ContextShield

Write freely. Scrub sensitive data and improve prompt clarity locally, before you send.

ContextShield is a VS Code extension that redacts PII and secrets from your editor selection, enhances the wording with a local Open Neural Network Exchange (ONNX) model, and replaces the text in place. No cloud required, nothing leaves your machine.

Why ContextShield

Developers frequently paste rough notes, stack traces, and config snippets into AI chats. That creates two recurring problems:

• Prompts are noisy, vague, or typo-heavy, which reduces answer quality
• Prompts may contain internal IPs, emails, tokens, or credentials

ContextShield solves both in one workflow:

1. Scrub sensitive data first
2. Optionally enhance wording and structure with a local model
3. Replace the selected text with a safer, clearer version

What It Is

• A local-first pre-send prompt scrubber and enhancer for VS Code
• Built for Cursor/Copilot-style chat workflows and editor selections
• Best for day-to-day debugging, code review prompts, and incident triage messages

What It Is Not

• Not a cloud prompt service
• Not a replacement for your AI assistant
• Not a post-send redaction tool — it works before you send text

Supported Workflows

• AI chat boxes in editor environments (selection + replace flow)
• Code and prose mixed selections
• Scrub-only, enhance-only, or combined mode through settings toggles

Core Features

• Redacts PII: email, phone, IPv4, UUID
• Redacts secrets: JWT-like strings, AWS-style key IDs, high-entropy tokens, password-in-context patterns
• Supports custom regex patterns for org-specific secrets
• Enhances prompt wording with a local ONNX model (@huggingface/transformers)
• Sidebar settings panel for model control, prompt settings, and scrubbing options
• Clipboard-safe replace flow — original clipboard is restored after transformation

Local-First By Design

• Inference runs locally in an isolated worker process (dist/llmWorker.js)
• No cloud inference API in the scrub or enhance pipeline
• Model artifacts are cached under VS Code extension storage (globalStorageUri/.../hf-cache)
• Includes a sidebar action to delete local model cache

Screenshot

Demo

Quick Start

For users

1. Install the extension
2. Open the ContextShield sidebar (ContextShield: Open Sidebar)
3. Download the local model from the Local model section
4. Select text in your editor or AI chat input
5. Press Cmd+Shift+2 (Mac) / Ctrl+Shift+2 (Windows/Linux)
6. Review the replaced text, then send

Architecture Overview

ContextShield runs as three isolated runtimes: an extension host, a sandboxed React webview, and a separate Node worker process for local inference.

System Architecture

User actions flow from the editor or webview into the extension host, then into the scrubbing engine and optionally into the model manager and isolated worker for enhancement.

Scrub and Enhance Flow

The primary business flow. Scrubbing always runs first; enhancement is conditional on LLM being enabled and the model being ready.

If the model is unavailable or generation fails, the command degrades gracefully to the scrubbed text — it never hard-fails on the enhancement step.

Settings and Model Control

• Download model triggers ModelManager.load(), which spawns the worker and initializes the ONNX model
• Delete cache stops the worker process and removes the local hf-cache directory
• Status updates are pushed to the webview as modelStatus messages: unknown, downloading, ready, or error

Why The Architecture Matters

• Worker isolation — native model and ONNX runtime failures cannot crash the extension host process
• Scrub-first pipeline — sensitive values are always redacted before any text reaches the model
• Typed protocol — the shared webviewProtocol.ts contract catches integration regressions at compile time
• Separation of concerns — webview UI, host message routing, scrubbing engine, and model runtime are fully isolated from each other

Trust And Boundaries

ContextShield is a tool for safer, clearer prompts — not a correctness oracle.

• Scrubbing is pattern-based and can produce false positives or negatives
• Enhancement improves wording and structure; it is not guaranteed to preserve intent perfectly
• You should review transformed output before sending to external AI systems

Installation

From Marketplace

Install ContextShield on Visual Studio Marketplace.

From source / VSIX

npm install
npm run build

Then package and install as a VSIX in VS Code.

Model Notes

• Default model: onnx-community/Llama-3.2-3B-Instruct-ONNX
• Initial download may take several minutes depending on network and hardware
• Cache persists across extension reloads until manually deleted from the sidebar
• If the model ID changes in settings, a reload is attempted automatically when LLM is enabled

Development

npm install
npm run check
npm run test
npm run build

Press F5 to launch the Extension Development Host.

Project Structure

License

MIT — see LICENSE for details.