[rohita5l]

Stacked on #1245 (base ci-flake-hardening); this diff contains only the follow-up commit. Retarget to main once #1245 merges.

What

Fixes the two remaining CUJ 4 flake causes from the CI analysis in #1245. The diagnosis replays the raw PTY transcripts from CI's CUJ evidence artifacts (runs 38801294447 and 39696214480, with passing siblings) through pyte, plus the final screens from 19 failing jobs.

1. wait_for_background_tasks (tests/integration/utils/terminal.py). "TUI did not show Claude's task view reporting no running tasks within 180s" was 17 of 43 flaky CUJ 4 jobs, and the same wait ends the Claude skill-toggle tests. The work had finished every time. The check was never flickering: in the failing transcript it is False continuously until the final repaint. There are two real causes:

   • No task view on screen (14 of 19). Claude was back at its idle prompt with every agent finished, either because a /tasks typed mid-turn was swallowed or because the view closed itself. The old wait never reopened it.
   • Frozen running snapshot (5 of 19). /tasks opened mid-turn and froze on a running state ("Background task update waiting while this panel is open"). The running header names the subagent type (ug-smart-router:… › Read file contents, footer … · x to stop · …), so the old agent › regex never even recognized it.

   The wait is now a bounded loop:

   • It resubmits /tasks when no task view is shown at an idle prompt, at most every 3 s.
   • It sends Esc, then /tasks, when the view hasn't reported completion for 10 s.
   • The detail regex accepts any <name> › <title> header and stays within one line.
   • Success still requires Claude's own task view to report that nothing is running, held for 0.3 s within the same 180 s deadline. Completion is never inferred from scrolled history.

2. CUJ 4 "exactly one completed child" (about 9 failures with assert 2 == 1). These were genuine second subagents, not one subagent counted twice: every failing job has two distinct agent-<id>.jsonl transcripts. The parent sometimes started a "relay" subagent just to message the first one for its result, e.g. Agent(subagent_type="fork", description="Get subagent result", …) in job 119554254666.

   • The delegation prompt now adds "Use only that one subagent: do not start any other subagent."
   • assert len(completed_children) == 1 and all routing assertions are unchanged.

Both changes stay within tests/AGENTS.md. Reopening /tasks is real keystrokes through the visible UI, and the prompt change is scenario input. There are no mocks, skips, retries of the task, or weaker assertions.

Not covered / risks

• The prompt change is policy-sensitive and unverified live. If relay spawns persist, the next option would be counting only children whose first prompt names the task file. That changes the assertion, so it isn't done here.
• Esc race. Esc is sent when the view was open on the latest screen. If Claude closes the view in that instant, Esc lands on the main prompt. The old final close step had the same race.
• First matching header wins. The detail regex matches the first <name> › <title> header on screen. A scrollback line containing " › " could make it misread the panel. That can only cause a missed completion, meaning another reopen, never a false success.
• The 14/19 shape (swallowed or self-closed view) is confirmed from final screens only; the transcripts for those runs were unavailable.

Docs

• tests/e2e_cuj/README.md: describes the reopen behaviour and the one-subagent prompt.

Tests

• tests/test_e2e_cuj_helpers.py:
  • Predicate checks on the real CI screens: a view that closed itself, a stale running panel, and a completed panel.
  • Idle-prompt checks.
  • Fake-ug PTY cases: a swallowed /tasks followed by completion; a stale running view that gets reopened; both together; and a never-completing case that must still raise at the deadline without sending /exit.
  • The existing background-wait cases still pass.
• tests/test_integration_contract.py, tests/test_e2e_cuj_helpers.py, tests/test_cuj_evidence.py and tests/test_tui_request_recorder.py: 339 passed.
• tests/e2e_cuj collects 111 tests; tests/integration collects 151 of 163. ruff check and format are clean.
• Not run live. This PR's CI run is the first real exercise of the reopen path.

This pull request and its description were written by Isaac.