NiceMove

A VS Code extension that classifies Move smart-contract code using a local CodeBERT + LoRA model (trained as part of an MSc thesis), and then asks Claude for a concrete fix when the classifier flags a real issue.

Architecture

VS Code Extension (TS)
        │  in-process — no network call, no local server
        ▼
CodeBERT + LoRA, merged and exported to a quantized ONNX model
        │  run via transformers.js (onnxruntime-node)
        ▼
   {label, confidence}
        │  if label != "Perfect":
        ▼
Anthropic API  (your key, your billing)
        ▼
Fix suggestion shown in a side panel (with Apply Fix button)

Classification runs entirely inside the extension host — there is no Python process to start and no local server to keep running. The model is bundled with the extension (model/) as a quantized ONNX export of the CodeBERT + LoRA adapter trained for this project; see scripts/export_onnx_model.py in the repo root if you need to regenerate it.

Setup

1. Install dependencies and compile:

   cd extension && npm install && npm run compile
   

2. Open the extension/ folder in VS Code and press F5 to launch an Extension Development Host.
3. In the host window, set your Anthropic API key in Settings → Extensions → NiceMove → Anthropic API Key.
4. Open a .move file, select code (or none for the whole file), and run Move: Diagnose Selection (Cmd+Alt+M / Ctrl+Alt+M).

Features

• Inline diagnostics — predictions appear as squiggles in the editor and entries in the Problems tab.
• Multi-label diagnostics — when the classifier is uncertain, the second-most likely prediction is shown as a hint alongside the primary diagnosis.
• Quick Fix lightbulb — every diagnostic carries a Fix … with Claude code action; trigger it with Cmd+..
• Streaming Claude output — the side panel streams tokens as Claude generates them.
• Apply Fix inline — after Claude suggests a fix, click "Apply Fix" in the panel to patch the code directly in the editor.
• Status bar indicator — shows ✓ NiceMove when idle, a spinner while classifying, and ⚠ NiceMove when issues are detected.
• Workspace scan — run Move: Scan All Move Files from the command palette to classify every .move file in the workspace at once.
• Classification cache — unchanged code is not re-classified, making repeated diagnoses instant.
• OOD guard — a Move-likeness heuristic suppresses predictions on non-Move snippets so the model's over-confidence on out-of-distribution input does not surface as false diagnostics.
• Auto-classify on save (opt-in) — set moveClassifier.classifyOnSave to true to refresh diagnostics every time a .move file is saved.

Settings

Packaging

To produce a .vsix for sideloading or marketplace upload:

cd extension
npx @vscode/vsce package --no-yarn

Privacy

• Classification runs entirely in-process inside VS Code — no network call, no local server, no external process.
• Your code is sent over the network only when an issue is detected, to api.anthropic.com using your API key, to request a fix.
• The extension does not send telemetry, analytics, or code anywhere else.
• We never see your code or your key.

Billing

You bring your own Anthropic API key. You are billed by Anthropic directly based on your usage of their API. The maintainers of this extension do not process payments and do not see your key or your invoices.