Claude Code review

Reproducible practice

Review a controlled diff before allowing any automatic fix

Freeze the intended diff, run a read-first local review, reproduce findings, and decide fixes separately.

Prerequisites

• A trusted repository and known base branch.
• Tests that can reproduce the changed behavior.
• No unrelated uncommitted work in the review scope.

Steps

1. Freeze scope evidence

   Record commits and working-tree changes before asking for review.

   git status --short
   git diff --stat
   git diff --check

2. Run the local review read-first

   Invoke `/code-review` without `--fix`; model output and cost were not exercised in this audit.

3. Reproduce each candidate

   Open cited lines, run the smallest relevant test, and reject findings that do not survive code evidence.

4. Fix in a separate step

   Apply only accepted findings, rerun targeted and broad gates, then compare the final diff with the frozen scope.

Expected result

Every accepted finding has a cited location and reproducible failure or strong code-path proof; review output alone does not modify the tree.

Validation

• Compare pre- and post-review `git status --short`.
• Record the command or code path that confirms each accepted finding.
• Run repository tests after fixes, not merely another model review.

Failure handling

Cleanup or rollback

• Discard only reviewed automatic edits after inspecting the patch; preserve unrelated user work.
• Remove temporary review notes that contain private code or identifiers.

Boundaries and when not to use it

• Managed Code Review is a research preview with plan and provider limits documented by Anthropic.
• A clean review is evidence from one method, not proof that the change has no defects.