ivcained

Picks the reasoning effort for every prompt: a fast judge (Jev/TypeSafe, a local OpenAI-compatible model, or any custom API) reads the next message once per turn and rewrites the request's reasoning effort before it ships — easy questions run cheap, hard jobs get high. Port of effortless for Claude Code. Disclosure — sends the next message (≤4000 chars) plus a short recent-conversation excerpt (≤1600 chars) to the configured judge endpoint (default api.typesafe.ai for Jev; a local or custom OpenAI-compatible API otherwise), solely to pick the effort; no telemetry; secrets resolved only through Hermes' secret scope; stdlib-only, no pip dependencies.