version: "1.0.0" name: flint-interaction-author description: "Use when: a Flint chart should respond to the reader (highlight, zoom, brush, legend toggle, reorder, annotate, explore, link charts); an application or an agent changes a mounted chart from outside or reads what it shows; or a gesture no preset gives must be built. Produce one valid interaction_spec, one ChartUpdate, or code against flint-chart/interactive, with no invented names."

flint-chart: interactions and updates on a chart

A Flint chart is shared state. A reader's gesture becomes an event, a handler turns the event into an update { id, ops }, updates compose as layers, and the layers render into a state the host and the agent read back. An application and an agent write the same kind of update directly. The full model is in the documentation: https://microsoft.github.io/flint-chart/#/documentation/interaction-introduction

The chart itself (chart_spec, semantic_types) is the flint-chart-author skill. Everything here runs on the Vega-Lite backend; a static render and the other backends leave interaction_spec and updates untouched.

What you produce (and what you do NOT)

• DO use only the names in this skill: preset types, options, ops, states,

triggers, cursors, event actions. Every name here is one the library ships. When the request asks for a behaviour no preset gives, say which preset comes closest and what it leaves out, then offer a bespoke definition if the host runs code.

• DO add behaviour only when the request or the session asks for it. A

static image or a bare spec gets no interaction_spec.

Which workflow the user is asking for

• Presets through the MCP tools: put the object beside chart_spec as

interaction_spec in the input of create_chart, which reports its warnings; validate_chart reports the same without rendering, for when you do not draw. list_interaction_presets with chartType lists the presets that chart type supports, what each does, and its options.

• An update through the MCP tools: there is no handle on the chart. Pass

the layers as the updates argument of create_chart or update_chart, beside the input; see "Step 4 — send the layer and read the result".

• Code, only when the user asks for code: mount with mountChart, or

<FlintChart> in React; presets as factory calls, updates with applyUpdate, state with getState and onChange. The programming API is documented at https://microsoft.github.io/flint-chart/#/documentation/interaction-api

• A bespoke interaction needs a host that runs code. In the MCP view, the

answer is the closest preset and what it leaves out.

Add interactions to the chart

Step 1 — read the request and the chart

Three things in the request decide the entries, read in this order:

1. What the user named. A gesture or a preset in their words: "brush a

period", "zoom", "toggle the legend", "highlight on click". Each is one job.

1. The goal they stated. "Explore", "compare the regions", "find

outliers", "present this", "a control in my app". A goal is the set of questions the reader will ask of this chart; each question is one job.

1. Nothing named, only "interactive", a live view, a dashboard. The chart

and the data say what the reader can do with it.

Then read the chart spec: the chart type, the field on each channel, and the shape of the data. What a chart invites a reader to do:

Step 2 — pick presets

Before any entry, a mounted chart already shows a tooltip with the mark's values on hover and nothing else. An entry adds one behaviour.

Take one row of the map per job. A named gesture takes the row whose "Gesture" is that gesture. A question takes the row whose "Answers" is that question, on a chart whose "Needs" it offers; a broad goal takes two or three rows, the ones the chart invites most. When nothing is named, the rows the chart invites, two or three. A row that answers no job on the request stays out.

Within a family the rows answer neighbouring questions, so one chart takes one of them: inspect reads the marks near the pointer, inspect-index reads every series at one index; select takes an area, brush-x a period, lasso a shape; navigate moves continuously, brush-zoom jumps into a box; double-activate, long-press, and context-activate are three spellings of "hand this mark to the app" for a mouse, a finger, and a menu.

"Needs" is what the chart type offers; list_chart_types has already applied it to the list it returns. The data confirms a few at mount, in Step 3. A viewport a gesture commits, and a reset, fly over 400 ms; "transition": { "duration": 0 } or "resetTransition": { "duration": 0 } jumps. Panning and wheel zooming follow the pointer.

`filter-controls`. It draws real controls beside the chart and changes the rows the chart shows, so it answers a request that names a field to filter by ("filter by region", "one year at a time", "let me pick the group"), or a data column the chart does not encode whose values are versions of the same chart: a year in a snapshot scatter, a population group, a scenario. A field the chart already draws has its own preset: the colour field legend-toggle, a continuous axis brush-x or navigate, a category on an axis click-highlight; and the measure the chart plots is the chart. List fields explicitly, one or two, three at most: unlisted, the preset picks up to four categorical and temporal fields itself. One entry per chart. A fixed subset belongs in the data, and slices side by side are a facet in the flint-chart-author skill.

Step 3 — resolve conflicts

Each trigger has one owner. At mount, admission gives a trigger to the first entry that asks for it; a later entry that asks for the same trigger yields it with an info warning when it can keep the rest, or is dropped with a warning when it cannot. The chart still renders. Four triggers are asked for by more than one preset:

click-highlight yields a legend or an axis click on its own, with an info warning; targets keeps the spec silent. accessible-navigation takes no pointer trigger, so it sits beside any entry.

The data confirms what the chart type cannot promise. An entry that fails one of these at mount is dropped with a warning:

• legend-toggle needs a colour field bound to a discrete legend.
• navigate and brush-zoom need a continuous, unfaceted axis.
• drag-reorder needs a discrete axis in the bound encodings.
• inspect-index needs an index axis, which it finds itself: the temporal

axis, else the discrete axis beside a measure, else x. A horizontal bar chart reads along y with no option; axis overrides the choice. seriesBy names the series field when the chart draws one series at a time.

• click-group-focus, hover-group-focus, and linked-brush need groupBy,

a bound field.

Step 4 — write and validate

The object is { "interactions": [ ... ] }, beside chart_spec as interaction_spec. Every entry is { "type": <preset>, "id"?: <string>, "options"?: { ... } }:

• type is a preset from the map, supported by the chart type per

list_chart_types.

• options holds the preset's own options; list_interaction_presets lists

every one with its type (see "Preset options"). reset is one of them: omit it for the default list in the map, set it to free a trigger. A preset that keeps nothing (hover-group-focus, inspect, context-activate) accepts no reset.

• id sits on the entry and is needed only when one preset appears twice.
• In code, the same entry is a factory call with the same options:

clickHighlight({ targets: ['mark'] }), brushX(), navigate({ axes: 'x' }).

{
  "interactions": [
    { "type": "click-highlight", "options": { "targets": ["mark"] } },
    { "type": "navigate", "options": { "axes": "x" } },
    { "type": "legend-toggle" }
  ]
}

create_chart checks the object when it builds the chart; validate_chart runs the same check in code or CI. Do not run both for the same chart. A malformed entry is an error: an unknown type, an option beside type, an id inside options, a bare string entry, a missing required option, an unsupported reset gesture, a duplicate id. A dropped or yielded entry is a warning that names the entry, what it needed, and the chart type:

Interaction "legend-toggle" requires a discrete legend; Bar Chart has none. The interaction was dropped.
Interaction "click-highlight" yields legend clicks to "legend-toggle".

Fix or drop what it warns about. Return the bare object, strict JSON, unless the user asks for commentary.

When no preset fits

A host that runs code mounts a CanvasInteractionDef in options.interactions beside presets: a trigger, what it affords, and a handler that returns the layer for its id.

import { rectangleTrigger, type CanvasInteractionDef } from 'flint-chart/interactive';
const timebox: CanvasInteractionDef = {
  id: 'timebox',
  eventSource: rectangleTrigger('contain'),
  affordances: { plot: { cursor: 'region' } },
  reset: ['click-none', 'escape'],
  handle(event) {
    if (event.action !== 'select-region' || event.phase !== 'commit') return null;
    const x = event.geometry.domain?.x;
    const y = event.geometry.domain?.y;
    if (x?.kind !== 'interval' || y?.kind !== 'interval') return null;
    const keep = seriesInsideBox(x.start, x.end, Number(y.start), Number(y.end)); // host logic over host rows
    return {
      id: 'timebox',
      ops: [{ op: 'set-style',
        targets: keep.map((series) => ({ select: { key: { Series: series } } })),
        value: { state: 'emphasized' } }],
    };
  },
};

Triggers.

match is 'intersect' (default) or 'contain' and decides which marks the region's target lists. guide styles the drawn shape; false hides it. A region is ephemeral: its guide and its preview end at the commit. A brush or select with mode: 'stateful' keeps its region on screen and editable, and event.operation reads create, move, resize-leading, resize-trailing, or clear. Spread a trigger to tune it: { ...hoverTrigger, defaultAssistDistance: 28 } acquires the nearest mark within 28 px. A chart admits a trigger when it offers what the trigger needs, per the map's "Needs"; a code definition the chart cannot honour throws at mount with a message that names what it needed.

The event.

interface CanvasInteractionEvent {
  action: string;                                     // from the table
  phase: 'start' | 'preview' | 'commit' | 'cancel';   // clicks and hovers skip start
  operation?: string;                                 // a stateful brush's edit, or pan | zoom | reset
  geometry: { plot?, domain?, projection? };
  target: { visual: { kind, role }, elements: [{ value, records? }] } | null;
  modifiers?: { shift, ctrl, meta };
  description?: { kind, type, content, text };  // on focus-element from accessible-navigation
}

• target.elements[i].value holds the mark's encoded values, derived fields

included; records holds the source rows when the runtime can prove them. A region target lists every mark inside. A legend hit's value is a LegendTargetValue: { channel, field, domain: { kind: 'value', value } }, or { kind: 'interval', start, end } on a continuous legend. An axis-label hit's value is an AxisTargetValue: { axis, field, value }.

• geometry.domain is inverted through the plot's scales: a brush on a time

axis gives domain.x = { kind: 'interval', start: Date, end: Date }, a linear axis gives numbers, a discrete axis gives the category.

• description arrives on focus-element while accessible-navigation walks

the chart: what the focused element is (type, such as "Bar") and what it represents (content), with the spoken text. The walk's emphasis reports through onChange as a preview.

• geometry.projection.kind === 'path' during a drag over an overlay with

projectable: true: segment.start.value and segment.end.value are the overlay rows on each side of the pointer and segment.t the fraction between them.

The types. flint-chart/interactive exports these, and flint-chart/react re-exports them. onChange receives a ChartChange: the gesture's own target and geometry, and the chart's state after it. Generated from the source by npm run gen:reference:

<!-- event-types:start -->

interface ChartChange {
  // `preview` while a gesture runs, `commit` for a committed change, `cancel` when a gesture ends with none.
  phase: ChartChangePhase;
  // `host` for `updates`, `applyUpdate`, `clearUpdate` and `dispatch`; always `commit`.
  source: ChartChangeSource;
  changed: ChartStateFacet[];
  // The interaction behind the change: a reader's gesture, or an external interaction run by `dispatch`.
  interactionId?: string;
  action?: CanvasInteractionAction;
  // The gesture's own hit: the hovered point, or the marks at an inspected index.
  target?: SemanticTarget | null;
  // The gesture's geometry: a brushed range, an index value, or the viewport after a move.
  geometry?: CanvasInteractionEvent['geometry'];
  // The reset gesture a reader made: a click on nothing, Escape, or a double-click. Reported even when the chart held nothing to reset, so a host that shows a selection of its own can let it go.
  reset?: InteractionResetGesture;
  state: ChartState;
  previous: ChartState;
}
interface ChartState {
  chartType: string;
  // The marks the chart emphasizes or focuses now, previews included.
  selected: SemanticElement[];
  // The emphasized marks of each update, by update id; `selected` is their union.
  entries?: ReadonlyMap<string, ChartStateEntry>;
  // The legend values a toggle hides now.
  hidden?: ChartHiddenValue[];
  // The domain the plot shows now, when the chart navigates.
  viewport?: DomainGeometry;
  // The category window of each rail.
  windows?: Partial<Record<'x' | 'y', ChartCategoryWindow>>;
  categoryOrder?: unknown[];
  // The notes the chart holds now, previews included; `inView` tells which ones it draws.
  annotations?: ChartAnnotation[];
  // The active filter of each field, set by filter controls; absent fields show every value.
  filters?: Readonly<Record<string, FilterValue>>;
}
interface SemanticTarget {
  visual: {
    kind: 'mark' | 'path' | 'region' | 'widget' | 'handle' | 'legend' | 'axis' | 'chart' | 'title' | 'header';
    role: string;
  };
  elements: SemanticElement[];
}
interface SemanticElement {
  // Values represented by the mark's channels, or by a semantic control such as a legend item.
  value: Record<string, unknown>;
  // Contributing input records when provenance is available; zero or many may support one value.
  records?: Record<string, unknown>[];
}
interface LegendTargetValue {
  channel?: string;
  field?: string;
  // The entry's value on a reader's hit; read `domain`, which every legend target carries.
  value?: unknown;
  domain: LegendDomain;
}
type LegendDomain =
  | { kind: 'value'; value: unknown }
  | { kind: 'interval'; start?: number; end?: number };
interface AxisTargetValue {
  axis: 'x' | 'y';
  field: string;
  value: unknown;
}
interface DomainGeometry {
  x?: DomainCoordinate;
  y?: DomainCoordinate;
  points?: DomainPoint[];
  // The detail level a multi-level projected chart draws for this viewport.
  level?: string;
  // On a multi-level projected chart, the coarsest-level region under the plot centre: its feature id and properties plus the joined row fields.
  focus?: Record<string, unknown>;
}
type DomainCoordinate =
  | { kind: 'value'; value: DomainValue }
  | { kind: 'interval'; start: DomainValue; end: DomainValue };
type DomainValue = number | Date | string;
interface DomainPoint {
  x?: DomainValue;
  y?: DomainValue;
}
interface ChartHiddenValue {
  channel: string;
  value: unknown;
}
type FilterValue =
  | { in: unknown[] }
  | { range: [unknown, unknown] };

<!-- event-types:end -->

• context is the chart state plus available, every drawn mark, for

building refs, and selected, the current emphasis, for a toggle rule.

The handler.

• It returns the complete layer for the id on every preview and commit;

the runtime drops the preview on cancel. A null on a preview shows nothing until the commit, which is how a rule applies to the committed box only. A region gesture keeps a preview layer under the interaction's id while the drag runs, so a commit after previews restates the layer; a null there lets the last preview stand.

• It maps geometry to the host's model and the model to ops. Host side

effects (state, panels, overlays the host draws) come from flint-interaction or onChange, which carry the same event; the handler itself returns ops and leaves the DOM to the runtime.

• Triggers have one owner for code definitions as for presets; two code

definitions on one trigger throw at mount. navigate({ pan: false }) goes beside a region trigger.

• An element drag owns the press on the marks it affords. A press there is a

drag with phase start; target is the pressed mark through every phase. The layer returned at start is kept as the gesture's preview, and a release without movement commits that preview with no further call. No click definition fires on those marks, so the definition that drags a mark also answers its press.

• A click definition that affords hover on marks receives preview events

for the hover cue. Return null or an emphasis for them, and the layer on commit.

Patterns.

Host-drawn detail, the shape of it:

const probe: CanvasInteractionDef = {
  id: 'probe',
  eventSource: { ...hoverTrigger, defaultAssistDistance: 28 },
  affordances: { mark: { hover: 'target' } },
  handle() { return null; },
};
container.addEventListener('flint-interaction', (e) => {
  const { interactionId, event } = (e as CustomEvent<FlintInteractionEventDetail>).detail;
  if (interactionId !== 'probe' || event.phase !== 'preview') return;
  const row = event.target?.elements[0]?.records?.[0];
  if (row) lens.show(row);                 // the host draws; the chart is unchanged
});

Update the chart

Step 1 — read the state

A host hands the model the chart input and the state, as text or as data. The MCP chart view does this after every build (first mount, new updates, a panel edit) and every committed gesture, and each send replaces the last:

Chart "Life expectancy": Line Chart, theme economist.
Edited in panel: theme default → economist.
Encodings: x = year, y = life, color = continent.
Interactions: legend-toggle, navigate.
Agent updates (yours):
- asia:
  ✓ emphasizes 3 marks where continent = Asia
- oceania:
  ✗ emphasizes marks where continent = Oceania: no row has continent = Oceania.
User interactions:
- Hidden: Africa.
The y axis shows life from 70 to 86.

The state is in field terms: what the chart is after the reader's panel edits, each op of your updates as applied (✓) or not and why (✗), what the reader selected, hid or filtered, and the viewport on view when the chart navigates. validate_chart, create_chart and update_chart return the same report before the reader touches the chart, with User interactions: none. From it, decide what the reader should see next: which rows stand out, what one note says, which range frames them. The rows you were given are the source of every value; the chart hands over no copy of the data, and nothing is computed that the chart cannot verify.

Step 2 — choose ops

Every change from outside is one update, { id, ops }. Each op does one kind of thing:

StyleSpec is { state?, opacity?, visible?, fill?, stroke?, strokeWidth?, mutedOpacity? }. Two states matter from outside: emphasized keeps the targets at full ink and mutes every other mark; normal with targets: [] clears the layer. opacity on the targets fades a subset and leaves the rest alone; opacity: 0 hides marks and keeps the axes, where visible: false removes the rows and shrinks the scales. The mute level is one per chart, fixed at mount from the lowest mutedOpacity the mounted interactions carry (0.25 when none).

Overlay marks are line, point, rule, rect, text. encodings name fields of the overlay's own rows, projected through the chart's x, y, and colour scales. style takes stroke, strokeWidth, strokeDash, fill, fillOpacity, opacity, pointRadius, fontSize, fontWeight, textAlign (start, middle, end), dx, dy. A line takes its stroke and a text its fill from the color encoding; a point takes style.fill. A text sits on the data, so it draws a halo in the chart's background; stroke sets another colour, strokeWidth its width, and stroke: 'none' drops it, as a faint watermark wants.

Two ops need a preset the chart mounts: set-viewport needs navigate or brush-zoom on that axis, set-order needs drag-reorder. The host says which are mounted, or the state shows viewport and categoryOrder. On a chart that mounts neither, say so and offer an emphasis; set-data replaces what the chart shows and is not a frame. Legend values belong to legend-toggle: no update hides or restores one, and a series fades with opacity instead.

Step 3 — name targets and values

A target is one of three forms:

{ select: { key: { Country: 'Japan' } } }                 // marks with these values
{ select: { where: { Region: { in: ['Asia', 'Europe'] }, Month: { range: ['2025-03', '2025-06'] } } } }  // a set of marks
{ visual: { kind: 'mark', role: 'mark' }, elements }      // elements from an event or the state

• A where describes a set, in the words filters use: each field in a list

or within a range (ends included; null leaves an end open), and every field must hold. Use it for a period, a band of values, or several categories at once, where a key would need one target per row. Only set-style takes a where; a note or an order names one mark, so it takes a key.

• A selector key names fields the chart encodes (x, y, color, detail,

size, …) and matches rows by equality, so '2007' and 2007 differ. One key may cover many marks ({ Continent: 'Africa' } takes every African mark); several fields narrow it; an annotation's key covers one. On a line or area chart, a key on the series field covers the whole line and a key on the x field covers the segment that starts at that row, so a run of segments is one target per row. The last row of a line starts no segment; its key takes the segment that ends there. A note on an x key marks the segment, for a change from one row to the next ("2019 → 2020: +4%"); anchor: "point" marks the row's own point, for its value ("Dec 2022: $0.87B"). A note on the last row is always at its point.

• A field the chart does not encode resolves to nothing. Encode the identity

field (detail on a scatter), or pass elements from event.target, state.selected, or context.available through as { visual, elements }, as they are: an element's value carries derived fields on a bar, pie, heatmap, or waterfall mark (a stack start, a sort index), so the element resolves where a key copied from it does not.

• A target that matches nothing is listed in the result as unresolved, and

nothing nearby is taken in its place.

A key value, an overlay row, and a viewport bound hold what the parsed row holds: a CSV 120 is the number 120. A temporal field (Year, YearMonth, Date, …) takes a date written the way the data writes it ("2014-01", 2014 on a Year field) or UTC epoch milliseconds. On a discrete axis the value is the category. Overlay rows need the fields their mark reads: line, point, text read x and y. A rule with only x is a vertical line across the plot, and one with only y a horizontal threshold; a rect with only x and x2 is a full-height band, and one with only y and y2 a full-width band. With all four channels, rule and rect run from (x, y) to (x2, y2). An overlay whose rows do not project is reported as unsupported.

{ "id": "agent", "ops": [
  { "op": "set-style", "value": { "state": "emphasized" },
    "targets": [{ "select": { "key": { "month": "2014-01" } } }, { "select": { "key": { "month": "2014-02" } } }] },
  { "op": "set-overlay", "name": "threshold", "value": { "mark": "rule", "role": "reference",
    "data": { "values": [{ "y": 1e9 }] }, "encodings": { "y": { "field": "y" } } } },
  { "op": "set-overlay", "name": "period", "value": { "mark": "rect", "role": "band",
    "data": { "values": [{ "start": "2022-11", "end": "2026-08" }] },
    "encodings": { "x": { "field": "start" }, "x2": { "field": "end" } }, "style": { "fillOpacity": 0.12 } } }
] }

Step 4 — send the layer and read the result

Ids are layers. One id holds one concern: the agent's update, a story step, a linked selection. Use one stable id, such as agent, and send the whole layer each time:

• An update with an id already retained replaces that layer, and carries only

what it lists: an omitted viewport flies home, an omitted order returns to the input order, an omitted annotation disappears. The ops that should hold are repeated; the ops the request replaces are dropped.

• Layers compose in insertion order, and a replaced layer keeps its place:

styles accumulate across layers; for an annotation, an order, a data swap, or an overlay name, the layer inserted last wins. One kind of op lives in one layer: a drag and a playback that both swap the rows write the same id.

• A canvas interaction retains its layer under its own id; a host may write

or clear that layer by the id, as a Reset button does.

• A gesture reset (click-none, escape, double-click) clears preset

layers only. A host layer stays until the host clears it, or until a layer with set-style, targets: [], state: 'normal' replaces it.

• After set-data the marks are new; a style layer is sent again with the

data.

The result. applyUpdate, setUpdates, and dispatch return a ChartUpdateResult: { status: 'applied' | 'partially-applied' | 'unsupported', resolvedTargets, unresolvedTargets, unsupportedOps }. An unresolved target has a wrong field or value; an unsupported op needs a preset the chart does not mount. What the chart shows is what the result says it shows: fix an unresolved key or drop an unsupported op before describing the chart.

From the MCP view. There is no applyUpdate to call. Pass the same layers as the updates argument of create_chart, beside the chart input, not inside it:

{
  "data": { "values": [ ... ] },
  "chart_spec": { "chartType": "Bar Chart", "encodings": { "x": "region", "y": "revenue" } },
  "updates": [
    { "id": "agent", "ops": [
      { "op": "set-style", "targets": [{ "select": { "key": { "region": "East" } } }], "value": { "state": "emphasized" } },
      { "op": "set-annotation", "target": { "select": { "key": { "region": "West" } } }, "value": { "text": "Lowest in Q3" } }
    ] }
  ]
}

show opens the chart with the layers applied, with the same look as a user action. To show the chart in a new state, call update_chart with its id and the layers wanted now, then show again. Each show opens a new card with exactly those layers; earlier cards keep what they showed. The argument is state, and interaction_spec is behaviour, so each stays where it is. The reader's own brush or selection does not survive the new view; the context message reports it in field terms, and it is carried as a layer when it still serves the reader.

Call create_chart directly; it validates the input and updates itself, so do not call validate_chart first. A malformed layer or an invalid spec is an error and keeps nothing. Otherwise the result is the report above, one line per op, led by a count when an op did not apply. The chart draws without a ✗ op: fix it with update_chart (a key no row has, an overlay encoding written as "x": "start" instead of { "field": "start" }, a viewport with no navigation preset) before you call show, or tell the reader what is missing. The reader's later brushes and clicks reach you with their next message, not in the same turn.

Use Flint in an application

Mount

npm install flint-chart vega vega-lite vega-tooltip

import { mountChart, navigate, legendToggle } from 'flint-chart/interactive';
const surface = mountChart(container, input, {
  backend: 'vegalite',
  renderer: 'svg',
  interactions: [navigate({ axes: 'x' }), legendToggle()],
  chartId: 'sales',
});
await surface.ready;
const warnings = await surface.warnings; // interactions the chart could not honour
// …
surface.destroy();

• input is a ChartAssemblyInput. Presets go in input.interaction_spec as

JSON or in options.interactions as factory calls. Both may appear; spec entries mount first, and a code definition with a spec entry's id replaces it with an interaction_overridden info warning.

• A malformed interaction_spec, a repeated code id, or a code definition

the chart cannot honour fails the mount with an error that names it. A spec entry the chart cannot honour is dropped with a warning instead.

• options.availableSize: { width?, height? } is the room the host gives the

chart. It becomes the canvasSize ceiling on the sides given, so the chart lays out for that room rather than being scaled into it.

• options.updates holds retained updates present at the first render, for a

chart that starts with a part hidden or framed.

• renderer: 'svg' for charts whose marks a host reads from the DOM, or under

a few thousand marks; 'canvas' for maps and dense charts.

• A Year or Date field makes a temporal axis whose rows hold Dates or

ISO strings; a bare number such as 2013 reads as milliseconds after 1970. A semantic_types entry may be an object, { semanticType, intrinsicDomain: [lo, hi] }; the intrinsic domain holds an axis still while the rows change.

• options: { addTooltips: false } on the input is the switch for the

tooltip, for a chart whose hidden values it would give away.

• Build with a bundler. The interactive entry reaches every backend through

lazy imports with no file extension; when echarts, chart.js, and plotly.js-dist-min are not installed, mark them external (esbuild --splitting --external:echarts --external:chart.js --external:plotly.js-dist-min; Vite build.rollupOptions.external).

• In React, use <FlintChart> from flint-chart/react rather than mounting

in an effect; see "In React" below.

In React

import { FlintChart, type FlintChartHandle } from 'flint-chart/react';
import { clickHighlight } from 'flint-chart/interactive';
const chartRef = useRef<FlintChartHandle>(null);
<FlintChart
  ref={chartRef}
  spec={input}
  interactions={[clickHighlight({ targets: ['mark'] })]}
  width="100%"
  fit="relayout"
  onChange={(change) => setSelected(change.state.selected)}
  onError={(error) => report(error)}
/>;

The ref's FlintChartHandle holds surface (the mounted surface or null), applyUpdate, clearUpdate, dispatch, getState, and refresh. A remount keeps the old chart on screen until the new one is ready.

Read the chart

Each element's value is the mark in field terms: a bar's category and measure, a legend item's { channel, field, domain }, a histogram bar's { field, range }. The host owns the rows and queries them from the value and the geometry.

onChange fires after every render with a ChartChange: phase (preview while a gesture runs, commit for a committed change or a host call, cancel when a gesture ends with nothing), interactionId and action (absent for a host call and for a reset), the gesture's own target and geometry, and the state after. An overlay or a data swap leaves no trace in the state, so the host keeps that model itself.

const valuesOf = (elements: readonly SemanticElement[]) => elements.map((element) => element.value);
// A panel follows the gesture live, else shows what the chart shows.
surface.onChange(({ phase, target, state }) => {
  const hit = valuesOf(target?.elements ?? []);
  panel.show(phase === 'preview' && hit.length > 0 ? hit : valuesOf(state.selected));
});
// An agent context reads committed state only: values and ranges, never rows.
surface.onChange(({ phase, state, geometry }) => {
  if (phase !== 'commit') return;
  sendContext({ selected: valuesOf(state.selected), range: geometry?.domain, hidden: state.hidden, viewport: state.viewport });
});

The container also emits a bubbling DOM event, flint-interaction, for every canvas interaction, before the chart reacts: detail.chartId, detail.interactionId, and detail.event, the same CanvasInteractionEvent a handler gets. It fires for a definition whose handle returns null, so it is the path for a host that draws its own response to a gesture.

Link charts, and drive a chart over time

linkCharts links mounted charts on the fields they share. A selection on one chart (a click, a brush, a group focus, a hover) becomes one condition, a part per field; every chart emphasizes the marks that meet it, the source included, through one linked-selection layer; a reset on any chart clears it everywhere.

import { linkCharts, mountChart, clickHighlight, brushX } from 'flint-chart/interactive';
const link = linkCharts({
  bars: { input: barsInput, surface: mountChart(barsEl, barsInput, { backend: 'vegalite', interactions: [clickHighlight()] }) },
  trend: { input: trendInput, surface: mountChart(trendEl, trendInput, { backend: 'vegalite', interactions: [brushX()] }) },
}, { fields: ['region', 'month'] });
for (const [id, chart] of Object.entries(await link.links)) console.log(id, chart.notes); // what each chart left out
link.onCondition(({ condition }) => table.filter(condition)); // { region: { in: ['Asia'] }, month: { range: [start, end] } }
// link.clear(); link.destroy();

A table, a filter or any control beside linked charts joins the same condition: link.select(parts) sets parts as if a chart had selected them, and onCondition gives the table what to show. One condition, so a row click, a region picker and a brush combine (AND) and a reset on any chart clears them all.

row.addEventListener('click', () => link.select({ country: { in: [row.dataset.country] } }));
regionPicker.addEventListener('change', (e) => link.select({ region: e.target.value ? { in: [e.target.value] } : null })); // null lets go
link.onCondition(({ condition }) => renderTable(rows.filter((r) => meets(r, condition))));

• select only takes fields in fields; a chart shows a part when its marks

carry the field (encode it, or put it on detail).

• The link emphasizes the matching marks; to drop rows instead (a filter), give

the charts filtered data as below.

import { FlintChart, useLinkedCharts } from 'flint-chart/react';
const link = useLinkedCharts({ bars: barsInput, trend: trendInput }, { fields: ['region', 'month'] });
<FlintChart spec={barsInput} {...link.chart('bars')} />;
// link.select({ region: { in: ["Asia"] } }) from a table or control; link.clear()
<FlintChart spec={trendInput} {...link.chart('trend', { onChange: setPageState, updates: myUpdates })} />;
// the chart's own onChange, onRender and updates go through chart(id, own); a prop after the spread replaces the link's
// link.condition.condition is the same condition; filtered rows in `spec` keep it

• A chart receives a field when its marks carry it (encode it, on detail if

nothing else shows it), and sends through each selecting interaction whose subject includes a connected field. { charts: { trend: { receive: false } } } keeps a chart as drawn; send: false keeps its selections to itself.

• The link emphasizes; it never filters or re-aggregates. To filter, read the

condition (onCondition, link.condition) and give each chart its rows: beside linkCharts, a set-data update (a new spec remounts the chart, and the link holds the old surface); beside useLinkedCharts, a spec with the filtered rows.

• setUpdates replaces every layer, the link's included, until the condition

next changes; use applyUpdate and clearUpdate beside a link.

• linkCharts checks the charts with the validator, loaded on first use; pass

{ links } from validateLinks (flint-chart/validate) to skip that load.

For a link linkCharts does not express (a payload each chart reads its own way), wire onChange to an externalInteraction on the other charts:

const linked = externalInteraction<{ countries: string[] }>({
  id: 'linked',
  handle: ({ countries }) => ({
    id: 'linked',
    ops: [{ op: 'set-style', targets: [{ select: { where: { Country: { in: countries } } } }], value: { state: 'emphasized' } }],
  }),
});
source.onChange(({ phase, interactionId, state }) => {
  if (phase !== 'commit' || !interactionId) return;       // host writes carry no id: no loop
  void other.dispatch('linked', { countries: state.selected.map((element) => String(element.value.Country)) });
});

dispatch is also how an application drives a chart over time: a playback loop dispatches one frame per animation frame, and the handler returns the whole layer for that frame. Linked charts share one theme_spec so a value has one colour everywhere.

Things the surface holds fixed: no preset changes the layout, so a zoom never recomputes the bar step, the ticks, or the domain (filter the rows and mount again for that); the mute level is set at mount; Escape reaches a chart after a press on it, and only that chart; a preset's toggle (select, click-highlight) needs Shift, Ctrl, or Meta, and a plain second click replaces; a press on an overlay path resolves to no mark, so a click-none reset on the definition that drew the overlay clears it at that press.

Worked examples

In each example data is a placeholder; the host binds real rows.

Presets: a named gesture

User: "An interactive chart that I can brush." A line chart of sales by month, one series.

{ "interactions": [ { "type": "brush-x" } ] }

The user named one job, a brush along the period axis, and the row for it is brush-x. Hover already reads the values.

Presets: a goal on a category chart

User: "Which regions stand out? Let me look." A bar chart of revenue by region, no colour field.

The reader singles out a region and compares orders. click-highlight singles one out; drag-reorder lets the reader rank by hand. No colour field is bound, so legend-toggle has nothing to work with. Both presets ask for a different trigger, a click on a bar and a drag on an axis label.

{
  "interactions": [
    { "type": "click-highlight" },
    { "type": "drag-reorder" }
  ]
}

Presets: a multi-series line chart in an explore session

User: "Help me understand how these three regions moved." The chart:

{
  "data": { "values": [] },
  "semantic_types": { "month": "YearMonth", "sales": "Amount", "region": "Region" },
  "chart_spec": { "chartType": "Line Chart", "encodings": { "x": { "field": "month" }, "y": { "field": "sales" }, "color": { "field": "region" } } }
}

The reader compares the series at one month, hides a region, and points at a period. inspect-index reads every series at one month; legend-toggle has a colour field to work with; brush-x focuses a period. inspect-index with show: "all" asks for no legend click, so it sits beside legend-toggle.

{
  "interactions": [
    { "type": "inspect-index" },
    { "type": "legend-toggle" },
    { "type": "brush-x" }
  ]
}

Presets: a chart that drives the app

User: "In my dashboard, a click selects the country and a double-click opens its page. Zooming the y axis would help too." A scatter of gdp by life with color: continent and detail: country.

navigate holds the double-click by default for its reset, and double-activate needs it; reset: ["escape"] frees it. click-highlight takes targets: ["mark"] so the app's selection comes from marks only.

{
  "interactions": [
    { "type": "click-highlight", "options": { "targets": ["mark"] } },
    { "type": "double-activate" },
    { "type": "navigate", "options": { "axes": "y", "reset": ["escape"] } }
  ]
}

The app reads the selection from onChange on a commit whose action is click-element, and the activation from one whose action is double-activate-element.

ChartUpdate: an agent points out rows

User: "Which Asian countries are past 80 years? Point them out." The chart is the scatter above, with navigate on y, and the state says nothing is emphasized. The agent finds Japan and Singapore in its rows and answers:

{
  "id": "agent",
  "ops": [
    { "op": "set-style",
      "targets": [
        { "select": { "key": { "country": "Japan" } } },
        { "select": { "key": { "country": "Singapore" } } }
      ],
      "value": { "state": "emphasized" } },
    { "op": "set-annotation",
      "target": { "select": { "key": { "country": "Japan" } } },
      "value": { "text": "Japan, 82.6 years" } },
    { "op": "set-viewport", "axes": "y", "value": { "y": [70, 86] } }
  ]
}

country resolves because it is encoded on detail. The result returns applied with two resolved targets; a key with "Country" would come back unresolved, and the agent fixes the spelling before it describes the chart.

Code: two linked charts

Two bar charts share a Country field: GDP per country, and population per country. A click on one emphasizes the same country on both.

import { mountChart, clickHighlight, linkCharts } from 'flint-chart/interactive';
const mount = (el: HTMLElement, input: ChartAssemblyInput) =>
  ({ input, surface: mountChart(el, input, { backend: 'vegalite', interactions: [clickHighlight({ targets: ['mark'] })] }) });
const link = linkCharts({ gdp: mount(gdpEl, gdpInput), population: mount(popEl, popInput) }, { fields: ['Country'] });

The clicked chart drops its own emphasis once the link shows the condition, so both charts show one selection. A click-none, Escape, or double-click on either chart clears it on both; the link's own writes are host changes, so no chart echoes them back.

Bespoke: a rule over the brushed box

A line chart of Series by Month. The reader drags a box; the series whose every point in that month range lies inside the y range stay emphasized. No preset applies a rule over series, so the definition is the timebox above: rectangleTrigger('contain'), a plot affordance with the region cursor, a handler that acts on select-region at commit, and reset: ['click-none', 'escape'] so a click on empty plot clears it. It mounts beside navigate({ pan: false }) when the chart also zooms, since both ask for the plot drag otherwise.

Validation checklist

For an interaction_spec:

1. It parses as strict JSON and is a bare object with interactions.
2. Every entry answers a job from Step 1: a gesture the user named, a

question their goal raises, or something the chart invites.

1. Every type is in the chart type's list from list_chart_types, and every

option is under options with id on the entry.

1. groupBy and seriesBy are present where required, and name bound fields.
2. At most one entry owns the plot drag, the double-click, the legend click,

and the retained mark focus.

1. create_chart (or validate_chart, in code or CI) returns no warning.

For a ChartUpdate:

1. Every op and state is one from the op table.
2. Every selector key names encoded fields with values that exist in the rows.
3. The update is the whole layer for its id; clearing uses targets: [] with

state: 'normal'.

1. Ops that need a mounted preset have it, or unsupportedOps has been read

and acted on.

1. The result came back applied before you said what the chart shows.

For code:

1. Every trigger, cursor, and field name is one from this skill or the

installed flint-chart/interactive types.

1. Every definition declares affordances for the hits its trigger yields.
2. No two definitions, and no definition and spec entry, share a trigger or

an id.

1. The surface is destroyed on unmount, and every applyUpdate or dispatch

result is read.

Preset options

list_interaction_presets returns every preset with what it needs from the chart, its reset gestures, and every option it accepts under options, with each option's type and note; type narrows it to one preset and chartType to the presets a chart type supports. In code, the same options are the factory's parameter type in flint-chart/interactive. id is never an option: it sits on the entry. A ResetGesture is click-none, double-click, or escape. dimOpacity is the opacity of the muted marks, 0.25 by default. A guide is { visible?, style? }, or false to draw nothing while the gesture runs.