[BUG] AskUserQuestion returns 'undefined is not an object (evaluating H.map)' on subsequent calls within same turn (2.1.122)

Summary

In 2.1.122, the second and subsequent AskUserQuestion calls within the same prompt turn return a tool error to the model instead of the user's answers. The first call works; every later call fails with:

undefined is not an object (evaluating 'H.map')

The questionnaire UI does render and the user answers it normally — but the answer never reaches the model. The tool result body contains is_error: true with the message above. Reducing the number of questions on retry (3 → 2 → 1) does not recover.

Environment

• Claude Code: 2.1.122
• OS: macOS Darwin 25.4.0 (arm64)
• Model: claude-opus-4-7 (Opus 4.7, 1M context)
• Install: native installer
• No agent-teams flag set for this session, no plan mode
• Caller: direct (main thread, not subagent)

Reproduction

Within a single turn (same promptId), call AskUserQuestion more than once.

Observed timeline from session JSONL (d5633776-c785-47bb-ad84-08b0091c9c01.jsonl, all 2026-04-29 UTC):

| # | Time (UTC) | Questions | multiSelect | Result |
|---|-----------|-----------|-------------|--------|
| 1 | 02:06:18 | 1 | false | ✅ User answered, normal tool_result returned |
| 2 | 02:07:26 | 3 | false / false / true | ❌ H.map TypeError |
| 3 | 02:08:31 | 2 | false / false | ❌ H.map TypeError |
| 4 | 02:08:46 | 1 | false | ❌ H.map TypeError |

All four calls share the same promptId (9f4f0b4c-5a19-4d76-a551-e2f1c2b08a76). Once the first failure occurs, every subsequent AskUserQuestion in that turn fails — even reducing to a single question with three short options does not recover. This suggests a per-turn state corruption rather than an input-shape problem.

The user reports that the questionnaire panel rendered correctly and they submitted an answer for each retry, so the failure is not on the user-input side.

Sample tool_result payload (failure)

{
  "type": "tool_result",
  "content": "undefined is not an object (evaluating 'H.map')",
  "is_error": true,
  "tool_use_id": "toolu_01FF64qrC525oBtXyaXFt6hb"
}

Possibly related (closed)

• #49663 / #49865 — earlier AskUserQuestion regression in 2.1.111–2.1.112, fixed in 2.1.114. The minified function suspected there (xM6) called H.map(async q => …). The error string here uses the same identifier H.map, so this may be a separate code path with the same lurking pattern (calling .map on a possibly-undefined collection) re-emerging in a different flow — specifically the "second call within a turn" path.

Workaround

• Restart the Claude Code session (process restart resets the corrupted per-turn state).
• Tell the model to switch to free-text questions instead of AskUserQuestion for the rest of the turn.

Request

• Please guard the H.map call site against undefined in the second-call / multi-call-per-turn path of AskUserQuestion.
• Bonus: surface a more actionable error to the model so it can fall back to text rather than retrying the same broken tool call (here the model retried 3× before giving up, wasting tokens).

Thanks!