RunAgentTestSuite executes a test suite using the agent loop approach
(ctx context.Context, testCases []models.TestCase)
| 38 | |
| 39 | // RunAgentTestSuite executes a test suite using the agent loop approach |
| 40 | func (tr *TestRunner) RunAgentTestSuite(ctx context.Context, testCases []models.TestCase) (*models.AgentReport, error) { |
| 41 | fmt.Printf("Starting agent test suite with %d test cases\n", len(testCases)) |
| 42 | |
| 43 | var wg sync.WaitGroup |
| 44 | resultsChan := make(chan models.AgentTestResult, len(testCases)) |
| 45 | |
| 46 | // Execute tests concurrently |
| 47 | for _, testCase := range testCases { |
| 48 | wg.Add(1) |
| 49 | go func(tc models.TestCase) { |
| 50 | defer wg.Done() |
| 51 | |
| 52 | fmt.Printf("Running agent test: %s\n", tc.Name) |
| 53 | result := tr.runAgentTest(ctx, tc) |
| 54 | resultsChan <- result |
| 55 | }(testCase) |
| 56 | } |
| 57 | |
| 58 | // Wait for all tests to complete |
| 59 | go func() { |
| 60 | wg.Wait() |
| 61 | close(resultsChan) |
| 62 | }() |
| 63 | |
| 64 | // Collect results and aggregate LLM metrics |
| 65 | var results []models.AgentTestResult |
| 66 | var totalTime time.Duration |
| 67 | var totalLLMRequests int |
| 68 | var totalLLMTime time.Duration |
| 69 | passedTests := 0 |
| 70 | failedTests := 0 |
| 71 | |
| 72 | for result := range resultsChan { |
| 73 | results = append(results, result) |
| 74 | totalTime += result.ResponseTime |
| 75 | |
| 76 | // Aggregate LLM metrics from successful responses |
| 77 | if result.Response != nil { |
| 78 | totalLLMRequests += result.Response.LLMRequests |
| 79 | totalLLMTime += result.Response.LLMTotalTime |
| 80 | } |
| 81 | |
| 82 | if result.Success { |
| 83 | passedTests++ |
| 84 | } else { |
| 85 | failedTests++ |
| 86 | } |
| 87 | } |
| 88 | |
| 89 | // Calculate average times |
| 90 | var averageTime time.Duration |
| 91 | var avgTimePerReq time.Duration |
| 92 | if len(results) > 0 { |
| 93 | averageTime = totalTime / time.Duration(len(results)) |
| 94 | } |
| 95 | if totalLLMRequests > 0 { |
| 96 | avgTimePerReq = totalLLMTime / time.Duration(totalLLMRequests) |
| 97 | } |