Autofix card not dismissed by Esc (intermittent; E2E-discovered)

Discovered by the ItE2E end-to-end suite (test/e2e/), running against the Store package Microsoft.IntelligentTerminal_8wekyb3d8bbwe (v0.1.1681.0) with the copilot agent. Filing for triage — it is unclear yet whether this is a product bug or an E2E assertion weakness; both hypotheses are below.

Summary

The autofix suggestion card ("[ Run command ] / Insert in Terminal") sometimes fails to disappear after the user presses Esc. The E2E "reject/dismiss" check pressed Esc up to 5 times (800 ms apart) and the card text was still present, so the assertion that the card is gone failed.

This is intermittent: the same test has passed in prior runs (it is counted among the green feature cases in test/e2e/README.md), and failed in this session's run — i.e. it is flaky, which on its own is worth fixing because a release-gating E2E should be deterministic.

Failing test

• File: test/e2e/tests/Feature.AutofixPane.Tests.ps1:41-43
• Checklist item: doc/release-check-list.md §3 "Reject/dismiss works" (tagged [UT✓] [E2E]).

It 'Reject/dismiss works (Esc closes the card)' {
    for ($i = 0; $i -lt 5 -and (& $script:CardShown); $i++) { Send-AgentKey -App $script:app -Key Escape | Out-Null; Start-Sleep -Milliseconds 800 }
    (& $script:CardShown) | Should -BeFalse
}
# $script:CardShown = { (Get-AgentPaneText -App $app -MaxLines 60) -match 'Run command|Insert in Terminal' }

Observed vs expected

Run evidence (this session): Tests Passed: 9, Failed: 1 for Feature.AutofixPane.Tests.ps1; the only failure was this case. All other autofix cases (render, insert, run, stashed, split, cleanup) passed.

Reproduction

pwsh -File test/e2e/Invoke-ItE2EReport.ps1 -Path test/e2e/tests/Feature.AutofixPane.Tests.ps1
# (requires the Store/Dev IT package installed + an authenticated `copilot` on PATH)

Because autofix de-dupes identical failures and a displayed card suppresses the next autofix, reproduction is timing-sensitive; re-run a few times.

Code pointers / hypotheses

The agent pane (wta-helper ratatui TUI) handles Esc as an ordered match with guards in App::handle_key (tools/wta/src/app.rs). The relevant arms, in evaluation order:

1. KeyCode::Esc if self.show_notification_banner → dismiss_notifications() — app.rs:6567
2. KeyCode::Esc if recommendations().is_some() || (autofix.pane_id.is_some() && !turn.is_idle()) → turn_cancel() (dismisses the armed fix card) — app.rs:6570
3. KeyCode::Esc if autofix.suggested_pane_id.is_some() → clears suggested_pane_id — app.rs:6606
4. KeyCode::Esc => → clear_input() — app.rs:6611

Because arm (1) precedes the card-dismiss arm (2), the first Esc only dismisses the notification banner; the card needs a second Esc. That matches the known "Esc needs up to 2 presses" behavior. But 5 presses still not clearing the card points to one of:

• Hypothesis A — product bug: the bottom-bar notification banner (show_notification_banner) is re-raised every frame while the card is armed, so every Esc is consumed by arm (1) and arm (2) is never reached → the card is undismissable by Esc. If so, fix the arm ordering / guard so a single Esc that has an armed card routes to turn_cancel (or don't re-arm the banner while the card owns dismissal).
• Hypothesis B — E2E false positive: Esc does dismiss the active card, but the dismissed card leaves its text ("Run command" / "Insert in Terminal") in the chat transcript / scrollback, so the CardShown regex over the last 60 lines still matches stale text. If so, the test should assert on the active card region (e.g. the bordered card / button row) rather than a substring of the whole buffer — and the product is fine.

Triage should determine which. A quick check: after pressing Esc, capture the pane and see whether the matched text is inside the live card border or in the scrolled-back transcript.

Suggested labels

Issue-Bug, Area-AutoFix, Area-AgentPane

Notes for whoever picks this up

• Related UT (logic half is unit-covered): trigger_echo_pane_clears_when_state_returns_to_idle and turn_cancel in tools/wta/src/app.rs. The gap is specifically the interactive Esc → dismiss path end-to-end.
• If Hypothesis B is confirmed, please also update test/e2e/tests/Feature.AutofixPane.Tests.ps1:41-43 to scope CardShown to the active card so the E2E stops flaking.

Start with tools/wta/src/app.rs:6567-6611 and the Esc handling order, then run test/e2e/Invoke-ItE2EReport.ps1 -Path test/e2e/tests/Feature.AutofixPane.Tests.ps1. Capture the pane after Esc to determine whether the active card remains or only stale transcript text matches. Done means identifying whether the product or Feature.AutofixPane.Tests.ps1:41-43 needs correction and making the reject/dismiss check deterministic.