Error handling

The compiler never throws away the document over one bad line. It is forgiving by design: it always produces the best diagram it can and reports what it could not use, so a typo never leaves you with a blank canvas.

An unparseable line is skipped, with a line N: … diagnostic.
An unknown attribute word, icon, shape, or colour is ignored with a diagnostic; the rest of the line still applies.
An edge to an undeclared name auto-declares the node (graph and sequence); ERD and class relations to unknown entities are dropped, since those families require a declaration.
Duplicate ids or names: the first declaration wins, and later ones fill in attributes the first did not set.
The result passes through the same sanitizer the streamed AI output goes through.

Diagnostics keep drawing#

The example below has three problems: an unknown colour, an unknown shape, and a self-edge. The diagram still renders, and each problem reports underneath.

Server [chartreuse]
Cache [octahedron, blue]
Server -> Server
Server -> Cache

diagnostic · line 1: unknown attribute "chartreuse"

diagnostic · line 2: unknown attribute "octahedron"

diagnostic · line 3: self-edges are not drawn

Diagnostics report the bad words; the good parts draw anyway.

Partial lines still apply#

An unknown word in an attribute list is dropped on its own; the words around it are still used. Here the colour survives even though the shape word is unknown.

Node [octahedron, blue, bold]
Other [rectangle, green]
Node -> Other

diagnostic · line 1: unknown attribute "octahedron"

octahedron is unknown, so the node falls back to a default shape, but blue and bold still apply.

Severity#

Diagnostics come in two severities. A warning (amber) means something was ignored but the diagram is intact: an unknown word, a dropped self-edge. An error (red) means a whole line could not be parsed. Both surface in the code panel, keyed to the line, and neither stops the diagram from rendering.