Agent tooling
A three-part Claude Code plugin uses Jev to retain or discard old tool results, choose model routes from project rules, and shortlist relevant skills.
Only you can see your notes.
Screenshot unavailable. Open the experiment ↗
Context removal can discard useful information, and model confidence does not guarantee correctness. Relevant state goes to TypeSafe. Skill shortlists can miss useful skills; fallback can preserve the session or strongest model instead of savings. Cost statistics use estimated comparisons and a fixed price constant, not a verified current invoice.
In a long Claude Code project session, inspect how often context shortening ran, what it did overall and its estimated cost. These statistics do not list each removed result. Separately check the retained conversation against information you still need: an earlier file listing may be stale after edits, while an old test failure may still matter.
Model routing and skill shortlists are optional, initially off. Set project preferences before enabling them and inspect their explanations. Relevant session state goes to TypeSafe; a shortlist may omit useful guidance. Fallback may retain the session or stronger model, so neither removal nor lower cost is guaranteed.