Assistant text before AskUserQuestion is nondeterministically dropped from the TUI (user sees only the picker)

Summary

Assistant text emitted in the same turn before an AskUserQuestion tool call is sometimes not displayed in the terminal — whole text chunks silently disappear from the conversation flow (not collapsed, not in scrollback at their position). The behavior is nondeterministic: structurally identical turns sometimes show the text and sometimes swallow it. This makes it impossible to reliably present decision context (e.g. a design comparison or a draft to approve) before asking a structured question — the user is left with a picker they cannot answer.

Observed with a Fable-class model; we did not reproduce-test other models, but the issue surfaced only after switching from Opus.

Environment

• Claude Code 2.1.173 (also observed on earlier 2.1.x)
• macOS (Darwin 25.5.0), terminal TUI
• Model: claude-fable-5

Reproduction

Have the model produce a turn shaped like: [text block] → [Bash tool call] → [text block] → [AskUserQuestion], e.g. by prompting:

Write 3 numbered lines of text (block A), then run date +%s with Bash, then write 3 more numbered lines (block B), then ask me via AskUserQuestion which blocks I can see.

We ran four controlled variations in one session and the user reported visibility while the question was on screen:

| # | Turn shape | Result |
|---|------------|--------|
| 1 | text A → Bash → text B → AskUserQuestion | A visible, B hidden |
| 2 | text C → AskUserQuestion | C hidden |
| 3 | text D → Bash → AskUserQuestion | D hidden |
| 4 | text E → Bash → one-line text → AskUserQuestion | E hidden, one-liner visible |

Note runs 1 and 4 have the same shape but opposite outcomes (first chunk visible in 1, hidden in 4; trailing chunk hidden in 1, visible in 4) — so this looks like a rendering race when the question widget opens, not a deterministic rule. In run 4, a screenshot shows the previous turn's content and the Bash call rendered normally, with block E simply absent between them — i.e. the chunk is dropped from the flow, not scrolled away.

Expected

All assistant text emitted before the AskUserQuestion call renders in the conversation and remains readable while the question is displayed, so the question can reference an explanation "above".

Actual

One or more pre-question text chunks are silently dropped, nondeterministically. Option preview fields are not a workaround for longer context since they fold after a few lines (… N lines hidden) and only the focused option's preview is shown.

Possibly related

• 2.1.152 changelog: "AskUserQuestion popup now hides the last line of preceding chat content less intrusively" — suggests the popup interacting with preceding content is a known area; what we observe goes beyond the last line and beyond visual overlap (entire chunks missing).

Impact

Any agent/skill workflow that explains a decision and then asks a structured question is unreliable: users see only the picker. We currently work around it by writing the decision context to a file and putting the file path into the question field (which always renders).