The expressiveness gap traced to vocabulary: draw.io has hundreds of shapes,
the declarative layer allowed six. Opening it, with the failure modes from
design review handled:
- shapes.ts: a ~20-entry curated catalog (full style fragment incl. the
matching perimeter — required or edges connect to the bounding box; a
text-scale factor verified in the real editor — the same sentence overflows
a 1.0x rhombus and fits a 1.5x one; labelOutside+glyph for umlActor-style
figures). Any other token passes through verbatim: draw.io degrades unknown
shapes to rectangles safely (verified). Injection-capable tokens (;/=) are
rejected outright.
- Pass-through emits a WARNING with a nearest-catalog hint (bounded edit
distance), so a typo'd 'cyclinder' is a one-turn fix instead of a silently
rectangular node forever. set_shape/set_role/set_group operations make the
fix possible without remove+re-add (which would drop links).
- mergeStyle(): style fragments merge per-key (later wins, bare shape classes
displace each other) instead of string concatenation. This is what makes
shape and theme composable by rule — shape owns geometry keys, theme owns
colour/type keys, and an overlap resolves by order instead of emitting
contradictory duplicates.
- dai_shape marker carries the declared token through the round trip:
appearance-based reverse mapping cannot distinguish aliases (diamond vs
decision) or a rotated queue from a cylinder.
- dai_auto marker separates engine-measured size from user-fixed size: parsed
boxes no longer freeze the first layout's numbers, so changing a label
re-measures. Pinned nodes keep everything, as before.
- draw_graph's closed shape enum opened to match (first instance of the
schema-drift problem the review predicted).
14 new tests: merge ownership, injection rejection, near-match hints,
alias-preserving round trip, re-measure on label change, mxgraph.* tokens
staying boxes with role/group intact. 556 unit tests green; acceptance
diagram (person/hexagon/cylinder/queue/cloud/decision/callout) verified in
the real editor.
Extends the declarative engine past cloud architecture. The tool routing was
divided by icon library — AWS through the engine, everything else hand-written
XML — which is the wrong axis. What matters is the LAYOUT SHAPE.
Measured first: a six-step approval flow declared in its natural order comes out
as one column, because the layout only arranges what nesting tells it to and
never looked at the arrows. That forces the arrow from the decision to its second
branch to jump over the first branch.
graph.ts computes what the layout should have looked at: layer assignment by
longest path, cycle breaking so a loop is drawn without setting the order, and
barycentre sweeping to cut edge crossings. It emits ordinary container
operations, so layout, routing and round-tripping are unchanged — reaching zero
arrows-through-boxes on a 14-node pipeline and zero crossings on a bipartite
graph whose declared order forces three.
Three new container kinds, each because one layout rule cannot serve them all:
pool — swimlanes. Lanes are real cells and each step is parented to its
band, so dragging a step to another role records the change.
sequence — participants across the top, one lifeline cell per participant so
head and line stay together on a drag. Messages bypass the router:
a message's height IS its order.
radial — mind maps and org charts. Children are a flat list and the
hierarchy comes from the links, because a branch is a box and a box
cannot hold children.
Flowchart box shapes (diamond, stadium, parallelogram, document) so a reader can
tell a branch from a step.
Two bugs the new tests caught: the duplicate-link guard blocked a sequence
diagram from having two messages between the same pair, and the fallback message
numbering was shared across containers, pushing a second diagram's messages off
its own lifelines.
523 unit tests and 17 diagram e2e tests pass. Every kind verified round-trip
stable to a fixed point, and rendered in a real browser — draw.io keeps the
lifeline shape and the lane markers.
Closes the loop: the model can now build and edit AWS architecture diagrams by
declaring structure, and never writes an mxCell again.
catalog.ts — 983 AWS icon and 19 group stencils as a name→style map, generated from
drawio-ai-kit's catalog (itself generated from jgraph's draw.io shape index). Styles
are verbatim, so the official category colours, connection points and aspect=fixed
come along for free and nothing is hand-assembled. An invented name is rejected with
suggestions instead of rendering as a blank square, which is what draw.io does with
an unknown resIcon today.
operations.ts — what the model actually sends: add_icon / add_container / move /
link / set_dir and so on, applied in order against the tree. Guards the things that
break a diagram quietly: duplicate ids (draw.io drops one of the two cells), edges
left pointing at a removed node, and moving a container inside itself.
index.ts — the entry point. current XML → parse → apply ops → check names → layout →
render → new XML. The tree is not stored between calls; it is re-derived from the
canvas every time, so a user's manual edits are input to the next layout rather than
state to reconcile.
Token cost, measured with Claude's tokenizer rather than estimated:
- build a VPC diagram: 515 tok as operations vs 3180 as XML (6.2x)
- add one icon: 27 tok as an operation vs 3823 re-emitting (142x)
- read current state: 216 tok as an outline vs 3180 as XML (14.7x)
The 142x is the one that matters day to day: "add a Redis" is one operation, not a
rewrite of the whole diagram.
Routing in the system prompt sends AWS architecture through this path and leaves
flowcharts, BPMN, sequence diagrams, mind maps and Azure/GCP on display_diagram —
the layout engine's primitives (nested rows, columns, grids) do not model a sequence
diagram's lifelines or a mind map's radial spread, and pretending otherwise would
make those worse rather than better.
Also: added a NOTICE recording the MIT port and the AWS Architecture Icons terms,
and a narrow .gitignore exception so the generated catalog is tracked while the
root data/ directory (admin settings, contains secrets) stays ignored.
403 unit tests + 3 new e2e. Verified in the real app: a structural tool call renders
with real stencils and container markers; a second call adds one node and keeps
everything from the first; an invented name is refused and nothing is drawn. The 13
existing diagram e2e tests still pass.
