Summary

Severity: High
Area: Gemini adapter
Affected files: agentkit/llm/gemini.py:104-118, :172-184
Suggested branch: fix/gemini-parallel-function-calls
Suggested PR title: fix(gemini): group parallel function responses into one turn

This is finding AUDIT-05 of 18 from a consolidated audit of agentkit/. It is deliberately self-contained so it can be fixed on its own branch with its own PR.

Related findings: Related: AUDIT-04 (signature persistence), AUDIT-06 (thought parts).

AUDIT-05 — High — Gemini parallel function calls are broken

File — agentkit/llm/gemini.py:104-118 and :172-184

Current code — agentkit/llm/gemini.py:104-118

            elif msg.role == Role.TOOL:
                tool_name = (
                    tool_id_to_name.get(msg.tool_call_id or "") or msg.tool_call_id or "tool"
                )
                contents.append(
                    types.Content(
                        role="user",
                        parts=[
                            types.Part.from_function_response(
                                name=tool_name,
                                response={"result": msg.content},
                            )
                        ],
                    )
                )

Current code — agentkit/llm/gemini.py:172-184

                        if getattr(part, "function_call", None):
                            fc = part.function_call
                            raw_sig = getattr(part, "thought_signature", None)
                            sig = raw_sig if isinstance(raw_sig, bytes) else None
                            tool_calls.append(
                                ToolCall(
                                    # Full UUID: tool_call_id is the join key between an
                                    # assistant tool call and its tool result, so a truncated
                                    # 32-bit id risks collisions within a long session.
                                    id=str(uuid.uuid4()),
                                    name=fc.name,
                                    arguments=dict(fc.args or {}),
                                    thought_signature=sig,
                                )
                            )

Why this is a bug

1. Every tool result becomes its own Content(role="user") with a single function_response part. Gemini validates the call turn: "Please ensure that the number of function response parts is equal to the number of function call parts of the function call turn" → parallel calls produce 400 INVALID_ARGUMENT.
2. FunctionResponse.id is never set, and the provider's own FunctionCall.id (present in google-genai 2.25: FunctionCall.id, FunctionResponse.id) is discarded in favour of a fresh UUID. Two parallel calls to the same tool therefore cannot be correlated to their results.
3. The provider id is thrown away, so any future provider-side correlation is impossible.

Reproduced

provider parts: [function_call(name="calculator", id="fc1"), function_call(name="calculator", id="fc2")]
adapter out   : ToolCall(id=<uuid4>), ToolCall(id=<uuid4>)          # 'fc1'/'fc2' discarded
outbound shape: [Content(role=user, 1 part), Content(role=user, 1 part)]  # 2 separate user contents
function_response.id = None

Pass cases

• All tool results for one assistant turn are sent in one Content containing one function_response part per call.
• FunctionResponse.id is set to the id of the FunctionCall it answers, and FunctionCall.id (the provider's) is preserved end-to-end (not replaced by a random UUID).
• A test with two parallel calls to the same tool asserts the responses are emitted in call order and carry the matching ids.
• A test asserts the outbound request contains exactly as many function_response parts as the assistant turn had function_call parts.
• Single-tool-call turns keep working unchanged (no regression), and the existing thought-signature forwarding test still passes.

Definition of done (per AGENTS.md)

• New/updated unit and integration tests covering the pass cases above, all passing (pytest).
• ruff check ., ruff format --check . and mypy agentkit clean.
• Docstrings updated where behaviour changed; docs/PROGRESS.md updated (status, PR link, commit count, notes).
• A new ADR in docs/DECISIONS.md if an architectural decision was made.
• Branch pushed and PR opened (never commit directly to main).

Baseline at the time of writing: pytest is green on main (all tests pass; tests/real/* skipped unless AGENTKIT_REAL_INFRA=1).