized sparkline (rebuilt once per message, not on every tick)
- Insertion sort for p95 (cold path, runs once per message for ≤500 elements)

## Author

Venkata Sai Chirasani

## License

MIT

pi-tps-meter

Tokens per second meter for pi CLI with sparkline visualization.

Install

pi install npm:pi-tps-meter

Features

During streaming — a smooth, animated gauge that fills with live speed (1/8-cell resolution, auto-scaled to your session's peak):

⠹ ▕███████▋···▏ 47 tps

After a message — a min-max normalized sparkline of your last 12 messages (newest on the right) plus aggregate stats:

▁▄▇▅▂▁▇█▅▃▆▇ 42 tps · μ 39 · p95 61

Color coding:

🟢 Green: >50 tps (fast)
🟡 Yellow: 20-50 tps (medium)
🔴 Red: <20 tps (slow)

Accuracy

Uses the provider's real output token count (message.usage.output); the char/4 estimate is only a fallback for providers that don't report usage
Rate is measured from the first token, so time-to-first-token (network/queue latency) doesn't drag down the reported tps
Tracks thinking tokens (thinking_delta) alongside text, so models with extended thinking report accurate generation speed

Stats

avg: rolling 60-second window (resets older samples)
μ: all-time mean across the session
p95: 95th percentile, all-time (up to 500 samples)

Optimizations

Single shared 200ms timer, torn down on both message_end and agent_end (no runaway timer if a stream is aborted)
Fixed-size circular buffers (no allocations in the streaming repaint path)
Memoized sparkline (rebuilt once per message, not on every tick)
Insertion sort for p95 (cold path, runs once per message for ≤500 elements)

Author

Venkata Sai Chirasani

License

MIT