aurora-agent: Skill for Claude Code

Prove a scanner fires

Some rules cannot be carried by a type. "No file in this crate cites a module it did not implement" and "no source line names this guarded type" are properties of text, and the only way to enforce them is a test that reads the crate and greps it. This workspace has a dozen such scanners.

crates/oraclex articulated the rule they all live under:

A scanner that detects nothing is worse than no scanner. It reports a clean bill of health forever, including after the rule it checks stops being true — and the passing test is what stops anyone from looking.

A workspace audit found seven scanners across six crates with no proof they fired — bioevalx, interweave, megafactory, modalities, oraclex and sweep. They all have one now. Two of them were genuinely broken, and neither would have been found by reading the code.

The four things every scanner needs

1. A planted violation, with negative cases

Name it <thing>_scanner_can_actually_see_one or the_<thing>_scanner_sees_a_planted_violation, and assert both directions in the same test. crates/oraclex/tests/hygiene.rs:

#[test]
fn the_hardcoded_constant_scanner_can_actually_see_one() {
    assert_eq!(
        numeric_constants("synthetic.rs", "const MIXTURE_FLOOR: f64 = 0.15;").len(),
        1,
        "a scanner that detects nothing is worse than no scanner"
    );
    assert!(numeric_constants("synthetic.rs", "pub const ALL: [Family; 12] = [").is_empty());
}

The negative case is not decoration. [Family; 12] is an array length and says nothing about biology; a scanner that flags it will be allowlisted into uselessness within a month. crates/bioevalx/tests/plane_and_zero.rs does the same for imputation — it fires on cell.score().unwrap_or(0.0) and on a fn score_or_zero, and deliberately does not fire on label.unwrap_or("unnamed") (a defaulted string is not an imputed score) or on a comment describing the rule.

2. Planted ids assembled from digits, never written out