Method runAgentTest

services/test_runner.go:116–172

runAgentTest executes a single test case using the agent loop

(ctx context.Context, testCase models.TestCase)

// runAgentTest executes a single test case using the agent loop
func (tr *TestRunner) runAgentTest(ctx context.Context, testCase models.TestCase) models.AgentTestResult {
	startTime := time.Now()

	// Generate a unique session ID for this test
	sessionID := fmt.Sprintf("test_%s_%d", testCase.Name, time.Now().UnixNano())

	// Create a session for the test
	session := &models.ChatSession{
		SessionID: sessionID,
		Messages:  []models.ChatMessage{},
		CreatedAt: time.Now(),
		UpdatedAt: time.Now(),
	}

	// Initialize cart state if specified in the test case
	if testCase.InitialCartState != nil {
		err := tr.openaiService.InitializeCartForTest(sessionID, testCase.InitialCartState)
		if err != nil {
			return models.AgentTestResult{
				TestCase:     testCase,
				ModelName:    tr.getModelName(),
				Success:      false,
				ErrorMessage: fmt.Sprintf("Failed to initialize cart state: %v", err),
				Timestamp:    time.Now(),
				ResponseTime: time.Since(startTime),
			}
		}
	}

	// Execute the test using the agent loop
	response, err := tr.openaiService.ProcessChatMessage(ctx, testCase.Prompt, session, testCase.Name)
	responseTime := time.Since(startTime)

	if err != nil {
		return models.AgentTestResult{
			TestCase:     testCase,
			ModelName:    tr.getModelName(),
			Success:      false,
			ErrorMessage: err.Error(),
			Timestamp:    time.Now(),
			ResponseTime: responseTime,
		}
	}

	// Evaluate if the test was successful by checking tool calls
	success, matchedPath := tr.evaluateAgentResponse(testCase, response)

	return models.AgentTestResult{
		TestCase:     testCase,
		ModelName:    tr.getModelName(),
		Response:     response,
		Success:      success,
		MatchedPath:  matchedPath,
		Timestamp:    time.Now(),
		ResponseTime: responseTime,
	}
}