[BUG] /feedback submissions should allow for redaction/scrubbing/audit of PII/Secrets before submission

Found 3 possible duplicate issues:

1. https://github.com/anthropics/claude-code/issues/67719
2. https://github.com/anthropics/claude-code/issues/63734
3. https://github.com/anthropics/claude-code/issues/71486

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

Came over from #72393, which the dedupe bot flagged as a likely duplicate of this - fair call, the core mechanism is the same.

Strong +1 on forcing a local PII/secret scrub with a mandatory review step before anything leaves for Anthropic.

One extension worth folding in here: that same client-side sanitize-and-review gate would also enable an opt-in path to contribute curated data for training, not just to harden /feedback. Today model-training data sharing is a single all-or-nothing toggle, so privacy-conscious users keep it off entirely and none of their (often high-signal) data ever helps. If I could export a batch, see exactly what the scrubber caught, edit it, and only then opt in per-batch, I'd contribute - and the privacy burden stays at the edge, before data leaves the device.

Same building block (local scrub + mandatory review), usable for two things: feedback hygiene and deliberate, user-curated training contribution.

i am also interested in this newsletter

Absolutely! You nailed it. Providing a better opt-in/opt-in(redact/sanitize) path would be wonderful.

The UX of the current mid-session CTA (or user initiated /feedback) can interrupt your thought process when reviewing code. Having to stop in the middle of development to decide about submitting unsanitized feedback may have other developers (like me) just NOT contribute useful training data, because we're forced to interrupt our though process right now.

From a mid-session feedback CTA (or user initiated /feedback), it would be awesome to be able to:

• submit feedback now, as-is, or choose to defer submission (so you don't lose momemtum)
• optionally provide on-device sanition/redaction feedback process
• optionally initiate an user-review of data to be submitted
• optionally store a copy of submitted data

I'm sure other devs are running into this exact issue where they wish they could help teach the model some juicy nugget o' wisdom, but you just can't lose your train of thought, so you don't post. There's not a good way to defer and get back to the submission out of band, or when your brain isn't working on your current project.

Possible Dupes/Related: #71486 #72393 #67719

I'm sure other devs are running into this exact issue where they wish they could help teach the model some juicy nugget o' wisdom, but you just can't lose your train of thought, so you don't post. There's not a good way to defer and get back to the submission out of band, or when your brain isn't working on your current project.

yeeeepppp, the reason i'm here is i already started building a special pii-phi-etc datastore that is just for agent stuff, specifically so i can handle data without it being ingested into claude's context st all. it's frustrating that i feel like i need to do this and i wish there were a way to do this in the software itself (i mean, it's fine, i like writing software so i'll just write my own)

Writing a skill for your AI/Agent may assist with that where you could identify, isolate PII/PHI/secrets locally and replace them with one-time-use placeholders (never to be reused). Such that when a coud providers utilize the data, they'll handle the prompt/context correctly, without knowing the actuall PII/PHI/secrets value. For example, "Patient Cruz ,Jp, DOB 2026.06.29" is replaced by "Patient Lastname4lkjd84.dxu, Firstname4duu5uy8.3456kx, DOB date.recentdate.redacted.id.28589m3.i4" so when Claude/OpenAI/Z/Grok get it, the divulged data isn't easily associated with the patient. On subsequnt turns, new IDs are utilized, so cross-turn data inference is hard(er). Not perfect, yet, but the skill can be built in stages:

Stage 1) automatic PII/PHI/secrets redaction before submission for each turn. (relatively easy. can be done heuristically)
Stage 2) automatic PII/PHI/secrets replacement (for return inference, or agent handoff) so that the downstream LLMs can handle the request correctly but not actually know PII/PHI/secrets.

I have another use case that this "transparent" or "automatic PII/secrets handling can be really useful. I run local and coud models and would like to be able to automatedly be able to do hand-offs that either obfuscate/redact/dynamically-replace PII/Secrets so that the various models can interact without leaking PII/Secrets to cloud models. So you don't end up actually caring if an email is used, and the cloud models can use placeholders for PII/secrets and on the return trip, those data are still used correctly in context.

Agent Collaboration session1
(LLM A local) bob.lastname@secret.com -> (handoff to LLM B - cloud) x3.d38nf5@d8439.do3b85 -> (handoff back to LLM A local) bob.lastname@secret.com

Agent Collaboration session2
(LLM A local) bob@lastname@secret.com > (handoff to LLMs C and D - cloud) 3kfdy58.29kf4@94kf.d4i9 -> (handoff back from external LLMs to LLM A Local) bob.lastname@secret.com

so secrets/pii can be processed correctly contextually by cloud models as one-time-string IDs, uesed externally then returned back to the caller without the external system being knowledgeable about the obfuscated PII/secret.

In any case, I think you have an enhancement opportunity to ask for anthropic to proactively identifiy PII/PHI/secrets and have it redact/replace that data automatically before every turn. An option you can turn on, that after opting-in to local PII/PHI/secrets protection you can "just work," confident in the knowledge that your information isn't getting leaked. Right now, the burden is on the end user to be proactive about working with sensitive information and handling it appropriately. All it takes is just one tired accidental keypress, cut-n-paste, or mistype in the wrong window and you've now posted the keys to the kingdom. How many times have you been there? "Noooo, Claude! I didn't mean to paste that in there. Thank you for reminding me to rotate my secrets. I promise to be more careful!"

Intentional or not, sensitive information will leak and I think we can create tools to at least mitigate some of the fallout.