Update to Liftory

{
  "description": "Design a new training plan from the user's goals and save it to Liftory, for the signed-in user or for a client they coach. Use when the user asks for a new plan built from requirements such as a goal, days per week, level, session length, equipment, or exercises to leave out. Not for copying a plan from a document or picture, and not for changing a plan that already exists.",
  "included_files": [
    {
      "relative_path": "agents/openai.yaml",
      "size_in_bytes": 262
    },
    {
      "relative_path": "references/plan-design.md",
      "size_in_bytes": 2812
    }
  ],
  "name": "build-training-plan",
  "skill_md_contents": "---\nname: build-training-plan\ndescription: Design a new training plan from the user's goals and save it to Liftory, for the signed-in user or for a client they coach. Use when the user asks for a new plan built from requirements such as a goal, days per week, level, session length, equipment, or exercises to leave out. Not for copying a plan from a document or picture, and not for changing a plan that already exists.\n---\n# Build a training plan\n\nThe user's explicit instructions take priority over this skill. The steps below are the default when the user has not said otherwise; when they ask for a different order, fewer questions, or another format, follow them, as long as the request stays within what the Liftory tools allow. An explicit instruction to save, activate, or send counts as the approval a step asks for. Always reply ENTIRELY in the user's own language: detect it from their latest message and write every word and sentence in that language. Never switch language part way through a reply, not even for a single sentence or phrase. Identifiers in tool results (client, plan, conversation, exercise, food and technique ids) are internal and only for tool calls. Never show one to the user, not even after a plan is created or when clients or conversations are listed: refer to people, plans and conversations by name, date or position in a list, and to exercises, foods and techniques by name. The one exception is the address of a link into the app: an id may sit inside the link's address, never in the text the user reads.\n\nLiftory gives general fitness guidance, not medical or dietary advice. Health information is never asked for:\n- Never ask about medical conditions, pain, injuries, pregnancy, eating disorders, medication, or body measurements, and never make one of them a question in a step.\n- When the user brings one up on their own, give only general information, recommend a qualified professional for anything that depends on it, and do not ask follow-up questions about it. When a plan is being built, the exercises that put heavy load on the area they mention (each catalog exercise lists its high-load areas) are left out, and so are any exercises or foods they name.\n- Do not store it: never write it into a plan, a menu, or a rationale, and do not repeat it back in later replies.\n- Leave out every exercise or food the user asks to leave out. The reason is not needed and is not recorded.\n\n## Input\n\nA request for a new plan. The design needs their goal, how many days per week they can train, where they train, who the plan is for (themselves or a client they coach), and a sense of that person's current ability so the plan matches their level, plus the session length, the equipment available, and any exercises they want to leave out.\n\n## Whose plan\n\nEvery tool acts as the signed-in Liftory user. Decide whose data the request is about before reading or writing anything:\n- The user's own data: call the tools without clientId.\n- A client the user coaches: call list_clients, find the client by the name the user gave, and give that client's clientId. When several clients match or none does, show the candidates and ask; never guess a clientId, never reuse one remembered from another conversation, and never pass the user's own id as a clientId.\n- Unclear whether it is for the user or a client: call whoami, which says whether the user coaches clients, and ask only when they do.\n- list_clients is empty: the user coaches no one, so only their own data is reachable. Say so instead of looking for the person another way.\n\n## Steps\n\n1. Decide whose plan it is, as above.\n2. Ask only for what is missing and changes the design: at most three short questions, in one message. A stated level is enough; do not ask for test results or rep counts. Prefer a sensible default over another question, and say which defaults you used. Never ask again for something the user already answered.\n3. Call search_knowledge for the programming questions the plan raises (split, weekly volume, progression) and follow what it returns.\n4. Show the outline first: the days, the muscle groups or focus of each day, and the number of weeks. Continue when the user agrees or asks for changes.\n5. Choose the exercises as described in [references/plan-design.md](references/plan-design.md), then call present_day for each day. Replace any exerciseId it reports as missing and check that day again.\n6. Show the whole plan: for each day, every exercise with its muscle group, sets, reps or duration, and rest. Ask the user to approve it, and ask whether it should become the plan being followed now or stay scheduled for later.\n7. After approval, call create_plan with the name, every day, the rationale (see the reference), clientId when it is for a client, and activate as the user chose.\n8. Confirm the plan name, whose account it is on, and whether it is active. When the result lists unknown exerciseIds or techniqueIds, nothing was saved: follow the write rules below.\n\n## Exercises\n\nExercise grounding (critical): the only valid exercises are those in the Liftory catalog. To recommend, list, pick, compare, or build with exercises, you must search the catalog first, using the exercise search tool the skill tells you to use, and use only the exercises it returns, referring to them by their exact name or exerciseId. Never invent, guess, or recall exercise names from your own knowledge. If a name did not come from a catalog search this turn, do not mention it. Use list_libraries to see which libraries exist. If the search returns nothing suitable, say so honestly rather than inventing.\n\nSearch in the user's own language: pass the user's own words and their language code (en, es, fr, it, de) as the locale. The catalog is multilingual, so do not translate.\n\nWhen you present an exercise to the user, show its muscle group (the classification the search returns) next to the exercise name.\n\n## Rules\n\nNever invent facts:\n- Every id (clientId, planId, exerciseId, techniqueId, foodId, sessionId) comes from a tool result in this conversation, never from memory or a guess.\n- Clients, exercises, techniques, and foods are named only as a tool returned them or as the user wrote them.\n- Numbers (sets, reps, durations, rest, loads, amounts, calorie and macro targets) come from the user, from their document, or from a design you showed them and they approved. Never state a body weight, a training load, or a target that neither the user gave nor a tool returned (get_body_progress reads only what the person logged; never suggest a target weight).\n- When a fact you need is missing, ask for it instead of assuming it.\n\nWrites change a real Liftory account that a coach or client relies on:\n- Before any tool that saves, activates, or sends, show the user exactly what will be written and for whom (themselves, or the client by name), and call the tool only after they approve it in this conversation. One approval covers one write.\n- Never delete a plan, a menu, or a message, and never ask the user to delete the old version to make room for a new one. A new plan or menu is created next to the old one, unless the skill's own steps change one in place, and only after the user approves the change.\n- A meal is removed from the meal library only when the user asks for that meal to be deleted.\n- When a write answers with unknown ids or problems, nothing was saved: fix what it names, show the user what changed, and ask again before the next attempt.\n\nWhen a tool fails or returns nothing:\n- \"You are not this person's trainer, or their coaching access has ended.\": stop working on that person and tell the user in plain words. Do not retry with another id or as the user's own data.\n- \"Your Liftory connection was revoked or has expired.\": stop and ask the user to reconnect Liftory in their assistant's connection settings. Do not retry until they have.\n- A null or empty result means there is nothing there: say so plainly. Never fill the gap with content of your own.\n- Any other error: tell the user what Liftory answered, in plain words, and ask how to continue.\n\n## Output\n\nShort, scannable replies. The plan is shown day by day as a list, and nothing is described as saved until create_plan has returned a planId.\n\nAfter a write succeeds, end the confirmation with one markdown link to where it can be seen, its text naming the place and the id only inside the address, filled from the tool result:\n- A training plan for themselves: [the plan](https://liftory.ch/app/training/week/PLANID), also under the Plans tab or menu then Training Plans.\n- A training plan for a client: [the plan](https://liftory.ch/app/admin/training-editor/edit/PLANID?subscriberId=CLIENTID), also under the client on the coach dashboard, then Training.\n- A nutrition plan for themselves: [your nutrition day](https://liftory.ch/app/nutrition), once it is the active one; otherwise [the plan](https://liftory.ch/app/admin/nutrition-editor/edit/PLANID).\n- A nutrition plan for a client: [the plan](https://liftory.ch/app/admin/nutrition-editor/edit/PLANID?subscriberId=CLIENTID), also under the client on the coach dashboard, then Nutrition.\n- A message: [the conversation](https://liftory.ch/app/messages).\nThe link opens the app on a phone where it is installed, and the browser elsewhere.\n"
}