coverage-diff-reporter

Overview

A whole-repo coverage gate is necessary but not sufficient. A drop of 0.4pp overall might hide a 12pp drop in one critical file. A new file at 35% line coverage might pass an aggregate gate but leave a regression-risk hot spot.

This skill builds a coverage diff report that solves the "reviewer can see what to look at" problem:

1. Parse the current run's coverage (LCOV / Cobertura).
2. Parse a baseline from the merge target.
3. Emit a per-file delta table sorted by absolute drop.
4. Highlight new files below threshold, files with disproportionate drops, and files that gained coverage (positive feedback).
5. Post a single-line summary as the PR top-of-comment, with click-through to the full table.

When to use

• A PR is missing reviewer signal about which file's coverage changed.
• The repo has coverage but no PR-time visualization (the data is in the artifacts but nobody opens them).
• A coverage SaaS isn't an option (compliance, cost) and the team wants a self-hosted equivalent.

This skill does not decide pass/fail - that's the gate's job (see lcov-analysis or lcov-analysis references/cobertura.md). This skill just makes the diff legible.

Step 1 - Pick the parser

Match the existing CI's reporter:

Existing reporter	Use
LCOV .info	lcov-analysis parser
Cobertura XML	lcov-analysis (references/cobertura.md) parser
Jest JSON / V8 coverage	Convert to LCOV first (jest --coverageReporters=lcov)
JaCoCo XML	Use jacoco-analysis, or convert to Cobertura
coverage.py	coverage xml → Cobertura, OR py2lcov → LCOV

The reporter writes to current.json (parsed). The same parser runs against the baseline → baseline.json.

Step 2 - Get the baseline

Two patterns:

Pattern A - cached artifact (recommended)

The main branch's last successful CI run uploaded its coverage as an artifact. PR jobs download it.

- name: Restore baseline
  uses: dawidd6/action-download-artifact@v3
  with:
    workflow: coverage.yml
    branch: main
    name: coverage-baseline
    path: baseline/

- name: Parse current
  run: python scripts/parse_lcov.py coverage/lcov.info > current.json

- name: Parse baseline
  run: python scripts/parse_lcov.py baseline/lcov.info > baseline.json

- name: Generate diff
  run: python scripts/coverage_diff.py current.json baseline.json > diff.md

Pattern B - recompute the baseline in the PR job

The PR job checks out main, runs tests + coverage, then checks out the PR head. Slower (~2x runtime) but always-fresh.

- name: Checkout main
  run: git fetch origin main && git checkout origin/main

- name: Run tests on main
  run: npm test -- --coverage && cp coverage/lcov.info baseline.lcov

- name: Checkout PR head
  run: git checkout ${{ github.event.pull_request.head.sha }}

- name: Run tests on PR head
  run: npm test -- --coverage

Pattern A is the default. Pattern B is the fallback when artifact retention has expired or main coverage is non-deterministic.

Step 3 - Compute the per-file delta

# scripts/coverage_diff.py
def compute_diff(current, baseline):
    base_idx = {f['path']: f for f in baseline}
    rows = []
    for f in current:
        b = base_idx.get(f['path'])
        line_now    = pct(f.get('lh', 0), f.get('lf', 0))
        branch_now  = pct(f.get('brh', 0), f.get('brf', 0))
        line_then   = pct(b.get('lh', 0), b.get('lf', 0)) if b else None
        branch_then = pct(b.get('brh', 0), b.get('brf', 0)) if b else None
        rows.append({
            'path': f['path'],
            'is_new': b is None,
            'line_now': line_now,    'line_delta':   delta(line_now, line_then),
            'branch_now': branch_now,'branch_delta': delta(branch_now, branch_then),
        })
    # Also catch deletions - files in baseline but not current.
    for path, b in base_idx.items():
        if path not in {f['path'] for f in current}:
            rows.append({'path': path, 'is_deleted': True, 'line_now': None, 'line_then': pct(b.get('lh', 0), b.get('lf', 0))})
    return rows

Step 4 - Sort and classify

Reviewers care most about big drops. Sort by line_delta ascending (most-negative first), with new sub-threshold files at the top:

def classify(row):
    if row.get('is_deleted'):                      return 'deleted'
    if row.get('is_new') and row['line_now'] < 80: return 'new_below_threshold'
    if row.get('is_new'):                          return 'new_ok'
    if row['line_delta'] is not None and row['line_delta'] <= -5:   return 'regressed'
    if row['line_delta'] is not None and row['line_delta'] <  0:    return 'declined'
    if row['line_delta'] is not None and row['line_delta'] >  0:    return 'improved'
    return 'unchanged'

The thresholds (80% for new files, -5pp for regression) are tunable per repo.

Step 5 - Render the report

## Coverage diff - `<sha>` vs `main` `<base-sha>`

**Overall:** line 84.2% (-0.3pp) | branch 71.5% (-0.1pp)
**Files changed:** 7 (3 regressed, 1 new, 2 improved, 1 deleted)

### ⚠ Regressions (4)

| File                                  | Line%       | Branch%     |
|---------------------------------------|-------------|-------------|
| `src/checkout/cart.ts`                | 65.4 (-12.8 ⬇) | 50.0 (-25.0 ⬇) |
| `src/checkout/promo.ts`               | 78.0 (-8.5 ⬇)  | 60.0 (-15.0 ⬇) |

### 🆕 New files (1)

| File                                  | Line%       | Branch%     |
|---------------------------------------|-------------|-------------|
| `src/checkout/discount-stack.ts`      | 35.0 (NEW, below 80% threshold) | 25.0 |

### ✅ Improvements (2)

| File                                  | Line%       | Branch%     |
|---------------------------------------|-------------|-------------|
| `src/orders/list.ts`                  | 92.0 (+4.5 ⬆) | 85.0 (+10.0 ⬆) |

### 🗑 Deleted (1)

| File                                  | Was line%   |
|---------------------------------------|-------------|
| `src/legacy/old-checkout.ts`          | 22.0        |

The four-section split (Regressions / New / Improvements / Deleted) matches reviewer attention budget. Improvements get airtime - positive feedback prevents the gate from feeling adversarial.

Step 6 - One-line summary for the PR top

PR comment APIs render long markdown by default; the summary line sits at the top so the reviewer doesn't have to scroll:

📉 Coverage 84.2% (-0.3pp) - 3 files regressed, 1 new file below threshold. See full report below.

Or if all-clear:

✅ Coverage 84.5% (+0.2pp) - no regressions, 2 files improved.

Step 7 - Post to the PR

- name: Generate diff report
  run: python scripts/coverage_diff.py current.json baseline.json > diff.md

- name: Post / update PR comment
  uses: marocchino/sticky-pull-request-comment@v2
  with:
    header: coverage-diff
    path: diff.md

sticky-pull-request-comment uses the header to update the same comment across pushes - the reviewer doesn't see N copies of the report as the PR evolves.

Anti-patterns

Anti-pattern	Why it fails	Fix
Posting only the aggregate (overall ± Xpp)	Hides which file regressed; reviewer can't act.	Per-file table sorted by drop (Step 4 - 5).
One thread per push (new comment per commit)	PR conversation drowns in coverage churn; nobody reads.	Sticky comment updated in place (Step 7).
Showing every unchanged file	500-row tables; the 3 regressions are buried.	Filter to only changed files; one summary line for unchanged count.
Adversarial framing ("FAIL: coverage dropped")	Reviewer associates coverage tool with friction; team disables.	Show improvements too (Step 5). Gate failures are the gate's job; this report is informational.
Using PR's merge-base coverage (re-runs main coverage)	Doubles CI cost; flake risk on the main re-run.	Cache main coverage as artifact (Step 2 Pattern A).
Hiding new files because they "don't have a baseline"	New files are exactly where regressions enter the codebase.	Always show new files; flag the sub-threshold ones explicitly (Step 4).
Ignoring deleted files	Coverage went up because high-coverage code was deleted; aggregate misleads.	Show deletions (Step 3); explain in summary if they cause aggregate movement.

Limitations

• Per-line uncovered detail isn't shown. This skill is the file-level summary; for per-line drilldown, generate the language-native HTML report (Allure / genhtml / coverage html).
• No semantic awareness. A 50% drop because the file got 2x bigger (more code, same number of tests) reads the same as a drop because tests were deleted. The reviewer still has to look.
• Sub-1pp deltas are noise. Coverage tools have measurement jitter from non-deterministic test ordering; show only deltas above a threshold (typical: ±0.5pp).
• Baseline staleness. Pattern A assumes main's last coverage artifact is recent. If main is stale, surface the baseline age in the report header.

References

• lcov-analysis - LCOV parser this skill consumes.
• lcov-analysis - LCOV + Cobertura parser this skill consumes.
• jacoco-analysis, coverage-py-analysis, and js-unit-tests (qa-unit-tests-js) for Jest/Vitest - language-specific parsers; convert to LCOV / Cobertura before feeding this skill.
• test-coverage-targeter - downstream skill that reads the same data to suggest which uncovered branches to target next.