macOS Real-Run Verification Report (PR #4919)

The PR asks for macOS verification with a real window drag (the author has no macOS machine). This report provides exactly that, on macOS Darwin 25.5.0 / Node v22.22.2 / Terminal.app, at two levels: byte-level (what escape sequences the TUI actually writes on resize, captured from a real pty) and window-level (real Terminal.app windows resized by scripted bounds changes — real SIGWINCH bursts, real scrollback reflow — then the tab's full scrollback inspected for #4891's mixed-width fragments).

Because main has moved 43 commits past the PR base (including commits touching AppContainer.tsx), all runtime tests ran on the auto-merged tree (main @ 963fc543d + PR decea67e8; merges cleanly) — i.e. the tree that would actually land — with vanilla origin/main as the before build.

Static checks (PR head decea67e8)

• Author-listed suites: 98/98 — useResizeSettleRepaint 6, AppContainer 79, MainContent 13 (matches the PR's numbers).
• npm run typecheck --workspace=@qwen-code/qwen-code clean; eslint clean on all 4 changed files.
• Hook logic read-through: mount no-op, trailing-edge timer with cleanup-cancel, return-to-settled-width no-op, unmount cancel — all four pinned by the new tests; the "referentially stable refreshStatic" requirement the hook documents is satisfied at the AppContainer call site (useCallback whose deps don't change on resize).
• CI at time of writing: macOS / Ubuntu tests, Lint, CodeQL all green; Windows test still pending.

1. Byte-level: what the TUI writes on resize (real pty, tmux harness)

Harness: the built TUI runs in a real pty; a driver issues real TIOCSWINSZ + SIGWINCH width-change bursts and records every output chunk with a monotonic timestamp. Markers: full clear = ESC[3J (only emitted by clearTerminal); old per-event repaint = cursorTo(0,0)+eraseDown (ESC[1;1H ESC[J]).

This is the PR's contract verified on the wire: per-event viewport-only erases are gone; one full clear (including scrollback wipe) fires once per settled gesture at the documented ~200 ms trailing edge; a gesture that returns to the starting width emits nothing.

2. Window-level: real Terminal.app, real resizes, scrollback inspected (#4891 repro)

Each build ran --approval-mode yolo -i "<prompt streaming 100 numbered lines>" in a fresh Terminal.app window; the window was then resized through 6 real bounds changes (~120 ms apart, ending at a different width than it started); after completion the tab's entire scrollback (history) was dumped and analyzed. (tmux can't reproduce #4891 — the PR notes this correctly — hence real windows for the visual half; the after burst fired while the response was still streaming, the before burst right after streaming; the stranding mechanism is per-resize-event and showed up regardless.)

The before run is a faithful reproduction of #4891 on macOS — alternating-width duplicated segments stranded above the viewport. The after run leaves a single clean final-width history: the settle-time ESC[3J genuinely wipes the stranded fragments in a real terminal, not just in unit tests.

Notes for the record

• Trade-offs behave as the PR describes: the post-settle clear is visible as one brief flash per gesture, and pre-session scrollback is cleared on settle (same as /clear). Observed, not measured further.
• VP mode (ui.useTerminalBuffer=true) was not runtime-tested here; the new AppContainer unit test pins that refreshStatic skips the physical clear in VP mode.
• The runtime "after" build being the merge tree also means the PR composes correctly at runtime with the main UI changes that landed since its base (e.g. feat(daemon): merge daemon-mode feature batch into main #4490's AppContainer touches) — built, booted, and exercised without issues.
• Process nit, zero impact on the code: the PR body is pasted as raw HTML (with styling classes/SVG fragments), which makes the rendered description and any future gh pr view parsing noisier than the repo's usual markdown template. Worth a quick re-paste as plain markdown if you get the chance.

Verdict

✅ Verified on macOS — recommend merge (pending the Windows CI leg finishing). The mechanism is exactly as designed at the byte level (single trailing-edge full clear per gesture, return-to-width no-op, old per-event erase gone), and the user-visible #4891 symptom — reproduced on main in a real Terminal.app window — is eliminated on the merged tree. This closes the macOS verification gap called out in the PR's own test plan.

# builds
git worktree add /tmp/qwen4919/before origin/main --detach
git worktree add /tmp/qwen4919/after  <pr-head> --detach && (cd /tmp/qwen4919/after && git merge origin/main)
# npm install && npm run build in each

# byte level (inside tmux): python pty harness
#   - spawn `node packages/cli/dist/index.js` in a pty (TERM=xterm-256color)
#   - bursts: TIOCSWINSZ + SIGWINCH; widths 110/100/90/80 @60ms, then 120, then 90→120 @100ms
#   - log timestamped chunks; count ESC[3J vs ESC[1;1H+ESC[J per phase

# window level (real Terminal.app):
#   osascript 'do script' a window running:
#     node <build>/packages/cli/dist/index.js --approval-mode yolo -i \
#       'Print the numbers 1 to 100 ... then print DONE-MARKER. Do not use any tools.'
#   6 × `set bounds of window id W to {...}` @120ms while streaming
#   dump `history of tab 1 of window id W`; count banner copies, rule widths, duplicate lines