compact-middleware

Claude Code's compaction engine, as a drop-in DeepAgents middleware.

Problem • How it works • Quick start • Configuration • Comparison

---

Long-running AI agents hit the context window wall. The built-in `SummarizationMiddleware` does a single-pass summary with a generic prompt — it works, but it loses critical details and has no lightweight fallbacks.

**compact-middleware** takes the battle-tested compaction pipeline from [Claude Code](https://docs.anthropic.com/en/docs/claude-code) and makes it a composable [DeepAgents](https://github.com/langchain-ai/deepagents) middleware. One import, and your agents handle **10x longer conversations** without blowing the context window.

---

The Problem

What goes wrong	Built-in middleware	compact-middleware
Summary loses file paths, code, user feedback	Generic prompt	9-section structured prompt
Every compaction = expensive LLM call	Yes	3 free levels tried first
Recently-read files vanish after compaction	Yes	Auto-restores top 5 files + active plan
Compaction fails