User prompt ↓ build_df_schema(df) → schema_block ↓ LLM_Synthesis.evaluate_and_generate_code( schema_block, picture, research, prompt ) ↓ returns decision.python_code (a plain string of Python) ↓ slim_app2 stores it in router_response ↓ CodeSandbox.execute_pandas_code(router_response, df, ticker) ↓ exec(router_response) inside the sandbox ↓ compact SandboxOutputSchema JSON + optional chart + optional parquet ↓ final synthesis LLM narrates only that compact result
