What Is Claude Code's SendFeedback Tool?
SendFeedback is a tool Claude Code can invoke itself, mid-session, to draft a feedback report about what just went wrong and put it in front of you for review. It requires Claude Code v2.1.238 or later. It fires when a command or tool fails repeatedly, when Claude can't complete a request after sustained attempts, when you tell it directly that it made a mistake, or when you explicitly ask it to file a report. In every case, invoking the tool only builds a draft. It doesn't send anything on its own.
This is a genuinely different mechanism from typing /feedback or /bug yourself. Those two commands still exist and still open the standard feedback window directly, bypassing the auto-draft entirely. SendFeedback is the new piece: Claude noticing something went wrong and writing the first draft of the report for you, instead of you reconstructing what happened from memory after the session's already moved on.
What Actually Gets Sent (and What Doesn't)
When SendFeedback fires, the session pauses and an interactive review form appears. You get four real choices: send the draft as written, edit the report before sending, toggle whether the current transcript is attached, or cancel and send nothing at all. The draft itself is built from context already available in the session, not from a fresh scan of your whole machine.
The three paths into a feedback report
| Path | What triggers it | What you see |
|---|---|---|
| SendFeedback (automatic) | Repeated tool failure, a stuck request, or you saying it made an error | A pre-filled draft report and a review form, before anything is sent |
| /feedback with text | You type it yourself | The standard feedback window, opened directly, no draft step |
| /bug | You type it yourself | The same standard feedback window, reporting through the same path |
Why Transcript Sharing Starts Off
Transcript sharing defaults to off on every SendFeedback draft. You have to actively turn it on for that submission, and Claude Code remembers your last choice going forward, though the setting can still be changed per submission. That default exists because a session transcript is everything that happened in the conversation, including file contents Claude read or wrote along the way.
The Off-by-Default Check We Use on Client Work
Inside the club we call this the Off-by-Default Check, and it's three questions, not a policy document. Run it in the ten seconds you already have while the review form is sitting on screen.
- Glance at what's actually in view before you toggle the transcript on: client API keys, real customer data, or anything under an NDA you signed changes the answer.
- Default to sending the written draft alone. It's usually enough for a real product issue to get acted on, and it carries none of the session's raw content.
- If the bug genuinely needs the full transcript to reproduce, say so, attach it deliberately, and treat that choice the same way you'd treat pasting a client's code into any third-party tool.
This isn't about distrust of the feature. It's the same instinct as reviewing a diff before you ship it, covered in [how I review Claude Code's output before I ship it](/blog/review-claude-code-output-before-you-ship): a review step only works if you actually use it instead of clicking through on autopilot.
When It's Actually Worth Sending
Send it when something genuinely broke and kept breaking, not every time Claude Code gets a little slow or a prompt needs rewording. A tool that fails the same way three times in a row, or a request that dies after real effort, is exactly the kind of thing a feedback report is meant to catch, and reporting it is how a rough edge actually gets fixed instead of just worked around by everyone who hits it.
Where This Works (and Where It Doesn't)
SendFeedback works in the interactive Claude Code app, the terminal, and Claude Code on the web, anywhere a human is actually present to see the review form and make a choice. It is not available through the Claude Agent SDK, which makes sense: a programmatic session running unattended has no one there to click send, so there's no draft-and-review step to offer.
Short, practical drops on skills, MCP, agents, prompts, and more. No spam, unsubscribe anytime.
Frequently asked questions
What is SendFeedback in Claude Code?
SendFeedback is a tool Claude Code can call on its own when a command keeps failing, a request can't be completed, or you say it made a mistake. It drafts a feedback report and opens a review form. It requires Claude Code v2.1.238 or later.
Does Claude Code send my code automatically when something fails?
No. SendFeedback only builds a draft and opens a review form. Nothing is sent until you choose to send it, and you can edit the draft or cancel it entirely before that happens.
Is transcript sharing turned on by default with SendFeedback?
No, it starts off by default on every draft. You have to actively enable it for that submission. Claude Code remembers your last choice for future reports, but you can change it each time.
How long does Anthropic keep a submitted transcript?
Transcripts shared through SendFeedback, /feedback, or /bug are retained for 5 years, since all three report through the same submission path.
What's the difference between SendFeedback and typing /feedback myself?
/feedback with text, and /bug, open the standard feedback window directly with no draft step. SendFeedback is Claude noticing a problem mid-session and writing the first draft of that report for you to review, rather than you typing one from a cold start after the fact.
Does SendFeedback work with the Claude Agent SDK?
No. It's available in the interactive Claude Code app, the terminal, and Claude Code on the web, but not through the Agent SDK, since an unattended programmatic session has no one present to review the draft before it's sent.
Last reviewed by Duncan Rogoff on September 2, 2026
