Problem
When Claude Code is actively processing a task, any message the user sends gets queued and processed sequentially after the current operation completes. This means if a user types something like "wait, don't do that", "stop, I changed my mind", or "hold on, let me clarify" — the agent continues executing the original task to completion, then reads the interruption after the fact. By that point, unwanted work may already be done (files written, commands run, etc.).
Expected Behavior
The agent should detect interrupt signals — messages that begin with phrases indicating the user wants to halt or redirect the current operation — and treat them differently from regular queued input.
When an interrupt signal is detected, the agent should:
- Pause the in-progress task as soon as possible
- Skim the new message to understand the user's intent
- Check in with the user before continuing — e.g., "I see you want me to stop. Should I abandon what I'm working on, or finish this step first?"
Interrupt Signal Examples
Messages starting with phrases like:
- "Wait" / "Wait,"
- "Stop" / "Stop,"
- "Hold on"
- "Actually, don't"
- "Cancel that"
- "Nevermind" / "Never mind"
- "Scratch that"
- "Hang on"
- "Pause"
Why This Matters
Users often realize mid-execution that they gave incomplete instructions, picked the wrong approach, or need to course-correct. The current queue-and-continue behavior forces them to either watch unwanted work happen or use the harder interrupt mechanism (Escape key). A soft interrupt via natural language would match how human collaboration actually works — you say "hold on" and your colleague pauses to listen.
Want me to adjust the tone, add anything, or post this somewhere specific (like a GitHub issue)?
Environment Info
- Platform: darwin
- Terminal: tmux
- Version: 2.1.62
- Feedback ID: 4104edb4-42b2-47f6-9242-ea91bc603ecb
Errors
n/a
Problem
When Claude Code is actively processing a task, any message the user sends gets queued and processed sequentially after the current operation completes. This means if a user types something like "wait, don't do that", "stop, I changed my mind", or "hold on, let me clarify" — the agent continues executing the original task to completion, then reads the interruption after the fact. By that point, unwanted work may already be done (files written, commands run, etc.).
Expected Behavior
The agent should detect interrupt signals — messages that begin with phrases indicating the user wants to halt or redirect the current operation — and treat them differently from regular queued input.
When an interrupt signal is detected, the agent should:
Interrupt Signal Examples
Messages starting with phrases like:
Why This Matters
Users often realize mid-execution that they gave incomplete instructions, picked the wrong approach, or need to course-correct. The current queue-and-continue behavior forces them to either watch unwanted work happen or use the harder interrupt mechanism (Escape key). A soft interrupt via natural language would match how human collaboration actually works — you say "hold on" and your colleague pauses to listen.
Want me to adjust the tone, add anything, or post this somewhere specific (like a GitHub issue)?
Environment Info
Errors
n/a