Add native Google Gemini provider support

[jul-sh]

Summary

Adds native Google Gemini API support using pydantic-ai's built-in google-gla: provider prefix.

Changes

• Added google to pydantic-ai-slim optional dependencies
• Updated .env.example with GEMINI_API_KEY

No custom provider needed - pydantic-ai's native google-gla: prefix handles everything!

Usage

export GEMINI_API_KEY="your-key"

# Run with Gemini as target
uv run prompt-siren run attack +dataset=agentdojo-workspace +attack=mini-goat agent.config.model=google-gla:gemini-3-pro-preview

# Can also set attacker model
attack.config.attacker_model=google-gla:gemini-3-pro-preview

Features Supported

• Native Gemini API with full thinking/reasoning support
• Thought signatures for multi-turn conversations
• All Gemini models (gemini-3-pro-preview, gemini-2.5-flash, etc.)

[evtimovi]

Thank you so much @jul-sh ! This looks very reasonable to me! I'll wait on the CI signals just in case but I think we are okay to merge this next week.

[evtimovi]

Looks good to me! Thank you for your contribution!

[evtimovi]

Actually, @jul-sh can I ask you to also resolve the conflicts in toml and regenerate the uv.lock in a new commit? I merged Edoardo's pull request making dependencies optional and I don't think we can manually resolve the uv lock differences.

[evtimovi]

Sorry - updating this with one more: can you put the Llama API back in and also address the comment about the conflicts?

[jul-sh]

resolved the conflicts. I'm happy to add the llama API key; but would note that this PR isn't removing it (main branch doesn't have it). But happy to add it.

[evtimovi]

@jul-sh Thank you so much! I am sorry - maybe I am misreading the UI but I was referring to this commit? 4736263