- LLMClient : default_model scope par provider — tester nvidia alors que deepseek est le provider actif ne lui envoie plus 'deepseek-v4-flash' (cause du 404 'model not found' de NVIDIA)
- presets NVIDIA actualises (anciens modeles retires de la plateforme / premium)
- test de regression ajoute (46 passed)
- menu de mentions : template Alpine avec racine unique (les elements s'affichent maintenant sous chaque section; avant, seuls les titres de sections etaient rendus)
- menu des skills : style .focus ajoute (surlignage fond + barre d'accent) pour rendre la selection clavier visible
- /api/agent/mentions : reponse en sections (fichiers de l'espace courant ouvert, collections, pages de collections, autres documents), workspace_id transmis par le frontend (localWsId) et prioritaire
- Navigation clavier up/down : la liste defile pour garder la selection visible (menus @ et /)
- Deduplication des propositions (meme type + meme titre affiches une seule fois)
- Skills : plus d'injection du texte du skill dans la boite d'edition; requete composee en arriere-plan (contexte + skills + texte) a l'envoi; les skills integres injectent leur template, les enregistres partent via skill_ids
- Tests : shape des mentions, dedup, scoping workspace
- Les skills (menu /, integres et enregistres) s'epinglent comme des bulles dans le composer, au meme titre que les mentions @; plusieurs peuvent coexister sur un meme post
- L'API /run accepte skill_ids (tableau) et l'engine injecte tous leurs prompts dans le system prompt (skill_id conserve pour retro-compat)
- Clic sur une bulle de contexte (@ document/page/base) ouvre l'element comme document courant (/pages/<id> ou /db/<id>); le bouton x retire sans ouvrir
- Envoi possible avec des skills seuls sans texte
- test_engine_multiple_skills_applied
- L'engine envoie le message assistant avec ses tool_calls (id) et chaque resultat d'outil avec son tool_call_id, y compris en cas de refus; la passe suivante n'est plus refusee par l'API
- Un fournisseur reel configure qui echoue remonte maintenant une erreur (SSE error) au lieu de repeter la question via le mock hors-ligne (mock reserve a offline/sans cle)
- llm_client conserve id + arguments_raw des tool_calls
- test de regression test_engine_tool_protocol_messages
- Composeur contenteditable : mentions inline (pastilles icône+nom), barre de contexte groupée documents/skills, sélecteur @ (icône+titre+chemin), skills intégrés + agent_skills + commandes admin via '/'
- Actions sous chaque réponse agent : copier, insérer dans la page, feedback 👍/👎
- Backend : table agent_feedback, GET /api/agent/mentions, POST /api/agent/feedback, contexte document:<id>/page:<id>/collection:<id> résolu en contenu réel
- Inclut le travail v4.14.0 (documents/espace de travail + outils + auto-titre de conversation)
- Tableaux GFM rendus comme texte brut dans les exports HTML/PDF. Ajout d'un
parseur de tableaux pipe (bloc 'table') + rendu <table> (thead/tbody,
alignement gauche/centre/droite, bordures .ftable) dans blocks_to_html;
blocks_to_markdown reconstruit un tableau pipe valide.
- Emojis 'carrés noirs' en PDF: xhtml2pdf n'embarque que des polices de base
sans glyphes emoji. Moteur PDF -> WeasyPrint (tables CSS + emojis couleur via
pango + fonts-noto-color-emoji installes dans l'image). Repli automatique sur
xhtml2pdf quand weasyprint n'a pas ses libs natives (dev Windows).
- Dockerfile: libs weasyprint (pango/harfbuzz/gdk-pixbuf/shared-mime-info) +
fonts-dejavu-core + fonts-noto-color-emoji. requirements: + weasyprint==69.0.
- Verifie en reel sur README.md: HTML = <table class=ftable> (thead/th, center);
PDF 10 pages, texte de table present, Noto-Color-Emoji embarque + pixels
colores confirmes. 199/199 tests (4 nouveaux).
Le service d'export ne lisait que les pages content_format='blocks'. Les docs
stockees autrement sortaient avec le seul titre:
- content_format='file' (.md/code uploades): content=JSON meta, le vrai texte
est sur disque (/data/uploads/workspace_*) -> n'etait jamais lu.
- content_format='markdown': HTML/PDF enveloppait chaque ligne en <p> (headings
et listes aplatis).
Resolution de la vraie source pour les 3 formats (app/services/export.py):
- lit le fichier upload sur disque pour les pages file (repertoire via
FLOWDECK_DATA_DIR, defaut /data),
- rend les pages markdown / fichiers .md en blocs (headings, listes, code,
quote, todo) pour un HTML/PDF riche,
- fichiers texte non-markdown -> bloc de code,
- binaires (PDF/images) ignores.
195/195 tests (5 nouveaux v4.7.2). Verifie en reel via HTTP sur README.md et
l'arborescence Base de Connaissances (contenu complet dans les 3 formats).
- share-dialog/more-menu/activity-popover/move-dialog: enfants de .page-editor-wrapper
(overflow-y:auto) ancores en absolute top:100% -> invisibles sous le viewport.
Repositionnes en fixed sous la topbar, scope .page-editor-wrapper > (pas de
regression sur .more-menu de library/local_workspace) + variante mobile <768px.
- Item Export du menu More faisait moreOpen=false avant d'ouvrir exportOpen:
les 4 formats etaient inaccessibles. Devient un toggle, le menu reste ouvert.
- Cache busting app.css v=4.7.1; VERSION/main.py bump 4.7.1; CHANGELOG+ROADMAP a jour.
- Verifie Playwright headless: Share/More/Export/download Markdown/toast OK, 0 pageerror; 190/190 tests.
- Service serveur app/services/export.py: conversion blocs vers Markdown / HTML / PDF
- 4 endpoints: /api/export/markdown|html|pdf|site/{page_id}
- Export Markdown complet (tous les blocs + images + sous-pages recursives)
- Export PDF via xhtml2pdf (pur Python, aucune lib systeme)
- Export HTML standalone self-contained (styles inline)
- Export Site: zip multi-pages (index.html + une page par sous-page)
- UI: menu 'More > Export' dans l'editeur (Markdown, HTML, PDF, Site .zip)
- Tests: 190 passing (6 nouveaux)
loadSubdir() créait les icônes des enfants avec textContent au lieu de
innerHTML : le markup SVG (<svg width=...>) retourné par getSvgIcon() était
affiché comme texte brut quand on ouvrait un répertoire de la section Gitea.
Le rendu racine (loadSidebarTree) utilisait déjà innerHTML ; aligné.
Le /auth/login construisait le redirect_uri avec le schéma http:// en dur
et dupliquait l'expression dans login/callback : toute URL d'accès non
enregistrée (https, reverse proxy, hostname, port différent) était rejetée
par Gitea avec 'Unregistered Redirect URI' — les 2 liens 'Connect Gitea'
et 'Register with Gitea' de Workspaces → Gitea Projects étaient touchés.
- nouveau helper get_redirect_uri(request) : override explicite
OAUTH_REDIRECT_URI (si défini), sinon schéma depuis X-Forwarded-Proto
(fallback request scheme) + hôte depuis X-Forwarded-Host (fallback Host)
- redirect_uri stocké en session à l'authorize et réutilisé tel quel dans
l'échange de code (plus de dérive entre les deux étapes)
- même correctif dans GitHubProvider.exchange_code (ignorait le paramètre)
- config : oauth_redirect_uri par défaut vide (dynamique) au lieu de
localhost:8080 en dur
- .env.example documente OAUTH_REDIRECT_URI
- 4 tests de régression (host header, X-Forwarded-Proto/Host, override env,
URL d'authorize) — 178/179 OK, l'échec restant (test_views_calendar) est
pré-existant et dépend de la date
The save() function's blocks.map() only copied 7 properties, missing:
embed_type, collection_id, dbs, file_name, file_size, file_mime.
Embed blocks lost their collection reference on save → page reload
showed empty page instead of database table.
- Property modal: dynamically created per DBInstance (no global ID dependency)
- Column context menu: dynamically created (no global ID dependency)
- Side peek panel: auto-created on first open (no template dependency)
- All UI elements now work for both inline embeds and full-page views
- Cell editing + save use correct CSRF token via getCsrf()
Scripts in innerHTML don't execute per HTML spec. Instead, the render
function now scans for .block-embed-collection elements and calls
FlowDeckDB.renderInto() on each one via setTimeout after DOM update.
Clicking Database button now creates the table directly in the page
without asking for a name (uses 'Database' as default). Same pattern
as Notion: one click, inline embed, no popup.
The </script> tag inside a JS template literal was prematurely closing
the outer <script> block, causing 'Unexpected end of input' syntax error
and cascading Alpine.js 'not defined' errors.
- Page title placeholder: new pages created with empty title, CSS :empty::before shows 'New page' in gray
- Database is now inline embed within page (not full-page replacement)
- Embed block 'collection' renders database table via FlowDeckDB.renderInto()
- Open button in column 1 of each row opens page in side peek panel
- Side peek: iframe loading /pages/{id}?embed=1, close/fullscreen/resize
- DB scripts refactored as reusable DBInstance + global FlowDeckDB API
Replace delta-based approach with direct mouse-position-to-column-edge calculation:
newW = ev.clientX - thLeft. No setPointerCapture needed, no coordinate drift.
- Column resize: use getBoundingClientRect() + setPointerCapture for proper tracking
- Context menu 'Name' field: click to edit, Enter to save, Escape to cancel
- Rename persists to server via PUT /db/properties/{id}/api
- Column resize: drag handle between headers to resize column width
- SVG outline icons per property type in header (Aa, #, ☰, 📅, ☑, 🔗, ✉, 📞)
- Header context menu on click: Show page icon toggle, AI Autofill, Filter,
Sort (asc/desc), Group, Calculate (None/Count/Percent), Freeze, Wrap,
Insert left/right
- Submenus for AI Autofill, Sort, Calculate with nested options
- New Page button moved below last data row (bottom of table)
- New rows inserted at end of table
Alpine 3.14.9 evaluates x-data expressions before Alpine.data() registration,
causing 'pages is not defined', 'emptyRows is not defined', etc.
Solution: pure vanilla JS DOM rendering (same pattern as Library fix).
- _database_table.html: static DOM with IDs only, no Alpine directives
- _database_table_scripts.html: vanilla JS state, render, cell editing, modals
- Removes all x-data, x-for, x-show, x-model from database table
- Migration: add collection_id column to pages table
- POST /api/pages/{id}/convert-to-database: transforms page into database
- GET /api/collections/{id}/table-data: returns props + pages for table view
- POST /api/collections/{id}/pages: creates new row in collection
- New page_editor_collection.html template for database view
- _database_table.html: Notion-style table with columns, rows, New Page, Add Property
- Alpine.js component for cell editing, new page creation, property mgmt
- CSS: complete database table styling (view bar, table, cells, modals, dropdowns)
- Frontend: createDatabase() now converts page to full-page DB (was inline embed)
When dragging the resize handle rightward to narrow the panel, the
mouse cursor moved over the iframe which captured mouse events.
document-level mousemove stopped firing, making narrowing impossible.
Fix: replace mouse events with pointer events + setPointerCapture.
The handle element captures ALL pointer events during drag regardless
of where the cursor moves (over iframe, outside window, etc).
Applied to both local_workspace.html and library.html.
Now 3 shared modules (zero code duplication):
- _page_editor_content.html — editor HTML (title block, slash menu, toolbar etc)
- _page_editor_scripts.html — 100% original editorState() JS (CMDS, render,
save, slash, database, share/publish, favorites, formatting, etc)
- Used by page_editor.html (full page) and page_editor_embed.html (peek iframe)
Embed page_editor_embed.html now has the exact same editor as full page.
One change to _page_editor_scripts.html updates both views.
Root cause: dashboard.router registered BEFORE board.router in main.py,
so dashboard's /pages/{page_id} handler got all requests and ignored
the embed parameter. board.py's embed-aware route was never reached.
Fix: add embed support to dashboard.py's view_page_root too:
- Detect ?embed=1 query param
- Select page_editor_embed.html when embed=True
- Pass embed_mode to template context
The embed_mode variable set in board.py context wasn't reaching
base.html through Jinja2 extends. Added document.body.classList.add
in page_editor_embed.html content block to force the class, which
triggers the CSS rules hiding sidebar, topbar, and header.
- board.py: pass embed_mode=True to template context when ?embed=1
- base.html: <body class="embed-mode"> when embed_mode is true
- CSS: body.embed-mode hides .sidebar, .topbar, .unified-header,
.header-actions; removes app-layout margin/padding; adjusts editor
padding and max-width for peek panel
- Sidebar and header now properly hidden in iframe via CSS
Architecture for zero code duplication:
- Extract page editor HTML into _page_editor_content.html (shared include)
- page_editor.html includes it (same behavior as before)
- page_editor_embed.html includes it with empty topbar (no header in peek)
- board.py: ?embed=1 renders page_editor_embed.html
Peek panels now load /pages/{id}?embed=1 in iframe:
- Editor renders without sidebar or header breadcrumb/actions
- Full editing capability (blocks, markdown, slash commands, save)
- Same code — one change to _page_editor_content.html updates everywhere
Reverted the iframe-based peek panels (d13f137, 99fee56) which caused:
- Sidebar/header still visible inside iframe despite embed CSS
- Resize handle broken by iframe layout
Restored:
- Vanilla JS content loading via API (_loadPreviewContentVanilla)
- Peek-body div instead of iframe
- Resize handles on both library and local-workspace panels
(click=close, drag=resize, width persisted in localStorage)
- Hide .unified-header and .header-actions in embed-mode (not just .topbar)
- Hide .topbar-right specifically (Edited, Share, Copy link, Favorite, ...)
- Set .page-editor-wrapper padding-top to 12px (no header gap)
- Only the document editor/content is visible in the iframe now
The .peek-header class from library.html CSS was not available on the
local-workspace page, causing header buttons to stack vertically.
Added inline display:flex;align-items:center;gap:4px on the header div.
- Same 560px width, same header layout (close | icon+title | open | fav | copy | more)
- Header uses peek-header class from library CSS for identical styling
- Meta info (type, size, date) moved inline into body as subtle header
- Tags moved inline into body below meta
- More menu with Open, Favorite, Copy link, Close actions
- Copy link button uses navigator.clipboard API with toast feedback
- Bottom action buttons removed (now in header + more menu)
- Empty state matches library (file icon + 'Select a file to preview')
Root cause: blocks format content is a direct array [{id,type,content}],
not wrapped in {blocks:[...]}. Old code checked blocks.blocks which was
always undefined, so internal pages showed nothing.
Fixes:
- Handle blocks as raw array or {blocks:[...]} wrapper (both formats)
- Add heading styling (h3-h6 based on heading_1/2/3), code blocks (pre)
- Add proper markdown format handler (renders as pre-wrapped text)
- Add format=file handler showing file metadata + Open file link
- Add unknown format fallback as pre-wrapped text with char limit
- Clean up escaped quotes — no more double-escaped \" in strings
- Better empty state messages (Empty document, No content)
Alpine 3.14.9 x-if and x-show are fundamentally broken for reactive
visibility toggling of the preview panel. Complete rewrite:
- Replace all Alpine directives in preview panel with static HTML + IDs
- selectForPreview(): vanilla JS to populate fields and show panel
- closePreviewPanel(): vanilla JS to hide panel
- _loadPreviewContentVanilla(): async content loader using innerHTML
- Open button wired via onclick to navigate to /pages/{id}
- No Alpine reactivity dependencies - panel just works
- Keep outer x-show on preview-panel div for transition animation
- Wrap all inner content referencing previewItem in <template x-if=previewItem>
to prevent Alpine from evaluating expressions when previewItem is null
- Clean up debug console.logs from earlier debugging
- This combines x-show (reliable visibility toggle) with x-if (null guard)
Root cause: Alpine 3.14.9 x-if on template element doesn't reliably re-render
when a reactive property changes. previewItem was being set (confirmed by logs)
but x-if never re-evaluated.
Fix: replace <template x-if> wrapper with direct x-show on the preview-panel div.
x-show with x-transition is the reliable pattern for this Alpine version.
- Add console.log in selectForPreview and openSidePeek to trace button clicks
- Add console.log in Alpine OPEN button click handler
- Remove redundant x-show from preview-panel (was conflicting with x-if transition)
- This should fix the preview panel not appearing when Open button is clicked
- Add CSS rule .ws-tree-item:hover .hover-only for tree item hover visibility
- Move OPEN button OUT of .actions div, place it right after title span with
margin-left:auto and .hover-only class (matches Library layout exactly)
- Fix renderChildren() static HTML: button now between name span and .actions
- Remove duplicate button from inside .actions div
- Button calls selectForPreview(node) or openSidePeek(id) for side preview panel
- Add openSidePeek(id) helper method to find node by id and open in side preview
- Update renderChildren() static HTML OPEN button: replace full page nav with selectForPreview
- Add OPEN button in Alpine x-for tree template actions div for file items
- Uses existing preview-panel (right side panel) instead of navigating away
- 'Get started with' Database/Form buttons now POST to /db/inline/api
- Creates real collections linked to the page via parent_page_id
- Inline DBs rendered as embed blocks with links to collection views
- Persisted in page blocks JSON (embed_type='inline_dbs')
- Templates button opens the 'More' dropdown with all view types
- 175 tests pass
- Add is_task flag on collections (toggle-task API)
- New page_dependencies table with auto_shift config (blocks/blocked_by/related)
- API: GET/POST/DELETE page dependencies
- API: POST auto-shift dates based on blocking dependencies (skip_weekends option)
- 5 new tests (171 total)
- Dropdown Filter dans la toolbar (icône filtre, à côté de Sort)
- Menu avec Folders, Pages, PDF, Images, Code, Text + Clear filter
- doFilter() applique maintenant filterType au tree view via _filterTree()
- Filter chips visibles dans toutes les vues comme indicateur
- L'enhanced table prend le relais quand un filtre est actif dans les vues non-tree
143 tests passent
- dashboard.py: quote(filename) dans file_url (espaces/caractères spéciaux)
- page_editor.html: iframe PDF → onerror fallback avec lien Open in new tab
- 143 tests passent
Maintenant au bon endroit: base.html sidebar-section-actions (L229-233)
Pas dans local_workspace.html tree (où il n'était pas visible).
143 tests passent
Alpine ne résolvait pas wsId dans le scope x-for.
Maintenant: href='/local-workspace' (route qui utilise le cookie ws actif).
+ hard refresh nécessaire (Ctrl+Shift+R) pour vider cache ancien JS.
143 tests passent
_header.html L54: {{ page_icon }} affichait le mot 'home'/'file'/'paperclip'
au lieu de l'icône SVG. Maintenant détecte les noms connus → fd_icon().
Fallback pour les valeurs non-reconnues: affiche tel quel.
143 tests passent
page_editor.html L5: le bouton star utilisait x-text avec fd_icon('star',14)
contenant des guillemets doubles SVG → Expression Error.
→ x-html + getSvgIcon('star',14) (JS, pas de guillemets dans l'expression)
→ share/lock/link: utilisent déjà la concaténation ~ (safe)
143 tests passent
bruno