A window's state is derived by the model from what the daemon sends, since the daemon sends none¶
TUI decision 129 · accepted · 2026-09-27 · supersedes 4
Supersedes the branch_state of Decision 4, which only the
old in-process harness logged: since Decision 100 every window said running for
ever, so the status line never counted what needs you, done or failed, no window
was unread, and Enter on the command line opened the first window rather than the
one asking. The fold now says needs_input while anything in a window is pending,
a subagent's request as much as its own agent's, and otherwise where the window's
own agent is: running, or at rest once the log says its turn is over (the
durable agent_state that turn_ended, cancelled and agent_done become; the
live one says idle once before the task is taken). A rest is failed_unread
when the headless printer would exit 1 for it (a model request that failed, a
tool that kept failing, an agent that ended other than finished), except that a
cancel stays done_unread as Decision 7 has it; otherwise done_unread. Either is
unread until the window is opened, and a window that ends while it is open is
read. The view's own count of what is pending (Decision 126) agrees with the
model's now. Proof: test/troupe/window_attention_test.exs, which failed on the
chunk's tip, and the installed TUI on a delegation whose subagent asks, on the
pull request.
Governs: clients/tui/lib/troupe/ui/tui/model.ex