### Outage handling
- Show an error toast when the OmniGen service is unavailable (5xx)
instead of failing silently.
- Distinct messages for the two failure points: couldn't load the model
list vs. couldn't enqueue a generation.
- Model-fetch toast fires once per failed attempt (singleton fetch +
message dedup); generation toast is gated to 5xx so content-policy/4xx
rejections still just show the failed card and 402 still opens the
credits modal.
### Prompt box
- Fix prompt box overflowing past the viewport at some desktop widths
(Generate button hidden) - removed `w-full` so it sizes to the content
area beside the sidebar (create-image + create-video).
### Recreate action
- Add **Recreate** to in-progress jobs in the create-page gallery (grid
+ list) and the task queue - previously only on completed/failed.
- Task queue had no Recreate at all; added it to all three states
(in-progress/completed/failed) for consistency.
- Recreate on the in-progress grid card sits bottom-right inline with
the prompt; in-progress list row reveals it on hover (matching completed
rows).
- Extract shared `useRecreateFromPromptToken` hook; refactor
FailedCard/FailedRow onto it (DRY).
### Model display names
- Consolidate to a single source of truth: webapp `getModelDisplayName`
now delegates to `@storyteller/model-list` (removed the duplicate local
map). Gallery, task queue, and lightbox now agree.
- Fix missing names: `gpt_image_1` (gallery) and Midjourney in the
lightbox - added `midjourney_7` / `midjourney_7_niji` / `midjourney_8`
to the model-list name + creator-icon maps; aligned `veo_2` → "Google
Veo 2".
### Sticky prompt settings across model switches
- Prompt-box settings no longer reset when switching models - the user's
choice persists and is resolved against the new model (kept if
supported, else falls back for display/generation only, without losing
the preference).
- Covers aspect ratio, count, resolution, quality (image) and aspect
ratio, resolution, count, duration, sound (video).
- Add `lib/resolve-model-setting.ts` (`resolveModelOption` /
`resolveModelCount`); `handleModelChange` now only switches the model (+
capability-bound input mode on video).
The frontend migrated to pnpm workspaces in #1606, which introduced
"workspace:*" dependencies. npm cannot resolve that protocol, so the
macOS and Windows publish workflows fail at "npm ci" with
EUNSUPPORTEDPROTOCOL, breaking production release builds on main.
Add a pinned pnpm setup step and replace the npm install block with
"pnpm install --frozen-lockfile". Verified the frozen install resolves
cleanly against the committed pnpm-lock.yaml.
---------
Co-authored-by: bflatastic <bombayapirat@gmail.com>
- Renamed the "Supercharge Credits" button to **Use on Web** (globe
icon) in the hero and bottom CTA. Routes subscribers to the webapp
homepage, everyone else to `/pricing`.
- Removed the "Create image / Create video" links from both sections.
- Split Discord + GitHub into a separate "Open source &
community-driven" CTA (subtle container) above the "Ready to craft?"
CTA.
MentionTextarea's DOM-sync effect rewrote innerHTML under the live caret
whenever buildHTML's identity changed, detaching the cursor on iOS
Safari mid-typing. Skip the rewrite when focused and the DOM already
matches value, and memoize colorMap in MobilePromptForm to stop the
per-render churn.
Brandon Thomas