Extends the Compliance Advisor from a Q&A chatbot into a full drafting engine that can generate, validate, and refine compliance documents within Scope Engine constraints. Includes intent classifier, state projector, constraint enforcer, SOUL templates, Go backend endpoints, and React UI components. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Breakpilot Agent Core
Multi-Agent Architecture Infrastructure für Breakpilot.
Übersicht
Das agent-core Modul stellt die gemeinsame Infrastruktur für Breakpilots Multi-Agent-System bereit:
- Session Management: Agent-Sessions mit Checkpoints und Recovery
- Shared Brain: Langzeit-Gedächtnis und Kontext-Verwaltung
- Orchestration: Message Bus, Supervisor und Task-Routing
Architektur
┌─────────────────────────────────────────────────────────────────┐
│ Breakpilot Services │
│ ┌─────────────┐ ┌─────────────┐ ┌─────────────────────────┐ │
│ │Voice Service│ │Klausur Svc │ │ Admin-v2 / AlertAgent │ │
│ └──────┬──────┘ └──────┬──────┘ └───────────┬─────────────┘ │
│ │ │ │ │
│ └────────────────┼──────────────────────┘ │
│ │ │
│ ┌───────────────────────▼───────────────────────────────────┐ │
│ │ Agent Core │ │
│ │ ┌─────────────┐ ┌─────────────┐ ┌───────────────────┐ │ │
│ │ │ Sessions │ │Shared Brain │ │ Orchestrator │ │ │
│ │ │ - Manager │ │ - Memory │ │ - Message Bus │ │ │
│ │ │ - Heartbeat │ │ - Context │ │ - Supervisor │ │ │
│ │ │ - Checkpoint│ │ - Knowledge │ │ - Task Router │ │ │
│ │ └─────────────┘ └─────────────┘ └───────────────────┘ │ │
│ └───────────────────────────────────────────────────────────┘ │
│ │ │
│ ┌───────────────────────▼───────────────────────────────────┐ │
│ │ Infrastructure │ │
│ │ Valkey (Redis) PostgreSQL Qdrant │ │
│ └───────────────────────────────────────────────────────────┘ │
└─────────────────────────────────────────────────────────────────┘
Verzeichnisstruktur
agent-core/
├── __init__.py # Modul-Exports
├── README.md # Diese Datei
├── requirements.txt # Python-Abhängigkeiten
├── pytest.ini # Test-Konfiguration
│
├── soul/ # Agent SOUL Files (Persönlichkeiten)
│ ├── tutor-agent.soul.md
│ ├── grader-agent.soul.md
│ ├── quality-judge.soul.md
│ ├── alert-agent.soul.md
│ └── orchestrator.soul.md
│
├── brain/ # Shared Brain Implementation
│ ├── __init__.py
│ ├── memory_store.py # Langzeit-Gedächtnis
│ ├── context_manager.py # Konversations-Kontext
│ └── knowledge_graph.py # Entity-Beziehungen
│
├── sessions/ # Session Management
│ ├── __init__.py
│ ├── session_manager.py # Session-Lifecycle
│ ├── heartbeat.py # Liveness-Monitoring
│ └── checkpoint.py # Recovery-Checkpoints
│
├── orchestrator/ # Multi-Agent Orchestration
│ ├── __init__.py
│ ├── message_bus.py # Inter-Agent Kommunikation
│ ├── supervisor.py # Agent-Überwachung
│ └── task_router.py # Intent-basiertes Routing
│
└── tests/ # Unit Tests
├── conftest.py
├── test_session_manager.py
├── test_heartbeat.py
├── test_message_bus.py
├── test_memory_store.py
└── test_task_router.py
Komponenten
1. Session Management
Verwaltet Agent-Sessions mit State-Machine und Recovery-Fähigkeiten.
from agent_core.sessions import SessionManager, AgentSession
# Session Manager erstellen
manager = SessionManager(
redis_client=redis,
db_pool=pg_pool,
namespace="breakpilot"
)
# Session erstellen
session = await manager.create_session(
agent_type="tutor-agent",
user_id="user-123",
context={"subject": "math"}
)
# Checkpoint setzen
session.checkpoint("task_started", {"task_id": "abc"})
# Session beenden
session.complete({"result": "success"})
Session States:
ACTIVE- Session läuftPAUSED- Session pausiertCOMPLETED- Session erfolgreich beendetFAILED- Session fehlgeschlagen
2. Heartbeat Monitoring
Überwacht Agent-Liveness und triggert Recovery bei Timeout.
from agent_core.sessions import HeartbeatMonitor, HeartbeatClient
# Monitor starten
monitor = HeartbeatMonitor(
timeout_seconds=30,
check_interval_seconds=5,
max_missed_beats=3
)
await monitor.start_monitoring()
# Agent registrieren
monitor.register("agent-1", "tutor-agent")
# Heartbeat senden
async with HeartbeatClient("agent-1", monitor) as client:
# Agent-Arbeit...
pass
3. Memory Store
Langzeit-Gedächtnis für Agents mit TTL und Access-Tracking.
from agent_core.brain import MemoryStore
store = MemoryStore(redis_client=redis, db_pool=pg_pool)
# Erinnerung speichern
await store.remember(
key="evaluation:math:student-1",
value={"score": 85, "feedback": "Gut gemacht!"},
agent_id="grader-agent",
ttl_days=30
)
# Erinnerung abrufen
result = await store.recall("evaluation:math:student-1")
# Nach Pattern suchen
similar = await store.search("evaluation:math:*")
4. Context Manager
Verwaltet Konversationskontext mit automatischer Komprimierung.
from agent_core.brain import ContextManager, MessageRole
ctx_manager = ContextManager(redis_client=redis)
# Kontext erstellen
context = ctx_manager.create_context(
session_id="session-123",
system_prompt="Du bist ein hilfreicher Tutor...",
max_messages=50
)
# Nachrichten hinzufügen
context.add_message(MessageRole.USER, "Was ist Photosynthese?")
context.add_message(MessageRole.ASSISTANT, "Photosynthese ist...")
# Für LLM API formatieren
messages = context.get_messages_for_llm()
5. Message Bus
Inter-Agent Kommunikation via Redis Pub/Sub.
from agent_core.orchestrator import MessageBus, AgentMessage, MessagePriority
bus = MessageBus(redis_client=redis)
await bus.start()
# Handler registrieren
async def handle_message(msg):
return {"status": "processed"}
await bus.subscribe("grader-agent", handle_message)
# Nachricht senden
await bus.publish(AgentMessage(
sender="orchestrator",
receiver="grader-agent",
message_type="grade_request",
payload={"exam_id": "exam-1"},
priority=MessagePriority.HIGH
))
# Request-Response Pattern
response = await bus.request(message, timeout=30.0)
6. Agent Supervisor
Überwacht und koordiniert alle Agents.
from agent_core.orchestrator import AgentSupervisor, RestartPolicy
supervisor = AgentSupervisor(message_bus=bus, heartbeat_monitor=monitor)
# Agent registrieren
await supervisor.register_agent(
agent_id="tutor-1",
agent_type="tutor-agent",
restart_policy=RestartPolicy.ON_FAILURE,
max_restarts=3,
capacity=10
)
# Agent starten
await supervisor.start_agent("tutor-1")
# Load Balancing
available = supervisor.get_available_agent("tutor-agent")
7. Task Router
Intent-basiertes Routing mit Fallback-Ketten.
from agent_core.orchestrator import TaskRouter, RoutingRule, RoutingStrategy
router = TaskRouter(supervisor=supervisor)
# Eigene Regel hinzufügen
router.add_rule(RoutingRule(
intent_pattern="learning_*",
agent_type="tutor-agent",
priority=10,
fallback_agent="orchestrator"
))
# Task routen
result = await router.route(
intent="learning_math",
context={"grade": 10},
strategy=RoutingStrategy.LEAST_LOADED
)
if result.success:
print(f"Routed to {result.agent_id}")
SOUL Files
SOUL-Dateien definieren die Persönlichkeit und Verhaltensregeln jedes Agents.
| Agent | SOUL File | Verantwortlichkeit |
|---|---|---|
| TutorAgent | tutor-agent.soul.md | Lernbegleitung, Fragen beantworten |
| GraderAgent | grader-agent.soul.md | Klausur-Korrektur, Bewertung |
| QualityJudge | quality-judge.soul.md | BQAS Qualitätsprüfung |
| AlertAgent | alert-agent.soul.md | Monitoring, Benachrichtigungen |
| Orchestrator | orchestrator.soul.md | Task-Koordination |
Datenbank-Schema
agent_sessions
CREATE TABLE agent_sessions (
id UUID PRIMARY KEY,
agent_type VARCHAR(50) NOT NULL,
user_id UUID REFERENCES users(id),
state VARCHAR(20) NOT NULL DEFAULT 'active',
context JSONB DEFAULT '{}',
checkpoints JSONB DEFAULT '[]',
created_at TIMESTAMPTZ DEFAULT NOW(),
updated_at TIMESTAMPTZ DEFAULT NOW(),
last_heartbeat TIMESTAMPTZ DEFAULT NOW()
);
agent_memory
CREATE TABLE agent_memory (
id UUID PRIMARY KEY,
namespace VARCHAR(100) NOT NULL,
key VARCHAR(500) NOT NULL,
value JSONB NOT NULL,
agent_id VARCHAR(50) NOT NULL,
access_count INTEGER DEFAULT 0,
created_at TIMESTAMPTZ DEFAULT NOW(),
expires_at TIMESTAMPTZ,
UNIQUE(namespace, key)
);
agent_messages
CREATE TABLE agent_messages (
id UUID PRIMARY KEY,
sender VARCHAR(50) NOT NULL,
receiver VARCHAR(50) NOT NULL,
message_type VARCHAR(50) NOT NULL,
payload JSONB NOT NULL,
priority INTEGER DEFAULT 1,
correlation_id UUID,
created_at TIMESTAMPTZ DEFAULT NOW()
);
Integration
Mit Voice-Service
from services.enhanced_task_orchestrator import EnhancedTaskOrchestrator
orchestrator = EnhancedTaskOrchestrator(
redis_client=redis,
db_pool=pg_pool
)
await orchestrator.start()
# Session für Voice-Interaktion
session = await orchestrator.create_session(
voice_session_id="voice-123",
user_id="teacher-1"
)
# Task verarbeiten (nutzt Multi-Agent wenn nötig)
await orchestrator.process_task(task)
Mit BQAS
from bqas.quality_judge_agent import QualityJudgeAgent
judge = QualityJudgeAgent(
message_bus=bus,
memory_store=memory
)
await judge.start()
# Direkte Evaluation
result = await judge.evaluate(
response="Der Satz des Pythagoras...",
task_type="learning_math",
context={"user_input": "Was ist Pythagoras?"}
)
if result["verdict"] == "production_ready":
# Response ist OK
pass
Tests
# In agent-core Verzeichnis
cd agent-core
# Alle Tests ausführen
pytest -v
# Mit Coverage
pytest --cov=. --cov-report=html
# Einzelnes Test-Modul
pytest tests/test_session_manager.py -v
# Async-Tests
pytest tests/test_message_bus.py -v
Metriken
Das Agent-Core exportiert folgende Metriken:
| Metrik | Beschreibung |
|---|---|
agent_session_duration_seconds |
Dauer von Agent-Sessions |
agent_heartbeat_delay_seconds |
Zeit seit letztem Heartbeat |
agent_message_latency_ms |
Latenz der Inter-Agent Kommunikation |
agent_memory_access_total |
Memory-Zugriffe pro Agent |
agent_error_total |
Fehler pro Agent-Typ |
Nächste Schritte
- Migration ausführen:
psql -f backend/migrations/add_agent_core_tables.sql - Voice-Service erweitern: Enhanced Orchestrator aktivieren
- BQAS integrieren: Quality Judge Agent starten
- Monitoring aufsetzen: Metriken in Grafana integrieren