dos-extract

dos-extract

Reverse-engineer a Design Ontology Spec from a repository. The product is two files: dos.yaml (the 13-section ontology: a core model of ≤7 objects + rules, and the complete vocabulary of everything else the team says) + decisions.md (why Topic, not TopicNode). This skill describes what a DOS is, the judgments that separate a real ontology from a scan dump, the primitive that does the mechanical scanning, the human seam, and the exit that certifies the result. It prescribes no step order — the engine sequences the work; what follows are the gaps and the gates.

The gap (why a scan is not an ontology)

A composite of three atoms: Knowledge (what a DOS is — the 13-section, two-layer format, the code-vs-docs signal priority), Capability (the mechanical noun/verb scan — a primitive the engine otherwise mis-improvises), Judgment (the four classification calls below).

The load-bearing reason this is not free:

Code contains the current implementation choices, not the ontology that should exist. A naive "scan code → emit objects" pass freezes mistakes into the contract: a CommentCard React component becomes a CommentCard object — wrong twice (it is UI, and Card is presentation, not domain).

Deletion test: remove this skill and ask the engine to "extract a DOS from this repo." It greps nouns and emits a polluted ontology — UI elements and *Repository/*Service names promoted to objects, no ≤7 discipline, code vocabulary winning over the team's language. The gap is the semantic judgment, plus the knowledge of what a clean DOS looks like.

The second gap (found on this plugin's own DOS, 2026-09-18) is the mirror image: a skill that only knows the ≤7 discipline produces a clean sliver. The 2026-09-05 dogfood run classified 45 real terms and shipped 7 objects; the other ~38 — AC, 指纹, 层, 三档, 隐藏集, Finding, Verdict, PSL, Territory… — survived only as prose in decisions.md and a workspace table, where no closure check, no card linter and no drift sensor could see them. The docs even carried an explicit glossary (ARCHITECTURE.md §16 术语, 13 terms) that the DOS did not lift as a glossary. From outside that reads as "it extracts a small part and aligns nothing" — and it is right. ≤7 caps the core model, not the language. No ontology methodology caps concepts: Ontology 101 enumerates every term first (step 3), METHONTOLOGY and NeOn open with a glossary, Evans's Highlighted Core flags the core inside the full model instead of deleting the rest, and every AI-facing semantic layer (Palantir, dbt, Snowflake, DataHub) treats coverage as the accuracy lever. They layer; they do not truncate. The origin of the ≤7 figure (Jorgenson's "no more than 5 objects") was a consolidation exercise for a greenfield prototype — never a completeness criterion for an existing repo. Full grounding: references/methodology.md §"Industry grounding".