nannyagent/agent.go

package main

import (
	"bytes"
	"encoding/json"
	"fmt"
	"io"
	"net/http"
	"os"
	"strings"
	"time"

	"nannyagentv2/internal/ebpf"
	"nannyagentv2/internal/executor"
	"nannyagentv2/internal/logging"
	"nannyagentv2/internal/system"
	"nannyagentv2/internal/types"

	"github.com/sashabaranov/go-openai"
)

// AgentConfig holds configuration for concurrent execution (local to agent)
type AgentConfig struct {
	MaxConcurrentTasks int  `json:"max_concurrent_tasks"`
	CollectiveResults  bool `json:"collective_results"`
}

// DefaultAgentConfig returns default configuration
func DefaultAgentConfig() *AgentConfig {
	return &AgentConfig{
		MaxConcurrentTasks: 10,   // Default to 10 concurrent forks
		CollectiveResults:  true, // Send results collectively when all finish
	}
}

//
// LinuxDiagnosticAgent represents the main diagnostic agent

// LinuxDiagnosticAgent represents the main diagnostic agent
type LinuxDiagnosticAgent struct {
	client      *openai.Client
	model       string
	executor    *executor.CommandExecutor
	episodeID   string                // TensorZero episode ID for conversation continuity
	ebpfManager *ebpf.BCCTraceManager // eBPF tracing manager
	config      *AgentConfig          // Configuration for concurrent execution
	authManager interface{}           // Authentication manager for TensorZero requests
	logger      *logging.Logger
}

// NewLinuxDiagnosticAgent creates a new diagnostic agent
func NewLinuxDiagnosticAgent() *LinuxDiagnosticAgent {
	// Get Supabase project URL for TensorZero proxy
	supabaseURL := os.Getenv("SUPABASE_PROJECT_URL")
	if supabaseURL == "" {
		logging.Warning("SUPABASE_PROJECT_URL not set, TensorZero integration will not work")
		supabaseURL = "https://gpqzsricripnvbrpsyws.supabase.co" // fallback
	}

	model := os.Getenv("NANNYAPI_MODEL")
	if model == "" {
		model = "tensorzero::function_name::diagnose_and_heal"
		logging.Warning("Using default model '%s'. Set NANNYAPI_MODEL environment variable for your specific function", model)
	}

	// Note: We don't use the OpenAI client anymore, we use direct HTTP to Supabase proxy
	agent := &LinuxDiagnosticAgent{
		client:   nil, // Not used anymore
		model:    model,
		executor: executor.NewCommandExecutor(10 * time.Second), // 10 second timeout for commands
		config:   DefaultAgentConfig(),                          // Default concurrent execution config
	}

	// Initialize eBPF manager
	agent.ebpfManager = ebpf.NewBCCTraceManager()
	agent.logger = logging.NewLogger()

	return agent
}

// NewLinuxDiagnosticAgentWithAuth creates a new diagnostic agent with authentication
func NewLinuxDiagnosticAgentWithAuth(authManager interface{}) *LinuxDiagnosticAgent {
	// Get Supabase project URL for TensorZero proxy
	supabaseURL := os.Getenv("SUPABASE_PROJECT_URL")
	if supabaseURL == "" {
		logging.Warning("SUPABASE_PROJECT_URL not set, TensorZero integration will not work")
		supabaseURL = "https://gpqzsricripnvbrpsyws.supabase.co" // fallback
	}

	model := os.Getenv("NANNYAPI_MODEL")
	if model == "" {
		model = "tensorzero::function_name::diagnose_and_heal"
		logging.Warning("Using default model '%s'. Set NANNYAPI_MODEL environment variable for your specific function", model)
	}

	// Note: We don't use the OpenAI client anymore, we use direct HTTP to Supabase proxy
	agent := &LinuxDiagnosticAgent{
		client:      nil, // Not used anymore
		model:       model,
		executor:    executor.NewCommandExecutor(10 * time.Second), // 10 second timeout for commands
		config:      DefaultAgentConfig(),                          // Default concurrent execution config
		authManager: authManager,                                   // Store auth manager for TensorZero requests
	}

	// Initialize eBPF manager
	agent.ebpfManager = ebpf.NewBCCTraceManager()
	agent.logger = logging.NewLogger()

	return agent
}

// DiagnoseIssue starts the diagnostic process for a given issue
func (a *LinuxDiagnosticAgent) DiagnoseIssue(issue string) error {
	logging.Info("Diagnosing issue: %s", issue)
	logging.Info("Gathering system information...")

	// Gather system information
	systemInfo := system.GatherSystemInfo()

	// Format the initial prompt with system information
	initialPrompt := system.FormatSystemInfoForPrompt(systemInfo) + "\n" + issue

	// Start conversation with initial issue including system info
	messages := []openai.ChatCompletionMessage{
		{
			Role:    openai.ChatMessageRoleUser,
			Content: initialPrompt,
		},
	}

	for {
		// Send request to TensorZero API via OpenAI SDK
		response, err := a.SendRequestWithEpisode(messages, a.episodeID)
		if err != nil {
			return fmt.Errorf("failed to send request: %w", err)
		}

		if len(response.Choices) == 0 {
			return fmt.Errorf("no choices in response")
		}

		content := response.Choices[0].Message.Content
		logging.Debug("AI Response: %s", content)

		// Parse the response to determine next action
		var diagnosticResp types.EBPFEnhancedDiagnosticResponse
		var resolutionResp types.ResolutionResponse

		// Try to parse as diagnostic response first (with eBPF support)
		logging.Debug("Attempting to parse response as diagnostic...")
		if err := json.Unmarshal([]byte(content), &diagnosticResp); err == nil && diagnosticResp.ResponseType == "diagnostic" {
			logging.Debug("Successfully parsed as diagnostic response with %d commands", len(diagnosticResp.Commands))
			// Handle diagnostic phase
			logging.Debug("Reasoning: %s", diagnosticResp.Reasoning)

			// Execute commands and collect results
			commandResults := make([]types.CommandResult, 0, len(diagnosticResp.Commands))
			if len(diagnosticResp.Commands) > 0 {
				logging.Info("Executing %d diagnostic commands", len(diagnosticResp.Commands))
				for i, cmdStr := range diagnosticResp.Commands {
					// Convert string command to Command struct (auto-generate ID and description)
					cmd := types.Command{
						ID:          fmt.Sprintf("cmd_%d", i+1),
						Command:     cmdStr,
						Description: fmt.Sprintf("Diagnostic command: %s", cmdStr),
					}
					result := a.executor.Execute(cmd)
					commandResults = append(commandResults, result)

					if result.ExitCode != 0 {
						logging.Warning("Command '%s' failed with exit code %d", cmd.ID, result.ExitCode)
					}
				}
			}

			// Execute eBPF programs if present - support both old and new formats
			var ebpfResults []map[string]interface{}
			if len(diagnosticResp.EBPFPrograms) > 0 {
				logging.Info("AI requested %d eBPF traces for enhanced diagnostics", len(diagnosticResp.EBPFPrograms))

				// Convert EBPFPrograms to TraceSpecs and execute concurrently using the eBPF service
				traceSpecs := a.ConvertEBPFProgramsToTraceSpecs(diagnosticResp.EBPFPrograms)
				ebpfResults = a.ExecuteEBPFTraces(traceSpecs)
			}

			// Prepare combined results as user message
			allResults := map[string]interface{}{
				"command_results":   commandResults,
				"executed_commands": len(commandResults),
			}

			// Include eBPF results if any were executed
			if len(ebpfResults) > 0 {
				allResults["ebpf_results"] = ebpfResults
				allResults["executed_ebpf_programs"] = len(ebpfResults)

				// Extract evidence summary for TensorZero
				evidenceSummary := make([]string, 0)
				for _, result := range ebpfResults {
					target := result["target"]
					eventCount := result["event_count"]
					summary := result["summary"]
					success := result["success"]

					status := "failed"
					if success == true {
						status = "success"
					}

					summaryStr := fmt.Sprintf("%s: %v events (%s) - %s", target, eventCount, status, summary)
					evidenceSummary = append(evidenceSummary, summaryStr)
				}
				allResults["ebpf_evidence_summary"] = evidenceSummary
			}

			resultsJSON, err := json.MarshalIndent(allResults, "", "  ")
			if err != nil {
				return fmt.Errorf("failed to marshal command results: %w", err)
			}

			// Add AI response and command results to conversation
			messages = append(messages, openai.ChatCompletionMessage{
				Role:    openai.ChatMessageRoleAssistant,
				Content: content,
			})
			messages = append(messages, openai.ChatCompletionMessage{
				Role:    openai.ChatMessageRoleUser,
				Content: string(resultsJSON),
			})

			continue
		} else {
			logging.Debug("Failed to parse as diagnostic. Error: %v, ResponseType: '%s'", err, diagnosticResp.ResponseType)
		}

		// Try to parse as resolution response
		if err := json.Unmarshal([]byte(content), &resolutionResp); err == nil && resolutionResp.ResponseType == "resolution" {
			// Handle resolution phase
			logging.Info("=== DIAGNOSIS COMPLETE ===")
			logging.Info("Root Cause: %s", resolutionResp.RootCause)
			logging.Info("Resolution Plan: %s", resolutionResp.ResolutionPlan)
			logging.Info("Confidence: %s", resolutionResp.Confidence)
			break
		}

		// If we can't parse the response, treat it as an error or unexpected format
		logging.Error("Unexpected response format or error from AI: %s", content)
		break
	}

	return nil
}

// sendRequest sends a request to TensorZero via Supabase proxy (without episode ID)
func (a *LinuxDiagnosticAgent) SendRequest(messages []openai.ChatCompletionMessage) (*openai.ChatCompletionResponse, error) {
	return a.SendRequestWithEpisode(messages, "")
}

// ExecuteCommand executes a command using the agent's executor
func (a *LinuxDiagnosticAgent) ExecuteCommand(cmd types.Command) types.CommandResult {
	return a.executor.Execute(cmd)
}

// sendRequestWithEpisode sends a request to TensorZero via Supabase proxy with episode ID for conversation continuity
func (a *LinuxDiagnosticAgent) SendRequestWithEpisode(messages []openai.ChatCompletionMessage, episodeID string) (*openai.ChatCompletionResponse, error) {
	// Convert messages to the expected format
	messageMaps := make([]map[string]interface{}, len(messages))
	for i, msg := range messages {
		messageMaps[i] = map[string]interface{}{
			"role":    msg.Role,
			"content": msg.Content,
		}
	}

	// Create TensorZero request
	tzRequest := map[string]interface{}{
		"model":    a.model,
		"messages": messageMaps,
	}

	// Add episode ID if provided
	if episodeID != "" {
		tzRequest["tensorzero::episode_id"] = episodeID
	}

	// Marshal request
	requestBody, err := json.Marshal(tzRequest)
	if err != nil {
		return nil, fmt.Errorf("failed to marshal request: %w", err)
	}

	// Get Supabase URL
	supabaseURL := os.Getenv("SUPABASE_PROJECT_URL")
	if supabaseURL == "" {
		return nil, fmt.Errorf("SUPABASE_PROJECT_URL not set")
	}

	// Create HTTP request to TensorZero proxy (includes OpenAI-compatible path)
	endpoint := fmt.Sprintf("%s/functions/v1/tensorzero-proxy/openai/v1/chat/completions", supabaseURL)
	logging.Debug("Calling TensorZero proxy at: %s", endpoint)
	req, err := http.NewRequest("POST", endpoint, bytes.NewBuffer(requestBody))
	if err != nil {
		return nil, fmt.Errorf("failed to create request: %w", err)
	}

	// Set headers
	req.Header.Set("Content-Type", "application/json")
	req.Header.Set("Accept", "application/json")

	// Add authentication if auth manager is available (same pattern as investigation_server.go)
	if a.authManager != nil {
		// The authManager should be *auth.AuthManager, so let's use the exact same pattern
		if authMgr, ok := a.authManager.(interface {
			LoadToken() (*types.AuthToken, error)
		}); ok {
			if authToken, err := authMgr.LoadToken(); err == nil && authToken != nil {
				req.Header.Set("Authorization", fmt.Sprintf("Bearer %s", authToken.AccessToken))
			}
		}
	}

	// Send request
	client := &http.Client{Timeout: 30 * time.Second}
	resp, err := client.Do(req)
	if err != nil {
		return nil, fmt.Errorf("failed to send request: %w", err)
	}
	defer resp.Body.Close()

	// Check status code
	if resp.StatusCode != 200 {
		body, _ := io.ReadAll(resp.Body)
		return nil, fmt.Errorf("TensorZero proxy error: %d, body: %s", resp.StatusCode, string(body))
	}

	// Parse response
	var tzResponse map[string]interface{}
	if err := json.NewDecoder(resp.Body).Decode(&tzResponse); err != nil {
		return nil, fmt.Errorf("failed to decode response: %w", err)
	}

	// Convert to OpenAI format for compatibility
	choices, ok := tzResponse["choices"].([]interface{})
	if !ok || len(choices) == 0 {
		return nil, fmt.Errorf("no choices in response")
	}

	// Extract the first choice
	firstChoice, ok := choices[0].(map[string]interface{})
	if !ok {
		return nil, fmt.Errorf("invalid choice format")
	}

	message, ok := firstChoice["message"].(map[string]interface{})
	if !ok {
		return nil, fmt.Errorf("invalid message format")
	}

	content, ok := message["content"].(string)
	if !ok {
		return nil, fmt.Errorf("invalid content format")
	}

	// Create OpenAI-compatible response
	response := &openai.ChatCompletionResponse{
		Choices: []openai.ChatCompletionChoice{
			{
				Message: openai.ChatCompletionMessage{
					Role:    openai.ChatMessageRoleAssistant,
					Content: content,
				},
			},
		},
	}

	// Update episode ID if provided in response
	if respEpisodeID, ok := tzResponse["episode_id"].(string); ok && respEpisodeID != "" {
		a.episodeID = respEpisodeID
	}

	return response, nil
}

// ConvertEBPFProgramsToTraceSpecs converts old EBPFProgram format to new TraceSpec format
func (a *LinuxDiagnosticAgent) ConvertEBPFProgramsToTraceSpecs(ebpfPrograms []types.EBPFRequest) []ebpf.TraceSpec {
	var traceSpecs []ebpf.TraceSpec

	for _, prog := range ebpfPrograms {
		spec := a.convertToTraceSpec(prog)
		traceSpecs = append(traceSpecs, spec)
	}

	return traceSpecs
}

// convertToTraceSpec converts an EBPFRequest to a TraceSpec for BCC-style tracing
func (a *LinuxDiagnosticAgent) convertToTraceSpec(prog types.EBPFRequest) ebpf.TraceSpec {
	// Determine probe type based on target and type
	probeType := "p" // default to kprobe
	target := prog.Target

	if strings.HasPrefix(target, "tracepoint:") {
		probeType = "t"
		target = strings.TrimPrefix(target, "tracepoint:")
	} else if strings.HasPrefix(target, "kprobe:") {
		probeType = "p"
		target = strings.TrimPrefix(target, "kprobe:")
	} else if prog.Type == "tracepoint" {
		probeType = "t"
	} else if prog.Type == "syscall" {
		// Convert syscall names to kprobe targets
		if !strings.HasPrefix(target, "__x64_sys_") && !strings.Contains(target, ":") {
			if strings.HasPrefix(target, "sys_") {
				target = "__x64_" + target
			} else {
				target = "__x64_sys_" + target
			}
		}
		probeType = "p"
	}

	// Set default duration if not specified
	duration := prog.Duration
	if duration <= 0 {
		duration = 5 // default 5 seconds
	}

	return ebpf.TraceSpec{
		ProbeType: probeType,
		Target:    target,
		Format:    prog.Description, // Use description as format
		Arguments: []string{},       // Start with no arguments for compatibility
		Duration:  duration,
		UID:       -1, // No UID filter (don't default to 0 which means root only)
	}
}

// executeEBPFTraces executes multiple eBPF traces using the eBPF service
func (a *LinuxDiagnosticAgent) ExecuteEBPFTraces(traceSpecs []ebpf.TraceSpec) []map[string]interface{} {
	if len(traceSpecs) == 0 {
		return []map[string]interface{}{}
	}

	a.logger.Info("Executing %d eBPF traces", len(traceSpecs))

	results := make([]map[string]interface{}, 0, len(traceSpecs))

	// Execute each trace using the eBPF manager
	for i, spec := range traceSpecs {
		a.logger.Debug("Starting trace %d: %s", i, spec.Target)

		// Start the trace
		traceID, err := a.ebpfManager.StartTrace(spec)
		if err != nil {
			a.logger.Error("Failed to start trace %d: %v", i, err)
			result := map[string]interface{}{
				"index":   i,
				"target":  spec.Target,
				"success": false,
				"error":   err.Error(),
			}
			results = append(results, result)
			continue
		}

		// Wait for the trace duration
		time.Sleep(time.Duration(spec.Duration) * time.Second)

		// Get the trace result
		traceResult, err := a.ebpfManager.GetTraceResult(traceID)
		if err != nil {
			a.logger.Error("Failed to get results for trace %d: %v", i, err)
			result := map[string]interface{}{
				"index":   i,
				"target":  spec.Target,
				"success": false,
				"error":   err.Error(),
			}
			results = append(results, result)
			continue
		}

		// Build successful result
		result := map[string]interface{}{
			"index":             i,
			"target":            spec.Target,
			"success":           true,
			"event_count":       traceResult.EventCount,
			"events_per_second": traceResult.Statistics.EventsPerSecond,
			"duration":          traceResult.EndTime.Sub(traceResult.StartTime).Seconds(),
			"summary":           traceResult.Summary,
		}
		results = append(results, result)

		a.logger.Debug("Completed trace %d: %d events", i, traceResult.EventCount)
	}

	a.logger.Info("Completed %d eBPF traces", len(results))
	return results
}