Claude 4.6 models fail: consecutive assistant messages treated as prefill

[shaul-saitowitz-pango]

Description

Claude 4.6 models (Sonnet 4.6, Opus 4.6) return 400 Bad Request errors when used with CrewAI agents. Anthropic made a breaking API change: Claude 4.6 no longer supports assistant message prefilling, and any request with a trailing assistant message is rejected.

Root Cause

CrewAgentExecutor._append_message() appends every LLM response as role: "assistant" (crew_agent_executor.py#L273):

self._append_message(formatted_answer.text)  # role defaults to "assistant"

This builds conversations with consecutive assistant messages:

user: [system prompt + task]
assistant: [LLM response with tool call]
assistant: [tool result + observation]   ← consecutive assistant
assistant: [next iteration]              ← consecutive assistant

Claude 4.6 interprets the trailing assistant message as a prefill attempt and rejects the request. Earlier Claude models (3.5, 4.5) tolerated this, but 4.6 is strict.

Steps to Reproduce

1. Create a CrewAI agent with llm=LLM(model="anthropic/claude-sonnet-4-6-20250514")
2. Give it a task that requires tool use (multiple iterations)
3. The agent fails on the second LLM call with:

   400 Bad Request: Prefilling assistant messages is no longer supported
   

Expected Behavior

CrewAI should normalize the message history before sending to the LLM — either by:

• Merging consecutive same-role messages
• Inserting a user message between consecutive assistant messages
• Using _format_model_specific_messages() to handle this (it already has Mistral/Ollama handling but not Anthropic)

Environment

• CrewAI version: 1.6.1 (also reproduced conceptually against 1.10.1 source)
• Python: 3.13
• LLM provider: OpenRouter (LiteLLM path)
• Models affected: anthropic/claude-sonnet-4-6-20250514, anthropic/claude-opus-4-6-20250514

Related Issues

• [BUG] Error with alternating messages structure in instruct mode #1454 (alternating messages in instruct mode — fixed via LiteLLM ensure_alternating_roles, but that param is no longer available)
• fix: ensure proper message formatting for Anthropic models #2063 (Anthropic message formatting — fixed first-message ordering but not consecutive assistant messages)
• [FEATURE] Enable CrewAI agents to use Claude's extended thinking #3964 (extended thinking — fixed thinking block preservation)
• External: livekit/agents#4907, strands-agents/sdk-python#1694

Suggested Fix

Add Anthropic handling to _format_model_specific_messages() in llm.py (around line 1508) to merge consecutive assistant messages, similar to the existing Mistral/Ollama handling:

# Handle Anthropic models — merge consecutive same-role messages
if self.is_anthropic:
    merged = [messages[0]]
    for msg in messages[1:]:
        if msg["role"] == merged[-1]["role"]:
            merged[-1]["content"] += "\n\n" + str(msg["content"])
        else:
            merged.append(msg)
    messages = merged

Workaround

Using a LiteLLM input callback to merge consecutive messages before they reach the API:

import litellm

def merge_consecutive_messages(kwargs, **_):
    messages = kwargs.get("messages", [])
    if len(messages) < 2:
        return
    merged = [messages[0]]
    for msg in messages[1:]:
        if msg["role"] == merged[-1]["role"]:
            merged[-1]["content"] += "\n\n" + str(msg["content"])
        else:
            merged.append(msg)
    kwargs["messages"] = merged

litellm.input_callback = [merge_consecutive_messages]

[ehansis]

Note: The line number above refers to some previous version of the CrewAI package (before the split into native/react tool calls?).

In the current version, this problem still exists, but only for the react tool calling pattern (_invoke_loop_react).

In the native tool calling pattern (_invoke_loop_native_tools) the tool results are appended with role tool:

crewAI/lib/crewai/src/crewai/agents/crew_agent_executor.py

Lines 1084 to 1089 in 542afe6

	tool_message: LLMMessage = {
	"role": "tool",
	"tool_call_id": call_id,
	"name": func_name,
	"content": result,
	}

An accompanying reasoning prompt is appended with role "user"

crewAI/lib/crewai/src/crewai/agents/crew_agent_executor.py

Lines 789 to 793 in 542afe6

	reasoning_message: LLMMessage = {
	"role": "user",
	"content": reasoning_prompt,
	}
	self.messages.append(reasoning_message)

crewAI/lib/crewai/src/crewai/agents/crew_agent_executor.py

Lines 812 to 816 in 542afe6

	reasoning_prompt = self._i18n.slice("post_tool_reasoning")
	reasoning_message = {
	"role": "user",
	"content": reasoning_prompt,
	}

So it looks like the issue should not occur here.