()
| 15 | ) |
| 16 | |
| 17 | func main() { |
| 18 | // Command line flags |
| 19 | var ( |
| 20 | apiKey = flag.String("api-key", "DMR", "OpenAI API key (or set OPENAI_API_KEY env var)") |
| 21 | baseURL = flag.String("base-url", "http://localhost:12434/engines/v1", "OpenAI API base URL (or set OPENAI_BASE_URL env var)") |
| 22 | model = flag.String("model", "", "Model to use (or set OPENAI_MODEL env var, defaults to gpt-4o-mini)") |
| 23 | configFile = flag.String("config", "config/test_cases.json", "Path to test cases configuration file") |
| 24 | testCase = flag.String("test-case", "", "Run only the specified test case by name") |
| 25 | ) |
| 26 | flag.Parse() |
| 27 | |
| 28 | // Load test cases |
| 29 | testCases, err := loadTestCases(*configFile, *testCase) |
| 30 | if err != nil { |
| 31 | log.Fatalf("Failed to load test cases: %v", err) |
| 32 | } |
| 33 | |
| 34 | // Generate output filenames with model name |
| 35 | sanitizedModel := sanitizeModelName(*model) |
| 36 | timestamp := time.Now().Format("20060102_150405") |
| 37 | outputFile := fmt.Sprintf("results/agent_test_results_%s_%s.json", sanitizedModel, timestamp) |
| 38 | logFile := fmt.Sprintf("logs/agent_test_logs_%s_%s.log", sanitizedModel, timestamp) |
| 39 | |
| 40 | // Ensure directories exist |
| 41 | if err := os.MkdirAll("results", 0755); err != nil { |
| 42 | log.Fatalf("Failed to create results directory: %v", err) |
| 43 | } |
| 44 | if err := os.MkdirAll("logs", 0755); err != nil { |
| 45 | log.Fatalf("Failed to create logs directory: %v", err) |
| 46 | } |
| 47 | |
| 48 | // Create request logger |
| 49 | logger, err := services.NewRequestLogger(logFile) |
| 50 | if err != nil { |
| 51 | log.Fatalf("Failed to create request logger: %v", err) |
| 52 | } |
| 53 | defer logger.Close() |
| 54 | |
| 55 | // Create test runner with logger |
| 56 | runner := services.NewTestRunnerWithLogger(*apiKey, *baseURL, *model, logger) |
| 57 | |
| 58 | // Print test configuration |
| 59 | fmt.Printf("🚀 Starting Agent Loop Tool Efficiency Test\n") |
| 60 | fmt.Printf("📊 Configuration:\n") |
| 61 | fmt.Printf(" Base URL: %s\n", *baseURL) |
| 62 | modelName := *model |
| 63 | if modelName == "" { |
| 64 | modelName = "gpt-4o-mini (default)" |
| 65 | } |
| 66 | fmt.Printf(" Model: %s\n", modelName) |
| 67 | if *testCase != "" { |
| 68 | fmt.Printf(" Single Test Case: %s\n", *testCase) |
| 69 | } |
| 70 | fmt.Printf(" Test Cases: %d\n", len(testCases)) |
| 71 | fmt.Printf(" Output: %s\n", outputFile) |
| 72 | fmt.Printf(" Log File: %s\n", logFile) |
| 73 | fmt.Println() |
| 74 |
nothing calls this directly
no test coverage detected