use std::sync::Arc;
use cudarc::driver::CudaDevice;
mod attention; mod macrogrok; mod memory; mod tavily; mod transformer;
use memory::AttentionMemory;
use tavily::TavilyClient;

#[tokio::main]
async fn main() -> anyhow::Result<()> {
    let device = Arc::new(CudaDevice::new(0)?);
    let config = transformer::TransformerConfig { hidden_size: 4096, num_layers: 32, num_heads: 32, head_dim: 128, max_seq_len: 8192, vocab_size: 32000 };
    let mut mem = AttentionMemory::new(device.clone(), config.num_layers, config.num_heads, config.head_dim, config.max_seq_len)?;
    let tavily = TavilyClient::new(std::env::var("TAVILY_API_KEY").unwrap_or_default());

    // MACROGROK demo
    let (out, score, flags) = macrogrok::infer4([macrogrok::Q1_14(16384), macrogrok::Q1_14(0), macrogrok::Q1_14(0), macrogrok::Q1_14(0)]);
    println!("MACROGROK INFER4 out={} score={} flags={:03b}", out, score, flags);

    // Transformer loop placeholder: tokenize → forward → tool call → Tavily → continue
    let _ = (&config, &mut mem, &tavily);
    println!("Harness ready. Set TAVILY_API_KEY and wire model weights.");
    Ok(())
}