An agent that reads files, runs commands, and edits code produces a lot of activity. Shown in full, every read, write, and command output turns the chat view into a wall of expanded cards — even when most of it is exactly what you'd expect and doesn't need a second look.
Tool calls, collapsed by default
Every tool call — reading a file, running a command, editing a diff — now renders as a single plain-text line: the action, a short summary, and a result once it finishes. Click it any time, including mid-run, to expand the full command output, file diff, or live-streaming content.
This applies uniformly to whatever state a tool is in. A running command shows a small spinner in place of the usual arrow; nothing pops into a bordered card until you ask it to.
Reasoning gets its own dropdown
Sweet's reasoning trace — the model thinking through a problem before it acts — now lives in a collapsed "Thinking" dropdown under each response, separate from the final answer. It shows a small spinner while reasoning is actively streaming in, and settles into a plain arrow once it's done. Expand it whenever you want to see the model's actual reasoning; leave it closed and the chat reads like a normal conversation.
Reasoning also now survives a page refresh — previously it only lived in memory for the current tab and disappeared the moment you reloaded.
Net effect
The chat view now reads closer to a conversation with occasional, inspectable detail, instead of a transcript where every tool call and thought is expanded by default. Nothing about what Sweet does changed — just how much of it you have to look at to follow along.