The Agent Skills Directory

COMMAND_EXECUTION (LOW): The skill implements methods like DEEP-PROBE and CURSOR-PROBE that utilize the run_terminal_cmd tool to collect system and IDE diagnostics. Additionally, the configuration templates for sub-skills (e.g., Home Assistant) include shell snippets for setup tasks, such as fetching credentials via the 1Password CLI (op) and testing network connectivity with curl.
PROMPT_INJECTION (LOW): The skill generates dynamic agent instructions (.mdc rules) from templates using a process described as "empathic templating." This creates an indirect prompt injection surface:
Ingestion points: Untrusted data enters the context via .moollm/hot.yml, .moollm/startup.yml, and various CARD.yml files.
Boundary markers: Absent; templates do not use explicit delimiters or "ignore" instructions to isolate variable content from the resulting rules.
Capability inventory: The skill is granted run_terminal_cmd and write_file capabilities, which are used to compile and apply these rules.
Sanitization: No explicit sanitization or validation of template variables (e.g., {{probe_results}}) is performed before they are interpolated into system-level instructions.
DATA_EXFILTRATION (SAFE): While the skill manages sensitive API tokens and configurations, it mandates storing them in the gitignored .moollm/ local directory. It also utilizes a "trekify" protocol to mask personal identifiers and absolute paths in shared logs and example files, demonstrating a privacy-centric design for handling sensitive environmental data.

bootstrap