From 9527fe931d4f03eb28cc7d64cdba99e207ad1477 Mon Sep 17 00:00:00 2001 From: Pawan-Shahane Date: Sun, 20 Sep 2026 11:58:13 +0530 Subject: [PATCH] fix(anthropic): return 400 for a content list holding non-dict items Building the Anthropic beta headers walks each message's content list and membership-tests or indexes every item, assuming a dict. A list holding a bare string, a null or a number therefore raised TypeError from inside header construction, which the caller saw as a 500 with no mention of their payload Three helpers on the validate_environment path each failed on a different input, so guarding one only moves the crash: a string containing "type" broke is_pdf_used, a string without it broke get_file_ids_from_messages, and a null broke is_cache_control_set. Skip non-dict items in all three, which lets the request reach the normal transformation and come back as a 400 A bare-string content list is not valid per the OpenAI message schema, so rejecting it is right; doing it as an internal error was not --- .../prompt_templates/common_utils.py | 2 +- litellm/llms/anthropic/common_utils.py | 4 +- .../anthropic/test_anthropic_common_utils.py | 61 +++++++++++++++++++ 3 files changed, 64 insertions(+), 3 deletions(-) diff --git a/litellm/litellm_core_utils/prompt_templates/common_utils.py b/litellm/litellm_core_utils/prompt_templates/common_utils.py index 7fedefa4025..cf1cc73bd89 100644 --- a/litellm/litellm_core_utils/prompt_templates/common_utils.py +++ b/litellm/litellm_core_utils/prompt_templates/common_utils.py @@ -1668,7 +1668,7 @@ def get_file_ids_from_messages(messages: list[AllMessageValues]) -> list[str]: if isinstance(content, str): continue for c in content: - if c["type"] == "file": + if isinstance(c, dict) and c["type"] == "file": file_object = cast(ChatCompletionFileObject, c) file_object_file_field = file_object.get("file") if not isinstance(file_object_file_field, dict): diff --git a/litellm/llms/anthropic/common_utils.py b/litellm/llms/anthropic/common_utils.py index 98e2f6d5bde..8ad0c12dbba 100644 --- a/litellm/llms/anthropic/common_utils.py +++ b/litellm/llms/anthropic/common_utils.py @@ -291,7 +291,7 @@ class AnthropicModelInfo(BaseLLMModelInfo): _message_content = message.get("content") if _message_content is not None and isinstance(_message_content, list): for content in _message_content: - if "cache_control" in content: + if isinstance(content, dict) and "cache_control" in content: return True return False @@ -336,7 +336,7 @@ class AnthropicModelInfo(BaseLLMModelInfo): for message in messages: if "content" in message and message["content"] is not None and isinstance(message["content"], list): for content in message["content"]: - if "type" in content and content["type"] != "text": + if isinstance(content, dict) and "type" in content and content["type"] != "text": return True return False diff --git a/tests/test_litellm/llms/anthropic/test_anthropic_common_utils.py b/tests/test_litellm/llms/anthropic/test_anthropic_common_utils.py index 945033c5cac..f2627ad3aed 100644 --- a/tests/test_litellm/llms/anthropic/test_anthropic_common_utils.py +++ b/tests/test_litellm/llms/anthropic/test_anthropic_common_utils.py @@ -2275,3 +2275,64 @@ def test_create_anthropic_model_list_response_lists_ids_as_told(): assert (gpt["id"], gpt["display_name"], gpt["max_input_tokens"]) == ("claude-router-gpt-4o[1m]", "GPT 4o", 1000000) assert (haiku["id"], haiku["display_name"]) == ("claude-haiku-4-5", "claude-haiku-4-5") assert (response["first_id"], response["last_id"]) == ("claude-router-gpt-4o[1m]", "claude-haiku-4-5") + + +class TestMalformedContentListItems: + """A `content` list holding bare strings is not valid per the OpenAI message + schema, but it reaches the Anthropic beta-header helpers before any validation + runs. Those helpers membership-test each item, so a non-dict item turned into a + substring test and then an unhandled TypeError, surfacing as a 500.""" + + @pytest.mark.parametrize( + "content", + [ + pytest.param(["what type of file is this?"], id="string_containing_type"), + pytest.param(["how do I set cache_control?"], id="string_containing_cache_control"), + pytest.param([None], id="none_item"), + pytest.param([5], id="int_item"), + pytest.param([["nested"]], id="list_item"), + ], + ) + def test_beta_headers_resolve_for_non_dict_content_items(self, content): + """The header path completes instead of raising, so the caller gets the + provider's own validation error rather than an internal failure.""" + from litellm.llms.anthropic.common_utils import AnthropicModelInfo + + config = AnthropicModelInfo() + messages = [{"role": "user", "content": content}] + + headers = config.validate_environment( + headers={}, + model="claude-sonnet-4-5", + messages=messages, + optional_params={}, + litellm_params={}, + api_key=FAKE_REGULAR_KEY, + ) + + assert headers["x-api-key"] == FAKE_REGULAR_KEY + + def test_real_content_parts_still_set_their_beta_headers(self): + """The guard skips malformed items only; a genuine media part must still be + detected, otherwise the pdf beta header would silently stop being sent.""" + from litellm.llms.anthropic.common_utils import AnthropicModelInfo + + config = AnthropicModelInfo() + + assert config.is_pdf_used([{"role": "user", "content": [{"type": "image", "source": {}}]}]) is True + assert config.is_pdf_used([{"role": "user", "content": [{"type": "text", "text": "hi"}]}]) is False + assert ( + config.is_cache_control_set( + [{"role": "user", "content": [{"type": "text", "text": "hi", "cache_control": {"type": "ephemeral"}}]}] + ) + is True + ) + + def test_mixed_list_keeps_detecting_the_valid_part(self): + """A malformed item earlier in the list must not mask a real media part after it.""" + from litellm.llms.anthropic.common_utils import AnthropicModelInfo + + config = AnthropicModelInfo() + messages = [{"role": "user", "content": ["what type of file is this?", {"type": "image", "source": {}}]}] + + assert config.is_pdf_used(messages) is True