Python: fix(lab-tau2): check the over-max warning before the half-of-max warning

python/packages/lab/tau2/agent_framework_lab_tau2/_sliding_window.py

@@ -100,16 +100,16 @@ def _get_token_count(self, messages: list[Message]) -> int:
                 # Simple text message
                 total_tokens += self._estimate_any_object_token_count(msg.text)
 
-        if total_tokens > self.max_tokens / 2:
+        if total_tokens > self.max_tokens:
+            logger.opt(colors=True).warning(
+                f"<red>Total tokens {total_tokens} is over max tokens {self.max_tokens}. Will truncate messages.</red>"
+            )
+        elif total_tokens > self.max_tokens / 2:
             logger.opt(colors=True).warning(
                 f"<yellow>Total tokens {total_tokens} is "
                 f"{total_tokens / self.max_tokens * 100:.0f}% "
                 f"of max tokens {self.max_tokens}</yellow>"
             )
-        elif total_tokens > self.max_tokens:
-            logger.opt(colors=True).warning(
-                f"<red>Total tokens {total_tokens} is over max tokens {self.max_tokens}. Will truncate messages.</red>"
-            )
 
         return total_tokens
 

python/packages/lab/tau2/tests/test_sliding_window.py

@@ -211,3 +211,39 @@ async def test_real_world_scenario():
 
     token_count = provider._get_token_count(truncated)
     assert token_count <= provider.max_tokens * 1.1
+
+
+def test_get_token_count_warns_over_max():
+    """The over-max warning must win over the half-of-max warning (#9338)."""
+    import io
+
+    from loguru import logger
+
+    provider = SlidingWindowHistoryProvider(max_tokens=10)
+    messages = [
+        Message(
+            role="user",
+            contents=[
+                Content.from_text(
+                    text=(
+                        "hello world, this message is intentionally long "
+                        "so it clearly exceeds the tiny ten token budget"
+                    )
+                )
+            ],
+        )
+    ]
+
+    total = provider._get_token_count(messages)
+    assert total > provider.max_tokens
+
+    buffer = io.StringIO()
+    handler_id = logger.add(buffer, format="{message}")
+    try:
+        provider._get_token_count(messages)
+    finally:
+        logger.remove(handler_id)
+
+    logged = buffer.getvalue()
+    assert "over max tokens" in logged
+    assert "Will truncate messages" in logged