Optional ReadonlyadapterWhich harness drives it — the adapter itself (default defaultAdapter). A harness whose firing signal is INFERRED rather than a discrete event reports n/a: the matrix asks WHICH skill fired, which an inference cannot answer.
Optional ReadonlyconcurrencyParallel runs across the prompts × trials grid (default 1).
Optional ReadonlydialectOptional ReadonlyeffortReasoning budget (claude --effort) for the selector, or undefined for the
harness default. Present for measureSelectionMatrix, the ASSERTABLE
test primitive, where the configuration a number came from has to be pinnable.
Deliberately NOT exposed as an audit CLI flag: audit is a local report,
not a reproducibility surface, and a flag nobody can act on is surface without
a use. The audit probe therefore leaves this unset and runs at the default.
Optional ReadonlyharnessPass SelectionOptions.adapter instead. Kept for one
release because this options type is public API (vigiles/claude-code);
a name is resolved through the registry, never compared to a literal.
Optional ReadonlylayoutLayout + dialect for candidate discovery, mirroring ProbeOptions — so a Codex repo's skills are found under the Codex layout, not silently missed by the Claude Code one.
Optional ReadonlymodelSelector model — defaults to Sonnet (a weaker model under-selects).
ReadonlypluginThe plugin root whose skills are measured against each other.
Optional ReadonlypromptsPer-skill recall prompts. Omit for ZERO-SETUP — prompts are auto-derived
from each skill's description (the same generator the audit trigger tier
uses). Supply your own for a curated collision benchmark; only the prompts
array per skill is read (any irrelevant bank is ignored here).
Optional ReadonlytrialsRepeats per prompt (default 1).
measureSelectionMatrix's two arguments as one declarable object.