fix: preserve provider on LiteLLM-routed models (#6849)

* fix: preserve provider on LiteLLM-routed models

LiteLLM construction computed the real provider in `__new__` but never
passed it into init, so `BaseLLM` silently defaulted every shared-path
model to `openai`. Infer the provider from a `provider/model` prefix
when none is supplied so groq, cohere, mistral, and the rest report
themselves correctly to callers like the policy engine.

* fix: avoid double-prefixing instructor model strings

With LiteLLM models now carrying a real `provider` while `model` keeps
its `provider/name` form, `InternalInstructor` was building
`groq/groq/...` for `instructor.from_provider`. Skip the prefix when
the model string is already qualified.

* fix: format LiteLLM multimodal content as OpenAI-shaped blocks

Preserving the real provider on the LiteLLM path made
`format_multimodal_content` emit Anthropic-native blocks for
`anthropic/...` models, which LiteLLM rejects. Keep `provider` as the
model identity for policies, but format multimodal blocks with the
OpenAI chat schema when `is_litellm` is set. Expose the formatter helper
on `BaseLLM` so native OpenAI/Azure completions share the same API.

Author: lucasgomide

lib/crewai/src/crewai/llm.py

@@ -2225,14 +2225,14 @@ def _process_message_files(self, messages: list[LLMMessage]) -> list[LLMMessage]
                 )
             return messages
 
-        provider = self.provider or self.model
+        formatter = self._multimodal_formatter_name()
 
         for msg in messages:
             files = msg.get("files")
             if not files:
                 continue
 
-            content_blocks = format_multimodal_content(files, provider)
+            content_blocks = format_multimodal_content(files, formatter)
             if not content_blocks:
                 msg.pop("files", None)
                 continue
@@ -2250,6 +2250,13 @@ def _process_message_files(self, messages: list[LLMMessage]) -> list[LLMMessage]
 
         return messages
 
+    def _multimodal_formatter_name(self) -> str:
+        # Identity (`self.provider`) stays e.g. anthropic. LiteLLM's completion()
+        # API is OpenAI-shaped and translates blocks to the vendor on the wire.
+        if self.is_litellm:
+            return "openai"
+        return self.provider or self.model
+
     async def _aprocess_message_files(
         self, messages: list[LLMMessage]
     ) -> list[LLMMessage]:
@@ -2276,14 +2283,14 @@ async def _aprocess_message_files(
                 )
             return messages
 
-        provider = self.provider or self.model
+        formatter = self._multimodal_formatter_name()
 
         for msg in messages:
             files = msg.get("files")
             if not files:
                 continue
 
-            content_blocks = await aformat_multimodal_content(files, provider)
+            content_blocks = await aformat_multimodal_content(files, formatter)
             if not content_blocks:
                 msg.pop("files", None)
                 continue

lib/crewai/src/crewai/llms/base_llm.py

@@ -274,7 +274,10 @@ def _validate_init_fields(cls, data: Any) -> Any:
             data["stop"] = list(stop)
 
         if not data.get("provider"):
-            data["provider"] = "openai"
+            model = data.get("model") or ""
+            data["provider"] = (
+                cls._extract_provider(model) if isinstance(model, str) else "openai"
+            )
 
         known_fields = set(cls.model_fields.keys())
         extras = {k: v for k, v in data.items() if k not in known_fields}
@@ -507,6 +510,10 @@ def supports_multimodal(self) -> bool:
         """
         return False
 
+    def _multimodal_formatter_name(self) -> str:
+        # Content-block schema key for crewai_files. Identity stays on self.provider.
+        return self.provider or self.model
+
     def format_text_content(self, text: str) -> dict[str, Any]:
         """Format text as a content block for the LLM.
 
@@ -866,7 +873,7 @@ def _process_message_files(self, messages: list[LLMMessage]) -> list[LLMMessage]
                 )
             return messages
 
-        provider = getattr(self, "provider", None) or getattr(self, "model", "openai")
+        formatter = self._multimodal_formatter_name()
         api = getattr(self, "api", None)
 
         for msg in messages:
@@ -878,7 +885,7 @@ def _process_message_files(self, messages: list[LLMMessage]) -> list[LLMMessage]
             text = existing_content if isinstance(existing_content, str) else None
 
             content_blocks = format_multimodal_content(
-                files, provider, api=api, prefer_upload=self.prefer_upload, text=text
+                files, formatter, api=api, prefer_upload=self.prefer_upload, text=text
             )
             if not content_blocks:
                 msg.pop("files", None)

lib/crewai/src/crewai/utilities/internal_instructor.py

@@ -105,7 +105,12 @@ def _create_instructor_client(self) -> Any:
                 if value is not None:
                     extra_kwargs[attr] = value
 
-        return instructor.from_provider(f"{provider}/{model_string}", **extra_kwargs)
+        qualified_model = (
+            model_string
+            if not provider or model_string.startswith(f"{provider}/")
+            else f"{provider}/{model_string}"
+        )
+        return instructor.from_provider(qualified_model, **extra_kwargs)
 
     def _extract_provider(self) -> str:
         """Extract provider from LLM model name.

lib/crewai/tests/llms/test_multimodal_integration.py

@@ -77,8 +77,8 @@ def test_audio_bytes() -> bytes:
 
 def _build_multimodal_message(llm: LLM, prompt: str, files: dict) -> list[dict]:
     """Build a multimodal message with text and file content."""
-    provider = getattr(llm, "provider", None) or llm.model
-    content_blocks = format_multimodal_content(files, provider)
+    formatter = llm._multimodal_formatter_name()
+    content_blocks = format_multimodal_content(files, formatter)
     return [
         {
             "role": "user",

lib/crewai/tests/test_llm.py

@@ -860,16 +860,19 @@ def test_prefixed_models_with_invalid_constants_use_litellm():
     llm = LLM(model="openai/gemini-2.5-flash", is_litellm=False)
     assert llm.is_litellm is True
     assert llm.model == "openai/gemini-2.5-flash"
+    assert llm.provider == "openai"
 
     # Test openai/ prefix with model that doesn't match patterns (e.g. no gpt- prefix) → LiteLLM
     llm2 = LLM(model="openai/custom-finetune-model", is_litellm=False)
     assert llm2.is_litellm is True
     assert llm2.model == "openai/custom-finetune-model"
+    assert llm2.provider == "openai"
 
     # Test anthropic/ prefix with non-Anthropic model → LiteLLM
     llm3 = LLM(model="anthropic/gpt-4o", is_litellm=False)
     assert llm3.is_litellm is True
     assert llm3.model == "anthropic/gpt-4o"
+    assert llm3.provider == "anthropic"
 
 
 def test_prefixed_models_with_valid_patterns_use_native_sdk():
@@ -893,11 +896,38 @@ def test_prefixed_models_with_non_native_providers_use_litellm():
     llm = LLM(model="groq/llama-3.3-70b", is_litellm=False)
     assert llm.is_litellm is True
     assert llm.model == "groq/llama-3.3-70b"
+    assert llm.provider == "groq"
 
     # Test together/ prefix (not a native provider) → LiteLLM
     llm2 = LLM(model="together/qwen-2.5-72b", is_litellm=False)
     assert llm2.is_litellm is True
     assert llm2.model == "together/qwen-2.5-72b"
+    assert llm2.provider == "together"
+
+
+@pytest.mark.parametrize(
+    ("model", "expected_provider"),
+    [
+        ("groq/llama-3.3-70b", "groq"),
+        ("cohere/command-r", "cohere"),
+        ("sambanova/Meta-Llama-3.1-70B-Instruct", "sambanova"),
+        ("mistral/mistral-large", "mistral"),
+        ("vertex_ai/gemini-1.5-pro", "vertex_ai"),
+        ("openai/custom-finetune-model", "openai"),
+        ("anthropic/gpt-4o", "anthropic"),
+    ],
+)
+def test_litellm_path_preserves_provider_from_model_prefix(model, expected_provider):
+    llm = LLM(model=model, is_litellm=False)
+    assert llm.is_litellm is True
+    assert llm.provider == expected_provider
+    assert llm.model == model
+
+
+def test_litellm_keeps_provider_but_formats_multimodal_as_openai_schema():
+    llm = LLM(model="anthropic/claude-3-5-haiku-20241022", is_litellm=True)
+    assert llm.provider == "anthropic"
+    assert llm._multimodal_formatter_name() == "openai"
 
 
 def test_unprefixed_models_use_native_sdk():

lib/crewai/tests/utilities/test_converter.py

@@ -1004,3 +1004,21 @@ def test_internal_instructor_omits_unset_base_url_and_api_key() -> None:
         InternalInstructor(content="x", model=SimpleModel, llm=mock_llm)
 
         mock_from_provider.assert_called_once_with("openai/gpt-4o")
+
+
+def test_internal_instructor_does_not_double_prefix_qualified_models() -> None:
+    from crewai.utilities.internal_instructor import InternalInstructor
+
+    mock_llm = Mock()
+    mock_llm.is_litellm = False
+    mock_llm.model = "groq/llama-3.3-70b"
+    mock_llm.provider = "groq"
+    mock_llm.base_url = None
+    mock_llm.api_key = None
+
+    with patch("instructor.from_provider") as mock_from_provider:
+        mock_from_provider.return_value = Mock()
+
+        InternalInstructor(content="x", model=SimpleModel, llm=mock_llm)
+
+        mock_from_provider.assert_called_once_with("groq/llama-3.3-70b")