[FEATURE] Sensitive data redaction in transcripts shared for debugging

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

I often run into unexpected model behavior / issues with Claude Code that I want to report. But my session transcripts almost always contain sensitive data eg. file paths, company names, personal information, document content that I can't share as-is.

Problem: Currently there is no way to redact sensitive content before sharing, so usually I just don't share the transcript. Not having access to the transcript of a Claude Code session where an issue arose likely makes it much harder for Anthropic to reproduce the issue, diagnose, resolve, and track the impact across the user base.

<img width="618" height="150" alt="Image" src="https://github.com/user-attachments/assets/f7fc112c-5aea-4089-abdb-7a5d4542b85a" />

Proposed Solution

When Claude Code prompts the user to share a session with Anthropic, there should be an option to share a redacted version instead. This could also be exposed as an
/export-redacted command (or --redact flag on /export) for manual use.

Redaction should happen locally, and the user should be able to review the redacted transcript before anything is sent to Anthropic.

Alternative Solutions

This issue is blocking me from sharing transcripts. As a result, I almost never share transcripts.

Priority

Medium - Would be very helpful

Feature Category

Interactive mode (TUI)

Use Case Example

Context: I'm using Claude Code to build automation tools to help people with their job search. In order to test the custom skills I built, my sessions include info like: list of companies applied to and status, location and salary preferences, and personal documents.

1. I'm using Claude Code.
2. Claude Code behaves unexpectedly (e.g. fails to use built-in functionality for PDF reading, does unnecessary work like try to run homebrew to install packages it doesn't need, tells me it can't do something I've had the same model has do in other sessions, hallucinates content, ignores instructions, makes a wrong decision).
3. I want to report it. The feedback popup appears and asks if I want to share the session transcript.
4. Instead of declining (because the raw transcript is sensitive), I choose "Share redacted version."
5. Claude Code redacts named entities, file contents, and credentials locally and shows me a preview.
6. I confirm the redacted version looks safe and it's sent to Anthropic with the bug report.

Additional Context

_No response_