K2 Study Decks & Algorithm

The catalog of K2 study decks: eight curated datasets spanning Japanese kanji and vocabulary, HSK Chinese, Korean Hangul, Spanish, French, classical Latin and medical Latin. Returns each deck id (the value other tools take), what it teaches, how many entries it has, and what its fields mean. Call this before search_deck.

{
  "type": "object",
  "properties": {}
}

Search one deck. `query` substring-matches the term, its reading/pronunciation, its English meanings, and its category, so "bone" finds entries glossed as bone and "ㄱ" finds that jamo. Narrow with `category` (exact, from list_decks) or `difficulty` (1-5; for the kanji deck this is the school grade). Paginated via `offset`.

{
  "type": "object",
  "properties": {
    "deck": {
      "type": "string",
      "description": "Deck id from list_decks",
      "enum": [
        "japanese-kanji",
        "japanese-vocabulary",
        "chinese",
        "hangul",
        "spanish",
        "french",
        "latin",
        "medica"
      ]
    },
    "query": {
      "type": "string",
      "description": "Substring to match against term, reading, meanings, category"
    },
    "category": {
      "type": "string",
      "description": "Exact category filter (see list_decks)"
    },
    "difficulty": {
      "type": "number",
      "description": "Exact difficulty/grade filter (1-5; kanji deck = grade 1-6)"
    },
    "limit": {
      "type": "number",
      "description": "Max results (default 20, max 50)"
    },
    "offset": {
      "type": "number",
      "description": "Skip this many matches (paging)"
    }
  },
  "required": [
    "deck"
  ]
}

One deck entry in full, matched on the exact term (the kanji, hanzi, jamo, or word itself). Case-insensitive. Returns every field the deck carries, including the deck-specific extras: stroke counts and JLPT level for kanji, component kanji for compounds, declension and English derivatives for Latin, IPA and pronunciation audio for French.

{
  "type": "object",
  "properties": {
    "deck": {
      "type": "string",
      "description": "Deck id from list_decks",
      "enum": [
        "japanese-kanji",
        "japanese-vocabulary",
        "chinese",
        "hangul",
        "spanish",
        "french",
        "latin",
        "medica"
      ]
    },
    "term": {
      "type": "string",
      "description": "Exact term to fetch"
    }
  },
  "required": [
    "deck",
    "term"
  ]
}

Rubik's cube last-layer algorithm sets used by the K2 cube trainer: full OLL (57 cases), full PLL (21), the two-look beginner subsets, and full ZBLL. Each case returns its primary algorithm plus alternates, with sources and finger-trick notes where recorded. Every algorithm in these sets is verified against a cube simulator by the repo's audit script (app/cube/scripts/audit-algorithms.mjs), run before data changes land. Filter by `case_id` or `group`; omit both for the set's structure and levels.

{
  "type": "object",
  "properties": {
    "set": {
      "type": "string",
      "description": "Algorithm set id",
      "enum": [
        "oll",
        "pll",
        "two-look-oll",
        "two-look-pll",
        "zbll"
      ]
    },
    "case_id": {
      "type": "string",
      "description": "Single case id, e.g. 'oll-21' or 'pll-t'"
    },
    "group": {
      "type": "string",
      "description": "Case group filter, e.g. 'T Shapes', 'OCLL'"
    },
    "limit": {
      "type": "number",
      "description": "Max cases (default 20, max 50)"
    }
  },
  "required": [
    "set"
  ]
}

The K2 spaced-repetition engine, described. K2 stores no review dates: each item holds a probability p that halves on a correct answer and doubles on a wrong one, and items are drawn by Monte Carlo rejection sampling, so the pool self-organizes around current weaknesses and expands when it cools. Returns the update rule (including the deliberate hinted-answer exception), the selection and pool-expansion mechanics, the live tuning constants, and notes for porting it. Read this before simulate_k2_session.

{
  "type": "object",
  "properties": {}
}

Run the real K2 engine over your own items. Give it a list of terms (any strings — your flashcards, cube cases, anything), a study mode, and a sequence of answers, and it returns the p-value trajectory per answer, which items are now mastered, when the pool expanded, and the probability distribution over what K2 would serve next. Use this to drive your own drilling with K2 as the scheduler instead of reimplementing it.

{
  "type": "object",
  "properties": {
    "terms": {
      "type": "array",
      "items": {
        "type": "string"
      },
      "description": "The item pool. Any strings; the engine only needs identity."
    },
    "mode": {
      "type": "string",
      "enum": [
        "meaning",
        "reading",
        "writing",
        "listening"
      ],
      "description": "Which dimension is being drilled (each carries its own p)."
    },
    "steps": {
      "type": "array",
      "description": "Answers in order. `outcome` is correct | incorrect | hinted (hinted = answered with the crutch visible; cools more weakly).",
      "items": {
        "type": "object",
        "properties": {
          "term": {
            "type": "string"
          },
          "outcome": {
            "type": "string",
            "enum": [
              "correct",
              "incorrect",
              "hinted"
            ]
          }
        },
        "required": [
          "term",
          "outcome"
        ]
      }
    },
    "starting_level": {
      "type": "number",
      "description": "Treat the first N terms as already known (pre-cooled), as the app's head-start option does. Default 0."
    }
  },
  "required": [
    "terms",
    "steps"
  ]
}

What K2 is, which apps it runs, what this server does and does not expose, and every machine-readable surface. Also returns the full capability map — every tool, prompt, and resource — so a client with a stale or tools-only roster can still see the whole surface.

{
  "type": "object",
  "properties": {}
}