Dash/docs/specs/2026-09-16-dash-design.md

Dash — Design Spec & Roadmap

Date: 2026-09-16 Status: proposed — pending UI direction pick (see mockups/)

1. Product

Self-hosted homelab dashboard. A visual launchpad for services (local + external) with live widgets. CasaOS-grade ease of use, Vercel-grade design. Single Docker container, no auth in v1.

Differentiator: UI-first management (no YAML) + modern design. Homepage requires YAML, Homarr forces auth and complexity, Dashy is overloaded, CasaOS is an OS not a dashboard.

2. Requirements (from prompt.md)

Req	Notes
Service cards	name, icon (URL or uploaded file), status indicator
Multi-URL services	each URL tagged local / external / custom label; 1 URL = direct nav, 2+ = popover chooser
Sections	named groups; collapse/expand; reorder sections and items via drag-drop; items move between sections
Add service	+ button, dialog: name, URLs, icon
Widgets	live on the grid; v1: clock/timezones, Pi-hole, AdGuard Home, Immich; extensible registry
Themes	dark + light
No auth	v1; optional auth is post-1.0
Lightweight	single container, single binary, <50 MB image target

3. Architecture

/apps/frontend   React 18 + Vite + TS strict + Tailwind + shadcn/ui + dnd-kit
/apps/backend    Go + Gin + zap, CGO-free
/packages/api-client   generated from OpenAPI (single source of truth)
/infra           Dockerfile multi-stage → ONE image; docker-compose.yml (1 service)
/data            SQLite file + uploaded icons (bind-mounted)

Approved deviations from locked stack

• SQLite (modernc.org/sqlite, CGO-free) instead of Postgres. Approved 2026-09-16. Reason: ~100 rows of state, single-binary deployment is the product's core install feature. goose migrations still apply.
• No DragonflyDB. Widget responses cached in-process with a small TTL map. Add a real cache only if profiling demands it.

Everything else per tdvorak-fullstack locked stack.

Key decisions

• Go embeds the frontend via embed.FS → docker run -v ./data:/data -p 3000:3000 dash is the entire install.
• Widget proxy server-side. Pi-hole/AdGuard/etc. APIs need secrets and are not CORS-safe. Backend exposes GET /api/widgets/:id/data; a registry maps widget type → fetcher. New integration = one Go file + one React component.
• Ordering via fractional position column (position REAL) — drag-drop writes one row, not a renumber of the set.
• Icons stored in data/icons/; dashboard-icons pack integration for auto-suggest in Phase 3.

4. Data model (SQLite)

sections(id TEXT PK, name TEXT, position REAL, collapsed INT, created_at)
items(id TEXT PK, section_id FK, kind TEXT,      -- 'service' | 'widget'
      name TEXT, icon TEXT,                      -- icon: path or URL
      position REAL, config JSON, created_at)    -- widget cfg or null
urls(id TEXT PK, item_id FK, url TEXT, label TEXT, position REAL)
settings(key TEXT PK, value JSON)                -- theme, wallpaper, widget secrets ref

Widget secrets (API keys) live in config JSON for v1; encrypted-at-rest only if auth lands later.

5. API (OpenAPI contract — sketch)

GET/POST/PATCH/DELETE  /api/sections
POST                   /api/sections/reorder
GET/POST/PATCH/DELETE  /api/items
POST                   /api/items/reorder
POST                   /api/items/:id/icon       (upload)
GET                    /api/widgets/types        (registry metadata for UI)
GET                    /api/widgets/:id/data     (live fetch, cached ~30s)
GET/PUT                /api/settings
GET                    /api/export               (full JSON backup)
POST                   /api/import
GET                    /api/healthz

Frontend never hand-writes API types — generated client from openapi.yaml.

6. Frontend structure

• ServiceCard — icon, name, host, status dot, multi-URL badge → UrlPopover
• Section — header (label, count, collapse chevron, drag handle) + SortableContext grid
• WidgetHost — renders widget component by config.type, polls /api/widgets/:id/data
• AddServiceDialog, EditServiceDialog (shadcn Dialog + Form)
• CommandPalette (⌘K) — Phase 3
• dnd-kit DndContext at board level; sections and items both sortable

7. UI system — DECIDED 2026-09-16

The app ships multiple layouts, not one. Same board state, same interactions (drag-drop, collapse, URL popover, add/edit), different structure per style. This is a core feature, not a theme afterthought.

Primary renderer: d-bento-mono.html — bento grid on the monochrome zinc palette. Geist + Geist Mono, hairline borders, no glow, status chips carry the only color (green/amber semantic).

Full renderer set:

Renderer	Source mockup	Structure
Bento (default)	d-bento-mono.html	mixed-size grid, widgets inline as large tiles
Cards	a-vercel-mono.html	uniform compact card grid, widgets as cards
Index	e-editorial.html	serif section headers, full-width rows, masthead stat strip
Console	f-console.html	bordered panels/tables, JetBrains Mono, [ up ] tags

Rejected and deleted: bento-with-glow (B), casa glass (C).

Renderer contract (architectural requirement)

renderers/
  <name>/
    BoardView      (grid/list arrangement, section placement)
    SectionView    (header chrome: label, count, collapse, drag handle)
    ItemView       (tile/card/row anatomy: icon, name, host, status, popover anchor)
    WidgetView     (widget tile/panel/strip-cell anatomy)

• Board state (sections, items, order, collapse) lives in one store, renderer-agnostic.
• Interactions (dnd-kit, popover, dialogs) are provided by shared primitives the renderer composes — renderers never reimplement dnd or popover logic.
• Design tokens (--bg, --surface, --border, --text, …) are shared CSS vars; each renderer maps them to its own chrome. Dark/light stays orthogonal to renderer choice.
• Setting: settings.renderer — user picks layout in settings; persisted like theme.

Building the second renderer (Cards) inside Phase 1 is deliberate: it validates the seam while the code is still small.

8. Roadmap

Canonical roadmap lives in ROADMAP.md at repo root — single source of truth, kept versioned with the code.

Explicit non-goals for v1

Auth, Postgres, multi-user, mobile app, container management, YAML-as-source-of-truth.

9. Testing

• Go: table-driven tests for widget fetchers + reorder logic (go test ./...)
• TS: tsc --noEmit + oxlint gate; dnd reducer unit tests only if ordering logic grows
• Smoke: docker compose up → create section → add service → drag → export JSON