fast-jev-compaction: A compressor that removes only stale tool results, preserving the original text without summarization

When reducing long conversation logs, the LLM summarization approach easily misses details such as file paths, exact error messages, and specific constraints. This tool does not generate summaries; instead, it deletes or truncates only unnecessary tool calls and results as determined by the Jev model. The remaining messages are kept exactly as they are, minimizing information loss.

The entire conversation state is sent to Jev to score the necessity of each tool call. Items that do not exceed the threshold are removed or truncated to keep only the beginning. The latest and first messages are left untouched, and if the token limit is exceeded, the state is progressively condensed.

It can be integrated directly as an npm library or installed as a Claude Code plugin to replace the default compression functionality. A TypeSafe API key is required, and if Jev fails to reduce sufficiently or fails, it automatically falls back to the existing summarization method. It is suitable for agent development requiring precise control or tasks where long context management is critical.