run_10_real_episodes.py

"""Run 10 real SupportDesk episodes and stream inference logs.

This does not fabricate output. It simply invokes `inference.py` repeatedly
with real task ids from the registry so stdout contains valid `[START]`,
`[STEP]`, and `[END]` lines for each completed run.
"""

from __future__ import annotations

import os
import subprocess
import sys
from itertools import cycle, islice

from tasks import list_task_ids

TOTAL_RUNS = 10


def main() -> int:
    base_env = os.environ.copy()
    task_ids = list(list_task_ids())

    if not task_ids:
        print("No tasks registered.", file=sys.stderr)
        return 1

    # Repeat the real task ids until we have 10 actual runs.
    run_plan = list(islice(cycle(task_ids), TOTAL_RUNS))

    for idx, task_id in enumerate(run_plan, start=1):
        env = base_env.copy()
        env["SUPPORTDESK_TASK_ID"] = task_id
        env.setdefault("PYTHONUTF8", "1")

        print(
            f"# run {idx}/{TOTAL_RUNS} task={task_id}",
            file=sys.stderr,
            flush=True,
        )

        completed = subprocess.run(
            [sys.executable, "inference.py"],
            env=env,
            check=False,
        )
        if completed.returncode != 0:
            print(
                f"Run {idx} failed for task {task_id} with exit code {completed.returncode}.",
                file=sys.stderr,
            )
            return completed.returncode

    return 0


if __name__ == "__main__":
    raise SystemExit(main())