Designed around small local models first, so it is careful with context, fails more gracefully, and says when it's stuck. It doesn't make a small model smarter. It makes working with one less frustrating.
A real local run of Qwen3.6 35B A3B (Unsloth UD-IQ2_M) via llama.cpp, sped up 3×. The model's first test fails and it fixes it from the error.
npm i -g @alexwkleung/reika
pnpm add -g @alexwkleung/reika
brew install alexwkleung/tap/reikaThen point it at a model server and run reika in a project. Install and first run covers requirements and setup.
Old tool output collapses to short summaries and requests stay append-only, so a slow local engine's prompt cache survives.
Repeated reads and runaway reasoning are detected and stopped, ending in an honest stop instead of a long spiral.
Blind edits go back for a read first, TypeScript edits are typechecked, and plan progress is tracked from what actually happened.
Plan before anything changes, put a harder change through a test-and-review procedure, keep a small window lean, or just chat. Shift+Tab switches.
Dangerous commands still prompt, and on macOS shell commands run in a kernel sandbox. No telemetry.
Default-on features keep an off switch for A/B runs, and what they did and didn't fix is published.