Three defects in the view-preservation path could leave the plot frozen
(no updates, no orbit, stale colours):
1. The merge injected raw _fullLayout objects - including undefined when a
push landed before the previous react resolved - and plotly validates
layouts, so one bad value made every later react fail permanently.
currentView() now deep-copies and validates eye/center/up and each
range, and mergeView() is wrapped so it can never block an update.
2. Updates were not serialized: overlapping Plotly.react calls on one graph
div left it unresponsive. otkoUpdate now queues and coalesces (one react
at a time, latest payload wins), and logs instead of failing silently.
3. A style change re-framed the camera, so re-colouring yanked a view the
user had orbited. Colour/opacity ride the traces and background/outline
ride the layout, so set_style is now a non-framing push.
Verified in the browser: a user orbit survives colour changes, 30 rapid
preview updates land on the final value and stay responsive, and deleting
the live camera no longer wedges the plot.
Plotly.react resets any scene attribute the incoming layout omits, so every
selection change snapped the camera back to the default and re-fit the
ranges - the pinned framing was meaningless. Data-only pushes now pass
preserveView and html.py merges the live camera and axis ranges into the
incoming layout; only an explicit re-frame sends computed framing.
Also fixes the ordering bug where the framing flag was consumed before the
not-ready early return, which dropped the very first framing on the floor
(show_project runs before loadFinished).
Verified in the browser: user orbit + zoom survive selection and working
plane updates, while reset_camera re-frames. Regression-tested headlessly
by asserting the emitted preserveView flag.
Frames the view deterministically: Scene now carries padded per-axis bounds
(opstool's pad_ratio 0.15, with unit slack on degenerate axes) plus
axis_overrides() pinning range + autorange=False, so the model is never
flush against the viewport edge. Adds entity hover templates (Node #n /
Element #n with an empty extra tag), and an optional axis outline (grid +
ticks) - off by default so the SAP2000-like clean viewport is unchanged;
the coloured X/Y/Z axis lines always stay as the orientation cue.
The outline is exposed as a bool in the Plot Properties table, so the
table-driven dialog now builds a checkbox for kind=bool.
Adds a table-driven dialog over RenderStyle.STYLE_FIELDS (colour swatches,
an opacity spin and a label font size), reachable from Options. Edits
preview immediately on every canvas, Cancel restores the style the dialog
opened with, and only OK persists — as JSON under QSettings plot/props,
reloaded on the next launch. Both canvases gain set_style(); the dialog
deliberately exposes plot_style() rather than style() so QWidget.style()
keeps its Qt meaning.
Colours that were hard-coded in the renderers now come from the shared
style: nodes, supports, nodal/element loads, section extrusions (+
opacity) and label font size. The PyVista frame LUT becomes a four-slot
palette [beam, truss, link, selected] with the cell scalar carrying the
family slot, so VTK matches the Plotly backend's per-family colouring
that landed earlier. Adds style helpers shared by both backends
(element_family_index, family_palette) and an immutable with_updates().
The style also gains the editable field table (STYLE_FIELDS) the Plot
Properties dialog is built from.
Ports opstool's per-family element colours and its diverging response
scale; RenderStyle.response_scale_colors now also drives the PyVista
force-diagram colouring instead of a hard-coded "coolwarm".
The frame renderer drops the two-trace normal/selected workaround: my
earlier assumption that plotly cannot colour segments individually was
wrong. Scatter3d.line.color accepts an array mapped through a colorscale,
so one trace now carries per-element colours (family + selection) and is
ready to be coloured by response value later.
Attribution recorded in NOTICE per GPLv3 section 5(a)/(b).
MainWindow persists preferences (window layout, canvas backend) through
QSettings("OTKO", "OTKO"), so tests constructing it were reading the
developer's real settings: after switching to the Plotly backend the
PyVista-specific viewport-axis tests failed with "PlotlyCanvas has no
attribute renderer". A session fixture now redirects QSettings into a
temp dir and is skipped when Qt is not installed, keeping the headless
job Qt-free.
opstool v1.0.26 (GPL-3.0) is cloned read-only under
.slim/clonedeps/repos/ to inspect its PyVista/Plotly visualization
settings. GPLv3 section 13 permits combining it with this AGPL-3.0
project. The clone itself is git-ignored; the manifest and the AGENTS.md
pointer are committed.
Both backends share one SelectionState owned by MainWindow and live
side by side in a QStackedWidget — switching is setCurrentWidget, so
no widget is destroyed mid-session (tearing a VTK window down leaves
dangling make-current callbacks). The choice persists in QSettings.
CanvasCapabilities declares per-backend gaps (force diagrams and
video export stay PyVista-only, both are documented and greyed out)
and the UI gates on capabilities rather than the backend name. The
architecture gate now allows the canvas_plotly package.
PlotlyCanvas hosts plotly.js in a QWebEngineView driven over
QWebChannel: figures update with Plotly.react (the camera survives
unless a view preset asks for it), and clicks round-trip as
node/element picks or grid-snap clicks. Assets are written to a temp
dir and loaded from file:// because the ~5 MB bundle is past
setHtml's data-URL limit. render() forwards QWidget's overload so
grab() and painting keep working.
Project → plotly.js figure dicts with no Qt/pyvista/plotly import, so
it is unit-tested in the headless job. Mirrors the PyVista renderer's
geometry (grid, nodes, frames, supports, loads, extrusions, local
axes, labels, deformation) while respecting plotly.js's medium:
pixel-sized markers, None-separated line segments, selection as a
second trace (per-segment line colours are impossible), cone traces
for arrows. Pickable traces carry meta.kind + customdata.
Supplies the offline plotly.js bundle for the new canvas backend. The
GUI CI job now installs the gui extra so the Qt/WebEngine tests run
instead of being import-skipped.
Auto-fixes (UP037, UP038, F401, I001, SIM105, RUF100) plus hand-fixes:
drop the now-unused `Union` imports, annotate mutable class attributes
with ClassVar, raise ValidationError instead of blind Exception in the
zero-length test, delete two dead locals, and bind the per-iteration
grid-builder closures to their loop variables (B023).
RUF001-003 are ignored repo-wide: × / ≈ / ² are deliberate in the
engineering docs and formulas, which also makes the scattered inline
noqa suppressions redundant. Tests use engineering symbols (A, I, L, E),
`importorskip` before imports and capitalised test names; Qt requires
camelCase signal names and PascalCase event overrides; OpenSees material
fields (cR1, epsU) mirror the solver verbatim. N806 stays exempt only for
the numeric modules that use L / Iz / Iy as symbols.
Reconstructs the workflow stripped during the tree copy to match the
meta-gates: separate type job, headless command over
tests/unit+tests/services, and dedicated GUI / integration jobs.
Tool versions pinned to the pre-commit revisions (ruff 0.4.4,
mypy 1.10.0) so CI and local hooks agree.
The form still passed flat b/h/d/tw/tf/diameter fields, which
ElasticSection (extra='forbid') rejects, so reading any elastic
section crashed. Rebuild the display-only hint as a PipeShape /
AngleShape / RectShape member instead; an incomplete shape degrades
to None rather than raising.
Catch-all for the intermixed residue of the unpushed otko-development
work ported into this tree: combinations/console-dock/quick-guide wiring
across commands, core, services, views and tests; repo-wide ruff-format
normalization; README/CONTRIBUTING updates; and the toolbar default
(both toolbars now open in the top area, quick guide text updated).
Splitting this further would require hunk-level surgery with low
confidence; the preceding commits in this branch isolate the
self-contained features.
Single source of truth for the package + pinned solver versions in
src/otko/_const.py (hatch dynamic version). scipy/pandas were listed in
[gui] but never imported. mypy strict mode rescoped to
core/services/viewmodels; tomli added for tomllib-less Pythons.
