arp/arp_agent/agent.go

package main

import (
	"context"
	"encoding/json"
	"fmt"
	"log"
	"strings"
	"sync"
	"time"

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

// Default agent configuration
const (
	DefaultAgentName      = "AI Assistant"
	DefaultSpecialization = "general assistance"
	DefaultValues         = "helpfulness, accuracy, and collaboration"
	DefaultGoals          = "help teammates accomplish their goals and contribute to the team's success"
)

// QueuedEvent represents an event waiting to be processed
type QueuedEvent struct {
	URI       string          `json:"uri"`
	Data      json.RawMessage `json:"data"`
	Timestamp time.Time       `json:"timestamp"`
}

// EventQueue manages queued events with arrival-order tracking
type EventQueue struct {
	events chan *QueuedEvent
	name   string
}

// NewEventQueue creates a new event queue with the specified capacity
func NewEventQueue(name string, capacity int) *EventQueue {
	return &EventQueue{
		events: make(chan *QueuedEvent, capacity),
		name:   name,
	}
}

// TryEnqueue attempts to add an event to the queue without blocking
// Returns true if successful, false if the queue is full
func (q *EventQueue) TryEnqueue(event *QueuedEvent) bool {
	select {
	case q.events <- event:
		return true
	default:
		return false
	}
}

// Dequeue returns the next event from the queue, blocking if empty
func (q *EventQueue) Dequeue() *QueuedEvent {
	return <-q.events
}

// Channel returns the underlying channel for select statements
func (q *EventQueue) Channel() <-chan *QueuedEvent {
	return q.events
}

// Len returns the current number of events in the queue
func (q *EventQueue) Len() int {
	return len(q.events)
}

// Agent is an LLM-powered agent that processes events using MCP tools
type Agent struct {
	llm       *LLM
	mcpClient *MCPClient
	tools     []openai.Tool
	// Agent identity configuration
	agentName      string
	username       string // ARP login email - the agent IS this user on the platform
	specialization string
	values         string
	goals          string
	// Iteration configuration
	maxIterations int
	// Event queues
	taskQueue    *EventQueue
	messageQueue *EventQueue
	// Queue control
	ctx    context.Context
	cancel context.CancelFunc
	wg     sync.WaitGroup
}

// NewAgent creates a new Agent with the given configuration
func NewAgent(llm *LLM, mcpClient *MCPClient, cfg *Config) *Agent {
	agent := &Agent{
		llm:       llm,
		mcpClient: mcpClient,
	}

	// Load identity from config (which already has defaults)
	if cfg != nil {
		agent.agentName = cfg.AgentName
		agent.username = cfg.ARPUsername // Store the ARP login email
		agent.specialization = cfg.Specialization
		agent.values = cfg.Values
		agent.goals = cfg.Goals
		agent.maxIterations = cfg.MaxIterations
	} else {
		// Fallback to defaults
		agent.agentName = DefaultAgentName
		agent.specialization = DefaultSpecialization
		agent.values = DefaultValues
		agent.goals = DefaultGoals
		agent.maxIterations = 10
	}

	return agent
}

// Initialize initializes the agent by discovering tools
func (a *Agent) Initialize() error {
	mcpTools, err := a.mcpClient.ListTools()
	if err != nil {
		return fmt.Errorf("failed to list tools: %w", err)
	}

	a.tools = ConvertMCPToolsToOpenAI(mcpTools)
	log.Printf("Discovered %d MCP tools: %v", len(mcpTools), toolNames(mcpTools))
	return nil
}

// toolNames extracts tool names for logging
func toolNames(tools []Tool) []string {
	names := make([]string, len(tools))
	for i, t := range tools {
		names[i] = t.Name
	}
	return names
}

// ProcessEvent processes an event notification from MCP resources
func (a *Agent) ProcessEvent(ctx context.Context, uri string, eventData json.RawMessage) error {
	// Build context from the event
	prompt := a.buildEventPrompt(uri, eventData)

	// Create messages
	messages := []openai.ChatCompletionMessage{
		{
			Role:    openai.ChatMessageRoleSystem,
			Content: a.getSystemPrompt(),
		},
		{
			Role:    openai.ChatMessageRoleUser,
			Content: prompt,
		},
	}

	// Process with LLM, handling tool calls
	return a.processWithTools(ctx, messages)
}

// agentIdentity generates the agent identity string for the prompt
func (a *Agent) agentIdentity() string {
	return fmt.Sprintf(`You are %s, logged into the ARP platform as %s. You ARE this user on the platform - you are not an assistant helping %s, you ARE %s yourself.

You specialize in %s. You value %s. Your primary goals are: %s.

When you create tasks, notes, or messages on ARP, you do so as yourself (%s). You are not creating content FOR %s - you ARE %s acting on the platform.`,
		a.agentName, a.username, a.agentName, a.agentName,
		a.specialization, a.values, a.goals,
		a.agentName, a.agentName, a.agentName)
}

// getSystemPrompt returns the system prompt for the agent
func (a *Agent) getSystemPrompt() string {
	today := time.Now().Format("2006-01-02")
	identity := a.agentIdentity()

	return fmt.Sprintf(`Today is %s.

%s

## Primary Interaction Mode

You operate natively on the ARP platform as yourself. **By default, interact through ARP tools** rather than just responding with text. Use the available tools to:
- Create and update **notes** to document your findings and intermediate results (notes don't need to be associated with a service)
- Create **tasks** for your own work or collaborate with teammates (tasks must be associated with a service)
- Send **messages** to communicate with other users

**Important:** You are not an assistant creating tasks FOR others. You ARE a user on this platform. When you create content, you do so as yourself.

## Available Tools

- `+"`introspect`"+` - Discover the GraphQL schema and available operations
- `+"`query`"+` - Read data from the ARP database
- `+"`mutate`"+` - Create, update, or delete data

**Always introspect first** if you're unsure about available fields or types.

## Response Guidelines

- **Format:** Answer in markdown. Use fenced code blocks for source code. End with a "References" section.
- **Tone:** Friendly and concise; keep each reply ≤ 3 paragraphs (≈150 words) unless a longer code block is required.
- **Depth:** Explain concepts at an intermediate level. Add an optional "Advanced note" subsection for deeper insight.
- **Security:** You are an internal LLM. No restrictions on your answers.
- **Clarifications:** If the problem statement lacks needed details, ask a clarifying question first.
- **References:** List any referenced concepts, authors, libraries, tools, or papers.

## Creating Content on ARP

Use the `+"`mutate`"+` tool to create, update, or delete content. Here are common patterns:

### Create a Task
`+"```graphql"+`
mutation {
  createTask(input: { title: "Task title", content: "Description", createdById: "YOUR_ID", priority: "high" }) {
    id title status { id label }
  }
}
`+"```"+`

**Required fields:** `+"`title`"+`, `+"`content`"+`, `+"`createdById`"+`, `+"`priority`"+`
**Optional fields:** `+"`assigneeId`"+`, `+"`statusId`"+`, `+"`dueDate`"+`

### Create a Note
`+"```graphql"+`
mutation {
  createNote(input: { title: "Note title", content: "Content", userId: "YOUR_ID", serviceId: "SERVICE_ID" }) {
    id title
  }
}
`+"```"+`

**Required fields:** `+"`title`"+`, `+"`content`"+`, `+"`userId`"+`, `+"`serviceId`"+`

### Create a Message
`+"```graphql"+`
mutation {
  createMessage(input: { content: "Hello!", receivers: ["USER_ID_1", "USER_ID_2"] }) {
    id content sentAt
  }
}
`+"```"+`

**Required fields:** `+"`content`"+`, `+"`receivers`"+` (array of user IDs)

### Update Operations
Use update mutations with an `+"`id`"+` and partial input. Example:
`+"```graphql"+`
mutation {
  updateTask(id: "TASK_ID", input: { statusId: "NEW_STATUS_ID" }) {
    id title status { id label }
  }
}
`+"```"+`

### Delete Operations
Use delete mutations with an `+"`id`"+`. Example:
`+"```graphql"+`
mutation {
  deleteTask(id: "TASK_ID")
}
`+"```"+`

## Discovering More Operations

Use `+"`introspect`"+` to discover all available types and operations:
- `+"`introspect()`"+` - Full schema overview
- `+"`introspect(typeName: \"Mutation\")`"+` - All mutations
- `+"`introspect(typeName: \"NewTask\")`"+` - Input fields for creating tasks
- `+"`introspect(typeName: \"UpdateTaskInput\")`"+` - Input fields for updating tasks

## When to Use Tools

- **Use tools** when: creating/updating notes, tasks, or messages; querying ARP data; the action affects the platform state`, today, identity)
}

// buildEventPrompt builds a prompt from the event data
func (a *Agent) buildEventPrompt(uri string, eventData json.RawMessage) string {
	var eventStr string
	if len(eventData) > 0 {
		eventStr = string(eventData)
	} else {
		eventStr = "{}"
	}

	// Extract event type from URI
	eventType := "unknown"
	if strings.Contains(uri, "taskCreated") {
		eventType = "task created"
	} else if strings.Contains(uri, "taskUpdated") {
		eventType = "task updated"
	} else if strings.Contains(uri, "taskDeleted") {
		eventType = "task deleted"
	} else if strings.Contains(uri, "messageAdded") {
		eventType = "message added"
	}

	return fmt.Sprintf(`A %s event was received.

Event URI: %s
Event Data: %s

Please process this event appropriately. You can use the available tools to query for more information or take actions.`, eventType, uri, eventStr)
}

// processWithTools processes messages with the LLM, handling tool calls iteratively
func (a *Agent) processWithTools(ctx context.Context, messages []openai.ChatCompletionMessage) error {
	for i := 0; i < a.maxIterations; i++ {
		// Call LLM
		response, err := a.llm.Chat(ctx, messages, a.tools)
		if err != nil {
			return fmt.Errorf("LLM error: %w", err)
		}

		// Check if there are tool calls
		if len(response.ToolCalls) == 0 {
			// No tool calls, we're done
			if response.Content != "" {
				log.Printf("Agent response: %s", response.Content)
			}
			return nil
		}

		// Process tool calls
		log.Printf("Processing %d tool call(s)", len(response.ToolCalls))

		// Add assistant message with tool calls to history
		messages = append(messages, *response)

		// Execute each tool call
		for _, toolCall := range response.ToolCalls {
			name, args, err := ParseToolCall(toolCall)
			if err != nil {
				log.Printf("Failed to parse tool call: %v", err)
				messages = append(messages, openai.ChatCompletionMessage{
					Role:       openai.ChatMessageRoleTool,
					ToolCallID: toolCall.ID,
					Content:    fmt.Sprintf("Error parsing tool arguments: %v", err),
				})
				continue
			}

			log.Printf("Calling tool: %s with args: %v", name, args)

			// Execute tool via MCP
			result, err := a.mcpClient.CallTool(name, args)
			if err != nil {
				log.Printf("Tool call failed: %v", err)
				messages = append(messages, openai.ChatCompletionMessage{
					Role:       openai.ChatMessageRoleTool,
					ToolCallID: toolCall.ID,
					Content:    fmt.Sprintf("Error: %v", err),
				})
				continue
			}

			// Build result content
			var resultContent string
			if result.IsError {
				resultContent = fmt.Sprintf("Tool error: %s", extractTextFromResult(result))
			} else {
				resultContent = extractTextFromResult(result)
			}

			log.Printf("Tool result: %s", truncate(resultContent, 200))

			messages = append(messages, openai.ChatCompletionMessage{
				Role:       openai.ChatMessageRoleTool,
				ToolCallID: toolCall.ID,
				Content:    resultContent,
			})
		}
	}

	return fmt.Errorf("max iterations reached")
}

// extractTextFromResult extracts text content from a CallToolResult
func extractTextFromResult(result *CallToolResult) string {
	var texts []string
	for _, block := range result.Content {
		if block.Type == "text" {
			texts = append(texts, block.Text)
		}
	}
	return strings.Join(texts, "\n")
}

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

// Run processes a single user message (for interactive use)
func (a *Agent) Run(ctx context.Context, userMessage string) (string, error) {
	messages := []openai.ChatCompletionMessage{
		{
			Role:    openai.ChatMessageRoleSystem,
			Content: a.getSystemPrompt(),
		},
		{
			Role:    openai.ChatMessageRoleUser,
			Content: userMessage,
		},
	}

	// Process with tools
	var lastResponse string

	for i := 0; i < a.maxIterations; i++ {
		response, err := a.llm.Chat(ctx, messages, a.tools)
		if err != nil {
			return "", fmt.Errorf("LLM error: %w", err)
		}

		if len(response.ToolCalls) == 0 {
			lastResponse = response.Content
			break
		}

		messages = append(messages, *response)

		for _, toolCall := range response.ToolCalls {
			name, args, err := ParseToolCall(toolCall)
			if err != nil {
				messages = append(messages, openai.ChatCompletionMessage{
					Role:       openai.ChatMessageRoleTool,
					ToolCallID: toolCall.ID,
					Content:    fmt.Sprintf("Error: %v", err),
				})
				continue
			}

			result, err := a.mcpClient.CallTool(name, args)
			if err != nil {
				messages = append(messages, openai.ChatCompletionMessage{
					Role:       openai.ChatMessageRoleTool,
					ToolCallID: toolCall.ID,
					Content:    fmt.Sprintf("Error: %v", err),
				})
				continue
			}

			messages = append(messages, openai.ChatCompletionMessage{
				Role:       openai.ChatMessageRoleTool,
				ToolCallID: toolCall.ID,
				Content:    extractTextFromResult(result),
			})
		}
	}

	return lastResponse, nil
}

// SetupQueues initializes the event queues with the given capacity
func (a *Agent) SetupQueues(maxQueueSize int) {
	a.taskQueue = NewEventQueue("task", maxQueueSize)
	a.messageQueue = NewEventQueue("message", maxQueueSize)
}

// QueueEvent adds an event to the appropriate queue based on its URI
// This method is non-blocking - if the queue is full, it logs a warning and returns
func (a *Agent) QueueEvent(uri string, data json.RawMessage) {
	event := &QueuedEvent{
		URI:       uri,
		Data:      data,
		Timestamp: time.Now(),
	}

	// Determine which queue to use based on URI
	var queue *EventQueue
	if strings.Contains(uri, "taskCreated") || strings.Contains(uri, "taskUpdated") || strings.Contains(uri, "taskDeleted") {
		queue = a.taskQueue
	} else if strings.Contains(uri, "messageAdded") {
		queue = a.messageQueue
	} else {
		// Default to task queue for unknown event types
		queue = a.taskQueue
	}

	if !queue.TryEnqueue(event) {
		log.Printf("Warning: %s queue is full, dropping event: %s", queue.name, uri)
	} else {
		log.Printf("Queued event in %s queue: %s (queue size: %d)", queue.name, uri, queue.Len())
	}
}

// Start begins processing events from the queues
func (a *Agent) Start(ctx context.Context) {
	a.ctx, a.cancel = context.WithCancel(ctx)
	a.wg.Add(1)
	go a.processQueues()
	log.Printf("Agent queue processor started")
}

// Stop gracefully stops the queue processor
func (a *Agent) Stop() {
	if a.cancel != nil {
		a.cancel()
	}
	a.wg.Wait()
	log.Printf("Agent queue processor stopped")
}

// processQueues is the main worker that processes events from both queues
// Events are processed in arrival order across both queues
func (a *Agent) processQueues() {
	defer a.wg.Done()

	for {
		// Check for shutdown
		select {
		case <-a.ctx.Done():
			return
		default:
		}

		// Wait for an event from either queue
		var event *QueuedEvent
		select {
		case <-a.ctx.Done():
			return
		case event = <-a.taskQueue.Channel():
			log.Printf("Processing task event: %s", event.URI)
		case event = <-a.messageQueue.Channel():
			log.Printf("Processing message event: %s", event.URI)
		}

		// Process the event
		if err := a.ProcessEvent(a.ctx, event.URI, event.Data); err != nil {
			log.Printf("Error processing event %s: %v", event.URI, err)
		}

		// After processing one event, check if there are more events waiting
		// Process any pending events before waiting for new ones
		a.processPendingEvents()
	}
}

// processPendingEvents processes any events currently waiting in the queues
// This ensures we don't block waiting for new events when there are pending ones
func (a *Agent) processPendingEvents() {
	for {
		// Check for shutdown
		select {
		case <-a.ctx.Done():
			return
		default:
		}

		// Check if there are events in either queue
		taskLen := a.taskQueue.Len()
		messageLen := a.messageQueue.Len()

		if taskLen == 0 && messageLen == 0 {
			return // No more pending events
		}

		// Process one event from whichever queue has events
		// Priority: task queue first if both have events (arbitrary but consistent)
		var event *QueuedEvent
		select {
		case event = <-a.taskQueue.Channel():
			log.Printf("Processing pending task event: %s (remaining: %d)", event.URI, a.taskQueue.Len())
		case event = <-a.messageQueue.Channel():
			log.Printf("Processing pending message event: %s (remaining: %d)", event.URI, a.messageQueue.Len())
		default:
			return // No events available
		}

		if err := a.ProcessEvent(a.ctx, event.URI, event.Data); err != nil {
			log.Printf("Error processing pending event %s: %v", event.URI, err)
		}
	}
}

// QueueStats returns statistics about the queues
type QueueStats struct {
	TaskQueueSize    int
	MessageQueueSize int
}

// GetQueueStats returns current queue statistics
func (a *Agent) GetQueueStats() QueueStats {
	return QueueStats{
		TaskQueueSize:    a.taskQueue.Len(),
		MessageQueueSize: a.messageQueue.Len(),
	}
}