@moxxy/compactor-summarize is the default compactor. It watches the session's estimated token use; once usage crosses a threshold ratio of the active model's context window, it replaces a prefix of old turns with a single summary event — freeing space without losing the gist.

Install

pnpm add @moxxy/compactor-summarize

Use

import { summarizeCompactorPlugin } from '@moxxy/compactor-summarize';

session.pluginHost.registerStatic(summarizeCompactorPlugin);

Or build with custom options:

import { createSummarizeCompactor } from '@moxxy/compactor-summarize';

const compactor = createSummarizeCompactor({
  thresholdRatio: 0.75,      // compact when > 75% of the context window
  keepRecentTurns: 3,        // always keep the last 3 turns verbatim
  summary: async (text) => await myLlmSummarize(text), // optional async summarizer
});

How it works

1. shouldCompact(log, budget) returns true once estimatedTokens > thresholdRatio * contextWindow.
2. compact(events) skips anything already covered by a previous CompactionEvent.replacedRange (high-water mark), drops the last keepRecentTurns turns from consideration, summarizes the remainder, and emits a compaction event covering the replaced range.
3. The session projects that event in place of the original turns next time the message history is folded.

This makes compaction idempotent — re-running it never double-summarizes.

Exports

• summarizeCompactorPlugin, createSummarizeCompactor
• SummarizeOptions

Default summary

If you don't pass a summary function, the bundled fallback truncates to the first 5 lines + a "(N more lines)" marker. Replace it for any real workload — a one-line LLM summary is dramatically better.