worker/checks/calls/init.go

// Package calls provides functionality.
package calls

import (
	"bytes"
	"context"
	"encoding/base64"
	"encoding/json"
	"errors"
	"fmt"
	"io"
	"log"
	"os"
	"strings"
	"time"

	"github.com/chromedp/cdproto/network"
	"github.com/chromedp/cdproto/runtime"
	"github.com/chromedp/chromedp"
	"github.com/google/uuid"
	"github.com/minio/minio-go/v7"
	"github.com/openai/openai-go/v3"
	"github.com/openai/openai-go/v3/option"

	"rsgit.ru/rsmon/rsmon/app/models"
	"rsgit.ru/rsmon/rsmon/internal/checkresult"
	"rsgit.ru/rsmon/rsmon/storage"
)

const (
	verdictError   = "error"
	verdictWarning = "warning"
	verdictNormal  = "normal"
)

// Result is a result of an AI health check
type Result struct {
	checkresult.CheckResult

	// Captured data
	Screenshot  string `json:"screenshot"`   // base64 encoded screenshot
	DOM         string `json:"dom"`          // HTML content
	NetworkLogs string `json:"network_logs"` // JSON array of network logs
	ConsoleLogs string `json:"console_logs"` // JSON array of console logs
	URL         string `json:"url"`          // Final URL after redirects
	StatusCode  int    `json:"status_code"`  // HTTP status code
	Title       string `json:"title"`        // Page title

	// LLM analysis
	LLMResponse  string `json:"llm_response"`  // Full LLM response
	LLMVerdict   string `json:"llm_verdict"`   // "normal", "error", "warning"
	LLMReasoning string `json:"llm_reasoning"` // LLM's explanation
	HasChanges   bool   `json:"has_changes"`   // Whether changes were detected from previous

	// Previous screenshot comparison
	PreviousURL  string `json:"previous_url"`  // URL of previous screenshot (if any)
	PreviousTime string `json:"previous_time"` // Timestamp of previous screenshot

	// Metadata
	CheckID      uint `json:"check_id"`      // For database reference
	ScreenshotID uint `json:"screenshot_id"` // ID of stored screenshot
}

// NetworkLog represents a network request log
type NetworkLog struct {
	URL        string  `json:"url"`
	Method     string  `json:"method"`
	StatusCode int     `json:"status_code"`
	Type       string  `json:"type"`
	Size       int64   `json:"size"`
	Time       float64 `json:"time"`
}

// ConsoleLog represents a console log entry
type ConsoleLog struct {
	Type  string `json:"type"`
	Value string `json:"value"`
}

// ScreenshotRecord represents a stored screenshot in the database
type ScreenshotRecord struct {
	ID        uint      `gorm:"primarykey"`
	CreatedAt time.Time `json:"created_at"`
	CheckID   uint      `json:"check_id"`
	MonitorID uint      `json:"monitor_id"`
	URL       string    `json:"url"`
	// ScreenshotPath stores the S3 object path (e.g., "screenshots/monitor_123/2025-01-15_123456.jpg")

	ScreenshotPath string `json:"screenshot_path" gorm:"type:varchar(500)"`
	DOM            string `gorm:"type:longtext" json:"dom"`
	NetworkLogs    string `gorm:"type:longtext" json:"network_logs"`
	ConsoleLogs    string `gorm:"type:longtext" json:"console_logs"`
	StatusCode     int    `json:"status_code"`
	Title          string `json:"title"`
	LLMVerdict     string `json:"llm_verdict"`
	LLMReasoning   string `json:"llm_reasoning"`
}

// Perform executes the AI health check
func Perform(c *models.Check) *Result {
	result := &Result{
		CheckID: uint(c.ID),
	}
	result.Warnings = []string{}
	result.Infos = []string{}

	start := time.Now()

	// Get the target URL
	targetURL := getTargetURL(c)
	result.URL = targetURL

	// Step 1: Capture screenshot, DOM, and logs
	screenshotData, domData, networkLogs, consoleLogs, statusCode, title, err := captureWebData(targetURL)
	if err != nil {
		result.Error = fmt.Errorf("failed to capture web data: %w", err)
		result.Duration = time.Since(start)
		result.State = "FAIL"
		return result
	}

	result.Screenshot = screenshotData
	result.DOM = domData
	result.NetworkLogs = networkLogs
	result.ConsoleLogs = consoleLogs
	result.StatusCode = statusCode
	result.Title = title

	// Step 2: Store current screenshot in database (if available)
	var screenshotID uint
	if isDBAvailable() {
		screenshotID, err = storeScreenshot(uint(c.MonitorID), uint(c.ID), result)
		if err != nil {
			result.Warnings = append(result.Warnings, "Failed to store screenshot: "+err.Error())
			log.Printf("[ai-check] Failed to store screenshot: %v", err)
		} else {
			result.ScreenshotID = screenshotID
			result.Infos = append(result.Infos, "Screenshot stored for reference")
		}
	} else {
		result.Infos = append(result.Infos, "Database not available - running in standalone mode")
	}

	// Step 3: Get previous screenshots for comparison (if DB available)
	var previousScreenshots []ScreenshotRecord
	if isDBAvailable() {
		previousScreenshots, err = getPreviousScreenshots(uint(c.MonitorID), 3)
		if err != nil {
			result.Warnings = append(result.Warnings, "Could not retrieve previous screenshots: "+err.Error())
		}
	}

	if len(previousScreenshots) > 0 {
		result.PreviousURL = previousScreenshots[0].URL
		result.PreviousTime = previousScreenshots[0].CreatedAt.Format(time.RFC3339)
		result.Infos = append(result.Infos, fmt.Sprintf("Comparing with %d previous screenshot(s)", len(previousScreenshots)))
	}

	// Step 4: Analyze with LLM
	llmVerdict, llmReasoning, hasChanges, err := analyzeWithLLM(targetURL, result, previousScreenshots)
	if err != nil {
		result.Error = fmt.Errorf("LLM analysis failed: %w", err)
		result.LLMResponse = err.Error()
		result.Duration = time.Since(start)
		result.State = "FAIL"
		return result
	}

	result.LLMVerdict = llmVerdict
	result.LLMReasoning = llmReasoning
	result.HasChanges = hasChanges

	// Update the screenshot record with LLM analysis
	if screenshotID > 0 {
		_ = updateScreenshotWithLLM(screenshotID, llmVerdict, llmReasoning)
	}

	// Determine final state based on LLM verdict
	switch llmVerdict {
	case verdictError:
		result.State = "ERR"
		result.Error = errors.New(llmReasoning)
	case verdictWarning:
		result.State = "WARN"
		result.Warnings = append(result.Warnings, llmReasoning)
	default:
		result.State = "OK"
		if hasChanges {
			result.Infos = append(result.Infos, "Visual changes detected from previous screenshot")
		}
	}

	result.Duration = time.Since(start)
	return result
}

// getTargetURL constructs the target URL from the check
func getTargetURL(c *models.Check) string {
	if c.URL != nil && *c.URL != "" {
		return *c.URL
	}

	// Construct from monitor host
	host := c.Monitor.Host
	if !strings.HasPrefix(host, "http://") && !strings.HasPrefix(host, "https://") {
		return "https://" + host
	}
	return host
}

// captureWebData captures screenshot, DOM, network and console logs using chromedp
func captureWebData(targetURL string) (screenshot, dom, networkLogsJSON, consoleLogsJSON string, statusCode int, title string, err error) {
	// Create a context with timeout
	ctx, cancel := context.WithTimeout(context.Background(), 30*time.Second)
	defer cancel()

	// Create chromedp context with options for headless chrome
	allocCtx, allocCancel := chromedp.NewExecAllocator(ctx,
		chromedp.NoDefaultBrowserCheck,
		chromedp.NoFirstRun,
		chromedp.DisableGPU,
		chromedp.IgnoreCertErrors,
		chromedp.Flag("headless", "new"),
		chromedp.Flag("no-sandbox", true),
		chromedp.Flag("disable-dev-shm-usage", true),
		chromedp.Flag("hide-scrollbars", true),
		chromedp.Flag("mute-audio", true),
		chromedp.WindowSize(1920, 1080),
	)
	defer allocCancel()

	taskCtx, taskCancel := chromedp.NewContext(allocCtx)
	defer taskCancel()

	// Collect network logs
	networkLogs := []NetworkLog{}
	consoleLogs := []ConsoleLog{}

	var finalStatusCode int

	// Listen for network events
	chromedp.ListenTarget(taskCtx, func(ev interface{}) {
		switch e := ev.(type) {
		case *network.EventRequestWillBeSent:
			// Track the main request
			if e.Type == network.ResourceTypeDocument {
				networkLogs = append(networkLogs, NetworkLog{
					URL:    e.Request.URL,
					Method: e.Request.Method,
					Type:   string(e.Type),
				})
			}
		case *network.EventResponseReceived:
			// Update the main request with status code
			for i, log := range networkLogs {
				if log.URL == e.Response.URL && log.Method == "" { // Not yet updated
					networkLogs[i].StatusCode = int(e.Response.Status)
					networkLogs[i].Size = int64(e.Response.EncodedDataLength)

					// Capture final status code for main document
					if e.Type == network.ResourceTypeDocument && finalStatusCode == 0 {
						finalStatusCode = int(e.Response.Status)
					}
					break
				}
			}
		case *runtime.EventConsoleAPICalled:
			// Capture console logs
			value := ""
			for _, arg := range e.Args {
				// Get the string representation of the argument
				value += arg.Value.String()
			}
			if value != "" && value != "undefined" {
				consoleLogs = append(consoleLogs, ConsoleLog{
					Type:  "console." + string(e.Type),
					Value: strings.TrimSpace(value),
				})
			}
		}
	})

	// Run the browser tasks
	var buf []byte
	var html string
	var pageTitle string

	tasks := chromedp.Tasks{
		network.Enable(),
		chromedp.Navigate(targetURL),
		chromedp.WaitReady("body", chromedp.ByQuery),
		chromedp.Sleep(2 * time.Second), // Wait for page to fully render
		chromedp.Title(&pageTitle),
		chromedp.OuterHTML(":root", &html, chromedp.ByQueryAll),
		chromedp.FullScreenshot(&buf, 90), // 90% quality JPEG
	}

	if err := chromedp.Run(taskCtx, tasks); err != nil {
		// Check if it's a timeout error
		if errors.Is(err, context.DeadlineExceeded) {
			return "", "", "", "", 0, "", fmt.Errorf("page load timeout after 30s")
		}
		return "", "", "", "", 0, "", fmt.Errorf("chromedp error: %w", err)
	}

	// Encode screenshot as base64
	screenshot = base64.StdEncoding.EncodeToString(buf)
	dom = html

	// Convert logs to JSON
	networkBytes, _ := json.Marshal(networkLogs)
	networkLogsJSON = string(networkBytes)

	consoleBytes, _ := json.Marshal(consoleLogs)
	consoleLogsJSON = string(consoleBytes)

	// Set status code if we got it from network events
	if finalStatusCode == 0 {
		finalStatusCode = 200 // Assume OK if we got this far
	}

	return screenshot, dom, networkLogsJSON, consoleLogsJSON, finalStatusCode, pageTitle, nil
}

// storeScreenshot stores the screenshot data in S3 and metadata in the database
func storeScreenshot(monitorID, checkID uint, result *Result) (uint, error) {
	// Initialize storage if not already initialized
	if !storage.IsAvailable() {
		if err := storage.Init(); err != nil {
			return 0, fmt.Errorf("failed to initialize storage: %w", err)
		}
	}

	// Generate a unique filename for the screenshot
	// Format: screenshots/monitor_<monitor_id>/YYYY-MM-DD/<uuid>.jpg
	timestamp := time.Now().Format("2006-01-02")
	screenshotID := uuid.New().String()
	objectPath := fmt.Sprintf("screenshots/monitor_%d/%s/%s.jpg", monitorID, timestamp, screenshotID)

	// Decode base64 screenshot data
	screenshotBytes, err := base64.StdEncoding.DecodeString(result.Screenshot)
	if err != nil {
		return 0, fmt.Errorf("failed to decode screenshot: %w", err)
	}

	// Upload screenshot to S3
	ctx := context.Background()
	reader := bytes.NewReader(screenshotBytes)
	objectSize := int64(len(screenshotBytes))

	_, err = storage.PutObject(ctx, objectPath, reader, objectSize, minio.PutObjectOptions{
		ContentType: "image/jpeg",
	})
	if err != nil {
		return 0, fmt.Errorf("failed to upload screenshot to S3: %w", err)
	}

	log.Printf("[ai-check] Stored screenshot at S3 path: %s (size: %d bytes)", objectPath, objectSize)

	// Store metadata in database
	record := &ScreenshotRecord{
		MonitorID:      monitorID,
		CheckID:        checkID,
		URL:            result.URL,
		ScreenshotPath: objectPath,
		DOM:            truncateString(result.DOM, 50000), // Limit DOM size
		NetworkLogs:    result.NetworkLogs,
		ConsoleLogs:    result.ConsoleLogs,
		StatusCode:     result.StatusCode,
		Title:          result.Title,
	}

	// Auto migrate the table
	if err := models.DB().AutoMigrate(&ScreenshotRecord{}); err != nil {
		return 0, fmt.Errorf("failed to auto-migrate: %w", err)
	}

	if err := models.DB().Create(record).Error; err != nil {
		return 0, fmt.Errorf("failed to insert screenshot metadata: %w", err)
	}

	return record.ID, nil
}

// loadScreenshotData loads screenshot data from S3 for a given record
func loadScreenshotData(record *ScreenshotRecord) (string, error) {
	if record.ScreenshotPath == "" {
		return "", errors.New("no screenshot path in record")
	}

	if !storage.IsAvailable() {
		if err := storage.Init(); err != nil {
			return "", fmt.Errorf("failed to initialize storage: %w", err)
		}
	}

	ctx := context.Background()
	object, err := storage.GetObject(ctx, record.ScreenshotPath, minio.GetObjectOptions{})
	if err != nil {
		return "", fmt.Errorf("failed to get screenshot from S3: %w", err)
	}
	defer object.Close() //nolint:errcheck // accepted lint exception

	// Read the object data
	var buf bytes.Buffer
	_, err = io.Copy(&buf, object)
	if err != nil {
		return "", fmt.Errorf("failed to read screenshot data: %w", err)
	}

	// Encode as base64 for LLM analysis
	return base64.StdEncoding.EncodeToString(buf.Bytes()), nil
}

// updateScreenshotWithLLM updates a screenshot record with LLM analysis
func updateScreenshotWithLLM(screenshotID uint, verdict, reasoning string) error {
	return models.DB().Model(&ScreenshotRecord{}).
		Where("id = ?", screenshotID).
		Updates(map[string]interface{}{
			"llm_verdict":   verdict,
			"llm_reasoning": reasoning,
		}).Error
}

// isDBAvailable checks if the database is available
func isDBAvailable() bool {
	return models.IsDBAvailable()
}

// getPreviousScreenshots retrieves previous screenshots for comparison
func getPreviousScreenshots(monitorID uint, limit int) ([]ScreenshotRecord, error) {
	var records []ScreenshotRecord

	err := models.DB().Model(&ScreenshotRecord{}).
		Where("monitor_id = ? AND llm_verdict != ''", monitorID).
		Order("created_at DESC").
		Limit(limit).
		Find(&records).Error
	if err != nil {
		return nil, err
	}

	return records, nil
}

// truncateString truncates a string to a maximum length
func truncateString(s string, maxLen int) string {
	if len(s) <= maxLen {
		return s
	}
	return s[:maxLen] + "... (truncated)"
}

// analyzeWithLLM sends the screenshot and data to the LLM for analysis
func analyzeWithLLM(targetURL string, current *Result, previous []ScreenshotRecord) (verdict, reasoning string, hasChanges bool, err error) { //nolint:lll
	// Get LLM configuration from environment
	llmKey := firstEnv("LLM_APIKEY", "LLAMA_KEY")
	llmURL := firstEnv("LLM_URL", "LLAMA_URL")
	llmModel := firstEnv("LLM_MODEL", "LLAMA_MODEL")

	if llmKey == "" || llmURL == "" {
		return "", "", false, errors.New("LLM credentials not configured (LLAMA_KEY, LLAMA_URL)")
	}

	if llmModel == "" {
		llmModel = "llama3.2-vision" // Default model
	}

	// Create OpenAI client with custom base URL
	client := openai.NewClient(
		option.WithBaseURL(llmURL),
		option.WithAPIKey(llmKey),
	)

	// Build the system prompt
	systemPrompt := buildSystemPrompt(len(previous) > 0)

	// Build user message with current screenshot
	userContent := buildUserMessage(targetURL, current, previous)

	// Prepare messages
	messages := []openai.ChatCompletionMessageParamUnion{
		openai.SystemMessage(systemPrompt),
		openai.UserMessage(userContent),
	}

	// Call LLM with timeout
	ctx, cancel := context.WithTimeout(context.Background(), 60*time.Second)
	defer cancel()

	params := openai.ChatCompletionNewParams{
		Messages:    messages,
		Model:       llmModel,
		MaxTokens:   openai.Int(2000),
		Temperature: openai.Float(0.3), // Lower temperature for more consistent analysis
	}

	completion, err := client.Chat.Completions.New(ctx, params)
	if err != nil {
		return "", "", false, fmt.Errorf("LLM request failed: %w", err)
	}

	if len(completion.Choices) == 0 {
		return "", "", false, errors.New("LLM returned no choices")
	}

	response := completion.Choices[0].Message.Content

	// Parse the LLM response to extract verdict and reasoning
	return parseLLMResponse(response)
}

func firstEnv(keys ...string) string {
	for _, key := range keys {
		if value := os.Getenv(key); value != "" {
			return value
		}
	}
	return ""
}

// buildSystemPrompt creates the system prompt for the LLM
func buildSystemPrompt(hasPrevious bool) string {
	basePrompt := `
You are a web application health monitoring assistant. Your task is to analyze website screenshots
and determine if the website appears to be functioning normally or if there are issues.


Consider the following aspects:
1. Visual layout - Is the page rendered correctly? Are elements properly aligned?
2. Error messages - Are there any visible error messages, 404s, 500s, or similar?
3. Loading states - Is the page stuck loading or showing spinners?
4. Broken elements - Are there broken images, missing styles, or layout shifts?
5. Content availability - Is the expected content visible?

Respond in the following JSON format:
{
  "verdict": "normal" | "warning" | "error",
  "reasoning": "Brief explanation of your assessment",
  "has_changes": true | false
}

Verdict guidelines:
- "error": Page is clearly broken (error messages, completely failed to load, blank page)
- "warning": Page loads but has issues (partial content, broken elements, degraded appearance)
- "normal": Page appears to be functioning correctly`

	if hasPrevious {
		basePrompt += `

You will also be shown previous screenshots of the same website. Compare the current screenshot with previous ones and report any significant visual changes that might indicate problems (layout shifts, missing elements, color changes, etc.).` //nolint:lll
	}

	return basePrompt
}

// buildUserMessage creates the user message with screenshot data
func buildUserMessage(targetURL string, current *Result, previous []ScreenshotRecord) string {
	var sb strings.Builder

	_, _ = fmt.Fprintf(&sb, "Analyze the screenshot of: %s\n\n", targetURL)
	_, _ = fmt.Fprintf(&sb, "HTTP Status Code: %d\n", current.StatusCode)
	_, _ = fmt.Fprintf(&sb, "Page Title: %s\n\n", current.Title)

	// Add console errors if any
	var consoleErrors []ConsoleLog
	json.Unmarshal([]byte(current.ConsoleLogs), &consoleErrors) //nolint:errcheck // accepted lint exception
	errorCount := 0
	for _, log := range consoleErrors {
		if strings.Contains(log.Type, "error") || strings.Contains(log.Type, "warn") {
			errorCount++
		}
	}
	if errorCount > 0 {
		_, _ = fmt.Fprintf(&sb, "Console Errors/Warnings: %d\n\n", errorCount)
	}

	// Add network errors
	var networkLogs []NetworkLog
	json.Unmarshal([]byte(current.NetworkLogs), &networkLogs) //nolint:errcheck // accepted lint exception
	failedRequests := 0
	for _, req := range networkLogs {
		if req.StatusCode >= 400 {
			failedRequests++
		}
	}
	if failedRequests > 0 {
		_, _ = fmt.Fprintf(&sb, "Failed Network Requests: %d\n\n", failedRequests)
	}

	sb.WriteString("Current screenshot (base64):\n")
	sb.WriteString(current.Screenshot)
	sb.WriteString("\n\n")

	// Add previous screenshots if available (load from S3)
	if len(previous) > 0 {
		_, _ = fmt.Fprintf(&sb, "For comparison, here are %d previous screenshot(s):\n\n", len(previous))
		for i, prev := range previous { //nolint:gocritic // range copy is acceptable here
			// Load screenshot data from S3
			screenshotData, err := loadScreenshotData(&prev)
			if err != nil {
				log.Printf("[ai-check] Failed to load previous screenshot %d: %v", i+1, err)
				_, _ = fmt.Fprintf(&sb, "Previous screenshot #%d (from %s): [error loading screenshot]\n\n",
					i+1, prev.CreatedAt.Format("2006-01-02 15:04:05"))
				continue
			}

			_, _ = fmt.Fprintf(&sb, "Previous screenshot #%d (from %s):\n", i+1, prev.CreatedAt.Format("2006-01-02 15:04:05"))
			sb.WriteString(screenshotData)
			sb.WriteString("\n\n")
		}
	}

	return sb.String()
}

// parseLLMResponse parses the LLM response to extract structured data
func parseLLMResponse(response string) (verdict, reasoning string, hasChanges bool, err error) {
	// Try to extract JSON from the response
	response = strings.TrimSpace(response)

	// Look for JSON block
	jsonStart := strings.Index(response, "{")
	jsonEnd := strings.LastIndex(response, "}")

	if jsonStart == -1 || jsonEnd == -1 {
		// No JSON found, try to parse text response
		return parseTextResponse(response)
	}

	jsonStr := response[jsonStart : jsonEnd+1]

	var parsed struct {
		Verdict    string `json:"verdict"`
		Reasoning  string `json:"reasoning"`
		HasChanges bool   `json:"has_changes"`
	}

	if err := json.Unmarshal([]byte(jsonStr), &parsed); err != nil {
		// JSON parse failed, fall back to text parsing
		return parseTextResponse(response)
	}

	// Validate verdict
	switch parsed.Verdict {
	case "normal", "ok", "healthy":
		return verdictNormal, parsed.Reasoning, parsed.HasChanges, nil
	case "warning", "warn":
		return verdictWarning, parsed.Reasoning, parsed.HasChanges, nil
	case "error", "fail", "unhealthy":
		return verdictError, parsed.Reasoning, parsed.HasChanges, nil
	default:
		// Unknown verdict, default to normal with warning
		if parsed.Verdict != "" {
			return verdictWarning, parsed.Reasoning, parsed.HasChanges, nil
		}
		return verdictNormal, "Unable to determine specific issues from analysis", false, nil
	}
}

// parseTextResponse parses a non-JSON response
func parseTextResponse(response string) (verdict, reasoning string, hasChanges bool, err error) {
	responseLower := strings.ToLower(response)

	// Look for keywords
	hasError := strings.Contains(responseLower, "error") ||
		strings.Contains(responseLower, "broken") ||
		strings.Contains(responseLower, "failed") ||
		strings.Contains(responseLower, "not working")

	hasWarning := strings.Contains(responseLower, "warning") ||
		strings.Contains(responseLower, "issue") ||
		strings.Contains(responseLower, "problem") ||
		strings.Contains(responseLower, "degraded")

	hasChange := strings.Contains(responseLower, "change") ||
		strings.Contains(responseLower, "different") ||
		strings.Contains(responseLower, "modified")

	if hasError {
		return verdictError, response, hasChange, nil
	} else if hasWarning {
		return verdictWarning, response, hasChange, nil
	}

	return verdictNormal, response, hasChange, nil
}