[BUG] Fable 5.1: final answer emitted as a thinking block (not text) before AskUserQuestion, so the user never sees it
Environment: Claude Desktop 1.44121.4 (Windows 11), Claude Code 2.1.255 / 2.1.258 bundled, model claude-fable-5-1, effort high.
Symptom: When a turn ends with AskUserQuestion, the assistant's explanation above the form is frequently missing entirely. It does not appear after answering either.
Evidence from session transcripts (~/.claude/projects/**/*.jsonl): In the failing turns there is no text content block at all between the last user prompt and the AskUserQuestion tool_use. Instead the message contains a second thinking block of 250-560 characters whose content is clearly the user-facing answer in summarized form (e.g. "Fix pass complete except for the FOMO recategorization still running in the background: ..."), followed directly by the tool_use. Output token usage (~1,300) shows the full answer was generated; it was generated inside the thinking channel.
Frequency (turn-scoped, one user's transcripts, 2026-07-17 to 09-03): on days dominated by claude-fable-5 or claude-fable-5-1, 14-36% of AskUserQuestion turns have zero text blocks since the last user prompt (e.g. 08-09: 9 of 43 forms; 08-29: 4 of 11; 08-31: 7 of 23; 09-02: 7 of 22). On claude-opus-5 days the rate is 0-10% (08-27: 0 of 57). First observed 2026-07-20. Present since Fable 5, not specific to 5.1. No local settings changed across the window.
Not reproducible with thinking off: thinking cannot be disabled on Fable 5 / 5.1, so there is no user-side mitigation short of a PreToolUse hook that denies AskUserQuestion until a text block exists in the turn.
Expected: the answer that precedes a question is emitted as a text block; or the desktop Code tab renders the summarized thinking block when it is the only content before a question.
Related (different symptom, same user-facing effect): #67475, #64776.
Showing cached comments. Read the full discussion on GitHub ↗
3 Comments
Truly one of the bigger you actually issues that I have to tolerate. It makes me swear at the computer program
Same thing on macOS: Claude Desktop 1.46388.1, Claude Code 2.1.258, claude-fable-5-1.
In one session today, 16 of 34 AskUserQuestion calls had a thinking block as the last block before the tool_use, and the summarized thinking reads as the user-facing answer ("I've updated the rule at ... I've also listed out the eight open decisions currently on the page for reference"). Several of those carried the whole context the question depended on, so I answered blind or dismissed.
One pattern from the transcript, in case it helps whoever debugs this: text written at the start of a turn, right after my prompt, was always persisted as a text block. Text written after a tool result and followed by AskUserQuestion in the same turn went into thinking every time. Turn-ending text after a tool result was fine.
Same failure mode observed on
claude-fable-5, Claude Code 2.1.269 (Linux, subscription): a ~30-turn working session where every AskUserQuestion turn rendered as a bare question card with the preceding analysis invisible — the prose exists only as a signed thinking-block paraphrase, notextblock in the turn. Full evidence (message-shape histogram, cross-model control onclaude-opus-5showing zero occurrences, per-session all-or-nothing pattern, and a possibly relevantx-cc-atisassignment timing) posted at https://github.com/anthropics/claude-code/issues/74558#issuecomment-5664690841 — same underlying fingerprint as that issue.(Investigation AI-assisted: evidence gathered and verified with Claude Code on the affected machine.)