Claw Mission Killer
v0.1.0Interrupt a running agent task and rollback its session to the state before the last triggering user message. Use when an agent is stuck, running the wrong t...
⭐ 0· 146·0 current·0 all-time
bymaomao@dorianyoung7702
MIT-0
Download zip
LicenseMIT-0 · Free to use, modify, and redistribute. No attribution required.
Security Scan
OpenClaw
Benign
high confidencePurpose & Capability
Name/description (interrupt + rollback) match the included scripts. The bundle contains an installer, a wrapper to register PIDs (run.py), a marker utility (mark.py), the main interrupt/rollback logic (interrupt.py) and a watcher (watch.py) to auto-inject integration — all relevant to implementing an agent-kill-and-rollback feature.
Instruction Scope
The runtime instructions and scripts intentionally read and modify OpenClaw config and agent workspace files (openclaw.json, AGENTS.md), create marker files under ~/.openclaw/agents/<id>/running.json, scan system processes, kill PIDs, backup and edit session transcript files, and write interrupt-logs. Those actions are expected for this purpose, but they are broad: they modify other agents' AGENTS.md, touch transcripts and markers, and terminate OS processes — which can cause data loss or collateral kills if used incorrectly.
Install Mechanism
There is no external installer or network download; this is instruction + local Python scripts only (stdlib usage). No package pulls or remote fetches are present in the provided files.
Credentials
The skill requires no cloud credentials or external secrets and reads only local OpenClaw files and agent workspaces. That access is appropriate for a tool that needs to modify agent transcripts and markers. It does rely on the presence of an OpenClaw home (defaults to ~/.openclaw) and read/write access there.
Persistence & Privilege
always is false (not forced). The skill writes persistent artifacts: it injects blocks into AGENTS.md, writes running.json markers and interrupt-logs, and watch.py is intended to be run periodically (cron). That file-modifying behavior is consistent with the described functionality but may be surprising to users who don't expect automatic edits to AGENTS.md.
Assessment
This skill appears to implement exactly what it claims: local utilities to register long-running agent processes, kill them, and roll back the last user message. Before installing: 1) Backup ~/.openclaw and your agents' AGENTS.md and session transcripts so you can recover if something goes wrong. 2) Review openclaw.json to confirm agent workspaces are correct — the tool locates processes by workspace path and will append to AGENTS.md. 3) Prefer running install.py and watch.py in dry-run mode first, and test interrupt.py with --dry-run on a non-critical agent. 4) Beware collateral damage: process matching falls back to substring-matching the workspace in ps output and kill is aggressive (kill -9 / Stop-Process). If a workspace path is a common substring, unrelated processes could be matched. 5) mark.py lets any process register a PID for an agent; ensure only trusted agent code runs the marker to avoid accidental/unauthorized kills. 6) The tool is local-only (no network exfiltration seen), but it modifies files and can terminate OS processes — use in a controlled environment and verify backups and permissions first.Like a lobster shell, security has layers — review code before you run it.
latestvk9736nbhxvp892djfk3g76dc418339ga
License
MIT-0
Free to use, modify, and redistribute. No attribution required.