v0.4.8
v0.4.8
Video is now something the agent can actually look at, plans are easier to take with you, and a turn's work survives an interruption.
New
- Videos can be read: the agent samples frames from a video file and looks at them, reporting the timestamp of each frame.
- Set how many frames a video read uses —
/video-frames(CLI) or "Video frames" under Context Management in the settings popup (VS Code). The agent can request more frames, or zoom into a few seconds, when it missed something. - Copy a plan as Markdown:
/copy-plan(CLI) and a "Copy plan" button on the plan card (VS Code). - Compaction summaries can be copied, or moved into a brand new chat to continue from (VS Code).
Improved
- Turns are written to disk as they happen — the assistant's replies, each tool result and your steering messages now survive a crash, a dropped connection, or closing the window mid-turn.
- The conversation sidebar shows messages you typed or queued during a running turn instead of going blank until the turn finishes (VS Code).
- Starting a new chat shows a loading state instead of appearing to do nothing (VS Code).
- Questions the agent asks you render as Markdown, so code and lists in a prompt are readable (VS Code).