Автоматический цикл рецензирования исследований

★ 7.1 · research

auto-review-loop is a Claude Code skill that autonomously runs a multi-round research review cycle — review → implement fixes → re-review — until an independent reviewer issues a policy-approved positive assessment or MAX_ROUNDS (4) is reached. It supports multiple reviewer backends: Copilot CLI sessions use a native rubber-duck subagent, other hosts default to Codex (model gpt-5.6-sol), with oracle-pro and Manual Review also available as explicit overrides. The loop stops only when both conditions hold simultaneously: score ≥ 6/10 AND verdict is "ready" or "almost". All outputs accumulate in review-stage/AUTO_REVIEW.md and are auto-rendered to HTML on termination. Reviewer difficulty is configurable (medium / hard / nightmare), and optional human_checkpoint mode lets you inspect scores and approve fixes between rounds.