Repository navigation
What's new
-
Chat tool modes keep every request lean:
/videoadds video and audio tools,/docsadds document tools, and/generalresets to core tools. Modes stack (a PDF-to-video task can use both) and stay on for follow-ups. If a tool is hidden, the agent can still find and enable it on demand. -
Small-context models, including 4k on-device and Ollama models, now fit their window: optional instructions and tool families are trimmed automatically instead of pushing the prompt past the limit.
-
Tool steps in the chat trace now use purpose-built preview cards — checklists for tasks, Q&A for questionnaires, file and schedule cards, voice and narration lists, link previews, and search results — instead of a plain text summary.
-
LFM3 1.7B is now available in the on-device model catalog.
-
The composer now continues bulleted and numbered lists when you press return, including nested lists.
-
Long videos can now be storyboarded and mix-checked before export: scene previews work across the whole video, and audio checks cover the full 10-minute render limit instead of stopping at 45 seconds.
Fixes
- Downloading an on-device model no longer throws a text-rendering error in the model list, and the download button now shows the live percentage so progress is visible while it runs.
- Tapping "Use model" on a model that needs more RAM than the device has no longer kills the app: it asks first, then loads with the CPU backend instead of the GPU delegate that copies the whole graph into memory.
- Running tools now show a live "Running X" row in the chat trace as soon as the call starts, instead of appearing only after it finishes (or showing just a generic Thinking indicator), so long or stuck calls are visible.
- Prior reasoning steps are no longer replayed to the model on every follow-up, removing SDK warnings and unnecessary context use.
- On-device models now run on a newer LiteRT-LM runtime, fixing loads that failed with a weights error.
- On-device models are pinned to exact, verified file revisions, so a catalog refresh can no longer point at weights the app cannot run.
- On Android, downloaded on-device models that no longer match the catalog are cleaned up automatically instead of failing to load.
- Stopping a reply now dismisses the keyboard, and canceled runs no longer raise unhandled promise rejections.
Improvements
- Improved typograhy in tool trace and overall UI/UX of trace previews.