[BUG] Transcript-share consent (post-rating) can be auto-accepted by a stray keystroke, uploading a full session with plaintext secrets

Summary

On v2.1.198, the post-rating transcript-share consent prompt ("Can Anthropic look at your session transcript to help improve Claude Code?" — Yes/No) can be accepted by a single stray keystroke intended for the message input box, silently uploading the entire session transcript to Anthropic.

Repro

1. Rate a session (the quality prompt).
2. The transcript-share Yes/No consent appears.
3. Start typing your next message. If it begins with "y", that keystroke is captured as "Yes".
4. The full session transcript (plus subagent transcripts and raw session log) uploads to Anthropic before you can react.

Why this is more serious than the existing key-collision reports

The already-closed issues (#53310, #65137, #63558) frame this as a general UX key-collision annoyance. This report is specifically about the transcript-share data-upload consent, which has a privacy consequence:

• A single intercepted keystroke consented to uploading the entire session, including plaintext secrets — in my case a machine login password appeared 100+ times in sudo -S command echoes, plus an auth (Bearer) token.
• The known-token redaction does not catch passwords typed/echoed in shell commands or non-standard tokens.
• There is no self-serve way to un-upload — deletion requires emailing privacy@anthropic.com.

So a UX focus-stealing bug, in this specific prompt, becomes an irreversible privacy exposure.

Requests

1. Data-consent prompts must not consume keystrokes intended for the input box — buffer/ignore input-box characters while the prompt is open, or require focus on the prompt.
2. Transcript-upload consent should require a deliberate confirmation (e.g. type "yes", or a two-step confirm) — not a single key that stray input can trigger.
3. Add an undo/cancel window for transcript uploads, or a self-serve way to revoke a just-submitted upload.

Environment

• Claude Code v2.1.198 (native binary, macOS, iTerm2)
• Plan: Max 20x

Related (all closed): #53310, #65137, #63558. Related feature request: #70622 (option to disable clickable Yes/No prompts).