Reconnect yhdellä rivillä: ei floodata terminaalia

Sama rivi päivittyy laskurilla: '↻ Yhdistetään uudelleen... (3)'
Rivi poistetaan kun yhteys palautuu.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

network-poc/Dockerfile.native-node

@@ -1,7 +1,7 @@
 FROM rust:slim AS builder
 
 RUN apt-get update && apt-get install -y \
-    pkg-config libssl-dev g++ \
+    pkg-config libssl-dev g++ libvulkan-dev \
     && rm -rf /var/lib/apt/lists/*
 
 WORKDIR /app
@@ -9,22 +9,27 @@ COPY Cargo.toml Cargo.lock ./
 COPY hub/Cargo.toml hub/Cargo.toml
 COPY node/Cargo.toml node/Cargo.toml
 COPY native-node/Cargo.toml native-node/Cargo.toml
+COPY cli/Cargo.toml cli/Cargo.toml
 
 # Tyhjät src-tiedostot riippuvuuksien esikääntämistä varten
-RUN mkdir -p hub/src node/src native-node/src \
+RUN mkdir -p hub/src node/src native-node/src cli/src \
     && echo "fn main(){}" > hub/src/main.rs \
     && echo "" > node/src/lib.rs \
     && echo "fn main(){}" > native-node/src/main.rs \
+    && echo "fn main(){}" > cli/src/main.rs \
     && cargo build --release -p native-node 2>/dev/null || true
 
 COPY native-node/src native-node/src
-RUN cargo build --release -p native-node
+# Touch pakottaa rekompilauksen dummy-binaryn yli
+RUN touch native-node/src/main.rs && cargo build --release -p native-node
 
 FROM debian:bookworm-slim
-RUN apt-get update && apt-get install -y ca-certificates && rm -rf /var/lib/apt/lists/*
+RUN apt-get update && apt-get install -y ca-certificates libvulkan1 && rm -rf /var/lib/apt/lists/*
 COPY --from=builder /app/target/release/native-node /usr/local/bin/native-node
 
-ENV HUB_URL=ws://hub:3000/ws
+ENV HUB_URL=ws://agentic-poc:3000/ws
+ENV OLLAMA_URL=http://ollama:11434
+ENV OLLAMA_MODEL=qwen2.5-coder:7b
 ENV ALLOCATED_GB=4
 
 CMD ["native-node"]

network-poc/docker-compose.yml

@@ -11,18 +11,14 @@ services:
     # Käännetään aina käynnistyksen yhteydessä varmuuden vuoksi Wasm uusimmista koodeista, ja päälle pyöräytetään Hub!
     command: bash -c "cd node && wasm-pack build --release --target web --out-dir ../static/pkg && cd ../hub && cargo run"
 
-  # Valinnainen natiivi-solmu — kerää oikeat laitteistotiedot (nvidia-smi-taso)
-  native-node:
-    build:
-      context: .
-      dockerfile: Dockerfile.native-node
-    container_name: kipina_native_node
-    environment:
-      - HUB_URL=ws://agentic-poc:3000/ws
-      - ALLOCATED_GB=4
-    depends_on:
-      - agentic-poc
-    # GPU passthrough (valinnainen — toimii myös ilman)
+  # Ollama — LLM-inferenssi GPU:lla (NVIDIA/AMD/Apple)
+  ollama:
+    image: ollama/ollama:latest
+    container_name: kipina_ollama
+    ports:
+      - "11434:11434"
+    volumes:
+      - ollama-models:/root/.ollama
     deploy:
       resources:
         reservations:
@@ -32,3 +28,23 @@ services:
               capabilities: [gpu]
     profiles:
       - native
+
+  # Natiivisolmu — yhdistää hubiin ja käyttää Ollamaa inferenssiin
+  native-node:
+    build:
+      context: .
+      dockerfile: Dockerfile.native-node
+    container_name: kipina_native_node
+    environment:
+      - HUB_URL=ws://agentic-poc:3000/ws
+      - OLLAMA_URL=http://ollama:11434
+      - OLLAMA_MODEL=qwen2.5-coder:7b
+      - ALLOCATED_GB=4
+    depends_on:
+      - agentic-poc
+      - ollama
+    profiles:
+      - native
+
+volumes:
+  ollama-models:

network-poc/hub/src/main.rs

@@ -39,6 +39,7 @@ struct AppState {
     ip_connections: Mutex<HashMap<IpAddr, u32>>,
     node_ips: Mutex<HashMap<u64, IpAddr>>,
     node_tasks: Mutex<HashMap<u64, String>>,  // node_id → selected_task
+    node_types: Mutex<HashMap<u64, String>>,  // node_id → "native" | "browser"
     node_busy: Mutex<std::collections::HashSet<u64>>,  // Solmut joilla on aktiivinen tehtävä
     pending_task_ids: Mutex<std::collections::HashSet<String>>,  // Hubin jakamat task_id:t (gamification-validointi)
     api_rate_limits: Mutex<HashMap<IpAddr, (std::time::Instant, u32)>>,  // IP → (ikkuna-alku, pyyntömäärä)
@@ -260,6 +261,7 @@ async fn main() {
         ip_connections: Mutex::new(HashMap::new()),
         node_ips: Mutex::new(HashMap::new()),
         node_tasks: Mutex::new(HashMap::new()),
+        node_types: Mutex::new(HashMap::new()),
         node_busy: Mutex::new(std::collections::HashSet::new()),
         pending_task_ids: Mutex::new(std::collections::HashSet::new()),
         api_rate_limits: Mutex::new(HashMap::new()),
@@ -382,6 +384,8 @@ async fn main() {
         .route("/api/pairs", get(api_pairs))
         .route("/api/stats", get(api_stats))
         .route("/api/v1/chat/completions", axum::routing::post(api_chat_completions))
+        .route("/api/v1/model", axum::routing::post(api_change_model))
+        .route("/api/v1/hardware", get(api_hardware))
         .route("/admin", get(admin_page))
         .nest_service("/", {
             let static_dir = std::env::var("STATIC_DIR").unwrap_or_else(|_| "../static".to_string());
@@ -677,6 +681,7 @@ async fn handle_socket(socket: WebSocket, state: Arc<AppState>, ip: IpAddr) {
                     state.db.insert_session(node_id, &ip.to_string(), node_type, &json);
                 }
                 state.node_tasks.lock().unwrap().insert(node_id, selected_task);
+                state.node_types.lock().unwrap().insert(node_id, node_type.to_string());
 
                 if node_type == "native" {
                     let sys = json.get("system");
@@ -934,6 +939,7 @@ async fn handle_socket(socket: WebSocket, state: Arc<AppState>, ip: IpAddr) {
         ips.remove(&node_id);
         vram.remove(&node_id);
     }
+    state.node_types.lock().unwrap().remove(&node_id);
     tracing::info!("Solmu {} ({}) poistui verkosta.", node_id, ip);
     broadcast_stats(&state).await;
     sender_task.abort();
@@ -943,6 +949,8 @@ struct ChatCompletionRequest {
     model: String,
     prompt: String,
     task_id: String,
+    #[serde(default)]
+    max_tokens: Option<u64>,
 }
 
 #[derive(serde::Serialize)]
@@ -952,6 +960,47 @@ struct ChatCompletionResponse {
     tokens_generated: u64,
 }
 
+async fn api_hardware(
+    axum::extract::State(state): axum::extract::State<Arc<AppState>>,
+) -> axum::response::Response {
+    // Etsitään natiivisolmun GPU-tiedot sessiosta
+    let sessions = state.db.get_sessions(50);
+    let native = sessions.iter().find(|s| {
+        s.get("node_type").and_then(|v| v.as_str()) == Some("native")
+    });
+
+    let (vram_mb, gpu_name, ram_mb) = if let Some(s) = native {
+        let gpus = s.get("gpus").and_then(|v| v.as_array());
+        let gpu = gpus.and_then(|g| g.first());
+        let vram = gpu.and_then(|g| g.get("vram_total_mb")).and_then(|v| v.as_u64()).unwrap_or(0);
+        let name = gpu.and_then(|g| g.get("name")).and_then(|v| v.as_str()).unwrap_or("?");
+        let ram = s.get("system").and_then(|v| v.get("ram_total_mb")).and_then(|v| v.as_u64()).unwrap_or(0);
+        (vram, name.to_string(), ram)
+    } else {
+        (0, "ei natiivisolmua".to_string(), 0)
+    };
+
+    axum::Json(serde_json::json!({
+        "gpu_name": gpu_name,
+        "vram_mb": vram_mb,
+        "ram_mb": ram_mb,
+    })).into_response()
+}
+
+async fn api_change_model(
+    axum::extract::State(state): axum::extract::State<Arc<AppState>>,
+    axum::Json(payload): axum::Json<serde_json::Value>,
+) -> axum::response::Response {
+    let model = payload.get("model").and_then(|v| v.as_str()).unwrap_or("");
+    if model.is_empty() {
+        return (axum::http::StatusCode::BAD_REQUEST, "model puuttuu").into_response();
+    }
+    tracing::info!("Mallin vaihto: {}", model);
+    let msg = serde_json::json!({ "type": "change_model", "model": model });
+    let _ = state.stats_tx.send(msg.to_string());
+    axum::Json(serde_json::json!({ "status": "ok", "model": model })).into_response()
+}
+
 async fn api_chat_completions(
     axum::extract::State(state): axum::extract::State<Arc<AppState>>,
     ConnectInfo(addr): ConnectInfo<SocketAddr>,
@@ -972,10 +1021,11 @@ async fn api_chat_completions(
         }
     }
 
-    // Etsitään vapaa tai varattu solmu, joka vastaa pyydettyä mallia
+    // Etsitään vapaa solmu — priorisoidaan natiivisolmut (GPU) selaimen edelle
     let (target_node_free, target_node_any, total_matching) = {
         let tasks = state.node_tasks.lock().unwrap();
         let busy = state.node_busy.lock().unwrap();
+        let node_types = state.node_types.lock().unwrap();
         let matching: Vec<u64> = tasks.iter().filter(|(_, task)| {
             if payload.model == "qwen-coder" {
                 task.starts_with("qwen-coder")
@@ -983,7 +1033,12 @@ async fn api_chat_completions(
                 **task == payload.model
             }
         }).map(|(k, _)| *k).collect();
-        let free = matching.iter().find(|id| !busy.contains(id)).copied();
+        // Vapaat solmut: natiivi ensin, sitten selain
+        let free_native = matching.iter().find(|id| {
+            !busy.contains(id) && node_types.get(id).map(|t| t == "native").unwrap_or(false)
+        }).copied();
+        let free_any = matching.iter().find(|id| !busy.contains(id)).copied();
+        let free = free_native.or(free_any);
         let any = matching.first().copied();
         (free, any, matching.len())
     };
@@ -1059,12 +1114,15 @@ async fn api_chat_completions(
     state.node_busy.lock().unwrap().insert(target_node_id);
     state.pending_task_ids.lock().unwrap().insert(payload.task_id.clone());
 
-    let msg = serde_json::json!({
+    let mut msg = serde_json::json!({
         "type": "llm_prompt",
         "prompt": payload.prompt,
         "model": payload.model,
         "task_id": payload.task_id,
     });
+    if let Some(mt) = payload.max_tokens {
+        msg.as_object_mut().unwrap().insert("max_tokens".to_string(), serde_json::json!(mt));
+    }
 
     // Odotuskanava valmiiksi (solmu palauttaa tuloksen stats_tx kautta)
     let mut rx = state.stats_tx.subscribe();

network-poc/native-node/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "native-node"
-version = "0.1.0"
+version = "0.2.0"
 edition = "2024"
 
 [dependencies]
@@ -12,10 +12,6 @@ serde_json = "1.0"
 sysinfo = "0.30"
 nvml-wrapper = "0.10"
 wgpu = "24"
-candle-core = { version = "0.8", features = ["cuda"] }
-candle-nn = "0.8"
-candle-transformers = "0.8"
-hf-hub = "0.4"
-tokenizers = "0.19"
+reqwest = { version = "0.12", features = ["json"] }
 tracing = "0.1"
 tracing-subscriber = { version = "0.3", features = ["env-filter"] }

network-poc/native-node/src/inference.rs

@@ -1,261 +1,114 @@
-use candle_core::{Device, Tensor, DType};
-use candle_nn::VarBuilder;
-use candle_transformers::models::qwen2::{Config as QwenConfig, ModelForCausalLM as QwenModel};
-use hf_hub::{api::sync::Api, Repo, RepoType};
 use std::time::Instant;
-
-/// Top-k sampling with temperature and repetition penalty
-fn sample_top_k(logits: &Tensor, k: usize, temperature: f64, generated_tokens: &[u32], repetition_penalty: f64, rng_state: &mut u64) -> Result<u32, String> {
-    let mut logits_vec: Vec<f32> = logits.to_vec1::<f32>().map_err(|e| format!("to_vec1: {}", e))?;
-    if logits_vec.is_empty() { return Err("Tyhjä logits".to_string()); }
-
-    // Repetition penalty: rankaisee jo generoituja tokeneita
-    for &token_id in generated_tokens {
-        if (token_id as usize) < logits_vec.len() {
-            let logit = &mut logits_vec[token_id as usize];
-            if *logit > 0.0 {
-                *logit /= repetition_penalty as f32;
-            } else {
-                *logit *= repetition_penalty as f32;
-            }
-        }
-    }
-
-    // Temperature scaling
-    if temperature > 0.0 && temperature != 1.0 {
-        for logit in logits_vec.iter_mut() {
-            *logit /= temperature as f32;
-        }
-    }
-
-    // Top-k: etsitään k suurinta
-    let mut indexed: Vec<(usize, f32)> = logits_vec.iter().enumerate().map(|(i, &v)| (i, v)).collect();
-    indexed.sort_by(|a, b| b.1.partial_cmp(&a.1).unwrap_or(std::cmp::Ordering::Equal));
-    indexed.truncate(k);
-
-    if k == 1 || temperature == 0.0 {
-        return Ok(indexed[0].0 as u32);
-    }
-
-    // Softmax top-k:lle
-    let max_logit = indexed[0].1;
-    let exps: Vec<f32> = indexed.iter().map(|x| (x.1 - max_logit).exp()).collect();
-    let sum: f32 = exps.iter().sum();
-    let probs: Vec<f32> = exps.iter().map(|e| e / sum).collect();
-
-    // XorShift64 RNG
-    *rng_state ^= *rng_state << 13;
-    *rng_state ^= *rng_state >> 7;
-    *rng_state ^= *rng_state << 17;
-    let rand_val = (*rng_state % 10000) as f32 / 10000.0;
-
-    let mut cumulative = 0.0;
-    for (i, p) in probs.iter().enumerate() {
-        cumulative += p;
-        if rand_val < cumulative {
-            return Ok(indexed[i].0 as u32);
-        }
-    }
-
-    Ok(indexed[0].0 as u32)
-}
+use std::cell::RefCell;
 
 pub struct LlmEngine {
-    tokenizer: tokenizers::Tokenizer,
-    model: QwenModel,
-    device: Device,
-    eos_token: u32,
+    ollama_url: String,
+    model: RefCell<String>,
+    client: reqwest::Client,
 }
 
 impl LlmEngine {
     pub fn load() -> Result<Self, String> {
-        let device = Device::cuda_if_available(0).map_err(|e| format!("Device: {}", e))?;
-        let device_name = if device.is_cuda() { "CUDA" } else { "CPU" };
-        tracing::info!("LLM device: {}", device_name);
+        let ollama_url = std::env::var("OLLAMA_URL").unwrap_or_else(|_| "http://localhost:11434".to_string());
+        let model = std::env::var("OLLAMA_MODEL").unwrap_or_else(|_| "qwen2.5-coder:7b".to_string());
 
-        let dtype = if device.is_cuda() { DType::F16 } else { DType::F32 };
+        tracing::info!("Ollama backend: {} | malli: {}", ollama_url, model);
 
-        tracing::info!("Ladataan Qwen2.5-Coder-0.5B-Instruct...");
-        let api = Api::new().map_err(|e| format!("HF API: {}", e))?;
-        let repo = api.repo(Repo::with_revision(
-            "Qwen/Qwen2.5-Coder-0.5B-Instruct".to_string(),
-            RepoType::Model,
-            "main".to_string(),
-        ));
+        let client = reqwest::Client::builder()
+            .timeout(std::time::Duration::from_secs(600))
+            .build()
+            .map_err(|e| format!("HTTP client: {}", e))?;
 
-        let tokenizer_path = repo.get("tokenizer.json").map_err(|e| format!("Tokenizer lataus: {}", e))?;
-        let model_path = repo.get("model.safetensors").map_err(|e| format!("Malli lataus: {}", e))?;
+        Ok(LlmEngine { ollama_url, model: RefCell::new(model), client })
+    }
 
-        tracing::info!("Ladataan tokenizer: {:?}", tokenizer_path);
-        let tokenizer = tokenizers::Tokenizer::from_file(&tokenizer_path)
-            .map_err(|e| format!("Tokenizer: {}", e))?;
+    pub fn model_name(&self) -> String {
+        self.model.borrow().clone()
+    }
 
-        let config = QwenConfig {
-            vocab_size: 151936,
-            hidden_size: 896,
-            intermediate_size: 4864,
-            num_hidden_layers: 24,
-            num_attention_heads: 14,
-            num_key_value_heads: 2,
-            max_position_embeddings: 32768,
-            sliding_window: 32768,
-            max_window_layers: 21,
-            tie_word_embeddings: true,
-            rope_theta: 1000000.0,
-            rms_norm_eps: 1e-6,
-            use_sliding_window: false,
-            hidden_act: candle_nn::Activation::Silu,
-        };
+    pub fn set_model(&self, new_model: String) {
+        *self.model.borrow_mut() = new_model;
+    }
+
+    /// Varmistaa että malli on ladattu Ollamaan (ollama pull)
+    pub async fn ensure_model(&self) -> Result<(), String> {
+        let model = self.model.borrow().clone();
+        tracing::info!("Tarkistetaan malli {}...", model);
+        let resp = self.client.post(format!("{}/api/pull", self.ollama_url))
+            .json(&serde_json::json!({ "name": model, "stream": false }))
+            .send()
+            .await
+            .map_err(|e| format!("Ollama pull: {}", e))?;
+
+        if resp.status().is_success() {
+            tracing::info!("Malli {} valmis", model);
+            Ok(())
+        } else {
+            Err(format!("Ollama pull epäonnistui: {}", resp.status()))
+        }
+    }
+
+    pub async fn generate(&self, prompt: &str, max_tokens: usize) -> Result<GenerateResult, String> {
+        let system = "You are a coding assistant. Respond with ONLY code. No explanations, no markdown, no comments unless asked.";
+        let model = self.model.borrow().clone();
 
         let start = Instant::now();
-        let vb = unsafe {
-            VarBuilder::from_mmaped_safetensors(&[model_path.clone()], dtype, &device)
-                .map_err(|e| format!("VarBuilder: {}", e))?
-        };
-        let model = QwenModel::new(&config, vb).map_err(|e| format!("Malli: {}", e))?;
-        tracing::info!("Malli ladattu ({:.1}s) — {}", start.elapsed().as_secs_f64(), device_name);
+        let resp = self.client.post(format!("{}/api/generate", self.ollama_url))
+            .json(&serde_json::json!({
+                "model": model,
+                "prompt": prompt,
+                "system": system,
+                "stream": false,
+                "options": {
+                    "num_predict": max_tokens,
+                    "temperature": 0.7,
+                    "top_k": 40,
+                    "repeat_penalty": 1.15,
+                    "stop": ["<|im_end|>", "\n###", "\nExplanation", "\nNote:"]
+                }
+            }))
+            .send()
+            .await
+            .map_err(|e| format!("Ollama generate: {}", e))?;
 
-        Ok(LlmEngine {
-            tokenizer,
-            model,
-            device,
-            eos_token: 151645,
-        })
+        if !resp.status().is_success() {
+            return Err(format!("Ollama HTTP {}", resp.status()));
         }
 
-    pub fn generate(&mut self, prompt: &str, max_tokens: usize) -> Result<GenerateResult, String> {
-        // Prefill: aloitetaan vastaus ```-koodiblokkilla → malli jatkaa suoraan koodilla
-        let formatted = format!("<|im_start|>system\nYou are a coding assistant. Respond with ONLY code. No explanations, no markdown, no comments unless asked.<|im_end|>\n<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n```\n", prompt);
+        let body: serde_json::Value = resp.json().await
+            .map_err(|e| format!("Ollama JSON: {}", e))?;
 
-        let encoding = self.tokenizer.encode(formatted.as_str(), true)
-            .map_err(|e| format!("Encode: {}", e))?;
-        let input_ids: Vec<u32> = encoding.get_ids().to_vec();
-        let input_len = input_ids.len();
+        let text = body["response"].as_str().unwrap_or("").to_string();
+        let total_duration_ns = body["total_duration"].as_u64().unwrap_or(0);
+        let eval_count = body["eval_count"].as_u64().unwrap_or(0) as usize;
+        let eval_duration_ns = body["eval_duration"].as_u64().unwrap_or(1);
 
-        // Nollataan KV-cache edellisestä promptista
-        self.model.clear_kv_cache();
-
-        // Sampling-parametrit
-        let temperature = 0.7;
-        let top_k = 40;
-        let repetition_penalty = 1.15;
-        let mut rng_state: u64 = std::time::SystemTime::now()
-            .duration_since(std::time::UNIX_EPOCH)
-            .unwrap()
-            .as_nanos() as u64;
-
-        let start = Instant::now();
-
-        // Prefill
-        let input = Tensor::new(input_ids.as_slice(), &self.device)
-            .and_then(|t| t.unsqueeze(0))
-            .map_err(|e| format!("Tensor: {}", e))?;
-
-        let logits = self.model.forward(&input, 0)
-            .map_err(|e| format!("Forward prefill: {}", e))?;
-
-        let logits = logits.squeeze(0).map_err(|e| format!("Squeeze: {}", e))?;
-        let logits = if logits.dims().len() == 2 {
-            let seq_len = logits.dim(0).map_err(|e| format!("Dim: {}", e))?;
-            if seq_len == 0 { return Err("Tyhjä tensori".to_string()); }
-            logits.get(seq_len - 1).map_err(|e| format!("Get: {}", e))?
-        } else {
-            logits
-        };
-
-        let mut generated_text = String::new();
-        let mut tokens_generated: usize = 0;
-        let mut all_tokens: Vec<u32> = Vec::new();
-
-        let mut next_token = sample_top_k(&logits, top_k, temperature, &all_tokens, repetition_penalty, &mut rng_state)?;
-
-        if next_token != self.eos_token {
-            if let Ok(text) = self.tokenizer.decode(&[next_token], true) {
-                generated_text.push_str(&text);
-            }
-            all_tokens.push(next_token);
-            tokens_generated += 1;
-        }
-
-        // Autoregressive
-        let mut pos = input_len;
-        for _ in 1..max_tokens {
-            if next_token == self.eos_token { break; }
-
-            let input = Tensor::new(&[next_token], &self.device)
-                .and_then(|t| t.unsqueeze(0))
-                .map_err(|e| format!("Tensor: {}", e))?;
-
-            let logits = self.model.forward(&input, pos)
-                .map_err(|e| format!("Forward pos {}: {}", pos, e))?;
-
-            let logits = logits.squeeze(0).map_err(|e| format!("Squeeze: {}", e))?;
-            let logits = if logits.dims().len() == 2 {
-                let seq_len = logits.dim(0).map_err(|e| format!("Dim: {}", e))?;
-                if seq_len == 0 { break; }
-                logits.get(seq_len - 1).map_err(|e| format!("Get: {}", e))?
-            } else {
-                logits
-            };
-            next_token = sample_top_k(&logits, top_k, temperature, &all_tokens, repetition_penalty, &mut rng_state)?;
-            pos += 1;
-
-            if next_token == self.eos_token { break; }
-
-            if let Ok(text) = self.tokenizer.decode(&[next_token], true) {
-                generated_text.push_str(&text);
-
-                // Stop-sekvenssit: katkaistaan kun malli alkaa selittää
-                let lower = generated_text.to_lowercase();
-                if lower.contains("\n###") || lower.contains("\nexplanation") || lower.contains("\nnote:") || lower.contains("\noutput:") || lower.contains("\n```\n\n") || lower.contains("\n// example") || lower.contains("\n# example") {
-                    for stop in &["\n###", "\nExplanation", "\nNote:", "\nOutput:", "\n```\n\n", "\n// Example", "\n// example", "\n# Example", "\n# example"] {
-                        if let Some(pos) = generated_text.find(stop) {
-                            generated_text.truncate(pos);
-                        }
-                    }
-                    break;
-                }
-            }
-            all_tokens.push(next_token);
-            tokens_generated += 1;
-        }
-
-        let gen_time = start.elapsed();
-        let tokens_per_sec = if gen_time.as_secs_f64() > 0.0 {
-            tokens_generated as f64 / gen_time.as_secs_f64()
+        let duration_ms = start.elapsed().as_millis() as f64;
+        let tokens_per_sec = if eval_duration_ns > 0 {
+            eval_count as f64 / (eval_duration_ns as f64 / 1_000_000_000.0)
         } else { 0.0 };
 
         Ok(GenerateResult {
-            text: strip_markdown_wrapper(&generated_text),
-            tokens_generated,
-            duration_ms: gen_time.as_millis() as f64,
+            text: strip_code_fences(&text),
+            tokens_generated: eval_count,
+            duration_ms,
             tokens_per_sec,
         })
     }
 }
 
-const LANG_TAGS: &[&str] = &[
-    "python", "py", "rust", "rs", "javascript", "js", "typescript", "ts",
-    "java", "kotlin", "scala", "go", "ruby", "rb", "php", "swift",
-    "c", "cpp", "c++", "c#", "csharp", "r", "sql", "bash", "sh", "zsh",
-    "html", "css", "json", "yaml", "yml", "toml", "xml", "markdown", "md",
-    "lua", "perl", "dart", "elixir", "haskell", "hs", "ocaml", "zig",
-    "plaintext", "text", "txt",
-];
-
-/// Siivoa mallin tuottama vastaus (prefill-yhteensopiva).
-fn strip_markdown_wrapper(text: &str) -> String {
+/// Siivoa mahdolliset markdown-koodiblokki-merkit
+fn strip_code_fences(text: &str) -> String {
     let mut result = text.trim().to_string();
 
-    // 1. Kielitunniste — VAIN tunnettu kieli
+    // Poista aloittava ```lang
+    if result.starts_with("```") {
         if let Some(nl) = result.find('\n') {
-        let first = result[..nl].trim().to_lowercase();
-        if LANG_TAGS.contains(&first.as_str()) {
             result = result[nl + 1..].to_string();
         }
     }
 
-    // 2. Sulkeva ``` — VAIN omalla rivillään lopussa
+    // Poista sulkeva ```
     let trimmed = result.trim_end();
     if trimmed.ends_with("```") {
         let before = &trimmed[..trimmed.len() - 3];
@@ -264,29 +117,7 @@ fn strip_markdown_wrapper(text: &str) -> String {
         }
     }
 
-    // 3. Johdantolauseet
-    let lower = result.trim().to_lowercase();
-    for prefix in &["sure!", "here is", "here's", "certainly!", "below is"] {
-        if lower.starts_with(prefix) {
-            if let Some(nl) = result.find('\n') { result = result[nl + 1..].to_string(); }
-            break;
-        }
-    }
-
-    // 4. Selityskommentit alusta
-    let mut lines: Vec<&str> = result.trim().lines().collect();
-    while !lines.is_empty() {
-        let first = lines[0].trim();
-        let is_preamble = first.starts_with("# ") && !first.starts_with("#!")
-            && (first.to_lowercase().contains("this is")
-                || first.to_lowercase().contains("simple")
-                || first.to_lowercase().contains("program that")
-                || first.to_lowercase().contains("here is")
-                || first.to_lowercase().contains("the following")
-                || first.to_lowercase().contains("below"));
-        if is_preamble { lines.remove(0); } else { break; }
-    }
-    lines.join("\n").trim().to_string()
+    result
 }
 
 pub struct GenerateResult {

network-poc/native-node/src/main.rs

@@ -285,15 +285,19 @@ async fn main() {
         }
     }
 
-    // Ladataan LLM-malli
-    tracing::info!("Ladataan LLM-mallia...");
-    let mut llm = match inference::LlmEngine::load() {
+    // Ollama-backend
+    tracing::info!("Alustetaan Ollama-yhteyttä...");
+    let llm = match inference::LlmEngine::load() {
         Ok(engine) => {
-            tracing::info!("LLM valmis inferenssiin!");
+            // Varmistetaan malli (ollama pull) — odotetaan kunnes valmis
+            match engine.ensure_model().await {
+                Ok(()) => tracing::info!("Ollama valmis inferenssiin!"),
+                Err(e) => tracing::warn!("Mallin lataus: {} — yritetään silti", e),
+            }
             Some(engine)
         }
         Err(e) => {
-            tracing::warn!("LLM-lataus epäonnistui: {} — toimitaan ilman inferenssiä", e);
+            tracing::warn!("Ollama-alustus epäonnistui: {} — toimitaan ilman inferenssiä", e);
             None
         }
     };
@@ -324,11 +328,13 @@ async fn main() {
                                 
                                 if !prompt.is_empty() && msg_model.starts_with("qwen-coder") {
 
-                                    if let Some(ref mut engine) = llm {
+                                    if let Some(ref engine) = llm {
                                         busy = true;
-                                        tracing::info!("Generoidaan (task_id: {}): \"{}\"", task_id, prompt);
+                                        let max_tokens = task.get("max_tokens").and_then(|v| v.as_u64()).unwrap_or(512) as usize;
+                                        tracing::info!("Generoidaan (task_id: {}, max_tokens: {}): \"{}\"", task_id, max_tokens, &prompt[..prompt.len().min(100)]);
 
-                                        match engine.generate(prompt, 64) {
+                                        let model_name = engine.model_name();
+                                        match engine.generate(prompt, max_tokens).await {
                                             Ok(result) => {
                                                 tracing::info!(
                                                     "Tulos: {} tokenia | {:.0}ms | {:.1} tok/s | \"{}\"",
@@ -341,7 +347,7 @@ async fn main() {
                                                 let done = json!({
                                                     "type": "llm_done",
                                                     "prompt": prompt,
-                                                    "model": "Qwen2.5-Coder-0.5B (native/GPU)",
+                                                    "model": format!("{} (Ollama)", model_name),
                                                     "response": result.text,
                                                     "tokens_generated": result.tokens_generated,
                                                     "duration_ms": result.duration_ms,
@@ -360,7 +366,21 @@ async fn main() {
                                 }
                             }
                         }
-                        // Ohitetaan pair_task, stats jne.
+                        // Mallin vaihto lennossa
+                        if text.contains("change_model") {
+                            if let Ok(task) = serde_json::from_str::<serde_json::Value>(&text) {
+                                if let Some(new_model) = task.get("model").and_then(|v| v.as_str()) {
+                                    if let Some(ref engine) = llm {
+                                        tracing::info!("Vaihdetaan malli: {}", new_model);
+                                        engine.set_model(new_model.to_string());
+                                        match engine.ensure_model().await {
+                                            Ok(()) => tracing::info!("Malli {} valmis!", new_model),
+                                            Err(e) => tracing::error!("Mallin lataus epäonnistui: {}", e),
+                                        }
+                                    }
+                                }
+                            }
+                        }
                     }
                 }
                 tracing::warn!("Yhteys hubiin katkesi — yritetään uudelleen 5s...");

network-poc/static/index.html

@@ -1157,7 +1157,7 @@
             coder: { name: 'Koodari — System Prompt', model: 'qwen-coder', default: 'Olet kokenut ohjelmistokehittäjä. Kirjoita selkeää, testattavaa koodia ja vastaa aina koodilla.' },
             data: { name: 'Data-Agentti — System Prompt', model: 'qwen-coder', default: 'Olet tietokanta-asiantuntija. Vastaat skeemojen suunnittelusta, SQL-kyselyiden optimoinnista ja datamalleista.' },
             qa: { name: 'QA — System Prompt', model: 'qwen-coder', default: 'Olet laadunvarmistaja (QA). Kirjoitat testejä, etsit virheitä ja varmistat, että kaikki reunatapaukset on huomioitu.' },
-            tester: { name: 'DevOps — System Prompt', model: 'qwen-coder', default: 'Olet DevOps-insinööri. Vastaat koodin julkaisuputkista, serveri-infrastruktuurista ja ympäristön suorituskyvystä.' },
+            tester: { name: 'DevOps — System Prompt', model: 'qwen-coder', default: 'Olet DevOps-insinööri. Kirjoitat Dockerfile- ja docker-compose.yml-tiedostot, README:t ja käynnistysohjeet. Käytä aina multi-stage Docker buildia ja docker compose -orkestrointia.' },
         };
         const selectedAgents = new Set();
         let sharedPrompt = localStorage.getItem('kpn-shared-prompt') || '';
@@ -1672,15 +1672,28 @@
             if (e.key === 'Enter') sendUserText();
         });
 
-        // Kytkemme sivuston UI-puolen (JS) omaan passiiviseen WebSocket-kuuntelijaan.
-        const uiSocket = new WebSocket(`${window.location.protocol === 'https:' ? 'wss:' : 'ws:'}//${window.location.host}/ws`);
+        // WebSocket-yhteys hubiin — automaattinen reconnect
+        const wsUrl = `${window.location.protocol === 'https:' ? 'wss:' : 'ws:'}//${window.location.host}/ws`;
+        let uiSocket = null;
+        let wsReconnectTimer = null;
+
+        function connectHub() {
+            if (uiSocket && (uiSocket.readyState === 0 || uiSocket.readyState === 1)) return;
+            uiSocket = new WebSocket(wsUrl);
             window._uiSocket = uiSocket;
+            // Kytketään onmessage uudelleen (handler määritellään myöhemmin, asetetaan kun valmis)
+            setTimeout(() => {
+                if (window._wsMessageHandler) uiSocket.onmessage = window._wsMessageHandler;
+            }, 0);
         uiSocket.onopen = async () => {
             // Päivitetään agents-näkymän hub-status
             const hubDot = document.getElementById('agent-hub-dot');
             const hubLabel = document.getElementById('agent-hub-label');
             const hubStatus = document.getElementById('agent-hub-status');
             if (hubDot) hubDot.style.background = '#3fb950';
+            // Poistetaan reconnect-rivi
+            const reconnLine = document.getElementById('agent-terminal')?.querySelector('.term-reconnect');
+            if (reconnLine) reconnLine.remove();
             if (hubLabel) { hubLabel.textContent = 'Yhdistetty'; hubLabel.style.color = '#3fb950'; }
             if (hubStatus) hubStatus.title = 'Yhdistetty Kipinä Hubiin — tehtävien jakelu ja solmujen koordinointi aktiivinen';
 
@@ -1746,7 +1759,31 @@
                 coderEl.textContent = 'Disconnected';
                 coderEl.style.color = '#f85149';
             }
+            // Automaattinen reconnect 3s kuluttua
+            if (!wsReconnectTimer) {
+                // Päivitetään samaa riviä eikä floodata uusia
+                let reconnLine = termPanel?.querySelector('.term-reconnect');
+                let reconnCount = 0;
+                if (!reconnLine) {
+                    reconnLine = document.createElement('div');
+                    reconnLine.className = 'terminal-line term-reconnect';
+                    termPanel?.appendChild(reconnLine);
+                } else {
+                    reconnCount = parseInt(reconnLine.dataset.count || '0');
+                }
+                wsReconnectTimer = setTimeout(() => {
+                    wsReconnectTimer = null;
+                    reconnCount++;
+                    reconnLine.dataset.count = reconnCount;
+                    reconnLine.innerHTML = `  <span style="color:#d29922">↻ Yhdistetään uudelleen...${reconnCount > 1 ? ' (' + reconnCount + ')' : ''}</span>`;
+                    termPanel.scrollTop = termPanel.scrollHeight;
+                    connectHub();
+                }, 3000);
+            }
         };
+        } // connectHub()
+        connectHub();
+
         // Terminaalin komentorivi
         const termInput = document.getElementById('term-input');
         const termPanel = document.getElementById('agent-terminal');
@@ -1767,7 +1804,7 @@
         const activeStreams = {};
 
         // Lähettää promptin mallille ja palauttaa vastauksen (tai null virhetilanteessa)
-        async function kpnRun(model, prompt, silent) {
+        async function kpnRun(model, prompt, silent, maxTokens) {
             const taskId = crypto.randomUUID();
             // Yksittäinen status-rivi jota päivitetään läpi pyynnön elinkaaren
             const statusDiv = document.createElement('div');
@@ -1801,7 +1838,7 @@
                 const res = await fetch('/api/v1/chat/completions', {
                     method: 'POST',
                     headers: { 'Content-Type': 'application/json' },
-                    body: JSON.stringify({ model, prompt: fullPrompt, task_id: taskId }),
+                    body: JSON.stringify({ model, prompt: fullPrompt, task_id: taskId, ...(maxTokens ? { max_tokens: maxTokens } : {}) }),
                 });
 
                 if (!res.ok) {
@@ -1905,11 +1942,15 @@
         }
 
         // Projektikortti: tiedostovälilehdet + kopioi + lataa ZIP
+        // Globaali storage projektikorttien tiedostoille (välttää JSON data-attribuuttien ongelmat)
+        const projectFiles = {};
+
         function renderProjectCard(files, projectName) {
             const fileEntries = Object.entries(files);
             if (fileEntries.length === 0) return;
 
             const cardId = 'proj-' + Date.now();
+            projectFiles[cardId] = files;
             const tabsHtml = fileEntries.map(([name], i) =>
                 `<span class="proj-tab" data-card="${cardId}" data-idx="${i}" style="padding:4px 10px;cursor:pointer;border-radius:4px 4px 0 0;font-size:12px;${i === 0 ? 'background:#161b22;color:#58a6ff;border:1px solid #30363d;border-bottom:none' : 'color:#8b949e'}" onclick="switchProjectTab('${cardId}',${i})">${esc(name)}</span>`
             ).join('');
@@ -1926,7 +1967,7 @@
             const allText = fileEntries.map(([name, code]) => `# --- ${name} ---\n${code}`).join('\n\n');
 
             const cardHtml = `
-                <div id="${cardId}" style="margin:8px 0;border:1px solid #30363d;border-radius:6px;background:#161b22;overflow:hidden" data-files='${esc(JSON.stringify(files))}'>
+                <div id="${cardId}" style="margin:8px 0;border:1px solid #30363d;border-radius:6px;background:#161b22;overflow:hidden">
                     <div style="display:flex;align-items:center;justify-content:space-between;padding:8px 12px;background:#0d1117;border-bottom:1px solid #30363d">
                         <span style="color:#a371f7;font-weight:600;font-size:13px">${esc(projectName || 'Projekti')} <span style="color:#8b949e;font-weight:normal">(${fileEntries.length} tiedostoa)</span></span>
                         <span style="display:flex;gap:6px">
@@ -1960,7 +2001,7 @@
         window.copyFileContent = function(cardId, idx) {
             const card = document.getElementById(cardId);
             if (!card) return;
-            const files = JSON.parse(card.dataset.files);
+            const files = projectFiles[cardId];
             const entries = Object.entries(files);
             if (entries[idx]) {
                 navigator.clipboard.writeText(entries[idx][1]);
@@ -1973,7 +2014,7 @@
         window.copyAllFiles = function(cardId) {
             const card = document.getElementById(cardId);
             if (!card) return;
-            const files = JSON.parse(card.dataset.files);
+            const files = projectFiles[cardId];
             const text = Object.entries(files).map(([name, code]) => `# --- ${name} ---\n${code}`).join('\n\n');
             navigator.clipboard.writeText(text);
             const btn = card.querySelector('[onclick*="copyAllFiles"]');
@@ -1983,9 +2024,20 @@
         window.downloadZip = async function(cardId) {
             const card = document.getElementById(cardId);
             if (!card) return;
-            const files = JSON.parse(card.dataset.files);
+            const files = projectFiles[cardId];
+
+            // CRC-32 laskenta ZIP-tiedostoille
+            function crc32(bytes) {
+                let crc = 0xFFFFFFFF;
+                for (let i = 0; i < bytes.length; i++) {
+                    crc ^= bytes[i];
+                    for (let j = 0; j < 8; j++) {
+                        crc = (crc >>> 1) ^ (crc & 1 ? 0xEDB88320 : 0);
+                    }
+                }
+                return (crc ^ 0xFFFFFFFF) >>> 0;
+            }
 
-            // Luodaan ZIP ilman ulkoisia kirjastoja (yksinkertainen uncompressed ZIP)
             const entries = Object.entries(files);
             const parts = [];
             const centralDir = [];
@@ -1994,6 +2046,7 @@
             for (const [name, content] of entries) {
                 const nameBytes = new TextEncoder().encode(name);
                 const contentBytes = new TextEncoder().encode(content);
+                const crc = crc32(contentBytes);
 
                 // Local file header
                 const header = new Uint8Array(30 + nameBytes.length);
@@ -2001,8 +2054,9 @@
                 view.setUint32(0, 0x04034b50, true);  // Signature
                 view.setUint16(4, 20, true);           // Version needed
                 view.setUint16(8, 0, true);            // Method: store
-                view.setUint32(18, contentBytes.length, true); // Compressed size
-                view.setUint32(22, contentBytes.length, true); // Uncompressed size
+                view.setUint32(14, crc, true);         // CRC-32
+                view.setUint32(18, contentBytes.length, true);
+                view.setUint32(22, contentBytes.length, true);
                 view.setUint16(26, nameBytes.length, true);
                 header.set(nameBytes, 30);
 
@@ -2012,6 +2066,7 @@
                 cdView.setUint32(0, 0x02014b50, true);
                 cdView.setUint16(4, 20, true);
                 cdView.setUint16(6, 20, true);
+                cdView.setUint32(16, crc, true);       // CRC-32
                 cdView.setUint32(20, contentBytes.length, true);
                 cdView.setUint32(24, contentBytes.length, true);
                 cdView.setUint16(28, nameBytes.length, true);
@@ -2055,17 +2110,18 @@
             termLog(`\n<span style="color:#d29922;font-weight:bold">[1] Manageri</span> — projektin suunnittelu`);
             pipelineStep('manager', 'Suunnittelu', 'active', task);
             const managerPrompt = `List the source files needed for this project. One file per line, format:
-filename.py: what this file contains
+filename.py: one-line description
 
-Rules:
-- Max 4 files
-- Only .py, .toml, .json, .html files
+CONSTRAINTS — the coder can only generate ~400 tokens per file:
+- Max 3 files (keep it minimal)
+- Each file must be SHORT: one clear responsibility, no boilerplate
+- Only .py and pyproject.toml files
 - No directories, no paths, just filenames
-- List dependencies first, then main app (e.g. models.py before main.py)
-- Use pyproject.toml for dependencies (not requirements.txt)
+- List dependencies first, then main app
+- Prefer fewer, focused files over many small ones
 
 Project: ${task}`;
-            const plan = await kpnRun(agentPrompts.manager.model, managerPrompt);
+            const plan = await kpnRun(agentPrompts.manager.model, managerPrompt, false, 200);
             if (!plan) { termLog('  ✗ Pipeline keskeytyi (manageri)', '#f85149'); return; }
             pipelineStep('manager', 'Suunnittelu', 'done', task, plan);
 
@@ -2133,7 +2189,7 @@ start = "uvicorn main:app --reload"`;
 
                 const coderPrompt = `${context}Project: ${task}
 Write ONLY the file "${file.name}"${file.desc ? ': ' + file.desc : ''}.${extraInstructions}
-Use the exact libraries mentioned in the project description. Write correct, working code.`;
+IMPORTANT: Keep the code SHORT and focused. Max ~50 lines. No comments, no docstrings, no type hints unless essential. Write minimal, working code.`;
                 const code = await kpnRun(agentPrompts.coder.model, coderPrompt);
                 if (!code) {
                     termLog(`  ✗ Pipeline keskeytyi (${file.name})`, '#f85149');
@@ -2154,7 +2210,7 @@ Use the exact libraries mentioned in the project description. Write correct, wor
 If the code is correct, say "LGTM".
 
 ${allCode}`;
-            const review = await kpnRun(agentPrompts.tester.model, reviewPrompt);
+            const review = await kpnRun(agentPrompts.tester.model, reviewPrompt, false, 200);
             pipelineStep('tester', 'Review', 'done', `${Object.keys(generatedFiles).length} tiedostoa`, review);
 
             // Vaihe 4: Korjausluuppi — jos testaaja löysi ongelmia
@@ -2173,11 +2229,75 @@ Write the corrected code.`;
                 if (fixedCode) {
                     termLog(`\n<span style="color:#58a6ff;font-weight:bold">[${fileList.length + 4}] Testaaja</span> — uudelleenarviointi`);
                     pipelineStep('tester', 'Re-review', 'active', fixedCode);
-                    const reReview = await kpnRun(agentPrompts.tester.model, `Review the corrected code briefly:\n${fixedCode}`);
+                    const reReview = await kpnRun(agentPrompts.tester.model, `Review the corrected code briefly:\n${fixedCode}`, false, 128);
                     pipelineStep('tester', 'Re-review', 'done', fixedCode, reReview);
                 }
             }
 
+            // Vaihe 5: QA kirjoittaa testit
+            const step5 = fileList.length + (review && !review.toLowerCase().includes('lgtm') ? 5 : 3);
+            termLog(`\n<span style="color:#3fb950;font-weight:bold">[${step5}] QA</span> — testit`);
+            pipelineStep('qa', 'Testit', 'active', 'Kirjoitetaan testejä');
+            const qaPrompt = `Write a short test file (test_app.py) for this project. Use pytest. Max 3 test functions. Keep it minimal.
+
+${Object.entries(generatedFiles).map(([n, c]) => `--- ${n} ---\n${c}`).join('\n\n')}`;
+            const tests = await kpnRun(agentPrompts.qa.model, qaPrompt, false, 512);
+            if (tests) generatedFiles['test_app.py'] = tests;
+            pipelineStep('qa', 'Testit', 'done', 'test_app.py', tests);
+
+            // Vaihe 6: DevOps — Dockerfile
+            const step6 = step5 + 1;
+            termLog(`\n<span style="color:#d29922;font-weight:bold">[${step6}] DevOps</span> — Dockerfile`);
+            pipelineStep('tester', 'Dockerfile', 'active', 'Dockerfile');
+            const mainFile = Object.keys(generatedFiles).find(f => f.includes('main') || f.includes('app')) || Object.keys(generatedFiles)[0];
+            const dockerPrompt = `Write a Dockerfile for this Python project using uv package manager.
+
+RULES:
+- Base: python:3.12-slim
+- Install uv: COPY --from=ghcr.io/astral-sh/uv:latest /uv /usr/local/bin/uv
+- COPY pyproject.toml and then: RUN uv sync --no-dev
+- COPY all .py files
+- EXPOSE 8000
+- CMD ["uv", "run", "uvicorn", "${mainFile.replace('.py','')}:app", "--host", "0.0.0.0", "--port", "8000"]
+- Only output Dockerfile content, no explanations
+
+Files: ${Object.keys(generatedFiles).join(', ')}`;
+            const dockerfile = await kpnRun(agentPrompts.tester.model, dockerPrompt, false, 256);
+            if (dockerfile) generatedFiles['Dockerfile'] = dockerfile;
+            pipelineStep('tester', 'Dockerfile', 'done', 'Dockerfile', dockerfile);
+
+            // Vaihe 7: DevOps — docker-compose.yml
+            const step7 = step6 + 1;
+            termLog(`\n<span style="color:#d29922;font-weight:bold">[${step7}] DevOps</span> — docker-compose.yml`);
+            pipelineStep('tester', 'Compose', 'active', 'docker-compose.yml');
+            const composePrompt = `Write a docker-compose.yml for this project. Include:
+- app service (build from Dockerfile, port mapping, restart: unless-stopped)
+- db service if SQLite/PostgreSQL is used (volume for data persistence)
+- Named volumes for persistent data
+Only output the YAML content, nothing else.
+
+Files: ${Object.keys(generatedFiles).join(', ')}`;
+            const compose = await kpnRun(agentPrompts.tester.model, composePrompt, false, 256);
+            if (compose) generatedFiles['docker-compose.yml'] = compose;
+            pipelineStep('tester', 'Compose', 'done', 'docker-compose.yml', compose);
+
+            // Vaihe 8: DevOps — README
+            const step8 = step7 + 1;
+            termLog(`\n<span style="color:#d29922;font-weight:bold">[${step8}] DevOps</span> — README`);
+            pipelineStep('tester', 'README', 'active', 'README.md');
+            const readmePrompt = `Write a minimal README.md. Include ONLY:
+1. One-line description
+2. Quick start: docker compose up
+3. Development: uv sync && uv run uvicorn main:app --reload
+4. API endpoints (if applicable)
+5. Testing: uv run pytest
+Max 20 lines.
+
+Files: ${Object.keys(generatedFiles).join(', ')}`;
+            const readme = await kpnRun(agentPrompts.tester.model, readmePrompt, false, 256);
+            if (readme) generatedFiles['README.md'] = readme;
+            pipelineStep('tester', 'README', 'done', 'README.md', readme);
+
             termLog(`\n<span style="color:#a371f7;font-weight:bold">━━━ Pipeline valmis (${Object.keys(generatedFiles).length} tiedostoa) ━━━</span>`);
             renderProjectCard(generatedFiles, task);
         }
@@ -2196,11 +2316,73 @@ Write the corrected code.`;
             termLog(`\n<span style="color:#a371f7;font-weight:bold">━━━ Pipeline valmis ━━━</span>`);
         }
 
+        // Autokorjaus: tunnetut kirjoitusvirheet ja lähimmän komennon ehdotus
+        function autocorrect(input) {
+            const typos = {
+                'knp': 'kpn', 'kpb': 'kpn', 'kpm': 'kpn', 'kn': 'kpn', 'kp': 'kpn',
+                'kpn rnu': 'kpn run', 'kpn rn': 'kpn run', 'kpn ru': 'kpn run',
+                'kpn laod': 'kpn load', 'kpn lod': 'kpn load', 'kpn loa': 'kpn load',
+                'kpn porject': 'kpn project', 'kpn projcet': 'kpn project', 'kpn proejct': 'kpn project',
+                'kpn pipelien': 'kpn pipeline', 'kpn pipline': 'kpn pipeline',
+                'kpn staus': 'kpn status', 'kpn stauts': 'kpn status',
+                'kpn modles': 'kpn models', 'kpn mdoels': 'kpn models',
+                'kpn hlep': 'kpn help', 'kpn hep': 'kpn help',
+                'kpn clera': 'kpn clear', 'kpn claer': 'kpn clear',
+                'kpn helo': 'kpn hello', 'kpn hell': 'kpn hello',
+            };
+            // Tarkista koko komento ja ensimmäinen sana + alikomento
+            const lower = input.toLowerCase();
+            for (const [typo, fix] of Object.entries(typos)) {
+                if (lower === typo || lower.startsWith(typo + ' ')) {
+                    return fix + input.slice(typo.length);
+                }
+            }
+            // Levenshtein-etäisyys ensimmäiselle sanalle
+            const words = input.trim().split(/\s+/);
+            const firstWord = words[0].toLowerCase();
+            if (firstWord !== 'kpn' && firstWord.length >= 2 && firstWord.length <= 5) {
+                const dist = levenshtein(firstWord, 'kpn');
+                if (dist <= 2) return 'kpn' + input.slice(firstWord.length);
+            }
+            // Fuzzy-korjaus alikomentotasolla: "kpn rnu" → "kpn run"
+            if (firstWord === 'kpn' && words.length >= 2) {
+                const sub = words[1].toLowerCase();
+                const subCommands = ['help', 'run', 'project', 'pipeline', 'load', 'status', 'models', 'hello', 'clear'];
+                let bestMatch = null, bestDist = 3;
+                for (const cmd of subCommands) {
+                    const d = levenshtein(sub, cmd);
+                    if (d > 0 && d < bestDist) { bestDist = d; bestMatch = cmd; }
+                }
+                if (bestMatch) {
+                    words[1] = bestMatch;
+                    return words.join(' ');
+                }
+            }
+            return null;
+        }
+
+        function levenshtein(a, b) {
+            const m = a.length, n = b.length;
+            const d = Array.from({length: m + 1}, (_, i) => [i]);
+            for (let j = 1; j <= n; j++) d[0][j] = j;
+            for (let i = 1; i <= m; i++)
+                for (let j = 1; j <= n; j++)
+                    d[i][j] = Math.min(d[i-1][j] + 1, d[i][j-1] + 1, d[i-1][j-1] + (a[i-1] !== b[j-1] ? 1 : 0));
+            return d[m][n];
+        }
+
         function termExec(cmd) {
             termLog(`<span class="terminal-prompt">$</span> ${esc(cmd)}`);
             termHistory.unshift(cmd);
             termHistIdx = -1;
 
+            // Autokorjaus
+            const corrected = autocorrect(cmd.trim());
+            if (corrected && corrected !== cmd.trim()) {
+                cmd = corrected;
+                termLog(`  <span style="color:#d29922">→ korjattu: ${esc(cmd)}</span>`);
+            }
+
             const parts = cmd.trim().split(/\s+/);
             if (parts[0] !== 'kpn') {
                 termLog('kpn: tuntematon komento. Kokeile: kpn help', '#f85149');
@@ -2227,38 +2409,81 @@ Write the corrected code.`;
 
             if (sub === 'load') {
                 const arg = parts[2];
-                const btn = document.getElementById('agent-compute-btn');
-                // Mallikatalogista valinta numerolla tai nimellä
-                const loadModels = [
-                    { id: '1', key: '05b', name: 'Qwen2.5-Coder:0.5B', size: '~990 MB', coderSize: '05b' },
-                    { id: '2', key: '3b',  name: 'Qwen2.5-Coder:1.5B Q4',  size: '~1 GB', coderSize: '3b' },
+                const ollamaModels = [
+                    { id: '1', name: 'qwen2.5-coder:0.5b', size: '~400 MB', vram_mb: 0, type: 'selain + Ollama' },
+                    { id: '2', name: 'qwen2.5-coder:1.5b', size: '~1 GB', vram_mb: 1500, type: 'Ollama GPU' },
+                    { id: '3', name: 'qwen2.5-coder:7b', size: '~4.7 GB', vram_mb: 5500, type: 'Ollama GPU', default: true },
+                    { id: '4', name: 'qwen2.5-coder:14b', size: '~9 GB', vram_mb: 10000, type: 'Ollama GPU' },
+                    { id: '5', name: 'qwen2.5-coder:32b', size: '~20 GB', vram_mb: 21000, type: 'Ollama GPU' },
                 ];
                 if (!arg) {
-                    // Näytetään lista
-                    termLog('  Ladattavat mallit:', '#c9d1d9');
-                    for (const m of loadModels) {
-                        const active = (btn?.dataset.state === 'ready' && coderSize === m.coderSize) ? ' <span style="color:#3fb950">✓ ladattu</span>' : '';
-                        termLog(`    <span style="color:#58a6ff">${m.id}</span>  ${m.name} <span style="color:#8b949e">(${m.size})</span>${active}`);
+                    // Haetaan laitteistotiedot ja näytetään sopivat mallit
+                    fetch('/api/v1/hardware').then(r => r.json()).then(hw => {
+                        const vram = hw.vram_mb || 0;
+                        const ram = hw.ram_mb || 0;
+                        const gpu = hw.gpu_name || '?';
+                        const available = vram || ram; // CPU-fallback käyttää RAM:ia
+                        if (vram > 0) {
+                            termLog(`  <span style="color:#8b949e">GPU: ${gpu} | VRAM: ${Math.round(vram/1024)} GB | RAM: ${Math.round(ram/1024)} GB</span>`);
+                        } else if (ram > 0) {
+                            termLog(`  <span style="color:#8b949e">Ei GPU:ta | RAM: ${Math.round(ram/1024)} GB (CPU-moodi)</span>`);
+                        }
+                        termLog('  Mallit:', '#c9d1d9');
+                        for (const m of ollamaModels) {
+                            const fits = m.vram_mb === 0 || m.vram_mb < available;
+                            const active = m.default ? ' <span style="color:#3fb950">← aktiivinen</span>' : '';
+                            const icon = fits ? `<span style="color:#58a6ff">${m.id}</span>` : `<span style="color:#8b949e;text-decoration:line-through">${m.id}</span>`;
+                            const warn = !fits ? ' <span style="color:#f85149">⚠ ei mahdu</span>' : '';
+                            termLog(`    ${icon}  ${fits ? '' : '<span style="color:#8b949e">'}${m.name}  ${m.size} | ${m.type}${fits ? '' : '</span>'}${active}${warn}`);
                         }
                         termLog('  Käyttö: kpn load &lt;numero&gt;', '#8b949e');
+                    }).catch(() => {
+                        termLog('  Mallit:', '#c9d1d9');
+                        for (const m of ollamaModels) {
+                            const active = m.default ? ' <span style="color:#3fb950">← aktiivinen</span>' : '';
+                            termLog(`    <span style="color:#58a6ff">${m.id}</span>  ${m.name}  <span style="color:#8b949e">${m.size} | ${m.type}</span>${active}`);
+                        }
+                        termLog('  Käyttö: kpn load &lt;numero&gt;', '#8b949e');
+                    });
                     return;
                 }
-                const selected = loadModels.find(m => m.id === arg || m.key === arg || m.coderSize === arg);
+                const selected = ollamaModels.find(m => m.id === arg || m.name === arg);
                 if (!selected) {
                     termLog(`  Tuntematon malli "${esc(arg)}". Kokeile: kpn load`, '#f85149');
                     return;
                 }
-                if (btn?.dataset.state === 'ready' && coderSize === selected.coderSize) {
-                    termLog(`  ✓ ${selected.name} on jo ladattu ja valmis`, '#3fb950');
+                // Selain-WASM (vain 0.5b)
+                if (selected.id === '1') {
+                    const btn = document.getElementById('agent-compute-btn');
+                    if (btn?.dataset.state === 'ready') {
+                        termLog('  ✓ Qwen2.5-Coder:0.5B on jo ladattu (selain)', '#3fb950');
                         return;
                     }
-                coderSize = selected.coderSize;
-                localStorage.setItem('kpn-coder-size', coderSize);
-                termLog(`  Alustetaan ${selected.name} (${selected.size})...`, '#d29922');
+                    coderSize = '05b';
+                    termLog('  Ladataan Qwen2.5-Coder:0.5B selaimeen...', '#d29922');
                     if (btn) btn.click();
                     else ensureCoderNode();
                     return;
                 }
+                // Ollama: vaihdetaan malli hubin kautta
+                termLog(`  Vaihdetaan Ollama-malli: ${selected.name} (${selected.size})...`, '#d29922');
+                fetch('/api/v1/model', {
+                    method: 'POST',
+                    headers: { 'Content-Type': 'application/json' },
+                    body: JSON.stringify({ model: selected.name }),
+                }).then(r => r.json()).then(data => {
+                    if (data.status === 'ok') {
+                        termLog(`  <span style="color:#3fb950">✓</span> Malli vaihdettu: ${selected.name}`, '#3fb950');
+                        termLog('  <span style="color:#8b949e">Ollama lataa mallin ensimmäisellä pyynnöllä</span>');
+                        // Päivitetään aktiivinen default
+                        ollamaModels.forEach(m => m.default = false);
+                        selected.default = true;
+                    } else {
+                        termLog(`  ✗ Mallin vaihto epäonnistui`, '#f85149');
+                    }
+                }).catch(e => termLog(`  ✗ ${e.message}`, '#f85149'));
+                return;
+            }
 
             if (sub === 'status') {
                 const nodes = statNodes.textContent || '0';
@@ -2268,14 +2493,14 @@ Write the corrected code.`;
             }
 
             if (sub === 'models') {
-                termLog('  Käytettävissä olevat mallit:', '#c9d1d9');
-                termLog('    <span style="color:#58a6ff">1</span>  qwen-coder     Qwen2.5-Coder:0.5B  <span style="color:#8b949e">~990 MB  | koodin generointi</span>');
-                termLog('    <span style="color:#58a6ff">2</span>  qwen-coder-3b  Qwen2.5-Coder:1.5B Q4 <span style="color:#8b949e">~1 GB  | kvantisoidtu, parempi laatu</span>');
-                termLog('    <span style="color:#58a6ff">3</span>  smollm-135m    SmolLM 135M         <span style="color:#8b949e">~270 MB  | kevyt, nopea</span>');
-                termLog('    <span style="color:#58a6ff">4</span>  qwen-05b       Qwen2.5:0.5B        <span style="color:#8b949e">~990 MB  | yleismalli</span>');
-                termLog('    <span style="color:#58a6ff">5</span>  phi3-mini      Phi-3 Mini          <span style="color:#8b949e">~2.2 GB  | Microsoftin malli</span>');
-                termLog('  Käyttö: kpn run &lt;malli&gt; "&lt;prompti&gt;"', '#8b949e');
-                termLog('  Lataus: kpn load &lt;numero&gt;', '#8b949e');
+                termLog('  <span style="color:#d29922">Selain (kpn load):</span>', '#c9d1d9');
+                termLog('    qwen-coder:0.5b   <span style="color:#8b949e">~990 MB  | WASM ~0.4 tok/s</span>');
+                termLog('  <span style="color:#3fb950">Natiivi (Ollama + GPU):</span>', '#c9d1d9');
+                termLog('    qwen2.5-coder:7b  <span style="color:#8b949e">~4.7 GB  | NVIDIA ~80 tok/s | AMD ~40 tok/s | Apple ~30 tok/s</span>');
+                termLog('    qwen2.5-coder:3b  <span style="color:#8b949e">~1.9 GB  | NVIDIA ~120 tok/s</span>');
+                termLog('    qwen2.5-coder:1.5b <span style="color:#8b949e">~1 GB   | NVIDIA ~150 tok/s</span>');
+                termLog('  Vaihda malli: <span style="color:#58a6ff">OLLAMA_MODEL=qwen2.5-coder:7b</span>', '#8b949e');
+                termLog('  Hub reitittää automaattisesti nopeimmalle solmulle', '#8b949e');
                 return;
             }
 
@@ -2352,6 +2577,13 @@ Write the corrected code.`;
         };
 
         function tabComplete(input) {
+            // Autokorjaus ensin: korjaa typo ja palauta true jos korjattiin
+            const corrected = autocorrect(input.value.trim());
+            if (corrected && corrected !== input.value.trim()) {
+                input.value = corrected;
+                return true;
+            }
+
             const val = input.value;
             const words = val.trimEnd().split(/\s+/);
 
@@ -2512,7 +2744,14 @@ Write the corrected code.`;
                 }
             } else if (e.key === 'Tab') {
                 e.preventDefault();
-                // Näytä dropdown tai täydennä jos vain yksi vaihtoehto
+                // 1. Autokorjaus ensin
+                const corrected = autocorrect(termInput.value.trim());
+                if (corrected && corrected !== termInput.value.trim()) {
+                    termInput.value = corrected;
+                    hideDropdown();
+                    return;
+                }
+                // 2. Dropdown / täydennys
                 const { items, prefix } = getCandidates(termInput.value);
                 if (items.length === 1) {
                     termInput.value = prefix + items[0] + (items[0].startsWith('"') ? '' : ' ');
@@ -2551,7 +2790,8 @@ Write the corrected code.`;
         // Klikkaa terminaalipaneelia → fokusoi input
         termPanel?.addEventListener('click', () => termInput?.focus());
 
-        uiSocket.onmessage = (event) => {
+        // Tallennetaan message-handler funktioon jotta reconnect voi käyttää samaa
+        const _wsHandler = (event) => {
             try {
                 const raw = event.data;
                 if (raw.includes('"single_tokenize"')) return;
@@ -2913,6 +3153,8 @@ Write the corrected code.`;
                 }
             } catch(e) {}
         };
+        window._wsMessageHandler = _wsHandler;
+        if (uiSocket) uiSocket.onmessage = _wsHandler;
 
         btn.addEventListener('click', async () => {
             // Käytetään viewer-authissa jo tunnistettua WebGPU-tilaa