fix: filter empty user message content in AnthropicCompletion

[Chase-Xuu]

Fixes #4427

Summary

The Anthropic API requires all messages to have non-empty content except for the optional final assistant message. When messages with empty string content are passed to the API, it returns a 400 error:

```
messages.0: all messages must have non-empty content except for the optional final assistant message
```

Changes

This PR adds validation in `_format_messages_for_anthropic()` to:

• Filter out messages with empty or whitespace-only content
• Allow empty content only for the final assistant message (per API spec)
• Fall back to a default "Hello" message if all messages are filtered out

Testing

Added regression tests to verify:

• Empty user messages are filtered
• Whitespace-only messages are filtered
• Mixed empty and valid messages are handled correctly
• Final empty assistant message is preserved (as allowed by API)

---

Note

Medium Risk
Touches core message pre-processing for all Anthropic calls; filtering and placeholder insertion could subtly change conversation history and tool-call flows, though covered by targeted tests.

Overview
Prevents Anthropic 400s by tightening _format_messages_for_anthropic() validation: drops empty/whitespace-only messages, only permits empty content for the final assistant message, and inserts placeholder turns when filtering would break required role alternation.

Also strips trailing whitespace from the final assistant message (Anthropic rejects it), and adds regression tests covering empty-message filtering, alternation placeholder insertion, and final-assistant whitespace/emptiness edge cases.

^{Written by Cursor Bugbot for commit f86b84d. This will update automatically on new commits. Configure here.}

[Chase-Xuu]

The Cursor Bugbot concerns have been addressed:

1. Tests asserting whitespace stripping: The implementation now includes actual .rstrip() logic for trailing whitespace in final assistant messages.

2. Empty message filtering breaking role alternation: Fixed by implementing role alternation maintenance after filtering - the code now inserts placeholder messages to maintain user/assistant alternation.

Note: This PR is related to #4430 which addresses the same issues from a different angle. You may want to consider merging one after the other to avoid conflicts.

[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.}

[cursor]

Synthetic placeholders alter conversation meaning

Medium Severity

_format_messages_for_anthropic() now injects synthetic {"content": "..."} messages whenever two adjacent roles match. These fabricated user/assistant turns change the actual conversation payload, so the model receives content that never existed in the original chat history.

 

[Chase-Xuu]

Thanks for catching this! Good point about the synthetic placeholders altering conversation meaning.

Fixed: Instead of inserting {"content": "..."} placeholders, consecutive same-role messages are now merged together. This:

• Preserves all original content
• Maintains Anthropic's role alternation requirement
• Handles both string and list content formats

Example: Two consecutive user messages "Hello" and "World" become a single "Hello\n\nWorld" message.