Numbered feedback prompts conflict with numbered user responses

[pentafive]

What's Wrong?

The feedback rating system prompts for numeric input (0-5 scale). When users respond to Claude's questions using numbered lists (e.g., "1. answer 2. question 3. task"), the input parser interprets leading numbers as feedback ratings instead of message content. This causes:

• Incorrect feedback being submitted
• User prompts being mangled or truncated
• Loss of intended message content

What Should Happen?

Numbered list responses should not be intercepted as feedback ratings. The feedback system should require explicit confirmation or use non-numeric input to avoid collision with common response formats.

Steps to Reproduce

1. Engage in a conversation where Claude asks a multi-part question
2. Wait for or trigger a feedback prompt (0-5 rating request)
3. Respond with a numbered list: "1. First answer 2. Second point 3. Additional context"
4. Observe that the leading number is captured as feedback rating
5. Notice the remaining text is either truncated or mishandled

Error Messages/Logs

No explicit error - the feedback is silently submitted with the wrong value and the user's actual response is corrupted.

Session Reference

Session ID with feedback activity: 14209255-d94a-4caf-98b8-f03350ca7daa (Available for Anthropic team to review if needed)

Claude Model

Opus 4.5

Is this a regression?

Yes - this behavior has been observed in previous versions but the exact version is unknown.

Claude Code Version

2.0.61 (Claude Code)

Platform

Anthropic Direct

Operating System

Ubuntu 22.04.5 LTS (GNU/Linux 6.6.87.2-microsoft-standard-WSL2 x86_64)

Terminal/Shell

Windows Terminal

Additional Information

This is a UX/input parsing issue where the feedback collection mechanism collides with common user input patterns. Numbered lists are a natural way to respond to multi-part questions, making this collision likely to occur frequently.

[github-actions[bot]]

Found 1 possible duplicate issue:

1. https://github.com/anthropics/claude-code/issues/10897

This issue will be automatically closed as a duplicate in 3 days.

• If your issue is a duplicate, please close it and 👍 the existing issue instead
• To prevent auto-closure, add a comment or 👎 this comment

🤖 Generated with Claude Code

[pentafive]

This is related but not a duplicate. #10897 and this issue describe the same root cause (survey/feedback prompts intercepting numbered input), but this issue provides:

• Session ID reference (14209255-d94a-4caf-98b8-f03350ca7daa) for debugging
• Specific repro steps with numbered list format collision
• Regression confirmation with version history
• Active tracking - I'm maintaining this in my local issue tracker and will provide follow-up testing

Related issues worth consolidating:

• #10897 - Same bug, different description (survey vs numbered lists)
• #8743 - Related UX (focus stealing during typing)
• #7944 - Different bug (option ordering inconsistency)

I suggest keeping this issue as the canonical tracker due to the debugging evidence provided. Happy to test fixes when available.

---

👎 Downvoting auto-close - this is not a duplicate, it's better documented. See related issues above.

[pentafive]

Live Reproduction Evidence (2025-12-08)

Environment: Claude Code v2.0.61, Ubuntu 22.04.5 LTS (WSL2), Windows Terminal

Scenario captured:

Feedback prompt appeared while composing a numbered response:

How did that compaction go? (optional)
1: Bad    2: Fine   3: Good   0: Dismiss

User was typing response to Claude's question:

2. build in progress 5 any other bug/issues we should track...

The leading 2 would be captured as "Fine" rating, corrupting the intended message.

Screenshot attached below - shows the exact moment of collision with both prompts visible simultaneously.

[pentafive]

Screenshot added showing the exact collision:

• Top: User input starting with 2. (numbered response to Claude)
• Bottom: Feedback prompt 1: Bad 2: Fine 3: Good 0: Dismiss
• The 2 at the start of the user's message would be captured as a "Fine" rating

This demonstrates the input parsing conflict in real-time use.

[pentafive]

Possible Fix Observed (2025-12-21)

Environment: Claude Code v2.0.75, Ubuntu 22.04.5 LTS (WSL2)

Tested the exact collision scenario today. With feedback prompt visible:

How is Claude doing this session? (optional)
1: Bad    2: Fine   3: Good   0: Dismiss

Entered a numbered response:

1. okay it's been commented out
2. testing if claude will mangle my reply since it's asking how it's doing in the session.

Result: Message was accepted as normal input — Claude processed the full text instead of capturing 1 as a "Bad" rating. Feedback prompt also appeared again after Claude's reply, confirming the input wasn't consumed as a rating.

This appears to be fixed. Will continue testing and confirm with additional observations before closing.