feat(plotly): selected()/brushed() are state accessors, seeded on bridge bind by cpsievert · Pull Request #419 · posit-dev/py-shinywidgets

cpsievert · 2026-11-04T17:54:16Z

Stacked on #412.

Summary

A plotly selection is state on the graph div (trace.selectedpoints, plus the region on gd._fullLayout.selections), not just an event. The event bridge (js/src/plotly.ts) now rebuilds and sends the current selection when it binds plotly_selected — as a synthetic plotly_selected message (a no-data one when empty, the same message plotly.js sends for an empty selection; deliberately not plotly_deselect, which would wrongly clear clicked()).

That lets selected()/brushed() (and so selected_rows()) become kind="state" accessors: they attach at render via #412's mechanism, and a box/lasso selection made before the accessor's first read — for example, before a chat handler or LLM tool first reads it — is reported instead of silently lost. This was the original motivating failure behind #412, which #412 alone could not fix (these accessors were event-kind, and events can't seed).

clicked() stays kind="event"; so do selecting()/brushing() (drag streams) and relayout() (autoranged axes have no explicit range, so the empty state isn't seedable).

Trade-off: state accessors don't reset on re-render, so selected()/brushed() keep their value until the new plot reports its own (initially empty) selection, at which point they become None. Documented in the render_plotly docstring and CHANGELOG.

Evidence

plotly.js retains selectedpoints post-lasso and clears on deselect, verified empirically (Playwright-driven standalone pages) for scatter, scattergl, bar, and histogram; the state-rebuilt payload matches the real plotly_selected payload point-for-point (including customdata and histogram pointNumbers flattening; lasso → parsed lassoPoints; box → ascending range).
Driving the compiled bridge with a lasso before bind produces exactly one correct seed message; the seed is sent once per graph div (re-binds from other accessors don't re-seed).

Verification

New committed Playwright tests (tests/playwright/plotly_selection_seed/): box-select before a button-gated first read → selected()/selected_rows(df)/brushed() reflect it; double-click deselect clears to None. 2 passed × 4 runs, no flakes; existing plotly Playwright suites green (plotly_accessors, plotly_accessors_rerender, plotly_config, plotly_rerender_cleanup).
make check: ruff ✓ pyright ✓ unit tests 842 passed (the one test_geolibre failure is pre-existing on this base and already fixed on main by feat(geolibre): use geolibre's async command methods when available #360).

Note: plotly.js emits plotly_selected a few hundred ms after mouseup (held back for double-click disambiguation), so the fixture waits on a sentinel output for browser→server sync before triggering the gated read.

Related hygiene issues found during this investigation: #417, #418.

… state accessors A plotly selection is state on the graph div (trace selectedpoints, plus the region on the last layout selection), so when the event bridge binds plotly_selected on a graph div it now also reports the current selection as a synthetic plotly_selected message (a no-data one when nothing is selected, the same message plotly.js sends for an empty selection). That lets selected()/brushed() (and so selected_rows()) become state accessors: they attach at render, and a box/lasso selection made before the accessor's first read -- or before the bind request reaches the browser -- is reported instead of silently lost. clicked() stays an event. State accessors don't reset on re-render, so selected()/brushed() now keep their value until the new plot reports its own (initially empty) selection; documented in the render_plotly docstring.

…te accessors Covers the selection-seed scenario end to end: a box selection made before anything reads selected()/brushed() (gated behind an action button), then a gated first read -- which must report the pre-read selection -- and a deselect (double-click on the plot background) after the read, which must clear the accessors to None. plotly.js emits plotly_selected a few hundred ms after mouseup, so a button click right after the drag can overtake the selection message; the fixture follows the sentinel pattern (an ungated text output reading the same accessor state, like ipyleaflet_eager_attach's synced_zoom) so the test waits for the browser-to-server sync before triggering the gated read.

Copilot

🟡 Changes recommended

Selection seeding drops valid empty regions, misparses coordinates with negative exponents, and omits implicit coordinates, while the tests do not verify selection before bridge binding.

4 open issues

1 file not reviewed

shinywidgets/static/output.js: Generated file

What changed in this PR

Makes Plotly selection accessors state-based, building on #412’s eager attachment so selections made before the first accessor read can be recovered.

Changes:

Seeds the current selection and brush region when the browser bridge first binds.
Changes selected() and brushed() to state accessors and documents their re-render behavior.
Adds unit and browser tests for selection retention, gated reads, and deselection.

File	Description
`tests/unit/test_plotly_accessors.py`	Tests eager attachment, early selection messages, and re-render behavior.
`tests/playwright/plotly_selection_seed/test_plotly_selection_seed.py`	Adds gated-read and deselection browser tests.
`tests/playwright/plotly_selection_seed/app.py`	Adds the browser-test fixture and synchronization output.
`shinywidgets/static/output.js`	Updates the bundled browser bridge.
`shinywidgets/_render_widget.py`	Documents selection state across re-renders.
`shinywidgets/_plotly.py`	Declares selection and brush accessors as state-based.
`js/src/plotly.ts`	Reconstructs and sends selection state on initial binding.
`CHANGELOG.md`	Records the accessor and selection-seeding changes.

_{Review effort: Balanced}

…ure read - Coordinates the trace doesn't carry explicitly (implicit x, bar x0/dx) fall back to calcdata, matching what live event point objects supply. - An empty drag selection seeds points: [] plus the region, the same message plotly.js sends for one; only no-points-and-no-region remains the no-data case. - The lasso vertex regex accepts signed exponents (1e-7), which it previously failed to match or misparsed. - The fixture's ungated synced_selection read attached the accessor at startup, so the gated reads weren't genuinely late and the test could pass pre-change via live forwarding. Removed; tests now wait on graph-div selection state plus plotly.js's event-emission delay, and would fail pre-change.

… spike/plotly-selection-state

Copilot

🟡 Changes recommended

Selection seeding can assign unrelated bin counts as histogram y values and drops valid typed-array selections, producing incorrect or missing selected-row data.

3 open issues

3 resolved since last review

1 file not reviewed

shinywidgets/static/output.js: Generated file

_{Review effort: Balanced}

- A selection preconfigured from plotly.py arrives as a typed array, which Array.isArray rejected, dropping the seed; normalize with jsonable() first. - Histogram selectedpoints are sample indices but histogram calcdata is bins, so the calcdata fallback seeded bin counts as y; the live event reads only the original arrays for aggregated traces, so skip the fallback for histograms to match.

Copilot

🟡 Changes recommended

Selection seeding misreports Scattergl coordinates, corrupts date-axis lasso vertices, and ignores subplot axis references, producing results inconsistent with live events.

4 open issues

2 resolved since last review

1 file not reviewed

shinywidgets/static/output.js: Generated file

_{Review effort: Balanced}

- The calcdata coordinate fallback is restricted to verified layouts and moved into calcCoord(): scatter and bar align per data index; scattergl keeps resolved coords in typed arrays on the single calc entry's .t (its top-level x/y are false sentinels); other types are skipped, so unverified layouts omit keys rather than seed wrong values. - Lasso vertices are parsed generically (M/L segments split on the comma), which also restores date-axis values: plotly stores them as date strings with '_' for the date/time separator, live lassoPoints use a space. - The seeded region is keyed by the saved selection's xref/yref (x2/y2 on later subplots), matching live range/lassoPoints keys. Each verified against the compiled bridge: scattergl y-only seeds x=0,1; a date-axis lasso seeds vertices identical to the live event; a second-subplot selection seeds range keyed x2/y2 with the subplot trace's points.

Copilot

🟡 Changes recommended

The selection-path regex includes the closing Z/z command in the final coordinate, corrupting regions returned by brushed(); the capture must be corrected and the bundled output regenerated.

2 open issues

3 resolved since last review

1 file not reviewed

shinywidgets/static/output.js: Generated file

_{Review effort: Balanced}

A preconfigured closed path (M0,0L1,0L1,1Z) captured '1Z' as the final coordinate, which Number() rejects and the date fallback returned as the string '1Z'.

Copilot

🟡 Changes recommended

Existing selections report incorrect multicategory coordinates, and seeded date-axis brush coordinates retain SVG whitespace, requiring source fixes and bundle regeneration.

3 open issues

1 resolved since last review

Exclude closing path commands from the final coordinate capture

1 file not reviewed

shinywidgets/static/output.js: Generated file

_{Review effort: Balanced}

- Multicategory axes store one array per category level, so trace[key][i] is a level array, not the coordinate; reconstruct it level-wise (verified: seed x becomes ['A', 'two'], identical to the live event). - Lasso vertex tokens are trimmed before conversion, so SVG formatting whitespace doesn't survive into date-valued region coordinates.

Copilot

🟡 Changes recommended

The multicategory selection logic skips points beyond the number of category levels, producing an incorrect coordinate; the source fix and regenerated output.js are still needed.

1 open issue

Handle selected points beyond multicategory levels · New

3 resolved since last review

1 file not reviewed

shinywidgets/static/output.js: Generated file

_{Review effort: Balanced}

A multicategory trace array's length is the category level count, so for a selected index beyond it trace[key][i] is undefined and the per-point check fell through to the wrong path; detect array-of-arrays on the trace attribute itself (verified: index 2 seeds ['B', 'one']).

Copilot

🟡 Changes recommended

The selection seed incorrectly treats selectedpoints=[] as absent, causing selected() to return None instead of an empty DataFrame; preserve the empty selection and regenerate the bundled output.

1 open issue

Preserve active empty selections from preconfigured selectedpoints · New

1 resolved since last review

Handle selected points beyond multicategory levels

1 file not reviewed

shinywidgets/static/output.js: Generated file

_{Review effort: Balanced}

plotly treats selectedpoints: [] as an active empty selection (every point gets unselected styling), distinct from an absent selectedpoints; the seed now emits points: [] for it (empty DataFrame, matching the empty-drag case) while a plot with no selection arrays anywhere still seeds the no-data message (None).

Copilot

🟢 Approval recommended

No active review issues were identified for the changes preserving Plotly selections made before the first accessor read.

No open issues 🎉

1 resolved since last review

Preserve active empty selections from preconfigured selectedpoints

1 file not reviewed

shinywidgets/static/output.js: Generated file

_{Review effort: Balanced}

cpsievert added 2 commits November 4, 2026 11:52

cpsievert requested a balanced review from Copilot November 4, 2026 17:55

Copilot started reviewing on behalf of cpsievert November 4, 2026 17:56

View session