[LiteLLM] wrong tool role for Gemma4 model

[claud42]

🔴 Required Information

Describe the Bug:
When using ADK's LiteLLM with any Gemma 4 model on any LLM serving backend, the message for a tool call result should have role tool_responses (specified in the model card).

This is a known issue in #5650 . However, the bugfix only handled the situation of using Ollama as the serving backend, by checking if the model name contains string gemma4:

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

Google published Gemma 4 models in huggingface using gemma-4 as the prefix of model names. As a result, when serving with other backends (e.g., vLLM, llama.cpp, etc.), the corresponding tool_role will still be tool. This would still cause Gemma 4 misbehavior as either dead loop till OOM, or hallucinate.

Steps to Reproduce:
Please provide a numbered list of steps to reproduce the behavior:

1. Using vLLM or llama.cpp to serve gemma4 model from hugging face
2. Run any ADK agent with a tool use, using the vLLM/llama.cpp via ADK's built-in litellm as default model

Expected Behavior:
After tool use, the result message being sent to the model should have role being tool_responses and the model should response based on tool results.

Observed Behavior:
After tool use, the result message being sent to the model had role tool, and the model start produce emojis, weird symbols, or dead loops.

Environment Details:

• ADK Library Version (pip show google-adk): 2.3.0
• Desktop OS:** macOS
• Python Version (python -V): 3.13.14

Model Information:

• Are you using LiteLLM: Yes
• Which model is being used: google/gemma-4-26B-A4B

---

🟡 Optional Information

Regression:
No

Additional Context:
Relevant issue: #5650
Relevant bugfix: 3d07960

How often has this issue occurred?:
Wrong role always occurred (100%). The model often respond in unexpected low quality (50%+.)

[abhiramArise]

I can confirm this — the "gemma4" in model.lower() check from #5650 only
matches Ollama-style names (gemma4:e2b), not the HF/vLLM/llama.cpp naming
convention (google/gemma-4-26B-A4B, with a hyphen), so the wrong tool role
still gets sent on non-Ollama backends.

I'm going to work on a fix (matching both naming patterns) and will open a PR shortly.

[surajksharma07]

@claud42 Thanks for pinning down the exact HF-style model name — confirmed that the gemma-4 naming from vLLM/llama.cpp slips right past the substring check #5650 added, so this is a real gap, not a dupe.

There's already an open PR (#6340) that swaps it for a regex matching both gemma4 and gemma-4 — could you pull that branch as a workaround and confirm it clears up the role/looping behavior on your setup?

Once you've tested it end-to-end and it's holding up, that'll help get it merged faster!

[claud42]

@surajksharma07 I've tested the PR, and it fixed the issue as expected at my end.