Merge 95947f29fe into 4e48d74455

litellm/utils.py

@@ -7117,7 +7117,10 @@ def shorten_message_to_fit_limit(message, tokens_needed, model: str | None, rais
 
         half_length = new_length // 2
         left_half = content[:half_length]
-        right_half = content[-half_length:]
+        # content[-0:] is content[0:], i.e. the whole string, so a zero half has to be
+        # spelled out. Otherwise every further attempt prepends ".." to the full content
+        # and the message grows by two characters instead of shrinking.
+        right_half = content[-half_length:] if half_length else ""
 
         trimmed_content = left_half + ".." + right_half
         message["content"] = trimmed_content

tests/test_litellm/test_utils.py

@@ -5661,3 +5661,21 @@ def test_snapshot_exception_for_hook_preserves_suppress_context_flag() -> None:
     snapshot = _snapshot_exception_for_hook(e)
     assert snapshot.__suppress_context__ is False
     assert snapshot.__context__ is e.__context__
+
+
+def test_shorten_message_to_fit_limit_never_grows_content():
+    """A zero half_length must not turn the trim into a two-character prefix.
+
+    `content[-0:]` is `content[0:]`, so with half_length == 0 the "right half" is the
+    whole string and each attempt returns `".." + content`. The loop then runs its full
+    attempt budget growing the message two characters at a time.
+    """
+    from litellm.utils import shorten_message_to_fit_limit
+
+    content = "hello world " * 40
+    message = {"role": "user", "content": content}
+
+    result = shorten_message_to_fit_limit(message, tokens_needed=1, model="claude-3-5-sonnet-20240620")
+
+    assert len(result["content"]) < len(content)
+    assert not result["content"].startswith("....")