Images carried inside a `.fig` message (`image.dataBlob` indexing the
message's `blobs`) were copied out before the message is dropped, one copy
per image hash. Nothing stops a file from pointing many hashes at one
blob: 100 hashes on a 250 MB blob made 25 GB of copies, an out-of-memory
abort from a small file.
The walk now records each hash's blob index, and `embedded_copies` makes
one shared copy per blob for every hash that names it, so the copies are
bounded by the message's own size. Each resolved image is still copied
once on its way into the document, where images are kept by content hash.
Regression: fifty hashes on one blob share one copy; a hash naming a
missing blob is left out.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LvG6Dyzof21BCrUsByhx2w
Image paints that carry a current filter container now import their exposure, contrast, vibrance (as saturation), temperature, tint, highlights and shadows as native image filters. Neutral values leave the paint untouched. The older color-adjust container alone, both containers together, filters that have no native equivalent (brightness, detail, vignette and unknown fields) and malformed or non-finite values are skipped and reported as import notes. A positive tint in .fig data moves toward magenta, so it is flipped once at the .fig boundary; public REST image filters keep their sign and are not reported as approximated. Imported filters use the approximate native renderer, which the import says once.
Figma stores where a numbered list starts counting on each text line
(listStartOffset; a list shown as "2." has offset 1). The import dropped
it, so every list counted from 1.
A paragraph's list settings now carry a start number (written only when
set, so older files open and save unchanged). Numbering counts on from
it where a list starts; items that go on counting ignore it. The .fig
import reads it, the .fig export writes it on a list's first line, and
plain paragraphs drop it. Toggling a list off ignores differing starts.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Copies written by plugins and tools carry their image bytes in the
message: a paint's image.dataBlob indexes the message's blobs. A bare
canvas (what the clipboard holds, see #647) has no images/, so these
fills came in as grey placeholders. The importer now falls back to the
embedded bytes, keyed by image hash, after the archive's images/.
Blobs over the image size cap and indexes past the blobs are skipped and
stay placeholders. Tests: a bare canvas carrying its image in a blob, and
a dataBlob past the blobs.
A real 705 MB .fig file (401 644 nodes, 894 004 layers once instances are
expanded) didn't open: "message: too many values". Lifting that cap alone
made it import only its first 200 000 layers, and lifting that one took
41 GB of memory.
The global caps stay as they are. The .fig import gets budgets from the
size of its compressed canvas instead (counted up to the file's own size),
not from the message it unpacks to, since chunks inflate up to 512 MB:
- the kiwi value budget is one value per 2 bytes of the canvas, at least
the old 20 million (kiwi is only read for .fig files);
- the layer cap is one layer per 32 bytes of the canvas, at least the REST
import's 200 000 and at most 2 000 000. A small or crafted file can't
expand its instances into millions of layers, however far it unpacks; the
REST JSON import keeps 200 000.
Memory:
- The mapper no longer passes the document through `json!`, which serialises
its arguments: the whole tree was deep-copied twice.
- Vector networks are built once per distinct network and handed to the REST
importer beside the JSON (`vectorNetworkRef`), not written into it as JSON
objects for every instance that shows them.
- Decoded kiwi objects are trimmed, and each import stage's input is dropped
as soon as the next stage is built.
On that file: peak footprint 41 GB -> 11.3 GB, 67 s -> 25 s.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YQxjkNSUJhTkTfY74aNeki
Grid columns and rows keep fixed pixel sizes, Hug sizing and fractional Fill weights when the minimum and maximum sizing definitions agree. Track GUIDs join the sizing records to the ordered tracks, so reordered records do not change cell placement. Unsupported min/max pairs, invalid sizes and duplicate records fall back with a reported note, and imports are bounded to 1,000 tracks per axis. Public REST snapshots ignore the private track metadata.
Figma's schema has leadingTrim on NodeChange (NONE, CAP_HEIGHT), and
the REST importer already reads style.leadingTrim, but the .fig mapping
never copied it, so text set to "Cap height to baseline" imported
untrimmed in its trimmed box. text_style now copies it with the other
text fields, from the layer or its text style.
Vector points carry their own stroke join in a .fig file. Import now keeps each point's join as the point's own override instead of dropping it for the layer's join.
Joins equal to the layer join no longer force styled stroke runs, and an absent layer join counts as miter. Unresolved point styles, unknown joins, and joins that dashed or variable-width strokes cannot draw each produce a warning.
A collection a saved file references by its asset key (the form library
collections keep) instead of a guid resolved to nothing, so every
explicitVariableModes entry was dropped silently and bound layers rendered
in the collection's default mode (#267). The .fig import now resolves such
references against the file's own nodes (the collection node carries the
same key) for mode entries and the variables' collection link, and reports
modes of collections it can't resolve.
- The document and its pages now say they are visible and opaque, and a page carries its background opacity and enabled flag. Left out, a reader took them as hidden and drew nothing: every frame was blank and the page background transparent.
- The schema written into the file uses the format's own names, kinds (struct or message), field ids and enum values for everything the export writes. The schema text parser takes `NAME=n` and `type field=n`.
- Rectangles are written as rounded rectangles with their four radii. Frames, components, instances and sections carry their fill outline; plain strokes carry their stroke outline (a ring around boxes and whole ellipses, the widened path of other centred strokes).
- Text carries its layout settings and one line entry per paragraph. Font styles are written under their standard names ("SemiBold" becomes "Semi Bold", "Regular Italic" becomes "Italic"); the import reads both spellings as the same style.
- A line height that is a share of the font size is written as a raw multiplier: percent units count from the font's own line height, so 140 % was laid out taller than in FigCraft.
- Masks keep their type (alpha, outline, luminance), written and read back.
- The import reads per-side stroke weights whose zero sides are left out (a bottom-only border no longer comes back on all four sides).
- `meta.json` gives the page colour and the thumbnail size; archive members are in the usual order. Texture effects are written under the format's grain type; WebP export settings are written as PNG with a warning (the format has none).
- Docs: `docs/file-format-parity.md`, `docs/gaps.md` (gap 1), `ROADMAP.md`, `docs/target-app-parity.md` (wording only; no number moves).
An experimental, best-effort `.fig` export, so work done in FigCraft can be handed back as a `.fig` file. It is written from the same public format write-ups the importer uses and is the inverse of the importer's mapping. Only round trips through FigCraft's own importer are verified; other apps may not open every file.
- **Encoder** (`crates/format/src/fig/kiwi.rs`): the kiwi encoder that test fixtures used is now shipped code. It never panics and checks every value against the schema: unknown fields or enum names, missing struct fields, wrong types, out-of-range numbers and non-finite floats are errors, and NULs inside strings are dropped. Varints and floats are written exactly the way the decoder reads them. `Schema::parse` reads the compact text schema form, and the importer's tests use it too.
- **Writer** (`crates/format/src/fig/export.rs`): writes the zip (`canvas.fig`, `meta.json`, `thumbnail.png`, `images/<hash>`). `canvas.fig` holds the fig-kiwi header, then a raw-deflate binary schema and a raw-deflate message (flate2's pure-Rust backend, already in the tree; no new dependencies). The schema covers only the fields we write. The canvas, images, blobs, node count and nesting depth are capped, and floats are made finite before they are written.
- **Mapping:** pages; frames, groups, sections, rectangles, ellipses (arcs), polygons, stars, lines and vectors (vector network blobs with per-point radius, mirroring and caps, plus fill outlines); text with styled runs (UTF-16 indexed) and list types; slices; boolean groups; solid, gradient and image paints (image bytes included); strokes; effects; corners; constraints; masks; auto layout, including grids with tracks and cells; child layout; layout grids; export settings; components, component sets and their properties. Instances are written as their main plus the fields their layers change, by id path. This covers nested instances and swaps, and fields the main sets but the instance clears. Prototype interactions (triggers, navigation and overlay actions, transitions, easing), scrolling and flow starts are included. Everything skipped is reported in a warnings list: style links, variables, comments, Dev Mode data, plugin data, guides, slots, brush and width-profile strokes, pattern fills, image adjustments, variable and conditional actions, and more.
- **Import fixes found by the round trips** (`map.rs`): ellipse arcs, boolean operation types and component set descriptions are now read. Durations are read as whole milliseconds, because seconds stored as 32-bit floats used to truncate by 1 ms.
- **File writes:** a failed write no longer leaves its `.tmp-figcraft` file behind. The junk-parameter sweep found this.
- **Commands and UI:**
- `file.export {format: "fig"}`, or any path ending in `.fig`, writes the whole file and returns `warnings`. Without a path it returns base64 `data`.
- File > Export... has a new "The whole file" section with a ".fig (experimental)" row. A note under it says other apps may not open every exported file. After export, a toast reports how many kinds of content were left out.
- `figcraft-cli export --format fig` (or `-o x.fig`) and `convert` to `.fig` print the warnings as `note:` lines.
- MCP's `export` accepts `format: "fig"` and passes the warnings through.
- New UI strings have English and Spanish entries.
- **Docs:** file-format-parity.md has a new export section covering what is written, what is skipped, and that only round trips through our own importer are verified. Also updated: mcp.md, cli.md, gaps.md (gap 1), target-app-parity.md and ROADMAP.md. Grading is kept low. Import/export goes to 86%/58% and file formats to 60%. The file-exchange discount stays at x0.85, because nothing shows that another app opens these files.
The `.fig` import's tail (gap 1):
- **Editable vector networks.** Vector layers now open as vector networks instead of their outlines: points, curved segments with their tangents, filled areas (loops) with their winding rule, per-point corner radius and handle mirroring, and end caps taken from the end points' own styles (an arrow drawn on one end keeps its arrowhead on that end only). The network blob is decoded defensively (`crates/format/src/fig/network.rs`: bounds-checked reads, capped counts, indices checked, floats finite) and scaled from its normalized size to the layer's size. When the blob is missing or damaged, or an open path shows a fill that regions can't hold, the outline import is used as before (damage is reported in the warnings).
- **Layer kinds that used to be dropped** now import as the closest editable layers, each kind with one warning: sticky notes, shapes with text, code blocks, tables and table cells, stamps, media, widgets and slide containers become frames with their paints and children, their text added as a text layer (notes top-left, shapes centred); a non-rectangular shape with text gets its saved outline as a vector under the frame; connectors, highlights and washi tape become vectors (the network or outline when saved, else a straight line between the connector's ends: free points or a point of a sibling layer) with the connector's end caps; text on a path becomes plain text.
- **Variable prototype actions:** set-variable (literal, variable or expression values), set-variable-mode and conditional actions (if / else blocks with expression conditions) are mapped onto the REST importer's existing support. A condition that can't be read makes the whole conditional skip with a warning rather than turning into an "else".
- The REST importer gained two small extensions the `.fig` mapping writes: `vectorNetwork` (FigCraft's own network JSON, validated: sizes capped, coordinates finite and in range, references sanitized, else the outline with a warning) and per-end `strokeCapStart` / `strokeCapEnd`.
- Docs: gaps entry 1, file-format parity, target-app parity, ROADMAP and the MCP guide's `.fig` section.
.fig files open through file.open, file.import, drag and drop, the dialogs, the CLI and MCP. A generic reader written from the public format description decodes the zip, the deflate/zstd chunks and the kiwi schema and message, then maps nodes by field name onto the REST importer: layer tree, geometry, paints with styles resolved, strokes, effects, text runs, vectors as outlines, auto layout including grid, constraints, components, sets, instances expanded from their main with overrides, styles, variables with modes and bindings, prototyping and image fills. Unsupported layers come back as warnings. Fuzzed; tested on fixtures generated in code. New dependency: ruzstd (pure-Rust zstd decoder).
bflatastic