PlanReActPlanner often fails to generate a final answer in the second turn of a conversation.

[ChrisCodeNation]

Describe the bug
When using PlanReActPlanner, the first round of question and answer works as expected, with the final answer being generated after the /FINAL_ANSWER/ tag. However, in the subsequent turn of the conversation, the planner frequently only outputs the /FINAL_ANSWER/ tag without the accompanying final response.

Upon inspecting the source code, it appears that the final answer from the first turn is removed from the conversation history before being passed to the LLM in the second turn. This likely causes the LLM to imitate the pattern of the previous turn, resulting in the output of only the /FINAL_ANSWER/ tag without the actual answer.

To Reproduce

root_agent = LlmAgent(
    model=LiteLlm(
        model=f"hosted_vllm/{MODEL_NAME}",
        api_base=BASE_URL,
        # Pass authentication headers if needed
        extra_headers={"Authorization": f"Bearer {BEARER_TOEKN}"},
        temperature=0.3,
        top_k=50,
        top_p=0.95,
        chat_template_kwargs={
            "enable_thinking": False
            }
    ),
    name="你的智慧好夥伴",
    description=ROLE,
    static_instruction=SYSTEM_PROMPT5,
    planner=PlanReActPlanner(), 
    tools=[mcp_tools] # mcp_tools only includes a tool that retrieve documents from milvus.
)

And Run adk web --port 8888

Expected behavior
Generate the fully answer to the question after /FINAL_ANSWER/ tag .

Screenshots
First pic shows that the conversation ends only with /FINAL_ANSWER/

Second pic shows that the final answer from the first turn is removed from the conversation history before being passed to the LLM in the second turn

Third pic shows that LLM log did generate the final answer after /FINAL_ANSWER/ tag in the first turn

Desktop (please complete the following information):

• OS: Linux
• Python version(python -V):3.14.0
• ADK version(pip show google-adk):1.18.0

Model Information:

• Are you using LiteLLM: Yes
• Which model is being used: Qwen3-30B-A3B-Instruct-2ㄌ507-FP8

Additional context

• This behavior happens every time, even I try other model.

[llalitkumarrr]

Hello @ChrisCodeNation,

We tried to reproduce your issue with Google search as tool and it worked as expected.

Code:

from google.adk.agents.llm_agent import Agent
import os
from dotenv import load_dotenv
from google.adk.models.lite_llm import LiteLlm
from google.adk.tools import google_search
from google.adk.planners import PlanReActPlanner


# Load variables from .env file
load_dotenv()
api = os.environ.get("GEMINI_API_KEY")
api_openai = os.environ.get("openai_key_personal")

os.environ["api_key"] = api
os.environ["OPENAI_API_KEY"] = api_openai

root_agent = Agent(
    model='gemini-2.5-flash',
    name='root_agent',
    instruction="You are an expert researcher. You always stick to the facts.",
    description="Agent to answer questions using Google Search.",
    tools=[google_search],
    planner=PlanReActPlanner()

)

Here are some pictures of response:

In these screenshots if you look closely you can see that the conversation history from previous turns is maintained and used in the following turns. The final response is also generated correctly.
However it is important to note that we reproduced this behavior using gemini-2.5-flash along with Google Search tools.
Could you please try running your setup with Gemini and see if you get the expected behavior? You might also want to experiment with different tools, to pin point the issue.