JackAILab's picture
Add files using upload-large-folder tool
6aa6c12 verified
Raw History Blame Contribute Delete
5.69 kB
#!/usr/bin/env python3
"""Copy videos from a closed runner result, preserving originals and score context."""
from __future__ import annotations
import argparse
import hashlib
import json
import re
import shutil
from datetime import datetime, timezone
from pathlib import Path
def component(value: str) -> str:
return re.sub(r"[^A-Za-z0-9_.-]", "_", value).strip(".") or "episode"
def resolve_video(run: Path, episode: dict, source: str) -> Path:
"""Absolute artifacts or bound-worker paths relative to that episode only."""
path = Path(source)
if not path.is_absolute():
matches = []
for result_path in (run / "episodes").glob("*/result.json"):
if not result_path.resolve().is_relative_to(run):
raise ValueError("episode result escaped the closed run")
recorded = json.loads(result_path.read_text())
if recorded["episode"]["id"] == episode["episode"]["id"]:
if recorded != episode:
raise ValueError("episode index and result disagree")
matches.append(result_path.parent / "artifacts")
if len(matches) != 1:
raise ValueError("relative video requires one matching closed episode result")
base = matches[0].resolve(strict=True)
path = (base / path).resolve(strict=True)
if not base.is_relative_to(run) or not path.is_relative_to(base):
raise ValueError("relative worker video escaped its bound artifact directory")
path = path.resolve(strict=True)
if not path.is_relative_to(run) or not path.is_file() or path.suffix.lower() != ".mp4":
raise ValueError("video artifact must be an MP4 file inside the closed run")
return path
def main() -> int:
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("--run-dir", type=Path, required=True)
parser.add_argument("--output-dir", type=Path, required=True, help="new archive directory; never overwritten")
args = parser.parse_args()
run = args.run_dir.resolve(strict=True)
status = json.loads((run / "status.json").read_text())
if status.get("status") not in ("completed", "completed_with_errors"):
parser.error("only closed runs may be exported; do not inspect or publish an active attempt")
manifest = json.loads((run / "manifest.json").read_text())
episodes = [json.loads(line) for line in (run / "episodes.jsonl").read_text().splitlines() if line]
if not episodes:
parser.error("closed run contains no episode records")
names = [component(item["episode"]["id"]) for item in episodes]
if len(names) != len(set(names)):
parser.error("episode archive paths would collide")
# Resolve all sources before creating output. Artifact references must not
# copy arbitrary files outside this completed run.
sources = []
destinations = set()
for index, episode in enumerate(episodes):
for key, source in episode.get("artifacts", {}).items():
worker_video = key.startswith("video_") and key not in ("video_manifest", "video_validation", "video_errors")
if key == "video" or key.endswith("_video") or worker_video:
destination_key = (index, component(key))
if destination_key in destinations:
parser.error("video artifact archive paths would collide")
destinations.add(destination_key)
try:
path = resolve_video(run, episode, source)
except (OSError, ValueError) as error:
parser.error(f"invalid video artifact for {episode['episode']['id']}: {key}: {error}")
sources.append((index, key, path))
output = args.output_dir.resolve()
output.mkdir(parents=True, exist_ok=False)
report = {
"schema_version": "starvla-rsi.video-archive/v1",
"exported_at": datetime.now(timezone.utc).isoformat(),
"source_run": str(run), "run_id": manifest["run_id"],
"benchmark": manifest["benchmark"], "protocol": manifest["protocol"],
"policy": manifest["policy"], "episodes": [],
}
for index, episode in enumerate(episodes):
target = output if len(episodes) == 1 else output / names[index]
target.mkdir(exist_ok=True)
videos = []
for source_index, key, source in sources:
if source_index != index:
continue
destination = target / (component(key) + ".mp4")
shutil.copy2(source, destination)
source_digest = hashlib.sha256(source.read_bytes()).hexdigest()
if hashlib.sha256(destination.read_bytes()).hexdigest() != source_digest:
raise RuntimeError("copied video digest differs from its source")
videos.append({"path": str(destination.relative_to(output)), "sha256": source_digest,
"source": str(source), "bytes": destination.stat().st_size})
report["episodes"].append({
"episode": episode["episode"], "status": episode["status"],
"success": episode["success"], "metrics": episode["metrics"],
"termination_reason": episode["termination_reason"], "error": episode["error"],
"environment_steps": episode["environment_steps"], "policy_calls": episode["policy_calls"],
"usage": episode["usage"], "videos": videos,
"video_status": "available" if videos else "not_produced",
})
(output / "index.json").write_text(json.dumps(report, indent=2, sort_keys=True) + "\n")
print(output)
return 0
if __name__ == "__main__":
raise SystemExit(main())