# pi-dumb 🧠

> **D**egradation **U**sage **M**eter **B**ar — a context rot indicator for [Pi](https://github.com/mariozechner/pi-coding-agent)

Shows your current token usage in the status bar with research-backed color warnings for when your LLM is getting dumber.

## Why?

LLM quality degrades based on **absolute token count**, not context window percentage. A 200k-window model at 100k tokens performs just as poorly as a 1M-window model at 100k tokens.

Research:
- **Chroma (2025)**: All 18 frontier models degrade as context grows
- **Liu et al. (Stanford)**: 30%+ accuracy drop from "lost-in-the-middle" effect
- Attention dilution is **quadratic**: 100k tokens = 10 billion pairwise relationships

## Thresholds

| Icon | Tokens | Zone | Meaning |
|------|--------|------|---------|
| 🧠 | 0–30k | Fresh | Sweet spot — peak reliability |
| 🟡 | 30–60k | Warm | Past recommended reset point |
| 🔥 | 60–100k | Hot | Danger zone for coding tasks |
| 💀 | 100k+ | Rot | Broadly degraded quality |

## Install

```bash
pi install npm:pi-dumb
```

## What it looks like

After each turn, you'll see something like this in your status bar:

```
🧠 12.4k     # you're good
🟡 45k       # consider compacting
🔥 82k       # quality is dropping, compact soon
💀 142k      # you should have compacted yesterday
```

## License

MIT