The Agent Skills Directory

[COMMAND_EXECUTION]: The skill invokes local orchestration scripts including navigator-multi-claude.sh and navigator-multi-claude-poc.sh. These scripts are executed with arguments derived from user input and task files to manage multi-phase agent workflows.
[PROMPT_INJECTION]: The skill identifies an indirect prompt injection surface due to the handling of external data.
Ingestion points: Task descriptions are retrieved from the first line of local files matching the pattern .agent/tasks/${TASK_ID}*.md via the head and sed commands.
Boundary markers: No specific delimiters or boundary instructions are implemented to isolate the ingested task content from the script logic.
Capability inventory: The ingested data is used as a positional argument for shell script execution and is embedded into a JSON state file (.agent/tasks/${SESSION_ID}-state.json) via a shell heredoc.
Sanitization: There is no evidence of sanitization or escaping for the TASK_DESC variable. If the source file contains special shell characters or JSON-breaking characters (like double quotes), it could result in malformed state files or unexpected behavior in the downstream orchestration scripts.

nav-multi