nvim-stride :: AI-Powered Next-Edit Predictions

neovim ai lua refactoring plugin
GitHub

nvim-stride

AI-Powered Next-Edit Suggestions for Neovim

nvim-stride is a Neovim plugin that brings intelligent next-edit prediction directly into your workflow. It analyzes recent code changes and predicts where you’ll edit next—rename a variable on line 10, and stride suggests updating line 50. This eliminates tedious manual search-and-replace while maintaining full developer control.

View on GitHub

The Problem

During refactoring sessions, developers make related changes across multiple locations. Traditional find-and-replace is blunt and risky, while manually tracking every related edit is tedious and error-prone. nvim-stride bridges this gap by intelligently predicting related edits based on coding patterns, suggesting changes only where they’re contextually relevant.

Key Features

Intelligent Refactor Mode

Stride’s refactor mode understands the intent behind edits:

  • Next-edit prediction based on recent changes
  • Multi-edit predictions: a rename touching several call sites is predicted in a single request, and Tab steps through the edits with a [n/m] counter
  • Automatic triggering on InsertLeave and normal mode edits
  • Remote suggestions with strikethrough highlighting for replaced text
  • Insert detection for new parameters, properties, or arguments
  • Incremental change tracking via nvim_buf_attach
  • Simple dismissal with Esc or :StrideClear

Ghost Text Completion Mode

Complementary inline completions for faster coding:

  • Real-time ghost text suggestions as you type
  • Suggestion retention: keep typing and characters matching the ghost text are consumed without a new request
  • Partial accepts for the next word (<M-w>) or first line (<M-l>) of a suggestion
  • Focused completions that target the current statement or expression, keeping them fast and non-intrusive
  • Comment-intent completion: type // log the id and get console.log(id)
  • Treesitter-aware context capture for accurate completions
  • Project context: include rules from files like AGENTS.md in prompts
  • Tab-to-accept workflow integration

Developer-Centric UX

Seamless integration with existing workflows:

  • Configurable debounce timing (500ms default)
  • Customizable keybindings for accept, partial accept and dismiss
  • Gutter icon indicator when suggestions are active
  • Global enable/disable commands (:StrideEnable, :StrideDisable)
  • :StrideStatus to inspect the active suggestion and connection state
  • Automatic race condition handling for concurrent edits
  • Bottom-center notifications showing which direction the next suggestion is in

Technical Implementation

nvim-stride leverages several technologies for intelligent prediction:

  • Lua core for Neovim integration and real-time event handling
  • nvim_buf_attach with on_bytes callback for incremental edit tracking
  • Treesitter integration for semantic context awareness
  • Cerebras API for ultra-low latency inference
  • Shared transport layer with real request cancellation, throttling and 429 backoff
  • Word-level diffs and deduplicated context to keep prompts small
  • Token budget management for efficient prompt construction
  • Smart context extraction with configurable line limits
  • Deterministic completion caching

Architecture Highlights

What makes stride’s prediction engine different:

  • Incremental Tracking - Real-time edit capture without blocking the editor, storing buffer state incrementally rather than full snapshots per keystroke
  • Smart Context - Treesitter-powered context expansion for accurate predictions
  • Token Budgeting - Automatic history trimming to stay within model limits
  • Dual Mode Design - Completion and refactor modes can run simultaneously
  • Race Protection - Stale responses discarded if cursor has moved
  • Rate-Limit Aware - Honors Retry-After on 429s and kills in-flight requests on cancel so no tokens are wasted
  • Remote Rendering - Preview changes inline without applying them

User Workflow

Completion Mode

  1. Type - Start coding in insert mode
  2. Pause - After 500ms, ghost text suggestion appears
  3. Accept - Press Tab to accept the suggestion, or <M-w> / <M-l> to accept just the next word or line
  4. Continue - Any other key dismisses and resumes typing

Refactor Mode

  1. Edit - Make a change (e.g., rename a variable)
  2. Exit - Leave insert mode to trigger prediction
  3. Review - Related edits shown with strikethrough and replacement preview
  4. Apply - Press Tab to accept and step to the next predicted edit, or Esc to dismiss

Configuration & Integration

Customization Options

Stride offers extensive customization:

  • API endpoint and model selection
  • Debounce timing and context window size
  • Rate limiting: minimum request interval and cooldown after a 429
  • Project context files such as AGENTS.md
  • Mode selection (completion, refactor, or both)
  • Filetype exclusions for specific workflows
  • Custom highlight groups for visual consistency
  • Token budget and change history limits
  • Gutter sign customization or disabling

Tool Integration

Works alongside existing tools:

  • blink.cmp - Smart Tab key handling with priority checking
  • snacks.nvim - Animated fade in/out for notifications
  • fidget.nvim - Optional notification backend
  • nvim-treesitter - Optional semantic context enhancement
  • plenary.nvim - Reliable async operations

Development & Roadmap

Current Status

nvim-stride is under active development:

  • Early release with stable core features
  • Support for Neovim 0.10+
  • Comprehensive documentation and examples
  • Open to feedback and bug reports

Future Enhancements

Planned improvements include:

  • LSP integration for diagnostics and symbol awareness
  • Enhanced Treesitter integration for scope-aware predictions
  • Multi-file context awareness for project-wide refactoring
  • Custom prompt templates for specialized workflows
  • Additional LLM provider support

Why nvim-stride?

What sets stride apart from traditional refactoring tools:

  • Context-Aware - Understands coding patterns, not just text matching
  • Non-Intrusive - Suggestions appear only when relevant
  • Ultra-Fast - Powered by Cerebras for minimal latency
  • Full Control - Review before applying, dismiss anytime
  • Dual Mode - Completion and refactoring in one plugin
  • Open Source - MIT licensed with transparent implementation