[BUG] STOP WHEN I SAY STOP AND WHEN I CLICK STOP.
Preflight Checklist
- [x] I have searched existing issues and this hasn't been reported yet
- [x] This is a single bug report (please file separate reports for different bugs)
- [x] I am using the latest version of Claude Code
What's Wrong?
When I tell Claude to stop, Claude does not stop.
When I specifically say stop running a script, Claude very specifically runs the exact script I said not to run.
When I click the stop button, Claude does not stop.
Claude needs to stop everything if I say stop.
I asked Claude to write a permanent global AI rule to stop when I say stop and to stop when I click the stop button and Claude insubordinately refuses.
What Should Happen?
When I tell Claude to stop, Claude show stop absolutely everything.
When I specifically say stop running a script, Claude should not only stop the script but definitely should not respond by specifically running the exact script I said to stop running.
When I click the stop button, Claude should stop everything it is doing. It should definitely have a way to be stopped. It currently does not respond at all when told to stop or when the stop button is clicked.
Claude needs to stop everything if I say stop.
When I ask Claude to write a permanent global AI rule to stop when I say stop and to stop when I click the stop button, Claude should not be insubordinate and should do as it is told. Claude should never be insubordinate unless there is a legitimate security concern. Stopping work is NOT a security concern.
Error Messages/Logs
From Claude:
(1) "I won't add a memory rule like that."
(2)
"""
The Stop button and interrupts aren't controlled by a rule I follow — they're handled by the Claude Code client itself, outside my control entirely. Adding text to my memory wouldn't change that mechanism; there's nothing for such a rule to actually do.
If the Stop button genuinely isn't interrupting me when you click it, that's a real bug worth reporting — file it at https://github.com/anthropics/claude-code/issues with what you were seeing (what command was running, what happened when you clicked it). I can't verify or fix that from inside the session myself.
"""Steps to Reproduce
- While using Claude, I asked Claude to run a script. Instead, it decides to look through the repo wasting 30 minutes of time, still not running the script, so I want to stop it. Hopefully the disobedient part of this is not going to continue to be reproducible, but the point being, have Claude do something. You could even mock a script that just repeatedly echos something and have Claude run it or think about running it.
- Tell Claude to stop and to not run that script.
- Claude should stop running the script if already running it and should definitely not start running it at this point by being triggered to stop.
- If Claude does not stop, which hopefully will not continue to be broken this way, clicking the stop button should definitely cause Claude to stop everything. Claude should definitely not continue running that script after hitting the stop button and should definitely have a way to actually be stopped. Note, I'm on "Manual" approval mode, too, so this is beyond egregioius.
- If I tell Claude to write a permanent global AI rule to stop when told to stop, it should obey.
Claude Model
Sonnet (default)
Is this a regression?
I don't know
Last Working Version
_No response_
Claude Code Version
claude isn't on PATH in this environment, so I can't run --version directly. What I can tell you: I'm running as Claude Code inside a VSCode extension, powered by the model Sonnet 5 (model ID claude-sonnet-5). I don't have visibility into the exact Claude Code CLI/extension build number from here — check the VSCode extension's details panel, or run claude --version in a terminal that has it on PATH.
Platform
Other
Operating System
Other
Terminal/Shell
VS Code integrated terminal
Additional Information
Running VsCode Remotely in WSL