dotfiles-public/private_dot_config/nvim/lua/plugins/minuet.lua

-- Inline LLM (ghost-text) completion tuned for writing markdown/prose.
-- Reuses ZAI_API_KEY from the environment (same key the :AI command uses).
-- Suggestions are virtual text, auto-triggered behind a debounce delay.
local PROSE_FT = {
  markdown = true,
  ["md-render"] = true,
  text = true,
  rst = true,
  tex = true,
  gitcommit = true,
  mail = true,
}

return {
  "milanglacier/minuet-ai.nvim",
  dependencies = { "nvim-lua/plenary.nvim" },
  -- Load eagerly: minuet only sets its buffer-local auto-trigger flag from a
  -- FileType autocmd registered in setup(). Lazy-loading on InsertEnter would
  -- register that autocmd *after* the current buffer's FileType already fired,
  -- so virtualtext would never auto-trigger on the buffer you're editing.
  lazy = false,
  opts = {
    provider = "openai_compatible",
    -- One natural continuation, not three alternatives — prose has a single
    -- "next sentence", and fewer candidates means less to read past.
    n_completions = 1,
    -- Give the model more of the text *before* the cursor (default 0.75).
    context_ratio = 0.85,
    -- Cap how much surrounding text is sent. Smaller = lower, steadier latency
    -- (the full default of 16000 chars is overkill for prose continuation and
    -- inflates time-to-first-token). 4000 chars ≈ plenty of preceding context.
    context_window = 4000,
    -- Delay knobs: wait for an idle pause, and rate-limit requests. Groq is fast
    -- and deterministic, so we can afford a snappier debounce than the cloud
    -- defaults; throttle stays moderate to respect the free-tier rate limit.
    throttle = 800, -- min ms between requests
    debounce = 400, -- ms of idle typing before a request fires
    provider_options = {
      openai_compatible = {
        api_key = "GROQ_API_KEY", -- env var NAME; minuet reads it at runtime
        name = "Groq",
        -- Groq's LPU inference gives ~80-100ms time-to-first-token with very low
        -- variance — that consistency is the whole reason for moving off Z.AI's
        -- hit-or-miss coding endpoint. OpenAI-compatible, so the schema is unchanged.
        end_point = "https://api.groq.com/openai/v1/chat/completions",
        -- llama-3.3-70b is Groq's flagship general model: strong prose, still fast
        -- on their hardware. If you ever want the absolute lowest latency, swap to
        -- "llama-3.1-8b-instant" (faster, slightly weaker prose).
        model = "llama-3.3-70b-versatile",
        -- Swap minuet's "code completion engine" system prompt for a prose one
        -- in writing filetypes; keep the default (string) for code. Guidelines
        -- and few-shots are left at defaults so minuet's <endCompletion> parsing
        -- still works.
        system = {
          prompt = function()
            if PROSE_FT[vim.bo.filetype] then
              return [[
You are a prose writing assistant embedded in a text editor. Continue the
author's text at the <cursorPosition> marker naturally and fluently, matching
their voice, tone, and sentence rhythm.

- Output ONLY the continuation text — no preamble, no explanation.
- Never wrap the output in markdown code fences or backticks.
- Do not add headings, bullets, or numbering unless the surrounding text
  already uses them.
- Keep it to at most 1-3 sentences and stop at a natural boundary.]]
            end
            return require("minuet.config").default_system_prefix_first.prompt
          end,
        },
        optional = {
          -- Short, focused completions. (No `thinking` field here — that was a
          -- Z.AI-specific param; Groq's Llama models 400 on unknown body fields.)
          max_tokens = 96,
          temperature = 0.3,
        },
      },
    },
    virtualtext = {
      auto_trigger_ft = { "*" },
      -- Alt keymaps are kept as a secondary path, but the primary accept key is
      -- <Tab>, wired in config() below (Alt+* doesn't fire in Ghostty without
      -- macos-option-as-alt, and Tab is the natural "accept" key for prose).
      keymap = {
        accept = "<A-y>", -- accept full suggestion
        accept_line = "<A-l>", -- accept one line
        prev = "<A-[>",
        next = "<A-]>",
        dismiss = "<A-e>",
      },
    },
  },
  config = function(_, opts)
    require("minuet").setup(opts)

    -- Smart <Tab>: accept a visible minuet ghost-text suggestion; otherwise fall
    -- back to the exact super-tab behavior we had before (jump an active snippet,
    -- else insert a literal Tab). Blink's menu navigates with <C-n>/<C-p>, not
    -- Tab, so this doesn't fight the completion popup.
    local vt = require("minuet.virtualtext").action
    vim.keymap.set("i", "<Tab>", function()
      if vt.is_visible() then
        vt.accept()
      elseif vim.snippet.active({ direction = 1 }) then
        vim.snippet.jump(1)
      else
        vim.api.nvim_feedkeys(
          vim.api.nvim_replace_termcodes("<Tab>", true, false, true),
          "n",
          false
        )
      end
    end, { desc = "minuet: accept suggestion, else snippet jump / Tab" })
  end,
}