Roadmap

Where ChartCraft goes next. Items are grouped by theme, not strictly ordered; the API contract governs every addition — each of these lands as a contract change first, then core, then wrappers, then docs (see CONTRIBUTING.md in the repository root).

Shipped in 0.4

Dogfooding, not more tests. 0.3 shipped with 1,984 passing tests, and then five real sample dashboards were built against the published packages. They found four defects those tests could not: every one was an experience bug — the API accepted the options, returned no error, and drew something wrong or lost user state. All four are fixed, with 39 new tests pinning the behaviour.

• A log value axis derives its domain from the positive data only. A boxplot of all-positive data (1.2 … 260) with yAxis: { type: 'log' } used to draw a 1e-12 … 1e3 axis with every box squashed into the top tenth of the plot. Zero anchoring is now skipped, rounding is to whole decades, a stacked series contributes its cumulative tops rather than its zero floor, and a non-positive bound from any source — min/max, a chart type's own widening, a decorator, a zoom-viewport edge — is discarded rather than clamped to an epsilon. A value ≤ 0 becomes a null gap with one console.warn, never an error: a live dashboard must not go blank when a linear axis is flipped to log. Applies to a log x axis and to a horizontal chart's value axis too.
• An options update no longer silently destroys the zoom viewport. The old rule reset the window whenever the payload merely contained data — and since every framework wrapper re-sends the whole options object, a pure theme change threw away the user's brush and emitted no event. The discriminator is now the computed domains, per axis, so new values on the same timestamps keep an x-zoom. A reset always emits zoom: null.
• PointEvent carries the colour of the mark that was clicked.PointEvent.color resolves through the same path as TooltipPoint.color, so a click-driven detail panel's swatch can never disagree with the chart — and no one has to re-derive it from a series index, which is wrong whenever a series is hidden or a type assigns slots per slice. It is required, which is why 0.4 is a minor: constructing a PointEvent (a test fixture, a mock) now needs the field.
• theme.warning, and gauge bands that need no hardcoded hex.up/down covered two of the three states a status mark has; the middle one forced every consumer to hardcode a colour. theme.warning (#fab219, both schemes) is that step, and it is the one optional slot on Theme so every hand-written custom theme still compiles — and still gets themed. gauge.bands[].color is now optional, filled by position.

Wrapper DX, from the same dogfooding.

• One package, not two. All four wrappers re-export core's complete public value surface as well as its types, so @chartcraft/core is no longer a second direct dependency. Named re-exports only, verified to tree-shake.
• ChartSpec (Omit<ChartOptions, 'type'>) is exported under that one name from all four, for the "options in their own module" pattern every sample uses. TypedChartOptions survives as a deprecated alias in Vue and Angular.
• React: memoising option props is a correctness requirement, not an optimisation — there is now a development-only warning when the wrapper sees a new-but-deeply-equal option prop three renders in a row, and ref is populated before any parent effect runs.
• Angular: a (ready) output and whenReady(), removing the two-level this.hero()?.chart()?.… nullability dance, plus the same dev-mode warning.
• Svelte: all six events work as Svelte 5 callback props (onpointclick, onzoom, …) as well as on: directives, on Svelte 4 and 5 — so a Svelte 5 app can drop the deprecated directive entirely. Plus onready/on:ready.
• Vue: no dev warning by design (it deep-watches, so a mutation is never silently lost), but the guidance to hold options in a ref/computed is now documented.

Still open, and stated honestly in all four framework guides: the per-type components share one loose options type, so <GaugeChart> accepts a sankey block. They buy the correct type string, not a narrowed options shape. Narrowing was assessed and deliberately deferred — it would break the shared ChartSpec pattern, and it is a 1.0-shaped change.

Shipped in 0.3

Twenty new chart types, for 39 total: range area, bullet, dumbbell, lollipop, slope, streamgraph, marimekko, pyramid, calendar, radial bar, rose, violin, parallel coordinates, icicle, circle packing, word cloud, sankey, gantt, choropleth and network. Every one ships with the full shared feature set — tooltip, legend policy, keyboard navigation, a data table, theming, animation, reduced motion, resize — and its own example page with an honest "when not to use this" section.

Audit-driven hardening. An adversarial quality audit of all 39 types drove a round of fixes that shipped in 0.3: the accessible table and exportData() now carry the full series rather than the downsampled render set; sequential ramps are scheme-aware, so the highest-magnitude cell clears 13.16:1 on dark instead of vanishing at 1.46:1; candlesticks encode rise/fall with fill as well as colour; forced-colors: active is genuinely implemented; series past palette slot 8 gain a dash/marker channel; network links became reachable by assistive tech; and zoomTo at 1M points went 76.8 ms → 9.9 ms while a 1M-point mount went 4.70 s → 1.21 s.

Six cross-cutting features, each with a page of its own:

• Error bars — the v0.2 roadmap's one open chart-type item, now a series decoration on line/area/bar/scatter/bubble, included in the value domain, the tooltip and the data table.
• Trendlines — least squares, centered moving average, exponential; dashed and legend-labeled so a fit can never read as data.
• Data labels — with measured selectivity: 'auto' drops labels that would collide.
• Annotations — reference lines, bands, labeled points, free text, plus an annotationclick event.
• Zoom, pan & brush — and downsampling that re-runs inside the visible window, so zooming into a million points reveals real detail.
• Export — exportImage() (PNG at any scale) and exportData() (CSV/JSON that mirrors the accessibility table exactly).

A public plugin surface. The previous roadmap's "plugin system" shipped as the decorator API: the five features above are implemented on it with no special treatment from the pipeline. It is exported and documented but marked experimental — four of its hooks were added during v0.3 in response to real feature needs, so the shape may still move in a minor release.

Also new: SeriesData/GraphData (graph payloads typecheck without a cast), DataPoint.value (so a real TreeNode[] needs no cast either), the zoom and annotationclick events, and three new Chart methods.

Shipped in 0.2

Thirteen new types plus combo (per-series mark mixing), for 19 at the time: bubble, sparkline, histogram, boxplot, candlestick, OHLC, waterfall, heatmap, treemap, sunburst, funnel, radar, and gauge — each with palette slots by identity, a data-table representation, and full keyboard navigation. The financial types got their volume-pane story as a small multiple, not a dual axis.

Renderers

The Renderer interface exists precisely so these slot in without touching chart code — and v0.3 added a reason to want them:

• SVG renderer — for print/export pipelines and style-by-CSS use cases (design systems that must restyle marks with stylesheets). Same visual spec; one retained node per mark, so intended for small/medium data. It is also what exportImage({ format: 'svg' }) needs: today that call rejects with "SVG renderer not available".
• WebGL renderer — for the 1M+ point regime where Canvas 2D fill rate becomes the ceiling. Targets scatter and line first. Automatic renderer selection by data size is a possible follow-on once both exist.

Frameworks

• Angular wrapper (@chartcraft/angular) — shipped in v0.3. Standalone, signal-based, zoneless-ready, all 39 chart types. See the Angular guide.
• Solid wrapper (@chartcraft/solid) — same thin-wrapper contract as React/Vue/Svelte/Angular: lifecycle, resize observation, event bridging, nothing else. Feature parity comes from core, as always.

Server-side & export

• SSR snapshot rendering — render a chart to markup/image on the server so the HTML payload contains real chart content (today, wrappers are SSR-safe but the chart appears at hydration). Also unlocks static-site and email use.
• Vector export — exportImage({ format: 'svg' }) and a PDF path, both built on the SVG renderer above, for vector-quality print output.

Interaction

• Parallel-coordinates brushing — filtering lines by dragging on an axis. The seam is already pure and documented (a decorator can map a pointer x to a dimension and a pixel back to that axis's value); the decorator itself is not written yet.
• Category-axis zoom — the viewport applies to continuous axes only, because windowing a band scale would desynchronize band indices from categories. A band-aware viewport is possible but needs a contract decision first.

Data

• Streaming data API — an append-oriented path (appendData-style) cheaper than setData for live feeds: no re-normalization of unchanged history, windowing, and downsample-aware appends.

Extensibility

• Stabilizing the decorator API — promoting it out of experimental: freezing the hook set, deciding whether decorators can be scoped per chart instead of per build, and giving LegendItem / A11yTableSpec public type exports so a decorator's hooks are fully typed from the package root.

Quality infrastructure

• Visual regression harness — pixel-diff snapshots of every example and mark-spec fixture across themes and DPRs, wired into CI, so the visual quality bar in the contract is enforced by machines rather than reviewers' eyes. The v0.3 determinism rules (no Math.random() in any layout) exist partly to make this possible.

Non-goals

For clarity, some things stay out regardless of demand:

• Dual y-axes — two measures of different scale get two charts or an indexed common base (why). This includes Pareto charts: a cumulative line belongs on the same normalized scale as its bars, or in a small multiple.
• Runtime dependencies in core — zero now, zero later. That is why no map topology is bundled: choropleth.geojson is always yours.
• Unvalidated palette growth — no 9th categorical slot, ever. Beyond 8 the hue order is reused with a composite encoding (dash pattern + marker shape) and a one-time console recommendation; the real fix is a design change on the caller's side (fold the tail into "Other", or small multiples). The library will not generate a colour, and it will not fold your data for you.
• Radius-linear encodings — rose radius, bubble and network node radius are always √value. Area-true or not at all.
• A project planner — the gantt type draws spans. Dependencies, critical paths and resource levelling belong to a planning tool, not a chart library.