The /chat/send SSE stream was building a stripped-down system prompt
without any Frankenbot command documentation (@CREATE_SUBTASK, @SEND_EMAIL,
@SEND_TELEGRAM etc.), so agents never knew to emit these commands during
web chat sessions — tasks were never created, emails never sent.
Extract build_agent_prompt() as shared function used by both
execute_agent_task() and the streaming chat generator.
- @SEND_TELEGRAM: fix regex To: -> TelegramID:, fix DB lookup telegram_chat_id -> telegram_id,
use send_telegram_message() (sync HTTP) instead of asyncio.run(), support direct numeric IDs
- @ADD_TEAM_MEMBER: fix field order in regex (Name/Role/Responsibilities/Email) to match prompt
- System prompt: document all Claude Code built-in tools (WebFetch, Read, Write, Edit,
Glob, Grep, Bash) so agents know what is available, clean up command documentation
- Add sent_emails table to DB for persistent outbox logging
- send_email() now writes every outgoing mail (incl. errors) to sent_emails
- parse_agent_commands() passes agent_key/task_id as triggered_by metadata
- Fix @UPDATE_TEAM_MEMBER parser: now matches Identifier/TelegramID/Role/etc.
format from system prompt (was expecting Email/Field/Value — never matched)
- update_team_member() called correctly via **kwargs (was positional args bug)
- Set Piotr telegram_id=1578034974 directly in DB
- email_log.html: two-tab UI (Inbox Journal + Outbox), click-to-expand body
- emails.html: per-message delete button in inbox list
- New routes: DELETE inbox (IMAP expunge), journal entry, sent entry
- Wire up Telegram bot with token, allowed users and username from .env
- Fix TaskBeat to handle direct tasks (Telegram/email) without sub_tasks
- Fix send_telegram_message to use run_coroutine_threadsafe (avoid event loop clash)
- Add TaskBeat watchdog thread for auto-restart on crash
- Reset stuck in_progress tasks on startup
- Add task detail page (/tasks/<id>) with full response/log view and auto-refresh
- Add task delete route (/tasks/delete/<id>) with confirmation
- Include agent sender info in Telegram task prompts
- Orchestrator self-updated knowledge base with Telegram contact info
- App-level password auth via Flask session (APP_PASSWORD in .env)
- login_required decorator on all routes
- Login page, logout button in navbar, 7-day session lifetime
- Upgrade musik_rechte_advisor and negotiator from Opus 4.0 to Opus 4.6
- Fix orchestrator session cookie overflow (kb/prompts no longer stored in session)
- Change app port from 5000 to 5050 (5000 occupied by Zou/Kitsu)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Add task creation notifications in Orchestrator Chat
- New messages appear at top (newest first)
- Add Clear Chat button with confirmation
- Improve timestamps: DD.MM.YYYY HH:MM:SS format
- Support for system notifications (task created, sub-agent responses)
- Add add_orchestrator_notification() helper function
- Auto-notify when agents create sub-tasks
- Clear Chat route: POST /orchestrator/clear
- Better visual distinction between messages and notifications
- Move knowledge base from root to agents/orchestrator/knowledge/
- Move orphaned files (AKM, Event-Plan) to orchestrator/work/
- Update all kb_file paths to new location
- Change orchestrator stream to use work_dir instead of root
- Now ALL agents (including orchestrator) work in their own folders
- No more files created in project root
- Consistent work directory behavior across all agents
- Add global AGENTS reload after agent deletion
- Ensures agent immediately disappears from all dropdowns/UIs
- Matches agent creation behavior (which already reloads)
- Fixes issue where deleted agents still appeared in chat/task forms
- Replace blocking execute_agent_task() with live subprocess streaming
- Use Popen() to read opencode output line-by-line in real-time
- Send 'chunk' events to frontend as agent thinks
- Frontend appends chunks incrementally for live response
- Matches Orchestrator's streaming UX
- No more waiting for complete response before seeing output
Problem:
- Task #1 was stuck in 'in_progress' status
- TaskBeat expects 'sub_tasks' array for orchestrated tasks
- Tasks created via old API didn't have sub_tasks defined
- TaskBeat only processes 'pending' tasks, so in_progress tasks stay stuck
Solution:
- Added validation: if orchestrated task has no sub_tasks, mark as completed with error message
- Prevents infinite loop with stuck tasks
- Manually completed Task #1 in database
Note: This handles legacy tasks from deprecated /api/orchestrate endpoint.
UX Improvements:
- Tasks page: Remove 'Create new task' form (orchestrator handles all)
- Orchestrator: Simplified to single 'Prompt' section, removed 'Tasks verteilen'
- Orchestrator: Removed 'Klassisch senden' button (live-only responses)
- Orchestrator: Removed 'Aktive Agenten' display (redundant with dashboard)
Critical Fix:
- TaskBeat now reads pending tasks from DATABASE instead of in-memory array
- All status updates (pending→in_progress→completed) now persist to DB
- Fixes issue where tasks created via API/UI were not being processed
- Agent_key updates also synced to DB
This fixes the bug where Task #1 was stuck in pending status.
Changed all 'files_page' to 'files' in redirect calls:
- delete_agent_file() - 4 occurrences
- delete_project_file() - 4 occurrences
Fixes Internal Server Error (500) when deleting files.
Error was: BuildError: Could not build url for endpoint 'files_page'
Corrected team member roles and responsibilities:
- Eric Fischer: Programmer → Funktionserweiterung
- Georg Tschare: CEO → Personalmanagement, Kundenkommunikation, Kommunikation mit Behörden
- Piotr Dyderski: Tech, 3D Art, RnD → Technische Infrastruktur, AI-Agenten, Automatisierung, 3D Avatare, Research and Development
DB data is the source of truth, code now matches.
- Edit modal with Bootstrap for inline editing
- Update name, email, role, responsibilities, telegram_id
- Form validation and error handling
- Success/error flash messages
- Team member data passed to frontend via JSON
- New /team route to view all team members
- Add/activate/deactivate team members via UI
- Display name, role, email, responsibilities, telegram_id
- Integrated into main navigation
- Self-learning system info displayed
- Add command parsing for @SEND_EMAIL, @SEND_TELEGRAM, @UPDATE_TEAM_MEMBER, @ADD_TEAM_MEMBER
- Migrate all tasks.append() calls (13 occurrences) to use create_task() for DB persistence
- Update task routes to read from database instead of in-memory array
- Orchestrator beat now executes parsed commands (email/telegram notifications)
- Maintain legacy task_queue compatibility for email processing
- All tasks now persist across app restarts
Team-Member Updates:
- Korrigierte Rollen und Verantwortlichkeiten:
* Eric Fischer: Lead Programming & Tech
* Georg Tschare: CEO & Kundenkommunikation
* Piotr Dyderski: SysAdmin, TechLead, 3D Artist & R&D
Neue Funktionen:
- update_team_member() - Aktualisiert Team-Member-Daten per Email/Name
- is_known_team_member() - Prüft ob Email in Team-DB existiert
- @UPDATE_TEAM_MEMBER Kommando für Orchestrator
- @ADD_TEAM_MEMBER Kommando für Orchestrator
Intelligentes Lernsystem:
- Erkennt neue Email-Absender automatisch
- Leitet unbekannte Absender automatisch an Orchestrator
- Orchestrator bittet um Vorstellung (Name, Rolle, Aufgaben)
- Informationen werden in Team-DB gespeichert
- Verbessert zukünftige Koordination
Email-Workflow für neue Absender:
1. Email von unbekanntem @diversityball.at kommt an
2. System erkennt: "Nicht in Team-DB"
3. Orchestrator übernimmt automatisch
4. Fragt freundlich nach Vorstellung
5. Beantwortet auch die eigentliche Anfrage
6. Speichert Infos mit @ADD_TEAM_MEMBER
Vorteile:
✓ Orchestrator baut automatisch Wissensdatenbank auf
✓ Bessere Koordination durch bekannte Verantwortlichkeiten
✓ Keine manuellen Team-Member Einträge nötig
✓ Freundlicher, professioneller Erstkontakt
✓ Lernfähiges System das besser wird
Context-Weiterleitung:
- extra_context aus Tasks wird jetzt durchgereicht
- Kombiniert Email-Kontext + Team-Member-Status
- Agents haben volle Informationen über Absender
Security Fixes:
- Fix XSS vulnerability in orchestrator.html (escapeHtml für user input)
- Verbesserte Error-Handling: 4 bare except clauses mit spezifischen Exception-Typen
Code Quality:
- Logging für alle Exception-Handler hinzugefügt
- Timeout für Agent-Tasks von 300s auf 600s erhöht (10 Min)
- Bessere Kommentare für Exception-Handling
Performance:
- Wissensdatenbank aus Systemprompt entfernt
- Agents nutzen @READ_KNOWLEDGE für on-demand Zugriff
- Reduziert Prompt-Größe um ~15KB pro Task
UI Improvements (aus vorherigem Work):
- Tasks: Auto-Refresh Info statt Toggle
- Tasks: Status-Anzeigen statt manuelle Buttons
- Konsistentes Auto-Refresh (15s) wenn Tasks aktiv
