evals-code-run

Run the evals on the current branch; it diffs HEAD against the upstream (or a base ref you pass as an argument) using a three-dot diff, then scores the added lines for duplication, bloat, inefficiency, and security. Results are appended to ~/.claude/evals/code.json.

The duplication eval picks distinctive identifiers, string literals, and logic shapes from the added code and greps the rest of the repo for matches, counting near-copied blocks of roughly four or more contiguous lines. Each finding cites both the new location and the pre-existing code it duplicates.

No. The skill is strictly read-only with respect to the repo: it never modifies, stages, commits, checks out, or pushes project files. Its only write is the eval log at ~/.claude/evals/code.json, plus a timestamped backup if that log is corrupt.

Without an upstream or origin/HEAD, the skill stops and asks you to re-run with an explicit base ref. If the diff against the base is empty, it stops without writing anything to the log, since a run that evaluates nothing must not produce entries.

No. Only committed-but-unpushed changes are evaluated by design; uncommitted working-tree changes are mentioned in the summary but excluded from scoring. The expected flow is that each implementation stage is committed before evals run.

Machine-written lines such as lockfiles, minified bundles, and build output say nothing about the quality of the authored change and would dilute every score toward zero. The denominator counts only added lines in non-excluded files, and exclusions are recorded in the summary.