Link to Issue or Description of Change

Closes: #5650

Problem

Gemma4 models served via Ollama do not recognise role: "tool" in the message history. When tool results were appended using this role, the model entered an infinite loop re-calling the tool instead of generating a final response.

Two separate code paths were affected:

1. _content_to_message_param — builds tool result messages from function_response parts.
2. _ensure_tool_results — heals incomplete histories by injecting placeholder messages when a tool call has no matching result. This function had both the trigger condition and the fallback block hardcoded to role: "tool", so it neither recognised incoming tool_responses messages as valid resolutions nor generated the correct role in the placeholder it injected.

Solution

Introduced a model-aware role selector in both affected functions. When the model name contains gemma4, the role is set to "tool_responses"; otherwise it falls back to the standard "tool".

In _content_to_message_param:
tool_role = "tool_responses" if "gemma4" in model.lower() else "tool"

In _ensure_tool_results, a single expected_tool_role variable is computed once before the loop and reused consistently in three places: the healing trigger condition, the elif branch that clears resolved call IDs, and the post-loop fallback block. This eliminates the previous dual-condition pattern where the two if blocks were not mutually exclusive and could both fire on the same message.

Testing Plan

Unit Tests:

• I have added or updated unit tests for my change.
• All unit tests pass locally.

Added tests/unittests/test_lite_llm_gemma_tool_role.py.

Covers:

• Single tool response: verifies that when the model name contains gemma4, the generated message uses role: "tool_responses" instead of role: "tool".
• Multiple tool responses: verifies the same behaviour when multiple function_response parts are present in a single content object.
• Non-gemma4 models: verifies that role: "tool" is preserved for all other model names.

Full test suite result: 5709 passed, 2300 warnings.

Manual End-to-End (E2E) Tests:

Setup:

• Ollama running locally with gemma4 (e2b and e4b variants tested), qwen2.5-coder:b and qwen2.5-coder:7b
• Google ADK agent with a single tool: get_comprehensive_refugee_help
• LiteLLM as the provider bridge

Steps:

1. Send a vague message → agent asks for clarification (no tool call). ✓
2. Send a message with explicit need and city → agent calls the tool once, receives tool_responses, generates final answer in the user's language. ✓
3. Confirmed via LiteLLM debug logs that the outgoing message history contains only role: "tool_responses" for gemma4 models and "tool" for qwen models, without infinite loops.

Before the fix for gemma4 models: history contained both a phantom role: "tool" error message and the real role: "tool_responses" result, causing the model to process the error last and loop.
After the fix: history contains only the real role: "tool_responses" result.

Checklist

• I have read the CONTRIBUTING.md document.
• I have performed a self-review of my own code.
• I have commented my code, particularly in hard-to-understand areas.
• I have added tests that prove my fix is effective or that my feature works.
• New and existing unit tests pass locally with my changes.
• I have manually tested my changes end-to-end.