fast-jev-compaction

Canonical version: fast-jev-compaction.

fast-jev-compaction is an MIT-licensed plugin for Claude Code by Tamara Tran that replaces summary-based compaction with Jev decisions. The question behind it is a good one: why is compaction still a summarization prompt in 2026?

Instead of asking an LLM to rewrite old messages into a summary (slow, and you lose details), it:

1. Scores every tool call/result pair with Jev in a single request
2. Decides per pair: keep, truncate or remove
3. Compresses progressively if needed (truncate inputs, abridge long texts, then collapse old messages)
4. Rebuilds the original message list with the deletions applied

Everything that's kept stays verbatim. Nothing gets rewritten. If Jev's decisions don't free enough space, it falls back to the built-in summarization.

The results made it go viral: Alex Volkov went from a session of nearly 1M tokens to 86K in about one second. The repo passed 6k stars within days.

I like this one because it shows exactly where System One Models fit. "Is this tool result still relevant?" is a bounded question asked hundreds of times per session; it doesn't need a writer, it needs a fast judge. It's also a neat example of Context Engineering: keeping the Context Window useful rather than just shorter.

References

• fast-jev-compaction on GitHub
• Alex Volkov's demo on X

Related

• Jev
• Claude Code
• Context Engineering
• Context Window

About Sébastien

Ready to get to the next level?