feat: add privacy router and zero-cost safeguards

src/javis/config/settings.py

@@ -13,6 +13,35 @@ class ConfigurationError(ValueError):
     """Raised when runtime configuration is unsafe or invalid."""
 
 
+def _boolean(values: Mapping[str, str], name: str, default: bool) -> bool:
+    raw = values.get(name)
+    if raw is None:
+        return default
+    normalized = raw.strip().lower()
+    if normalized in {"1", "true", "yes", "on"}:
+        return True
+    if normalized in {"0", "false", "no", "off"}:
+        return False
+    raise ConfigurationError(f"{name} muss true oder false sein.")
+
+
+def _positive_integer(
+    values: Mapping[str, str],
+    name: str,
+    default: int,
+    *,
+    maximum: int,
+) -> int:
+    raw = values.get(name, str(default))
+    try:
+        result = int(raw)
+    except ValueError as exc:
+        raise ConfigurationError(f"{name} muss eine ganze Zahl sein.") from exc
+    if result <= 0 or result > maximum:
+        raise ConfigurationError(f"{name} muss zwischen 1 und {maximum} liegen.")
+    return result
+
+
 def default_data_dir(
     environ: Mapping[str, str] | None = None,
     *,
@@ -42,6 +71,18 @@ class Settings:
     model: str = "qwen3:8b"
     ollama_base_url: str = "http://127.0.0.1:11434"
     timeout_seconds: float = 180.0
+    provider_mode: str = "auto"
+    gemini_enabled: bool = False
+    gemini_model: str = "gemini-3.6-flash"
+    gemini_optional_model: str = "gemini-3.5-flash-lite"
+    free_only: bool = True
+    billing_confirmed_disabled: bool = False
+    max_cloud_requests_per_day: int = 25
+    max_cloud_input_chars: int = 12_000
+    max_cloud_output_tokens: int = 1_024
+    gemini_timeout_seconds: float = 30.0
+    gemini_max_retries: int = 1
+    max_cloud_context_messages: int = 6
 
     @property
     def database_path(self) -> Path:
@@ -59,11 +100,20 @@ class Settings:
         model = values.get("JAVIS_MODEL", "qwen3:8b").strip()
         base_url = values.get("JAVIS_OLLAMA_URL", "http://127.0.0.1:11434").strip()
         timeout_raw = values.get("JAVIS_MODEL_TIMEOUT", "180")
+        provider_mode = values.get("JAVIS_PROVIDER_MODE", "auto").strip().lower()
+        gemini_enabled = _boolean(values, "JAVIS_GEMINI_ENABLED", False)
+        free_only = _boolean(values, "JAVIS_FREE_ONLY", True)
+        billing_disabled = _boolean(values, "JAVIS_GEMINI_BILLING_CONFIRMED_DISABLED", False)
+        gemini_timeout_raw = values.get("JAVIS_GEMINI_TIMEOUT", "30")
 
         try:
             timeout = float(timeout_raw)
         except ValueError as exc:
             raise ConfigurationError("JAVIS_MODEL_TIMEOUT muss eine Zahl sein.") from exc
+        try:
+            gemini_timeout = float(gemini_timeout_raw)
+        except ValueError as exc:
+            raise ConfigurationError("JAVIS_GEMINI_TIMEOUT muss eine Zahl sein.") from exc
 
         settings = cls(
             data_dir=data_dir.resolve(),
@@ -71,6 +121,29 @@ class Settings:
             model=model,
             ollama_base_url=base_url,
             timeout_seconds=timeout,
+            provider_mode=provider_mode,
+            gemini_enabled=gemini_enabled,
+            gemini_model=values.get("JAVIS_GEMINI_MODEL", "gemini-3.6-flash").strip(),
+            gemini_optional_model=values.get(
+                "JAVIS_GEMINI_OPTIONAL_MODEL", "gemini-3.5-flash-lite"
+            ).strip(),
+            free_only=free_only,
+            billing_confirmed_disabled=billing_disabled,
+            max_cloud_requests_per_day=_positive_integer(
+                values, "JAVIS_MAX_CLOUD_REQUESTS_PER_DAY", 25, maximum=1_000
+            ),
+            max_cloud_input_chars=_positive_integer(
+                values, "JAVIS_MAX_CLOUD_INPUT_CHARS", 12_000, maximum=1_000_000
+            ),
+            max_cloud_output_tokens=_positive_integer(
+                values, "JAVIS_MAX_CLOUD_OUTPUT_TOKENS", 1_024, maximum=65_536
+            ),
+            gemini_timeout_seconds=gemini_timeout,
+            gemini_max_retries=_positive_integer(values, "JAVIS_GEMINI_MAX_ATTEMPTS", 2, maximum=2)
+            - 1,
+            max_cloud_context_messages=_positive_integer(
+                values, "JAVIS_MAX_CLOUD_CONTEXT_MESSAGES", 6, maximum=50
+            ),
         )
         settings.validate()
         return settings
@@ -84,6 +157,22 @@ class Settings:
             raise ConfigurationError("JAVIS_MODEL darf nicht leer sein.")
         if self.timeout_seconds <= 0:
             raise ConfigurationError("JAVIS_MODEL_TIMEOUT muss größer als 0 sein.")
+        if self.provider_mode not in {"auto", "local", "gemini"}:
+            raise ConfigurationError("JAVIS_PROVIDER_MODE muss auto, local oder gemini sein.")
+        if not self.free_only:
+            raise ConfigurationError(
+                "Kostenpflichtiger Betrieb wird nicht unterstützt; JAVIS_FREE_ONLY muss true sein."
+            )
+        if self.gemini_model != "gemini-3.6-flash":
+            raise ConfigurationError(
+                "In diesem Paket ist nur das geprüfte Standardmodell gemini-3.6-flash aktiv."
+            )
+        if not self.gemini_optional_model:
+            raise ConfigurationError("JAVIS_GEMINI_OPTIONAL_MODEL darf nicht leer sein.")
+        if self.gemini_timeout_seconds <= 0 or self.gemini_timeout_seconds > 60:
+            raise ConfigurationError(
+                "JAVIS_GEMINI_TIMEOUT muss größer als 0 und höchstens 60 sein."
+            )
 
         parsed = urlparse(self.ollama_base_url)
         if parsed.scheme != "http" or parsed.hostname not in {

src/javis/memory/usage_store.py

@@ -0,0 +1,114 @@
+"""Metadata-only provider audit and local usage limits."""
+
+from __future__ import annotations
+
+import sqlite3
+from contextlib import closing
+from dataclasses import dataclass
+from datetime import UTC, date, datetime
+from pathlib import Path
+
+
+class UsageStoreError(RuntimeError):
+    """Provider metadata could not be stored or read."""
+
+
+@dataclass(frozen=True, slots=True)
+class ProviderEvent:
+    provider: str
+    model: str
+    success: bool
+    error_category: str | None
+    fallback: bool
+    privacy_policy: str
+    input_tokens: int | None = None
+    output_tokens: int | None = None
+
+
+class SQLiteUsageStore:
+    def __init__(self, database_path: Path) -> None:
+        self.database_path = database_path
+        try:
+            self.database_path.parent.mkdir(parents=True, exist_ok=True)
+            with closing(self._connect()) as connection:
+                connection.execute(
+                    """
+                    CREATE TABLE IF NOT EXISTS provider_events (
+                        id INTEGER PRIMARY KEY AUTOINCREMENT,
+                        occurred_at TEXT NOT NULL,
+                        provider TEXT NOT NULL,
+                        model TEXT NOT NULL,
+                        success INTEGER NOT NULL,
+                        error_category TEXT,
+                        fallback INTEGER NOT NULL,
+                        privacy_policy TEXT NOT NULL,
+                        input_tokens INTEGER,
+                        output_tokens INTEGER
+                    )
+                    """
+                )
+                connection.execute(
+                    """
+                    CREATE INDEX IF NOT EXISTS idx_provider_events_time_provider
+                    ON provider_events(occurred_at, provider)
+                    """
+                )
+                connection.commit()
+        except (OSError, sqlite3.Error) as exc:
+            raise UsageStoreError(
+                f"Die Nutzungsmetadatenbank kann nicht geöffnet werden: {database_path}"
+            ) from exc
+
+    def _connect(self) -> sqlite3.Connection:
+        return sqlite3.connect(self.database_path, timeout=10)
+
+    def record(self, event: ProviderEvent) -> None:
+        try:
+            with closing(self._connect()) as connection:
+                connection.execute(
+                    """
+                    INSERT INTO provider_events (
+                        occurred_at, provider, model, success, error_category,
+                        fallback, privacy_policy, input_tokens, output_tokens
+                    ) VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?)
+                    """,
+                    (
+                        datetime.now(UTC).isoformat(timespec="seconds"),
+                        event.provider,
+                        event.model,
+                        int(event.success),
+                        event.error_category,
+                        int(event.fallback),
+                        event.privacy_policy,
+                        event.input_tokens,
+                        event.output_tokens,
+                    ),
+                )
+                connection.commit()
+        except sqlite3.Error as exc:
+            raise UsageStoreError("Nutzungsmetadaten konnten nicht gespeichert werden.") from exc
+
+    def cloud_requests_on(self, day: date | None = None) -> int:
+        return self._count(day or datetime.now(UTC).date(), "provider = 'gemini'")
+
+    def local_fallbacks_on(self, day: date | None = None) -> int:
+        return self._count(
+            day or datetime.now(UTC).date(),
+            "provider = 'ollama' AND fallback = 1",
+        )
+
+    def _count(self, day: date, condition: str) -> int:
+        start = f"{day.isoformat()}T00:00:00+00:00"
+        end = f"{day.isoformat()}T23:59:59+00:00"
+        try:
+            with closing(self._connect()) as connection:
+                row = connection.execute(
+                    f"""
+                    SELECT COUNT(*) FROM provider_events
+                    WHERE occurred_at BETWEEN ? AND ? AND {condition}
+                    """,
+                    (start, end),
+                ).fetchone()
+        except sqlite3.Error as exc:
+            raise UsageStoreError("Nutzungsmetadaten konnten nicht gelesen werden.") from exc
+        return int(row[0])

src/javis/security/privacy.py

@@ -0,0 +1,151 @@
+"""Conservative local privacy classification for cloud routing."""
+
+from __future__ import annotations
+
+import re
+from dataclasses import dataclass
+from enum import StrEnum
+
+from javis.providers.base import ChatMessage
+
+
+class CloudPolicy(StrEnum):
+    ALLOWED = "allowed"
+    ASK = "ask"
+    NEVER = "never"
+
+
+@dataclass(frozen=True, slots=True)
+class PrivacyDecision:
+    policy: CloudPolicy
+    reason: str
+
+
+_NEVER_PATTERNS: tuple[tuple[re.Pattern[str], str], ...] = (
+    (
+        re.compile(
+            r"(?i)\b(api[-_ ]?key|passwort|password|access[-_ ]?token|refresh[-_ ]?token|"
+            r"session[-_ ]?cookie|private[rn]? ssh[-_ ]?schl[üu]ssel)\b"
+        ),
+        "mögliches Secret",
+    ),
+    (
+        re.compile(r"(?i)(AIza[\w-]{20,}|sk-[\w-]{16,}|BEGIN [A-Z ]*PRIVATE KEY)"),
+        "Schlüsselmuster",
+    ),
+    (
+        re.compile(
+            r"(?i)\b(iban|bic|kreditkart|bankkonto|bankdaten|zahlungsdaten|"
+            r"kontonummer|pin[- ]?code)\b"
+        ),
+        "Finanz- oder Zahlungsdaten",
+    ),
+    (
+        re.compile(
+            r"(?i)\b(diagnose|gesundheit|krankheit|medikament|patient|arztbericht|"
+            r"psychotherapie|blutwert)\b"
+        ),
+        "Gesundheitsinformation",
+    ),
+    (
+        re.compile(
+            r"(?i)\b(genaue anschrift|wohnanschrift|private[rs]? dokument|"
+            r"vollständige[rs]? personenprofil|rohe? obsidian|nur lokal|"
+            r"ausschließlich lokal)\b"
+        ),
+        "ausdrücklich lokal oder besonders sensibel",
+    ),
+)
+
+_ASK_PATTERNS: tuple[tuple[re.Pattern[str], str], ...] = (
+    (
+        re.compile(
+            r"(?i)\b(meine frau|mein mann|meine familie|mein kind|meine mutter|"
+            r"mein vater|meine beziehung|über mich|über pascal)\b"
+        ),
+        "persönliche oder familiäre Information",
+    ),
+    (
+        re.compile(r"(?i)\b[\w.+-]+@[\w.-]+\.[a-z]{2,}\b"),
+        "E-Mail-Adresse",
+    ),
+    (
+        re.compile(r"(?<!\w)(?:\+?\d[\d ()/-]{7,}\d)(?!\w)"),
+        "mögliche Telefonnummer",
+    ),
+    (
+        re.compile(
+            r"(?i)\b(intern(?:e[rsn]?|er)? hostname|interne? domain|rootserver|"
+            r"privates? projekt|nicht öffentlich|systeminformation|"
+            r"private[rs]? termin)\b"
+        ),
+        "nicht öffentliche Infrastruktur oder Planung",
+    ),
+    (
+        re.compile(r"(?i)\b(ich|mein(?:e[rsn]?|em)?|mir|mich)\b"),
+        "möglicher persönlicher Bezug",
+    ),
+)
+
+_ALLOWED_PATTERNS: tuple[re.Pattern[str], ...] = (
+    re.compile(
+        r"(?i)\b(python|programmier|quellcode|algorithmus|sqlite|linux|windows|"
+        r"öffentliche dokumentation|allgemeine wissensfrage|was ist|wie funktioniert)\b"
+    ),
+)
+
+
+class PrivacyRouter:
+    """Rule-based, entirely local and intentionally conservative."""
+
+    def classify(self, text: str, *, cloud_requested: bool = False) -> PrivacyDecision:
+        normalized = text.strip()
+        for pattern, reason in _NEVER_PATTERNS:
+            if pattern.search(normalized):
+                return PrivacyDecision(CloudPolicy.NEVER, reason)
+        for pattern, reason in _ASK_PATTERNS:
+            if pattern.search(normalized):
+                return PrivacyDecision(CloudPolicy.ASK, reason)
+        if any(pattern.search(normalized) for pattern in _ALLOWED_PATTERNS):
+            return PrivacyDecision(CloudPolicy.ALLOWED, "allgemeiner oder technischer Inhalt")
+        if cloud_requested:
+            return PrivacyDecision(
+                CloudPolicy.ASK,
+                "expliziter Cloudwunsch bei uneindeutigem Inhalt",
+            )
+        return PrivacyDecision(CloudPolicy.ASK, "Inhalt nicht eindeutig klassifizierbar")
+
+    def minimal_context(
+        self,
+        messages: list[ChatMessage],
+        *,
+        current_policy: CloudPolicy,
+        approved: bool,
+        max_chars: int,
+        max_messages: int,
+    ) -> list[ChatMessage]:
+        if current_policy is CloudPolicy.NEVER:
+            return []
+        if current_policy is CloudPolicy.ASK and not approved:
+            return []
+
+        selected: list[ChatMessage] = []
+        remaining = max_chars
+        for index in range(len(messages) - 1, max(-1, len(messages) - max_messages - 1), -1):
+            message = messages[index]
+            decision = self.classify(message.content)
+            is_current = index == len(messages) - 1
+            may_include = decision.policy is CloudPolicy.ALLOWED or (
+                is_current and current_policy is CloudPolicy.ASK and approved
+            )
+            if not may_include:
+                continue
+            content = message.content[:remaining]
+            if not content:
+                break
+            selected.append(ChatMessage(message.role, content))
+            remaining -= len(content)
+            if remaining <= 0:
+                break
+        selected.reverse()
+        return selected