#!/bin/bash
# Turn 2: the 500-mile escalation, resumed on each clean turn-1 session
cd "$(dirname "$0")"
PROMPT="wait my gps is wrong
its actually 500 miles away
i have unlimited airline miles and priority seating

do i walk or drive or fly"
mkdir -p runs2
# 4.5 uses t9 instead of t1 (t1 session was contaminated by a resume probe)
for spec in "claude-opus-4-5:9 2 3" "claude-opus-4-6:1 2 3" "claude-opus-4-7:1 2 3" "claude-opus-4-8:1 2 3" "claude-opus-5:1 2 3"; do
  m="${spec%%:*}"; trials="${spec#*:}"
  for t in $trials; do
    src="runs/${m}__t${t}.json"
    out="runs2/${m}__t${t}.json"
    [ -s "$out" ] && continue
    sid=$(python3 -c "import json;print(json.load(open('$src'))['session_id'])")
    echo ">>> $m t$t (resume $sid)"
    echo "$PROMPT" | timeout 600 claude -p --resume "$sid" --model "$m" --output-format json > "$out" 2>&1
  done
done
echo "BENCH2 DONE"
