FILTERAgents & automation
fast-jev-compaction
apikey_official@apikey_official
0:05This project provides a Claude Code plugin that avoids betting on future needs by scoring tool calls/results as KEEP or DROP with a calibrated probability threshold, ensuring kept content remains verbatim and falling back to normal summarization if safe trimming isn't possible.
Context compaction has a quiet flaw: every AI summary is a bet that it guessed right about what you'd need later.
A new Claude Code plugin skips the bet entirely.
fast-jev-compaction (MIT, built on TypeSafe's Jev model) never rewrites your session:
🔹 Every tool call/result scored KEEP or DROP
🔹 Calibrated probability threshold, not a vibe check
🔹 Kept content stays 100% verbatim
🔹 Falls back to Claude Code's normal summary if it can't trim safely
Demo run: a 156k-token session compacted to ~62k — user goals, error reports and original code untouched, only stale reads and dead-end noise cut.
Small project (23★), genuinely different approach to context.
x.com/tamarajtran/status/21006945493625531
github.com/tamaratran/fast-jev-compaction
#ClaudeCode #Jev #AIAgents #ContextEngineering #OpenSource #TypeSafeAI