beffroi-tools/docs/interfaces/canvas-pilot-exchange.md

28 KiB
Raw Blame History

Canvas Pilot geometry exchange

Canvas Pilot was formerly named Midra Layout. The rename preserves existing wire-format identifiers, .midra.json projects, settings/recovery locations and MIDRA_VECTORCRAFT_CLI configuration. No schema version changes are needed. New exports use Canvas Pilot in human-readable titles; legacy files remain readable.

The Export menu writes SVG v2, draw.io v1, geometry JSON v1, geometry CSV v1, DXF/DWG, HTML/CSS v1, PPTX, Figma FIG / FigCraft and Illustrator AI (the latter via a local VectorCraft CLI). Project JSON is version 15 and is the full-fidelity save format. These exchanges preserve layer visible-box geometry, not sources, crop settings, zoom, pivots, Midra links, locks, visibility, camera or selection. Pictures illustrate current cropping; importing a picture does not reconstruct its crop or source settings.

The starter-file catalog centralizes ready-to-open two-layer templates under docs/templates/canvas-pilot/ for Full HD (1920 × 1080) and UHD 4K (3840 × 2160), in all eleven exchange formats (including PPTX, FIG and AI). The 4K geometry is an exact 2× scale of the Full HD layout. Regenerate them with:

cargo run --locked -p canvas-pilot --example check_geometry_exchange -- templates docs/templates/canvas-pilot
# Optional native codecs (default build); AI requires VectorCraft CLI:
MIDRA_VECTORCRAFT_CLI=/path/to/vectorcraft-cli cargo run --locked -p canvas-pilot --example check_geometry_exchange -- templates docs/templates/canvas-pilot pptx fig ai

The generator verifies imported geometry before writing each template and replaces existing templates in the explicitly supplied directory. Samples contain only two editable layer boxes, with no background or placeholder images. PPTX uses one slide, FIG one frame, and AI one artboard. Their canvas sizes are pixels (PPTX uses the documented 96 pixels/inch conversion). These files are geometry examples, not full-fidelity project saves. AI samples were generated and round-trip checked with VectorCraft CLI 0.24.1; Adobe Illustrator itself was not used for validation. Native fixtures are checked semantically rather than byte-for-byte because encoder metadata can vary. The AI regression check is opt-in and requires the real CLI:

MIDRA_VECTORCRAFT_CLI=/path/to/vectorcraft-cli cargo test --locked -p canvas-pilot --test exchange illustrator_readme_templates -- --ignored

Application and format matrix

Workflow Exchange format Import/export Background Placeholder pictures
Inkscape, Illustrator, Affinity Designer, Figma, web vector tools SVG Yes Embedded PNG Embedded PNG
diagrams.net / draw.io draw.io XML Yes Embedded PNG Embedded PNG
Scripts, automation and future adapters Geometry JSON Yes Embedded PNG Embedded PNG
Excel, LibreOffice Calc and data tools Geometry CSV Yes Geometry row No
AutoCAD, LibreCAD and QCAD DXF Yes Rectangle entity No
Browsers and web layout tools HTML/CSS Yes Embedded PNG Embedded PNG
PowerPoint, LibreOffice Impress, DeckCraft PPTX Yes; slide picker on import Embedded PNG Embedded PNG
Figma, FigCraft Native FIG Yes; page/frame picker on import Embedded PNG Embedded PNG
FigCraft Native FIGCRAFT Yes; page/frame picker on import Embedded PNG Embedded PNG
Illustrator, VectorCraft Native AI via VectorCraft CLI Yes; artboard picker on import Embedded PNG Embedded PNG

PSD, VSDX and ODG remain future adapters rather than implied support. PDF is a useful final visual but does not retain a reliable editable-rectangle contract for import, so it is not presented as a round-trip format.

Scope, coordinates and artwork

All formats offer Current layer (exactly one selected), Selected layers, Specific layer (including hidden), All layers and Visible layers only. Stable IDs are layer-N; order is document stacking order, independent of selection order. All chosen layers are exported visibly, even if hidden in the editor. No layers, unknown/duplicate IDs, invalid project geometry or dimensions rounding below one pixel reject export. X/Y are top-left output pixels, relative to the whole output screen, including for single-layer exports. X, Y, width and height round independently to nearest integer; halves round away from zero. Negative positions are retained. No stroke expansion is necessary: exported boxes have no outline.

Export layers… is one window: choose a format, artwork options and layer scope. All codecs share project validation and shape preparation; the byte-oriented interchange registry supplies formats, extensions and capabilities to the UI. Importing files uses the same mapping review as pasted text. Multi-page files require a page/frame/slide choice; changing it resets mappings. File parsing and conversion run on workers so the viewport remains responsive. Files are written only after encoding succeeds; existing destinations require replacement approval.

Options are shared for the current session, default off; they do not modify the project/settings schema:

  • Include background layer adds a first, full-screen shape with ID midra-background, explicitly marked as background. SVG, draw.io, JSON, HTML, PPTX, FIG and AI embed all visible references composited in project order over a dark backdrop using each image's fit/stretch/native-size, offset and opacity; otherwise they embed a plain dark backdrop. This is a baked visual reference, not an editable copy of reference settings. Image raster dimensions are proportional to the screen, capped at 1920 pixels on the longest side; the exported shape still has the exact screen dimensions. CSV and DXF carry only the background rectangle.
  • Include placeholder pictures uses embedded PNG source-grid patterns instead of flat layer fills. The pattern reflects native input cropping and advanced cropping, fitted into each visible box. Each raster is capped at 512 pixels on its longest side. No live source image is available/captured by this exporter. The images contain no external URLs and need no network access. Imported picture layers are still geometry-only. Without this option, the layers are rectangles. CSV and DXF disable this option because they cannot embed self-contained raster pictures in the supported profile.

The background is detected and ignored by default on import and cannot be mapped as a content layer. A legitimate full-screen layer is kept as a layer. Every placeholder is one importable layer, not an image plus duplicate rectangle.

Optional native document codecs and licensing builds

The default vendor_formats Cargo feature uses the published, pinned deckcraft-pptx 0.6.1 and figcraft-format 0.5.1 crates with their document/geometry types. Their licenses are MIT OR Apache-2.0; vendor file formats do not imply a proprietary implementation license. No upstream app UI, font bundle, media engine or online service is embedded. VectorCraft's AI readers are unpublished and pull in rendering, font and PDF stacks, so AI uses a separate optional executable rather than embedding that stack. Upstream licenses still apply to separately distributed tools; Midra does not bundle VectorCraft.

# Normal build: native PPTX/FIG/DWG codecs and the optional AI converter interface.
cargo build --locked --release -p canvas-pilot
# Excludes vendor codecs, their optional dependency trees and the AI interface.
cargo build --locked --release -p canvas-pilot --no-default-features --features no_proprietary

no_proprietary takes precedence in code and UI if combined with other features. Cargo features are additive: the --no-default-features part is required to remove the dependencies themselves. An --all-features build disables the adapters but still compiles optional dependencies. CI tests both configurations; production coverage uses defaults so native-codec code remains included in the denominator. This is a format/dependency policy switch, not a legal certification.

PPTX

Standard OOXML ZIP, one slide matching the screen; 96 screen pixels per inch, or 0.75 points per pixel (9525 EMU/px). Native shapes/pictures preserve editable names and boxes. The background is named midra-background, a reserved marker excluded from mapping. PNGs are embedded as media; no source URLs are used. PPTX/FIG/FIGCRAFT export asks to rename a content layer with that reserved name so it cannot silently become an ignored background on reimport. Stable Midra device/source IDs are not represented in the native file.

Import lists each readable slide independently. Shapes and pictures use their explicit transforms, recursively composing group coordinate transforms. Rotation and flips produce axis-aligned enclosing boxes. Hidden, zero-sized, unpositioned and unsupported opaque/connector objects are omitted with a count in the review. Masters, inherited placeholders, text content, animations, media playback and source/crop semantics are not imported. The note reports slide dimensions and the fixed unit convention; importing never silently resizes the Midra output.

Figma FIG

User-validated import/export on 2026-11-15.

FigCraft writes a native .fig ZIP containing the Kiwi-encoded canvas and embedded image assets. Export creates a Canvas Pilot frame with screen dimensions and rectangle children (image fills when requested). Coordinates are Figma pixels. The background uses the same reserved midra-background name. The upstream codec is best effort; warnings on export are errors rather than silent losses.

Import offers top-level frames/components as separate canvases, relative to their own origin, and pages with loose objects. Visible leaf objects are mapped by their transformed enclosing boxes; nested containers are traversed without duplicating their bounding rectangle. Hidden children and zero-sized objects are omitted. Upstream approximation warnings appear in the mapping review. Constraints, components, auto-layout rules, text and styling do not become Midra parameters. Native round-trip tests verify geometry/artwork with the pinned upstream codec; they do not constitute validation against every Figma service version.

FigCraft FIGCRAFT

User-validated import/export on 2026-11-15.

The separate .figcraft option uses the pinned figcraft-format crate's native version 1 reader/writer, not a renamed .fig. Its ZIP contains document.json (with format: "figcraft" and version: 1) and image assets under images/. Import also accepts the native bare JSON representation with a .figcraft extension; unsupported future versions are rejected by the upstream reader.

Geometry, page/frame selection, reserved background names and artwork follow the FIG contract above. Export rounds positions and sizes to whole screen pixels; import retains fractional coordinates. One output-sized frame contains the named layer rectangles. Background and placeholder pictures are optional. Midra source, crop, mapping and workspace settings are omitted; extra-parameter embedding is not enabled. FigCraft history, branches and library data are not imported into Midra or recreated on export. Reader repair warnings appear in the import review. This option shares the FIG vendor feature and is excluded by no_proprietary. Tests cover native resaving, bare JSON, geometry, artwork and both starter sizes.

Illustrator AI

Optional compatibility target: prefer SVG for Illustrator exchange. AI import and export require a separately installed VectorCraft CLI; it remains external and is not embedded or bundled. See the manual setup instructions.

Uses VectorCraft CLI, discovered through MIDRA_VECTORCRAFT_CLI, beside the Midra executable, in its bin/ subdirectory, then on PATH. Conversion runs locally in safe mode in a temporary directory, with a 60-second timeout and 16 MiB result limit. Nothing is installed automatically. The CLI reads AI and converts to SVG for the shared geometry importer, or writes PDF-compatible AI from Midra SVG. Import enumerates up to 64 artboards and presents each separately in the same canvas selector. Converter import warnings are shown in the mapping review; export warnings reject the result instead of silently losing content. SVG's rectangle/path/image restrictions still apply; this is layer-box exchange, not arbitrary illustration reconstruction.

Manual native-file generation and round-trip diagnostics (no device access):

cargo run --locked -p canvas-pilot --example check_geometry_exchange -- native /tmp/midra-native pptx fig
MIDRA_VECTORCRAFT_CLI=/path/to/vectorcraft-cli cargo run --locked -p canvas-pilot --example check_geometry_exchange -- native /tmp/midra-ai ai

These generate plain and illustrated documents, with backgrounds, and verify the imported layer boxes. Existing diagnostic files are not overwritten. VectorCraft 0.24.1 native AI round trips were checked for both variants. DeckCraft 0.8.0 rendered the exported PPTX independently of the pinned 0.6.1 codec; the resulting Full HD image was visually inspected.

SVG v2

Example: canvas-pilot-svg-v2.example.svg. UTF-8 SVG; SVG namespace and optional XLink namespace. Root attributes:

  • data-midra-layout-geometry="2"
  • width, height: output screen dimensions in pixels
  • viewBox="0 0 W H"

Each layer is a <rect> (plain) or <image> (placeholder), with id="layer-N", data-midra-role="layer", x, y, width, height. A <title> carries its name for external viewers; Midra import does not rename destinations. Images use xlink:href="data:image/png;base64,..." and preserveAspectRatio="none". The optional background is an <image> with data-midra-role="background" and full-page bounds. No visible text geometry, transforms, clipping or CSS is needed.

The importer supports v1 and v2 and explicitly rejects other marked versions. VectorCraft-derived SVG may omit the root version and retain data-midra-role on wrapping groups. The importer inherits these role markers and their stable IDs; marked image layers remain importable and backgrounds remain excluded. Explicit unsupported root versions are still refused. Legacy Project::to_svg/to_svg_ids continue producing v1 rectangle-only SVG; the GUI uses v2. Old applications cannot import v2 placeholder images; use JSON or a plain rectangle-only export where required. Editors should retain role markers. If they remove them, ordinary image import remains intentionally limited to a first full-page background. Midra never fetches image URLs during import.

draw.io v1

Example: canvas-pilot-drawio-v1.example.drawio. UTF-8 uncompressed <mxfile compressed="false"> with one <diagram> containing an <mxGraphModel data-midra-layout-geometry="1">. pageWidth, pageHeight, pageScale="1", and page="1" establish the output canvas; diagram units are pixels. Root cells 0 and 1 provide the required graph root and default layer.

Each exported shape is an <mxCell vertex="1" parent="1"> with stable id, escaped plain-text value (layer name), and data-midra-role="layer" or "background". Its single <mxGeometry as="geometry"> stores pixel X/Y/width/ height. Styles disable outlines and rounding. Pictures use shape=image, imageAspect=0, and draw.io's image=data:image/png,BASE64 style syntax (without ;base64, which would split the style). Background label text is empty.

Import accepts raw/once-percent-encoded mxGraphModel clipboard XML, single-page mxfile XML, and native compressed diagram text: Base64 → raw DEFLATE → percent decode → mxGraphModel. Native re-saving may remove the model's Midra version attribute; role attributes and stable IDs survive in the tested version. The reserved midra-background ID also identifies the exported background. Unknown marked draw.io exchange versions are rejected. Multiple pages are refused with an instruction to export one page, rather than silently importing the wrong page. Groups, edges, rotation, flips, relative/alternate bounds and nonrectangular shapes remain unsupported; ordinary unrotated image cells now import their box.

Geometry JSON v1

UTF-8 JSON with root fields format: "beffroi.midra-geometry", version: 1, units: "px", canvas: {width,height} and layers. Each layer has id, name, role, x, y, width and height; image_png_base64 is optional. Roles are layer or background. Unknown fields, versions and units are rejected. IDs must be unique and at most 255 characters; names are limited to 1024 characters without controls. This JSON is the preferred small machine interface. It is not the version-10 project save file and does not contain sources, crop or device links.

Geometry CSV v1

RFC 4180 CSV with CRLF records and this fixed header:

midra_geometry_version,canvas_width,canvas_height,id,name,role,x,y,width,height

Every row repeats version 1 and the same canvas so a selected row remains self-describing. Names use normal CSV quoting. Roles and geometry have the same meaning as JSON. Empty files, mismatched canvases, extra/missing fields and non-integer geometry are rejected. CSV carries background geometry but no image data or placeholder pictures.

CAD: DXF and DWG profiles

Both DXF and DWG import/export were user-validated on 2026-11-15.

DXF uses the official shared craft-dxf document model and serializer, pinned to Craft libraries revision 79e97b6febcb62de406e5bcf5a724bd3f7333065 (MIT OR Apache-2.0). It writes UTF-8 ASCII DXF AC1021 (AutoCAD 2007), with complete symbol tables, model/paper-space block records, dictionaries, layouts, unique handles, owner references and subclass markers. Handles reserve the low 0x100 range to avoid QCAD stock-object collisions; Craft remaps definitions and pointers together. This replaces the incomplete hand-written DXF skeleton rejected by QCAD. The rectangle/coordinate contract remains unchanged; older Midra DXF files are still readable.

Drawings are unitless ($INSUNITS=0): one CAD unit equals one screen pixel. Whole-pixel $LIMMIN/$LIMMAX describe the canvas, with matching initial extents. Every stable ID has a layer-table entry; every box is one closed four-vertex LWPOLYLINE. CAD Y increases upward, so cad_y = canvas_height - screen_y. Import reverses this conversion and requires explicit canvas limits (falling back to extents), rather than guessing scale or page origin. The special midra-background layer is excluded from destination mapping. Human names are layer descriptions; imported IDs remain CAD layer names. Pictures are omitted. DXF import uses Craft's bounded ASCII/binary tag reader and retains the existing axis-aligned rectangle restrictions.

DWG uses the optional craft-dwg byte bridge from the same pinned revision, backed by unmodified acadrust 0.6.6 (MPL-2.0). It converts the same DXF model to AutoCAD 2007 DWG and back; canvas limits, IDs, pixel rounding and background rules are identical. A writer report of omitted records rejects export. Reader notes appear in the import review. Both DWG input and expanded DXF are limited to 16 MiB. DWG is available with vendor_formats, excluded by no_proprietary; DXF remains available in all builds. The bridge reports unsupported platforms explicitly (WebAssembly and FreeBSD ARM64 at this pin).

Geometry exchange is not a general CAD editor: arbitrary drawing entities, blocks and drafting annotations are not a supported round-trip profile. Keep original CAD drawings; import/export the rectangular layer layout separately.

HTML/CSS v1

The XHTML-compatible document is a directly viewable page. Root data-midra-layout-geometry="1" marks the version; #midra-canvas stores data-width/data-height. Each <div> or <img> stores authoritative data-midra-role, data-name, data-x, data-y, data-width and data-height; inline CSS renders the same box. Images use self-contained PNG data URLs. Import reads the data attributes, not computed CSS, scripts or browser layout. The file contains no script or external resource.

Limits, import and round trips

Operator acceptance (2026-11-14): the user confirmed PPTX working flawlessly, and draw.io and SVG working. Their reported format issues are marked solved in the task list.

Import files and clipboard text, and generated exports, are capped at 16 MiB (to allow self-contained images). Decompression is bounded: at most 48 MiB of percent-encoded text and 16 MiB decoded XML. Up to 256 layer boxes plus one background are allowed. No imported image is decoded or downloaded by the geometry importer. Export reference decoding retains the existing 8192-pixel per-side/256 MiB decoder bounds. Oversized exports fail before writing a file.

Use Import → Geometry file…, drop supported files, or paste supported text, then review and map shapes to editor layers. Drops route case-insensitive registered extensions (including draw.io XML and vendor formats) through the same asynchronous byte importer as the picker; native-codec build restrictions still apply. Path-backed and in-memory file drops are supported. Raster images retain reference-image behavior. Several geometry files become named choices in the file/page/frame/slide selector, with a combined 16 MiB input limit. Parsing is atomic across the geometry batch; errors report the failed filename, and no geometry is applied before mapping confirmation. Destination count need not match current selection. The review offers All / Selected / Folder destination scopes without changing canvas selection; hidden and locked layers are unavailable. Every included item needs a distinct destination, while excluded items and unmapped editor layers remain unchanged. Existing mappings survive scope changes when still eligible. Clipboard text and per-item geometry details are collapsed, read-only previews. These review choices are session-only and do not change any file schema. Import with size and position enabled restores the rounded geometry in one Undo step. Output screen resolution is not imported. Mapped names, sources and Midra links stay with the destination layers. File replacement retains the existing explicit overwrite confirmation and atomic-write path.

SVG files rendered by draw.io Desktop are also accepted. The importer recognizes the data-cell-id wrapper groups retained by draw.io, reads the first rectangle, image or symbol use in each cell, and ignores generated HTML labels, raster text fallbacks, symbols and the draw.io help link. The standard SVG 1.1 public DOCTYPE emitted by draw.io is accepted without fetching external content; XML parsing is limited to 200,000 nodes. The reserved midra-background cell remains an ignored background.

Draw.io's SVG export normally crops the canvas to the diagram content, adds one pixel to its bounds and rebases cell coordinates. Those rebased coordinates are the only positions present in that SVG. Include Canvas Pilot's background layer before exporting through draw.io when the full output-screen origin must survive the round trip.

Tests cover each format/artwork combination, the twelve text templates and eight native PPTX/FIG/FIGCRAFT/DWG templates (plus two AI templates in the opt-in CLI check), full-screen layers vs backgrounds, negative/fractional coordinates, repeat round trips, compressed clipboard/files, embedded references, invalid IDs/versions and no-overwrite file handling. A manual native-editor check is available:

cargo run --locked -p canvas-pilot --example check_geometry_exchange -- generate /tmp/midra-exchange-check
drawio -x --size page -o /tmp/midra-exchange-check/render.png /tmp/midra-exchange-check/illustrated.drawio
drawio -x -o /tmp/midra-exchange-check/resaved.xml /tmp/midra-exchange-check/illustrated.drawio
drawio -x -f svg -o /tmp/midra-exchange-check/drawio-rendered.svg /tmp/midra-exchange-check/illustrated.drawio
cargo run --locked -p canvas-pilot --example check_geometry_exchange -- verify /tmp/midra-exchange-check/resaved.xml
cargo run --locked -p canvas-pilot --example check_geometry_exchange -- verify /tmp/midra-exchange-check/drawio-rendered.svg

On 2026-11-14, installed draw.io Desktop 31.7.0 (/usr/bin/drawio, not Flatpak) rendered the illustrated diagram correctly, and its re-saved XML returned both layer boxes unchanged. Its SVG export, including <symbol>/<use> image cells and HTML/raster label fallbacks, also returned both layer boxes unchanged when the background preserved the page. Generated plain/illustrated SVG and draw.io files passed geometry verification. Native output artifacts remain in /tmp, not Git.

Protocol references: draw.io diagram format and desktop CLI, SVG 2, CSV RFC 4180, and Autodesk's DXF group-code reference.

Multiple references are flattened into the same single background image in artwork-capable exports; no format version changes. Reimport reconstructs that background as a geometry/artwork item, not the editable reference collection. Use project JSON v15 to preserve independent images and visibility.

QCAD interoperability check

Validated with QCAD Professional 3.28.2: both 1080p and 4K DXF/DWG templates render two correctly placed rectangles without import/handle warnings. Saving both Full HD samples through QCAD as R21 (2007) and reimporting preserves both layer boxes. Example manual check (temporary output only):

QT_QPA_PLATFORM=offscreen /path/to/qcad/dwg2dwg -r R21 -o /tmp/layout.dwg docs/templates/canvas-pilot/canvas-pilot-1080p-2-layers.dwg
cargo run --locked -p canvas-pilot --example check_geometry_exchange -- verify /tmp/layout.dwg

The verifier now uses the unified byte importer for text and binary formats. The .ai starter ZIP is only a download wrapper, not a new import format; its two entries must remain byte-identical to the .ai fixtures (Python regression).

Optional extra Midra parameters

Export layers offers Include extra Midra parameters (experimental) for SVG, draw.io, JSON, CSV, DXF, HTML and PPTX. This embeds an inert snapshot separately from the artwork. JSON/CSV use geometry version 2 only when enabled; plain exports remain version 1. Other carrier versions remain unchanged. Existing omissions in this document describe plain geometry export; the optional snapshot preserves source/crop, identity, links and workspace settings, excluding image bytes and capture bindings. See the complete carrier contract and format audit.

Restore the original snapshot with Project → Restore parameters from export…. Normal geometry import continues to use the artwork's current boxes, including any external edits. Third-party editors may strip metadata. DWG, FIG, FIGCRAFT and AI are not enabled until their current codec paths can preserve it reliably.

Embedded project snapshots now contain project version 15 and family-aware device_link objects. Geometry representation versions are unchanged. Existing embedded versions 1–13 migrate using the project contract; plain geometry continues to omit device destinations.

Project v14 additionally preserves layer_fit in embedded snapshots. Geometry alone describes the layer container when explicit fit is enabled and omits crop and fit behavior. Optional placeholder artwork renders the fitted image clipped to the container, including transparent margins for Centered or small 1:1 content.

Project v15 also preserves normalized layer masks in embedded snapshots. Placeholder artwork blanks pixels transparently without resampling. Plain rectangle geometry continues to omit mask appearance; use project JSON for exact round trips.