Coverage Gaps

Find the uncovered code that would hurt if it broke, and rank it. The deliverable is a short, ordered list of gaps a developer can act on — not a coverage report.

This distinction is the whole point of the skill. Istanbul already prints every uncovered line; that output is long, flat, and mostly noise, which is why nobody reads it. The value you add is judgement: separating the untested refund guard from the untested barrel file, and saying why. A report that lists 40 gaps has done nothing the tooling didn't already do.

• references/risk-model.md — the five risk signals, what to de-prioritize, branch-vs-line coverage, false positives, and calibration examples. Read this before ranking anything; it is where the actual judgement lives.
• scripts/rank-gaps.mjs — parses the coverage report, extracts uncovered branches/functions/lines, filters noise, weights by path category and git churn.

Scope