GitHub radar
Jev, smart context compression for Claude Code plugin
An alternative compaction implementation for Claude Code that uses Jev (a TypeSafe model) to decide if tool invocations are still necessary. Rather than summarizing previous tool invocations with a lossy summary, it evaluates each previous invocation and preserves necessary information exactly.
This is an npm package and Claude Code plugin. Rather than using Claude Code’s default compaction implementation, which summarizes previous tool invocations with a lossy summary, this plugin forwards the entire conversation history to Jev, a TypeSafe model, and asks if each previous tool invocation is still necessary. Previous tool invocations that are still necessary are preserved exactly in context. Previous tool invocations that are unnecessary are dropped or summarized with a brief note. Neither user nor assistant messages are ever dropped or rephrased. This provides a more accurate context window for lengthy conversations where specific file paths, error messages, and previous commands are important.
Why a vibe-coder should care
Claude Code’s default compaction implementation summarizes previous tool invocations with a lossy summary, leading to inaccuracies in lengthy conversations. Using this plugin, you can preserve necessary information exactly, reducing the chance that the agent will repeat a step, refer to the wrong file path, or forget about existing constraints.
How to install
Copy this and send it to your agent — Claude Code, Codex, any of them:
Install this Claude Code plugin from https://github.com/tamaratran/fast-jev-compaction — follow the README, set the TYPESAFE_API_KEY variable and enable the compaction hooks, ask me if you need keys.
Any computer with Claude Code installed, plus a TypeSafe API key — all processing happens in the cloud via Jev
Open on GitHub▌ More finds