What problem does it solve? Raw git diff output mixes hand-written source changes with regenerated lockfiles, rebuilt bundles, and snapshot churn, forcing reviewers to read every hunk in file order with no signal about which ones are dangerous. diffscope classifies files by category, excludes generated noise by default, and ranks hunks by a deterministic, local, rule-based risk score so the hunks that matter surface first. ## Core Features & Use Cases - File classification: Categorizes files as source, lockfile, generated, snapshot, manifest, ci, test, docs, binary, conflict, or submodule, excluding generated noise by default. - Risk-ranked hunks: Scores every hunk with documented integer rule weights (path keywords, deletion-heavy, large hunk, CI file, manifest file, deleted test) and lists them highest-risk first with concrete reasons. - Auditable and offline: The scorer never calls an LLM or network; scores are reproducible byte-for-byte, and diffscope categories exposes the glob table for auditing. - Use Case: Before reviewing a pull request that touches source code plus a regenerated package-lock.json, run diffscope hunks --limit 5 to read only the five highest-risk hunks instead of all 40. ## Quick Start Ask the agent to run diffscope on the current working tree and show the highest-risk hunks before reviewing the change.