Supported formats

mehen parses six report formats in-house — no external tools, no network — covering the default or one-flag-away output of the mainstream test runners in every language mehen analyzes. The parsers are streaming (large monorepo reports never sit fully in memory) and were adapted from the MIT-licensed covrs project.

​

The formats

​

Detection is content-based

Filenames lie: Pester writes JaCoCo XML into coverage.xml, coverage.py writes Cobertura XML into the same name, and .info files are occasionally GNU documentation. mehen therefore sniffs the first 4 KiB of every candidate and requires format-specific content markers before believing a file, in a fixed priority order chosen so no format can false-positive on another’s output:

1. LCOV — SF: plus DA:/FN: records.
2. Go coverprofile — a mode: set|count|atomic header or file.go:N.N,N.N … block lines.
3. Istanbul — a JSON object containing "statementMap" and "fnMap".
4. JaCoCo — <report plus a JaCoCo DTD reference or <package.
5. Clover — <coverage carrying a clover="…" version attribute.
6. Cobertura — <coverage without the Clover marker.

An explicit --coverage=<path> argument goes through the same detection; a file no parser recognizes is a hard error, while an auto-discovered candidate that fails sniffing is silently recorded as rejected (that is business as usual for e.g. coverage.txt text summaries).

​

Normalization and merge semantics

Formats disagree about granularity, so parsed records are normalized into one model — per-file line hits, branch arms, and function records:

• Statements → lines. Istanbul statements and Go blocks map onto lines; when several statements share a line, the maximum hit count wins.
• Branch encodings → arms. LCOV BRDA records, Cobertura condition-coverage="50% (1/2)" fractions, JaCoCo mb/cb counters, Clover truecount/falsecount pairs, and Istanbul b arrays all become flat branch arms (see branch coverage for what that folding means).
• Duplicate records collapse. lcov -a-merged tracefiles repeat DA lines; Cobertura emits the same line under both <method> and <class> — duplicates keep the maximum.
• Cross-report merge. All ingested reports fold into one dataset: union of files, saturating-max hits for records shared between reports (“covered anywhere ⇒ covered”). Max is commutative and associative, so the merged result is independent of discovery order — determinism is a hard requirement for reproducible CI gates. Hit-count summing was rejected because re-running the same suite twice would double every count; newest-file-wins was rejected because git checkouts do not preserve mtimes.

​

Hardening

Coverage artifacts are ingested from build directories that other tools write into, so the parsers are defensive by construction: branch expansion is capped at 1,024 arms per line, Go block spans at 100,000 lines, report files at 256 MiB; XML parsing never resolves DTDs or external entities (billion-laughs and XXE are structurally inert, and a regression test pins that); and a malformed report is a per-file diagnostic, never a crash.

​

What about raw instrumentation output?

.profraw/.gcda (LLVM/GCC counters), SimpleCov’s .resultset.json, c8’s raw V8 dumps in coverage/tmp/, and coverlet’s proprietary coverage.json are not report formats — decoding them requires the compiled binaries or the producing tool’s internals. Export a report instead (cargo llvm-cov --lcov, coverage xml, --coverageReporters=lcov, …); grcov takes the same report-level stance for Firefox-scale ingestion.

​

References

• geninfo(1) — the LCOV tracefile format. Linux Test Project.
• The cover story — Go’s coverage design and the coverprofile format. The Go Blog.
• Istanbul.js and the istanbuljs coverage object.
• JaCoCo XML report documentation.
• OpenClover documentation — the Clover XML schema.
• Cobertura — the original tool whose XML schema became the de-facto interchange format.
• mozilla/grcov — prior art for multi-format coverage aggregation at scale.