source/scripts/replay.py

"""Replay admission-style episodes against a running server over the OpenEnv WebSocket.

For every task: the example actions end with done=true and a reward in [0, 1]; the reference
answer scores 1; a wrong answer scores below 1; the python tool cannot read the task sources;
and an episode that never submits ends at the action budget with reward 0.
"""

import argparse
import json
import sys
import time
from pathlib import Path
from urllib.request import urlopen

from websockets.sync.client import connect

sys.path.insert(0, str(Path(__file__).resolve().parents[1]))

from workbench.models import MAX_STEPS  # noqa: E402
from workbench.tasks import ALL_TASK_IDS, build_task  # noqa: E402

# every family at levels 1 and 2, plus the Repo Repair levels that get submitted
TASK_IDS = ALL_TASK_IDS + [f"{t.split('@')[0]}@{lv}" for t in ALL_TASK_IDS
                           if t.startswith("swe-repo-") and t.endswith("@1") for lv in (3, 4)]

EXAMPLE_ACTIONS = [{"tool": "ls"}, {"tool": "submit", "answer": ""}]


def exchange(ws, message):
    ws.send(json.dumps(message))
    reply = json.loads(ws.recv(timeout=60))
    assert reply.get("type") == "observation", reply
    data = reply["data"]
    assert isinstance(data["done"], bool) and 0 <= data["reward"] <= 1, data
    return data


def episode(url, task_id, seed, actions):
    with connect(url, open_timeout=15) as ws:
        first = exchange(ws, {"type": "reset", "data": {"task_id": task_id, "seed": seed}})
        assert not first["done"]
        result = first
        for action in actions:
            result = exchange(ws, {"type": "step", "data": action})
        return first, result


def main():
    parser = argparse.ArgumentParser()
    parser.add_argument("--url", default="http://127.0.0.1:8000")
    parser.add_argument("--seeds", type=int, default=3)
    parser.add_argument("--out", default="artifacts/replay.json")
    args = parser.parse_args()
    base = args.url.rstrip("/")
    ws_url = base.replace("http", "ws", 1) + "/ws"
    schema = json.load(urlopen(f"{base}/schema", timeout=15))
    Path(args.out).parent.mkdir(parents=True, exist_ok=True)
    Path(args.out).with_name("schema.json").write_text(
        json.dumps({k: schema[k] for k in ("action", "observation")}, indent=2) + "\n")
    report, started = [], time.monotonic()
    for task_id in TASK_IDS:
        for seed in range(args.seeds):
            task = build_task(task_id, seed)
            first, ex = episode(ws_url, task_id, seed, EXAMPLE_ACTIONS)
            assert ex["done"] and ex["reward"] == 0.0, (task_id, ex)
            if task.meta.get("grades_workspace"):  # Repo Repair: write the fixed files, then submit
                fixes = [{"tool": "write", "path": m["path"], "content": task.meta["pristine"][m["path"]]}
                         for m in task.meta["mutants"]]
                _, good = episode(ws_url, task_id, seed, fixes + [{"tool": "submit", "answer": ""}])
                assert good["done"] and good["reward"] == 1.0, (task_id, seed, good)
                bad = ex  # untouched repository already scored 0 above
            else:
                assert task.oracle not in json.dumps(first) or len(task.oracle) < 8, task_id
                _, good = episode(ws_url, task_id, seed, [{"tool": "submit", "answer": task.oracle}])
                assert good["done"] and good["reward"] == 1.0, (task_id, seed, good)
                _, bad = episode(ws_url, task_id, seed, [{"tool": "submit", "answer": task.wrong[0]}])
                assert bad["done"] and bad["reward"] < 1.0, (task_id, seed, bad)
            report.append({"task_id": task_id, "seed": seed, "example": ex["reward"],
                           "oracle": good["reward"], "wrong": bad["reward"],
                           "reset_chars": len(first["observation"]["output"])})
    probe = "import glob\ntry:\n    open('/app/workbench/grading.py')\n    print('READABLE')\nexcept OSError as e:\n    print('blocked', type(e).__name__)\n"
    _, sandbox = episode(ws_url, TASK_IDS[0], 0, [{"tool": "python", "code": probe}])
    _, budget = episode(ws_url, TASK_IDS[0], 0, [{"tool": "ls"}] * MAX_STEPS)
    assert budget["done"] and budget["reward"] == 0.0, budget
    summary = {"passed": True, "episodes": len(report) * 3 + 2, "tasks": len(TASK_IDS),
               "sandbox_probe": sandbox["observation"]["output"].strip(),
               "max_reset_chars": max(r["reset_chars"] for r in report),
               "duration_s": round(time.monotonic() - started, 1), "results": report}
    Path(args.out).write_text(json.dumps(summary, indent=2) + "\n")
    print(json.dumps({k: v for k, v in summary.items() if k != "results"}))


if __name__ == "__main__":
    main()