Update app.py

app.py

@@ -7,7 +7,8 @@ import json
 # 1. KONFIGURATION & AUFGABEN-PROFILE
 # ==========================================
 
-MODERATOR_MODEL = "meta-llama/Llama-3.3-70B-Instruct"
+MODERATOR_MODEL = "openai/gpt-oss-120b"
+MODERATOR_PROVIDER = "groq"  # 435 tok/s — wird oft aufgerufen, Geschwindigkeit summiert sich
 
 # Jedes Profil definiert, wie Experten denken und was am Ende rauskommt
 TASK_PROFILES = {
@@ -79,7 +80,7 @@ COUNCIL_MEMBERS = [
     },
     {
         "name": "🧐 Experte II",
-        "model": "deepseek-ai/DeepSeek-V3.2",
+        "model": "deepseek-ai/DeepSeek-R1",  # Reasoning-Modell — ideal für Kritik & Fehlersuche
         "tag": "KRITIK",
         "color": "#7c3aed",
         "role_hint": (
@@ -89,7 +90,8 @@ COUNCIL_MEMBERS = [
     },
     {
         "name": "🛠️ Experte III",
-        "model": "meta-llama/Llama-4-Maverick-17B-128E-Instruct",
+        "model": "moonshotai/Kimi-K2.5",  # andere Trainings-Lineage → echte Diversität im Council
+        "code_model": "Qwen/Qwen3-Coder-480B-A35B-Instruct",  # Swap für CODE-Tasks
         "tag": "UMSETZUNG",
         "color": "#b45309",
         "role_hint": (
@@ -224,11 +226,12 @@ class LLMService:
             api_key=os.getenv("HF_TOKEN"),
         )
 
-    def ask(self, model_id, system_prompt, user_input):
+    def ask(self, model_id, system_prompt, user_input, provider=None):
         messages = [
             {"role": "system", "content": system_prompt},
             {"role": "user", "content": user_input},
         ]
+        extra_headers = {"x-provider": provider} if provider else {}
         try:
             response = self.client.chat.completions.create(
                 model=model_id,
@@ -236,6 +239,7 @@ class LLMService:
                 max_tokens=4000,
                 temperature=0.4,
                 stream=False,
+                extra_headers=extra_headers,
             )
             return response.choices[0].message.content or ""
         except Exception as e:
@@ -274,6 +278,7 @@ class PlenumOrchestrator:
             MODERATOR_MODEL,
             self.pm.task_detection_sys(),
             self.pm.task_detection_user(user_prompt),
+            provider=MODERATOR_PROVIDER,
         )
         try:
             clean = raw.strip().removeprefix("```json").removeprefix("```").removesuffix("```").strip()
@@ -317,6 +322,7 @@ class PlenumOrchestrator:
             MODERATOR_MODEL,
             self.pm.moderator_kickoff_sys(task_profile),
             self.pm.moderator_kickoff_user(user_prompt, task_info),
+            provider=MODERATOR_PROVIDER,
         )
         history.append({"role": "assistant", "content": self.ui.message("🎤 Moderator", kickoff, "#FF5A4D")})
         yield history
@@ -335,6 +341,7 @@ class PlenumOrchestrator:
                     MODERATOR_MODEL,
                     self.pm.moderator_steering_sys(),
                     self.pm.moderator_steering_user(current_draft, r),
+                    provider=MODERATOR_PROVIDER,
                 )
                 history.append({
                     "role": "assistant",
@@ -352,10 +359,17 @@ class PlenumOrchestrator:
             for idx, expert in enumerate(COUNCIL_MEMBERS):
                 focus = task_profile["expert_focuses"][idx]
 
+                # Task-aware Model Swap: Experte III nutzt Coder-Modell bei CODE-Tasks
+                model_id = (
+                    expert.get("code_model", expert["model"])
+                    if task_type == "CODE" and expert["tag"] == "UMSETZUNG"
+                    else expert["model"]
+                )
+
                 sys_msg = self.pm.expert_sys(expert, task_profile, focus, r + 1)
                 usr_msg = self.pm.expert_user(user_prompt, round_draft, steering_instruction if r > 0 else "", expert["tag"])
 
-                answer = self.llm.ask(expert["model"], sys_msg, usr_msg)
+                answer = self.llm.ask(model_id, sys_msg, usr_msg, provider=expert.get("provider"))
 
                 # Dieser Experte liefert den Entwurf für den Nächsten
                 round_draft = answer
@@ -380,6 +394,7 @@ class PlenumOrchestrator:
             MODERATOR_MODEL,
             self.pm.final_sys(task_profile),
             self.pm.final_user(user_prompt, current_draft, task_info),
+            provider=MODERATOR_PROVIDER,
         )
         history.append({"role": "assistant", "content": final})
         yield history