fast-jev-compaction is a Claude Code plugin that replaces context compaction with lossless message pruning based on relevance scoring.
The problem it solves is that traditional context compaction summarizes old conversation turns, which loses exact details like file paths, error messages, or constraints that may matter later. This tool takes a different approach: it never rewrites anything. Instead, it uses a language model to score each tool call and its result for relevance, then deletes only those marked as no longer needed while keeping everything else verbatim and in original order. The full conversation is shown to the scoring model, with tool results replaced by brief notes, so decisions are made with complete context.
The tool works by pairing each tool call with its result, pinning recent messages to preserve them, and fitting the conversation into token limits through staged truncation strategies. It sends the full conversation state to the scoring model with two questions per tool call: whether the call itself should stay and whether its result should be kept verbatim or truncated. Decisions are made against a configurable threshold, and messages that lose all content are removed while untouched messages are returned unchanged. The project is available both as an npm package and as a Claude Code plugin. Anyone adopting it should know that it requires careful configuration of token limits and thresholds, and that it will throw errors rather than silently degrade if the conversation cannot be fitted within constraints. It suits teams using Claude Code who need to preserve exact technical details across long conversations without losing information to summarization.
The project shows active maintenance with recent commits addressing core functionality. Development includes both the plugin implementation and the underlying library, with careful attention to token estimation and the mechanics of message reconstruction. The codebase handles edge cases like malformed responses and missing keys by throwing errors that let callers decide on fallback behavior.
