The Agent Skills Directory

[COMMAND_EXECUTION]: The skill interacts with the system using the scripts/mo-watchdog.sh script, which executes the orca CLI tool to fetch session data and send notifications.
Evidence: The script invokes commands like orca orchestration worker-show and orca terminal list. These operations use quoted variables and validated patterns to prevent command injection.
[INDIRECT_PROMPT_INJECTION]: The skill ingests data from external session outputs which could potentially contain malicious instructions intended to influence the agent's behavior.
Ingestion points: The skill captures JSON output from orca CLI commands within scripts/mo-watchdog.sh.
Boundary markers: The script uses jq to enforce strict schema validation and is instructed to fail closed if the native data is malformed or unexpected.
Capability inventory: The skill has the ability to send text messages (nudges) to specific backend sessions via the orca send command.
Sanitization: It uses jq to filter out non-semantic fields (such as terminal previews and titles) and maps the remaining machine-readable status data to a predefined set of classification enums, effectively preventing data-driven control flow manipulation.
[PERSISTENCE]: The skill maintains a local state directory to track sent messages and avoid duplicate notifications.
Evidence: Local state is managed via the state_file() function in scripts/mo-watchdog.sh, storing digests of prior nudges.
Safety: The script ensures that these state files are protected by setting restrictive access permissions (umask 077 and chmod 600), ensuring they are only accessible by the current user.

mo-watchdog