November 2026 Summaries

OpenRouter’s guide describes a 233-line Python AI code-review script that examines Git changes for bugs and security issues by uploading changed files, context files, and diffs to a hosted Linux sandbox, where models can run checks to confirm findings and generate V4A patches. The reviewer uses the Responses API with Files, Shell, Apply Patch, and structured submit-review tools; it records confirmed findings with command evidence, labels non-reproducible issues as unconfirmed, saves reports in `.git/ai-review.json`, optionally applies tested patches, and returns exit codes suitable for blocking pushes or CI workflows. It can run locally, as a Git pre-push hook, or in GitHub Actions, with the workflow designed to use the base branch’s reviewer script and avoid sending forked pull-request code to the service. Tests on a sample Python change containing five planted bugs found that Claude Sonnet 5.5 consistently produced patches fixing all issues, while Claude Haiku 5.5 was cheaper but less reliable; reviews generally took one to four minutes and cost roughly $0.015 to $0.045, including sandbox time. The guide cautions that uploaded code is shared with OpenRouter and the selected model provider, the beta tools work only through the global endpoint, reviews are limited to 19 source or context files plus the diff, sandbox internet access is disabled by default, and human review remains necessary because models may miss issues, misjudge severity, or produce faulty patches.