Short answer: the Claude SDK sends everything these designs need, but conduit currently throws it away. Today conduit keeps only "this session has background work, yes or no". Showing names, kinds, a count or ages needs a small plumbing change. No SDK upgrade is needed, and nothing has to be invented.
Checked against the installed Claude Agent SDK (0.3.280) type definitions and conduit's Claude translator. Task names below are the kind Claude really writes; I sampled them from this session's own background tasks.
| Shown in the design | Where the SDK has it | Conduit today |
|---|---|---|
| Task count | Yes. background_tasks_changed sends the full live list on every change (and again on reconnect). | Dropped. Only "any tasks?" survives. |
| Task name | Yes. description on each task, e.g. "Run codex to implement ticket .1". Sentences, so they need truncating. | Dropped. |
| Kind (shell, subagent, workflow) | Yes. task_type. | Dropped. |
| Age ("6m") | Not directly. The list has no start time. Conduit can note the time it first sees each task. That's accurate, because conduit starts the Claude process and sees every task begin. | Not recorded. |
| Shell command | Indirect. Each task points at the tool call that started it, and that call holds the command. | In the transcript, not linked. |
| Stop one task | Yes. stopTask(id). | Not wired. Stop interrupts the turn, which kills every task. |
Each of 3, 4 and 7 brought its own input-area signal. All three at once is too much, so pick one. Chosen: B, one dot per task. The combination above and the desktop mock use it.
There's room for up to four chips with ages. On desktop the pull-down is a popover under the row instead of a full-width sheet. Tap the row to toggle it.