fix: merge consecutive same-role messages for Anthropic Claude 4.6+

Claude 4.6+ rejects requests containing consecutive assistant messages,
treating the trailing one as an unsupported prefill. This occurs when
CrewAgentExecutor appends multiple assistant messages during tool-use
iterations in the ReAct loop.

Added _merge_consecutive_messages() static method to BaseLLM that
merges consecutive same-role messages by joining their text content
with double newlines. Messages with list content (tool-use blocks,
multimodal) are left untouched.

The merge is called in both:
- LLM._format_messages_for_provider() for the LiteLLM path
- AnthropicCompletion._format_messages_for_anthropic() for native path

Added 7 unit tests covering:
- Consecutive assistant message merging (LiteLLM and native paths)
- Alternating messages preserved
- List content not merged
- Original messages not mutated
- Non-Anthropic models unaffected
- Consecutive user messages also merged

Fixes #4798

Co-Authored-By: João <joao@crewai.com>

lib/crewai/src/crewai/llm.py

@@ -2113,6 +2113,12 @@ class LLM(BaseLLM):
         if not self.is_anthropic:
             return messages  # type: ignore[return-value]
 
+        # Merge consecutive same-role messages to avoid Anthropic API errors.
+        # Claude 4.6+ rejects consecutive assistant messages (treated as prefill).
+        # This can happen when CrewAgentExecutor appends multiple assistant messages
+        # during tool-use iterations.
+        messages = self._merge_consecutive_messages(messages)
+
         # Anthropic requires messages to start with 'user' role
         if not messages or messages[0]["role"] == "system":
             # If first message is system or empty, add a placeholder user message

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

@@ -581,6 +581,43 @@ class BaseLLM(ABC):
 
             return None
 
+    @staticmethod
+    def _merge_consecutive_messages(
+        messages: list[LLMMessage],
+    ) -> list[LLMMessage]:
+        """Merge consecutive messages that share the same role.
+
+        Anthropic (Claude 4.6+) rejects requests containing consecutive
+        assistant messages, treating the trailing one as an unsupported
+        prefill. This helper collapses runs of same-role messages into a
+        single message by joining their text content with double newlines.
+
+        Messages whose content is already a list (e.g. tool-use blocks or
+        multimodal content) are left untouched — only plain-string content
+        is merged.
+
+        Args:
+            messages: The message list to process.
+
+        Returns:
+            A new list with consecutive same-role messages merged.
+        """
+        if not messages:
+            return messages
+
+        merged: list[LLMMessage] = [messages[0].copy()]  # type: ignore[union-attr]
+        for msg in messages[1:]:
+            prev = merged[-1]
+            if (
+                msg["role"] == prev["role"]
+                and isinstance(msg.get("content"), str)
+                and isinstance(prev.get("content"), str)
+            ):
+                prev["content"] = prev["content"] + "\n\n" + msg["content"]
+            else:
+                merged.append(msg.copy())  # type: ignore[union-attr]
+        return merged
+
     def _format_messages(self, messages: str | list[LLMMessage]) -> list[LLMMessage]:
         """Convert messages to standard format.
 

lib/crewai/src/crewai/llms/providers/anthropic/completion.py

@@ -623,6 +623,12 @@ class AnthropicCompletion(BaseLLM):
         if pending_tool_results:
             formatted_messages.append({"role": "user", "content": pending_tool_results})
 
+        # Merge consecutive same-role messages to avoid Anthropic API errors.
+        # Claude 4.6+ rejects consecutive assistant messages (treated as prefill).
+        # This can happen when CrewAgentExecutor appends multiple assistant
+        # messages during tool-use iterations in the ReAct loop.
+        formatted_messages = self._merge_consecutive_messages(formatted_messages)
+
         # Ensure first message is from user (Anthropic requirement)
         if not formatted_messages:
             # If no messages, add a default user message

lib/crewai/tests/test_llm.py

@@ -481,6 +481,150 @@ def test_anthropic_message_formatting(anthropic_llm, system_message, user_messag
         anthropic_llm._format_messages_for_anthropic([{"invalid": "message"}])
 
 
+def test_anthropic_merges_consecutive_assistant_messages_litellm():
+    """Test that consecutive assistant messages are merged for Anthropic models via LiteLLM.
+
+    Claude 4.6+ rejects consecutive assistant messages (treated as prefill).
+    This test ensures that _format_messages_for_provider merges them.
+    Reproduces the scenario described in issue #4798.
+    """
+    llm = LLM(model="anthropic/claude-3-sonnet", is_litellm=True)
+    messages = [
+        {"role": "user", "content": "Do something"},
+        {"role": "assistant", "content": "I will use a tool"},
+        {"role": "assistant", "content": "Observation: tool result"},
+        {"role": "assistant", "content": "Now I have the answer"},
+    ]
+
+    formatted = llm._format_messages_for_provider(messages)
+
+    # All three consecutive assistant messages should be merged into one
+    assert len(formatted) == 2
+    assert formatted[0]["role"] == "user"
+    assert formatted[0]["content"] == "Do something"
+    assert formatted[1]["role"] == "assistant"
+    assert "I will use a tool" in formatted[1]["content"]
+    assert "Observation: tool result" in formatted[1]["content"]
+    assert "Now I have the answer" in formatted[1]["content"]
+
+
+def test_anthropic_merges_consecutive_assistant_messages_native():
+    """Test that consecutive assistant messages are merged for native Anthropic provider.
+
+    Claude 4.6+ rejects consecutive assistant messages (treated as prefill).
+    This test ensures that _format_messages_for_anthropic merges them.
+    Reproduces the scenario described in issue #4798.
+    """
+    llm = AnthropicCompletion(model="claude-3-sonnet", is_litellm=False)
+    messages = [
+        {"role": "user", "content": "Do something"},
+        {"role": "assistant", "content": "I will use a tool"},
+        {"role": "assistant", "content": "Observation: tool result"},
+        {"role": "assistant", "content": "Now I have the answer"},
+    ]
+
+    formatted, system_msg = llm._format_messages_for_anthropic(messages)
+
+    # All three consecutive assistant messages should be merged into one
+    assert len(formatted) == 2
+    assert formatted[0]["role"] == "user"
+    assert formatted[0]["content"] == "Do something"
+    assert formatted[1]["role"] == "assistant"
+    assert "I will use a tool" in formatted[1]["content"]
+    assert "Observation: tool result" in formatted[1]["content"]
+    assert "Now I have the answer" in formatted[1]["content"]
+    assert system_msg is None
+
+
+def test_merge_consecutive_messages_preserves_alternating():
+    """Test that already-alternating messages are not modified."""
+    llm = LLM(model="anthropic/claude-3-sonnet", is_litellm=True)
+    messages = [
+        {"role": "user", "content": "Hello"},
+        {"role": "assistant", "content": "Hi there"},
+        {"role": "user", "content": "How are you?"},
+        {"role": "assistant", "content": "Great!"},
+    ]
+
+    formatted = llm._format_messages_for_provider(messages)
+
+    assert len(formatted) == 4
+    assert formatted[0]["role"] == "user"
+    assert formatted[1]["role"] == "assistant"
+    assert formatted[2]["role"] == "user"
+    assert formatted[3]["role"] == "assistant"
+
+
+def test_merge_consecutive_messages_skips_list_content():
+    """Test that messages with list content (e.g. tool-use blocks) are not merged."""
+    llm = LLM(model="anthropic/claude-3-sonnet", is_litellm=True)
+    messages = [
+        {"role": "user", "content": "Hello"},
+        {"role": "assistant", "content": [{"type": "tool_use", "id": "1", "name": "test", "input": {}}]},
+        {"role": "assistant", "content": "text after tool use"},
+    ]
+
+    formatted = llm._format_messages_for_provider(messages)
+
+    # The list-content assistant message cannot be merged, so we get 3 messages
+    assert len(formatted) == 3
+    assert formatted[1]["role"] == "assistant"
+    assert isinstance(formatted[1]["content"], list)
+    assert formatted[2]["role"] == "assistant"
+    assert formatted[2]["content"] == "text after tool use"
+
+
+def test_merge_consecutive_messages_does_not_mutate_original():
+    """Test that merging does not mutate the original message list."""
+    llm = LLM(model="anthropic/claude-3-sonnet", is_litellm=True)
+    messages = [
+        {"role": "user", "content": "Hello"},
+        {"role": "assistant", "content": "Response 1"},
+        {"role": "assistant", "content": "Response 2"},
+    ]
+    original_len = len(messages)
+    original_content = messages[1]["content"]
+
+    llm._format_messages_for_provider(messages)
+
+    # Original messages should not be modified
+    assert len(messages) == original_len
+    assert messages[1]["content"] == original_content
+
+
+def test_merge_consecutive_messages_non_anthropic_unchanged():
+    """Test that non-Anthropic models do not merge consecutive messages."""
+    llm = LLM(model="gpt-4", is_litellm=True)
+    messages = [
+        {"role": "user", "content": "Hello"},
+        {"role": "assistant", "content": "Response 1"},
+        {"role": "assistant", "content": "Response 2"},
+    ]
+
+    formatted = llm._format_messages_for_provider(messages)
+
+    # Non-Anthropic models should not merge
+    assert len(formatted) == 3
+
+
+def test_merge_consecutive_user_messages():
+    """Test that consecutive user messages are also merged for Anthropic."""
+    llm = LLM(model="anthropic/claude-3-sonnet", is_litellm=True)
+    messages = [
+        {"role": "user", "content": "Part 1"},
+        {"role": "user", "content": "Part 2"},
+        {"role": "assistant", "content": "Response"},
+    ]
+
+    formatted = llm._format_messages_for_provider(messages)
+
+    assert len(formatted) == 2
+    assert formatted[0]["role"] == "user"
+    assert "Part 1" in formatted[0]["content"]
+    assert "Part 2" in formatted[0]["content"]
+    assert formatted[1]["role"] == "assistant"
+
+
 def test_deepseek_r1_with_open_router():
     if not os.getenv("OPEN_ROUTER_API_KEY"):
         pytest.skip("OPEN_ROUTER_API_KEY not set; skipping test.")