A user's approval-workflow flowchart came out with the return arrow drawn
straight through two unrelated steps, and a second report showed arrows
leaving a box and bending straight back across it. Measured over 250
generated flowcharts (2722 edges): 347 arrows crossed an unrelated box and
215 waypoints landed inside a shape. Four defects, each measured in
isolation:
1. The invisible layer containers draw_graph emits were handed to the
router as frames, so every clean return path was rejected for
"trespassing" on a border that is not drawn, and the fallback cut
through two boxes. Excluding invisible containers: 347 -> 218 crossing
arrows, no diagram made worse.
2. The router chose the horizontal-vs-vertical axis BEFORE searching, so
when the only clean corridor ran along the other axis it was never
looked at. A complete two-bend candidate generator that tries both
trunk axes, all four sides at each end, and the port fractions:
218 -> 27. (An independent ablation measured the axis pre-choice alone
at a 40% per-edge failure rate.)
3. Nothing stopped a route's first leg from turning back across its own
source shape - the obstacle test exempts an edge's own endpoints, and
must, since the line has to touch them. A terminal-leg rule refuses
such routes outright: 215 -> 4 hooks.
4. A two-bend search cannot express the staircase needed when a box sits
directly between two vertically aligned nodes (21 of the last 27
crossings). Added the orthogonal visibility graph + A* from Wybrow,
Marriott & Stuckey, "Orthogonal Connector Routing" (GD 2009) - the
libavoid algorithm - as the backstop when the candidate search finds
nothing. The interesting-points grid is provably sufficient: any valid
route shrinks onto it without getting longer or gaining bends. The A*
state is (point, incoming direction) with libavoid's bend cost of 10,
and the admissible bends-remaining heuristic, so it returns a cheapest
route, not merely a route. Implemented from the paper, not ported.
After all four: 0 crossing arrows and 0 hooks over the same 250 diagrams,
page area unchanged (494k px^2 mean), 260ms for the whole corpus, mean
0.82 bends per edge. The shape ladder still runs first, so routes that
were already clean are byte-identical.
Also post-nudge validation now checks the whole path (the nudge pass only
reverts the single segment it moved, judged in isolation) and restores the
search's route if nudging made it dirty.
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.
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.
Fixes the reported problem: arrows overlapped on top of icons and ran through shapes
they had nothing to do with.
The cause was a missing layer, not a bug. render.ts emitted only source and target, so
draw.io routed every edge itself — and its router sees the two terminals' bounds and
nothing else, not where the other icons are. It therefore ran lines straight through
whatever was in between and left several edges leaving one node at the same point.
Ported from drawio-ai-kit's router (MIT, see NOTICE):
- Port de-collision. Edges leaving the same side of the same node spread along it,
ordered by where their far end sits so they do not cross on the way out. An edge
with a clean straight shot keeps the centre.
- Obstacle avoidance. Candidate shapes in order of directness — straight, a Z through
the gap, an L, a two-bend detour — each tested against every icon on the page.
- Frame placement. A frame is passable (an edge into a VPC must cross its border) but
not free: running alongside a border, or cutting through a frame that holds only one
of the two endpoints, is penalised.
- Port snapping. On a bent route, each port moves to the side its leg actually arrives
from. Without this the terminal segment can pierce the icon to reach a far-side port.
- Lane claiming. Each routed edge records the lanes it occupies so later edges avoid
them, rather than colliding and being pulled apart afterwards.
- Global nudge, three passes, reverting any move that makes a path worse.
Two things I got wrong on the way, both caught by looking at real geometry:
- The Z corridor was computed as min/max of both nodes' edges, which spans the whole
distance between them — including anything parked in between. So the lane sweep
would place the detour's middle leg on top of the very icon it was avoiding. It has
to be the gap: trailing edge of the first node to the leading edge of the other.
- The router was fed layout's slot rectangles, but an icon's cell is the glyph square
centred in a slot roughly twice as wide. Collision tests against slots both missed
real overlaps and invented false ones. Extracted cellRect() so the router and the
emitted XML cannot diverge.
Accept-or-reject was not enough on its own. When one endpoint is inside a VPC and the
other outside, EVERY path trespasses on that frame, so the strict rule always fails and
the relaxed pass took whatever it tried first — which is how a line ended up cutting
across a whole VPC. Candidates are now scored (frame offences 500, lane sharing 700,
bends 80, length 1) and the cheapest wins, so an edge that must trespass still gets the
least-bad route.
Waypoints are still written only when load-bearing — a labelled bend or a deliberate
detour — so an unobstructed edge stays drag-friendly.
455 unit tests (27 new, asserting produced geometry rather than algorithm shape).
Verified by rendering the reported diagram in the real editor: the two arrows that
overlapped on the EC2 icon now leave from different sides, and the load balancer's
fan-out leaves from three distinct points.
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.
Completes the tree → coordinates → XML direction, so the model can declare nesting
and never write a coordinate or an mxCell again.
layout.ts — measure bottom-up, place top-down, the same shape as flexbox. A
container sums its children along the flow axis and adds padding, so "child spills
out of its frame" and "siblings overlap" cannot happen by construction rather than
being caught afterwards. Slack from sibling equalisation is shared between children
instead of left as dead margin, capped at one gap so a stretched frame reads as
spaced rather than sparse.
render.ts — writes the mxCells, stamping container=1 and the dai_* markers so
parse.ts can read the structure back. Edges carry no waypoints: draw.io's own router
recomputes the route on every edit, so a user who moves a node never has to re-link
an arrow. Cells the parser could not interpret are re-emitted verbatim, so a
re-layout never deletes a user's annotations.
Phantoms are gone (task #5). The reference project's layout-only wrapper emits no
cell, which makes the round-trip lossy by construction — measured on its own
build_vpc.mjs, a phantom erased a container's "col" direction for good. An
unlabelled frame here emits a real cell with fillColor/strokeColor=none instead:
invisible, but present in the XML and therefore recoverable.
Two bugs the round-trip test caught, both real:
- An icon's cell was being emitted at its measured slot size, which includes room
for the label underneath. Parsing read that width back as the glyph size, so the
icon grew on every round-trip. The cell is now the glyph square and the label
renders outside it via verticalLabelPosition, as the reference does.
- An Azure or GCP icon is an embedded base64 image whose style contains no name
anywhere, so the catalog name was unrecoverable. Added a dai_name marker.
Verified in a real browser (3 Playwright tests, not mocks): engine output renders in
draw.io; dragging a shape into a frame makes draw.io rewrite its parent and the
engine reads the new structure back; re-laying out from that structure PRESERVES the
user's move instead of undoing it, and leaves untouched nodes alone; and the
re-laid-out XML still renders.
That last point is the whole design: there is no second copy of the state, so a
manual edit is an input to the next layout rather than a conflict to reconcile.
304 unit tests + 3 e2e.
Groundwork for a declarative diagram engine where the model declares nesting and
the engine computes every coordinate, instead of the model emitting raw mxCell XML.
The design keeps the canvas as the SINGLE source of truth: the node tree is never
persisted, it is re-derived from the current canvas XML whenever needed. A user's
manual edits are therefore an input to the next re-layout, not a second copy of
the state that has to be reconciled.
Two behaviours this relies on, both verified against the real embedded editor with
a Playwright mouse drag (reading the editor's own autosave payload):
1. An AWS group stencil WITHOUT container=1 does not get a dragged shape
reparented — parent stays "1" and geometry stays absolute. WITH container=1
it does: parent becomes the frame, geometry becomes parent-relative. So the
engine must stamp container=1 on every container it emits.
2. draw.io preserves style keys it does not understand, and resolves a duplicate
key last-wins. So dai_* markers survive a user edit, and container=1 can be
appended to a catalog style without first parsing out an existing value —
which matters because the AWS catalog is inconsistent about it (group_vpc,
group_region, group_subnet ship without it; group_account ships with it).
markers.ts — dai_kind / dai_dir / dai_gap / dai_cols / dai_pin, and the container
token normalisation.
types.ts — the node tree contract, plus a `foreign` bucket so cells the engine
does not understand (user annotations, imported shapes) round-trip
verbatim rather than being destroyed by a re-layout.
parse.ts — XML → tree. Handles all four icon encodings (resIcon=, bare shape=,
shape=image data URI, grIcon=), resolves nesting from parent with a
geometry fallback for frames that lack container=1, recovers layout
direction from the marker or infers it from child positions, and
survives cycles, compressed files and multi-page decks.
75 unit tests, including a round-trip against real output from the reference
project's build_vpc.mjs.
One finding worth recording: the reference project's "phantom" node (a wrapper that
participates in layout but emits no cell) makes the round-trip lossy by
construction. In build_vpc.mjs a phantom erased a container's col direction — its
children were reparented onto the grandparent, leaving a 2-D arrangement the parser
can only read as a grid. 26 of the reference project's 31 examples use phantoms, so
our engine needs an invisible-but-real container instead. Tracked separately.
- Try legacy port (61337) first to preserve existing users' localStorage,
fall back to 13370 which is below the Windows Hyper-V ephemeral range (#705)
- Bind server and all URL references to 127.0.0.1 instead of localhost
to fix IPv4/IPv6 mismatch on Linux (#684)
- Add OS-assigned port fallback (port 0) so startup never throws
- Log error codes in port checks for easier debugging
- Update localhost guards in index.ts and window-manager.ts to also
match 127.0.0.1
Related: #705, #684
Previously, Ollama only used the OLLAMA_BASE_URL environment variable.
Now client-provided base URL from settings takes priority, allowing
users to configure custom Ollama endpoints (e.g., remote servers).
Fixes#652
Allows users to select Ollama as a provider from client settings.
Previously, Ollama was blocked with "Invalid provider" error even
though the UI supported it.
Fixes#652
Ollama is a local/self-hosted model that doesn't require API keys.
The SSRF protection was incorrectly blocking Ollama connections
when users provided a custom base URL without an API key.
Fixes#652
The previous logic only restored on iframe ready, but chartXML might
be set AFTER the iframe is ready (session loaded after iframe).
Now we track the last restored XML and load whenever chartXML changes
to a new real diagram while iframe is ready.
Fixes diagram not restoring after refresh. The issue was:
1. DrawIoEmbed rendered with canPersist=false
2. Session restore loaded diagram into iframe
3. canPersist changed to true, causing iframe remount
4. New iframe loaded empty, losing the diagram
Now we wait for isIndexedDBUsable() to complete before rendering,
avoiding the remount entirely.
Address review comments: canPersist is set async after mount, so DrawIoEmbed
needs to remount when it resolves to apply correct configuration and URL params.
- Enable draw.io autosave and handle autosave events to update chartXML
- Clear modified state after autosave to avoid beforeunload prompts
- Disable confirmExit in draw.io configuration
- Set modified=false and keepmodified=false URL parameters
- Fix session save condition to also save when only diagram exists
- Fix: Do not close shared IndexedDB connection in isIndexedDBUsable()
- Replace forwardRef/useImperativeHandle with prop-based focus control
- Add shouldFocus and onFocused props to ChatInput
- Use useEffect with setTimeout for proper cleanup
- Removes unnecessary complexity while maintaining same functionality
Refactored all 11 providers to use the resolveBaseURL() utility function
instead of inline ternary expressions. This ensures:
1. The security fix is centralized in one testable function
2. Unit tests actually validate the production code path
3. Future changes only need to modify one location
Providers refactored: openai, anthropic, google, azure, openrouter,
deepseek, siliconflow, sglang, gateway, doubao, modelscope
Add comprehensive tests for the resolveBaseURL utility function:
- Tests for user-provided API key scenarios
- Tests for server credential scenarios
- Edge case tests for empty strings and undefined values
This addresses the Copilot review suggestion to add test coverage
for the critical security fix.
When users provide their own API key but not a custom baseURL,
the code was incorrectly falling back to the server's environment
variable for baseURL. This caused user API keys to be sent to
the server's custom proxy endpoint instead of the provider's
official endpoint, resulting in 'API key format incorrect' errors.
This fix ensures that when a user provides their own API key:
- Only the user's baseUrl is used (if provided)
- Otherwise, the provider's official/default endpoint is used
- Server's baseURL env vars are never mixed with user credentials
Affected providers: openai, anthropic, google, azure, openrouter,
deepseek, siliconflow, sglang, gateway, doubao, modelscope
Also fixes Azure's resourceName to not leak server config to user keys.
Fixes#577
Keep only essential history integration:
- Import addHistory from history.js
- Remove unused getServerPort import
- Add browser state sync and history saving in display_diagram
- Add history saving in edit_diagram
No changes to prompts, descriptions, or code style.
- Sync browser state before saving history in display_diagram
- Save AI result to history (in addition to state before)
- Add SVG capture after browser loads AI diagrams
- Add /api/history-svg endpoint to update last entry's SVG
- Add updateLastHistorySvg() function to history module
- Reduce history.ts from 169 to 51 lines
- Remove AI tools (list_history, restore_version, get_version)
- Remove /api/update-svg endpoint
- Remove 10-second history polling
- Simplify HistoryEntry to just {xml, svg}
- Use array index instead of version numbers
Total reduction: 1936 → 923 lines (-52%)
- Add history.ts module with circular buffer (max 50 entries)
- Add history UI with floating button and modal
- Add HTTP endpoints: /api/history, /api/restore
- Add MCP tools: list_history, restore_version, get_version
- Save history before and after AI changes
- Track source (ai/human) for each entry
When images are included in chat messages, the AI SDK telemetry with
recordInputs: true sends base64 image data to Langfuse. Langfuse then
attempts to upload these images to media storage, causing 1m31s timeouts.
Setting recordInputs: false prevents this while still capturing user
text input via setTraceInput().
The React state update (setChartXML) is async, so chartXMLRef wasn't updated
when edit_diagram tool callback checked it. Now we update the ref directly
in onFormSubmit, handleRegenerate, and handleEditMessage before sending.
DrawIO iframe export was unreliable on Vercel due to network latency,
causing edit_diagram tool to hang. Now uses chartXML from context directly,
falling back to export only when no cached XML exists.
Bedrock streaming responses don't auto-report token usage to OpenTelemetry.
This fix manually sets span attributes (ai.usage.promptTokens, gen_ai.usage.input_tokens)
from the AI SDK onFinish callback to ensure Langfuse captures token counts.
- Add Zod schema validation for log-feedback and log-save endpoints
- Create singleton LangfuseClient to avoid per-request instantiation
- Simplify log-save to only flag trace (no XML content sent)
- Use generic error messages to prevent info leakage
- Update log-feedback API to find existing chat trace by sessionId and attach score to it
- Update log-save API to create span on existing chat trace instead of standalone trace
- Add thumbs up/down feedback buttons on assistant messages
- Add message regeneration and edit functionality
- Add save dialog with format selection (drawio, png, svg)
- Pass sessionId through components for Langfuse linking
- Install @ai-sdk/deepseek package
- Add DeepSeek provider support to lib/ai-providers.ts
- Add DeepSeek configuration to env.example
- Update README.md with DeepSeek in provider list
- Support both default and custom base URL for DeepSeek
- Explicitly require mcp inline comment tool for each issue
- Clarify gh pr comment is only for final summary
- Forbid dismissing issues as minor/harmless
- Add Examples section to README with 2-column grid layout
- Include demo images for GCP, AWS, Azure, animated connectors, and cat
- Update example panel buttons with clearer labels
- Add animated connector example button
- Add instruction for animated connectors in chat route
- Remove header from main page for clean editor-only interface
- Create /app/about/page.tsx with comprehensive SEO content (1000+ words)
- Add About link next to 'Next-AI-Drawio' title in chat panel
- Add GitHub icon link to /about page navigation
- Update sitemap.ts to include /about page (priority: 0.8)
SEO improvements following industry best practices:
- Separate marketing content from app interface (Figma/Canva/Miro approach)
- Server-rendered /about page for optimal crawlability
- Clean URL structure for better internal linking
- Multiple indexable pages for broader keyword coverage
- Proper semantic HTML: H1, H2, H3, article, section tags
- 1000+ words of keyword-rich content
/about page includes:
- AI diagram generator overview with value proposition
- 6 detailed feature sections (AI creation, AWS diagrams, image replication, etc.)
- 3 popular use cases (AWS architecture, flowcharts, system design)
- Step-by-step usage guide (4 steps)
- Benefits section (save time, precision, free, privacy)
- Clear call-to-action with link back to editor
- GitHub link in navigation for social proof
This follows Google-approved architecture and avoids hidden content penalties.
- Update all URLs from vercel.app to next-ai-drawio.jiang.jp
- Remove fake aggregateRating from JSON-LD (better SEO credibility)
- Move JSON-LD script to <head> for better SEO crawler detection
- Enhance JSON-LD description with specific features
- Add 'free diagram generator' and 'online diagram maker' keywords
- Improve Twitter Card description with call-to-action
Addresses critical issues identified in SEO review:
- Artificial rating data removed
- JSON-LD placement improved
- Domain consistency across all files
- Enhanced metadata with Open Graph and Twitter Card tags
- Added robots.ts for search engine crawling directives
- Added sitemap.ts for site structure
- Improved image alt text for better accessibility and SEO
- Added JSON-LD structured data (Schema.org SoftwareApplication)
- Configured proper meta tags including keywords, authors, and robots settings
- Set metadataBase for proper URL resolution
- Add OpenRouter provider support with @openrouter/ai-sdk-provider
- Fix input not disabling during 'submitted' state for fast providers
- Apply disable logic to all interactive elements (textarea, buttons, handlers)
- Clean up env.example by removing model examples and separator blocks
- Upgrade zod to v4.1.12 for compatibility with ollama-ai-provider-v2
- Add debug logging for status changes in chat components
- Add toggle button in chat panel header
- Implement collapsed state with thin vertical strip
- Add Ctrl+B keyboard shortcut to toggle visibility
- Canvas expands to full width when chat is hidden
- Smooth 300ms transition animation
- Updated system prompt to allow up to 3 retry attempts with adjusted search patterns
- Simplified error response to provide current diagram XML and reference retry policy
- AI model self-manages retries based on system instructions
Add responsive detection to show a message prompting users to access the application from desktop or laptop when viewing on mobile devices (screen width < 768px)
dayuan.jiang