The Agent Skills Directory

PROMPT_INJECTION (LOW): The script scripts/init_plan.py creates an indirect prompt injection surface by interpolating unsanitized user input (plan_id) into markdown templates. If the agent later reads these generated files, it could be influenced by malicious instructions embedded in the input. * Ingestion points: scripts/init_plan.py takes input via the plan_id argument. * Boundary markers: Absent. Templates in the assets/ directory lack delimiters or warnings to ignore embedded instructions. * Capability inventory: The skill facilitates file creation and project organization. * Sanitization: None; the script uses simple string replacement.
DATA_EXFILTRATION (LOW): The plan_id argument is used directly in file path construction without sanitization, allowing for potential path traversal. Evidence: In scripts/init_plan.py, plan_dir = base_dir / str(plan_id) (line 60) and output_path = plan_dir / filename (line 116) allow relative path segments to manipulate output locations.

implementation-planner