# Running OnCallVal from this dataset

Self-contained: the environment, scorer, generator, and runners are here.

```bash
pip install pyarrow   # only to read test.parquet; the code is stdlib-only

# 1. inspect a task
python - <<'PY'
import json, pyarrow.parquet as pq
rows = pq.read_table("../test.parquet").to_pylist()
task = rows[0]
print(task["task_id"], task["difficulty"], task["alert"])
PY

# 2. wire up an episode by hand
python - <<'PY'
import json, sys, pyarrow.parquet as pq
sys.path.insert(0, ".")
from oncallval.env import Episode
from oncallval.score import score_episode, truth_from_spec

row = pq.read_table("../test.parquet").to_pylist()[0]
task = dict(row); task["spec"] = json.loads(row["spec"])
ep = Episode(task)
print(ep.system_prompt())
ep.step({"action": "list_services"})
ep.step({"action": "submit_report", "cause": "bad_deploy:public-gw", "fix_summary": "x"})
print(score_episode(ep, truth_from_spec(task)))
PY

# 3. export to jsonl (the harness reads jsonl, not parquet), then run any
#    OpenAI-compatible endpoint
python - <<'PY'
import json, pyarrow.parquet as pq
rows = pq.read_table("../test.parquet").to_pylist()
with open("tasks.jsonl", "w", encoding="utf-8") as f:
    for r in rows:
        f.write(json.dumps(r) + "
")
PY
OPENAI_BASE_URL=https://api.openai.com/v1 OPENAI_API_KEY=sk-...   python -m oncallval.harness --tasks tasks.jsonl --agent llm --model <model> --out run.jsonl
```

Notes:
- `truth_from_spec` derives the answer key from the task spec, so there is no
  answers file to leak or lose.
- Never pass `spec` to the agent. It sees `alert` and action observations only.