From 811768a38d08d0e7c852ebff7557d4cf745c5fc1 Mon Sep 17 00:00:00 2001
From: taltal-beep <ttlgptinc@gmail.com>
Date: Sun, 11 Oct 2026 09:57:31 +0300
Subject: [PATCH] test(utils): cover trim_messages return_response_tokens when
 trimming fails

Also say in the docstring which limit response_tokens is measured against.
---
 litellm/utils.py         |  5 +++--
 tests/unit/test_utils.py | 15 +++++++++++++++
 2 files changed, 18 insertions(+), 2 deletions(-)

diff --git a/litellm/utils.py b/litellm/utils.py
index 97590fafe64..899f1cf80f3 100644
--- a/litellm/utils.py
+++ b/litellm/utils.py
@@ -7700,8 +7700,9 @@ def trim_messages(
 
     Returns:
         Trimmed messages. If return_response_tokens is True, a `(messages, response_tokens)` tuple instead,
-        whether or not anything had to be trimmed. `response_tokens` is None when there is no token limit to
-        measure against (unknown model and no `max_tokens`).
+        whether or not anything had to be trimmed. `response_tokens` is what is left of the limit the messages
+        were trimmed to (`max_tokens`, or the model's max input tokens times `trim_ratio`). It is None when
+        there is no token limit to measure against (unknown model and no `max_tokens`) or trimming failed.
     """
     # Initialize max_tokens
     # if users pass in max tokens, trim to this amount
diff --git a/tests/unit/test_utils.py b/tests/unit/test_utils.py
index 29f8180fb5e..e4c9849ac4e 100644
--- a/tests/unit/test_utils.py
+++ b/tests/unit/test_utils.py
@@ -7626,6 +7626,21 @@ def test_trim_messages_return_response_tokens_is_none_when_no_token_limit_is_kno
     assert response_tokens is None
 
 
+def test_trim_messages_return_response_tokens_returns_a_tuple_when_trimming_fails(monkeypatch):
+    def failing_token_count(*args, **kwargs):
+        raise RuntimeError("tokenizer unavailable")
+
+    monkeypatch.setattr(litellm.utils, "get_token_count", failing_token_count)
+    messages = [{"role": "user", "content": "hi"}]
+
+    assert trim_messages(messages, max_tokens=1000) == messages
+
+    trimmed_messages, response_tokens = trim_messages(messages, max_tokens=1000, return_response_tokens=True)
+
+    assert trimmed_messages == messages
+    assert response_tokens is None
+
+
 @pytest.mark.parametrize("model", ["gpt-5.4-mini", "claude-sonnet-4-6"])
 def test_trimming_with_model_cost_max_input_tokens(model):
     messages = [