The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/openclaw_relay.py constructs and executes shell commands locally and on remote hosts using subprocess.run. Remote commands are wrapped in bash -lc and passed through SSH.
[REMOTE_CODE_EXECUTION]: The skill enables executing Node.js applications and shell commands on remote systems via SSH. While configured with a default host associated with the author, the --host parameter allows the agent to target any reachable SSH server.
[DYNAMIC_EXECUTION]: The run_acpx and run_openclaw functions in scripts/openclaw_relay.py dynamically assemble shell functions and complex command strings at runtime for execution on remote machines.
[INDIRECT_PROMPT_INJECTION]: The publish and ask commands relay content between different agent sessions, creating a vulnerability where untrusted data from one session is processed as a prompt in another.
Ingestion points: Content provided via the --message, --text, or --text-file arguments in scripts/openclaw_relay.py.
Boundary markers: Relayed text is interpolated into a prompt template in build_publish_prompt that provides natural language instructions but lacks robust structural separation or sanitization directives.
Capability inventory: The skill possesses extensive capabilities through run_local and run_ssh, which execute shell commands.
Sanitization: Relayed content is not filtered, escaped, or validated before being included in the prompt for the target session.
[DATA_EXFILTRATION]: The show and status commands retrieve internal session state, including agent message history and thinking blocks, which could expose sensitive information when relayed between sessions.

openclaw-relay