Image generation reused the single (text-endpoint) API key, which breaks the
common 'local LLM with no key + OpenAI for images' setup. Add an optional
image_api_key (encrypted, write-only, never returned); generate-design uses it
for image calls and falls back to the main key when blank (all-OpenAI setups).
Local sd.cpp / ComfyUI still need no key. Schema column + migration.
A prompt now produces a full sign: the LLM writes the design AND image prompts,
the server generates the images and composites them with the crisp text layer.
- lib/image-gen.js: text-to-image with 3 BYO/self-hostable backends, all behind
the SSRF guard: 'sdcpp' (local stable-diffusion.cpp OpenAI-compatible server,
exact small sizes that fit VRAM), 'openai' (cloud / OpenAI-compatible, snapped
sizes), 'comfyui' (prompt/history/view API).
- ai.js: prompt asks for a background_prompt (preferred — full-bleed atmosphere)
and an optional foreground image element; after the design is normalized, the
bg + fg images are generated best-effort (a failed image never fails the sign)
and returned as data URLs. New image_* settings (provider/base_url/model),
image_provider whitelist, schema column + migration.
- designer.js: AI-images section in settings; generate applies the background
image; publish bakes the background image into the HTML so it survives.
- server.js: raise JSON body limit to 12mb for embedded image data URLs.
Verified end-to-end on local Vulkan SDXL (RTX 5090): prompt -> bg+fg images on
the canvas -> publish creates a widget with the images embedded. 63/63.
Note: prod (not self-hosted) requires a PUBLIC image endpoint (e.g. OpenAI); the
SSRF guard blocks localhost there. Follow-up: upload generated images to the
content store and reference by URL to avoid multi-MB widget configs.
Models sometimes stacked text lines at the same y (unreadable) and emitted accent
shapes after text, so a band could hide the words.
- deoverlapTexts: push a line down only when it also overlaps horizontally
(leaves side-by-side text alone), with conservative line-height clearance so
real rendering doesn't re-overlap; shift the stack up if it ran past the bottom.
- Order shapes before text in the output so accent bands always render behind the
words.
Verified: 0 text-on-text overlaps across multiple prompts (Playwright DOM check);
unit test asserts overlapping lines get separated + shapes precede text. 63/63.
Text could run off the edge (long/large headlines, nowrap) and shapes placed at
the far edge (e.g. a bottom band at y=100) spilled over.
- Server-side fit pass on every generated element: shrink text fontSize so it
fits the canvas width (chars*fontSize*0.075, tuned for bold/uppercase
headlines) and height (incl. line-height), then nudge x/y within 4% margins;
clamp shapes so x+width<=100 and y+height<=100. Deterministic - doesn't rely on
the model getting layout right.
- Designer preview: vw -> cqw (+ container-type on the canvas) so text scales to
the canvas, not the browser window. The preview was overstating size vs what
actually publishes; now it matches. Published widget keeps vw (scales on the
player).
Verified: Playwright DOM check shows zero elements overflowing the canvas after
generation; unit test asserts long text is shrunk + repositioned in-bounds. 62/62.
- POST /api/ai/models lists the configured endpoint's models (OpenAI-compatible
/models) so the settings modal can populate a 'Load models' dropdown instead of
requiring users to type the model name. Combobox (datalist) so they can still
type a custom one. Admin only; same SSRF guard; uses the posted or saved key.
- Bump generate-design timeout 120s -> 180s for slow local endpoints.
Competitor pressure (Mandoe 'AI Magic Create'): prompt -> signage. We answer it
in a way that's actually BETTER for signage and costs the operator nothing.
Key idea: don't generate raw images (AI garbles text - fatal for menus/promos).
The LLM returns a STRUCTURED design spec (headline, supporting text, accent
shapes, palette) that the existing Designer renders with real fonts - crisp and
fully editable. Reuses the whole Designer.
BYOK, fully under the customer's control: each workspace configures its own
OpenAI-COMPATIBLE endpoint + key - OpenAI cloud OR self-hosted (Ollama / LM Studio
/ llama.cpp). Operator bears zero AI cost/liability.
- server/lib/secretbox.js: AES-256-GCM for the key at rest (never returned).
- routes/ai.js: GET/PUT /api/ai/settings (admin; key write-only) + POST
/generate-design (editor+). Output is strictly validated/normalized (cap count,
clamp ranges, px->%, strip HTML, validate colors) - never trust the model.
SSRF guard: hosted instances block private/internal targets; self-hosted (the
whole point of local AI) may point at localhost/LAN.
- Designer: an 'AI generate' panel (prompt + Generate) + a settings modal.
Verified end-to-end against local Ollama (llama3.1:8b): prompt -> editable design
on the canvas. Unit tests cover normalization + the SSRF guard. Suite 61/61.
Phase 2 (next): AI background images (OpenAI images / AUTOMATIC1111).
The logo/title/theme/favicon are static 'ScreenTinker' in index.html, and
applyBranding() only overrode them AFTER an async /api/white-label fetch - that
network delay was the flash, on every load and on switch (which reloads).
Now applyBranding caches the resolved white-label per workspace (keyed by the
JWT's current_workspace_id), and a tiny same-origin brand-prime.js loads
render-blocking right after the logo - so it applies the cached colors/name/
title/favicon/custom-css BEFORE first paint. CSP-safe (external 'self' script,
not inline). applyBranding still runs to refresh + re-cache. First-ever visit to
an uncached branded workspace still shows the default once; every load after is
flash-free.
After uploading, content thumbnails were blank until the item was added to a
playlist/widget. The public /api/content/:id/thumbnail (and /file) endpoints are
reference-gated (an anonymous player with a UUID must not pull arbitrary tenants'
media), and a plain <img> can't send a Bearer token - so a just-uploaded item 403'd.
- Backend: add an authenticated bypass - a logged-in user who can access the
content's workspace (verified from the Bearer token) may view its file/thumbnail
even when unreferenced. Anonymous players still hit the reference gate.
- Frontend: the content library lazy-fetches thumbnails/previews WITH the token
and swaps in an object URL (IntersectionObserver keeps it under the rate limit;
the URL is revoked after load).
Verified: unreferenced thumbnail now 200 with a bearer token, still 403 anonymous.
The upload input already has 'multiple' and the click handler shares handleFiles()
with drag-drop, so picking multiple files (shift/ctrl-click) already works - it just
wasn't discoverable ('click to upload' read as single-file). Reword to 'click to
select one or more'.
Saving a layout grew its zone count on every server restart. Root cause: the
editor saved zones with a per-zone delete-then-POST loop, and POST /zones minted
a NEW uuid for every zone - so each save replaced the seeded ids (z-sh-1, ...)
with fresh uuids. schema.sql re-seeds template zones via INSERT OR IGNORE on every
boot, so the next restart re-added the now-missing canonical zone alongside the
renamed copy -> a 2-zone template became 4, 6, ... (worse for self-hosters who
rebuild often).
Fix:
- PUT /api/layouts/:id now accepts a zones[] and replaces them atomically in one
transaction, REUSING each zone's id when supplied. The editor sends the full
set in a single call, so the layout ends up with exactly those zones and ids
stay stable (also fixes fit_mode not persisting, and stops device->zone
assignments being orphaned by id churn).
- One-time dedupe migration removes positional-duplicate template zones, keeping
the canonical 'z-...' seeded id so the re-seed stays an idempotent no-op.
Verified: 2 atomic saves keep count + ids stable with fit updated; dedupe restores
a polluted 4-zone split template to its 2 canonical zones. Suite 56/56.
The boot summary counted any non-throwing statement, so UPDATE/index migrations
(which always succeed) made a healthy DB report 'applied N new column migration(s)'
every boot. Count only a successful ALTER ... ADD COLUMN (genuinely new), so the
line appears only when a column was actually added.
Self-hosters rebuilding could end up schema-behind-code, failing only at runtime
(a missing users.must_change_password locked out all logins). Two root causes:
1. The migration loop swallowed EVERY error (catch {}), so a real ALTER failure
was indistinguishable from the benign 'duplicate column' on an already-migrated
DB. Now only 'duplicate column'/'already exists' is treated as a no-op; any
other error is logged loudly, and a one-line summary reports how many new
column migrations actually applied this boot.
2. Nothing verified the schema after migrations. Added lib/schema-check.js:
verifyAndRepairSchema() checks the tables + columns the request path REQUIRES,
idempotently repairs missing repairable columns (logging each), and if anything
required is STILL missing, prints a loud FATAL block and exits - failing fast at
boot instead of at the first authed request.
Note: the reported 'audit_log missing' was a misdiagnosis - the code uses
activity_log (0 refs to audit_log), created by schema.sql on every boot.
Tests: healthy (no-op), auto-repair of must_change_password, missing-table report.
Platform admins can now cleanly remove a customer org (account ends) or a stray
workspace from the UI, instead of raw SQL that risks orphaning resources.
The tenant cascade isn't pure DB CASCADE - workspace-scoped tables (devices,
content, playlists, ...) are NO ACTION and must be purged before the workspace.
Extracted that logic out of deleteUserCascade into shared deleteWorkspaceCascade /
deleteOrgCascade helpers (one tested implementation; deleteUserCascade now reuses
the purgeWorkspaces extraction).
Backend (platform-admin only): GET /api/admin/orgs (list + owner + counts +
workspaces), DELETE /api/admin/orgs/:id, DELETE /api/admin/workspaces/:id.
UI: an Organizations section in Admin listing every org/workspace with a
type-the-name confirmation before the irreversible delete.
Tests: org/workspace cascade (real FKs) + endpoint gating/404. Suite 53/53.
MSPs onboarding customers as separate orgs had no way to create one with
AUTO_CREATE_ORG_ON_SIGNUP=false (the only path was signup auto-org). Add a
platform-admin 'Create organization' action.
POST /api/admin/orgs (requirePlatformAdmin) creates the org + its first 'Default'
workspace. organizations.owner_user_id is NOT NULL, so an org can't be ownerless;
the creating admin becomes org_owner + workspace_admin (mirrors the signup
bootstrap in routes/auth.js) - which also surfaces the org in their switcher.
Customer users are then added via the existing Add User / manage-memberships flow.
UI: 'Create organization' button + single-field modal in the Admin area (gated).
Tests: create (201 + memberships + audit), empty-name 400, non-admin/operator 403.
Multi-zone videos/images were cropped: every template zone inherited fit_mode
'cover' (fill+crop) and the layout editor had no control to change it, so a
landscape video in a tall split zone showed only a center strip. The player
already honors fit_mode (web object-fit, Android scaleType) - the gap was the UI
and the default. Add a per-zone Fit selector (Contain/Cover/Stretch) to the layout
editor, and make 'contain' (show the whole frame) the default for new zones, the
schema column, and the save fallbacks. Existing built-in templates are migrated
separately.
Follow-up to the layout cache. On a cold start with a cached playlist but no cached
layout yet (first run after shipping, or cleared cache), the player still rendered
fullscreen and flashed before the payload arrived. Now gate the optimistic cached
render on the layout being KNOWN (cache key present — null=fullscreen vs object=
zoned, both fine); if unknown, wait ~1s for the payload to drive the first render.
Eliminates the fullscreen flash on the very first pass too.
The player cached only the playlist, not the layout. On cold start it restored the
playlist and rendered immediately with layout=null -> fullscreen, then re-rendered
into zones once the server payload arrived (the 'fullscreen first, then split'
flash). Cache the layout alongside the playlist and restore it before the first
render; cleared on reset.
A widget (e.g. directory board) assigned to a 'content' zone rendered as a black
zone: showZoneItem gated the widget branch on zone.zone_type==='widget', so the
widget was skipped and (mime_type null) nothing else matched either. Key off the
assignment's widget_id instead - matching the Android ZoneManager, which is why
the same layout worked on the APK but not the web player.
The render had no Cache-Control. A copy cached before the X-Frame-Options fix keeps
showing blank, and widget data (clock/weather/rss/directory) is dynamic anyway, so
mark the render no-store. Pairs with the X-Frame-Options removal.
The web player embeds widget/kiosk renders in a sandboxed (allow-scripts, no
allow-same-origin) iframe = a null origin. The global helmet X-Frame-Options:
SAMEORIGIN refuses that (null != same-origin), so every widget rendered blank in
the web player (video worked since it isn't an iframe). Drop X-Frame-Options on
just the /render endpoints - the sandbox, not X-Frame-Options, is what isolates
the widget from the dashboard (it still can't read the JWT). Dashboard keeps its
clickjacking protection. Verified: directory board now renders in a sandboxed
iframe with no refusal.
Mirror of the Android fix. The web player showed only the FIRST assignment per
zone (playlist.find) and an image zone set the GLOBAL advanceTimer->nextItem, so
the whole layout re-rendered on one global tick instead of each zone cycling its
own content. Now each zone groups its assignments (by zone_id, sorted), renders
the first, and advances on its OWN timer (images/widgets/youtube: duration;
videos: on end; single-item zones loop). Cleared in teardown. Also render zones
before the single-item 'renderable?' bail so an empty current item can't blank
the screen.
sw-admin.js (scope '/') intercepted every non-API GET with clone+cache+respond.
Video requests are Range requests -> 206 Partial Content, which can't be cached;
cache.put threw and the handler errored ('ServiceWorker encountered an unexpected
error'), so .mp4s never loaded on any page this SW controls - including the web
player at /player, which then thrashed between items.
Now bypass (network-only) non-GET, Range requests, and /uploads//player/api/
socket.io; only cache same-origin 200s. CACHE bumped to v4 so clients pick up the
new SW + drop the stale bucket.
- YouTube: load the embed via loadDataWithBaseURL with a youtube.com base URL so
the iframe has a valid origin/referer (a bare loadUrl of /embed/ID gives
'player misconfigured, Error 153'). Applies to zone + fullscreen YouTube.
- Web frames: shared WebViewSupport.configure() enables mixed-content (self-hosted
http LAN servers) and pipes WebView load/HTTP/JS-console errors to DebugLog, so a
failing web frame surfaces the real error in the live panel instead of a black
broken-page view.
After stopping the fullscreen controller in multi-zone, the only switch logs went
away - each zone now logs every item it renders (initial + each rotation) so the
live debug panel shows each zone advancing on its own interval.
From Chris's live debug logs on the L-Bar layout:
- ZoneManager only rendered the FIRST assignment per zone -> the Main zone (3
images) never rotated ('says it's switching but it's not'). Now each zone
cycles its own assignments: images/widgets on a duration timer, videos on
end (single-item zones still loop).
- The fullscreen PlaylistController kept running BEHIND the zones (playItem every
10s, would leak audio for a zone video) because startIfNeeded() ran after every
playlist update. Now only start it when not in multi-zone (zoneManager.hasZones).
- renderAssignments still called container.removeAllViews() (the same static-view
nuke the cleanup() fix addressed) -> now removes only its own zone views.
