What problem does it solve?
The fixer consumes evaluator outputs and turns them into a deterministic, action-ready repair plan, gating costs and ensuring no paid-model calls occur before user approval.
Core Features & Use Cases
- Reads eval output (eval-deep-vision.json or eval.json) to classify findings by ownership (art-director / scenarist / editor) and orders them by severity.
- Generates and presents a grouped repair plan via the repair-plan verb, then waits for user approval before applying fixes through existing ralphy verbs, re-renders, and re-evaluates.
- Provides a safe-guarded workflow for batch or single-item repair with append-only versioning and zero-model-invocations until approved.
Quick Start
Run the evaluator to generate a report, then use ralphy project repair-plan <id> to generate a repair plan and apply approved fixes via the existing ralphy verbs.