pi-early-compact

A Pi extension that compacts large-context models before they get dumb.

Many models advertise 1M-token context windows, but output quality often degrades well before that — in practice around 400k tokens. Pi's built-in auto-compaction only kicks in when the context nearly fills the window (contextWindow - 16k), so a 1M-window model keeps running past the point where it stops being useful.

This extension watches context usage after every agent run and triggers compaction early:

• Large-window models (contextWindow > 400k, e.g. 1M-class): compact at 350k tokens — the 400k effective limit minus a 50k headroom, so the next run cannot push the context past the limit where quality drops.
• Small-window models (contextWindow ≤ 400k, e.g. 200k-class): left untouched; Pi's built-in threshold compaction already handles them closer to the window.

Compaction fires on agent_settled — only when the agent has fully finished and no retry, queued continuation, or built-in compaction is pending — so it never interrupts work in progress. It shows a single info notification when it triggers, and Pi's built-in overflow recovery remains as a safety net.

While a large-window model is active, the extension also adds a footer status slot like early 352k/400k (88%) — context usage measured against the effective limit instead of the advertised window, so you can see how close an early compaction is. Small-window models get no slot; Pi's own footer percentage already covers them.

Install

pi install npm:pi-early-compact

Then restart Pi (or run /reload). No configuration needed.

How it works

| Setting | Value | | --- | --- | | Effective context limit | 400,000 tokens | | Headroom for the next run | 50,000 tokens | | Compaction trigger | ≥ 350,000 tokens, window > 400k | | Trigger point | after each settled agent run |

1. After every settled run, the extension reads ctx.getContextUsage().
2. If the estimated context is at or above 350k on a model whose window exceeds 400k, it calls ctx.compact() — Pi's own native compaction pipeline and summary prompt, identical to a manual /compact — and notifies the user.
3. Small-window models and fresh post-compaction contexts (unknown token count) are ignored.

Development

npm install
npm run check   # tsc --noEmit
npm test        # node --test

Requires Node.js ≥ 22.18 (Pi runs extensions with TypeScript support built in — there is no build step).

License

MIT