feat(diagram-engine): wire up restructure_diagram + stencil catalog
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.
2026-08-09 12:13:54 +09:00
/ * *
* Structural operations — what the model sends instead of XML .
*
* The tree is re - derived from the canvas on every call , so an operation names existing
* nodes by id and says what to change . Adding one node costs a few dozen tokens ; the
* equivalent as raw mxCell XML is hundreds , and re - emitting the whole diagram to add one
* icon costs thousands .
*
* Operations are applied in order , each against the result of the last , so a sequence
* like "add a frame, then move two nodes into it" works in a single call .
* /
import { z } from "zod"
import {
type ContainerNode ,
type DiagramNode ,
type DiagramTree ,
findNode ,
findParent ,
isContainer ,
type LinkSpec ,
walkTree ,
} from "./types"
export const OperationSchema = z . discriminatedUnion ( "op" , [
z . object ( {
op : z.literal ( "add_icon" ) ,
id : z.string ( ) . describe ( "New unique id for this node" ) ,
parent : z
. string ( )
. optional ( )
. describe ( "Container id to add into; omit for top level" ) ,
name : z.string ( ) . describe ( "Catalog stencil name, e.g. 's3' or 'ec2'" ) ,
label : z.string ( ) . optional ( ) ,
feat(diagram-engine): flowcharts, swimlanes, sequence diagrams and mind maps
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.
2026-08-09 13:47:23 +09:00
lane : z
. number ( )
. optional ( )
. describe (
"Inside a pool: which lane (0-based row) this belongs to" ,
) ,
col : z
. number ( )
. optional ( )
. describe (
"Inside a pool: which column (0-based step) this sits in" ,
) ,
feat(diagram-engine): wire up restructure_diagram + stencil catalog
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.
2026-08-09 12:13:54 +09:00
after : z
. string ( )
. optional ( )
. describe ( "Insert after this sibling id; omit to append" ) ,
} ) ,
z . object ( {
op : z.literal ( "add_box" ) ,
id : z.string ( ) ,
parent : z.string ( ) . optional ( ) ,
label : z.string ( ) ,
feat(diagram-engine): design tokens + role/group composition, engine-wide theming
Paper-summary posters previously required hand-written XML: every engine
box rendered identically (white, 11px), so anything whose meaning lives
in visual hierarchy came out flat. This makes presentation a first-class,
generalised part of the declaration - not a poster feature.
Structure/presentation separation, the same split HTML and CSS settled on:
- ROLE says what a node IS: banner, heading, body, callout, good, bad,
metric, muted. Maps to a type scale and an emphasis (filled / tinted /
outlined / ghost), never to a colour.
- GROUP says which semantic zone a node belongs to. Each distinct group
name gets one hue ramp (tint / base / dark), assigned in document
order. Promoted from a draw_graph-only field to BoxNode and GroupNode,
round-tripped via dai_group.
- themedStyle(role, hue, kind) composes the two by rule - there is no
per-combination table to extend, so a new diagram kind gets full
theming by tagging nodes. The model never sees a hex value.
A heading container plus a group yields the tinted section panel with a
dark title; a grouped body box takes its zone's tint; verdict roles stay
green/red regardless of zone; the banner is the page's one dark field.
Also fixed, found while building the acceptance poster:
- autoBoxSize only counted explicit newlines, so a long single-line label
wrapped to six lines in draw.io but got a one-line-tall box, and the
text overflowed the cell.
- Marker stamping appended without replacing, so every render of a
recovered style grew it by one duplicate dai_* token per key -
unnoticed because draw.io resolves duplicates last-wins. dai_* keys
are now replaced in place; mxGraph keys still append, because
last-wins is load-bearing for container=1 normalisation.
- Banner/heading/metric roles stretch across their container's cross
axis, the way a masthead spans its page.
- Prompt: a poster's banner IS its title (no set_title alongside), and
sections get their colour by naming groups.
537 unit tests, 250-flowchart corpus still zero crossing arrows, 5 e2e
tests in a real browser. Verified visually: the Transformer-paper poster
renders with a navy masthead, three hue-coded section panels, metric,
verdict and callout boxes - all engine-computed geometry.
2026-08-09 19:48:01 +09:00
role : z
. enum ( [
"banner" ,
"heading" ,
"body" ,
"callout" ,
"good" ,
"bad" ,
"metric" ,
"muted" ,
] )
. optional ( )
. describe (
"What this IS: banner=masthead, heading=section title, callout=must-not-miss, good/bad=verdict, metric=key number, muted=fine print. The theme decides how each looks" ,
) ,
group : z
. string ( )
. optional ( )
. describe (
"Semantic zone name; nodes and panels sharing a group get the same hue from the engine's palette. Never pick colours" ,
) ,
feat(diagram-engine): label avoidance, paired opposite edges, semantic group colours
A git-workflow flowchart rendered with no overlaps but read poorly. Three
distinct causes, each fixed and measured:
1. Edge labels sat on boxes and on each other (4 collisions on the
reported diagram; 280 across 250 generated flowcharts). The router
keeps LINES off the boxes but a label renders at its edge's midpoint,
which on a long edge is beside exactly the things the line was routed
around. placeLabels slides each label along its own edge to a clear
spot — longest edges first, midpoint-outward tries — written as the
geometry's relative x, which draw.io natively supports. Corpus: 280
label collisions -> 8.
2. A->B and B->A were routed independently, so "git add" ran straight
while "git reset" wandered through a different corridor with a kink.
Opposite edges that agree on axis now get two absolute parallel tracks
in the strip where the two boxes overlap, a constant 24px apart,
converted back to port fractions. Zero crossing regressions.
3. All boxes rendered the same white, because the render layer's
fill/stroke support was never reachable: neither add_box's schema nor
draw_graph's nodes exposed it. Rather than exposing raw hex (the model
picks mismatched saturations, differently every time), nodes take a
semantic group name and the engine maps groups to a fixed palette of
six paired fill/strokes in order of first appearance. The model names
the zones - remote vs local vs temp - and never touches a colour.
532 unit tests pass; the 5 diagram e2e tests pass in a real browser.
2026-08-09 18:59:34 +09:00
fill : z
. string ( )
. optional ( )
. describe (
"Fill colour, e.g. #DAE8FC. Prefer draw_graph's group field over picking colours" ,
) ,
stroke : z
. string ( )
. optional ( )
. describe ( "Border colour; pair it with fill" ) ,
feat(diagram-engine): flowcharts, swimlanes, sequence diagrams and mind maps
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.
2026-08-09 13:47:23 +09:00
shape : z
. enum ( [
"box" ,
"decision" ,
"terminator" ,
"round" ,
"data" ,
"document" ,
] )
. optional ( )
. describe (
"Flowchart outline: decision=diamond, terminator=start/end, data=input/output, document=report. Omit for a plain rectangle" ,
) ,
lane : z
. number ( )
. optional ( )
. describe (
"Inside a pool: which lane (0-based row) this belongs to" ,
) ,
col : z
. number ( )
. optional ( )
. describe (
"Inside a pool: which column (0-based step) this sits in" ,
) ,
feat(diagram-engine): wire up restructure_diagram + stencil catalog
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.
2026-08-09 12:13:54 +09:00
after : z.string ( ) . optional ( ) ,
} ) ,
z . object ( {
op : z.literal ( "add_container" ) ,
id : z.string ( ) ,
parent : z.string ( ) . optional ( ) ,
label : z
. string ( )
. describe ( "Frame title; empty string means invisible wrapper" ) ,
feat(diagram-engine): design tokens + role/group composition, engine-wide theming
Paper-summary posters previously required hand-written XML: every engine
box rendered identically (white, 11px), so anything whose meaning lives
in visual hierarchy came out flat. This makes presentation a first-class,
generalised part of the declaration - not a poster feature.
Structure/presentation separation, the same split HTML and CSS settled on:
- ROLE says what a node IS: banner, heading, body, callout, good, bad,
metric, muted. Maps to a type scale and an emphasis (filled / tinted /
outlined / ghost), never to a colour.
- GROUP says which semantic zone a node belongs to. Each distinct group
name gets one hue ramp (tint / base / dark), assigned in document
order. Promoted from a draw_graph-only field to BoxNode and GroupNode,
round-tripped via dai_group.
- themedStyle(role, hue, kind) composes the two by rule - there is no
per-combination table to extend, so a new diagram kind gets full
theming by tagging nodes. The model never sees a hex value.
A heading container plus a group yields the tinted section panel with a
dark title; a grouped body box takes its zone's tint; verdict roles stay
green/red regardless of zone; the banner is the page's one dark field.
Also fixed, found while building the acceptance poster:
- autoBoxSize only counted explicit newlines, so a long single-line label
wrapped to six lines in draw.io but got a one-line-tall box, and the
text overflowed the cell.
- Marker stamping appended without replacing, so every render of a
recovered style grew it by one duplicate dai_* token per key -
unnoticed because draw.io resolves duplicates last-wins. dai_* keys
are now replaced in place; mxGraph keys still append, because
last-wins is load-bearing for container=1 normalisation.
- Banner/heading/metric roles stretch across their container's cross
axis, the way a masthead spans its page.
- Prompt: a poster's banner IS its title (no set_title alongside), and
sections get their colour by naming groups.
537 unit tests, 250-flowchart corpus still zero crossing arrows, 5 e2e
tests in a real browser. Verified visually: the Transformer-paper poster
renders with a navy masthead, three hue-coded section panels, metric,
verdict and callout boxes - all engine-computed geometry.
2026-08-09 19:48:01 +09:00
role : z
. enum ( [
"banner" ,
"heading" ,
"body" ,
"callout" ,
"good" ,
"bad" ,
"metric" ,
"muted" ,
] )
. optional ( )
. describe (
"Section role: heading=titled tinted panel, banner=masthead strip, good/bad=verdict panel" ,
) ,
group : z
. string ( )
. optional ( )
. describe (
"Semantic zone name; the panel and everything sharing this group take one hue" ,
) ,
feat(diagram-engine): wire up restructure_diagram + stencil catalog
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.
2026-08-09 12:13:54 +09:00
dir : z.enum ( [ "row" , "col" ] ) . describe ( "How children stack" ) ,
gname : z
. string ( )
. optional ( )
. describe (
"Group stencil name, e.g. 'group_vpc'; omit for a plain frame" ,
) ,
gap : z.number ( ) . optional ( ) ,
after : z.string ( ) . optional ( ) ,
} ) ,
z . object ( {
op : z.literal ( "add_grid" ) ,
id : z.string ( ) ,
parent : z.string ( ) . optional ( ) ,
label : z.string ( ) ,
cols : z.number ( ) . describe ( "Number of columns" ) ,
gap : z.number ( ) . optional ( ) ,
after : z.string ( ) . optional ( ) ,
} ) ,
feat(diagram-engine): flowcharts, swimlanes, sequence diagrams and mind maps
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.
2026-08-09 13:47:23 +09:00
z . object ( {
op : z.literal ( "add_pool" ) ,
id : z.string ( ) ,
parent : z.string ( ) . optional ( ) ,
label : z.string ( ) . describe ( "Pool title, e.g. the process name" ) ,
lanes : z
. array ( z . string ( ) )
. describe (
"Role names, one per lane, top to bottom. Steps go in these lanes via add_box lane/col" ,
) ,
phases : z
. array ( z . string ( ) )
. optional ( )
. describe ( "Milestone labels spanning the columns; omit for none" ) ,
orientation : z
. enum ( [ "horizontal" , "vertical" ] )
. optional ( )
. describe (
"horizontal (default): lanes stack down, flow goes right" ,
) ,
gap : z.number ( ) . optional ( ) ,
after : z.string ( ) . optional ( ) ,
} ) ,
z . object ( {
op : z.literal ( "add_sequence" ) ,
id : z.string ( ) ,
parent : z.string ( ) . optional ( ) ,
label : z.string ( ) . describe ( "Diagram title; empty string for none" ) ,
gap : z
. number ( )
. optional ( )
. describe ( "Horizontal spacing between participants" ) ,
step : z
. number ( )
. optional ( )
. describe ( "Vertical spacing between messages" ) ,
after : z.string ( ) . optional ( ) ,
} ) ,
z . object ( {
op : z.literal ( "add_radial" ) ,
id : z.string ( ) ,
parent : z.string ( ) . optional ( ) ,
label : z.string ( ) . describe ( "Frame title; empty string for none" ) ,
spread : z
. enum ( [ "radial" , "down" ] )
. optional ( )
. describe (
"radial (default): branches on both sides, for a mind map. down: everything below the centre, for an org chart" ,
) ,
gap : z.number ( ) . optional ( ) ,
after : z.string ( ) . optional ( ) ,
} ) ,
feat(diagram-engine): wire up restructure_diagram + stencil catalog
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.
2026-08-09 12:13:54 +09:00
z . object ( {
op : z.literal ( "remove" ) ,
id : z
. string ( )
. describe ( "Node to delete; its descendants and edges go too" ) ,
} ) ,
z . object ( {
op : z.literal ( "move" ) ,
id : z.string ( ) ,
parent : z
. string ( )
. optional ( )
. describe ( "New container id; omit to move to top level" ) ,
after : z.string ( ) . optional ( ) ,
} ) ,
z . object ( {
op : z.literal ( "set_label" ) ,
id : z.string ( ) ,
label : z.string ( ) ,
} ) ,
z . object ( {
op : z.literal ( "set_dir" ) ,
id : z.string ( ) . describe ( "Container to re-orient" ) ,
dir : z.enum ( [ "row" , "col" ] ) ,
} ) ,
z . object ( {
op : z.literal ( "set_gap" ) ,
id : z.string ( ) ,
gap : z.number ( ) ,
} ) ,
z . object ( {
op : z.literal ( "link" ) ,
source : z.string ( ) ,
target : z.string ( ) ,
label : z.string ( ) . optional ( ) ,
dashed : z
. boolean ( )
. optional ( )
. describe ( "Dashed line — replication, sync, policy" ) ,
step : z
. number ( )
. optional ( )
. describe ( "Step number, shown as an 'N. ' prefix" ) ,
} ) ,
z . object ( {
op : z.literal ( "unlink" ) ,
source : z.string ( ) ,
target : z.string ( ) ,
feat(diagram-engine): flowcharts, swimlanes, sequence diagrams and mind maps
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.
2026-08-09 13:47:23 +09:00
step : z
. number ( )
. optional ( )
. describe (
"Remove only the edge with this step number; omit to remove every edge between the two" ,
) ,
feat(diagram-engine): wire up restructure_diagram + stencil catalog
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.
2026-08-09 12:13:54 +09:00
} ) ,
z . object ( {
op : z.literal ( "set_title" ) ,
title : z.string ( ) ,
} ) ,
] )
export type Operation = z . infer < typeof OperationSchema >
export interface ApplyResult {
tree : DiagramTree
/** One entry per operation that could not be applied, in order. */
errors : string [ ]
}
feat(diagram-engine): flowcharts, swimlanes, sequence diagrams and mind maps
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.
2026-08-09 13:47:23 +09:00
/ * *
* The pool cell an add operation declared , if any .
*
* ` lane ` alone is enough — a step in a lane with no column given goes to column 0 — so the
* cell is recorded whenever either is present rather than requiring both .
* /
function cellOf ( op : { lane? : number ; col? : number } ) : {
cell ? : { lane : number ; col : number }
} {
if ( op . lane == null && op . col == null ) return { }
return {
cell : {
lane : Math.max ( 0 , Math . round ( op . lane ? ? 0 ) ) ,
col : Math.max ( 0 , Math . round ( op . col ? ? 0 ) ) ,
} ,
}
}
/** Are these two nodes participants of the same sequence diagram? */
function sameSequence ( tree : DiagramTree , a : string , b : string ) : boolean {
for ( const n of walkTree ( tree ) ) {
if ( n . kind !== "sequence" ) continue
const ids = new Set ( n . children . map ( ( c ) = > c . id ) )
if ( ids . has ( a ) && ids . has ( b ) ) return true
}
return false
}
feat(diagram-engine): wire up restructure_diagram + stencil catalog
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.
2026-08-09 12:13:54 +09:00
/** Insert into a child list, after a named sibling or at the end. */
function insert (
list : DiagramNode [ ] ,
node : DiagramNode ,
after : string | undefined ,
) : void {
if ( after ) {
const i = list . findIndex ( ( c ) = > c . id === after )
if ( i >= 0 ) {
list . splice ( i + 1 , 0 , node )
return
}
}
list . push ( node )
}
/** Detach a node from wherever it currently sits. Returns it, or null if not found. */
function detach ( tree : DiagramTree , id : string ) : DiagramNode | null {
const rootIdx = tree . roots . findIndex ( ( r ) = > r . id === id )
if ( rootIdx >= 0 ) return tree . roots . splice ( rootIdx , 1 ) [ 0 ]
const parent = findParent ( tree , id )
if ( ! parent ) return null
const i = parent . children . findIndex ( ( c ) = > c . id === id )
return i >= 0 ? parent . children . splice ( i , 1 ) [ 0 ] : null
}
/** Would making `id` a descendant of `parentId` create a cycle? */
function wouldCycle ( tree : DiagramTree , id : string , parentId : string ) : boolean {
if ( id === parentId ) return true
const node = findNode ( tree , id )
if ( ! node || ! isContainer ( node ) ) return false
for ( const d of walkTree ( { . . . tree , roots : [ node ] } ) )
if ( d . id === parentId ) return true
return false
}
/ * *
* Resolve where a new or moved node goes . Returns the child list to insert into , or an
* error string .
* /
function targetList (
tree : DiagramTree ,
parentId : string | undefined ,
) : DiagramNode [ ] | string {
if ( ! parentId ) return tree . roots
const p = findNode ( tree , parentId )
if ( ! p ) return ` No node with id " ${ parentId } " `
if ( ! isContainer ( p ) )
return ` " ${ parentId } " is a ${ p . kind } , not a container — it cannot hold children `
return p . children
}
/ * *
* Apply operations to a tree , in order .
*
* The input tree is deep - copied first : a partially - applied batch must not leave the
* caller ' s tree half - mutated when a later operation fails .
* /
export function applyOperations (
input : DiagramTree ,
ops : Operation [ ] ,
) : ApplyResult {
const tree : DiagramTree = structuredClone ( input )
const errors : string [ ] = [ ]
const exists = ( id : string ) = > findNode ( tree , id ) !== null
for ( const op of ops ) {
switch ( op . op ) {
case "add_icon" :
case "add_box" :
case "add_container" :
feat(diagram-engine): flowcharts, swimlanes, sequence diagrams and mind maps
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.
2026-08-09 13:47:23 +09:00
case "add_grid" :
case "add_pool" :
case "add_sequence" :
case "add_radial" : {
feat(diagram-engine): wire up restructure_diagram + stencil catalog
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.
2026-08-09 12:13:54 +09:00
if ( exists ( op . id ) ) {
errors . push ( ` ${ op . op } : id " ${ op . id } " is already taken ` )
break
}
const list = targetList ( tree , op . parent )
if ( typeof list === "string" ) {
errors . push ( ` ${ op . op } : ${ list } ` )
break
}
let node : DiagramNode
if ( op . op === "add_icon" )
node = {
kind : "icon" ,
id : op.id ,
name : op.name ,
label : op.label ? ? "" ,
feat(diagram-engine): flowcharts, swimlanes, sequence diagrams and mind maps
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.
2026-08-09 13:47:23 +09:00
. . . cellOf ( op ) ,
feat(diagram-engine): wire up restructure_diagram + stencil catalog
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.
2026-08-09 12:13:54 +09:00
}
else if ( op . op === "add_box" )
feat(diagram-engine): flowcharts, swimlanes, sequence diagrams and mind maps
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.
2026-08-09 13:47:23 +09:00
node = {
kind : "box" ,
id : op.id ,
label : op.label ,
. . . ( op . shape && op . shape !== "box"
? { shape : op.shape }
: { } ) ,
feat(diagram-engine): label avoidance, paired opposite edges, semantic group colours
A git-workflow flowchart rendered with no overlaps but read poorly. Three
distinct causes, each fixed and measured:
1. Edge labels sat on boxes and on each other (4 collisions on the
reported diagram; 280 across 250 generated flowcharts). The router
keeps LINES off the boxes but a label renders at its edge's midpoint,
which on a long edge is beside exactly the things the line was routed
around. placeLabels slides each label along its own edge to a clear
spot — longest edges first, midpoint-outward tries — written as the
geometry's relative x, which draw.io natively supports. Corpus: 280
label collisions -> 8.
2. A->B and B->A were routed independently, so "git add" ran straight
while "git reset" wandered through a different corridor with a kink.
Opposite edges that agree on axis now get two absolute parallel tracks
in the strip where the two boxes overlap, a constant 24px apart,
converted back to port fractions. Zero crossing regressions.
3. All boxes rendered the same white, because the render layer's
fill/stroke support was never reachable: neither add_box's schema nor
draw_graph's nodes exposed it. Rather than exposing raw hex (the model
picks mismatched saturations, differently every time), nodes take a
semantic group name and the engine maps groups to a fixed palette of
six paired fill/strokes in order of first appearance. The model names
the zones - remote vs local vs temp - and never touches a colour.
532 unit tests pass; the 5 diagram e2e tests pass in a real browser.
2026-08-09 18:59:34 +09:00
. . . ( op . fill ? { fill : op.fill } : { } ) ,
. . . ( op . stroke ? { stroke : op.stroke } : { } ) ,
feat(diagram-engine): design tokens + role/group composition, engine-wide theming
Paper-summary posters previously required hand-written XML: every engine
box rendered identically (white, 11px), so anything whose meaning lives
in visual hierarchy came out flat. This makes presentation a first-class,
generalised part of the declaration - not a poster feature.
Structure/presentation separation, the same split HTML and CSS settled on:
- ROLE says what a node IS: banner, heading, body, callout, good, bad,
metric, muted. Maps to a type scale and an emphasis (filled / tinted /
outlined / ghost), never to a colour.
- GROUP says which semantic zone a node belongs to. Each distinct group
name gets one hue ramp (tint / base / dark), assigned in document
order. Promoted from a draw_graph-only field to BoxNode and GroupNode,
round-tripped via dai_group.
- themedStyle(role, hue, kind) composes the two by rule - there is no
per-combination table to extend, so a new diagram kind gets full
theming by tagging nodes. The model never sees a hex value.
A heading container plus a group yields the tinted section panel with a
dark title; a grouped body box takes its zone's tint; verdict roles stay
green/red regardless of zone; the banner is the page's one dark field.
Also fixed, found while building the acceptance poster:
- autoBoxSize only counted explicit newlines, so a long single-line label
wrapped to six lines in draw.io but got a one-line-tall box, and the
text overflowed the cell.
- Marker stamping appended without replacing, so every render of a
recovered style grew it by one duplicate dai_* token per key -
unnoticed because draw.io resolves duplicates last-wins. dai_* keys
are now replaced in place; mxGraph keys still append, because
last-wins is load-bearing for container=1 normalisation.
- Banner/heading/metric roles stretch across their container's cross
axis, the way a masthead spans its page.
- Prompt: a poster's banner IS its title (no set_title alongside), and
sections get their colour by naming groups.
537 unit tests, 250-flowchart corpus still zero crossing arrows, 5 e2e
tests in a real browser. Verified visually: the Transformer-paper poster
renders with a navy masthead, three hue-coded section panels, metric,
verdict and callout boxes - all engine-computed geometry.
2026-08-09 19:48:01 +09:00
. . . ( op . role ? { role : op.role } : { } ) ,
. . . ( op . group ? { group : op.group } : { } ) ,
feat(diagram-engine): flowcharts, swimlanes, sequence diagrams and mind maps
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.
2026-08-09 13:47:23 +09:00
. . . cellOf ( op ) ,
}
feat(diagram-engine): wire up restructure_diagram + stencil catalog
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.
2026-08-09 12:13:54 +09:00
else if ( op . op === "add_container" )
node = {
kind : "group" ,
id : op.id ,
gname : op.gname ? ? null ,
label : op.label ,
dir : op.dir ,
gap : op.gap ? ? 20 ,
children : [ ] ,
feat(diagram-engine): design tokens + role/group composition, engine-wide theming
Paper-summary posters previously required hand-written XML: every engine
box rendered identically (white, 11px), so anything whose meaning lives
in visual hierarchy came out flat. This makes presentation a first-class,
generalised part of the declaration - not a poster feature.
Structure/presentation separation, the same split HTML and CSS settled on:
- ROLE says what a node IS: banner, heading, body, callout, good, bad,
metric, muted. Maps to a type scale and an emphasis (filled / tinted /
outlined / ghost), never to a colour.
- GROUP says which semantic zone a node belongs to. Each distinct group
name gets one hue ramp (tint / base / dark), assigned in document
order. Promoted from a draw_graph-only field to BoxNode and GroupNode,
round-tripped via dai_group.
- themedStyle(role, hue, kind) composes the two by rule - there is no
per-combination table to extend, so a new diagram kind gets full
theming by tagging nodes. The model never sees a hex value.
A heading container plus a group yields the tinted section panel with a
dark title; a grouped body box takes its zone's tint; verdict roles stay
green/red regardless of zone; the banner is the page's one dark field.
Also fixed, found while building the acceptance poster:
- autoBoxSize only counted explicit newlines, so a long single-line label
wrapped to six lines in draw.io but got a one-line-tall box, and the
text overflowed the cell.
- Marker stamping appended without replacing, so every render of a
recovered style grew it by one duplicate dai_* token per key -
unnoticed because draw.io resolves duplicates last-wins. dai_* keys
are now replaced in place; mxGraph keys still append, because
last-wins is load-bearing for container=1 normalisation.
- Banner/heading/metric roles stretch across their container's cross
axis, the way a masthead spans its page.
- Prompt: a poster's banner IS its title (no set_title alongside), and
sections get their colour by naming groups.
537 unit tests, 250-flowchart corpus still zero crossing arrows, 5 e2e
tests in a real browser. Verified visually: the Transformer-paper poster
renders with a navy masthead, three hue-coded section panels, metric,
verdict and callout boxes - all engine-computed geometry.
2026-08-09 19:48:01 +09:00
. . . ( op . role ? { role : op.role } : { } ) ,
. . . ( op . group ? { group : op.group } : { } ) ,
feat(diagram-engine): wire up restructure_diagram + stencil catalog
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.
2026-08-09 12:13:54 +09:00
}
feat(diagram-engine): flowcharts, swimlanes, sequence diagrams and mind maps
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.
2026-08-09 13:47:23 +09:00
else if ( op . op === "add_grid" )
feat(diagram-engine): wire up restructure_diagram + stencil catalog
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.
2026-08-09 12:13:54 +09:00
node = {
kind : "grid" ,
id : op.id ,
gname : null ,
label : op.label ,
cols : Math.max ( 1 , op . cols ) ,
gap : op.gap ? ? 14 ,
children : [ ] ,
}
feat(diagram-engine): flowcharts, swimlanes, sequence diagrams and mind maps
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.
2026-08-09 13:47:23 +09:00
else if ( op . op === "add_pool" ) {
if ( op . lanes . length === 0 ) {
errors . push (
` add_pool: " ${ op . id } " needs at least one lane — a swimlane diagram with no roles has nothing to divide ` ,
)
break
}
node = {
kind : "pool" ,
id : op.id ,
label : op.label ,
lanes : op.lanes ,
phases : op.phases ? ? [ ] ,
orientation : op.orientation ? ? "horizontal" ,
gap : op.gap ? ? 40 ,
children : [ ] ,
}
} else if ( op . op === "add_sequence" )
node = {
kind : "sequence" ,
id : op.id ,
label : op.label ,
gap : op.gap ? ? 60 ,
step : Math.max ( 24 , op . step ? ? 44 ) ,
children : [ ] ,
}
else
node = {
kind : "radial" ,
id : op.id ,
label : op.label ,
spread : op.spread ? ? "radial" ,
gap : op.gap ? ? 40 ,
children : [ ] ,
}
feat(diagram-engine): wire up restructure_diagram + stencil catalog
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.
2026-08-09 12:13:54 +09:00
insert ( list , node , op . after )
break
}
case "remove" : {
const node = findNode ( tree , op . id )
if ( ! node ) {
errors . push ( ` remove: no node with id " ${ op . id } " ` )
break
}
// Collect the subtree's ids first — edges touching any of them go too,
// otherwise draw.io renders an arrow pointing at nothing.
const doomed = new Set < string > ( )
for ( const d of walkTree ( { . . . tree , roots : [ node ] } ) )
doomed . add ( d . id )
detach ( tree , op . id )
tree . links = tree . links . filter (
( l ) = > ! doomed . has ( l . source ) && ! doomed . has ( l . target ) ,
)
break
}
case "move" : {
if ( ! exists ( op . id ) ) {
errors . push ( ` move: no node with id " ${ op . id } " ` )
break
}
if ( op . parent && ! exists ( op . parent ) ) {
errors . push ( ` move: no node with id " ${ op . parent } " ` )
break
}
if ( op . parent && wouldCycle ( tree , op . id , op . parent ) ) {
errors . push (
` move: cannot move " ${ op . id } " into " ${ op . parent } " — that is inside itself ` ,
)
break
}
const list = targetList ( tree , op . parent )
if ( typeof list === "string" ) {
errors . push ( ` move: ${ list } ` )
break
}
const node = detach ( tree , op . id )
if ( ! node ) {
errors . push ( ` move: could not detach " ${ op . id } " ` )
break
}
insert ( list , node , op . after )
break
}
case "set_label" : {
const node = findNode ( tree , op . id )
if ( ! node ) {
errors . push ( ` set_label: no node with id " ${ op . id } " ` )
break
}
if ( node . kind === "title" ) {
errors . push (
` set_label: use set_title to change the page title ` ,
)
break
}
node . label = op . label
break
}
case "set_dir" : {
const node = findNode ( tree , op . id )
if ( ! node || ! isContainer ( node ) ) {
errors . push ( ` set_dir: " ${ op . id } " is not a container ` )
break
}
refactor(diagram-engine): apply review findings, fix vertical pool phases
Four reviewers went over the previous commit (three Claude, one Codex). Their
findings, verified independently before applying:
A REAL BUG. A vertical pool with milestone labels drew the label strip outside
the pool frame. The measure pass reserves width as padding + content + strip with
no gap between the last two; the renderer placed the strip one gap further out.
No test caught it because every vertical case omitted phases and every phases
case was horizontal — both regression cases added.
Duplicated logic, now single-sourced:
- messageCount existed byte-identically in layout.ts and render.ts. Two copies
that had to agree or the lifelines stop reaching the last message.
- sequenceMetrics was called twice per sequence container, once inside the
chrome builder and again for the message positions. Same drift hazard, in the
file whose own comment warns about it.
Dead code, each verified unreachable rather than assumed:
- Placed.extent: declared and documented, never written or read. Every .extent
access belongs to RadialTree.
- SequenceMetrics.top: computed, returned, no reader.
- spread()'s level parameter: threaded through the recursion, never used.
- radialReach's .slice(0, generations): widestPerLevel writes one entry per
generation, so its length IS the depth. Confirmed over 20,000 random trees;
removing it made RadialTree.depth dead too.
- Two of three cycle guards in radialHierarchy: self-links are already skipped
when the parent map is built, and that map holds one parent per node, so the
structure is a forest and the visited-set filter cannot fire. The rootOf
guard does fire and stays.
- GraphOptions.layerGap/nodeGap/idPrefix: no caller, not in the tool schema.
Simplifications:
- LayoutContext wrapped a single field; the link array now passes directly,
which also removes the NO_CONTEXT default no call site ever took.
- stretches() and the mirror-image check five lines below it expressed one rule
two ways; unified, with the rationale stated once.
- hasStencilFrame/isDirectional: one caller each, and isDirectional's name
contradicted its body, which the guarded branch then re-discriminated anyway.
- poolFrameStyle() took no arguments and had one caller.
- poolCellOf clamped a value already clamped at the model boundary and
unreachable-by-construction from the parser.
- A comment on stampPoolDecoration described container behaviour the function
does not implement.
Kept deliberately, with evidence:
- The best-arrangement tracking in the crossing reducer. Two reviewers
suspected it was dead weight. Measured: barycentre sweeping regressed below
its own running best in 180 of 500 random graphs, so without it a third of
flowcharts would keep a worse arrangement than one already found.
- Vertical pools. Two reviewers recommended deleting the feature as
undiscoverable. The bug was one line, and vertical swimlanes are a real
convention — documented to the model instead, which is what was actually
missing.
- styleValue duplicating readMarker, isLeaf, findPageIndex: all genuinely
redundant, all predating this branch. Left alone to keep the diff scoped.
525 unit tests and 11 diagram e2e tests pass.
2026-08-09 14:47:46 +09:00
if ( node . kind !== "group" ) {
feat(diagram-engine): flowcharts, swimlanes, sequence diagrams and mind maps
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.
2026-08-09 13:47:23 +09:00
// A grid, pool, sequence or radial container arranges its children by its
// own rule; "row or column" is not a property they have.
feat(diagram-engine): wire up restructure_diagram + stencil catalog
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.
2026-08-09 12:13:54 +09:00
errors . push (
feat(diagram-engine): flowcharts, swimlanes, sequence diagrams and mind maps
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.
2026-08-09 13:47:23 +09:00
node . kind === "grid"
? ` set_dir: " ${ op . id } " is a grid — change its column count instead `
: ` set_dir: " ${ op . id } " is a ${ node . kind } , which arranges its children by its own rule and has no row/column direction ` ,
feat(diagram-engine): wire up restructure_diagram + stencil catalog
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.
2026-08-09 12:13:54 +09:00
)
break
}
node . dir = op . dir
break
}
case "set_gap" : {
const node = findNode ( tree , op . id )
if ( ! node || ! isContainer ( node ) ) {
errors . push ( ` set_gap: " ${ op . id } " is not a container ` )
break
}
; ( node as ContainerNode ) . gap = Math . max ( 0 , op . gap )
break
}
case "link" : {
if ( ! exists ( op . source ) ) {
errors . push ( ` link: no node with id " ${ op . source } " ` )
break
}
if ( ! exists ( op . target ) ) {
errors . push ( ` link: no node with id " ${ op . target } " ` )
break
}
feat(diagram-engine): flowcharts, swimlanes, sequence diagrams and mind maps
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.
2026-08-09 13:47:23 +09:00
// A second arrow between the same pair is normally a mistake — two identical
// lines drawn on top of each other — EXCEPT between two participants of a
// sequence diagram, where a back-and-forth conversation is the whole point.
// There the messages are distinguished by their step, not by their endpoints.
const conversation = sameSequence ( tree , op . source , op . target )
const dup =
! conversation &&
tree . links . some (
( l ) = > l . source === op . source && l . target === op . target ,
)
feat(diagram-engine): wire up restructure_diagram + stencil catalog
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.
2026-08-09 12:13:54 +09:00
if ( dup ) {
errors . push (
` link: " ${ op . source } " → " ${ op . target } " already exists ` ,
)
break
}
const link : LinkSpec = { source : op.source , target : op.target }
if ( op . label ) link . label = op . label
if ( op . dashed ) link . dashed = true
if ( op . step != null ) link . step = op . step
tree . links . push ( link )
break
}
case "unlink" : {
const before = tree . links . length
feat(diagram-engine): flowcharts, swimlanes, sequence diagrams and mind maps
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.
2026-08-09 13:47:23 +09:00
// With a step given, remove only that message: two participants of a sequence
// diagram can exchange several, and dropping all of them would delete messages
// the caller did not ask about.
feat(diagram-engine): wire up restructure_diagram + stencil catalog
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.
2026-08-09 12:13:54 +09:00
tree . links = tree . links . filter (
feat(diagram-engine): flowcharts, swimlanes, sequence diagrams and mind maps
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.
2026-08-09 13:47:23 +09:00
( l ) = >
! (
l . source === op . source &&
l . target === op . target &&
( op . step == null || l . step === op . step )
) ,
feat(diagram-engine): wire up restructure_diagram + stencil catalog
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.
2026-08-09 12:13:54 +09:00
)
if ( tree . links . length === before )
errors . push (
feat(diagram-engine): flowcharts, swimlanes, sequence diagrams and mind maps
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.
2026-08-09 13:47:23 +09:00
op . step == null
? ` unlink: no edge from " ${ op . source } " to " ${ op . target } " `
: ` unlink: no edge from " ${ op . source } " to " ${ op . target } " with step ${ op . step } ` ,
feat(diagram-engine): wire up restructure_diagram + stencil catalog
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.
2026-08-09 12:13:54 +09:00
)
break
}
case "set_title" :
tree . title = op . title
break
}
}
return { tree , errors }
}
/** Every icon/group name in a tree, for validating against the catalog. */
export function collectNames (
tree : DiagramTree ,
) : { id : string ; name : string ; kind : "icon" | "group" } [ ] {
const out : { id : string ; name : string ; kind : "icon" | "group" } [ ] = [ ]
for ( const n of walkTree ( tree ) ) {
if ( n . kind === "icon" && n . name )
out . push ( { id : n.id , name : n.name , kind : "icon" } )
refactor(diagram-engine): apply review findings, fix vertical pool phases
Four reviewers went over the previous commit (three Claude, one Codex). Their
findings, verified independently before applying:
A REAL BUG. A vertical pool with milestone labels drew the label strip outside
the pool frame. The measure pass reserves width as padding + content + strip with
no gap between the last two; the renderer placed the strip one gap further out.
No test caught it because every vertical case omitted phases and every phases
case was horizontal — both regression cases added.
Duplicated logic, now single-sourced:
- messageCount existed byte-identically in layout.ts and render.ts. Two copies
that had to agree or the lifelines stop reaching the last message.
- sequenceMetrics was called twice per sequence container, once inside the
chrome builder and again for the message positions. Same drift hazard, in the
file whose own comment warns about it.
Dead code, each verified unreachable rather than assumed:
- Placed.extent: declared and documented, never written or read. Every .extent
access belongs to RadialTree.
- SequenceMetrics.top: computed, returned, no reader.
- spread()'s level parameter: threaded through the recursion, never used.
- radialReach's .slice(0, generations): widestPerLevel writes one entry per
generation, so its length IS the depth. Confirmed over 20,000 random trees;
removing it made RadialTree.depth dead too.
- Two of three cycle guards in radialHierarchy: self-links are already skipped
when the parent map is built, and that map holds one parent per node, so the
structure is a forest and the visited-set filter cannot fire. The rootOf
guard does fire and stays.
- GraphOptions.layerGap/nodeGap/idPrefix: no caller, not in the tool schema.
Simplifications:
- LayoutContext wrapped a single field; the link array now passes directly,
which also removes the NO_CONTEXT default no call site ever took.
- stretches() and the mirror-image check five lines below it expressed one rule
two ways; unified, with the rationale stated once.
- hasStencilFrame/isDirectional: one caller each, and isDirectional's name
contradicted its body, which the guarded branch then re-discriminated anyway.
- poolFrameStyle() took no arguments and had one caller.
- poolCellOf clamped a value already clamped at the model boundary and
unreachable-by-construction from the parser.
- A comment on stampPoolDecoration described container behaviour the function
does not implement.
Kept deliberately, with evidence:
- The best-arrangement tracking in the crossing reducer. Two reviewers
suspected it was dead weight. Measured: barycentre sweeping regressed below
its own running best in 180 of 500 random graphs, so without it a third of
flowcharts would keep a worse arrangement than one already found.
- Vertical pools. Two reviewers recommended deleting the feature as
undiscoverable. The bug was one line, and vertical swimlanes are a real
convention — documented to the model instead, which is what was actually
missing.
- styleValue duplicating readMarker, isLeaf, findPageIndex: all genuinely
redundant, all predating this branch. Left alone to keep the diff scoped.
525 unit tests and 11 diagram e2e tests pass.
2026-08-09 14:47:46 +09:00
else if ( ( n . kind === "group" || n . kind === "grid" ) && n . gname )
feat(diagram-engine): wire up restructure_diagram + stencil catalog
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.
2026-08-09 12:13:54 +09:00
out . push ( { id : n.id , name : n.gname , kind : "group" } )
}
return out
}
feat(diagram-engine): flowcharts, swimlanes, sequence diagrams and mind maps
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.
2026-08-09 13:47:23 +09:00
/** How a container arranges its children, in one short phrase for the outline. */
function containerMeta ( n : ContainerNode ) : string {
switch ( n . kind ) {
case "grid" :
return ` grid cols= ${ n . cols } `
case "pool" :
return ` pool lanes=[ ${ n . lanes . join ( " | " ) } ] ${
n . phases . length ? ` phases=[ ${ n . phases . join ( " | " ) } ] ` : ""
} $ { n . orientation === "vertical" ? " vertical" : "" } `
case "sequence" :
return "sequence"
case "radial" :
return ` radial ${ n . spread } `
default :
return n . dir
}
}
feat(diagram-engine): wire up restructure_diagram + stencil catalog
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.
2026-08-09 12:13:54 +09:00
/ * *
* A compact text outline of the tree , for showing the model what is on the canvas .
*
* Sending the tree as JSON would cost several times more for the same information , and
* the model does not need coordinates — it needs to know what exists and how it nests so
* it can name ids in the next operation .
* /
export function outline ( tree : DiagramTree ) : string {
const lines : string [ ] = [ ]
if ( tree . title ) lines . push ( ` title: ${ tree . title } ` )
const walk = ( n : DiagramNode , depth : number ) = > {
const pad = " " . repeat ( depth )
feat(diagram-engine): flowcharts, swimlanes, sequence diagrams and mind maps
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.
2026-08-09 13:47:23 +09:00
// A node inside a pool reports its cell: that is how the model knows which lane a
// step ended up in, which is exactly what it needs to move one.
const at = ( x : DiagramNode ) = >
( x . kind === "icon" || x . kind === "box" ) && x . cell
? ` @lane ${ x . cell . lane } ,col ${ x . cell . col } `
: ""
feat(diagram-engine): wire up restructure_diagram + stencil catalog
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.
2026-08-09 12:13:54 +09:00
if ( n . kind === "icon" )
lines . push (
feat(diagram-engine): flowcharts, swimlanes, sequence diagrams and mind maps
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.
2026-08-09 13:47:23 +09:00
` ${ pad } ${ n . id } : icon ${ n . name } ${ n . label ? ` " ${ n . label } " ` : "" } ${ at ( n ) } ` ,
)
else if ( n . kind === "box" )
lines . push (
` ${ pad } ${ n . id } : box ${ n . shape ? ` ${ n . shape } ` : "" } " ${ n . label } " ${ at ( n ) } ` ,
feat(diagram-engine): wire up restructure_diagram + stencil catalog
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.
2026-08-09 12:13:54 +09:00
)
else if ( n . kind === "title" ) lines . push ( ` ${ pad } ${ n . id } : title ` )
else {
lines . push (
feat(diagram-engine): flowcharts, swimlanes, sequence diagrams and mind maps
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.
2026-08-09 13:47:23 +09:00
` ${ pad } ${ n . id } : ${ containerMeta ( n ) } ${ n . label ? ` " ${ n . label } " ` : " (wrapper)" } ` ,
feat(diagram-engine): wire up restructure_diagram + stencil catalog
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.
2026-08-09 12:13:54 +09:00
)
for ( const c of n . children ) walk ( c , depth + 1 )
}
}
for ( const r of tree . roots ) walk ( r , 0 )
for ( const l of tree . links ) {
const bits = [ l . label , l . dashed ? "dashed" : null ]
. filter ( Boolean )
. join ( ", " )
lines . push ( ` link ${ l . source } -> ${ l . target } ${ bits ? ` ( ${ bits } ) ` : "" } ` )
}
if ( tree . foreign . length )
lines . push (
` ${ tree . foreign . length } cell(s) kept as-is: ${ tree . foreign . map ( ( f ) = > f . id ) . join ( ", " ) } ` ,
)
return lines . join ( "\n" )
}