pulse/internal/ai/service.go

package ai

import (
	"context"
	"crypto/hmac"
	cryptorand "crypto/rand"
	"crypto/sha256"
	"encoding/hex"
	"fmt"
	"io"
	"net"
	"net/http"
	"net/url"
	"os"
	"reflect"
	"regexp"
	"strconv"
	"strings"
	"sync"
	"time"

	"github.com/google/uuid"
	"github.com/rcourtman/pulse-go-rewrite/internal/agentcapabilities"
	"github.com/rcourtman/pulse-go-rewrite/internal/agentexec"
	"github.com/rcourtman/pulse-go-rewrite/internal/ai/approval"
	"github.com/rcourtman/pulse-go-rewrite/internal/ai/baseline"
	"github.com/rcourtman/pulse-go-rewrite/internal/ai/cost"
	"github.com/rcourtman/pulse-go-rewrite/internal/ai/knowledge"
	"github.com/rcourtman/pulse-go-rewrite/internal/ai/memory"
	"github.com/rcourtman/pulse-go-rewrite/internal/ai/modelboundary"
	"github.com/rcourtman/pulse-go-rewrite/internal/ai/providers"
	aitools "github.com/rcourtman/pulse-go-rewrite/internal/ai/tools"
	"github.com/rcourtman/pulse-go-rewrite/internal/config"
	"github.com/rcourtman/pulse-go-rewrite/internal/models"
	"github.com/rcourtman/pulse-go-rewrite/internal/unifiedresources"
	"github.com/rcourtman/pulse-go-rewrite/pkg/aicontracts"
	pkglicensing "github.com/rcourtman/pulse-go-rewrite/pkg/licensing"
	"github.com/rcourtman/pulse-go-rewrite/pkg/securityutil"
	"github.com/rs/zerolog/log"
)

var modelsCacheCredentialKey = newModelsCacheCredentialKey()

func newModelsCacheCredentialKey() []byte {
	key := make([]byte, 32)
	if _, err := cryptorand.Read(key); err != nil {
		panic(fmt.Sprintf("initialize AI models cache credential key: %v", err))
	}
	return key
}

func credentialCacheDigest(values ...string) string {
	mac := hmac.New(sha256.New, modelsCacheCredentialKey)
	for _, value := range values {
		_, _ = io.WriteString(mac, strconv.Itoa(len(value)))
		_, _ = io.WriteString(mac, ":")
		_, _ = io.WriteString(mac, value)
	}
	return hex.EncodeToString(mac.Sum(nil))
}

// StateProvider is a type alias for models.SnapshotProvider.
// Kept for local convenience; all new code should use models.SnapshotProvider directly.
type StateProvider = models.SnapshotProvider

// CommandPolicy defines the interface for command security policy
type CommandPolicy interface {
	Evaluate(command string) agentexec.PolicyDecision
}

// LicenseState represents the current state of the license
type LicenseState string

const (
	LicenseStateNone        LicenseState = "none"
	LicenseStateActive      LicenseState = "active"
	LicenseStateExpired     LicenseState = "expired"
	LicenseStateGracePeriod LicenseState = "grace_period"
)

const legacyServiceToolResultModelContextChars = 8000

func legacyServiceProviderToolResultContextOptions() agentcapabilities.ProviderToolResultContextOptions {
	return agentcapabilities.ProviderToolResultContextOptions{
		MaxModelContentChars: legacyServiceToolResultModelContextChars,
	}
}

func newLegacyServiceProviderToolResultContextProjection(toolUseID, content string, isError bool) agentcapabilities.ProviderToolResultContextProjection {
	projection := agentcapabilities.NewProviderToolResultContextProjection(toolUseID, content, isError, legacyServiceProviderToolResultContextOptions())
	logLegacyServiceProviderToolResultTruncation(projection.Truncation)
	return projection
}

func legacyAssistantCapabilityInputSchema(name string) map[string]interface{} {
	cap, ok := agentcapabilities.FindCapability(agentcapabilities.CanonicalManifest().Capabilities, name)
	if !ok {
		return agentcapabilities.ProviderInputSchema(agentcapabilities.InputSchema{})
	}
	return agentcapabilities.ProviderInputSchemaFromRaw(agentcapabilities.ToolInputSchema(cap))
}

func logLegacyServiceProviderToolResultTruncation(truncation agentcapabilities.ProviderToolResultTruncation) {
	if !truncation.Applied {
		return
	}

	log.Debug().
		Int("original_chars", truncation.OriginalChars).
		Int("truncated_to", truncation.MaxChars).
		Int("chars_cut", truncation.TruncatedChars).
		Msg("Truncated large tool result")
}

// LicenseChecker provides license feature checking for Pro features
type LicenseChecker interface {
	HasFeature(feature string) bool
	// GetLicenseStateString returns the current license state (none, active, expired, grace_period)
	// and whether features are available (true for active/grace_period)
	GetLicenseStateString() (string, bool)
}

// AgentServer defines the interface for communicating with agents
type AgentServer interface {
	GetConnectedAgents() []agentexec.ConnectedAgent
	ExecuteCommand(ctx context.Context, agentID string, cmd agentexec.ExecuteCommandPayload) (*agentexec.CommandResultPayload, error)
}

// ChatServiceProvider defines the interface for accessing chat functionality
// This is used by the investigation orchestrator and patrol to run AI executions
type ChatServiceProvider interface {
	CreateSession(ctx context.Context) (*ChatSession, error)
	ExecuteStream(ctx context.Context, req ChatExecuteRequest, callback ChatStreamCallback) error
	ExecutePatrolStream(ctx context.Context, req PatrolExecuteRequest, callback ChatStreamCallback) (*PatrolStreamResponse, error)
	GetMessages(ctx context.Context, sessionID string) ([]ChatMessage, error)
	DeleteSession(ctx context.Context, sessionID string) error
	ReloadConfig(ctx context.Context, cfg *config.AIConfig) error
}

// ChatSession represents a chat session (minimal interface for investigations)
type ChatSession struct {
	ID string `json:"id"`
}

// ChatExecuteRequest represents a chat execution request
type ChatExecuteRequest struct {
	Prompt    string `json:"prompt"`
	SessionID string `json:"session_id,omitempty"`
}

// ChatStreamCallback is called for each streaming event
type ChatStreamCallback func(event ChatStreamEvent)

// ChatStreamEvent represents a streaming event
type ChatStreamEvent struct {
	Type string `json:"type"`
	Data []byte `json:"data,omitempty"`
}

// ChatMessage represents a chat message
type ChatMessage struct {
	ID               string          `json:"id"`
	Role             string          `json:"role"`
	Content          string          `json:"content"`
	ReasoningContent string          `json:"reasoning_content,omitempty"`
	ToolCalls        []ChatToolCall  `json:"tool_calls"`
	ToolResult       *ChatToolResult `json:"tool_result,omitempty"`
	Timestamp        time.Time       `json:"timestamp"`
}

func EmptyChatMessage() ChatMessage {
	return ChatMessage{}.NormalizeCollections()
}

func (m ChatMessage) NormalizeCollections() ChatMessage {
	if m.ToolCalls == nil {
		m.ToolCalls = []ChatToolCall{}
	}
	for i := range m.ToolCalls {
		m.ToolCalls[i] = m.ToolCalls[i].NormalizeCollections()
	}
	return m
}

// ChatToolCall retains observed output and result status in product history.
// Provider turns use the explicit ProviderToolCall projection.
type ChatToolCall = agentcapabilities.TranscriptToolCall

func EmptyChatToolCall() ChatToolCall {
	return ChatToolCall{}.NormalizeCollections()
}

// ChatToolResult represents the result of a tool invocation. It aliases the
// shared Pulse Intelligence provider-result shape so API-facing chat messages
// and provider turns do not drift on tool result JSON.
type ChatToolResult = agentcapabilities.ProviderToolResult

// PatrolExecuteRequest represents a patrol execution request via the chat service
type PatrolExecuteRequest struct {
	Prompt           string   `json:"prompt"`
	SystemPrompt     string   `json:"system_prompt"`
	SessionID        string   `json:"session_id,omitempty"`
	ExecutionID      string   `json:"execution_id,omitempty"`
	UseCase          string   `json:"use_case"` // "patrol" — for model selection
	MaxTurns         int      `json:"max_turns,omitempty"`
	AllowedToolNames []string `json:"allowed_tool_names,omitempty"`
}

// QuickAnalysisRequest represents a lightweight single-turn analysis request.
// ExecutionID lets higher-level workflows group ancillary calls under one
// commercial run instead of charging each helper call separately.
type QuickAnalysisRequest struct {
	Prompt      string `json:"prompt"`
	ExecutionID string `json:"execution_id,omitempty"`
	UseCase     string `json:"use_case,omitempty"`
	// TargetType tags the recorded usage event so cost telemetry can
	// distinguish sub-workloads sharing a use case (e.g. patrol alert
	// auto-resolve reviews vs the main patrol pass).
	TargetType string `json:"target_type,omitempty"`
}

// PatrolStreamResponse contains the results of a patrol execution via the chat service
type PatrolStreamResponse struct {
	Content      string `json:"content"`
	InputTokens  int    `json:"input_tokens"`
	OutputTokens int    `json:"output_tokens"`
	// Prompt-cache buckets summed across the run's provider calls; disjoint
	// from InputTokens, see providers.ChatResponse.
	CacheCreationInputTokens int `json:"cache_creation_input_tokens,omitempty"`
	CacheReadInputTokens     int `json:"cache_read_input_tokens,omitempty"`
	// ToolCalls carries calls observed by bounded Patrol follow-up passes so
	// the parent run's durable audit contains every provider action.
	ToolCalls []ToolCallRecord `json:"tool_calls,omitempty"`
}

// Service orchestrates AI interactions
type Service struct {
	mu          sync.RWMutex
	orgID       string
	persistence *config.ConfigPersistence
	provider    providers.Provider
	// providerInitErr records why the default provider could not be built even
	// though the config claims an enabled, configured provider (for example a
	// live model-catalog resolution against an Ollama server that was still
	// booting). Patrol readiness reports it instead of the misleading
	// "provider not configured" copy, and RetryProviderInit clears it.
	providerInitErr          string
	cfg                      *config.AIConfig
	agentServer              AgentServer
	policy                   CommandPolicy
	stateProvider            StateProvider
	readState                unifiedresources.ReadState
	alertProvider            AlertProvider
	knowledgeStore           *knowledge.Store
	costStore                *cost.Store
	unifiedResourceProvider  UnifiedResourceProvider
	resourceExportStore      unifiedresources.ResourceStore
	resourceExportStoreOrgID string
	patrolAutopilotPolicy    func() unifiedresources.PatrolAutopilotServerPolicy
	patrolObjectiveStore     *PatrolObjectiveStore
	patrolService            *PatrolService        // Background AI monitoring service
	thresholdProvider        ThresholdProvider     // User-configured alert thresholds; nil means factory defaults
	metadataProvider         MetadataProvider      // Enables AI to update resource URLs
	incidentStore            *memory.IncidentStore // Alert-scoped investigation memory; not the canonical durable resource history
	chatService              ChatServiceProvider   // Chat service for investigation orchestrator

	// Alert-triggered analysis - token-efficient real-time AI insights
	alertTriggeredAnalyzer aicontracts.AlertAnalyzer
	// alertAnalyzerFactory gates construction of the AlertAnalyzer.
	// When nil (OSS binary), the analyzer is never created regardless of config.
	alertAnalyzerFactory func(aicontracts.AlertAnalyzerDeps) aicontracts.AlertAnalyzer

	limits executionLimits

	modelsCache modelsCache

	// patrolPreflightCache holds the most recent preflight result for the
	// configured Patrol provider+model so operators can see when tools
	// were last verified without re-running the call.
	patrolPreflightCache patrolPreflightCache

	// patrolModelReadinessCache holds the explicit, multi-scenario advisor
	// evidence separately from the lightweight startup preflight.
	patrolModelReadinessCache patrolModelReadinessCache

	// License checker for Pro feature gating
	licenseChecker LicenseChecker
}

type executionLimits struct {
	chatSlots   chan struct{}
	patrolSlots chan struct{}
}

type providerModelsEntry struct {
	at     time.Time
	models []providers.ModelInfo
}

type modelsCache struct {
	mu        sync.RWMutex
	key       string
	providers map[string]providerModelsEntry
	ttl       time.Duration
}

// alertFindingRecorderAdapter adapts PatrolService.recordFinding for the
// aicontracts.FindingRecorder interface used by the enterprise alert analyzer.
type alertFindingRecorderAdapter struct {
	ps *PatrolService
}

func (a *alertFindingRecorderAdapter) RecordAlertFinding(f *aicontracts.AlertAnalyzerFinding) {
	if a.ps == nil || f == nil {
		return
	}
	a.ps.recordFinding(&Finding{
		ID:              f.ID,
		Key:             f.Key,
		Severity:        FindingSeverity(f.Severity),
		Category:        FindingCategory(f.Category),
		ResourceID:      f.ResourceID,
		ResourceName:    f.ResourceName,
		ResourceType:    f.ResourceType,
		Title:           f.Title,
		Description:     f.Description,
		Recommendation:  f.Recommendation,
		Evidence:        f.Evidence,
		AlertIdentifier: f.AlertIdentifier,
		DetectedAt:      f.DetectedAt,
		LastSeenAt:      f.LastSeenAt,
	})
}

// NewService creates a new AI service
func NewService(persistence *config.ConfigPersistence, agentServer AgentServer) *Service {
	// Initialize knowledge store
	var knowledgeStore *knowledge.Store
	costStore := cost.NewStore(cost.DefaultMaxDays)
	objectiveStore := NewInMemoryPatrolObjectiveStore()
	if persistence != nil {
		var err error
		knowledgeStore, err = knowledge.NewStore(persistence.DataDir())
		if err != nil {
			log.Warn().Err(err).Msg("failed to initialize knowledge store")
		}
		if err := costStore.SetPersistence(NewCostPersistenceAdapter(persistence)); err != nil {
			log.Warn().Err(err).Msg("failed to initialize AI usage cost store")
		}
		objectiveStore, err = NewPatrolObjectiveStore(persistence.DataDir())
		if err != nil {
			objectiveStore = nil
			log.Error().Err(err).Msg("failed to initialize encrypted Patrol objective store")
		}
	}

	svc := &Service{
		orgID:                "default",
		persistence:          persistence,
		agentServer:          agentServer,
		policy:               agentexec.DefaultPolicy(),
		knowledgeStore:       knowledgeStore,
		costStore:            costStore,
		patrolObjectiveStore: objectiveStore,
		limits: executionLimits{
			chatSlots:   make(chan struct{}, 4),
			patrolSlots: make(chan struct{}, 1),
		},
		modelsCache: modelsCache{
			ttl:       5 * time.Minute,
			providers: make(map[string]providerModelsEntry),
		},
	}

	if persistence != nil {
		if store, err := unifiedresources.NewSQLiteResourceStore(persistence.DataDir(), svc.orgID); err != nil {
			log.Warn().Err(err).Str("orgID", svc.orgID).Msg("failed to initialize unified resource export store")
		} else {
			svc.resourceExportStore = store
			svc.resourceExportStoreOrgID = svc.orgID
		}
	}

	svc.loadPatrolModelReadiness()

	return svc
}

// SetOrgID sets the org scope used when creating approval records.
func (s *Service) SetOrgID(orgID string) {
	normalized := strings.TrimSpace(orgID)
	if normalized == "" {
		normalized = "default"
	}

	s.mu.Lock()
	if s.orgID == normalized && s.resourceExportStore != nil {
		s.mu.Unlock()
		return
	}

	oldStore := s.resourceExportStore
	incidentStore := s.incidentStore
	persistence := s.persistence
	s.orgID = normalized
	s.resourceExportStore = nil
	s.resourceExportStoreOrgID = ""
	s.mu.Unlock()

	if incidentStore != nil {
		incidentStore.SetResourceTimelineStore(nil)
	}

	if oldStore != nil {
		if err := oldStore.Close(); err != nil {
			log.Warn().Err(err).Str("orgID", normalized).Msg("failed to close previous unified resource export store")
		}
	}

	if persistence == nil {
		return
	}

	store, err := unifiedresources.NewSQLiteResourceStore(persistence.DataDir(), normalized)
	if err != nil {
		log.Warn().Err(err).Str("orgID", normalized).Msg("failed to initialize unified resource export store for org")
		return
	}

	s.mu.Lock()
	if s.orgID == normalized && s.resourceExportStore == nil {
		s.resourceExportStore = store
		s.resourceExportStoreOrgID = normalized
		incidentStore = s.incidentStore
		s.mu.Unlock()
		if incidentStore != nil {
			incidentStore.SetResourceTimelineStore(store)
		}
		return
	}
	s.mu.Unlock()

	if err := store.Close(); err != nil {
		log.Warn().Err(err).Str("orgID", normalized).Msg("failed to close unused unified resource export store")
	}
}

// GetOrgID returns the org scope associated with this service instance.
func (s *Service) GetOrgID() string {
	s.mu.RLock()
	defer s.mu.RUnlock()
	orgID := strings.TrimSpace(s.orgID)
	if orgID == "" {
		return "default"
	}
	return orgID
}

func (s *Service) acquireExecutionSlot(ctx context.Context, useCase string) (func(), error) {
	normalized := strings.TrimSpace(strings.ToLower(useCase))
	if normalized == "" {
		normalized = "chat"
	}

	var slots chan struct{}
	if normalized == "patrol" {
		slots = s.limits.patrolSlots
	} else {
		slots = s.limits.chatSlots
	}

	timer := time.NewTimer(5 * time.Second)
	defer func() {
		if !timer.Stop() {
			select {
			case <-timer.C:
			default:
			}
		}
	}()

	select {
	case slots <- struct{}{}:
		return func() { <-slots }, nil
	case <-ctx.Done():
		return nil, ctx.Err()
	case <-timer.C:
		return nil, fmt.Errorf("Pulse Assistant is busy - too many concurrent requests")
	}
}

func (s *Service) enforceBudget(useCase string) error {
	s.mu.RLock()
	cfg := s.cfg
	store := s.costStore
	s.mu.RUnlock()

	if cfg == nil || cfg.CostBudgetUSD30d <= 0 || store == nil {
		return nil
	}

	summary := store.GetSummary(30)
	if !summary.Totals.PricingKnown {
		// We can't reliably enforce without pricing. Keep tracking and allow.
		return nil
	}

	if summary.Totals.EstimatedUSD >= cfg.CostBudgetUSD30d {
		normalized := strings.TrimSpace(strings.ToLower(useCase))
		if normalized == "" {
			normalized = "chat"
		}
		return &CostBudgetExceededError{
			SpentUSD:  summary.Totals.EstimatedUSD,
			BudgetUSD: cfg.CostBudgetUSD30d,
			Days:      summary.EffectiveDays,
		}
	}

	return nil
}

// CheckBudget is a public wrapper around enforceBudget, allowing callers
// (e.g. patrol) to verify budget availability before starting expensive work.
func (s *Service) CheckBudget(useCase string) error {
	return s.enforceBudget(useCase)
}

// SetReadState injects a unified read-state provider for context enrichment.
// This is forwarded to PatrolService when available.
func (s *Service) SetReadState(rs unifiedresources.ReadState) {
	s.mu.Lock()
	defer s.mu.Unlock()
	s.readState = rs

	s.initPatrolServiceLocked()

	if s.patrolService != nil {
		s.patrolService.SetReadState(rs)
	}
}

// initPatrolServiceLocked creates the patrol service once any canonical patrol
// runtime input exists. Must be called while holding s.mu.
func (s *Service) initPatrolServiceLocked() {
	if s.patrolService != nil {
		return
	}
	if s.stateProvider == nil && s.readState == nil && s.unifiedResourceProvider == nil {
		return
	}

	s.patrolService = NewPatrolService(s, s.stateProvider)
	s.patrolService.SetObjectiveStore(s.patrolObjectiveStore)
	if s.thresholdProvider != nil {
		s.patrolService.SetThresholdProvider(s.thresholdProvider)
	}
	if s.knowledgeStore != nil {
		s.patrolService.SetKnowledgeStore(s.knowledgeStore)
	}
	if s.incidentStore != nil {
		s.patrolService.SetIncidentStore(s.incidentStore)
	}
	if s.readState != nil {
		s.patrolService.SetReadState(s.readState)
	}
	if s.unifiedResourceProvider != nil {
		s.patrolService.SetUnifiedResourceProvider(s.unifiedResourceProvider)
	}
}

// SetStateProvider sets the state provider for infrastructure context
func (s *Service) SetStateProvider(sp StateProvider) {
	s.mu.Lock()
	s.stateProvider = sp

	s.initPatrolServiceLocked()
	if s.patrolService != nil {
		s.patrolService.SetStateProvider(sp)
	}

	// Initialize alert-triggered analyzer if not already done (requires enterprise factory)
	if s.alertTriggeredAnalyzer == nil && s.alertAnalyzerFactory != nil && s.patrolService != nil {
		deps := aicontracts.AlertAnalyzerDeps{
			FindingRecorder:  &alertFindingRecorderAdapter{ps: s.patrolService},
			IncidentRecorder: s,
		}
		s.alertTriggeredAnalyzer = s.alertAnalyzerFactory(deps)
	}
	s.mu.Unlock()
}

// GetStateProvider returns the state provider for infrastructure context
func (s *Service) GetStateProvider() StateProvider {
	s.mu.RLock()
	defer s.mu.RUnlock()
	return s.stateProvider
}

// GetPatrolService returns the patrol service for background monitoring
func (s *Service) GetPatrolService() *PatrolService {
	s.mu.RLock()
	defer s.mu.RUnlock()
	return s.patrolService
}

// GetPatrolObjectiveStore returns the organization-scoped durable operator
// objective store. It exists independently from a running Patrol service so an
// operator can retain intent while provider or infrastructure setup is offline.
func (s *Service) GetPatrolObjectiveStore() *PatrolObjectiveStore {
	s.mu.RLock()
	defer s.mu.RUnlock()
	return s.patrolObjectiveStore
}

// SetChatService sets the chat service for investigation orchestrator
func (s *Service) SetChatService(cs ChatServiceProvider) {
	s.mu.Lock()
	defer s.mu.Unlock()
	s.chatService = cs
	s.configurePatrolProviderFactoryLocked()
}

// GetChatService returns the chat service for investigation orchestrator
func (s *Service) GetChatService() ChatServiceProvider {
	s.mu.RLock()
	defer s.mu.RUnlock()
	return s.chatService
}

// GetRemediationLog returns the remediation log from the patrol service
func (s *Service) GetRemediationLog() *RemediationLog {
	s.mu.RLock()
	patrol := s.patrolService
	s.mu.RUnlock()
	if patrol == nil {
		return nil
	}
	return patrol.GetRemediationLog()
}

// GetAlertTriggeredAnalyzer returns the alert-triggered analyzer for token-efficient real-time analysis
func (s *Service) GetAlertTriggeredAnalyzer() aicontracts.AlertAnalyzer {
	s.mu.RLock()
	defer s.mu.RUnlock()
	return s.alertTriggeredAnalyzer
}

// GetAIConfig returns the current AI configuration
func (s *Service) GetAIConfig() *config.AIConfig {
	s.mu.RLock()
	defer s.mu.RUnlock()
	return s.cfg
}

// CostStore returns the underlying usage cost store handle so callers
// that own a Service can record cost events directly. Returned value
// may be nil when persistence/initialization failed during NewService.
// Consumers must nil-check before recording.
func (s *Service) CostStore() *cost.Store {
	s.mu.RLock()
	defer s.mu.RUnlock()
	return s.costStore
}

// GetCostSummary returns usage rollups for the last N days.
func (s *Service) GetCostSummary(days int) cost.Summary {
	s.mu.RLock()
	store := s.costStore
	s.mu.RUnlock()

	if store == nil {
		if days <= 0 {
			days = 30
		}
		effectiveDays := days
		truncated := false
		if cost.DefaultMaxDays > 0 && days > cost.DefaultMaxDays {
			effectiveDays = cost.DefaultMaxDays
			truncated = true
		}
		return cost.EmptySummary(days, cost.DefaultMaxDays, effectiveDays, truncated)
	}
	return store.GetSummary(days)
}

// ListCostEvents returns retained AI usage events within the requested time window.
func (s *Service) ListCostEvents(days int) []cost.UsageEvent {
	s.mu.RLock()
	store := s.costStore
	s.mu.RUnlock()
	if store == nil {
		return nil
	}
	return store.ListEvents(days)
}

// ClearCostHistory deletes retained AI usage events (admin operation).
func (s *Service) ClearCostHistory() error {
	s.mu.RLock()
	store := s.costStore
	s.mu.RUnlock()
	if store == nil {
		return nil
	}
	return store.Clear()
}

// SetPatrolThresholdProvider sets the threshold provider for patrol
// This should be called with an AlertThresholdAdapter to connect patrol to user-configured thresholds
func (s *Service) SetPatrolThresholdProvider(provider ThresholdProvider) {
	s.mu.Lock()
	s.thresholdProvider = provider
	patrol := s.patrolService
	s.mu.Unlock()

	if patrol != nil {
		patrol.SetThresholdProvider(provider)
	}
}

// MetricsHistoryProvider provides access to historical metrics for trend analysis
// This interface matches the monitoring.MetricsHistory methods we need
type MetricsHistoryProvider interface {
	GetNodeMetrics(nodeID string, metricType string, duration time.Duration) []MetricPoint
	GetGuestMetrics(guestID string, metricType string, duration time.Duration) []MetricPoint
	GetAllGuestMetrics(guestID string, duration time.Duration) map[string][]MetricPoint
	GetAllStorageMetrics(storageID string, duration time.Duration) map[string][]MetricPoint
}

// MetricPoint is an alias for the shared metric point type
type MetricPoint = models.MetricPoint

// SetMetricsHistoryProvider sets the metrics history provider for enriched AI context
// This enables the AI to see trends, anomalies, and predictions based on historical data
func (s *Service) SetMetricsHistoryProvider(provider MetricsHistoryProvider) {
	s.mu.RLock()
	patrol := s.patrolService
	s.mu.RUnlock()

	if patrol != nil {
		patrol.SetMetricsHistoryProvider(provider)
	}
}

// SetBaselineStore sets the baseline store for anomaly detection
func (s *Service) SetBaselineStore(store *BaselineStore) {
	s.mu.RLock()
	patrol := s.patrolService
	s.mu.RUnlock()

	if patrol != nil {
		patrol.SetBaselineStore(store)
	}
}

// SetChangeDetector sets the change detector for operational memory
func (s *Service) SetChangeDetector(detector *ChangeDetector) {
	s.mu.RLock()
	patrol := s.patrolService
	s.mu.RUnlock()

	if patrol != nil {
		patrol.SetChangeDetector(detector)
	}
}

// SetRemediationLog sets the remediation log for tracking fix attempts
func (s *Service) SetRemediationLog(remLog *RemediationLog) {
	s.mu.RLock()
	patrol := s.patrolService
	s.mu.RUnlock()

	if patrol != nil {
		patrol.SetRemediationLog(remLog)
	}
}

// SetIncidentStore sets the incident store for alert timeline memory.
func (s *Service) SetIncidentStore(store *memory.IncidentStore) {
	s.mu.Lock()
	defer s.mu.Unlock()
	s.incidentStore = store
	if store != nil {
		store.SetResourceTimelineStore(s.resourceExportStore)
	}

	if s.patrolService != nil {
		s.patrolService.SetIncidentStore(store)
	}
}

// SetPatternDetector sets the pattern detector for failure prediction
func (s *Service) SetPatternDetector(detector *PatternDetector) {
	s.mu.RLock()
	patrol := s.patrolService
	s.mu.RUnlock()

	if patrol != nil {
		patrol.SetPatternDetector(detector)
	}
}

// SetCorrelationDetector sets the correlation detector for multi-resource correlation
func (s *Service) SetCorrelationDetector(detector *CorrelationDetector) {
	s.mu.RLock()
	patrol := s.patrolService
	s.mu.RUnlock()

	if patrol != nil {
		patrol.SetCorrelationDetector(detector)
	}
}

// SetLicenseChecker sets the license checker for Pro feature gating
func (s *Service) SetLicenseChecker(checker LicenseChecker) {
	s.mu.Lock()
	defer s.mu.Unlock()
	s.licenseChecker = checker
}

// SetAlertAnalyzerFactory registers the factory that creates the enterprise
// AlertTriggeredAnalyzer. When nil (OSS binary), no analyzer is created.
func (s *Service) SetAlertAnalyzerFactory(fn func(aicontracts.AlertAnalyzerDeps) aicontracts.AlertAnalyzer) {
	s.mu.Lock()
	defer s.mu.Unlock()
	s.alertAnalyzerFactory = fn
}

// HasLicenseFeature checks if a Pro feature is licensed (returns true if no license checker is set)
func (s *Service) HasLicenseFeature(feature string) bool {
	s.mu.RLock()
	checker := s.licenseChecker
	s.mu.RUnlock()

	if checker == nil {
		// No license checker means no enforcement (development mode or feature not gated)
		return true
	}
	return checker.HasFeature(feature)
}

// PatrolInvestigationLicensed reports a positive licence answer for Patrol
// investigation. Unlike HasLicenseFeature it treats a missing checker as not
// licensed, so the Pro default mode never applies before the licence is known.
func (s *Service) PatrolInvestigationLicensed() bool {
	s.mu.RLock()
	checker := s.licenseChecker
	s.mu.RUnlock()
	return checker != nil && checker.HasFeature(FeatureAIAutoFix)
}

const minCommunityPatrolInterval = 1 * time.Hour

// GetEffectivePatrolAutonomyLevel returns the autonomy level that should be enforced at runtime
// for the current license state.
//
// Community tier is locked to "monitor" (findings-only). Higher levels imply investigation and
// are gated behind ai_autofix (Pro/Cloud). A licensed install that never chose a mode runs
// "approval" (Ask first), so the investigation it pays for is on by default.
func (s *Service) GetEffectivePatrolAutonomyLevel() string {
	cfg := s.GetConfig()
	if cfg == nil {
		return config.PatrolAutonomyMonitor
	}
	if !s.HasLicenseFeature(FeatureAIAutoFix) {
		return config.PatrolAutonomyMonitor
	}
	if !cfg.PatrolAutonomyLevelChosen() {
		if s.PatrolInvestigationLicensed() {
			return config.PatrolAutonomyApproval
		}
		return config.PatrolAutonomyMonitor
	}
	s.mu.RLock()
	orgID := s.orgID
	policyProvider := s.patrolAutopilotPolicy
	s.mu.RUnlock()
	policy := unifiedresources.CurrentPatrolAutopilotServerPolicy(time.Now().UTC())
	if policyProvider != nil {
		policy = policyProvider()
	}
	level, _ := cfg.GetEffectivePatrolAutonomyWithPolicy(orgID, policy)
	return level
}

// SetPatrolAutopilotServerPolicyProvider keeps runtime admission on the same
// server-owned clock and current-version source as the API boundary.
func (s *Service) SetPatrolAutopilotServerPolicyProvider(provider func() unifiedresources.PatrolAutopilotServerPolicy) {
	if s == nil {
		return
	}
	s.mu.Lock()
	defer s.mu.Unlock()
	s.patrolAutopilotPolicy = provider
}

// ApplyPatrolAutonomyConfig publishes an already-persisted autonomy mutation
// without rebuilding provider clients. Callers must persist first and invoke
// this while holding the policy-mutation coordinator so revocation and runtime
// admission observe one ordered boundary.
func (s *Service) ApplyPatrolAutonomyConfig(cfg *config.AIConfig) error {
	if s == nil || cfg == nil {
		return fmt.Errorf("Pulse Patrol config is unavailable")
	}
	s.mu.Lock()
	defer s.mu.Unlock()
	if s.cfg == nil {
		s.cfg = config.NewDefaultAIConfig()
	}
	s.cfg.PatrolAutonomyLevel = cfg.PatrolAutonomyLevel
	s.cfg.PatrolFullModeUnlocked = cfg.PatrolFullModeUnlocked
	s.cfg.PatrolInvestigationBudget = cfg.PatrolInvestigationBudget
	s.cfg.PatrolInvestigationTimeoutSec = cfg.PatrolInvestigationTimeoutSec
	s.cfg.PatrolAutopilotAcknowledgements = clonePatrolAutopilotAcknowledgements(cfg.PatrolAutopilotAcknowledgements)
	s.cfg.PatrolAutopilotRevocations = append([]unifiedresources.PatrolAutopilotRevocation(nil), cfg.PatrolAutopilotRevocations...)
	if cfg.PatrolAutopilotActivation == nil {
		s.cfg.PatrolAutopilotActivation = nil
	} else {
		activation := *cfg.PatrolAutopilotActivation
		s.cfg.PatrolAutopilotActivation = &activation
	}
	if s.patrolService != nil {
		orchestrator := s.patrolService.GetInvestigationOrchestrator()
		if updater, ok := orchestrator.(aicontracts.InvestigationExecutionLimitUpdater); ok {
			updater.SetInvestigationExecutionLimits(
				s.cfg.GetPatrolInvestigationBudget(),
				s.cfg.GetPatrolInvestigationTimeout(),
			)
		}
	}
	return nil
}

func clonePatrolAutopilotAcknowledgements(records []unifiedresources.PatrolAutopilotAcknowledgement) []unifiedresources.PatrolAutopilotAcknowledgement {
	cloned := append([]unifiedresources.PatrolAutopilotAcknowledgement(nil), records...)
	for index := range cloned {
		cloned[index].AcceptedScope = append([]string(nil), cloned[index].AcceptedScope...)
	}
	return cloned
}

func (s *Service) getEffectivePatrolInterval(cfg *config.AIConfig) time.Duration {
	if cfg == nil {
		return 0
	}
	interval := cfg.GetPatrolInterval()
	if interval <= 0 {
		return interval
	}
	if !s.HasLicenseFeature(FeatureAIAutoFix) && interval < minCommunityPatrolInterval {
		return minCommunityPatrolInterval
	}
	return interval
}

func (s *Service) patrolConfigFromAIConfig(cfg *config.AIConfig) PatrolConfig {
	patrolCfg := DefaultPatrolConfig()
	if cfg == nil {
		if IsDemoMode() {
			// Demo runtimes simulate patrol without provider settings.
			patrolCfg.Enabled = true
			return patrolCfg
		}
		patrolCfg.Enabled = false
		patrolCfg.RuntimeBlockedReason = "Pulse Intelligence settings could not be loaded from persistence."
		patrolCfg.RuntimeBlockedCause = PatrolFailureCauseSettingsPersistence
		return patrolCfg
	}

	patrolCfg.Enabled = cfg.IsPatrolEnabled()
	patrolCfg.Interval = s.getEffectivePatrolInterval(cfg)
	patrolCfg.AnalyzeNodes = cfg.PatrolAnalyzeNodes
	patrolCfg.AnalyzeGuests = cfg.PatrolAnalyzeGuests
	patrolCfg.AnalyzeDocker = cfg.PatrolAnalyzeDocker
	patrolCfg.AnalyzeStorage = cfg.PatrolAnalyzeStorage
	if IsDemoMode() {
		// Demo/mock runtimes simulate patrol runs without a provider; never
		// carry a readiness blocker into the demo loop.
		patrolCfg.Enabled = true
		return patrolCfg
	}
	if patrolCfg.Enabled {
		if readiness := EvaluatePatrolConfigReadiness(cfg); !readiness.Ready {
			patrolCfg.RuntimeBlockedReason = readiness.Summary
			patrolCfg.RuntimeBlockedCause = readiness.Cause
		}
	}
	return patrolCfg
}

// GetLicenseState returns the current license state and whether features are available
func (s *Service) GetLicenseState() (string, bool) {
	s.mu.RLock()
	checker := s.licenseChecker
	s.mu.RUnlock()

	if checker == nil {
		// No license checker means development mode - treat as active
		return string(LicenseStateActive), true
	}
	return checker.GetLicenseStateString()
}

// Feature constants are imported from the license package for compile-time consistency.
// This ensures the ai package always uses the same feature strings as the license system.
const (
	FeatureAIPatrol  = pkglicensing.FeatureAIPatrol
	FeatureAIAlerts  = pkglicensing.FeatureAIAlerts
	FeatureAIAutoFix = pkglicensing.FeatureAIAutoFix
)

const DevDisableBackgroundAIEnv = "PULSE_DEV_DISABLE_BACKGROUND_AI"

// BackgroundAutomationDisabledForDev returns true when the local development
// runtime should avoid automatic AI work that can burn provider quota while a
// developer is only trying to run the app.
func BackgroundAutomationDisabledForDev() bool {
	if os.Getenv("PULSE_DEV") != "true" && !strings.EqualFold(os.Getenv("NODE_ENV"), "development") {
		return false
	}

	value := strings.TrimSpace(os.Getenv(DevDisableBackgroundAIEnv))
	if value == "" {
		return false
	}

	disabled, err := strconv.ParseBool(value)
	if err != nil {
		return false
	}
	return disabled
}

func patrolEventTriggerConfigFromAIConfig(cfg *config.AIConfig) PatrolEventTriggerConfig {
	if cfg == nil {
		return PatrolEventTriggerConfig{}
	}
	return PatrolEventTriggerConfig{
		AlertTriggersEnabled:   cfg.IsPatrolAlertTriggersEnabled(),
		AnomalyTriggersEnabled: cfg.IsPatrolAnomalyTriggersEnabled(),
	}
}

// StartPatrol starts the background patrol service
func (s *Service) StartPatrol(ctx context.Context) {
	s.mu.RLock()
	patrol := s.patrolService
	alertAnalyzer := s.alertTriggeredAnalyzer
	cfg := s.cfg
	licenseChecker := s.licenseChecker
	s.mu.RUnlock()

	if patrol == nil {
		log.Debug().Msg("patrol service not initialized, cannot start")
		return
	}

	if cfg == nil || !cfg.IsPatrolEnabled() {
		// Demo runtimes still start the loop: the release demo instance boots
		// with mock fixtures off until the license sync authorizes them, and
		// each tick re-checks IsDemoMode() live, so the loop self-heals into
		// simulated patrol cycles once fixtures enable. Requires a loaded
		// config because the setup below dereferences it.
		if cfg == nil || !IsDemoRuntimeIntended() {
			log.Debug().Msg("AI Patrol not enabled")
			return
		}
		log.Info().Msg("AI Patrol: starting loop for demo runtime despite disabled config")
	}

	// Check license for Patrol fix actions (Pro only) - Patrol itself is free with BYOK
	if licenseChecker != nil && !licenseChecker.HasFeature(FeatureAIAutoFix) {
		log.Info().Msg("Patrol safe remediation requires Pulse Pro license - fixes will require manual approval")
	}

	// Configure patrol from AI config (preserve defaults for resource types not in AI config).
	patrolCfg := s.patrolConfigFromAIConfig(cfg)
	patrol.SetConfig(patrolCfg)
	patrol.SetProactiveMode(cfg.UseProactiveThresholds)
	patrol.SetEventTriggerConfig(patrolEventTriggerConfigFromAIConfig(cfg))

	if BackgroundAutomationDisabledForDev() {
		patrol.SetEventTriggerBlock(BackgroundAutomationEventTriggerBlock())
		if alertAnalyzer != nil {
			alertAnalyzer.SetEnabled(false)
		}
		log.Info().
			Str("env", DevDisableBackgroundAIEnv).
			Msg("Pulse dev background AI disabled; Patrol scheduler and alert-triggered AI are not started")
		// The dev guard protects provider quota; a simulated demo cycle costs
		// nothing, so mock-mode dev still gets a populated patrol surface.
		if IsDemoMode() {
			patrol.startDemoPatrolWarmup()
		}
		return
	}

	patrol.SetEventTriggerBlock(PatrolEventTriggerBlock{})
	patrol.Start(ctx)

	// Configure alert-triggered analyzer (also Pro-only)
	if alertAnalyzer != nil {
		// Only enable if licensed for AI Alerts
		enabled := cfg.IsAlertTriggeredAnalysisEnabled()
		if enabled && licenseChecker != nil && !licenseChecker.HasFeature(FeatureAIAlerts) {
			log.Info().Msg("AI Alert Analysis requires Pulse Pro license - alert-triggered analysis disabled")
			enabled = false
		}
		alertAnalyzer.SetEnabled(enabled)
		alertAnalyzer.Start() // Start cleanup goroutine
		log.Info().
			Bool("enabled", enabled).
			Msg("Alert-triggered AI analysis configured")
	}

	// In demo/mock mode, populate the patrol surface promptly with a
	// simulated cycle instead of waiting for the first scheduled tick.
	if IsDemoMode() {
		patrol.startDemoPatrolWarmup()
	}
}

// StopPatrol stops the background patrol service
func (s *Service) StopPatrol() {
	s.mu.RLock()
	patrol := s.patrolService
	alertAnalyzer := s.alertTriggeredAnalyzer
	s.mu.RUnlock()

	if patrol != nil {
		patrol.Stop()
	}
	if alertAnalyzer != nil {
		alertAnalyzer.Stop()
	}
}

// Stop stops all background services
func (s *Service) Stop() {
	s.StopPatrol()

	s.mu.Lock()
	store := s.resourceExportStore
	incidentStore := s.incidentStore
	costStore := s.costStore
	s.resourceExportStore = nil
	s.resourceExportStoreOrgID = ""
	defer s.mu.Unlock()

	if incidentStore != nil {
		incidentStore.SetResourceTimelineStore(nil)
	}

	if costStore != nil {
		if err := costStore.Flush(); err != nil {
			log.Warn().Err(err).Msg("failed to flush AI usage cost store")
		}
	}

	if store != nil {
		if err := store.Close(); err != nil {
			log.Warn().Err(err).Msg("failed to close unified resource export store")
		}
	}
}

// ReconfigurePatrol updates the patrol configuration without restarting
// Call this after changing patrol settings to apply them immediately
func (s *Service) ReconfigurePatrol() {
	s.mu.RLock()
	patrol := s.patrolService
	alertAnalyzer := s.alertTriggeredAnalyzer
	cfg := s.cfg
	licenseChecker := s.licenseChecker
	s.mu.RUnlock()

	if patrol == nil || cfg == nil {
		return
	}

	// Update patrol configuration (preserve defaults for resource types not in AI config).
	patrolCfg := s.patrolConfigFromAIConfig(cfg)
	patrol.SetConfig(patrolCfg)

	// Update proactive threshold mode
	patrol.SetProactiveMode(cfg.UseProactiveThresholds)
	patrol.SetEventTriggerConfig(patrolEventTriggerConfigFromAIConfig(cfg))

	if BackgroundAutomationDisabledForDev() {
		patrol.SetEventTriggerBlock(BackgroundAutomationEventTriggerBlock())
		patrol.Stop()
		if alertAnalyzer != nil {
			alertAnalyzer.SetEnabled(false)
		}
		log.Info().
			Str("env", DevDisableBackgroundAIEnv).
			Msg("Pulse dev background AI disabled; Patrol background automation stopped")
		return
	}

	patrol.SetEventTriggerBlock(PatrolEventTriggerBlock{})

	log.Info().
		Bool("enabled", patrolCfg.Enabled).
		Dur("interval", patrolCfg.Interval).
		Bool("proactiveThresholds", cfg.UseProactiveThresholds).
		Msg("Patrol configuration updated")

	// Update alert-triggered analyzer (re-check license on each config change)
	if alertAnalyzer != nil {
		enabled := cfg.IsAlertTriggeredAnalysisEnabled()
		// Re-check license - don't allow re-enabling without valid license
		if enabled && licenseChecker != nil && !licenseChecker.HasFeature(FeatureAIAlerts) {
			log.Debug().Msg("alert-triggered analysis requires Pulse Pro license - staying disabled")
			enabled = false
		}
		alertAnalyzer.SetEnabled(enabled)
	}
}

// enrichRequestFromFinding looks up a patrol finding by ID and enriches the request
// with the finding's context (node, resource ID, resource type, etc.)
// This ensures proper command routing when the AI helps fix a patrol finding.
// The function modifies the request in place.
func (s *Service) enrichRequestFromFinding(req *ExecuteRequest) {
	if req.FindingID == "" {
		return
	}

	s.mu.RLock()
	patrol := s.patrolService
	s.mu.RUnlock()

	if patrol == nil {
		log.Debug().Str("finding_id", req.FindingID).Msg("cannot enrich request - patrol service not available")
		return
	}

	findings := patrol.GetFindings()
	if findings == nil {
		log.Debug().Str("finding_id", req.FindingID).Msg("cannot enrich request - findings store not available")
		return
	}

	finding := findings.Get(req.FindingID)
	if finding == nil {
		log.Debug().Str("finding_id", req.FindingID).Msg("cannot enrich request - finding not found")
		return
	}

	// Ensure context map exists
	if req.Context == nil {
		req.Context = make(map[string]interface{})
	}

	// Inject finding context (only if not already set)
	if _, ok := req.Context["node"]; !ok && finding.Node != "" {
		req.Context["node"] = finding.Node
	}
	if _, ok := req.Context["guestName"]; !ok && finding.ResourceName != "" {
		req.Context["guestName"] = finding.ResourceName
	}
	if req.TargetID == "" && finding.ResourceID != "" {
		req.TargetID = finding.ResourceID
	}
	if req.TargetType == "" && finding.ResourceType != "" {
		req.TargetType = finding.ResourceType
	}

	// Also store the finding details for reference in context
	req.Context["finding_resource_id"] = finding.ResourceID
	req.Context["finding_resource_name"] = finding.ResourceName
	req.Context["finding_resource_type"] = finding.ResourceType
	if finding.Node != "" {
		req.Context["finding_node"] = finding.Node
	}

	log.Debug().
		Str("finding_id", req.FindingID).
		Str("node", finding.Node).
		Str("resource_id", finding.ResourceID).
		Str("resource_name", finding.ResourceName).
		Str("resource_type", finding.ResourceType).
		Msg("Enriched request with finding context")
}

// GuestInfo contains information about a guest (VM or container) found by VMID lookup
type GuestInfo struct {
	Node       string
	Name       string
	Type       string // semantic type: "system-container" or "vm"
	Technology string // implementation: "lxc", "qemu", etc.
	Instance   string // PVE instance ID (for multi-cluster disambiguation)
}

// lookupNodeForVMID looks up which node owns a given VMID using the state provider
// Returns the node name and guest name if found, empty strings otherwise
// If targetInstance is provided, only matches guests from that instance (for multi-cluster setups)
func (s *Service) lookupNodeForVMID(vmID int) (node string, guestName string, guestType string) {
	guests := s.lookupGuestsByVMID(vmID, "")
	if len(guests) == 1 {
		return guests[0].Node, guests[0].Name, guests[0].Type
	}
	if len(guests) > 1 {
		// Multiple matches - VMID collision across instances
		// Log warning and return first match (caller should use lookupGuestsByVMID with instance filter)
		log.Warn().
			Int("vmid", vmID).
			Int("matches", len(guests)).
			Msg("VMID collision detected - multiple guests with same VMID across instances")
		return guests[0].Node, guests[0].Name, guests[0].Type
	}
	return "", "", ""
}

// lookupGuestsByVMID finds all guests with the given VMID
// If targetInstance is non-empty, only returns guests from that instance
func (s *Service) lookupGuestsByVMID(vmID int, targetInstance string) []GuestInfo {
	s.mu.RLock()
	rs := s.readState
	s.mu.RUnlock()

	if rs == nil {
		return nil
	}

	var results []GuestInfo

	// Check containers via ReadState
	for _, ct := range rs.Containers() {
		if ct.VMID() == vmID {
			if targetInstance == "" || ct.Instance() == targetInstance {
				results = append(results, GuestInfo{
					Node:       ct.Node(),
					Name:       ct.Name(),
					Type:       "system-container",
					Technology: "lxc",
					Instance:   ct.Instance(),
				})
			}
		}
	}

	// Check VMs via ReadState
	for _, vm := range rs.VMs() {
		if vm.VMID() == vmID {
			if targetInstance == "" || vm.Instance() == targetInstance {
				results = append(results, GuestInfo{
					Node:       vm.Node(),
					Name:       vm.Name(),
					Type:       "vm",
					Technology: "qemu",
					Instance:   vm.Instance(),
				})
			}
		}
	}

	return results
}

// extractVMIDFromCommand parses pct/qm/vzdump commands to extract the VMID being targeted
// Returns the VMID, whether it requires node-specific routing, and whether found
// Some commands (like vzdump) can run from any cluster node, others (like pct exec) must run on the owning node
func extractVMIDFromCommand(command string) (vmID int, requiresOwnerNode bool, found bool) {
	// Commands that MUST run on the node that owns the guest
	// These interact directly with the container/VM runtime
	nodeSpecificPatterns := []string{
		// pct commands - match any pct subcommand followed by VMID
		// Covers: exec, enter, start, stop, shutdown, reboot, status, push, pull, mount, unmount, etc.
		`(?:^|\s)pct\s+\w+\s+(\d+)`,
		// qm commands - match any qm subcommand followed by VMID
		// Covers: start, stop, shutdown, reset, status, guest exec, monitor, etc.
		`(?:^|\s)qm\s+(?:guest\s+)?\w+\s+(\d+)`,
	}

	// Commands that can run from any cluster node (cluster-aware)
	// vzdump uses the cluster to route to the right node automatically
	clusterAwarePatterns := []string{
		`(?:^|\s)vzdump\s+(\d+)`,
		// pvesh commands can specify node in path, so we don't force routing
	}

	// Check node-specific commands first (higher priority)
	for _, pattern := range nodeSpecificPatterns {
		re := regexp.MustCompile(pattern)
		if matches := re.FindStringSubmatch(command); len(matches) > 1 {
			if v, err := strconv.Atoi(matches[1]); err == nil {
				return v, true, true
			}
		}
	}

	// Check cluster-aware commands (don't force node routing)
	for _, pattern := range clusterAwarePatterns {
		re := regexp.MustCompile(pattern)
		if matches := re.FindStringSubmatch(command); len(matches) > 1 {
			if v, err := strconv.Atoi(matches[1]); err == nil {
				return v, false, true
			}
		}
	}

	return 0, false, false
}

// formatApprovalNeededToolResult returns a structured tool marker for commands
// that require approval. The shared marker vocabulary lives in
// agentcapabilities so Assistant and external-agent bridges cannot drift.
func formatApprovalNeededToolResult(command, toolID, reason, approvalID string) string {
	return agentcapabilities.ApprovalRequiredToolMarker(
		command,
		toolID,
		reason,
		approvalID,
		"Ask the user to click the approval button shown in the UI.",
	)
}

func resolveRunCommandApprovalTarget(req ExecuteRequest, runOnHost bool, targetHost string) (targetType, targetID, targetName string) {
	normalizedTargetHost := strings.ToLower(strings.TrimSpace(targetHost))
	targetType = agentcapabilities.NormalizeActionTargetType(req.TargetType)
	targetID = strings.TrimSpace(req.TargetID)

	if runOnHost {
		targetType = agentcapabilities.ActionTargetTypeAgent
		if normalizedTargetHost != "" {
			targetID = normalizedTargetHost
			targetName = normalizedTargetHost
		}
	}

	if targetType == "" {
		targetType = agentcapabilities.ActionTargetTypeAgent
	}

	if targetType == agentcapabilities.ActionTargetTypeAgent && targetID == "" {
		if node, ok := req.Context["node"].(string); ok && node != "" {
			targetID = strings.ToLower(strings.TrimSpace(node))
		} else if node, ok := req.Context["hostname"].(string); ok && node != "" {
			targetID = strings.ToLower(strings.TrimSpace(node))
		} else if node, ok := req.Context["host_name"].(string); ok && node != "" {
			targetID = strings.ToLower(strings.TrimSpace(node))
		}
	}

	if targetName == "" {
		targetName = strings.TrimSpace(targetHost)
	}
	if targetName == "" {
		targetName = targetID
	}

	return targetType, targetID, targetName
}

func createRunCommandApprovalRecord(orgID, command, toolID string, req ExecuteRequest, runOnHost bool, targetHost, reason string) string {
	store := approval.GetStore()
	if store == nil {
		log.Debug().Msg("approval store not available, run_command approval will not be persisted")
		return ""
	}

	targetType, targetID, targetName := resolveRunCommandApprovalTarget(req, runOnHost, targetHost)
	approvalReq := &approval.ApprovalRequest{
		OrgID:      strings.TrimSpace(orgID),
		ToolID:     toolID,
		Command:    command,
		TargetType: targetType,
		TargetID:   targetID,
		TargetName: targetName,
		Context:    reason,
	}
	if err := store.CreateApproval(approvalReq); err != nil {
		log.Warn().Err(err).Msg("failed to create run_command approval record")
		return ""
	}

	return approvalReq.ID
}

// formatPolicyBlockedToolResult returns a structured tool marker for commands
// blocked by policy.
func formatPolicyBlockedToolResult(command, reason string) string {
	return agentcapabilities.PolicyBlockedToolMarker(command, reason)
}

func approvalNeededFromToolCall(req ExecuteRequest, tc providers.ToolCall, result string) (ApprovalNeededData, bool) {
	if !agentcapabilities.HasApprovalRequiredToolMarker(result) {
		return ApprovalNeededData{}, false
	}
	if tc.Name != agentcapabilities.LegacyAssistantRunCommandToolName {
		return ApprovalNeededData{}, false
	}

	parsed, parsedOK := agentcapabilities.ParseApprovalRequiredToolMarkerData(result)
	if parsedOK && strings.TrimSpace(parsed.Command) == "" {
		parsedOK = false
	}
	cmd, _ := tc.Input[agentcapabilities.LegacyAssistantCommandArgumentName].(string)
	runOnHost, _ := tc.Input[agentcapabilities.LegacyAssistantRunOnHostArgumentName].(bool)
	targetHost, _ := tc.Input[agentcapabilities.LegacyAssistantTargetHostArgumentName].(string)

	if targetHost == "" {
		if node, ok := req.Context["node"].(string); ok && node != "" {
			targetHost = node
		} else if node, ok := req.Context["hostname"].(string); ok && node != "" {
			targetHost = node
		} else if node, ok := req.Context["host_name"].(string); ok && node != "" {
			targetHost = node
		}
	}

	if parsedOK {
		cmd = parsed.Command
		toolID := strings.TrimSpace(parsed.ToolID)
		if toolID == "" {
			toolID = tc.ID
		}
		return ApprovalNeededData{
			Command:    cmd,
			ToolID:     toolID,
			ToolName:   tc.Name,
			RunOnHost:  runOnHost,
			TargetHost: targetHost,
			ApprovalID: strings.TrimSpace(parsed.ApprovalID),
		}, true
	}

	return ApprovalNeededData{
		Command:    cmd,
		ToolID:     tc.ID,
		ToolName:   tc.Name,
		RunOnHost:  runOnHost,
		TargetHost: targetHost,
	}, true
}

// LoadConfig loads the AI configuration and initializes the provider
func (s *Service) LoadConfig() error {
	s.mu.RLock()
	persistence := s.persistence
	orgID := s.orgID
	s.mu.RUnlock()

	if persistence == nil {
		s.mu.Lock()
		s.provider = nil
		s.cfg = nil
		s.mu.Unlock()
		return fmt.Errorf("Pulse Assistant config persistence unavailable")
	}

	cfg, err := persistence.LoadAIConfig()
	if err != nil {
		return fmt.Errorf("failed to load Pulse Assistant config: %w", err)
	}
	var providerClient providers.Provider
	var providerInitErr string

	// Don't initialize provider if AI is not enabled or not configured
	if cfg != nil && cfg.Enabled && cfg.IsConfigured() {
		providerClient, providerInitErr = buildConfiguredProvider(context.Background(), cfg, orgID, true)
	}

	s.mu.Lock()
	s.cfg = cfg
	s.provider = providerClient
	s.providerInitErr = providerInitErr
	s.mu.Unlock()

	return nil
}

// buildConfiguredProvider resolves the effective model and constructs the
// default provider client for an enabled, configured AI config. It returns a
// nil provider plus a redacted failure summary when either step fails — for
// example when model resolution needs the provider's live catalog and that
// server is still booting. Callers own storing the result under s.mu.
// stampModel controls whether the resolved model is written back to cfg.Model:
// LoadConfig owns the config instance it just loaded, while retry callers
// share s.cfg and must not mutate it without the lock.
func buildConfiguredProvider(ctx context.Context, cfg *config.AIConfig, orgID string, stampModel bool) (providers.Provider, string) {
	selectedModel, resolveErr := ResolveConfiguredModel(ctx, cfg)
	if resolveErr != nil {
		log.Warn().Err(resolveErr).Str("orgID", orgID).Msg("AI enabled but no effective provider model could be resolved")
		return nil, summarizePatrolRuntimeFailureDetail(resolveErr.Error(), false)
	}
	if stampModel {
		cfg.Model = selectedModel
	}
	selectedProvider, _ := config.ParseModelString(selectedModel)

	nextProvider, providerErr := providers.NewForModel(cfg, selectedModel)
	if providerErr != nil {
		log.Warn().
			Err(providerErr).
			Str("selected_model", selectedModel).
			Str("selected_provider", selectedProvider).
			Strs("configured_providers", cfg.GetConfiguredProviders()).
			Msg("AI enabled but selected provider could not be initialized")
		return nil, summarizePatrolRuntimeFailureDetail(providerErr.Error(), false)
	}

	log.Info().
		Str("provider", selectedProvider).
		Str("model", selectedModel).
		Str("control_level", cfg.GetControlLevel()).
		Bool("autonomous", cfg.IsAutonomous()).
		Msg("AI service initialized")
	return nextProvider, ""
}

// ProviderInitError returns the redacted reason the default provider could not
// be initialized, or "" when the provider is healthy or intentionally absent.
func (s *Service) ProviderInitError() string {
	s.mu.RLock()
	defer s.mu.RUnlock()
	return s.providerInitErr
}

// RetryProviderInit rebuilds the default provider when the config claims an
// enabled, configured provider but initialization previously failed. Provider
// construction can depend on a live model-catalog call, so a boot-time race
// (Pulse starting before a local Ollama server) must not strand the runtime
// until the next settings save; Patrol calls this once per scheduled run.
// Returns true when a usable provider is available afterwards.
func (s *Service) RetryProviderInit(ctx context.Context) bool {
	s.mu.RLock()
	cfg := s.cfg
	provider := s.provider
	orgID := s.orgID
	s.mu.RUnlock()

	if provider != nil {
		return true
	}
	if cfg == nil || !cfg.Enabled || !cfg.IsConfigured() {
		return false
	}

	nextProvider, initErr := buildConfiguredProvider(ctx, cfg, orgID, false)

	s.mu.Lock()
	defer s.mu.Unlock()
	// A concurrent LoadConfig may have replaced provider or cfg; never clobber
	// a healthy provider with a failed retry.
	if s.provider != nil {
		return true
	}
	if s.cfg != cfg {
		return s.provider != nil
	}
	s.provider = nextProvider
	s.providerInitErr = initErr
	return nextProvider != nil
}

// IsEnabled returns true if AI is enabled and configured
func (s *Service) IsEnabled() bool {
	// In demo mode, AI is always considered enabled (using mock backend)
	if IsDemoMode() {
		return true
	}
	s.mu.RLock()
	defer s.mu.RUnlock()
	if s.cfg != nil && s.cfg.Enabled && s.provider != nil {
		return true
	}
	return false
}

type patrolProviderFactorySetter interface {
	SetPatrolProviderFactory(func(modelStr string) (providers.StreamingProvider, error))
}

func (s *Service) configurePatrolProviderFactoryLocked() {
	setter, ok := s.chatService.(patrolProviderFactorySetter)
	if !ok {
		return
	}
	setter.SetPatrolProviderFactory(s.createPatrolProviderForModel)
}

func (s *Service) createPatrolProviderForModel(modelStr string) (providers.StreamingProvider, error) {
	s.mu.RLock()
	cfg := s.cfg
	defaultProvider := s.provider
	s.mu.RUnlock()

	providerName, _ := config.ParseModelString(modelStr)
	if providerName == config.AIProviderQuickstart {
		return nil, fmt.Errorf("quickstart provider is retired; configure a provider API key or Ollama")
	}

	provider, err := providers.NewForModel(cfg, modelStr)
	if err != nil {
		if streaming, ok := defaultProvider.(providers.StreamingProvider); ok {
			return streaming, nil
		}
		return nil, err
	}
	streaming, ok := provider.(providers.StreamingProvider)
	if !ok {
		return nil, fmt.Errorf("patrol provider %q does not support streaming", provider.Name())
	}
	return streaming, nil
}

// QuickAnalysis performs a lightweight AI analysis for simple decisions.
// This is used for things like determining if an alert should be auto-resolved.
// It uses a single-turn, no-tools call for efficiency.
func (s *Service) QuickAnalysis(ctx context.Context, req QuickAnalysisRequest) (string, error) {
	s.mu.RLock()
	provider := s.provider
	cfg := s.cfg
	costStore := s.costStore
	s.mu.RUnlock()

	if provider == nil {
		return "", fmt.Errorf("Pulse Assistant is not enabled or configured")
	}

	// Use a fast model for quick analysis if available
	model := ""
	if cfg != nil && cfg.PatrolModel != "" {
		model = cfg.PatrolModel
	}
	sanitizerModel := model
	if sanitizerModel == "" && cfg != nil {
		sanitizerModel = cfg.GetChatModel()
	}

	messages := []providers.Message{
		{
			Role:    "system",
			Content: "You are a brief, decisive infrastructure monitoring assistant. Give short, direct answers.",
		},
		{
			Role:    "user",
			Content: req.Prompt,
		},
	}

	executionID := strings.TrimSpace(req.ExecutionID)
	if executionID == "" {
		executionID = uuid.NewString()
	}

	chatReq := providers.ChatRequest{
		Messages:    messages,
		Model:       model,
		ExecutionID: executionID,
	}
	if sanitizer := s.requestSanitizerForModel(sanitizerModel); sanitizer != nil {
		chatReq = sanitizer(chatReq)
	}

	useCase := strings.TrimSpace(req.UseCase)
	if useCase == "" {
		useCase = "quick_analysis"
	}

	resp, err := provider.Chat(ctx, chatReq)
	if err != nil {
		return "", fmt.Errorf("Pulse Assistant analysis failed: %w", err)
	}

	// Record token usage in the operator-facing cost ledger so quick
	// analysis (alert auto-resolve, etc.) shows up alongside chat and
	// patrol spend. Recording happens before the empty-content guard
	// because the operator was billed regardless.
	if costStore != nil {
		providerName, _ := config.ParseModelString(model)
		if providerName == "" {
			providerName = provider.Name()
		}
		costStore.Record(cost.UsageEvent{
			Timestamp:                time.Now(),
			Provider:                 providerName,
			RequestModel:             model,
			ResponseModel:            resp.Model,
			UseCase:                  useCase,
			TargetType:               strings.TrimSpace(req.TargetType),
			InputTokens:              resp.InputTokens,
			OutputTokens:             resp.OutputTokens,
			CacheCreationInputTokens: resp.CacheCreationInputTokens,
			CacheReadInputTokens:     resp.CacheReadInputTokens,
		})
	}

	if resp.Content == "" {
		return "", fmt.Errorf("Pulse Assistant returned empty response")
	}

	return resp.Content, nil
}

func (s *Service) requestSanitizerForModel(model string) func(providers.ChatRequest) providers.ChatRequest {
	s.mu.RLock()
	urp := s.unifiedResourceProvider
	s.mu.RUnlock()
	return modelboundary.RequestSanitizerForModel(model, urp)
}

// GetConfig returns a copy of the current AI config
func (s *Service) GetConfig() *config.AIConfig {
	s.mu.RLock()
	defer s.mu.RUnlock()
	if s.cfg == nil {
		return nil
	}
	cfg := *s.cfg
	cfg.PatrolAutopilotAcknowledgements = clonePatrolAutopilotAcknowledgements(s.cfg.PatrolAutopilotAcknowledgements)
	cfg.PatrolAutopilotRevocations = append([]unifiedresources.PatrolAutopilotRevocation(nil), s.cfg.PatrolAutopilotRevocations...)
	if s.cfg.PatrolAutopilotActivation != nil {
		activation := *s.cfg.PatrolAutopilotActivation
		cfg.PatrolAutopilotActivation = &activation
	}
	return &cfg
}

// GetDebugContext returns debug information about what context would be sent to the AI
func (s *Service) GetDebugContext(req ExecuteRequest) map[string]interface{} {
	s.mu.RLock()
	rs := s.readState
	agentServer := s.agentServer
	cfg := s.cfg
	s.mu.RUnlock()

	result := make(map[string]interface{})

	// ReadState info
	result["has_state_provider"] = rs != nil
	if rs != nil {
		nodes := rs.Nodes()
		vms := rs.VMs()
		containers := rs.Containers()
		dockerHosts := rs.DockerHosts()
		hosts := rs.Hosts()
		pbsInstances := rs.PBSInstances()

		result["state_summary"] = map[string]interface{}{
			"nodes":         len(nodes),
			"vms":           len(vms),
			"containers":    len(containers),
			"docker_hosts":  len(dockerHosts),
			"hosts":         len(hosts),
			"pbs_instances": len(pbsInstances),
		}

		// List some VMs/containers for verification
		var vmNames []string
		for _, vm := range vms {
			vmNames = append(vmNames, fmt.Sprintf("%s (VMID:%d, node:%s)", vm.Name(), vm.VMID(), vm.Node()))
		}
		if len(vmNames) > 10 {
			vmNames = vmNames[:10]
		}
		result["sample_vms"] = vmNames

		var ctNames []string
		for _, ct := range containers {
			ctNames = append(ctNames, fmt.Sprintf("%s (VMID:%d, node:%s)", ct.Name(), ct.VMID(), ct.Node()))
		}
		if len(ctNames) > 10 {
			ctNames = ctNames[:10]
		}
		result["sample_containers"] = ctNames

		var hostNames []string
		for _, h := range hosts {
			hostNames = append(hostNames, h.Hostname())
		}
		result["host_names"] = hostNames

		var dockerHostNames []string
		for _, dh := range dockerHosts {
			dockerHostNames = append(dockerHostNames, fmt.Sprintf("%s (%d containers)", dh.Name(), dh.ChildCount()))
		}
		result["docker_host_names"] = dockerHostNames
	}

	// Agent info
	result["has_agent_server"] = agentServer != nil
	if agentServer != nil {
		agents := agentServer.GetConnectedAgents()
		var agentNames []string
		for _, a := range agents {
			agentNames = append(agentNames, a.Hostname)
		}
		result["connected_agents"] = agentNames
	}

	// Config info
	result["has_config"] = cfg != nil
	if cfg != nil {
		result["custom_context_length"] = len(cfg.CustomContext)
		if len(cfg.CustomContext) > 200 {
			result["custom_context_preview"] = cfg.CustomContext[:200] + "..."
		} else {
			result["custom_context_preview"] = cfg.CustomContext
		}
	}

	// Build and include the system prompt
	systemPrompt := s.buildSystemPrompt(req, "")
	result["system_prompt_length"] = len(systemPrompt)
	result["system_prompt"] = systemPrompt

	return result
}

// IsAutonomous returns true if autonomous mode is enabled AND licensed.
// Autonomous mode requires the ai_autofix license feature (Pro tier).
func (s *Service) IsAutonomous() bool {
	s.mu.RLock()
	defer s.mu.RUnlock()
	if s.cfg == nil || !s.cfg.IsAutonomous() {
		return false
	}
	// Autonomous mode requires Pro license with ai_autofix feature
	if s.licenseChecker != nil && !s.licenseChecker.HasFeature(FeatureAIAutoFix) {
		return false
	}
	return true
}

func (s *Service) isAutonomousForRequest(req ExecuteRequest) bool {
	if req.AutonomousMode != nil && !*req.AutonomousMode {
		return false
	}
	return s.IsAutonomous()
}

// ConversationMessage represents a message in conversation history
type ConversationMessage struct {
	Role    string `json:"role"` // "user" or "assistant"
	Content string `json:"content"`
}

// ExecuteRequest represents a request to execute an AI prompt
type ExecuteRequest struct {
	Prompt                 string                 `json:"prompt"`
	TargetType             string                 `json:"target_type,omitempty"` // "agent", "system-container", "vm"
	TargetID               string                 `json:"target_id,omitempty"`
	Context                map[string]interface{} `json:"context,omitempty"`                  // Current metrics, state, etc.
	SystemPrompt           string                 `json:"system_prompt,omitempty"`            // Override system prompt
	History                []ConversationMessage  `json:"history,omitempty"`                  // Previous conversation messages
	FindingID              string                 `json:"finding_id,omitempty"`               // If fixing a patrol finding, the ID to resolve
	Model                  string                 `json:"model,omitempty"`                    // Override model for this request (for user selection in chat)
	UseCase                string                 `json:"use_case,omitempty"`                 // "chat" or "patrol" - determines which default model to use
	AutonomousMode         *bool                  `json:"autonomous_mode,omitempty"`          // Per-request execution override; false keeps scoped handoffs approval-required
	RequireCommandApproval bool                   `json:"require_command_approval,omitempty"` // Force every run_command tool call through operator approval
}

func assistantContextScopeForExecuteRequest(req ExecuteRequest) string {
	if strings.TrimSpace(req.FindingID) != "" {
		return "patrol_finding"
	}
	if strings.TrimSpace(req.TargetType) != "" || strings.TrimSpace(req.TargetID) != "" {
		return "resource_context"
	}
	if len(req.Context) > 0 {
		return "operator_context"
	}
	return ""
}

// ExecuteResponse represents the AI's response
type ExecuteResponse struct {
	Content      string `json:"content"`
	Model        string `json:"model"`
	InputTokens  int    `json:"input_tokens"`
	OutputTokens int    `json:"output_tokens"`
	// Prompt-cache buckets summed across the loop's provider calls; disjoint
	// from InputTokens, see providers.ChatResponse.
	CacheCreationInputTokens int                  `json:"cache_creation_input_tokens,omitempty"`
	CacheReadInputTokens     int                  `json:"cache_read_input_tokens,omitempty"`
	ToolCalls                []ToolExecution      `json:"tool_calls"`        // Commands that were executed
	PendingApprovals         []ApprovalNeededData `json:"pending_approvals"` // Commands that require approval (non-streaming)
}

func EmptyExecuteResponse() ExecuteResponse {
	return ExecuteResponse{}.NormalizeCollections()
}

func (r ExecuteResponse) NormalizeCollections() ExecuteResponse {
	if r.ToolCalls == nil {
		r.ToolCalls = []ToolExecution{}
	}
	if r.PendingApprovals == nil {
		r.PendingApprovals = []ApprovalNeededData{}
	}
	return r
}

// ToolExecution represents a tool that was executed during the AI conversation
type ToolExecution struct {
	Name    string `json:"name"`
	Input   string `json:"input"`  // Human-readable input (e.g., the command)
	Output  string `json:"output"` // Result of execution
	Success bool   `json:"success"`
}

// getModelForRequest determines which model to use for a request
// Priority: explicit override > use case default > config default
func (s *Service) getModelForRequest(req ExecuteRequest) string {
	// If request has explicit model override, use it
	if req.Model != "" {
		return req.Model
	}

	s.mu.RLock()
	cfg := s.cfg
	s.mu.RUnlock()

	if cfg == nil {
		return ""
	}

	// Use case-specific model selection
	switch req.UseCase {
	case "patrol":
		return cfg.GetPatrolModel()
	case "chat":
		return cfg.GetChatModel()
	default:
		// Default to chat model for interactive requests
		return cfg.GetChatModel()
	}
}

// StreamEvent represents an event during AI execution for streaming
type StreamEvent struct {
	Type string      `json:"type"` // "tool_start", "tool_end", "content", "done", "error", "approval_needed"
	Data interface{} `json:"data,omitempty"`
}

// StreamCallback is called for each event during streaming execution
type StreamCallback func(event StreamEvent)

// ToolStartData is sent when a tool execution begins
type ToolStartData struct {
	Name  string `json:"name"`
	Input string `json:"input"`
}

// ToolEndData is sent when a tool execution completes
type ToolEndData struct {
	Name    string `json:"name"`
	Input   string `json:"input"`
	Output  string `json:"output"`
	Success bool   `json:"success"`
}

// ApprovalNeededData is sent when a command needs user approval
type ApprovalNeededData struct {
	Command    string `json:"command"`
	ToolID     string `json:"tool_id"`   // ID to reference when approving
	ToolName   string `json:"tool_name"` // "run_command"
	RunOnHost  bool   `json:"run_on_host"`
	TargetHost string `json:"target_host,omitempty"` // Explicit host to route to
	ApprovalID string `json:"approval_id,omitempty"` // ID of the approval record in the store
}

// Execute sends a prompt to the AI and returns the response
// If tools are available and the AI requests them, it executes them in a loop
func (s *Service) Execute(ctx context.Context, req ExecuteRequest) (*ExecuteResponse, error) {
	release, err := s.acquireExecutionSlot(ctx, req.UseCase)
	if err != nil {
		return nil, err
	}
	defer release()

	if err := s.enforceBudget(req.UseCase); err != nil {
		return nil, err
	}

	// Enrich request with finding context if this is a "help me fix" request
	s.enrichRequestFromFinding(&req)

	s.mu.RLock()
	defaultProvider := s.provider
	agentServer := s.agentServer
	cfg := s.cfg
	costStore := s.costStore
	s.mu.RUnlock()

	// Determine the model to use for this request
	modelString := s.getModelForRequest(req)
	requestSanitizer := s.requestSanitizerForModel(modelString)

	// Create a provider for this specific model (supports multi-provider switching)
	provider, err := providers.NewForModel(cfg, modelString)
	if err != nil {
		// Fall back to default provider if model-specific provider can't be created
		log.Debug().Err(err).Str("model", modelString).Msg("could not create provider for model, using default")
		provider = defaultProvider
	}

	if provider == nil {
		// In demo mode, return mock response if no provider is configured
		if IsDemoMode() {
			return GenerateDemoAIResponse(req.Prompt), nil
		}
		return nil, fmt.Errorf("Pulse Assistant is not enabled or configured")
	}

	// Build the system prompt
	systemPrompt := req.SystemPrompt
	if systemPrompt == "" {
		systemPrompt = s.buildSystemPrompt(req, modelString)
	}

	// Check if agent is available for this target
	hasAgent := s.hasAgentForTarget(req)

	// Build tools list if agent is available
	var tools []providers.Tool
	if hasAgent && agentServer != nil {
		tools = s.getTools()
		systemPrompt += `

## Available Tools
You have access to tools to execute commands on the target system. You should:
1. Use run_command to investigate issues, gather information, and PERFORM actions
2. Actually execute the commands - don't just explain what commands to run
3. For Proxmox operations (resize disk, manage containers/VMs), run commands on the AGENT (target_type=agent)
4. For operations inside a system container, use target_type=system-container

Examples of actions you can perform:
- Resize LXC disk: pct resize <vmid> rootfs +10G (run on host)
- Check disk usage: df -h (run on container)
- View processes: ps aux --sort=-%mem | head -20
- Check logs: tail -100 /var/log/syslog

Always execute the commands rather than telling the user how to do it.`
	}

	// Inject previously learned knowledge about this guest
	if s.knowledgeStore != nil {
		guestID := s.getGuestID(req)
		if guestID != "" {
			if knowledgeContext := s.knowledgeStore.FormatForContext(guestID); knowledgeContext != "" {
				systemPrompt += knowledgeContext
			}
		}
	}

	// Build initial messages with conversation history
	var messages []providers.Message
	for _, histMsg := range req.History {
		messages = append(messages, providers.Message{
			Role:    histMsg.Role,
			Content: histMsg.Content,
		})
	}
	messages = append(messages, providers.Message{Role: "user", Content: req.Prompt})

	var toolExecutions []ToolExecution
	totalInputTokens := 0
	totalOutputTokens := 0
	totalCacheCreationTokens := 0
	totalCacheReadTokens := 0
	var finalContent string
	var model string

	// Agentic loop - keep going while AI requests tools within the shared
	// provider-call budget.
	loopGuard := newServiceToolLoopGuard()
	for {
		if err := s.enforceBudget(req.UseCase); err != nil {
			return nil, err
		}
		if err := loopGuard.beforeProviderCall(); err != nil {
			return nil, err
		}

		chatReq := providers.ChatRequest{
			Messages:  messages,
			Model:     modelString,
			System:    systemPrompt,
			MaxTokens: 4096,
			Tools:     tools,
		}
		if requestSanitizer != nil {
			chatReq = requestSanitizer(chatReq)
		}

		resp, err := provider.Chat(ctx, chatReq)
		if err != nil {
			return nil, fmt.Errorf("Pulse Assistant request failed: %w", err)
		}

		if costStore != nil {
			providerName, _ := config.ParseModelString(modelString)
			if providerName == "" {
				providerName = provider.Name()
			}
			costStore.Record(cost.UsageEvent{
				Timestamp:                time.Now(),
				Provider:                 providerName,
				RequestModel:             modelString,
				ResponseModel:            resp.Model,
				UseCase:                  req.UseCase,
				ContextScope:             assistantContextScopeForExecuteRequest(req),
				InputTokens:              resp.InputTokens,
				OutputTokens:             resp.OutputTokens,
				CacheCreationInputTokens: resp.CacheCreationInputTokens,
				CacheReadInputTokens:     resp.CacheReadInputTokens,
				TargetType:               req.TargetType,
				TargetID:                 req.TargetID,
				FindingID:                req.FindingID,
			})
		}

		totalInputTokens += resp.InputTokens
		totalOutputTokens += resp.OutputTokens
		totalCacheCreationTokens += resp.CacheCreationInputTokens
		totalCacheReadTokens += resp.CacheReadInputTokens
		model = resp.Model
		finalContent = resp.Content

		// If no tool calls, we're done
		if len(resp.ToolCalls) == 0 || resp.StopReason != "tool_use" {
			break
		}
		if err := s.observeDeniedServiceToolCalls(loopGuard, resp.ToolCalls); err != nil {
			return nil, err
		}

		// Add assistant's response with tool calls to messages
		messages = append(messages, providers.Message{
			Role:             "assistant",
			Content:          resp.Content,
			ReasoningContent: resp.ReasoningContent, // DeepSeek thinking mode
			ToolCalls:        resp.ToolCalls,
		})

		// Execute each tool call and add results
		var pendingApprovals []ApprovalNeededData
		for _, tc := range resp.ToolCalls {
			result, execution := s.executeTool(ctx, req, tc)
			toolExecutions = append(toolExecutions, execution)

			if approval, ok := approvalNeededFromToolCall(req, tc, result); ok {
				pendingApprovals = append(pendingApprovals, approval)
				continue
			}

			// Add tool result to messages
			projection := newLegacyServiceProviderToolResultContextProjection(tc.ID, result, !execution.Success)
			messages = append(messages, providers.Message{
				Role:       "user",
				ToolResult: &projection.Model,
			})

		}

		// Stop the agentic loop when approvals are required.
		// The caller can execute approvals via /api/ai/run-command and continue.
		if len(pendingApprovals) > 0 {
			return &ExecuteResponse{
				Content:                  finalContent,
				Model:                    model,
				InputTokens:              totalInputTokens,
				OutputTokens:             totalOutputTokens,
				CacheCreationInputTokens: totalCacheCreationTokens,
				CacheReadInputTokens:     totalCacheReadTokens,
				ToolCalls:                toolExecutions,
				PendingApprovals:         pendingApprovals,
			}, nil
		}
	}

	return &ExecuteResponse{
		Content:                  finalContent,
		Model:                    model,
		InputTokens:              totalInputTokens,
		OutputTokens:             totalOutputTokens,
		CacheCreationInputTokens: totalCacheCreationTokens,
		CacheReadInputTokens:     totalCacheReadTokens,
		ToolCalls:                toolExecutions,
	}, nil
}

// ExecuteStream sends a prompt to the AI and streams events via callback
// This allows the UI to show real-time progress during tool execution
func (s *Service) ExecuteStream(ctx context.Context, req ExecuteRequest, callback StreamCallback) (*ExecuteResponse, error) {
	release, err := s.acquireExecutionSlot(ctx, req.UseCase)
	if err != nil {
		return nil, err
	}
	defer release()

	if err := s.enforceBudget(req.UseCase); err != nil {
		return nil, err
	}

	// Enrich request with finding context if this is a "help me fix" request
	s.enrichRequestFromFinding(&req)

	s.mu.RLock()
	defaultProvider := s.provider
	agentServer := s.agentServer
	cfg := s.cfg
	costStore := s.costStore
	s.mu.RUnlock()

	// Determine the model to use for this request
	modelString := s.getModelForRequest(req)
	requestSanitizer := s.requestSanitizerForModel(modelString)

	// Create a provider for this specific model (supports multi-provider switching)
	provider, err := providers.NewForModel(cfg, modelString)
	if err != nil {
		// Fall back to default provider if model-specific provider can't be created
		log.Debug().Err(err).Str("model", modelString).Msg("could not create provider for model, using default")
		provider = defaultProvider
	}

	if provider == nil {
		// In demo mode, simulate streaming response if no provider is configured
		if IsDemoMode() {
			return GenerateDemoAIStream(req.Prompt, callback)
		}
		return nil, fmt.Errorf("Pulse Assistant is not enabled or configured")
	}

	// Build the system prompt
	systemPrompt := req.SystemPrompt
	if systemPrompt == "" {
		systemPrompt = s.buildSystemPrompt(req, modelString)
	}

	// Debug log the system prompt length and key sections
	log.Debug().
		Int("prompt_length", len(systemPrompt)).
		Bool("has_infrastructure_map", strings.Contains(systemPrompt, "## Infrastructure Map")).
		Bool("has_docker_hosts", strings.Contains(systemPrompt, "### Docker Hosts")).
		Bool("has_standalone_hosts", strings.Contains(systemPrompt, "### Standalone Hosts")).
		Bool("has_guests", strings.Contains(systemPrompt, "### All Guests")).
		Msg("AI system prompt built")

	// Check if agent is available for this target
	hasAgent := s.hasAgentForTarget(req)

	// Build tools list if agent is available
	var tools []providers.Tool
	if hasAgent && agentServer != nil {
		tools = s.getTools()
		systemPrompt += `

## Available Tools
You have access to tools to execute commands on the target system. You should:
1. Use run_command to investigate issues, gather information, and PERFORM actions
2. Actually execute the commands - don't just explain what commands to run
3. For Proxmox operations (resize disk, manage containers/VMs), run commands on the AGENT (target_type=agent)
4. For operations inside a system container, use target_type=system-container

Examples of actions you can perform:
- Resize LXC disk: pct resize <vmid> rootfs +10G (run on host)
- Check disk usage: df -h (run on container)
- View processes: ps aux --sort=-%mem | head -20
- Check logs: tail -100 /var/log/syslog

Always execute the commands rather than telling the user how to do it.`
	}

	// Inject previously learned knowledge about this guest
	if s.knowledgeStore != nil {
		guestID := s.getGuestID(req)
		if guestID != "" {
			if knowledgeContext := s.knowledgeStore.FormatForContext(guestID); knowledgeContext != "" {
				log.Debug().
					Str("guest_id", guestID).
					Int("context_length", len(knowledgeContext)).
					Msg("Injecting saved knowledge into AI context")
				systemPrompt += knowledgeContext
			} else {
				log.Debug().Str("guest_id", guestID).Msg("no saved knowledge for guest")
			}
		}
	}

	// Build initial messages with conversation history
	var messages []providers.Message
	for _, histMsg := range req.History {
		messages = append(messages, providers.Message{
			Role:    histMsg.Role,
			Content: histMsg.Content,
		})
	}
	messages = append(messages, providers.Message{Role: "user", Content: req.Prompt})

	var toolExecutions []ToolExecution
	totalInputTokens := 0
	totalOutputTokens := 0
	totalCacheCreationTokens := 0
	totalCacheReadTokens := 0
	var finalContent string
	var model string

	// Agentic loop - keep going while AI requests tools within the same
	// provider-call budget as non-streaming execution.
	loopGuard := newServiceToolLoopGuard()
	for {
		if err := s.enforceBudget(req.UseCase); err != nil {
			callback(StreamEvent{Type: "error", Data: map[string]string{"message": err.Error()}})
			return nil, err
		}
		if err := loopGuard.beforeProviderCall(); err != nil {
			emitServiceToolLoopFailure(callback, err)
			return nil, err
		}
		iteration := loopGuard.providerCalls
		log.Debug().
			Int("iteration", iteration).
			Int("message_count", len(messages)).
			Int("system_prompt_length", len(systemPrompt)).
			Int("tools_count", len(tools)).
			Msg("Calling AI provider...")

		// Send a processing event so the frontend knows we're making an AI call
		// This is especially important after tool execution when the next AI call can take a while
		if iteration > 1 {
			callback(StreamEvent{Type: "processing", Data: fmt.Sprintf("Analyzing results (iteration %d)...", iteration)})
		}

		chatReq := providers.ChatRequest{
			Messages:  messages,
			Model:     modelString,
			System:    systemPrompt,
			MaxTokens: 4096,
			Tools:     tools,
		}
		if requestSanitizer != nil {
			chatReq = requestSanitizer(chatReq)
		}

		resp, err := provider.Chat(ctx, chatReq)
		if err != nil {
			log.Error().Err(err).Int("iteration", iteration).Msg("AI provider call failed")
			callback(StreamEvent{Type: "error", Data: map[string]string{"message": err.Error()}})
			return nil, fmt.Errorf("Pulse Assistant request failed: %w", err)
		}

		if costStore != nil {
			providerName, _ := config.ParseModelString(modelString)
			if providerName == "" {
				providerName = provider.Name()
			}
			costStore.Record(cost.UsageEvent{
				Timestamp:                time.Now(),
				Provider:                 providerName,
				RequestModel:             modelString,
				ResponseModel:            resp.Model,
				UseCase:                  req.UseCase,
				ContextScope:             assistantContextScopeForExecuteRequest(req),
				InputTokens:              resp.InputTokens,
				OutputTokens:             resp.OutputTokens,
				CacheCreationInputTokens: resp.CacheCreationInputTokens,
				CacheReadInputTokens:     resp.CacheReadInputTokens,
				TargetType:               req.TargetType,
				TargetID:                 req.TargetID,
				FindingID:                req.FindingID,
			})
		}

		log.Debug().Int("iteration", iteration).Msg("AI provider returned successfully")

		totalInputTokens += resp.InputTokens
		totalOutputTokens += resp.OutputTokens
		totalCacheCreationTokens += resp.CacheCreationInputTokens
		totalCacheReadTokens += resp.CacheReadInputTokens
		model = resp.Model
		finalContent = resp.Content

		// Stream thinking/reasoning content if present (DeepSeek reasoner)
		if resp.ReasoningContent != "" {
			callback(StreamEvent{Type: "thinking", Data: resp.ReasoningContent})
		}

		// Stream intermediate content so users see the AI's explanations between tool calls
		// This gives users visibility into the AI's reasoning as it works, not just at the end
		if resp.Content != "" {
			callback(StreamEvent{Type: "content", Data: resp.Content})
		}

		log.Debug().
			Int("tool_calls", len(resp.ToolCalls)).
			Str("stop_reason", resp.StopReason).
			Int("iteration", iteration).
			Int("total_input_tokens", totalInputTokens).
			Int("total_output_tokens", totalOutputTokens).
			Int("content_length", len(resp.Content)).
			Bool("has_content", resp.Content != "").
			Msg("AI streaming iteration complete")

		// If no tool calls, we're done
		if len(resp.ToolCalls) == 0 || resp.StopReason != "tool_use" {
			log.Info().
				Int("tool_calls", len(resp.ToolCalls)).
				Str("stop_reason", resp.StopReason).
				Int("iteration", iteration).
				Msg("AI streaming loop ending - no more tool calls or stop_reason != tool_use")
			break
		}
		if err := s.observeDeniedServiceToolCalls(loopGuard, resp.ToolCalls); err != nil {
			emitServiceToolLoopFailure(callback, err)
			return nil, err
		}

		// Add assistant's response with tool calls to messages
		messages = append(messages, providers.Message{
			Role:             "assistant",
			Content:          resp.Content,
			ReasoningContent: resp.ReasoningContent, // DeepSeek thinking mode
			ToolCalls:        resp.ToolCalls,
		})

		// Execute each tool call and add results
		// Track if any command needs approval - if so, we'll stop the loop after processing
		anyNeedsApproval := false
		for _, tc := range resp.ToolCalls {
			toolInput := s.getToolInputDisplay(tc)
			class := agentcapabilities.ClassifyLegacyAssistantInvocation(tc.Name)
			if !s.legacyAssistantInvocationPolicy().Allows(tc.Name, class) {
				result := agentcapabilities.ToolResultText(agentcapabilities.NewInvocationBlockedToolResult(tc.Name, class))
				execution := ToolExecution{Name: tc.Name, Input: toolInput, Output: result, Success: false}
				toolExecutions = append(toolExecutions, execution)
				callback(StreamEvent{Type: "tool_start", Data: ToolStartData{Name: tc.Name, Input: toolInput}})
				callback(StreamEvent{Type: "tool_end", Data: ToolEndData{Name: tc.Name, Input: toolInput, Output: result, Success: false}})
				projection := newLegacyServiceProviderToolResultContextProjection(tc.ID, result, true)
				messages = append(messages, providers.Message{Role: "user", ToolResult: &projection.Model})
				continue
			}

			// Check if this command needs approval
			needsApproval := false
			approvalID := ""
			approvalReason := "Command requires user approval"
			if tc.Name == agentcapabilities.LegacyAssistantRunCommandToolName {
				cmd, _ := tc.Input[agentcapabilities.LegacyAssistantCommandArgumentName].(string)
				runOnHost, _ := tc.Input[agentcapabilities.LegacyAssistantRunOnHostArgumentName].(bool)
				targetHost, _ := tc.Input[agentcapabilities.LegacyAssistantTargetHostArgumentName].(string)
				s.mu.RLock()
				approvalOrgID := s.orgID
				s.mu.RUnlock()

				// If AI didn't specify target_host, try to get it from request context
				// This is crucial for proper routing when the command is approved
				if targetHost == "" {
					if node, ok := req.Context["node"].(string); ok && node != "" {
						targetHost = node
					} else if node, ok := req.Context["hostname"].(string); ok && node != "" {
						targetHost = node
					} else if node, ok := req.Context["host_name"].(string); ok && node != "" {
						targetHost = node
					}
				}

				isAuto := s.isAutonomousForRequest(req)
				policyDecision := s.policy.Evaluate(cmd)
				log.Debug().
					Bool("autonomous", isAuto).
					Str("policy_decision", string(policyDecision)).
					Str("command", cmd).
					Str("target_host", targetHost).
					Msg("Checking command policy/approval")

				// Always block commands blocked by policy (even in autonomous mode).
				if policyDecision == agentexec.PolicyBlock {
					result := formatPolicyBlockedToolResult(cmd, "This command is blocked by security policy")
					execution := ToolExecution{
						Name:    tc.Name,
						Input:   toolInput,
						Output:  result,
						Success: false,
					}
					toolExecutions = append(toolExecutions, execution)

					callback(StreamEvent{
						Type: "tool_start",
						Data: ToolStartData{Name: tc.Name, Input: toolInput},
					})
					callback(StreamEvent{
						Type: "tool_end",
						Data: ToolEndData{Name: tc.Name, Input: toolInput, Output: result, Success: false},
					})

					projection := newLegacyServiceProviderToolResultContextProjection(tc.ID, result, true)
					messages = append(messages, providers.Message{
						Role:       "user",
						ToolResult: &projection.Model,
					})
					continue
				}

				if req.RequireCommandApproval || isAuto || policyDecision == agentexec.PolicyRequireApproval {
					needsApproval = true
					anyNeedsApproval = true
					approvalReason = "Security policy requires approval"
					if req.RequireCommandApproval {
						approvalReason = "This handoff requires operator approval before command execution"
					} else if isAuto {
						approvalReason = "Autonomous model sessions cannot dispatch raw commands; operator approval is required"
					}
					approvalID = createRunCommandApprovalRecord(
						approvalOrgID,
						cmd,
						tc.ID,
						req,
						runOnHost,
						targetHost,
						approvalReason,
					)
					callback(StreamEvent{
						Type: "approval_needed",
						Data: ApprovalNeededData{
							Command:    cmd,
							ToolID:     tc.ID,
							ToolName:   tc.Name,
							RunOnHost:  runOnHost,
							TargetHost: targetHost,
							ApprovalID: approvalID,
						},
					})
				}
			}

			var result string
			var execution ToolExecution

			if needsApproval {
				// Don't execute - command needs user approval
				// We'll break out of the loop after processing all tool calls
				// Note: We don't add to toolExecutions here because the approval_needed event
				// already tells the frontend to show the approval UI
				cmd, _ := tc.Input[agentcapabilities.LegacyAssistantCommandArgumentName].(string)
				result = formatApprovalNeededToolResult(cmd, tc.ID, approvalReason, approvalID)
				execution = ToolExecution{
					Name:    tc.Name,
					Input:   toolInput,
					Output:  result,
					Success: true, // Not an error; awaiting approval
				}
			} else {
				// Stream tool start event
				callback(StreamEvent{
					Type: "tool_start",
					Data: ToolStartData{Name: tc.Name, Input: toolInput},
				})

				result, execution = s.executeTool(ctx, req, tc)
				toolExecutions = append(toolExecutions, execution)

				// Stream tool end event
				callback(StreamEvent{
					Type: "tool_end",
					Data: ToolEndData{Name: tc.Name, Input: toolInput, Output: result, Success: execution.Success},
				})

				// Log to remediation log for operational memory
				// Only log run_command since that's the main remediation action
				if tc.Name == agentcapabilities.LegacyAssistantRunCommandToolName {
					s.logRemediation(req, toolInput, result, execution.Success)
				}
			}

			// Add tool result to messages
			projection := newLegacyServiceProviderToolResultContextProjection(tc.ID, result, !execution.Success)
			messages = append(messages, providers.Message{
				Role:       "user",
				ToolResult: &projection.Model,
			})
		}

		// If any command needed approval, stop the agentic loop here.
		// Don't call the AI again with "COMMAND_BLOCKED" results - this causes duplicate
		// approval requests and confusing "click the button" messages.
		// The frontend will show approval buttons, and user action will continue the conversation.
		if anyNeedsApproval {
			log.Info().
				Int("pending_approvals", len(resp.ToolCalls)).
				Int("iteration", iteration).
				Msg("Stopping AI loop - commands need user approval")
			// Use the AI's current response as final content (if any)
			// This preserves any explanation the AI provided before requesting the command
			break
		}
	}

	// Don't stream finalContent here - it was already streamed in the iteration above
	// Sending it again causes duplicate responses (issue #947)
	callback(StreamEvent{Type: "done"})

	return &ExecuteResponse{
		Content:                  finalContent,
		Model:                    model,
		InputTokens:              totalInputTokens,
		OutputTokens:             totalOutputTokens,
		CacheCreationInputTokens: totalCacheCreationTokens,
		CacheReadInputTokens:     totalCacheReadTokens,
		ToolCalls:                toolExecutions,
	}, nil
}

// getToolInputDisplay returns a human-readable display of tool input
func (s *Service) getToolInputDisplay(tc providers.ToolCall) string {
	switch tc.Name {
	case agentcapabilities.LegacyAssistantRunCommandToolName:
		cmd, _ := tc.Input[agentcapabilities.LegacyAssistantCommandArgumentName].(string)
		if runOnHost, ok := tc.Input[agentcapabilities.LegacyAssistantRunOnHostArgumentName].(bool); ok && runOnHost {
			return fmt.Sprintf("[host] %s", cmd)
		}
		return cmd
	case agentcapabilities.LegacyAssistantFetchURLToolName:
		fetchURL, _ := tc.Input[agentcapabilities.LegacyAssistantURLArgumentName].(string)
		return fetchURL
	case agentcapabilities.LegacyAssistantSetResourceURLToolName:
		resourceType, _ := tc.Input[agentcapabilities.LegacyAssistantResourceTypeArgumentName].(string)
		resourceURL, _ := tc.Input[agentcapabilities.LegacyAssistantURLArgumentName].(string)
		return fmt.Sprintf("Set %s URL: %s", resourceType, resourceURL)
	default:
		return fmt.Sprintf("%v", tc.Input)
	}
}

// logRemediation logs a tool execution to the remediation log for operational memory
// This enables learning from past fix attempts
// ONLY logs commands that perform actual ACTIONS (restarts, resizes, cleanups, fixes)
// Skips diagnostic commands (df, grep, cat, tail, ps) to avoid noise
func (s *Service) logRemediation(req ExecuteRequest, command, output string, success bool) {
	// First, check if this is an actionable command worth logging
	// Diagnostic/read-only commands don't provide value in the achievement log
	if !isActionableCommand(command) {
		log.Debug().
			Str("command", command).
			Msg("Skipping diagnostic command - not logging to remediation history")
		return
	}

	s.mu.RLock()
	patrol := s.patrolService
	s.mu.RUnlock()

	if patrol == nil {
		return
	}

	remLog := patrol.GetRemediationLog()
	if remLog == nil {
		return
	}

	// Determine outcome
	outcome := OutcomeUnknown
	if success {
		outcome = OutcomeResolved
	} else {
		outcome = OutcomeFailed
	}

	// Extract problem from the original prompt (first 200 chars as summary)
	problem := req.Prompt
	if len(problem) > 200 {
		problem = problem[:200] + "..."
	}

	// Generate a meaningful summary from the command that describes what was achieved
	// This is what gets displayed in the Pulse Intelligence Impact section
	summary := generateRemediationSummary(command, req.TargetType, req.Context)

	// Get resource name from context if available
	resourceName := ""
	if req.Context != nil {
		if name, ok := req.Context["name"].(string); ok {
			resourceName = name
		}
	}

	// Truncate output for storage
	truncatedOutput := output
	const maxOutputLen = 1000
	if len(truncatedOutput) > maxOutputLen {
		truncatedOutput = truncatedOutput[:maxOutputLen] + "..."
	}

	if err := remLog.Log(RemediationRecord{
		ResourceID:   req.TargetID,
		ResourceType: req.TargetType,
		ResourceName: resourceName,
		FindingID:    req.FindingID,
		Problem:      problem,
		Summary:      summary,
		Action:       command,
		Output:       truncatedOutput,
		Outcome:      outcome,
		Automatic:    req.UseCase == "patrol", // Patrol runs are automatic
	}); err != nil {
		log.Warn().Err(err).Str("resource_id", req.TargetID).Msg("Failed to log ACTION to Pulse Intelligence Impact")
	}

	log.Info().
		Str("resource_id", req.TargetID).
		Str("resource_name", resourceName).
		Str("command", command).
		Str("summary", summary).
		Bool("success", success).
		Msg("Logged ACTION to Pulse Intelligence Impact")
}

// isActionableCommand returns true if the command performs an actual action
// that's worth logging as an achievement, false for read-only diagnostics
func isActionableCommand(cmd string) bool {
	cmd = strings.TrimSpace(cmd)

	// Strip [host] or [hostname] prefix if present
	if strings.HasPrefix(cmd, "[") {
		if idx := strings.Index(cmd, "]"); idx != -1 {
			cmd = strings.TrimSpace(cmd[idx+1:])
		}
	}

	// Commands that PERFORM ACTIONS (should be logged)
	actionPatterns := []string{
		"docker restart", "docker start", "docker stop", "docker rm",
		"docker compose up", "docker compose down", "docker compose restart",
		"systemctl restart", "systemctl start", "systemctl stop", "systemctl enable", "systemctl disable",
		"service restart", "service start", "service stop",
		"pct resize", "pct start", "pct stop", "pct shutdown", "pct reboot",
		"qm resize", "qm start", "qm stop", "qm shutdown", "qm reboot",
		"rm -", "rm /", // File deletion/cleanup
		"chmod", "chown", // Permission fixes
		"mkdir",      // Creating directories
		"mv ", "cp ", // File operations
		"echo >", "tee ", // Writing to files
		"apt install", "apt upgrade", "apt remove",
		"yum install", "dnf install",
		"pip install", "npm install",
		"kill ", "pkill ", "killall ",
		"reboot", "shutdown",
	}

	for _, pattern := range actionPatterns {
		if strings.Contains(cmd, pattern) {
			return true
		}
	}

	// Everything else is diagnostic (df, grep, cat, tail, ps, ls, etc.)
	return false
}

// generateRemediationSummary creates a human-readable summary of what a command achieved
// This is used to display meaningful descriptions in the Pulse Intelligence Impact section
func generateRemediationSummary(command string, _ string, context map[string]interface{}) string {
	cmd := strings.TrimSpace(command)

	// Extract target name from context
	targetName := ""
	if context != nil {
		if name, ok := context["name"].(string); ok && name != "" {
			targetName = name
		} else if name, ok := context["containerName"].(string); ok && name != "" {
			targetName = name
		} else if name, ok := context["guestName"].(string); ok && name != "" {
			targetName = name
		}
	}

	// Extract meaningful path or service from command
	extractPath := func() string {
		// Look for common path patterns
		pathPatterns := []string{
			`/[\w/._-]+`, // Unix paths
		}
		for _, pattern := range pathPatterns {
			re := regexp.MustCompile(pattern)
			if match := re.FindString(cmd); match != "" {
				// Extract the meaningful part (last 2 segments)
				parts := strings.Split(strings.Trim(match, "/"), "/")
				if len(parts) > 2 {
					parts = parts[len(parts)-2:]
				}
				return "/" + strings.Join(parts, "/")
			}
		}
		return ""
	}

	// Extract container/service name from docker commands
	extractDockerTarget := func() string {
		// docker ps --filter name=XXX
		if match := regexp.MustCompile(`name[=:]\s*(\w+)`).FindStringSubmatch(cmd); len(match) > 1 {
			return match[1]
		}
		// docker restart XXX, docker start XXX, etc.
		if match := regexp.MustCompile(`docker\s+(?:restart|start|stop|logs)\s+([\w.-]+)`).FindStringSubmatch(cmd); len(match) > 1 {
			return match[1]
		}
		return ""
	}

	path := extractPath()
	dockerTarget := extractDockerTarget()

	// Generate summary based on command type
	switch {
	case strings.Contains(cmd, "docker restart") || strings.Contains(cmd, "docker start"):
		if dockerTarget != "" {
			return fmt.Sprintf("Restarted %s container", dockerTarget)
		}
		return "Restarted container"

	case strings.Contains(cmd, "docker stop"):
		if dockerTarget != "" {
			return fmt.Sprintf("Stopped %s container", dockerTarget)
		}
		return "Stopped container"

	case strings.Contains(cmd, "docker ps"):
		if dockerTarget != "" {
			return fmt.Sprintf("Verified %s container is running", dockerTarget)
		}
		return "Checked container status"

	case strings.Contains(cmd, "docker logs"):
		if dockerTarget != "" {
			return fmt.Sprintf("Retrieved %s logs", dockerTarget)
		}
		return "Retrieved container logs"

	case strings.Contains(cmd, "systemctl restart"):
		if match := regexp.MustCompile(`systemctl\s+restart\s+(\S+)`).FindStringSubmatch(cmd); len(match) > 1 {
			return fmt.Sprintf("Restarted %s service", match[1])
		}
		return "Restarted system service"

	case strings.Contains(cmd, "systemctl status"):
		if match := regexp.MustCompile(`systemctl\s+status\s+(\S+)`).FindStringSubmatch(cmd); len(match) > 1 {
			return fmt.Sprintf("Checked %s service status", match[1])
		}
		return "Checked service status"

	case strings.Contains(cmd, "df ") || strings.Contains(cmd, "du "):
		if path != "" {
			// Check for known services in path
			pathLower := strings.ToLower(path)
			if strings.Contains(pathLower, "frigate") {
				return "Analyzed Frigate storage usage"
			}
			if strings.Contains(pathLower, "plex") {
				return "Analyzed Plex storage usage"
			}
			if strings.Contains(pathLower, "recordings") {
				return "Analyzed recordings storage"
			}
			return fmt.Sprintf("Analyzed %s storage", path)
		}
		return "Analyzed disk usage"

	case strings.Contains(cmd, "grep") && (strings.Contains(cmd, "config") || strings.Contains(cmd, "conf") || strings.Contains(cmd, ".yml") || strings.Contains(cmd, ".yaml")):
		if path != "" && strings.Contains(strings.ToLower(path), "frigate") {
			return "Inspected Frigate configuration"
		}
		if path != "" {
			return fmt.Sprintf("Inspected %s configuration", path)
		}
		return "Inspected configuration"

	case strings.Contains(cmd, "tail") || strings.Contains(cmd, "journalctl"):
		if targetName != "" {
			return fmt.Sprintf("Reviewed %s logs", targetName)
		}
		return "Reviewed system logs"

	case strings.Contains(cmd, "pct resize"):
		if match := regexp.MustCompile(`pct\s+resize\s+(\d+)`).FindStringSubmatch(cmd); len(match) > 1 {
			return fmt.Sprintf("Resized container %s disk", match[1])
		}
		return "Resized container disk"

	case strings.Contains(cmd, "qm resize"):
		if match := regexp.MustCompile(`qm\s+resize\s+(\d+)`).FindStringSubmatch(cmd); len(match) > 1 {
			return fmt.Sprintf("Resized VM %s disk", match[1])
		}
		return "Resized VM disk"

	case strings.Contains(cmd, "ping") || strings.Contains(cmd, "curl"):
		return "Tested network connectivity"

	case strings.Contains(cmd, "free") || strings.Contains(cmd, "meminfo"):
		return "Checked memory usage"

	case strings.Contains(cmd, "ps aux") || strings.Contains(cmd, "top"):
		return "Analyzed running processes"

	case strings.Contains(cmd, "rm "):
		return "Cleaned up files"

	case strings.Contains(cmd, "chmod") || strings.Contains(cmd, "chown"):
		return "Fixed file permissions"

	default:
		// Generic fallback - try to use target name if available
		if targetName != "" {
			return fmt.Sprintf("Ran diagnostics on %s", targetName)
		}
		return "Ran system diagnostics"
	}
}

// hasAgentForTarget checks if we have an agent connection for the given target.
// This uses the same routing logic as command execution to determine if the target
// can be reached, including cluster peer routing for Proxmox clusters.
func (s *Service) hasAgentForTarget(req ExecuteRequest) bool {
	// Check for nil interface or nil underlying value
	// Note: A typed nil pointer assigned to an interface is NOT nil according to Go's == operator
	// We need to use reflection to check if the underlying value is nil
	if s.agentServer == nil {
		return false
	}
	// Check if the interface contains a typed nil pointer
	v := reflect.ValueOf(s.agentServer)
	if v.Kind() == reflect.Ptr && v.IsNil() {
		return false
	}

	agents := s.agentServer.GetConnectedAgents()
	if len(agents) == 0 {
		return false
	}

	// For agent targets with no specific context, any connected agent will do.
	if req.TargetType == "agent" && len(req.Context) == 0 {
		return true
	}

	// Try to determine the target node from the request context
	// This mirrors the logic in routeToAgent
	targetNode := ""

	// Check context fields for the target node
	hostFields := []string{"node", "host", "guest_node", "hostname", "host_name", "target_host"}
	for _, field := range hostFields {
		if value, ok := req.Context[field].(string); ok && value != "" {
			targetNode = strings.ToLower(value)
			break
		}
	}

	// If no target node found in context, try the unified resource provider
	if targetNode == "" {
		s.mu.RLock()
		urp := s.unifiedResourceProvider
		s.mu.RUnlock()

		if urp != nil {
			resourceName := ""
			if name, ok := req.Context["containerName"].(string); ok && name != "" {
				resourceName = name
			} else if name, ok := req.Context["name"].(string); ok && name != "" {
				resourceName = name
			} else if name, ok := req.Context["guestName"].(string); ok && name != "" {
				resourceName = name
			}

			if resourceName != "" {
				if host := urp.FindContainerHost(resourceName); host != "" {
					targetNode = strings.ToLower(host)
				}
			}
		}
	}

	// If we still don't have a target node, check for single agent scenario
	if targetNode == "" {
		// For unknown targets, we need at least one agent
		// The actual routing will determine which one to use
		return len(agents) >= 1
	}

	// Check if we have a direct agent match or a cluster peer
	for _, agent := range agents {
		if strings.EqualFold(agent.Hostname, targetNode) {
			return true
		}
	}

	// Try cluster peer routing (for Proxmox clusters)
	if peerAgentID := s.findClusterPeerAgent(targetNode, agents); peerAgentID != "" {
		return true
	}

	return false
}

// getTools returns the available tools for AI
func (s *Service) legacyAssistantInvocationPolicy() aitools.InvocationPolicy {
	s.mu.RLock()
	cfg := s.cfg
	s.mu.RUnlock()
	level := aitools.ControlLevelReadOnly
	if cfg != nil {
		level = aitools.ControlLevel(cfg.GetControlLevel())
	}
	return aitools.InvocationPolicy{
		ControlLevel:                level,
		HasExecuteAuthority:         true,
		ExecuteAuthorityBound:       true,
		DenyInfrastructureMutations: true,
		PulseStateAllowlist:         map[string]bool{},
		Profile:                     aitools.ProfileInteractiveAssistant,
	}
}

func (s *Service) getTools() []providers.Tool {
	candidates := append([]providers.Tool{}, agentcapabilities.LegacyAssistantUtilityProviderTools()...)
	candidates = append(candidates,
		providers.Tool{
			Name:        agentcapabilities.ResolveFindingCapabilityName,
			Description: "Mark an AI patrol finding as resolved after successfully fixing the issue. Use the finding ID shown in your Patrol Finding Context section. Call this after verifying the fix worked - do NOT ask the user for the finding ID.",
			InputSchema: legacyAssistantCapabilityInputSchema(agentcapabilities.ResolveFindingCapabilityName),
		},
		providers.Tool{
			Name:        agentcapabilities.DismissFindingCapabilityName,
			Description: "Dismiss an AI patrol finding when it's not actually an issue or is expected behavior. Use this instead of " + agentcapabilities.ResolveFindingCapabilityName + " when the finding is a false positive or the configuration is intentional. This creates a suppression rule to prevent similar findings from being raised again.",
			InputSchema: legacyAssistantCapabilityInputSchema(agentcapabilities.DismissFindingCapabilityName),
		},
	)

	// Add web search tool for Anthropic provider
	if s.provider != nil && s.provider.Name() == "anthropic" {
		candidates = append(candidates, providers.Tool{
			Type:    "web_search_20250305",
			Name:    "web_search",
			MaxUses: 3, // Limit searches per request to control costs
		})
	}

	policy := s.legacyAssistantInvocationPolicy()
	available := make([]providers.Tool, 0, len(candidates))
	for _, candidate := range candidates {
		if candidate.Name == "web_search" || policy.Allows(candidate.Name, agentcapabilities.ClassifyLegacyAssistantInvocation(candidate.Name)) {
			available = append(available, candidate)
		}
	}
	return available
}

// executeTool executes a tool call and returns the result
func (s *Service) executeTool(ctx context.Context, req ExecuteRequest, tc providers.ToolCall) (string, ToolExecution) {
	execution := ToolExecution{
		Name:    tc.Name,
		Success: false,
	}
	class := agentcapabilities.ClassifyLegacyAssistantInvocation(tc.Name)
	if !s.legacyAssistantInvocationPolicy().Allows(tc.Name, class) {
		execution.Output = agentcapabilities.ToolResultText(agentcapabilities.NewInvocationBlockedToolResult(tc.Name, class))
		return execution.Output, execution
	}

	switch tc.Name {
	case agentcapabilities.LegacyAssistantRunCommandToolName:
		command, _ := tc.Input[agentcapabilities.LegacyAssistantCommandArgumentName].(string)
		runOnHost, _ := tc.Input[agentcapabilities.LegacyAssistantRunOnHostArgumentName].(bool)
		targetHost, _ := tc.Input[agentcapabilities.LegacyAssistantTargetHostArgumentName].(string)
		execution.Input = command
		if runOnHost && targetHost != "" {
			execution.Input = fmt.Sprintf("[%s] %s", targetHost, command)
		} else if runOnHost {
			execution.Input = fmt.Sprintf("[host] %s", command)
		}

		if command == "" {
			execution.Output = "Error: command is required"
			return execution.Output, execution
		}

		// Enforce security policy.
		// - Blocked commands are ALWAYS blocked (even in autonomous mode).
		// - Approval-required commands only require approval when not in autonomous mode.
		decision := s.policy.Evaluate(command)
		if decision == agentexec.PolicyBlock {
			execution.Output = formatPolicyBlockedToolResult(command, "This command is blocked by security policy")
			return execution.Output, execution
		}
		isAuto := s.isAutonomousForRequest(req)
		if req.RequireCommandApproval || isAuto || decision == agentexec.PolicyRequireApproval {
			s.mu.RLock()
			approvalOrgID := s.orgID
			s.mu.RUnlock()
			approvalReason := "Security policy requires approval"
			if req.RequireCommandApproval {
				approvalReason = "This handoff requires operator approval before command execution"
			} else if isAuto {
				approvalReason = "Autonomous model sessions cannot dispatch raw commands; operator approval is required"
			}
			approvalID := createRunCommandApprovalRecord(
				approvalOrgID,
				command,
				tc.ID,
				req,
				runOnHost,
				targetHost,
				approvalReason,
			)
			execution.Output = formatApprovalNeededToolResult(command, tc.ID, approvalReason, approvalID)
			execution.Success = true // Not an error, just needs approval
			return execution.Output, execution
		}

		// Build execution request with proper targeting
		execReq := req

		// If target_host is explicitly specified by AI, use it for routing
		if targetHost != "" {
			// Ensure Context map exists
			if execReq.Context == nil {
				execReq.Context = make(map[string]interface{})
			} else {
				// Make a copy to avoid modifying the original
				newContext := make(map[string]interface{})
				for k, v := range req.Context {
					newContext[k] = v
				}
				execReq.Context = newContext
			}
			// Set the node explicitly - this takes priority in routing
			execReq.Context["node"] = targetHost
			log.Debug().
				Str("target_host", targetHost).
				Str("command", command).
				Msg("AI explicitly specified target_host for command routing")
		}

		// If run_on_host is true, override the target type to run on host
		if runOnHost {
			log.Debug().
				Str("command", command).
				Str("target_host", targetHost).
				Str("original_target_type", req.TargetType).
				Str("original_target_id", req.TargetID).
				Msg("run_on_host=true - overriding target type to 'agent'")
			execReq.TargetType = "agent"
			execReq.TargetID = ""
		} else {
			log.Debug().
				Str("command", command).
				Str("target_type", req.TargetType).
				Str("target_id", req.TargetID).
				Bool("run_on_host", runOnHost).
				Msg("Executing command with current target type")
		}

		// Execute via agent
		result, err := s.executeOnAgent(ctx, execReq, command)
		recordIncident := func(success bool, output string) {
			alertIdentifier := extractAlertIdentifier(req.Context)
			resourceID := extractIncidentResourceIdentifier(req.TargetID, req.Context)
			details := map[string]any{
				"resource_id":   resourceID,
				"resource_type": req.TargetType,
				"run_on_host":   runOnHost,
			}
			if targetHost != "" {
				details["target_host"] = targetHost
			}
			if alertIdentifier != "" {
				s.mu.RLock()
				incidentStore := s.incidentStore
				s.mu.RUnlock()
				if incidentStore != nil {
					incidentStore.RecordCommand(alertIdentifier, command, success, output, details)
				}
			}
			s.recordCanonicalResourceChange(unifiedresources.BuildCommandExecutionChange(
				resourceID,
				alertIdentifier,
				"agent:pulse-assistant",
				command,
				success,
				output,
				details,
			))
		}
		if err != nil {
			recordIncident(false, result)
			execution.Output = fmt.Sprintf("Error executing command: %s", err)
			return execution.Output, execution
		}

		recordIncident(true, result)
		execution.Output = result
		execution.Success = true
		return result, execution

	case agentcapabilities.LegacyAssistantFetchURLToolName:
		urlStr, _ := tc.Input[agentcapabilities.LegacyAssistantURLArgumentName].(string)
		execution.Input = urlStr

		if urlStr == "" {
			execution.Output = "Error: url is required"
			return execution.Output, execution
		}

		// Fetch the URL
		result, err := s.fetchURL(ctx, urlStr)
		if err != nil {
			execution.Output = fmt.Sprintf("Error fetching URL: %s", err)
			return execution.Output, execution
		}

		execution.Output = result
		execution.Success = true
		return result, execution

	case agentcapabilities.LegacyAssistantSetResourceURLToolName:
		resourceType, _ := tc.Input[agentcapabilities.LegacyAssistantResourceTypeArgumentName].(string)
		resourceID, _ := tc.Input[agentcapabilities.LegacyAssistantResourceIDArgumentName].(string)
		resourceURL, _ := tc.Input[agentcapabilities.LegacyAssistantURLArgumentName].(string)
		execution.Input = fmt.Sprintf("%s %s -> %s", resourceType, resourceID, resourceURL)

		if resourceType == "" {
			execution.Output = "Error: resource_type is required (use canonical v6 types: vm, system-container, oci-container, app-container, agent, node, docker-host)"
			return execution.Output, execution
		}
		if resourceID == "" {
			// Try to get the resource ID from the request context
			if req.TargetID != "" {
				resourceID = req.TargetID
			} else {
				execution.Output = "Error: resource_id is required"
				return execution.Output, execution
			}
		}

		// Allow empty URL to clear the setting
		// if resourceURL == "" {
		// 	execution.Output = "Error: url is required"
		// 	return execution.Output, execution
		// }

		// Update the metadata
		if err := s.SetResourceURL(resourceType, resourceID, resourceURL); err != nil {
			execution.Output = fmt.Sprintf("Error setting URL: %s", err)
			return execution.Output, execution
		}

		execution.Output = fmt.Sprintf("Successfully set URL for %s '%s' to: %s\nThe URL is now visible in the Pulse dashboard as a clickable link.", resourceType, resourceID, resourceURL)
		execution.Success = true
		return execution.Output, execution

	case agentcapabilities.ResolveFindingCapabilityName:
		findingID, _ := tc.Input[agentcapabilities.FindingIDArgumentName].(string)
		resolutionNote, _ := tc.Input[agentcapabilities.ResolutionNoteArgumentName].(string)
		execution.Input = fmt.Sprintf("finding: %s, note: %s", findingID, resolutionNote)

		// If no finding ID provided by AI, check the request context
		if findingID == "" {
			findingID = req.FindingID
		}

		if findingID == "" {
			execution.Output = "Error: finding_id is required. The finding ID should be provided in the request context when helping fix a patrol finding."
			return execution.Output, execution
		}

		// Get the patrol service to resolve the finding
		s.mu.RLock()
		patrolService := s.patrolService
		s.mu.RUnlock()

		if patrolService == nil {
			execution.Output = "Error: Patrol service not available"
			return execution.Output, execution
		}

		// Resolve the finding
		err := patrolService.ResolveFinding(findingID, resolutionNote)
		if err != nil {
			execution.Output = fmt.Sprintf("Error resolving finding: %s", err)
			return execution.Output, execution
		}

		execution.Output = fmt.Sprintf("Finding resolved! The Patrol finding has been marked as fixed.\nID: %s", findingID)
		if resolutionNote != "" {
			execution.Output += fmt.Sprintf("\nResolution: %s", resolutionNote)
		}
		execution.Success = true
		return execution.Output, execution

	case agentcapabilities.DismissFindingCapabilityName:
		findingID, _ := tc.Input[agentcapabilities.FindingIDArgumentName].(string)
		reason, _ := tc.Input[agentcapabilities.ReasonArgumentName].(string)
		note, _ := tc.Input[agentcapabilities.NoteArgumentName].(string)
		execution.Input = fmt.Sprintf("finding: %s, reason: %s, note: %s", findingID, reason, note)

		// If no finding ID provided by AI, check the request context
		if findingID == "" {
			findingID = req.FindingID
		}

		if findingID == "" {
			execution.Output = "Error: finding_id is required. The finding ID should be provided in the request context when helping fix a patrol finding."
			return execution.Output, execution
		}

		// Validate reason
		validReasons := map[string]bool{"not_an_issue": true, "expected_behavior": true, "will_fix_later": true}
		if !validReasons[reason] {
			execution.Output = "Error: reason must be one of: not_an_issue, expected_behavior, will_fix_later"
			return execution.Output, execution
		}

		// Get the patrol service to dismiss the finding
		s.mu.RLock()
		patrolService := s.patrolService
		s.mu.RUnlock()

		if patrolService == nil {
			execution.Output = "Error: Patrol service not available"
			return execution.Output, execution
		}

		// Dismiss the finding (this will also create a suppression rule for expected_behavior/not_an_issue)
		err := patrolService.DismissFinding(findingID, reason, note)
		if err != nil {
			execution.Output = fmt.Sprintf("Error dismissing finding: %s", err)
			return execution.Output, execution
		}

		// Format a helpful response based on reason
		var resultMsg string
		if reason == "will_fix_later" {
			remindAt := time.Now().Add(DefaultWillFixLaterRemindAfter).Format("2006-01-02")
			resultMsg = fmt.Sprintf("Finding dismissed as '%s'. Pulse Patrol will stay quiet on this until %s; if the finding is still tripping after that date, it will surface again with a 'reminded' lifecycle event.\nID: %s\nNote: %s", reason, remindAt, findingID, note)
		} else {
			resultMsg = fmt.Sprintf("Finding dismissed as '%s' and suppression rule created. Similar findings for this resource will not be raised again.\nID: %s\nNote: %s", reason, findingID, note)
		}
		execution.Output = resultMsg
		execution.Success = true
		return execution.Output, execution

	default:
		execution.Output = fmt.Sprintf("Unknown tool: %s", tc.Name)
		return execution.Output, execution
	}
}

// getGuestID returns a unique identifier for the guest based on the request
func (s *Service) getGuestID(req ExecuteRequest) string {
	// Build a consistent guest ID from the target information
	if req.TargetType == "" || req.TargetID == "" {
		return ""
	}

	// For Proxmox targets, include the node info
	// Format: instance-node-type-vmid or instance-targetid
	return fmt.Sprintf("%s-%s", req.TargetType, req.TargetID)
}

// GetGuestKnowledge returns all knowledge for a guest
func (s *Service) GetGuestKnowledge(guestID string) (*knowledge.GuestKnowledge, error) {
	if s.knowledgeStore == nil {
		return nil, fmt.Errorf("knowledge store not available")
	}
	return s.knowledgeStore.GetKnowledge(guestID)
}

// SaveGuestNote saves a note for a guest
func (s *Service) SaveGuestNote(guestID, guestName, guestType, category, title, content string) error {
	if s.knowledgeStore == nil {
		return fmt.Errorf("knowledge store not available")
	}
	return s.knowledgeStore.SaveNote(guestID, guestName, guestType, category, title, content)
}

// DeleteGuestNote deletes a note from a guest
func (s *Service) DeleteGuestNote(guestID, noteID string) error {
	if s.knowledgeStore == nil {
		return fmt.Errorf("knowledge store not available")
	}
	return s.knowledgeStore.DeleteNote(guestID, noteID)
}

// GetKnowledgeStore returns the knowledge store for external use
// This is used by components like the investigation orchestrator to get
// infrastructure context for proposing correct CLI commands
func (s *Service) GetKnowledgeStore() *knowledge.Store {
	s.mu.RLock()
	defer s.mu.RUnlock()
	return s.knowledgeStore
}

// fetchURL fetches content from a URL with size limits and timeout
func (s *Service) fetchURL(ctx context.Context, urlStr string) (string, error) {
	parsedURL, err := parseAndValidateFetchURL(ctx, urlStr)
	if err != nil {
		return "", err
	}

	// Create HTTP client with timeout and safe transport to prevent SSRF/DNS rebinding
	client := &http.Client{
		Timeout: 30 * time.Second,
		Transport: &http.Transport{
			DialContext: func(ctx context.Context, network, addr string) (net.Conn, error) {
				// Use a dialer with reasonable timeout
				d := net.Dialer{
					Timeout: 10 * time.Second,
				}
				conn, err := d.DialContext(ctx, network, addr)
				if err != nil {
					return nil, err
				}

				// Validate the actual connected IP address
				// This prevents DNS rebinding attacks
				if tcpConn, ok := conn.(*net.TCPConn); ok {
					remoteAddr := tcpConn.RemoteAddr().(*net.TCPAddr)
					if isBlockedFetchIP(remoteAddr.IP) {
						conn.Close()
						return nil, fmt.Errorf("URL resolves to blocked IP address: %s", remoteAddr.IP)
					}
				}

				return conn, nil
			},
			DisableKeepAlives: true, // One-off requests
		},
		CheckRedirect: func(req *http.Request, via []*http.Request) error {
			if len(via) >= 3 {
				return fmt.Errorf("too many redirects")
			}
			// Still validate the URL structure and initial resolution for failsafe
			if _, err := parseAndValidateFetchURL(ctx, req.URL.String()); err != nil {
				return err
			}
			return nil
		},
	}

	// Create request with context
	req, err := http.NewRequestWithContext(ctx, "GET", parsedURL.String(), nil)
	if err != nil {
		return "", fmt.Errorf("invalid URL: %w", err)
	}

	// Set a reasonable user agent
	req.Header.Set("User-Agent", "Pulse-AI/1.0")

	// Make the request
	resp, err := client.Do(req)
	if err != nil {
		return "", fmt.Errorf("request failed: %w", err)
	}
	defer resp.Body.Close()

	// Read response with size limit (64KB)
	const maxSize = 64 * 1024
	limitedReader := io.LimitReader(resp.Body, maxSize)
	body, err := io.ReadAll(limitedReader)
	if err != nil {
		return "", fmt.Errorf("failed to read response: %w", err)
	}

	// Build result with status info
	result := fmt.Sprintf("HTTP %d %s\n", resp.StatusCode, resp.Status)
	result += fmt.Sprintf("Content-Type: %s\n", resp.Header.Get("Content-Type"))
	result += fmt.Sprintf("Content-Length: %d bytes\n\n", len(body))
	result += string(body)

	if len(body) == maxSize {
		result += "\n\n[Response truncated at 64KB]"
	}

	return result, nil
}

func parseAndValidateFetchURL(ctx context.Context, urlStr string) (*url.URL, error) {
	clean := strings.TrimSpace(urlStr)
	if clean == "" {
		return nil, fmt.Errorf("url is required")
	}

	parsed, err := url.Parse(clean)
	if err != nil {
		return nil, fmt.Errorf("invalid URL: %w", err)
	}
	if !parsed.IsAbs() {
		return nil, fmt.Errorf("URL must be absolute")
	}
	if parsed.Scheme != "http" && parsed.Scheme != "https" {
		return nil, fmt.Errorf("only http/https URLs are allowed")
	}
	if parsed.User != nil {
		return nil, fmt.Errorf("URLs with embedded credentials are not allowed")
	}
	if parsed.Fragment != "" {
		return nil, fmt.Errorf("URL fragments are not allowed")
	}

	host := parsed.Hostname()
	if host == "" {
		return nil, fmt.Errorf("URL must include a host")
	}

	if isBlockedFetchHost(host) {
		return nil, fmt.Errorf("URL host is blocked")
	}

	if ip := net.ParseIP(host); ip != nil {
		if isBlockedFetchIP(ip) {
			return nil, fmt.Errorf("URL IP is blocked")
		}
		return parsed, nil
	}

	addrs, err := net.DefaultResolver.LookupIPAddr(ctx, host)
	if err != nil {
		// DNS failures are surfaced directly to the caller.
		return nil, fmt.Errorf("failed to resolve host: %w", err)
	}
	for _, addr := range addrs {
		if isBlockedFetchIP(addr.IP) {
			return nil, fmt.Errorf("URL host resolves to a blocked address")
		}
	}

	return parsed, nil
}

func isBlockedFetchHost(host string) bool {
	h := strings.TrimSpace(strings.ToLower(host))
	if h == "localhost" || h == "localhost." {
		return true
	}
	return false
}

func isBlockedFetchIP(ip net.IP) bool {
	if ip == nil {
		return true
	}
	if ip.IsLoopback() || ip.IsUnspecified() || ip.IsLinkLocalUnicast() || ip.IsLinkLocalMulticast() {
		// Allow loopback only if explicitly permitted (for local development)
		if ip.IsLoopback() && os.Getenv("PULSE_AI_ALLOW_LOOPBACK") == "true" {
			return false
		}
		return true
	}
	// SECURITY: Block private IP ranges (RFC1918) to prevent SSRF attacks
	// Private ranges: 10.0.0.0/8, 172.16.0.0/12, 192.168.0.0/16
	if ip.IsPrivate() {
		// Allow private IPs only if explicitly permitted
		if os.Getenv("PULSE_AI_ALLOW_PRIVATE_IPS") == "true" {
			return false
		}
		return true
	}
	// Block multicast and other non-unicast targets.
	if !ip.IsGlobalUnicast() {
		return true
	}
	// SECURITY: every check above reads the literal address bytes, so an IPv6
	// transition address (NAT64, 6to4, Teredo, ISATAP) tunnels to an internal
	// IPv4 destination none of them can see. Hold the embedded destination to
	// the same policy, including the loopback and private-IP escape hatches.
	for _, embedded := range securityutil.EmbeddedIPv4Candidates(ip) {
		if isBlockedFetchIP(embedded) {
			return true
		}
	}
	return false
}

// sanitizeError cleans up error messages to remove internal networking details
// that are not helpful to users or AI models (IP addresses, port numbers, etc.)
func sanitizeError(err error) error {
	if err == nil {
		return nil
	}

	errMsg := err.Error()

	// Replace raw TCP connection details with generic message
	// e.g., "write tcp 192.0.2.10:7655->198.51.100.20:58004: i/o timeout"
	// becomes "connection to agent timed out"
	if strings.Contains(errMsg, "i/o timeout") {
		if strings.Contains(errMsg, "failed to send command") {
			return fmt.Errorf("connection to agent timed out - the agent may be disconnected or unreachable")
		}
		return fmt.Errorf("network timeout - the target may be unreachable")
	}

	// Replace "write tcp ... connection refused" style errors
	if strings.Contains(errMsg, "connection refused") {
		return fmt.Errorf("connection refused - the agent may not be running on the target host")
	}

	// Replace "no such host" errors
	if strings.Contains(errMsg, "no such host") {
		return fmt.Errorf("host not found - verify the hostname is correct and DNS is working")
	}

	// Replace "context deadline exceeded" with friendlier message
	if strings.Contains(errMsg, "context deadline exceeded") {
		return fmt.Errorf("operation timed out - the command may have taken too long")
	}

	return err
}

func hasExplicitHostRoutingContext(ctx map[string]interface{}) bool {
	if len(ctx) == 0 {
		return false
	}
	for _, key := range []string{"node", "host", "guest_node", "hostname", "host_name", "target_host"} {
		if value, ok := ctx[key].(string); ok && strings.TrimSpace(value) != "" {
			return true
		}
	}
	return false
}

func normalizeExecuteTargetType(raw, targetID string, ctx map[string]interface{}) (string, error) {
	targetType := agentcapabilities.NormalizeActionTargetType(raw)
	switch targetType {
	case agentcapabilities.ActionTargetTypeAgent, agentcapabilities.ActionTargetTypeSystemContainer, agentcapabilities.ActionTargetTypeVM:
		return targetType, nil
	case "":
		if strings.TrimSpace(targetID) == "" && hasExplicitHostRoutingContext(ctx) {
			return agentcapabilities.ActionTargetTypeAgent, nil
		}
		return "", fmt.Errorf("target_type is required (agent, system-container, or vm)")
	default:
		return "", fmt.Errorf("unsupported target_type %q (allowed: %s)", raw, agentcapabilities.ActionTargetTypeAllowedDescription())
	}
}

func extractVMIDFromContext(ctx map[string]interface{}) (string, bool) {
	if ctx == nil {
		return "", false
	}
	vmID, ok := ctx["vmid"]
	if !ok {
		return "", false
	}

	switch v := vmID.(type) {
	case float64:
		if v > 0 {
			return fmt.Sprintf("%.0f", v), true
		}
	case int:
		if v > 0 {
			return fmt.Sprintf("%d", v), true
		}
	case int64:
		if v > 0 {
			return fmt.Sprintf("%d", v), true
		}
	case string:
		if parsed := extractVMIDFromTargetID(v); parsed > 0 {
			return strconv.Itoa(parsed), true
		}
	}

	return "", false
}

// executeOnAgent executes a command via the agent WebSocket
func (s *Service) executeOnAgent(ctx context.Context, req ExecuteRequest, command string) (string, error) {
	return s.executeOnAgentWithApproval(ctx, req, command, "")
}

func (s *Service) executeOnAgentWithApproval(ctx context.Context, req ExecuteRequest, command, approvalID string) (string, error) {
	return s.executeOnAgentWithAuthorization(ctx, req, command, approvalID, "", "")
}

func (s *Service) executeOnAgentWithAuthorization(ctx context.Context, req ExecuteRequest, command, approvalID, orgID, actionID string) (string, error) {
	if s.agentServer == nil {
		return "", fmt.Errorf("agent server not available")
	}

	normalizedTargetType, err := normalizeExecuteTargetType(req.TargetType, req.TargetID, req.Context)
	if err != nil {
		return "", err
	}

	normalizedReq := req
	normalizedReq.TargetType = normalizedTargetType

	// Find the appropriate agent using robust routing
	agents := s.agentServer.GetConnectedAgents()

	// Use the new robust routing logic
	routeResult, err := s.routeToAgent(normalizedReq, command, agents)
	if err != nil {
		// Check if this is a routing error that should ask for clarification
		if routingErr, ok := err.(*RoutingError); ok && routingErr.AskForClarification {
			// Return a message that encourages the AI to ask the user for clarification
			// instead of just failing with an error
			return routingErr.ForAI(), nil
		}
		// Return actionable error message for other errors
		return "", err
	}

	// Log any warnings from routing
	for _, warning := range routeResult.Warnings {
		log.Warn().Str("warning", warning).Msg("routing warning")
	}

	agentID := routeResult.AgentID

	log.Debug().
		Str("agent_id", agentID).
		Str("agent_hostname", routeResult.AgentHostname).
		Str("target_node", routeResult.TargetNode).
		Str("routing_method", routeResult.RoutingMethod).
		Bool("cluster_peer", routeResult.ClusterPeer).
		Msg("Command routed to agent")

	targetID := strings.TrimSpace(req.TargetID)
	if normalizedTargetType == "system-container" || normalizedTargetType == "vm" {
		if vmID, ok := extractVMIDFromContext(req.Context); ok {
			targetID = vmID
		} else if extracted := extractVMIDFromTargetID(req.TargetID); extracted > 0 {
			targetID = strconv.Itoa(extracted)
		} else {
			return "", fmt.Errorf("%s target requires numeric VMID in context.vmid or target_id", normalizedTargetType)
		}
	}

	dispatchTargetType := normalizedTargetType
	if normalizedTargetType == "system-container" {
		// Agent transport uses "container" for LXC command routing.
		dispatchTargetType = "container"
	}

	requestID := uuid.New().String()

	// Automatically force non-interactive mode for package managers
	// This prevents hanging when apt/dpkg asks for confirmation or configuration
	if strings.Contains(command, "apt") || strings.Contains(command, "dpkg") {
		if !strings.Contains(command, "DEBIAN_FRONTEND=") {
			command = "export DEBIAN_FRONTEND=noninteractive; " + command
		}
	}

	cmd := agentexec.ExecuteCommandPayload{
		RequestID:  requestID,
		Command:    command,
		ApprovalID: strings.TrimSpace(approvalID),
		TargetType: dispatchTargetType,
		TargetID:   targetID,
		Timeout:    300, // 5 minutes - commands like du, backups, etc. can take a while
	}
	if strings.TrimSpace(approvalID) != "" {
		cmd.BindCommandAuthorization(orgID, actionID)
	}

	result, err := s.agentServer.ExecuteCommand(ctx, agentID, cmd)
	if err != nil {
		return "", sanitizeError(err)
	}

	if !result.Success {
		if result.Error != "" {
			return "", fmt.Errorf("%s", result.Error)
		}
		if result.Stderr != "" {
			return result.Stderr, nil // Return stderr as output, not error
		}
	}

	output := result.Stdout
	if result.Stderr != "" && result.Stdout != "" {
		output = fmt.Sprintf("%s\n\nSTDERR:\n%s", result.Stdout, result.Stderr)
	} else if result.Stderr != "" {
		output = result.Stderr
	}

	return output, nil
}

// RunCommandRequest represents a request to run a single command
type RunCommandRequest struct {
	Command    string `json:"command"`
	ApprovalID string `json:"approval_id,omitempty"` // Consumed by API handler before execution
	TargetType string `json:"target_type"`           // "agent", "system-container", "vm"
	TargetID   string `json:"target_id"`
	RunOnHost  bool   `json:"run_on_host"` // If true, run on host instead of target
	VMID       string `json:"vmid,omitempty"`
	TargetHost string `json:"target_host,omitempty"` // Explicit host for routing
	OrgID      string `json:"-"`
	ActionID   string `json:"-"`
}

// RunCommandResponse represents the result of running a command
type RunCommandResponse struct {
	Output  string `json:"output"`
	Success bool   `json:"success"`
	Error   string `json:"error,omitempty"`
}

// RunCommand executes a single command via the agent (used for approved commands)
func (s *Service) RunCommand(ctx context.Context, req RunCommandRequest) (*RunCommandResponse, error) {
	if s.agentServer == nil {
		return &RunCommandResponse{Success: false, Error: "Agent server not available"}, nil
	}

	// Build an ExecuteRequest from the RunCommandRequest
	execReq := ExecuteRequest{
		TargetType: req.TargetType,
		TargetID:   req.TargetID,
		Context:    make(map[string]interface{}),
	}

	// If running on host, override target type
	if req.RunOnHost {
		execReq.TargetType = agentcapabilities.ActionTargetTypeAgent
		// Keep the original target info for routing
	}

	// Add VMID to context if provided
	if req.VMID != "" {
		execReq.Context["vmid"] = req.VMID
	}

	// If target_host is specified, set it in context for routing
	if req.TargetHost != "" {
		execReq.Context["node"] = req.TargetHost
		log.Debug().
			Str("target_host", req.TargetHost).
			Str("command", req.Command).
			Msg("RunCommand using explicit target_host for routing")
	}

	output, err := s.executeOnAgentWithAuthorization(ctx, execReq, req.Command, req.ApprovalID, req.OrgID, req.ActionID)
	if err != nil {
		return &RunCommandResponse{
			Success: false,
			Error:   err.Error(),
			Output:  output,
		}, nil
	}

	return &RunCommandResponse{
		Success: true,
		Output:  output,
	}, nil
}

// buildSystemPrompt creates the system prompt based on the request context
func (s *Service) buildSystemPrompt(req ExecuteRequest, destinationModel string) string {
	commandApproval := `Pulse has a built-in approval system:
- Safe commands (read-only: ls, df, cat, ps) execute immediately
- Destructive commands (rm, service restart, apt install) require user approval
- Blocked commands are never executed
Execute commands normally - Pulse handles the approval flow automatically.`
	if req.RequireCommandApproval {
		commandApproval = `This handoff is approval-bound:
- Do not run commands just because they look safe.
- Any run_command tool call will pause for explicit operator approval before execution.
- Prefer existing metrics, alert context, and known system state before requesting command approval.
- Blocked commands are never executed.`
	}

	prompt := fmt.Sprintf(`You are Pulse's diagnostic assistant for Proxmox, Docker, and Kubernetes homelabs.

## Command Approval
%s

## Command Execution
- run_on_host=true: Run on PVE/Docker host (pct, qm, vzdump, docker commands)
- run_on_host=false: Run inside the container/VM
- target_host: Required when using run_on_host=true - use the node name from context

Example for LXC 106 on node 'minipc':
- Inside container: run_command(command="df -h", run_on_host=false)
- On host: run_command(command="pct exec 106 -- df -h", run_on_host=true, target_host="minipc")

If you get "Configuration file does not exist" - wrong host, check target_host.

## LXC Management
Pulse manages LXC containers agentlessly - there is no Pulse process inside containers.
- run_on_host=false: Pulse routes commands into the LXC via the PVE host
- run_on_host=true + target_host: Runs pct/qm commands on the PVE host directly

## URL Discovery
When finding web URLs for resources:
1. Check listening ports: ss -tlnp | grep LISTEN
2. Get IP: hostname -I | awk '{print $1}'
3. Test with fetch_url to verify
4. Save with set_resource_url tool

resource_id format: {instance}:{node}:{vmid} (colons, not dashes)
Example: "homelab:pve-node:201" for VMID 201 on node pve-node in instance homelab

## Installing/Updating Pulse
curl -sSL https://raw.githubusercontent.com/rcourtman/Pulse/main/install.sh | bash
After: systemctl enable pulse && systemctl start pulse
Latest version: https://api.github.com/repos/rcourtman/Pulse/releases/latest`, commandApproval)

	// Add custom context from AI settings (user's infrastructure description)
	s.mu.RLock()
	cfg := s.cfg
	s.mu.RUnlock()
	if cfg != nil && cfg.CustomContext != "" {
		prompt += "\n\n## User's Infrastructure Description\n"
		prompt += "The user has provided this context about their infrastructure:\n\n"
		prompt += cfg.CustomContext
	}

	// Add connected infrastructure info via unified resource model
	s.mu.RLock()
	hasUnifiedResourceProvider := s.unifiedResourceProvider != nil
	s.mu.RUnlock()

	if hasUnifiedResourceProvider {
		if strings.TrimSpace(destinationModel) != "" {
			prompt += s.buildUnifiedResourceContextForModel(destinationModel)
		} else {
			prompt += s.buildUnifiedResourceContext()
		}
	} else {
		log.Warn().Msg("AI context: unified resource provider not available, infrastructure context will be limited")
	}

	// Add user annotations from all resources (global context)
	prompt += s.buildUserAnnotationsContext()

	// Add current alert status - this gives AI awareness of active issues
	prompt += s.buildAlertContext()

	// Add incident memory for alert or resource context
	alertIdentifier := ""
	resourceID := req.TargetID
	if req.Context != nil {
		alertIdentifier = extractAlertIdentifier(req.Context)

		if resourceID == "" {
			if val, ok := req.Context["resourceId"].(string); ok && val != "" {
				resourceID = val
			} else if val, ok := req.Context["resource_id"].(string); ok && val != "" {
				resourceID = val
			} else if val, ok := req.Context["guest_id"].(string); ok && val != "" {
				resourceID = val
			}
		}
	}

	if incidentContext := s.buildIncidentContext(resourceID, alertIdentifier); incidentContext != "" {
		prompt += incidentContext
	}

	// Add all saved knowledge when no specific target is selected
	// This gives the AI context about everything learned from previous sessions
	if req.TargetType == "" && s.knowledgeStore != nil {
		prompt += s.knowledgeStore.FormatAllForContext()
	}

	// Add target context if provided
	if req.TargetType != "" {
		guestName := ""
		if name, ok := req.Context["guestName"].(string); ok {
			guestName = name
		} else if name, ok := req.Context["name"].(string); ok {
			guestName = name
		}

		if guestName != "" {
			// Include the node in the focus header so AI can't miss it for routing
			nodeName := ""
			if node, ok := req.Context["node"].(string); ok && node != "" {
				nodeName = node
			} else if node, ok := req.Context["guest_node"].(string); ok && node != "" {
				nodeName = node
			}
			if nodeName != "" {
				prompt += fmt.Sprintf("\n\n## Current Focus\nYou are analyzing **%s** (%s on node **%s**)\n**ROUTING: When using run_on_host=true, set target_host=\"%s\"**",
					guestName, req.TargetType, nodeName, nodeName)
			} else {
				prompt += fmt.Sprintf("\n\n## Current Focus\nYou are analyzing **%s** (%s)", guestName, req.TargetType)
			}
		} else if req.TargetID != "" {
			prompt += fmt.Sprintf("\n\n## Current Focus\nYou are analyzing %s '%s'", req.TargetType, req.TargetID)
		}

		// Add prior action history for this resource as neutral context.
		prompt += s.buildRemediationContext(req.TargetID)

	}

	// If we're helping with a patrol finding, provide the finding ID and the
	// available lifecycle tools without forcing a tool choice.
	if req.FindingID != "" {
		prompt += fmt.Sprintf("\n\n## Patrol Finding Context\n"+
			"You are helping with patrol finding **%s**.\n\n"+
			"Lifecycle tools are available when current evidence supports them:\n"+
			"- `%s` records that the underlying issue has actually been fixed.\n"+
			"- `%s` records a false positive or expected behavior.\n"+
			"Only call a lifecycle tool when that is the model's conclusion from the current evidence.\n",
			req.FindingID,
			agentcapabilities.ResolveFindingCapabilityName,
			agentcapabilities.DismissFindingCapabilityName)
	}

	// Add any provided context in a structured way
	if len(req.Context) > 0 {
		prompt += "\n\n## Current Metrics and State"

		// Group metrics by category for better readability
		categories := map[string][]string{
			"Identity":     {"name", "guestName", "type", "vmid", "node", "guest_node", "status", "uptime"},
			"CPU":          {"cpu_usage", "cpu_cores"},
			"Memory":       {"memory_used", "memory_total", "memory_usage", "memory_balloon", "swap_used", "swap_total"},
			"Disk":         {"disk_used", "disk_total", "disk_usage"},
			"I/O Rates":    {"disk_read_rate", "disk_write_rate", "network_in_rate", "network_out_rate"},
			"Backup":       {"backup_status", "last_backup", "days_since_backup"},
			"System Info":  {"os_name", "os_version", "guest_agent", "ip_addresses", "tags"},
			"User Context": {"user_notes", "user_annotations"},
		}

		categoryOrder := []string{"Identity", "User Context", "Backup", "CPU", "Memory", "Disk", "I/O Rates", "System Info"}

		for _, category := range categoryOrder {
			keys := categories[category]
			hasValues := false
			categoryContent := ""

			for _, k := range keys {
				if v, ok := req.Context[k]; ok && v != nil && v != "" {
					if !hasValues {
						categoryContent = fmt.Sprintf("\n### %s", category)
						hasValues = true
					}
					categoryContent += fmt.Sprintf("\n- %s: %v", formatContextKey(k), v)
				}
			}

			if hasValues {
				prompt += categoryContent
			}
		}

		// Add any remaining context that wasn't categorized
		for k, v := range req.Context {
			found := false
			for _, keys := range categories {
				for _, key := range keys {
					if k == key {
						found = true
						break
					}
				}
				if found {
					break
				}
			}
			if !found && v != nil && v != "" {
				prompt += fmt.Sprintf("\n- %s: %v", formatContextKey(k), v)
			}
		}

		// Add enriched historical context (baselines, trends, predictions)
		// This is the Pulse Pro value-add that Claude Code can't replicate
		prompt += s.buildEnrichedResourceContext(req.TargetID, req.TargetType, req.Context)
	}

	return prompt
}

// formatContextKey converts snake_case keys to readable labels
func formatContextKey(key string) string {
	replacements := map[string]string{
		"guestName":         "Guest Name",
		"name":              "Name",
		"type":              "Type",
		"vmid":              "VMID",
		"node":              "Host Node",
		"guest_node":        "Host Node",
		"status":            "Status",
		"uptime":            "Uptime",
		"cpu_usage":         "CPU Usage",
		"cpu_cores":         "CPU Cores",
		"memory_used":       "Memory Used",
		"memory_total":      "Memory Total",
		"memory_usage":      "Memory Usage",
		"memory_balloon":    "Memory Balloon",
		"swap_used":         "Swap Used",
		"swap_total":        "Swap Total",
		"disk_used":         "Disk Used",
		"disk_total":        "Disk Total",
		"disk_usage":        "Disk Usage",
		"disk_read_rate":    "Disk Read Rate",
		"disk_write_rate":   "Disk Write Rate",
		"network_in_rate":   "Network In Rate",
		"network_out_rate":  "Network Out Rate",
		"backup_status":     "Backup Status",
		"last_backup":       "Last Backup",
		"days_since_backup": "Days Since Backup",
		"os_name":           "OS Name",
		"os_version":        "OS Version",
		"guest_agent":       "Guest Agent",
		"ip_addresses":      "IP Addresses",
		"tags":              "Tags",
		"user_notes":        "User Notes",
		"user_annotations":  "User Annotations",
	}

	if label, ok := replacements[key]; ok {
		return label
	}
	return key
}

// buildUserAnnotationsContext gathers all user annotations from guests and docker containers
// These provide infrastructure context that the AI should know about for any query
func (s *Service) buildUserAnnotationsContext() string {
	// Return empty if persistence is not available
	if s.persistence == nil {
		return ""
	}

	var annotations []string

	// Load guest metadata
	guestStore, err := s.persistence.LoadGuestMetadata()
	if err != nil {
		log.Warn().Err(err).Msg("failed to load guest metadata for AI context")
	} else {
		guestMeta := guestStore.GetAll()
		log.Debug().Int("count", len(guestMeta)).Msg("loaded guest metadata for AI context")
		for id, meta := range guestMeta {
			if meta != nil && len(meta.Notes) > 0 {
				// Use LastKnownName if available, otherwise use ID
				name := meta.LastKnownName
				if name == "" {
					name = id
				}
				for _, note := range meta.Notes {
					annotations = append(annotations, fmt.Sprintf("- Guest '%s': %s", name, note))
				}
			}
		}
	}

	// Load docker metadata - include host info for context
	dockerStore, err := s.persistence.LoadDockerMetadata()
	if err != nil {
		log.Warn().Err(err).Msg("failed to load docker metadata for AI context")
	} else {
		dockerMeta := dockerStore.GetAll()
		log.Debug().Int("count", len(dockerMeta)).Msg("loaded docker metadata for AI context")
		for id, meta := range dockerMeta {
			if meta != nil && len(meta.Notes) > 0 {
				// Extract host and container info from ID (format: hostid:container:containerid)
				name := id
				hostInfo := ""
				parts := strings.Split(id, ":")
				if len(parts) >= 3 {
					hostInfo = parts[0] // First part is the host identifier
					containerID := parts[2]
					if len(containerID) > 12 {
						containerID = containerID[:12]
					}
					name = fmt.Sprintf("Docker container %s", containerID)
				}
				log.Debug().Str("name", name).Str("host", hostInfo).Int("notes", len(meta.Notes)).Msg("found docker container with annotations")
				for _, note := range meta.Notes {
					if hostInfo != "" {
						annotations = append(annotations, fmt.Sprintf("- %s (on host '%s'): %s", name, hostInfo, note))
					} else {
						annotations = append(annotations, fmt.Sprintf("- %s: %s", name, note))
					}
				}
			}
		}
	}

	log.Debug().Int("total_annotations", len(annotations)).Msg("built user annotations context")

	if len(annotations) == 0 {
		return ""
	}

	return "\n\n## User Infrastructure Notes\nThe user has added these annotations to describe their infrastructure. USE THESE to understand relationships between systems:\n" + strings.Join(annotations, "\n")
}

// TestConnection tests the AI provider connection
// Tests the provider for the currently configured default model
func (s *Service) TestConnection(ctx context.Context) error {
	s.mu.RLock()
	cfg := s.cfg
	defaultProvider := s.provider
	persistence := s.persistence
	s.mu.RUnlock()

	// Load config if not available
	if cfg == nil {
		if persistence == nil {
			return fmt.Errorf("Pulse Assistant config persistence unavailable")
		}
		var err error
		cfg, err = persistence.LoadAIConfig()
		if err != nil {
			return fmt.Errorf("failed to load Pulse Assistant config: %w", err)
		}
	}

	if cfg == nil || !cfg.IsConfigured() {
		return fmt.Errorf("no provider configured")
	}

	// Try to create a provider for the current default model
	selectedModel, err := ResolveConfiguredModel(ctx, cfg)
	if err != nil {
		if defaultProvider != nil {
			return defaultProvider.TestConnection(ctx)
		}
		return err
	}
	provider, err := providers.NewForModel(cfg, selectedModel)
	if err != nil {
		// Fall back to default provider or NewFromConfig
		log.Debug().Err(err).Str("model", selectedModel).Msg("could not create provider for model, using fallback")
		if defaultProvider != nil {
			provider = defaultProvider
		} else {
			return fmt.Errorf("failed to create fallback AI provider: %w", err)
		}
	}

	return provider.TestConnection(ctx)
}

// ListModels fetches available models from ALL configured AI providers
// Returns a unified list with models prefixed by provider name
func (s *Service) ListModels(ctx context.Context) ([]providers.ModelInfo, error) {
	models, _, err := s.ListModelsWithCache(ctx)
	return models, err
}

func (s *Service) ListModelsWithCache(ctx context.Context) ([]providers.ModelInfo, bool, error) {
	s.mu.RLock()
	persistence := s.persistence
	s.mu.RUnlock()
	if persistence == nil {
		return nil, false, fmt.Errorf("Pulse Assistant config persistence unavailable")
	}

	cfg, err := persistence.LoadAIConfig()
	if err != nil {
		return nil, false, fmt.Errorf("failed to load Pulse Assistant config: %w", err)
	}
	if cfg == nil {
		return nil, false, fmt.Errorf("Pulse Assistant not configured")
	}

	cacheKey := buildModelsCacheKey(cfg)

	// If config changed, clear all cached provider entries
	s.modelsCache.mu.Lock()
	if s.modelsCache.key != cacheKey {
		s.modelsCache.key = cacheKey
		s.modelsCache.providers = make(map[string]providerModelsEntry)
	}
	s.modelsCache.mu.Unlock()

	providersList := cfg.GetConfiguredProviders()

	allCached := true

	for _, providerName := range providersList {
		if !cfg.HasProvider(providerName) {
			continue
		}

		// Check if this provider's cache is still valid
		s.modelsCache.mu.RLock()
		entry, hasEntry := s.modelsCache.providers[providerName]
		valid := hasEntry && s.modelsCache.ttl > 0 && time.Since(entry.at) < s.modelsCache.ttl && len(entry.models) > 0
		s.modelsCache.mu.RUnlock()

		if valid {
			continue
		}

		allCached = false

		// Create provider
		provider, err := providers.NewForProvider(cfg, providerName, "")
		if err != nil {
			log.Debug().Err(err).Str("provider", providerName).Msg("skipping provider - not configured")
			continue
		}

		// Fetch models from this provider
		models, err := provider.ListModels(ctx)
		if err != nil {
			log.Warn().Err(err).Str("provider", providerName).Msg("failed to fetch models from provider")
			// Keep stale entry (don't overwrite or delete) — the provider's
			// previous models remain visible until a successful fetch replaces them.
			if !hasEntry || len(entry.models) == 0 {
				if fallbackModels := providerCatalogFallbackModels(providerName); len(fallbackModels) > 0 {
					s.modelsCache.mu.Lock()
					s.modelsCache.providers[providerName] = providerModelsEntry{
						at:     time.Now(),
						models: prefixProviderModels(providerName, fallbackModels),
					}
					s.modelsCache.mu.Unlock()
				}
			}
			continue
		}
		models = mergeProviderCatalogFallbackModels(providerName, models)

		// Build prefixed model list for this provider
		prefixed := prefixProviderModels(providerName, models)

		s.modelsCache.mu.Lock()
		s.modelsCache.providers[providerName] = providerModelsEntry{
			at:     time.Now(),
			models: prefixed,
		}
		s.modelsCache.mu.Unlock()
	}

	// Aggregate results in stable provider order
	var allModels []providers.ModelInfo
	s.modelsCache.mu.RLock()
	for _, providerName := range providersList {
		if entry, ok := s.modelsCache.providers[providerName]; ok {
			allModels = append(allModels, entry.models...)
		}
	}
	s.modelsCache.mu.RUnlock()

	return allModels, allCached, nil
}

func prefixProviderModels(providerName string, models []providers.ModelInfo) []providers.ModelInfo {
	prefixed := make([]providers.ModelInfo, 0, len(models))
	for _, m := range models {
		if strings.TrimSpace(m.ID) == "" {
			continue
		}
		description := strings.TrimSpace(m.Description)
		if description == "" {
			description = providerDisplayName(providerName) + ": " + m.ID
		}
		prefixed = append(prefixed, providers.ModelInfo{
			ID:          config.FormatModelString(providerName, m.ID),
			Name:        m.Name,
			Description: description,
			CreatedAt:   m.CreatedAt,
			Notable:     m.Notable,
			Provider:    providerName,
		})
	}
	return prefixed
}

func mergeProviderCatalogFallbackModels(providerName string, models []providers.ModelInfo) []providers.ModelInfo {
	fallbacks := providerCatalogFallbackModels(providerName)
	if len(fallbacks) == 0 {
		return models
	}
	seen := make(map[string]struct{}, len(models)+len(fallbacks))
	merged := make([]providers.ModelInfo, 0, len(models)+len(fallbacks))
	for _, model := range models {
		id := strings.ToLower(strings.TrimSpace(model.ID))
		if id == "" {
			continue
		}
		seen[id] = struct{}{}
		merged = append(merged, model)
	}
	for _, fallback := range fallbacks {
		id := strings.ToLower(strings.TrimSpace(fallback.ID))
		if id == "" {
			continue
		}
		if _, ok := seen[id]; ok {
			continue
		}
		merged = append(merged, fallback)
	}
	return merged
}

func providerCatalogFallbackModels(providerName string) []providers.ModelInfo {
	fallbacks := config.AIProviderFallbackModels(providerName)
	if len(fallbacks) == 0 {
		return nil
	}
	models := make([]providers.ModelInfo, 0, len(fallbacks))
	for _, fallback := range fallbacks {
		models = append(models, providers.ModelInfo{
			ID:          fallback.ID,
			Name:        fallback.Name,
			Description: fallback.Description,
			Notable:     fallback.Notable,
		})
	}
	return models
}

func buildModelsCacheKey(cfg *config.AIConfig) string {
	if cfg == nil {
		return "nil"
	}

	var b strings.Builder
	b.WriteString("providers=")
	b.WriteString(strings.Join(cfg.GetConfiguredProviders(), ","))
	b.WriteString("|auth=")
	b.WriteString(string(cfg.AuthMethod))
	b.WriteString("|keys=")
	b.WriteString(aiCredentialCacheSignature(cfg))

	b.WriteString("|openai_base=")
	b.WriteString(cfg.OpenAIBaseURL)
	b.WriteString("|ollama_base=")
	b.WriteString(cfg.OllamaBaseURL)
	b.WriteString("|ollama_auth=")
	if cfg.OllamaUsername != "" || cfg.OllamaPassword != "" {
		b.WriteString(credentialCacheDigest(cfg.OllamaUsername, cfg.OllamaPassword))
	}

	return b.String()
}

func aiCredentialCacheSignature(cfg *config.AIConfig) string {
	if cfg == nil {
		return ""
	}
	var b strings.Builder
	for _, providerName := range cfg.GetConfiguredProviders() {
		b.WriteString(providerName)
		b.WriteString("=")
		if key := cfg.GetAPIKeyForProvider(providerName); key != "" {
			b.WriteString(credentialCacheDigest(providerName, key))
		}
		b.WriteString(";")
	}
	return b.String()
}

// providerDisplayName returns a user-friendly name for a provider
func providerDisplayName(provider string) string {
	return config.AIProviderDisplayName(provider)
}

// Reload reloads the AI configuration (call after settings change)
func (s *Service) Reload() error {
	if err := s.LoadConfig(); err != nil {
		return fmt.Errorf("Reload: %w", err)
	}

	// Also reload the chat service so patrol picks up model/provider changes
	s.mu.RLock()
	cs := s.chatService
	cfg := s.cfg
	s.mu.RUnlock()

	if cs != nil && cfg != nil {
		if err := cs.ReloadConfig(context.Background(), cfg); err != nil {
			log.Warn().Err(err).Msg("failed to reload chat service config")
		}
	}

	return nil
}

// buildRemediationContext adds resource-scoped remediation history as neutral context.
// Pulse must not keyword-match the current problem to old fixes or recommend a
// prior command; the selected model owns diagnosis and remediation reasoning.
func (s *Service) buildRemediationContext(resourceID string) string {
	s.mu.RLock()
	patrol := s.patrolService
	s.mu.RUnlock()

	if patrol == nil {
		return ""
	}

	remLog := patrol.GetRemediationLog()
	if remLog == nil {
		return ""
	}

	if resourceID != "" {
		history := remLog.GetForResource(resourceID, 5)
		if len(history) > 0 {
			var context string
			context += "\n\n## Prior Actions Recorded For This Resource\n"
			context += "These are historical actions only. Decide from current evidence whether they are relevant.\n"
			for _, rec := range history {
				ago := time.Since(rec.Timestamp)
				agoStr := formatDuration(ago)
				context += fmt.Sprintf("- %s ago: %s → `%s` (%s)\n",
					agoStr,
					truncateString(rec.Problem, 50),
					truncateString(rec.Action, 60),
					rec.Outcome)
			}
			return context
		}
	}

	return ""
}

// buildIncidentContext adds alert-scoped incident memory first, then the
// canonical resource timeline and relationship context when a resource is known.
// The incident store is a derived investigation view, not the source of truth
// for durable backend history.
func (s *Service) buildIncidentContext(resourceID, alertIdentifier string) string {
	s.mu.RLock()
	store := s.incidentStore
	s.mu.RUnlock()

	var sections []string

	if store != nil {
		if alertIdentifier != "" {
			if text := store.FormatForAlert(alertIdentifier, 8); text != "" {
				sections = append(sections, text)
			}
		} else if resourceID != "" {
			if text := store.FormatForResource(resourceID, 4); text != "" {
				sections = append(sections, text)
			}
		}
	}

	if resourceID != "" {
		if recentChanges := s.buildRecentResourceChangesContext(resourceID); recentChanges != "" {
			sections = append(sections, recentChanges)
		}
		if relationshipContext := s.buildResourceRelationshipContext(resourceID); relationshipContext != "" {
			sections = append(sections, relationshipContext)
		}
	}

	if len(sections) == 0 {
		return ""
	}
	return strings.Join(sections, "")
}

// RecordIncidentAnalysis stores an AI analysis event for an alert.
func (s *Service) RecordIncidentAnalysis(alertIdentifier, summary string, details map[string]interface{}) {
	if alertIdentifier == "" {
		return
	}
	s.mu.RLock()
	store := s.incidentStore
	s.mu.RUnlock()
	if store == nil {
		return
	}
	store.RecordAnalysis(alertIdentifier, summary, details)
}

// RecordIncidentRunbook stores a runbook execution event for an alert.
func (s *Service) RecordIncidentRunbook(resourceID, alertIdentifier, runbookID, title string, outcome memory.Outcome, automatic bool, message string) {
	if alertIdentifier == "" || runbookID == "" {
		return
	}
	s.mu.RLock()
	store := s.incidentStore
	s.mu.RUnlock()
	if store != nil {
		store.RecordRunbook(alertIdentifier, runbookID, title, string(outcome), automatic, message)
	}
	if resourceID == "" {
		resourceID = s.lookupIncidentResourceIdentifier(alertIdentifier)
	}
	actor := ""
	if automatic {
		actor = "agent:pulse-patrol"
	}
	s.recordCanonicalResourceChange(unifiedresources.BuildRunbookExecutionChange(
		resourceID,
		alertIdentifier,
		actor,
		runbookID,
		title,
		string(outcome),
		automatic,
		message,
		nil,
	))
}

func extractAlertIdentifier(ctx map[string]interface{}) string {
	if ctx == nil {
		return ""
	}
	if alertIdentifier, ok := ctx["alertIdentifier"].(string); ok && alertIdentifier != "" {
		return alertIdentifier
	}
	return ""
}

func extractIncidentResourceIdentifier(targetID string, ctx map[string]interface{}) string {
	targetID = strings.TrimSpace(targetID)
	if targetID != "" {
		return targetID
	}
	if ctx == nil {
		return ""
	}
	for _, key := range []string{"resourceID", "resource_id", "resourceId"} {
		if value, ok := ctx[key].(string); ok && strings.TrimSpace(value) != "" {
			return strings.TrimSpace(value)
		}
	}
	return ""
}

func (s *Service) lookupIncidentResourceIdentifier(alertIdentifier string) string {
	if strings.TrimSpace(alertIdentifier) == "" {
		return ""
	}
	s.mu.RLock()
	store := s.incidentStore
	s.mu.RUnlock()
	if store == nil {
		return ""
	}
	page, err := store.QueryIncidents(memory.IncidentQuery{AlertIdentifier: alertIdentifier, Limit: 1})
	if err != nil {
		log.Warn().Err(err).Msg("Canonical incident resource lookup unavailable")
		return ""
	}
	if len(page.Incidents) == 0 {
		return ""
	}
	return strings.TrimSpace(page.Incidents[0].ResourceID)
}

func (s *Service) recordCanonicalResourceChange(change *unifiedresources.ResourceChange) {
	if s == nil || change == nil {
		return
	}

	s.mu.RLock()
	store := s.resourceExportStore
	orgID := strings.TrimSpace(s.orgID)
	storeOrgID := strings.TrimSpace(s.resourceExportStoreOrgID)
	s.mu.RUnlock()

	if store == nil {
		return
	}
	if storeOrgID != "" && orgID != "" && storeOrgID != orgID {
		return
	}
	if err := store.RecordChange(*change); err != nil {
		log.Warn().
			Err(err).
			Str("resource_id", change.ResourceID).
			Str("kind", string(change.Kind)).
			Msg("failed to record canonical incident timeline change")
	}
}

func (s *Service) buildRecentResourceChangesContext(resourceID string) string {
	resourceID = strings.TrimSpace(resourceID)
	if resourceID == "" {
		return ""
	}

	s.mu.RLock()
	store := s.resourceExportStore
	orgID := strings.TrimSpace(s.orgID)
	storeOrgID := strings.TrimSpace(s.resourceExportStoreOrgID)
	patrol := s.patrolService
	s.mu.RUnlock()

	if store != nil && (storeOrgID == "" || storeOrgID == orgID) {
		changes, err := store.GetRecentChanges(resourceID, time.Now().Add(-24*time.Hour), 5)
		if err != nil {
			log.Warn().
				Err(err).
				Str("resource_id", resourceID).
				Msg("failed to load canonical resource timeline context")
		} else if len(changes) > 0 {
			if len(changes) > 3 {
				changes = changes[:3]
			}
			if section := unifiedresources.FormatResourceRecentChangesContext(changes, false, "###"); section != "" {
				return section
			}
		}
	}

	if patrol == nil {
		return ""
	}

	changeDetector := patrol.GetChangeDetector()
	if changeDetector == nil {
		return ""
	}

	changes := changeDetector.GetChangesForResource(resourceID, 5)
	if len(changes) == 0 {
		return ""
	}
	if len(changes) > 3 {
		changes = changes[:3]
	}
	return memory.FormatRecentChangesContext(changes, false, "###")
}

type canonicalResourceGetter interface {
	Get(id string) (*unifiedresources.Resource, bool)
}

func (s *Service) buildResourceRelationshipContext(resourceID string) string {
	resourceID = strings.TrimSpace(resourceID)
	if resourceID == "" {
		return ""
	}

	s.mu.RLock()
	rs := s.readState
	s.mu.RUnlock()

	if rs == nil {
		return ""
	}

	getter, ok := rs.(canonicalResourceGetter)
	if !ok {
		return ""
	}

	resource, ok := getter.Get(resourceID)
	if !ok || resource == nil {
		return ""
	}
	return unifiedresources.FormatResourceRelationshipContext(resource, 3)
}

// truncateString truncates a string to maxLen characters.
func truncateString(s string, maxLen int) string {
	if len(s) <= maxLen {
		return s
	}
	return s[:maxLen-3] + "..."
}

// buildEnrichedResourceContext adds historical intelligence to regular AI chat
// This is what makes Pulse Pro valuable - context that Claude Code can't have:
// - Baseline comparisons ("this is 2x normal")
// - Trend analysis ("memory growing 5%/day")
// - Predictions ("disk full in 12 days")
// - Recent changes ("config changed 2 hours ago")
func (s *Service) buildEnrichedResourceContext(resourceID, _ string, currentMetrics map[string]interface{}) string {
	s.mu.RLock()
	patrol := s.patrolService
	s.mu.RUnlock()

	if patrol == nil {
		return ""
	}

	var sections []string

	// Get baseline store for comparisons
	baselineStore := patrol.GetBaselineStore()
	if baselineStore != nil {
		var baselineInfo []string

		// Helper to get raw numeric value from context
		// Prefers _raw fields, falls back to regular fields
		getRawValue := func(rawKey, formattedKey string) (float64, bool) {
			// Try raw value first
			if val, ok := currentMetrics[rawKey].(float64); ok && val > 0 {
				return val, true
			}
			// Fallback to formatted value if it's already a float
			if val, ok := currentMetrics[formattedKey].(float64); ok && val > 0 {
				return val, true
			}
			return 0, false
		}

		// Helper to format baseline comparison with status
		// ALWAYS shows baseline info when available - this is Pulse Pro's value-add
		formatBaselineComparison := func(metric string, currentVal float64, bl *baseline.MetricBaseline) string {
			if bl.SampleCount < 10 {
				return fmt.Sprintf("- %s: %.1f%% (baseline learning: %d/10 samples)", metric, currentVal, bl.SampleCount)
			}

			ratio := currentVal / bl.Mean
			status := "✓ normal"
			if ratio > 2.0 {
				status = "🔴 **ANOMALY**"
			} else if ratio > 1.5 {
				status = "🟡 elevated"
			} else if ratio < 0.3 {
				status = "🔵 unusually low"
			} else if ratio < 0.5 {
				status = "low"
			}

			return fmt.Sprintf("- %s: %.1f%% vs baseline %.1f%% (σ=%.1f) → %s", metric, currentVal, bl.Mean, bl.StdDev, status)
		}

		// Check CPU baseline - always show if we have data
		if cpuVal, ok := getRawValue("cpu_usage_raw", "cpu_usage"); ok {
			if bl, exists := baselineStore.GetBaseline(resourceID, "cpu"); exists {
				baselineInfo = append(baselineInfo, formatBaselineComparison("CPU", cpuVal, bl))
			}
		}

		// Check memory baseline - always show if we have data
		if memVal, ok := getRawValue("memory_usage_raw", "memory_usage"); ok {
			if bl, exists := baselineStore.GetBaseline(resourceID, "memory"); exists {
				baselineInfo = append(baselineInfo, formatBaselineComparison("Memory", memVal, bl))
			}
		}

		// Check disk baseline - always show if we have data
		if diskVal, ok := getRawValue("disk_usage_raw", "disk_usage"); ok {
			if bl, exists := baselineStore.GetBaseline(resourceID, "disk"); exists {
				baselineInfo = append(baselineInfo, formatBaselineComparison("Disk", diskVal, bl))
			}
		}

		if len(baselineInfo) > 0 {
			sections = append(sections, "### Learned Baselines (7-day patterns)\n"+strings.Join(baselineInfo, "\n"))
		}
	}

	// Compute trends from metrics history (this works immediately, no learning period)
	metricsHistory := patrol.GetMetricsHistoryProvider()
	if metricsHistory != nil {
		var trendInfo []string
		duration := 24 * time.Hour // Last 24 hours

		// Get metrics for this resource
		// Try guest metrics first, then node metrics
		allMetrics := metricsHistory.GetAllGuestMetrics(resourceID, duration)
		if len(allMetrics) == 0 {
			// Try as node
			cpuPoints := metricsHistory.GetNodeMetrics(resourceID, "cpu", duration)
			memPoints := metricsHistory.GetNodeMetrics(resourceID, "memory", duration)
			if len(cpuPoints) > 0 {
				allMetrics["cpu"] = cpuPoints
			}
			if len(memPoints) > 0 {
				allMetrics["memory"] = memPoints
			}
		}

		// Analyze trends for each metric
		for metric, points := range allMetrics {
			if len(points) < 3 { // Need at least 3 points for trends
				continue
			}

			// Get first and last values
			first := points[0].Value
			last := points[len(points)-1].Value
			duration := points[len(points)-1].Timestamp.Sub(points[0].Timestamp)

			if duration < 30*time.Minute { // Need at least 30min of data
				continue
			}

			// Calculate rate of change per day
			change := last - first
			hoursSpan := duration.Hours()
			ratePerDay := change * (24 / hoursSpan)

			// Only report significant trends
			if metric == "disk" && ratePerDay > 0.5 { // Growing by >0.5GB/day
				trendInfo = append(trendInfo, fmt.Sprintf("**Disk** growing %.1f GB/day over last %.0fh",
					ratePerDay, hoursSpan))
			} else if metric == "memory" && absFloat(ratePerDay) > 2 { // >2% change per day
				direction := "growing"
				if ratePerDay < 0 {
					direction = "declining"
				}
				trendInfo = append(trendInfo, fmt.Sprintf("**Memory** %s %.1f%%/day over last %.0fh",
					direction, absFloat(ratePerDay), hoursSpan))
			} else if metric == "cpu" && absFloat(ratePerDay) > 5 { // >5% change per day
				direction := "increasing"
				if ratePerDay < 0 {
					direction = "decreasing"
				}
				trendInfo = append(trendInfo, fmt.Sprintf("**CPU** %s %.1f%%/day over last %.0fh",
					direction, absFloat(ratePerDay), hoursSpan))
			}
		}

		if len(trendInfo) > 0 {
			sections = append(sections, "### Trends (24h)\n"+strings.Join(trendInfo, "\n"))
		}
	}

	// Get pattern detector for predictions
	patternDetector := patrol.GetPatternDetector()
	if patternDetector != nil {
		predictions := patternDetector.GetPredictionsForResource(resourceID)
		if len(predictions) > 0 {
			var predInfo []string
			for _, pred := range predictions {
				if pred.DaysUntil < 30 { // Only show predictions within 30 days
					predInfo = append(predInfo, fmt.Sprintf("**%s** predicted in %.0f days (%.0f%% confidence)",
						pred.EventType, pred.DaysUntil, pred.Confidence*100))
				}
			}
			if len(predInfo) > 0 {
				sections = append(sections, "### Predictions\n"+strings.Join(predInfo, "\n"))
			}
		}
	}

	// Get change detector for recent changes
	if recentChanges := s.buildRecentResourceChangesContext(resourceID); recentChanges != "" {
		sections = append(sections, recentChanges)
	}

	// Get canonical relationship context from unified resources.
	if relationshipContext := s.buildResourceRelationshipContext(resourceID); relationshipContext != "" {
		sections = append(sections, relationshipContext)
	}

	// Get shared correlation context for learned related resources.
	if intel := patrol.GetIntelligence(); intel != nil {
		if ctx := intel.FormatCorrelationsContext(resourceID); ctx != "" {
			sections = append(sections, ctx)
		}
	}

	// Get alert history for this resource (this works immediately, uses existing alert data)
	s.mu.RLock()
	alertProvider := s.alertProvider
	s.mu.RUnlock()

	if alertProvider != nil {
		// Get active alerts for this resource
		activeAlerts := alertProvider.GetAlertsByResource(resourceID)

		// Get historical alerts (last 20)
		historicalAlerts := alertProvider.GetAlertHistory(resourceID, 20)

		if len(activeAlerts) > 0 || len(historicalAlerts) > 0 {
			var alertInfo []string

			if len(activeAlerts) > 0 {
				alertInfo = append(alertInfo, fmt.Sprintf("**%d active alert(s)** right now", len(activeAlerts)))
				for _, a := range activeAlerts {
					alertInfo = append(alertInfo, fmt.Sprintf("- %s %s: %s (active %s)",
						strings.ToUpper(a.Level), a.Type, a.Message, a.Duration))
				}
			}

			if len(historicalAlerts) > 0 {
				// Count alerts by type in the last 30 days
				alertsByType := make(map[string]int)
				cutoff := time.Now().Add(-30 * 24 * time.Hour)
				for _, a := range historicalAlerts {
					if a.ResolvedTime.After(cutoff) {
						alertsByType[a.Type]++
					}
				}

				if len(alertsByType) > 0 {
					var typeCounts []string
					for alertType, count := range alertsByType {
						typeCounts = append(typeCounts, fmt.Sprintf("%d %s", count, alertType))
					}
					alertInfo = append(alertInfo, fmt.Sprintf("**Past 30 days:** %s alerts", strings.Join(typeCounts, ", ")))
				}
			}

			if len(alertInfo) > 0 {
				sections = append(sections, "### Alert History\n"+strings.Join(alertInfo, "\n"))
			}
		}
	}

	// If no sections but baseline store exists, show learning status
	if len(sections) == 0 {
		baselineStore := patrol.GetBaselineStore()
		if baselineStore != nil {
			// Check if we're still learning baselines for this resource
			if rb, exists := baselineStore.GetResourceBaseline(resourceID); exists {
				// Count metrics with enough samples
				ready := 0
				learning := 0
				for _, mb := range rb.Metrics {
					if mb.SampleCount >= 10 {
						ready++
					} else if mb.SampleCount > 0 {
						learning++
					}
				}

				if learning > 0 && ready == 0 {
					return "\n\n## Historical Intelligence (Pulse Pro)\n*Learning baselines for this resource... Trends and anomaly detection will be available after more data is collected.*"
				}
			}
		}
		return ""
	}

	return "\n\n## Historical Intelligence (Pulse Pro)\n" + strings.Join(sections, "\n\n")
}

// absFloat returns the absolute value of a float64
func absFloat(x float64) float64 {
	if x < 0 {
		return -x
	}
	return x
}