Changelog#

`3.0.0` (unreleased)#

Breaking changes#

Tools no longer replace the pipe’s pending flow when they run. Tool merges its reports into tee.inlet.flow (order preserved, duplicates dropped) through the new Tool.record_reports(), so two tools triggered in the same turn – or a tool triggered while the initial ("new",) flow is still unconsumed – keep every writer’s pending reports. ToggleCollapsedTool.run no longer assigns the flow itself (#156).
Viewer.fit() / Viewer.center() now carry the full SprottyViewer signatures (model_ids: str | Sequence[str] | None, plus the animate/zoom/padding keywords) and are documented no-ops on the generic viewer. model_ids accepts one id or any sequence of ids (a str is one id, never its characters); list[str] still works. The base Viewer’s Fit/Center buttons no longer raise TypeError when clicked (#156).
Tool invocation and lifecycle names each have one role (#156):
- Tool.trigger() requests execution (cancels a running request, schedules run(), returns the asyncio.Task); Tool.handler is removed.
- ToolButton.on_click replaces the handler trait; it is called as on_click(tool).
- SetTool’s active observer is the private _update_active; it is not an execution entry point.
- Tool.on_start(callback) replaces on_run. It fires when the work actually starts (2.x on_run fired after a successful run, despite its name).
- Tool.on_done(callback) is a registration method (several listeners; the owning Diagram registers its refresh there), no longer an assignable single-slot trait. Callbacks receive the tool and fire only after run() completes successfully and its change reports are recorded – never on failure, cancellation, or a request rejected by trigger().
- Reports are recorded when the work finishes (successfully or not), so partial changes from failed or cancelled work still reach the next refresh; a request superseded by a newer trigger() never overwrites the newer request’s state.
- trigger() on a tool without a run() (state-only tools such as Selection, callback tools such as ToolButton) raises NotImplementedError before anything is scheduled; nothing is faked.
- Every callback receives the emitting object: Toolbar.on_close is typed Callable and called as on_close(toolbar).
- Diagram.refresh(sender=None) ignores its argument (it was change).
Tool.disabled replaces the inert, synced disable trait (nothing ever read it). It is an execution guard: trigger() and ToolButton clicks are rejected with RuntimeError while the tool is disabled (nothing is scheduled, no callback fires), and every control under tool.ui that has a disabled trait follows it – buttons and the buttons inside SetTool’s box, but not boxes or progress bars, which have no such trait (#156).
Unbound tools are a supported state: PipelineProgressBar.pipe, ToggleCollapsedTool.selection and SetTool.selection are None until bound (they used to raise TraitError on read), so a tool can be constructed first and bound later (Diagram.register_tool, assignment, or the progress bar’s first update). While a dependency listed in the tool’s _dependencies is None, Tool.missing_dependencies() names it, the tool’s controls are disabled, and trigger() raises RuntimeError naming the missing dependency (#156).
Selection.elements() stays strict by default (an unknown id raises NotFoundError naming it) and gains explicit tolerant resolution: elements(strict=False) skips unknown ids and missing_ids() reports them, both in ids order. Requested ids are never filtered on assignment, so selecting before the first layout keeps working. get_index() now builds the index when the pipe’s MarkIndex has no elements yet (it only checked for a missing MarkIndex, which always exists) (#156).
Toolbar.order() and Toolbar.tool_order() are annotated as returning tool UIs (dict[int, list[DOMWidget]] / list[DOMWidget]), which is what they always returned, and document that tools without a ui are omitted. Selection.ids is documented as the tuple it is; the 01_Linking example compares it to a tuple (its list comparison was always true), and the 12/13 examples drop commented-out references to the long-gone toolbar.commands (#156).
Hover.ids is replaced by Hover.hovered_id: str | None (default None). No alias exists: code that reads or observes ids on the hover tool must switch to hovered_id. Despite its name, ids only ever held one id (a string, never a tuple), so the value shape is unchanged; only the trait name and the None state are new. Selection.ids is unchanged and still a tuple (#155).
Leaving an element now clears the hover: hovered_id becomes None when the pointer leaves the diagram’s elements (it previously kept the last hovered id forever). Observers that only expect strings must handle None (#155).
Viewer.control_overlay is opt-in and defaults to None; assign a ControlOverlay (e.g. view.control_overlay = ControlOverlay()) before setting its children. The browser renders no overlay container for None or for an overlay without children (previously every selection rendered an empty <id>_entropy node), and re-renders when the overlay’s children change (#156).
ipyelk.tools.contol_overlay (sic) is renamed to ipyelk.tools.control_overlay without an alias (#156).
Viewer.viewport (an ipyelk.tools.Viewport) replaces the never-written Zoom and Pan tools and the Viewer.viewed trait with real viewport synchronization (#156). The browser reports the camera of the most recently reporting view as one snapshot – view_id, origin, zoom, canvas_size, viewed_ids – so an observer of any field sees all five updated together. It is not a global viewport: several views of one diagram stay independent and each report names its view. The traits are read-only in the kernel (the browser is the only writer) and None until the first report. viewed_ids lists the model elements whose bounds touch the visible rectangle, in model order; edges, renderer artifacts and the slack ports/edges that stand in for hidden elements are never listed.
SprottyViewer.set_viewport(*, origin=None, zoom=None, animate=True, view_id=None) moves the camera (a command, like fit/center, never a trait): None keeps the view’s current origin/zoom, view_id=None moves every connected view (#156).
Viewer.painter (an ipyelk.tools.Painter) is live: temporary, view-only styling (#156). painter.paint(ids, *css_classes) adds classes to the rendered elements in every connected view without touching properties.cssClasses – a re-layout re-applies them, the ELK/model JSON never changes, and exported SVG shows them. unpaint(ids, *css_classes) removes classes (with none given, drops the ids), clear() removes everything the painter applied and never a class the model set, painted_ids, elements(strict=True) and missing_ids() mirror Selection. Targets are ids (a str is one id) or elements (stored as get_id()); the synced styles trait is dict[str, tuple[str, ...]]. The unfinished 2.x placeholders cssClasses, marks and name are removed.
Removed names are hard errors for the whole 3.x line, not aliases or warnings: Tool.handler, ToolButton.handler, Tool.on_run, Tool.disable, Hover.ids, Viewer.zoom, Viewer.pan, Viewer.viewed, ipyelk.tools.Zoom, ipyelk.tools.Pan, Painter.cssClasses, Painter.marks, Painter.name, and assigning Tool.on_start or Tool.on_done raise ipyelk.exceptions.DeprecatedAPIError (an AttributeError, so hasattr, getattr(..., default) and introspection keep working) on read, assignment, and as constructor keywords, with the reason and the replacement in the message (mechanism: ipyelk.exceptions.RemovedAPI / RegistrationMethod / check_removed). The removed module-level names raise ipyelk.exceptions.DeprecatedImportError instead, which is an ImportError and not an AttributeError, because from ipyelk.tools import Zoom would otherwise lose the message to a bare “cannot import name”. Importing the misspelled ipyelk.tools.contol_overlay module raises it too, naming the new path; it re-exports nothing. Both flavors share the ipyelk.exceptions.DeprecatedAPI base, so one except catches either.

Migration#

# 2.x: a Tool subclass that hand-rolled the merge to avoid clobbering the flow
inlet.flow = tuple(dict.fromkeys((*inlet.flow, *self.reports)))
# 3.0: the base class does this; drop the hand-rolled merge or call
self.record_reports()

# 2.x                                   # 3.0
button.on_click(tool.handler)           button.on_click(tool.trigger)
await tool.handler()                    await tool.trigger()
FitTool(handler=lambda: view.fit())     FitTool(on_click=lambda tool: view.fit())
tool.on_run(callback)                   tool.on_start(callback)   # when work starts
                                        tool.on_done(callback)    # success only
tool.on_done = callback                 tool.on_done(callback)    # + remove=True
toolbar.on_close = lambda: ...          toolbar.on_close = lambda toolbar: ...
tool.disable = True                     tool.disabled = True   # now actually disables
# a Tool subclass that did its work synchronously and refreshed by hand:
inlet.flow = (...); self.on_done()      self.trigger()  # run() may be a no-op

# 2.x
hover.ids  # last hovered id, never reset
# 3.0
hover.hovered_id  # id under the pointer, or None
hover.hovered_id = "n1"  # highlight from the kernel
hover.hovered_id = None  # clear the highlight

# 2.x: every viewer allocated an overlay, rendered even when empty
view.control_overlay.children = [button]
# 3.0: opt in first
from ipyelk.tools import ControlOverlay  # was ipyelk.tools.contol_overlay

view.control_overlay = ControlOverlay()
view.control_overlay.children = [button]
view.control_overlay.children = []  # renders nothing

# 2.x: placeholders nothing ever wrote
view.zoom.zoom, view.pan.origin, view.viewed
# 3.0: the browser's latest report (None until a view reports), one snapshot
view.viewport.zoom, view.viewport.origin, view.viewport.viewed_ids
view.viewport.view_id  # which view reported
view.viewport.observe(callback, "origin")  # every field is already updated
# move a view (a command, not state)
view.set_viewport(origin=(0, 0), zoom=1.5)  # every view, animated
view.set_viewport(zoom=2, animate=False, view_id=view.viewport.view_id)

# 2.x: Painter(cssClasses=..., marks=..., name=...) painted nothing
# 3.0: view-only classes per id, applied in every view, never in the model
view.painter.paint(["n1", "n2"], "highlight")  # or elements
view.painter.unpaint("n1", "highlight")  # drop a class; unpaint("n1") drops the id
view.painter.clear()
view.painter.styles  # {"n2": ("highlight",)}

The removed names (handler, on_run, disable, Hover.ids, assigning on_start or on_done, the contol_overlay module, Viewer.zoom/pan/viewed, ipyelk.tools.Zoom/Pan, Painter.cssClasses/marks/name) raise DeprecatedAPIError – or DeprecatedImportError for the module-level names – throughout 3.x with the replacement in the message.

`@jupyrdf/jupyter-elk 3.0.0`#

Write hovered_id back on pointer leave, but only when the departed element is still the hovered one, so a stale leave cannot erase a newer enter; never dispatch null to sprotty as an element id; re-wiring the hover tool releases the previous tool’s listener (#155).
Skip the control overlay when the viewer’s control_overlay is null or has no children, and re-render when its children change (#156).
Report the viewport: every ELKViewerView owns a stable view_id and, 100 ms after the last SetViewport/Center/FitToScreen/InitializeCanvasBounds action (plus sprotty’s animation time for an animated move) or re-layout, writes view_id, origin, zoom, canvas_size and viewed_ids to the kernel’s viewport tool in one state update. A gather superseded by a newer one is dropped, a snapshot the kernel could not tell from the last write is skipped, and a detached view never reports. The viewport custom message applies SprottyViewer.set_viewport to the addressed view (or every view) (#156).
Apply the kernel painter.styles at transform time: the ELK -> sprotty transform merges the painted classes after each element’s model classes (no duplicates, symbols excluded), so every render – including re-layouts – carries them, and a change:styles re-renders the same layout through sprotty’s model update, which keeps the selection and the camera. UpdateModelCommand2 now also carries hoverFeedback over to the updated element, so a re-render under a resting pointer no longer drops the mouseover (#156).

`2.1.3` (unreleased)#

Fixed#

Hiding a root-level node with visible edges no longer crashes elkjs: iter_visible reported every sibling after a hidden one as hidden and projected a hidden node’s slack port onto its preceding sibling (for a hidden first child of the root, onto the root itself, which elkjs rejects). Hidden compound children now project onto the compound, and an edge whose projection would touch the root is dropped from the wire while its Edge stays in the index (#161, #169)
Never serialize an id as null: an id-less element mints a wire id once and keeps it, the index adopts it, and an id the Registry mints first becomes the wire id, so edges no longer go on the wire as sources: [null] and the id first seen is the id for the object’s life. Port ids compose from the parent’s id in and out of a Registry (was <uuid>.<uuid> under Node(id="N")), a copied element mints its own wire id instead of sharing the original’s, and Label.wrap no longer copies a generated id onto every line (#160, #169)
Add nested label spacing regardless of the sublabel width: size_nested_label parsed ls.width or 0 + spacing as ls.width or (0 + spacing) (#161, #169)
Fix the first-frame culling bug while culling stays off: validCanvasBounds was the inverse of its comment, so nothing was culled on a sized canvas and almost everything on the 0x0 canvas before the first InitializeCanvasBoundsAction; the guard now skips culling while the canvas size is unknown, behind an explicit CULLING_ENABLED constant (turning it on is #170). The SVG exporter sizes its viewBox from the model’s content extent instead of the live group’s bounding rect, which under-sized by the layout’s left margin and made the export viewport-dependent (#161, #169)
Stop echoing browser-written state back to the frontends: the selection, the hover and the laid-out graph were re-sent to every attached frontend as soon as the browser wrote them, and the graph was then sent a second time as a re-serialized copy, so every layout arrived back twice and the diagram re-rendered on each. Known limitation, documented on the traits: a second frontend attached to the same kernel no longer learns browser-made selections from that echo (#164, #161)
Keep a refresh that was requested while a run was already in flight: a style change, or a tool that failed after changing the model, recorded itself by overwriting the pipeline’s pending flow, and a completing run cleared that flow by overwriting it in turn — so the request was lost and the change never redrawn. Runs now take the pending flow when they start and put it back if they fail or are cancelled (#164, #160)
Stop throwing away browser work a refresh has already started: a burst of refresh() calls cancelled the in-flight run and started another, so the browser went on laying out graphs whose results were discarded, and a cancelled run’s answer could be consumed by its successor. Requests are now counted rather than cancelled and one runner serves them on the trailing edge (ten calls in one tick cost one layout), and each roundtrip carries a generation counter, so a re-sent request is recognized as the same work instead of queueing a duplicate layout in the browser and an answer for an abandoned generation is ignored. Pipe.cancel() is now the only thing that stops a run, and pipe._task is the pipe’s current runner rather than “the run this call just scheduled” (#164, #161)
Stop swapping the browser’s laid-out copy into the pipeline’s input: it carried nothing that persisting the result does not already merge into the user’s own elements, but the next run rebuilt the element index from that copy, so a selection resolved to a copy and a hidden node resolved to the slack port sharing its id (#164, #161)
Make edge-owner validation linear in the hierarchy instead of quadratic in it: validating 1000 nodes with 3000 root-owned edges fell from 14.4 s to 0.010 s (500 nodes / 1500 edges, 3.5 s to 0.005 s; 100 / 300, 0.146 s to 0.001 s), and validation no longer re-reports on its outlet when it changed nothing (#164, #161)
Settle a browser roundtrip on the event loop that awaits it: with ipykernel 7 and a frontend that uses kernel subshells (JupyterLab 4.5+, ipywidgets 8.1.8+), widget messages are handled on a subshell thread with its own loop, so an answer written from there did not wake a run started from a cell until the resend timer fired, and a refresh() from a widget callback reused a runner from the other loop and failed silently. Answers are now delivered through the future’s own loop, and a refresh from another loop hands the runner over to the caller’s loop (#164)
Stop leaking widgets on every refresh. Each pipe status change created a new PipeStatus widget, and a displayed pipeline status view rebuilt its rows. Neither closed what it replaced: 9 widgets leaked per refresh, 145 with the view shown. The status view now rebuilds its rows only when the sub-pipes or their views change, and closes the rows it replaces (#176)
- PipeStatus is no longer an ipywidgets.Widget: it is an immutable value with the same attributes and constructors, compared by identity, so every new status still notifies status observers
- PipelineStatusView.statuses is removed. It held the view’s row widgets, was never synced, and is now a private list; the rows still reach the frontend as children
Stop a diagram from freezing for good on an update that removes elements. Sprotty’s fade-out removed the faded elements when the eased animation time reached 1, which easeInOut rounds to within 1e-8 of the end, while the animation only stops at the raw time 1. A frame landing a hair before the end (evenly spaced frame timestamps, 15 frames of 1000/60 ms for the 250 ms animation) removed them early, and the real last frame removed them again and threw. The throw happened inside an animation-frame callback, so the animation never finished and sprotty’s command stack waited on it forever: every later layout reached the browser and was never drawn (#95). Reported upstream as eclipse-sprotty/sprotty#573; the workaround can go once that is fixed
Close the old pipeline when diagram.pipe is replaced. Only its run was cancelled: its sub-pipes, intermediate marks and status views stayed open, and the diagram’s style link kept it alive, so every replacement left 37 widgets behind, 99 with the status view shown. A diagram now owns its pipe: replacing it closes the old pipe, its sub-pipes and the widgets they created (including a status_widget displayed elsewhere), never diagram.source or an inlet/outlet passed in. Pipe.close() closes what the pipe created (a Pipeline also closes its sub-pipes). style (also for a text sizer in a nested pipeline), the progress bar and the tools now follow the new pipe (#180)
- A pipe belongs to one diagram, with every pipe nested in it. Assigning a pipe that nests a pipe another open diagram owns raises TraitError: both diagrams used to share its outlet, so one silently showed the other’s graph. Two diagrams sharing a source is a separate question (#178)
- A replaced pipe is closed, so assigning a pipe that nests a closed pipe, or any part of the current pipe (wrapping it, or reusing its stages), raises TraitError too: build a new pipe instead
- The rule also holds when a pipeline’s pipes changes in place, and before a Pipeline is wired. Adding another open diagram’s stage to an assigned pipeline was accepted, so the stage ran in both, and Pipeline(pipes=[...]) rewired a live diagram’s stage before anything could refuse it. Both now raise TraitError and change nothing, as does a pipes list with a closed stage or a stage listed twice (it was wired into a cycle). A stage added in place belongs to the diagram (and gets its style); a stage removed in place is released, not closed (#191)
Close what a diagram owns when the diagram is closed. Diagram.close() only released its pipe, so the pipe tree, the view, the tools and the toolbar stayed open: every closed diagram left 68 widgets behind, 130 with the status view shown. It now closes them with their layout and style widgets and drops its links to them, and leaves open diagram.source, an inlet or outlet passed to a pipe, the viewer’s control_overlay, and a layout or tool ui passed in. A closed diagram’s pipe is closed, so another diagram can no longer take it. Tool.close() also closes the ui the tool built, and Viewer.close() its tools (#191)
Log the routine “fixing N ids” message from ValidationPipe at debug level instead of warning: assigning ids to id-less elements is the normal path (fix_null_id=True), so every diagram built from id-less elements logged it
Keep the elements behind a collapsed view intact across layouts (#171, #187):
- A hidden port, or a port on a hidden node, keeps its size, position, layout options and properties. The slack port that stands in for it in the layout used to overwrite them with its own 5×5 geometry, permanently. Any other port, including one that moved to another node, still takes its layout from the browser.
- Each line of a wrapped label gets its own ids, and so do its sub-labels. The lines of a label with id L get the ids L#0, L#1, … (sub-label sub gets sub#0, …). If another element already uses one of these derived ids, indexing gives that line a fresh id; ids you set yourself never change. Before, every line had the same id (L, or one generated id) and the same sub-label ids. A wrapped line shares its label’s metadata.
- The ELK JSON schema accepts the slack-port key and hidden properties, so a projection validates after a collapse and an expand.
Center a Circle in its node: Circle no longer serializes x and y as its radius. The node:round renderer takes its radii from the laid-out node, so a node larger than the circle drew the ellipse off center. x and y that you set are still serialized (#200)

Development#

Add scripts/bench_pipeline.py, a headless benchmark of Diagram.refresh(): it runs the real kernel pipeline and real elkjs in node with only the two browser stages stubbed, over deterministic graphs from scripts/bench_graphs.py, and reports wall and per-stage time, comm messages and bytes in both directions, layout runs, a collapse refresh and a ten-refresh burst (#164, #167), and the comms opened per refresh (#176)
Gate CI on the benchmark’s deterministic counts: pixi run bench-check compares them to scripts/bench_baseline.json and fails on any change until the baseline is updated with pixi run bench-update (#176), including the live widgets left by replacing diagram.pipe (#180) and by closing a diagram (#191)
Run the benchmark on the oldest supported ipywidgets too: on 8.0.1 it crashed in its first refresh, because ipywidgets 8.0 opens an ipykernel comm directly and the harness never saw a message. pixi run bench-check-oldest gates it against the same baseline in the test-oldest (ubuntu-latest) job (#191)
Add ast-grep structural lint rules for Python and TypeScript (sgconfig.yml, scripts/ast-grep/), each with test cases, run by pixi run lint-ast-grep as part of pixi run lint; a suppression must name its rule and still match something (#147)
Lint prose with Vale in the lint job: pixi run lint-vale checks US English spelling and the proselint and write-good styles in the root *.md files, docs/, docstrings and comments in src/, tests/ and scripts/, and notebook markdown, and fails on any warning or error. It replaces check-spelling, which linted the built HTML and never failed (#181)
The nbconvert kernel watchdog writes its stack dump to the kernel’s real stderr, not to the descriptor ipykernel captures, where the dump woke the capture thread and made it look busy. A second dump lists the pending asyncio tasks of the kernel’s event loops, with their await chains, and whether the shell locks are held (#177)
Fix the itest-nbconvert hang in 03_App.ipynb (#177). ipykernel 7.0 to 7.3 send shell replies raw on the socket their shell stream reads from, which can swallow the wake-up for a request that arrives during the send. The request then waits unread while the kernel idles. In nbconvert that happens when a cell prints into an ipywidgets.Output: nbclient answers with a comm_msg just as the cell’s reply goes out, and the next cell never runs.
- The itest and atest environments now lock ipykernel 7.4, which fixed it (ipykernel#1529).
- ipykernel 7.4 needs Python 3.11 or newer. For the Python 3.10 environments, the kernels of the nbconvert-- task and of the robot tests load scripts/kernel_shell_reply_fix.py, a startup file that backports the fix and does nothing on ipykernel 7.4 or newer. utest-oldest tests it
Lower the nbconvert-- cell timeout from 1200 s to 40 s. The slowest healthy notebook in 4061 notebook runs from the last 45 CI runs took 8.2 s. The kernel watchdog now restarts with every cell and fires after 30 s without a new cell (#177)
Fail CI when the committed elkschema.json differs from what jlpm schema generates, and rebuild it when elkgraph-json.ts changes (#187)

Documentation#

Document the shape API: a docstring for each class in ipyelk.elements.shapes, and a Shapes page in the API reference with one table row for each shape (#66)

`2.1.2`#

Development#

Migrate to native Pydantic 2 validation and serialization; require pydantic >=2.12,<3. Use model_dump() / model_dump_json() instead of dict() / json(). Custom models now use model_config, Field(exclude=True), and native serializers instead of Config, merge_excluded, and dict() overrides. Nested subclass fields and graph references are preserved; shape serialization no longer mutates dimensions. Explicit serialization field selections are now respected.
Fix Python typing throughout the package and run mypy as part of pixi run lint.
Minimum supported Python is now 3.10
Upgrade the pinned pixi from 0.34.0 to 0.67.0 (and setup-pixi to v0.10.0); relocking updates libgfortran5 13.2.0→14.2.0, fixing a macOS arm64 dyld failure that broke numpy/bqplot in the example notebooks
Only reinstall requirements in the 07_Simulation example when ipyelk is missing, and add tooltips to its control widgets
Preserve the normal labextension in src/_d during the coverage build (build-js-ext-cov), which emptied it and left pixi run build with an unloadable extension (_build.load: "static")

`@jupyrdf/jupyter-elk 2.1.2`#

Add optional label tooltips and full-width separators above labels.
Observe source rewiring and apply initial kernel selections after model submission completes.
Report missing browser state for re-sync and surface text-measurement errors through the kernel error channel.
Discard widget views whose attachment host disappears while they load.
Attribute label hover feedback to the nearest hoverable element.
Fix the SVG exporter enabled flag, which was always true (F5)
Report browser-side layout failures to the kernel instead of silently emitting an empty layout (F6)
Make ELKLayoutModel.layout() re-entrant: it stripped element properties (incl. cssClasses) off the shared inlet value in place, so a duplicate run message or an overlapping refresh re-laid-out the stripped graph and pushed a style-less (black-and-white) diagram (F7)
Render edge labels where ELK placed them: ElkLabel carried sprotty’s edgeLayoutFeature, whose EdgeLayoutPostprocessor re-anchors edge labels along the route and treats ELK’s absolute coordinates as a relative offset, shifting every edge label by roughly its edge’s origin (F8)
Orient edge end symbols along the visible route: the adjacent-segment tangent collapsed to atan2(0, 0) on the duplicated control points of elkjs SPLINES sections (arrowheads drawn 180° wrong, inside the target node) and followed the short exit stub of POLYLINE routes instead of the visible diagonal; interior bends under a symbol’s footprint no longer make the trimmed shaft double back (F9)
Drop the selection write-back race: getSelection() resolves one action-queue slot later, so two SelectActions dispatched close together each read the other action’s resulting state and wrote it back, flipping the selection tool’s ids forever — a self-sustaining oscillation that pegged the renderer’s main thread (F11)
Render nested JupyterLab widgets in the pass that reveals them: the overlay was built from a registry populated by the previous render’s snabbdom hooks, so a widget node that had just become visible produced no container until some later, unrelated re-render happened to run (F14)
Make the selection write-back set-based so linked views cannot bounce: F11’s generation stamp is per view, but change:ids is observed by every view of a shared model, and each view gathers the same selection in a different order, so a reordering kept dispatching SelectActions between two views until the browser ran out of memory (F16)
Add a vitest unit-test harness (F5)
Upgrade ELK.js from 0.9.3 to 0.12.0 (#140). ElkEdge gains an optional container field in the generated schema, and a vitest lays out a graph with the real bundle to pin the port, edge-section and properties round-trip contract.
Upgrade sprotty and sprotty-protocol from 1.3.0 to 1.4.0, which requires inversify ^6.1.3 (pinned to 6.2.2) and reflect-metadata ^0.2.2. The inversify shared module declares its version statically: its lib/esm entry ships a version-less package.json, so webpack module federation would otherwise register it as 0 and warn on every page load; a unit test keeps the static version exact, equal to the dependency pin and inside the declared requiredVersion.
Drop the unused sprotty-elk dependency: nothing in js/ ever imported it, sprotty and sprotty-protocol do not depend on it, and the built extension has no reference to it. Its elkjs ^0.8.2 range was the only thing that conflicted with ELK.js 0.12.0, so the resolutions.elkjs override goes with it (#140).

`ipyelk 2.1.2`#

Allow diagram construction without a running event loop. Pipe.schedule_run() and Diagram.refresh() return None when they cannot schedule a task.
Add optional LabelProperties.tooltip and LabelProperties.separator fields.
Re-send widget state after a browser reports missing state, with throttling.
Echo terminal progress state twice, cancelling pending echoes on updates and close.
Fix IDReport.message() printing literal {eid}/{el} placeholders (F1)
Fix Pipeline.check() / get_progress_value() crashing on an empty pipeline (F2)
Give each Tool its own on_run callback dispatcher (was shared across all tools) (F3)
Surface pipe/diagram exceptions via an on_error callback instead of silently dropping them in the asyncio done-callback (F4)
Add a configurable timeout and a browser→kernel error channel to ElkJS/BrowserTextSizer so a failed or silent browser layout no longer hangs the diagram (F6)
Re-send the browser run request with backoff until a frontend answers, so a pipe run before its diagram is displayed no longer waits on a message nobody received; a browser-reported layout error stops the retries immediately (F10)
Treat an errored run as terminal for progress reporting: PipeStatus.step() returned None for errored pipes, so Pipeline.get_progress_value() raised TypeError inside the error path — on_error saw the TypeError instead of the layout error, and the PipelineProgressBar sat “in progress” forever; the bar now fills as a visible warning (F10)
Keep hidden elements in the shared MarkIndex across browser round trips: Node.dict() always drops hidden children, so rebuilding the index from a value that has been through the browser erased them and ToggleCollapsedTool could never reveal them again — nested widgets (such as the 15_Nesting_Plots figures) never appeared. ElementIndex.update() now merges instead of requiring every id to be known: existing ids update in place, unknown ids are added, which also fixes a NotFoundError when slack ports introduced by VisibilityPipe come back from a layout (F12)
Log a failed layout in Diagram.refresh() instead of returning silently: a silently failed pipeline is indistinguishable from a hung diagram (F12)
Fix the local TextSizer fallback, which read self.source.value, assigned self.outlet.changes and returned self.value — none of which exist on Pipe — and therefore raised on its first statement (F13)
Do not wait on a browser that cannot answer: with IPYELK_NO_BROWSER set (as nbconvert --execute does) a pipe gives up its roundtrip immediately instead of keeping a task alive across cell boundaries re-sending run requests for 30 s, which wedged kernels on slower CI runners (F15)

`2.1.1`#

`@jupyrdf/jupyter-elk 2.1.1`#

restore license files

`ipyelk 2.1.1`#

restore license files

`2.1.0` (broken)#

`@jupyrdf/jupyter-elk 2.1.0`#

support JupyterLab 4.1-4.3

`ipyelk 2.1.0`#

improve type hints

`2.1.0a0`#

`@jupyrdf/jupyter-elk 2.1.0-alpha0`#

Update dependencies elkjs 0.9.3, sprotty 1.3, jupyterlab 4.2
Add shim for reflect-metadata vs fast-foundation

`ipyelk 2.1.0a0`#

Support pydantic >=1,<3

`2.0.0`#

`@jupyrdf/jupyter-elk 2.0.0`#

Added control layer to allow jupyterlab widgets to exist on top of the diagram based on current node selection.
Adding controllable render delay for jupyterlab widgets used in diagram nodes.
Updated dependencies to elkjs 0.8.2.
Fixed diagram bounding box issue affecting node visibility (#94).
Improved test sizing that takes into account css properties (#97)

`ipyelk 2.0.0`#

Migrated to ipywidgets >=8.0.1,<9
Added simple visualizer widget for the diagram pipe status.
Fixed edge parent ownership affecting self edges (#101)

`2.0.0a0`#

`@jupyrdf/jupyter-elk 2.0.0-alpha0`#

Label Schema fix ([#73])
Element API overhaul (#88)
- Add mypy for type checking
- Use pydantic for Element base models
Overhaul top level interface (#89)
- Backporting Sprotty Duplicate ID (#17)
- Generalize the processing stages to use a common interface of Marks and simplify processing to composable pipes
- Refactoring top level APIs and attempt and more streamlined Diagram creation

`ipyelk 2.0.0a0`#

`1.0.1`#

`@jupyrdf/jupyter-elk 1.0.1`#

hides some browser console messages

`ipyelk 1.0.1`#

`1.0.0`#

`@jupyrdf/jupyter-elk 1.0.0`#

updates for JupyterLab 3 (#6)
- uses @lumino components

`ipyelk 1.0.0`#

supports (and depends on) JupyterLab 3 (#6)
- labextension is delivered as part of the ipyelk python package, no more lab build
- npm tarballs will still be uploaded

`0.3.0`#

`@jupyrdf/jupyter-elk 0.3.0`#

`ipyelk 0.3.0`#

Custom shapes (#60)
- Ability to add custom SVG symbols and use as a reference for other elements
- Custom node shapes
- Custom connector end shapes for edges
- Custom shapes for ports
- Custom node label shapes
- JupyterLab widgets rendering inside Node
- Node compartments
- Initial level of detail checks for labels
- Rendering checks for nodes outside of view bounding box
Initial documentation (#64)

`0.2.1`#

`@jupyrdf/jupyter-elk 0.2.1`#

fix ElkTransformer handling of custom properties (#46)
add ElkTextSizer passing through of custom CSS style when sizing labels (#48)

`ipyelk 0.2.1`#

update Elk schema to allow for properties (and c) on edge labels and port labels (#48)
Merge layout options if specified in a given node’s data with default layout options (#48)

`0.2.0`#

`@jupyrdf/jupyter-elk 0.2.0`#

provides in-browser text measurement against ground-truth CSS (#15)
upgrades to sprotty-elk 0.9.0 (#15)
adds basic browser testing with Robot Framework (#21)
adds SVG export with ElkExporter (#27)
handles multiple views of the same ELK model more robustly (#36)

`ipyelk 0.2.0`#

adds optional node label positioning with NodeLabelPlacement (#15)
- vertical/horizontal alignment
- inside/outside the node
improves evented updates of networkx to diagram with ElkDiagram.connect(XElk) (#15)
adds optional ElkTextSizer for interacting with browser text sizing (#15)
add layout options widgets to control various layout parameters (#24)
add support for multiline node labels, port labels, and edge labels ([#35])
- adds possibility of passing CSS classes through to the final DOM elements

`0.1.3`#

`@jupyrdf/jupyter-elk 0.1.3`#

includes all files using npm publish

`ipyelk 0.1.3`#

updates some metadata for pypi

`0.1.2`#

`ipyelk 0.1.2`#

`@jupyrdf/jupyter-elk 0.1.2` (broken)#

(failed) fix npm release process

`0.1.1`#

`ipyelk 0.1.1`#

initial release

`@jupyrdf/jupyter-elk 0.1.1` (broken)#

initial release

Changelog#

3.0.0 (unreleased)#

Breaking changes#

Migration#

@jupyrdf/jupyter-elk 3.0.0#

2.1.3 (unreleased)#

Fixed#

Development#

Documentation#

2.1.2#

Development#

@jupyrdf/jupyter-elk 2.1.2#

ipyelk 2.1.2#

2.1.1#

@jupyrdf/jupyter-elk 2.1.1#

ipyelk 2.1.1#

2.1.0 (broken)#

@jupyrdf/jupyter-elk 2.1.0#

ipyelk 2.1.0#

2.1.0a0#

@jupyrdf/jupyter-elk 2.1.0-alpha0#

ipyelk 2.1.0a0#

2.0.0#

@jupyrdf/jupyter-elk 2.0.0#

ipyelk 2.0.0#

2.0.0a0#

@jupyrdf/jupyter-elk 2.0.0-alpha0#

ipyelk 2.0.0a0#

1.0.1#

@jupyrdf/jupyter-elk 1.0.1#

ipyelk 1.0.1#

1.0.0#

@jupyrdf/jupyter-elk 1.0.0#

ipyelk 1.0.0#

0.3.0#

@jupyrdf/jupyter-elk 0.3.0#

ipyelk 0.3.0#

0.2.1#

@jupyrdf/jupyter-elk 0.2.1#

ipyelk 0.2.1#

0.2.0#

@jupyrdf/jupyter-elk 0.2.0#

ipyelk 0.2.0#

0.1.3#

@jupyrdf/jupyter-elk 0.1.3#

ipyelk 0.1.3#

0.1.2#

ipyelk 0.1.2#

@jupyrdf/jupyter-elk 0.1.2 (broken)#

0.1.1#

ipyelk 0.1.1#

@jupyrdf/jupyter-elk 0.1.1 (broken)#

`3.0.0` (unreleased)#

`@jupyrdf/jupyter-elk 3.0.0`#

`2.1.3` (unreleased)#

`2.1.2`#

`@jupyrdf/jupyter-elk 2.1.2`#

`ipyelk 2.1.2`#

`2.1.1`#

`@jupyrdf/jupyter-elk 2.1.1`#

`ipyelk 2.1.1`#

`2.1.0` (broken)#

`@jupyrdf/jupyter-elk 2.1.0`#

`ipyelk 2.1.0`#

`2.1.0a0`#

`@jupyrdf/jupyter-elk 2.1.0-alpha0`#

`ipyelk 2.1.0a0`#

`2.0.0`#

`@jupyrdf/jupyter-elk 2.0.0`#

`ipyelk 2.0.0`#

`2.0.0a0`#

`@jupyrdf/jupyter-elk 2.0.0-alpha0`#

`ipyelk 2.0.0a0`#

`1.0.1`#

`@jupyrdf/jupyter-elk 1.0.1`#

`ipyelk 1.0.1`#

`1.0.0`#

`@jupyrdf/jupyter-elk 1.0.0`#

`ipyelk 1.0.0`#

`0.3.0`#

`@jupyrdf/jupyter-elk 0.3.0`#

`ipyelk 0.3.0`#

`0.2.1`#

`@jupyrdf/jupyter-elk 0.2.1`#

`ipyelk 0.2.1`#

`0.2.0`#

`@jupyrdf/jupyter-elk 0.2.0`#

`ipyelk 0.2.0`#

`0.1.3`#

`@jupyrdf/jupyter-elk 0.1.3`#

`ipyelk 0.1.3`#

`0.1.2`#

`ipyelk 0.1.2`#

`@jupyrdf/jupyter-elk 0.1.2` (broken)#

`0.1.1`#

`ipyelk 0.1.1`#

`@jupyrdf/jupyter-elk 0.1.1` (broken)#