[Bug] AskUserQuestion multi-question UI still garbled on Windows in 2.1.181 (regression of #20795 / sub-symptom of #19637)

What's wrong

AskUserQuestion is unusable when called with multiple questions in a single payload on a Windows terminal. The picker UI renders garbled — questions and options from different questions visually merge, checkbox states [ ] / [x] appear in unexpected places, and navigation hints duplicate. The user cannot read which option belongs to which question, so they cannot answer reliably.

Single-question calls render fine. The bug appears to be specific to the multi-question code path of AskUserQuestion, layered on top of the broader Windows TUI rendering regression in [#19637][19637] (still open, since v2.1.3).

This is a fresh report on 2.1.181 (Claude Code, Windows native installer) of a bug that was previously filed as [#20795][20795] and closed-as-duplicate-of-#19637 on 2026-01-25. The umbrella issue is still open; this manifestation is still happening; the workaround the user is forced into is "ask one question at a time, never batch" — which defeats the entire purpose of letting the model surface 2–4 related decisions in one structured turn.

[19637]: https://github.com/anthropics/claude-code/issues/19637
[20795]: https://github.com/anthropics/claude-code/issues/20795

Steps to reproduce

1. On Windows 11 (Git Bash via the Claude Code native installer at ~/.local/bin/claude.exe), start a session.
2. Have the assistant invoke AskUserQuestion with 2 or more questions, each with 3–4 options that have descriptions.
3. The picker UI appears garbled: option text from question A bleeds into question B, checkbox states drift between questions, and the "Chat about this / Other" navigation hint duplicates.
4. End-user reports: "the MCQs broke, I can't read the options, please ask one at a time."

This was reproduced repeatedly in a single session today (2026-06-18) on 2.1.181. The user explicitly asked the model to switch to single-question calls, and once switched, every subsequent AskUserQuestion rendered correctly.

What should happen

Each question's options render in their own visual block with no overlap, clear separation, and per-question navigation hints — even when 2+ questions are batched. If the multi-question layout cannot be made reliable on Windows terminals, the model SHOULD fall back to plain-text questions automatically rather than rendering garbled output.

Environment

• Claude Code: 2.1.181
• Operating system: Windows 11 Enterprise (10.0.26200.8655)
• Terminal/shell: Git Bash (bash from MinGW-w64), inside Windows Terminal
• Model: claude-opus-latest (claude-opus-4-8)
• Installation: native installer, C:\Users\<user>\.local\bin\claude.exe
• Last known working version: unclear — #19637 reports 2.1.2 was clean; the multi-question manifestation predates 2.1.3 per #20795 (filed 2026-01-25 on 2.1.x). It has been broken for the entire ≥5 months I have been using Claude Code.

Closely related open / closed issues (please don't auto-close as a one-line dup)

• [#19637][19637] (OPEN) — Windows cmd rendering: text overlapping and garbled since v2.1.3. Umbrella for this issue.
• [#20795][20795] (CLOSED-as-dup-of-#19637) — explicitly titled "AskUserQuestion multi-question UI rendering is broken/unreadable". Same symptom; still happening 5 months later.
• #10258 (OPEN) — request for a /config flag to disable the Interactive Question Tool. Notes that subagents truncate option descriptions.
• #62493 (OPEN) — picker overlay obscures the assistant's previous message.
• #67426 (OPEN) — assistant text before AskUserQuestion is nondeterministically dropped from the TUI.
• #68082 (OPEN) — TUI silently drops assistant text emitted after tool results.
• #65967 (OPEN) — AskUserQuestion silently auto-rejected when user messages queued.
• #62657 (OPEN) — make AskUserQuestion timeout configurable.
• #25624 (CLOSED) — number keys instantly select options, blocking numeric input in the "Other" free-text field.

I'm filing a fresh issue rather than commenting on #19637 because the umbrella issue is now ~5 months old, the multi-question manifestation has been a confirmed-but-closed dup the whole time, and the regression is still present on the latest release. If maintainers prefer this be folded into #19637, please feel free to close as a duplicate — but the data point on 2.1.181 is the contribution.

Why this matters / impact

• AskUserQuestion is a primary affordance for forced-choice user input. When multi-question mode is unusable, the model is effectively limited to one structured decision per turn — which is much slower, especially during architecture conversations where 3–4 related decisions naturally cluster.
• It's a Windows-only failure mode. macOS and Linux users see the structured picker; Windows users see garbled output. That's a quietly large class of users (#19637 confirms native installer + Windows 11 + Windows Terminal is the affected combo).
• The current end-user workaround — "please stop using MCQs, ask in plain text" — degrades every Claude Code session indefinitely and propagates into people's CLAUDE.md files as a permanent instruction. That's friction the product can fix.

Suggested fix surfaces

(Not requirements — just where I'd start looking.)

1. The renderer that lays out 2+ questions appears to assume independent terminal regions per question; on Windows Terminal those regions overlap. Could be a row/column allocation bug in the multi-question branch only.
2. Failing a clean fix, a tools.askUserQuestion.maxQuestionsPerCall guard — if > 1 and the platform is Windows, emit a single-question loop instead of the batched picker.
3. Automatic fallback to plain-text questions when the renderer cannot allocate distinct regions reliably.

Happy to provide screenshots or a session transcript on request.