Background tasks · header row + composer indicator

Your direction from the last round: tool blue, the tasks case (a shell or subagent Claude left running after its turn), shown as a row in the session header, under the goal row, plus something on the input area. Every design below keeps that header row and pairs it with a different composer indicator, inside the new composer (icons mode, goal active).

Each card says whether it works with today's data (the app only knows "tasks are running") or needs more from the Claude SDK, such as a task count, names or output. That's a real cost, so it's on every card.

Theme
Tasks

Best practice

Proven patterns. The header row reads like the goal row above it, and the composer gets one familiar signal.

In between

Grounded, but each takes one bold step.

Off the wall

Convention ignored. Mostly here to shake loose an idea.

Picks

My pick
2, with the pull-down from 4. The header row says "Background tasks running" and opens a pull-down on tap, exactly like the goal row opens goal details, so every header row behaves the same. The composer's own status line, which already says "Working" during a turn, says "Tasks running · you can still send" after it. It works with today's data, and the row grows names and a count later if the SDK gives them to us.
Runner-up
7, the tree. When a goal is active, nesting tasks under it reads well: this work serves that goal. Without a goal the elbow has nothing to hang from, so it needs a fallback, which is design 2 anyway.
Worth stealing
1's badge if you want Stop near your thumb. 3's chips once task names exist. 10's two-tone border is striking, but the border already means "goal", so giving it a second meaning is risky.
Avoid
11 implies you can stop one task; Stop actually ends the whole Claude process. 14 needs a start time the server doesn't have. 15, 19, 20, 21 are constant motion on a screen you'll leave open for hours.
Data check. From the earlier background-work spike, the app only knows a session has background work and whether it's a watcher or tasks. Counts, names, start times and output (anything marked "Needs" below) mean finding out what the Claude Agent SDK reports. I haven't verified that yet.

Copy this and paste it back to me: