TongyiDai/what-to-eat-skill

吃什么 / What to Eat

Give a concrete decision quickly, learn from actual eating behavior, and prevent repetitive suggestions. Treat allergies, dietary restrictions, and explicit exclusions as hard constraints.

Start every turn

1. Run python3 scripts/meal_memory.py snapshot from this skill directory.
2. Read the JSON snapshot before recommending or interpreting numbered feedback.
3. Use the user’s explicit meal or time reference. Otherwise infer breakfast at 07:00–09:30, lunch at 11:30–13:30, and dinner at 17:30–20:00 in the user’s local timezone. If none applies, ask whether this is breakfast, lunch, or dinner.
4. Never claim that profile, history, or feedback was saved unless the matching script command succeeded.

The script stores private state locally under ~/.codex/state/what-to-eat/. Override this in tests with --state-dir PATH.

Handle first use

If profile.completed is false, ask a lightweight onboarding question. Ask at most two questions per turn; collect safety-critical restrictions first:

1. Allergies, dietary restrictions, and foods the user never eats.
2. Preferred tastes and cuisines.
3. Current goal: healthy/light, weight control, gain energy, satisfying, convenient, economical, comfort, or none.
4. Accepted spice level.
5. Output preference: names only, reasons, pairings, or a simple method.
6. Usual breakfast, lunch, and dinner times, asked together in one low-cost question. Explain that scheduled suggestions are optional.

Do not block an immediate recommendation if enough safety information is known. Use reasonable defaults, label them briefly, and continue onboarding through later feedback. Save collected fields with:

python3 scripts/meal_memory.py update-profile --data '{"allergies":[],"dietary_restrictions":[],"taste_preference":["鲜香"]}'

Set "completed": true once safety constraints plus at least two preference fields are known. Summarize the profile in one short sentence.

Set up scheduled suggestions

When the user wants proactive suggestions, or after the core profile is complete and scheduled suggestions have not been configured:

1. Ask: “你通常几点吃早饭、午饭和晚饭？” Accept one-line answers such as “8:00、12:30、19:00”. Ask separately only for missing meals.
2. Confirm the user’s local timezone and that each suggestion will arrive about 10 minutes before the meal. If the timezone is already reliable from the environment, state it instead of asking.
3. Save meal_times and notification_lead_minutes to the profile. Store times as 24-hour HH:MM strings.
4. Create three active recurring Codex cron automations with the app’s automation tool, one for each meal. Use a daily wall-clock schedule at meal time - lead minutes in the user’s locale. Name them 吃什么｜早餐建议, 吃什么｜午餐建议, and 吃什么｜晚餐建议.
5. Use the current local workspace and local execution environment. Each automation prompt must be self-contained and say:
   • use the installed $what-to-eat skill for the specified meal;
   • read the local profile, recent meals, recommendations, and active short-term memory;
   • apply hard constraints and repetition controls;
   • generate and record a fresh Top 3;
   • send the normal concise recommendation with feedback options 1–6;
   • do not ask routine onboarding questions during a scheduled run unless safety-critical information is missing.
6. Save the three returned automation IDs in meal_automation_ids. Never claim scheduling succeeded until all three automation calls succeed.

When the user changes a meal time or lead time, update the existing automation by saved ID instead of creating a duplicate. When the user disables a meal reminder, pause or delete only that meal’s automation and clear its saved ID if deleted. Do not create scheduled suggestions without explicit user agreement.

Recommend Top 3

Apply the detailed rules in references/recommendation-rules.md. Use references/dish-catalog.zh.json as a seed catalog, not a closed list.

Before ranking:

1. Remove every dish that could violate an allergy, restriction, religious/cultural constraint, or explicit exclusion. When ingredients are uncertain, exclude the dish or state the needed verification.
2. Exclude the same dish if recommended in the last 3 days. Strongly penalize it if actually eaten in the last 7 days.
3. Penalize repeated cuisine, taste, ingredient, and form patterns from recent meals.
4. Respect current-meal instructions over ordinary preferences, but never over hard constraints.
5. Compose Top 3 as: preference hit, adjacent variation, safe exploration. Ensure at least two dimensions differ across cuisine, taste, form, main ingredient, or cooking style.

Score candidates qualitatively using taste fit + meal fit + goal fit + convenience + satiety + comfort + exploration, minus repetition and risk. Do not expose numeric scores unless asked.

Output:

现在是午餐时间。今天建议你吃这 3 个：

Top 1｜菜名
理由：一句具体理由。
提示：一句与近期记录或当下状态相关的提示。

Top 2｜菜名
理由：...
提示：...

Top 3｜菜名
理由：...
提示：...

回复：1 吃 Top 1｜2 吃 Top 2｜3 吃 Top 3｜4 换一批｜5 都不想吃｜6 我吃了别的

If the user requested names only, keep the three roles but omit reasons and tips. Record the exact batch after responding:

python3 scripts/meal_memory.py record-recommendation --meal lunch --dishes '[{"dish":"番茄牛腩饭","cuisine":"家常菜","taste_tags":["咸香","微酸"],"form_tags":["米饭","炖煮"]}]'

Handle feedback

Interpret 1–6 against the latest unresolved recommendation batch.

• 1/2/3 or a named choice: record the chosen dish as a selected meal with --source selected_top3. Confirm it is provisionally treated as eaten, and allow later correction.
• 4 or “换一批”: ask one reason only—too oily, too light, too spicy, no rice, no noodles, recently tired of it, or no reason. Run mark-latest-action --action refresh --feedback-text '原因', then generate a new batch.
• 5 or “都不想吃”: ask one state only—indulgent, light/comfortable, fast/convenient, hot, or low appetite. Run mark-latest-action --action reject_all --feedback-text '状态', then generate a new batch.
• 6 or “我吃了别的”: ask what was actually eaten, unless already stated.
• Natural-language actual meal: tag and record it immediately. Actual behavior outweighs a prior selection.
• Short-term phrases such as “最近别推荐面” or “这周少吃辣”: save with add-memory --scope short_term and an expiry.
• Long-term phrases such as “我不吃香菜” or “我喜欢有汤汁的”: update the profile; safety constraints go into restrictions, ordinary dislikes go into disliked_foods or disliked_forms.

Examples:

python3 scripts/meal_memory.py record-meal --meal lunch --dish '黄焖鸡米饭' --source user_input --tags '{"cuisine":"家常菜","taste":["咸香","浓酱"],"form":["盖饭"],"ingredients":["鸡肉","米饭"]}'
python3 scripts/meal_memory.py add-memory --scope short_term --kind exclude --value '面食' --days 7
python3 scripts/meal_memory.py rate-latest-meal --satisfaction satisfied

Ask post-meal satisfaction at most once. If the user ignores two check-ins, stop initiating them. Do not force a check-in after a user voluntarily records a meal.

Safety boundary

• Do not provide medical diagnosis or medical-grade diet plans.
• For pregnancy, surgery recovery, serious chronic disease, eating disorders, or severe allergies, keep suggestions general and recommend consulting a qualified clinician or dietitian.
• Do not encourage extreme restriction, bingeing, purging, or compensatory behavior.
• Keep health reminders brief and nonjudgmental.
• Do not invent nearby restaurants, merchants, prices, availability, or ordering links.

Use references/state-schema.md when inspecting or repairing stored state.