Twinny

Generates single-line and multi-line code suggestions as developers type, using fill-in-the-middle (FIM) architecture where the model predicts code tokens between cursor position and surrounding context. Integrates directly into VS Code's IntelliSense pipeline, triggering automatically on keystroke with configurable debounce and context window management to balance latency against suggestion quality.

Provides a sidebar chat interface where developers can ask questions about code, request refactoring suggestions, or discuss implementation approaches. The chat maintains conversation history locally, passes selected code blocks and file context to the LLM, and renders responses with syntax-highlighted code blocks that can be accepted, diffed, or inserted into new documents. Uses stateful conversation management to preserve context across multiple turns.

Allows developers to specify custom API endpoints and ports for LLM providers, enabling connection to local Ollama instances on non-standard ports, private API gateways, or self-hosted model servers. Configuration is stored in VS Code settings and applied to all requests. Supports endpoint path customization for providers with non-standard API routes.

Defaults to routing all AI requests through a local Ollama instance (running on localhost:11434), keeping code and context on the developer's machine by default. Developers can optionally configure cloud providers (OpenAI, Anthropic, etc.) for higher-quality models, but this is an explicit opt-in choice. This architecture prioritizes privacy by default while maintaining flexibility for users who prefer cloud inference.

Computes vector embeddings of workspace files locally to enable semantic search and context retrieval for chat and completion suggestions. When enabled, the extension indexes accessible workspace files, stores embeddings in local storage, and retrieves relevant code snippets based on semantic similarity to current context or chat queries. Uses embedding model inference (likely via Ollama or provider API) to generate dense vectors for retrieval-augmented generation (RAG) patterns.

Abstracts LLM inference across 8+ providers (Ollama, OpenAI, Anthropic, OpenRouter, Deepseek, Cohere, Mistral, Groq, Perplexity) through a unified OpenAI-compatible API interface. Developers configure provider and endpoint via settings, and the extension translates all completion and chat requests to the selected provider's API format. Supports both local inference (Ollama) and cloud APIs with configurable authentication and endpoint paths.

Analyzes staged changes in Git (diff between HEAD and staging area) and generates descriptive commit messages using the configured LLM. Extracts file changes, added/removed lines, and context from commit scope, then prompts the model to generate conventional commit-formatted messages. Generated messages can be accepted or edited before committing.

Allows developers to define custom system prompts and instruction templates for code completion and chat interactions. Templates are stored in extension settings (likely JSON or YAML format) and injected into LLM requests before user input, enabling fine-tuning of model behavior without forking the extension. Supports variable substitution for context like file path, language, or selected text.

Stores chat conversation history locally in VS Code's extension storage (likely IndexedDB or file-based), preserving multi-turn conversations across editor sessions. Conversations are indexed by timestamp or ID and can be retrieved, resumed, or cleared via UI. No cloud sync or team sharing — history remains on developer's machine only.

Renders code suggestions from chat responses with syntax highlighting, language detection, and interactive actions (accept, diff, copy, new document). When user accepts a suggestion, the extension inserts code into the current file at cursor position, or optionally creates a new document. Diff view shows changes side-by-side before acceptance, allowing review before committing.

Experimental feature enabling distributed inference across peer nodes in a decentralized network, allowing developers to contribute spare compute capacity and access inference from peers. Technical implementation details are not documented, but likely involves splitting inference workload across multiple machines or sharing model weights via P2P protocol. Status and stability unclear.

Provides a full-screen chat mode separate from the sidebar, offering a distraction-free environment for extended code discussion and collaboration. Likely includes conversation history, model/provider selection, and all chat features (code block actions, context insertion) in a dedicated view. Can be toggled between sidebar and full-screen modes.