Web Search

Sending the message starts the trail, the sources stream in under the user turn, and the answer waits on onComplete. It shares its reveal, its clipping edge and its curved guide with Task List, so the two can sit in one thread without looking like two different products.

Give a step sources and it branches once more, into a Sources row carrying the marks as an overlapping stack. A table of five URLs is the heaviest thing that can appear in a thread and nobody reads it mid-run; what a reader wants at a glance is where the agent looked, which recognisable marks answer instantly. Each carries its title and domain on hover, lifts clear of its neighbours, and becomes a link when you pass an href. Past six, the rest collapse into a count.

branddraws a site's real logo from the same 24-mark set the social buttons use, in the site's own colour. The handful whose brand colour is black take the text colour instead and invert with the theme, which is what their own guidelines ask for. Anything not in that set takes a plain icon.

A search that hits cache and one that crawls five pages do not take the same time, and a log that marches at a fixed tempo reads as a list being printed rather than as work happening. Give a step dwell and the trail holds there that much longer, shimmering the whole while, before the next one lands. A step with sources splits in two: the step, then its Sources row a beat later.

While the search runs, an Agent Thinking indicator sits at the tail of the trail and moves down as each step lands, so the log always ends on what is being worked on rather than on the last thing finished. It leaves when the run completes. Pass working to relabel it, or false to drop it.

Dwells are still a stand-in for real latency. Once you have a live stream, pass revealed and advance it as each query resolves; the internal timer switches off and every animation stays the same. It counts rows, so a step carrying sources is two of them.