Valentin KleinandClaude Opus 5.5
3aeedf4856
Squash of PR #625 by @vangel76. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
12 KiB
File-format parity
Last reviewed: 2026-10-10 · Last updated: 2026-10-11 · Change: minor (code pages, #505; native MULTILEADER, #459; foreign data kept, #528; PDF text, layers and sheets, #567, #571; templates and user files, #552; plotter devices, #564) · Target: Autodesk AutoCAD 2027
Every format AutoCAD 2027 reads or writes, with CADCraft's support, fidelity and tests. Work items are in gaps.md; the weighted score feeds target-app-parity.md.
Headline (estimated): file formats ≈ 30% of AutoCAD's. Opening the commonest 2D DXF/DWG drawings works, but saving is lossy: everything CADCraft doesn't model is dropped, and no test uses a real third-party file. DXF and DWG save at 2000–2018 (2018 by default, #421). Remaining: ≈ 170–280 Opus 5.5 hours: ≈ 40–60 h for DXF/DWG fidelity (the part that decides beta), ≈ 110–180 h for the other formats below, and ≈ 20–40 h for 3D exchange formats that wait for 3D modelling.
How this was measured
- CADCraft: read from
crates/dxf,crates/io(dxf_read.rs,dxf_write.rs,pdf.rs,svg.rs,tests.rs) andcrates/dwgon 2026-10-10. Nothing was run (the measurement pass did no builds). - AutoCAD: the formats in its Open/Save As/Import/Export/Attach/Plot dialogs, from Autodesk's
published documentation and the menu tree in
xtask/data/menu-catalog.txt. The installed AutoCAD 2027 for Mac bundle was inspected by listing names only (clean-room rule: nothing inside it is read):Resources/Supportlistsacad.lin,acadiso.lin,acad.pat,acadiso.pat,acad.mln,.dctspell dictionaries,.pssand.shxfiles, which confirms the supporting formats below exist in the target. None of them may be copied. - No AutoCAD-produced file is committed or used as a fixture. A real-file corpus needs owner
action (see gaps.md): files made by people who license them to us, kept in a
separate
storytold/cadcraft-corpusrepo.
DWG, version by version
AutoCAD 2027 opens every DWG version from R2.0 onward and saves 2018 (AC1032, still the current
DWG format), 2013, 2010, 2007, 2004, 2000 and R14. CADCraft reads DWG by converting it to DXF
with the acadrust 0.6.3 crate and reading that DXF; it saves by writing its own R2000 DXF and
converting it to a DWG of the chosen version (#421): SAVEAS version, else the drawing's own
version (the file's $ACADVER, 2018 for a new drawing), at least 2004 unless 2000 is asked for.
The bridge repairs two acadrust 0.6.3 gaps in 2010+ DWG: merged table cells and table-style
title/header suppression.
| DWG version | Code | AutoCAD 2027 | CADCraft read | CADCraft write | Tested how |
|---|---|---|---|---|---|
| R1.0–R2.6, R9, R10, R11/R12 | MC0.0–AC1009 | read | no (is_dwg accepts only AC10xx magic; acadrust doesn't target pre-R13) |
no | — |
| R13 | AC1012 | read | acadrust claims support; untested | no | none |
| R14 | AC1014 | read, write | acadrust claims support; untested | no | none |
| 2000–2002 | AC1015 | read, write | yes via bridge | yes, when asked for (no true colours, transparency or table styles: the format has none) | synthetic round trips (dwg_at_every_version_reads_back, dwg_roundtrip_keeps_non_ascii_text) |
| 2004–2006 | AC1018 | read, write | via bridge | yes | dwg_at_every_version_reads_back, dwg_roundtrip_keeps_extension_data |
| 2007–2009 | AC1021 | read, write | via bridge | yes | dwg_at_every_version_reads_back |
| 2010–2012 | AC1024 | read, write | via bridge | yes | dwg_at_every_version_reads_back, ac1024_dwg_with_many_entities_opens |
| 2013–2017 | AC1027 | read, write | via bridge | yes | dwg_at_every_version_reads_back |
| 2018–2027 | AC1032 | read, write (default) | via bridge; untested on real files | yes (default for new drawings) | dwg_at_every_version_reads_back (synthetic only) |
| Web build | — | — | no (error: "save as DXF") | no | — |
Everything lost in the DXF read/write paths below is also lost through DWG, plus whatever
acadrust's own DXF output omits (proxy objects, ACIS data, custom objects: not measured). User
report #57: a 10 MB DWG failed after about two minutes; limits now stop runaway conversions
early (#175), but the throughput problem remains. DWG ≈ 30% read, ≈ 15% write.
DXF, version by version
| DXF version | AutoCAD 2027 | CADCraft read | CADCraft write |
|---|---|---|---|
| R12 (AC1009), ASCII | read, write | yes: no version gating; old-style POLYLINE/VERTEX read (polyface meshes flattened to 3D polylines) | no |
| R12 binary | read, write | no (binary parser expects R13+ 2-byte group codes) | no |
| R13–R2018, ASCII | read, write each of 2000/2004/2007/2010/2013/2018 | yes, one reader for all versions | yes, each of 2000–2018 (#421): R2000 structure (later versions add only optional groups) with that $ACADVER; UTF-8 text from 2007; no CLASS instance counts (91) |
| R13–R2018, binary | read, write | yes | yes, each of 2000–2018 (SAVEAS binary; a binary file stays binary on QSAVE) |
Code pages ($DWGCODEPAGE, ANSI_932/936/949/950/125x) |
yes | yes (#505): ASCII and binary DXF before 2007 decode from their code page (ANSI_874/932/936/949/950/1250–1258, DOS437/850/852/866, ISO 8859, Macintosh); a file that is valid UTF-8 reads as UTF-8; Johab and other DOS code pages read as ANSI_1252 (dxf/src/codepage.rs) |
writes \U+ escapes (#169), which is correct; declares ANSI_1252 |
\U+/\M+ escapes |
yes | yes (\M+ since #505; Johab \M+4 stays literal) |
\U+ yes |
DXF content: read and write by object type
| Content | Read | Write | Notes |
|---|---|---|---|
| LINE, POINT, CIRCLE, ARC, ELLIPSE, LWPOLYLINE (widths, bulges, elevation), SPLINE, RAY, XLINE | yes | yes | round-trip tests |
| POLYLINE 2D/3D | yes | yes (as LWPOLYLINE / 3D POLYLINE) | polyface and polygon meshes flattened: lossy |
| TEXT, MTEXT, ATTDEF, INSERT + ATTRIB | yes | yes | ATTDEF prompt kept (#137) |
| DIMENSION (all 7 kinds) | yes | yes, with anonymous *D blocks |
a reopened dimension's *D block is regenerated under its name, unreferenced ones dropped (#495); DIMASSOC objects written |
| LEADER | yes | yes | |
| MULTILEADER | yes (#459) | yes, native MULTILEADER with MLEADERSTYLE (#459) | several landings per multileader are folded into one; per-object overrides and annotative scale contexts not kept |
| HATCH pattern / solid | yes | yes | |
| HATCH gradient | yes | yes (#162, fixes #80) | |
| SOLID, TRACE, 3DFACE | yes | yes | |
| ACAD_TABLE | yes | yes (entity + block) | |
| VIEWPORT | yes | yes; frozen-layer lists kept (#150) | VP layer colour overrides lost (#99, PR #197) |
| IMAGE | yes, with IMAGEDEF | yes, with IMAGEDEF and IMAGEDEF_REACTOR | no attach command or display of the picture yet (FF10) |
| WIPEOUT | yes | yes (#147, fixes #79) | |
| MLINE (+ MLINESTYLE), HELIX, TOLERANCE | yes (#510, #420) | yes | |
| REGION | other programs' REGIONs (ACIS data) kept as Unknown | our regions are written as closed polylines tagged with CADCraft xdata (#510; ACIS has no public spec) | |
| 3DSOLID, BODY, SURFACE, MESH, SHAPE, UNDERLAY, OLE2FRAME, LIGHT, CAMERA, SECTION, POINTCLOUD, proxy entities | kept as Unknown (raw tags) | written back as read, with their CLASS records (#528); dropped by DWG save (acadrust) | only their layer follows edits |
| Header variables | all read | 36 written | AutoCAD writes about 280 |
| Tables: LAYER (+ transparency, description), LTYPE, STYLE, DIMSTYLE, BLOCK_RECORD | yes | yes | |
| Tables: VPORT, VIEW, UCS, APPID | yes | yes: saved viewport configurations (#504), named views, UCSs | UCS elevation lost |
| Objects: LAYOUT, DICTIONARY, XRECORD, TABLESTYLE, DIMASSOC | yes | yes | |
| Objects: GROUP, MLINESTYLE, MLEADERSTYLE, IMAGEDEF, SCALE | yes | yes (groups #417, scale list #468) | |
| Objects: VISUALSTYLE, PLOTSETTINGS, materials, field objects and other unmodelled objects | kept raw (#528) | written back as read | not editable |
| Foreign xdata, extension dictionaries, reactors, unmodelled objects, colour books | kept raw (#528) | written back, handles remapped, references to deleted objects dropped | through DWG: xdata on table records and blocks, entity colour books and reactors, custom-class objects lost (acadrust) |
| Handles and owners | entity handles kept; others regenerated | consistent | stable across CADCraft re-saves (#495); other programs' table and object handles renumbered once |
DXF ≈ 50% read, ≈ 30% write (estimated; weights: common 2D entities 50%, annotation objects 20%, tables/objects 15%, foreign data preservation 15%).
Other formats
| Format | AutoCAD 2027 | CADCraft | Hours to parity (est.) |
|---|---|---|---|
| DWT templates | open, save | NEW/QNEW from .dwt and .dxf templates, SAVEAS .dwt (#552) |
0–1 |
| DWS standards | open, save, CHECKSTANDARDS | no | 6–10 |
| PDF output (PLOT, EXPORTPDF) | multi-sheet, layers, searchable text, hyperlinks | vector PDF with lineweights, real text in embedded TrueType subsets, layers (OCG), several sheets per file and hyperlinks (#567, #571) | 1–2 |
| PDF import (PDFIMPORT) and PDF underlays (PDFATTACH) | yes | no | 15–25 |
| Raster images attach (PNG, JPEG, TIFF, BMP …) | yes | IMAGE entity exists, no attach/loading/display path | 6–10 |
| PNG / JPEG / BMP / TIFF export | PNGOUT, JPGOUT, BMPOUT, TIFOUT | PNG; SVG too (AutoCAD doesn't export SVG) | 1–2 |
| DWF/DWFx export, DWF underlay | yes | no | 10–15 |
| DGN import/export/underlay | yes | no (needs a public spec; owner decision) | 25–40 |
| WMF, EPS, ACIS SAT, STL, FBX, IGES/STEP (Windows), 3DS | various | no; most are 3D and wait for M11 | 20–40 |
SHX fonts (.shx, big fonts) |
read | not parsed: our stroke font or FONTALT stands in, so text metrics differ from AutoCAD | 8–12 |
Linetype (.lin) and hatch pattern (.pat) files |
load user files | user files load (#552); linetype shape (SHX) elements are skipped | 1–2 |
Plot styles (.ctb, .stb), plotter configs (.pc3), .pmp |
yes | plot style tables in our own JSON format (plot-styles.md) and kept in the drawing; plotter devices with papers and printable areas in our own JSON format (plotters.md, #564); binary .ctb/.stb/.pc3/.pmp not read (no public spec) |
4–8 |
Scripts (.scr) |
SCRIPT | script text through the CLI, MCP and control channel; no SCRIPT command reading a .scr file |
1–2 |
Aliases (.pgp), CUI(x), workspaces |
yes | user alias files and ALIASEDIT (#552); no CUI(x) or workspaces | 4–8 |
SVG import (.svg, .svgz) |
yes (via scripts) | OPEN reads shapes, paths, groups and Inkscape layers into entities (#636); no CSS classes, use, fills, text or images |
2–4 |
AutoLISP (.lsp, .fas, .vlx), .dll/.arx plug-ins |
yes | no | see gaps.md (M12) |
Data extraction (.csv, .xls) and data links |
yes | DATAEXTRACTION/EATTEXT to CSV or a table (#433); table data links read CSV (#534); no .xls, no .dxe templates |
3–5 |
eTransmit (.zip packages) |
yes | no | 4–6 |
Tests
crates/io/src/tests.rs has 36 tests (round trips of entities, styles, associativity, tables,
constraints, non-ASCII text, DWG round trips, hostile input). All inputs are synthetic:
inline strings or our own writer's output. No test opens a file made by another program, and no
external validator runs in CI (ezdxf's audit was run by hand once, 2026-10-06). The next step is
a licensed real-file corpus with an open → save → re-open → compare test per file, and checking
our output in AutoCAD black-box.
Revision history
| Date | Change | Summary |
|---|---|---|
| 2026-10-11 | minor | Rows refreshed for merged work: code pages and \M+ (#505), MULTILEADER (#459), IMAGE, MLINE/HELIX/REGION/TOLERANCE, VPORT/VIEW/UCS tables, modelled and raw objects, foreign data (#528), DWT templates and user .lin/.pat/.pgp files (#552), PDF output (#567, #571), plotter devices (#564), data extraction (#433, #534); weighted percentages not re-measured |
| 2026-10-11 | minor | DXF (ASCII and binary) and DWG saved at 2000–2018, 2018 by default (#421) |
| 2026-10-10 | minor | WIPEOUT and gradients written since #147/#162; other save losses remain (#201, #204, #208, #209, #213, #223) |
| 2026-10-10 | major | First file-format parity doc: DWG and DXF version by version, per-object read/write, other formats |