mirror of
https://github.com/storytold/artcraft.git
synced 2026-10-20 17:02:53 -04:00
BFlatandClaude Opus 5.5
b5614ad333
## What
A new workspace member, `artcraft-native/`. It is the ArtCraft desktop
app in pure Rust: an egui UI with no Tauri, no webview and no
JavaScript. It covers the two main product pages, **Create Image** and
**Create Video**, and reuses the Tauri-free API crates
(`artcraft_client`, `artcraft_api_defs`, `tokens`, `enums`).
It also adds:
- a Home page and a Library page
- a lightbox
- sign-in and settings dialogs
- a sidebar and top bar in the style of the brutalist
**artcraft-webapp**
Tools that aren't ported yet show a "coming soon" page that links to the
web app.
```bash
cargo run -p artcraft_native --release
```
## Why
To get a native, webview-free desktop app. Generation goes through
**OmniGen** (`/v1/omni_gen/{models,cost,generate}`), the same API as the
webapp. Model capabilities come from the server, so the pickers stay
current without app updates.
## Prompt box
Recreated feature for feature from `PromptBoxImage` / `PromptBoxVideo`:
- **Editor:** auto-grow, focus mode, resize grip, length counter with
the model's limit, Enter-to-generate setting, and `@Image1` / `@Video1`
/ `@Audio1` mentions with highlighting and autocomplete.
- **References:** the fanned reference deck (hover-expand, reorder, add
menus with limits) and the first/last keyframe slots with swap. Files
can be uploaded, picked from the library, dragged and dropped, or pasted
as images.
- **Model selector:** families with a flyout, maker logos and capability
badges.
- **Pickers:** aspect ratio, resolution, quality, count, bitrate, output
format, duration slider, sound and input mode (Keyframe / Omni
Reference).
- **Generate:** a live credit cost estimate, validation (starting frame,
text-only models, limits, uploads in progress), and clear-all with
confirmation.
The feed shows running jobs (with progress and time left), failures
(dismissable), finished results and the paginated library, in a masonry
grid or a list. It supports Recreate, Make Video, Share and Download,
plus select mode for batch downloads.
## Reviewer notes
- **Sign-in.** On first launch the app reuses an existing desktop login
by reading the session cookie from the Tauri app's jar, and keeps its
own copy in `~/Artcraft/credentials/native_session.json`. Browser
(device-code) and password sign-in are also available.
- **Decoding.**
- Generate and cost requests are sent as JSON maps, like the Tauri app's
omni path.
- Jobs, library, batch and prompt responses use lenient structs
(`backend/wire.rs`), so new server enum values can't break decoding.
- **Renderer.** It uses eframe's **glow** renderer. `wgpu-hal` fails to
build against the `windows` crate versions this workspace pins for
Tauri.
- **Model names and logos.** Known families use curated names: "Flux 1
Dev" rather than the server's "FLUX.1 [dev]". When the server reports
`artcraft` as a model's maker, the logo comes from the id prefix instead
(Nano Banana → Google), matching the webapp.
- **Fonts and assets.** Inter, Archivo and Geist Mono are bundled under
the OFL and listed in `NOTICE`. The wordmark, icon and maker logos are
the existing in-repo assets.
- **Diagnostics.** `ARTCRAFT_SCREENSHOT=<png>` (optionally with
`ARTCRAFT_PAGE` and `ARTCRAFT_SCREENSHOT_EXIT=1`) saves a screenshot
once the window has settled.
## Testing
- `cargo test -p artcraft_native`: 29 tests pass. Unit tests cover the
catalog, feed store, wire decoding and backend helpers. Headless
`egui_kittest` tests drive the prompt box: generate gating,
Enter-to-generate, pickers, `@` mentions and clear-all.
- `cargo clippy -p artcraft_native --all-targets --no-deps` is clean.
- Run against a real account: the session import, model lists, cost
estimates, job polling, library paging and thumbnails all worked.
- **Not exercised: a real paid generation and real uploads.** I held
back to avoid spending credits, so they need a manual check.
## Not done yet
- `@Character` mentions for Seedance 2.0
- Provider choice (Midjourney direct, fal); everything runs through
ArtCraft
- In-app video and audio playback; these open in the system player
- Animated previews in the feed
- The remaining tools (Audio, 3D, editors)
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
ArtCraft Native
ArtCraft's desktop app in pure Rust: an egui UI over the same ArtCraft API client crates the Tauri app uses. There is no Tauri, no webview and no JavaScript.
The look follows the brutalist webapp (artcraft-services/frontend/apps/artcraft-webapp). That
means near-black surfaces with white/15 hairlines and 3 px corners on controls. Type is Archivo for
display, Inter for body text and Geist Mono for labels. The brand blue marks the primary actions.
Create Image and Create Video are fully built. The other tools show a "coming soon" page that links to the same tool in the web app.
Running
cargo run -p artcraft_native --release
The workspace's default member is the Tauri app, so pass -p artcraft_native.
- Sign-in. On first launch the app reuses an existing ArtCraft login. It reads the session
cookie from the Tauri app's jar (
<cache dir>/ai.artcraft.app/.cookies) and keeps its own copy in~/Artcraft/credentials/native_session.json. Otherwise use "Sign in with browser" (device code) or a username and password. - API host.
~/Artcraft/settings/env_configs.jsonselects it, exactly as for the Tauri app.
Architecture
| Module | What it does |
|---|---|
backend |
Tokio runtime and every API call; results come back to the UI as Events |
backend::wire |
Lenient response shapes; new server enum values can't break decoding |
backend::session |
Credentials: own file, Tauri cookie import, env config |
backend::media_cache |
Thumbnails and previews: downloaded and decoded off-thread, evicted when unused |
models |
Model catalog from OmniGen listings, plus names, families and limits |
prompt_box |
The shared prompt box: editor, reference deck, keyframes, pickers, model selector |
pages |
Create Image, Create Video, Home, Library, and the shared page shell |
feed |
Generation feed: running, failed and finished jobs, library paging, grid and list |
overlays |
Lightbox, library picker, sign-in and settings dialogs |
shell |
Sidebar and top bar |
theme, ui |
Colour and type tokens, Lucide-style vector icons, widgets, toasts |
How the backend works:
- Generation. It goes through OmniGen:
/v1/omni_gen/{models,cost,generate}/{image,video}. Model capabilities come from the server, so new models and options appear without an app update. - Typed client functions. Model lists, uploads, credits, login and deletes use
artcraft_client. - Lenient decoding. Generate and cost requests are JSON maps, as in the Tauri app. Job polling
(
/v1/jobs/session), the library list, batches and prompts are decoded with lenient structs. - Identity headers. Every request sends the desktop identity headers:
OriginandUser-Agent.
Prompt box parity
| Feature | Status |
|---|---|
| Auto-growing prompt (~4 lines), expand/collapse, resize grip | Done |
| Focus mode (fullscreen editor with deck and toolbar) | Done |
| Length counter, red over the model's limit | Done |
| Enter to generate (setting; Shift+Enter for a new line) | Done |
@Image1 / @Video1 / @Audio1 highlight and autocomplete |
Done |
| Reference deck: fan, +N badge, hover-expand, reorder, clear all | Done |
| Keyframe slots: first/last frame, tilt, swap | Done |
| Upload, pick from library, drag & drop, paste image | Done (video uploads must be MP4) |
| Model selector with families, logos, capability badges | Done |
| Aspect ratio, resolution, quality, count | Done |
| Bitrate, output format, duration slider, sound, input mode | Done |
| Live credit cost on the generate button | Done |
| Validation: starting frame, text-only banner, limits, uploads | Done |
| Clear all (confirms when references are attached) | Done |
@Character mentions and the Characters modal (Seedance 2.0) |
Not yet |
| Provider choice (Midjourney direct, fal) | Not yet; everything runs via ArtCraft |
| In-app video and audio playback | Opens in the system player instead |
| Animated video previews in the feed | Still frames only |
Feed and lightbox:
- Feed. Pending cards show progress, time left and a batch banner. Failed cards show the reason and can be dismissed. The grid is masonry, with a list view alongside. Hover actions are Recreate, Make Video, Share and Download. Select mode does batch downloads. The last viewed item is marked.
- Lightbox. Prev/next with the arrow keys, prompt copy and details. Actions are Recreate, Make Video, Download, Share and Delete.
Diagnostics
Set ARTCRAFT_SCREENSHOT=<file.png> to save a screenshot once the window has settled.
ARTCRAFT_PAGE (home, create-image, create-video or library) picks the page, and
ARTCRAFT_SCREENSHOT_EXIT=1 closes the app afterwards.
Tests
cargo test -p artcraft_native
There are unit tests for the catalog, feed store, wire decoding and backend helpers. Headless UI
tests (egui_kittest) drive the prompt box: generate gating, Enter to generate, the pickers,
@ mentions and clear-all.
Fonts and brand assets
assets/fonts holds Inter, Archivo and Geist Mono, all under the SIL Open Font License; the
licences sit next to the files. The wordmark, app icon and creator logos are the Tauri app's own
(frontend/apps/artcraft/app/public/resources, crates/desktop/artcraft/icons); see NOTICE
for their terms.