Merge acf73c271a into dab2deb5ed

litellm/utils.py

@@ -7536,7 +7536,8 @@ 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 the whole string
+        right_half = content[-half_length:] if half_length else ""
 
         trimmed_content = left_half + ".." + right_half
         message["content"] = trimmed_content

tests/unit/test_utils.py

@@ -6012,6 +6012,23 @@ def test_load_credentials_from_list_fills_kwargs_from_the_loaded_credential_with
     }
     assert _credential_warnings(caplog) == []
 
+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("....")
+
 
 _MOCK_STREAM_ID: Final = "chatcmpl-mock-stream"
 _ChunkSnapshot = tuple[str, tuple[str | None, ...], Usage | None]