Context Compression Plugin 'Fast Jev Compaction' for Claude Code Released

Individual developer Tamara Tran has released the 'Fast Jev Compaction' plugin, which improves the context compression method of Claude Code. While the default compression in existing Claude Code involved the LLM rewriting the entire conversation into a summary, risking the loss of details such as file paths and error messages, this plugin uses TypeSafe AI's Jev model to preserve the original text and delete only unnecessary tool call results.

How It Works and Technical Features

The plugin takes the conversation history as input and evaluates the importance of each tool call and the irreplaceability of its result with a probability between 0 and 1. Calls that do not meet the keepThreshold (default 0.5) are deleted, and if partially important, only the beginning of the result (truncateHeadChars, default 300 characters) is retained while the rest is truncated. Text messages from the user and assistant are kept as-is.

Safety Mechanisms and Application Conditions

A Fallback mechanism is built in to automatically switch to the existing summary method if the Jev request fails or the reduction amount is less than minReductionRatio (default 25%). This tool is suitable for sessions with frequent tool calls and long results, such as code exploration or build logs, but is unsuitable for environments where external API calls are prohibited or for text-heavy conversations. It requires the early access function hook feature of Claude Code 2.1.274+ and the mandatory configuration of the TYPESAFE_API_KEY environment variable.