Add global research verification policy

router/ai_profile_router.py

@@ -121,6 +121,8 @@ POLL_INTERVAL = float(os.environ.get("POLL_INTERVAL", "2"))         # s, Polling
 MAX_GENERATION_TOKENS = int(os.environ.get("MAX_GENERATION_TOKENS", "8192"))
 DEFAULT_REASONING_EFFORT = os.environ.get(
     "DEFAULT_REASONING_EFFORT", "off").strip().lower()
+GLOBAL_SYSTEM_POLICY_FILE = os.environ.get(
+    "GLOBAL_SYSTEM_POLICY_FILE", "").strip()
 
 # --- Bildgenerierung und Referenzbild-Bearbeitung (FLUX.2 Klein 4B) ---
 LLAMA_SERVICE = os.environ.get("LLAMA_SERVICE", "mike-ai-llama-ui.service")
@@ -1270,6 +1272,57 @@ _REASONING_EFFORT_MAP = {
 _REASONING_OFF = {"", "none", "off", "disabled", "false"}
 
 
+def _load_global_system_policy() -> str:
+    """Load the static cross-client platform policy.
+
+    The file is read for every request so operators can revise the policy
+    without rebuilding or restarting the router. Its contents remain stable
+    between edits and therefore remain friendly to upstream prompt caches.
+    """
+    if not GLOBAL_SYSTEM_POLICY_FILE:
+        return ""
+    try:
+        with open(GLOBAL_SYSTEM_POLICY_FILE, encoding="utf-8") as handle:
+            return handle.read().strip()
+    except OSError as exc:
+        raise ValueError(
+            f"globale Systemrichtlinie nicht lesbar: {exc}") from exc
+
+
+def _inject_global_system_policy(data: dict, path: str) -> dict:
+    """Prepend the shared policy to OpenAI chat and Responses requests."""
+    policy = _load_global_system_policy()
+    if not policy:
+        return data
+
+    if path == "/v1/chat/completions":
+        messages = data.get("messages")
+        if not isinstance(messages, list):
+            return data
+        if messages and isinstance(messages[0], dict) and (
+                messages[0].get("role") == "system"
+                and isinstance(messages[0].get("content"), str)):
+            existing = messages[0]["content"]
+            if policy not in existing:
+                messages[0]["content"] = f"{policy}\n\n{existing}"
+        elif not any(
+                isinstance(message, dict)
+                and message.get("role") == "system"
+                and message.get("content") == policy
+                for message in messages):
+            messages.insert(0, {"role": "system", "content": policy})
+        return data
+
+    if path == "/v1/responses":
+        instructions = data.get("instructions")
+        if isinstance(instructions, str) and instructions:
+            if policy not in instructions:
+                data["instructions"] = f"{policy}\n\n{instructions}"
+        elif instructions is None or instructions == "":
+            data["instructions"] = policy
+    return data
+
+
 def _normalize_llamacpp_reasoning(data: dict) -> dict:
     """Mappt OpenAI/Hermes-Reasoning auf llama.cpp-Template-Parameter.
 
@@ -2203,6 +2256,15 @@ class Handler(BaseHTTPRequestHandler):
                                  "invalid_request_error", "unknown_model")
                 return
 
+            if path in {"/v1/chat/completions", "/v1/responses"}:
+                try:
+                    data = _inject_global_system_policy(data, path)
+                except ValueError as exc:
+                    self._send_error(500, str(exc), "server_error",
+                                     "system_policy_unavailable")
+                    return
+                body = json.dumps(data).encode()
+
         # Alle modellbezogenen Requests erhalten eine atomare Lease. Damit
         # kann kein zweiter Client zwischen Profilwahl und Upstream-Request das
         # Modell austauschen. Vision-Vorbereitung gehört zur selben Transaktion.