Incomplete runs and limits

What incomplete means

incomplete is true when some tests could not be compared: an unreadable config, a dynamic title, a table that cannot be resolved statically, a limit reached. The absence of a finding is then not evidence of its absence.

How each reporter says so:

Reporter	Where
terminal, html	A banner before the findings.
json, jsonl	The incomplete flag, plus incompleteCauses, which enumerates the reasons, counted.
sarif	invocations[0].properties.incomplete. SARIF does not carry the causes.

--fail-on-incomplete turns the flag into a non-zero exit. A file in which every test was excluded is a blind spot; see Blind spots.

Choosing which causes count

--fail-on-incomplete or --fail-on-incomplete=any counts every cause, because the classes call for different responses.

• =reads counts only causes meaning the project could not be fully read: unreadable, oversized or partially read files, undetermined config, and tests whose context could not be resolved.
• =limits counts only the analyzer’s own bounded work being capped: the pair budget, the hook-scan budget, the memory budget, and truncated findings.

The value needs the =; with a space, reads would be read as the path.

Blind spots

A file in which every test was excluded produced no findings, so it appears nowhere in the report. Blind spots are named on stderr, worst first, counted in the blindFiles field of the json and jsonl summaries, and gated by --fail-on-incomplete=reads, since a wholly excluded file is the extreme form of an unresolved-context cause.

Source limits

Source files are limited to 8 MiB. Before parsing, the frontend also refuses lexical bracket nesting deeper than 256 and more than 8,192 raw syntax units.

• The units count opening brackets, operator, member, label and escape punctuation, and recursive control and type keywords, including occurrences in strings, comments and regexes.
• Nonnegative integer payloads and commas do not consume units.
• The bound is conservative: lexical ambiguity must not hide stack-exhausting input. A wide generated file or punctuation-heavy literal can therefore exceed it even when its actual syntax is shallow.
• These are analyzer resource limits, not JavaScript or Vitest restrictions.

Refused files make analysis incomplete and produce diagnostics; findings in other files remain available. First-party imported helpers use the same preflight limits. Use --fail-on-incomplete=reads (or --fail-on-incomplete) to make refused input an operational failure. Admitted extraction runs on an explicitly sized worker stack; a larger stack alone would not protect against unbounded input.

Declaration limits

Each file emits at most 10,000 declarations (tests, .each rows and suite records, including those from imported runners) and unrolls at most 10,000 literal for...of iterations. A .each table or loop array longer than what remains of its limit is skipped without being read. Past either limit the file stops emitting and counts as not read in full; the declarations already emitted are compared as usual.

Memory budget

The run reads whole files until the analyzer holds more than its memory budget: 2,048 MiB by default, set with --max-memory <MiB> and removed with --no-memory-cap. The figure is the retained heap --debug-memory reports plus source text kept for the HTML report.

A second backstop checks current process RSS after each file and stops above 1.5 times the budget on Linux, macOS and Windows.

• RSS includes allocator overhead and other resident allocations; it is separate from the lifetime peak reported by --debug-memory.
• It is an absolute process measurement: it includes the startup footprint and memory still resident after the baseline pass, even after baseline data is dropped.
• A budget whose RSS ceiling is below that footprint can stop the current pass after its first file, even when little test data is retained.
• If the OS sample is unavailable, the retained-byte check still applies.
• --no-memory-cap disables both checks.

The checks run at file boundaries, so a single file or the later analysis phase can exceed either threshold. Files after that point are not analyzed and are counted in filesOverMemoryBudget (JSON and JSONL; the terminal and HTML banners name the cause). Every rule runs over the files read, with the same findings as a run over exactly those files. --baseline-from-ref applies the same budget to the base revision.

Back to the documentation index.