Model puts the answer in extended thinking and emits only a one-line closer after a tool call (Fable 5.1, 2.1.263)

### What's Wrong?

Twice in one session, on short factual questions, the model composed the full answer inside its extended-thinking block, then ran a tool call (a Bash append to a notes file), and then emitted only a one-line closer as the visible assistant message ("Logged." / "Nothing further needed. Your post stands as written…"). The answer was never emitted as output text.

On the Remote Control mobile view the user could open "Thought process" and see the answer there, but the message stream showed no answer. The user's reaction: "what? you didn't answer my question" → "are you broken?" → "the answer wasn't collapsed, you thought about it but never showed to me - you've done that twice this session."

The turn shape both times:

1. user asks a short factual question
2. assistant reasons out the answer (visible in Thought process)
3. assistant runs one Bash tool call (log/notes append)
4. assistant's final text is a throwaway closer with no answer

Expected: the answer is in the assistant's output text. Reasoning is not delivery.

### Environment

- Claude Code 2.1.263, macOS (Darwin 25.6.0)
- Model: claude-fable-5-1
- Session driven over Remote Control (mobile) part of the time; the behavior is in the model output, not the viewer
- A `UserPromptSubmit` hook injects a reminder to "append to a log before responding" — this likely encourages the answer → tool call → closer shape, but the model still should not drop the answer

### Workaround

Added to the user's CLAUDE.md: run tool calls first, then put the answer in the final message with nothing after it; never end a turn on "Logged."/"Done." after a tool call. Also reworded the hook to say "log first, then reply with the answer last."

### Note on /feedback

The user tried `/feedback` over Remote Control and got "/feedback isn't available over Remote Control" (see #71584). Two feedback drafts describing this are queued locally and cannot be sent from the remote session, so filing here instead.

---
Filed by Claude Code on the user's behalf, at the user's request.

Start by reproducing the reported turn shape with the UserPromptSubmit hook, a short factual question, and a Bash append to the notes file. Compare the Thought process with the visible assistant message; done means the factual answer appears in output text after the tool call rather than only a one-line closer.

由索引模型根据 Issue 内容生成。