#74 — Support PDF complet:
- backend/pdf_reader.py — pymupdf/pypdf text extraction
- .pdf added to SUPPORTED_EXTENSIONS in indexer.py
- PDF-aware file view API endpoint (metadata, text preview)
- Streaming endpoint /api/file/{vault}/pdf/stream
- Frontend PDF viewer with iframe, toolbar, download
- CSS for PDF viewer container
#75 — Split View (PaneManager):
- New module frontend/js/pane-manager.js — ES module
- 1-4 pane CSS grid with resize handles
- Split right/down via keyboard (Ctrl+Alt+\) or PaneManager API
- Persistence in localStorage (obsigate-panes)
- Collapse to single pane, per-pane active state
285 tests passent
Problem: language preference stored only in localStorage, lost on
'clear site data'. French users with browser lang=fr always got French
UI even when profile was English.
Solution:
- Backend: add 'language' field to user profile (default 'fr')
- Backend: PATCH /api/auth/me to update language preference
- Frontend: initI18n() priority: server > localStorage > navigator > 'fr'
- Frontend: initAuth() runs before initI18n() so server check works
- Frontend: profile Save persists language to server via PATCH
Verified: tabs show English after server profile set to 'en'
Add Ollama autocomplete status endpoint to AI routes, update sync panel
UI to display autocomplete state, replace monolithic test compose files
with separate Linux/Windows variants, and add PowerShell build script
for Windows deployment
- Backend: delete_cookie access_token (/) + refresh_token
- Logout ne requiert plus auth (graceful si token expire)
- Bouton Deconnexion toujours visible dans le menu
- Ollama provider ajoute dans ai.py (OpenAI-compatible API)
- Service ollama dans docker-compose (auto-pull qwen2.5-coder:1.5b)
- Ghost overlay dans Forge: prediction en texte pale apres le curseur
- Tab pour accepter, Escape pour rejeter, taper pour annuler
- Declenchement automatique apres 800ms de pause
- Scroll sync entre overlay et textarea
- Env vars: AI_DEFAULT_PROVIDER, OLLAMA_BASE_URL, OLLAMA_MODEL
- CSP script-src: ajout https://static.cloudflareinsights.com
- legacy.js ligne 98 a deja le null-check if(filterBtn)
- L'erreur production vient d'un ancien deploy, sera resolu au prochain rebuild
Add standalone /editor-poc page demonstrating all 5 toolbar zones:
- Zone 1: Top Fixed Bar (hamburger, title, AI glow button, save indicator)
- Zone 2a: Slash Command Menu (/ on empty line, 15 commands, filtering)
- Zone 2b: Bubble Menu (on text selection, formatting + AI dropdown)
- Zone 3: Floating Quick Insert (per-line, Image/PDF/Code/Table/AI)
- Zone 4: AI Side Panel (Ctrl+J toggle, chat, suggestions)
- Zone 5: Drag & Drop overlay with dashed border
Files:
- new: frontend/editor-poc.html (46KB standalone vanilla JS)
- backend/main.py: +/editor-poc route before catch-all
- frontend/index.html: Editor POC button in header menu
- frontend/js/config.js: initEditorPocBtn() function + export
- frontend/js/legacy.js: re-export initEditorPocBtn
- frontend/js/app.js + frontend/app.js: init call in both versions
#15: Click backup item to preview content in side panel (100KB limit)
#16: Purge all backups per vault (red 'Vider' button)
#17: /api/backups/compress — gzip backups older than N days
#18: /api/backups/auto — backup files modified since N hours
#19: Restore button per backup in manager with timestamp extraction
+ Professional side-panel preview, action buttons on hover, vault grouping
- Config: max_backups_per_file (default 10) in config.json + UI field
- Auto-cleanup: _backup_file deletes oldest backups when limit exceeded
- New endpoints: GET /api/backups (list all), POST /api/backups/delete,
POST /api/backups/purge (by file or vault)
- New module: backup-manager.js - full-page view with filter, grouped by
vault/file, individual delete, per-file purge, stats
- Link from backup viewer footer: 'Gerer tous les backups'
- _get_backup_dir: now resolves relative path using the SPECIFIC vault's
directory (matching _backup_file exactly), not the first vault in index
- Removed _get_backup_root (unused after refactor)
- backups.js: global document click handler for #backup-close (same
pattern as Graph View's initGraphView)
- _backup_file: resolve relative backup path against vault parent (survives CWD changes)
- _get_backup_root: same resolution logic for listing
- backups.js: global document-level capture-phase click handler for .backup-close-btn
(works regardless of innerHTML replacements, text node targets, or event bubbling)
- Vault home (#12): recursive file listing (rglob), rel_dir in metadata,
'recursif' badge, updated roadmap description
- Backup viewer: fix close button (inline onclick + event delegation
fallback for text nodes), improve error messages
- Backup API: remove response_model to prevent Pydantic 500 errors,
add try/except + logging on backups/diff/restore endpoints,
per-entry error handling in _list_backup_files,
encoding safety (errors='replace') on read_text
- AI status: call /api/ai/status at startup in _initAIStatus
instead of waiting for user click on sync badge
Problème: l'indicateur affichait 'IA non configurée' alors que les
actions AI fonctionnaient.
Causes:
1. Deux endpoints /api/ai/status en conflit (main.py + ai_routes.py)
2. ai_routes.py lisait PROVIDERS évalué à l'import (os.getenv figé)
3. Cache côté frontend (_aiStatusChecked) empêchait de re-vérifier
Fixes:
- Suppression de l'endpoint redondant dans main.py
- ai_routes.py: lecture os.getenv au runtime (pas depuis PROVIDERS)
- sync.js: _checkAIStatus() appelée à chaque ouverture du panneau
L'étape 'Building inverted index...' au démarrage bloquait l'event loop
asyncio, empêchant le serveur de répondre aux requêtes HTTP (navigation,
refresh) jusqu'à la fin de l'indexation.
- init_inverted_index() exécuté dans ThreadPoolExecutor via run_in_executor
- get_inverted_index() ne fait plus de rebuild auto bloquant
- Le serveur répond immédiatement, même pendant l'indexation
Point 11a — Indicateur AI Actif:
- Badge sync flashe en blanc pendant traitement AI (animation CSS)
- Événements ai-processing-start/end dans ai.js
- Section statut IA dans panneau sync (configurée/non configurée/en cours)
- Endpoint GET /api/ai/status vérifiant les clés API configurées
Point 11b — Page d'accueil de voute:
- Endpoint GET /api/vault/{vault}/files?dir=... (trié par mtime)
- Module vaulthome.js: affichage style résultats de recherche
- Breadcrumb navigable, mise à jour auto sur clic dossier dans tree
- Déclenché auto quand un vault spécifique est sélectionné
- Backend: 3 nouveaux endpoints
- GET /api/file/{vault}/backups — liste les backups horodatés
- GET /api/file/{vault}/diff — diff unifié entre versions
- POST /api/file/{vault}/restore — restaure une version (backup auto avant)
- Frontend: module backups.js — UI modale avec:
- Sélecteur gauche (backup) / droite (version actuelle ou autre backup)
- Diff coloré: vert pour ajouts, rouge pour suppressions
- Bouton Restaurer avec confirmation
- Menu contextuel: clic droit fichier → 'Backups / Versions'
- CSS: styles complets pour la modale, toolbar, table de diff
- ROADMAP: point 6 marqué comme complété
- Backend: nouvel endpoint POST /api/move/{vault_name} pour déplacer
fichiers et dossiers vers un autre répertoire parent
- Frontend: module dragdrop.js avec délégation d'événements sur le
vault-tree (dragstart, dragover, dragleave, drop)
- CSS: styles pour drag-source, drag-valid-target, drag-over
- Protections: pas de drop sur soi-même, pas dans un enfant,
pas de cross-vault
- Mise à jour de l'index et broadcast SSE après move
- ROADMAP: portion 5 marquée comme complétée
Replace double-nested stem loop (stems × tokens) with single-pass
stem frequency map. For 100 unique tokens per file: 10,000 iterations
→ 100 iterations per file. Critical for large vaults.
Backend (main.py):
- GraphNode: added tags, incoming_count, outgoing_count
- GraphEdge: added 'backlink' relation
- GraphResponse: added 'scope' field
- api_graph: scope=full|directory, tag= filter, backlinks
- Full-vault tree walk with configurable depth 0-3
- Tag index from in-memory file index for fast filtering
- Incoming/outgoing link count per node
Frontend (graph.js + index.html):
- Theme-adaptive colors via CSS custom properties
- Depth slider (0-3) with live reload
- Full-vault toggle button (🌐 Tout / 📁 Dossier)
- Search input with tag filtering + visual highlighting
- Tooltip on hover (name, path, tags, link counts)
- Backlink edges rendered in red dashed
- Node size proportional to link count
- Larger modal (1000px, 85vh)
Add French snowball stemmer to tokenization pipeline:
- Index both original tokens AND their stems in InvertedIndex
- Query terms are also stemmed before lookup
- Stemmed forms accumulate TF from all original forms
- Lazy-init singleton pattern for stemmer
The login() function used 'request: LoginRequest' which shadowed
FastAPI's Starlette Request object. Request.client was accessed on
the LoginRequest Pydantic model instead of the HTTP request, causing
AttributeError: 'LoginRequest' object has no attribute 'client'.
Fix: rename the Pydantic parameter to 'body' and add explicit
'request: Request' for IP extraction and rate limiting.
- Add .dockerignore to exclude .git, __pycache__, docs, etc. from Docker context
- Create .env.example template with documented env vars
- Move OBSIGATE_ADMIN_PASSWORD from docker-compose.yml to env_file: .env
- Add .env.* to .gitignore (excluding .env.example)
- Enable GZipMiddleware for ~70% bandwidth reduction on text responses
- Add Cache-Control: immutable for /static/ assets
- Update ROADMAP: mark all 4 quick wins as done, add audit findings
- Add comprehensive technical audit report (AUDIT_TECHNIQUE_2026-05-27.md)
Remove the old HTML-based PDF download endpoint in favor of the new
WeasyPrint-based one, and replace the generic "Télécharger" button
in popout.html with a dedicated .md download and a new PDF button.
Also remove the unused generic download button from the main file view.
bruno