No in-session way to list tracked background shells (TaskList returns todos, not shells) → false "all clear"

Summary

There is no reliable in-session way for Claude (or the user) to enumerate the background shells the harness is tracking. The exit dialog clearly knows this list ("Background work is running" — with human-readable names like XLSR-53 F01 sweep (background)), but nothing surfaces it during the session:

• TaskList returns the todo task list, not background shells. It reported No tasks found while three background shells were still running, so an empty TaskList is easily (mis)read as "nothing is running."
• The task IDs needed for TaskStop only appear in the one-off completion notification for each shell. If that notification has scrolled off or a context compaction happened, there is no command to re-list live shells and recover their IDs.

The practical failure mode: to answer the user's "is anything still running / did it all close?", the assistant falls back to a hand-written pgrep pattern over script names it happens to remember. That silently misses any job whose command doesn't match the guessed pattern, producing a confident but wrong "all clear." In my case a genuinely-hung job (sweep_ssl.py, stuck ~17h at model-load, 0.2% CPU) plus two orphaned until … pgrep … sleep watcher loops were still alive, but my pgrep filter (built from other script names) didn't cover sweep_ssl.py, so I reported everything closed. The exit dialog was right; my check was wrong.

Environment

• Claude Code 2.1.202
• Platform: Linux
• Non-interactive/agent shell

Steps to reproduce

1. Start one or more long-running background shells (run_in_background), e.g. a slow job plus an until ! pgrep -f "<thing>"; do sleep 5; done watcher loop.
2. Let a context compaction occur (or simply lose the original completion notifications).
3. Ask the assistant to verify whether background work is still running.
4. TaskList → No tasks found, even though the exit dialog lists live shells. There is no command that returns the same list the exit dialog uses (task ID + name + status).

Expected

A first-class way to list the harness-tracked background shells in-session — the same data the exit dialog renders — including each shell's task ID, display name, and running/exited status, so that:

• an empty result unambiguously means "no background work is running," and
• the returned IDs can be passed straight to TaskStop.

Suggestions

• Have TaskList (or /tasks) include background shells, not just todo tasks; or add a dedicated "list background shells" affordance.
• Bonus: surface basic liveness (CPU%/last-output time) so a hung shell (sleeping at 0% CPU, e.g. a stalled model/network fetch) is distinguishable from an actively-running one.
• Doc note: clarify that TaskList covers todos, not background shells, so an empty TaskList is not evidence that background work has stopped.

Workaround

Enumerate the user's own processes directly (ps -o pid,ppid,etime,%cpu,stat,args -u "$(id -un)") instead of grepping for remembered script names, and treat the exit dialog's list — not TaskList — as the source of truth.