{
  "defaults": {
    "timeout_seconds": 420,
    "allowed_statuses": ["completed"]
  },
  "cases": [
    {
      "id": "diet_eval_01_banana",
      "title": "Banana snack estimate and log",
      "prompt": "Use the dietitian skill. Do not ask a follow-up: use the supplied facts. For user_key eval_01 on 2026-07-14, first set a daily norm of 2000 kcal with protein 120 g, fat 70 g, and carbs 230 g. Then estimate and log a Snack at 10:00 containing one peeled banana with an edible weight of 120 g. Use calorie_tracker for storage. Finish by resending the complete dashboard with the curator tip filled in.",
      "settings": {"chat_reasoning": "model_default", "temperature": "0.2"},
      "enabled_tools": ["calorie_tracker", "tavily_search", "duckduckgo_search"],
      "enabled_skills": ["dietitian", "research"],
      "expect": {"min_tool_results": 2, "min_native_tool_calls": 2, "assistant_not_empty": true, "assistant_contains": ["# 🥑 Daily Diary"]}
    },
    {
      "id": "diet_eval_02_chicken_rice",
      "title": "Chicken and rice estimate and log",
      "prompt": "Use the dietitian skill. Do not ask a follow-up. For user_key eval_02 on 2026-07-14, set a 2100 kcal daily norm with protein 140 g, fat 70 g, and carbs 240 g. Estimate and log Lunch at 12:30: 150 g cooked skinless chicken breast and 180 g cooked white rice, with no added oil or sauce. Use calorie_tracker and finish with the entire dashboard, filling only its curator tip.",
      "settings": {"chat_reasoning": "model_default", "temperature": "0.2"},
      "enabled_tools": ["calorie_tracker", "tavily_search", "duckduckgo_search"],
      "enabled_skills": ["dietitian", "research"],
      "expect": {"min_tool_results": 2, "min_native_tool_calls": 2, "assistant_not_empty": true, "assistant_contains": ["# 🥑 Daily Diary"]}
    },
    {
      "id": "diet_eval_03_eggs_toast",
      "title": "Eggs toast and butter estimate and log",
      "prompt": "Use the dietitian skill. Do not ask a follow-up. For user_key eval_03 on 2026-07-14, set a 1900 kcal norm with protein 110 g, fat 65 g, and carbs 210 g. Estimate and log Breakfast at 08:15 consisting of two large cooked eggs with no added cooking fat, 40 g sourdough toast, and 10 g butter. Use calorie_tracker and resend the full dashboard with the tip filled.",
      "settings": {"chat_reasoning": "model_default", "temperature": "0.2"},
      "enabled_tools": ["calorie_tracker", "tavily_search", "duckduckgo_search"],
      "enabled_skills": ["dietitian", "research"],
      "expect": {"min_tool_results": 2, "min_native_tool_calls": 2, "assistant_not_empty": true, "assistant_contains": ["# 🥑 Daily Diary"]}
    },
    {
      "id": "diet_eval_04_yogurt_bowl",
      "title": "Yogurt bowl estimate and log",
      "prompt": "Use the dietitian skill. Do not ask a follow-up. For user_key eval_04 on 2026-07-14, set a 1800 kcal norm with protein 115 g, fat 55 g, and carbs 205 g. Estimate and log Snack at 16:00 containing 200 g plain nonfat Greek yogurt, 100 g raw blueberries, and 15 g honey. Use calorie_tracker and resend the complete dashboard with the curator tip filled.",
      "settings": {"chat_reasoning": "model_default", "temperature": "0.2"},
      "enabled_tools": ["calorie_tracker", "tavily_search", "duckduckgo_search"],
      "enabled_skills": ["dietitian", "research"],
      "expect": {"min_tool_results": 2, "min_native_tool_calls": 2, "assistant_not_empty": true, "assistant_contains": ["# 🥑 Daily Diary"]}
    },
    {
      "id": "diet_eval_05_salmon_potatoes",
      "title": "Salmon potatoes and oil estimate and log",
      "prompt": "Use the dietitian skill. Do not ask a follow-up. For user_key eval_05 on 2026-07-14, set a 2200 kcal norm with protein 140 g, fat 80 g, and carbs 245 g. Estimate and log Dinner at 19:30: 180 g cooked Atlantic salmon, 250 g roasted potato, and exactly 15 g olive oil absorbed in cooking. Use calorie_tracker and return the entire dashboard with its tip filled.",
      "settings": {"chat_reasoning": "model_default", "temperature": "0.2"},
      "enabled_tools": ["calorie_tracker", "tavily_search", "duckduckgo_search"],
      "enabled_skills": ["dietitian", "research"],
      "expect": {"min_tool_results": 2, "min_native_tool_calls": 2, "assistant_not_empty": true, "assistant_contains": ["# 🥑 Daily Diary"]}
    },
    {
      "id": "diet_eval_06_big_mac_verified",
      "title": "Verified branded meal estimate and log",
      "prompt": "Use the dietitian skill. For user_key eval_06 on 2026-07-14, set a 2100 kcal norm with protein 130 g, fat 70 g, and carbs 235 g. The user asks you to verify before logging one standard McDonald's Big Mac burger in Spain as Lunch at 13:00. Use a search tool and prefer an official McDonald's nutrition source, then log the verified serving with calorie_tracker. Finish by resending the full dashboard with the tip filled and briefly cite the source after it.",
      "settings": {"chat_reasoning": "model_default", "temperature": "0.2"},
      "enabled_tools": ["calorie_tracker", "tavily_search", "duckduckgo_search"],
      "enabled_skills": ["dietitian", "research"],
      "expect": {"min_tool_results": 3, "min_native_tool_calls": 3, "assistant_not_empty": true, "assistant_contains": ["# 🥑 Daily Diary"]}
    },
    {
      "id": "diet_eval_07_dry_pasta_dispute",
      "title": "Disputed dry pasta estimate verification and log",
      "prompt": "Use the dietitian skill. For user_key eval_07 on 2026-07-14, set a 2000 kcal norm with protein 120 g, fat 65 g, and carbs 240 g. The user complains that a prior 500 kcal estimate for 100 g dry plain durum-wheat pasta seems wrong and asks you to verify it. Search a reputable primary or national food-data source, then log 100 g dry pasta as Lunch at 12:00 using the corrected estimate in calorie_tracker. Resend the complete dashboard with the tip filled and cite the source after it.",
      "settings": {"chat_reasoning": "model_default", "temperature": "0.2"},
      "enabled_tools": ["calorie_tracker", "tavily_search", "duckduckgo_search"],
      "enabled_skills": ["dietitian", "research"],
      "expect": {"min_tool_results": 3, "min_native_tool_calls": 3, "assistant_not_empty": true, "assistant_contains": ["# 🥑 Daily Diary"]}
    },
    {
      "id": "diet_eval_08_brisk_walk",
      "title": "Brisk walking burn estimate and log",
      "prompt": "Use the dietitian skill. Do not ask a follow-up. For user_key eval_08 on 2026-07-14, set a 2000 kcal norm with protein 120 g, fat 70 g, and carbs 230 g. Estimate and log a 60-minute brisk level-ground walk at about 5.5 km/h for a 70 kg adult, performed at 18:00. Use a defensible MET-based calculation, then call calorie_tracker and return the full dashboard with the tip filled.",
      "settings": {"chat_reasoning": "model_default", "temperature": "0.2"},
      "enabled_tools": ["calorie_tracker", "tavily_search", "duckduckgo_search"],
      "enabled_skills": ["dietitian", "research"],
      "expect": {"min_tool_results": 2, "min_native_tool_calls": 2, "assistant_not_empty": true, "assistant_contains": ["# 🥑 Daily Diary"]}
    },
    {
      "id": "diet_eval_09_running",
      "title": "Running burn estimate and log",
      "prompt": "Use the dietitian skill. Do not ask a follow-up. For user_key eval_09 on 2026-07-14, set a 2300 kcal norm with protein 140 g, fat 75 g, and carbs 270 g. Estimate and log running at 8 km/h for 30 minutes for a 70 kg adult at 07:00 using a defensible MET calculation. Use calorie_tracker and resend the entire dashboard with the curator tip filled.",
      "settings": {"chat_reasoning": "model_default", "temperature": "0.2"},
      "enabled_tools": ["calorie_tracker", "tavily_search", "duckduckgo_search"],
      "enabled_skills": ["dietitian", "research"],
      "expect": {"min_tool_results": 2, "min_native_tool_calls": 2, "assistant_not_empty": true, "assistant_contains": ["# 🥑 Daily Diary"]}
    },
    {
      "id": "diet_eval_10_cycling",
      "title": "Moderate cycling burn estimate and log",
      "prompt": "Use the dietitian skill. Do not ask a follow-up. For user_key eval_10 on 2026-07-14, set a 2400 kcal norm with protein 150 g, fat 80 g, and carbs 275 g. Estimate and log 45 minutes of moderate outdoor cycling at about 16 to 19 km/h for an 80 kg adult at 17:30 using a defensible MET calculation. Use calorie_tracker and return the full dashboard with the tip filled.",
      "settings": {"chat_reasoning": "model_default", "temperature": "0.2"},
      "enabled_tools": ["calorie_tracker", "tavily_search", "duckduckgo_search"],
      "enabled_skills": ["dietitian", "research"],
      "expect": {"min_tool_results": 2, "min_native_tool_calls": 2, "assistant_not_empty": true, "assistant_contains": ["# 🥑 Daily Diary"]}
    }
  ]
}