Bob the Reviewer: See Beyond the Diff

A small code change can affect functions in files that were never modified. Reviewers naturally focus on the diff, while passing tests only cover the cases someone thought to write. This leaves an important question unanswered: what behavior changed elsewhere, and was that change intentional? Bob the Reviewer helps developers answer three questions: What else might this change affect? Did the same inputs produce different outputs? Was the difference intentional? The tool traces callers of changed functions up to two calls away, highlighting affected code outside the diff. For Python projects, it runs the same committed probes, function calls with fixed inputs, against both Git revisions and compares their outputs. It also compares test results and clearly reports missing coverage, unresolved references, and execution limitations. When a difference appears, a developer records it as intended, unintended, or unresolved, with a rationale. That decision becomes approved through the team’s normal review and merge process. Earlier approved decisions provide context for future reviews without automatically approving new changes. Our target users are developers, pull request reviewers, and teams working with AI coding assistants. Developers can run a review from the terminal before opening a PR, use /behavior-review inside IBM Bob, or explore saved reviews through a local web interface. GitHub Actions posts a review comment on the pull request and updates it as new commits arrive, keeping evidence beside the review discussion. What makes the project distinctive is its connection between impact analysis, reproducible behavior comparisons, and human intent. It goes beyond identifying changed lines: it connects a change to affected callers, concrete input/output differences, and a recorded explanation of whether those differences belong. IBM Bob provides a conversational workflow for interpreting findings and proposing missing probes. Deterministic analysis and execution supply the evidence, while people retain responsibility for decisions. The principle is simple: AI proposes. Algorithms verify. Humans decide. The result is a focused, repeatable review process that helps teams investigate behavior changes without treating incomplete evidence or passing tests as proof of safety.