lightcraft

Audit basis. Refreshed 2026-11-07 against tag v0.19.0, commit 16636c45a63bff3065a6fd31aecc22a6fc15ca53 — git rev-parse HEAD in the pinned checkout confirms, and the installed /Applications/LightCraft.app matches (its binary prints lightcraft 0.19.0). The previous release this skill described was v0.18.0 (32d569a).

v0.18.0 → v0.19.0 delta. 479 files changed (+50,232/−3,046). The engine registry gains 9 command ids, zero removed (multiline extract 379 → 388): app.hotPixelLimit, lens.databaseStatus, lens.downloadDatabase, library.refreshLensProfiles, prefs.backgroundPrep, prefs.setBackgroundPrep, preset.saveRecipe, profiles.camera, profile.deleteCamera (crates/engine/src/cmd/), while UI_COMMANDS loses modelSetup.cancel (195 → 194) and lightcraft-cli gains an 11th subcommand, bench-cull (apps/lightcraft-cli/src/main.rs). New surfaces: the user's own DNG camera profiles (.dcp) import and replace a raw's colour model (crates/raw/src/dcp.rs, crates/engine/src/cmd/dcp_profiles.rs); the lensfun lens database is offered on first launch and downloaded on request — 56 XML files, about 5 MB, pinned by size and SHA-256 — with the profile pinned per photo (crates/engine/src/lensfun_download.rs, crates/engine/src/cmd/lens_refresh.rs, docs/lens-profiles.md); compressed Olympus/OM ORF now decodes from sensor data (crates/raw/src/vendor/orfc.rs), lifting coverage to 559 verified / 721 unverified / 129 preview / 37 unsupported of 1,446 (use-weighted ~97% fully, ~86% on a verified body); AI Denoise needs no setup by default because a built-in BM3D filter — no weights, GPU-accelerated — is the default denoiser (crates/denoise-core/src/bm3d.rs, crates/gpu/src/bm3d.rs, docs/denoise.md); Remove Hot Pixels is on for newly imported raws, bounded by app.hotPixelLimit; process versions extend through V6 (V3 pixel-scale sharpening, V4 calibrated Contrast/Blacks, V5 highlight roll-off, V6 colour-noise reduction) (crates/develop/src/settings.rs); Windows now reads the monitor's ICC profile (crates/wincolor, docs/display-profiles.md); Culling's loupe uses built previews and prepares photos ahead in the background (prefs.backgroundPrep, crates/engine/src/cmd/previews.rs); and the web build gains a page-host mode (?host=parent, apps/lightcraft-web/src/host.rs). Correctness fixes touch documented surfaces: masks stay on the photo under lens corrections/Upright/Transform/Rotate, lens corrections crop inside the corrected photo, Auto Sync applies As Shot/Auto/Temp to each photo's own white balance, export subfolders stay under the destination, and Korean/Vietnamese watermark text draws whole.

v0.17.0 → v0.18.0 delta. 13 files changed; the command set is unchanged (274 ids). A correctness-only wave: histogram percentile when the target truncates to 0, shutter-fraction validation, case-insensitive date search, LIGHTCRAFT_GPU_LIMITS overflow, settings merging with linkMargins, and Cmd/Ctrl-click taking a photo out of a survey (crates/ui-egui).

Verify against your installed version — the ArtCraft Launcher auto-updates, so a newer build may differ; every count quoted here was re-measured at this tag, and code paths that moved were re-cited. The measurement method for each count is in evidence.md.

Overview

LightCraft is a photo library plus a non-destructive raw developer — an open-source, clean-room reimplementation of Adobe Lightroom Classic 15.5 / Lightroom 9.6 (ROADMAP.md:3, docs/target-app-parity.md), written from scratch in pure Rust with no Adobe assets (README.md, AGENTS.md). Its surfaces are the desktop app (apps/lightcraft, eframe + wgpu + native macOS menus, migrated to egui 0.37 with the shared craft-ui docking in this release), lightcraft-cli (headless render, batch commands, MCP server, headless UI snapshots), and a WebAssembly build of the same UI (apps/lightcraft-web).

Maturity: alpha. At this tag the generated parity summary reports 83.5% of 543 in-scope rows (P0 98.3%, P1 94.4%, P2 56.7%, docs/parity-checklist.md:58), while ROADMAP.md:15 and docs/target-app-parity.md:15 still print 81.6% of 532 — they lag the generated block. The project's readiness estimate is unchanged at ~61% overall, ~51% for a mainstream working photographer and ~65% for an essentials user; beta is an estimated ~340–600 agent-hours away (ROADMAP.md:3-22). The gap is quality and coverage — camera colour, render fidelity, some raw variants, AI models — not missing buttons. Source: GitHub storytold/lightcraft, MIT OR Apache-2.0 (Cargo.toml).

Capabilities (verified)

Library, catalog and views — implemented

• The library is a folder, not a document. catalog.snap (JSON snapshot) + catalog.log (append-only JSON-lines op log: seq, CRC-32, fsynced; a torn final record is repaired; a mid-log is preserved as catalog.log.damaged-<seq>), plus presets.json, view.json, prefs.json, thumbs/ and Originals/ for copy-imported photos (crates/engine/src/library.rs:1-22; crates/catalog/src/journal.rs:1-40). Default: LightCraft Library under the user's Pictures folder; LIGHTCRAFT_LIBRARY overrides (crates/engine/src/library.rs:63-70).
• Crash-safe and single-writer. Every top-level command that changes the catalog persists its ops (fsynced) before replying; a failed write keeps the change in memory, reports unsaved and retries. One process per library: catalog.lock refuses a second writer with "This library is already open in …" (crates/catalog/src/lock.rs:86; crates/ui-egui/src/control.rs ui.inspect). On Windows a hard link to an original now counts as that original.
• Library model. Albums in folders, smart albums (rule trees incl. File Path rules; "in the last…" date rules refresh when the session clock crosses a minute), stacks/auto-stack, virtual copies, versions and history, ratings/flags/colour labels + configurable label sets, filter bar with Enable Filters (⌘L, library.toggleFilters sets a filter aside and brings it back), fielded search (including people named on faces), sort, Photo Grid / Square Grid / People, filmstrip, Compare and Survey culling with synced zoom, Reference view (⇧R, photo.setReference), Local folder browsing without import (edits go to XMP sidecars; "Add to My Photos" adopts the photo), Recently Deleted, missing-file relink (matches a library made on Windows by file name), hold view (library.holdView) and Assisted Culling (classical focus score, clipping, burst grouping; no expression/eye detection) (crates/engine/src/cmd/organize.rs, crates/engine/src/cmd/library.rs, crates/engine/src/cmd/filters.rs, crates/engine/src/cmd/query.rs, crates/engine/src/cmd/browse.rs, crates/engine/src/cmd/cull.rs; docs/parity-checklist.md rows LRC-LIB-REFVIEW, LR-LIB-CULL).
• Choosing a library at launch. Settings ▸ General ▸ At launch: Open the last library or Ask me; with Ask me (or Alt/Option held as the window appears) a window lists recent libraries with name, folder, last-opened time and a background existence check. app.recentLibraries returns {path, name, exists, checking, open, lastOpened} and {remove: path}/{clear: true}; app.newLibrary makes and opens one (crates/ui-egui/src/dialogs/settings/libraries.rs — moved from panels/settings.rs in v0.16.0; docs/control-protocol.md "Choosing the library at launch").
• Workspace. Detail view with Edit / Info / Keywords / History panels, full-screen preview (F), Lights Out (⇧L), second window (⌘F11), Before/After (Y, ⇧Y, ⌥Y), soft proofing (S; partial — see limitations), info overlay (⌘I) and View Options (⌘J), a loupe layout-image overlay, multi-select Info panel (view.infoSelectedPhotos shows the values the selection shares and marks mixed fields), and dockable/splittable panels via ui.dock / ui.dock.reset — side panels now dock in columns on craft-ui's shared docking (crates/ui-egui/src/menus.rs; docs/control-protocol.md "Methods").
• Long-running tasks (imports, exports, merges, downloads, preview builds, the face scan) live in one activity stack, visible to agents as activity.list / activity.cancel; quitting asks first (crates/engine/src/cmd/activity.rs; docs/background-tasks.md:1-10; ui.inspect → quitPrompt).
• Previews and proxies. library.buildPreviews|cancelPreviews|previewProgress build the disk thumbnail cache; photo.smartPreview plus library.smartPreviews|smartPreviewsLocation build/relocate smart previews that keep working when originals are offline (crates/engine/src/cmd/previews.rs; docs/camera-preview-colour.md "smart previews").
• Faces and focus points. XMP face names, editable face boxes and the People view work with no model; detection (YuNet) and recognition (SFace/AuraFace) are opt-in model installs run by LightCraft's own pure-Rust CPU interpreter (crates/engine/src/cmd/face_models.rs, crates/engine/src/cmd/face_detect.rs, crates/engine/src/cmd/face_recognize.rs; docs/faces.md:1-24). photo.inspect carries meta.focus; View ▸ Focus Point (⌥F), zoom-to-focus and a background photo.readFocusPoints sweep — only Sony ILCE-1 (α1) and ILCE-7SM3 (A7S III) at this tag (docs/focus-points.md:36,66-67; crates/engine/src/focus_scan.rs).

Import and raw decoding — implemented, with coverage limits

• Importable extensions are one list of 39 in crates/engine/src/import/files.rs (EXTENSIONS): jpg jpeg png tif tiff webp dng cr2 cr3 nef nrw arw sr2 srf raf orf rw2 rwl raw pef srw psd psb jxl gif bmp heic heif avif iiq crw mrw x3f kdc mos erf 3fr fff mef. Support is decided by extension (is_supported), then the decoder may still refuse. Modes: Add (reference in place), Copy (into Originals/ or a destination, with date/month/flat/folder-template filing and rename templates), Move (copy, verify, then remove the source and its sidecars); duplicates are detected by path and by content; copy imports can convert raws to DNG (crates/engine/src/import.rs; crates/engine/src/cmd/library.rs).
• Raw decoded from sensor data (docs/raw-parity.md; crates/raw/src/lib.rs): DNG (uncompressed, LJ92, lossy/Smart Preview, LinearRaw, Deflate, DNG 1.7 JPEG XL tiles, Apple ProRAW, row/column interleave), CR2, CR3 (lossless CRX Bayer + version 0x100/0x200 C-RAW; R5 Mark II), ARW (uncompressed, cRAW/ARW2, lossless L/M/S, A900 packed, ARW6 Compressed/HQ on the ILCE-7RM6), NEF (uncompressed, lossless/lossy, lossy-after-split, table-less 34713 strips as packed 12/14-bit), RAF (uncompressed and lossless/lossy compressed, Bayer + X-Trans), RW2/RWL/Panasonic RAW (formats 2/4/5/6/7/8 incl. tiled high-resolution format 8), PEF, ORF (uncompressed incl. 12.8 bpp, and compressed — OM-D E-M1 and E-M10 Mark III verified, other bodies unverified), SRW (uncompressed), ERF, MEF, Minolta MRW, 3FR/FFF, Kodak DCS520C TIFF, and Sony SR2. Hostile-file memory bounds were hardened across the raw decoders in this release.
• Per-model coverage is measured, not implied. Of Adobe's 1,446 listed models: 559 verified from sensor data (39%), 721 decoder-path-unverified, 129 preview-only, 37 unsupported (docs/raw/coverage.tsv sums re-checked at this tag; the docs/raw-parity.md headline now agrees). Use-weighted: ~97% of photographers' cameras fully decode, ~86% on a verified body (docs/raw-parity.md:20-25).
• Anything it cannot decode falls back to the embedded camera JPEG and is flagged previewOnly; the app rechecks those photos in the background on version change, and headless callers use library.recheckPreviewOnly (crates/engine/src/cmd/recheck.rs; docs/mcp.md). Sparse header + embedded-JPEG reads avoid pulling sensor data over a network share (56.5 MB → 5.9 MB on the test share; docs/raw-preview-reads.md:1-13).
• HEIC/HEIF decode exists behind the heif build feature; it is off by default for HEVC-patent reasons and on in official packaging (crates/codecs/Cargo.toml still claims releases don't ship it — packaging/macos/package.sh:65 passes --features lightcraft/heif, and the installed 0.19.0 binary contains lightcraft_heif::decode|probe|decode_thumbnail). The heic-rs path is hardened against hostile HEIF containers. AVIF is import-listed but never decodes (export only).
• Lightroom Classic catalog import (library.inspectLightroom / library.importLightroom, File ▸ Import Lightroom Catalog…): reads .lrcat directly (read-only pure-Rust SQLite incl. committed WAL; 16 GiB / 16,777,216-page cap), one-way and approximate, imports ratings/flags/labels/keywords/collections/virtual copies/mapped develop settings. The app first asks whether the catalog merges into the open library or goes into a new library — file.importLightroom {into?: "open"|"new", name?, location?} makes that choice headlessly, and Settings ▸ General can turn the question off (crates/ui-egui/src/dialogs/lightroom_destination.rs; docs/lightroom-catalog-import.md:87). Also folder.scanChanges / folder.synchronize and library.autoImport / library.autoImportScan (crates/engine/src/cmd/sync.rs, crates/engine/src/cmd/library.rs).

Develop, masks and merge — implemented

• 122 named develop sliders in one spec table plus 10 indexed specs (8 Point Color + 2 Red Eye), addressed as pointColor.<i>.<control> / redEye.<control>; 16 sections: Light, Curve, Color, Mixer, BwMix, Grading, Effects, Vignette, Grain, Detail, Optics, Geometry, Profile, Calibration, PointColor, RedEye (crates/develop/src/controls.rs; count it with lightcraft-cli controls --json).
• 22 built-in develop profiles (looks, crates/engine/src/presets.rs:500 PROFILES) — Color, Neutral, Vivid, Landscape, Portrait, Monochrome plus film/cinema/muted/B&W families — a profile browser with recent profiles; the profile is separate from the raw camera profile below.
• The pipeline is scene-referred linear Rec.2020 32-bit float with OkLCh-based colour and a filmic shoulder for raw, pass-through for untouched JPEGs (README.md). Process versions V1–V6 (V6 current, V1 frozen) make re-renders deterministic — V2 highlights/shadows, V3 pixel-scale sharpening with working Radius/Detail, V4 calibrated Contrast/Blacks, V5 highlight roll-off, V6 colour-noise reduction; develop.updateProcess moves photos forward (docs/process-versions.md; crates/develop/src/settings.rs Process).
• develop.set / develop.adjust / develop.auto / develop.wb / develop.wbPick / develop.matchExposure / develop.copy|paste|sync|autoSync / develop.merge / develop.reset* / develop.beginInteraction|endInteraction; white balance is Kelvin on raws with colour matrices or measured presets, relative (6500 = as shot) elsewhere (crates/engine/src/cmd/develop.rs; docs/control-protocol.md "White balance").
• Masking: brush, linear/radial gradients, luminance range, colour range, Subject / Sky / Background heuristics, plus Object and Describe via an optional SAM 3 model; Add/Subtract/Intersect, invert, and 17 local sliders + a colour overlay + Amount per mask. Straight agent brush strokes: mask.brushStroke {connect: true} (crates/develop/src/settings.rs:829-852; crates/engine/src/cmd/masks.rs; docs/ai-masks.md).
• Crop, geometry and repair: crop/straighten/rotate/flip with aspect ratios that stay portrait after rotation, aspect lock (crop.lockAspect, A), auto-straighten (refuses turns > 10° and needs ≥ 3 agreeing lines), Upright Auto/Level/Vertical/Full/Guided, Constrain Crop; Heal/Clone spots with auto source, non-generative Fill (PatchMatch), editable pins, Visualize Spots, Red Eye and Pet Eye (crates/engine/src/cmd/develop.rs; crates/inpaint/src/lib.rs:1-22).
• Photo Merge: HDR (auto-align, deghost incl. moving lights, deterministic order, bracket batching), Panorama (spherical/cylindrical/perspective, boundary warp, auto crop, Lanczos-3 resampling; hard maxima 512 MP / 65,000 px; default output at the photos' own resolution unless maxMegapixels caps it), HDR Panorama → DNG; commands merge.hdr|panorama|hdrPanorama|hdrBatch; CLI merge hdr|panorama|hdr-panorama|hdr-batch (docs/merge.md:20-31; crates/merge/src/pano.rs:54-62; crates/merge/src/flow.rs, crates/merge/src/refine.rs). Full-size exports got faster in this release (fewer copies and serial passes).
• GPU: the develop pipeline as wgpu compute kernels on Metal/Vulkan/DX12, checked against the CPU pipeline to within 1/255; CPU fallback on device limits or errors (README.md; docs/gpu-pipeline.md). AI Denoise is a per-photo switch, Bayer RAW only: a built-in BM3D filter needs no weights and runs on the GPU where one is faster, and an optional user-installed ONNX model (denoise-model.json) can replace it; results are cached under <library>/denoise/; lightcraft-cli denoise-serve serves tiles to other machines over a LAN/VPN with token auth (docs/denoise.md:1-40).

Presets, colour and lens data — implemented

• 41 built-in presets with an Amount slider 0–200%; copy/paste/sync of chosen settings groups; .lcpreset is LightCraft's own preset file. Import also accepts XMP presets, legacy .lrtemplate, Luminar .lmp/.mplumpack, .zip bundles, .cube LUT profiles and local-adjustment effect presets (.lrtemplate, .xmp, .lceffect) (crates/engine/src/presets.rs; crates/engine/src/cmd/preset_files.rs, crates/engine/src/cmd/lut_profiles.rs, crates/engine/src/cmd/mask_effect_files.rs).
• Camera colour starts from DNG matrices, guarded per-file fits to the camera JPEG, 10 bundled profiles (Canon EOS D60; Sony ILCA-99M2, ILCE-7CR, ILCE-7M4, ILCE-7RM5; Fujifilm X-H2S, X-T4; Nikon Z 6II, Z 8, D5300 — crates/engine/src/camera_profiles.rs:27 BUNDLED; assets/camera-profiles/), the user's own imported .dcp camera profiles (profiles.camera / profile.deleteCamera, crates/raw/src/dcp.rs), 52 spectral matrices (crates/raw/src/spectral_table.rs, 52 Camera entries), or a neutral fallback. lightcraft-cli calibrate fits per-camera profiles from a user's own raws (Sony ARW, Nikon NEF, Fujifilm RAF). Measured ΔE against the camera JPEG is 2.9–10 (docs/gaps.md:18).
• Lens corrections: embedded DNG opcodes, Panasonic/Leica RW2/RWL distortion, Nikon Z 6II / Z 8 / Z f distortion, Sony ILCE-7RM4A / ILCE-7CR / ILCE-7RM6 / ILCE-6700 (new in v0.16.0) distortion, and Fujifilm distortion + vignetting for 59 validated bodies — X-S20 added in v0.16.0, plus lens profiles for compact cameras with a built-in lens; all go through Optics ▸ Enable Lens Corrections (crates/raw/src/vendor/; docs/*-lens-corrections.md). For everything else, a lensfun XML database adds distortion/TCA/vignetting: the app offers to download it on first launch (56 files, ~5 MB, pinned by size and SHA-256) or reads a user folder via LIGHTCRAFT_LENSFUN_DIR; no database is bundled, Adobe .lcp is never used, matching is automatic, the profile is pinned per photo, and fisheye is not corrected (crates/lensdb/src/lib.rs:1-14; crates/engine/src/lensfun_download.rs; docs/lens-profiles.md).

Export and output modules — implemented

• 10 export formats (ExportFormat, crates/engine/src/export/options.rs:5): jpeg png tiff psd webp avif jxl original dng rendereddng. psd is a flat PSD (8/16-bit, no layers); original copies the file plus an XMP sidecar carrying the edits; dng writes a raw photo as DNG with edits embedded; rendereddng bakes the render into a linear 16-bit float DNG.
• Validated export options (ExportOptions, crates/engine/src/export.rs, parsed by from_params): quality 1-100, bitDepth 8|10|16|32 per format, colorSpace srgb|displayP3|adobeRgb|proPhoto|rec2020 with embedded ICC, metadata all|allExceptCamera|copyright|none, removeLocation (an original export is refused when a PSD/PSB embeds GPS it cannot strip), sharpen none|screen|matte|glossy + amount, tiffCompression, dngCompression, conflict unique|overwrite|skip, resize by longEdge|shortEdge|width|height|megapixels|percent with dontEnlarge, fit dimensions|widthHeight, ppi, limitKb (JPEG, incl. HDR gain-map JPEG), text/image watermark (vertical, rotation, nine anchors, image orientation honoured), hdr with hdrTransfer pq|hlg, jxlLossless|jxlEffort|jxlProgressive, destinationMode: "sameAsOriginal", and macOS-only addToPhotos + photosAlbum.
• File-naming and folder templates share one token set — {name} {num} {seq} {seq:N} {date} {date:%Y%m%d} {folder} {camera} {lens} {iso} {rating} {title} {creator} {ext}; unknown tokens stay as typed, illegal characters become -, collisions get a suffix (crates/engine/src/rename.rs; photo.renameTokens lists them).
• Contact Sheet PDF: A4/Letter, rows/columns, captions, 150 dpi sRGB; limits 1,000 photos / 100 pages / 256 MiB; atomic replace; command export.contactSheet (crates/engine/src/contact_sheet.rs:87-92).
• Print module as print.* commands (print.get/set/reset, print.cell.*, print.layout.*, print.exportPdf, print.template.list|save|apply|delete|export|import): Single Image, Contact Sheet, Picture Package (Auto Layout), Custom Package; zoom-to-fill / rotate-to-fit; borders/strokes; A3–A5, Letter/Legal/Tabloid, 4×6/5×7/8×10/13×19, Japanese photo/postcard sizes or custom; margins; guides/rulers; saved templates as .lcprint files; printer/paper ICC + intent + black-point compensation; 8/16-bit colour-managed PDF; Page Options (page numbers, page info, crop marks) on Single Image and Contact Sheet (crates/engine/src/cmd/print.rs:385-401; crates/engine/src/print/template.rs:34 FILE_EXT = "lcprint"; crates/engine/src/print_pdf.rs).
• Slideshow: playback (fades, random order, repeat, prepared previews), layouts, overlays (identity plate, rating stars, text tokens, export watermark), intro/ending title screens, built-in + user templates with import/export, and background export to PDF (page per slide), numbered JPEG sequence, or AV1 video in MP4/WebM (no audio, at most 2 hours); commands slideshow.*, progress via activity.list, slideshow.exportStatus (crates/engine/src/cmd/slideshow.rs; crates/engine/src/slideshow_export.rs:39-42; docs/slideshow.md).
• Apple Photos (macOS only): app.export {addToPhotos: true, photosAlbum} or export.addToPhotos {paths}; the first run triggers a macOS Automation permission prompt (docs/apple-photos.md).

Interchange — implemented

• XMP sidecars read/write, manual (⌘S, photo.saveMetadataToFile) or automatic (library.xmpPreferences {autoWrite, naming: stem|full}); LightCraft's fields plus the full develop state as lc:settings under http://ns.lightcraft.app/lc/1.0/ (crates/meta/src/xmp.rs:18); owned properties are replaced, capture time/GPS only when LightCraft has a value, foreign namespaces kept byte for byte (including when they use lc:/xmp: prefixes), an existing sidecar is merged into, writes are atomic, failed writes retry and library.retrySidecars forces them (docs/xmp-interop.md; crates/engine/src/cmd/xmp.rs).
• Reading Adobe crs: develop settings is approximate and reports per-photo warnings; Lightroom radial masks apply inside the circle as Lightroom does (docs/xmp-interop.md).
• PSD/PSB metadata: imports embedded EXIF (resource 1058) and XMP (resource 1060) image resources, so capture dates, GPS and descriptive fields survive; PSD edit copies keep their capture date (crates/meta/src/psd.rs:1-20).
• GPX track logs, offline reverse geocoding (GeoNames towns ≥15k; metadata.reverseGeocode {maxDistance}), keyword lists with keyword sets (nine keywords per set, ⌥1–⌥9; next/previous set ⌥0/⌥⇧0), XMP face names (face boxes are catalog-only and never written to XMP), and capture-time edits (photo.setCaptureTime) (crates/engine/src/cmd/geocode.rs, crates/engine/src/cmd/keywords.rs, crates/engine/src/cmd/manage.rs).
• macOS edit-in: Edit in PhotoCraft / Edit in Other Editor renders <name>-Edit.tif (16-bit Adobe RGB by default; flat .psd or 8-bit optional), stacks it on the original and opens it; at most 50 photos per call (crates/engine/src/cmd/convert.rs).

Automation surface — implemented (the strongest part of the app)

• Everything is a command. At this tag: 388 engine command specs (counted from the cmd! literals in crates/engine/src/cmd/*.rs, wired by 41 specs() sites in crates/engine/src/cmd.rs) and 194 UI commands in UI_COMMANDS (crates/ui-egui/src/menus/ui_commands.rs) plus 18 language commands, i.e. 212 UI entries, so a running app's engine.commands reports 600 commands. engine.commands returns id, label, menu path, shortcut, parameter doc and current enablement; UI, CLI, control channel and MCP all dispatch by id. Import extensions (39), export formats (10), develop controls (122 + 10), presets (41), develop profiles (22), spectral matrices (52) and interface languages (18) are unchanged at v0.19.0.
• MCP server: lightcraft-cli mcp [--demo | --library DIR | FILES…] headless, or lightcraft-cli mcp --connect 127.0.0.1:7980 to drive the live app. Hand-written JSON-RPC over stdio, protocol 2025-06-18 (older revisions accepted), 7 resources (lightcraft://document, //commands, //library, //photos, //photo/active, //develop/active, //controls). Helpers: 17 headless / 28 connected (crates/mcp/src/tools.rs:79 helper_tools: 17 always-on tools + 11 UI-only when a desktop app is attached): command_list/run/batch, doc_inspect, render_preview, import, query_photos, select_photos, list_controls, get_develop, set_develop, apply_preset, crop, render_photo, export (plus the documented legacy list_commands / run_command); connect mode adds ui_inspect, ui_screenshot, screenshot, inspect_ui, set_ui, list_widgets, click, press_key, type_text, clipboard, pointer_gesture. One generated cmd_* tool per reported command, so tools/list totals 405 tools headless (17 helpers + 388 engine commands) and 628 connected (28 helpers + 600 commands); --compact lists helpers only (crates/mcp/src/tools.rs; crates/mcp/src/server.rs; docs/mcp.md).
• Control channel: lightcraft --control 7980 (or LIGHTCRAFT_CONTROL_PORT) serves JSON-lines on 127.0.0.1:PORT, loopback only, no authentication, at most 16 connections, 4 MiB line cap. Requests are answered on the UI thread with a 60 s timeout; a request that has not started and draws no frame for 5 s is failed and never retried; an idle connection closes after 5 minutes; a malformed line closes the connection (apps/lightcraft/src/control_server.rs:24-77; docs/control-protocol.md:1-40). Same dispatch surface as the previous skill (24 arms exposing 29 callable method names, list at crates/ui-egui/src/control.rs): engine.execute (aliases ui.menu.invoke, command), engine.commands, ui.menu.list/tree, ui.inspect, ui.widgets, ui.clickWidget|dragWidget|hoverWidget, raw ui.move|click|drag, image-coordinate ui.pointer, ui.key, ui.text, ui.clipboard, ui.scroll, ui.zoom (pinch), ui.set, ui.dialog.confirm|cancel, ui.resize, ui.zoomFactor, ui.screenshot, ui.render, ui.dropFiles (drag-and-drop imports, with hover for the drop card), app.quit.
• CLI subcommands (apps/lightcraft-cli/src/main.rs USAGE + dispatch, 11 at this tag): run, mcp, render, snapshot, merge, synth-merge, denoise-serve, commands, controls, calibrate, and bench-cull (a culling benchmark harness). run takes --demo|--library DIR|--connect [ADDR], repeatable --import, --script FILE|-, --keep-going; render takes --set control=value, --settings look.json|.lcpreset|.xmp, --preset ID, --size, --quality, and any export option via --opt KEY=VALUE; merge writes <first>-HDR.dng / -Pano.dng / -HDR-Pano.dng next to the first file (--preview OUT.png writes only a preview); commands lists ids sorted for readable families.
• Headless UI snapshots need no window and no GPU: lightcraft-cli snapshot --demo --script tour.jsonl -o out.png --size 1600x1000 --scale 2 runs the full egui UI on the CPU. In a running app use ui.screenshot {"path": …, "headless": true}; windowed capture falls back to headless after 2 s. In snapshot scripts ui.settle {timeoutMs} waits for renders (docs/control-protocol.md "Headless rendering").
• Update check (new this release): a startup update prompt modal with self-update, delegated to the shared craft-update crate (apps/lightcraft, Cargo.toml craft-update pin).
• Web build automation: the browser build exposes await globalThis.lightcraft.command("<id>", "<json params>") returning JSON text; params must be a JSON object (a typo such as {ids: [3]} is refused instead of running the command on the selection) (apps/lightcraft-web/index.html:39-40; apps/lightcraft-web/src/automation.rs:1-20).

Platform — implemented

• Native macOS / Windows / Linux (deb, rpm, AppImage, Flatpak, tarball; riscv64 tarball), a FreeBSD tarball, and a WebAssembly build of the same UI (WebGL2 via eframe's glow backend) with a persistent browser library in OPFS/IndexedDB; WebGPU and Safari/Firefox testing are in progress; JXL export is native-only and merges have no file access in the browser (docs/web.md; README.md Downloads). Windows installs per-user by default now (MSI for your account, no administrator prompt), and previews follow the monitor's ICC profile there (crates/wincolor).
• 18 interface languages in code: English, Simplified/Traditional Chinese, Japanese, Brazilian Portuguese, Spanish, German, Russian, French, Ukrainian, Polish, Czech, Italian, Georgian, Swedish, Finnish, Indonesian, Vietnamese (crates/ui-egui/src/i18n.rs; one LANGUAGE_COMMANDS entry each). Coverage is partial and lower-layer errors stay English; no RTL or complex-script shaping.
• Display colour management: macOS screen profile, Linux colord, Windows WCS monitor profile (crates/wincolor), or a chosen ICC file via Settings ▸ Display / app.displayProfile (docs/display-profiles.md). Portable mode (portable.txt beside the executable) and isolated automation profiles via LIGHTCRAFT_CONFIG_DIR (apps/lightcraft/src/main.rs).
• Tests in CI: Linux x64 runs cargo xtask ci on every push/PR (.github/workflows/ci.yml), FreeBSD runs a nightly build/package check; macOS/Windows package but do not run the suite (docs/gaps.md:26).

Known limitations and not-implemented

Read this before assigning work. Each item is the project's own admission or a code path that is absent.

Raw and colour

• Camera colour is estimated, not measured. Raws start from a per-file fit to the camera JPEG (run on the vignetting-corrected sensor since v0.16.0), the 10 bundled profiles, the user's own .dcp profiles, 52 spectral matrices, or a neutral fallback; measured ΔE against the camera JPEG is 2.9–10 (docs/gaps.md:18). The old "5 bundled profiles" doc bug is fixed in the docs at this tag.
• No bundled lens database. In-file data plus a lensfun XML database the app can download on request (56 files, ~5 MB, pinned by size and SHA-256) or read from LIGHTCRAFT_LENSFUN_DIR; Adobe's 3,635 profiles are never used — deliberately. lensfun matching is automatic (no manual picker), raws only, no fisheye, and not validated against real photos (docs/gaps.md:25; docs/lens-profiles.md).
• Preview-only raws: Nikon High-Efficiency (HE / HE★) NEF, several CRX variants (vertical-tile C-RAW, partial subband, multi-tile 0x200, adaptive low-frequency gains), compressed Samsung SRW, and the long tail IIQ / Hasselblad-compressed / X3F / CRW / GPR / DCR (docs/gaps.md:19,23-24,37; docs/raw/coverage.tsv). Compressed Olympus/OM ORF now decodes from sensor data (OM-D E-M1 and E-M10 Mark III verified; most other bodies still unverified). All 110 public CC0 CR3s decode; 26 CR3 bodies are verified, but some variants have no public sample (docs/cr3.md).
• AVIF cannot be imported (no permissive pure-Rust AV1 decoder); AVIF is export-only. .HIF HEIC stills are not CR3; HEIC 4:2:2/4:4:4 frames are refused (crates/engine/src/import/files.rs; docs/file-format-parity.md).
• PSD/PSB import is the merged composite only — no layers or layer styles; LightCraft writes flat PSDs (crates/codecs/src/psd.rs:1-5).
• Render fidelity is unmeasured: tone, highlights, texture/clarity/dehaze, noise reduction and sharpening are tuned by eye, so edits imported from Lightroom will not look identical (docs/gaps.md:20).

AI and enhancement

• AI masks: Subject, Sky, Background (and the model's People/Landscape kinds) are classical heuristics, not models; only Object and Describe use SAM 3, and the 3.4 GB model is not bundled. The built-in download-mirror list is empty, so the in-app download needs a user-configured LIGHTCRAFT_SAM3_MIRRORS or a manually installed model folder (docs/ai-masks.md; docs/gaps.md:22).
• AI Denoise ships no trained-network weights: it works out of the box only because the default is a built-in classical BM3D filter; a user-installed ONNX model + denoise-model.json can replace it. Bayer RAW only, no X-Trans, and it does not run on merged linear DNGs. The optional RawNIND weights are GPL-3.0 and gated behind a release policy (docs/denoise.md; docs/gaps.md:28).
• Absent: Super Resolution, Raw Details, content-aware/generative remove, reflections and distraction removal (docs/gaps.md:28-29).

Output and interchange

• Print lacks identity plate / watermark / photo-info overlays (beyond Page Options), Cut Guides, Auto Layout for Custom Package, printer-driver Page Setup, direct printer output and JPEG output; it produces a colour-managed PDF (docs/gaps.md:32).
• HDR display (EDR / Windows HDR) is not implemented: HDR editing, the headroom limit and SDR rendition work, but you edit blind on an HDR monitor (docs/gaps.md:31).
• Slideshow has playback, layouts, overlays, titles, templates and PDF/JPEG/AV1 export, but no music and no dedicated module editor (docs/gaps.md:33).
• Lightroom catalog migration is one-way and approximate; unsupported Adobe profiles/AI settings are archived (docs/lightroom-catalog-import.md; docs/gaps.md:30).
• No custom ICC output profile, and Render to DNG exists only as the export format rendereddng, not as a catalog command (docs/file-format-parity.md; docs/gaps.md:42).

Do not assign

• Video (playback, trim, edits, export): a Video media kind exists in the catalog model, but the importer's extension list is stills-only and there is no playback or editing path (crates/catalog/src/model.rs; crates/engine/src/import/files.rs; docs/gaps.md:35).
• Map view, Book module, publish services, tethered capture, Lightroom plug-ins / SDK; no HDR display (map/book docs/gaps.md:33, tethered :34, plug-ins :36, HDR display :31).
• No document file format. There is no "save as" for an edited photo: the library folder is the unit of persistence and edits live in the catalog and/or an XMP sidecar. The macOS bundle declares no CFBundleDocumentTypes (/Applications/LightCraft.app/Contents/Info.plist; crates/engine/src/library.rs:1-22).

Platform, quality and docs

• Per-model verification is thin: 551 of 1,446 Adobe models verified (38%), ~85% of photographers on a verified body (docs/gaps.md:27). macOS/Windows/Linux package but do not run the test suite. Windows/Linux GPU paths are only partly proven (docs/gaps.md:44).
• Localization is partial; lower-layer errors stay English; no RTL or complex-script shaping (docs/localization-parity.md). Accessibility is partial: AccessKit tree audited, but brush/red-eye keyboard access and a real screen-reader session remain (docs/gaps.md:39).
• Docs that disagree with code (trust the code): ROADMAP 81.6%/532 vs checklist summary 83.1%/544; crates/codecs/Cargo.toml says releases don't ship the heif feature while packaging/macos/package.sh:65 passes it and the installed binary decodes HEIC; docs/gaps.md:38 names Vietnamese as absent although it ships in crates/ui-egui/src/i18n.rs.

How to operate

Launch (macOS). open -a /Applications/LightCraft.app for the GUI, or the binary directly when you need flags: /Applications/LightCraft.app/Contents/MacOS/LightCraft. Real flags, parsed in apps/lightcraft/src/main.rs (HELP at :793):

lightcraft [OPTIONS] [FILES/FOLDERS…]      # positional paths are imported at launch
  --library DIR    open (or create) this library (env LIGHTCRAFT_LIBRARY; default: last library opened,
                   else ~/Pictures/LightCraft Library; Settings ▸ General ▸ At launch: Ask me shows a list)
  --no-demo        don't seed a new library with the procedural demo photos
  --memory         throwaway in-memory session (alias --demo); writes nothing
  --control PORT   serve the JSON-lines control channel on 127.0.0.1:PORT
  --version, --help

Useful environment: LIGHTCRAFT_LIBRARY, LIGHTCRAFT_CONTROL_PORT, LIGHTCRAFT_CONFIG_DIR (isolated settings/logs/model profile), LIGHTCRAFT_LANGUAGE=de|ja|zh-hans|…, LIGHTCRAFT_GPU_BACKEND=dx12|vulkan|metal|auto|off, LIGHTCRAFT_GPU=0, LIGHTCRAFT_LOG=info|debug / RUST_LOG, LIGHTCRAFT_MEMORY_MB, LIGHTCRAFT_NO_PREFS, LIGHTCRAFT_NO_NATIVE_MENU, LIGHTCRAFT_SAM3_DIR|_MIRRORS|_TOKEN|_REMOTE, LIGHTCRAFT_LENSFUN_DIR, LIGHTCRAFT_CAMERA_PROFILES, LIGHTCRAFT_DENOISE_MODELS|_URL|_TOKEN, LIGHTCRAFT_HDR_DISPLAY=0, LIGHTCRAFT_FPS=1 (start with the frame-rate overlay on), LIGHTCRAFT_PROFILE=1 (apps/lightcraft/src/main.rs:806-815; other env vars in crates/engine and crates/segment). Logs go to the settings folder's logs/lightcraft.log, rotated on each start.

The CLI is a separate download on macOS. The .app contains only the desktop app; there is no lightcraft-cli inside it and none on a stock PATH. Get lightcraft-cli-<ver>-macos-universal.zip from the release assets or build it with cargo build --release -p lightcraft-cli. Use the CLI and the app from the same release (docs/mcp.md "From an installed release").

Window/document model. One process = one open library; a main window (grid or Detail) plus an optional second window (⌘F11). There are no per-photo documents and no tabs. Switching libraries is File ▸ Open Library… (app.openLibrary, ⌘O) or File ▸ New Library… (app.newLibrary); if a library cannot be opened, the app starts an empty session and asks (ui.inspect → libraryProblem). Pointing Open Library at a Lightroom catalog folder offers the importer instead (docs/control-protocol.md; crates/ui-egui/src/panels/library_problem.rs).

Keyboard (verified in the command registry and crates/ui-egui/src/shortcuts.rs — do not invent others). Views: G grid, D detail, N survey, ⇧C compare, ⇧R reference, F full-screen preview, ⇧L Lights Out, ⇧F filter bar, / filmstrip, F6/F7/F8 filmstrip/left/right panel, \ Filter Bar in grids (Show Original in Detail). v0.16.0 additions: holding Space over any tool pans (click-zoom as usual); Alt/Option-drag shows a live clipping or sharpening preview; keyword painting is the tool.keywordPainter {keyword?} command (no default shortcut — assign one with app.setShortcut); Clear Color Label can carry a user shortcut. Panels/tools: E Edit, C Crop & Rotate, M Masking, H Remove, I Info, K Keywords, ⌘K Add Keywords, ⇧P Presets, ⇧V Versions, B/L/R/W brush/linear/radial/WB picker, A aspect lock (crop) / Visualize Spots, S soft proof (in grids S toggles the stack), Shift+S stack set-top. View aids: Z zoom, J clipping, Shift+O crop overlay, ⇧H tool overlay, ⌘J View Options, ⌘I info overlay, ⌥F focus point, Y/⇧Y/⌥Y before/after. Rating: 0–5, 6–9 labels, P pick, X reject, U unflag, ` cycle flag, [/] step rating, ⌘↑/⌘↓ flag up/down, ⇧0-9/P/X/U apply-and-advance. Keywords: ⌥1–⌥9 apply a set member, ⌥0/⌥⇧0 next/previous set. Misc: Enter open/done, Esc back, ⌘Z undo, ⌘L Enable Filters, ⌘⌥⇧C/⌘⌥⇧V copy/paste metadata, ⌘⇧I import, ⌘⇧E Edit in PhotoCraft, ⌘⌥E Edit in Other Editor, ⌘P print, ⌘⌥↵ slideshow. All keys are user-remappable (app.setShortcut); ⌘, and ⌘Q are reserved (crates/ui-egui/src/shortcuts.rs:17; docs/library-shortcuts.md).

Open / save / export. Import with File ▸ Import Photos… (⌘⇧I, file.addPhotos), Import from Folder… (file.addFolder), or by passing paths on the command line; files dropped on the window go through the import review (ui.dropFiles reproduces that headlessly). There is nothing to "save": each command persists. ⌘S (photo.saveMetadataToFile, Photo ▸ Save Metadata to File) writes sidecars; automatic sidecar writing is File ▸ Automatically Write Changes into XMP (library.toggleAutoWriteXmp), both driving library.xmpPreferences. Export with File ▸ Export… (dialog.export), Photo ▸ Edit In (⌘⇧E / ⌘⌥E), File ▸ Contact Sheet PDF…, File ▸ Print… (⌘P; its dialog is titled "Print to PDF…" and writes a colour-managed PDF), View ▸ Export Slideshow…, or the equivalent commands. The photo export itself is always app.export with the options above (crates/engine/src/export.rs; crates/engine/src/cmd/xmp.rs).

Drive it — the three supported routes

1. One-shot CLI (best for batch work, no session state):

   lightcraft-cli run --import ./shoot develop.set control=light.exposure value=0.7 \
       develop.auto app.export dir=out longEdge=2048 colorSpace=displayP3
   lightcraft-cli run --library ./lib library.select ids=[3] develop.get
   lightcraft-cli run --demo --script steps.jsonl --keep-going    # JSON lines; stdin works too
   lightcraft-cli run --connect 127.0.0.1:7980 ui.set view=detail ui.screenshot path=ui.png
   lightcraft-cli render in.dng -o out.jpg --set light.exposure=0.5 --size 2048 --quality 92
   lightcraft-cli merge hdr-batch ./brackets                      # every bracket set in the folder

   One JSON line per command ({command, ok, result|error, ms}); non-zero exit on failure; a bare word starts the next command, key=value values parse as JSON when they can (docs/mcp.md "One-shot commands").
2. Control channel (best for interacting with a live UI or an external controller):

   {"id":1,"method":"engine.execute","params":{"command":"develop.set","params":{"values":{"light.highlights":-45}}}}
   {"id":2,"method":"ui.dropFiles","params":{"paths":["./shoot"]}}

   Discover identifiers with engine.commands, develop.controls, ui.widgets; never hard-code translated labels or widget positions. Group slider bursts in develop.beginInteraction / develop.endInteraction (one undo step). On macOS ui.key injects egui events and does not synthesize Cocoa menu accelerators; use ui.menu.invoke with the command id when testing menu shortcuts (docs/control-protocol.md "Methods").
3. MCP (lightcraft-cli mcp): helpers plus command_list / command_run / command_batch / doc_inspect / render_preview, 7 resources, and one generated tool per command unless --compact. In connect mode the UI tools are added; in headless mode they return an error explaining how to start the app (docs/mcp.md).

Look at the result. Any UI work must be verified visually: lightcraft-cli snapshot --demo --script tour.jsonl -o out.png or {"method":"ui.screenshot","params":{"path":"a.png","headless":true}}. Before capturing, poll ui.inspect until the expected loupe.photo is active, loupe.source is render or cached, loupe.pending is false and the render queues are empty — or loupe.failed is set (docs/control-protocol.md "Before capturing a finished Detail view").

Working with an AI agent

Legitimate agent work

• Pre-generate or convert inputs into supported formats before touching the library. There is no ingest conversion step for unsupported files: an unsupported raw/AVIF stays unsupported unless you convert it externally. HEIC works only in a build with the heif feature (official macOS packaging has it) (crates/engine/src/import/files.rs; crates/codecs/Cargo.toml).
• Batch develop and export through lightcraft-cli run --import …, or against a persistent library with --library DIR so ratings and edits survive between calls (docs/mcp.md).
• Batch-render one image with a settings JSON: lightcraft-cli render in.dng -o out.jpg --settings look.json (a partial develop-settings object, or a single-preset .lcpreset/XMP file) plus --set control=value and --opt for export options (apps/lightcraft-cli/src/main.rs USAGE).
• Inspect and repair library files: catalog.snap and catalog.log are documented, CRC-checked JSON; the loader already handles torn appends and damaged records. Back the folder up before editing anything by hand (crates/catalog/src/journal.rs:1-40).
• Verify exports by reading the returned {path, width, height, bytes, sidecars} and then the file itself; the encoder is pure (bytes in, bytes out) so it is testable without a window (crates/engine/src/export.rs:1-5).
• Read XMP sidecars to understand another tool's edits; LightCraft preserves foreign namespaces. Do not expect LightCraft to write face regions to XMP (docs/xmp-interop.md).
• Rebuild camera colour profiles from a user's own raws with lightcraft-cli calibrate; enable lens corrections from the user's own lensfun database by setting LIGHTCRAFT_LENSFUN_DIR before launch (apps/lightcraft-cli/src/main.rs USAGE; crates/engine/src/lens_profiles.rs).
• Drive output modules as commands: export.contactSheet, the print.* family (including .lcprint template import/export), slideshow.* and export.addToPhotos, with activity.list progress (crates/engine/src/cmd/export.rs, crates/engine/src/cmd/print.rs, crates/engine/src/cmd/slideshow.rs; crates/engine/src/cmd/activity.rs).
• Choose where a Lightroom catalog goes before importing: file.importLightroom {into: "open"|"new"} (UI command) or run library.importLightroom after opening the target library (crates/ui-egui/src/dialogs/lightroom_destination.rs).
• Query and edit through MCP or the control channel using command ids and control ids from the registry, not translated labels; use ui.inspect/ui.widgets and automation ids for UI state (docs/control-protocol.md; docs/mcp.md).
• Install/point at optional models only with the user's consent: SAM 3 (Hugging Face gated; 3.4 GB) and denoise ONNX weights; face models are separate (docs/ai-masks.md; docs/denoise.md; docs/faces.md).

Do not

• Do not launch the GUI and drive it with OS-level synthetic input (osascript/System Events, cliclick): it goes to whatever window is frontmost. Use the control channel, lightcraft-cli snapshot, or MCP connect mode (repo AGENTS.md: "Never send OS-level synthetic input").
• Do not assume the CLI exists — on macOS it is a separate download, not inside the .app (docs/mcp.md "From an installed release").
• Do not assume a library can be opened by two programs at once (catalog.lock); use --connect instead (crates/catalog/src/lock.rs:86).
• Do not assume Adobe assets or formats work: .lcp, Lightroom plug-ins, .lrlibrary. .lrtemplate is import-only; a user's own .dcp camera profile can be imported, but none ship and Adobe .lcp is never used (docs/file-format-parity.md).
• Do not promise AI subject/sky masks, AI denoise, face recognition or super resolution out of the box — they need optional models, and the default SAM mirror list is empty (docs/ai-masks.md; docs/denoise.md; docs/gaps.md:28-29).
• Do not promise that a raw will develop from sensor data; check the format list above and expect preview-only for compressed ORF, HE Nikon NEF, some CR3 variants and compressed SRW (docs/raw-parity.md).
• Do not treat a returned command result as proof of a correct image. Nothing here validates the pixels (docs/gaps.md:20).
• Do not overwrite a catalogued original or its XMP sidecar: MCP export/render_photo/screenshot refuse such a path; use a new output folder (docs/mcp.md export helper).

Verification protocol

LightCraft is alpha (v0.19.0); its library is crash-safe but its output is not self-checking. For every non-trivial task:

1. Throwaway round-trip first. Make a scratch library (--library pointed at a fresh folder, or lightcraft-cli run --demo), import one file, change one slider, export, then reopen the export in LightCraft or any other viewer and confirm dimensions and pixels. Only then run the real batch. Use --memory when you only want to experiment — it writes nothing (apps/lightcraft/src/main.rs).
2. Never trust a silent export. app.export reports per-file {path, width, height, bytes, sidecars} and can report gpuFallback; a size/resolution mismatch or a zero-byte file is the failure you would otherwise miss. format: "original" also writes an XMP sidecar — check for it before assuming the folder is unchanged (crates/engine/src/export.rs; docs/mcp.md export helper).
3. Check the render, not just the command. ok: true means the command ran. Poll ui.inspect for loupe.pending == false (or loupe.failed for the reason) and take a screenshot/PNG before declaring success (docs/control-protocol.md "Before capturing a finished Detail view").
4. Version and back up per session. Copy the whole library folder before a bulk operation; the catalog is an append-only log, so a copied folder is a complete restore point. On the desktop there is no in-app backup command — file.backupLibrary / file.restoreLibrary are wired only where the platform supplies a backup service (the web build); on desktop the command returns "the library is a folder; back it up with your other files". Keep the app and CLI from the same release.
5. Update between projects, never mid-project. Re-run step 1 after any update — the launcher auto-updates, and a new build can change rendering; process versions V1–V6 exist for exactly that reason (docs/process-versions.md).
6. Know your single-writer rule. Any --library headless run fails while the desktop app holds that library. Reserve it in your plan, or drive the open app with --connect (docs/mcp.md:20; crates/catalog/src/lock.rs:86).
7. Treat the control port as a local privilege. It has no authentication and can run any command; only enable it for a task that needs it, on loopback (docs/control-protocol.md:39).

Sources & evidence

• Repo: https://github.com/storytold/lightcraft — inspected tag v0.19.0, commit 16636c45a63bff3065a6fd31aecc22a6fc15ca53 (git rev-parse HEAD in the pinned checkout), release dated 2026-11-06; inspection date 2026-11-07. Target: Lightroom Classic 15.5 / Lightroom 9.6 (ROADMAP.md:3). License MIT OR Apache-2.0.
• Installed /Applications/LightCraft.app: CFBundleShortVersionString 0.19.0, binary --version prints lightcraft 0.19.0, bundle id ai.storyteller.lightcraft — the installed bundle matches this pin. The bundle has no CFBundleDocumentTypes. The 0.19.0 binary contains lightcraft_heif::decode|probe|decode_thumbnail (HEIC enabled in the official build). Verify against your installed version.
• Key code files read for this skill: README.md, ROADMAP.md, Cargo.toml, apps/lightcraft/src/main.rs, apps/lightcraft/src/control_server.rs, apps/lightcraft-cli/src/main.rs, crates/engine/src/{library,import,import/files,camera_profiles,rename,contact_sheet,presets,export,export/options}.rs, crates/engine/src/cmd.rs (specs() wiring) and crates/engine/src/cmd/{library,develop,masks,keywords,print,export,slideshow,merge,sync,xmp,convert,previews,activity,cull,browse,query,organize,recheck}.rs, crates/develop/src/{controls,settings}.rs, crates/catalog/src/{journal,lock,model}.rs, crates/raw/src/{lib,spectral_table,dcp}.rs, crates/raw/src/vendor/, crates/denoise-core/src/{builtin,bm3d}.rs, crates/engine/src/{lensfun_download,lens_profiles}.rs, crates/engine/src/cmd/{dcp_profiles,lens_refresh,similar}.rs, crates/wincolor/, crates/mcp/src/{tools,server}.rs, crates/lensdb/src/lib.rs, crates/codecs/src/{psd,sniff}.rs, crates/codecs/Cargo.toml, crates/meta/src/{psd,xmp}.rs, crates/merge/src/, crates/inpaint/src/lib.rs, crates/ui-egui/src/{menus,menus/ui_commands,shortcuts,i18n,control,frame_rate}.rs, crates/ui-egui/src/dialogs/lightroom_destination.rs, crates/ui-egui/src/dialogs/settings/ (libraries.rs; was panels/settings.rs), packaging/macos/Info.plist.in, packaging/macos/package.sh, assets/camera-profiles/, .github/workflows/{ci,freebsd}.yml.
• Docs read for this skill: docs/{target-app-parity,raw-parity,file-format-parity,gaps,control-protocol,mcp,merge,slideshow,denoise,ai-masks,faces,focus-points,xmp-interop,process-versions,lens-profiles,camera-preview-colour,web,localization,localization-parity,background-tasks,raw-preview-reads,lightroom-catalog-import,display-profiles,apple-photos,library-shortcuts,parity-checklist,cr3}.md, docs/raw/coverage.tsv, docs/user/ trees.
• Counts were measured from source at this tag (not from README/docs or the previous skill): 388 engine command specs (41 specs() wiring sites in cmd.rs), 194 UI_COMMANDS + 18 language commands = 212 UI entries (600 total), 17/28 MCP helper tools, 7 MCP resources, 405/628 tools/list entries, 24 control arms / 29 method names (set unchanged), 11 CLI subcommands, 39 import extensions, 10 export formats, 122 + 10 develop controls, 41 presets, 22 develop profiles, 10 camera profiles, 52 spectral matrices, 18 languages, 1,446 Adobe raw models (559/721/129/37 at this tag).
• Shared libraries storytold/craft-libs (read-only reference): craft-docpath, craft-ui and craft-window pinned to c5a1cbfb5f9049f9270955caffb5bee5997aebf7, craft-crw to f3b62ec61f45534392376492fc5ad33047375205; craft-update pinned to a231e517c3b371e1bf0ee8af71ed807424a9a05d (Cargo.toml); LightCraft's own formats are documented in this repo, so nothing outside it was needed.
• Evidence tiers: implemented (the code path exists and is reachable from UI/CLI/command), partial/stub (exists but visibly incomplete — what is missing is stated), not implemented (absent or explicitly out of scope).

Uncertain / not verified in this pass

• Runtime GUI behaviour, menus and dialogs were not launched: every GUI claim comes from the command registry, the control-protocol document, the parity checklist and the packaging metadata. The installed binary was inspected read-only (--version, Info.plist).
• The exact visual/colour fidelity of renders is not compared against Lightroom; the project itself says it is unmeasured (docs/gaps.md:20).
• HEIC availability depends on the build: official packaging passes --features heif and the installed binary has the symbols; a source build without the feature does not decode HEIC/HEIF.
• MCP tool totals are derived statically from helper_tools and the measured command counts, not captured from a running tools/list.
• Whether a specific raw variant decodes can only be settled file-by-file; the per-model table (docs/raw/coverage.tsv) is evidence, not a guarantee for every body that writes the container.
• Where project docs disagree with code at this tag, code was followed and the drift is listed above.