agent-interrupt

One command. Agent stops. Memory rolled back. Like it never happened.

When an agent goes off-track — running the wrong task, stuck in a loop, or executing something you didn't intend — agent-interrupt kills it cleanly and rewinds its memory to before it received that task.

Quick Start

After installing, run once to set up all your agents:

python -X utf8 scripts/install.py

Then just tell your assistant:

kill dev1
kill all

What It Does

1. Kills the process — Finds and terminates the agent's running subprocess (precise PID or workspace path fallback)
2. Verifies it's dead — Aborts if any process survives; transcript stays untouched
3. Backs up memory — Saves the removed messages to interrupt-logs/ for recovery
4. Rolls back transcript — Deletes the last user message (the task trigger) and everything after it
5. Agent wakes up clean — It enters standby with no memory of the interrupted task

Fuzzy Agent Matching

No need to remember exact agent IDs. It understands names:

kill dev1        → matches zero_dev1
kill pm          → matches zero_pm
kill 开发工程师   ��� matches by Chinese name
kill all         → interrupts every agent except main

Precise Kill (Recommended)

For agents running long scripts, use run.py wrapper so they can be killed precisely:

# Instead of: python your_script.py
python -X utf8 scripts/run.py --agent <your_agent_id> -- python your_script.py

install.py automatically adds this protocol to all your agents' AGENTS.md files. New agents added later are picked up automatically via a background watcher.

Recovery

Every rollback is backed up:

~/.openclaw/agents/<id>/interrupt-logs/rollback-YYYYMMDD-HHMMSS.jsonl

Append the backup lines back to the transcript to restore.

Scripts

Platform

Works on Windows, Linux, macOS. No external dependencies beyond Python stdlib.