optim-jev plugin uses Jev to decide what Claude Code forgets
An experimental plugin has Jev decide which old tool calls to keep or prune, which model and effort to use, and which skills to load; benchmarks are still pending.
Someone put another AI inside Claude Code to decide what it should forget. optim-jev is an experimental plugin that uses Jev for three decisions: • Keep, truncate, or remove older tool calls and outputs while protecting recent messages. • Choose a model and effort level across Haiku, Sonnet, and Opus. • Select relevant skills for the current prompt. The pruning runs before Claude Code’s built-in compactor. The interesting part: Jev needs its own context management too. The author caps its input, trims lower-priority history first, and caches repeat decisions. No final benchmark conclusions yet. The author is still testing token usage, decision quality, cost, and interactions with prompt caching. Would you trust another model to decide what your coding agent forgets?