- Create app/core/worker.py with WorkerSettings for ARQ
- Register all job functions: reindex, index_*, embedding_batch, deep_analysis
- Update prestart.sh to start ARQ worker in background before uvicorn
- Worker runs with max_jobs=10, job_timeout=300s
- Ollama Cloud has no embedding endpoint, OpenRouter does
- text-embedding-3-small with dimensions=768 matches DB vector(768) column
- API_KEY_OPENROUTER env var is primary, DB provider is fallback
- Added dimensions parameter for text-embedding-3 models
- mail.ts: add explicit types to decodeMimeHeader callback params
- SessionList.tsx: use React.MouseEvent instead of nativeEvent
- ComposeModal.tsx: construct name from first_name + last_name
- MailDetail.tsx: destructure onReply/onForward from props
- RichTextEditor.tsx: use SetContentOptions object instead of boolean
- MailSearchBar.tsx: add optional value prop to interface
- EmptyState.tsx: add optional children prop
- Badge.tsx: add secondary variant to BadgeVariant, variantClasses, dotColors
- ConfirmDialog.tsx: add optional children prop
- hooks.ts: extend SearchResult type to include mail/file/event
- Remove duplicate vi.mock(@/api/hooks) in SettingsRoles.test.tsx that overrode forceUpdate logic
- SettingsRoles mock: use vi.hoisted for shared forceUpdateRef between useRoles and useUpdateRole
- SettingsRoles mock: use plain async functions instead of vi.fn().mockImplementation for mutations
- query_understanding.py: get API key/base_url/provider_type from ai_providers DB
- embedding.py: get API key from DB, pass db+tenant_id through call chain
- routes.py: pass db+tenant_id to llm_analyze_query and llm_aggregate_results
- search_engine.py: pass db+tenant_id to generate_embedding
- unified_search/jobs.py: pass db+tenant_id to generate_embedding
- Fix all default model names: ollama/deepseek-v4 -> ollama/deepseek-v4-flash
- Ollama Cloud has no embedding endpoint; embedding calls fail gracefully
- _get_llm_api_key now returns provider_type
- Model string built with provider_type prefix (openai/deepseek-v4 instead of ollama/deepseek-v4)
- api_base set from provider.base_url (https://ollama.com/v1)
- Fixes: litellm tried localhost:11434 instead of Ollama Cloud
- ai_proactive now reads API key from ai_providers table (like ai_assistant)
- Falls back to API_KEY_OLLAMA_CLOUD env var if no provider found
- Fixes: proactive engine could not generate suggestions because API key was empty
- AISidebar: full height, right of TopBar and Toolbar (all pages except AI Assistant)
- AISidebar: collapsed mode shows narrow 48px icon strip with KI button
- AISidebar: expanded mode uses ResizablePanel with drag-to-resize (240-600px)
- AppShell: moved AISidebar outside inner content container to full-height flex
- uiStore: renamed aiSidebarOpen→aiSidebarCollapsed (default: true=collapsed)
- TopBar: AI button highlights when sidebar is expanded
- MailFolderTree: reduce padding/gap ~25% (py-2→py-1, gap-2→gap-1.5)
- MailList: text-sm→text-xs for from and subject lines
- ResizablePanel: use DOM ref for width during drag instead of React state
- No re-renders during drag, only single setWidth on mouseup
- panelRef on outer div, contentRef on inner div
- pointer-events:none on content during drag
- contain:strict on content area
- Add imap_uid column + unique index for proper deduplication
- Use deterministic message_id for emails without Message-ID header
- Dedup by (account_id, folder_id, imap_uid) first, message_id fallback
- Track folder moves: update folder_id when email appears in different folder
- Fix Sent folder lookup: check Sent/INBOX.Sent/Sent Items/Sent Mail + ILIKE
- Detect IMAP delimiter from LIST response instead of hardcoding dot
- Use detected delimiter in stale folder name migration
- Migration 0004_imap_uid.sql with imap_uid column and unique index
Agent Zero