Cap runaway chat generations in router

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: