Function main

tools/provenance-eval/multi_session.py:196–227

()

def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--binary", action="append", required=True, help="mode=/absolute/path")
    ap.add_argument("--output", required=True)
    ap.add_argument("--workers", default="1,4,8")
    ap.add_argument("--repetitions", type=int, default=5)
    ap.add_argument("--scenarios", default="balanced,read-heavy")
    ap.add_argument("--events", type=int, default=16)
    ap.add_argument("--turns", type=int, default=2)
    ap.add_argument("--serialize-stop", action="store_true")
    args = ap.parse_args()
    binaries = dict(x.split("=", 1) for x in args.binary)
    configs = {"balanced": {"reads_per_event": 1}, "read-heavy": {"reads_per_event": 4},
               "paced": {"reads_per_event": 1, "think_ms": 100}}
    result = {"kind": "deterministic real-CLI multi-session replay; no model calls", "platform": platform.platform(),
              "binaries": {k: {"path": str(pathlib.Path(v).resolve()), "sha256": hashlib.sha256(pathlib.Path(v).read_bytes()).hexdigest()} for k, v in binaries.items()}, "runs": []}
    # Reuse stopped, fully recorded fixture templates. Setup is never timed.
    # TemporaryDirectory cleanup also removes copied local owner metadata.
    with tempfile.TemporaryDirectory(prefix="atomic-multi-templates-") as templates:
        roots = {}
        for mode, binary in binaries.items():
            root = pathlib.Path(templates) / mode
            root.mkdir()
            fixture = SeededFixture(pathlib.Path(binary).resolve(), root)
            try:
                fixture.start()
                fixture.hook("user-prompt-submit", {"session_id": base.SESSION, "prompt": "Prepare fixed benchmark context"})
                fixture.hook("stop", {"session_id": base.SESSION, "reason": "end_turn", "response": "Context prepared"})
            finally:
                fixture.close()
            roots[mode] = root
        execute(args, binaries, configs, result, roots)


def execute(args, binaries, configs, result, roots):