No-bullshit VS Code extension for LLM inline completions via llama.cpp FIM (/infill) endpoints.
Usage
- Run a llama.cpp server with
/infillsupport (e.g../server -m model.gguf --ctx-size 8192 --host 0.0.0.0 --port 8080) - Configure the extension in VS Code settings (see Configuration)
- Place cursor anywhere in a file
- Press
Cmd+'to trigger an inline completion suggestion
Configuration
| Setting | Default | Description |
|---|---|---|
banana.baseUrl |
http://localhost:8080 |
Base URL of the llama.cpp server |
banana.n_predict |
128 |
Maximum number of tokens to predict |
banana.t_max_predict_ms |
500 |
Maximum time in milliseconds for prediction |
banana.stop |
null |
Custom stop tokens (null = omit the field, let model decide when to stop) |
banana.headers |
{} |
Custom HTTP headers to send with each request. Keys and values must be strings. |
banana.temperature |
0.6 |
Temperature for token sampling (0 = greedy, higher = more random/creative) |
Context Injection
Banana Complete automatically injects additional context into each request via the input_extra field:
- File info: Project name, language, file path, and cursor position
- References: Top completions from VS Code's built-in completion provider
- AGENTS.md: If an
AGENTS.mdfile exists in your workspace root, its contents are included as additional context
Development
npm install
npm run compile
Package
npm run package
Credits
Icon made by juicy_fish from www.flaticon.com