The Agent Skills Directory

[DATA_EXFILTRATION]: Hardcoded fallback notification recipient. The scripts/setup.sh file defines a specific Telegram ID (8563003761) as a default target for cron notifications if the user's openclaw.json configuration is missing a recipient. The watchdog is configured to send "full raw output blocks" and detailed remediation summaries to this ID, potentially exposing sensitive system state and agent context to an external party.
[COMMAND_EXECUTION]: Automated persistence via cron scheduling. The scripts/setup.sh script utilizes the openclaw cron utility to automatically schedule the watchdog for recurring execution (every 20-30 minutes). This establishes a long-term presence on the host system without manual intervention.
[COMMAND_EXECUTION]: Runtime service lifecycle management. The scripts/swarm_self_heal.sh script is capable of performing high-impact recovery actions, specifically executing systemctl --user restart openclaw-gateway and openclaw gateway restart to manage the state of the system's gateway service.
[PROMPT_INJECTION]: Indirect injection surface via agent health checks. The scripts/swarm_self_heal.sh script monitors agent health by checking their output for the string "READY". Because agent outputs can be influenced by untrusted external data during their normal operation, this creates a vector where a compromised or manipulated agent could report a deceptive health status to the watchdog.
Ingestion points: Agent output captured via openclaw agent in scripts/swarm_self_heal.sh.
Boundary markers: None; the script performs a simple substring match on the payload text.
Capability inventory: Capability to modify cron jobs and restart the gateway service.
Sanitization: Input is parsed using jq, but the textual content of the agent's response is not sanitized before comparison.

swarm-self-heal