Автоматическое итеративное исследование с оценщиком

★ 8.5 · vibe-coding

autoresearch is a Claude Code skill that runs a stateful, single-mission improvement loop driven by a strict evaluator contract, with markdown decision logs and a max-runtime stop condition. On each iteration it executes one experiment cycle, invokes the evaluator — which must return structured JSON with a required boolean `pass` field and an optional numeric `score` — then writes machine-readable evaluation JSON and appends a human-readable entry to a decision log under `.omc/autoresearch/<mission-slug>/runs/<run-id>/`. The mission definition and evaluator must be created beforehand via `/deep-interview --autoresearch`; autoresearch itself does not generate them. Claude Code native cron is a supported integration point for periodic reruns, with each run appending new artifacts rather than overwriting prior ones. It is the right choice for bounded, evaluator-driven experimentation that needs durable logs, resume support, and explicit termination semantics.