BEHAVIOR-003: learn causal shadow triggers

app/behavior/engine.py

@@ -22,6 +22,7 @@ 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"})
@@ -198,12 +199,18 @@ class BehaviorEngine:
             )
             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,
         )
         behavior = record.behavior.model_copy(
@@ -326,7 +333,12 @@ class BehaviorEngine:
             if _matches_own_execution(point, own_executions):
                 continue
             source, weight = _action_source(point, logbook)
-            if source == "automation":
+            trigger = _recent_context_transition(
+                context_history,
+                context_ids,
+                point.timestamp,
+            )
+            if source == "automation" and trigger is None:
                 continue
             contexts = {
                 entity_id: state
@@ -340,6 +352,9 @@ class BehaviorEngine:
                     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,
@@ -373,14 +388,52 @@ def predict_behavior(
     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
@@ -414,6 +467,7 @@ def predict_behavior(
         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
@@ -423,7 +477,12 @@ def predict_behavior(
         generated_at=now,
         matching_patterns=support,
         reason=(
-            f"{support} ähnliche Handlungsmuster passen zu Zeit und aktuellem Kontext."
+            (
+                f"{causal_support} historische Handlungen folgten demselben "
+                "frischen Sensorwechsel."
+            )
+            if causal_support
+            else f"{support} ähnliche Handlungsmuster passen zu Zeit und aktuellem Kontext."
         ),
     )
 
@@ -476,6 +535,32 @@ def _matches_own_execution(
     )
 
 
+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)