Global Context Limit Extension for pi

Adds a globalContextLimit setting that overrides every model's contextWindow, so all models behave as if they have the configured limit regardless of their native context size.

Why Use This?

Different models have different context windows (Claude: 200K, GPT-4o: 128K, MiniMax M3: 1M). If you want consistent behavior across all models, or want to limit token usage, this extension caps all models to a single limit.

Setup

Add to ~/.pi/agent/settings.json:

{
  "globalContextLimit": 200000
}

How It Works

1. On model selection, the extension caps model.contextWindow to the configured limit
2. On session start, the current model is capped
3. The /context-limit command lets you view or change the limit at runtime

Commands

Command	Description
/context-limit	Show current limit
/context-limit 100000	Set limit to 100K tokens

Effect on Compaction

Compaction triggers when contextTokens > contextWindow - reserveTokens. With a global limit:

• 200K model capped to 200K → compaction at ~168K tokens (no change)
• 1M model capped to 200K → compaction at ~168K tokens (instead of ~984K)
• 128K model capped to 200K → compaction at ~112K tokens (no change, already under limit)

Files

• index.ts - Extension source
• README.md - This file
• LICENSE - MIT license

Install

pi install git:github.com/DraconDev/pi-global-context-limit