auto-review-loop

An autonomous review loop automates iterative refinement by scoring research artifacts, applying minimal fixes, and re-evaluating until a positive threshold is reached. It reduces manual review time by persisting state across rounds and recording reviewer responses automatically.

The iterative review loop uses Codex MCP as an external reviewer to score artifacts and propose fixes, while an executor model implements the changes. It persists state across rounds in REVIEW_STATE.json and records responses in AUTO_REVIEW.md until acceptance or maximum rounds.

Yes, this setup supports cross-model review configurations where Codex MCP serves as the reviewer and an external model acts as the executor. It enables autonomous review-implement cycles with optional human checkpoints when enabled.

When MAX_ROUNDS are reached without meeting the POSITIVE_THRESHOLD, the loop terminates. The detailed state and reviewer responses remain persisted in REVIEW_STATE.json and AUTO_REVIEW.md for human inspection and intervention.

The autonomous review loop applies to multi-round review scenarios for academic manuscripts, code, and experiments. It handles any technical project artifacts where an executor and external reviewer collaborate across iterations for iterative refinement.