auto-review-loop

Automated iterative peer review repeatedly obtains an external review, implements the minimum recommended fixes across code and documentation, and re-submits the updated work for re-assessment until a score threshold is met.

You set up an autonomous review loop by defining YAML parameters for round limits and stop conditions, then running the process to parse reviewer verdicts, execute fixes, and re-check progress until submission-ready.

Yes, you can route reviewer verification via codex exec for adversarial repo-reading scrutiny, or use Codex MCP for review with thread persistence to match different risk profiles.

The best way to ensure consistency is using deterministic, execution-grounded repair that applies concrete changes across experiments, metrics, analysis, and figures rather than only revising the narrative.

State recovery works by persisting progress in REVIEW_STATE.json and appending verbatim reviewer outputs to cumulative logs, maintaining full traceability and allowing the review loop to resume from the last saved round.

The limitation is that reaching the MAX_ROUNDS threshold stops the loop regardless of acceptance status, meaning work that fails to earn a positive decision within the round limit will exit the review cycle unfinished.