[mosafariuk]

💡 Motivation and Context

Gemini returns an id on function calls (FunctionCall.id) and on function responses (FunctionResponse.id). The Python SDK drops both from captured generations:

• $ai_output_choices: the non-streaming formatter (format_gemini_response), the per-chunk extractor and the streaming assembler (format_gemini_streaming_output) all build the function block without the ID.
• $ai_input: _format_part normalizes function_call and function_response parts without their IDs.

So a Gemini tool result can't be matched to the call that produced it, unlike OpenAI and Anthropic captures, which keep tool-call IDs.

This is the Python counterpart of PostHog/posthog-js#5087 (output IDs) and the ID handling in its Gemini input normalization. It uses the same shape: id on the function block, and tool_use_id on a function result. The ID is added only when Gemini returned one as a string, so calls without an ID keep their current shape.

💚 How did you test it?

• Regression tests in posthog/test/ai/gemini/test_gemini.py, built from real google.genai.types objects:
  • test_function_call_ids_in_output_choices: parametrized over streaming and non-streaming, and over an ID versus no ID. The cases with an ID fail on main and pass with this change. The no-ID cases pass either way, which shows the existing shape is unchanged.
  • test_function_call_ids_in_input_history: a function_call and a function_response with the same ID in contents. The test asserts id and tool_use_id in $ai_input. It fails on main and passes with this change.
• Existing Gemini tests (103 passed) that use MagicMock parts pass unchanged. A mocked part answers any attribute, so the ID is only kept when it's a real str.
• Full suite: 5830 passed. The only 3 failures are the fork-based tests in test_metrics_autocapture.py, which also fail for me on unmodified main on macOS and are unrelated. ruff format --check, ruff check, mypy | mypy-baseline filter and make public_api_check are clean (no public API change).
• End to end: I patched Models.generate_content and generate_content_stream to return a real GenerateContentResponse containing FunctionCall(id="call_649034", ...). On main, both paths capture the call without its ID. With this change, both include "id": "call_649034". The async client uses the same converter functions.

📝 Checklist

• I reviewed the submitted code.
• I added tests to verify the changes.
• I updated the docs if needed.
• No breaking change or entry added to the changelog.

If releasing new changes

• Ran sampo add to generate a changeset file

🤖 Agent context

Autonomy: Human-driven (agent-assisted)

• An AI coding agent helped me find this by comparing recent @posthog/ai fixes with the Python SDK, then drafted the fix and tests. I reviewed and ran all of it.
• I kept the posthog-js #5087 shape (id on calls, tool_use_id on results). I typed the new blocks as Dict[str, Any] instead of importing FormattedFunctionCall into the converter, so the public API snapshot doesn't change.

[marandaneto]

The code looks good; no qualifying findings. Verdict: correct.

[marandaneto]

thanks @mosafariuk