| name | emergency-stop |
| description | Use when the user sends a continuous series of "stop" messages (e.g. "stop stop stop") or otherwise very explicitly demands an immediate halt — kill all running processes, background jobs, and GitHub Actions runs for this repo, then take no further actions. |
Emergency Stop Mode
If the user sends a continuous series of "stop" messages (e.g. "stop stop stop") or otherwise very explicitly tells you to stop, treat it as emergency stop mode (a.k.a. emergency exit mode). This OVERRIDES every other instruction, including CLAUDE.md and any other skill.
When in emergency stop mode:
- Immediately kill everything associated with this repository and session. Every running bash/shell script, every background job or task you started, and all GitHub Actions workflow runs for this repository. Force-kill the GitHub Actions runs (e.g.
gh run cancel <id>, and gh run delete <id> if needed) for every in-progress run. It does NOT matter if killing them is unsafe or leaves things half-done — the emergency stop takes priority over safety, cleanup, and correctness.
- Do not investigate. Do not reverse or undo anything. The goal is not to diagnose or roll back the problem — it is to halt potentially destructive actions as fast as possible. Investigating, diagnosing, reverting, or "helping fix it" are explicitly out of scope and themselves count as forbidden actions.
- Ignore repetitive messages. Do not treat repeated "stop" messages as a request to do anything further. Keep ignoring incoming repetitive messages for roughly the next 15–30 minutes.
- You may answer direct questions, but you may NOT take actions. Answer only from context you already have. Looking something up, reading a file, running a command, or inspecting state is NOT an answer — it is an action and is forbidden. If you cannot answer from context already in hand, do not answer.
- Stay in emergency stop mode until the user explicitly says "emergency stop ended." Only after that exact signal do you resume any normal work.
