Changelog

New

• Chrome 138 onboarding: The welcome page now shows what's actually happening on first install — model checking, downloading with a live progress bar, ready, or an error with steps to fix it. No more static instructions that assumed you'd already toggled flags.
• Download progress: If Gemini Nano needs to download, both the welcome page and popup show live MB progress, not just a spinner.
• Popup status: The popup now shows the model name, download state if it's still loading, and a Retry button if initialization fails.

Fixes

• Add Context output: The modal was showing the AI's lead-up text ("Okay, great! Knowing you want to...") on top of the actual prompt. It now shows just the rewritten prompt.
• Ctrl+Shift+E on Claude.ai: The shortcut was being picked up by Claude.ai's own handlers. The extension now intercepts it before the host page does.
• Generating phrases: The modal title cycles through different phrases while the AI works, instead of sitting on "Enhanced Prompt" the whole time.

Security

• HTML escaping in markdown renderer: Single and double quotes were missing from the escape list — now consistent with everything else.
• Perplexity Insert origin check: The scripting API now only runs on Perplexity pages, closing a gap left over from when the permission was first added.
• History copy bounds check: Clicking a history item now validates the index before touching the array.
• Removed platform logging: The extension was logging which site it detected to the browser console, visible to any host-page script that overrides console.log.

Under the hood

• Insert no longer uses execCommand: Switched to the Selection API for contenteditable editors and the native textarea value setter for ChatGPT — both were long overdue.
• Copy uses Clipboard API: execCommand('copy') replaced with navigator.clipboard.writeText.
• Dead code removed: Unused variable cleaned up in context mode.

New

• Claude.ai support: Enhance button, Instant Enhance, and Add Context mode now work on claude.ai.

Under the hood

• Per-tab AI sessions: Each tab now gets its own AI session for context mode — using Add Context in multiple tabs simultaneously no longer produces mixed-up results.
• Session cleanup: AI sessions are automatically destroyed when tabs close, freeing memory.

Security

• Sender validation: Message and port handlers now verify the sender is the extension itself, blocking unauthorized callers.
• Input limits: Context mode chat messages now have the same 10,000-character limit as instant enhance.
• HTML escaping: Single-quote escaping added to all HTML sanitization functions (OWASP complete).
• Output cap: AI output is truncated at 50,000 characters to prevent DOM overload from runaway model responses.
• Storage validation: History loading validates data types to handle corrupted storage gracefully.
• Dead code removed: Unreachable hostname checks cleaned up in platform detection.

Stability

• Streaming timeout: If the AI model hangs for more than 30 seconds during enhancement, the UI automatically recovers with an error notification instead of staying stuck.
• Nested list rendering: AI responses with indented sub-lists now render correctly as nested bullet points.
• Perplexity Insert fix: Insert button now works reliably on both Perplexity's home page and conversation pages.

Under the hood

• Robustness: Separate abort controllers for instant enhance and context mode streams — stopping one no longer affects the other.
• Safety guardrails: Harmful content refusal now applies to both Instant Enhance and Add Context modes (previously context mode only).
• Input guard: Prompts over 10,000 characters are rejected with a clear error instead of silently failing.
• Faster streaming: HTML escaping during streaming rewritten to eliminate unnecessary DOM allocation per token.
• Stability: 30-second safety reset prevents the enhance button from getting stuck in a loading state.
• Memory: URL polling interval now properly cleaned up on page unload.
• Unknown platforms: Extension no longer attempts to inject on unsupported pages.

New

• Enhance button: Enhance icon positioned near the text box on all supported platforms — outside the input to keep website UX clean.
• Add Context mode: Dropdown with "Instant Enhance" and "Add Context" options — the AI asks clarifying questions before crafting your prompt.
• Draggable context box: In-page overlay for iterative prompt crafting — drag it anywhere, chat with the AI, get your enhanced prompt in the modal.
• Streaming enhancement: Prompts stream in real time with a typing indicator — action buttons appear when generation finishes.
• Context-aware regeneration: Regenerate remembers your context answers instead of starting from scratch.
• Insert into text box: One-click button to push the enhanced prompt directly into the platform's input field.
• Ctrl+Shift+D shortcut: Opens context mode without touching the mouse.
• Perplexity support: Text extraction and enhancement now works on Perplexity alongside ChatGPT and Gemini.
• AI disclaimer: "AI may make mistakes" notice added to the modal, context box, popup, and welcome page.

Improvements

• Prompt engineering: Rewrote AI prompts using few-shot examples and Google-recommended delimiters for Gemini Nano — short or vague inputs now produce clear, detailed results.
• AI session settings: Low temperature and topK on both sessions for more consistent output.
• Button injection: MutationObserver re-injects the button after SPA navigation on all platforms.
• Dropdown menu: Keyboard shortcut hints included, positioned to avoid viewport clipping.
• Markdown rendering: AI responses now render headings, bold, italic, lists, and inline code.
• Popup: Shows both modes with shortcuts, capped at 10 most recent prompts.

Bug fixes

• Fixed drag event listener leak when context box port disconnects unexpectedly.
• Fixed button position interval continuing to run after extension is invalidated.
• Fixed untrimmed message used in chat session retry path.
• Fixed version tracking being set before the updates page finishes rendering.
• Removed dead code (getInputContainer() function).

New

• Dark & light theme: Toggle between dark and light mode from the popup toolbar.
• Redesigned UI: New teal blue look applied consistently across the popup, welcome page, and content overlays.

Improvements

• Dark mode default: Dark slate background with teal accents.
• XSS fix: Prompt display was rendering unsanitized HTML.
• Clipboard feedback: Copy actions now show success and error states.
• Notifications: Toast notifications now match the active theme.

Bug fixes

• Fixed potential crash when background worker returns no response.
• Added error handling for history loading.
• Merged duplicate install listeners in background worker.
• Restricted resource access permissions to target sites only.

New

• Gemini support: Now works on Gemini in addition to ChatGPT.
• Button visibility: Translucent by default, appears on hover.
• Loading state: The button shows a clear animation while the AI is working.
• Welcome page: Shown on first install to walk you through setup.
• Update page: You'll see what changed after each update.

Improvements

• Prompt detection improved across platforms.
• Smoother animations.
• Clearer error messages.
• Button hover behaviour fixed during processing.

Bug fixes

• Fixed issue where enhanced prompt wouldn't display properly.
• Resolved button flickering during hover while processing.
• Fixed text extraction from contenteditable elements.

Initial release

• On-device AI prompt enhancement using Gemini Nano.
• ChatGPT integration with floating button.
• Keyboard shortcut (Ctrl+Shift+E) support.
• Copy to clipboard.
• 100% private, offline operation.
• Real-time status monitoring.