scripts/job_entry.sh

#!/usr/bin/env bash
# HF Job entrypoint. Runs from the repo root (downloaded via `hf download` into the
# current working dir at launch, to avoid FUSE-mount read races). Runs the CausalGame
# LLM-agent matrix via HF Inference Providers, LLM-judges the reports, and copies all
# outputs to the mounted bucket at /out.
set -uo pipefail

MODELS="${MODELS:-hf-deepseek-v3 hf-qwen3-235b hf-glm-4.6 hf-kimi-k2 hf-gpt-oss-120b hf-llama-3.3-70b}"
SCENARIOS="${SCENARIOS:-antenna_trap antenna_trap_no_selection_bias deployment_zone_trap_categorical weather_noise}"
MODES="${MODES:-legacy}"
REPEATS="${REPEATS:-1}"
TIMEOUT="${TIMEOUT:-900}"
JUDGE_MODEL="${JUDGE_MODEL:-Qwen/Qwen3-235B-A22B-Instruct-2507}"

echo "== CausalGame repro job =="
echo "cwd=$(pwd)"; ls -la
echo "models=$MODELS"
echo "scenarios=$SCENARIOS ; repeats=$REPEATS ; modes=$MODES"

pip install -q -r CausalGame/requirements.txt openai plotly 2>&1 | tail -3 || pip install -q -r CausalGame/requirements.txt openai

mkdir -p outputs
python scripts/run_llm_agents.py \
  --models $MODELS \
  --scenarios $SCENARIOS \
  --modes $MODES \
  --repeats "$REPEATS" \
  --timeout "$TIMEOUT" \
  --out outputs/llm_results.jsonl || echo "runner exited non-zero"

echo "== judging reports =="
python scripts/judge_reports.py --in outputs/llm_results.jsonl \
  --judge-model "$JUDGE_MODEL" \
  --out outputs/llm_results_judged.jsonl || echo "judge failed"

echo "== copying outputs to bucket /out =="
mkdir -p /out
cp -v outputs/llm_results*.jsonl /out/ 2>/dev/null || true
cp -v outputs/llm_summary.csv /out/ 2>/dev/null || true
cp -v outputs/.sess_*.log /out/ 2>/dev/null || true
echo "== done =="
ls -la /out || true