Fix/content publishing media links

.github/workflows/fork-image.yml

@@ -0,0 +1,43 @@
+name: Fork staging image
+
+on:
+  workflow_dispatch:
+  push:
+    branches: [main, chore/mittwald-staging]
+
+permissions:
+  contents: read
+  packages: write
+
+concurrency:
+  group: fork-image-${{ github.ref }}
+  cancel-in-progress: false
+
+jobs:
+  image:
+    if: github.repository == 'Hubertoink/Instatic'
+    runs-on: ubuntu-latest
+    timeout-minutes: 30
+    steps:
+      - uses: actions/checkout@v4
+      - uses: docker/setup-buildx-action@v3
+      - uses: docker/login-action@v3
+        with:
+          registry: ghcr.io
+          username: ${{ github.actor }}
+          password: ${{ secrets.GITHUB_TOKEN }}
+      - name: Build and publish fork image
+        uses: docker/build-push-action@v6
+        with:
+          context: .
+          platforms: linux/amd64
+          push: true
+          tags: ghcr.io/hubertoink/instatic:staging-${{ github.sha }}
+          labels: |
+            org.opencontainers.image.source=https://github.com/Hubertoink/Instatic
+            org.opencontainers.image.url=https://github.com/Hubertoink/Instatic
+          build-args: |
+            INSTATIC_VERSION=staging-${{ github.sha }}
+            INSTATIC_REVISION=${{ github.sha }}
+          cache-from: type=gha
+          cache-to: type=gha,mode=max

docs/deployment/hochstaett-fork.md

@@ -0,0 +1,70 @@
+# Hochstaett deployment fork
+
+This fork preserves the locally tested CMS customizations used by the Hochstaett site.
+
+- Fork: https://github.com/Hubertoink/Instatic
+- Upstream: https://github.com/CoreBunch/Instatic
+- Initial integration branch: `feat/hochstaett-deployment-baseline`
+- Upstream base: `f92e8dc1` (0.0.20)
+
+## Source and runtime data
+
+Git contains application code, tests and documentation. Website content lives in
+the CMS database and uploads volume. Database files, uploaded media, backups,
+environment files and secret keys are not deployment source and must be
+transferred separately through a private backup/restore process.
+
+The integration branch includes relation/repeater loops, preview improvements,
+publication-date handling, animation presets, and content image sizing,
+lightbox and link insertion. Review these local changes through the fork's draft
+pull request before merging; the integration branch is not a production release.
+
+## Git remotes
+
+`origin` points to this fork; `upstream` points to CoreBunch. Feature branches
+and pull requests belong to the fork unless a contribution to CoreBunch is
+explicitly intended.
+
+```sh
+git clone https://github.com/Hubertoink/Instatic.git
+cd Instatic
+git remote add upstream https://github.com/CoreBunch/Instatic.git
+git switch --track origin/feat/hochstaett-deployment-baseline
+```
+
+Fetch upstream updates and integrate them on a separate review branch. Do not
+reset the fork to upstream when preserving these customizations.
+
+## Container handoff
+
+Build the image from the reviewed fork commit:
+
+```sh
+docker build -t instatic-hochstaett:staging .
+```
+
+The fork image workflow, `.github/workflows/fork-image.yml`, builds Linux amd64
+images on pushes to `main` and `chore/mittwald-staging`, and supports manual
+dispatch after merging. Images are tagged
+`ghcr.io/hubertoink/instatic:staging-<full-commit-sha>`. The Docker build runs
+TypeScript checking and the production frontend build. It does not run tests
+or lint; those remain separate verification steps.
+
+The upstream release workflow still targets `ghcr.io/corebunch/instatic`;
+do not create release tags to publish fork images. The fork workflow only
+publishes images and does not change a running Mittwald container.
+
+Mittwald needs an image available in a registry, persistent storage, HTTPS
+routing and the application environment described in [docker-image.md](docker-image.md).
+For a copy of the existing SQLite installation, preserve the data and uploads
+volumes and its secret encryption key using [backup-restore.md](backup-restore.md).
+
+Publishing inside a local CMS instance does not deploy that instance to Mittwald.
+An online staging instance publishes to its own staging site.
+
+## Validation at initial handoff
+
+The production Docker build (TypeScript and Vite), ESLint, targeted content tests,
+and desktop/mobile lightbox checks passed locally. The full Windows test run
+was not green; unrelated architecture, platform and UI failures remain for
+review. Consult the draft pull request for the handoff results.

docs/editor.md

@@ -8,6 +8,9 @@ The frontend is a single React 19 + Vite SPA mounted at `/admin`. Inside it, two
 
 ## TL;DR
 
+- **Entrance animations:** select a class, then **Styles → Effects → Animation** to choose Fade in, Slide up/down/from left/from right, or Zoom in. Duration, delay (milliseconds), and easing are editable. Choose Trigger: On load or On scroll into view. Scroll effects run once when the element intersects the viewport; the configured delay starts at that point. Design mode keeps scroll effects visible for editing; Live mode and published pages use the same IntersectionObserver runtime, including entries inserted by loops. Without JavaScript/Web Animations support, or with reduced motion, scroll content remains visible. None disables an inherited animation; Inherited clears the local value. Custom CSS preserves arbitrary imported animation shorthands. Preset keyframes are emitted by the shared canvas/publisher class-CSS generator only when used. Reduced-motion preferences keep content visible without entrance movement, including during delays.
+- **Local links:** the URL control accepts `/page#section`, `#section`, `./page`, `../page`, query references, and `tel:` as well as HTTP(S)/mailto links. Protocol-relative URLs and unsafe schemes remain rejected.
+
 - **Entry:** `src/admin/main.tsx` mounts `<Router><AdminRoutes /></Router><AdminContextMenuGuard />` with React 19 root-level error callbacks. `flushSync` forces the initial render synchronous to cut LCP.
 - **Router:** `src/admin/lib/routing/` — in-house router replacing `react-router-dom`. Ten workspace/page routes are wrapped in a per-route `<ErrorBoundary>` and `<Suspense>`, with root redirects plus a final `path="/admin/*"` catch-all redirecting unknown admin URLs to `/admin/dashboard` (login form when unauthenticated) instead of rendering an empty tree. Public-site 404s are NOT claimed — the publish pipeline's NotFound handling owns those.
 - **Cold path:** entry chunk is tiny. `AuthenticatedAdmin` is `React.lazy` and only loads post-login. Each workspace page is wrapped in `prewarmedLazy(...)`: the active page fires its import at module evaluation; the remaining pages pre-warm via `requestIdleCallback` after first paint so subsequent nav is synchronous (no Suspense flicker).
@@ -337,6 +340,7 @@ The store is composed of **12 slices**, each created by a factory in `store/slic
 | `selectionSlice`       | `selectedNodeId`, `hoveredNodeId`                                          |
 | `canvasSlice`          | Zoom, pan, `activeBreakpointId`, `activeConditionId`, `canvasMode` ('select'|'pan'|'insert'), `canvasView` ('design'|'live'), `runScripts` |
 | `uiSlice`              | Site editor panel visibility, unsaved-changes flag, insert picker, `componentizeEditorRequest` |
+| `previewSelectionSlice` | Session-only template and component preview source selections |
 | `classSlice`           | Style-rule CRUD, node ↔ class assignment, ambient selector creation         |
 | `filesSlice`           | `SiteFile` CRUD                                                            |
 | `visualComponentsSlice`| Visual Component CRUD                                                      |
@@ -390,6 +394,14 @@ Selectors are pure reads. Mutations go through actions (`useEditorStore.getState
 
 ### 1. Design mode and live mode
 
+Native `details`/`summary` accordions can be opened and closed by clicking their
+summary in either view, or with Enter/Space while the summary is focused.
+Selection still works normally. Selecting a layer inside a collapsed body
+reveals its ancestor accordions automatically. This state belongs only to the
+rendered editor frame: it does not change the page tree, collaboration document,
+undo history, or the published initial `open` attribute. Reloading the editor
+restores the authored state. Read-only composed regions are not toggled.
+
 `CanvasRoot` switches between two rendering surfaces based on `canvasView`:
 
 - **Design mode** (`canvasView === 'design'`): `CanvasRoot` → `CanvasTransformLayer` → `BreakpointFrame` → `IframeFrameSurface` → `NodeRenderer`. Each breakpoint gets its own iframe rendered side-by-side inside the pan/zoom transform layer. The author sees all breakpoints at once and can zoom in/out. The canvas opens at 50% (`INITIAL_ZOOM`) so several frames fit in view; reset (Cmd/Ctrl+0, the toolbar % button) goes to 100% (`RESET_ZOOM`).
@@ -788,3 +800,4 @@ See [docs/features/plugin-system.md](features/plugin-system.md) for the plugin S
   - `src/__tests__/architecture/canvas-aware-selectors.test.ts`
   - `src/__tests__/architecture/spotlight-no-direct-store-mutation.test.ts`
   - `src/__tests__/architecture/keybindings-registry-single-source.test.ts`
+

docs/features/agent.md

@@ -2,6 +2,10 @@
 
 The AI Agent is a model-powered assistant integrated into the Site editor and Content workspace. The shared Agent Panel owns conversation state, provider selection, streaming, history, and the browser bridge; each workspace supplies its own snapshot builder and tool executor.
 
+Relation and repeater catalogs are exposed through the existing tools. `site_list_loop_sources` preserves relation `targetTableSlug` and `allowMultiple`, media metadata, and repeater item fields with their binding tokens. An `entry.field` loop uses `<instatic-loop data-source-id="entry.field" data-field-id="teammembers" data-direction="asc">...</instatic-loop>` inside an entry context. Relation children bind the target table's fields; repeater children bind item fields. The HTML importer preserves the selected field and defaults contextual loops to authored order.
+
+`content_get_collection_schema` includes relation target table IDs/cardinality and repeater item schemas (including select option IDs). Content writes use bare row ID strings or `null` for single relations and string arrays for multi-relations; media uses the same shape with asset IDs. Repeaters use ordered `{ id, cells }` items. The Content bridge edits post-type entries; schema management and a Data-workspace toolset are not provided.
+
 In the Site editor, the agent reads the current page snapshot, plans a sequence of edits, and executes them by calling tools. Structure is written as semantic HTML (`site_insert_html` / `site_replace_node_html`); styling is written as CSS — a `<style>` block and/or `class=` attributes inside the insert, or the dedicated `site_apply_css` tool for authoring/editing any CSS on its own. There is one CSS path and it accepts every selector; `site_assign_class` / `site_remove_class` attach existing classes to nodes.
 
 In the Content workspace, the agent works against content collections and entries. It reads collection schemas and document state server-side, then mutates the live content editor through a browser bridge so the open draft, Tiptap body editor, and sidebar selection stay authoritative.

docs/features/content-workspace.md

@@ -12,6 +12,10 @@ The Content workspace renders a three-pane shell — explorer sidebar, document
 - **Body editor:** `TiptapBodyEditor` — one ProseMirror document, not a block list. Body persists as markdown text in the `body` cell.
 - **Inline marks:** bubble menu (B / I / code / strike / link). Block inserts: slash menu (`/`) + notch quick-actions.
 - **Canvas modes:** `write` (bare editor surface) and `live` (entry rendered inside its template with real site styles).
+- **Body images:** toolbar sizes Original (intrinsic width), L (content width), M (560px), S (320px), capped to the available width. Zoom uses the active button variant. Markdown titles persist size and lightbox metadata; stored `size=sm` maps to M.
+- **Link insertion:** the block-options menu opens `InsertLinkDialog` for link text and a safe target URL. Both link editors offer the media library (all file types) and Open/Download. Download links store the reserved Markdown title `instatic:download` and publish with the HTML `download` attribute; browsers support forced downloads for same-origin media.
+- **Text alignment:** paragraph and heading alignment persists through the Markdown bridge as restricted `<p align="center">…</p>` / `<h2 align="right">…</h2>` wrappers. Their contents remain Markdown; the publisher emits semantic elements with `instatic-content-align--*` classes backed by `publisher/reset.ts`, surviving rich-text sanitization. Alignment-only edits therefore mark the entry dirty and can be published.
+- **Lightbox runtime:** `src/core/imageLightboxRuntime.ts` opens a native modal dialog in a shadow root to isolate it from site styles. Fade/scale animation respects reduced motion; closing by Escape, backdrop or the high-contrast button restores scroll and focus without fragment navigation. The publisher includes the same-origin runtime for lightbox content and deferred fragments. Without JavaScript the link opens the image file.
 - **Settings panel:** `ContentSettingsPanel` — entry-specific; hidden when no entry is selected. Reopened via the top-right notch when collapsed.
 - **Hooks:** `useContentWorkspace` (CRUD + selection), `useContentEntryDraft` (field state + save/publish), `useContentMediaPicker` (media modal + featured media).
 - **AI assistant:** `ContentAgentMount` docks the shared Agent Panel in the content rail. `useContentToolBridge`, mounted by `ContentPage`, exposes the live workspace to both the built-in agent and scoped MCP relay even when the panel is closed, so writes mutate the open draft/editor state rather than stale database rows.
@@ -195,3 +199,17 @@ Body content is exchanged with the model as markdown. The browser bridge convert
   - `src/admin/pages/content/hooks/useContentEntryDraft.ts` — field draft state
   - `src/core/markdown/markdownDocument.ts` — markdown ↔ ProseMirror round-trip
   - `src/admin/layouts/AdminWorkspaceCanvasLayout/AdminWorkspaceCanvasLayout.tsx` — workspace shell + notch
+
+---
+
+# Post publication dates
+
+The seeded **Posts** collection includes an editable `date` field labelled
+**Publication date**. Existing installations add it through migration 031 and
+backfill existing posts and published versions from their stored publication
+timestamp. Publishing a post with an empty date fills the current calendar day;
+an author-entered date is preserved. Site loops can order posts by this value
+with `orderBy: "cell:date"`, which keeps imported or intentionally backdated
+posts in editorial order instead of the order in which they were migrated.
+Migration 032 also recognises the opening `**Veröffentlicht am DD.MM.YYYY.**`
+line used by archive imports and restores that original date automatically.

docs/features/data-workspace.md

@@ -131,6 +131,12 @@ the ordinary row draft. Nested relation and media fields reuse their shared
 pickers. Multi-media fields place `MediaPickerModal` in true multi-selection
 mode: plain clicks toggle assets and the footer commits the entire selection.
 
+Relation pickers show checkboxes for multiple selection, a selection count, and
+the target entry's display title and publication status. Confirm commits the
+selection; Cancel leaves the stored value unchanged. Relation editor fields
+show the selected names in relation order, with the full list in a tooltip when
+space is limited. Titles use the target table's configured primary field.
+
 When a repeater contains exactly one single-value media field,
 `MediaRepeaterGallery.tsx` replaces the generic structured-item cards with the
 Media workspace presentation. It renders the shared `AssetTile` / `AssetRow`

docs/features/html-import.md

@@ -96,7 +96,7 @@ Callers splice the fragment into the page tree via `insertImportedNodes(parentId
 | Selector | Module | Props set | Recurse |
 |---|---|---|---|
 | `instatic-outlet` | `base.outlet` | none (the CMS content outlet) | **No** |
-| `instatic-loop` | `base.loop` | `sourceId`, `filters.tableId`, `orderBy`, `direction`, `limit`, `offset`, `pagination`, `pageSize`, optional `tag` / `customTag` from `data-*` attrs | Yes |
+| `instatic-loop` | `base.loop` | `sourceId`, `filters.tableId`, `filters.fieldId`, `orderBy`, `direction`, `limit`, `offset`, `pagination`, `pageSize`, optional `tag` / `customTag` from `data-*` attrs | Yes |
 | `h1`–`h6`, `p`, `span`, `small`, `strong`, `em` | `base.text` | `text` = `el.textContent`, `tag` = tag name | No |
 | `a` with class `btn` | `base.button`, or `base.link` when it wraps element children | `label` (`text` on `base.link`) = `el.textContent`, `href`, `target` | No for text-only; yes when it wraps elements |
 | `a` (no `btn` class) | `base.link` | `text` = `el.textContent`, `href`, `target` | No for text-only; yes when it wraps elements |
@@ -117,7 +117,7 @@ Callers splice the fragment into the page tree via `insertImportedNodes(parentId
 **Key details:**
 
 - **`<instatic-outlet>` → `base.outlet`.** The custom element marks where matched content flows in a CMS template. It maps to a childless `base.outlet` node (any inner markup is ignored — the composer fills it). This rule lets the AI agent and hand-authored template HTML place the single content outlet inline via the normal import path. See [templates.md](templates.md) and [agent.md](agent.md).
-- **`<instatic-loop>` → `base.loop`.** The custom element lets the AI agent and hand-authored snippets create a real Loop through the same HTML import path. Children recurse normally and become loop variants. Loop configuration is read from attributes: `data-source-id`, `data-table-id` (stored as `filters.tableId`), `data-order-by`, `data-direction`, `data-limit`, `data-offset`, `data-pagination`, `data-page-size`, and optional `data-tag` / `data-custom-tag`. See [loops.md](loops.md) and [agent.md](agent.md).
+- **`<instatic-loop>` → `base.loop`.** The custom element lets the AI agent and hand-authored snippets create a real Loop through the same HTML import path. Children recurse normally and become loop variants. Loop configuration is read from attributes: `data-source-id`, `data-table-id` (stored as `filters.tableId`), `data-field-id` (stored as `filters.fieldId` for contextual relation/repeater loops), `data-order-by`, `data-direction`, `data-limit`, `data-offset`, `data-pagination`, `data-page-size`, and optional `data-tag` / `data-custom-tag`. Without an explicit direction, `entry.field` loops preserve authored order (`asc`); other sources default to `desc`. See [loops.md](loops.md) and [agent.md](agent.md).
 - `base.text` uses `tag` (not a separate `level` or heading prop) — the tag name is passed through directly. Imported bare DOM text uses `tag: 'none'`, which publishes text without an element wrapper.
 - **Direct text inside a recursing container is preserved.** The walker iterates `childNodes` (not just `children`): element children route through the rules, and each significant text node becomes a synthesized `base.text` child with `tag: 'none'` in document order. That no-wrapper text mode publishes back to bare text, so `<div class="num">98%</div>` and `<li>Buy milk</li>` import as containers holding their original text without adding selector-visible wrapper elements. Whitespace-only text (indentation between tags) is skipped; internal whitespace runs collapse to single spaces, and boundary spaces are kept when the text run sits between element siblings.
 - **`<body>` metadata is preserved separately.** Classes, safe HTML attributes (`id`, ARIA, `data-*`, etc.), and harvested inline styles on `<body>` are returned as `fragment.body` rather than inserted into `rootIds`. Full-site import applies them to `base.body`; paste-style HTML import can ignore them without changing the fragment structure.