Files
nofx/assistant/agent.go
tinklefund 01ba348841 feat: Add Telegram AI Assistant (moltbot-nofx integration)
- Add assistant package with AI Agent runtime
  - agent.go: Core agent loop with tool calling
  - session.go: Conversation memory management
  - tool.go: Tool interface and base implementation
  - trading_tools.go: Trading-specific tools (13 tools)
  - prompts.go: Trading expert system prompts (EN/ZH)

- Add telegram package for Telegram bot integration
  - bot.go: Telegram bot with rate limiting & access control
  - config.go: Environment-based configuration

- Update main.go to initialize Telegram bot on startup
- Update .env.example with new configuration options
- Add gopkg.in/telebot.v3 dependency

Trading tools available:
- Query: get_balance, get_positions, list_traders, get_trader_status
- Control: start_trader, stop_trader
- Trading: get_market_price, open_long, open_short, close_position
- Config: list_strategies, list_exchanges, list_ai_models
2026-01-30 03:29:22 +08:00

351 lines
8.7 KiB
Go

// Package assistant implements the AI Agent runtime with tool calling capabilities
// Inspired by moltbot's agent architecture, specialized for trading
package assistant
import (
"context"
"encoding/json"
"fmt"
"nofx/logger"
"nofx/mcp"
"strings"
"sync"
"time"
)
// Agent represents an AI assistant with tool-calling capabilities
type Agent struct {
// AI client for LLM calls
aiClient mcp.AIClient
// Tool registry
tools map[string]Tool
toolsLock sync.RWMutex
// Session/memory management
sessions map[string]*Session
sessionsLock sync.RWMutex
// Configuration
config AgentConfig
// System prompt
systemPrompt string
}
// AgentConfig holds agent configuration
type AgentConfig struct {
// Max tool calls per turn (prevent infinite loops)
MaxToolCalls int `json:"max_tool_calls"`
// Max conversation history to keep
MaxHistoryMessages int `json:"max_history_messages"`
// Timeout for single AI call
AITimeout time.Duration `json:"ai_timeout"`
// Model to use
Model string `json:"model"`
}
// DefaultAgentConfig returns sensible defaults
func DefaultAgentConfig() AgentConfig {
return AgentConfig{
MaxToolCalls: 10,
MaxHistoryMessages: 50,
AITimeout: 120 * time.Second,
Model: "deepseek-chat",
}
}
// NewAgent creates a new AI agent
func NewAgent(aiClient mcp.AIClient, config AgentConfig) *Agent {
agent := &Agent{
aiClient: aiClient,
tools: make(map[string]Tool),
sessions: make(map[string]*Session),
config: config,
}
// Set default system prompt
agent.systemPrompt = DefaultTradingSystemPrompt()
return agent
}
// RegisterTool adds a tool to the agent's toolkit
func (a *Agent) RegisterTool(tool Tool) {
a.toolsLock.Lock()
defer a.toolsLock.Unlock()
a.tools[tool.Name()] = tool
logger.Infof("🔧 Registered tool: %s", tool.Name())
}
// RegisterTools adds multiple tools
func (a *Agent) RegisterTools(tools ...Tool) {
for _, tool := range tools {
a.RegisterTool(tool)
}
}
// SetSystemPrompt sets the agent's system prompt
func (a *Agent) SetSystemPrompt(prompt string) {
a.systemPrompt = prompt
}
// GetSession returns or creates a session for the given ID
func (a *Agent) GetSession(sessionID string) *Session {
a.sessionsLock.Lock()
defer a.sessionsLock.Unlock()
if session, ok := a.sessions[sessionID]; ok {
return session
}
session := NewSession(sessionID, a.config.MaxHistoryMessages)
a.sessions[sessionID] = session
return session
}
// Chat processes a user message and returns the agent's response
// This is the main entry point for the agent loop
func (a *Agent) Chat(ctx context.Context, sessionID string, userMessage string) (*AgentResponse, error) {
session := a.GetSession(sessionID)
// Add user message to history
session.AddMessage(Message{
Role: "user",
Content: userMessage,
Timestamp: time.Now(),
})
// Build the full prompt with tools
systemPrompt := a.buildSystemPromptWithTools()
conversationPrompt := a.buildConversationPrompt(session)
// Agent loop - keep calling AI until it's done or max iterations
var finalResponse string
toolCallCount := 0
for {
// Check context cancellation
if ctx.Err() != nil {
return nil, ctx.Err()
}
// Check max tool calls
if toolCallCount >= a.config.MaxToolCalls {
logger.Warnf("⚠️ Max tool calls reached (%d), stopping agent loop", a.config.MaxToolCalls)
break
}
// Call AI
response, err := a.aiClient.CallWithMessages(systemPrompt, conversationPrompt)
if err != nil {
return nil, fmt.Errorf("AI call failed: %w", err)
}
// Parse response for tool calls
toolCalls, textResponse, err := a.parseResponse(response)
if err != nil {
// If parsing fails, treat entire response as text
finalResponse = response
break
}
// If no tool calls, we're done
if len(toolCalls) == 0 {
finalResponse = textResponse
break
}
// Execute tool calls
toolResults := a.executeToolCalls(ctx, toolCalls)
toolCallCount += len(toolCalls)
// Add tool calls and results to conversation for next iteration
conversationPrompt += fmt.Sprintf("\n\nAssistant called tools:\n%s\n\nTool results:\n%s\n\nBased on the tool results, please provide your response to the user:",
formatToolCalls(toolCalls),
formatToolResults(toolResults))
// If there's also a text response, capture it
if textResponse != "" {
finalResponse = textResponse
}
}
// Add assistant response to history
session.AddMessage(Message{
Role: "assistant",
Content: finalResponse,
Timestamp: time.Now(),
})
return &AgentResponse{
Text: finalResponse,
SessionID: sessionID,
}, nil
}
// buildSystemPromptWithTools creates the system prompt including tool definitions
func (a *Agent) buildSystemPromptWithTools() string {
a.toolsLock.RLock()
defer a.toolsLock.RUnlock()
var toolDefs []string
for _, tool := range a.tools {
toolDef := fmt.Sprintf(`- **%s**: %s
Parameters: %s`, tool.Name(), tool.Description(), tool.ParameterSchema())
toolDefs = append(toolDefs, toolDef)
}
toolsSection := ""
if len(toolDefs) > 0 {
toolsSection = fmt.Sprintf(`
## Available Tools
You can call tools by responding with JSON in this format:
{"tool_calls": [{"name": "tool_name", "arguments": {"param": "value"}}]}
After receiving tool results, provide a natural language response to the user.
Tools:
%s
`, strings.Join(toolDefs, "\n"))
}
return a.systemPrompt + toolsSection
}
// buildConversationPrompt builds the conversation history as a prompt
func (a *Agent) buildConversationPrompt(session *Session) string {
messages := session.GetMessages()
var parts []string
for _, msg := range messages {
parts = append(parts, fmt.Sprintf("%s: %s", strings.Title(msg.Role), msg.Content))
}
return strings.Join(parts, "\n\n")
}
// parseResponse extracts tool calls and text from AI response
func (a *Agent) parseResponse(response string) ([]ToolCall, string, error) {
// Try to find JSON tool calls in response
// Look for {"tool_calls": [...]} pattern
var toolCalls []ToolCall
textResponse := response
// Try to parse as JSON
if strings.Contains(response, "tool_calls") {
// Find JSON block
start := strings.Index(response, "{")
end := strings.LastIndex(response, "}")
if start >= 0 && end > start {
jsonStr := response[start : end+1]
var parsed struct {
ToolCalls []struct {
Name string `json:"name"`
Arguments json.RawMessage `json:"arguments"`
} `json:"tool_calls"`
}
if err := json.Unmarshal([]byte(jsonStr), &parsed); err == nil {
for _, tc := range parsed.ToolCalls {
toolCalls = append(toolCalls, ToolCall{
Name: tc.Name,
Arguments: tc.Arguments,
})
}
// Extract text before/after JSON
textResponse = strings.TrimSpace(response[:start] + response[end+1:])
}
}
}
return toolCalls, textResponse, nil
}
// executeToolCalls runs the requested tools
func (a *Agent) executeToolCalls(ctx context.Context, calls []ToolCall) []ToolResult {
a.toolsLock.RLock()
defer a.toolsLock.RUnlock()
var results []ToolResult
for _, call := range calls {
tool, ok := a.tools[call.Name]
if !ok {
results = append(results, ToolResult{
Name: call.Name,
Error: fmt.Sprintf("unknown tool: %s", call.Name),
})
continue
}
logger.Infof("🔧 Executing tool: %s", call.Name)
result, err := tool.Execute(ctx, call.Arguments)
if err != nil {
logger.Errorf("❌ Tool %s failed: %v", call.Name, err)
results = append(results, ToolResult{
Name: call.Name,
Error: err.Error(),
})
} else {
logger.Infof("✅ Tool %s completed", call.Name)
results = append(results, ToolResult{
Name: call.Name,
Result: result,
})
}
}
return results
}
// ToolCall represents a tool invocation request from the AI
type ToolCall struct {
Name string `json:"name"`
Arguments json.RawMessage `json:"arguments"`
}
// ToolResult represents the result of a tool execution
type ToolResult struct {
Name string `json:"name"`
Result interface{} `json:"result,omitempty"`
Error string `json:"error,omitempty"`
}
// AgentResponse is the final response from the agent
type AgentResponse struct {
Text string `json:"text"`
SessionID string `json:"session_id"`
}
func formatToolCalls(calls []ToolCall) string {
var parts []string
for _, c := range calls {
parts = append(parts, fmt.Sprintf("- %s(%s)", c.Name, string(c.Arguments)))
}
return strings.Join(parts, "\n")
}
func formatToolResults(results []ToolResult) string {
var parts []string
for _, r := range results {
if r.Error != "" {
parts = append(parts, fmt.Sprintf("- %s: ERROR: %s", r.Name, r.Error))
} else {
resultJSON, _ := json.Marshal(r.Result)
parts = append(parts, fmt.Sprintf("- %s: %s", r.Name, string(resultJSON)))
}
}
return strings.Join(parts, "\n")
}