AlwinXavier19/gaia-benchmark-agent

1"""2test_tools.py — Quick smoke tests for all GAIA agent tools.3 4Run:5    python test_tools.py6"""7 8import os9import sys10 11# Force UTF-8 output on Windows so arrow/emoji chars don't crash12if sys.stdout.encoding and sys.stdout.encoding.lower() != "utf-8":13    import io14    sys.stdout = io.TextIOWrapper(sys.stdout.buffer, encoding="utf-8", errors="replace")15 16from dotenv import load_dotenv17 18load_dotenv()19 20# ──────────────────────────────────────────────────────────────────────────────21# Colour helpers (works on Windows too)22# ──────────────────────────────────────────────────────────────────────────────23 24GREEN  = "\033[92m"25RED    = "\033[91m"26YELLOW = "\033[93m"27RESET  = "\033[0m"28 29PASS = f"{GREEN}PASS{RESET}"30FAIL = f"{RED}FAIL{RESET}"31SKIP = f"{YELLOW}SKIP{RESET}"32 33 34def test(name: str, fn, *args, expect_contains: str = "", **kwargs):35    try:36        result = fn(*args, **kwargs)37        result_str = str(result)38        if expect_contains and expect_contains.lower() not in result_str.lower():39            print(f"[{FAIL}] {name}")40            print(f"       Expected to contain: '{expect_contains}'")41            print(f"       Got: {result_str[:200]}")42        else:43            print(f"[{PASS}] {name}")44            print(f"       → {result_str[:120]}")45    except Exception as exc:46        print(f"[{FAIL}] {name}: {exc}")47 48 49# ──────────────────────────────────────────────────────────────────────────────50# Tests51# ──────────────────────────────────────────────────────────────────────────────52 53from tools import (54    reverse_text,55    calculate,56    decode_text,57    web_search,58    wikipedia_search,59    execute_python_code,60)61 62 63def run_tests():64    print("\n" + "=" * 60)65    print("  GAIA Agent Tool Tests")66    print("=" * 60 + "\n")67 68    # ── Pure utility tests (no API needed) ────────────────────────────────69    test("reverse_text basic", reverse_text, ".rewsna eht si siht",70         expect_contains="this is the answer")71 72    test("reverse_text short", reverse_text, "hello",73         expect_contains="olleh")74 75    test("calculate add", calculate, "2 + 2",76         expect_contains="4")77 78    test("calculate complex", calculate, "sqrt(144) + 3**2",79         expect_contains="21")80 81    test("calculate pi", calculate, "round(pi, 4)",82         expect_contains="3.1416")83 84    test("decode_text reverse", decode_text, "!dlrow olleh",85         expect_contains="hello world")86 87    test("execute_python_code simple", execute_python_code,88         "print(2 ** 10)", expect_contains="1024")89 90    test("execute_python_code multiline", execute_python_code,91         "x = [i**2 for i in range(5)]\nprint(sum(x))",92         expect_contains="30")93 94    # ── Network tests (require internet) ──────────────────────────────────95    print(f"\n--- Network Tests ---")96 97    test("web_search python", web_search, "Python programming language",98         expect_contains="python")99 100    test("wikipedia_search france", wikipedia_search, "France",101         expect_contains="france")102 103    test("wikipedia_search albert einstein", wikipedia_search, "Albert Einstein",104         expect_contains="einstein")105 106    print("\n" + "=" * 60)107    print("Tests complete!")108    print("=" * 60 + "\n")109 110 111if __name__ == "__main__":112    run_tests()113