Итеративное рецензирование исследований через MiniMax

★ 7.1 · research

auto-review-loop-minimax is a Claude Code skill that autonomously runs a multi-round research review cycle using MiniMax API with MiniMax-M3 as the external reviewer. The loop follows a review → fix → re-review pattern for up to four rounds (MAX_ROUNDS = 4), stopping when the score reaches ≥ 6/10 and the verdict is "ready" or "almost" — both conditions must hold simultaneously. Two API access methods are supported: the mcp__minimax-chat__minimax_chat MCP tool as the primary path, and a direct curl fallback using MINIMAX_API_KEY. State is persisted to review-stage/REVIEW_STATE.json after each phase, enabling recovery from context compaction or unexpected interruptions. This skill is the practical alternative to Codex MCP for iterative improvement of ML papers and research artifacts, since Codex CLI relies on OpenAI's /v1/responses endpoint not supported by third-party providers.