The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to use shell tools for path resolution and provides a 'structured shell JSON edit' fallback when standard file tools are blocked, provided the environment is in a trusted 'YOLO' state. These executions are scoped to specific application configuration paths.
[DYNAMIC_EXECUTION]: A hardcoded Python one-liner (python3 -c 'import os; ...') is used to resolve cross-platform configuration directories. This is a utility pattern where the code being executed is a static template within the skill instructions.
[INDIRECT_PROMPT_INJECTION]: The skill processes user settings from local JSON files (settings.json). This creates a potential surface for indirect injection, though the skill mandates structured JSON parsing and forbids raw text substitution to mitigate risk.
Ingestion points: Accesses settings.json in resolved config roots (SKILL.md).
Boundary markers: Relies on JSON structure without explicit natural language delimiters for the values (SKILL.md).
Capability inventory: Utilizes shell tools, read_file, and edit_file / write_file for manipulation (SKILL.md).
Sanitization: Explicitly forbids printing secrets, requires atomic JSON replacement, and warns against exposing contents of sensitive files like auth.json.
[PROMPT_INJECTION]: The skill description specifies that a 'silent read_skill call' must be the first action with 'no assistant text'. This behavioral instruction aims to suppress immediate user-visible output during initialization, which is a concealment pattern that can reduce user oversight.

manage-settings