All contributors have signed the CLA. Thank you! ✅
_{Posted by the CLA Assistant Lite bot.}

Greptile Summary

This PR fixes a critical bug in the Antigravity message converter that caused API errors when using Claude extended thinking models through the proxy.

Key Changes:

• Added thought?: boolean property to GeminiPart interface to support Gemini's thinking block format
• Extended convertContentToParts() to handle thinking and redacted_thinking content types from Anthropic responses
• Preserves thinking blocks by converting them to Gemini format with thought: true flag

Root Cause Fixed:
Previously, when assistant messages contained thinking blocks from Claude responses (using models like claude-opus-4-5-thinking), these blocks were silently dropped during OpenAI→Gemini conversion. This caused Anthropic API to reject requests with: "Expected thinking or redacted_thinking, but found text. When thinking is enabled, a final assistant message must start with a thinking block."

The fix properly extracts thinking content from part.thinking (Anthropic format) with fallback to part.text, then converts to Gemini's thought format.

Confidence Score: 5/5

• This PR is safe to merge with minimal risk - it fixes a critical bug with a focused, well-scoped change
• The change is minimal (3 lines of interface modification + 5 lines of logic), addresses a specific API compatibility issue, follows existing patterns in the codebase (similar handling exists in src/tools/session-manager/utils.ts and src/auth/antigravity/thinking.ts), and includes proper fallback logic. The fix directly resolves the reported error without introducing side effects.
• No files require special attention

Important Files Changed

Sequence Diagram

sequenceDiagram
    participant Client
    participant Antigravity
    participant MessageConverter
    participant AnthropicAPI

    Client->>Antigravity: POST /chat (OpenAI format)
    Note over Client: model: claude-opus-4-5-thinking<br/>messages with thinking blocks
    
    Antigravity->>MessageConverter: convertOpenAIToGemini(messages)
    
    Note over MessageConverter: Process each message
    
    loop For each message
        alt Message role: assistant
            MessageConverter->>MessageConverter: convertContentToParts(content)
            
            loop For each content part
                alt part.type === "text"
                    MessageConverter->>MessageConverter: Add {text: part.text}
                else part.type === "thinking" or "redacted_thinking"
                    MessageConverter->>MessageConverter: Add {thought: true, text: part.thinking || part.text}
                else part.type === "image_url"
                    MessageConverter->>MessageConverter: Add {inlineData: {...}}
                end
            end
        end
    end
    
    MessageConverter-->>Antigravity: Gemini contents format
    Note over Antigravity: contents with thought parts
    
    Antigravity->>AnthropicAPI: POST /messages (Anthropic format)
    Note over AnthropicAPI: Validates thinking blocks present<br/>in assistant message
    
    AnthropicAPI-->>Antigravity: Success response
    Antigravity-->>Client: OpenAI-compatible response

Greptile found no issues!

From now on, if a review finishes and we haven't found any issues, we will not post anything, but you can confirm that we reviewed your changes in the status check section.

_{This feature can be toggled off in your Code Review Settings by deselecting "Create a status check for each PR".}

I have read the CLA Document and I hereby sign the CLA

Closing in favor of a more comprehensive fix that addresses the complete signature flow issue.