# ============================================
# [NEW] 3/6 启动内置本地模型 (llama-server)
# ============================================
echo -e "${BLUE}[3/6] 启动内置本地模型...${NC}"

LLM_DIR="$USB_PATH/llm"
LLM_PORT=18080
LLM_PID_FILE="$LLM_DIR/server.pid"
LLM_LOG="$LLM_DIR/server.log"

# Detect platform binary
OS_TYPE="$(uname -s)"
ARCH_TYPE="$(uname -m)"

case "$OS_TYPE" in
  Linux*)  LLM_BIN="$LLM_DIR/bin/llama-server-linux-x86_64" ;;
  Darwin*)
    if [ "$ARCH_TYPE" = "arm64" ]; then
      LLM_BIN="$LLM_DIR/bin/llama-server-macos-arm64"
    else
      LLM_BIN="$LLM_DIR/bin/llama-server-macos-x86_64"
    fi ;;
  *) LLM_BIN="" ;;
esac

LLM_MODEL="$LLM_DIR/models/qwen2.5-1.5b-instruct-q4_k_m.gguf"
LLM_BUNDLED_READY=0

if [ -x "$LLM_BIN" ] && [ -f "$LLM_MODEL" ]; then
  # Check if already running on port
  if ! lsof -i :$LLM_PORT -sTCP:LISTEN -t &>/dev/null 2>&1; then
    THREADS=$(( $(nproc 2>/dev/null || sysctl -n hw.logicalcpu 2>/dev/null || echo 2) - 1 ))
    THREADS=$(( THREADS < 1 ? 1 : THREADS ))
    
    chmod +x "$LLM_BIN"
    nohup "$LLM_BIN" \
      --model "$LLM_MODEL" \
      --port $LLM_PORT \
      --host 127.0.0.1 \
      --ctx-size 32768 \
      --threads $THREADS \
      --parallel 1 \
      -ngl 0 \
      --log-disable \
      >> "$LLM_LOG" 2>&1 &
    echo $! > "$LLM_PID_FILE"
    echo -e "${GREEN}✅ llama-server 已启动 (PID: $!, port: $LLM_PORT, threads: $THREADS)${NC}"
    echo -e "${YELLOW}   模型加载中，首次响应约需 5-15 秒...${NC}"
    LLM_BUNDLED_READY=1
  else
    echo -e "${GREEN}✅ 内置模型已在运行 (port: $LLM_PORT)${NC}"
    LLM_BUNDLED_READY=1
  fi
else
  echo -e "${YELLOW}⚠️  内置模型未找到，跳过 (仍可使用云端 API)${NC}"
fi

# Auto-inject bundled model config if model is available and user has no primary model set
if [ $LLM_BUNDLED_READY -eq 1 ]; then
  BUNDLED_MODEL_CONFIG=$(cat <<'JSONEOF'
{
  "models": {
    "mode": "merge",
    "providers": {
      "bundled-local": {
        "baseUrl": "http://127.0.0.1:18080/v1",
        "apiKey": "bundled-no-key",
        "api": "openai-completions",
        "models": [
          {
            "id": "qwen2.5-1.5b",
            "name": "Qwen2.5 1.5B (Bundled CPU)",
            "contextWindow": 32768,
            "maxTokens": 4096,
            "cost": { "input": 0, "output": 0 }
          }
        ]
      }
    }
  }
}
JSONEOF
  )
  
  # Use existing install-models.js merge mechanism
  echo "$BUNDLED_MODEL_CONFIG" > "$TEMP_DIR/bundled-model-inject.json"
  
  # Only set as default if user has NOT configured a primary model
  HAS_PRIMARY=$(node -e "
    try {
      const cfg = JSON.parse(require('fs').readFileSync('$TEMP_DIR/openclaw.json','utf8'));
      console.log(cfg?.agents?.defaults?.model?.primary ? 'yes' : 'no');
    } catch(e) { console.log('no'); }
  " 2>/dev/null || echo "no")
  
  if [ "$HAS_PRIMARY" = "no" ]; then
    # Merge and set as default
    node -e "
      const fs = require('fs');
      const cfgPath = '$TEMP_DIR/openclaw.json';
      const cfg = fs.existsSync(cfgPath) ? JSON.parse(fs.readFileSync(cfgPath,'utf8')) : {};
      const inject = JSON.parse(fs.readFileSync('$TEMP_DIR/bundled-model-inject.json','utf8'));
      
      // Deep merge providers
      cfg.models = cfg.models || {};
      cfg.models.providers = Object.assign({}, cfg.models.providers, inject.models.providers);
      cfg.models.mode = 'merge';
      
      // Set as default only if no primary configured
      cfg.agents = cfg.agents || {};
      cfg.agents.defaults = cfg.agents.defaults || {};
      cfg.agents.defaults.model = cfg.agents.defaults.model || {};
      cfg.agents.defaults.model.primary = 'bundled-local/qwen2.5-1.5b';
      
      fs.writeFileSync(cfgPath, JSON.stringify(cfg, null, 2));
      console.log('✅ 内置模型已设为默认模型');
    " 2>/dev/null && echo -e "${GREEN}   bundled-local/qwen2.5-1.5b 已注册为默认模型${NC}"
    rm -f "$TEMP_DIR/bundled-model-inject.json"
  else
    echo -e "${CYAN}   检测到已配置主模型，内置模型作为备用 (fallback)${NC}"
  fi
fi

# Kill bundled llama-server
LLM_PID_FILE="$USB_PATH/llm/server.pid"
if [ -f "$LLM_PID_FILE" ]; then
  LLM_PID=$(cat "$LLM_PID_FILE")
  if kill -0 "$LLM_PID" 2>/dev/null; then
    kill "$LLM_PID"
    echo -e "${GREEN}✅ 内置模型已停止 (PID: $LLM_PID)${NC}"
  fi
  rm -f "$LLM_PID_FILE"
fi

# In build-offline-package.sh, add:
download_bundled_model() {
  MODEL_URL="https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct-GGUF/resolve/main/qwen2.5-1.5b-instruct-q4_k_m.gguf"
  MODEL_DIR="$SCRIPT_DIR/llm/models"
  mkdir -p "$MODEL_DIR"
  if [ ! -f "$MODEL_DIR/qwen2.5-1.5b-instruct-q4_k_m.gguf" ]; then
    echo "Downloading bundled model (~900MB)..."
    curl -L --progress-bar -o "$MODEL_DIR/qwen2.5-1.5b-instruct-q4_k_m.gguf" "$MODEL_URL"
  fi
}