## Changes
- **Prompt character counters on web**: the website and webapp
promptboxes (image + video) now show a `123 / 2000` counter driven by
the API's `text_prompt_max_length`, matching the desktop app. It turns
red when over the limit and generation is blocked with a toast; typing
is never clamped. The webapp's fullscreen prompt editor shows the
counter too.
- **Seedance CJK waiver**: seedance video models normally use the API
limit, but the limit is waived (shows `∞`) while the prompt contains
Chinese characters. Applies on desktop and web via a shared
`effectivePromptMaxLength` helper.
- **API type parity**: added the new `model_creator` field to the web
`OmniGenImageModelInfo` / `OmniGenVideoModelInfo` types (desktop types
were already up to date).
- **Cleanup**: deleted the manual model-to-creator map
(`MODEL_TYPE_TO_CREATOR`, ~150 lines) now that creator info comes from
the API; lightbox model icons use the shared
`getCreatorIconPathForModelId` lookup like every other surface.
### New shared lib: `@storyteller/omni-gen`
- The `omni-gen-hooks.ts` file was copy-pasted between `artcraft-webapp`
and
`artcraft-website` and had drifted (website copy was missing outage
toasts,
carried a stale hardcoded display-name map, and had weaker icon
resolution).
- Extracted the newer (webapp) version into a new lib
`frontend/libs/omni-gen`
(generated via `nx g @nx/react:library`): singleton session cache,
`useOmniGenImageModels()` / `useOmniGenVideoModels()`, outage message
constants.
- App-local toasts are injected via `setOmniGenErrorNotifier()`, called
once in
each app's `main.tsx`, since the hooks are singleton-cache based and
each app
has its own toast component.
### Consolidated model metadata into `@storyteller/model-list`
- `getModelDescription()` / `getModelInfo()` (tagline + info-blurb
fallbacks)
moved to `ModelDescriptions.ts`; API values still always win.
- `getModelDisplayName()` gained an optional `fullName` param (API name
wins);
the app-local wrapper is gone.
- Both apps' local `getModelCreatorIconPath()` replaced with the
existing,
stronger `getCreatorIconPathForModelId()` (checks `ALL_MODELS_LIST`
first,
then prefix match). The loader's duplicate `CREATOR_BY_PREFIX` map now
reuses
the same shared `MODEL_ID_PREFIX_CREATORS` table.
### Dead code removed
- `modelReconciliation.ts`: deprecated `reconcile*()` helpers (never
called) and
the empty, retired `BACKEND_TO_TAURI_*` alias maps.
`buildModelsFromListing()`
dropped its alias-map indirection accordingly (backend ids == tauriIds
since
the 2026-07 id migration). `modelCreatorFromBackend()` moved inline into
the
loader, its only consumer.
- Six unused static `*_PAGE_MODEL_LIST` constants in
`model-selectors-for-pages.tsx`
(superseded by the `use*PageModelList` hooks in #1710). The two
`IMAGE_TO_3D_*`
constants stay: 3D models are not served by the omni backend.
- Untracked the accidentally committed
`apps/artcraft/tsconfig.tsbuildinfo` and
gitignored `*.tsbuildinfo`.
### Website-visible improvements (from adopting the newer hooks)
- Model outage toasts now appear on the website (previously
webapp-only).
- Current display names from the canonical map (previously-unmapped ids
like
`gpt_image_2` and Midjourney models rendered as raw ids).
- Model taglines/descriptions now render in the website model picker.
- Midjourney and SwitchX creator icons now resolve.
Add a footer to the fullscreen prompt modal (desktop lib + webapp) so
focus mode is a complete editing surface, not just the textarea:
- Model selector, always-on reference-image row, and a Done button
(closes the modal) instead of Generate.
- Video: also surfaces the input-mode (keyframe/reference) picker and
the @Characters button (seedance_2p0). Shared image-row / character /
input-mode elements are reused inline and in the modal (DRY).
- Webapp: model selector + leftToolbar + image row in the footer.
Layout: editor area is flex-1/min-h-0/overflow-hidden with the image row
and footer shrink-0, so the editor takes only the leftover space and
scrolls internally — resizing the window shorter no longer overflows.
MentionTextarea gains a `style` prop so the modal can cap resize/height.
### Fullscreen focus mode (desktop + webapp)
- Added a top-right expand icon on every prompt editor that opens a
"Prompt" focus-mode overlay (blurred backdrop, `max-w-4xl` centered
card) for more prompting space.
- Same editor state — overlay binds to the same value + onChange, so
text/@-mentions persist. Auto-focuses with the caret at the end on open.
- Editor fills full height; @-mention dropdown no longer clipped; closes
on Esc / backdrop click.
- Reused `@storyteller/ui-modal`; shared `PromptFullscreenModal` +
button + `useFullscreenPrompt` hook across all 5 desktop promptboxes,
mirrored in the webapp's own PromptBox.
### 3D editor
- Model selector now renders inside the promptbox (was standalone on the
canvas).
- Pose button hides on deselect; pressing "Done" keeps it (flips back to
"Enter Pose Mode") while the character is still selected.
- Drag and drop into folders fix within gallery modal - doesn't close
immediately unless on 3D editor page.
### Tooltips & expand
- Fixed top-left menu tooltips getting stuck open after a tab change.
- Reworked image promptbox auto-grow/expand into a shared
`useAutoGrowEditorHeight` hook (no per-render style writes) with a
smooth height transition.
Reveal/RevealGroup called `motion(as)` inside the render body. Unlike
the cached `motion.div` accessor, calling `motion()` as a function
returns a brand-new component type on every invocation. So each
re-render produced a different element type at that position, React
unmounted and remounted the whole subtree, and any focused <input>
inside (the login/signup/forgot-password forms) lost focus on every
keystroke — you could only type one character at a time.
Cache one motion component per element type via a module-level map so
the component reference stays stable across renders. React now
reconciles in place and focus is preserved.
Verified with a Playwright A/B test on the login form: before the fix
typing "test@example.com" left only "t" with focus on <body> (1 blur
mid-type); after the fix the full value is retained, focus stays on the
input, zero blurs.
handleUseSelected only searched groupedItems (root library), so picks
made
while browsing inside a folder — which live in folderGroupedItems —
resolved
to nothing and an empty array was passed to onUseSelected. Search both
sources, deduping by id, so folder selections attach correctly.
Co-authored-by: Claude <noreply@anthropic.com>
Brandon Thomas