Text in the same assistant message as an AskUserQuestion call is never displayed or persisted

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

When the assistant produces a text block in the same assistant message as an AskUserQuestion tool call, that text is never displayed in the terminal, and it also never gets written to the session transcript jsonl under ~/.claude/projects/. The model still has it in the live conversation, so it keeps answering as if you read it.

How I confirmed this: I saved a full copy of my terminal scrollback and diffed it against the transcript jsonl. The transcript stores one row per content block, with rows of the same assistant message sharing a message id (thinking rows, tool_use rows, text rows). Text the assistant wrote before ordinary tool calls (Bash, Read) shows up both on screen and as text block rows. The multi-paragraph answers it attached to AskUserQuestion calls appear in neither: no text on screen, no text row anywhere in the jsonl. The only trace of them is that the model's later messages quote them verbatim, which is also how we caught it.

In practice this went bad fast: the model wrote a long analysis, asked a question about it in the same turn, and treated my answer as informed by an analysis I never saw. Took several rounds to figure out I was answering blind.

Since the transcript is what survives the session, resume and compaction presumably lose that text for the model too, not just for the user.

What Should Happen?

The text block should render before the question widget, the same way text renders before any other tool call, and it should be persisted to the transcript like any other text block.

Error Messages/Logs

None. The text is silently absent from both the terminal and the transcript jsonl.

Steps to Reproduce

1. In a terminal session, get Claude to produce a long text answer and an AskUserQuestion call in one turn. A prompt like "Explain your reasoning in detail, then ask me a multiple-choice question about it in the same turn" does it.
2. Answer the question, then ask the model to quote the explanation it just gave.
3. It quotes text that appeared nowhere on screen. Grep the session's transcript jsonl in ~/.claude/projects/<project>/<session-id>.jsonl for a distinctive phrase from the quote: no text block row contains it.

Claude Model

Claude Fable 5

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.1.201

Platform

macOS

Operating System

macOS 26.2

Terminal/Shell

VS Code integrated terminal / zsh

Additional Information

Related but different: #74237 is about the VS Code plugin's question modal covering the conversation. Here the text is never rendered at all in the plain terminal UI, and never persisted.