Cap runaway chat generations in router

compose.yaml

@@ -567,6 +567,10 @@ services:
       PROFILE_CONTROL_TOKEN: "${CONTROLLER_TOKEN:?CONTROLLER_TOKEN is required}"
       SWITCH_TIMEOUT: "600"
       REQUEST_TIMEOUT: "600"
+      # Last-resort guard for every OpenAI-compatible client. Without a
+      # request limit llama.cpp uses n_predict=-1 and a reasoning loop can
+      # consume the complete context before yielding visible output.
+      MAX_GENERATION_TOKENS: "8192"
       IMAGE_DIR: /data/images
       IMAGE_WORKER_URL: http://flux-worker:8086
       IMAGE_WORKER_TOKEN: "${CONTROLLER_TOKEN:?CONTROLLER_TOKEN is required}"

dev/test_router_support.py

@@ -24,6 +24,7 @@ from router_support import (  # noqa: E402
     load_profile_registry,
 )
 from ai_profile_router import (  # noqa: E402
+    _cap_chat_generation,
     _context_matches,
     _normalize_chat_image,
     _normalize_chat_images,
@@ -176,6 +177,25 @@ class LlamaCppReasoningTests(unittest.TestCase):
         self.assertNotIn("chat_template_kwargs", request)
 
 
+class ChatGenerationLimitTests(unittest.TestCase):
+    def test_missing_limit_gets_platform_cap(self) -> None:
+        request = {"messages": []}
+        self.assertEqual(_cap_chat_generation(request)["max_tokens"], 8192)
+
+    def test_smaller_explicit_limit_is_preserved(self) -> None:
+        request = {"max_tokens": 512}
+        self.assertEqual(_cap_chat_generation(request)["max_tokens"], 512)
+
+    def test_unlimited_and_oversized_values_are_capped(self) -> None:
+        for value in (-1, 0, 999999):
+            with self.subTest(value=value):
+                request = {"max_completion_tokens": value}
+                self.assertEqual(
+                    _cap_chat_generation(request)["max_completion_tokens"],
+                    8192,
+                )
+
+
 class RetentionTests(unittest.TestCase):
     def test_oldest_pairs_are_removed(self) -> None:
         with tempfile.TemporaryDirectory() as temp:

router/ai_profile_router.py

@@ -113,6 +113,7 @@ SWITCH_TIMEOUT = float(os.environ.get("SWITCH_TIMEOUT", "600"))     # s, Warten
 REQUEST_TIMEOUT = float(os.environ.get("REQUEST_TIMEOUT", "600"))   # s, Read-Timeout Upstream
 CONNECT_TIMEOUT = float(os.environ.get("CONNECT_TIMEOUT", "10"))    # s, Connect-Timeout
 POLL_INTERVAL = float(os.environ.get("POLL_INTERVAL", "2"))         # s, Polling-Intervall
+MAX_GENERATION_TOKENS = int(os.environ.get("MAX_GENERATION_TOKENS", "8192"))
 
 # --- Bildgenerierung (FLUX.2 [klein] 4B Base) ---
 LLAMA_SERVICE = os.environ.get("LLAMA_SERVICE", "mike-ai-llama-ui.service")
@@ -1166,6 +1167,39 @@ def _normalize_llamacpp_reasoning(data: dict) -> dict:
     return data
 
 
+def _cap_chat_generation(data: dict) -> dict:
+    """Apply a client-independent upper bound to one chat generation.
+
+    Some OpenAI-compatible clients omit both token-limit fields. llama.cpp
+    interprets that as ``n_predict=-1`` and can remain in hidden reasoning
+    until the context is exhausted. Smaller explicit limits are preserved.
+    """
+    cap = MAX_GENERATION_TOKENS
+    if cap <= 0:
+        return data
+
+    fields = ("max_tokens", "max_completion_tokens")
+    present = False
+    for field in fields:
+        if field not in data:
+            continue
+        present = True
+        value = data[field]
+        if isinstance(value, bool):
+            data[field] = cap
+            continue
+        try:
+            parsed = int(value)
+        except (TypeError, ValueError):
+            data[field] = cap
+            continue
+        data[field] = min(parsed, cap) if parsed > 0 else cap
+
+    if not present:
+        data["max_tokens"] = cap
+    return data
+
+
 # ---------------------------------------------------------------------------
 # HTTP-Handler
 # ---------------------------------------------------------------------------
@@ -1975,6 +2009,7 @@ class Handler(BaseHTTPRequestHandler):
                     switch_profile(profile, implicit=True)
 
                 data = _normalize_llamacpp_reasoning(data)
+                data = _cap_chat_generation(data)
 
                 if _request_has_image(data):
                     data = _normalize_chat_images(data)