fast-jev-compaction

fast-jev-compaction replaces the usual context-compaction summary with decisions made by a model called Jev. Instead of asking an LLM to summarize old conversation turns, which can lose an exact file path, error, or command, it only deletes tool calls and tool results that Jev says are no longer needed; everything else, including all user and assistant text, stays verbatim and in order.

Every tool call is paired with its result, and calls in the first message or the most recent pinned messages are never touched. The conversation state is fitted into a token budget in stages: truncating tool inputs, abridging long texts, collapsing old non-pinned messages, and reducing old tool calls to one line each. For every remaining call, Jev is asked whether the call and its result should still be kept, split across as many requests as the token budget requires. It ships both as an npm library and as a Claude Code plugin that uses it to replace the built-in compaction summary.