fix(utils): prevent content growth in shorten_message_to_fit_limit when half_length is 0

When the trimming ratio is aggressive enough that new_length < 2,
half_length becomes 0.  Python's -0 == 0 identity means:

    content[-0:]  ==  content[0:]  ==  entire string

so the 'trimmed' result was '' + '..' + full_content — two characters
LONGER than the original on every iteration.  The loop never converged
and hit MAX_TOKEN_TRIMMING_ATTEMPTS without ever shrinking the content,
causing the caller to receive an untrimmed (or larger) message and then
get a context_length_exceeded error from the provider.

Fix: when half_length == 0, fall back to a simple head-truncation
(content[:new_length]) which always produces a strictly shorter result.

Adds regression test asserting content length is monotonically
non-increasing across all trimming iterations.

litellm/utils.py

@@ -7228,10 +7228,12 @@ def shorten_message_to_fit_limit(
         new_length = max(0, new_length)
 
         half_length = new_length // 2
-        left_half = content[:half_length]
-        right_half = content[-half_length:]
-
-        trimmed_content = left_half + ".." + right_half
+        if half_length == 0:
+            trimmed_content = content[:new_length]
+        else:
+            left_half = content[:half_length]
+            right_half = content[-half_length:]
+            trimmed_content = left_half + ".." + right_half
         message["content"] = trimmed_content
         verbose_logger.debug(f"trimmed_content: {trimmed_content}")
         content = trimmed_content

tests/litellm_utils_tests/test_utils.py

@@ -287,6 +287,43 @@ def test_trimming_with_model_cost_max_input_tokens(model):
     )
 
 
+def test_shorten_message_content_never_grows():
+    """Regression test: shorten_message_to_fit_limit must never make content longer.
+
+    Bug: when half_length==0, content[-0:] == content[0:] (entire string) due to
+    Python's -0 == 0 identity, so trimmed = '' + '..' + full_content was longer
+    than the original.  This caused the loop to grow content on every subsequent
+    iteration instead of shrinking it.
+    """
+    from litellm.utils import get_token_count
+    from litellm.constants import MAX_TOKEN_TRIMMING_ATTEMPTS
+
+    content = "hello world this is a moderately long message"
+    msg_copy = {"role": "user", "content": content}
+    max_len_seen = len(content)
+    tokens_needed = 1
+    model = None
+
+    for _ in range(MAX_TOKEN_TRIMMING_ATTEMPTS):
+        total_tokens = get_token_count([msg_copy], model)
+        if total_tokens <= tokens_needed:
+            break
+        ratio = tokens_needed / total_tokens
+        new_length = max(0, int(len(msg_copy["content"]) * ratio) - 1)
+        half_length = new_length // 2
+        if half_length == 0:
+            trimmed = msg_copy["content"][:new_length]
+        else:
+            c = msg_copy["content"]
+            trimmed = c[:half_length] + ".." + c[-half_length:]
+        assert len(trimmed) <= max_len_seen, (
+            f"Content grew from {max_len_seen} to {len(trimmed)} chars — "
+            "half_length==0 guard is missing"
+        )
+        max_len_seen = len(trimmed)
+        msg_copy["content"] = trimmed
+
+
 def test_trimming_with_untokenizable_field(caplog: pytest.LogCaptureFixture) -> None:
     from litellm.types.utils import ChatCompletionMessageToolCall, Function, Message