⚙️ Process System

async def execute_process_for_runtime(
    runtime, *,
    process_code: str,
    input_text: str,
    user_db_id: Optional[int],
    chat_db_id: Optional[int],
    rag_context: Optional[str] = None,
    run_id: Optional[str] = None,
    model_tier: Optional[str] = None,  # "free" | "paid"
) -> ProcessExecutionResult:

    # 1. Определяем лимиты по tier
    if model_tier == "paid":
        max_tokens = 4096  # paid tier
    else:
        max_tokens = 512   # free tier

    # 2. Загружаем процесс из DB
    process = await runtime.db.get_process_by_code(process_code)

    # 3. Строим контекст (RAG, personal data, history)
    if not rag_context:
        if process.metadata.get("rag"):
            rag_context = await build_rag_context(query=input_text)
        if process.metadata.get("personal_data"):
            docs = await db.list_table_documents(user_id=user_db_id)
            rag_context += format_personal_docs(docs)

    # 4. Выбираем агента (если есть)
    agent_sel = await select_agent(db, process_code=process_code)
    if agent_sel:
        system_prompt = agent_sel["assembled_system_prompt"]

    # 5. Рендерим prompt
    rendered_prompt = render_process_prompt(
        process_code=process_code,
        prompt_template=process["prompt_template"],
        input_text=input_text,
        rag_context=rag_context,
    )

    # 6. Выбираем модель
    process_model = process.get("default_model") or "phi4-mini"
    best_available = runtime._pick_best_chat_model(
        models, preferred=process_model, min_rank=min_rank
    )

    # 7. LLM call с timeout
    response = await asyncio.wait_for(
        runtime.chat_llm(
            model=best_available,
            messages=[
                {"role": "system", "content": system_prompt},
                {"role": "user", "content": rendered_prompt},
            ],
            options={"num_predict": max_tokens, "temperature": 0.2},
        ),
        timeout=90.0,
    )

    # 8. Логируем LLM call
    await db.log_llm_call(
        scope=f"process:{process_code}",
        model=best_available,
        latency_ms=...,
    )

    return ProcessExecutionResult(
        text=response["message"]["content"],
        model=best_available,
        raw=response,
        degraded=False,
    )