A classifier + parallel render meta-skill. After scanning a target
path, an llm_classify step picks the most informative diagram kind
(one of class | sequence | component | deploy | flow), then two
independent render branches synthesize PlantUML source and draw.io
XML for the same scan — running in parallel because they are
independent tools serving different downstream uses (git-friendly text
review vs. editable canvas).
Trigger surface
Fire by saying diagram triangulation or one of the localized triggers listed
in the frontmatter, with the target path or module reference in the same turn.
Fallback
If either render step fails, compose_doc should still produce the
docx referencing whichever render succeeded; manually re-run the
failed render via sub-agent with the same scan output as input.