Merge pull request 'v0.7.0: sichere Steuerungsübergabe und klare Bedienung' (#40 ) from feature/control-handoff-v0.7.0 into main

CONTROL-001: add safe HA automation handoff
Merge pull request 'v0.6.2: HA-Automationen gleichwertig lernen' (#39 ) from feature/automation-equality-v0.6.2 into main
2026-06-14 16:22:18 +02:00 · 2026-06-14 16:21:57 +02:00 · 2026-06-14 15:58:44 +02:00 · 2026-06-14 15:58:05 +02:00 · 2026-06-14 15:38:26 +02:00 · 2026-06-14 15:38:11 +02:00
43 changed files with 3145 additions and 510 deletions
--- a/.env.example
+++ b/.env.example
@@ -7,3 +7,9 @@ SILLYHOME_HISTORY_DAYS=14
 SILLYHOME_MIN_TRAINING_POINTS=24
 SILLYHOME_RETRAIN_STALE_HOURS=24
 SILLYHOME_RECONCILE_INTERVAL_SECONDS=900
+SILLYHOME_MIN_BEHAVIOR_ACTIONS=3
+SILLYHOME_PREDICTION_CONFIDENCE=0.82
+SILLYHOME_PREDICTION_WINDOW_MINUTES=30
+SILLYHOME_PREDICTION_INTERVAL_SECONDS=60
+SILLYHOME_EXECUTION_COOLDOWN_SECONDS=900
+SILLYHOME_TIMEZONE=Europe/Berlin
--- a/.gitea/ISSUE_TEMPLATE/bug.md
+++ b/.gitea/ISSUE_TEMPLATE/bug.md
@@ -0,0 +1,33 @@
+---
+name: Fehler
+about: Reproduzierbaren SillyHome-Fehler melden
+title: "BUG: "
+---
+
+## Beobachtet
+
+Was ist tatsächlich passiert?
+
+## Erwartet
+
+Was sollte passieren?
+
+## Aktor und Kontext
+
+- Aktor:
+- Trigger/Kontext:
+- SillyHome-Modus:
+- Passende HA-Automation und Zustand:
+
+## Nachweise
+
+- Version:
+- Relevante Logs:
+- `activation_reason`:
+- `prediction.execution_reason`:
+
+## Reproduktion
+
+1.
+2.
+3.
--- a/.gitea/PULL_REQUEST_TEMPLATE.md
+++ b/.gitea/PULL_REQUEST_TEMPLATE.md
@@ -0,0 +1,25 @@
+## Ziel
+
+Welches konkrete Verhalten ändert sich?
+
+## Umsetzung
+
+- 
+
+## Sicherheit
+
+- Backup/Rollback:
+- Auswirkung auf bestehende HA-Automationen:
+- Shadow/Active-Verhalten:
+
+## Verifikation
+
+```bash
+.venv/bin/pytest -q
+.venv/bin/ruff check .
+.venv/bin/mypy app backend tests
+git diff --check
+```
+
+- Live-Health:
+- Live-Aktor:
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -0,0 +1,50 @@
+# AGENTS.md
+
+Diese Datei ist die kurze Arbeitsanweisung für Menschen und kleine Coding-Modelle.
+
+## Reihenfolge
+
+1. `README.md` lesen.
+2. Für Verhaltenslogik `docs/BEHAVIOR_ENGINE.md` lesen.
+3. Für Fehler `docs/DEBUGGING.md` abarbeiten.
+4. Für HA-Automationen `docs/CONTROL_HANDOFF.md` lesen.
+5. Vor Release oder Live-Update `docs/OPERATIONS.md` vollständig abarbeiten.
+
+## Verbindliche Regeln
+
+- Erst Zustand und Logs prüfen, dann Ursache formulieren, dann ändern.
+- Keine Annahme als Fakt darstellen.
+- Vor Live-Änderungen Backup oder klaren Rollback-Punkt erstellen.
+- Bestehende Nutzeränderungen nicht zurücksetzen.
+- Kleine, fokussierte Änderungen mit passenden Tests.
+- Eigene SillyHome-Schaltungen niemals als neues Nutzerverhalten lernen.
+- Ein Aktor darf nicht unbeabsichtigt ohne Steuerung bleiben:
+  - SillyHome aktiv: passende HA-Automation darf pausiert sein.
+  - SillyHome Shadow: HA-Automation muss auf Wunsch fortgesetzt werden können.
+- Keine Secrets in Code, Dokumentation, Commits oder Logs.
+
+## Pflichtprüfung
+
+```bash
+.venv/bin/pytest -q
+.venv/bin/ruff check .
+.venv/bin/mypy app backend tests
+git diff --check
+```
+
+## Versionsstellen
+
+Bei jedem Release dieselbe Version setzen:
+
+- `pyproject.toml`
+- `addon/config.yaml`
+- `app/main.py`
+- `CHANGELOG.md`
+
+Danach prüfen:
+
+```bash
+grep -R 'version.*0\\.7\\.0' -n pyproject.toml addon/config.yaml app/main.py
+```
+
+Die konkrete Zielversion im Befehl anpassen.
--- a/ARCHITECTURE.md
+++ b/ARCHITECTURE.md
@@ -1,13 +1,24 @@
 # SillyHome Next — Architekturübersicht

-Ziel ist ein lokales, datensparsames, erklärbares Smart-Home-Intelligenzsystem für Home Assistant. Es analysiert Historie, erkennt Gewohnheiten, erstellt Vorhersagen, empfiehlt Automationen und kann auf Wunsch einfach in Automationen übersetzen. Vier Intelligenzebenen sind vorgesehen: regelbasiert, ML-gestützt, LLM-unterstützt und autonomer Hausagent.
+Ziel ist ein lokales, datensparsames und erklärbares Smart-Home-Intelligenzsystem
+für Home Assistant. Nutzer wählen ausschließlich erlaubte Aktoren. Das System
+ordnet Kontext automatisch zu, erkennt historische Nutzerhandlungen, trainiert
+pro Aktor ein Verhaltensmodell und trifft zunächst nur Shadow-Vorhersagen.
+Autonomes Schalten wird separat pro Aktor freigegeben.

 ## Leitentscheidungen

 - Lokal-first und datensparsam; keine Cloudpflicht.
- Trennung von Datenintegration, Trainingspipeline, Vorhersageservice und Erklärungsschicht.
- Standardintegration über MQTT und Home Assistant WebSocket plus REST.
- Schnittstellen über FastAPI und OpenAI-kompatible Endpunkte.
- Langzeitdaten in PostgreSQL und TimescaleDB; Vektoren für semantische Suche optional.
- Deployment über Docker Compose; Kubernetes optional für erweiterte Betriebsgrößen.
+- Trennung von Datenintegration, Kontextzuordnung, Verhaltenslernen,
+  Vorhersage und Aktorausführung.
+- Logbook-basierte Herkunftserkennung; eindeutig erkannte HA-Automationen
+  zählen wie manuelle Bedienungen. Eigene SillyHome-Schaltungen werden nicht
+  zurückgelernt.
+- Ausführung nur für freigegebene, reversible Domains und Zustände sowie mit
+  Konfidenzschwelle und zustandsbezogenem Cooldown.
+- Eindeutig passende HA-Automationen können bei einer SillyHome-Übernahme
+  pausiert und beim Rückfall in den Shadow-Modus wieder fortgesetzt werden.
+- Standardintegration über die lokale Home-Assistant-REST-API.
+- Persistenz als atomische lokale Modell- und Aktorartefakte.
+- Deployment als Home-Assistant-Add-on oder über Docker Compose.
 - Tests, Docs und Changelog sind Pflichtbestandteil jeder Änderung.
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -1,5 +1,79 @@
 # Changelog

+## 0.7.0 - 2026-06-14
+- Freie Eingabe von Home-Assistant-Entitätsnamen mit Vorschlagsliste
+- Freigabestatus und Blockadegrund sind in Übersicht und Details immer sichtbar
+- Vorhersagen erklären konkret, warum sie ausgeführt oder nicht ausgeführt wurden
+- Cooldown blockiert nur Wiederholungen desselben Zielzustands; Gegenaktionen
+  wie `Licht an` gefolgt von `Licht aus` bleiben sofort möglich
+- Passende HA-Automationen werden aus ihren echten Konfigurationen erkannt und
+  können pausiert oder fortgesetzt werden
+- Sichere Steuerungsübergabe: SillyHome kann übernehmen und passende
+  HA-Automationen pausieren; beim Stoppen können sie gezielt fortgesetzt werden
+- Dashboard wird ohne Browser-Cache ausgeliefert
+- Reproduzierbare Runbooks für Debugging, Berechnung, Entwicklung, Tests,
+  Release, Add-on-Update, Live-Verifikation und Rollback
+
+## 0.6.2 - 2026-06-14
+- Eindeutig im Home-Assistant-Logbuch erkannte Automationen und Scripts zählen für
+  Lernen und Freigabe gleichwertig wie manuelle Bedienungen
+- Automationsmuster erhalten dieselbe Modellgewichtung wie manuelle Handlungen
+- Oberfläche zeigt die gemeinsame Zahl als `eindeutig geregelt`; eine
+  ausdrückliche Aktivierung pro Aktor bleibt weiterhin erforderlich
+
+## 0.6.1 - 2026-06-14
+- Manuelle Prüfung als `Aktuelle Situation auswerten` eindeutig von Simulation
+  oder Aktorschaltung abgegrenzt
+- Sichtbare Rückmeldung mit Prüfzeitpunkt, vorhergesagtem Zustand und Sicherheit
+  oder klarem Hinweis auf einen fehlenden frischen Sensorwechsel
+
+## 0.6.0 - 2026-06-14
+- Kausales Shadow-Lernen erkennt frische Kontextwechsel unmittelbar vor einer
+  Aktorhandlung, etwa `Tür geschlossen → offen` vor `Licht aus → an`
+- Historische Home-Assistant-Automationen dürfen Vorhersagen begründen, zählen
+  aber weiterhin niemals als eindeutige Benutzerhandlung oder Ausführungsfreigabe
+- Aktuelle `last_changed`-Zeitpunkte verhindern Vorhersagen aus längst
+  unveränderten Sensorzuständen
+- Oberfläche trennt gelernte Benutzerhandlungen und erkannte HA-Automationen
+
+## 0.5.4 - 2026-06-14
+- Tür-, Bewegungs- und andere belastbare Kontextsensoren werden auch ohne
+  numerischen Sensor als vollständige automatische Kontextzuordnung angezeigt
+- Status und Zuordnungssicherheit bilden das aktive Verhaltenslernen ab statt
+  eines optionalen numerischen Modells
+- Ausführungsfreigabe erscheint erst, wenn genügend eindeutig manuelle
+  Bedienungen vorliegen; bis dahin nennt die Oberfläche die noch fehlende Anzahl
+
+## 0.5.3 - 2026-06-14
+- Verhindert fachlich falsche Sensorzuordnungen nur aufgrund generischer Namen wie
+  `Licht` oder `Lichtschalter`
+- Übernimmt numerische Sensoren nur noch bei einem belastbaren absoluten Score und
+  einer eindeutigen Abgrenzung zum zweitbesten Kandidaten
+- Begrenzt Zusatzkontext auf relevante Sensoren und bevorzugt bei Lichtaktoren
+  echte Beleuchtungsstärke gegenüber fremden Leistungs- oder Energiezählern
+
+## 0.5.2 - 2026-06-14
+- Add-on-Build invalidiert den Docker-Cache bei jeder Versionsänderung, damit
+  Versionsmetadaten und tatsächlich ausgelieferter Anwendungscode übereinstimmen
+- Korrigierte Ingress-Oberfläche aus 0.5.1 dadurch erstmals zuverlässig ausgeliefert
+
+## 0.5.1 - 2026-06-14
+- Technische Modell-, Intervall- und Sicherheitsparameter aus der normalen
+  Home-Assistant-Add-on-Konfiguration entfernt; sichere Standardwerte bleiben aktiv
+- Ingress um einen klaren Ablauf mit Aktorauswahl, Beobachtungsphase und späterer
+  Ausführungsfreigabe ergänzt
+- Bedienelemente und Diagnosen in verständlicher Alltagssprache erklärt
+
+## 0.5.0 - 2026-06-14
+- Ingress auf reine Aktorauswahl, automatischen Lernstatus und Vorhersagen reduziert
+- Automatische Kontextzuordnung ohne Sensor-Overrides oder Review-Blockade
+- Historische Handlungserkennung aus HA-State-History und Logbook-Herkunft
+- Persistentes Verhaltensmodell pro Aktor mit Zeit-, Wochentags- und Kontextmustern
+- Shadow-Vorhersagen vor jeder Ausführungsfreigabe
+- Explizite Aktivierung pro Aktor, Konfidenzschwelle, Cooldown und enge Service-Whitelist
+- Schutz vor dem Lernen erkannter HA-Automationen und eigener Schaltvorgänge
+- Automation-Proposal- und Override-Endpunkte aus dem aktiven Produkt entfernt
+
 ## 0.4.0 - 2026-06-13
 - Aktuator-zentrierte Einrichtung: Nutzer wählen nur noch Aktuatoren, Sensoren werden deterministisch gefunden und bewertet
 - Persistente automatische und manuelle Sensorzuordnungen mit Evidenz, Confidence, Review-Gating und Neustart-Sicherheit
--- a/8
+++ b/8
@@ -9,7 +9,13 @@ ENV SILLYHOME_AUTOMATION_STORE=/app/data/automations \
    SILLYHOME_HISTORY_DAYS=14 \
    SILLYHOME_MIN_TRAINING_POINTS=24 \
    SILLYHOME_RETRAIN_STALE_HOURS=24 \
-    SILLYHOME_RECONCILE_INTERVAL_SECONDS=900
+    SILLYHOME_RECONCILE_INTERVAL_SECONDS=900 \
+    SILLYHOME_MIN_BEHAVIOR_ACTIONS=3 \
+    SILLYHOME_PREDICTION_CONFIDENCE=0.82 \
+    SILLYHOME_PREDICTION_WINDOW_MINUTES=30 \
+    SILLYHOME_PREDICTION_INTERVAL_SECONDS=60 \
+    SILLYHOME_EXECUTION_COOLDOWN_SECONDS=900 \
+    SILLYHOME_TIMEZONE=Europe/Berlin

 WORKDIR /app

--- a/README.md
+++ b/README.md
@@ -1,13 +1,25 @@
 # SillyHome Next

-Lokaler, datenschutzfreundlicher API-Prototyp für Home Assistant.
+SillyHome lernt aus Home Assistant, sagt Aktorhandlungen voraus und darf sie
+nach einer ausdrücklichen Freigabe ausführen.
+
+## Schnell orientieren
+
+- Fehler finden: [`docs/DEBUGGING.md`](docs/DEBUGGING.md)
+- Berechnung verstehen: [`docs/BEHAVIOR_ENGINE.md`](docs/BEHAVIOR_ENGINE.md)
+- Steuerung übernehmen/zurückgeben:
+  [`docs/CONTROL_HANDOFF.md`](docs/CONTROL_HANDOFF.md)
+- Entwickeln, testen, veröffentlichen und installieren:
+  [`docs/OPERATIONS.md`](docs/OPERATIONS.md)
+- Arbeitsregeln für Coding-Agenten: [`AGENTS.md`](AGENTS.md)

 ## Reifegrad

-Die aktuelle Entwicklungslinie ist aktor-zentriert: Nutzer konfigurieren nur
-noch Home-Assistant-Aktuatoren. SillyHome Next findet dazu passende numerische
-Sensoren und Kontext-Entities, zeigt Evidenz und Review-Bedarf an und hält
-passende Modelle lokal und autonom aktuell.
+Die aktuelle Entwicklungslinie ist vollständig aktor-zentriert: Nutzer wählen
+nur Home-Assistant-Aktuatoren aus. SillyHome Next findet Sensoren, Zustände und
+Kontext automatisch, wertet die vorhandene Historie aus und hält passende
+lokale Modelle autonom aktuell. Es gibt keinen Regel-, Trigger-, Sensor- oder
+YAML-Konfigurationsschritt.

 ## Motivation
 TheSillyHome zeigte die Idee: statt statischer Regeln das Zuhause aus Verhaltensmustern verstehen. Diese Architektur modernisiert den Ansatz in Richtung Explainable AI, hybride Intelligenzebenen und langlebige Wartbarkeit.
@@ -16,7 +28,7 @@ TheSillyHome zeigte die Idee: statt statischer Regeln das Zuhause aus Verhaltens
 - Home Assistant und Sensoren/Aktoren verstehen
 - Historie auswerten und Gewohnheiten erkennen
 - Vorhersagen erstellen und erklären
- Automationen vorschlagen und direkt generieren
+- Persönliches Verhalten pro Aktor lernen und zukünftige Handlungen vorhersagen
 - Lokal-first ohne Cloudpflicht
 - Erweiterbar, testbar, dokumentiert

@@ -46,12 +58,13 @@ uvicorn app.main:app --reload
 - `http://127.0.0.1:8000/v1/discovery` - klassifizierte, filterbare Entities
 - `http://127.0.0.1:8000/v1/history` - normalisierte numerische Zeitreihen
 - `http://127.0.0.1:8000/v1/actuators/discovery` - unterstützte Aktuatoren für den aktor-zentrierten Workflow
- `POST http://127.0.0.1:8000/v1/actuators` - Aktuator registrieren, Sensorzuordnung prüfen und Modell-Lebenszyklus starten
+- `POST http://127.0.0.1:8000/v1/actuators` - Aktor freigeben; Kontextzuordnung und Modell-Lebenszyklus starten automatisch
+- `POST http://127.0.0.1:8000/v1/actuators/{entity_id}/evaluate` - Shadow-Vorhersage aktualisieren
+- `POST http://127.0.0.1:8000/v1/actuators/{entity_id}/activation` - autonomes Schalten pro Aktor freigeben oder stoppen
 - `POST http://127.0.0.1:8000/v1/actuators/reconciliation/run` - globale Reconciliation manuell anstoßen
 - `http://127.0.0.1:8000/ml/health` - Registry-/Serving-Health
 - `POST http://127.0.0.1:8000/ml/retrain` - Modell-Metadaten aktualisieren
 - `POST http://127.0.0.1:8000/ml/evaluate` - MAE/RMSE/Coverage berechnen
- `POST http://127.0.0.1:8000/v1/automations/proposals` - sicheren Entwurf anlegen

 Ohne vollständige HA-Konfiguration liefert `/v1/entities` bewusst `503`.

@@ -70,12 +83,17 @@ dem Netz muss ein authentifizierender Reverse Proxy vorgeschaltet werden.
 - `SILLYHOME_HA_URL` – Basis-URL deiner Home-Assistant-Instanz (z. B. `http://homeassistant.local:8123`)
 - `SILLYHOME_HA_TOKEN` – Long-Lived Access Token eines dedizierten HA-Benutzers mit minimalen Rechten
 - `SILLYHOME_MODEL_STORE` – Verzeichnis für persistierte Modell-Metadaten
- `SILLYHOME_AUTOMATION_STORE` – Verzeichnis für Automation-Entwürfe
- `SILLYHOME_ACTUATOR_STORE` – Verzeichnis für persistente Aktuator-Zuordnungen, Overrides und Reconciliation-Status
+- `SILLYHOME_ACTUATOR_STORE` – Verzeichnis für persistente Aktor-Zuordnungen und Reconciliation-Status
 - `SILLYHOME_HISTORY_DAYS` – Trainingsfenster für HA-History (1 bis 31 Tage)
 - `SILLYHOME_MIN_TRAINING_POINTS` – Mindestanzahl nutzbarer numerischer Messpunkte vor einem Modelltraining
 - `SILLYHOME_RETRAIN_STALE_HOURS` – Staleness-Grenze für automatisches Retraining
 - `SILLYHOME_RECONCILE_INTERVAL_SECONDS` – Intervall für sichere periodische Reconciliation
+- `SILLYHOME_MIN_BEHAVIOR_ACTIONS` – Mindestzahl gelernter Handlungen vor einer Freigabe
+- `SILLYHOME_PREDICTION_CONFIDENCE` – Mindestkonfidenz für autonomes Schalten
+- `SILLYHOME_PREDICTION_WINDOW_MINUTES` – Zeitfenster um gelernte Handlungsmuster
+- `SILLYHOME_PREDICTION_INTERVAL_SECONDS` – Intervall für Shadow-/Aktiv-Vorhersagen
+- `SILLYHOME_EXECUTION_COOLDOWN_SECONDS` – Mindestabstand zwischen eigenen Schaltungen
+- `SILLYHOME_TIMEZONE` – lokale Zeitzone für Tages- und Wochenmuster

 Niemals Administrator-Tokens oder Passwörter eintragen. `.env` gehört nicht ins
 Versionskontrollsystem.
@@ -88,15 +106,26 @@ unter **Einstellungen → Add-ons → Add-on-Shop → Repositories** diese URL e
 `http://192.168.6.31:3000/pino/sillyhome-next`

 Danach **SillyHome Next** installieren und starten. Das Dashboard wird per Ingress
-geöffnet. Das Add-on nutzt die Supervisor-API nur lesend; Automation-Entwürfe werden
-lokal gespeichert und niemals automatisch ausgeführt.
+geöffnet. Dort werden ausschließlich erlaubte Aktoren ausgewählt; Kontext- und
+Lernentscheidungen erfolgen automatisch.

 ### Normaler Workflow
-1. Im Dashboard oder per API einen Aktuator auswählen, zum Beispiel `light.abstellkammer`.
-2. SillyHome Next bewertet passende numerische Sensoren und binäre Kontext-Entities anhand von Bereich, Gerät, Namen, Domain und `device_class`.
-3. Starke und eindeutige Zuordnungen werden automatisch genutzt; schwache oder knappe Kandidaten bleiben mit Review-Hinweis sichtbar.
-4. Manuelle Overrides haben Vorrang, bleiben persistent und überstehen Neustarts.
-5. Sobald genügend numerische HA-Historie vorhanden ist, trainiert das System automatisch ein lokales Modell pro Aktuator-Zuordnung und retrainiert es bei relevanten Datenänderungen oder Staleness.
+1. Im Dashboard einen Aktor auswählen, zum Beispiel `light.abstellkammer`.
+2. SillyHome Next bewertet automatisch Messwerte, Anwesenheit, Bewegung,
+   Bereiche, Gerätebeziehungen und weitere HA-Kontexte.
+3. Das System verwendet selbstständig die beste verfügbare Zuordnung.
+   Niedrige Sicherheit bleibt als Diagnose sichtbar, verlangt aber keine
+   manuelle Konfiguration.
+4. Sobald genügend Historie vorhanden ist, trainiert und aktualisiert das
+   System das lokale Modell automatisch.
+5. Vorhersagen laufen zunächst ausschließlich im Shadow-Modus.
+6. Erst nach ausdrücklicher Freigabe pro Aktor werden hochkonfidente,
+   erlaubte Zustände geschaltet. Eindeutig im HA-Logbuch erkannte Automationen
+   und Scripts zählen dabei gleichwertig wie manuelle Bedienungen. Eigene
+   Schaltungen von SillyHome werden nicht zurückgelernt.
+7. Bei der Freigabe kann SillyHome passende HA-Automationen pausieren und die
+   Steuerung übernehmen. Beim Stoppen können diese Automationen gezielt wieder
+   fortgesetzt werden.

 Vor einem Update sollte in Home Assistant unter **Einstellungen → System → Backups**
 eine Teil-Sicherung des Add-ons erstellt werden. Zur Wiederherstellung das gewünschte
@@ -107,5 +136,5 @@ Teststand `v0.3.0` wurde als HA-Backup `7df0fca0` gesichert.
 ```bash
 pytest
 ruff check .
-mypy
+mypy app backend tests
 ```
--- a/addon/Dockerfile
+++ b/addon/Dockerfile
@@ -4,13 +4,17 @@ ENV PYTHONDONTWRITEBYTECODE=1 \
    PYTHONUNBUFFERED=1 \
    PIP_NO_CACHE_DIR=1

+# The add-on version changes for every release. Copying its config before the
+# clone makes Docker invalidate the application layer instead of reusing old code.
+COPY config.yaml /tmp/addon-config.yaml
+
 RUN apt-get update \
    && apt-get install -y --no-install-recommends git \
    && git clone --depth 1 --branch main \
       http://192.168.6.31:3000/pino/sillyhome-next.git /app \
    && python -m pip install --upgrade pip \
    && python -m pip install /app \
-    && rm -rf /var/lib/apt/lists/* /app/.git
+    && rm -rf /var/lib/apt/lists/* /app/.git /tmp/addon-config.yaml

 COPY run.sh /run.sh
 RUN chmod 0755 /run.sh
--- a/addon/config.yaml
+++ b/addon/config.yaml
@@ -1,7 +1,7 @@
 name: SillyHome Next
-version: "0.4.0"
+version: "0.7.0"
 slug: sillyhome_next
-description: Lokale HA-Analyse, Vorhersagen und sichere Automation-Entwürfe
+description: Lernt automatisch aus deinem Verhalten und steuert freigegebene Aktoren
 url: http://192.168.6.31:3000/pino/sillyhome-next
 arch:
  - amd64
@@ -16,16 +16,6 @@ panel_admin: true
 homeassistant_api: true
 hassio_api: false
 auth_api: false
-options:
-  history_days: 14
-  min_training_points: 24
-  retrain_stale_hours: 24
-  reconcile_interval_seconds: 900
-schema:
-  history_days: "int(1,31)"
-  min_training_points: "int(2,10000)"
-  retrain_stale_hours: "int(1,720)"
-  reconcile_interval_seconds: "int(60,86400)"
 map:
  - type: addon_config
    read_only: false
--- a/addon/run.sh
+++ b/addon/run.sh
@@ -12,6 +12,12 @@ if [ -f /data/options.json ]; then
  export SILLYHOME_MIN_TRAINING_POINTS="$(python -c 'import json; print(json.load(open("/data/options.json")).get("min_training_points", 24))')"
  export SILLYHOME_RETRAIN_STALE_HOURS="$(python -c 'import json; print(json.load(open("/data/options.json")).get("retrain_stale_hours", 24))')"
  export SILLYHOME_RECONCILE_INTERVAL_SECONDS="$(python -c 'import json; print(json.load(open("/data/options.json")).get("reconcile_interval_seconds", 900))')"
+  export SILLYHOME_MIN_BEHAVIOR_ACTIONS="$(python -c 'import json; print(json.load(open("/data/options.json")).get("min_behavior_actions", 3))')"
+  export SILLYHOME_PREDICTION_CONFIDENCE="$(python -c 'import json; print(json.load(open("/data/options.json")).get("prediction_confidence", 0.82))')"
+  export SILLYHOME_PREDICTION_WINDOW_MINUTES="$(python -c 'import json; print(json.load(open("/data/options.json")).get("prediction_window_minutes", 30))')"
+  export SILLYHOME_PREDICTION_INTERVAL_SECONDS="$(python -c 'import json; print(json.load(open("/data/options.json")).get("prediction_interval_seconds", 60))')"
+  export SILLYHOME_EXECUTION_COOLDOWN_SECONDS="$(python -c 'import json; print(json.load(open("/data/options.json")).get("execution_cooldown_seconds", 900))')"
+  export SILLYHOME_TIMEZONE="$(python -c 'import json; print(json.load(open("/data/options.json")).get("timezone", "Europe/Berlin"))')"
 fi

 mkdir -p "$SILLYHOME_MODEL_STORE" "$SILLYHOME_AUTOMATION_STORE" "$SILLYHOME_ACTUATOR_STORE"
--- a/app/actuators/lifecycle.py
+++ b/app/actuators/lifecycle.py
@@ -13,7 +13,6 @@ from app.actuators.models import (
    AssignmentSource,
    LifecycleAuditEntry,
    LifecycleStatus,
-    ManualOverride,
    ModelLifecycleState,
    ReconciliationState,
    model_id_for_actuator,
@@ -46,6 +45,8 @@ _STOPWORDS = frozenset(
        "humidity",
        "illuminance",
        "light",
+        "licht",
+        "lichtschalter",
        "power",
        "sensor",
        "state",
@@ -55,9 +56,11 @@ _STOPWORDS = frozenset(
    }
 )
 _NUMERIC_AUTO_ACCEPT_SCORE = 0.82
+_NUMERIC_AUTO_ACCEPT_MIN_SCORE = 0.5
 _NUMERIC_MIN_MARGIN = 0.18
 _CONTEXT_AUTO_ACCEPT_SCORE = 0.78
-_MAX_CONTEXT_SELECTIONS = 3
+_CONTEXT_AUTO_ACCEPT_MIN_SCORE = 0.3
+_MAX_CONTEXT_SELECTIONS = 5
 _AUDIT_LIMIT = 20


@@ -85,21 +88,6 @@ class ActuatorReconciliationService:
    def get_actuator(self, actuator_entity_id: str) -> ActuatorRecord:
        return self._store.get(actuator_entity_id)

-    def set_override(
-        self,
-        actuator_entity_id: str,
-        override: ManualOverride | None,
-    ) -> ActuatorRecord:
-        record = self._store.get(actuator_entity_id)
-        updated = record.model_copy(
-            update={
-                "manual_override": override,
-                "updated_at": datetime.now(timezone.utc),
-            }
-        )
-        self._store.upsert(updated)
-        return self.reconcile_actuator(actuator_entity_id, trigger="override")
-
    def delete_actuator(self, actuator_entity_id: str) -> None:
        model_id = model_id_for_actuator(actuator_entity_id)
        self._registry.archive(model_id)
@@ -132,7 +120,7 @@ class ActuatorReconciliationService:
            last_summary=(
                f"{len(refreshed)} Aktuatoren geprüft, "
                f"{sum(1 for record in refreshed if record.assignment.review_required)} "
-                "mit Prüfbedarf."
+                "mit niedriger Zuordnungssicherheit."
            ),
        )
        self._store.save_reconciliation_state(summary)
@@ -214,7 +202,6 @@ class ActuatorReconciliationService:
            actuator=actuator,
            numeric_candidates=numeric_candidates,
            context_candidates=context_candidates,
-            override=record.manual_override,
        )
        lifecycle = self._reconcile_lifecycle(
            actuator=actuator,
@@ -225,6 +212,7 @@ class ActuatorReconciliationService:
        updated = record.model_copy(
            update={
                "assignment": assignment,
+                "manual_override": None,
                "numeric_candidates": numeric_candidates,
                "context_candidates": context_candidates,
                "lifecycle": lifecycle,
@@ -246,36 +234,40 @@ class ActuatorReconciliationService:
        actuator: HaEntitySummary,
        numeric_candidates: list[AssignmentCandidate],
        context_candidates: list[AssignmentCandidate],
-        override: ManualOverride | None,
    ) -> AssignmentSelection:
-        if override is not None:
-            selected_numeric = override.numeric_entity_id
-            selected_contexts = list(dict.fromkeys(override.context_entity_ids))
-            return AssignmentSelection(
-                selected_numeric_entity_id=selected_numeric,
-                selected_context_entity_ids=selected_contexts,
-                source=AssignmentSource.MANUAL,
-                confidence=1.0 if selected_numeric else 0.6,
-                review_required=False,
-                reason=(
-                    "Manuelle Zuordnung überschreibt die automatische Heuristik dauerhaft."
-                ),
-            )
-
-        top_numeric = numeric_candidates[0] if numeric_candidates else None
-        top_contexts = [
-            candidate.entity_id
+        top_numeric = next(
+            (candidate for candidate in numeric_candidates if candidate.auto_accepted),
+            None,
+        )
+        accepted_contexts = [
+            candidate
            for candidate in context_candidates
            if candidate.auto_accepted
        ][: _MAX_CONTEXT_SELECTIONS]
+        top_contexts = [candidate.entity_id for candidate in accepted_contexts]
        if top_numeric is None:
+            if accepted_contexts:
+                return AssignmentSelection(
+                    selected_numeric_entity_id=None,
+                    selected_context_entity_ids=top_contexts,
+                    source=AssignmentSource.AUTOMATIC,
+                    confidence=max(candidate.confidence for candidate in accepted_contexts),
+                    review_required=False,
+                    reason=(
+                        "Passender Schaltkontext automatisch erkannt. Für diese "
+                        "Verhaltensvorhersage ist kein numerischer Sensor erforderlich."
+                    ),
+                )
            return AssignmentSelection(
                selected_numeric_entity_id=None,
                selected_context_entity_ids=top_contexts,
                source=AssignmentSource.NONE,
                confidence=0.0,
                review_required=True,
-                reason=f"Kein numerischer Sensor konnte für {display_name(actuator)} bestimmt werden.",
+                reason=(
+                    f"Für {display_name(actuator)} ist noch kein nutzbarer numerischer "
+                    "Kontext verfügbar. Die Zuordnung wird automatisch erneut geprüft."
+                ),
            )

        return AssignmentSelection(
@@ -285,9 +277,9 @@ class ActuatorReconciliationService:
            confidence=top_numeric.confidence,
            review_required=not top_numeric.auto_accepted,
            reason=(
-                "Automatisch akzeptiert."
+                "Kontext automatisch und eindeutig zugeordnet."
                if top_numeric.auto_accepted
-                else "Top-Kandidat gefunden, aber Zuordnung ist noch nicht eindeutig genug."
+                else "Besten verfügbaren Kontext automatisch mit niedriger Sicherheit zugeordnet."
            ),
        )

@@ -306,14 +298,6 @@ class ActuatorReconciliationService:
                "Ohne numerische Sensorzuordnung wird kein Modell aktiv gehalten.",
                now=now,
            )
-        if assignment.review_required and assignment.source is not AssignmentSource.MANUAL:
-            return self._archive_state(
-                lifecycle,
-                "Zuordnung ist nicht eindeutig; Modell wartet auf Review.",
-                now=now,
-                status=LifecycleStatus.REVIEW_REQUIRED,
-            )
-
        sensor_id = assignment.selected_numeric_entity_id
        series = self._read_history(sensor_id, now)
        points = series.points if series is not None else []
@@ -327,7 +311,7 @@ class ActuatorReconciliationService:
                            f"{len(points)} von mindestens {self._settings.min_training_points} "
                            f"Messpunkten für {sensor_id} vorhanden."
                        ),
-                        "next_action": "Mehr Historie sammeln und Reconciliation erneut ausführen.",
+                        "next_action": "Historie wird automatisch weiter gesammelt.",
                        "last_history_point_count": len(points),
                    }
                ),
@@ -369,7 +353,7 @@ class ActuatorReconciliationService:
                        "last_history_signature": signature,
                        "last_history_point_count": len(points),
                        "reason": retrain_reason,
-                        "next_action": "Automatisch überwachen und bei neuen Daten neu trainieren.",
+                        "next_action": "Neue Daten automatisch überwachen und nachtrainieren.",
                    }
                ),
                action="retrained" if result.replaced else "trained",
@@ -384,8 +368,8 @@ class ActuatorReconciliationService:
                    "last_reconciled_at": now,
                    "last_history_signature": signature,
                    "last_history_point_count": len(points),
-                    "reason": "Modell ist aktuell und passt zur bestätigten Sensorzuordnung.",
-                    "next_action": "Auf neue Historie oder Staleness warten.",
+                    "reason": "Modell ist aktuell und passt zur automatischen Kontextzuordnung.",
+                    "next_action": "Neue Historie automatisch auswerten.",
                }
            ),
            action="kept",
@@ -423,7 +407,7 @@ class ActuatorReconciliationService:
                    "status": status,
                    "last_reconciled_at": now,
                    "reason": reason,
-                    "next_action": "Review oder neue Zuordnung erforderlich.",
+                    "next_action": "Bei neuen Home-Assistant-Daten automatisch erneut zuordnen.",
                }
            ),
            action="archived",
@@ -470,8 +454,15 @@ class ActuatorReconciliationService:
            confidence = candidate.score / highest if highest else 0.0
            margin = candidate.score - second_score if index == 0 else 0.0
            auto_score = _CONTEXT_AUTO_ACCEPT_SCORE if context else _NUMERIC_AUTO_ACCEPT_SCORE
-            auto_accepted = confidence >= auto_score and (
-                context or margin >= _NUMERIC_MIN_MARGIN
+            minimum_score = (
+                _CONTEXT_AUTO_ACCEPT_MIN_SCORE
+                if context
+                else _NUMERIC_AUTO_ACCEPT_MIN_SCORE
+            )
+            auto_accepted = (
+                candidate.score >= minimum_score
+                and confidence >= auto_score
+                and (context or margin >= _NUMERIC_MIN_MARGIN)
            )
            sorted_candidates[index] = candidate.model_copy(
                update={
@@ -547,6 +538,9 @@ def _score_candidate(
    if entity.device_class in preferred_device_classes:
        score += 0.2
        evidence.append(f"Passende device_class: {entity.device_class}")
+    if not context and actuator.domain == "light" and entity.device_class == "illuminance":
+        score += 0.2
+        evidence.append("Beleuchtungsstärke wird für Lichtaktoren bevorzugt.")
    if not context and entity.unit_of_measurement is not None:
        score += 0.05
        evidence.append(f"Numerische Einheit vorhanden: {entity.unit_of_measurement}")
--- a/app/actuators/models.py
+++ b/app/actuators/models.py
@@ -25,6 +25,18 @@ class LifecycleStatus(StrEnum):
    ARCHIVED = "archived"


+class BehaviorMode(StrEnum):
+    SHADOW = "shadow"
+    ACTIVE = "active"
+    PAUSED = "paused"
+
+
+class BehaviorStatus(StrEnum):
+    COLLECTING = "collecting"
+    TRAINED = "trained"
+    BLOCKED = "blocked"
+
+
 class AssignmentCandidate(BaseModel):
    entity_id: str
    domain: str
@@ -71,10 +83,64 @@ class ModelLifecycleState(BaseModel):
    last_history_signature: str | None = None
    last_history_point_count: int = Field(default=0, ge=0)
    reason: str = "Noch keine Trainingsdaten ausgewertet."
-    next_action: str = "Aktuator auswählen und Zuordnung prüfen."
+    next_action: str = "Aktor auswählen; Kontext und Historie werden automatisch geprüft."
    audit: list[LifecycleAuditEntry] = Field(default_factory=list)


+class BehaviorPattern(BaseModel):
+    target_state: str = Field(min_length=1, max_length=100)
+    minute_of_day: int = Field(ge=0, le=1439)
+    weekday: int = Field(ge=0, le=6)
+    context_states: dict[str, str] = Field(default_factory=dict)
+    trigger_entity_id: str | None = None
+    trigger_from_state: str | None = None
+    trigger_to_state: str | None = None
+    source: str = Field(default="observed", max_length=40)
+    weight: float = Field(default=1.0, ge=0.1, le=1.0)
+    observed_at: datetime
+
+
+class BehaviorPrediction(BaseModel):
+    target_state: str
+    confidence: float = Field(ge=0.0, le=1.0)
+    generated_at: datetime
+    reason: str
+    matching_patterns: int = Field(default=0, ge=0)
+    executed: bool = False
+    execution_reason: str = "Vorhersage wurde noch nicht ausgeführt."
+
+
+class ExecutionEvent(BaseModel):
+    target_state: str
+    executed_at: datetime
+
+
+class RelatedAutomation(BaseModel):
+    entity_id: str = Field(pattern=r"^automation\.[a-z0-9_]+$")
+    config_id: str = Field(min_length=1, max_length=120)
+    friendly_name: str = Field(min_length=1, max_length=200)
+    enabled: bool
+
+
+class BehaviorState(BaseModel):
+    mode: BehaviorMode = BehaviorMode.SHADOW
+    status: BehaviorStatus = BehaviorStatus.COLLECTING
+    approved_at: datetime | None = None
+    sample_count: int = Field(default=0, ge=0)
+    high_confidence_sample_count: int = Field(default=0, ge=0)
+    patterns: list[BehaviorPattern] = Field(default_factory=list)
+    prediction: BehaviorPrediction | None = None
+    last_trained_at: datetime | None = None
+    last_evaluated_at: datetime | None = None
+    last_executed_at: datetime | None = None
+    execution_events: list[ExecutionEvent] = Field(default_factory=list)
+    activation_ready: bool = False
+    activation_reason: str = "Noch nicht genügend Verhalten für eine Freigabe gelernt."
+    related_automations: list[RelatedAutomation] = Field(default_factory=list)
+    paused_automation_entity_ids: list[str] = Field(default_factory=list)
+    reason: str = "Historische Aktorhandlungen werden analysiert."
+
+
 class ActuatorRecord(BaseModel):
    actuator_entity_id: str = Field(pattern=r"^[a-z0-9_]+\.[a-z0-9_]+$")
    enabled: bool = True
@@ -85,6 +151,7 @@ class ActuatorRecord(BaseModel):
    numeric_candidates: list[AssignmentCandidate] = Field(default_factory=list)
    context_candidates: list[AssignmentCandidate] = Field(default_factory=list)
    lifecycle: ModelLifecycleState
+    behavior: BehaviorState = Field(default_factory=BehaviorState)


 class ReconciliationState(BaseModel):
--- a/app/api/v1/actuators.py
+++ b/app/api/v1/actuators.py
@@ -4,10 +4,12 @@ from fastapi import APIRouter, Depends, HTTPException, Query, Request, status
 from pydantic import BaseModel, Field

 from app.actuators.lifecycle import ActuatorReconciliationService
-from app.actuators.models import ActuatorRecord, ManualOverride, ReconciliationState
+from app.actuators.models import ActuatorRecord, ReconciliationState
 from app.actuators.store import ActuatorStore
+from app.behavior.engine import BehaviorEngine
 from app.dependencies import get_ha_reader
 from app.ha.discovery import EntityRole
+from app.ha.exceptions import HaClientError
 from app.ha.models import HaEntitySummary
 from app.ha.reader import HaReader

@@ -19,11 +21,15 @@ class ConfigureActuatorRequest(BaseModel):
    enabled: bool = True


-class OverrideRequest(BaseModel):
-    numeric_entity_id: str | None = Field(default=None, pattern=r"^[a-z0-9_]+\.[a-z0-9_]+$")
-    context_entity_ids: list[str] = Field(default_factory=list)
-    note: str | None = Field(default=None, max_length=300)
-    clear: bool = False
+class ActivationRequest(BaseModel):
+    active: bool
+    pause_matching_automations: bool = False
+    restore_paused_automations: bool = False
+
+
+class AutomationControlRequest(BaseModel):
+    automation_entity_id: str = Field(pattern=r"^automation\.[a-z0-9_]+$")
+    enabled: bool


@router.get("/discovery", response_model=list[HaEntitySummary])
@@ -44,10 +50,12 @@ def list_configured(request: Request) -> list[ActuatorRecord]:
@router.post("", response_model=ActuatorRecord, status_code=201)
 def configure(payload: ConfigureActuatorRequest, request: Request) -> ActuatorRecord:
    try:
-        return _service(request).configure_actuator(
+        record = _service(request).configure_actuator(
            payload.actuator_entity_id,
            enabled=payload.enabled,
        )
+        _behavior(request).train(record.actuator_entity_id)
+        return _behavior(request).evaluate(record.actuator_entity_id)
    except KeyError as exc:
        raise HTTPException(status_code=404, detail=str(exc)) from exc

@@ -65,34 +73,86 @@ def delete_actuator(actuator_entity_id: str, request: Request) -> None:
    _service(request).delete_actuator(actuator_entity_id)


-@router.post("/{actuator_entity_id}/override", response_model=ActuatorRecord)
-def set_override(
-    actuator_entity_id: str,
-    payload: OverrideRequest,
-    request: Request,
-) -> ActuatorRecord:
-    override = None if payload.clear else ManualOverride(
-        numeric_entity_id=payload.numeric_entity_id,
-        context_entity_ids=payload.context_entity_ids,
-        note=payload.note,
-    )
-    try:
-        return _service(request).set_override(actuator_entity_id, override)
-    except KeyError as exc:
-        raise HTTPException(status_code=404, detail=str(exc)) from exc
-
-
@router.post("/{actuator_entity_id}/reconcile", response_model=ActuatorRecord)
 def reconcile_actuator(
    actuator_entity_id: str,
    request: Request,
 ) -> ActuatorRecord:
    try:
-        return _service(request).reconcile_actuator(actuator_entity_id, trigger="manual")
+        _service(request).reconcile_actuator(actuator_entity_id, trigger="manual")
+        _behavior(request).train(actuator_entity_id)
+        return _behavior(request).evaluate(actuator_entity_id)
    except KeyError as exc:
        raise HTTPException(status_code=404, detail=str(exc)) from exc


+@router.post("/{actuator_entity_id}/evaluate", response_model=ActuatorRecord)
+def evaluate_actuator(
+    actuator_entity_id: str,
+    request: Request,
+) -> ActuatorRecord:
+    try:
+        return _behavior(request).evaluate(actuator_entity_id)
+    except KeyError as exc:
+        raise HTTPException(status_code=404, detail=str(exc)) from exc
+
+
+@router.post("/{actuator_entity_id}/activation", response_model=ActuatorRecord)
+def set_activation(
+    actuator_entity_id: str,
+    payload: ActivationRequest,
+    request: Request,
+) -> ActuatorRecord:
+    try:
+        return _behavior(request).set_active(
+            actuator_entity_id,
+            active=payload.active,
+            pause_matching_automations=payload.pause_matching_automations,
+            restore_paused_automations=payload.restore_paused_automations,
+        )
+    except KeyError as exc:
+        raise HTTPException(status_code=404, detail=str(exc)) from exc
+    except ValueError as exc:
+        raise HTTPException(status_code=409, detail=str(exc)) from exc
+
+
+@router.post(
+    "/{actuator_entity_id}/related-automations/refresh",
+    response_model=ActuatorRecord,
+)
+def refresh_related_automations(
+    actuator_entity_id: str,
+    request: Request,
+) -> ActuatorRecord:
+    try:
+        return _behavior(request).refresh_related_automations(actuator_entity_id)
+    except KeyError as exc:
+        raise HTTPException(status_code=404, detail=str(exc)) from exc
+    except (ValueError, HaClientError) as exc:
+        raise HTTPException(status_code=409, detail=str(exc)) from exc
+
+
+@router.post(
+    "/{actuator_entity_id}/related-automations/control",
+    response_model=ActuatorRecord,
+)
+def control_related_automation(
+    actuator_entity_id: str,
+    payload: AutomationControlRequest,
+    request: Request,
+) -> ActuatorRecord:
+    try:
+        return _behavior(request).set_automation_enabled(
+            actuator_entity_id,
+            payload.automation_entity_id,
+            enabled=payload.enabled,
+        )
+    except KeyError as exc:
+        raise HTTPException(status_code=404, detail=str(exc)) from exc
+    except (ValueError, HaClientError) as exc:
+        raise HTTPException(status_code=409, detail=str(exc)) from exc
+
+
@router.get("/reconciliation/state", response_model=ReconciliationState)
 def get_reconciliation_state(request: Request) -> ReconciliationState:
    store = getattr(request.app.state, "actuator_store", None)
@@ -109,7 +169,10 @@ def run_reconciliation(
    request: Request,
    trigger: str = Query(default="manual", pattern=r"^[a-z0-9_-]{1,32}$"),
 ) -> ReconciliationState:
-    return _service(request).reconcile_all(trigger=trigger)
+    state = _service(request).reconcile_all(trigger=trigger)
+    _behavior(request).train_all()
+    _behavior(request).evaluate_all()
+    return state


 def _service(request: Request) -> ActuatorReconciliationService:
@@ -120,3 +183,13 @@ def _service(request: Request) -> ActuatorReconciliationService:
            detail="Actuator-Reconciliation nicht initialisiert.",
        )
    return service
+
+
+def _behavior(request: Request) -> BehaviorEngine:
+    engine = getattr(request.app.state, "behavior_engine", None)
+    if not isinstance(engine, BehaviorEngine):
+        raise HTTPException(
+            status_code=status.HTTP_503_SERVICE_UNAVAILABLE,
+            detail="Verhaltenslernen ist nicht initialisiert.",
+        )
+    return engine
--- a/app/behavior/init.py
+++ b/app/behavior/init.py
@@ -0,0 +1 @@
+"""Learning and prediction for actuator behavior."""
--- a/app/behavior/engine.py
+++ b/app/behavior/engine.py
@@ -0,0 +1,794 @@
+from __future__ import annotations
+
+import logging
+from datetime import datetime, timedelta, timezone
+from zoneinfo import ZoneInfo
+
+from app.actuators.models import (
+    ActuatorRecord,
+    BehaviorMode,
+    BehaviorPattern,
+    BehaviorPrediction,
+    BehaviorState,
+    BehaviorStatus,
+    ExecutionEvent,
+    RelatedAutomation,
+)
+from app.actuators.store import ActuatorStore
+from app.config import Settings
+from app.ha.exceptions import HaClientError
+from app.ha.history import LogbookEntry, StateHistoryPoint, StateHistorySeries
+from app.ha.reader import HaReader
+
+_MAX_PATTERNS = 500
+_MAX_EXECUTION_EVENTS = 100
+_ACTION_LOGBOOK_TOLERANCE = timedelta(seconds=10)
+_CONTEXT_TRIGGER_TOLERANCE = timedelta(seconds=3)
+_OWN_ACTION_TOLERANCE = timedelta(seconds=20)
+_SAFE_ACTIVE_DOMAINS = frozenset({"cover", "fan", "humidifier", "light", "switch"})
+_AUTOMATION_CONTEXT_DOMAINS = frozenset({"automation", "script"})
+logger = logging.getLogger(__name__)
+
+
+class BehaviorEngine:
+    def __init__(
+        self,
+        *,
+        ha_reader: HaReader,
+        store: ActuatorStore,
+        settings: Settings,
+    ) -> None:
+        self._ha_reader = ha_reader
+        self._store = store
+        self._settings = settings
+
+    def train_all(self) -> list[ActuatorRecord]:
+        results: list[ActuatorRecord] = []
+        for record in self._store.list():
+            try:
+                results.append(self.train(record.actuator_entity_id))
+            except Exception:
+                logger.exception("Behavior training failed for %s", record.actuator_entity_id)
+                results.append(record)
+        return results
+
+    def train(self, actuator_entity_id: str) -> ActuatorRecord:
+        record = self._store.get(actuator_entity_id)
+        now = datetime.now(timezone.utc)
+        raw_context_ids = list(
+            dict.fromkeys(
+                [
+                    record.assignment.selected_numeric_entity_id,
+                    *record.assignment.selected_context_entity_ids,
+                ]
+            )
+        )
+        context_ids = [
+            entity_id for entity_id in raw_context_ids if isinstance(entity_id, str)
+        ]
+        if not context_ids:
+            return self._save_behavior(
+                record,
+                record.behavior.model_copy(
+                    update={
+                        "status": BehaviorStatus.COLLECTING,
+                        "activation_ready": False,
+                        "activation_reason": (
+                            "Freigabe gesperrt: Noch kein geeigneter Kontext erkannt."
+                        ),
+                        "last_trained_at": now,
+                        "reason": "Noch kein geeigneter Kontext für Verhaltenslernen vorhanden.",
+                    }
+                ),
+            )
+
+        start = now - timedelta(days=self._settings.history_days)
+        history_ids = [actuator_entity_id, *context_ids]
+        try:
+            history = {
+                series.entity_id: series
+                for series in self._ha_reader.read_state_history(history_ids, start, now)
+            }
+        except (HaClientError, ValueError) as exc:
+            logger.warning("Behavior history unavailable for %s: %s", actuator_entity_id, exc)
+            return self._save_behavior(
+                record,
+                record.behavior.model_copy(
+                    update={
+                        "status": BehaviorStatus.BLOCKED,
+                        "last_trained_at": now,
+                        "reason": f"Home-Assistant-Historie konnte nicht gelesen werden: {exc}",
+                    }
+                ),
+            )
+        actuator_history = history.get(actuator_entity_id)
+        if actuator_history is None or len(actuator_history.points) < 2:
+            return self._save_behavior(
+                record,
+                record.behavior.model_copy(
+                    update={
+                        "status": BehaviorStatus.COLLECTING,
+                        "sample_count": 0,
+                        "high_confidence_sample_count": 0,
+                        "activation_ready": False,
+                        "activation_reason": (
+                            "Freigabe gesperrt: Noch keine historischen "
+                            "Aktorhandlungen gefunden."
+                        ),
+                        "patterns": [],
+                        "last_trained_at": now,
+                        "reason": "Noch keine historischen Aktorhandlungen gefunden.",
+                    }
+                ),
+            )
+
+        try:
+            logbook = list(self._ha_reader.read_logbook(actuator_entity_id, start, now))
+        except (HaClientError, ValueError) as exc:
+            logger.warning("Logbook unavailable for %s: %s", actuator_entity_id, exc)
+            logbook = []
+        patterns = self._build_patterns(
+            actuator_history=actuator_history,
+            context_history=history,
+            context_ids=context_ids,
+            logbook=logbook,
+            own_executions=record.behavior.execution_events,
+        )
+        trusted_actions = sum(
+            1 for pattern in patterns if pattern.source in {"user", "automation"}
+        )
+        status = (
+            BehaviorStatus.TRAINED
+            if len(patterns) >= self._settings.min_behavior_actions
+            else BehaviorStatus.COLLECTING
+        )
+        reason = (
+            f"{len(patterns)} Handlungen mit automatisch erfasstem Kontext gelernt."
+            if status is BehaviorStatus.TRAINED
+            else (
+                f"{len(patterns)} von mindestens {self._settings.min_behavior_actions} "
+                "benötigten Handlungen gelernt."
+            )
+        )
+        activation_ready = (
+            status is BehaviorStatus.TRAINED
+            and trusted_actions >= self._settings.min_behavior_actions
+        )
+        activation_reason = (
+            "Freigabe bereit: Genügend eindeutig zugeordnete Handlungen gelernt."
+            if activation_ready
+            else (
+                "Freigabe gesperrt: "
+                f"{max(0, self._settings.min_behavior_actions - trusted_actions)} "
+                "eindeutig zugeordnete Handlungen fehlen."
+            )
+        )
+        behavior = record.behavior.model_copy(
+            update={
+                "status": status,
+                "sample_count": len(patterns),
+                "high_confidence_sample_count": trusted_actions,
+                "activation_ready": activation_ready,
+                "activation_reason": activation_reason,
+                "patterns": patterns[-_MAX_PATTERNS:],
+                "last_trained_at": now,
+                "reason": reason,
+            }
+        )
+        return self._save_behavior(record, behavior)
+
+    def evaluate_all(self) -> list[ActuatorRecord]:
+        results: list[ActuatorRecord] = []
+        for record in self._store.list():
+            try:
+                results.append(self.evaluate(record.actuator_entity_id))
+            except Exception:
+                logger.exception("Behavior evaluation failed for %s", record.actuator_entity_id)
+                results.append(record)
+        return results
+
+    def evaluate(self, actuator_entity_id: str) -> ActuatorRecord:
+        record = self._store.get(actuator_entity_id)
+        now = datetime.now(timezone.utc)
+        try:
+            entities = {entity.entity_id: entity for entity in self._ha_reader.read_entities()}
+        except HaClientError as exc:
+            logger.warning("Current HA state unavailable for %s: %s", actuator_entity_id, exc)
+            return self._save_behavior(
+                record,
+                record.behavior.model_copy(
+                    update={
+                        "last_evaluated_at": now,
+                        "prediction": None,
+                        "reason": f"Aktueller Home-Assistant-Zustand ist nicht verfügbar: {exc}",
+                    }
+                ),
+            )
+        actuator = entities.get(actuator_entity_id)
+        if actuator is None:
+            return self._save_behavior(
+                record,
+                record.behavior.model_copy(
+                    update={
+                        "last_evaluated_at": now,
+                        "prediction": None,
+                        "reason": "Aktor ist aktuell nicht in Home Assistant verfügbar.",
+                    }
+                ),
+            )
+        current_context = {
+            entity_id: entities[entity_id].state
+            for entity_id in (
+                [
+                    record.assignment.selected_numeric_entity_id,
+                    *record.assignment.selected_context_entity_ids,
+                ]
+            )
+            if entity_id and entity_id in entities and entities[entity_id].state is not None
+        }
+        current_context_changed_at = {
+            entity_id: entities[entity_id].last_changed
+            for entity_id in current_context
+        }
+        prediction = predict_behavior(
+            record.behavior.patterns,
+            current_context=current_context,
+            current_context_changed_at=current_context_changed_at,
+            now=now,
+            min_support=self._settings.min_behavior_actions,
+            window_minutes=self._settings.prediction_window_minutes,
+            causal_window_seconds=self._settings.prediction_interval_seconds * 2,
+            timezone_name=self._settings.timezone,
+        )
+        if prediction is not None:
+            prediction = prediction.model_copy(
+                update={
+                    "execution_reason": self._prediction_execution_reason(
+                        record,
+                        actuator.state,
+                        prediction,
+                        now,
+                    )
+                }
+            )
+        behavior = record.behavior.model_copy(
+            update={
+                "last_evaluated_at": now,
+                "prediction": prediction,
+                "reason": (
+                    prediction.reason
+                    if prediction is not None
+                    else "Aktuell ist kein gelerntes Handlungsmuster fällig."
+                ),
+            }
+        )
+        if (
+            prediction is not None
+            and behavior.mode is BehaviorMode.ACTIVE
+            and prediction.confidence >= self._settings.prediction_confidence
+            and actuator.state != prediction.target_state
+            and self._cooldown_elapsed(
+                behavior,
+                now,
+                prediction.target_state,
+            )
+        ):
+            domain = actuator_entity_id.split(".", 1)[0]
+            service = service_for_state(domain, prediction.target_state)
+            if service is not None:
+                try:
+                    self._ha_reader.call_service(
+                        domain,
+                        service,
+                        {"entity_id": actuator_entity_id},
+                    )
+                except (HaClientError, ValueError) as exc:
+                    logger.error(
+                        "Predicted action failed for %s: %s",
+                        actuator_entity_id,
+                        exc,
+                    )
+                    behavior = behavior.model_copy(
+                        update={
+                            "reason": f"Vorhersage wurde aus Sicherheitsgründen nicht ausgeführt: {exc}"
+                        }
+                    )
+                    return self._save_behavior(record, behavior)
+                event = ExecutionEvent(
+                    target_state=prediction.target_state,
+                    executed_at=now,
+                )
+                behavior = behavior.model_copy(
+                    update={
+                        "prediction": prediction.model_copy(
+                            update={
+                                "executed": True,
+                                "execution_reason": (
+                                    f"Ausgeführt mit {prediction.confidence:.0%} Sicherheit."
+                                ),
+                            }
+                        ),
+                        "last_executed_at": now,
+                        "execution_events": [
+                            *behavior.execution_events,
+                            event,
+                        ][-_MAX_EXECUTION_EVENTS:],
+                        "reason": (
+                            f"Vorhersage mit {prediction.confidence:.0%} Sicherheit ausgeführt."
+                        ),
+                    }
+                )
+            else:
+                behavior = behavior.model_copy(
+                    update={
+                        "reason": (
+                            f"Der vorhergesagte Zustand {prediction.target_state!r} "
+                            "ist für autonomes Schalten nicht freigegeben."
+                        )
+                    }
+                )
+        return self._save_behavior(record, behavior)
+
+    def refresh_related_automations(self, actuator_entity_id: str) -> ActuatorRecord:
+        record = self._store.get(actuator_entity_id)
+        related = [
+            RelatedAutomation(
+                entity_id=item.entity_id,
+                config_id=item.config_id,
+                friendly_name=item.friendly_name,
+                enabled=item.enabled,
+            )
+            for item in self._ha_reader.find_automations_for_entity(
+                actuator_entity_id
+            )
+        ]
+        behavior = record.behavior.model_copy(
+            update={"related_automations": related}
+        )
+        return self._save_behavior(record, behavior)
+
+    def set_automation_enabled(
+        self,
+        actuator_entity_id: str,
+        automation_entity_id: str,
+        *,
+        enabled: bool,
+    ) -> ActuatorRecord:
+        record = self.refresh_related_automations(actuator_entity_id)
+        if automation_entity_id not in {
+            item.entity_id for item in record.behavior.related_automations
+        }:
+            raise ValueError(
+                "Die Automation ist diesem Aktor nicht eindeutig zugeordnet."
+            )
+        self._ha_reader.call_service(
+            "automation",
+            "turn_on" if enabled else "turn_off",
+            {"entity_id": automation_entity_id},
+        )
+        related = [
+            item.model_copy(update={"enabled": enabled})
+            if item.entity_id == automation_entity_id
+            else item
+            for item in record.behavior.related_automations
+        ]
+        paused = [
+            entity_id
+            for entity_id in record.behavior.paused_automation_entity_ids
+            if entity_id != automation_entity_id
+        ]
+        behavior = record.behavior.model_copy(
+            update={
+                "related_automations": related,
+                "paused_automation_entity_ids": paused,
+            }
+        )
+        return self._save_behavior(record, behavior)
+
+    def set_active(
+        self,
+        actuator_entity_id: str,
+        *,
+        active: bool,
+        pause_matching_automations: bool = False,
+        restore_paused_automations: bool = False,
+    ) -> ActuatorRecord:
+        record = self.refresh_related_automations(actuator_entity_id)
+        now = datetime.now(timezone.utc)
+        if active:
+            domain = actuator_entity_id.split(".", 1)[0]
+            if domain not in _SAFE_ACTIVE_DOMAINS:
+                raise ValueError(
+                    f"Automatisches Schalten ist für die Domain {domain} nicht freigegeben."
+                )
+            if record.behavior.status is not BehaviorStatus.TRAINED:
+                raise ValueError("Das Verhaltensmodell hat noch nicht genügend Handlungen gelernt.")
+            if not record.behavior.activation_ready:
+                raise ValueError(record.behavior.activation_reason)
+            mode = BehaviorMode.ACTIVE
+            approved_at = now
+            behavior = record.behavior.model_copy(
+                update={
+                    "mode": mode,
+                    "approved_at": approved_at,
+                    "reason": (
+                        "Autonomes Lernen und Schalten wurde ausdrücklich freigegeben."
+                    ),
+                }
+            )
+            record = self._save_behavior(record, behavior)
+            if pause_matching_automations:
+                paused: list[str] = []
+                try:
+                    for automation in record.behavior.related_automations:
+                        if not automation.enabled:
+                            continue
+                        self._ha_reader.call_service(
+                            "automation",
+                            "turn_off",
+                            {"entity_id": automation.entity_id},
+                        )
+                        paused.append(automation.entity_id)
+                except (HaClientError, ValueError):
+                    for entity_id in paused:
+                        try:
+                            self._ha_reader.call_service(
+                                "automation",
+                                "turn_on",
+                                {"entity_id": entity_id},
+                            )
+                        except (HaClientError, ValueError):
+                            logger.exception(
+                                "Failed to restore automation %s after handoff error",
+                                entity_id,
+                            )
+                    rollback = record.behavior.model_copy(
+                        update={
+                            "mode": BehaviorMode.SHADOW,
+                            "approved_at": None,
+                            "reason": (
+                                "Übernahme fehlgeschlagen; SillyHome bleibt im "
+                                "Shadow-Modus."
+                            ),
+                        }
+                    )
+                    self._save_behavior(record, rollback)
+                    raise
+                related = [
+                    automation.model_copy(update={"enabled": False})
+                    if automation.entity_id in paused
+                    else automation
+                    for automation in record.behavior.related_automations
+                ]
+                behavior = record.behavior.model_copy(
+                    update={
+                        "related_automations": related,
+                        "paused_automation_entity_ids": paused,
+                        "reason": (
+                            "SillyHome steuert aktiv; passende HA-Automationen "
+                            "wurden pausiert."
+                        ),
+                    }
+                )
+                return self._save_behavior(record, behavior)
+            return record
+        else:
+            if restore_paused_automations:
+                for entity_id in record.behavior.paused_automation_entity_ids:
+                    self._ha_reader.call_service(
+                        "automation",
+                        "turn_on",
+                        {"entity_id": entity_id},
+                    )
+            mode = BehaviorMode.SHADOW
+            approved_at = None
+            reason = (
+                "Shadow-Modus aktiv; pausierte HA-Automationen wurden fortgesetzt."
+                if restore_paused_automations
+                else "Shadow-Modus aktiv; Vorhersagen werden nicht ausgeführt."
+            )
+        behavior = record.behavior.model_copy(
+            update={
+                "mode": mode,
+                "approved_at": approved_at,
+                "related_automations": [
+                    automation.model_copy(update={"enabled": True})
+                    if (
+                        restore_paused_automations
+                        and automation.entity_id
+                        in record.behavior.paused_automation_entity_ids
+                    )
+                    else automation
+                    for automation in record.behavior.related_automations
+                ],
+                "paused_automation_entity_ids": (
+                    []
+                    if restore_paused_automations
+                    else record.behavior.paused_automation_entity_ids
+                ),
+                "reason": reason,
+            }
+        )
+        return self._save_behavior(record, behavior)
+
+    def _prediction_execution_reason(
+        self,
+        record: ActuatorRecord,
+        current_state: str | None,
+        prediction: BehaviorPrediction,
+        now: datetime,
+    ) -> str:
+        if record.behavior.mode is not BehaviorMode.ACTIVE:
+            return "Nicht ausgeführt: SillyHome ist im Shadow-Modus."
+        if prediction.confidence < self._settings.prediction_confidence:
+            return (
+                "Nicht ausgeführt: Sicherheit liegt unter der "
+                f"Schaltschwelle von {self._settings.prediction_confidence:.0%}."
+            )
+        if current_state == prediction.target_state:
+            return "Nicht ausgeführt: Zielzustand ist bereits erreicht."
+        if not self._cooldown_elapsed(
+            record.behavior,
+            now,
+            prediction.target_state,
+        ):
+            return "Nicht ausgeführt: Sicherheits-Cooldown ist noch aktiv."
+        return "Ausführung ist freigegeben."
+
+    def _build_patterns(
+        self,
+        *,
+        actuator_history: StateHistorySeries,
+        context_history: dict[str, StateHistorySeries],
+        context_ids: list[str],
+        logbook: list[LogbookEntry],
+        own_executions: list[ExecutionEvent],
+    ) -> list[BehaviorPattern]:
+        patterns: list[BehaviorPattern] = []
+        previous_state = actuator_history.points[0].state
+        for point in actuator_history.points[1:]:
+            if point.state == previous_state:
+                continue
+            previous_state = point.state
+            if _matches_own_execution(point, own_executions):
+                continue
+            source, weight = _action_source(point, logbook)
+            trigger = _recent_context_transition(
+                context_history,
+                context_ids,
+                point.timestamp,
+            )
+            contexts = {
+                entity_id: state
+                for entity_id in context_ids
+                if (state := _state_at(context_history.get(entity_id), point.timestamp)) is not None
+            }
+            local = point.timestamp.astimezone(ZoneInfo(self._settings.timezone))
+            patterns.append(
+                BehaviorPattern(
+                    target_state=point.state,
+                    minute_of_day=local.hour * 60 + local.minute,
+                    weekday=local.weekday(),
+                    context_states=contexts,
+                    trigger_entity_id=trigger[0] if trigger else None,
+                    trigger_from_state=trigger[1] if trigger else None,
+                    trigger_to_state=trigger[2] if trigger else None,
+                    source=source,
+                    weight=weight,
+                    observed_at=point.timestamp,
+                )
+            )
+        return patterns
+
+    def _cooldown_elapsed(
+        self,
+        behavior: BehaviorState,
+        now: datetime,
+        target_state: str,
+    ) -> bool:
+        if behavior.last_executed_at is None:
+            return True
+        last_event = behavior.execution_events[-1] if behavior.execution_events else None
+        if last_event is not None and last_event.target_state != target_state:
+            return True
+        return (now - behavior.last_executed_at) >= timedelta(
+            seconds=self._settings.execution_cooldown_seconds
+        )
+
+    def _save_behavior(
+        self,
+        record: ActuatorRecord,
+        behavior: BehaviorState,
+    ) -> ActuatorRecord:
+        updated = record.model_copy(
+            update={
+                "behavior": behavior,
+                "updated_at": datetime.now(timezone.utc),
+            }
+        )
+        return self._store.upsert(updated)
+
+
+def predict_behavior(
+    patterns: list[BehaviorPattern],
+    *,
+    current_context: dict[str, str | None],
+    now: datetime,
+    min_support: int,
+    window_minutes: int,
+    current_context_changed_at: dict[str, datetime | None] | None = None,
+    causal_window_seconds: int = 120,
+    timezone_name: str = "Europe/Berlin",
+) -> BehaviorPrediction | None:
+    if not patterns:
+        return None
+    local = now.astimezone(ZoneInfo(timezone_name))
+    minute_of_day = local.hour * 60 + local.minute
+    changed_at = current_context_changed_at or {}
+    by_state: dict[str, list[float]] = {}
+    causal_support_by_state: dict[str, int] = {}
+    for pattern in patterns:
+        if pattern.trigger_entity_id and pattern.trigger_to_state:
+            trigger_changed_at = changed_at.get(pattern.trigger_entity_id)
+            trigger_age = (
+                (now - trigger_changed_at).total_seconds()
+                if trigger_changed_at is not None
+                else None
+            )
+            if not (
+                current_context.get(pattern.trigger_entity_id)
+                == pattern.trigger_to_state
+                and trigger_age is not None
+                and 0 <= trigger_age <= causal_window_seconds
+            ):
+                continue
+            comparable = [
+                (entity_id, expected)
+                for entity_id, expected in pattern.context_states.items()
+                if entity_id in current_context
+            ]
+            context_score = (
+                sum(
+                    current_context[entity_id] == expected
+                    for entity_id, expected in comparable
+                )
+                / len(comparable)
+                if comparable
+                else 0.5
+            )
+            score = pattern.weight * (0.85 + 0.15 * context_score)
+            by_state.setdefault(pattern.target_state, []).append(score)
+            causal_support_by_state[pattern.target_state] = (
+                causal_support_by_state.get(pattern.target_state, 0) + 1
+            )
+            continue
+        distance = _circular_minute_distance(minute_of_day, pattern.minute_of_day)
+        if distance > window_minutes:
+            continue
+        time_score = 1.0 - (distance / max(window_minutes, 1))
+        weekday_score = (
+            1.0
+            if local.weekday() == pattern.weekday
+            else 0.5
+            if (local.weekday() >= 5) == (pattern.weekday >= 5)
+            else 0.0
+        )
+        comparable = [
+            (entity_id, expected)
+            for entity_id, expected in pattern.context_states.items()
+            if entity_id in current_context
+        ]
+        context_score = (
+            sum(current_context[entity_id] == expected for entity_id, expected in comparable)
+            / len(comparable)
+            if comparable
+            else 0.5
+        )
+        score = pattern.weight * (
+            0.45 * time_score + 0.45 * context_score + 0.10 * weekday_score
+        )
+        by_state.setdefault(pattern.target_state, []).append(score)
+    if not by_state:
+        return None
+    target_state, scores = max(
+        by_state.items(),
+        key=lambda item: (sum(item[1]), len(item[1]), item[0]),
+    )
+    support = len(scores)
+    causal_support = causal_support_by_state.get(target_state, 0)
+    confidence = min(1.0, (sum(scores) / support) * min(1.0, support / min_support))
+    if confidence <= 0:
+        return None
+    return BehaviorPrediction(
+        target_state=target_state,
+        confidence=round(confidence, 4),
+        generated_at=now,
+        matching_patterns=support,
+        reason=(
+            (
+                f"{causal_support} historische Handlungen folgten demselben "
+                "frischen Sensorwechsel."
+            )
+            if causal_support
+            else f"{support} ähnliche Handlungsmuster passen zu Zeit und aktuellem Kontext."
+        ),
+    )
+
+
+def service_for_state(domain: str, target_state: str) -> str | None:
+    if domain in {"fan", "humidifier", "light", "switch"}:
+        return {"on": "turn_on", "off": "turn_off"}.get(target_state)
+    if domain == "cover":
+        return {"open": "open_cover", "closed": "close_cover"}.get(target_state)
+    return None
+
+
+def _state_at(series: StateHistorySeries | None, timestamp: datetime) -> str | None:
+    if series is None:
+        return None
+    state: str | None = None
+    for point in series.points:
+        if point.timestamp > timestamp:
+            break
+        state = point.state
+    return state
+
+
+def _action_source(
+    point: StateHistoryPoint,
+    logbook: list[LogbookEntry],
+) -> tuple[str, float]:
+    nearest = min(
+        logbook,
+        key=lambda item: abs(item.timestamp - point.timestamp),
+        default=None,
+    )
+    if nearest is None or abs(nearest.timestamp - point.timestamp) > _ACTION_LOGBOOK_TOLERANCE:
+        return "physical_or_unknown", 0.7
+    if nearest.context_user_id:
+        return "user", 1.0
+    if nearest.context_domain in _AUTOMATION_CONTEXT_DOMAINS:
+        return "automation", 1.0
+    return "physical_or_unknown", 0.7
+
+
+def _matches_own_execution(
+    point: StateHistoryPoint,
+    own_executions: list[ExecutionEvent],
+) -> bool:
+    return any(
+        event.target_state == point.state
+        and abs(event.executed_at - point.timestamp) <= _OWN_ACTION_TOLERANCE
+        for event in own_executions
+    )
+
+
+def _recent_context_transition(
+    history: dict[str, StateHistorySeries],
+    context_ids: list[str],
+    timestamp: datetime,
+) -> tuple[str, str, str] | None:
+    nearest: tuple[timedelta, str, str, str] | None = None
+    for entity_id in context_ids:
+        series = history.get(entity_id)
+        if series is None:
+            continue
+        previous_state: str | None = None
+        for point in series.points:
+            if point.timestamp > timestamp:
+                break
+            if previous_state is not None and point.state != previous_state:
+                age = timestamp - point.timestamp
+                if age <= _CONTEXT_TRIGGER_TOLERANCE and (
+                    nearest is None or age < nearest[0]
+                ):
+                    nearest = (age, entity_id, previous_state, point.state)
+            previous_state = point.state
+    if nearest is None:
+        return None
+    return nearest[1], nearest[2], nearest[3]
+
+
+def _circular_minute_distance(left: int, right: int) -> int:
+    direct = abs(left - right)
+    return min(direct, 1440 - direct)
--- a/app/config.py
+++ b/app/config.py
@@ -15,6 +15,12 @@ class Settings:
    min_training_points: int = 24
    retrain_stale_hours: int = 24
    reconcile_interval_seconds: int = 900
+    min_behavior_actions: int = 3
+    prediction_confidence: float = 0.82
+    prediction_window_minutes: int = 30
+    prediction_interval_seconds: int = 60
+    execution_cooldown_seconds: int = 900
+    timezone: str = "Europe/Berlin"

    @property
    def ha_configured(self) -> bool:
@@ -34,4 +40,19 @@ def load_settings() -> Settings:
        reconcile_interval_seconds=max(
            60, int(os.getenv("SILLYHOME_RECONCILE_INTERVAL_SECONDS", "900"))
        ),
+        min_behavior_actions=max(2, int(os.getenv("SILLYHOME_MIN_BEHAVIOR_ACTIONS", "3"))),
+        prediction_confidence=max(
+            0.5,
+            min(0.99, float(os.getenv("SILLYHOME_PREDICTION_CONFIDENCE", "0.82"))),
+        ),
+        prediction_window_minutes=max(
+            5, min(120, int(os.getenv("SILLYHOME_PREDICTION_WINDOW_MINUTES", "30")))
+        ),
+        prediction_interval_seconds=max(
+            30, int(os.getenv("SILLYHOME_PREDICTION_INTERVAL_SECONDS", "60"))
+        ),
+        execution_cooldown_seconds=max(
+            60, int(os.getenv("SILLYHOME_EXECUTION_COOLDOWN_SECONDS", "900"))
+        ),
+        timezone=os.getenv("SILLYHOME_TIMEZONE", "Europe/Berlin"),
    )
--- a/app/ha/client.py
+++ b/app/ha/client.py
@@ -5,6 +5,7 @@ from dataclasses import dataclass
 from datetime import datetime
 import json
 import re
+from typing import Any
 from urllib.parse import quote

 import requests
@@ -19,6 +20,7 @@ from app.ha.exceptions import (
 logger = logging.getLogger(__name__)

 _ENTITY_ID_PATTERN = re.compile(r"^[a-z0-9_]+\.[a-z0-9_]+$")
+_SERVICE_PART_PATTERN = re.compile(r"^[a-z0-9_]+$")
 _MAX_HISTORY_SECONDS = 31 * 24 * 60 * 60


@@ -84,6 +86,56 @@ class HaClient:
            )
        return payload

+    def get_logbook(
+        self,
+        entity_id: str,
+        start_time: datetime,
+        end_time: datetime,
+    ) -> list[object]:
+        self._validate_period([entity_id], start_time, end_time)
+        start = quote(start_time.isoformat(), safe=":+")
+        payload = self._get_json(
+            f"/api/logbook/{start}",
+            params={
+                "entity": entity_id,
+                "end_time": end_time.isoformat(),
+            },
+        )
+        if not isinstance(payload, list):
+            raise HaUnexpectedPayloadError(
+                "Logbook-Antwort von Home Assistant hat unerwartetes Format."
+            )
+        return payload
+
+    def get_automation_config(self, automation_id: str) -> dict[str, object]:
+        if not automation_id or len(automation_id) > 120:
+            raise ValueError("Ungültige Automation-ID.")
+        payload = self._get_json(
+            f"/api/config/automation/config/{quote(automation_id, safe='')}"
+        )
+        if not isinstance(payload, dict):
+            raise HaUnexpectedPayloadError(
+                "Automation-Konfiguration hat ein unerwartetes Format."
+            )
+        return payload
+
+    def call_service(
+        self,
+        domain: str,
+        service: str,
+        service_data: dict[str, object],
+    ) -> list[object]:
+        if not _SERVICE_PART_PATTERN.fullmatch(domain):
+            raise ValueError("Ungültige Service-Domain.")
+        if not _SERVICE_PART_PATTERN.fullmatch(service):
+            raise ValueError("Ungültiger Service-Name.")
+        payload = self._post_json(f"/api/services/{domain}/{service}", service_data)
+        if not isinstance(payload, list):
+            raise HaUnexpectedPayloadError(
+                "Service-Antwort von Home Assistant hat unerwartetes Format."
+            )
+        return payload
+
    def list_entity_metadata(self, entity_ids: list[str]) -> dict[str, dict[str, str | None]]:
        if not entity_ids:
            return {}
@@ -153,6 +205,36 @@ class HaClient:

        return payload

+    def _post_json(self, path: str, payload: Any) -> object:
+        try:
+            response = self._session.post(
+                f"{self._settings.url.rstrip('/')}{path}",
+                json=payload,
+                timeout=self._settings.timeout_seconds,
+            )
+        except requests.Timeout as exc:
+            raise HaTimeoutError("Zeitüberschreitung beim Zugriff auf Home Assistant.") from exc
+        except requests.RequestException as exc:
+            raise HaHttpError(
+                getattr(getattr(exc, "response", None), "status_code", 502),
+                "Netzwerkfehler beim Zugriff auf Home Assistant.",
+            ) from exc
+        if response.status_code in (401, 403):
+            raise HaAuthError(
+                response.status_code,
+                "Authentifizierung bei Home Assistant fehlgeschlagen.",
+            )
+        try:
+            response.raise_for_status()
+        except requests.HTTPError as exc:
+            raise HaHttpError(response.status_code, "Home Assistant meldet einen Fehler.") from exc
+        try:
+            return response.json()
+        except ValueError as exc:
+            raise HaUnexpectedPayloadError(
+                "Antwort von Home Assistant ist kein gültiges JSON."
+            ) from exc
+
    def _post_text(self, path: str, payload: dict[str, str]) -> str:
        try:
            response = self._session.post(
@@ -179,6 +261,25 @@ class HaClient:
            raise HaHttpError(response.status_code, "Home Assistant meldet einen Fehler.") from exc
        return response.text

+    @staticmethod
+    def _validate_period(
+        entity_ids: list[str],
+        start_time: datetime,
+        end_time: datetime,
+    ) -> None:
+        if not entity_ids:
+            raise ValueError("Mindestens eine entity_id ist erforderlich.")
+        if len(entity_ids) > 100:
+            raise ValueError("Es können höchstens 100 Entities abgefragt werden.")
+        if any(not _ENTITY_ID_PATTERN.fullmatch(entity_id) for entity_id in entity_ids):
+            raise ValueError("entity_id enthält ein ungültiges Format.")
+        if start_time.tzinfo is None or end_time.tzinfo is None:
+            raise ValueError("start_time und end_time müssen eine Zeitzone enthalten.")
+        if end_time <= start_time:
+            raise ValueError("end_time muss nach start_time liegen.")
+        if (end_time - start_time).total_seconds() > _MAX_HISTORY_SECONDS:
+            raise ValueError("History-Abfragen sind auf 31 Tage begrenzt.")
+

 def _metadata_template(entity_ids: list[str]) -> str:
    ids = json.dumps(entity_ids, ensure_ascii=True)
--- a/app/ha/history.py
+++ b/app/ha/history.py
@@ -18,6 +18,25 @@ class EntityHistorySeries(BaseModel):
    points: list[NumericHistoryPoint]


+class StateHistoryPoint(BaseModel):
+    timestamp: datetime
+    state: str
+
+
+class StateHistorySeries(BaseModel):
+    entity_id: str
+    points: list[StateHistoryPoint]
+
+
+class LogbookEntry(BaseModel):
+    entity_id: str
+    timestamp: datetime
+    message: str = ""
+    context_user_id: str | None = None
+    context_domain: str | None = None
+    context_service: str | None = None
+
+
 def normalize_history_payload(payload: object) -> list[EntityHistorySeries]:
    if not isinstance(payload, list):
        raise HaUnexpectedPayloadError("History-Payload muss eine Liste sein.")
@@ -33,6 +52,68 @@ def normalize_history_payload(payload: object) -> list[EntityHistorySeries]:
    return sorted(normalized, key=lambda item: item.entity_id)


+def normalize_state_history_payload(payload: object) -> list[StateHistorySeries]:
+    if not isinstance(payload, list):
+        raise HaUnexpectedPayloadError("History-Payload muss eine Liste sein.")
+    normalized: list[StateHistorySeries] = []
+    for raw_series in payload:
+        if not isinstance(raw_series, list):
+            raise HaUnexpectedPayloadError("History-Serie muss eine Liste sein.")
+        entity_id: str | None = None
+        points: list[StateHistoryPoint] = []
+        for raw_entry in raw_series:
+            if not isinstance(raw_entry, dict):
+                raise HaUnexpectedPayloadError("History-Eintrag muss ein Objekt sein.")
+            raw_entity_id = raw_entry.get("entity_id")
+            if raw_entity_id is not None:
+                if not isinstance(raw_entity_id, str) or "." not in raw_entity_id:
+                    raise HaUnexpectedPayloadError(
+                        "History-Eintrag enthält ungültige entity_id."
+                    )
+                if entity_id is not None and entity_id != raw_entity_id:
+                    raise HaUnexpectedPayloadError("History-Serie enthält mehrere Entities.")
+                entity_id = raw_entity_id
+            raw_state = raw_entry.get("state")
+            if not isinstance(raw_state, str) or raw_state in {"unknown", "unavailable"}:
+                continue
+            if entity_id is None:
+                raise HaUnexpectedPayloadError("History-Serie enthält keine entity_id.")
+            timestamp = _parse_timestamp(
+                raw_entry.get("last_changed") or raw_entry.get("last_updated")
+            )
+            if not points or points[-1].state != raw_state:
+                points.append(StateHistoryPoint(timestamp=timestamp, state=raw_state))
+        if entity_id is not None and points:
+            points.sort(key=lambda point: point.timestamp)
+            normalized.append(StateHistorySeries(entity_id=entity_id, points=points))
+    return sorted(normalized, key=lambda item: item.entity_id)
+
+
+def normalize_logbook_payload(payload: object, entity_id: str) -> list[LogbookEntry]:
+    if not isinstance(payload, list):
+        raise HaUnexpectedPayloadError("Logbook-Payload muss eine Liste sein.")
+    entries: list[LogbookEntry] = []
+    for raw_entry in payload:
+        if not isinstance(raw_entry, dict):
+            raise HaUnexpectedPayloadError("Logbook-Eintrag muss ein Objekt sein.")
+        raw_entity_id = raw_entry.get("entity_id")
+        if raw_entity_id != entity_id:
+            continue
+        entries.append(
+            LogbookEntry(
+                entity_id=entity_id,
+                timestamp=_parse_timestamp(raw_entry.get("when")),
+                message=str(raw_entry.get("message") or ""),
+                context_user_id=_optional_string(raw_entry.get("context_user_id")),
+                context_domain=_optional_string(
+                    raw_entry.get("context_domain") or raw_entry.get("domain")
+                ),
+                context_service=_optional_string(raw_entry.get("context_service")),
+            )
+        )
+    return sorted(entries, key=lambda item: item.timestamp)
+
+
 def _normalize_series(raw_series: list[object]) -> EntityHistorySeries | None:
    entity_id: str | None = None
    points: list[NumericHistoryPoint] = []
@@ -89,3 +170,9 @@ def _parse_timestamp(value: object) -> datetime:
    if parsed.tzinfo is None:
        raise HaUnexpectedPayloadError("History-Zeitstempel muss eine Zeitzone enthalten.")
    return parsed
+
+
+def _optional_string(value: object) -> str | None:
+    if value is None or value == "":
+        return None
+    return str(value)
--- a/app/ha/models.py
+++ b/app/ha/models.py
@@ -1,5 +1,7 @@
 from __future__ import annotations

+from datetime import datetime
+
 from pydantic import BaseModel


@@ -14,6 +16,8 @@ class HaState(BaseModel):
 class HaEntitySummary(BaseModel):
    entity_id: str
    domain: str
+    state: str | None = None
+    last_changed: datetime | None = None
    state_class: str | None = None
    device_class: str | None = None
    unit_of_measurement: str | None = None
@@ -22,3 +26,10 @@ class HaEntitySummary(BaseModel):
    area_name: str | None = None
    device_id: str | None = None
    device_name: str | None = None
+
+
+class HaAutomationSummary(BaseModel):
+    entity_id: str
+    config_id: str
+    friendly_name: str
+    enabled: bool
--- a/app/ha/reader.py
+++ b/app/ha/reader.py
@@ -1,7 +1,8 @@
 from __future__ import annotations

 from collections.abc import Sequence
-from datetime import datetime
+from datetime import datetime, timedelta, timezone
+from threading import RLock
 from typing import Any
 import logging

@@ -9,8 +10,15 @@ from app.ha.exceptions import HaClientError

 from app.ha.client import HaClient
 from app.ha.discovery import DiscoveredEntity, discover_entities
-from app.ha.history import EntityHistorySeries, normalize_history_payload
-from app.ha.models import HaEntitySummary
+from app.ha.history import (
+    EntityHistorySeries,
+    LogbookEntry,
+    StateHistorySeries,
+    normalize_history_payload,
+    normalize_logbook_payload,
+    normalize_state_history_payload,
+)
+from app.ha.models import HaAutomationSummary, HaEntitySummary

 logger = logging.getLogger(__name__)

@@ -18,6 +26,11 @@ logger = logging.getLogger(__name__)
 class HaReader:
    def __init__(self, client: HaClient) -> None:
        self._client = client
+        self._automation_cache: list[
+            tuple[HaAutomationSummary, dict[str, object]]
+        ] = []
+        self._automation_cache_at: datetime | None = None
+        self._automation_cache_lock = RLock()

    def read_entities(self) -> Sequence[HaEntitySummary]:
        entities = self._client.list_entities()
@@ -45,6 +58,8 @@ class HaReader:
                HaEntitySummary(
                    entity_id=entity_id,
                    domain=domain,
+                    state=_optional_str(item.get("state")),
+                    last_changed=_optional_datetime(item.get("last_changed")),
                    state_class=_optional_str(attributes.get("state_class")),
                    device_class=_optional_str(attributes.get("device_class")),
                    unit_of_measurement=_optional_str(attributes.get("unit_of_measurement")),
@@ -77,8 +92,127 @@ class HaReader:
        payload = self._client.get_history(entity_ids, start_time, end_time)
        return normalize_history_payload(payload)

+    def read_state_history(
+        self,
+        entity_ids: list[str],
+        start_time: datetime,
+        end_time: datetime,
+    ) -> Sequence[StateHistorySeries]:
+        payload = self._client.get_history(entity_ids, start_time, end_time)
+        return normalize_state_history_payload(payload)
+
+    def read_logbook(
+        self,
+        entity_id: str,
+        start_time: datetime,
+        end_time: datetime,
+    ) -> Sequence[LogbookEntry]:
+        payload = self._client.get_logbook(entity_id, start_time, end_time)
+        return normalize_logbook_payload(payload, entity_id)
+
+    def call_service(
+        self,
+        domain: str,
+        service: str,
+        service_data: dict[str, object],
+    ) -> Sequence[object]:
+        return self._client.call_service(domain, service, service_data)
+
+    def find_automations_for_entity(
+        self,
+        entity_id: str,
+    ) -> list[HaAutomationSummary]:
+        current_states = {
+            raw_entity_id: item.get("state") == "on"
+            for item in self._client.list_entities()
+            if isinstance((raw_entity_id := item.get("entity_id")), str)
+            and raw_entity_id.startswith("automation.")
+        }
+        matches = [
+            summary.model_copy(
+                update={
+                    "enabled": current_states.get(
+                        summary.entity_id,
+                        summary.enabled,
+                    )
+                }
+            )
+            for summary, config in self._read_automation_configs()
+            if _contains_exact_value(config, entity_id)
+        ]
+        return sorted(matches, key=lambda item: item.entity_id)
+
+    def _read_automation_configs(
+        self,
+    ) -> list[tuple[HaAutomationSummary, dict[str, object]]]:
+        now = datetime.now(timezone.utc)
+        with self._automation_cache_lock:
+            if (
+                self._automation_cache_at is not None
+                and now - self._automation_cache_at < timedelta(minutes=10)
+            ):
+                return list(self._automation_cache)
+            configs: list[tuple[HaAutomationSummary, dict[str, object]]] = []
+            for item in self._client.list_entities():
+                raw_entity_id = item.get("entity_id")
+                if not isinstance(raw_entity_id, str) or not raw_entity_id.startswith(
+                    "automation."
+                ):
+                    continue
+                attributes = item.get("attributes")
+                if not isinstance(attributes, dict):
+                    continue
+                config_id = attributes.get("id")
+                if not isinstance(config_id, str) or not config_id:
+                    continue
+                try:
+                    config = self._client.get_automation_config(config_id)
+                except (HaClientError, ValueError) as exc:
+                    logger.warning(
+                        "Automation config unavailable for %s: %s",
+                        raw_entity_id,
+                        exc,
+                    )
+                    continue
+                configs.append(
+                    (
+                        HaAutomationSummary(
+                            entity_id=raw_entity_id,
+                            config_id=config_id,
+                            friendly_name=str(
+                                attributes.get("friendly_name") or raw_entity_id
+                            ),
+                            enabled=item.get("state") == "on",
+                        ),
+                        config,
+                    )
+                )
+            self._automation_cache = configs
+            self._automation_cache_at = now
+            return list(configs)
+

 def _optional_str(value: object) -> str | None:
    if value is None or value == "":
        return None
    return str(value)
+
+
+def _optional_datetime(value: object) -> datetime | None:
+    if not isinstance(value, str) or not value:
+        return None
+    try:
+        parsed = datetime.fromisoformat(value.replace("Z", "+00:00"))
+    except ValueError:
+        return None
+    return parsed if parsed.tzinfo is not None else None
+
+
+def _contains_exact_value(value: object, expected: str) -> bool:
+    if value == expected:
+        return True
+    if isinstance(value, dict):
+        return any(_contains_exact_value(item, expected) for item in value.values())
+    if isinstance(value, list):
+        return any(_contains_exact_value(item, expected) for item in value)
+    return False
--- a/app/main.py
+++ b/app/main.py
@@ -12,8 +12,7 @@ from app.actuators.lifecycle import ActuatorReconciliationService
 from app.actuators.store import ActuatorStore
 from app.api.v1.actuators import router as actuators_router
 from app.api.v1.entities import router as entities_router
-from app.api.v1.automations import router as automations_router
-from app.automations.store import AutomationStore
+from app.behavior.engine import BehaviorEngine
 from app.config import load_settings
 from app.core.exception_handlers import register_exception_handlers
 from app.ha.client import HaClient, HaClientSettings
@@ -27,13 +26,15 @@ async def lifespan(app: FastAPI) -> AsyncIterator[None]:
    settings = app.state.settings
    client: HaClient | None = None
    reconcile_task: asyncio.Task[None] | None = None
+    prediction_task: asyncio.Task[None] | None = None
    app.state.registry = ModelRegistry(settings.model_store)
-    app.state.automation_store = AutomationStore(settings.automation_store)
    app.state.actuator_store = ActuatorStore(settings.actuator_store)
    if hasattr(app.state, "ha_reader"):
        del app.state.ha_reader
    if hasattr(app.state, "actuator_service"):
        del app.state.actuator_service
+    if hasattr(app.state, "behavior_engine"):
+        del app.state.behavior_engine
    if settings.ha_configured:
        client = HaClient(
            settings=HaClientSettings(
@@ -48,8 +49,16 @@ async def lifespan(app: FastAPI) -> AsyncIterator[None]:
            registry=app.state.registry,
            settings=settings,
        )
+        app.state.behavior_engine = BehaviorEngine(
+            ha_reader=app.state.ha_reader,
+            store=app.state.actuator_store,
+            settings=settings,
+        )
        await asyncio.to_thread(app.state.actuator_service.reconcile_all, "startup")
+        await asyncio.to_thread(app.state.behavior_engine.train_all)
+        await asyncio.to_thread(app.state.behavior_engine.evaluate_all)
        reconcile_task = asyncio.create_task(_periodic_reconciliation(app))
+        prediction_task = asyncio.create_task(_periodic_prediction(app))
    try:
        yield
    finally:
@@ -57,6 +66,10 @@ async def lifespan(app: FastAPI) -> AsyncIterator[None]:
            reconcile_task.cancel()
            with suppress(asyncio.CancelledError):
                await reconcile_task
+        if prediction_task is not None:
+            prediction_task.cancel()
+            with suppress(asyncio.CancelledError):
+                await prediction_task
        if client is not None:
            client.close()

@@ -64,13 +77,12 @@ async def lifespan(app: FastAPI) -> AsyncIterator[None]:
 app = FastAPI(
    title="SillyHome Next API",
    description="Lokales Smart-Home-Intelligenzsystem für Home Assistant.",
-    version="0.4.0",
+    version="0.7.0",
    lifespan=lifespan,
 )
 app.state.settings = load_settings()
 register_exception_handlers(app)
 app.include_router(entities_router)
-app.include_router(automations_router)
 app.include_router(actuators_router)
 init_ml_routes(app, model_store=app.state.settings.model_store)

@@ -85,7 +97,10 @@ def health() -> dict[str, str]:

@app.get("/")
 def root() -> FileResponse:
-    return FileResponse(STATIC_DIR / "index.html")
+    return FileResponse(
+        STATIC_DIR / "index.html",
+        headers={"Cache-Control": "no-store, max-age=0"},
+    )


 async def _periodic_reconciliation(app: FastAPI) -> None:
@@ -95,3 +110,15 @@ async def _periodic_reconciliation(app: FastAPI) -> None:
        if not isinstance(service, ActuatorReconciliationService):
            continue
        await asyncio.to_thread(service.reconcile_all, "scheduled")
+        engine = getattr(app.state, "behavior_engine", None)
+        if isinstance(engine, BehaviorEngine):
+            await asyncio.to_thread(engine.train_all)
+
+
+async def _periodic_prediction(app: FastAPI) -> None:
+    while True:
+        await asyncio.sleep(app.state.settings.prediction_interval_seconds)
+        engine = getattr(app.state, "behavior_engine", None)
+        if not isinstance(engine, BehaviorEngine):
+            continue
+        await asyncio.to_thread(engine.evaluate_all)
--- a/app/static/index.html
+++ b/app/static/index.html
@@ -7,97 +7,136 @@
  <style>
    :root { color-scheme: dark; font-family: system-ui, sans-serif; background: #10151c; color: #eaf1f8; }
    body { margin: 0; }
-    header { padding: 20px; background: linear-gradient(135deg,#142b3a,#193f36); }
+    header { padding: 22px; background: linear-gradient(135deg,#142b3a,#193f36); }
    h1,h2,h3 { margin: 0 0 12px; }
-    header p { margin: 4px 0; color: #b9c9d6; }
+    header p { margin: 5px 0; color: #c3d1dc; }
    main { display: grid; grid-template-columns: repeat(auto-fit,minmax(320px,1fr)); gap: 14px; padding: 14px; }
    section { background: #18212b; border: 1px solid #2d3a47; border-radius: 12px; padding: 16px; }
    .wide { grid-column: 1 / -1; }
+    .steps { display:grid; grid-template-columns:repeat(auto-fit,minmax(220px,1fr)); gap:12px; }
+    .step { background:#111a23; border:1px solid #31404d; border-radius:10px; padding:14px; }
+    .step-number { display:inline-grid; place-items:center; width:28px; height:28px; border-radius:50%; background:#23715b; font-weight:700; margin-bottom:8px; }
+    .step p { margin:5px 0; }
    .ok { color: #66dfa9; }
    .warn { color: #f3c969; }
    .bad { color: #ff8f8f; }
    label { display: block; margin: 9px 0 4px; color: #b9c9d6; }
-    input,select,textarea,button { box-sizing: border-box; width: 100%; border-radius: 7px; border: 1px solid #3b4b5b; padding: 9px; background: #101820; color: #fff; }
+    select,input,button { box-sizing: border-box; width: 100%; border-radius: 7px; border: 1px solid #3b4b5b; padding: 10px; background: #101820; color: #fff; }
    button { margin-top: 10px; background: #23715b; border: 0; font-weight: 700; cursor: pointer; }
    button.secondary { background: #37495c; }
-    pre { white-space: pre-wrap; overflow: auto; background: #0d141b; padding: 10px; border-radius: 7px; }
-    table { width: 100%; border-collapse: collapse; font-size: .9rem; }
-    td,th { padding: 7px; border-bottom: 1px solid #2d3a47; text-align: left; vertical-align: top; }
+    button.danger { background: #7b3434; }
+    table { width: 100%; border-collapse: collapse; font-size: .92rem; }
+    td,th { padding: 8px; border-bottom: 1px solid #2d3a47; text-align: left; vertical-align: top; }
    ul { margin: 8px 0; padding-left: 18px; }
-    .notice { border-left: 4px solid #e8b34b; padding-left: 10px; }
-    .grid-two { display:grid; grid-template-columns:repeat(auto-fit,minmax(220px,1fr)); gap:8px; }
+    .notice { border-left: 4px solid #66dfa9; padding-left: 10px; }
+    .grid-two { display:grid; grid-template-columns:repeat(auto-fit,minmax(240px,1fr)); gap:12px; }
    .chips { display:flex; flex-wrap:wrap; gap:6px; margin-top:8px; }
    .chip { padding:4px 8px; border-radius:999px; background:#22303c; border:1px solid #31404d; font-size:.85rem; }
+    .muted { color:#9fb0be; }
  </style>
 </head>
 <body>
 <header>
  <h1>SillyHome Next</h1>
-  <p>Aktuator-zentrierte Home-Assistant-Analyse mit nachvollziehbarer Sensorzuordnung und kontrolliertem Modell-Lebenszyklus.</p>
-  <p class="notice">Sicherheitsmodus: SillyHome führt niemals selbst Aktor-Services aus. Automationen bleiben manuell freizugebende YAML-Entwürfe.</p>
+  <p>Hier wählst du nur Geräte aus, deren Bedienung SillyHome lernen soll. Sensoren, Zusammenhänge und Modelle werden automatisch verwaltet.</p>
+  <p class="notice">Sicherer Start: Zuerst wird nur beobachtet und vorhergesagt. Ohne deine spätere Freigabe wird nichts geschaltet.</p>
 </header>
 <main>
-  <section>
-    <h2>Systemstatus</h2>
-    <div id="status">Prüfung läuft ...</div>
-    <div class="chips" id="status-chips"></div>
-    <button class="secondary" onclick="loadOverview()">Neu laden</button>
-    <button onclick="runReconciliation()">Reconciliation ausführen</button>
+  <section class="wide">
+    <h2>So gehst du vor</h2>
+    <div class="steps">
+      <div class="step">
+        <span class="step-number">1</span>
+        <h3>Aktor auswählen</h3>
+        <p><strong>Wo?</strong> Unten im Feld „Gerät auswählen“.</p>
+        <p><strong>Was passiert?</strong> SillyHome ordnet Raum, Sensoren, Zustände und vorhandene Historie automatisch zu.</p>
+      </div>
+      <div class="step">
+        <span class="step-number">2</span>
+        <h3>Wie gewohnt bedienen</h3>
+        <p><strong>Wo?</strong> Weiterhin in Home Assistant, an Schaltern oder über deine bisherigen Bedienwege.</p>
+        <p><strong>Was passiert?</strong> SillyHome lernt deine Handlungen und zeigt Vorhersagen an, schaltet aber noch nicht selbst.</p>
+      </div>
+      <div class="step">
+        <span class="step-number">3</span>
+        <h3>Später freigeben</h3>
+        <p><strong>Wo?</strong> In den Details des ausgewählten Geräts, sobald genug Verhalten gelernt wurde.</p>
+        <p><strong>Was passiert?</strong> Erst dann darf SillyHome passende Vorhersagen automatisch ausführen. Die Freigabe kann jederzeit gestoppt werden.</p>
+      </div>
+    </div>
  </section>

  <section>
-    <h2>Aktuator wählen</h2>
-    <label for="actuator-select">Home-Assistant-Aktor</label>
-    <select id="actuator-select"></select>
-    <button onclick="configureActuator()">Aktuator übernehmen</button>
-    <pre id="actuator-config-result">Noch kein Aktuator konfiguriert.</pre>
+    <h2>Systemstatus</h2>
+    <p class="muted">Zeigt, ob Verbindung, Lernsystem und automatische Prüfungen funktionieren. Hier musst du normalerweise nichts einstellen.</p>
+    <div id="status">Prüfung läuft ...</div>
+    <div class="chips" id="status-chips"></div>
+    <button class="secondary" onclick="loadOverview()">Status aktualisieren</button>
+  </section>
+
+  <section>
+    <h2>1. Gerät zum Lernen auswählen</h2>
+    <p class="muted">Wähle eine Lampe, einen Rollladen oder einen anderen unterstützten Aktor. Du wählst keine Sensoren und erstellst keine Regeln.</p>
+    <label for="actuator-input">Entitätsname oder Gerät aus Home Assistant</label>
+    <input id="actuator-input" list="actuator-options" placeholder="z. B. light.licht_abstellraum" autocomplete="off">
+    <datalist id="actuator-options"></datalist>
+    <button onclick="configureActuator()">Gerät hinzufügen und Beobachtung starten</button>
+    <p id="actuator-config-result" class="muted">Noch kein Aktor ausgewählt.</p>
  </section>

  <section class="wide">
-    <h2>Konfigurierte Aktuatoren</h2>
+    <h2>2. Beobachtete Geräte</h2>
+    <p class="muted">Öffne „Details“, um Lernfortschritt, aktuelle Vorhersage und den automatisch gefundenen Kontext zu sehen.</p>
    <div id="configured-actuators">Noch nicht geladen.</div>
  </section>

  <section class="wide">
-    <h2>Zuordnung und Modellstatus</h2>
-    <div id="actuator-detail">Einen konfigurierten Aktuator auswählen.</div>
-  </section>
-
-  <section class="wide">
-    <h2>Automation-Entwurf</h2>
-    <p>Der Entwurf muss explizit freigegeben werden. Auch danach wird nur YAML exportiert, nichts geschaltet.</p>
-    <div class="grid-two">
-      <div><label for="alias">Name</label><input id="alias" value="Licht bei Dunkelheit"></div>
-      <div><label for="trigger">Trigger-Entity</label><input id="trigger" placeholder="sensor.flur_illuminance"></div>
-      <div><label for="below">Unter Grenzwert</label><input id="below" type="number" value="10"></div>
-      <div><label for="service">Dienst</label><select id="service"><option>light.turn_on</option><option>light.turn_off</option><option>switch.turn_on</option><option>switch.turn_off</option></select></div>
-      <div><label for="target">Ziel-Entity</label><input id="target" placeholder="light.flur"></div>
-    </div>
-    <button onclick="createProposal()">Entwurf speichern</button>
-    <button class="secondary" onclick="loadProposals()">Entwürfe aktualisieren</button>
-    <div id="proposals"></div>
+    <h2>3. Lernfortschritt und Freigabe</h2>
+    <p class="muted">Die Freigabe erscheint erst, wenn genug eindeutig zugeordnete Handlungen gelernt wurden. Vorher bleibt das Gerät sicher im Beobachtungsmodus.</p>
+    <div id="actuator-detail" class="muted">Öffne bei einem beobachteten Gerät die Details.</div>
  </section>
 </main>
 <script>
-const pretty = value => JSON.stringify(value, null, 2);
+const escapeHtml = value => String(value ?? "")
+  .replaceAll("&", "&amp;")
+  .replaceAll("<", "&lt;")
+  .replaceAll(">", "&gt;")
+  .replaceAll('"', "&quot;")
+  .replaceAll("'", "&#039;");
 let currentActuatorId = null;

 async function api(path, options = {}) {
  const response = await fetch(path, {headers: {"Content-Type": "application/json"}, ...options});
-  const body = await response.json().catch(() => ({}));
-  if (!response.ok) throw new Error(body.detail || `${response.status} ${response.statusText}`);
+  const body = response.status === 204 ? null : await response.json().catch(() => ({}));
+  if (!response.ok) throw new Error(body?.detail || `${response.status} ${response.statusText}`);
  return body;
 }

+function lifecycleLabel(record) {
+  if (record.behavior.status === "trained") return "Kontext erkannt";
+  const labels = {
+    trained: "lernt",
+    pending_history: "sammelt Historie",
+    pending_assignment: "sucht Kontext",
+    review_required: "geringe Zuordnungssicherheit",
+    archived: "wartet auf Kontext",
+    orphaned: "Aktor nicht gefunden",
+  };
+  return labels[record.lifecycle.status] || record.lifecycle.status;
+}
+
 function statusClass(record) {
-  if (record.assignment.review_required) return "warn";
+  if (record.behavior.status === "trained") return "ok";
  if (record.lifecycle.status === "trained") return "ok";
-  if (record.lifecycle.status === "review_required" || record.lifecycle.status === "invalid") return "warn";
+  if (["pending_history", "pending_assignment", "archived"].includes(record.lifecycle.status)) return "warn";
  return "bad";
 }

-function renderEvidence(evidence) {
-  return evidence.length ? `<ul>${evidence.map(item => `<li>${item}</li>`).join("")}</ul>` : "<span class='bad'>Keine Evidenz</span>";
+function behaviorLabel(record) {
+  if (record.behavior.mode === "active") return "aktiv freigegeben";
+  if (record.behavior.status === "trained") return "Shadow-Vorhersage";
+  if (record.behavior.status === "blocked") return "Lernen blockiert";
+  return "sammelt Handlungen";
 }

 async function loadOverview() {
@@ -110,57 +149,53 @@ async function loadOverview() {
      api("v1/actuators/reconciliation/state"),
      api("v1/actuators"),
    ]);
-    status.innerHTML = `<p class="ok">API und ML bereit</p><p>Letzte Reconciliation: ${reconciliation.last_completed_at || "noch nie"}</p><p>${reconciliation.last_summary}</p>`;
+    status.innerHTML = `<p class="ok">System bereit</p><p>Letzte automatische Prüfung: ${escapeHtml(reconciliation.last_completed_at || "noch nie")}</p>`;
    chips.innerHTML = [
-      `<span class="chip">Health: ${health.status}</span>`,
-      `<span class="chip">ML: ${ml.status}</span>`,
-      `<span class="chip">Aktuatoren: ${actuators.length}</span>`,
-      `<span class="chip">Trainierte Modelle: ${reconciliation.trained_models}</span>`,
+      `<span class="chip">API: ${escapeHtml(health.status)}</span>`,
+      `<span class="chip">Lernsystem: ${escapeHtml(ml.status)}</span>`,
+      `<span class="chip">Aktoren: ${actuators.length}</span>`,
+      `<span class="chip">Lernbereite Geräte: ${reconciliation.trained_models}</span>`,
    ].join("");
  } catch (error) {
-    status.innerHTML = `<p class="bad">${error.message}</p>`;
+    status.innerHTML = `<p class="bad">${escapeHtml(error.message)}</p>`;
    chips.innerHTML = "";
  }
-  await Promise.all([loadActuatorDiscovery(), loadConfiguredActuators(), loadProposals()]);
+  await Promise.all([loadActuatorDiscovery(), loadConfiguredActuators()]);
 }

 async function loadActuatorDiscovery() {
-  const select = document.getElementById("actuator-select");
+  const options = document.getElementById("actuator-options");
  try {
-    const actuators = await api("v1/actuators/discovery");
-    select.innerHTML = actuators.length
-      ? actuators.map(entity => `<option value="${entity.entity_id}">${entity.friendly_name || entity.entity_id}${entity.area_name ? ` (${entity.area_name})` : ""}</option>`).join("")
-      : "<option value=''>Keine Aktuatoren gefunden</option>";
+    const [available, configured] = await Promise.all([
+      api("v1/actuators/discovery"),
+      api("v1/actuators"),
+    ]);
+    const configuredIds = new Set(configured.map(record => record.actuator_entity_id));
+    const choices = available.filter(entity => !configuredIds.has(entity.entity_id));
+    options.innerHTML = choices.map(entity =>
+      `<option value="${escapeHtml(entity.entity_id)}">${escapeHtml(entity.friendly_name || entity.entity_id)}${entity.area_name ? ` (${escapeHtml(entity.area_name)})` : ""}</option>`
+    ).join("");
  } catch (error) {
-    select.innerHTML = `<option value="">${error.message}</option>`;
+    options.innerHTML = "";
  }
 }

 async function configureActuator() {
-  const actuatorId = document.getElementById("actuator-select").value;
-  const box = document.getElementById("actuator-config-result");
+  const actuatorId = document.getElementById("actuator-input").value.trim();
+  const result = document.getElementById("actuator-config-result");
  if (!actuatorId) return;
+  result.textContent = "Kontext wird automatisch analysiert ...";
  try {
    const record = await api("v1/actuators", {
      method: "POST",
      body: JSON.stringify({actuator_entity_id: actuatorId}),
    });
    currentActuatorId = record.actuator_entity_id;
-    box.textContent = pretty(record);
+    result.textContent = `${record.actuator_entity_id}: ${lifecycleLabel(record)}.`;
    await loadOverview();
    await showActuator(record.actuator_entity_id);
  } catch (error) {
-    box.textContent = error.message;
-  }
-}
-
-async function runReconciliation() {
-  try {
-    await api("v1/actuators/reconciliation/run", {method: "POST"});
-    await loadOverview();
-    if (currentActuatorId) await showActuator(currentActuatorId);
-  } catch (error) {
-    alert(error.message);
+    result.textContent = error.message;
  }
 }

@@ -170,185 +205,188 @@ async function loadConfiguredActuators() {
    const rows = await api("v1/actuators");
    box.innerHTML = rows.length ? `
      <table>
-        <tr><th>Aktuator</th><th>Numerischer Sensor</th><th>Review</th><th>Modellstatus</th><th>Letztes Training</th><th>Aktion</th></tr>
+        <tr><th>Gerät</th><th>Lernstatus</th><th>Freigabe</th><th>Gelernte Handlungen</th><th>Letzte Vorhersage</th><th>Aktionen</th></tr>
        ${rows.map(record => `
          <tr>
-            <td>${record.actuator_entity_id}</td>
-            <td>${record.assignment.selected_numeric_entity_id || "-"}</td>
-            <td class="${record.assignment.review_required ? "warn" : "ok"}">${record.assignment.review_required ? "erforderlich" : "nein"}</td>
-            <td class="${statusClass(record)}">${record.lifecycle.status}</td>
-            <td>${record.lifecycle.last_trained_at || "-"}</td>
-            <td><button onclick="showActuator('${record.actuator_entity_id}')">Details</button></td>
+            <td>${escapeHtml(record.actuator_entity_id)}</td>
+            <td class="${record.behavior.status === "trained" ? "ok" : "warn"}">${escapeHtml(behaviorLabel(record))}</td>
+            <td class="${record.behavior.activation_ready ? "ok" : "warn"}">${escapeHtml(record.behavior.activation_ready ? "bereit" : record.behavior.activation_reason)}</td>
+            <td>${record.behavior.sample_count}</td>
+            <td>${record.behavior.prediction
+              ? `${escapeHtml(record.behavior.prediction.target_state)} (${Math.round(record.behavior.prediction.confidence * 100)} %)`
+              : "-"}</td>
+            <td>
+              <button onclick="showActuator('${escapeHtml(record.actuator_entity_id)}')">Details</button>
+              ${record.behavior.mode === "active"
+                ? `<button class="danger" onclick="setActivation('${escapeHtml(record.actuator_entity_id)}', false, false, true)">Stoppen + HA-Automationen fortsetzen</button>`
+                : record.behavior.activation_ready
+                  ? `<button onclick="setActivation('${escapeHtml(record.actuator_entity_id)}', true, true, false)">SillyHome übernehmen lassen</button>`
+                  : ""}
+              <button class="danger" onclick="removeActuator('${escapeHtml(record.actuator_entity_id)}')">Entfernen</button>
+            </td>
          </tr>
        `).join("")}
-      </table>` : "<p>Keine konfigurierten Aktuatoren.</p>";
+      </table>` : "<p>Noch keine Aktoren ausgewählt.</p>";
  } catch (error) {
    box.textContent = error.message;
  }
 }

-async function showActuator(actuatorId) {
+async function showActuator(actuatorId, evaluationMessage = "") {
  currentActuatorId = actuatorId;
  const box = document.getElementById("actuator-detail");
  try {
-    const record = await api(`v1/actuators/${encodeURIComponent(actuatorId)}`);
-    const numericRows = record.numeric_candidates.map(candidate => `
-      <tr>
-        <td>${candidate.entity_id}</td>
-        <td>${candidate.score.toFixed(3)} / ${candidate.confidence.toFixed(2)}</td>
-        <td>${candidate.auto_accepted ? "<span class='ok'>auto</span>" : "<span class='warn'>review</span>"}</td>
-        <td>${renderEvidence(candidate.evidence)}</td>
-      </tr>
-    `).join("");
-    const contextRows = record.context_candidates.map(candidate => `
-      <tr>
-        <td>${candidate.entity_id}</td>
-        <td>${candidate.score.toFixed(3)} / ${candidate.confidence.toFixed(2)}</td>
-        <td>${candidate.auto_accepted ? "<span class='ok'>auto</span>" : "<span class='warn'>optional</span>"}</td>
-        <td>${renderEvidence(candidate.evidence)}</td>
-      </tr>
-    `).join("");
+    let record;
+    try {
+      record = await api(`v1/actuators/${encodeURIComponent(actuatorId)}/related-automations/refresh`, {method: "POST"});
+    } catch (_) {
+      record = await api(`v1/actuators/${encodeURIComponent(actuatorId)}`);
+    }
+    const contexts = [
+      record.assignment.selected_numeric_entity_id,
+      ...record.assignment.selected_context_entity_ids,
+    ].filter(Boolean);
+    const evidence = [...record.numeric_candidates, ...record.context_candidates]
+      .filter(candidate => contexts.includes(candidate.entity_id))
+      .map(candidate => `<li><strong>${escapeHtml(candidate.friendly_name || candidate.entity_id)}</strong>: ${candidate.evidence.map(escapeHtml).join(", ") || "statistisch relevanter Kandidat"}</li>`)
+      .join("");
+    const prediction = record.behavior.prediction;
+    const learnedAutomationActions = record.behavior.patterns.filter(
+      pattern => pattern.source === "automation",
+    ).length;
+    const relatedAutomations = record.behavior.related_automations || [];
+    const activationButton = record.behavior.mode === "active"
+      ? `<button class="danger" onclick="setActivation('${escapeHtml(record.actuator_entity_id)}', false, false, true)">SillyHome stoppen und pausierte HA-Automationen fortsetzen</button>
+         <button class="secondary" onclick="setActivation('${escapeHtml(record.actuator_entity_id)}', false, false, false)">SillyHome stoppen; HA-Automationen pausiert lassen</button>`
+      : record.behavior.activation_ready
+        ? `<button onclick="setActivation('${escapeHtml(record.actuator_entity_id)}', true, true, false)">SillyHome übernehmen lassen und passende HA-Automationen pausieren</button>
+           <button class="secondary" onclick="setActivation('${escapeHtml(record.actuator_entity_id)}', true, false, false)">SillyHome parallel aktivieren</button>`
+        : `<p class='warn'>${escapeHtml(record.behavior.activation_reason)}</p>`;
+    const automationControls = relatedAutomations.length
+      ? `<ul>${relatedAutomations.map(automation => `
+          <li>
+            <strong>${escapeHtml(automation.friendly_name)}</strong>
+            <code>${escapeHtml(automation.entity_id)}</code>:
+            <span class="${automation.enabled ? "ok" : "warn"}">${automation.enabled ? "aktiv" : "pausiert"}</span>
+            <button class="secondary" onclick="setRelatedAutomation('${escapeHtml(record.actuator_entity_id)}', '${escapeHtml(automation.entity_id)}', ${automation.enabled ? "false" : "true"})">${automation.enabled ? "Pausieren" : "Fortsetzen"}</button>
+          </li>`).join("")}</ul>`
+      : "<p class='muted'>Keine eindeutig passende HA-Automation gefunden.</p>";
    box.innerHTML = `
      <div class="grid-two">
        <div>
-          <h3>Auswahl</h3>
-          <p><strong>Aktuator:</strong> ${record.actuator_entity_id}</p>
-          <p><strong>Numerischer Sensor:</strong> ${record.assignment.selected_numeric_entity_id || "-"}</p>
-          <p><strong>Kontext:</strong> ${record.assignment.selected_context_entity_ids.join(", ") || "-"}</p>
-          <p><strong>Quelle:</strong> ${record.assignment.source}</p>
-          <p><strong>Review:</strong> <span class="${record.assignment.review_required ? "warn" : "ok"}">${record.assignment.review_required ? "erforderlich" : "nicht erforderlich"}</span></p>
-          <p><strong>Begruendung:</strong> ${record.assignment.reason}</p>
+          <h3>${escapeHtml(record.actuator_entity_id)}</h3>
+          <p><strong>Status:</strong> <span class="${statusClass(record)}">${escapeHtml(lifecycleLabel(record))}</span></p>
+          <p><strong>Kontextzuordnung:</strong> automatisch erledigt</p>
+          <p><strong>Zuordnungssicherheit:</strong> ${Math.round(record.assignment.confidence * 100)} %</p>
+          <p class="muted">Dieser Wert beschreibt, wie sicher Raum, Sensoren und Zustände zu diesem Gerät passen.</p>
+          <p><strong>Ergebnis:</strong> ${escapeHtml(record.assignment.reason)}</p>
        </div>
        <div>
-          <h3>Modell-Lebenszyklus</h3>
-          <p><strong>Status:</strong> <span class="${statusClass(record)}">${record.lifecycle.status}</span></p>
-          <p><strong>Letztes Training:</strong> ${record.lifecycle.last_trained_at || "-"}</p>
-          <p><strong>Messpunkte:</strong> ${record.lifecycle.last_history_point_count}</p>
-          <p><strong>Grund:</strong> ${record.lifecycle.reason}</p>
-          <p><strong>Nächste Aktion:</strong> ${record.lifecycle.next_action}</p>
-          <button onclick="reconcileActuator('${record.actuator_entity_id}')">Diesen Aktuator erneut prüfen</button>
+          <h3>Lernfortschritt</h3>
+          <p><strong>Betriebsart:</strong> ${escapeHtml(behaviorLabel(record))}</p>
+          <p><strong>Gelernte Handlungen:</strong> ${record.behavior.sample_count}</p>
+          <p><strong>Davon eindeutig geregelt:</strong> ${record.behavior.high_confidence_sample_count}</p>
+          <p><strong>Davon erkannte HA-Automationen:</strong> ${learnedAutomationActions}</p>
+          <p><strong>Letztes Training:</strong> ${escapeHtml(record.behavior.last_trained_at || "noch nicht")}</p>
+          <p><strong>Was noch passiert:</strong> ${escapeHtml(record.behavior.reason)}</p>
+          <p><strong>Freigabestatus:</strong> <span class="${record.behavior.activation_ready ? "ok" : "warn"}">${escapeHtml(record.behavior.activation_reason)}</span></p>
+          ${activationButton}
+          <button class="secondary" onclick="evaluateActuator('${escapeHtml(record.actuator_entity_id)}')">Aktuelle Situation auswerten</button>
+          <p class="muted">Die Prüfung simuliert keinen Sensorwechsel und schaltet keinen Aktor.</p>
+          ${evaluationMessage ? `<p class="ok">${escapeHtml(evaluationMessage)}</p>` : ""}
        </div>
      </div>
-      <div class="grid-two">
-        <div>
-          <h3>Manuelle Overrides</h3>
-          <label for="override-numeric">Numerischer Sensor</label>
-          <input id="override-numeric" value="${record.manual_override?.numeric_entity_id || record.assignment.selected_numeric_entity_id || ""}">
-          <label for="override-context">Kontext-Entities (kommagetrennt)</label>
-          <textarea id="override-context">${(record.manual_override?.context_entity_ids || record.assignment.selected_context_entity_ids || []).join(", ")}</textarea>
-          <label for="override-note">Notiz</label>
-          <input id="override-note" value="${record.manual_override?.note || ""}">
-          <button onclick="saveOverride('${record.actuator_entity_id}')">Override speichern</button>
-          <button class="secondary" onclick="clearOverride('${record.actuator_entity_id}')">Override löschen</button>
-        </div>
-        <div>
-          <h3>Audit</h3>
-          <pre>${pretty(record.lifecycle.audit)}</pre>
-        </div>
-      </div>
-      <h3>Numerische Kandidaten</h3>
-      ${numericRows ? `<table><tr><th>Entity</th><th>Score / Confidence</th><th>Auto</th><th>Evidenz</th></tr>${numericRows}</table>` : "<p>Keine Kandidaten.</p>"}
-      <h3>Kontext-Kandidaten</h3>
-      ${contextRows ? `<table><tr><th>Entity</th><th>Score / Confidence</th><th>Auto</th><th>Evidenz</th></tr>${contextRows}</table>` : "<p>Keine Kandidaten.</p>"}
+      <h3>Was SillyHome aktuell vorhersagt</h3>
+      ${prediction
+        ? `<p><strong>${escapeHtml(prediction.target_state)}</strong> mit ${Math.round(prediction.confidence * 100)} % Sicherheit. ${escapeHtml(prediction.reason)} <span class="${prediction.executed ? "ok" : "muted"}">${escapeHtml(prediction.execution_reason)}</span></p>`
+        : "<p class='muted'>Aktuell ist kein gelerntes Handlungsmuster fällig.</p>"}
+      <h3>Passende Home-Assistant-Automationen</h3>
+      <p class="muted">Bei einer Übernahme pausiert SillyHome diese Automationen. Beim Stoppen können sie gezielt fortgesetzt werden.</p>
+      ${automationControls}
+      <h3>Welche Zusammenhänge automatisch verwendet werden</h3>
+      ${evidence ? `<ul>${evidence}</ul>` : "<p class='warn'>Noch kein geeigneter Kontext erkannt. SillyHome prüft bei neuen HA-Daten erneut.</p>"}
    `;
  } catch (error) {
    box.textContent = error.message;
  }
 }

-async function reconcileActuator(actuatorId) {
+async function evaluateActuator(actuatorId) {
  try {
-    await api(`v1/actuators/${encodeURIComponent(actuatorId)}/reconcile`, {method: "POST"});
-    await loadOverview();
-    await showActuator(actuatorId);
+    const record = await api(
+      `v1/actuators/${encodeURIComponent(actuatorId)}/evaluate`,
+      {method: "POST"},
+    );
+    const checkedAt = new Date(
+      record.behavior.last_evaluated_at || Date.now(),
+    ).toLocaleString("de-DE");
+    const message = record.behavior.prediction
+      ? `Prüfung ${checkedAt}: ${record.behavior.prediction.target_state} mit ${Math.round(record.behavior.prediction.confidence * 100)} % vorhergesagt.`
+      : `Prüfung ${checkedAt}: Kein frischer passender Sensorwechsel erkannt; aktuell ist keine Aktion fällig.`;
+    await loadConfiguredActuators();
+    await showActuator(actuatorId, message);
  } catch (error) {
    alert(error.message);
  }
 }

-async function saveOverride(actuatorId) {
-  const numeric = document.getElementById("override-numeric").value.trim() || null;
-  const contexts = document.getElementById("override-context").value
-    .split(",")
-    .map(item => item.trim())
-    .filter(Boolean);
-  const note = document.getElementById("override-note").value.trim() || null;
+async function setActivation(actuatorId, active, pauseMatchingAutomations, restorePausedAutomations) {
+  const question = active
+    ? pauseMatchingAutomations
+      ? `${actuatorId}: SillyHome aktivieren und passende HA-Automationen pausieren?`
+      : `${actuatorId}: SillyHome parallel zu den HA-Automationen aktivieren?`
+    : restorePausedAutomations
+      ? `${actuatorId}: SillyHome stoppen und pausierte HA-Automationen fortsetzen?`
+      : `${actuatorId}: SillyHome stoppen und HA-Automationen pausiert lassen?`;
+  if (!confirm(question)) return;
  try {
-    await api(`v1/actuators/${encodeURIComponent(actuatorId)}/override`, {
+    await api(`v1/actuators/${encodeURIComponent(actuatorId)}/activation`, {
      method: "POST",
      body: JSON.stringify({
-        numeric_entity_id: numeric,
-        context_entity_ids: contexts,
-        note,
+        active,
+        pause_matching_automations: pauseMatchingAutomations,
+        restore_paused_automations: restorePausedAutomations,
      }),
    });
-    await loadOverview();
+    await loadConfiguredActuators();
    await showActuator(actuatorId);
  } catch (error) {
    alert(error.message);
  }
 }

-async function clearOverride(actuatorId) {
+async function setRelatedAutomation(actuatorId, automationEntityId, enabled) {
+  const action = enabled ? "fortsetzen" : "pausieren";
+  if (!confirm(`${automationEntityId} wirklich ${action}?`)) return;
  try {
-    await api(`v1/actuators/${encodeURIComponent(actuatorId)}/override`, {
+    await api(`v1/actuators/${encodeURIComponent(actuatorId)}/related-automations/control`, {
      method: "POST",
-      body: JSON.stringify({clear: true}),
+      body: JSON.stringify({
+        automation_entity_id: automationEntityId,
+        enabled,
+      }),
    });
-    await loadOverview();
+    await loadConfiguredActuators();
    await showActuator(actuatorId);
  } catch (error) {
    alert(error.message);
  }
 }

-async function createProposal() {
+async function removeActuator(actuatorId) {
+  if (!confirm(`${actuatorId} aus SillyHome entfernen?`)) return;
  try {
-    await api("v1/automations/proposals", {method: "POST", body: JSON.stringify({
-      alias: document.getElementById("alias").value,
-      description: "Manuell im SillyHome-Dashboard erstellter und nicht automatisch ausgeführter Entwurf.",
-      trigger: {entity_id: document.getElementById("trigger").value, below: Number(document.getElementById("below").value)},
-      action: {service: document.getElementById("service").value, entity_id: document.getElementById("target").value, data: {}}
-    })});
-    await loadProposals();
+    await api(`v1/actuators/${encodeURIComponent(actuatorId)}`, {method: "DELETE"});
+    if (currentActuatorId === actuatorId) {
+      currentActuatorId = null;
+      document.getElementById("actuator-detail").textContent = "Öffne bei einem beobachteten Gerät die Details.";
+    }
+    await loadOverview();
  } catch (error) {
    alert(error.message);
  }
 }

-async function decide(id, revision, action) {
-  try {
-    await api(`v1/automations/proposals/${id}/${action}`, {method: "POST", body: JSON.stringify({expected_revision: revision})});
-    await loadProposals();
-  } catch (error) {
-    alert(error.message);
-  }
-}
-
-async function loadProposals() {
-  const box = document.getElementById("proposals");
-  try {
-    const rows = await api("v1/automations/proposals");
-    box.innerHTML = rows.length ? `
-      <table>
-        <tr><th>Name</th><th>Status</th><th>Aktion</th></tr>
-        ${rows.map(item => `
-          <tr>
-            <td>${item.alias}</td>
-            <td>${item.status}</td>
-            <td>${item.status === "draft"
-              ? `<button onclick="decide('${item.proposal_id}',${item.revision},'approve')">Freigeben</button><button class="secondary" onclick="decide('${item.proposal_id}',${item.revision},'reject')">Ablehnen</button>`
-              : item.status === "approved"
-                ? `<a href="v1/automations/proposals/${item.proposal_id}/yaml">YAML laden</a>`
-                : "-"}</td>
-          </tr>
-        `).join("")}
-      </table>` : "<p>Keine Entwürfe.</p>";
-  } catch (error) {
-    box.textContent = error.message;
-  }
-}
-
 loadOverview();
 </script>
 </body>
--- a/docker-compose.yml
+++ b/docker-compose.yml
@@ -14,6 +14,12 @@ services:
      SILLYHOME_MIN_TRAINING_POINTS: 24
      SILLYHOME_RETRAIN_STALE_HOURS: 24
      SILLYHOME_RECONCILE_INTERVAL_SECONDS: 900
+      SILLYHOME_MIN_BEHAVIOR_ACTIONS: 3
+      SILLYHOME_PREDICTION_CONFIDENCE: 0.82
+      SILLYHOME_PREDICTION_WINDOW_MINUTES: 30
+      SILLYHOME_PREDICTION_INTERVAL_SECONDS: 60
+      SILLYHOME_EXECUTION_COOLDOWN_SECONDS: 900
+      SILLYHOME_TIMEZONE: Europe/Berlin
    volumes:
      - model-data:/app/data/models
      - automation-data:/app/data/automations
--- a/docs/BEHAVIOR_ENGINE.md
+++ b/docs/BEHAVIOR_ENGINE.md
@@ -0,0 +1,73 @@
+# Verhaltensmodell und Berechnung
+
+## Datenfluss
+
+1. Nutzer wählt einen Aktor.
+2. `ActuatorReconciliationService` ordnet Kontext-Entities zu.
+3. `BehaviorEngine.train()` liest Aktor- und Kontexthistorie.
+4. Aktor-Zustandswechsel werden als `BehaviorPattern` gespeichert.
+5. `BehaviorEngine.evaluate()` vergleicht aktuelle Zustände mit den Mustern.
+6. Shadow zeigt nur die Vorhersage. Active darf sie ausführen.
+
+## Herkunft und Gewicht
+
+- HA-Benutzer: `source=user`, Gewicht `1.0`
+- eindeutig erkannte HA-Automation oder Script: `source=automation`, Gewicht `1.0`
+- physisch oder unbekannt: `source=physical_or_unknown`, Gewicht `0.7`
+- eigene SillyHome-Ausführung: wird verworfen
+
+Manuelle und eindeutig automatisierte Handlungen zählen für die Freigabe.
+
+## Kausale Muster
+
+Wechselt ein Kontextsensor höchstens drei Sekunden vor der Aktorhandlung, wird
+der Wechsel gespeichert:
+
+```text
+binary_sensor.tuer: off -> on
+light.raum: off -> on
+```
+
+Eine kausale Vorhersage gilt nur, wenn derselbe Kontextzustand frisch ist. Das
+Standardfenster ist zweimal `SILLYHOME_PREDICTION_INTERVAL_SECONDS`.
+
+## Nicht-kausale Bewertung
+
+Für Muster ohne frischen Trigger:
+
+```text
+score = weight * (
+  0.45 * time_score
+  + 0.45 * context_score
+  + 0.10 * weekday_score
+)
+```
+
+Die Confidence ist der mittlere Score, begrenzt durch die Mindestunterstützung:
+
+```text
+confidence = mean(scores) * min(1, support / min_behavior_actions)
+```
+
+## Ausführungsbedingungen
+
+Eine Vorhersage wird nur ausgeführt, wenn alle Bedingungen erfüllt sind:
+
+- Betriebsart `active`
+- Confidence mindestens `SILLYHOME_PREDICTION_CONFIDENCE`
+- Zielzustand ist noch nicht erreicht
+- Domain und Zustand sind erlaubt
+- Cooldown erlaubt die Aktion
+
+Der Cooldown sperrt nur eine schnelle Wiederholung desselben Zielzustands.
+Eine Gegenaktion, beispielsweise `on` gefolgt von `off`, bleibt sofort erlaubt.
+
+## Freigabe
+
+`activation_ready=true`, wenn:
+
+- Verhaltensstatus `trained`
+- mindestens `SILLYHOME_MIN_BEHAVIOR_ACTIONS` eindeutig zugeordnete manuelle
+  oder automatisierte Handlungen vorhanden sind
+
+Die UI zeigt `activation_reason` immer an.
--- a/docs/CONTROL_HANDOFF.md
+++ b/docs/CONTROL_HANDOFF.md
@@ -0,0 +1,47 @@
+# Übergabe zwischen SillyHome und HA-Automationen
+
+## Erkennung
+
+SillyHome liest aktive `automation.*`-Entities, lädt deren Konfiguration über
+die Home-Assistant-API und sucht darin nach der exakten Aktor-Entity-ID.
+Namensähnlichkeit allein reicht nicht.
+
+## Betriebsarten
+
+### Shadow
+
+- SillyHome lernt und prognostiziert.
+- SillyHome schaltet nicht.
+- HA-Automationen können normal weiterlaufen.
+
+### Active parallel
+
+- SillyHome darf schalten.
+- Passende HA-Automationen bleiben aktiv.
+- Diese Betriebsart kann doppelte Auslöser verursachen und ist nur für Tests.
+
+### Active mit Übernahme
+
+- SillyHome wird zuerst aktiviert.
+- Danach werden aktuell aktive, passend erkannte HA-Automationen pausiert.
+- Nur erfolgreich pausierte Automationen werden für eine spätere
+  Wiederherstellung gespeichert.
+- Scheitert die Pause, fällt SillyHome auf Shadow zurück und stellt bereits
+  pausierte Automationen wieder her.
+
+## Stoppen
+
+Zwei bewusste Optionen:
+
+- SillyHome stoppen und pausierte HA-Automationen fortsetzen.
+- SillyHome stoppen und HA-Automationen pausiert lassen.
+
+Einzelne passende Automationen können im Dashboard jederzeit pausiert oder
+fortgesetzt werden.
+
+In Home Assistant bedeutet:
+
+```text
+automation.turn_off = pausieren/deaktivieren
+automation.turn_on  = fortsetzen/aktivieren
+```
--- a/docs/DEBUGGING.md
+++ b/docs/DEBUGGING.md
@@ -0,0 +1,63 @@
+# Debugging
+
+## Vorhersage korrekt, aber keine Ausführung
+
+1. Aktor-Details öffnen.
+2. `Betriebsart` prüfen.
+3. `Freigabestatus` prüfen.
+4. Text hinter der Vorhersage lesen. `execution_reason` nennt exakt:
+   - Shadow-Modus
+   - Confidence unter Schaltschwelle
+   - Zielzustand bereits erreicht
+   - Cooldown aktiv
+   - ausgeführt
+5. Live-Zustand des Aktors und Triggers in HA prüfen.
+6. Add-on-Logs prüfen.
+
+## Weder SillyHome noch HA-Automation schaltet
+
+1. SillyHome-Modus prüfen.
+2. Unter `Passende Home-Assistant-Automationen` den Zustand prüfen.
+3. Bei Shadow mindestens eine gewünschte HA-Automation fortsetzen.
+4. Bei Active mit Übernahme müssen die passenden HA-Automationen pausiert sein.
+
+## Freigabe fehlt
+
+Die UI zeigt den Grund immer als `activation_reason`.
+
+Prüfen:
+
+```text
+behavior.status
+behavior.sample_count
+behavior.high_confidence_sample_count
+behavior.activation_ready
+behavior.activation_reason
+```
+
+## Entität fehlt in der Liste
+
+Den vollständigen Entitätsnamen direkt eingeben. Der Server akzeptiert nur
+existierende, unterstützte Aktoren. Ein unbekannter Name liefert `404`.
+
+## Standarddiagnose lokal
+
+```bash
+.venv/bin/pytest tests/behavior/test_engine.py -q
+.venv/bin/pytest tests/api/test_actuators.py -q
+.venv/bin/ruff check app tests
+.venv/bin/mypy app backend tests
+```
+
+## Standarddiagnose im HA-Add-on
+
+```bash
+ha apps info 58adbe1e_sillyhome_next
+ha apps logs 58adbe1e_sillyhome_next
+```
+
+Health aus einem Add-on mit Zugriff auf das interne Netz:
+
+```bash
+wget -qO- http://58adbe1e-sillyhome-next:8000/health
+```
--- a/docs/OPERATIONS.md
+++ b/docs/OPERATIONS.md
@@ -0,0 +1,87 @@
+# Entwicklung, Release und Betrieb
+
+## Lokales Setup
+
+```bash
+python3 -m venv .venv
+.venv/bin/pip install -e '.[dev]'
+cp .env.example .env
+.venv/bin/uvicorn app.main:app --reload
+```
+
+`SILLYHOME_HA_URL` und `SILLYHOME_HA_TOKEN` nur lokal in `.env` setzen.
+
+## Qualitätsprüfung
+
+```bash
+.venv/bin/pytest -q
+.venv/bin/ruff check .
+.venv/bin/mypy app backend tests
+git diff --check
+```
+
+## Release
+
+1. Version in allen vier Stellen ändern:
+   `pyproject.toml`, `addon/config.yaml`, `app/main.py`, `CHANGELOG.md`.
+2. Qualitätsprüfung ausführen.
+3. Feature-Branch committen und pushen.
+4. Pull Request nach `main` erstellen und mergen.
+5. Annotiertes Tag auf dem Merge-Commit erstellen.
+6. Gitea-Release aus demselben Tag erstellen.
+
+Beispiel:
+
+```bash
+git tag -a v0.7.0 -m 'SillyHome Next 0.7.0'
+git push origin v0.7.0
+```
+
+## Home-Assistant-Update
+
+Vorher Teil-Backup des Add-ons erstellen. Danach:
+
+```bash
+ha store reload
+ha apps info 58adbe1e_sillyhome_next
+ha apps update 58adbe1e_sillyhome_next
+ha apps info 58adbe1e_sillyhome_next
+ha apps logs 58adbe1e_sillyhome_next
+```
+
+Kein Home-Assistant-Neustart ist erforderlich.
+
+## Live-Verifikation
+
+Pflicht:
+
+```bash
+wget -qO- http://58adbe1e-sillyhome-next:8000/health
+wget -qO- http://58adbe1e-sillyhome-next:8000/v1/actuators
+```
+
+Für einen Aktor prüfen:
+
+- `behavior.mode`
+- `behavior.activation_ready`
+- `behavior.activation_reason`
+- `behavior.related_automations`
+- `behavior.paused_automation_entity_ids`
+- `behavior.prediction.execution_reason`
+
+Bei einer Übernahme testen:
+
+1. Passende HA-Automation ist vorher `on`.
+2. SillyHome übernimmt.
+3. SillyHome ist `active`.
+4. Passende HA-Automation ist `off`.
+5. Trigger erzeugt erwartete Aktoraktion.
+6. Gegenaktion wird trotz Cooldown ausgeführt.
+7. SillyHome stoppen und Automationen fortsetzen.
+8. SillyHome ist `shadow`, HA-Automation wieder `on`.
+
+## Rollback
+
+Bevorzugt das vor dem Update erstellte HA-Teil-Backup wiederherstellen.
+Alternativ vorherige Git-Version in `addon/config.yaml` veröffentlichen und das
+Add-on erneut aktualisieren.
--- a/docs/automations.md
+++ b/docs/automations.md
@@ -1,14 +1,6 @@
-# Automation-Vorschläge
+# Keine manuell erzeugten Automationen

-SillyHome Next führt Automationen niemals automatisch aus. Der Workflow ist:
-
-1. Vorschlag als `draft` erstellen.
-2. Inhalt und Ziel-Entity prüfen.
-3. Mit aktueller Revision explizit freigeben oder ablehnen.
-4. Nur freigegebene Vorschläge als Home-Assistant-YAML exportieren.
-5. Das YAML außerhalb von SillyHome Next in Home Assistant importieren.
-
-Erlaubt sind numerische Sensor-Trigger und Aktionsdienste aus den Domains
-`light`, `switch`, `climate`, `fan` und `cover`. Shell-Kommandos, Skripte und
-beliebige Service-Domains werden abgewiesen. Eine einmal getroffene Entscheidung
-kann nicht überschrieben werden; Änderungen benötigen einen neuen Vorschlag.
+Seit `v0.5.0` erstellt SillyHome Next keine YAML-Automationen und bietet keinen
+Regel- oder Trigger-Editor mehr an. Der produktive Ablauf besteht aus
+Aktorauswahl, automatischem Verhaltenslernen, Shadow-Vorhersage und einer
+separaten Ausführungsfreigabe pro Aktor.
--- a/docs/ml_api.md
+++ b/docs/ml_api.md
@@ -178,9 +178,10 @@ Listet unterstützte Aktuatoren mit angereicherter HA-Metadatenbasis.

 ### `POST /v1/actuators`

-Registriert einen Aktuator, ermittelt passende numerische Sensoren und
-Kontext-Entities, trainiert bei ausreichender History automatisch ein Modell und
-liefert Zuordnung, Confidence, Evidenz und Lifecycle-Status zurück.
+Registriert einen Aktor. Das System ermittelt passende Messwerte und
+Kontext-Entities vollständig automatisch, trainiert bei ausreichender Historie
+ein Modell und liefert Zuordnung, Confidence, Evidenz und Lifecycle-Status zur
+Diagnose zurück.

 **Request**
 ```json
@@ -190,21 +191,63 @@ liefert Zuordnung, Confidence, Evidenz und Lifecycle-Status zurück.
 }
 ```

-### `POST /v1/actuators/{actuator_entity_id}/override`
-
-Persistiert manuelle Overrides. Diese haben Vorrang vor der automatischen
-Heuristik und überstehen Neustarts.
-
 ### `POST /v1/actuators/reconciliation/run`

 Führt eine sichere globale Reconciliation aus. Die periodische Add-on-Schleife
 ruft denselben idempotenten Ablauf auf, startet aber keine Services in Home
 Assistant.

+### `POST /v1/actuators/{actuator_entity_id}/evaluate`
+
+Erstellt aus aktuellem Kontext eine neue Shadow- oder Aktiv-Vorhersage. Im
+Shadow-Modus wird niemals geschaltet.
+
+### `POST /v1/actuators/{actuator_entity_id}/activation`
+
+```json
+{
+  "active": true,
+  "pause_matching_automations": true,
+  "restore_paused_automations": false
+}
+```
+
+Aktiviert autonomes Schalten erst nach ausreichendem Training und nur für
+erlaubte Aktor-Domains. `pause_matching_automations` pausiert eindeutig
+zugeordnete HA-Automationen bei der Übernahme.
+
+Beim Stoppen:
+
+```json
+{
+  "active": false,
+  "pause_matching_automations": false,
+  "restore_paused_automations": true
+}
+```
+
+Damit wird der Aktor in den Shadow-Modus versetzt und zuvor von SillyHome
+pausierte Automationen werden fortgesetzt.
+
+### `POST /v1/actuators/{actuator_entity_id}/related-automations/refresh`
+
+Liest passende HA-Automationen anhand ihrer echten Konfiguration neu ein.
+
+### `POST /v1/actuators/{actuator_entity_id}/related-automations/control`
+
+```json
+{
+  "automation_entity_id": "automation.licht_abstellkammer",
+  "enabled": false
+}
+```
+
+Pausiert oder aktiviert eine eindeutig diesem Aktor zugeordnete Automation.
+
 ## Betrieb

-Die produktive App lädt Artefakte aus `SILLYHOME_MODEL_STORE`. Aktuator-,
-Override- und Reconciliation-Zustände liegen atomisch in
+Die produktive App lädt Artefakte aus `SILLYHOME_MODEL_STORE`. Aktor- und
+Reconciliation-Zustände liegen atomisch in
 `SILLYHOME_ACTUATOR_STORE`. Neue Artefakte werden über `/ml/retrain`,
 `RetrainingService` oder den aktor-zentrierten Lifecycle registriert. Die API
 sollte nur in einem vertrauenswürdigen Netz oder hinter einem
--- a/docs/ml_training.md
+++ b/docs/ml_training.md
@@ -1,92 +1,51 @@
-# ML Training- und Evaluations-Workflow
+# Verhaltenslernen und Vorhersage

-SillyHome Next trainiert ein lokales statistisches Baseline-Modell pro Sensor
-und Merkmal. Es benötigt keine Cloud und keine externe ML-Laufzeit.
+Seit `v0.5.0` ist der produktive Lernpfad aktor-zentriert. Nutzer wählen nur
+einen Aktor; Sensoren, Kontext und Modelle werden automatisch verwaltet.

-Seit `v0.4.0` ist der bevorzugte Weg aktor-zentriert: ein bestätigter Aktuator
-wird mit einem numerischen Primärsensor verknüpft, die Historie dieses Sensors
-wird automatisch geladen und in ein deterministisches Artefakt überführt.
+## Datengrundlage

-## 1. Daten sammeln
+Für jeden Aktor lädt SillyHome Next:

-Alle Trainingsvektoren werden über `FeatureStore.add(...)` oder `add_batch(...)` eingepflegt. Jeder Vektor enthält eine Sensor-ID sowie ein Dictionary mit Merkmalen.
+- dessen Zustandswechsel aus der Home-Assistant-Historie
+- Logbook-Einträge zur Herkunft der Handlung
+- automatisch zugeordnete Mess- und Kontext-Entities
+- deren Zustand zum Zeitpunkt der Handlung

-Im Normalbetrieb erzeugt die Reconciliation diese Vektoren selbst aus realer
-Home-Assistant-History. Das Trainingsmerkmal heißt dabei immer `value`.
-Binäre Kontextsensoren bleiben Kontext und werden nicht als numerische Samples
-missverstanden.
+Eindeutig einem Home-Assistant-Benutzer zugeordnete Handlungen und im Logbuch
+erkannte Automations- oder Script-Aktionen erhalten das höchste Gewicht.
+Physische oder nicht eindeutig zuordenbare Bedienungen dürfen das Shadow-Modell
+ergänzen, reichen allein aber nicht zur Aktivierung.

-## 2. Statistisches Artefakt erzeugen
+## Modell

-```python
-store = FeatureStore()
-store.add(FeatureVector(sensor_id="sensor.kitchen", values={"temperature": 21.0}))
-pipeline = TrainingPipeline(store)
-artifact = pipeline.run("my_artifact")
-pipeline.export("my_artifact")
-```
+Das lokale Modell speichert pro beobachteter Handlung:

-`TrainingPipeline.run(...)` berechnet für jedes numerische Merkmal:
+- Zielzustand
+- lokale Tageszeit
+- Wochentag
+- Kontextzustände
+- Herkunft und Gewicht

- Stichprobenzahl
- Mittelwert und Standardabweichung
- Minimum und Maximum
- linearen Trend mit Steigung und Achsenabschnitt
+Eine Vorhersage bewertet zeitliche Nähe, Wochentagsmuster und aktuellen
+Kontext. Mehrere passende historische Handlungen erhöhen die Confidence.

-Die nächste Vorhersage kombiniert den letzten beobachteten Wert mit der
-trainierten Trendsteigung. Die Confidence berücksichtigt Datenmenge und
-Stabilität.
+## Betriebsstufen

-## 3. Modell evaluieren
+1. `collecting`: Noch nicht genügend Handlungen vorhanden.
+2. `shadow`: Modell ist trainiert; Vorhersagen werden angezeigt, aber nicht ausgeführt.
+3. `active`: Nutzer hat den Aktor ausdrücklich freigegeben.

-```python
-evaluator = Evaluator(pipeline)
-report = evaluator.evaluate(artifact.artifact_id, validation_samples)
-```
+Die Aktivierung verlangt genügend eindeutig zugeordnete manuelle oder
+automatisierte Handlungen. Ausgeführt werden nur erlaubte Zustände reversibler Domains:
+`light`, `switch`, `fan`, `humidifier` und `cover`.

-Der Report enthält echte numerische Vergleichsmetriken:
- `artifact_id`
- `sample_size`
- `mae` (Mean Absolute Error)
- `rmse` (Root Mean Squared Error)
- `coverage` für den Anteil auswertbarer Merkmale
+## Schutzmechanismen

-## 4. Modell registrieren
-
-Das trainierte Artefakt kann anschließend über `ModelRegistry.register(artifact)` bereitgestellt werden. Die ML-Serving-API stellt es unter `/ml/predict` und `/ml/batch` zur Verfügung.
-
-## 5. Retraining ausführen
-
-`RetrainingService.retrain(...)` führt genau einen Trainingslauf aus und ersetzt
-ein vorhandenes Artefakt mit derselben ID atomisch in der Registry:
-
-```python
-service = RetrainingService(registry)
-result = service.retrain("home-model", vectors)
-```
-
-Scheduler, Cronjobs oder Home-Assistant-Automationen können alternativ die
-zustandslose Funktion `retrain_model(registry, artifact_id, vectors)` aufrufen.
-Der Service startet bewusst keinen eigenen Hintergrundprozess. Über
-`POST /ml/retrain` kann derselbe Ablauf per API angestoßen werden.
-
-## 6. Autonomer Lebenszyklus
-
-Der `ActuatorReconciliationService` verwaltet pro konfiguriertem Aktuator:
-
- die automatische Sensor- und Kontextzuordnung mit Score, Confidence und Evidenz
- persistente manuelle Overrides
- den Modellstatus (`trained`, `pending_history`, `review_required`, `archived`, ...)
- ein Audit-Protokoll mit Gründen für Training, Retraining oder Archivierung
-
-Retraining erfolgt nur, wenn:
-
- genügend nutzbare numerische Historie vorliegt
- die aktuelle Zuordnung eindeutig oder manuell bestätigt ist
- die Historie sich materiell verändert hat oder das Modell als stale gilt
-
-## Hinweise
- Für reproduzierbare Sensor-Reihenfolgen wird in `TrainingPipeline.run(...)` eine sortierte Sensor-Liste verwendet.
- Fehlende Trainingsdaten lösen `ValueError` aus; nicht registrierte Artefakte lösen `KeyError` aus.
- Nur endliche numerische Werte werden trainiert.
- `coverage` bleibt im Bereich 0 bis 1.
+- explizite Freigabe pro Aktor
+- konfigurierbare Mindestkonfidenz
+- Cooldown zwischen Schaltungen
+- keine Ausführung bei bereits erreichtem Zielzustand
+- keine Ausführung unbekannter Zustände oder riskanter Domains
+- eigene Schaltungen werden beim nächsten Training herausgefiltert
+- Automation-/Script-Aktionen zählen nur bei eindeutiger Herkunft im HA-Logbuch
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"

 [project]
 name = "sillyhome-next"
-version = "0.4.0"
+version = "0.7.0"
 description = "Lokales Smart-Home-Intelligenzsystem für Home Assistant"
 requires-python = ">=3.11"
 dependencies = [
--- a/tests/actuators/test_lifecycle.py
+++ b/tests/actuators/test_lifecycle.py
@@ -142,7 +142,7 @@ def test_reconciliation_auto_assigns_and_trains_numeric_model(tmp_path: Path) ->
    assert "binary_sensor.abstellkammer_motion" not in artifact.supported_sensors


-def test_reconciliation_requires_review_for_ambiguous_sensor_mapping(tmp_path: Path) -> None:
+def test_reconciliation_rejects_ambiguous_numeric_mapping(tmp_path: Path) -> None:
    start = datetime(2026, 6, 1, tzinfo=timezone.utc)
    entities = [
        HaEntitySummary(
@@ -182,10 +182,65 @@ def test_reconciliation_requires_review_for_ambiguous_sensor_mapping(tmp_path: P
    record = service.configure_actuator("switch.garage_pump")

    assert record.assignment.review_required is True
-    assert record.lifecycle.status is LifecycleStatus.REVIEW_REQUIRED
+    assert record.assignment.selected_numeric_entity_id is None
+    assert record.lifecycle.status is LifecycleStatus.ARCHIVED


-def test_manual_override_persists_and_wins_after_restart(tmp_path: Path) -> None:
+def test_reconciliation_does_not_cross_assign_other_room_light_energy(
+    tmp_path: Path,
+) -> None:
+    start = datetime(2026, 6, 1, tzinfo=timezone.utc)
+    entities = [
+        HaEntitySummary(
+            entity_id=(
+                "light.lichtschalter_abstellraum_"
+                "lichtschalter_abstellraum_s1"
+            ),
+            domain="light",
+            friendly_name="Licht Abstellraum",
+        ),
+        HaEntitySummary(
+            entity_id="sensor.licht_badezimmer_energy",
+            domain="sensor",
+            device_class="energy",
+            state_class="total_increasing",
+            unit_of_measurement="kWh",
+            friendly_name="Lichtschalter_Badezimmer Licht Badezimmer energy",
+        ),
+        HaEntitySummary(
+            entity_id="binary_sensor.abstellraum_ture",
+            domain="binary_sensor",
+            device_class="door",
+            friendly_name="Abstellraum Türe",
+        ),
+        HaEntitySummary(
+            entity_id="binary_sensor.briefkasten_open",
+            domain="binary_sensor",
+            device_class="opening",
+            friendly_name="Briefkasten open",
+        ),
+    ]
+    service = _service(
+        tmp_path,
+        entities,
+        {"sensor.licht_badezimmer_energy": _points(8, start, 1.0)},
+    )
+
+    record = service.configure_actuator(
+        "light.lichtschalter_abstellraum_lichtschalter_abstellraum_s1"
+    )
+
+    assert record.assignment.selected_numeric_entity_id is None
+    assert record.assignment.selected_context_entity_ids == [
+        "binary_sensor.abstellraum_ture"
+    ]
+    assert record.assignment.source is AssignmentSource.AUTOMATIC
+    assert record.assignment.confidence == 1.0
+    assert record.assignment.review_required is False
+    assert record.lifecycle.status is LifecycleStatus.ARCHIVED
+
+
+def test_legacy_manual_override_is_cleared_and_automatic_mapping_wins(tmp_path: Path) -> None:
    start = datetime(2026, 6, 1, tzinfo=timezone.utc)
    entities = [
        HaEntitySummary(
@@ -218,21 +273,21 @@ def test_manual_override_persists_and_wins_after_restart(tmp_path: Path) -> None
        "sensor.abstellkammer_power": _points(8, start, 30.0),
    }
    service = _service(tmp_path, entities, history)
-    service.configure_actuator("light.abstellkammer")
-
-    updated = service.set_override(
-        "light.abstellkammer",
-        ManualOverride(
-            numeric_entity_id="sensor.abstellkammer_power",
-            context_entity_ids=[],
-            note="Manuelle Leistungs-Zuordnung",
-        ),
+    configured = service.configure_actuator("light.abstellkammer")
+    legacy = configured.model_copy(
+        update={
+            "manual_override": ManualOverride(
+                numeric_entity_id="sensor.abstellkammer_power",
+                context_entity_ids=[],
+                note="Alte manuelle Zuordnung",
+            )
+        }
    )
+    service._store.upsert(legacy)

    restarted = _service(tmp_path, entities, history)
    record = restarted.reconcile_actuator("light.abstellkammer")

-    assert updated.assignment.source is AssignmentSource.MANUAL
-    assert record.assignment.selected_numeric_entity_id == "sensor.abstellkammer_power"
-    assert record.manual_override is not None
-    assert record.manual_override.numeric_entity_id == "sensor.abstellkammer_power"
+    assert record.assignment.selected_numeric_entity_id == "sensor.abstellkammer_illuminance"
+    assert record.assignment.source.value == "automatic"
+    assert record.manual_override is None
--- a/tests/api/test_actuators.py
+++ b/tests/api/test_actuators.py
@@ -7,11 +7,17 @@ from fastapi.testclient import TestClient

 from app.actuators.lifecycle import ActuatorReconciliationService
 from app.actuators.store import ActuatorStore
+from app.behavior.engine import BehaviorEngine
 from app.config import Settings
 from app.ha.discovery import DiscoveredEntity
 from app.ha.discovery import discover_entities
-from app.ha.history import EntityHistorySeries, NumericHistoryPoint
-from app.ha.models import HaEntitySummary
+from app.ha.history import (
+    EntityHistorySeries,
+    LogbookEntry,
+    NumericHistoryPoint,
+    StateHistorySeries,
+)
+from app.ha.models import HaAutomationSummary, HaEntitySummary
 from app.ha.reader import HaReader
 from app.main import app
 from app.ml.registry.model_registry import ModelRegistry
@@ -54,6 +60,36 @@ class FakeHaReader(HaReader):
            if entity_id in self._history
        ]

+    def read_state_history(
+        self,
+        entity_ids: list[str],
+        start_time: datetime,
+        end_time: datetime,
+    ) -> list[StateHistorySeries]:
+        return []
+
+    def read_logbook(
+        self,
+        entity_id: str,
+        start_time: datetime,
+        end_time: datetime,
+    ) -> list[LogbookEntry]:
+        return []
+
+    def call_service(
+        self,
+        domain: str,
+        service: str,
+        service_data: dict[str, object],
+    ) -> list[object]:
+        return []
+
+    def find_automations_for_entity(
+        self,
+        entity_id: str,
+    ) -> list[HaAutomationSummary]:
+        return []
+

 def _install_service(tmp_path: Path) -> None:
    entities = [
@@ -103,9 +139,14 @@ def _install_service(tmp_path: Path) -> None:
        registry=app.state.registry,
        settings=settings,
    )
+    app.state.behavior_engine = BehaviorEngine(
+        ha_reader=app.state.ha_reader,
+        store=app.state.actuator_store,
+        settings=settings,
+    )


-def test_actuator_api_configures_reconciles_and_overrides(tmp_path: Path) -> None:
+def test_actuator_api_configures_reconciles_and_removes(tmp_path: Path) -> None:
    with TestClient(app) as client:
        _install_service(tmp_path)

@@ -118,18 +159,34 @@ def test_actuator_api_configures_reconciles_and_overrides(tmp_path: Path) -> Non
        listed = client.get("/v1/actuators")
        assert listed.status_code == 200
        assert listed.json()[0]["lifecycle"]["status"] == "trained"
+        assert listed.json()[0]["behavior"]["mode"] == "shadow"

-        override = client.post(
-            "/v1/actuators/light.abstellkammer/override",
-            json={
-                "numeric_entity_id": "sensor.abstellkammer_illuminance",
-                "context_entity_ids": ["binary_sensor.abstellkammer_motion"],
-                "note": "Explizit bestaetigt",
-            },
+        evaluation = client.post("/v1/actuators/light.abstellkammer/evaluate")
+        assert evaluation.status_code == 200
+
+        premature_activation = client.post(
+            "/v1/actuators/light.abstellkammer/activation",
+            json={"active": True},
        )
-        assert override.status_code == 200
-        assert override.json()["assignment"]["source"] == "manual"
+        assert premature_activation.status_code == 409

        reconciliation = client.post("/v1/actuators/reconciliation/run")
        assert reconciliation.status_code == 200
        assert reconciliation.json()["trained_models"] == 1
+
+        removed = client.delete("/v1/actuators/light.abstellkammer")
+        assert removed.status_code == 204
+        assert client.get("/v1/actuators").json() == []
+
+
+def test_manual_override_endpoint_is_not_exposed(tmp_path: Path) -> None:
+    with TestClient(app) as client:
+        _install_service(tmp_path)
+        client.post("/v1/actuators", json={"actuator_entity_id": "light.abstellkammer"})
+
+        response = client.post(
+            "/v1/actuators/light.abstellkammer/override",
+            json={"numeric_entity_id": "sensor.abstellkammer_illuminance"},
+        )
+
+    assert response.status_code == 404
--- a/tests/api/test_automations.py
+++ b/tests/api/test_automations.py
@@ -1,59 +1,17 @@
-from pathlib import Path
-
 from fastapi.testclient import TestClient

-from app.automations.store import AutomationStore
 from app.main import app


-def _payload() -> dict[str, object]:
-    return {
-        "alias": "Licht bei Dunkelheit",
-        "description": "Schaltet das Flurlicht unter dem Helligkeitsgrenzwert ein.",
-        "trigger": {"entity_id": "sensor.hall_illuminance", "below": 10},
-        "action": {
-            "service": "light.turn_on",
-            "entity_id": "light.hall",
-            "data": {"brightness_pct": 40},
-        },
-    }
-
-
-def test_proposal_requires_explicit_approval_before_yaml(tmp_path: Path) -> None:
+def test_automation_api_is_not_exposed() -> None:
    with TestClient(app) as client:
-        app.state.automation_store = AutomationStore(tmp_path)
-        created = client.post("/v1/automations/proposals", json=_payload())
-        proposal_id = created.json()["proposal_id"]
-        blocked = client.get(f"/v1/automations/proposals/{proposal_id}/yaml")
-        approved = client.post(
-            f"/v1/automations/proposals/{proposal_id}/approve",
-            json={"expected_revision": 1},
+        response = client.post(
+            "/v1/automations/proposals",
+            json={
+                "alias": "Nicht mehr verfügbar",
+                "trigger": {"entity_id": "sensor.hall_illuminance", "below": 10},
+                "action": {"service": "light.turn_on", "entity_id": "light.hall"},
+            },
        )
-        exported = client.get(f"/v1/automations/proposals/{proposal_id}/yaml")
-    assert created.status_code == 201
-    assert created.json()["status"] == "draft"
-    assert blocked.status_code == 409
-    assert approved.json()["status"] == "approved"
-    assert "service: light.turn_on" in exported.text

-
-def test_proposal_rejects_unsafe_service_domain(tmp_path: Path) -> None:
-    payload = _payload()
-    payload["action"] = {
-        "service": "shell_command.run",
-        "entity_id": "light.hall",
-        "data": {},
-    }
-    with TestClient(app) as client:
-        app.state.automation_store = AutomationStore(tmp_path)
-        response = client.post("/v1/automations/proposals", json=payload)
-    assert response.status_code == 422
-
-
-def test_proposal_requires_numeric_threshold(tmp_path: Path) -> None:
-    payload = _payload()
-    payload["trigger"] = {"entity_id": "sensor.hall_illuminance"}
-    with TestClient(app) as client:
-        app.state.automation_store = AutomationStore(tmp_path)
-        response = client.post("/v1/automations/proposals", json=payload)
-    assert response.status_code == 422
+    assert response.status_code == 404
--- a/tests/api/test_entities.py
+++ b/tests/api/test_entities.py
@@ -73,9 +73,11 @@ def test_entities_returns_reader_data() -> None:
    assert response.status_code == 200
    assert response.json() == [
        {
-            "entity_id": "sensor.temperature",
-            "domain": "sensor",
-            "state_class": None,
+                "entity_id": "sensor.temperature",
+                "domain": "sensor",
+                "state": None,
+                "last_changed": None,
+                "state_class": None,
            "device_class": None,
            "unit_of_measurement": None,
            "friendly_name": None,
--- a/tests/behavior/test_engine.py
+++ b/tests/behavior/test_engine.py
@@ -0,0 +1,525 @@
+from __future__ import annotations
+
+from datetime import datetime, timedelta, timezone
+from pathlib import Path
+
+import pytest
+
+from app.actuators.models import (
+    BehaviorMode,
+    BehaviorPattern,
+    BehaviorState,
+    BehaviorStatus,
+    ExecutionEvent,
+)
+from app.actuators.store import ActuatorStore
+from app.behavior.engine import BehaviorEngine, predict_behavior, service_for_state
+from app.config import Settings
+from app.ha.history import (
+    LogbookEntry,
+    StateHistoryPoint,
+    StateHistorySeries,
+)
+from app.ha.models import HaAutomationSummary, HaEntitySummary
+from app.ha.reader import HaReader
+
+
+class FakeBehaviorReader(HaReader):
+    def __init__(
+        self,
+        *,
+        entities: list[HaEntitySummary],
+        history: list[StateHistorySeries],
+        logbook: list[LogbookEntry],
+    ) -> None:
+        self.entities = entities
+        self.history = history
+        self.logbook = logbook
+        self.service_calls: list[tuple[str, str, dict[str, object]]] = []
+        self.automations: list[HaAutomationSummary] = []
+
+    def read_entities(self) -> list[HaEntitySummary]:
+        return list(self.entities)
+
+    def read_state_history(
+        self,
+        entity_ids: list[str],
+        start_time: datetime,
+        end_time: datetime,
+    ) -> list[StateHistorySeries]:
+        return [series for series in self.history if series.entity_id in entity_ids]
+
+    def read_logbook(
+        self,
+        entity_id: str,
+        start_time: datetime,
+        end_time: datetime,
+    ) -> list[LogbookEntry]:
+        return [entry for entry in self.logbook if entry.entity_id == entity_id]
+
+    def call_service(
+        self,
+        domain: str,
+        service: str,
+        service_data: dict[str, object],
+    ) -> list[object]:
+        self.service_calls.append((domain, service, service_data))
+        return []
+
+    def find_automations_for_entity(
+        self,
+        entity_id: str,
+    ) -> list[HaAutomationSummary]:
+        return list(self.automations)
+
+
+def _settings(tmp_path: Path) -> Settings:
+    return Settings(
+        actuator_store=str(tmp_path / "actuators"),
+        model_store=str(tmp_path / "models"),
+        automation_store=str(tmp_path / "automations"),
+        history_days=14,
+        min_behavior_actions=3,
+        prediction_confidence=0.8,
+        prediction_window_minutes=30,
+        execution_cooldown_seconds=900,
+        timezone="Europe/Berlin",
+    )
+
+
+def _reader(now: datetime) -> FakeBehaviorReader:
+    actuator_points: list[StateHistoryPoint] = []
+    logbook: list[LogbookEntry] = []
+    for days_ago in (3, 2, 1):
+        action_at = now - timedelta(days=days_ago)
+        actuator_points.extend(
+            [
+                StateHistoryPoint(timestamp=action_at - timedelta(minutes=1), state="off"),
+                StateHistoryPoint(timestamp=action_at, state="on"),
+                StateHistoryPoint(timestamp=action_at + timedelta(hours=6), state="off"),
+            ]
+        )
+        logbook.extend(
+            [
+                LogbookEntry(
+                    entity_id="light.office",
+                    timestamp=action_at,
+                    message="turned on",
+                    context_user_id="user-1",
+                ),
+                LogbookEntry(
+                    entity_id="light.office",
+                    timestamp=action_at + timedelta(hours=6),
+                    message="turned off",
+                    context_domain="automation",
+                    context_service="trigger",
+                ),
+            ]
+        )
+    actuator_points.sort(key=lambda point: point.timestamp)
+    context_points = [
+        StateHistoryPoint(timestamp=now - timedelta(days=7), state="on"),
+    ]
+    return FakeBehaviorReader(
+        entities=[
+            HaEntitySummary(entity_id="light.office", domain="light", state="off"),
+            HaEntitySummary(
+                entity_id="binary_sensor.office_presence",
+                domain="binary_sensor",
+                state="on",
+            ),
+        ],
+        history=[
+            StateHistorySeries(entity_id="light.office", points=actuator_points),
+            StateHistorySeries(
+                entity_id="binary_sensor.office_presence",
+                points=context_points,
+            ),
+        ],
+        logbook=logbook,
+    )
+
+
+def _engine(tmp_path: Path, now: datetime) -> tuple[BehaviorEngine, FakeBehaviorReader]:
+    settings = _settings(tmp_path)
+    store = ActuatorStore(settings.actuator_store)
+    record = store.configure("light.office")
+    store.upsert(
+        record.model_copy(
+            update={
+                "assignment": record.assignment.model_copy(
+                    update={
+                        "selected_context_entity_ids": [
+                            "binary_sensor.office_presence"
+                        ],
+                    }
+                )
+            }
+        )
+    )
+    reader = _reader(now)
+    return (
+        BehaviorEngine(ha_reader=reader, store=store, settings=settings),
+        reader,
+    )
+
+
+def test_engine_trains_predicts_in_shadow_and_executes_only_after_approval(
+    tmp_path: Path,
+) -> None:
+    now = datetime.now(timezone.utc).replace(second=0, microsecond=0)
+    engine, reader = _engine(tmp_path, now)
+
+    trained = engine.train("light.office")
+    shadow = engine.evaluate("light.office")
+
+    assert trained.behavior.status is BehaviorStatus.TRAINED
+    assert trained.behavior.sample_count == 6
+    assert trained.behavior.high_confidence_sample_count == 6
+    assert shadow.behavior.mode is BehaviorMode.SHADOW
+    assert shadow.behavior.prediction is not None
+    assert shadow.behavior.prediction.target_state == "on"
+    assert reader.service_calls == []
+
+    engine.set_active("light.office", active=True)
+    active = engine.evaluate("light.office")
+
+    assert active.behavior.mode is BehaviorMode.ACTIVE
+    assert active.behavior.prediction is not None
+    assert active.behavior.prediction.executed is True
+    assert reader.service_calls == [
+        ("light", "turn_on", {"entity_id": "light.office"})
+    ]
+
+
+def test_engine_counts_known_automation_actions_like_manual_actions(
+    tmp_path: Path,
+) -> None:
+    now = datetime.now(timezone.utc).replace(second=0, microsecond=0)
+    engine, _ = _engine(tmp_path, now)
+
+    trained = engine.train("light.office")
+
+    assert {pattern.target_state for pattern in trained.behavior.patterns} == {
+        "on",
+        "off",
+    }
+    assert {pattern.source for pattern in trained.behavior.patterns} == {
+        "user",
+        "automation",
+    }
+    assert trained.behavior.high_confidence_sample_count == 6
+    assert {pattern.weight for pattern in trained.behavior.patterns} == {1.0}
+
+
+def test_engine_learns_causal_automation_with_activation_credit(
+    tmp_path: Path,
+) -> None:
+    now = datetime.now(timezone.utc).replace(second=0, microsecond=0)
+    actuator_points: list[StateHistoryPoint] = []
+    door_points: list[StateHistoryPoint] = []
+    logbook: list[LogbookEntry] = []
+    for days_ago in (3, 2, 1):
+        action_at = now - timedelta(days=days_ago)
+        actuator_points.extend(
+            [
+                StateHistoryPoint(
+                    timestamp=action_at - timedelta(minutes=1),
+                    state="off",
+                ),
+                StateHistoryPoint(timestamp=action_at, state="on"),
+            ]
+        )
+        door_points.extend(
+            [
+                StateHistoryPoint(
+                    timestamp=action_at - timedelta(minutes=1),
+                    state="off",
+                ),
+                StateHistoryPoint(
+                    timestamp=action_at - timedelta(seconds=1),
+                    state="on",
+                ),
+            ]
+        )
+        logbook.append(
+            LogbookEntry(
+                entity_id="light.storage",
+                timestamp=action_at,
+                message="turned on",
+                context_domain="automation",
+                context_service="trigger",
+            )
+        )
+    actuator_points.sort(key=lambda point: point.timestamp)
+    door_points.sort(key=lambda point: point.timestamp)
+    settings = _settings(tmp_path)
+    store = ActuatorStore(settings.actuator_store)
+    record = store.configure("light.storage")
+    store.upsert(
+        record.model_copy(
+            update={
+                "assignment": record.assignment.model_copy(
+                    update={
+                        "selected_context_entity_ids": [
+                            "binary_sensor.storage_door"
+                        ],
+                    }
+                )
+            }
+        )
+    )
+    reader = FakeBehaviorReader(
+        entities=[],
+        history=[
+            StateHistorySeries(
+                entity_id="light.storage",
+                points=actuator_points,
+            ),
+            StateHistorySeries(
+                entity_id="binary_sensor.storage_door",
+                points=door_points,
+            ),
+        ],
+        logbook=logbook,
+    )
+    engine = BehaviorEngine(ha_reader=reader, store=store, settings=settings)
+
+    trained = engine.train("light.storage")
+    automation_patterns = [
+        pattern
+        for pattern in trained.behavior.patterns
+        if pattern.source == "automation"
+    ]
+
+    assert len(automation_patterns) == 3
+    assert trained.behavior.high_confidence_sample_count == 3
+    assert {pattern.weight for pattern in automation_patterns} == {1.0}
+    assert {
+        (
+            pattern.trigger_entity_id,
+            pattern.trigger_from_state,
+            pattern.trigger_to_state,
+        )
+        for pattern in automation_patterns
+    } == {("binary_sensor.storage_door", "off", "on")}
+
+    active = engine.set_active("light.storage", active=True)
+
+    assert active.behavior.mode is BehaviorMode.ACTIVE
+
+
+def test_active_mode_rejects_unsafe_domains(tmp_path: Path) -> None:
+    settings = _settings(tmp_path)
+    store = ActuatorStore(settings.actuator_store)
+    record = store.configure("lock.front_door")
+    store.upsert(
+        record.model_copy(
+            update={
+                "behavior": record.behavior.model_copy(
+                    update={"status": BehaviorStatus.TRAINED}
+                )
+            }
+        )
+    )
+    reader = FakeBehaviorReader(entities=[], history=[], logbook=[])
+    engine = BehaviorEngine(ha_reader=reader, store=store, settings=settings)
+
+    with pytest.raises(ValueError, match="nicht freigegeben"):
+        engine.set_active("lock.front_door", active=True)
+
+
+def test_active_mode_requires_trusted_manual_or_automation_actions(tmp_path: Path) -> None:
+    settings = _settings(tmp_path)
+    store = ActuatorStore(settings.actuator_store)
+    record = store.configure("light.office")
+    store.upsert(
+        record.model_copy(
+            update={
+                "behavior": record.behavior.model_copy(
+                    update={
+                        "status": BehaviorStatus.TRAINED,
+                        "sample_count": 3,
+                        "high_confidence_sample_count": 0,
+                    }
+                )
+            }
+        )
+    )
+    reader = FakeBehaviorReader(entities=[], history=[], logbook=[])
+    engine = BehaviorEngine(ha_reader=reader, store=store, settings=settings)
+
+    with pytest.raises(ValueError, match="Freigabe"):
+        engine.set_active("light.office", active=True)
+
+
+def test_control_handoff_pauses_and_restores_matching_automation(
+    tmp_path: Path,
+) -> None:
+    settings = _settings(tmp_path)
+    store = ActuatorStore(settings.actuator_store)
+    record = store.configure("light.storage")
+    store.upsert(
+        record.model_copy(
+            update={
+                "behavior": record.behavior.model_copy(
+                    update={
+                        "status": BehaviorStatus.TRAINED,
+                        "sample_count": 3,
+                        "high_confidence_sample_count": 3,
+                        "activation_ready": True,
+                        "activation_reason": "Freigabe bereit.",
+                    }
+                )
+            }
+        )
+    )
+    reader = FakeBehaviorReader(entities=[], history=[], logbook=[])
+    reader.automations = [
+        HaAutomationSummary(
+            entity_id="automation.storage_light",
+            config_id="123",
+            friendly_name="Storage light",
+            enabled=True,
+        )
+    ]
+    engine = BehaviorEngine(ha_reader=reader, store=store, settings=settings)
+
+    active = engine.set_active(
+        "light.storage",
+        active=True,
+        pause_matching_automations=True,
+    )
+    shadow = engine.set_active(
+        "light.storage",
+        active=False,
+        restore_paused_automations=True,
+    )
+
+    assert active.behavior.mode is BehaviorMode.ACTIVE
+    assert active.behavior.paused_automation_entity_ids == [
+        "automation.storage_light"
+    ]
+    assert shadow.behavior.mode is BehaviorMode.SHADOW
+    assert shadow.behavior.paused_automation_entity_ids == []
+    assert reader.service_calls == [
+        (
+            "automation",
+            "turn_off",
+            {"entity_id": "automation.storage_light"},
+        ),
+        (
+            "automation",
+            "turn_on",
+            {"entity_id": "automation.storage_light"},
+        ),
+    ]
+
+
+def test_cooldown_allows_opposite_follow_up_action(tmp_path: Path) -> None:
+    settings = _settings(tmp_path)
+    store = ActuatorStore(settings.actuator_store)
+    reader = FakeBehaviorReader(entities=[], history=[], logbook=[])
+    engine = BehaviorEngine(ha_reader=reader, store=store, settings=settings)
+    now = datetime.now(timezone.utc)
+    behavior = BehaviorState(
+        mode=BehaviorMode.ACTIVE,
+        last_executed_at=now - timedelta(seconds=5),
+        execution_events=[
+            ExecutionEvent(target_state="on", executed_at=now - timedelta(seconds=5))
+        ],
+    )
+
+    assert engine._cooldown_elapsed(behavior, now, "off") is True
+    assert engine._cooldown_elapsed(behavior, now, "on") is False
+
+
+@pytest.mark.parametrize(
+    ("domain", "state", "service"),
+    [
+        ("light", "on", "turn_on"),
+        ("switch", "off", "turn_off"),
+        ("cover", "open", "open_cover"),
+        ("cover", "closed", "close_cover"),
+        ("lock", "unlocked", None),
+    ],
+)
+def test_service_for_state_is_strictly_allowlisted(
+    domain: str,
+    state: str,
+    service: str | None,
+) -> None:
+    assert service_for_state(domain, state) == service
+
+
+def test_prediction_requires_temporal_support() -> None:
+    assert predict_behavior(
+        [],
+        current_context={},
+        now=datetime.now(timezone.utc),
+        min_support=3,
+        window_minutes=30,
+    ) is None
+
+
+def test_prediction_uses_fresh_causal_context_transition_outside_time_window() -> None:
+    now = datetime.now(timezone.utc).replace(second=0, microsecond=0)
+    patterns = [
+        BehaviorPattern(
+            target_state="on",
+            minute_of_day=60,
+            weekday=0,
+            context_states={"binary_sensor.storage_door": "on"},
+            trigger_entity_id="binary_sensor.storage_door",
+            trigger_from_state="off",
+            trigger_to_state="on",
+            source="automation",
+            weight=0.7,
+            observed_at=now - timedelta(days=days_ago),
+        )
+        for days_ago in (3, 2, 1)
+    ]
+
+    prediction = predict_behavior(
+        patterns,
+        current_context={"binary_sensor.storage_door": "on"},
+        current_context_changed_at={
+            "binary_sensor.storage_door": now - timedelta(seconds=10)
+        },
+        now=now,
+        min_support=3,
+        window_minutes=30,
+    )
+
+    assert prediction is not None
+    assert prediction.target_state == "on"
+    assert prediction.matching_patterns == 3
+    assert prediction.confidence == 0.7
+    assert "frischen Sensorwechsel" in prediction.reason
+
+
+def test_prediction_ignores_stale_causal_context_state() -> None:
+    now = datetime.now(timezone.utc).replace(second=0, microsecond=0)
+    pattern = BehaviorPattern(
+        target_state="on",
+        minute_of_day=60,
+        weekday=0,
+        context_states={"binary_sensor.storage_door": "on"},
+        trigger_entity_id="binary_sensor.storage_door",
+        trigger_from_state="off",
+        trigger_to_state="on",
+        source="automation",
+        weight=0.7,
+        observed_at=now - timedelta(days=1),
+    )
+
+    assert predict_behavior(
+        [pattern],
+        current_context={"binary_sensor.storage_door": "on"},
+        current_context_changed_at={
+            "binary_sensor.storage_door": now - timedelta(minutes=5)
+        },
+        now=now,
+        min_support=1,
+        window_minutes=30,
+    ) is None
--- a/tests/ha/test_ha_client.py
+++ b/tests/ha/test_ha_client.py
@@ -108,6 +108,35 @@ def test_list_entity_metadata_calls_template_api() -> None:
    }


+def test_get_logbook_filters_entity_and_period() -> None:
+    response = _response(payload=[{"entity_id": "light.office"}])
+    client = _client_with_response(response)
+    start = datetime(2026, 6, 1, tzinfo=timezone.utc)
+    end = datetime(2026, 6, 2, tzinfo=timezone.utc)
+
+    payload = client.get_logbook("light.office", start, end)
+
+    assert payload == [{"entity_id": "light.office"}]
+    call = client._session.get.call_args  # type: ignore[attr-defined]
+    assert "/api/logbook/2026-06-01T00:00:00+00:00" in call.args[0]
+    assert call.kwargs["params"]["entity"] == "light.office"
+
+
+def test_call_service_posts_to_home_assistant() -> None:
+    response = _response(payload=[])
+    client = HaClient(HaClientSettings(url="http://ha.local", token="test-token"))
+    client._session.post = Mock(return_value=response)  # type: ignore[method-assign]
+
+    result = client.call_service("light", "turn_on", {"entity_id": "light.office"})
+
+    assert result == []
+    client._session.post.assert_called_once_with(
+        "http://ha.local/api/services/light/turn_on",
+        json={"entity_id": "light.office"},
+        timeout=10,
+    )
+
+
@pytest.mark.parametrize(
    ("entity_ids", "start", "end"),
    [
--- a/tests/ha/test_ha_reader.py
+++ b/tests/ha/test_ha_reader.py
@@ -15,6 +15,7 @@ class FakeHaClient(HaClient):
            {
                "entity_id": "sensor.temperature",
                "state": "21.5",
+                "last_changed": "2026-06-14T12:00:00+00:00",
                "attributes": {
                    "state_class": "measurement",
                    "device_class": "temperature",
@@ -54,6 +55,29 @@ class FakeHaClient(HaClient):
            }
        }

+    def get_logbook(
+        self,
+        entity_id: str,
+        start_time: datetime,
+        end_time: datetime,
+    ) -> list[object]:
+        return [
+            {
+                "entity_id": entity_id,
+                "when": start_time.isoformat(),
+                "message": "turned on",
+                "context_user_id": "user-1",
+            }
+        ]
+
+    def call_service(
+        self,
+        domain: str,
+        service: str,
+        service_data: dict[str, object],
+    ) -> list[object]:
+        return []
+

 def test_ha_reader_returns_summaries() -> None:
    reader = HaReader(FakeHaClient())
@@ -63,6 +87,8 @@ def test_ha_reader_returns_summaries() -> None:
    assert domains == {"sensor", "light"}
    sensor = next(item for item in summaries if item.entity_id == "sensor.temperature")
    assert sensor.unit_of_measurement == "°C"
+    assert sensor.state == "21.5"
+    assert sensor.last_changed == datetime(2026, 6, 14, 12, 0, tzinfo=timezone.utc)
    assert sensor.area_name == "Kueche"
    assert sensor.device_name == "Thermometer"

@@ -87,3 +113,40 @@ def test_ha_reader_normalizes_history() -> None:

    assert history[0].entity_id == "sensor.temperature"
    assert history[0].points[0].value == 21.5
+
+
+def test_ha_reader_normalizes_state_history_and_logbook() -> None:
+    reader = HaReader(FakeHaClient())
+    start = datetime(2026, 6, 1, tzinfo=timezone.utc)
+    end = datetime(2026, 6, 2, tzinfo=timezone.utc)
+
+    history = reader.read_state_history(["light.living_room"], start, end)
+    logbook = reader.read_logbook("light.living_room", start, end)
+
+    assert history[0].points[0].state == "21.5"
+    assert logbook[0].context_user_id == "user-1"
+
+
+def test_ha_reader_finds_automation_that_targets_entity() -> None:
+    client = FakeHaClient()
+    client.list_entities = lambda: [  # type: ignore[method-assign]
+        {
+            "entity_id": "automation.storage_light",
+            "state": "on",
+            "attributes": {
+                "id": "123",
+                "friendly_name": "Storage light",
+            },
+        }
+    ]
+    client.get_automation_config = lambda automation_id: {  # type: ignore[method-assign]
+        "id": automation_id,
+        "target": {"entity_id": "light.storage"},
+    }
+    reader = HaReader(client)
+
+    matches = reader.find_automations_for_entity("light.storage")
+
+    assert len(matches) == 1
+    assert matches[0].entity_id == "automation.storage_light"
+    assert matches[0].enabled is True
--- a/tests/ha/test_history.py
+++ b/tests/ha/test_history.py
@@ -5,7 +5,11 @@ from datetime import datetime, timezone
 import pytest

 from app.ha.exceptions import HaUnexpectedPayloadError
-from app.ha.history import normalize_history_payload
+from app.ha.history import (
+    normalize_history_payload,
+    normalize_logbook_payload,
+    normalize_state_history_payload,
+)


 def test_normalize_history_payload_groups_and_sorts_numeric_states() -> None:
@@ -90,3 +94,46 @@ def test_normalize_history_payload_rejects_malformed_structure(payload: object)

 def test_normalize_history_payload_accepts_empty_series() -> None:
    assert normalize_history_payload([[]]) == []
+
+
+def test_normalize_state_history_keeps_categorical_changes() -> None:
+    result = normalize_state_history_payload(
+        [
+            [
+                {
+                    "entity_id": "light.office",
+                    "state": "off",
+                    "last_changed": "2026-06-01T08:00:00+00:00",
+                },
+                {
+                    "state": "on",
+                    "last_changed": "2026-06-01T08:05:00+00:00",
+                },
+                {
+                    "state": "on",
+                    "last_changed": "2026-06-01T08:06:00+00:00",
+                },
+            ]
+        ]
+    )
+
+    assert [point.state for point in result[0].points] == ["off", "on"]
+
+
+def test_normalize_logbook_preserves_action_origin() -> None:
+    result = normalize_logbook_payload(
+        [
+            {
+                "entity_id": "light.office",
+                "when": "2026-06-01T08:05:00+00:00",
+                "message": "turned on",
+                "context_user_id": "user-1",
+                "context_domain": "light",
+                "context_service": "turn_on",
+            }
+        ],
+        "light.office",
+    )
+
+    assert result[0].context_user_id == "user-1"
+    assert result[0].context_service == "turn_on"
--- a/tests/test_addon_config.py
+++ b/tests/test_addon_config.py
@@ -0,0 +1,18 @@
+from pathlib import Path
+
+
+def test_addon_does_not_expose_internal_learning_parameters() -> None:
+    config = Path("addon/config.yaml").read_text(encoding="utf-8")
+
+    assert "\noptions:" not in config
+    assert "\nschema:" not in config
+    assert "prediction_confidence" not in config
+    assert "execution_cooldown_seconds" not in config
+
+
+def test_addon_version_invalidates_application_build_layer() -> None:
+    dockerfile = Path("addon/Dockerfile").read_text(encoding="utf-8")
+
+    config_copy = dockerfile.index("COPY config.yaml /tmp/addon-config.yaml")
+    repository_clone = dockerfile.index("git clone --depth 1 --branch main")
+    assert config_copy < repository_clone
--- a/tests/test_config.py
+++ b/tests/test_config.py
@@ -15,6 +15,12 @@ def test_load_settings_reads_documented_environment(monkeypatch: MonkeyPatch) ->
    monkeypatch.setenv("SILLYHOME_MIN_TRAINING_POINTS", "12")
    monkeypatch.setenv("SILLYHOME_RETRAIN_STALE_HOURS", "48")
    monkeypatch.setenv("SILLYHOME_RECONCILE_INTERVAL_SECONDS", "600")
+    monkeypatch.setenv("SILLYHOME_MIN_BEHAVIOR_ACTIONS", "4")
+    monkeypatch.setenv("SILLYHOME_PREDICTION_CONFIDENCE", "0.9")
+    monkeypatch.setenv("SILLYHOME_PREDICTION_WINDOW_MINUTES", "20")
+    monkeypatch.setenv("SILLYHOME_PREDICTION_INTERVAL_SECONDS", "45")
+    monkeypatch.setenv("SILLYHOME_EXECUTION_COOLDOWN_SECONDS", "1200")
+    monkeypatch.setenv("SILLYHOME_TIMEZONE", "Europe/Berlin")

    settings = load_settings()

@@ -27,4 +33,10 @@ def test_load_settings_reads_documented_environment(monkeypatch: MonkeyPatch) ->
    assert settings.min_training_points == 12
    assert settings.retrain_stale_hours == 48
    assert settings.reconcile_interval_seconds == 600
+    assert settings.min_behavior_actions == 4
+    assert settings.prediction_confidence == 0.9
+    assert settings.prediction_window_minutes == 20
+    assert settings.prediction_interval_seconds == 45
+    assert settings.execution_cooldown_seconds == 1200
+    assert settings.timezone == "Europe/Berlin"
    assert settings.ha_configured
--- a/tests/test_dashboard.py
+++ b/tests/test_dashboard.py
@@ -9,4 +9,21 @@ def test_dashboard_is_served_at_root() -> None:

    assert response.status_code == 200
    assert "SillyHome Next" in response.text
-    assert "Automation-Entwurf" in response.text
+    assert "So gehst du vor" in response.text
+    assert "Gerät zum Lernen auswählen" in response.text
+    assert "Entitätsname oder Gerät aus Home Assistant" in response.text
+    assert "Wie gewohnt bedienen" in response.text
+    assert "Ohne deine spätere Freigabe wird nichts geschaltet" in response.text
+    assert "Du wählst keine Sensoren und erstellst keine Regeln" in response.text
+    assert "Freigabestatus" in response.text
+    assert "SillyHome übernehmen lassen" in response.text
+    assert "Passende Home-Assistant-Automationen" in response.text
+    assert "Pausieren" in response.text
+    assert "Davon erkannte HA-Automationen" in response.text
+    assert "Aktuelle Situation auswerten" in response.text
+    assert "Die Prüfung simuliert keinen Sensorwechsel" in response.text
+    assert "Kein frischer passender Sensorwechsel erkannt" in response.text
+    assert "Vorhersage jetzt prüfen" not in response.text
+    assert "record.behavior.activation_ready" in response.text
+    assert "Automation-Entwurf" not in response.text
+    assert "Manuelle Overrides" not in response.text
Author	SHA1	Message	Date
Otto	f9c7c27e00	Merge pull request 'v0.7.0: sichere Steuerungsübergabe und klare Bedienung' (#40 ) from feature/control-handoff-v0.7.0 into main	2026-06-14 16:22:18 +02:00
Otto	b3cf68eade	CONTROL-001: add safe HA automation handoff Some checks failed quality / test (3.11) (push) Has been cancelled Details quality / test (3.13) (push) Has been cancelled Details	2026-06-14 16:21:57 +02:00
Otto	77f328c4a8	Merge pull request 'v0.6.2: HA-Automationen gleichwertig lernen' (#39 ) from feature/automation-equality-v0.6.2 into main	2026-06-14 15:58:44 +02:00
Otto	7ad97320a2	BEHAVIOR-004: trust HA automation actions Some checks failed quality / test (3.11) (push) Has been cancelled Details quality / test (3.13) (push) Has been cancelled Details	2026-06-14 15:58:05 +02:00
Otto	58d3126a35	Merge pull request 'v0.6.1: sichtbare Rückmeldung bei Situationsprüfung' (#38 ) from fix/evaluation-feedback-v0.6.1 into main	2026-06-14 15:38:26 +02:00
Otto	1c5eab14b6	UI-003: show prediction evaluation feedback	2026-06-14 15:38:11 +02:00
Otto	87ae051238	Merge pull request 'v0.6.0: kausales Shadow-Lernen aus Sensorwechseln' (#37 ) from feature/causal-shadow-v0.6.0 into main	2026-06-14 15:35:22 +02:00
Otto	fb76d89204	BEHAVIOR-003: learn causal shadow triggers Some checks failed quality / test (3.11) (push) Has been cancelled Details quality / test (3.13) (push) Has been cancelled Details	2026-06-14 15:35:07 +02:00
Otto	1370d02c15	Merge pull request 'v0.5.4: korrekter Kontext- und Freigabestatus' (#36 ) from fix/context-status-v0.5.4 into main	2026-06-14 15:29:07 +02:00
Otto	100f5af578	UI-002: align context and activation status	2026-06-14 15:28:52 +02:00
Otto	ede6b87dbd	Merge pull request 'v0.5.3: sichere Sensorzuordnung für Aktoren' (#35 ) from fix/sensor-assignment-v0.5.3 into main	2026-06-14 15:19:55 +02:00
Otto	47e8c7e549	ASSIGN-001: reject unrelated actuator sensors	2026-06-14 15:19:30 +02:00
Otto	ef7e0c5600	Merge pull request 'v0.5.2: Add-on-Build liefert zuverlässig aktuellen Code' (#34 ) from fix/addon-cache-v0.5.2 into main	2026-06-14 11:38:51 +02:00
Otto	8d070fc9ca	BUILD-001: invalidate addon application cache per release	2026-06-14 11:38:36 +02:00
Otto	ba15cc4d83	Merge pull request 'v0.5.1: verständliche Ingress-Führung und vereinfachte Add-on-Konfiguration' (#33 ) from fix/ingress-guidance-v0.5.1 into main	2026-06-14 11:03:20 +02:00
Otto	da51ac2063	UI-001: simplify addon setup and explain ingress workflow	2026-06-14 11:02:38 +02:00
Otto	ce568056fc	Merge pull request 'v0.5.0: behavior learning, shadow prediction and safe activation' (#32 ) from feature/actuator-sensor-lifecycle into main Merge pull request v0.5.0 behavior learning and safe activation (#32)	2026-06-14 10:41:16 +02:00
Otto	da4603be17	BEHAVIOR-002: isolate per-actuator runtime failures Some checks failed quality / test (3.11) (push) Has been cancelled Details quality / test (3.13) (push) Has been cancelled Details	2026-06-14 10:40:19 +02:00
Otto	b215f23dd9	Merge remote-tracking branch 'origin/main' into feature/actuator-sensor-lifecycle	2026-06-14 10:38:33 +02:00
Otto	fa250216be	BEHAVIOR-001: learn and predict actuator actions	2026-06-14 10:37:59 +02:00
				`@@ -0,0 +1 @@`
				`"""Learning and prediction for actuator behavior."""`