Repository navigation
day 1 fix - #150
Merged
Merged
day 1 fix#150
Conversation
… to presets The chart-author skill hands behaviour to the new skill, which reads the chart spec and the request and returns one interaction_spec from presets. The MCP server ships the skill as a resource and a prompt. The Vega-Lite reference drops its Interactions line per chart type, since the skill is where a chart's presets are read. The skill text and its tests are trimmed to the request and the spec.
…he MCP chart view hands it to the model surface.onSelection calls back with a ChartSelection on every committed gesture except a viewport move: the rows under the marks, from the traced records or else the mark values, and the brushed range on the brushed axis, named by its field. The dashboard lab subscribes through it instead of reading the DOM event. The MCP chart view sends each selection to the host with updateModelContext, as a text table and as structured content, and sends a cleared note when the selection is cleared.
The surface root sizes itself to the compiled chart, so the frame's max-width rule never reached the SVG and a chart with a legend, or with its own size, scrolled. The app clamps the root to the frame, and the SVG scales down through its viewBox as it does in the static path. The surface adds the rail row gap only when it appends a rail, which removes the few px of vertical overflow.
- Readings: overflow placeholders, number precision, ECDF ties, full dates on time axes, untitled legends, stroke-dash series names, aggregate legend fields, per-panel mark grouping, unknown vs. empty membership. - Controller: single tab stop with escape-reset presets, in-place focus on tab-in, activation announces the updated state. - Rebuild the tree only when the scene signature changes. - keyboardTargeting is ignored while accessible navigation is mounted. - Radar: unique layer names per facet panel (panels drew the last panel's vertices). - Add docs/design-accessible-navigation.md and regression tests. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
…n context The context a preset reads now extends a ChartState a host can read: the emphasized marks per update, the hidden legend values, the shown viewport, and the marks inside the plot area. The scene is scanned once for both available and visible.
…of onSelection ChartState is what the chart shows: the emphasized marks with their per-update entries, the hidden legend values, the viewport, and the rail windows. The runtime reports every change after the render with the state it produced; a click preset's hover cue reports as a preview, and a chart without navigation reports no viewport. A host reads each mark's value and the gesture's geometry and queries its own data, so there is no row helper and no scene scan for hosts.
…he site tabs read it The MCP view replaces the model context on each commit with a sentence: which marks are emphasized and the range or the values they share, the hidden values, and the zoomed axes. The chart-to-external tab reads onChange, follows a pan and a zoom, and queries its own rows inside the viewport; the dashboard routes commit changes by their interaction id. The map stages stop the page overscroll.
The design doc, the interactive README and the architecture illustration show the read side: the chart state leaves the update processor for the application, in the same semantic terms the gestures enter with.
… ChartUpdate, surface code, and bespoke interactions One skill for every output around a chart's behaviour: how an interaction works, presets as one interaction_spec, the ChartUpdate an application or an agent applies to a mounted chart, reading the state, linking charts, and a bespoke definition from a trigger and a handler. The intent groups and the scenario table give way to a chart-shape list. The preset option table is generated by gen:reference from the factory option types, whose JSDoc now carries the defaults. The MCP instructions, resource, and prompt name the wider scope, and the skill copies are synced.
…ry, and the skill The programming guide and the gallery's app section read the chart through getState and onChange, with the flint-interaction event kept for the gesture itself. The agent pointers name the skill's full scope, and the skill's event shape carries the description accessible-navigation sets on focus-element.
Chart state for hosts, and one interaction-author skill
An agent with no handle on the mounted chart passes ChartUpdate layers beside the input; the view applies them at mount. validateChartUpdates checks the list shape, the surface warns about a layer the chart cannot honour, and an overlay that does not project is reported as unsupported instead of crashing the chart.
The board strips furniture from its cells and draws it once on its frame, but matched only the earlier kind names, so the Economist tab, footer rules, and text or image furniture were never drawn. planFrameFurniture plans the frame with the chart's own header and closing-rule rules, and headerFurnitureItems, shared with applyHeaderFurniture, places the pieces; the board and its PNG export draw them.
A cell no longer lifts the chart's title and deck into HTML or strips the theme's furniture: the chart draws its own heading, deck, tab and rules, as it does on its own card. The chart/table switch and the edit handle float over the cell; the table view shows the table alone. Plots in a row no longer share a top line when headings wrap differently. A board shown in a pane (fullscreen or picture-in-picture) scrolls, with a stable gutter; inline it stays sized to the board.
…ts and keeps each chart's furniture The board export loaded each chart's SVG from a blob URL, which an MCP App sandbox's img-src blocks, so the board could not be exported in hosts such as Claude Code; it now uses a data URL. A chart's furniture is a 1px overlay layer, which the board export drew at 1px; it is drawn at the chart's size. The chart card painted its furniture on a canvas Vega had left scaled and offset, so the tab came out at the scale squared; the transform is reset before painting.
…art Step 4 sections
The bar and box templates band a temporal x such as a run of years, but the Grouped Bar Chart's local dodge picked its band axis from the semantic type, which stays temporal. It grouped the lane window by the y field, so bars sharing a value shared a lane group and the rest collapsed to zero width. The axis now comes from the offset channel the assembler chose from the banded types. The dodge option's check, in the Grouped Bar Chart and the Boxplot, made the same choice and offered or hid the control from the y field. It now asks for the band axis as declareLayoutMode does.
Bump flint-chart and flint-chart-mcp to 1.0.1 (flint-chart-mcp depends on flint-chart ^1.0.1), with the Copilot plugin manifest and MCP pin synced, and date the 1.0.1 changelog entry. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Every tool builds the chart or board again from the spec it is given, so the server runs stateless. validate_chart and the new validate_board check a spec and return the report without drawing; show takes the spec as chart or board, draws it, and returns the same report, or draws nothing for a spec with errors. create_chart, create_board, add_chart, update_chart, remove_chart, update_board and the card-only restore_board, set_board_theme and set_board_layout are removed, with the BoardStore and its versions. The card keeps the reader's theme, arrangement and panel edits itself and says in its context which fields keep them in the next show (theme_spec, grid, chart_spec); a board spec takes grid, fitted to its charts. A card the host draws again without its result calls show with its tool input. The chart, board and interaction skills follow the new loop.
A section matches its heading exactly, or else as the start of one heading only, ignoring case and punctuation, so "Link charts" or "step 4 write the spec" find their section and an ambiguous start still lists the headings. The interaction skill's "Link charts, and drive a chart over time" is now "Link charts", with the playback note inside it and the paragraph on what the surface holds fixed moved to "Use Flint in an application". The Copilot bundle is rebuilt with the locked MCP SDK (1.32.1).
…les into it The frame grows to the chart up to the canvas ceiling and hides overflow, so a scrollbar no longer narrows the chart into a relayout loop or cuts off its title. A chart laid out taller than the ceiling is given it as its box and scaled down; in fullscreen or pip the chart is laid out for the pane. The page hides its scrollbar inline as the board does, safe-area insets add to the card's padding instead of replacing it, and the not-drawn message is shorter.
xavier (xavier-shaw)
approved these changes
Oct 21, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.