Test Coverage Gap Finder

What this skill does

Parses an lcov or cobertura coverage report, joins it with a static call graph extracted from the source, and produces a ranked list of "test these next" targets. Targets that sit on hot paths (high in-degree) and have low coverage rank first.

Inputs

• coverage_path: path to coverage/lcov.info or coverage/cobertura.xml.
• repo_dir: source root used to build the call graph.
• language: one of js, ts, py, go. Drives which static-analysis CLI to run.
• Optional top_n (default 20).

Steps

1. Detect the report format by reading the first line; lcov starts with TN: or SF:, cobertura is XML.
2. For lcov, parse SF: (source file) and DA:<line>,<hit> records into {file, line, hits} tuples.
3. Identify uncovered lines (hits == 0) and group them by enclosing function. For JS/TS use ast-grep --pattern 'function $F($$$) { $$$ }'; for Python use python -m ast; for Go use go/parser.
4. Build a call graph: for JS/TS run madge --json src; for Python run pyan3 --json; for Go run go-callvis -format json.
5. Compute in-degree per function (number of distinct callers) from the call graph.
6. Score each uncovered function as score = log(1 + in_degree) * (uncovered_lines / total_lines).
7. Sort descending by score, take top N.
8. For each entry emit file:line — function — score — sample callers.
9. Write the markdown report to stdout.

Output

A markdown report with a header table (totals, average coverage, hottest uncovered file) and a numbered list of N suggested test targets. Each entry includes a fenced snippet of the function signature pulled from source.

Verification

After ranking, sample-check the top three entries by re-grepping the source with rg -n '<function name>' <file> and confirming the line numbers match the lcov report. Recount hot-path callers by re-running the call-graph CLI on just those files. If any sampled entry has zero callers in the static graph, demote it (it's likely dead code, not a coverage gap).

Edge cases

• Empty coverage report: exit with a single warning "no coverage data" and skip ranking.
• Mixed-language repo: run the analyzer for the requested language only; skip other extensions.
• Generated code paths (e.g., dist/, build/): exclude via a default ignore list before scoring.
• Report with line counts but no function metadata: fall back to file-level ranking and note the degradation in the report header.