aurora-agent: Skill for Claude Code

Measure section boilerplate

Sections of this blueprint are heavily repetitive. §42 carries about five distinguishing lines per module; §23 carries about fifty-four. That difference is the whole reason a coverage percentage is not a progress bar — twenty modules of one section are not twenty modules of another — so every crate that builds against a section is expected to measure it.

Agents measuring the same section have disagreed three times. Every disagreement was run down to a specific unstated definition, and none was a mistake. The resolutions are in docs/COVERAGE.md under "Three disagreements, three distinct causes". Read them before you measure, because you are about to make the same three choices.

State the method before the number

A percentage with no method attached is not reproducible and will start a fourth disagreement. State all six of these, every time. crates/megafactory/src/lib.rs and crates/sweep/src/lib.rs are the worked examples.

1. Corpus — which files. Exclude 00_SECTION_INDEX. Say the count.
2. Unit — line, heading-block, or character.
3. Normalisation — trailing whitespace trimmed, and anything else you did.
4. Threshold — a line is shared when it occurs in at least k of n modules. Give k.
5. Blank lines — in or out.
6. YAML front matter — in or out.

Then print the numerator and denominator, not just the ratio. crates/atlasx could not land on §33's published 75.4% under any variant it tried — it came out 0.1 point above, about two lines in 1,210 — so it printed both figures in lib.rs to make the gap checkable rather than arguable. That is the right response to an irreducible discrepancy.

Write percentages to one decimal place. Two decimals produce a token tools/coverage.sh reads as a blueprint citation; see the classify-blueprint-modules skill.

Run the sensitivity, not just the measurement

Three knobs. Report what each does, because which one moves is itself the finding.

Threshold is usually inert, and the reason is structural. Sharing in most sections is bimodal: a line is either in exactly one module or in all of them, with almost nothing between. crates/worldfactory found §27 gives 53.6% at k=22, k=20 and k=12 alike — 539 distinct lines in exactly one module, 28 in all twenty-two, six anywhere between. crates/megafactory found §35 identical at k = 50%, 80% and 100%. crates/bioevalx found §26 and §07 identical at 50, 80 and 100%. Five of crates/sweep's six sections do not move by a tenth of a point between t=0.3 and t=0.9.