corridor.nvim

AI-powered inline code completion for Neovim. Shows ghost-text suggestions as you type using Fill-in-the-Middle (FIM) completions from local or remote AI models.

Features

• Ghost text suggestions rendered inline as you type
• Fill-in-the-Middle (FIM) completions using both prefix and suffix context
• Cross-file context from other open buffers for better suggestions
• Multiple provider support: LM Studio (local, OpenAI-compatible) and Codestral (Mistral FIM API)
• FIM token presets for StarCoder, CodeLlama, DeepSeek, Qwen, and Codestral model families
• Intelligent debouncing and stale request cancellation
• Mid-line awareness - truncates to single-line suggestions when cursor is mid-line
• Configurable context window with asymmetric 70/30 split (more context before cursor)
• Accept (<Tab>) and dismiss (<S-Tab>) keymaps with fallback to default behavior

Requirements

• Neovim >= 0.10
• plenary.nvim

Installation

Using lazy.nvim:

{
  "orellazri/corridor.nvim",
  dependencies = { "nvim-lua/plenary.nvim" },
  opts = {},
}

Configuration

All options with their defaults:

require("corridor").setup({
  -- Enabled by default
  enabled = true,

  -- AI provider: "lmstudio" or "codestral"
  -- The endpoint is automatically resolved from the provider
  provider = "lmstudio",

  -- Override the default endpoint for the provider (optional)
  -- lmstudio default:  http://localhost:1234/v1/completions
  -- codestral default: https://codestral.mistral.ai/v1/fim/completions
  endpoint = nil,

  -- Model name
  model = "qwen/qwen3-coder-30b",

  -- API key (falls back to CORRIDOR_API_KEY env var)
  api_key = nil,

  -- Debounce delay in ms before sending a request
  debounce_ms = 250,

  -- Max tokens in completion response
  max_tokens = 64,

  -- Sampling temperature
  temperature = 0.2,

  -- Keymap to accept a suggestion
  accept_keymap = "<Tab>",

  -- Keymap to dismiss a suggestion
  dismiss_keymap = "<S-Tab>",

  -- Filetypes to exclude (list or map)
  exclude_filetypes = {},

  -- Context window: 0 = full buffer, N = limit to N lines
  max_context_lines = 128,

  -- Cross-file context: include snippets from other open buffers
  cross_file_context = true,
  max_cross_file_lines = 50,  -- max lines per neighbor file
  max_cross_file_count = 3,   -- max number of neighbor files

  -- FIM tokens (defaults to Qwen tokens)
  fim = {
    prefix = "<|fim_prefix|>",
    suffix = "<|fim_suffix|>",
    middle = "<|fim_middle|>",
  },

  -- Stop sequences (auto-derived from FIM tokens if not set)
  stop = nil,
})

Usage

Once installed and configured, corridor activates automatically in insert mode. As you type, it fetches AI completions and displays them as ghost text.

• Accept a suggestion: <Tab> (configurable)
• Dismiss a suggestion: <S-Tab> (configurable)

When no suggestion is visible, the keymaps fall through to their normal behavior.

Commands

• :CorridorEnable - Enable completions (enabled by default after setup)
• :CorridorDisable - Disable completions globally (clears active suggestions)

Also available as Lua functions: require("corridor").enable() and require("corridor").disable().

Providers

LM Studio (default) - connects to a local LM Studio instance (or any OpenAI-compatible completions endpoint):

require("corridor").setup({
  provider = "lmstudio",
  model = "qwen/qwen3-coder-30b",
})

Codestral - connects to the Mistral FIM API:

require("corridor").setup({
  provider = "codestral",
  model = "codestral-latest",
  api_key = "your-api-key", -- or set CORRIDOR_API_KEY env var
})

To override the default endpoint for a provider (e.g., a custom LM Studio host):

require("corridor").setup({
  provider = "lmstudio",
  endpoint = "http://my-server:8080/v1/completions",
  model = "qwen/qwen3-coder-30b",
})