The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/openclaw_relay.py utilizes subprocess.run to execute local commands such as node and pnpm. This is the primary method for interacting with the local OpenClaw and ACPX environments.
[REMOTE_CODE_EXECUTION]: The skill implements a remote transport feature using ssh to execute commands on a remote host (defaulting to mac-studio). This capability is documented as a way to reach OpenClaw agents on different machines.
[CREDENTIALS_UNSAFE]: The skill is configured to access a sensitive authentication token at ~/.openclaw/gateway.token. This token is required by the underlying OpenClaw system to authorize the relay operations.
[PROMPT_INJECTION]: The skill exhibits an attack surface for indirect prompt injection as it relays user-provided text to external sessions.
Ingestion points: Untrusted data enters the agent context via the --message and --text arguments in scripts/openclaw_relay.py.
Boundary markers: The skill uses basic JSON formatting for results but lacks explicit instructions to downstream sessions to ignore embedded commands within the relayed text.
Capability inventory: The skill can execute shell commands locally and remotely via subprocess.run across all relay functions.
Sanitization: User input is sanitized using shlex.quote() before being interpolated into shell commands, which effectively prevents direct command injection on the host running the script.

openclaw-relay