Local Inference OnDevice Agent

Stability and fixes. • Much lighter on memory. On phones with less RAM the app was loading the model at the largest possible context window, which left the system paging it in and out and could make the phone feel like it had frozen. It now sizes the model to what your device can actually hold — around 200 MB less memory in use, and far less swapping. • The app now genuinely frees the model when you switch away, instead of holding it in the background. • Photos work properly. You can send a picture without typing a caption, retrying a photo question keeps the photo, and attaching one no longer silently turns tools off. • Tools tell you why they failed instead of just saying "failed", and tools that need the internet are hidden when you're offline rather than timing out. • Chats can be renamed — press and hold a chat in the list. • Answers you stop early can now be copied and read aloud. • A reply interrupted by the app closing is no longer lost; you can retry it. • The top bar now blends into the background instead of sitting on it as a flat slab. • Web search results are no longer matched to the wrong page. • Attached images are cleaned up properly instead of taking up space forever.