fix: strip trailing whitespace from final assistant message for Anthropic

[Chase-Xuu]

Summary

Fixes #4413

Problem

When using LLM.call with an Anthropic model, if the final assistant message ends with trailing whitespace, the call fails with a 400 BadRequestError:

BadRequestError: Error code: 400 - {'type': 'error', 'error': {'type': 'invalid_request_error', 'message': 'messages: final assistant content cannot end with trailing whitespace'}, 'request_id': '...'}

Solution

Added a check in _format_messages_for_anthropic() method in the AnthropicCompletion provider to strip trailing whitespace from the final assistant message content before sending to the API.

The fix:

• Only applies to the final assistant message (not intermediate ones)
• Only affects string content (not list/structured content)
• Uses rstrip() to remove all trailing whitespace characters (spaces, tabs, newlines)

Changes

• lib/crewai/src/crewai/llms/providers/anthropic/completion.py: Added whitespace stripping logic

Testing

Added tests in lib/crewai/tests/llms/anthropic/test_anthropic.py (from related PR #4427) to verify:

• Trailing whitespace is stripped from final assistant messages
• Multiple whitespace characters (spaces, tabs, newlines) are handled
• Non-final assistant messages preserve their whitespace
• User messages are not affected

Reproduction

from crewai import LLM
import os

llm = LLM(
    model="anthropic/claude-3-haiku-20240307",
    api_key=os.getenv('ANTHROPIC_API_KEY'),
    max_tokens=50
)

messages = [
    {"role": "user", "content": "Hello. Say world"},
    {"role": "assistant", "content": "Say: "}  # trailing space triggers the error
]

result = llm.call(messages)  # Now works without error
print(result)

---

Note

Medium Risk
Moderate risk because it changes Anthropic message formatting by filtering/merging messages, which can subtly alter conversation history and tool-call sequencing, though the behavior is constrained and covered by new tests.

Overview
Fixes Anthropic API 400s by hardening _format_messages_for_anthropic() to drop empty/whitespace-only messages (except an allowed empty final assistant message) and then merge consecutive same-role messages to preserve required role alternation.

Also strips trailing whitespace from the final assistant message when its content is a string, and adds unit tests covering empty-message filtering, role-merge behavior, and whitespace stripping/preservation cases.

^{Written by Cursor Bugbot for commit 492ec2c. This will update automatically on new commits. Configure here.}

[Chase-Xuu]

The Cursor Bugbot concerns have been addressed in the latest commits:

1. Role alternation fix conflicts with test expectations: Fixed in commit c6e50f3d - updated test expectations to account for the role alternation placeholder messages.

2. Filtering empty messages breaks alternation: Fixed in commit 9f40c3fc - the code now maintains proper user/assistant alternation by inserting minimal placeholder messages when consecutive same-role messages would result.

The test suite passes with the updated expectations.

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{Bugbot Autofix is OFF. To automatically fix reported issues with Cloud Agents, enable Autofix in the Cursor dashboard.}