fast-jev-compaction

Claude Code compaction usually asks an LLM to summarize the chat. That is slow, and the summary can drop a path, error, or constraint.

fast-jev-compaction scores every tool call instead. Jev answers how important the call and its result still are. Above the threshold the original text stays; below it the result is stubbed or the pair is dropped. Nothing is rewritten.

The author measured about $0.04 per million tokens of context and noted the limits: Jev sees the calls more than the full outputs, its 32k window means long chats are chunked, and this is a pre-filter rather than a full replacement.

How it works

1. Pair each tool_use with its tool_result.
2. Ask Jev, in one or more concurrent requests, whether the call and the result still matter.
3. Rebuild the message list. Kept text is the original text.

The Destination is the plugin repo.