The idea
Tamara demonstrates context compaction driven by Jev decisions. The accompanying fast-jev-compaction project keeps relevant tool calls and results, truncates or drops stale ones, and preserves retained conversation content verbatim.
Inside the decision loop
Conversation + paired tool results
Score which calls and results to keep
Rebuild a smaller conversation
What to take from it
Treat compaction as a selection problem when exact filenames, errors and commands matter. Keep recent context pinned and preserve the relationship between each tool call and its result.
The project documents a fallback when scoring fails or removes too little context. A relevance score does not guarantee that a discarded result will never be needed again. The repository also includes a scripted visual demo that does not call the API.
Follow the original work
This case is an editorial interpretation of a public community demonstration. It has not been reproduced or benchmarked by this publication.
Original post by @tamarajtran