claude models fail all parallel reads after an image read through antigravity provider gemini do not

[Whamp]

What happened?

Any time Claude models through the Antigravity provider send a batch of parallel reads that involve an image read, all reads after the image read fail with:

  tool_use ids were found without tool_result blocks immediately after

Steps to reproduce

1. Use a Claude model via Antigravity/Cloud Code Assist
2. Trigger parallel tool calls where 2+ results contain images (e.g. reading multiple image files)
3. Can also trigger by telling clanker to read text file, then an image then a different text file.

Expected behavior

parallel reads work like through anthropic provider

Version

0.57.1

pkg:ai

[badlogic]

Fixed.

Actual issue was not "parallel reads" in general. It was our Google shared tool-result conversion splitting image results into a synthetic extra user message.

That breaks Antigravity Claude because all tool_result blocks for a tool_use batch have to be in the immediate next message.

What changed:

• keep image tool results inline in functionResponse.parts for:
  • Gemini 3+ models
  • non-Gemini models like Claude on Antigravity
• keep the old synthetic user image message only for Gemini < 3, because those models still reject multimodal functionResponse.parts

I tested this against live endpoints:

• Antigravity Claude: fixed
• Google API Gemini 3+: inline works
• Google API Gemini 2.5: still needs the old split behavior
• Gemini CLI 2.5 Flash: also still needs the old split behavior

Commit:

• 453b2239 fix(ai): keep image tool results inline for gemini 3+ and antigravity closes #2052

[Whamp]

Thank you! Sorry for the poor description. Very excited to have this bug fixed!