Claude Code chat compaction handed to probabilistic model Jev

The author of the Tips AI channel tested the open-source plugin fast-jev-compaction, which replaces lossy LLM summaries during chat compaction in Claude Code with probabilistic estimates from TypeSafe AI's Jev model.

Compiled by Sergey KostenchukPublished 2026-09-18Updated 2026-09-18

2026-09-18 Research Anthropic

🤖 Claude Code chat compaction handed to probabilistic model Jev

The author of the Tips AI channel tested the fast-jev-compaction plugin (MIT, 1739 stars): instead of lossy LLM summaries, Jev evaluates each tool_use and tool_result on two typed questions and removes outdated calls, while what is saved remains verbatim.

🌍 Context compaction is an expensive bottleneck in agentic systems: for the first time a probabilistic classifier is used instead of a generative LLM summary, $0.042 per million input tokens, cheaper, faster, and without losing verbatim messages.

👤 The plugin can be installed today: CLAUDE_CODE_ENABLE_FUNCTION_HOOKS=1 in ~/.claude/settings.json and installation from the marketplace. Free $5 is enough for ~125M input tokens, but 150k+ sessions will have to be cut due to the 32k state limit. The evidence base is a single test by the author and vendor documentation.

Source 1: https://github.com/tamaratran/fast-jev-compaction Source 2: https://docs.typesafe.ai/introduction

Sources