review-code

Automate PR code review by orchestrating subagents to generate deterministic, question-driven quality checks from git diffs and commit messages. The process classifies questions by type, routes them to specialized subsystems, and produces a synthesized REVIEW.md with supporting planning artifacts for human action.

A question-driven code review generates a backlog of questions from PR diffs, classifying them into FACT, RESEARCH, DESIGN, CORRECTNESS, and INTENT categories. Specialized subagents process each question type per commit, culminating in a final synthesized REVIEW.md that captures design considerations and rationale.

The code review assistant requires access to PR metadata, git diffs, and the repository's planning workspace. It is designed to work across Bottlerocket forest PRs, standardizing reviews by producing local planning artifacts in the planning/ directory that you can inspect and act on.

Generate a REVIEW.md from a git diff by feeding PR metadata and commit messages to the review workflow. Subagents orchestrate the creation of structured artifacts, analyzing per-commit correctness checks and design considerations, then summarize everything into a human-actionable final review document.

Standardize code reviews across multiple PRs by applying a deterministic, question-driven review process that generates consistent planning artifacts. This approach captures rationale and design considerations per commit, outputting a structured REVIEW.md to ensure every PR undergoes identical quality and correctness checks.

Orchestrating subagents for PR analysis is limited by its dependency on accessing the repository's planning workspace and PR metadata. Without the required git diffs and planning context, the subagents cannot generate the structured artifacts needed to produce the final synthesized REVIEW.md.