test_run.py

#!/usr/bin/env python
"""Free local checks for the Handicate framework (no GPU, no model download).

Verifies: config imports, every module parses, the rubric + feedback load and compose,
the GRPO reward func has the right shape, and the eval gate logic is correct.
Run: python test_run.py
"""
import ast
import os
import sys
from pathlib import Path

HERE = Path(__file__).resolve().parent
sys.path.insert(0, str(HERE))


def ok(m): print("  [PASS]", m)
def fail(m): print("  [FAIL]", m); sys.exit(1)


def main():
    print("LOCAL checks (free):")

    # every .py parses
    for p in HERE.rglob("*.py"):
        if p.name == "test_run.py":
            continue
        try:
            ast.parse(p.read_text(encoding="utf-8"))
        except SyntaxError as e:
            fail(f"{p.name} syntax error: {e}")
    ok("all .py modules parse")

    import config as C
    for k in ["BASE_MODEL", "JUDGE_MODEL", "ROUNDS", "ACCEPT_THRESHOLD", "EVAL_SET"]:
        if not hasattr(C, k):
            fail(f"config missing {k}")
    ok(f"config OK (base={C.BASE_MODEL})")

    # rubric + feedback compose (this is the 'requirements + examples + criticisms')
    from core.judge import load_rubric
    r = load_rubric()
    if "Requirements" not in r or "criticisms" not in r.lower():
        fail("rubric/feedback did not compose")
    ok("rubric + accumulated criticisms compose")

    # eval-gate logic
    from core.evalgate import keep_round
    assert keep_round(0.8, 0.7) and not keep_round(0.6, 0.7), "keep_round wrong"
    assert keep_round(0.69, 0.7, tolerance=0.02), "tolerance wrong"
    ok("eval gate keep/revert logic correct")

    print("LOCAL checks passed.\n")
    print("Next: add data/eval_heldout.jsonl + data/seed_prompts.jsonl, then")
    print("  hf upload AmongTheCouch23/handicate-code . . --repo-type dataset")
    print("  hf jobs uv run -d --flavor a100-large --timeout 12h --python 3.11 -s HF_TOKEN ./jobs/run_handicate.py")


if __name__ == "__main__":
    main()