Turns your codebase into an autoresearch loop that discovers what to measure, sets up the benchmark, and runs tree search with parallel subagents.