claude-code - 💡(How to fix) Fix Feedback rating prompt collides with assistant's numbered questions, fires mid-process, and leaves no transcript record

Summary

The Claude Code satisfaction-rating prompt (the numeric "rate this turn" UI) can fire (a) while a tool/process is still active and (b) while the user is mid-answer to a numbered clarifying question the assistant has asked. Both inputs accept a numeric reply, so the user cannot tell which UI their number routes to, and the rating modal also interrupts active concentration during exactly the moments that matter (long-running tool calls, destructive-action confirmations, etc.).

Compounding the problem: when the rating prompt fires inline like this, neither the prompt nor the user's response to it appears in the conversation transcript. So the user cannot scroll back and reconstruct which UI their number was routed to, and Anthropic has no audit trail of the collision either.

Repro (qualitative — collision is timing-dependent)

1. Ask the assistant something that prompts a numbered clarifying question back, e.g. "Which of these three options do you mean — 1, 2, or 3?"
2. Before answering, the satisfaction-rating prompt fires.
3. Both UIs are now competing for a numeric reply. Typing a digit could be answering the assistant's question or rating the previous turn — the user can't tell.
4. After the turn completes, scroll back through the conversation history. The rating prompt is not recorded; neither is whichever input the number was routed to. The user has no way to audit what just happened.

A second flavor of the same problem: the rating prompt fires while a long-running Bash/Agent tool call is still streaming output. The user is trying to monitor the output and is interrupted by a prompt that wants a rating.

Expected behavior

The rating prompt should be suppressed while either of these is true:

• A tool call is in flight (Bash, Agent, etc. — anything that has emitted a "running" status and not yet completed).
• The assistant's most recent message contains an unanswered question (especially one with numbered choices, which directly collides with the rating UI input shape).

Session start and session end are fine — that's the natural place for a satisfaction prompt anyway.

Additionally, the rating prompt and the user's response to it should be captured in the conversation transcript so the interaction is auditable after the fact.

Why this needs a harness fix, not a prompt fix

The rating prompt is emitted by the Claude Code client, not by the model. Memory rules and prompt-level guidance can suppress the model from behaviors that would invite the prompt to fire, but they cannot prevent the harness from showing the prompt at an inopportune time, and they cannot make the harness write the interaction to the transcript. Both fixes have to live in the client.

Environment

• Claude Code on Windows 11 (PowerShell harness)
• Issue is harness-side, so likely reproduces across platforms / IDE extensions as well — the rating UI is a client-level feature