Question regarding to pass state keyed data to context

[zmy1116]

Hello,

I am keep seeing adk examples where in system prompt we have lines like "user xxx data from the xxx state key". I assume the intention is to pass state[xxx] to the context. However, I don't think this works.

For example, in the gemini fullstack example
https://github.com/google/adk-samples/blob/main/python/agents/gemini-fullstack/app/agent.

section_planner = LlmAgent(
    model=config.worker_model,
    name="section_planner",
    description="Breaks down the research plan into a structured markdown outline of report sections.",
    instruction="""
    You are an expert report architect. Using the research topic and the plan from the 'research_plan' state key, design a logical structure for the final report.
    Note: Ignore all the tag nanes ([MODIFIED], [NEW], [RESEARCH], [DELIVERABLE]) in the research plan.
    Your task is to create a markdown outline with 4-6 distinct sections that cover the topic comprehensively without overlap.
    You can use any markdown format you prefer, but here's a suggested structure:
    # Section Name
    A brief overview of what this section covers
    Feel free to add subsections or bullet points if needed to better organize the content.
    Make sure your outline is clear and easy to follow.
    Do not include a "References" or "Sources" section in your outline. Citations will be handled in-line.
    """,
    output_key="report_sections",
)

It says Using the research topic and the plan from the 'research_plan' state key as if there is some implemented funcitonality that allows adk to load the state['research_plan'] into context at will.

However, when I tried it myself, it does not work. And I went through the full adk code base and I did not see adk has any functionality that load state['research_plan'] into context at will based on a textual prompt.

Can anyone confirm this or point me to the adk code portion that actually allow this to happen?