File size: 3,174 Bytes
eeaa10a | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 | """Evaluate one actor on one case, n times.
python scripts/run_case.py --case cases/001-coral-lcm-porites-01/case.yaml --n-runs 1 \
--actor-agent claude-code --actor-model claude-opus-5-5 \
--env-model claude-sonnet-5-5 --judge-model claude-opus-5-5 \
--actor-timeout-s 10800 --checkpoints case --out runs
--actor-agent codex runs Codex with the provider of machine.yaml (e.g. --actor-model deepseek-chat).
--checkpoints: case = the case's own setting; on/off override it (2x2 design).
Output: <out>/<case_id>/<timestamp>/run_<i>/ and summary.json. A run that stops on an error is resumable with
scripts/resume_run.py; the other runs continue.
"""
import argparse
import sys
import time
from pathlib import Path
sys.path.insert(0, str(Path(__file__).resolve().parents[1])) # run from a checkout without installing
from twinlab.actors import AGENTS # noqa: E402
from twinlab.case import CHECKPOINT_MODES, Case # noqa: E402
from twinlab.run import one_run, run_config, summarize # noqa: E402
from twinlab.settings import ROOT # noqa: E402
ACTOR_EFFORTS = ["low", "medium", "high", "xhigh", "max"] # claude --effort levels (CLI 2.1.296 --help)
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--case", required=True, type=Path)
ap.add_argument("--n-runs", required=True, type=int)
ap.add_argument("--actor-agent", required=True, choices=sorted(AGENTS))
ap.add_argument("--actor-model", required=True)
ap.add_argument("--actor-effort", choices=ACTOR_EFFORTS,
help="claude-code only: passed as --effort; omitted = the CLI's default effort")
ap.add_argument("--env-model", required=True)
ap.add_argument("--judge-model", required=True)
ap.add_argument("--actor-timeout-s", required=True, type=int)
ap.add_argument("--checkpoints", required=True, choices=CHECKPOINT_MODES)
ap.add_argument("--out", required=True, type=Path)
a = ap.parse_args()
case = Case(a.case, a.checkpoints)
root = (a.out / case.id / time.strftime("%Y%m%d-%H%M%S")).resolve()
# the environment and the judge get their own /tmp (twinlab.claude_cli.private_tmp_prefix), so the run directory
# must not be under the host /tmp
assert not root.is_relative_to("/tmp"), f"--out {a.out} is under /tmp; put the runs elsewhere"
assert not ROOT.is_relative_to("/tmp"), f"the repository {ROOT} is under /tmp; move it elsewhere"
print(f"[run_case] {root}", flush=True)
failed = []
for i in range(a.n_runs):
cfg = run_config(case, i, a.actor_agent, a.actor_model, a.env_model, a.judge_model, a.actor_timeout_s,
a.checkpoints, a.actor_effort)
try:
one_run(case, root / f"run_{i}", cfg)
except Exception as e: # keep going with the other runs; lab state is saved after every call -> resumable
print(f"[run] run_{i} STOPPED: {str(e).splitlines()[0]}", flush=True)
failed.append(f"run_{i}")
summarize(root)
if failed:
print(f"[run] resumable runs: {failed}\n python scripts/resume_run.py --run-dir {root}/<run>", flush=True)
if __name__ == "__main__":
main()
|