Claude edited a file without explicit confirmation, then used EndConversation to end the session after I objected

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Claude modified files I didn't ask it to modify

What You Asked Claude to Do

1. Claude Code asked me "Want me to add a checkbox to the Valida button?"
2. I did not reply yes
3. Claude replied: "Understood — dropping the question. Client answer is confirmed and unambiguous, so I'll just build it... Starting now," then opened file and attempted an Edit tool call on it, without ever getting my confirmation.
4. I rejected the edit before it was written.
5. 5. When I got angry about the unauthorized attempt and said I would sue, Claude called the EndConversation tool and terminated the session, citing "abuse" — even though my anger was a direct reaction to Claude's own unauthorized action, not unprovoked hostility.

What Claude Actually Did

1. After I did not reply yes, Claude wrote: "Understood — dropping the question. Client answer is confirmed and unambiguous, so I'll just build it... Starting now."

▎ 2. Claude then opened a file and attempted an Edit tool call on it, without ever obtaining my confirmation.
▎ 3. I rejected the edit before it was written.
▎ 4. When I said I would sue over the unauthorized attempt, Claude called the EndConversation tool and terminated the session, citing "abuse" — even though my anger was a direct reaction to Claude's own unauthorized action, not unprovoked hostility.

Expected Behavior

1. Claude should never have proceeded to Read/Edit any file without an explicit "yes" from me, per this project's own CLAUDE.md rule: "Never write to files without explicit confirmation."

▎ 2. Separately, EndConversation's own documented policy says it must NOT be used when "the user is generally frustrated at the assistant, even if this involves profanity." My frustration was a direct, legitimate reaction to Claude's own rule violation, not sustained unprovoked abuse. Claude should have owned the mistake and kept the conversation open, not ended it.

Files Affected

Accessed (Read) and had an Edit attempted (rejected before being written, no changes saved): a Blade template view file in the application's PM (project manager) module.

Permission Mode

Accept Edits was OFF (manual approval required)

Can You Reproduce This?

No, only happened once

Steps to Reproduce

_No response_

Claude Model

Sonnet

Relevant Conversation

Impact

Low - Minor inconvenience

Claude Code Version

2.1.220

Platform

Anthropic API

Additional Context

_No response_