PrivacyAI Shield 0.0.2 improves privacy protection for longer AI chat sessions.

What’s new:

• Added rolling context compaction so previous chat history is summarized instead of repeatedly sending the full conversation to the local privacy model.
• Improved prompt sanitization so only the latest user prompt is sanitized; previous context is treated as reference only and is not copied into the final prompt.
• Added stricter validation to prevent the local privacy model from answering the task, leaking context, or rewriting prompts unnecessarily.
• Added automatic repair and fallback behavior: if the local model returns invalid or unsafe sanitizer output, PrivacyAI retries once and then falls back to local regex-based redaction.
• Improved per-tab/session state handling for more consistent behavior across conversations.
• Updated tests and benchmarks for context leakage, prompt preservation, URL/path preservation, and task-answer prevention.

This update is focused on making PrivacyAI safer and more reliable with small local AI models while keeping prompts as close as possible to what the user originally typed.