update

backend/agents.py

@@ -82,11 +82,12 @@ async def _api_slot(key: str) -> None:
 
 
 class AgentErgebnis:
-    __slots__ = ("rc", "text", "err", "tokens", "wait_s")
+    __slots__ = ("rc", "text", "err", "tokens", "wait_s", "model")
 
     def __init__(self, rc: int, text: str, err: str, tokens: dict | None = None):
         self.rc, self.text, self.err, self.tokens = rc, text, err, tokens or {}
         self.wait_s = 0.0  # Queue-Zeit (Semaphore/Drossel) — run_agent füllt sie
+        self.model = ""    # aufgelöstes Modell — run_agent füllt es (Ledger)
 
     @property
     def ok(self) -> bool:
@@ -161,12 +162,13 @@ async def run_agent(key: str, prompt: str, timeout: int, *, provider: str,
                 await _api_slot(key)
                 wait_s = time.time() - warte_start
                 try:
-                    res = await _text_api(key, prompt, timeout, model)
+                    res = await _text_api(key, prompt, timeout, model, role)
                 finally:
                     _api_inflight -= 1
                 if res.rc == 0:
                     _erfolg_melden()
                 res.wait_s = wait_s
+                res.model = model
                 return res
             await _drossel_warten(key)
             wait_s = time.time() - warte_start
@@ -175,6 +177,7 @@ async def run_agent(key: str, prompt: str, timeout: int, *, provider: str,
             else:
                 res = await _claude_cli(key, prompt, timeout, model, capabilities)
             res.wait_s = wait_s
+            res.model = model
             return res
         finally:
             _active.pop(key, None)
@@ -335,12 +338,20 @@ _API_MODEL_OPTS = {
     # Antwort nach bis zu 44 min — Extraktion war 53 von 63 min der Ebene).
     "minimax-kalt/MiniMax-M2.7-highspeed": {"temperature": 0.3,
                                             "thinking": {"type": "disabled"}},
+    # native Route (judge/guide): ungedrosselt liefen Merge-Judges ins Output-Cap
+    # (58–71 % Parse-Fehler). role=guide (Writer) behält Thinking — _text_api
+    # nimmt die thinking-Option dort wieder raus.
+    "minimax/MiniMax-M3": {"temperature": 0.2, "thinking": {"type": "disabled"}},
 }
 
 
-async def _text_api(key: str, prompt: str, timeout: int, model: str) -> AgentErgebnis:
+async def _text_api(key: str, prompt: str, timeout: int, model: str,
+                    role: str = "judge") -> AgentErgebnis:
+    opts = dict(_API_MODEL_OPTS.get(model, {}))
+    if role == "guide":
+        opts.pop("thinking", None)  # Writer braucht Denkraum — nur Judges kalt
     body = {"model": model.split("/", 1)[1], "max_tokens": _API_MAX_TOKENS,
-            "messages": [{"role": "user", "content": prompt}], **_API_MODEL_OPTS.get(model, {})}
+            "messages": [{"role": "user", "content": prompt}], **opts}
     headers = {"x-api-key": os.environ.get("MINIMAX_API_KEY", ""), "anthropic-version": "2023-06-01"}
     try:
         async with httpx.AsyncClient(timeout=httpx.Timeout(timeout, connect=30)) as client: