[cpsievert]

Fixes #471.

Summary

@render_altair's selected_rows() silently mis-filtered selections whose Vega store spans several views: on a layered or concatenated chart whose subcharts use different data frames, every view's selection entries were applied to a single frame, returning the wrong rows when the datasets shared column names (or raising when they didn't).

Now the store is partitioned by the entries' unit (the view name, readable from the chart object), and each view's own data frame is filtered:

• A selection that touches several views with different data returns a dict of frames keyed by view name. Give subcharts name= for readable keys; altair's generated view_... names are used otherwise.
• df= also accepts such a dict; passing a single frame for a multi-dataset selection raises instead of silently mis-filtering.
• Selections touching a single view — or views sharing the identical data frame object — return a single frame, exactly as before. A per-layer selection on a multi-data layered chart (previously unsupported, since chart.data is undefined there) now filters that layer's data.
• Bare point selections (alt.selection_point() without fields=/encodings=) raise an actionable error on multi-dataset charts: their store holds positional _vgsid_ ids, which Vega assigns globally across the whole chart's dataflow (verified in a live browser: clicking the right chart's 6th point of a two-frame hconcat yields _vgsid_: 16), so they can't be mapped back to one view's rows. Declaring fields=/encodings= makes the selection value-based and fully supported.

Implementation notes

• named_subcharts() walks layer/concat charts to map view name → subchart (duplicate names raise; facet/repeat charts are out of scope and fail loudly rather than mis-filtering).
• selection_views() reads the hoisted selection param's views to learn which views the selection participates in, falling back to the store's units when the param isn't readable (e.g. alt.Chart.from_dict charts).
• The narwhals replay of Vega's selectionTest is unchanged — it's just applied per view via filter_selection_by_unit().

Verification

• 14 new unit tests: per-unit partitioning, dict return and empty-selection key stability, shared-data-stays-single-frame regression guard, facet regression guard, df= dict/scalar rules, duplicate and unresolvable view errors, vegafusion transformer.
• A new Playwright test drives real gestures on a multi-dataset hconcat: a brush in the left chart filters only the left frame, and a click in the right chart selects exactly one right row (tests/playwright/altair_accessors).
• make check is clean (ruff, pyright, 1157 unit tests passing).

import altair as alt
import pandas as pd
from shiny.express import render
from shinywidgets import render_altair

left_df = pd.DataFrame({"x": range(10), "y": range(10)})
right_df = pd.DataFrame({"x": range(10), "z": range(10)})

brush = alt.selection_interval(name="brush")


@render_altair
def chart():
    left = alt.Chart(left_df, name="left").mark_point().encode(x="x", y="y").add_params(brush)
    right = alt.Chart(right_df, name="right").mark_point().encode(x="x", y="z").add_params(brush)
    return alt.hconcat(left, right)


@render.text
def selected():
    rows = chart.selected_rows("brush")  # a dict of frames keyed by view name
    return str({name: len(frame) for name, frame in rows.items()})

[Copilot]

🟡 Changes recommended

Unresolved data-source handling can incorrectly reuse a parent frame and permit unsafe positional selections across mixed sources, causing selected_rows() to return incorrect or empty results.

2 open issues

• Do not inherit parent frame for explicit child data · New
• Count non-frame datasets when validating positional selections · New

3 resolved since last review

• Allow scalar df overrides for views sharing one dataset
• Reject positional selections across distinct chart datasets
• Inherit each view’s nearest ancestor data declaration

_{Review effort: Balanced}

[cpsievert]

@copilot please re-review

[Copilot]

@copilot please re-review

Re-review found two regressions:

• shinywidgets/_altair.py:106: repeat charts aren't traversed, so selections spanning repeated views now fail even with a scalar df=; previously they filtered that frame.
• shinywidgets/_render_widget.py:693: views sharing the same URL are treated as different datasets, rejecting a scalar df= that previously worked.

The current PR CI checks pass.

[cpsievert]

@copilot please re-review

[Copilot]

@copilot please re-review

Re-review found a remaining silent mis-filter in shinywidgets/_render_widget.py:704–711: if a concatenation has root data frame A and a nested repeat uses frame B, its generated views aren't in view_frames(), so they are assigned A. Brushing the repeat then returns rows from A instead of B. Resolve the repeat views against their own inherited/declared source, or fail rather than falling back to the root frame.

The latest CI run is still in progress; the preceding run passed.

[cpsievert]

@copilot please re-review

[Copilot]

@copilot please re-review

Re-review found one remaining mis-filter: in shinywidgets/_render_widget.py:668–697, a compound alt.Chart.from_dict() selection without readable views falls back to the currently populated store units. A global selection can have only the right view in its store even though it participates in both views; selected_rows(df=left_frame) then silently filters the left frame instead of rejecting an ambiguous scalar. Imported parameter names can also be ParameterName objects, so selection_views() may miss even explicitly declared views. Normalize those names and handle unknown participation conservatively on multi-dataset charts.

All Python and Playwright CI jobs passed on the current commit.

[cpsievert]

@copilot please re-review

[Copilot]

@copilot please re-review

Re-review found two remaining cases:

• shinywidgets/_render_widget.py:685–689: alt.SelectionParameter(...) naturally has views=[]. Vega applies it to both views of a concatenation, but a selection in the right view filters the root/left frame instead. Treat an empty list as unknown participation too.
• shinywidgets/_render_widget.py:668–672: with missing view metadata and no df=, inferring views from active store entries still produces an incomplete, unstable result (a single frame or an error before interaction rather than a per-view dict). Resolve participation or reject the ambiguity.

All Python and Playwright CI jobs passed on the current commit.

[cpsievert]

@copilot please re-review

[Copilot]

@copilot please re-review

Re-review found one compatibility regression in shinywidgets/_altair.py:171: two imported NamedData references to the same dataset are counted separately by identity. Consequently, a shared-source hconcat rejects selected_rows(df=frame) as multi-dataset, although Vega uses one source and the scalar frame worked before this PR. Deduplicate named references by dataset name.

CI’s ipydatagrid lifecycle test failed with a transient view count of 6 rather than 5; a retry is already in progress. The Python jobs passed.

[netlify]

❌ Deploy Preview for shinywidgets failed.

Name	Link
🔨 Latest commit	63dd16a
🔍 Latest deploy log	https://app.netlify.com/projects/shinywidgets/deploys/6af7726de496126f60e27c6f