lcov-analysis

Overview

LCOV is "a tool suite for manipulating and displaying code coverage information" with three command-line utilities: geninfo (creates LCOV data files from raw coverage data), lcov (captures, filters, manipulates, processes in parallel), and genhtml (HTML report generation) (lcov-readme (opens in new window)).

The toolchain is "language-agnostic (via converter scripts: llvm2lcov, py2lcov, perl2lcov, xml2lcov)" (lcov-readme (opens in new window)) - LCOV .info is the lingua franca every coverage UI (Coveralls, Codecov, Codacy, SonarQube, in-house dashboards) ingests.

This skill covers parsing the .info text format directly so the team can gate PRs without running the full HTML generation step.

When to use

• The CI already emits LCOV (or can via a converter) and the team wants PR-time per-file coverage gating.
• A coverage SaaS isn't an option (compliance, cost, air-gapped CI).
• A multi-language project needs one analyzer (Python via py2lcov, Java via xml2lcov, Node via nyc/jest's lcov reporter, C++ via gcov → lcov).
• A custom coverage UI / Slack bot needs structured input.

How to use

1. Confirm CI emits LCOV .info (natively or via a converter: py2lcov, xml2lcov, nyc / Jest's lcov reporter, gcov → lcov).
2. Parse the .info into per-file line / function / branch metrics (Parse below); recompute totals from the per-line records rather than trusting the summary fields.
3. Cache main's .info as the baseline; diff each PR's coverage against it.
4. Gate per file on the three rules below (whole-repo drop, per-file drop, new-file minimum) - the diff + gate reference implementation, the full record-format table, and the CI wiring are in references/format-diff-and-ci.md.

Worked example

A single source file's .info record, and how to read it:

TN:
SF:src/checkout/cart.ts
FN:10,addItem
FN:32,removeItem
FNDA:42,addItem
FNDA:0,removeItem
FNF:2
FNH:1
DA:11,42
DA:12,42
DA:13,0
DA:33,0
DA:34,0
LF:5
LH:2
BRDA:13,0,0,42
BRDA:13,0,1,0
BRF:2
BRH:1
end_of_record

Reading: cart.ts has 2 functions, 1 hit (50% function coverage); 5 lines, 2 hit (40% line); 2 branches, 1 hit (50% branch). removeItem was never called (FNDA:0). SF opens the record, DA:<line>,<count> gives per-line hits, LH/LF and BRH/BRF are the roll-ups, and end_of_record closes the file. The full record-keyword table is in references/format-diff-and-ci.md.

Parse

# scripts/parse_lcov.py
from collections import defaultdict

def parse_lcov(path):
    files = []
    cur = None
    with open(path) as f:
        for line in f:
            line = line.strip()
            if line.startswith('SF:'):
                cur = {
                    'path': line[3:],
                    'functions': [],
                    'lines': {},
                    'branches': defaultdict(list),
                    'fnf': 0, 'fnh': 0,
                    'lf': 0, 'lh': 0,
                    'brf': 0, 'brh': 0,
                }
            elif line.startswith('FN:'):
                lineno, name = line[3:].split(',', 1)
                cur['functions'].append({'line': int(lineno), 'name': name, 'hits': 0})
            elif line.startswith('FNDA:'):
                hits, name = line[5:].split(',', 1)
                for fn in cur['functions']:
                    if fn['name'] == name:
                        fn['hits'] = int(hits)
                        break
            elif line.startswith('DA:'):
                lineno, hits = line[3:].split(',', 1)
                cur['lines'][int(lineno)] = int(hits.split(',')[0])  # checksum optional
            elif line.startswith('BRDA:'):
                lineno, block, branch, taken = line[5:].split(',', 3)
                cur['branches'][int(lineno)].append({
                    'block': int(block),
                    'branch': int(branch),
                    'taken': 0 if taken == '-' else int(taken),
                })
            elif line.startswith(('FNF:', 'FNH:', 'LF:', 'LH:', 'BRF:', 'BRH:')):
                key, val = line.split(':', 1)
                cur[key.lower()] = int(val)
            elif line == 'end_of_record':
                files.append(cur)
                cur = None
    return files

Don't trust the FNF/FNH/LF/LH/BRF/BRH summary fields blindly - some buggy emitters produce summaries that don't match the per-line data. For correctness, recompute from lines, functions, branches.

Gate rules

A defensible per-file gate has three rules:

1. Whole-repo line coverage MAY drop by at most N pp (typically N = 0.5 - guards against runaway erosion without blocking refactors).
2. No file MAY drop more than M pp (typically M = 5 - calls out the specific file that lost coverage).
3. New files MUST hit threshold (typically 80% line, 70% branch).

Per-file gates beat whole-repo gates: an aggregate drop hides which file caused it; per-file output gives the reviewer a direct target. The coverage_diff + gate reference implementation and the PR coverage-gate CI job are in references/format-diff-and-ci.md.

Anti-patterns

Limitations

• .info is text, not standardized via a formal spec. The authoritative source is the lcov-readme (opens in new window) + the codebase itself; format extensions vary by emitter. Tolerant parsing is required.
• No file-level "this is a test file" marker. Filter by path convention.
• No PR-context awareness. The format doesn't know about git diff - pair with git diff --name-only to scope coverage changes to PR-touched files when needed.
• Branch coverage shape varies. Some emitters report per-condition (multi-arm BRDA); some report per-decision (single arm). Normalize before cross-tool comparisons.

References

• lcov-readme (opens in new window) - LCOV toolchain (geninfo / lcov / genhtml), .info format keywords, language-agnostic converters.
• references/format-diff-and-ci.md - full .info record-keyword table, the baseline-diff + gate reference implementation, and the PR coverage-gate CI job.
• cobertura-analysis - sibling for the Cobertura XML format (same PR-gating shape, different parser).
• coverage-diff-reporter - build-an-X workflow that consumes parsed coverage and emits a PR comment with file-level deltas.
• test-coverage-targeter - picks which uncovered branches to target first using the parsed output.