Gemini image model does not accept history with tool calls or thinking

[DouweM]

Initial Checks

• I confirm that I'm using the latest version of Pydantic AI
• I confirm that I searched for my issue in https://github.com/pydantic/pydantic-ai/issues before opening this issue

Description

When gemini-2.5-flash-image-preview with https://ai.pydantic.dev/builtin-tools/#image-generation-tool (see example code there) is passed message history from another model with tool calls or thinking, the API returns an error:

Tool calls:

Function calling is not enabled for models/gemini-2.5-flash-image-preview

Thinking signatures:

Request contains an invalid argument.

Thinking signatures we can simply omit (while leaving the thinking summaries in), but tool calls are a bit more tricky, as the tool call results may be relevant to image generation, and may not be repeated in text messages.

Example Code

Python, Pydantic AI & LLM client version

Latest

[towseef41]

@DouweM can i work on this?

[DouweM]

@towseef41 What are you thinking in terms of the solution to tool calls?

tool calls are a bit more tricky, as the tool call results may be relevant to image generation, and may not be repeated in text messages.

[sidg1215]

@towseef41 What are you thinking in terms of the solution to tool calls?

tool calls are a bit more tricky, as the tool call results may be relevant to image generation, and may not be repeated in text messages.

Hey @DouweM, might sound naive, but would transforming the tool calls and returns to UserPromptPart and TextPart be a valid solution (that way the tool results can still be somewhat kept as context)?

Btw, big fan of the library, keep up the good work!

[DouweM]

@sidg1215 Yeah I agree that may be the best we can do -- update GoogleModel._map_messages to represent ToolCallPart and ToolReturnPart as text messages if not self.profile.supports_tools. I'd accept a contribution that does that.