autoresearch-v2

Turns the Claude Code main thread into an event-driven coordinator for autonomous optimization. Use when the user asks to autonomously improve something measurable: optimize a file by a metric, run experiments on code, iteratively reduce loss / latency / bundle size, raise pass rate, improve a prompt eval score, "run experiments on this", "make this faster/smaller/better autonomously", "find a better version", "keep iterating until I stop you". Works for ML training (val_loss, val_bpb), API performance (p50_ms), bundle size, prompt pass-rate, LLM-judged quality — for ANY metric that reduces to a single shell command printing one parseable scalar. Also triggers on phrases like "run autoresearch", "start the experiment loop", "optimize X by metric Y", "iterate on this autonomously".