The user should approve the lane order, main path, branches, returns, and assumptions before a file is generated. Unknown business facts remain unresolved; the agent must not invent hidden steps or owners.
Language models identify meaning and topology. A deterministic engine handles coordinates, spacing, ports, routing, label carriers, and XML serialization.
The confirmed main path is routed first. Same-lane progress prefers a direct top-to-bottom connection. Returns and retries use separate channels where space permits.
The primary artifact is a native .drawio file. PNG or other exports are review and sharing formats, not replacements for the editable source.
The next revision starts from the latest user-saved Draw.io file. Stable IDs and semantic metadata support inspection, targeted patches, and preservation of compatible manual geometry.
Patches contain declared additions, updates, and deletions. Geometry changes require explicit authorization. Unrelated nodes, connectors, phases, and manual waypoints remain unchanged whenever compatibility allows.
Unknown schema fields, unsafe deletions, invalid main paths, output replacement, and unsupported files produce clear diagnostics instead of best-effort mutation.
Strict validation catches defined structural and routing problems. Preview export proves only that rendering succeeded. A multimodal review may find additional visual issues but is non-deterministic and does not replace human review for important diagrams.
This project optimizes editable product and business vertical swimlanes. It does not attempt to become a general-purpose generator for BPMN, UML, C4, ERD, network topology, or presentation graphics.
The Skill package contains no user data, organization-specific terminology, domain-specific sample flow, generated diagram, or dependency on a single agent runtime. Public examples live outside the Skill directory and use fictional, neutral content.