I run long sessions with Claude. For a long time I've had Claude in the middle of some difficult coding task, and then I see the auto-compact message pop up. Every time I wonder if Claude actually picks everything up correctly and nothing is lost. Claude told me 78/82 auto-compactions landed while Claude was working. For a while I couldn't do much. At first I made a skill that would make Claude explicitly save progress throughout long runs. But that didn't fix the core problem and cost more tokens as well. Now that mods came out, I fixed the actual problem.

catnap tells Claude at 50% context what the context is and that it should finish what it's working on, write a handoff, and then initialize compaction itself. When Claude finishes its task, catnap compacts, and Claude can then pick up work again. The handoff replaces Claude Code's default summary, so we don't do similar work twice. If Claude for some reason doesn't compact before 60%, it gets a stricter note. And Claude Code's default auto-compact kicks in at 70% as a backstop. The 3 numbers are all settings.

After I took on catnap, 0 of 129 naps came mid-task: 98 of them in benchmarks and 31 in my own sessions.

The cost is the same as the default auto-compact kicking in at the same point, in tokens and dollars, on 200K and 1M context windows. Compared to the default auto-compaction that kicks in when context is nearly full, catnap used 56% fewer tokens. But lowering the default auto-compact will get you most of the way there as well.

I measured memory today as well. Claude worked through tasks one message at a time while 20 instructions and facts were said in the chat, inside long chatty messages, and nowhere else. Asked right after compaction to list them from memory, no file reads: all 20 were listed with catnap and with the default compaction, in every run (15 with catnap, 5 with the default). So in my tests the memory was exactly the same, but catnap got there with a handoff 60% shorter (4.0K tokens against 10K).

I was surprised that catnap is actually faster than the default option. This wasn't even one of the reasons I wanted to make this, it just happened. Since Claude writes the handoff in one go instead of the separate summary call the default makes, and the note is shorter, tasks finish sooner: measured end to end, 406 seconds instead of 464.

Try it with nothing installed:

  npx github:LippInc/catnap

It prints how often auto-compact cut in mid-task in your own logs (reads numbers only, sends nothing).

Install (needs Claude Code 2.1.287 or later, where mods are on by default):

  claude plugin marketplace add LippInc/catnap
  claude plugin install catnap@lippinc

Currently Claude Code only, tested on Windows and Linux. The memory test used made-up coding tasks on a 200K window.

Some related tools: compact-adviser, handoff-compact, session-pilot and sub-agent-compact. What catnap adds: Claude finishes its piece first, and the numbers on where the cut lands.

Norm Brandinger on the XERJ Corpus Hub

SSH Keys on macOS with Touch ID

Building Persistent Agents

Voxtral 3

The number of monthly active Mastodon users is mostly going down

Making a VGA Driver in SystemVerilog

Why scraping isn't enough for AI agents and how DNS-verified profiles could help

Ask HN: As developers from low-income countries,how to cope with AI replacement?

Sixteen Years After Deepwater Horizon, Deep Sea Life Shows Few Signs of Recovery

Add Honest Achmed's root certificate

Show HN: Agent Archive – Session Handoffs and Backups for Claude Code and Codex

Effective Development (2013) [video]

I Took a Shower with My Personal Assistant

Roll your own, fork the rest

I used celestial vision navigation to solve a mystery place in Elite:Dangerous

24x faster Rust compilation with rust-analyzer

Build a Discord FAQ and Feedback Bot with Google Sheets

Steam Users Beg for "No AI" Filter as Vibecoders Flood Next Fest

Show HN: BodyCommit – movement breaks in VS Code while your agent works

OpenDE: Play Deus Ex 1999 demo in the browser

Should I build an MCP server or chat inside my app?

No Kings, Just a Few Techno Lords

AI Pricing Is About to Shock Everyone: Why the $20/Month Era Is Ending

Show HN: A layer on top llama.cpp enabling big MoE streaming on phone and laptop

12.8B destructible voxels rendered in the browser with Rust WASM custom engine

UK to deploy jets to defend Saudi Arabia against Houthi attacks

U.S. SMR stocks slump as short sellers reap $2B profit

IKEA – 3D Guide

Show HN: Stop Claude Code from compacting mid-task

Ask HN: How do you find meaningful peer discussions?

Comments