Google ADK Bug Report: before_model_callback LlmResponse Not Guaranteed to Skip LLM Call

Required Information

Describe the Bug:
When returning an LlmResponse from before_model_callback to skip the LLM call and return a deterministic response, the behavior is inconsistent. The LLM call is sometimes still made, resulting in non-deterministic output instead of the pre-generated content.

Additionally, the callback execution flow does not guarantee that returning LlmResponse will always prevent the LLM from being invoked - especially when other conditions in the callback may have already modified llm_request.config.tool_config.

Steps to Reproduce:

1. Create an agent with before_model_callback that tracks state across multiple tool calls
2. Implement a multi-phase workflow:
   • Phase 1: Load artifacts (images)
   • Phase 2: Process artifacts with save_datapoints tool
   • Phase 3: First save_final_corrected_datapoints call(we have multiple images from which different sections needs to be extracted and saved, so we have this extra final tool to give the final correct fields)
   • Phase 4: Self-reflection with load_artifacts for latest revision images ( for self reflecting to correct any errors, we pass latest revision of the image again)
   • Phase 5: Post-reflection save_final_corrected_datapoints
   • Phase 6: Return pre-generated validation tables via LlmResponse
3. In before_model_callback, when post_reflection_saved=True, return:

   from google.adk.models import LlmResponse
   from google.genai.types import Content, Part
   
   return LlmResponse(
       content=Content(
           role="model",
           parts=[Part.from_text(text=validation_tables)]
       ),
       turn_complete=True
   )

4. Observe that sometimes the LLM is still called and generates a free-form response instead of the deterministic tables and also in 99% of such cases, the second load_artifacts also gets failed. But I have found a case where even the second load_artifacts happend(in self-reflection) and stilll the final output was llm response generated by llm instead of the tables. In all these cases I could see the state for the tables correctly populated. Will there be any race conditions within the state read/writes?

Expected Behavior:
When before_model_callback returns an LlmResponse object with turn_complete=True, the LLM should never be called, and the response should be exactly what was returned from the callback.

Observed Behavior:

• Intermittently, the LLM is still invoked even after returning LlmResponse
• The agent produces free-form LLM-generated text instead of the pre-generated validation tables
• This breaks determinism in workflows that require strict output formatting

Environment Details:

• ADK Library Version: 1.22.1
• Desktop OS: macOS (Darwin)
• Python Version: 3.12.10

Model Information:

• Are you using LiteLLM: Yes
• Which model is being used: azure/gpt-5.1-mini (Azure OpenAI)