Claude Code prompts routed by Jev to pick model per turn

The plugin is a personal project rather than a product release, and its routing logic is deliberately simple. Jev does not pick a model; it only returns one of six tier labels (micro, low, medium, high, critical, max). The plugin then looks up that label in a hardcoded table to decide the model and effort level. This separation means the classifier's job is narrow, and the developer can adjust routing by editing the table rather than retraining anything.

The one exception in the code—switching from Haiku to Sonnet when a conversation passes 150,000 tokens—suggests the developer is guarding against a small model being kept on a long context where it might struggle. The plugin also avoids touching conversation history or memory files, which likely keeps it from disturbing Claude Code's normal session behavior.

Beyond cost and convenience, the developer sees potential in security uses such as preventing prompt injection, though no implementation or timeline is described.